Mathematical Modeling

in Biology
Mathematical Modeling in Biology: A Research Methods Approach is
a textbook written primarily for advanced mathematics and science
undergraduate students and graduate-level biology students. Although
the applications center on ecology, the expertise of the authors,
the methodology can be imported to any other science, including
social science and economics. The aim of this book, beyond being a
useful aid to teaching and learning the core modeling skills needed
for mathematical biology, is to encourage students to think deeply
and clearly about the meaning of mathematics in science and to
learn significant research methods. Most importantly, it is hoped that
students will experience some of the excitement of doing research.

Features

• Minimal prerequisties are a solid background in calculus, such as

a Calculus I course

• Suitable for upper division mathematics and science students and

graduate-level biology students

• Provides sample MATLAB codes and instruction in appendices

along with data sets available on https://bit.ly/3fcLF3D.
 Chapman & Hall/CRC Mathematical
Biology Series
Series Editors:
Ruth Baker, Mark Broom, Adam Kleczkowski, Doron Levy,
Sergei Petrovskiy
About the Series
This series aims to capture new developments in mathematical
biology, as well as high-quality work summarizing or contributing to
more established topics. Publishing a broad range of textbooks,
reference works, and handbooks, this series is designed to appeal to
students, researchers, and professionals in mathematical biology.
We will consider proposals on all topics and applications within the
field, including but not limited to stochastic modeling, differential
equation modeling, dynamical systems, game theory, machine
learning, data science, evolutionary biology, cell biology, oncology,
epidemiology, and ecology.
Systems Biology
Mathematical Modeling and Model Analysis
Andreas Kremling
Cellular Potts Models
Multiscale Extensions and Biological Applications
Marco Scianna, Luigi Preziosi
Quantitative Biology
From Molecular to Cellular Systems
Edited By Michael E. Wall
Game-Theoretical Models in Biology, Second Edition
Mark Broom, Jan Rychtáv̌
Mathematical Modeling in Biology
A Research Methods Approach
Shandelle M. Henson, James L. Hayward

For more information about the series, visit:

https://www.routledge.com/Chapman–HallCRC-Mathematical-
Biology-Series/book-series/CRCMBS
 Mathematical Modeling
in Biology
A Research Methods Approach

Shandelle M. Henson and James L. Hayward

MATLAB® is a trademark of The MathWorks, Inc. and is used with permission. The MathWorks
does not warrant the accuracy of the text or exercises in this book. This book’s use or discussion of
MATLAB® software or related products does not constitute endorsement or sponsorship by The
MathWorks of a particular pedagogical approach or particular use of the MATLAB® software.

Cover photo © by James L. Hayward

First edition published 2023

by CRC Press
4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN

and by CRC Press

6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742

© 2023 Shandelle M. Henson and James L. Hayward

CRC Press is an imprint of Taylor & Francis Group LLC

The right of Shandelle M. Henson and James L. Hayward to be identified as authors of this work
has been asserted in accordance with sections 77 and 78 of the Copyright, Designs and Patents
Act 1988.

All rights reserved. No part of this book may be reprinted or reproduced or utilised in any form
or by any electronic, mechanical, or other means, now known or hereafter invented, including
photocopying and recording, or in any information storage or retrieval system, without permission
in writing from the publishers.

For permission to photocopy or use material electronically from this work, access www.copyright.
com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers,
MA 01923, 978-750-8400. For works that are not available on CCC please contact
[email protected]

Trademark notice: Product or corporate names may be trademarks or registered trademarks,

and are used only for identification and explanation without intent to infringe.

British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library

Library of Congress Cataloging-in-Publication Data

Names: Henson, Shandelle Marie, 1964-author. | Hayward, James L., 1948-author.
Title: Mathematical modeling in biology : a research methods approach /
Shandelle M. Henson, Andrews University, USA, James L. Hayward, Andrews
University, USA.
Description: First edition. | Boca Raton : C&H/CRC Press, 2023. |
Series: Chapman and Hall/CRC mathematical biology series |
Includes bibliographical references and index.
Identifiers: LCCN 2022030360 (print) | LCCN 2022030361 (ebook) |
ISBN 9781032208213 (hardback) | ISBN 9781032206943 (paperback) |
ISBN 9781003265382 (ebook)
Subjects: LCSH: Biology–Mathematical models.
Classification: LCC QH323.5 .H46 2023 (print) | LCC QH323.5 (ebook) |
DDC 570.1/5118–dc23/eng/20220728
LC record available at https://lccn.loc.gov/2022030360
LC ebook record available at https://lccn.loc.gov/2022030361

ISBN: 9781032208213 (hbk)

ISBN: 9781032206943 (pbk)
ISBN: 9781003265382 (ebk)

DOI: 10.1201/9781003265382

Typeset in Minion
by codeMantra
 To our parents
Audrey Gackenheimer Henson and
John William Henson III
Jane Watson Hayward and James Lloyd Hayward, Sr.
Contents

ACKNOWLEDGMENTS xvii
FOR PROFESSORS AND STUDENTS xix
AUTHORS xxi

Section I Introduction to Modeling

Chapter 1 ■ Mathematical Modeling 3

1.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 3

1.2 THE MODELING CYCLE 4
1.2.1 Step 1: Translation into Mathematics 5
1.2.1.1 Choosing Variables and Parameters 5
1.2.1.2 Simplifying Assumptions 6
1.2.1.3 Parameterization 7
1.2.2 Step 2: Model Analysis 7
1.2.3 Step 3: Back-Translation 7
1.2.3.1 Model Selection and Validation 8
1.2.3.2 Test of Model Predictions 8
1.2.4 Step 4: Revising Model Assumptions 8
1.3 BIOLOGY 9
1.3.1 Ecology 10
1.4 MATHEMATICS 13
1.5 STATISTICS 16
1.6 EPISTEMOLOGY: HOW WE KNOW 19
1.7 EXERCISES 23
BIBLIOGRAPHY 28

vii
viii ■ Contents

Chapter 2 ■ Avian Bone Growth: A Case Study 29

2.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 29

2.2 SCIENTIFIC PROBLEM 30
2.2.1 Data 30
2.3 TRANSLATION INTO MATHEMATICS 32
2.3.1 Simplifying Assumptions 32
2.3.1.1 Deterministic Assumptions 33
2.3.1.2 Stochastic Assumptions 33
2.3.2 The Deterministic Model 33
2.3.3 The Stochastic Model 34
2.4 MODEL PARAMETERIZATION 36
2.4.1 Dividing the Data Set 36
2.4.2 Maximum Likelihood (ML) Method 38
2.4.3 Nonlinear Least Squares (LS) Method 39
2.4.4 Downhill Minimization Routine:
Nelder-Mead Algorithm 40
2.4.5 Implementing Parameterization in Code 41
2.4.6 Results of Parameterization 42
2.5 MODEL SELECTION 42
2.6 MODEL VALIDATION 45
2.7 EXERCISES 46
BIBLIOGRAPHY 51

Section II Discrete-Time Models

Chapter 3 ■ Discrete-Time Maps 55

3.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 55

3.2 COMPARTMENTAL MODELS 55
3.3 LINEAR MAPS 57
3.3.1 Malthusian Growth 57
3.4 NONLINEAR MAPS 60
 Contents ■ ix

3.5 LINEARIZATION 62
3.5.1 Linearization of Functions 62
3.5.2 Linearization of Discrete-Time Maps 64
3.5.3 Linearizing the Ricker Map 67
3.6 THE RICKER NONLINEARITY 70
3.7 EXERCISES 71
BIBLIOGRAPHY 74

Chapter 4 ■ Chaos: Simple Rules Can Generate

Complex Results 75

4.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 75

4.2 RICKER MODEL REVISITED 75
4.3 NEW PARADIGMS ARISE FROM CHAOS 80
4.3.1 Deterministic Unpredictability 80
4.3.2 Complex Results Can Arise from Simple Rules 80
4.4 MAY’S HYPOTHESIS 81
4.5 EXERCISES 81
BIBLIOGRAPHY 84

Chapter 5 ■ Higher-Dimensional Discrete-Time Models 85

5.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 85

5.2 INTRASPECIFIC INTERACTIONS 86
5.3 INTERSPECIFIC INTERACTIONS 86
5.4 EXAMPLE OF AN AGE-STRUCTURED
SINGLE-SPECIES MODEL 87
5.5 EXAMPLE OF A TWO-SPECIES MODEL 89
5.6 n-DIMENSIONAL LINEAR DIFFERENCE EQUATIONS 90
5.6.1 n-Dimensional Leslie Models 90
5.7 SOLVING LINEAR SYSTEMS OF DIFFERENCE
EQUATIONS 92
5.7.1 An Example 92
x ■ Contents

5.7.2 Solving the General Two-Dimensional System 96

5.7.3 Solving Higher-Dimensional Systems 97
5.8 NONLINEAR SYSTEMS 99
5.8.1 Linearization 100
5.8.2 An Example 102
5.9 EXERCISES 104
BIBLIOGRAPHY 109

Chapter 6 ■ Flour Beetle Dynamics: A Case Study 111

6.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 111

6.2 FLOUR BEETLES 111
6.3 DATA 113
6.4 ASSUMPTIONS 114
6.4.1 Deterministic Assumptions 114
6.4.2 Stochastic Assumptions 115
6.5 ALTERNATIVE DETERMINISTIC MODELS 116
6.6 STOCHASTIC MODELS 117
6.7 MODEL PARAMETERIZATION 119
6.7.1 Conditioned One-Step Residuals 119
6.7.2 Conditioned Least Squares (CLS) 120
6.8 MODEL SELECTION 122
6.9 MODEL VALIDATION 124
6.10 THE “HUNT FOR CHAOS” EXPERIMENTS 125
6.10.1 Results of the “Hunt for Chaos” Experiments 128
6.10.2 Manipulating the Parameters in the
Laboratory 128
6.11 EXERCISES 130
BIBLIOGRAPHY 136
 Contents ■ xi

Section III Continuous-Time Models

Chapter 7 ■ Introduction to Differential Equations 145

7.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 145

7.2 COMPARTMENTAL MODELS 146
7.2.1 A Tank Problem 146
7.2.2 The SIR Model 147
7.2.3 The Continuous-Time Logistic Model 150
7.3 EXERCISES 154
BIBLIOGRAPHY 156

Chapter 8 ■ Scalar Differential Equations 159

8.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 159

8.2 LINEAR EQUATIONS 159
8.2.1 Malthusian Growth 161
8.3 NONLINEAR EQUATIONS 162
8.3.1 Logistic Growth 162
8.3.2 Allee Effects 164
8.3.3 The “Doomsday Model” of Human
Population Growth 166
8.4 LINEARIZATION 170
8.5 BIFURCATIONS 175
8.5.1 Transcritical Bifurcation 176
8.5.2 Saddle-Node Bifurcation 177
8.5.3 Pitchfork Bifurcation 177
8.5.4 Hysteresis 178
8.6 EXERCISES 179
BIBLIOGRAPHY 185

Chapter 9 ■ Systems of Differential Equations 187

9.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 187

xii ■ Contents

9.2 LINEAR SYSTEMS OF ODES AND PHASE PLANE ANALYSIS 188

9.2.1 Unstable Node 189
9.2.2 Asymptotically Stable Node 191
9.2.3 Saddle 192
9.2.4 Center 193
9.2.5 Unstable Spiral 195
9.2.6 Asymptotically Stable Spiral 196
9.2.7 Summary: Eigenvalues Tell All 197
9.3 NONLINEAR SYSTEMS OF ODES 198
9.3.1 Linearization 201
9.4 LIMIT CYCLES, CYCLE CHAINS, AND BIFURCATIONS 202
9.5 LOTKA-VOLTERRA MODELS AND NULLCLINE ANALYSIS 204
9.5.1 Lotka-Volterra Competition 205
9.5.2 Lotka-Volterra Cooperation 205
9.5.3 Lotka-Volterra Predator-Prey 206
9.6 EXERCISES 208
BIBLIOGRAPHY 213

Chapter 10 ■ Seabird Behavior: A Case Study 215

10.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 215

10.2 THE SCIENTIFIC PROBLEM 215
10.3 HISTORICAL DATA 217
10.3.1 Count Data 217
10.3.2 Dividing the Data 217
10.3.3 Tide and Solar Elevation Data 217
10.4 GENERAL MODEL 218
10.5 ALTERNATIVE MODELS 220
10.6 MODEL PARAMETERIZATION 220
10.7 MODEL SELECTION 221
10.8 MODEL VALIDATION 222
10.9 TEST OF A PRIORI PREDICTIONS 222
10.10 STEADY-STATE MODEL 226
10.11 DISCUSSION 227
 Contents ■ xiii

10.11.1 Importance of Scale 227

10.11.2 Resource Management 228
10.12 EXERCISES 228
BIBLIOGRAPHY 230

Section IV Regression Models

Chapter 11 ■ Introduction to Regression 237

11.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 237

11.2 LINEAR REGRESSION 237
11.2.1 Simple Linear Regression (Single Factor) 238
11.2.2 Multiple Linear Regression (Multiple Factors) 238
11.2.3 Stochastic Model and Parameter Estimation 239
11.2.4 Confidence Intervals for Regression Coefficients 240
11.3 LOGISTIC REGRESSION 240
11.3.1 Odds Ratios (ORs) 241
11.3.2 OR Confidence Intervals 242
11.4 GENERALIZED LINEAR MODELS (GLMs) 242
11.5 INTERACTION TERMS 242
11.6 EXERCISES 243
BIBLIOGRAPHY 246

Chapter 12 ■ Climate Change and Seabird Cannibalism:

A Case Study 247

12.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER 247

12.2 THE SCIENTIFIC PROBLEM 248
12.3 DATA 251
12.4 LOGISTIC REGRESSION ANALYSIS 252
12.5 MODEL VALIDATION 253
12.6 OUTCOMES 253
12.7 CLIMATE CHANGE, CANNIBALISM, AND
REPRODUCTIVE SYNCHRONY 255
xiv ■ Contents

12.8 EXERCISES 257

BIBLIOGRAPHY 259

Section V Appendix

Appendix A ■ Linear Algebra Basics 265

A.1 MATRIX OPERATIONS 265

A.1.1 Matrix Addition 265
A.1.2 Scalar Multiplication 265
A.1.3 Matrix Subtraction 266
A.1.4 Matrix Multiplication 266
A.1.5 Determinants of Square Matrices 267
A.2 EXERCISES 267
A.3 SOLUTIONS 270
A.4 SUMMARY OF LINEAR ALGEBRA CONCEPTS 272

Appendix B ■ MATLAB: The Basics 273

B.1 PRELIMINARIES 273

B.2 SYNTAX AND PROGRAMMING 273
B.2.1 Command Line 273
B.2.2 Case-Sensitivity 274
B.2.3 Displaying the Current Value of a Variable 274
B.2.4 Clearing All Variables 274
B.2.5 Closing MATLAB 274
B.2.6 Variables and Arithmetic Operators 274
B.2.7 Programs 276
B.2.8 Comment Lines 277
B.2.9 Printing to the Screen 277
B.2.10 Numerical Format 278
B.2.11 Loops 278
B.2.12 Crashing a Program on Purpose 279
 Contents ■ xv

B.2.13 Logical Statements (If-Then-Else) 279

B.2.14 Input and Output Files 280
B.2.15 Creating Functions 281
B.2.16 Subroutines 282
B.2.17 Vectors, Matrices, and Arrays 283
B.2.18 Functions in the MATLAB Library 285
B.2.19 Plotting 286
B.2.20 Simulating Discrete-Time Models 286
B.2.21 Simulating Ordinary Differential
Equations (ODEs) 290
B.2.22 The Downhill Nelder-Mead Algorithm 292
B.3 EXERCISES 293

Appendix C ■ Connecting Models to Data: A Brief Summary

with Sample Codes 295

C.1 PARAMETERIZATION 295

C.2 RESIDUAL ERRORS (RESIDUALS) 296
2
C.3 RSS AND R 296
C.4 MAXIMUM LIKELIHOOD (ML) PARAMETERS 297
C.5 LEAST SQUARES (LS) PARAMETERS 297
C.6 IMPLEMENTATION IN CODE 298
C.6.1 Basic Structure of Program 298
C.6.2 Constructing Input Files 299
C.6.3 Example: Algebraic Model Using Vectors 299
C.6.4 Example: Algebraic Model Using Loop 300
C.6.5 Example: Scalar Map with One-Step
Predictions 302
C.6.6 Example: Higher-Dimensional Discrete-Time
Model with One-Step Predictions 303

Index 307
Acknowledgments

We thank the U. S. National Science Foundation for funding the

research of the Beetle Team and Seabird Ecology Team, whose data and
models form the case studies in this book. We also thank the Office of
Research and Creative Scholarship at Andrews University for funding
our research and students. Members of the Beetle Team have been
especially influential in the preparation of this book, namely Robert
F. Costantino, Jim M. Cushing, Brian Dennis, Robert A. Desharnais,
and Aaron A. King, as well as senior members of the Seabird Ecology
Team Gordon J. Atkins, Jim M. Cushing (again), Joseph G. Galusha,
and Lynelle M. Weldon. We also thank the co-authors of articles on
which the case study chapters are based. Our research students and
the students who have taken the mathematical modeling class deserve
much credit for inspiring and testing this book. We express special
thanks to our Ph.D. advisors, Thomas G. Hallam and Don E. Miller,
and to S.M.H.’s postdoc advisor Jim M. Cushing (again!), who turned
us into professionals and sharpened our thinking. Several institutions
have supported our work logistically, including Rosario Beach Marine
Laboratory, U. S. Fish and Wildlife Service, and Cape George Colony
which provided use of its marina for access to Protection Island. We
are particularly indebted to Ross Anderson of Cape George for his
friendship and numerous kindnesses during our research. It was a joy to
work with Jessica Rim, who prepared the figures. Finally, we appreciate
the kind editors at CRC Press/Taylor & Francis Group.

xvii
For professors and
students

This textbook has been inspired by our graduate and undergraduate

research students and the fascinating work with which they have
been engaged over the last few decades. It is a research methods
textbook for a mathematical modeling course, written for upper
division mathematics and science students and graduate-level biology
students. Although the methodology in this text is presented in the
context of ecology because that is the research area of the authors,
it can be imported to any other science, including social science and
economics.
The prerequisite for this textbook is a solid background in
differential calculus and an introduction to integral calculus at least
through the method of integration by substitution. This prerequisite is
equivalent to most one-semester university Calculus I courses. The text
contains three appendices that easily lead the student through series of
exercises to develop all other necessary skills from linear algebra and
computer programming. Although the programming language is not
specified in the text—an instructor can use any scientific programming
language—the sample codes and instruction in the appendices use
MATLAB. Thus, if students are first-time programmers, the easiest
approach for both teacher and student would be to use MATLAB,
unless the instructor wants to rewrite the sample instructional codes
in another language.
Our intent in this book is to integrate mathematical theory,
modeling, real data from the literature, and programming in such a
way that students gain significant research skills. The “Case Study”
chapters involve connecting models to data. Most of the exercises
are not “routine” but rather require students to engage in a way
that a researcher would, bringing together many skills to solve novel
problems. Although there is not a large number of exercises, they

xix
xx ■ For professors and students

are substantial. Many of them can be assigned as projects. Even

though Calculus I is the only prerequisite, this approach requires
a level of intellectual maturity in students. Only students who are
ready to engage in open-ended and sometimes frustrating problem
solving leading to unknown outcomes should take the course. At
our university, we have used the text for a research methods “swing
course”—a capstone course for undergraduate mathematics and science
majors and a beginning course for biology graduate students. Highly
motivated, curious, research-oriented freshmen and sophomores will be
successful, as well.
Each chapter has an annotated bibliography. In general we have
restricted the references to classic or foundational papers. In the “Case
Study” chapters, we also include the body of literature on which the
case study is based. The short annotations below most bibliography
entries allow the reader to understand the relevance at a glance.
We hope this text will help students think deeply and clearly
about the meaning of mathematics in science and learn some of the
research methods of applied mathematics. Most importantly, we hope
that students will experience some of the excitement of doing research.
Researchers are explorers who have a passion to find and understand
the deep structure of the universe. May each student find a passion for
learning and discovery.
Authors

Shandelle M. Henson is Professor of Mathematics and Professor of

Ecology at Andrews University, Michigan, the USA. She uses dynamical
systems theory and bifurcation theory to study nonlinear population dy-
namics and the effects of climate change on marine organisms. Shandelle
earned a Ph.D. in mathematics at the University of Tennessee, Knoxville,
and did several years of postdoctoral work at the University of Arizona.
She serves as Editor-in-Chief of the journal Natural Resource Modeling
and is a co-author of the book Chaos in Ecology: Experimental Nonlinear
Dynamics (Academic Press 2003), which documented chaotic dynamics
in laboratory populations of insects.

James L. Hayward is Professor Emeritus of Biology at Andrews

University, Michigan, the USA. He earned a Ph.D. in zoology from
Washington State University, and he has studied the behaviors and
population dynamics of marine birds and mammals in the Pacific
Northwest of North America since beginning graduate studies in
1972. In addition, Jim has studied the behavior of marine iguanas
and flightless cormorants on the remote and uninhabited island of
Fernandina, the westernmost island in the Galápagos. He also is a
recognized expert in the fossilization of eggshell in birds and dinosaurs.
Shandelle and Jim, a wife-husband research team, are widely
published in both the technical and popular literature, and both
have won awards for their teaching. They have applied mathematical
methods to the behavioral dynamics of seabirds at Protection Island
National Wildlife Refuge in the Strait of Juan de Fuca, Washington, the
USA, since 2002. They reside in Niles, Michigan, the USA, and they
have a daughter, son-in-law, and four grandsons. They enjoy hiking,
geology, art, music, and literature.

xxi
 I
Introduction to Modeling

1
 CHAPTER 1

Mathematical Modeling

1.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

We begin this textbook with an overview of the concept of

mathematical modeling. What is mathematical modeling, how is it
done, and how does it fit into the scientific method? How do the tools
of mathematics and science together fit into the larger context of the
humanities?
Mathematical models may be “proof-of-concept,” or quite realistic
or somewhere in between. A proof-of-concept model is a toy model,
not tied rigorously to data from a specific system, that incorporates
a mechanism(s) of interest. The mechanism is mathematically probed
and perturbed by the modeler in order to determine implications for
the outcomes. Mathematical results from such a model analysis suggest
hypotheses for the real-world system of interest. A realistic model, on
the other hand, is tied rigorously to data from a particular system. Of
course, even a realistic model has simplifications by design and is not
as complicated as the original system. In general, the more realistic a
mathematical model is, the less tractable it is. The “art of modeling” is
the skill of finding that middle ground in which the model is tractable,
but realistic enough to be informative.
In addition, mathematical models may be phenomenological or
mechanistic. Phenomenological models describe the observed systems
via curve fitting, but are not based on first principles. The equations
of mechanistic models, on the other hand, are based on the main
mechanisms thought to drive the dynamics of the system. The most

DOI: 10.1201/9781003265382-2 3
4 ■ Mathematical Modeling in Biology

powerful mathematical models not only describe, but also explain

and predict real-world systems. In general, models are most useful
scientifically if their formulations are mechanistic.
With these general ideas in mind, let’s take a closer look at
mathematical models.

1.2 THE MODELING CYCLE

A mathematical model is an equation or system of equations that
describes a physical or biological system. For example, the algebraic
model
v(t) = −gt (1.1)
is an equation that describes the velocity of an object dropped from
rest near the surface of a planet, where g > 0 is the acceleration due
to gravity and t is time.
Models often describe dynamical systems, that is, systems that
change through time. Since continuous-time rates of change are
derivatives, models often take the form of differential equations
(equations that involve derivatives). For example, the algebraic model
(1.1) can be written as the differential equation model

ds
= −gt, (1.2)
dt
where s(t) is the distance traveled by the object. The differential
equation model (1.2) is a continuous-time model.
Some physical and biological processes are better described by
discrete-time models. For example, if each individual in a population
of annual plants produces b seeds per year, a fraction p of which
germinate, then the number of individual plants the next year is

Nt+1 = pbNt , (1.3)

where Nt is the number of plants in the population at year t and

the parameters satisfy b > 0 and 0 < p < 1. Model (1.3) is called a
difference equation, a discrete-time map, or a recursion formula.
Mathematical models are effective tools for addressing scientific
problems. This is because a mathematical model is the translation
of a scientific problem into the precise language of mathematics,
 Mathematical Modeling ■ 5

Figure 1.1: The process of mathematical modeling.

where it becomes amenable to many powerful techniques. Figure 1.1

shows a conceptual diagram for the process of mathematical modeling.
Important issues arise in each of the four modeling steps.

1.2.1 Step 1: Translation into Mathematics

The first step in modeling is to translate the scientific problem into
the precise language of mathematics. This process requires a crisp
conceptual clarification that often proves valuable in and of itself.

1.2.1.1 Choosing Variables and Parameters

We must decide what quantitative aspects of the system we wish to
follow and name those quantities with variables. Input variables are
called independent variables. Time, location in space, and age are
examples of common independent variables. Dependent variables are
the functions of interest that depend on the independent variables. In
modeling, dependent variables often are called state variables, because
they track the state of the system. For example, suppose we want to
study how a population is distributed in time and space. We could take
N (t, x) to be the density of organisms at time t at location x. Here,
t and x are the independent variables and N is the state variable. In
equation (1.2), the state variable is s and the independent variable is
t. We know this because the presence of the derivative ds/dt alerts us
to the fact that s(t) is the function under consideration.
6 ■ Mathematical Modeling in Biology

Besides the independent variables, the system in question may

depend on other extrinsic (environmental) factors that are relatively
constant on the time scale of the problem. Typical examples are
temperature and humidity. We might include these factors in the
equation as constants. Such constants are called parameters. (Some
people call them “model coefficients,” but this is not the best
terminology because the constants do not always appear in the equation
as true coefficients; for example, they may appear as terms or as
exponents.) In equation (1.1), t is the independent variable, v is the
state variable, and g is an extrinsic parameter (the acceleration due
to gravity). Earth has one value of g (about 9.8 m/s2 ), while Mars
has another (about 3.7 m/s2 ). Other parameters may be intrinsic.
In equation (1.3), N is the state variable, t is the independent
variable, and b is an intrinsic parameter. The parameter p might be
a compendium of both extrinsic and intrinsic factors.

1.2.1.2 Simplifying Assumptions

Once the variables have been identified, they must be linked together
with equations. Constructing model equations always involves making
simplifying assumptions about the real system; for example, the object
is close to the surface of the planet; the force of gravity at the surface
of the planet is constant; gravity is the only force acting on the object
(no friction). Mathematical models do not perfectly describe the “real
world”; they are simplifications that capture the main mechanisms
driving the observed patterns. A major goal of modeling and the art of
modeling is to construct a simple, mechanistic, low-dimensional (few
state variables) model that describes enough of the observed variability
so that the leftover, “higher-order” variability is relatively small and
can be described statistically as noise (random effects). We can add
more and more mathematical complications (for example, the force due
to friction on the falling object) to a model in an attempt to make it
more realistic, but at some point the model becomes unwieldly and
intractable. There is little advantage in replacing a real system that is
too complicated to understand with a mathematical model that is too
complicated to understand. An important modeling mantra is therefore
KISS (keep it simple stupid). On the other hand, if the assumptions
oversimplify the problem, the mathematical model will not be able to
 Mathematical Modeling ■ 7

approximate the real system. Thus, in modeling there is always a trade-

off between realism and tractability. The goal is to capture the essential
behavior of a complicated real system with a simplified mathematical
system; the “art” consists in determining the main mechanisms driving
the system, building the assumptions around those mechanisms, and
consigning lesser influences to noise.

1.2.1.3 Parameterization

If model equations are to describe a real system, they must be

connected to real data through parameterization. Parameterization is
the statistical process by which data are used to estimate the values
of the parameters. For example, in equation (1.1) the value of g can
be estimated from experimental data on falling objects. You may have
done this experiment in a general physics lab. Parameterization is also
known as model fitting or parameter estimation. The word “estimation”
in this context is a technical term from statistics that refers to the result
of rigorous model fitting.
The parameterized model equation(s) may now be taken, tenta-
tively, as a surrogate for the real system. In fact, the model constitutes
a hypothesis about the scientific problem. More than one model can
be constructed, based on different hypotheses about what drives the
system. These competing models serve as alternative hypotheses, which
are then tested in Step 3.

1.2.2 Step 2: Model Analysis

The second step in modeling is mathematical analysis. All the powerful
concepts and tools of mathematics may now be brought to bear on the
scientific problem through its surrogate, the model. The analysis may
require solving differential equations, drawing bifurcation diagrams,
linearizing, or utilizing any of a host of other techniques you will learn
in this book or have learned in other mathematics classes.

1.2.3 Step 3: Back-Translation

The third step is to translate the mathematical results back into the
language of the original scientific problem. At this point, we must ask
8 ■ Mathematical Modeling in Biology

whether the results correspond to extant data, and whether the model
can accurately predict new data. If not, then we will need to revise the
model and “go around the loop again” (Figure 1.1).

1.2.3.1 Model Selection and Validation

Model selection is the process by which alternative parameterized

models are compared to each other in an effort to select the best
model. In Chapter 2, we will discuss a statistical tool called the Akaike
information criterion (AIC), one of the tools from information theory
that researchers often use for model selection. Model validation is the
process of evaluating a model by comparing its output to new data.
If there is a pronounced lack of correspondence with data, then the
modeling assumptions need to be revised. Model validation should
be a procedure distinct from model fitting. Ideally, models should be
validated on independent data sets, that is, on data sets that were
not used to estimate the model parameters. Model validation involves
computing the so-called goodness-of-fit statistics that measure how
closely model output corresponds to data.

1.2.3.2 Test of Model Predictions

A good model not only describes and explains, but also predicts. The
best way to further test a validated model is to make a priori
predictions and then collect data to test those predictions. Ideally, the
predictions should be unusual or unexpected. A successful test of such
predictions makes a strong case that the model really does capture the
essential behavior of the system.

1.2.4 Step 4: Revising Model Assumptions

The fourth step is ubiquitous in modeling. A model that doesn’t

correspond to data must be revised. One or more of the assumptions
were false or were oversimplifications; important factors may have been
left out of the model. In practice, model revision occurs more or less
continually throughout the modeling process as new insights are gained
and equations are tweaked.
 Mathematical Modeling ■ 9

You can see that modeling in biology requires a thorough

integration of biology, mathematics, and statistics. The next three
sections contain some important general ideas and terms from each
of these disciplines.

1.3 BIOLOGY

Biology is the study of living systems. Mathematical models can

describe living systems at subcellular, cellular, tissue, organ, organ
system, organism, population, community, and ecosystem levels of
organization. The level of organization chosen depends on the interest
of the investigator and the questions asked. For example, if an epidemi-
ologist (biologist who studies patterns of disease) wants to know how
fast a disease will spread through a region, the modeling process would
focus on the population inhabiting the region. Modeling subcellular
processes might be of little help. In other words, before modeling
begins, one must select an appropriate scale of interest (Levin 1992).
Living systems are the most complex systems known. In order
to model a living system at any scale, we must observe the system
thoroughly. In Chapter 10, we explain how we developed a model
to accurately predict the number of seabirds occupying a particular
habitat at a particular time. Our efforts were successful because we and
others spent a lot of time learning about the everyday lives of these
birds. Armed with this knowledge, we could ask the right questions,
identify the independent variables most likely to influence seabird
behavior, and develop an efficient data collection protocol suitable
for model parameterization and testing. There is no substitute for an
intimate observational knowledge of a living system whose mechanisms
you want to understand better through modeling. This is why an
applied mathematician who wants to model a biological system should
collaborate with a biologist who knows the system well.
Some biologists assume that living systems are too complicated to
be modeled mathematically. It may be that at some scales, a particular
living system is too complicated to model. It might not be feasible, for
example, to model how each of the millions of neurons in your cerebral
cortex are working in concert to process information conveyed by the
words in this paragraph. A neurobiologist, however, might be able to
model the activity of two or three interacting neurons, or entire sections
10 ■ Mathematical Modeling in Biology

of the brain. Mathematical modeling may not be able to answer all the
interesting questions a biologist might ask.
Evolutionary biologists and population geneticists have long taken
advantage of mathematical modeling. This is because people like G. H.
Hardy, Sir Ronald Fisher, Sewell Wright, and J. B. S. Haldane, trained
in mathematics, turned their attention to solving biological problems.
Biologists in other subdisciplines have also discovered the power of
modeling.
At the most basic level, mathematical modeling provides biologists
with a way to qualitatively and quantitatively describe the behavior of
living systems. At a higher level, modeling provides a tool to identify
factors that drive living systems. Finally, once these factors have been
identified, biologists can test predictions about how these systems will
function in the future.
Our interest and experience concern ecological modeling. Because
many of the examples used in this book come from ecology, we now
provide a brief description of this subdiscipline.

1.3.1 Ecology
Ecologists are concerned with interactions among organisms and their
environments. Both organisms and environments are very complex;
moreover, they constantly undergo change. Ecologists thus face
significant challenges as they try to uncover ecological patterns.
An interesting challenge that faces all ecologists is the problem of
scale. Patterns apparent at one scale may disappear at higher or lower
levels of organization, and new patterns may emerge at these levels.
Patterns at some scales exhibit emergent properties. This is another
way of saying the whole often seems more than the sum of the parts—
qualities emerge at higher levels that might not be predicted on the
basis of the known properties of lower levels. Emergent properties are
not magical, however. They are the result of the properties of entities
at smaller scales and all of their interactions at the larger scale.
Ecologists view ecosystems as consisting of progressively more
inclusive scales:

Individual organism: Individuals live more or less autonomous

lives. Members of solitary species interact little with conspecifics,
whereas members of social species interact frequently with
 Mathematical Modeling ■ 11

one another. Moreover, individuals are sometimes difficult to

distinguish. Examples include the assemblage of medusae and
polyps in a Portuguese man-o’-war, the collection of individuals
forming a sphere of Volvox, and a grove of aspen trees generated
through vegetative reproduction.

Population: Members of a single species that inhabit a particular

geographic region.

Species: A group of interbreeding or potentially interbreeding

organisms genetically isolated from other such groups. (Note: This
is only one of many competing species concepts. Species concepts
generate lots of debate among biologists.)

Community: An assemblage of populations that occupy the same

geographic region.

Ecosystem: A community plus all the non-living components of

the system. These non-living (abiotic) components include things
such as energy, moisture, and nutrients.

Ecology includes a number of subdisciplines:

Physiological ecologists (or environmental physiologists) deal with

how individual organisms cope with environmental change. For
example, they have investigated how camels conserve water in the
heat of the desert, and how seals and porpoises conserve oxygen
during long underwater dives.

Population ecologists are interested in the factors that influence

population growth.

Community ecologists are concerned with interactions between

and among species that live in the same community. These
interactions include food web relations, predator-prey interactions,
interspecific competition, mutualism, parasitism, speciation, and
coevolution.

Ecosystem ecologists examine how energy flows through and how

nutrients cycle within ecosystems. Thus, ecosystem ecologists work
with energy budgets and accounting schemes by which the precise
12 ■ Mathematical Modeling in Biology

nature of community interactions can be teased apart, and they

determine pathways taken by nutrients through ecosystems.

Behavioral ecologists study the causation, function, ontogeny, and

evolution of the behaviors of animals in their ecological setting.

Mathematical ecologists use theoretical methods such as mathe-

matical models and computational simulations to study ecosys-
tems. They try to understand how the dynamics of populations
and ecosystems depend on mechanistic processes as well as biotic
and abiotic factors.

Considerable overlap occurs among the areas of ecology. For

example, a population ecologist needs to understand the life history
characteristics of individual organisms in the population as well as
how community interactions such as competition and predator-prey
relations impact the population.
Ecologists, like other biologists, are often tempted to confuse
correlation with causation. The fact that we can mathematically model
a relationship between particular independent and dependent variables
is no proof that the variables are causally linked. Mathematical
implication is not the same as causation. For example, we might develop
a model based on temperature change that quite accurately predicts
when white-crowned sparrows migrate south for the winter. Yet we
know from experimental studies that photoperiod, not temperature,
is the environmental factor that causes white-crowned sparrows to
migrate. Because temperature covaries with the light-dark cycle, we see
how easy it would be to arrive at an invalid conclusion concerning the
cause of migration. Thus, if we are not careful, mathematical modeling
can sometimes lead us to incorrect conclusions. On the other hand,
mathematical modeling can often suggest possible causal relationships
that we might not have guessed otherwise.
Another point worth making in this context is the difference
between proximate and ultimate causes. In the example just given, the
proximate cause of migration is photoperiod change, which alters the
neuroendocrine system of white-crowned sparrows. The ultimate cause
in this example was natural selection which, over generations, favored
the fitness of birds that nested where food was abundant during the
spring and moved to warmer regions when nesting areas turned frigid.
 Mathematical Modeling ■ 13

1.4 MATHEMATICS

The most basic mathematical concept in modeling is that of the

function. A relationship between an input t and an output x(t) is called
a function if and only if the value of the output x(t) is completely
determined by the input t. Symbolically, x(t) is a function if and
only if
t1 = t2 =⇒ x (t1 ) = x (t2 ) ,

where the double arrow means “implies.” In words, if the inputs are
the same, then the outputs must be the same. Said another way, you
can’t get two different outputs from the same input. Algebra students
learn this as the “vertical line test.” Functions are deterministic. (By
contrast, a relation in which you can get two different outputs from
the same input is called stochastic; we will turn our attention to
stochasticity in the next section.)
In modeling dynamical systems, we are interested in the state of
the system (call it x) as a function of time t. That is, we are interested
in the behavior of x(t) as t changes. We can list the corresponding
values of x and t in a table, and we can also graph x vs. t. This list
of pairs of numbers (or its graph) is called a time series. You will
see many times series graphs in later chapters (e.g., Figures 4.2 and
8.2). If there are two state variables for the system of interest, say x(t)
and y(t), we must show two time series graphs together: x vs. t and
y vs. t.
In general, the two state variables x and y may be coupled, that
is, may depend on each other, so the two time series graphs must be
interpreted together. This is a visually difficult task. A better visual
tool in this case is that of state space, or phase space, in which the
state variables are graphed against each other in the x-y plane. The
current state of the system is represented by a point (x(t), y(t)) that
moves around in the plane as time progresses, tracing out an orbit.
You can see an example of orbits in a state space graph in Figure 9.1.
The arrows indicate the movement of the point (x(t), y(t)) in time.
Calculus provides us with many powerful tools for studying systems
whose state changes more or less continuously throughout time. The
main tool, of course, is the derivative dx/dt, which is the instantaneous
rate of change of x(t) with respect to time t.
14 ■ Mathematical Modeling in Biology

A dynamical system is at equilibrium if it does not change over time.

The time series graph for such a system is a horizontal line x(t) = xe .
In state space, a system at equilibrium is represented by an orbit that
consists of a single stationary point (xe , ye ) (for a two-dimensional state
space). Mathematically speaking, equilibria are constant solutions of
the model. Consider, for example, the logistic population model , which
we will derive in a later chapter:
 
dN N
= rN 1 − .
dt K

Here, N (t) is the population size at time t, and r, K > 0 are constant
parameters. A solution N (t) of the logistic model is a constant solution
if and only if it satisfies
dN
=0
dt
for all time t. Thus, the equilibria are the roots of the equilibrium
equation  
Ne
0 = rNe 1 − ;
K
that is, the equilibria are
Ne = 0, K.
For a discrete-time example of equilibria, consider the plant
population model (1.3). A solution of this difference equation is a
constant solution Ne if and only if it satisfies

Nt+1 = Nt = Ne

for all time t. Thus, the equilibra are the roots of the equilibrium
equation (or fixed point equation)

Ne = pbNe .

If bp ̸= 1, the only such constant Ne is the extinction equilibrium

Ne = 0. If bp = 1, then all constant values of Ne are equilibria.
Another important notion is that of stability. Consider a planar
pendulum. There are two equilibrium states: angle 0 (straight down)
with zero velocity and angle π (straight up) with zero velocity. If the
pendulum is at rest in the “down” position and is perturbed slightly
 Mathematical Modeling ■ 15

away from equilibrium, it will remain close to equilibrium. The “down”

equilibrium is called stable, because if the system starts close to it, the
system stays close to it. In fact, the system asymptotically returns to
equilibrium in this case, so the equilibrium is called asymptotically
stable. If, on the other hand, the pendulum is at rest in the “up”
position and is perturbed, the pendulum will move away from the “up”
equilibrium. This equilibrium is called unstable.
Consider now a large hemispherical bowl, sitting upright on a table,
with a tennis ball inside it. If the ball is placed on the bottom of the
bowl with zero velocity, it will remain there. This state is therefore an
equilibrium (“if it starts there, it stays there”). If the tennis ball is
released, perhaps with a small velocity, near the bottom of the bowl,
it will remain in the vicinity of the bottom of the bowl with small
velocity, and so the equilibrium state is stable (“if it starts close, it
stays close”). In fact, the tennis ball actually approaches the bottom
of the bowl and its velocity approaches zero, so the equilibrium state
is asymptotically stable. Now suppose the bowl is turned upside down.
The top point of the bowl gives rise to an unstable equilibrium for the
ball, because although it is true that “if you start there, you stay there,”
it is not true that “if you start close, you stay close.” Now remove the
bowl and set the tennis ball down on the table top with zero velocity.
If the table is flat, the ball remains where you placed it. Therefore,
it is at equilibrium. If you move the ball a small distance away and
release it with a small velocity, still on the table top, it probably will
not return to the old position, but it will not stray far. Thus, it is true
that “if you start close, you stay close” to the original position; hence,
the equilibrium is stable. However, since the ball does not necessarily
return to the original state, the stable equilibrium is not asymptotically
stable. We say it is neutrally stable. In fact, every point on the table
top is a neutrally stable equilibrium.
Finally, we list a few phrases in English that you should always be
able to translate directly into mathematics. We say y is proportional
to x if and only if we can write
y = ax
for some constant a. We say y is inversely proportional to x if and only
if we can write
a
y=
x
16 ■ Mathematical Modeling in Biology

for some constant a. Population growth rate is the change in population

size with respect to time. In continuous time, this is the derivative
dN/dt. A population’s per capita growth rate is defined to be

1 dN
.
N dt
For example, suppose a population’s growth rate is proportional to
the square root of the population size N , and inversely proportional
to the temperature T . A mathematical model that describes this
system is √
dN N
=a .
dt T
For another example, suppose a population’s per capita growth rate
is inversely proportional to the square of the population size. A
mathematical model that describes this system is

1 dN a
= 2,
N dt N
that is,
dN a
= .
dt N

1.5 STATISTICS
Deterministic models are approximations of real systems; a good model
captures the signal (main deterministic trend) in the data. Neverthe-
less, the data likely will deviate somewhat from the model prediction.
This deviation from the signal is called noise or stochasticity. The
two main types of noise in biological data are process error and
measurement error (observational error ). The process error occurs
because the real system is more complicated than the mathematical
model. The measurement error occurs because the real system cannot
be measured exactly. Stochasticity in ecological data can be handled
with statistical methods, several of which will be addressed in this
book.
There are two main types of process error in ecology: environmental
stochasticity and demographic stochasticity. Stochastic events in a
population can be likened to the toss of a fair coin. Imagine that
 Mathematical Modeling ■ 17

a single coin is tossed for a population of animals. The outcome of

the toss, although random, is the same for each individual member
of the population. This is environmental stochasticity. Such extrinsic
events as weather cause this type of noise. Now imagine that each
animal in the population tosses its own coin. This time there is a
random outcome for each individual. This is demographic stochasticity.
Individual variability in intrinsic parameters such as birth and death
rates cause this type of noise.
Systems in classical physics may have relatively little stochasticity,
and their mathematical models can be so precise that some people call
them “laws.” Some social science systems, on the other hand, may have
a lot of stochasticity—so much so that the signal may be swamped out
by noise and mathematical modeling may be impossible. In ecology,
deterministic and stochastic forces are more or less equally important.
Therefore, noise should—ideally—be incorporated explicitly into a
deterministic model to produce a stochastic version of the model. The
interaction of deterministic and stochastic forces can give rise to a rich
class of emergent dynamic phenomena that cannot occur in purely
deterministic or purely random systems.
Stochasticity is modeled mathematically through the notions of
random variable and distribution. Suppose you count the number of
seabirds on a beach. If the number of birds is large, repeated counts of
the exact same group of birds will probably yield different results due
to observational error. We can use the notion of a random variable X to
stand for the outcomes of trial observations (counts). Note that random
variables are usually denoted with uppercase letters. A particular
observation X = x, where x is an observed value, is called a realization
of the random variable. You would want to know if some observations
x are more likely than others, because you might want to know, for
example, whether the count errors are biased. Mathematically, you
would be asking, “what is the distribution of the random variable X?”
The answer to this question is, of course, situation dependent. One has
to make assumptions about how measurements, and hence errors, are
distributed. Ideally, such assumptions can be tested experimentally.
The main distribution used in this book is the normal distribution.
We will continue our discussion by supposing the bird count random
variable X is normally distributed about the true number of birds
µ. The normal distribution (Figure 1.2) probability density function
18 ■ Mathematical Modeling in Biology

Figure 1.2: The normal distribution with µ = 10 and σ = 3.

(PDF) is given by
1 1 x−µ 2
f (x) = √ e− 2 ( σ ) . (1.4)
σ 2π
Here, f (x) is not the probability of observing x birds. Rather, the
probability that the observational count X will fall between the values
a and b is the area under the normal curve that lies between the vertical
lines x = a and x = b; that is,
Z b
P [a ≤ X ≤ b] = f (x) dx.
a

The total area under the normal curve gives the probability that the
count lies between −∞ and +∞, which is, of course, equal to one.
Therefore, it must be true that (Exercise 10)
Z ∞
P [−∞ < X < +∞] = f (x) dx = 1. (1.5)
−∞

The expected value of the random variable X, often denoted E(X), is

the sum of all possible outcomes of X weighted by the probability of
obtaining that outcome. For the normal distribution, E[X] = µ; that
is (Exercise 11), Z ∞
E[X] = xf (x) dx = µ. (1.6)
−∞

The inflection points of the normal PDF f occur at x = µ±σ (Exercise

12). The parameter σ is called the standard deviation. It measures the
 Mathematical Modeling ■ 19

spread of the distribution about the mean µ. The square of the standard
deviation, σ 2 , is called the variance of the distribution.
You may have noticed that we have assumed X to be a continuous
random variable, while in the bird counting example, it is actually a
discrete random variable. That is, you can count only an integer number
of birds, while the normal distribution allows x to be any real number.
We could have used a bell-shaped histogram for the distribution
of a discrete random variable, but we often simplify problems with
continuous approximations.

1.6 EPISTEMOLOGY: HOW WE KNOW

There are two main kinds of logical inference: deduction and induction.
Deduction infers a particular conclusion from a general statement. For
example,

All adult male northern cardinals (Cardinalis cardinalis) are red.

The bird Jim is observing is an adult male northern cardinal.
Therefore, the bird Jim is observing is red.

Notice that if the first two statements are true, then we all agree
that the conclusion is guaranteed to be true. Deductive arguments are
conclusive. However, if one (or both) of the first two statements is false,
then the conclusion is not guaranteed to be true; it might be true or it
might be false.
Induction infers a general conclusion from a set of particular
statements or observations:

All the adult male northern cardinals I have observed are red.
Therefore, all adult male northern cardinals are red.

Notice that, despite your observations, the conclusion might still

be false, depending on the sample of data you observed. Perhaps you
did not have the chance to observe an albino cardinal. The only way
you can be completely sure of your conclusion is if your induction
was exhaustive, that is, if you observed all male northern cardinals.
Inductive arguments are not, in general, conclusive unless you are able
to observe all possible instances of the data.
20 ■ Mathematical Modeling in Biology

Here, we pause to speak to fellow aficionados of Sherlock Holmes.

Despite the frequent use of the word “deductive” in the Sherlockian
canon, Holmes’ methods were mostly inductive! We wonder how many
students have missed questions about deduction versus induction on
standardized exams because they thought back to the methods of that
master of induction!
Pure mathematics is deduction. Make no mistake: The activity of
doing mathematics is not purely deductive. Trial and error, hunches,
and experimentation lead mathematicians to pose the conjectures that
they then try to prove deductively as theorems. The final result,
however, must be purely deductive, or it is not pure mathematics.
Inductive arguments are not permitted as mathematical proof. (In your
mathematics classes, you may have learned a special kind of proof
called “mathematical induction” (MI). MI is actually a special type
of deduction. See Exercise 14.) Mathematics is the only discipline in
which arguments are conclusive. In one sense, deduction cannot create
“new” knowledge about physical reality, since the conclusions must be
implicit in the statements with which you begin. In another sense,
deduction can create new knowledge, because it is a powerful tool
that teases information out of more general statements—information
that you might never have guessed was implied by the original
statements. The “problem,” of course, is that the general statements
themselves must have come from even more general statements, and
so on. You can see that in pure deduction, there must be original
statements which must be taken without proof. Indeed, the edifices
of pure mathematics ideally rest on unproven original statements
called axioms. Now, pure mathematicians do not ask whether those
axioms are true with respect to the observed world. They only
want to know that the axioms are logically independent and logically
consistent, that is, that no axiom is derivable from the others, and
that they cannot generate a logical contradiction among themselves.
The axiomatic method is a fascinating and tremendously powerful
tool.
In general, scientists do want to know whether or not their starting
statements correspond to observations in the real world. Unlike pure
mathematicians, they are interested in the correspondence between
logical inferences and observations of nature. Science, therefore, utilizes
both deduction and induction. Science inductively infers hypotheses
 Mathematical Modeling ■ 21

Figure 1.3: Scientific methodology.

from data, draws deductive conclusions (called predictions) from

these hypotheses, and then tests the predictions against more data
(Figure 1.3). The inclusion of observation and induction in scientific
methodology makes science more powerful than pure mathematics
alone as a tool for understanding real systems. But there is a
price to pay for this increase in power, and that is a decrease in
certainty. In science, there is no proof in the mathematical sense of
the word.
The most powerful science is not purely observational or inductive,
but also uses deduction. Observation guides theory, and theory guides
observation. Note that the so-called hard sciences (physics is the prime
example) employ mathematics in the deductive step (Figure 1.3).
The richness of the humanities is a hallmark of what it means to
be human. The humanities ask some of the most important questions
about reality, questions science and mathematics cannot address, such
as questions about meaning, values, ethics, and God. The humanities
use deduction and induction, but they also employ other ways of
knowing. A great novel, for example, although not literally true, can
be more true than any transcript of reality. A great landscape painting
can convey more truth about nature than a snapshot or a data sheet.
This increase in scope of questions the humanities can address comes
at the price of a further decrease in certainty, in the sense of being able
to convince someone else of what you know.
22 ■ Mathematical Modeling in Biology

Figure 1.4: Nested epistemologies for pure mathematics, science, and

the humanities.

When we compare the types of epistemologies (ways of knowing)

of various disciplines, we can see that they are nested (Figure 1.4), as
in a Venn diagram.
We also see that the more rigorous the method of inference, the
more certain one can be, but the less one can learn about nature
(Figure 1.5). The least rigorous methods of “inference” include other,

Figure 1.5: Scope and conclusiveness of the epistemologies of the

disciplines.
 Mathematical Modeling ■ 23

extra-rational (not ir rational) ways of knowing. These epistemologies

are the most “powerful,” in the sense that they can address the
greatest range of questions about reality. But they are also the least
“compelling,” in the sense of being able to rationally convince someone
else. (See Exercise 26.)
In conclusion, systems of thought—mathematics, science, and the
humanities—have different epistemologies. One cannot do science
(using induction) and call it pure mathematics, even though it may
be perfectly good science. One cannot do theology (using scripture)
and call it science, even though it may be perfectly good theology.
Nevertheless, all of these disciplines may have a common goal:
understanding reality through discipline-specific types of models.
Science is not so much about certainty as it is about understanding.
In particular, mathematical modeling is not about absolute certainty.
It is about understanding. In the words of Alfred J. Lotka, an early
pioneer in mathematical ecology, “An empirical formula is therefore not
so much the solution of a problem as the challenge to such a solution.
It is a point of interrogation, an animated question mark”(Lotka 1925).
We hope you will find this book to be an “animated question mark.”

1.7 EXERCISES
1. Consider the population model

dx
= (b − d) x.
dt

Identify the independent variable(s), the state variable(s), and

the parameter(s).

2. Consider the LPA model of flour beetle (Tribolium) dynamics

L(t + 1) = bA(t)e−cel L(t)−cea A(t)

P (t + 1) = (1 − µl ) L(t)
A (t + 1) = P (t)e−cpa A(t) + (1 − µa ) A(t).

Identify the independent variable(s), the state variable(s), and

the parameter(s).
24 ■ Mathematical Modeling in Biology

3. Consider the logistic population model

 
dN N
= rN 1 − .
dt K

Identify the independent variable(s), the state variable(s), and

the parameter(s).

4. Consider the dynamical system

dx
= 2x(1 − x)(2x − 4)2 .
dt
Find all the equilibria.

5. Consider the discrete-time dynamical system known as the Ricker

model:
xt+1 = bxt e−cxt .
Here, b, c > 0. This famous model was first used in fisheries
(Ricker 1954). We will see it again several times in this book.
Find all the equilibria.

6. Consider the model

dx
= x + xy
dt
dy
= x + y.
dt
Find all the equilibria. Hint: The equilibria are constant solution
pairs (xe , ye ) and are obtained by setting both derivatives equal
to zero and solving the resulting algebraic system of equations.

7. Malthusian, or exponential, growth is modeled by

dx
= rx.
dt
Is the extinction equilibrium xe = 0 asymptotically stable,
neutrally stable, or unstable? Hint: Consider the cases r < 0,
r = 0, and r > 0 separately.
 Mathematical Modeling ■ 25

8. The per capita growth rate of a certain population is proportional

to both the population size and the temperature, and it is
inversely proportional to the humidity. Assume the temperature
and humidity are constants on the time scale of interest. Choose
variable names for the dependent and independent variables and
for the parameters. Translate the statement into a mathematical
model.

9. Identify the following random variables as environmental or

demographic: wind speed, per capita birth rate, outdoor
temperature, probability of death due to predatory encounter.

10. Note to instructor: This exercise requires a knowledge of vector

calculus (third-semester calculus). Verify equation (1.5). Hint:
Show that
Z ∞ 2
1 − 12 ( x− µ 2
)
√ e σ dx
−∞ σ 2π
Z ∞ Z ∞
1 − 21 ( x− µ 2
σ ) dx
1 y −µ 2
= 2
e e− 2 ( σ ) dy
2πσ −∞ −∞
Z ∞ Z ∞
1 1 2 1 2
= e− 2 u du e− 2 v dv
2π −∞ −∞
Z ∞Z ∞
1 2 2
e− 2 [u +v ] dudv,
1
=
2π −∞ −∞

where u = (x − µ) /σ and v = (y − µ) /σ. Now change from

rectangular to polar coordinates and carry out the integration.

11. Note to instructor: This exercise requires a knowledge of improper

integrals (second-semester calculus). Verify equation (1.6) for µ =
0. That is, verify that
Z ∞
1 x2
√ xe− 2σ2 dx = 0.
σ 2π −∞

12. Prove that the inflection points of the normal curve (1.4) occur
at x = µ ± σ.

13. Suppose you want to count a very large group of objects. Assume
the number of objects is fixed in time. There are so many objects
26 ■ Mathematical Modeling in Biology

that your repeated counts give different tallies. Suppose you

record a large number of your repeated counts and find they
are normally distributed with standard deviation s. Would you
expect s to be dependent or independent of group size? Explain.

14. Mathematical induction (MI) is a particular kind of mathematical

argument used to prove theorems of the form, “For all natural
numbers n = 1, 2, 3, . . . , the statement Q(n) is true.” MI is like
climbing an infinitely tall ladder, where each rung is a natural
number. Provided one can always move to the next highest rung
from any given rung, and provided one can get on the first rung to
begin with, one can ascend the whole ladder. MI uses a three-step
method. First, one proves that the statement is true for n = 1,
that is, Q(1). Second, one assumes the statement is true for an
arbitrary number n = k, that is, Q(k). Third, one shows that the
statement is therefore true for n = k + 1, that is, that Q(k + 1).
Here is an example: We will use MI to prove the statement “The
sum of the first n natural numbers is 1+2+3+. . .+n = n(n+1)/2.
Proof: (i) Clearly, this formula is true for n = 1. (ii) Assume it is
true for n = k; that is, assume 1 + 2 + 3 + · · · + k = k(k + 1)/2.
(iii) Then, using part (ii), we have 1 + 2 + 3 + · · · + k + (k + 1) =
k(k + 1)/2 + (k + 1) = (k + 1)(k/2 + 1) = (k + 1) ([k + 1] + 1) /2,
so the formula is true for n = k + 1. QED

a. Use MI to prove that the sum of the first n squares is 12 +

22 + 32 + · · · + n2 = n(n + 1)(2n + 1)/6.
b. Is MI induction or deduction, or could it be both? Explain
thoroughly.

15. Deduction is conclusive. Induction generally is not. Therefore,

scientific reasoning (induction + deduction) is not conclusive,
yet it is still a compelling and powerful type of argument. The
following questions are highly nontrivial.

a. Why do you believe good scientific reasoning is compelling?

b. Why do you believe deduction is conclusive?
c. Why is the word “believe” used in (a) and (b)?
 Mathematical Modeling ■ 27

16. Can one use deduction to prove conclusively that deduction is

conclusive?

17. Suppose you observe x = 4, 269 birds on the beach, whereas

the model predicted x = 0. Clearly, something is wrong, either
with the model or with your observation. You conclude there is
something wrong with the model. Why did your observation take
precedence over the model as the standard for reality in this case?
Can you prove that your observation depicts physical reality more
accurately than the model prediction? Explain.

18. Criticize or support the following statement: Every system of

thought must begin with unprovable presuppositions; however, a
good system of thought minimizes its presuppositions.

19. Criticize or support the following statement: If two models

describe the same phenomenon equally well, the most elegant
model is the one closest to the truth.

20. How does the statement in Exercise 19 relate to Occam’s Razor?

21. What presuppositions do all scientists have in common?

22. Criticize or support the following argument: Every system of

thought begins with unprovable presuppositions; thus, no system
of thought can yield absolute certainty about “reality.” Therefore,
all systems of thought are equally valid. Hint: Are astronomy and
astrology equally valid?

23. Do you think there is a difference between extra-rational and

irrational ? Explain.

24. Consider Figure 1.4. Can science logically contradict mathe-

matics? Can philosophy or theology logically contradict science
(or mathematics)? Note that it can be very difficult to prove
that an apparent contradiction is really a logical contradiction.
Some apparent contradictions are actually paradoxes that can be
logically resolved with more information.

25. Consider Figure 1.4. Can the epistemologies of the outer levels
ever be utilized legitimately at the inner levels? For example,
28 ■ Mathematical Modeling in Biology

can a pure mathematician use scientific induction when doing

mathematics? Can a mathematician or scientist use religious
scripture when doing mathematics or science? Explain your
answer carefully.

26. Can one hold extra-rational knowledge with internal certainty

even though the information cannot be transmitted to anyone else
using rational inference? Explain. If you say “yes,” give examples.

27. Criticize or support the following statement: My religion’s

scriptures are my set of axioms, and they trump all observations
made via the senses.

BIBLIOGRAPHY
Levin, S. A. 1992. The problem of pattern and scale in ecology. Ecology
73:1943–1967. DOI: 10.2307/1941447. [Classic paper on scale in ecology,
presented as the Robert H. MacArthur Award Lecture in 1989 in Toronto,
Canada. Every student of ecology and mathematical biology should read
this paper.]

Lotka, A. J. 1925. Elements of Physical Biology. Williams & Wilkins

Company, Baltimore. [Lotka was one of the first scientists to apply
mathematical methods to problems of population growth. He used
mathematics from the field of physical chemistry to create “physical
biology.” Students interested in mathematical biology, ecology, or the
history of ecology should be aware of this foundational book. In Chapter
9 you will study the famous Lotka-Volterra models of predator/prey,
competition, and cooperation dynamics.]

Ricker, W. E. 1954. Stock and recruitment. Journal of the Fisheries

Research Board of Canada 11:559–623. DOI: 10.1139/f54-039. [Classic
paper introducing the Ricker model.]
 CHAPTER 2

Avian Bone Growth: A

Case Study

2.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

In this chapter, we use bone growth data collected on a seabird colony
to illustrate the major aspects of modeling methodology. The chapter
is based on a paper published in the Journal of Morphology in 2009
by the Seabird Ecology Team, a National Science Foundation-funded
research group co-directed by the authors of this textbook. This paper
was co-authored by an undergraduate student (Hayward et al. 2009).
This is a crucial chapter for the student; many important ideas
and techniques are introduced within the context of this case study.
You should take time to master this material. Most of these techniques
require computer implementation. You should complete the self-guided
programming tutorial in Appendix B before beginning the exercises in
this chapter.
If you feel uncomfortable with basic matrix operations (addition,
subtraction, multiplication, scalar multiplication, and determinants) or
need a refresher, you should read Appendix A before working through
Appendix B.
Appendix C contains a concise summary of the methods of this
chapter along with samples of computer code, which students may
consult before attempting the exercises in this chapter.

DOI: 10.1201/9781003265382-3 29
30 ■ Mathematical Modeling in Biology

2.2 SCIENTIFIC PROBLEM

Growth is a fundamental process of life. Growth of multicellular
organisms involves the multiplication and differentiation of cells. Cells
can be arranged in different ways to create different body shapes,
just like bricks can be arranged in different ways to create different
buildings.
Size constrains an organism’s life. You would not be surprised to see
a spider walking up a vertical wall, but you would be surprised to see a
human doing this. Spider locomotion is constrained by electrostatic
forces, whereas human locomotion is constrained by gravitational
forces. Size makes the difference.
Shape also plays a defining role in life. A giraffe feeds on leaves
high in trees. By contrast, a hippo, shaped much differently, eats
different things and would find it impossible to feed like the giraffe.
As anatomists say, “structure follows function.”
Through development, multicellular organisms grow from single
cells into species-typical sizes and shapes. Organs within organisms
do the same thing. As organs grow, they also change shape. And as
they change shape, they change (or gain or lose) function.
Bones are organs of internal support in vertebrates. As a vertebrate
grows, the size and shape of its bones change as well. Different bones
change in different ways. By comparing the growth of various bones,
we can learn about where the animal is funneling its energy and what
is happening in its life.
Glaucous-winged gulls (Larus glaucescens) breed in large colonies in
North America’s Pacific Northwest. For many years, we have studied
these birds at Protection Island National Wildlife Refuge, Strait of
Juan de Fuca, Washington, the USA. We noticed that newly hatched
young behave differently than older juveniles, and that older juveniles
behave differently than adults. So we asked a simple question: How is
the development of a particular behavior related to the development
of a particular bone?

2.2.1 Data
Gull chicks hatch in late June and early July on Protection Island.
Many chicks die from predation, overheating, and dehydration, or are
 Avian Bone Growth: A Case Study ■ 31

killed by neighboring adults. If they are lucky enough to survive,

they will fledge at about 44 days old and reach maturity at four
years old.
During the hatching period, we banded 373 newly hatched chicks
(Hayward et al. 2009). Over the next several weeks, we collected 80
banded chicks that had died. Because we knew when each chick had
been banded, we could tell how old it was when it died. The dead chicks
ranged from 0 to 42 days old. We also collected 13 dead adults on the
colony.
We shipped the dead chicks and adults back to Michigan where
we prepared their skeletons—a tedious, time-consuming, and smelly
task. We decided to focus on three wing bones (humerus, ulna,
and carpometacarpus) and three leg bones (femur, tibiotarsus, and
tarsometatarsus), because these bones facilitate the easily observed
behaviors of flying and walking.
Using calipers, we measured the diaphyseal length and midshaft
diameter of each of the six bones from the right side of each skeleton
and recorded those data along with the chick age. Data for the humerus
and ulna lengths are given in Data Set 2.1. We graphed the lengths and
diameters of each of the six bones against age and noted that the bones
exhibited different growth patterns.
An individual female gull typically lays an egg every other day until
she has produced three eggs. The first two eggs, called the A and B
eggs, are generally bigger, are the first to hatch, and produce larger,
healthier chicks than the C egg. Our sample consisted of 52 (65% of
the sample) C chicks, probably because they are least likely to survive.
Thus, our data set is skewed in favor of smaller birds, something that
should be kept in mind when interpreting the results.
It is also important to note that we used bones from different birds
for each age level—in other words, ours was a cross-sectional study.
Had we followed the growth of individual bones in individual birds,
we would have done a longitudinal study. Cross-sectional studies are
limited by the fact that they tell us nothing about individual rates
of growth, but only provide estimates of mean rates of growth for a
population. Although it would be more informative, carrying out a
longitudinal study on bone growth in wild gulls would be logistically
difficult.
32 ■ Mathematical Modeling in Biology

2.3 TRANSLATION INTO MATHEMATICS

How do gull bones grow? A bit of thought will convince you that
this question cannot be translated into the language of mathematics,
because it is too vague. What does “grow” mean? We might define an
object to be “growing” if and only if its size is changing over time. (This
definition for growth includes shrinking as well as expanding.) But what
does “size” mean? Are we interested in length, diameter, volume, or
what? And which kind of bone are we talking about? Humerus? Ulna?
You can see that the very first step in the modeling process, which
is the translation into mathematics, typically requires careful thought.
This step can be quite fruitful in and of itself, even if you never
go any further in the modeling cycle, because the act of translating
forces you to clarify concepts and sharpen questions. Translation
into mathematics can help you ask whether your scientific question
makes sense and whether it can be expected to have a solution.
This is important, because some apparently meaningful questions are
actually nonsensical (Exercise 3). Indeed, some of the burning scientific
questions of history have simply “gone away” because they were
discovered to be meaningless.
Let’s pose our problem precisely. How does the length of the
humerus change in time over the life of a gull? That is, how does the
length of the humerus change as a function of age? Consider a single
“average” gull. Let

x = Age in days
f (x) = Length of humerus in cm.

Mathematically, the question becomes: How does f (x) depend on x?

2.3.1 Simplifying Assumptions

When we relate the variables x and f (x) through an equation, we
are making implicit biological assumptions. As far as possible, these
assumptions should be stated explicitly in the language of biology.
Ideally, the assumptions should address two kinds of questions. First,
what deterministic mechanisms are most important in driving the
system? Second, what kind of process error is in the system, and what
kind of measurement error is in the data?
 Avian Bone Growth: A Case Study ■ 33

2.3.1.1 Deterministic Assumptions

We pose two alternative hypotheses as the deterministic modeling

assumption. They will give rise to two competing deterministic models,
which we will test against one another using the data.

(A1a) The length of the humerus increases linearly with age until
the chick reaches some critical age, at which point the bone
abruptly ceases to grow.

(A1b) The length of the humerus grows first at an increasing

rate and then at a decreasing rate, asymptotically leveling off
(saturating) toward some maximal length. (This is called sigmoidal
growth because the length vs. age curve is s-shaped.)

2.3.1.2 Stochastic Assumptions

We also make assumptions about process and measurement errors in

the system:

(A2) The dominant type of noise in the system is demographic

stochasticity; environmental stochasticity and measurement error
are negligible.

2.3.2 The Deterministic Model

Assumption (A1a) says that humerus length f (x), when graphed
against age x, is a line with positive slope from age 0 (hatching)
to some critical age, at which point the graph of f (x) becomes a
horizontal line. Such a graph can be completely determined by three
parameters. We will certainly want parameters for the critical age and
the maximal bone length. The third parameter could be either the y-
intercept (humerus length at hatching) or the slope (rate of growth).
Let

b = Age at which humerus growth stops

K = Maximal humerus length
a = Slope (rate of growth).
34 ■ Mathematical Modeling in Biology

Then assumption (A1a) can be translated into mathematics as the

model (Exercise 4)

a (x − b) + K x < b
f (x) = (2.1)
K x ≥ b.

Here, x is the age in days, f (x) is the length of the humerus in cm at

age x, and a, b, K > 0 are parameters.
The alternative assumption (A1b) is not specific enough to yield
a unique equation, because there are several classes of functions that
produce sigmoidal curves. In this example, we will assume a Janoschek
curve (Gille and Salomon 1995)
c
f (x) = K − (K − a)e−bx , (2.2)

where a, b, c, K > 0 are parameters.

The two deterministic models (2.1) and (2.2) are competing
hypotheses that we are setting forth to explain the humerus data in
Data Set 2.1. It is important to note that the parameters a and b have
different biological meanings in the two different models, whereas the
parameter K has the same interpretation in both models. Note also
that model (2.1) has three parameters, whereas model (2.2) has four
parameters. The number of parameters will be important during model
selection.

2.3.3 The Stochastic Model

We wish to explicitly model the stochasticity in the system based on
assumption (A2) about the source of the noise. Let F (x) be a random
variable denoting the measurement of the humerus length in a chick
of known age x. We can think of the random variable F (x) as the
deterministic prediction f (x) plus a random perturbation (noise):

F (x) = f (x) + noise. (2.3)

Think of equation (2.3) as the deterministic skeleton f (x) “clothed”

with noise. The deterministic skeleton of a stochastic model is the part
of the model that would remain if all the noise could be tuned to zero.
Typically, however, noise is not additive as in equation (2.3).
Usually, one must first transform the observational data and the
 Avian Bone Growth: A Case Study ■ 35

deterministic predictions with a variance-stabilizing transformation ϕ

under which noise becomes additive:
ϕ (F (x)) = ϕ (f (x)) + noise. (2.4)
Here, we mention an important point from statistical theory:
Demographic noise is approximately additive on the square root scale,
whereas environmental noise is approximately additive on the log scale
(Cushing et al. 2003). That is, if demographic noise is dominant, then
√
ϕ(·) = ·, and if environmental noise is dominant, then ϕ(·) = ln(·).
Thus, in our current example, under assumption (A2), equation
(2.4) becomes
p p
F (x) = f (x) + σε, (2.5)
where σ > 0 is a parameter representing the standard deviation of the
noise and ε is a standard normal random variable (a normal random
variable with mean zero and standard deviation one). Equation (2.5)
is the stochastic model for our current example.
Let the pair of numbers (x, lx ) be an actual data point, that is, a
measurement lx of the length of the humerus in a particular chick of
age x. Since we are thinking of equation (2.5) as a surrogate for the
real system, we can think of lx as a realization of the random variable
F (x). The residual error of the deterministic prediction is the actual
measurement minus the predicted value. The residuals corresponding
to equation (2.3) are the values of
res = lx − f (x).
The residuals corresponding to the more general equation (2.4) are the
values of
res = ϕ (lx ) − ϕ (f (x)) ,
and the residuals corresponding to our particular example, with
equation (2.5), are the values
p p
res = lx − f (x).
Note that any particular residual is a realization of the random
variable σε in the stochastic model (2.5). Thus, on the square root scale
the residuals are hypothesized to be normally distributed with mean zero
and standard deviation σ.
36 ■ Mathematical Modeling in Biology

2.4 MODEL PARAMETERIZATION

It might be difficult to estimate the maximal adult humerus length K
from chick data, because none or very few of the dead chicks might
have approached maximal humerus length. From the 13 dead adults,
however, we have explicit data on the maximal humerus length K. The
lengths of the 13 adult humeri are listed at the bottom of Data Set 2.1.
We cannot include these adult data in estimating parameters a, b, and
c, because the ages of the adult birds were unknown. However, we can
use the adult data to estimate directly a value for K, thus reducing
the number of model parameters by one. A statistically sophisticated
way to estimate K is found in Hayward et al. (2009); here, however,
we simply estimate the value of K as the sample mean of the adult
humerus data in Data Set 2.1:

K
b = 11.96923077.

The “hat” in K b indicates that this is the estimated value of K. We

will save quite a few decimals for the computations that follow, even
though they are not all significant. In general, only the results of final
computations should be rounded.
Our next goal is to use the juvenile data to estimate a, b, and c.

2.4.1 Dividing the Data Set

Which data are to be used for parameter estimation? This often re-
quires careful thought; the choice depends on the scientific question be-
ing asked, as well as how much and what kind of data exist. In general, if
the data set is fairly large and robust, it is best to set aside (randomly)
some of the data for purposes of independent model validation.
We will divide the juvenile data set in Data Set 2.1 into two parts,
one for model fitting (parameterization) and one for model evaluation
(validation). We will call the two data sets the estimation data set and
the validation data set. We want to divide the data randomly, but at
the same time, we would like to have representative data points from
a variety of ages in each data set. We therefore use a technique called
stratified random sampling in the following steps:

1. Chick ages in Data Set 2.1 range from x = 0 to x = 42

with integer values. Bin the juvenile humerus data into the age
 Avian Bone Growth: A Case Study ■ 37

intervals [0, 4] , [5, 9] , [10, 14] , . . . , [35, 39] , [40, 44] (Data Set 2.2).
Our goal is to randomly select half of the data from each bin.

2. Number the data points in each bin (Data Set 2.2).

3. Consider the first bin. It has 18 data points numbered 1 through

18. Use a random number generator to flip an “18-sided coin”
50 times, and write down the sequence of random numbers
generated. That is, use the computer to generate a list of 50
integers from 1 to 18 chosen from a uniform random distribution.
Repeat this step for each bin. Such a procedure yielded the
sequences of numbers in Data Set 2.3. (Obviously, if you carried
this out again, you would get a different set of sequences of
random numbers.)

4. Consider the random sequence for the first bin: 17, 12, 16, 2, 15,
8, 16, 13, 12, 6, 3, 3, . . . Put data point number 17 into data set I,
number 12 into data set II, number 16 into data set I, number 2
into data set II, number 15 into data set I, number 8 into data set
II, (skip over the 2nd occurrence of number 16), number 13 into
data set I, (skip over the 2nd occurrence of number 12), number
6 into data set II, etc. Skip over any random number that has
already occurred in the sequence. Do this for each bin. The data
are now divided into two sets (I and II as shown in Data Set 2.3).
Notice that if a bin had an odd number of data points, then data
set I will always have one more data point than data set II, which
is not a problem.

5. Flip a (two-sided) coin for Bin 1. “Heads” means data set I for
that bin goes into the estimation data; “tails” means it goes into
the validation data. Do this for each bin. Such a procedure yielded
the data sets in Data Set 2.4.

The data are now divided. The estimation and validation data are
graphed together in Figure 2.1. Make sure you can reconstruct Data
Set 2.4 from the original Data Set 2.1, given the sequences of random
numbers in Data Set 2.3 (Exercise 6).
Our next goal is to use the estimation data to estimate the
parameters (other than K). In what follows, we discuss two methods
for doing this.
38 ■ Mathematical Modeling in Biology

Figure 2.1: Estimation and validation data shown as length vs. age.

2.4.2 Maximum Likelihood (ML) Method

√ p
Our model assumes that the residual errors res = lx − f (x) are
normally distributed with mean zero and standard deviation σ. Thus,
the likelihood that a given data point (and hence a given residual)
occurred as a realization of our stochastic model (2.5) is

1 1 res 2
√ e− 2 ( σ ) .
σ 2π

This is the normal distribution PDF from Chapter 1. The likelihood

that the entire data set occurred as independent realizations of our
stochastic model is the product of the likelihoods over the whole data
set:
Y 1 1 res 2
L= √ e− 2 ( σ ) , (2.6)
data
σ 2π

where the symbol Π denotes the product operator. If there are q

residuals, then by the properties of exponents, equation (2.6) is
equivalent to (Exercise 7)
 q !
1 1 X 2
L= √ exp − 2 (res) , (2.7)
σ 2π 2σ
data
 Avian Bone Growth: A Case Study ■ 39

where exp y means ey .

We want to maximize the likelihood L that the whole data set occurs
as a set of independent realizations of our stochastic model. The only
things that can be adjusted
√ p on the right-hand side of equation (2.7) are
the residuals res = lx − f (x), which depend on the deterministic
model prediction f (x). The values of f (x) can be adjusted only by
changing the values of the model parameters. Thus, for model (2.1),
the likelihood L is a function of parameters a and b:
 q !
1 1 X 2
L (a, b) = √ exp − 2 (res) . (2.8)
σ 2π 2σ
data

Equation (2.8) is called the likelihood function. Note that for equation
(2.2), L is a function of three parameters, L(a, b, c). For simplicity,
we will continue to write L (a, b) in the explanation that follows. The
goal is to maximize L (a, b) as a function of the parameters a and b. 
We want to find the maximizer, that is, the pair of parameters b a, bb
that maximizes the function L (a, b) . These maximizing parameters are
called the maximum likelihood (ML) parameter estimates.
Note that a maximizer for L will also be a maximizer for ln L, and
vice versa (Exercise 8). The properties of logarithms allow us to write
the log-likelihood function as (Exercise 9):
q 1 X 2
ln L (a, b) = −q ln σ − ln (2π) − 2 (res) . (2.9)
2 2σ
data
 
Thus, the ML parameter estimate b
a, bb is the parameter vector that
maximizes ln L (a, b) .

2.4.3 Nonlinear Least Squares (LS) Method

Nonlinear least squares (LS) parameter estimates are obtained by
minimizing the residual sum of squares (RSS)
X 2
RSS(a, b) = (res) (2.10)
data
 
as a function of the parameters (a, b). The LS parameter estimate b
a, bb
is the parameter vector that minimizes RSS.
40 ■ Mathematical Modeling in Biology

Although it would be more accurate to say “sum of squared

residuals” than to say “residual sum of squares,” RSS is the traditional
notation for equation (2.10).
Note that in our current example having one state variable, in
which noise is Gaussian (normally distributed) with mean zero and
constant variance σ 2 and the residuals are independent, maximizing the
log-likelihood (2.9) is equivalent to minimizing the RSS (2.10). (See
Exercise 10.)
One nice thing to know about LS, however, is that it loosens
the restrictive assumptions on the residuals; LS parameter estimates
converge to the true values even if the noise is non-normal and
autocorrelated, as long as the noise has a stationary distribution
(Cushing et al. 2003; Tong 1990).
Once a and b are estimated, the variance of the residuals is
estimated by
RSS
d
b2 =
σ ,
q
where RSS
d denotes the fitted value of RSS.

2.4.4 Downhill Minimization Routine: Nelder-Mead Algorithm

Typically, one cannot maximize the log-likelihood or minimize RSS
analytically; it must be done numerically on a computer. For equation
(2.1), think of RSS(a, b) as a surface suspended over parameter space.
In this case, parameter space is the horizontal plane spanned by the
a-axis and b-axis. The RSS-axis rises vertically out of the plane. Each
point on the plane corresponds to a parameter pair (a, b), and the
value of RSS(a, b) is plotted above each point on the plane,
 generating
a surface. We want to locate the minimizer point(s) b a, b on the plane
b
at which the surface attains a minimum value.
The Nelder-Mead algorithm begins with an initial “guess” (a0 , b0 )
in parameter space and then systematically checks around nearby in
parameter space to see if there is a lower point on the surface. In this
way, the routine “walks downhill” along the RSS surface and converges
on a local minimum. It is important to remember that a surface can
have more than one local minimum. There is no foolproof way of making
sure you have found a global minimum. Always try a variety of different
initial guesses to see if the Nelder-Mead algorithm always converges to
the same minimum.
 Avian Bone Growth: A Case Study ■ 41

2.4.5 Implementing Parameterization in Code

A code to estimate model parameters requires three basic parts: a main
program, a downhill search function, and a program that computes the
RSS.
The main program is a “front end” in which the user sets the initial
“guess” for the parameter vector. This program passes the initial vector
and the name of the RSS subroutine to the downhill search algorithm
(which attempts to converge on a minimizer for RSS) and, finally,
outputs the minimizer (parameter estimates) that it obtained from
the downhill algorithm.
The second part is the downhill search program. This is not
something you need to code yourself; it is likely in the library of
functions of the language you are using. For example, in MATLAB
the downhill search function is called fminsearch. When the main code
passes the initial parameter vector and the name of the RSS subroutine
to this search function, it uses those initial parameters to compute RSS
via the RSS subroutine. It then chooses a suite of nearby parameters,
computes RSS for them, and uses the parameters associated with the
smallest RSS as the next “guess.” It continues this process of “walking
downhill” until it finds a local minimum value of RSS. Finally, it returns
the minimizer vector of parameters to the main code.
The third part is the subroutine that computes the RSS as a
function of the parameters that are passed to it by the search algorithm.
This is where the model equations are coded, predictions and residuals
are computed, and the RSS value is computed.
Most downhill search functions search for negative as well as
positive minimizers. In population models, however, parameters are
usually positive and we do not want to search negative parameter space.
One way to address this is to pass the initial parameter vector v to the
downhill routine as the vector ln v so that it can search all of parameter
space. This works because the log function maps the interval (0, ∞) to
the interval (−∞, ∞). In the RSS routine, the vector ln v must be
converted back to v via exponentiation before the parameters are used
to compute model predictions. Also, when the final parameter estimates
are returned to the main program from the search algorithm, they are
returned as ln vb, and so they must be exponentiated in order to recover
the parameter estimates vb on the correct scale.
42 ■ Mathematical Modeling in Biology

TABLE 2.1 LS Parameters for Equations

(2.1) and (2.2) Estimated from the
Estimation Data on the Square Root
Scale

Model a
b b
b cb
A1a 0.2475 37.92 N/A
A1b 2.872 0.005402 1.661

The best way to understand all this is to work through sample code.
For codes similar to the ones you will write in the exercises concerning
the current example, see Appendix C.

2.4.6 Results of Parameterization

The LS parameters for models (2.1) and (2.2), estimated from the
estimation data on the square root scale under the assumption that
Kb = 11.96923077, are shown in Table 2.1 to four significant figures
(Exercise 11).
The first question we should ask is: Do these parameters make
biological sense (Exercise 12)? Note that if one or more parameters
persist in wandering off to zero or infinity during the parameterization
process no matter what starting parameter values you choose, then the
model is probably flawed and should be revised.
The predictions of both models at their LS parameter values are
shown with the estimation data in Figure 2.2a and b. Which of the two
alternative models best fits the data? Is it clear visually, or do we need
a quantitative method to determine which model is best? We turn to
this question in the next section.

2.5 MODEL SELECTION

Which of the two alternative models, (2.1) or (2.2), best describes the
data? The two models have different numbers of parameters. A curve
with many parameters is more flexible and easier to fit to data than
one with few parameters. Indeed, an overparameterized model can fit
just about any data, but has little explanatory power. Thus, the model
with more parameters should be penalized for overfitting. To do this,
 Avian Bone Growth: A Case Study ■ 43

(a)

(b)

(c)

Figure 2.2: Model predictions. (a) Model (A1a) as fitted to the

estimation data. (b) Model (A1b) as fitted to the estimation data.
(c) Model (A1a) as compared to the validation data without refitting.

researchers often use a tool from information theory called the Akaike
information criterion (AIC) to select the best model (Burnham and
Anderson 2002). The AIC is defined as
 
AIC = − ln L
b + 2κ, (2.11)
44 ■ Mathematical Modeling in Biology

where Lb is the maximized value for the likelihood function and κ is the
number of estimated parameters in the model. Model comparison is
based on the rank of the AIC values for the suite of alternative models.
The least AIC indicates the best model. The first term in equation
(2.11) is smallest for the model with the largest likelihood, but this is
discounted by the second term, which is largest for the model with the
most parameters.
In our current example, the residuals are Gaussian with mean zero
and constant variance σ 2 . In such a situation, it is possible to show that
ranking the AIC values (2.11) is equivalent to ranking the values of

AIC* = q ln RSS
d + 2κ, (2.12)

where q is the number of residuals, RSS

d is the fitted value of RSS, and κ
is the number of parameters being estimated, including σ (Exercise 18).
Note that if all of the alternative models have the same number of
parameters κ, then selecting the model with minimum AIC is equivalent
to selecting the model with the maximum likelihood (for ML estimation)
or the minimum RSS (for LS estimation), in which case there is no
reason to compute the AIC.
In our current example, using equation (2.12), we obtain the AIC
values shown in Table 2.2 (Exercise 16). The least AIC belongs to
model (2.1). Thus, we discard model (2.2) in favor of model (2.1). Our
deterministic model is therefore

a (x − b) + K x < b
f (x) =
K x≥b
a = 0.2475 (2.13)
b = 37.92
K = 11.97,

TABLE 2.2 AIC Values Obtained from Equation (2.12)

Model q b2
σ κ AIC* ∆ AIC
A1a 40 0.005266 3 −203.9 0.0000
A1b 40 0.006106 4 −195.9 8.000
 Avian Bone Growth: A Case Study ■ 45

and our demographic stochastic model is

 p
p a (x√− b) + K x < b
F (x) = σε + (2.14)
K x≥b
a = 0.2475
b = 37.92
K = 11.97
√
σ = 0.005266.

2.6 MODEL VALIDATION

To validate the selected model (2.13), we must evaluate how well
it explains the validation data, without reparameterizing. We will
compute a number called the goodness of fit on both the estimation
data set and the validation data set. If the model fits the validation
data as well as the estimation data, then we will consider the model
validated.
One measure of goodness of fit is the generalized R-squared value,
denoted R2 . Consider a data set with members (x, lx ). Let lx be the
sample mean of the data set. Then the R-squared value (when data
are not transformed by ϕ) is defined to be
P 2
2 data (lx − f (x))
R =1− P
2 .
data x l − lx
P 2
The residual sum of squares data (lx − f (x)) is in the numerator;
it measures the variability that is unexplained by the model. The
P
2
denominator data lx − lx measures the variability around the
mean of the data. Thus, the quotient of these two numbers measures
the fraction of variability not explained by the model, relative to the
use of the mean as a predictor. One minus the quotient measures the
fraction of variability that is explained by the model.
In general, for transformed data, the R-squared is
P 2
2 (ϕ (lx ) − ϕ (f (x)))
R = 1 − Pdata  2 ,
data ϕ (lx ) − ϕ (lx )
46 ■ Mathematical Modeling in Biology

TABLE 2.3 R2 Values for Model (2.13) as

Fitted to the Estimation Data and Computed
on the Validation Data without Refitting

Estimation Data Validation Data

2
R 0.9838 0.9813

where ϕ (lx ) denotes the mean of the transformed data. For our current
example, we have
P √ p 2
data lx − f (x)
R2 = 1 − P √ √ 2 ,
data lx − lx
√
where lx denotes the mean of the square roots of the humerus length
measurements.
Note that it is always true that R2 ≤ 1. The closer R2 is to one,
the better the model fit. Although R2 is usually between zero and one,
it can also be negative (Exercise 15).
The R2 values for model (2.13) on the two data sets are given
in Table 2.3 (Exercise 17). Successful validation is supported because
the R2 values are about the same. The predictions of model (2.13) are
shown in Figure 2.2a and c with the estimation and the validation data,
respectively. Visually, note that the model appears to fit the validation
data about as well as it fits the estimation data set upon which it was
parameterized.

2.7 EXERCISES
1. Work through the self-guided linear algebra tutorial in Appendix
A.

2. Work through the self-guided programming tutorial in Appendix

B.

3. Some questions that seem to make sense are actually nonsensical.

Explain why each of the following questions is nonsensical.

a. What is the basal metabolism rate of a jackalope?

b. What proportion of the mass of iron is due to phlogiston?
 Avian Bone Growth: A Case Study ■ 47

c. How many minutes are there in a kilometer?

d. Let A be the set of all sets which do not have themselves as
a member, that is, A = {B | B ∈ / B } . Is A ∈ A?

4. Given only the verbal description in assumption (A1a), how

would you write down a mathematical model equivalent to that
description? In other words, derive model (2.1) from assumption
(A1a). Explain each step as you go, and explain why the model
has to have at least three parameters.

5. Another well-known sigmoidal curve is the Holling type III curve,

which has the form

M x2
y= ; x ∈ [0, ∞); a, M > 0. (2.15)
a2+ x2

Use calculus to do the following problems.

a. Show that y(0) = 0 and limx→∞ y(x) = M.

b. Show that y(x) is an increasing function of x.
c. Where is y concave up? Concave down? Find all inflection
points.
d. Find the value of x at which the curve reaches one half of its
asymptotic limit. This is called the half-saturation constant.
e. Use the information gained above to graph y vs x.
f. The Holling type III curve cannot describe the length of
the humerus, because the humerus has a positive length b
upon hatching; that is, f (0) = b. We want to translate the
Holling type III curve vertically so that the initial value is
b instead of zero. Also, we want the saturation level to be
K. From equation (2.15), derive the following modification
of the Holling III model:

(K − b) x2
f (x) = + b.
a2 + x2

Explain each step as you go.

48 ■ Mathematical Modeling in Biology

6. Make sure you can reconstruct Data Set 2.4 from the original
Data Set 2.1, given the sequences of random numbers in Data
Set 2.3. You don’t need to show any work for this problem, but
indicate whether or not you were successful in reproducing Data
Set 2.4.

7. Show that equation (2.6) can be written as

 q
1 1
P 2
L= √ e− 2σ2 data (res) ,
σ 2π
where q is the number of residuals.

8. Let g(z) be any positive-valued function that is twice differen-

tiable. Use the “derivative tests” from calculus to prove that a
maximizer for g(z) will also be a maximizer for ln g(z), and vice
versa. Can you drop the “twice differentiable” hypothesis and
still prove the result?

9. Derive the log-likelihood function (2.9) from the likelihood

function (2.8).

10. Explain why the log-likelihood in equation (2.9) is maximized

when the residual sum of squares (equation 2.10) is minimized.

11. Reproduce the LS parameter estimates in Table 2.1. Attach your

programs, input files, and screenshots of your output.

12. Can you determine the biological meanings of the parameters in

each of models (2.1) and (2.2)? Are the LS parameter estimates in
Table 2.1 for the humerus model (2.13) biologically reasonable?
Explain. Hint: It may take some work to thoroughly understand
the biological meanings of parameters b and c in the Janoschek
model (2.2). Reading the 1995 paper by Gille and Salomon (1995)
will help.

13. Stochastic models are important for two reasons. First, they
are necessary for parameterization because they specify how the
residuals are distributed. Second, they can be used to simulate
the real (noisy) system. Write a program to produce a simulated
data set of humerus lengths using the stochastic model (2.14).
 Avian Bone Growth: A Case Study ■ 49

For each of the ages 0, 1, 2, . . . , 50, generate 10 simulated data

points. Present the output as a scatter plot of length against age.
Attach your program and the output graph.

14. What does R2 = 1 mean? What does R2 = 0 mean? Can R2 be

greater than one? Explain.

15. Consider the number of seals hauled out on a beach at hour t.

Suppose a wildlife refuge biologist models the number of seals
N (t) on the beach at hour t with the equation

N (t) = 3t − 7.

The biologist then collects 10 hours of data:

Time t Seals Model Prediction N (t)
5 10
6 9
7 11
8 10
9 10
10 9
11 8
12 10
13 11
14 12

a. Fill in the model predictions for each hour.

b. Compute R2 assuming demographic noise.
c. Under what mathematical conditions is R2 negative?
d. What does a negative R2 mean to a scientist?

16. Reproduce all the values in Table 2.2. Attach your programs,
input files, and screen shots of your output.

17. Reproduce the R2 values in Table 2.3. Attach your programs,

input files, and screen shots of your output.
50 ■ Mathematical Modeling in Biology

18. If the residuals for each of the alternative models are independent
and Gaussian with mean zero and constant variance, prove that
ranking the AIC values from equation (2.11) is equivalent to
ranking the values of

AIC* = q ln RSS
d + 2κ,

where q is the number of residuals, RSS

d is the fitted value of RSS,
and κ is the number of parameters being estimated, including σ.

19. Project: Model the growth of the ulna in glaucous-winged gulls.

That is, repeat the comprehensive analysis in this chapter for
the ulna length data in Data Set 2.1. Take your two alternative
deterministic models to be the modified linear model

a (x − b) + K x < b
f (x) =
K x≥b

and the modified Holling type III model

(K − b) x2
f (x) = + b.
a2 + x2
Use the same binning procedure as in Data Set 2.2, and use the
same sequences of random numbers in Data Set 2.3, but work
through all the details. Present your work in a complete, precise,
and organized fashion.

20. Project: We might suspect that the humerus in C chicks grows

more slowly than in A and B chicks. Given the work already done
in this chapter, we can now test this hypothesis. The idea is to
fit model (2.1) to A and B chick data only and then see if it also
fits the C chick data without reparameterizing.

a. Divide the humerus data in Data Set 2.1 into two sets: data
set AB (consisting of all the A and B chick data) and data
set C (consisting of all the C chick data).
b. Estimate the LS parameters for model (2.1) on the AB data
set. Record the parameter values and the RSS. Compute σb2
2
and R .
 Avian Bone Growth: A Case Study ■ 51

c. Without re-estimating the parameters, compute the R2 on

the independent data set C.
d. What is your conclusion? Does the humerus grow differently
in C chicks?
e. Present your work in a short scientific paper. Your paper
should have standard sections including title page, abstract,
introduction, methods, results, discussion, conclusion, and
references, as well as tables and graphs. Include your code
at the end of the paper in an appendix. Ask your instructor
which journal style you should follow in the preparation of
the manuscript.

BIBLIOGRAPHY
Burnham, K. P. and Anderson, D. R. 2002. Model Selection and Multi-Model
Inference: A Practical Information-Theoretic Approach, 2nd ed. Springer,
New York. [This is a practical, user-friendly book that shows scientists
how to make valid inferences from empirical data within an information-
theoretic framework.]

Cushing, J. M., Costantino, R. F., Dennis, B., Desharnais, R. A., and

Henson, S. M. 2003. Chaos in Ecology: Experimental Nonlinear Dynamics.
Academic Press, San Diego, CA. [In Chapter 6 we will draw heavily on
this book to showcase how dynamic population data can be connected to
mathematical models.]

Gille U. and Salomon F. V. 1995. Bone growth in ducks through mathematical

models with special reference to the Janoschek growth curve. Growth,
Development and Aging 59:207–214. [The authors compare five alternative
models for bone (weight) growth in white Pekin ducks: Janoschek,
Richards, Bertalanffy, Gompertz, and Logistic.]

Hayward, J. L., Henson, S. M., Banks, J. C., and Lyn, S. L. 2009.

Mathematical modeling of appendicular bone growth in glaucous-winged
gulls. Journal of Morphology 270:70–82. DOI: 10.1002/jmor.10669. [This
is the paper on which Chapter 2 is based. Both theoretical and empirical:
models connected to data.]

Tong, H. 1990. Non-Linear Time Series: A Dynamical System Approach.

Oxford University Press, Oxford. [Valuable introduction to the analysis
of nonlinear time series.]
 II
Discrete-Time Models

53
 CHAPTER 3

Discrete-Time Maps

3.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

Think about almost any area of biology—physiology, genetics,
development, ecology, or evolution. Could we study this area without
considering time? Certainly not. Look at any biology text or journal.
How much could you learn if you ignored all the time-based graphs?
Not much. Life happens in time. Biologists, then, concern themselves
with many time-based, or temporal, aspects of life.
Not surprisingly, mathematical modeling in biology often involves
equations that describe and predict temporal processes. Some
biological processes, like bone growth, occur continuously in time.
Others, like population growth in annual plants, occur more discretely
in time. Even though data collected from continuous processes are by
necessity taken at discrete time intervals, model predictions based on
these data assume temporal continuity and are commonly written in
the form of differential equations. By contrast, discrete-time processes
are modeled using difference equations, or maps, equations that take
into account the discontinuous nature of these processes.
In this chapter, we focus on difference equations. Although
difference equations themselves do not involve derivatives, the tools
you need in order to study them are from first-semester calculus.

3.2 COMPARTMENTAL MODELS

Let xt be the state of a system—say, the size of a population—at time t.
Think of the population as a compartment having inflows and outflows.

DOI: 10.1201/9781003265382-5 55
56 ■ Mathematical Modeling in Biology

Inflows are due to births and immigration. Outflows are due to deaths,
emigration, harvesting, etc. (Figure 3.1).
In general, compartmental models state that the net rate of change
of the quantity in a compartment is equal to the sum of its inflow rates
minus the sum of its outflow rates:
X X
Net rate of change = Inflow rates − Outflow rates.

For continuous-time processes, the net rate of change of a quantity

x(t) is the derivative dx/dt. What is the analogous concept for a
discrete-time process? In mathematics, the uppercase Greek letter delta
(∆) usually stands for “change.” We define ∆xt+1 to be the amount x
changes from time t to time t + 1. That is,

∆xt+1 = xt+1 − xt .

Now, in general, ∆xt+1 is not a rate of change; it is simply an amount

of change. However, in the context of discrete time steps, ∆xt+1 is the
rate of change per unit time step.
Thus, a discrete-time compartmental model has the form
X X
∆xt+1 = (Inflow rates at time t) − (Outflow rates at time t) ,
(3.1)
where each flow rate on the right-hand side of equation (3.1) is replaced
by a mathematical expression based on the modeling assumptions.
In discrete-time modeling, the time step is chosen to be appropriate
for the application. Discrete-time models can be particularly useful if
there is a temporal “pulse” in the biological system, in which case
you would choose the time step to equal the time between pulses.
For example, suppose you were studying an annual plant species that

Figure 3.1: Flows into and out of a population may be due to births,
immigration, deaths, and/or emigration.
 Discrete-Time Maps ■ 57

produced seeds each autumn, and you wanted to track the population
size. The model time step might be one year, measured in the autumn
after the seeds were produced in order to take advantage of the natural
reproductive pulse in the system.

3.3 LINEAR MAPS

3.3.1 Malthusian Growth
We illustrate linear maps with the following problem. Let xt be the
number of individuals in a population at time t. Assume
(A1) The average per capita fecundity (birth rate) is constant and
positive.

(A2) Generations are nonoverlapping. All deaths occur at the end

of the generational time period.

(A3) The only flows into or out of the population are due to births
and deaths.
We need a parameter for the per capita birth rate. Let

b = Number of births per individual per unit time.

Since the per capita birth rate is b, the total population birth rate will
be b times the number of individuals in the population:

bxt = Number of births in the population per unit time at time t.

The population death rate is exactly xt individuals per unit time

because all individuals die at the end of each time step. Following
equation (3.1), we can write

∆xt+1 = Birth rate at time t − Death rate at time t (3.2)

= bxt − xt .

By definition of ∆xt+1 , equation (3.2) becomes

xt+1 − xt = bxt − xt

or simply
xt+1 = bxt . (3.3)
58 ■ Mathematical Modeling in Biology

That is, the population size at time t + 1 is the per capita birth rate b
times the population size xt at time t.
Suppose the number of individuals at time t = 0 is x0 = p ≥ 0.
The statement x0 = p is called the initial value or initial condition for
equation (3.3). The population size is therefore modeled by the initial
value problem corresponding to assumptions (A1)–(A3):

xt+1 = bxt
x0 = p.

The difference equation xt+1 = bxt is called linear because the state
variable x appears linearly in the equation.
Suppose a petri dish culture initially consists of ten cells, and that
each cell divides into two daughter cells every hour. Then p = 10 and
b = 2, so the culture can be modeled by the initial value problem

xt+1 = 2xt
x0 = 10,

where the time step is one hour. Note that

x0 = 10 = 20 (10)
x1 = 2x0 = 21 (10)
x2 = 2x1 = 2 [2 (10)] = 22 (10)
x3 = 2 (2 [2 (10)]) = 23 (10)
..
.
xt = 2t (10) .

The function xt = 10 (2t ) is called the closed-form solution of the initial

value problem. “Closed form” means that, given any value of t, you
can compute the population size xt without iterating. Here, we see an
example of the fact that linear dynamical systems have exponentially
growing solutions. Systems that grow exponentially are also said to
grow geometrically, or to exhibit Malthusian growth.
In general, the solution of the initial value problem

xt+1 = bxt (3.4)

x0 = p
 Discrete-Time Maps ■ 59

is (Exercise 4)
xt = pbt . (3.5)
Note that the extinction state x = 0 is an equilibrium of model (3.4),
since it satisfies the equilibrium equation

xe = bxe .

That is, if the initial condition is x0 = p = 0, then solution (3.5) is the

constant solution xt = 0 for all t.
Now suppose p > 0. In this case, the behavior of solution (3.5)
depends on the value of b. If b > 1, the population grows exponentially
as bt grows without bound. Thus, the equilibrium state xe = 0 is
unstable. If b = 1, the population remains at its initial size p for all
time. Every value of x is an equilibrium solution in this case, and all
the equilibria are neutrally stable. If 0 < b < 1, the population declines
exponentially toward zero as bt declines, and so the equilibrium xe = 0
is asymptotically stable.
You can see that the most important expression in model (3.4)
is the number b; the value of b determines the long-term fate of the
population. The parameter b is called the intrinsic growth rate. It is
also called the eigenvalue.
Let xt be the number of individuals in a population at time t.
Assume

(A1) The average per capita birth rate is a constant b > 0 offspring
per individual per unit time.

(A2) Generations are overlapping.

(A3) The average per capita death rate is a constant d deaths per
individual per unit time, with 0 < d < 1.

(A4) The only flows into or out of the population are due to births
and deaths.

Then the population birth rate is bxt offspring per unit time and
the population death rate is dxt deaths per unit time. Thus,

∆xt+1 = Birth rate at time t − Death rate at time t

= bxt − dxt
60 ■ Mathematical Modeling in Biology

or
xt+1 − xt = bxt − dxt .
We can write the model in several equivalent forms to aid in various
interpretations. For example, we can write

xt+1 = xt + bxt − dxt

New census = Last census + Births − Deaths.

We can also write

xt+1 = bxt + (1 − d) xt
New census = Recruits + Survivors.

The term bxt is called the recruitment term. The number 1 − d is

called the survivorship, the probability that an individual will survive
one unit of time. We can also write the model as

xt+1 = (b + 1 − d) xt ,

that is,
xt+1 = rxt ,
where we define the new parameter r to be r = b + 1 − d. Note that
r > 0 since b > 0 and d < 1. This example is completed in Exercise 5.

3.4 NONLINEAR MAPS

What if a lack of resources due to overcrowding were to affect the birth
rate? Let xt be the size of a population at time t. Assume

(A1) The average per capita birth rate is a constant b > 0 offspring
per individual per unit time at low populations sizes (i.e., if there
are no crowding effects).

(A2) Because of crowding effects, the average per capita birth rate
is reduced by the factor e−cxt , where c > 0 quantifies the strength
of the crowding effect. (Note that 0 < e−cxt ≤ 1.)

(A3) Generations are nonoverlapping.

 Discrete-Time Maps ■ 61

By assumptions (A1) and (A2), the per capita birth rate in the
presence of xt individuals is be−cxt ; therefore, the population birth
rate is be−cxt xt offspring per unit time, and the model is
xt+1 = bxt e−cxt (3.6)
x0 = p
with parameters b, c > 0. Model (3.6) is the famous Ricker model ,
historically used in fisheries (Ricker 1954; Ricker 1975). It is called
nonlinear because the state variable x appears in a nonlinear way. This
simple deterministic model can have incredibly complex dynamics, as
we shall see in Chapter 4.
Let’s find the equilibrium states of the Ricker model. The
equilibrium equation is
xe = bxe e−cxe .
̸ 0. If we divide
Note that xe = 0 is a equilibrium solution. Suppose xe =
both sides by xe , we have
1 = be−cxe ,
which yields a nontrivial equilibrium solution xe = (ln b) /c.
It is often useful to graph the equilibria as a function of one of the
parameters. In this case, we graph the two equilibria xe = 0, (ln b) /c as
a function of b (Figure 3.2). Note that the nontrivial equilibrium xe =
(ln b) /c is positive if and only if b > 1. The value b = 1 at which the two
equilibrium branches cross is called a bifurcation value or bifurcation
point. The graph of xe vs b (Figure 3.2) is called a bifurcation diagram,
and b is called the bifurcation parameter.
A model such as xt+1 = bxt e−cxt is really a collection of infinitely
many models, one for each specific pair of values of b and c. When
looking at a bifurcation diagram such as Figure 3.2, one must remember
that any particular system has a fixed value of b, for example, b =
1.5. Its equilibria are given by the values of the equilibrium branches
directly above that specific value of b. Another way to say it is that
any particular system “lives on a vertical line” in Figure 3.2.
What are the stabilities of the two equilibria xe = 0, (ln b) /c of the
Ricker map? Do their stabilities depend on the value of b? Clearly, we
need to find a way to quantify and study stability in nonlinear maps.
To this end, we now turn to one of the most important subjects in
applied mathematics, the subject of linearization.
62 ■ Mathematical Modeling in Biology

Figure 3.2: Bifurcation diagram for the Ricker map shows the
equilibria xe as a function of the parameter b. The two equilibrium
branches are xe = 0 and xe = (ln b) /c. Here, c = 1.

3.5 LINEARIZATION
3.5.1 Linearization of Functions
Suppose f (x) is a function whose graph passes through the point
(a, f (a)). Close to the point (a, f (a)), we can approximate the graph of
f with its tangent line, that is, the line that is tangent to the graph of
f at the point (a, f (a)). What is the equation of this line? Recall that
the slope m of a line is the “rise over run.” Given a nearby point (x, y)
on the line (Figure 3.3), the rise from (a, f (a)) to (x, y) is y − f (a) ,
while the run is x − a. The slope is therefore
y − f (a)
m= ,
x−a
and so
y − f (a) = m (x − a) ,
which is the “point-slope” formula for a line from high school algebra.
From calculus, we know that the slope m of the line tangent to f
at the point (a, f (a)) is the derivative of f evaluated at x = a:
df
m= (a) .
dx
 Discrete-Time Maps ■ 63

Figure 3.3: The equation of the tangent line at point (a, f (a)) is y =
f (a) + f ′ (a)(x − a).

Thus, the equation of the tangent line is

 
df
y − f (a) = (a) (x − a) ,
dx
or  
df
y = f (a) + (a) (x − a) . (3.7)
dx
Equation (3.7) is called the linearization of f at x = a. It is also the
equation of the tangent line.
Because the linearization is a good approximation of the function
f near the point x = a, we have
 
df
f (x) ≈ f (a) + (a) (x − a) for x ≈ a.
dx

Note that you always linearize a function about a point. It doesn’t

make sense to ask what the linearization of a function is without
reference to a specific point x = a. Furthermore, the linearization of a
function is always a linear equation of the form y = mx + b, where m
and b are constants.
Linearization is one of the most important tools in applied math-
ematics and is extremely important in modeling. In fact, linearization
provides the key to studying stability, as we shall soon see.
64 ■ Mathematical Modeling in Biology

Let’s find the linearization of the curve f (x) = x3 − 4x2 + 7x − 1

at x = 2. In this problem, we have a = 2. The derivative of f is

df
= 3x2 − 8x + 7.
dx
Thus,
2
f (a) = f (2) = 23 − 4 (2) + 7 (2) − 1 = 5
and
df df 2
(a) = (2) = 3 (2) − 8 (2) + 7 = 3.
dx dx
Hence, the linearization of f at x = 2 is
 
df
y = f (2) + (2) (x − 2)
dx
= 5 + 3 (x − 2)
= 3x − 1,

and therefore, we have the linear approximation

f (x) ≈ 3x − 1 for x ≈ 2,

that is,
x3 − 4x2 + 7x − 1 ≈ 3x − 1 for x ≈ 2,
which is a neat result.

3.5.2 Linearization of Discrete-Time Maps

Consider the nonlinear discrete dynamical system

xt+1 = f (xt ) (3.8)

with equilibrium xe . Note that the constant xe is an equilibrium of the

difference equation xt+1 = f (xt ) if and only it satisfies the equilibrium
equation xe = f (xe ) . When we say equation (3.8) is nonlinear , we
mean that the right-hand side f is a nonlinear function of xt . An
example of such an equation would be

xt+1 = ax2t .
 Discrete-Time Maps ■ 65

What we would like to do is to replace the nonlinear f (xt ) on the

right-hand side of equation (3.8) with a simple linear function that
approximates it for values of xt ≈ xe .
In the previous section, we learned that
 
df
f (x) ≈ f (a) + (a) (x − a) for x ≈ a.
dx

In the current context, x is xt , and a is xe . Making those substitutions,

we have
 
df
f (xt ) ≈ f (xe ) + (xe ) (xt − xe ) for xt ≈ xe .
dx

Also, since xe is an equilibrium of equation (3.8), we know that xe =

f (xe ) .
Thus, we can write
 
df
f (xt ) ≈ xe + (xe ) (xt − xe ) for xt ≈ xe .
dx

It is traditional to define the Greek letter lambda (λ) to be the

derivative df /dx evaluated at xe :
df
λ= (xe ) .
dx
Note that λ is simply a constant, a number that you can compute. It
is called the eigenvalue.
Then

xt+1 = f (xt )
≈ xe + λ (xt − xe ) for xt ≈ xe ,

and hence
xt+1 − xe ≈ λ (xt − xe ) for xt ≈ xe . (3.9)
We can restate equation (3.9) in words: As long as the system is in
the neighborhood of the equilibrium, the displacement of the system
state from equilibrium at time t + 1 is approximately the number λ
times the displacement of the system from equilibrium at time t. That
is, xt+1 is about λ times as far from the equilibrium as xt was. You can
66 ■ Mathematical Modeling in Biology

see that if −1 < λ < 1, this is good news for the stability of equilibrium
xe , because the distance between the system state and the equilibri-
um is shrinking as time goes on—in fact, it appears that the equilibrium
xe would be asymptotically stable. If λ > 1 or λ < −1, however, the
distance between the system state and the equilibrium grows whenever
the system is near the equilibrium, and so it appears that xe would be
unstable.
We now clean up the notation a bit and state these observations in
a theorem. Define the displacement, or variation, of the system state
from equilibrium to be
z t = xt − x e . (3.10)
Using this as a change of variables, we can rewrite equation (3.9) as

zt+1 ≈ λzt for zt ≈ 0.

Note that the change of variables (equation 3.10) simply shifts the
equilibrium to zero.
The equation
zt+1 = λzt for zt ≈ 0
is called the variation equation. The variation equation approximates
the dynamic change in displacement z when the system is close to its
equilibrium. Note that the variation equation is linear, and its solution
is
zt = z0 λt .
Thus, the displacement of the population from equilibrium grows or
decays exponentially when the population size is near its equilibrium
value. Whether the displacement grows or decays depends on whether
the eigenvalue λ is greater than or less than one in absolute value, that
is, whether |λ| > 1 or |λ| < 1.

Definition 3.1 The linearization of xt+1 = f (xt ) at the

equilibrium xe is the linear map

zt+1 = λzt ,

where the eigenvalue λ is given by

df
λ= (xe ) .
dx
 Discrete-Time Maps ■ 67

Note that you must always linearize a difference equation about a

specific equilibrium. It does not make sense to “find the linearization”
without reference to a particular equilibrium. Furthermore, the
linearization of a nonlinear difference equation is always a linear
difference equation of the form zt+1 = λzt , where λ is a number.

Definition 3.2 An equilibrium xe of xt+1 = f (xt ) is hyperbolic if

̸ 1.
and only if |λ| =

Theorem 3.1 (Linearization Theorem) Let xt+1 = f (xt ) be a

nonlinear map with hyperbolic equilibrium xe . Suppose the function
f is continuously differentiable in x (that is, df /dx exists and is
continuous). Then

|λ| < 1 =⇒ xe is asymptotically stable

|λ| > 1 =⇒ xe is unstable,

where
df
λ= (xe ) .
dx

Note that the nonhyperbolic case |λ| = 1 is not covered by the

linearization theorem. If |λ| = 1, the theorem simply does not apply,
and no conclusion can be drawn from the theorem.

3.5.3 Linearizing the Ricker Map

Consider again the Ricker map

xt+1 = bxt e−cxt ,

where b, c > 0. Here, f (x) = bxe−cx . The equilibria are

ln b
xe = 0 and xe = .
c
Let’s linearize the Ricker map about each of its equilibria. In both
cases, we will need the derivative of f with respect to x:

df
= be−cx (1 − cx) .
dx
68 ■ Mathematical Modeling in Biology

We first linearize about xe = 0. Now,

df
λ= (0) = be0 (1 − 0) = b,
dx
and so the linearization at xe = 0 is

zt+1 = bzt .

Recalling that b > 0, we have

0 < b < 1 =⇒ xe = 0 is asymptotically stable

b > 1 =⇒ xe = 0 is unstable.
ln b
Now we linearize about the nontrivial equilibrium xe = c :
 
df ln b
λ =
dx c
  
= be−c(
ln b
c ) 1 − c ln b
c
−1
= beln b (1 − ln b)
−1
= bb (1 − ln b)
= 1 − ln b.
ln b
The linearization at xe = c is therefore

zt+1 = (1 − ln b) zt ,

and so
ln b
|1 − ln b| < 1 =⇒ xe = is asymptotically stable (3.11)
c
ln b
|1 − ln b| > 1 =⇒ xe = is unstable.
c
In order to obtain simple conditions on b for stability, we need to do
the algebra necessary to remove the absolute values. Recall that

|x| < p if and only if − p < x < p (3.12)

|x| > p if and only if x < −p or x > p.
 Discrete-Time Maps ■ 69

From inequalities (3.11) and (3.12), we obtain (Exercise 1):

ln b
1 < b < e2 =⇒ xe = is stable (3.13)
c
ln b
0 < b < 1 =⇒ xe = is unstable
c
ln b
b > e2 =⇒ xe = is unstable.
c

We can now indicate stability on our bifurcation diagram

(Figure 3.4).
Note that if b > e2 , both equilibria are unstable! How do the
solutions of the Ricker map behave for these values of b? Toward
what value do they tend, if any? What is the long-term fate of the
system? We will investigate this fascinating question in Chapter 4.
Before moving on, however, let’s look more carefully at the Ricker
nonlinearity. This type of nonlinearity shows up a lot in models of
population dynamics. We might ask: Is it simply phenomenological
and descriptive, or is there a mechanistic underpinning?

Figure 3.4: Bifurcation diagram for the Ricker map showing stability
of equilibria. Dotted equilibrium branches are unstable (u); solid are
stable (s). Here, c = 1.
70 ■ Mathematical Modeling in Biology

3.6 THE RICKER NONLINEARITY

The Ricker nonlinearity e−cx is typically a result of random deleterious
contacts between individuals in a population. For example, in fisheries
or insect models, the Ricker function can result from random
cannibalistic encounters between individuals within a population.
Suppose, for example, that we are modeling an insect population
with J juveniles and A adults using a discrete-time model. Suppose
the adults cannibalize juveniles through random encounters. Let ∆t
be a small fraction of one model time step, say one-mth of the time
step, so that 1 = m∆t. Assume that the probability that one adult
encounters (and hence cannibalizes) one juvenile during ∆t units of
time is proportional to the time elapsed ∆t, that is, c∆t. The coefficient
c > 0 can be thought of as the continuous-time “instantaneous”
encounter rate or cannibalism rate. The probability that one juvenile
survives cannibalism by one adult in ∆t units of time is therefore
1 − c∆t. The probability that a juvenile survives all of the A adults in
∆t units of time (assuming the encounters are independent events) is
the product
YA
A
(1 − c∆t) = (1 − c∆t) .
i=1

Now let’s consider what happens during one time step of the model.
This time period is composed of m = 1/∆t successive times of duration
∆t. The probability that a juvenile survives all of the adults during this
time is the product
m
Y A Am
(1 − c∆t) = (1 − c∆t)
i=1
A/∆t
= (1 − c∆t) .

If we let ∆t → 0, we find that (Exercise 2)

A/∆t
lim (1 − c∆t) = e−cA . (3.14)
∆t→0

So, the probability that a juvenile survives cannibalism during one

model time step in the presence of A adults is e−cA , which is the
Ricker nonlinearity.
 Discrete-Time Maps ■ 71

3.7 EXERCISES
1. Use inequalities (3.11) and (3.12) to obtain inequalities (3.13),
keeping in mind that b > 0.

2. Prove equation (3.14). Hint: You will need L’Hospital’s rule.

3. A population of annual plants has an average per capita fecundity

of 120.3 seeds per plant per year. Approximately 12% of these
seeds germinate. About 10% of the resulting seedlings survive
until maturity. The initial population census (year t = 0) is 52
plants.

a. How can the fecundity be 120.3? That is, what does 0.3
seeds mean?
b. Write down the population model (the initial value
problem).
c. Find the closed-form solution of the initial value problem.
d. Will the population grow or decline? Explain.
e. Is the extinction state stable or unstable?
f. According to the model, how many plants will there be at
year t = 5?
g. How many years will it take for the population to grow to
75 plants, according to the model?
h. How many years will it take for the population to double
its initial size, according to the model? This is called the
doubling time.
i. Explain why the doubling time in this model does not
depend on the initial population size.

4. Consider the linear model

xt+1 = bxt (3.15)

x0 = p.

a. Use mathematical induction (MI) to prove that the closed-

form solution is xt = pbt .
72 ■ Mathematical Modeling in Biology

b. If p > 0 and 0 < b < 1, what is limt→∞ xt ? What does this

mean in biological terms?
c. If p > 0 and b > 1, what is limt→∞ xt ? What does this mean
in biological terms?
d. If p > 0 and b = 1, what is limt→∞ xt ? What does this mean
in biological terms?
e. Write a program that iterates recursion formula (3.15) to
produce a time series of length n. In your program, set p =
50 and n = 5. Graph the time series for each of b = 0.5,
0.9, 1.0, 1.1, and 1.5. Display all five time series on the same
graph. Turn in the program and the graph.

5. Consider the linear population model

xt+1 = (b + 1 − d)xt
x0 = p > 0,

where b > 0 and 0 < d < 1.

a. Give the closed-form solution.

b. Use the closed-form solution to explain why

b < d =⇒ lim xt = 0
t→∞
b > d =⇒ lim xt = ∞
t→∞
b = d =⇒ xt = p for all t.

c. Give a biological interpretation for the results in (b). Does

this make sense biologically?

6. Find the linearization of the function f (x) = x4 −x+15 at x = 1.

Graph the function and its linearization on the same axes near
x = 1.

7. Find the linearization of the function f (x) = sin x at x = 0.

Graph the function and its linearization on the same axes near
x = 0.
 Discrete-Time Maps ■ 73

8. In this problem, you will carry out a complete equilibrium

stability analysis for the Beverton-Holt population model. This
model was introduced by Beverton and Holt in 1957 in the
context of fisheries (Beverton and Holt 1957):

bxt
xt+1 =
1 + cxt
x0 = p≥0 (3.16)
b, c > 0.

a. Is the Beverton-Holt model (3.16) linear or nonlinear? Why?

b. Find all the equilibria.
c. Graph the equilibria xe as functions of the parameter b.
d. Under what conditions is the nontrivial equilibrium posi-
tive?
e. Find the linearization of the Beverton-Holt model at each
equilibrium.
f. For what values of b is each equilibrium stable/unstable?
g. Indicate the stabilities on your bifurcation diagram in (c).
h. What are the bifurcation points on your diagram in (c)?
i. Write a program that iterates the Beverton-Holt map to
produce a time series of length n. In your program, set p =
15, c = 0.002, and n = 100. Graph the time series for each
of b = 0.5, 0.9, 1.0, 1.1, and 1.5. Display all five time series
on the same graph. Turn in the program and the graph.

9. In this problem, you will carry out a complete equilibrium

stability analysis for the so-called discrete logistic map

xt+1 = rxt (1 − xt )
0 < x0 < 1 (3.17)
0 < r < 4.

Here, xt is the density of organisms (number of individuals per

unit area or volume) at time t.
74 ■ Mathematical Modeling in Biology

a. Is the discrete logistic map (3.17) linear or nonlinear? Why?

b. Find all the equilibria.
c. Graph the equilibria xe as functions of the parameter r.
d. Under what conditions is the nontrivial equilibrium posi-
tive?
e. Find the linearization of the model at each equilibrium.
f. For what values of r is each equilibrium stable/unstable?
g. Indicate the stabilities on your bifurcation diagram in (c).
h. What are the bifurcation points on your diagram in (c)?
i. Write a program that iterates the discrete logistic map to
produce a time series of length n. In your program, set x0 =
0.5 and n = 30. Graph the time series for each of r = 0.5,
1.5, 2.5, 3.2, and 3.5. Display all five time series on the same
graph. Turn in the program and the graph. Also turn in the
five numerical time series as lists of numbers.
j. Explore the behavior of the discrete logistic map for 3.5 <
r < 4. You may have to increase r by small increments to
see how the time series change.
k. Mathematically, what happens to solutions if r > 4? Hint:
By hand, iterate xt+1 = 5xt (1 − xt ) starting with x0 = 0.5
to see what happens.

BIBLIOGRAPHY
Beverton, R. J. H. and Holt, S. J. 1957. On the dynamics of exploited
fish populations. Fishery Investigations (II) 19:1–533. [Introduces the
Beverton-Holt mathematical model.]

Ricker, W. E. 1954. Stock and recruitment. Journal of the Fisheries

Research Board of Canada 11:559–623. DOI:10.1139/f54-039. [Classic
paper introducing the Ricker model.]

Ricker, W. E. 1975. Computation and interpretation of biological statistics of

fish populations. Bulletin of the Fisheries Research Board of Canada, No.
119, Ottawa.
 CHAPTER 4

Chaos: Simple Rules

Can Generate Complex
Results

4.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

In this chapter, we return to the Ricker map from Chapter 3 and
explore its dynamics beyond the point at which all the equilibria
destabilize. To complete the exercises in this chapter, you will need a
program that plots bifurcation diagrams for discrete maps. An example
of such a program is the freeware E&F Chaos (see bibliography).
Alternately, you can write your own code (Exercise 1).
Chaos is a fascinating subject that has captured the interest of the
general public as well as that of scientists and mathematicians. If you
would like to read more about chaos, check out the popular book by
James Gleick (Gleick 1987).

4.2 RICKER MODEL REVISITED

In Chapter, 3 we studied the equilibria xe = 0 and xe = c−1 ln b of the
Ricker map

xt+1 = bxt e−cxt

x0 ≥ 0
b, c > 0.

DOI: 10.1201/9781003265382-6 75
76 ■ Mathematical Modeling in Biology

We used the linearization theorem to prove that

0 < b < 1 =⇒ xe = 0 is stable

b > 1 =⇒ xe = 0 is unstable

and that
ln b
1 < b < e2 =⇒ xe = is stable
c
ln b
0 < b < 1 =⇒ xe = is unstable
c
ln b
b > e2 =⇒ xe = is unstable.
c
We summarized this information in a bifurcation diagram, shown again
here as Figure 4.1.
The equilibrium analysis of the Ricker map begs the question,
“What happens to solutions if b > e2 ?” For these values of
b, nonequilibrium solutions cannot equilibrate, for there are no
stable equilibria for them to approach. Some exploratory computer
simulations are in order.
Let c = 0.01 and x0 = 75. Let’s set b at various values and inspect
the resulting time series. Figure 4.2 shows the time series generated
for b = 0.5, b = 1.3, b = 3.6, and b = 5.7. If b = 0.5, solutions

Figure 4.1: Bifurcation diagram of the Ricker map showing the

stability of equilibria. Dotted equilibria are unstable (u); solid are
stable (s). Here, c = 1.
 Chaos: Simple Rules Can Generate Complex Results ■ 77

Figure 4.2: Time series for the Ricker map with b = 0.5, 1.3, 3.6, 5.7.
In each case, c = 0.01.

equilibrate to zero, whereas if b = 1.3, solutions equilibrate to a positive

equilibrium xe ≈ 26.3. The attractors in these cases are the sets {0}
and {26.3}, respectively. For b = 3.6, the equilibrium is higher, about
xe ≈ 128.1, and for b = 5.7, the equilibrium value is even higher,
about xe ≈ 174.0. The attractors are the sets {128.1} and {174.0},
respectively. The interesting thing in this last case is that solutions
oscillate as they approach the equilibrium.
Figure 4.3 shows time series for b = 8, b = 14, b = 14.6, and
b = 17. We know from our previous linearization calculations that for
78 ■ Mathematical Modeling in Biology

Figure 4.3: Ricker time series for b = 8, 14, 14.6, 17. In each case,
c = 0.01.

b > e2 ≈ 7.389, the solutions no longer equilibrate. The computer

simulation for b = 8 in Figure 4.3 suggests solutions approach a two-
cycle; that is, they begin to oscillate between the two values x ≈ 138.6
and x ≈ 277.3. The attractor in this case is the set of two points
{138.6, 277.3} . Further numerical explorations near b = e2 suggest
that as b is increased through the value e2 , the point attractor (for
b < e2 ) bifurcates into two points (for b > e2 ). The value b = e2
is called a bifurcation value. For b = 14, solutions approach a four-
cycle attractor, and for b = 14.6, an eight-cycle. As b is increased, the
periods of these cycles continue to double, until at a critical value of
 Chaos: Simple Rules Can Generate Complex Results ■ 79

b, there is a transition to an aperiodic, complicated dynamic called

chaos. This sequence of bifurcations is called a period-doubling cascade
to chaos.
If the attractor, also called the final state, of each time series is
plotted against the parameter b, we can continue the Ricker bifurcation
diagram in Figure 4.1 beyond b = e2 (Figure 4.4). Note that unstable
equilibria and cycles are not shown in Figure 4.4, since computer
iterations will only identify the stable ones. Chaos has a complicated
mathematical definition which we will not address in this book. For
our purposes, here we simply note three important characteristics
of chaos. First, chaos is deterministic. Note that the chaotic time
series shown at the bottom of Figure 4.3 is completely determined
by the simple iterative rule xt+1 = 17xt e−0.01xt , where x0 = 75.
Second, chaotic dynamics, although deterministic, “look” random.
Third, chaotic dynamics exhibit sensitivity to initial conditions. In
Figure 4.5, we see two time series of the Ricker model with b = 17
and c = 0.01. The solid curve begins with the initial condition x0 =
75, whereas the dotted curve is generated from the initial condition
x0 = 75.1. The two time series are close together for several time steps,
but soon diverge and look nothing like each other. This is sensitivity to
initial conditions: A small change or error in initial conditions is quickly
magnified.

Figure 4.4: Bifurcation diagram for the Ricker map with c = 0.01.
80 ■ Mathematical Modeling in Biology

Figure 4.5: Two time series of the Ricker map with b = 17 and c =
0.01. The solid curve has initial condition x0 = 75, and the dashed
curve has initial condition x0 = 75.1. The two solutions soon diverge,
showing sensitivity to initial conditions.

4.3 NEW PARADIGMS ARISE FROM CHAOS

4.3.1 Deterministic Unpredictability

Determinism has generally been equated with predictability. If a system
is deterministic, then the output is completely determined by the input
and is thus predictable, given the input. Similarly, stochasticity has
been identified with unpredictability, for obvious reasons.
Chaos, however, significantly alters this classical paradigm. Chaos
is deterministic, but has sensitivity to initial conditions. Therefore, if
the initial conditions are not known precisely (and they never are in
the real world), then the future state of the system is, for all practical
purposes, unpredictable. So in a very real sense, there is such a thing
as “deterministic unpredictability.”

4.3.2 Complex Results Can Arise from Simple Rules

Another classical paradigm has been that simple causes give rise
to simple results, and (equivalently) complex results imply complex
causes. Chaos shows emphatically that this is not necessarily the
case. The Ricker model is a simple recursion formula, and yet its
dynamics are so complicated that a mathematician can spend a lifetime
 Chaos: Simple Rules Can Generate Complex Results ■ 81

exploring them. Thus, we see from this example that complex results
do not necessary imply complex causes. Very simple rules can generate
extraordinarily complicated results.

4.4 MAY’S HYPOTHESIS

In 1976, Robert May wrote (May 1976):

Quite apart from their intrinsic mathematical interest,

the above results raise very awkward biological questions.
They show that simple and fully deterministic models, in
which all biological parameters are exactly known, can
nonetheless (if the nonlinearities are sufficiently severe) lead
to population dynamics which are in effect indistinguishable
from the sample function of a random process. Apparently
chaotic population fluctuations need not necessarily be due
to random environmental fluctuations, or sampling errors,
but may reflect the workings of some deterministic, but
strongly density dependent, population model.

May’s hypothesis is the idea that the apparently random

fluctuations of population abundances can be explained largely by
low-dimensional, nonlinear, deterministic forces.
In 1997, chaos was documented in population dynamics for the first
time when it was experimentally induced in laboratory populations of
insects (Costantino et al. 1997; Cushing et al. 2003; Dennis et al. 2001).
We will discuss these experiments in Chapter 6. To date, no one knows
whether chaos occurs “naturally” in field populations; speculation and
argumentation abound.

4.5 EXERCISES
In this set of exercises, you will need a program that draws bifurcation
diagrams for discrete maps. An example is the freeware program E&F
Chaos (see bibliography entry for URL). Alternately, you can write
your own code (Exercise 1).

1. Write your own code to reproduce Figure 4.4, which is the

bifurcation diagram for the Ricker map. Hint: You will need to
82 ■ Mathematical Modeling in Biology

specify a “mesh” of values of the bifurcation parameter b along

the horizontal axis. For each of these fixed values of b, you will
iterate the map a fixed number of times, maybe 300, so that the
time series approaches the attractor. The early iterates, before the
time series is close to the attractor, are called “transients,” and
you will not plot them in the bifurcation (final state) diagram.
You will plot the later iterates, say iterates 200–300, as a scatter
plot above the associated value of b.

2. Use the computer to draw a bifurcation diagram for the Beverton-

Holt model
bxt
xt+1 =
1 + cxt
x0 > 0
b, c > 0

using b as the bifurcation parameter. Set c = 1, and let 0 ≤ b ≤

10.

3. Use the computer to draw a bifurcation diagram for the so-called

discrete logistic model

xt+1 = bxt (1 − xt )
0 < x0 < 1
0 < b<4

using b as the bifurcation parameter. What happens to solutions

if b > 4?

4. Consider the Ricker model with survivorship:

xt+1 = bxt e−cxt + (1 − µ) xt

x0 > 0
b, c > 0
0 ≤ µ ≤ 1.

Here, µ is the fraction of the population that does not survive

one time step and 1 − µ is the fraction that does survive one time
step.
 Chaos: Simple Rules Can Generate Complex Results ■ 83

a. Set c = 0.01. Use the computer to draw bifurcation

diagrams for each of the following values of µ, using b as
the bifurcation parameter with 0 ≤ b ≤ 100.
i. µ = 1.0 (Ricker map without survivorship)
ii. µ = 0.9
iii. µ = 0.8
iv. µ = 0.7
v. µ = 0.6
vi. µ = 0.5
vii. µ = 0.4
viii. µ = 0.3
ix. µ = 0.2
x. µ = 0.1
xi. µ = 0.0
b. Set c = 0.01 and b = 80. Use the computer to draw the
bifurcation diagram using µ as the bifurcation parameter
with 0 ≤ µ ≤ 1. Relate this diagram to the sequence of
diagrams in part (4a).

5. In this problem, you will use the computer to further explore the
bifurcation diagram of the Ricker model

xt+1 = bxt e−cxt

x0 > 0
b, c > 0

using b as the bifurcation parameter. Let c = 1.

a. Draw the bifurcation diagram for 0 ≤ b ≤ 50.

b. Draw the bifurcation diagram for 21 ≤ b ≤ 26.
c. Draw the bifurcation diagram for 23 ≤ b ≤ 25.
d. Draw the bifurcation diagram for 24.8 ≤ b ≤ 25. Zoom in on
the neighborhood of 0.3 ≤ x ≤ 1.8 and 24.88 ≤ b ≤ 24.95.
This illustrates the fractal nature of the bifurcation diagram,
in which patterns are repeated at ever-smaller scales. Fractal
structures are deeply related to chaos.
84 ■ Mathematical Modeling in Biology

BIBLIOGRAPHY
Costantino, R. F., Desharnais, R. A., Cushing, J. M., and Dennis, B.
1997. Chaotic dynamics in an insect population. Science 275:389–391.
DOI: 10.1126/science.275.5298.389. [“Announcement” paper heralding the
documentation of chaos in a laboratory insect population. A follow-up
paper gave the details. See Dennis et al. (2001) below. Both theoretical
and empirical: models connected to data.]

Cushing, J. M., Costantino, R. F., Dennis, B., Desharnais, R. A., and

Henson, S. M. 2003. Chaos in Ecology: Experimental Nonlinear Dynamics.
Academic Press, San Diego. [Summarizes the large body of work involved
in the documentation of chaos in a laboratory population. In Chapter 6
we will draw heavily on this book to showcase how dynamic population
data can be connected to mathematical models.]

Dennis, B., Desharnais, R. A., Cushing, J. M., Henson, S. M., and

Costantino, R. F. 2001. Estimating chaos and complex dynamics in an
insect population. Ecological Monographs 71:277–303. DOI: 10.1890/0012-
9615(2001)071[0277:ECACDI]2.0.CO;2. [Detailed follow-up to the an-
nouncement paper in Science. See Ref. Costantino et al. 1997]

E&F Chaos. https://cendef.uva.nl/software/ef-chaos/ef-chaos.html. [Useful

tool for drawing bifurcation diagrams.]

Gleick, J. 1987. Chaos: Making a New Science. Viking, New York. [This
popular book was a finalist for the Pulitzer Prize in 1987.]

May, R. M., ed. 1976. Theoretical Ecology: Principles and Applications. W.

B. Saunders, Philadelphia. [Played a foundational role in the development
of theoretical ecology.]
 CHAPTER 5

Higher-Dimensional
Discrete-Time Models

5.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

In this chapter, we consider how to model discrete-time systems whose
descriptions require more than one state variable. Before beginning
this chapter, you should review basic matrix operations and learn
(or review) some introductory linear algebra by working through the
self-guided tutorial in Appendix A. A course in linear algebra is
not a prerequisite for this textbook; everything you need to know
from linear algebra for this and subsequent chapters is contained in
Appendix A.
Also, note that in this chapter, you will need to know how to take
partial derivatives, a topic encountered in third-semester calculus. For
those who have not met partial derivatives, they are easy to learn. If
you have a function of (say) two variables f (x, y), then the partial
derivative of f with respect to x, denoted ∂f /∂x, is the derivative
of f with respect to x while holding y constant. So, for example, if
f (x, y) = x2 + y 2 + x3 y 3 , then the partial derivatives are (showing
zeros to make a point):

∂f
= 2x + 0 + y 3 3x2


∂x
∂f
= 0 + 2y + x3 3y 2 .


∂y
If this is your first experience with partial derivatives, work a few
problems in a calculus book to get comfortable with them.

DOI: 10.1201/9781003265382-7 85
86 ■ Mathematical Modeling in Biology

The main idea in this chapter is to learn how to solve and analyze
linear systems, and then to apply those techniques to nonlinear systems
through the method of linearization.

5.2 INTRASPECIFIC INTERACTIONS

Up to this point, we have modeled populations with one state variable,

usually the number of organisms (or density of organisms per unit
area or volume). In doing this, we have assumed that all members
of the population are similar. For many populations, this is a false
assumption. For example, flour beetles (Tribolium) undergo complete
metamorphosis and, in the process, exhibit four discrete life stages—
egg, larva, pupa, and adult. Life-cycle stages can have different
durations, behaviors, intraspecific interactions (interactions within
species), and vital rates (e.g., birth and death rates).
To model the population dynamics of organisms that pass through
various life-cycle stages, we typically need to use more than one
state variable. For example, flour beetles require at least three state
variables: one for larvae, one for pupae, and another for adults. The
egg stage can sometimes be ignored because of its relatively short
duration (Cushing et al. 2003). The resultant model is called an age-
structured model , or more generally, a stage-structured model . These
models are used to describe the dynamics of populations whose vital
rates and interactions depend on the ages or stages of the organisms.
Discrete-time age-structured models are called Leslie matrix models,
after biologist Patrick H. Leslie. The number of state variables is called
the dimension of the system. A well-known standard text about Leslie
matrix models is the one by Hal Caswell (Caswell 2001).

5.3 INTERSPECIFIC INTERACTIONS

When modeling a population, it is also important to consider whether

any interspecific interactions (interactions between species) play a
significant role in the dynamics. In terms of dynamics, the three main
categories of interspecific interactions are predator-prey, mutualism
(cooperation), and competition (Table 5.1). In predator-prey systems,
the presence of the predator species has a negative effect on the prey
species, while the prey species has a positive effect on the predator
 Higher-Dimensional Discrete-Time Models ■ 87

TABLE 5.1 Species Interactions

Effect of Species 1 Effect of Species 2

on Species 2 on Species 1
Predator-prey − +
Mutualism + +
Competition − −

species. Host-parasite systems also fit in this category. In mutualistic

systems (cooperative systems), each species has a positive effect on the
other. In competitive systems, each species has a negative effect on the
other.
Models of interacting species require more than one state variable.
For example, a multi-species system consisting of one predator and two
prey species would require at least a three-dimensional model.

5.4 EXAMPLE OF AN AGE-STRUCTURED SINGLE-SPECIES

MODEL
Consider an age-structured juvenile-adult model for a fish population,
with the following assumptions:

(A1) An adult fish has probability µA of dying during one year,

where 0 < µA < 1.

(A2) Adult fish deposit an average of b > 0 fertilized eggs per

adult per year.

(A3) A fertilized egg hatches with probability p > 0.

(A4) A juvenile fish matures sexually in one year.

(A5) A juvenile fish has probability µJ of dying before maturing,

where 0 < µJ < 1.

These assumptions are summarized in the Leslie diagram in

Figure 5.1.
Let Jt and At be the number of juvenile and adult fish, respectively,
at time t. The time step is one year, the duration of the juvenile stage.
88 ■ Mathematical Modeling in Biology

Figure 5.1: Leslie diagram for an age-structured fish population model

showing per capita flow rates.

In general, we can write

Jt+1 = Juvenile recruits + juveniles carried over from last census

At+1 = Adult recruits + adults surviving from last census.

From the assumptions, we know that

Juvenile recruits = (Eggs per adult) (adults at time t)

× (fraction eggs that hatch)
= bAt p
Juvenile carryovers = 0
Adult recruits = (Juveniles at time t)
× (fraction that survive until maturity)
= Jt (1 − µJ )
Adult survivors = (Adults at time t)
× (fraction that survive one year)
= At (1 − µA ) .

The Leslie model is therefore

Jt+1 = bpAt (5.1)

At+1 = (1 − µJ ) Jt + (1 − µA ) At .

We can also write equation (5.1) as a matrix equation:

    
Jt+1 0 bp Jt
= .
At+1 (1 − µJ ) (1 − µA ) At
 Higher-Dimensional Discrete-Time Models ■ 89

The matrix  
0 bp
M=
(1 − µJ ) (1 − µA )

is called a Leslie matrix , or projection matrix . System (5.1) is a two-

dimensional linear model: two-dimensional because it has two state
variables Jt and At , and linear because both equations in (5.1) are
linear in the state variables J and A.

5.5 EXAMPLE OF A TWO-SPECIES MODEL

Consider two similar species with nonoverlapping generations, compet-
ing for space, with the following assumptions:

(A1) In the absence of crowding, species 1 and 2 have per capita

birth rates of b1 and b2 , respectively.

(A2) Each species suffers fractional reductions in per capita birth

rate due to crowding by its own species and by the other species.
Each fractional reduction is of Ricker type, that is, of the form
e−cN , where N is the number of competing organisms.

Let xt and yt be the numbers of species 1 and 2, respectively, at

time t. Then

xt+1 = b1 xt e−c11 xt e−c12 yt

yt+1 = b2 yt e−c21 xt e−c22 yt ,

or

xt+1 = b1 xt e−c11 xt −c12 yt (5.2)

−c21 xt −c22 yt
yt+1 = b2 yt e .

The system (5.2) is a two-dimensional nonlinear model: two-

dimensional because it has two state variables xt and yt , and nonlinear
because the right-hand sides of the equations in (5.2) are nonlinear
functions of the state variables x and y.
In what follows, we first discuss linear models and how to solve
them, and then we use that knowledge to study nonlinear models.
90 ■ Mathematical Modeling in Biology

5.6 n-DIMENSIONAL LINEAR DIFFERENCE EQUATIONS

Consider the n-dimensional linear discrete-time system


 x1 (t + 1) = a11 x1 (t) + a12 x2 (t) + · · · + a1n xn (t)
 x2 (t + 1) = a21 x1 (t) + a22 x2 (t) + · · · + a2n xn (t)

.. .


 .
xn (t + 1) = an1 x1 (t) + an2 x2 (t) + · · · + ann xn (t)


This system is called linear because the equations are linear in the
state variables x1 , x2 , . . . , xn . It can be written as a matrix equation:
    
x1 (t + 1) a11 a12 ··· a1n x1 (t)
 x2 (t + 1)   a21 a22 ··· a2n  x2 (t) 
= .
    
 .. .. .. ..  ..
 .   . . .  . 
xn (t + 1) an1 an2 · · · ann xn (t)

The matrix equation can be written with the simpler notation

x (t + 1) = Mx (t) ,

where
   
x1 a11 a12 ··· a1n
 x2   a21 a22 ··· a2n 
x =  and M =  .
   
.. .. .. ..
 .   . . . 
xn an1 an2 · · · ann

5.6.1 n-Dimensional Leslie Models

Suppose a species has n life stages or “classes” of equal duration, which
we choose as the model time step. Suppose that in one time step,
individuals survive to move from class i to class i + 1 with probability
pi for i = 1, 2, . . . , n − 1, and that individuals in class n (the final
class) survive each time step to remain in class n with probability pn .
Suppose furthermore that each class i has a per capita fecundity of
fi . We can conceptualize a population of this species with the Leslie
diagram shown in Figure 5.2.
 Higher-Dimensional Discrete-Time Models ■ 91

Figure 5.2: n-dimensional Leslie diagram.

If xi (t) is the number of organisms in class i at time t, we can write



 x1 (t + 1) = f1 x1 (t) + f2 x2 (t) + · · · + fn xn (t)
 2 (t + 1) = p1 x1 (t)
x



x3 (t + 1) = p2 x2 (t)
 ..
.




xn (t + 1) = pn−1 xn−1 (t) + pn xn (t) ,


or, in matrix form,

 
  f1 f2 · · · fn  
x1 (t + 1)  p1 0 · · · 0  x1 (t)
x2 (t + 1)    x2 (t) 
 

  . .. ..
= 0 p2 . .
 
 ..  ..
. .

  . 

 .. 0 . ..  
xn (t + 1) 0 
xn (t)
0 0 pn−1 pn
The matrix  
f1 f2 · · · fn
 p1 0 · · · 0 
 
 . .. .. 
 0 p2
M = . 

 .. . .. 
 . 0 0 
0 0 pn−1 pn
is called a Leslie matrix . The birth rates are in the first row, while the
survivorships are on the subdiagonal.
92 ■ Mathematical Modeling in Biology

5.7 SOLVING LINEAR SYSTEMS OF DIFFERENCE EQUATIONS

5.7.1 An Example
We begin with an example. Consider the two-dimensional linear system

 xt+1 = 3xt + 2yt
yt+1 = 21 xt + 3yt


. (5.3)

 x0 = 0
y0 = 2


What are the equilibria of this system? The fixed point equations, or
equilibrium equations, are

x = 3x + 2y
.
y = 12 x + 3y

The first of these implies that y = −x. The second, however, implies
that y = −x/4. This is a contradiction unless x = y = 0. Thus, the
only equilibrium is the extinction state (xe , ye ) = (0, 0). Because the
given initial condition (x0 , y0 ) = (0, 2) is different from the equilibrium,
we know that the system is not initially at equilibrium, and so the
population size must change over time. How will it change? Will it tend
toward the extinction equilibrium? In order to answer this question, we
will find the closed-form solution of system (5.3).
Recall that the closed-form solution to the one-dimensional linear
difference equation
xt+1 = bxt
is
xt = cbt ,
where c is the initial condition. We might expect that the solution of
system (5.3) will be similar. Therefore, let’s look for nontrivial solutions
(solutions that are not constantly zero) of the form

xt = v1 λt (5.4)
t
yt = v2 λ ,

where v1 , v2 , and λ are fixed real numbers. Let’s require λ ̸= 0, and

also require that at least one of v1 or v2 be nonzero; otherwise, we
will simply recover the equilibrium solution (trivial solution) (0, 0).
 Higher-Dimensional Discrete-Time Models ■ 93

The proposed solution (5.4) is called the Ansatz , a German word for
“approach.” Because it is a noun in German, Ansatz is capitalized.
If we plug our proposed solution (5.4) into the dynamical system
(5.3), we arrive at the following equations:

v1 λt+1 = 3v1 λt + 2v2 λt


.
v2 λt+1 = 21 v1 λt + 3v2 λt

̸ 0, we can divide through by λt to obtain

Given λ =

v1 λ = 3v1 + 2v2
.
v2 λ = 12 v1 + 3v2

One further manipulation gives us the algebraic system of equations


(3 − λ) v1 + 2v2 = 0
1 . (5.5)
2 v1 + (3 − λ) v2 = 0

We are looking for a solution (v1 , v2 ) of system (5.5), where v1 , v2

are not both zero. Notice, however, that (v1 , v2 ) = (0, 0) is indeed a
solution of system (5.5). If you think of the two equations (3 − λ) v1 +
2v2 = 0 and 21 v1 + (3 − λ) v2 = 0 as lines in the v1 -v2 plane, this
means the two lines meet at the origin (0, 0). But system (5.5) has
a nontrivial solution (v1 , v2 ) if and only if the lines meet somewhere
other than the origin. This can happen if and only if the two lines are
coincident. The two lines are coincident if and only if the two equations
(3 − λ) v1 + 2v2 = 0 and 12 v1 + (3 − λ) v2 = 0 are linearly dependent,
that is, if and only if they are multiples of each other. A fact from
linear algebra says this is true if and only if the determinant of the
coefficient matrix is zero, that is, if and only if

(3 − λ) 2
1 = 0.
2 (3 − λ)

This holds if and only if

λ2 − 6λ + 8 = 0,

which is to say
(λ − 4) (λ − 2) = 0.
94 ■ Mathematical Modeling in Biology

The last equality holds if and only if λ = 2 or λ = 4. These values of λ

are called the eigenvalues for the system. We have shown that system
(5.5) has a nontrivial solution (v1 , v2 ) if and only if λ = 2 or λ = 4.
We now consider each of these eigenvalues, in turn, and find v1 and
v2 for each. If λ = 2, then system (5.5) becomes

v1 + 2v2 = 0
1 .
2 v1 + v 2 = 0

Note that these last two equations are indeed multiples of each other;
their graphs in the v1 -v2 plane are coincident lines. Thus, there
are infinitely many solutions (v1 , v2 ) . However, these solutions are
constrained by the relationship

v1 = −2v2 .

For example, if v2 = 1, then v1 = −2. We say an eigenvector associated

with eigenvalue λ = 2 is
   
v1 −2
= .
v2 1

Note that any scalar multiple of this eigenvector is also an eigenvector.

Looking back at our proposed solution (5.4), we see that we have found
an eigensolution
t
xt = v1 λt = −2 (2)
t
yt = v2 λt = 1 (2) ,

which can also be written

   
xt −2
= 2t ,
yt 1

for system (5.3).

Now let’s consider the other eigenvalue, λ = 4. If λ = 4, system
(5.5) becomes

−v1 + 2v2 = 0
1 .
2 v1 − v 2 = 0
 Higher-Dimensional Discrete-Time Models ■ 95

Thus, v1 = 2v2 , so we can let v2 = 1 and v1 = 2. Therefore, an

eigenvector associated with λ = 4 is
   
v1 2
= ,
v2 1
and the corresponding eigensolution for system (5.3) is
t
xt = 2 (4)
t
yt = 1 (4) ,
that is    
xt 2
= 4t .
yt 1
A fact from dynamical systems theory says that the eigensolutions
span the space of all solutions; that is, any solution of the original
system (5.3) is a linear combination of the eigensolutions. The general
solution (i.e., the set of all solutions) of system (5.3) is therefore the
closed-form solution
     
xt −2 t 2
= c1 2 + c2 4t
yt 1 1
 t t 
−2c1 (2) + 2c2 (4)
= t t ,
c1 (2) + c2 (4)
where c1 , c2 ∈ R. In component form, the general solution is
t t
xt = −2c1 (2) + 2c2 (4) (5.6)
t t
yt = c1 (2) + c2 (4) .
Now we want to find the particular solution corresponding to the
given initial values of x0 = 0 and y0 = 2. We set t = 0 in the general
solution (5.6) to obtain the algebraic system
0 = −2c1 + 2c2
2 = c1 + c2 .
From this, we see that c1 = 1 and c2 = 1. Plugging these values for c
back into the general solution (5.6) gives the particular solution
t t
xt = −2 (2) + 2 (4)
yt = 2t + 4t .
96 ■ Mathematical Modeling in Biology

This is the closed-form solution of the initial value problem (5.3).

Note that the dominant eigenvalue (λ = 4) determines the exponential
rate of growth for the system. In particular, the equilibrium (0, 0) is
unstable, because solutions starting near (0, 0) grow exponentially.

5.7.2 Solving the General Two-Dimensional System

We now solve the general two-dimensional linear system

xt+1 = axt + byt
. (5.7)
yt+1 = cxt + dyt

Clearly, this system has the extinction state (x, y) = (0, 0) as a solution.
In fact, the only equilibrium is (0, 0) as long as the determinant of the
coefficient matrix is nonzero, that is, as long as

a b
̸= 0.
c d

In what follows, we will always assume this holds.

We want to look for nontrivial solutions of the form

xt = v1 λt (5.8)
t
yt = v2 λ ,

with λ ̸= 0 and v1 , v2 not both zero. This is the Ansatz.

Plugging the Ansatz (5.8) into system (5.7) yields

v1 λt+1 = av1 λt + bv2 λt


,
v2 λt+1 = cv1 λt + dv2 λt

which can be simplified to


v1 λ = av1 + bv2
.
v2 λ = cv1 + dv2

Rearranging, we obtain the algebraic system of equations


(a − λ) v1 + bv2 = 0
. (5.9)
cv1 + (d − λ) v2 = 0
 Higher-Dimensional Discrete-Time Models ■ 97

System (5.9) has the trivial solution (v1 , v2 ) = (0, 0). It has a
nontrivial solution (v1 , v2 ) if and only if the two equations in (5.9)
are coincident lines, if and only if
(a − λ) b
= 0,
c (d − λ)
that is, if and only if
λ2 − (a + d) λ + (ad − bc) = 0.
This last equation is called the characteristic equation. The solutions λ1
and λ2 of the characteristic equation are the eigenvalues for the system.
These are the values of λ for which the Ansatz (5.8) is a solution;
hence, these are the exponential growth rates of the system. For each
eigenvalue, we can find an associated eigenvector from system (5.9)
by plugging in the eigenvalue and finding a relationship between v1
̸ λ2 and v = (v1 , v2 )⊤ and w = (w1 , w2 )⊤ are linearly
and v2 . If λ1 =
independent eigenvectors associated with λ1 and λ2 , respectively, then
the general solution of system (5.7) written in vector form is
xt = c1 vλt1 + c2 wλt2 ,
where xt = (xt , yt )⊤ and c1 , c2 ∈ R. In component form, the general
solution is
xt = c1 v1 λt1 + c2 w1 λ2t
yt = c1 v2 λt1 + c2 w2 λt2 .
A given particular solution is found by plugging the initial conditions
into the general solution, solving for the constants c1 , c2 , and plugging
these back into the general solution.

5.7.3 Solving Higher-Dimensional Systems

Linear systems with dimension higher than two are solved in the same
way as two-dimensional systems. A discrete linear model

 x1 (t + 1) = a11 x1 (t) + a12 x2 (t) + · · · + a1n xn (t)

 x2 (t + 1) = a21 x1 (t) + a22 x2 (t) + · · · + a2n xn (t)

.. (5.10)


 .
xn (t + 1) = an1 x1 (t) + an2 x2 (t) + · · · + ann xn (t)

98 ■ Mathematical Modeling in Biology

can be written as a matrix equation

x (t + 1) = Mx (t) ,

where
   
x1 a11 a12 ··· a1n
 x2   a21 a22 ··· a2n 
x =  and M =  .
   
 ...   ... ..
. ... 
xn an1 an2 · · · ann

The eigenvalues of the n × n matrix M are the solutions λ of the

characteristic equation
|M−λI| = 0.

If you do not know how to find the determinant of a n × n matrix for

n > 2, you can look this up in an introductory linear algebra text or
calculus text.
Given a particular eigenvalue λ, a solution v of the vector equation

(M−λI) v = 0

is called an eigenvector corresponding to λ.

Let λ1 , λ2 , . . . , λn be distinct eigenvalues of M with corresponding
eigenvectors v1 , v2 , · · · , vn . Suppose the eigenvectors are linearly
independent (i.e., no one of them can be written as a linear combination
of the others); this is true in many applications. Then the general
solution of system (5.10) can be written in vector form as

x (t) = c1 v1 λt1 + c2 v2 λt2 + · · · + cn vn λtn . (5.11)

Initial conditions give rise to particular values of the constants ci ,

leading to the particular solution of the initial value problem.
The fundamental theorem of demography basically says that the
magnitudes of the eigenvalues tell you whether or not the population
is going extinct. The theorem has two parts:
 Higher-Dimensional Discrete-Time Models ■ 99

1. If all the eigenvalues are inside the unit circle in the complex
plane, that is, if |λi | < 1 for all i = 1, 2, · · · n, then the extinction
equilibrium is asymptotically stable and all solutions approach
the zero vector.

2. If any one of the eigenvalues is outside the unit circle, that is, if
|λi | > 1 for some i, then the extinction equilibrium is unstable.

The strong ergodic property tells us how the age classes are
distributed in the long run: If there is a simple dominant eigenvalue
λ1 , that is, an unrepeated eigenvalue λ1 such that |λ1 | > |λi | for every
̸ 1, then from equation (5.11) we have (Exercise 4)
i=

1
lim x (t) = c1 v1 .
t→∞ λt1

The eigenvector v1 belonging to the simple dominant eigenvalue λ1

gives the stable age distribution (or stable stage distribution) of the
model, that is, the asymptotic relative distribution of stages as t → ∞.
This does not necessarily mean the population is equilibrating; it can
be growing exponentially or declining to extinction, depending on the
magnitude of λ1 . For example, if the dominant eigenvalue is λ1 = 1.5
⊤
and its eigenvector is (2, 1) , then in the long run, the population is
growing without bound with an asymptotic age distribution of twice
as many juveniles as adults.

5.8 NONLINEAR SYSTEMS

Let’s consider a nonlinear Leslie model for populations of the flour
beetle Tribolium castaneum. Under laboratory conditions, eggs quickly
hatch into larvae, larvae pupate after about 14 days, and pupae
metamorphize into adults after another 14 days. Larvae and adults
eat eggs, and adults eat pupae. A well-validated model of Tribolium
castaneum dynamics is the three-dimensional nonlinear Leslie model
known as the LPA model (Cushing et al. 2003):

Lt+1 = bAt e−cel Lt −cea At

Pt+1 = (1 − µl ) Lt
At+1 = Pt e−cpa At + (1 − µa ) At .
100 ■ Mathematical Modeling in Biology

The LPA model is nonlinear because of the Ricker nonlinearities in the

first and third equations.
Nonlinear models such as the LPA model do not, in general, have
simple closed-form solutions. Furthermore, it is often impossible to
solve explicitly for the equilibria; in such cases, the equilibria must
be found numerically. Given an equilibrium, however, we can study
its stability through the technique of linearization, just as we did for
one-dimensional maps.

5.8.1 Linearization
Consider the two-dimensional nonlinear system

xt+1 = f (xt , yt ) (5.12)

yt+1 = g (xt , yt ) .

Equilibria of this system are constant solutions, i.e., constant pairs

(xe , ye ) that simultaneously satisfy the fixed point system of equations

xe = f (xe , ye )
ye = g (xe , ye ) .

We want to approximate the behavior of the system near the equilibria

in order to investigate the stability of the equilibria.
In multivariate calculus, one learns to approximate a surface z =
f (x, y) near the point (x, y) = (a, b) with a plane that is tangent to
the surface at that point. The equation of the tangent plane is

∂f ∂f
z = f (a, b) + (a, b) (x − a) + (a, b) (y − b) ,
dx ∂y

and this is a good approximation for f (x, y) as long as the point

(x, y) is sufficiently close to the point (a, b) . Here, the notation ∂f /dx
refers to the partial derivative of f with respect to the variable x.
Computationally, ∂f /dx is simply the derivative of f (x, y) with respect
to x while treating y as a constant. To practice computing partial
derivatives, see any multivariate calculus text.
 Higher-Dimensional Discrete-Time Models ■ 101

If we approximate f and g with their tangent planes at equilibrium

(xe , ye ) , we have

∂f ∂f
f (x, y) ≈ f (xe , ye ) + (xe , ye ) (x − xe ) + (xe , ye ) (y − ye )
dx ∂y
∂f ∂f
= xe + (xe , ye ) (x − xe ) + (xe , ye ) (y − ye )
dx ∂y
∂g ∂g
g (x, y) ≈ g (xe , ye ) + (xe , ye ) (x − xe ) + (xe , ye ) (y − ye )
dx ∂y
∂g ∂g
= ye + (xe , ye ) (x − xe ) + (xe , ye ) (y − ye )
dx ∂y

for x ≈ xe and y ≈ ye . Thus, system (5.12) can be approximated by

∂f ∂f
xt+1 ≈ xe + (xe , ye ) (xt − xe ) + (xe , ye ) (yt − ye )
dx ∂y
∂g ∂g
yt+1 ≈ ye + (xe , ye ) (xt − xe ) + (xe , ye ) (yt − ye )
dx ∂y
for xt ≈ xe and yt ≈ ye .
The change of variables

ut = xt − xe (5.13)
vt = yt − ye

leads to the linear system

∂f ∂f
ut+1 ≈ (xe , ye ) ut + (xe , ye ) vt
dx ∂y
∂g ∂g
vt+1 ≈ (xe , ye ) ut + (xe , ye ) vt
dx ∂y

for ut ≈ 0 and vt ≈ 0. Note that, by their definition in equation (5.13),

the variables ut and vt measure the displacement, or variation, of the
system from equilibrium.

Definition 5.1 The linearization of a nonlinear system

xt+1 = f (xt , yt )
yt+1 = g (xt , yt )
102 ■ Mathematical Modeling in Biology

at an equilibrium (xe , ye ) is the linear system

∂f ∂f
ut+1 = (xe , ye ) ut + (xe , ye ) vt
dx ∂y
∂g ∂g
vt+1 = (xe , ye ) ut + (xe , ye ) vt .
dx ∂y
We can write the linearization as
ut+1 = J (xe , ye ) ut ,
where !
  ∂f ∂f
u dx ∂y
u= and J = ∂g ∂g .
v dx ∂y

The matrix J is called the Jacobian matrix .

5.8.2 An Example
Let’s carry out a complete equilibrium stability analysis for the
nonlinear system
xt+1 = yt (5.14)
3
yt+1 = yt (1 − xt )
2
at each of its equilibria.
First, we find the equilibria by solving the fixed point equation
x = y
3
y = y (1 − x) .
2
The equilibria are (0, 0) and (1/3, 1/3) . Then we compute the Jacobian:
!
∂f ∂f
dx ∂y
J = ∂g ∂g
dx ∂y
 
0 1
= .
− 32 y 3
2 (1 − x)
At the equilibrium (0, 0), the Jacobian is
 
0 1
J (0, 0) = ,
0 32
 Higher-Dimensional Discrete-Time Models ■ 103

so the linearization at (0, 0) is the linear system

ut+1 = vt
3
vt+1 = vt .
2
The eigenvalues of the linearized system are λ = 0, 3/2. Since λ = 3/2
is outside the unit circle in the complex plane (that is, |λ| > 1), the
(0, 0) equilibrium of the linearized system is unstable.
At the equilibrium (1/3, 1/3) , the Jacobian is
 
0 1
J= ,
− 21 1

so the linearization at (1/3, 1/3) is

ut+1 = vt
1
vt+1 = − ut + vt .
2
1
The eigenvalues of this linear system are λ = ± 12 i, which are both
2q

1 2

2
inside the unit circle in the complex plane since |λ| = 2 + 12 =
√
2
2 < 1. Thus, the (0, 0) equilibrium of the linearized u-v system is
stable. Since ut and vt measure the displacement of the system away
from the equilibrium (1/3, 1/3) , it seems reasonable to suspect that
the (1/3, 1/3) equilibrium of the nonlinear system (5.14) is stable as
well. The following theorem states that this is so.

Definition 5.2 An equilibrium (xe , ye ) of

xt+1 = f (xt , yt )
yt+1 = g (xt , yt )

is hyperbolic if and only if the Jacobian matrix J (xe , ye ) evaluated at

the fixed point has no eigenvalues of modulus one.

Theorem 5.1 (Linearization Theorem) Let f (x, y) and g (x, y) be

continuously differentiable functions of x and y (that is, have
104 ■ Mathematical Modeling in Biology

continuous derivatives in both variables). Consider the nonlinear

system

xt+1 = f (xt , yt )
yt+1 = g (xt , yt )

with hyperbolic fixed point (xe , ye ) .

1. If all the eigenvalues of the Jacobian matrix J (xe , ye ) at the fixed

point have moduli less than one, then the fixed point (xe , ye ) is
asymptotically stable.

2. If at least one of the eigenvalues of J (xe , ye ) has modulus greater

than one, then (xe , ye ) is unstable.

5.9 EXERCISES
1. Consider the discrete-time system

xt+1 = 2xt + 3yt

yt+1 = −xt − 2yt .

a. Find the general solution.

b. Find the particular solution subject to the initial conditions
x0 = 4 and y0 = 0.

2. Consider the discrete-time system

xt+1 = 2xt + 2yt

yt+1 = 7xt − 3yt .

a. Find the general solution.

b. Find the particular solution subject to the initial conditions
x0 = 0 and y0 = 9.

3. Consider the discrete-time system

xt+1 = 18xt − 11yt (5.15)

yt+1 = 8xt − yt .
 Higher-Dimensional Discrete-Time Models ■ 105

a. Write a program that iterates system (5.15). Use the initial

conditions x0 = 35 and y0 = 26, and compute xt and yt
for t = 1 to t = 5. In the program, create a 6 × 4 matrix
with the following four columns: a column for t which would
begin with 0 and end with 5; a column for the corresponding
value of xt ; a column for the corresponding value of yt ; and
a column for the value of the ratio xt /yt . Output this matrix
to the screen. Attach your program and the output matrix.
b. Find the general solution of system (5.15).
c. Find the dominant eigenvalue and the stable age distribu-
tion.
d. How does the stable age distribution from part (3c) relate
to the xt /yt column in your output from part (3a)?
e. Find the particular solution subject to the initial conditions
x0 = 35 and y0 = 26.
f. Evaluate the particular solution in part (3e) at t = 5 to
check the values of x5 and y5 that you obtained by iteration
in part (3a).
g. Consider the particular solution you found in part (3e). If
you wrote it in vector form, rewrite it in component form.
Use the closed-form expressions for xt and yt to write a
closed-form expression for xt /yt . Now use the methods of
calculus to compute limt→∞ xt /yt . How does this limit relate
to the stable age distribution you found in part (3c)?

4. Prove that if λ1 is the dominant eigenvalue for system (5.10), then

the strong ergodic property holds, that is, limt→∞ λ1t x (t) = c1 v1 .
1

5. Consider a population categorized into three age classes: juveniles

(ages zero to one year), one-year-olds (ages one to two years),
and two-year-olds (ages two to three years). Suppose individuals
mature sexually at one year and live no longer than three years.
Suppose that the average per capita number of births per year for
one-year-olds is 4 and the average per capita number of births per
year for two-year-olds is 2/3. Suppose also that 60% of juveniles
do not survive until sexual maturity, and that 15% of one-year-
olds do not survive to become two-year-olds.
106 ■ Mathematical Modeling in Biology

a. Draw and label a Leslie diagram for this population.

b. Give the Leslie matrix model for this population.
c. Suppose the population starts with 10 two-year-olds and
no individuals of any other age. Iterate the Leslie model to
find the number of animals in each age class and the total
population size after 3 years.
d. Find the long-term growth rate (dominant eigenvalue) and
the stable age distribution (corresponding eigenvector).
e. Find the general solution for the Leslie model.
f. Find the particular solution, given the initial condition in
(5c).
g. What is the long-term fate of the population?

6. Consider a stage-structured juvenile-adult system in which the

time step is chosen to be the duration of the juvenile stage.
Assume that after one time step, 75% of the juveniles survive
to become sexually mature adults. Suppose that the average per
capita birth rate for adults is b = 1/6, and that 25% of adults
survive each time step.

a. Draw and label a Leslie diagram for this system.

b. Give the Leslie matrix model for this system.
c. Find the general solution.
d. What is the long-term fate of this population, according to
the Leslie model prediction?
e. Suppose we begin with a juvenile density of 3, but with
0 adults. Find the particular solution for the system. Use
a computer to iterate the Leslie model to t = 20. Graph
the numbers of juveniles and adults in the population as a
function of time. Attach your graph.
f. What is the stable age distribution for this population?

7. In this problem, you will analyze a juvenile-adult model for

a cannibalistic insect population. The model is a discrete-time
model, with a time step of one week. The larvae (juveniles) of
 Higher-Dimensional Discrete-Time Models ■ 107

this insect damage agricultural crops. The model with initial

conditions is

Jt+1 = bAt exp (−cej Jt − cea At )

At+1 = Jt exp (−cja At ) + (1 − µa ) At (5.16)
J0 = 500 juveniles per unit area
A0 = 50 adults per unit area

a. Write down a list of assumptions on which model (5.16) is

based.
b. Draw and label a Leslie diagram for model (5.16).
c. List the parameters in model (5.16).
d. Write down the fixed point equations for model (5.16).
e. From your answer in (7d), prove that the extinction state
(J, A) = (0, 0) is an equilibrium.
f. Find the linearization of model (5.16) at the extinction state
(0, 0) by hand. Hint: Your answer must be a linear system
of difference equations. It will have parameters in it.
g. Find conditions on the parameters that guarantee the
stability/instability of the extinction state.
h. Define the inherent net reproductive number n to be n =
b/µa . Because the expected reproductive lifespan of an adult
is 1/µa , we can interpret n to be the expected number of
offspring of an adult over its lifetime. Rewrite your stability
conditions in (7g) in terms of the inherent net reproductive
number n, and give a biological interpretation of these
conditions.
i. Write a computer program to simulate (iterate) model
(5.16). Attach your program.
j. Suppose you gather data to parameterize the model, and
estimate that b = 30, cej = cea = cja = 0.01, and µa = 0.02.
Use your program to graph the predicted time series for
the number of juveniles over the first 30 weeks. Attach your
graph. What is the long-term fate of the population? Is this
consistent with your answer in (7g)?
108 ■ Mathematical Modeling in Biology

k. In an effort to reduce crop damage, an agricultural agent

applies an insecticide that kills adult insects. Suppose this
raises the adult death rate to µa = 0.20 while leaving
the other parameters unchanged. Re-graph the time series.
What change occurs in the population dynamics? What
changes in the insect infestation does the farmer observe?
How will these changes affect crop damage?
l. The agricultural agent now applies a higher level of
insecticide so that µa = 0.98. (The adult death rate is
98%!) Graph the time series again. What change occurs
in the population dynamics? What changes in the insect
infestation does the farmer observe?
m. Use the computer to draw a bifurcation diagram for larval
numbers, using µa as a bifurcation parameter with the range
0 ≤ µa ≤ 1. Use this diagram to explain why an insecticide
that affects only the adult death rate will probably not
reduce crop damage.
n. The agricultural agent now decides to use a different kind
of insecticide—one that reduces the birth rate b. Given that
the natural value of µa is 0.02, how much must the scientist
reduce the value of b to cause extinction of the insect pest?
Hint: Use your answer in (7g).
o. Explain mathematically, from inspecting model (5.16), why
reducing the number of adults can actually increase crop
damage.
p. Explain biologically why reducing the number of adults can
actually increase crop damage.
q. How did this exercise affect your thinking about population
dynamics and mathematical modeling?
r. Consider once again your fixed point equations in (7d) with
arbitrary parameter values. Prove that if b > µa , then there
exists a unique nonzero fixed point. Hint: Two expressions
are equal where their curves intersect.
s. Let b = 30, cej = cea = cja = 0.01, and µa = 0.02.
Numerically estimate the value of the nonzero fixed point.
Attach your program.
 Higher-Dimensional Discrete-Time Models ■ 109

t. Find the linearization of the system at the nonzero fixed

point in (7s). Your answer must be a linear system of
difference equations.
u. Is this nonzero equilibrium stable or unstable?

BIBLIOGRAPHY
Caswell, H. 2001. Matrix Population Models: Construction, Analysis, and
Interpretation, 2nd ed. Sinauer Associates, Sunderland, MA. [Caswell’s
classic book should be in the library of every ecologist and mathematical
biologist.]

Cushing, J. M., Costantino, R. F., Dennis, B., Desharnais, R. A., and

Henson, S. M. 2003. Chaos in Ecology: Experimental Nonlinear Dynamics.
Academic Press, San Diego, CA. [Summarizes the large body of work
involved in the documentation of chaos in a laboratory population. In
Chapter 6 we will draw heavily on this book to showcase how dynamic
population data can be connected to mathematical models.]
 CHAPTER 6

Flour Beetle Dynamics:

A Case Study

6.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

This chapter is based on a body of research conducted by a
collaboration of scientists and mathematicians known as the “Beetle
Team.” The Beetle Team used laboratory populations of flour beetles
(Tribolium castaneum) to test nonlinear dynamics theory. “The
Beetles” initially were composed of ecologist R. F. (Bob) Costantino
of the University of Rhode Island; mathematician J. M. (Jim) Cushing
of the University of Arizona; statistical ecologist Brian Dennis of the
University of Idaho; and ecologist R. A. (Bob) Desharnais of California
State University, Los Angeles. Later, mathematicians Shandelle M.
Henson and Aaron A. King, then postdocs at the University of Arizona,
joined the Team. Key research papers authored by the group, including
those on which this chapter is based, are found in this chapter’s
bibliography.
The ideas, techniques, and coding in this case study are similar to
those in the avian bone growth case study in Chapter 2, except that
here you will be applying the ideas to a discrete-time dynamical system
with three state variables. You will need the skills from Chapter 2 and
those from all three appendices.

6.2 FLOUR BEETLES

Flour beetles of the genus Tribolium are small insects of the family
Tenebrionidae. They can be found occasionally in their native forest

DOI: 10.1201/9781003265382-8 111

112 ■ Mathematical Modeling in Biology

Figure 6.1: T. castaneum pupa and adult. (Used with permission of

R. F. Costantino.)

habitats, but for thousands of years their preferred habitat has been
milled flour and other stored grain products. These insects are hardy
and prolific animals, traits which make them important agricultural
pests and quite resistant to eradication. These same traits, however,
make them easy to culture and manipulate in the laboratory. Flour
beetles have been used extensively as animal models in genetics and
population ecology. The details of their life cycle are well known. A
comprehensive reference for the life history of flour beetles is Alexander
Sokoloff’s three-volume book “The Biology of Tribolium” (Sokoloff
1972; Sokoloff 1975; Sokoloff 1978). A fairly comprehensive history
of the mathematics of Tribolium prior to the work presented in this
chapter can be found in Costantino and Desharnais (1991).
Flour beetles undergo complete metamorphosis, which encompasses
four life-cycle stages: egg, larva, pupa, and adult. The tiny eggs—as
many as 500 from a single female—hatch in approximately four days
under laboratory conditions and can barely be seen by the unaided
eye. The larvae feed voraciously and eventually grow to the size of rice
grains. After about 14 days, a larva encloses itself in a pupal case,
which consists of a thin outer covering, the pupal cuticle. During the
pupal stage, the individual completely reorganizes itself to become an
adult (Figure 6.1). The new adult emerges another 14 days later and
sheds its pupal case as little pieces of dried tissue called frass.
One might expect that populations of flour beetles, when cultured
under constant environmental conditions, would grow exponentially
 Flour Beetle Dynamics: A Case Study ■ 113

at first and then equilibrate due to resource limitations. But they do

not. Instead, their numbers typically exhibit sustained oscillations. The
mechanism that regulates population size and causes these oscillations
is cannibalism. In general, the mobile stages (larvae and adults)
cannibalize the immobile stages (eggs and pupae). These cannibalistic
encounters are crucial to the understanding of flour beetle population
dynamics.
Cannibalism occurs in a broad range of taxa. At least 1,300 species,
including protozoans, insects, reptiles, fish, birds, and mammals, en-
gage in cannibalism. Some key references on cannibalism include Dong
and Polis (1992); Elgar and Crespi (1992); Fox (1975); and Polis (1981).
Cannibalism often is associated with low food supply. In flour
beetles, however, cannibalism does not appear to be a significant source
of nutrition, nor does it appear to be caused by limitations of food or
space. Instead, the incidence of this behavior is under genetic control.
Varying degrees of cannibalism can be artificially selected for in the
laboratory to create different genetic strains, some highly cannibalistic,
and others essentially noncannibalistic. As far as the modeling process
is concerned, the Beetle Team assumed Tribolium cannibalism occurs
at random, as the mobile larvae and adults encounter immobile eggs
and pupae. Under this assumption, the rate of cannibalism in Tribolium
should be inversely proportional to habitat volume; this was established
theoretically in Henson and Cushing (1997) and tested empirically
using periodic habitat volumes in the laboratory (Henson et al. 1999).

6.3 DATA

The time series in Data Set 6.1 are the control replicates from
laboratory experiments conducted by R. A. Desharnais in 1978 at the
University of Rhode Island when he did his master of science in zoology
with R. F. Costantino (Desharnais and Costantino 1980). Populations
of the corn oil-sensitive strain of Tribolium castaneum (Herbst) were
cultured in half-pint milk bottles containing 20 g of corn oil media
(90% wheat flour, 5% dried brewer’s yeast, and 5% liquid corn oil).
The cultures were kept in an unlighted incubator at 33◦ C±1◦ C and
a relative humidity of 56% ± 11%. Every two weeks, all life-cycle
stages (except eggs) were censused, and the animals (including eggs)
were returned to the incubator in fresh media. Small (feeding) larvae
114 ■ Mathematical Modeling in Biology

were counted as L-stage; large (nonfeeding) larvae, pupae, and callow

(immature) adults were counted as P-stage; and sexually mature adults
were counted as A-stage. In this chapter, we refer to L-stage animals as
“larvae,” P-stage animals as “pupae,” and A-stage animals as “adults”
for the sake of simplicity. These data are also available in Appendix A
of the book by Cushing et al. (2003), along with other Tribolium data
sets.

6.4 ASSUMPTIONS
A major goal of modeling, often called the “art of modeling,” is
to construct a simple, mechanistic, low-dimensional (i.e., few state
variables) model that describes enough of the observed variability so
that the leftover, “higher-order” variability is relatively small and can
be described statistically as “noise.” The beetle system has four life
stages (egg, larval, pupal, and adult) and thus might require four state
variables. The egg stage duration, however, is much shorter than that
of the other three life stages. Following the KISS principle (keep it
simple stupid), here we model the beetle system with two alternative
three-dimensional models. We will select the best of these two models
and attempt to validate it (Exercises 5–8).
The Beetle Team used the following assumptions.

6.4.1 Deterministic Assumptions

The deterministic assumptions were as follows:

(A1) Given that the duration of the egg stage is much shorter than
the duration of the larval and pupal stages, we can ignore numbers
of eggs and treat egg cannibalism as a factor that dampens
recruitment into the larval stage.

(A2) In the absence of cannibalism, the per capita larval

recruitment rate is a constant b > 0 larvae per adult per unit
time.

(A3) In the presence of cannibalism, the larval recruitment rate is

decreased due to the cannibalism of eggs by larvae. It is decreased
by a Ricker factor of the form e−cel L in the presence of L larvae.
 Flour Beetle Dynamics: A Case Study ■ 115

(A4) In the presence of cannibalism, the larval recruitment rate is

decreased due to cannibalism of eggs by adults. It is decreased by
a Ricker factor of the form e−cea A in the presence of A adults.
(A5) A fraction µl of larvae die per unit time, where 0 < µl < 1.
(A6) In the absence of cannibalism, no pupae die per unit time.
(A7) In the presence of cannibalism, pupal survivorship per unit
time is decreased by a fraction e−cpa A in the presence of A adults
due to the cannibalism of pupae by adults.
(A8a) There is no cannibalism of pupae by larvae.
(A8b) (Alternate hypothesis) In the presence of cannibalism,
pupal survivorship per unit time is decreased by a fraction e−cpl L
in the presence of L larvae due to the cannibalism of pupae by
larvae.
(A9) A fraction µa of adults die per unit time, where 0 < µa < 1.

6.4.2 Stochastic Assumptions

The stochastic assumptions were as follows:
(A10) There is no measurement error.
(A11a) Environmental stochasticity is the main source of noise in
the system.
(A11b) (Alternate hypothesis) Demographic stochasticity is the
main source of noise in the system.
(A12) Stochastic perturbations are uncorrelated in time. We
also assume, for simplicity, that stochastic perturbations are
uncorrelated across life-cycle stages.
Although, for the sake of simplicity, we will assume in this chapter
that noise is uncorrelated across life-cycle stages, assumption (A12), it
is important to note that for environmental noise, covariances between
the life stages are likely. See Dennis et al. (1995) for details on
estimating a covariance matrix when parameterizing the LPA model
under the assumption of environmental stochasticity. Here, we assume
the covariance matrix is negligible.
116 ■ Mathematical Modeling in Biology

6.5 ALTERNATIVE DETERMINISTIC MODELS

Let
Lt = Number of larvae at time t
Pt = Number of pupae at time t
At = Number of adults at time t.
The assumptions (A1)–(A9), using (A8a), can be conceptualized in the
Leslie diagram shown in Figure 6.2. The corresponding nonlinear Leslie
model is known as the LPA model :

Lt+1 = bAt e−cel Lt −cea At

Pt+1 = (1 − µl ) Lt (6.1)
−cpa At
At+1 = Pt e + (1 − µa ) At
with time step two weeks, the duration of each of the larval and pupal
stages.
The assumptions (A1)–(A9), using the alternative hypothesis
(A8b), give rise to the slightly more complicated alternative model

Lt+1 = bAt e−cel Lt −cea At

Pt+1 = (1 − µl ) Lt (6.2)
−cpl Lt −cpa At
At+1 = Pt e + (1 − µa ) At ,
which we will call the “LPAalt model” for the purposes of this chapter.

Figure 6.2: Leslie diagram for T. castaneum. The per capita transition
rate is shown beside each arrow. The time step is 2 weeks.
 Flour Beetle Dynamics: A Case Study ■ 117

6.6 STOCHASTIC MODELS

Given a discrete-time model of the form

xt+1 = f (xt ) , (6.3)

where xt is a real number or a vector of real numbers, how does one

include stochasticity? If noise is additive Gaussian, then the stochastic
version of the model is

Xt+1 = f (Xt ) + σEt , (6.4)

where Xt is a random variable, the function f is the deterministic

predictor, σ > 0, and Et is a standard normal random variable (mean
zero and standard deviation one). The standard deviation of the noise
(that is, of the random variable σEt ) is σ. When σ is set to zero, the
deterministic skeleton (6.3) is recovered. Stochastic models of the form
(6.4) are called nonlinear autoregressive (NLAR) models and have nice
statistical properties.
As a matter of fact, however, process noise, whether environmental
or demographic, is not in general additive. In general, data and pre-
dictions must be transformed so that the noise becomes approximately
additive on some scale. A result from statistical theory, which we will
simply reference here, makes it possible to approximate environmental
and demographic noise with NLAR models. First, environmental noise
is approximately additive on the log scale, that is,

ln Xt+1 = ln (f (Xt )) + σEt ,

whereas demographic noise is approximately additive on the square root

scale (Cushing et al. 2003), that is,
p p
Xt+1 = f (Xt ) + σEt .

In general, if a stochastic system with deterministic skeleton (6.3)

can be approximated by an NLAR model of the form

ϕ (Xt+1 ) = ϕ (f (Xt )) + σEt

for some function ϕ, then ϕ is called a variance-stabilizing transforma-

tion for that system.
118 ■ Mathematical Modeling in Biology

Under assumption (A11a), environmental stochasticity is the main

source of variability in the system. The corresponding NLAR LPA
model is therefore

ln (Lt+1 ) = ln bAt e−cel Lt −cea At + σL ELt

ln (Pt+1 ) = ln ((1 − µl ) Lt ) + σP EP t
ln (At+1 ) = ln Pt e−cpa At + (1 − µa ) At + σA EAt ,

where the Eit are standard normal random variables (mean zero and
standard deviation one). This model can be rewritten as
Lt+1 = bAt e−cel Lt −cea At eσL ELt
Pt+1 = (1 − µl ) Lt eσP EP t (6.5)
−cpa At σA EAt


At+1 = Pt e + (1 − µa ) At e .
This is the “environmental noise LPA model.” The “environmental
noise LPAalt model” is similar (Exercise 3).
If the demographic stochasticity were the main source of variability
in the system, assumption (A11b), the appropriate NLAR LPA model
would be
p p
Lt+1 = bAt e−cel Lt −cea At + σL ELt
p p
Pt+1 = (1 − µl ) Lt + σP EP t
p q
At+1 = Pt e−cpa At + (1 − µa ) At + σA EAt ,
where again the Eit are standard normal random variables. This model
can be rewritten as
p 2
Lt+1 = bAt e−cel Lt −cea At + σL ELt
p 2
Pt+1 = (1 − µl ) Lt + σP EP t (6.6)
q 2
At+1 = Pt e−cpa At + (1 − µa ) At + σA EAt ,

where the quantities inside the squares on the right-hand sides of

equation (6.6) are set to zero if they are negative. This is the
“demographic noise LPA model.” The “demographic noise LPAalt
model” is similar (Exercise 3).
 Flour Beetle Dynamics: A Case Study ■ 119

6.7 MODEL PARAMETERIZATION

In order to decide whether the LPA model (6.1) or the LPAalt model
(6.2) is better, we must fit both models to the data in Data Set 6.1.
In this section, we will focus on the procedure for parameterizing the
LPA model.
There are six parameters in the deterministic LPA model. We write
them here in the order they appear in the model: b, cel , cea , µl , cpa , µa .
We will use the historical beetle data in Data Set 6.1 to estimate the
values of these parameters using the method of nonlinear conditioned
least squares (CLS). Ideally, we should randomly divide the data
set into two parts. One part would be used for model calibration
(parameter estimation), and the other part would be reserved for
independent model evaluation (validation). In this case, however, we
are estimating six parameters, so we need a lot of data to calibrate
the model. We will therefore use all of the historical data for model
calibration.

6.7.1 Conditioned One-Step Residuals

Let (lt , pt , at ) and (lt+1 , pt+1 , at+1 ) be the actual observed data triples
at times t and t + 1. Since we are thinking of the stochastic model as a
surrogate for the biological system, we can think of these data triples
as realizations of the random variable triple (L, P, A). The conditional
one-step residual errors at time t + 1 are the (transformed) actual
measurements (lt+1 , pt+1 , at+1 ) at time t + 1 minus the (transformed)
values predicted by the deterministic skeleton, given the data point
(lt , pt , at ) at time t. For example, for the environmental noise LPA
model, we compute the log residuals

εLt = ln (lt+1 ) − ln bat e−cel lt −cea at

εP t = ln (pt+1 ) − ln ((1 − µl ) lt )
εAt = ln (at+1 ) − ln pt e−cpa at + (1 − µa ) at .

(There is a computational caveat here: If a state variable is zero, the

program that computes the log residuals will crash. You will need to
replace the zeros with a positive number less than one, for example 0.5.)
For the demographic noise LPA model, we would compute the square
120 ■ Mathematical Modeling in Biology

root residuals. In general, the residuals on the transformed scale would

be

εLt = ϕ (lt+1 ) − ϕ bat e−cel lt −cea at

εP t = ϕ (pt+1 ) − ϕ ((1 − µl ) lt )
pt e−cpa at + (1 − µa ) at



εAt = ϕ (at+1 ) − ϕ .

Note that the residuals εit themselves are realizations of the random
variables σi Eit in the stochastic models.
Because each parameter appears in only one equation and we are
assuming no covariances, the likelihood function is the product

! ! !
ε2 ε2 ε2
Y 1 − Lt
2
Y 1 − P2t 1 Y− At2
L= √ e 2σL √ e 2σP √ e 2σA ,
σL 2π σP 2π σA 2π
(6.7)
assuming the residuals are normally distributed. As in Chapter 2,
maximizing the likelihood in this case is equivalent to using the method
of CLS to parameterize the model. One nice thing to know about CLS
is that it loosens the restrictive assumptions on the residuals; CLS
parameter estimates converge to the true values even if the noise is
non-normal and autocorrelated, as long as the noise has a stationary
distribution (Cushing et al. 2003; Tong 1990).

6.7.2 Conditioned Least Squares (CLS)

Suppose the data time series is of duration t = 0, 1, 2, . . . , q. The

residual sums of squares are

q−1
X
RSSL = ε2Lt
t=0
q−1
X
RSSP = ε2P t
t=0
q−1
X
RSSA = ε2At .
t=0
 Flour Beetle Dynamics: A Case Study ■ 121

Thus, for the environmental noise LPA model, we have

q−1
X
2
ln (lt+1 ) − ln bat e−cel lt −cea at

RSSL (b, cel , cea ) =
t=0
q−1
X 2
RSSP (µl ) = [ln (pt+1 ) − ln ((1 − µl ) lt )]
t=0
q−1
X
2
ln (at+1 ) − ln pt e−cpa at + (1 − µa ) at

RSSA (cpa , µa ) = .
t=0

Because each parameter appears in only one equation, the three RSS
values are minimized if and only if the sum

RSS (b, cel , cea , µl , cpa , µa ) = RSSL (b, cel , cea )+RSSP (µl )+RSSA (cpa , µa )

is minimized as a function of the six parameters. The minimizing

parameters are called the conditioned least squares (CLS) parameter
estimates. The estimated variances are given by

2 RSS
\ L RSS
\ P RSS
\ A
σ
bL = bP2 =
; σ ; σ
bA2
= ,
q q q

where RSS
d denotes the fitted value of RSS.
The parameters for the environmental noise LPA model and the
environmental noise LPAalt model, estimated from the data in Data
Set 6.1 and reported to four significant figures, are listed in Tables 6.1–
6.4. In Exercise 5, you will complete Tables 6.1–6.4 by parameterizing
the demographic noise models.
Note that the estimated values of the six parameters held in
common by the environmental noise LPA and LPAalt models are
similar; in fact, to four significant figures, they differ only in the values

TABLE 6.1 LPA Model Parameters Estimated from Data Set 6.1

LPA model b
b cbel cbea µ
bl cbpa µ
ba
Environmental
noise 10.57 0.009463 0.009649 0.5129 0.01817 0.1065
Demographic
noise
122 ■ Mathematical Modeling in Biology

TABLE 6.2 LPA Model Parameters Estimated from Data Set 6.1
2
LPA model RSS
\L RSS
\P RSS
\A RSS
d σ
bL b P2
σ σ
bA2

Environmental
noise 20.99 32.56 0.8440 54.39 0.2762 0.4284 0.01111
Demographic
noise

TABLE 6.3 LPAalt Model Parameters Estimated from Data Set 6.1
LPAalt model b
b cbel cbea µ
bl cbpl cbpa µ
ba
Environmental
noise 10.57 0.009463 0.009649 0.5129 0.0007201 0.01782 0.1025
Demographic
noise

TABLE 6.4 LPAalt Model Parameters Estimated from Data Set 6.1
2
LPAalt model RSS
\L RSS
\P RSS
\A RSS
d σ
bL b P2
σ σ
bA2

Environmental
noise 20.99 32.56 0.8401 54.39 0.2762 0.4284 0.01105
Demographic
noise

of cpa and µa . (This is because the cpl parameter affects only the third
equation.) The introduction of the parameter cpl in the LPAalt model
slightly decreases the value of cpa and also µa . This suggests that
there may be some cannibalism of pupae by larvae, but not very much
(because cpl is small compared to the other cannibalism coefficients).
The question becomes: Is cannibalism of pupae by larvae important
enough that we must include it in the model? To answer this question,
we next turn to the methods of model selection.

6.8 MODEL SELECTION

We will select the best of the two alternative models (LPA and LPAalt)
by using the Akaike information criterion (AIC). The AIC presupposes
that all alternative models under consideration have the same number
of state variables and that they are all parameterized on the same data
set. Both of these conditions apply for the LPA and LPAalt models.
 Flour Beetle Dynamics: A Case Study ■ 123

The AIC, in its most general formulation, is

AIC = −2 ln L + 2κ, (6.8)

where L is the maximized value of the likelihood function and κ is the

number of estimated parameters in the model, including the variance
and covariance parameters. The model with the smallest AIC is the
best of the alternatives.
Given assumption (A12) and the fact that each parameter in the
deterministic skeleton appears in only one equation, the AIC is equal
to (Exercise 4)

p
X
bi2 + 2κ,


AIC = pq(ln(2π) + 1) + q ln σ (6.9)
i=1

where p is the number of state variables (p = 3 in our case), q is

bi2 are the estimated variances
the number of one-step transitions, the σ
for the state variables, and κ is the number of parameters, which
includes the parameters in the skeleton plus the number of variances.
For the LPA model, κ = 6 + 3 = 9, whereas for the LPAalt model,
κ = 7 + 3 = 10.
In Exercise 6, the reader will complete Table 6.5. Here, ∆AIC =
AIC−AICmin . The best model (the one with AICmin ) will have ∆AIC =
0, and the second-best model will have a value of ∆AIC greater than
zero.
In completing Table 6.5 (Exercise 6), the reader will find that the
LPA model is the best of the two alternatives. Thus, although there
may be some cannibalism of pupae by larvae, that interaction can and
should be ignored in the model. This means that we now discard the
LPAalt model and focus our attention on validating the LPA model.

TABLE 6.5 Model Selection Based on Parameters

Estimated from Data Set 6.1

Model AIC ∆AIC

LPA, environmental noise 160.8
LPAalt, environmental noise
124 ■ Mathematical Modeling in Biology

6.9 MODEL VALIDATION

To validate a model, we first compute the goodness of fit on estimation
data, and then, without re-estimating parameters, we compute the
goodness of fit on an independent data set. To measure goodness of
fit, we define a generalized R2 for each state variable (Dennis et al.
2001). For example, for the environmental noise LPA model, the R2
for state variable L is
−1 
qP
2
ln (lt+1 ) − ln bat e−cel lt −cea at
2 t=0
RL =1− q−1 i2 ,
Ph
ln (lt+1 ) − ln (l)
t=0

where ln (l) denotes the sample mean of the log-transformed observed

L-stage abundances. A similar R2 formula holds for P and A. In
general, the R2 for state variable L on the ϕ-scale is
−1 
qP
2
ϕ (lt+1 ) − ϕ bat e−cel lt −cea at
2 t=0
RL =1− −1 h
qP i2 ,
ϕ (lt+1 ) − ϕ (l)
t=0

where ϕ (l) denotes the sample mean of the ϕ-transformed observed

L-stage abundances. A similar R2 formula holds for P and A.
Data Set 6.2 contains part of a data set collected in 1993–1994 by
R. F. Costantino when he was on sabbatical at the lab of Alan Hastings
at the University of California at Davis. In this experiment, the adult
death rate µa was manipulated at the values 0.04, 0.27, 0.5, 0.73, and
0.96. Data Set 6.2 contains the control and the treatments for µa = 0.5
and µa = 0.96. The entire data set is listed in Appendix B of the book
by Cushing et al. (2003).
Table 6.6 gives the R2 values for the estimation data in Data Set
6.1 and the R2 values (without refitting) for the independent data in
Data Set 6.2 (Exercise 8). Note that we do not compute validation R2
for the LPAalt model, because that model has already been discarded.
A comparison of the R2 values in Table 6.6 suggests that the
LPA model under the assumption of environmental noise is not well
validated for the L-stage. This is not surprising, because collection of
the estimation and validation data sets occurred over a decade apart
 Flour Beetle Dynamics: A Case Study ■ 125

TABLE 6.6 Model Validation

Environmental Noise LPA Model Estimation R2 Validation R2

L 0.80 0.42
P
A
Parameters were estimated on Data Set 6.1. Validation was attempted on Data Set
6.2 without re-estimating parameters.

in different laboratories using different culture protocols and different

strains of Tribolium castaneum. What is surprising, however, is that
the validation R2 values are positive and quite high for ecological data.
In Dennis et al. (1997), the entire data set on which Data Set 6.2
is based was randomly divided into estimation data and validation
data. The subsequently well-validated model documented a transition
between equilibria (control), 2-cycles (µa = 0.5), and aperiodic cycles
(µa = 0.96) for the treatments in Data Set 6.2 (Exercise 9).
In the next section, we revisit a rigorous historical validation of the
LPA model.

6.10 THE “HUNT FOR CHAOS” EXPERIMENTS

Let’s back up to the beginning of the story. After Lord Robert May
proposed the hypothesis that complex fluctuations in populations
could be due to deterministic feedback mechanisms (see Chapter 4),
researchers set out to search for chaos in population time series data.
Most of the data sets were from field systems (Cushing et al. 2003).
The multidisciplinary Beetle Team took a different approach. If they
located a biologically feasible region in parameter space in which a well-
validated model predicted chaotic population dynamics, and placed
experimental laboratory populations in that parameter region, would
the observed experimental dynamics be chaotic? If one could not
demonstrate chaos in laboratory populations, how could one hope to
demonstrate it in natural populations? Therefore, they decided that
the starting place for the hunt for chaos should be in the laboratory.
Costantino and Desharnais had studied flour beetle dynamics
in the lab for many years prior to the Beetle Team collaboration.
Based on their thorough understanding of the Tribolium system, the
Team constructed a mathematical model that incorporated the main
126 ■ Mathematical Modeling in Biology

mechanisms thought to drive the dynamics. This was the origin of the
LPA model. Using historical time series data, the Team parameterized
and validated the model (as explained in the previous part of this
chapter) and then used the fitted model to make predictions about
how the system dynamics would respond to various experimental
manipulations of the parameters. Based on the model predictions, they
designed and carried out experiments to test these predictions.
Suppose we set the LPA parameters to be the CLS estimates for
the demographic LPA model from Dennis et al. (2001): b = 10.45,
cel = 0.01731, cea = 0.01310, µl = 0.2000, cpa = 0.004619, µa =
0.007629. If the experimenter could fix the adult death rate µa at 0.96
and manipulate cpa at various values between 0 and 1 in the laboratory,
what dynamics would we predict to see in the beetle populations?
The easiest way to answer this question is to have the computer draw
a bifurcation diagram with cpa as the bifurcation parameter, using
µa = 0.96 and all other parameter values as listed directly above. The
bifurcation diagram for total population size (L + P + A) is shown in
Figure 6.3 (Exercise 10).

Figure 6.3: Demographic LPA model bifurcation diagram using CLS

parameter estimates from Dennis et al. (2001): b = 10.45, cel = 0.01731,
cea = 0.01310, µl = 0.2000, µa = 0.96. The bifurcation parameter cpa
ranges from 0 to 1.
 Flour Beetle Dynamics: A Case Study ■ 127

To test the predictions of the bifurcation diagram in Figure 6.3, the

Beetle Team designed a way to manipulate µa and cpa in the laboratory.
They chose seven values of cpa (0.00, 0.05, 0.10, 0.25, 0.35, 0.50,
and 1.00) at which to run experimental treatments, replicating each
treatment three times. They also had a control with three replicates,
in which cpa was not manipulated. The parameter µa was manipulated
to be 0.96 in all 24 populations. The LPA model, with all other
parameters set to those in the caption of Table 6.7, estimated under
the assumption of demographic noise, predicts the dynamics shown in
Table 6.7. Of particular interest are the transitions between dynamic
states, for example, from an equilibrium (control) to chaos (cpa = 0.35)
to a 3-cycle (cpa = 0.5). That is, the LPA model predicts not only
attractors, but also transitions, as cpa is “tuned” from 0 to 1. The idea
was to document transitions and to “bookend” complicated dynamics
in the middle range of cpa with simple dynamics at each end of the cpa
range.
The plan was for Bob Costantino to count L, P , and A numbers in
the 24 populations every two weeks for 80 weeks and then to compare
the observed dynamics to the predicted dynamics. At the end of this
section, we discuss the general idea of how µa and cpa were manipulated
in the laboratory; details are given in the papers by Costantino et al.
(1997) and Dennis et al. (2001).

TABLE 6.7 LPA Model Predictions. Parameters

are CLS estimates for the demographic LPA
model from Dennis et al. (2001): b = 10.45,
cel = 0.01731, cea = 0.01310, µl = 0.2000,
cpa = 0.004619, µa = 0.007629. Compare with
Figure 6.3.

cpa Prediction with µa = 0.96

Control Equilibrium
0.00 Double invariant loop
0.05 Chaos
0.10 26-cycle
0.25 8-cycle
0.35 Chaos
0.50 3-cycle
1.00 6-cycle
128 ■ Mathematical Modeling in Biology

6.10.1 Results of the “Hunt for Chaos” Experiments

The observed dynamics for the “Hunt” experiments are found in Data
Set 6.3. For a complete comparison of the data with model-predicted
dynamics, see Costantino et al. (1997) and Dennis et al. (2001). Here we
show data for the control and three treatments (cpa = 0.0, cpa = 0.35,
and cpa = 0.5; Figure 6.4). For the control, we can see that the data
are clustered around the model-predicted equilibrium (Figure 6.4a). For
cpa = 0.5, we see that the data are clustered about the three points of
the model-predicted 3-cycle (Figure 6.4d). These are the “bookends.”
For cpa = 0.0, we see that the data have parted into two clusters around
the model predicted double invariant loop (Figure 6.4b). For cpa = 0.35,
the data are strung out along the model-predicted chaotic attractor
(Figure 6.4c). The data patterns in each panel are quite different. The
strongest evidence that the model is working comes from the fact that
it predicted the transitions between a simple equilibrium, complicated
dynamics, and a simple 3-cycle.
The Beetle Team announced the results of the Hunt for Chaos
experiments in Science in 1997 (Costantino et al. 1997), followed
by a comprehensive paper published in Ecological Monographs in
2001 (Dennis et al. 2001). Costantino continued the Hunt for Chaos
experiment for 8 years; the data set eventually reached a length of 424
weeks. The Team conducted many more experiments of various types
in which they further tested the nonlinear dynamic predictions of the
LPA model; you can find these in the annotated references at the end
of this chapter. The LPA model has been the most rigorously validated
mathematical population model in the history of ecology to date.

6.10.2 Manipulating the Parameters in the Laboratory

Here, we give the general idea of how µa and cpa were manipulated in
the laboratory; details are given in Dennis et al. (2001).
The adult mortality rate was manipulated to be µa = 0.96 and
the adult recruitment rate was manipulated so that it would equal
Pt e−cpa At . To manipulate the adult death rate µa , at each census
Costantino counted dead adults and then removed or added older live
adults to the population as necessary. To manipulate cpa , he further
removed or added younger adults to make the third equation in model
(6.1) exact. For example, at census t + 1 for a cpa = 0.10 culture, he
 Flour Beetle Dynamics: A Case Study ■ 129

(a) (b)

(c) (d)

Figure 6.4: Data (open circles) and deterministic model predictions

(solid circles) for the LPA model. Parameters are CLS estimates for the
demographic noise LPA model from Dennis et al. (2001): b = 10.45,
cel = 0.01731, cea = 0.01310, µl = 0.2000, cpa = 0.004619, µa =
0.007629. (a) Control. Model predicts an equilibrium. (b) cpa = 0.
Model predicts a double invariant loop. (c) cpa = 0.35. Model predicts
chaos. (d) cpa = 0.5. Model predicts a 3-cycle.

would compute the value of

Pt e−0.10At + (1 − 0.96)At
based on the previous census values of Pt and At . This new value was
the target value for At+1 . Given the observed number of adults at time
t + 1, Costantino removed or added adults to the culture as needed. Of
course, the observed number of adults (not the manipulated number)
was the number recorded in the data time series.
When the experimental results were announced in Science
(Costantino et al. 1997) and the Team began giving talks about the
results at research conferences, there was often a question from the
130 ■ Mathematical Modeling in Biology

audience as to whether “making the third equation exact” was, in fact,

driving the dynamics and forcing them to fit the model predictions.
This is a very important question. It is right to be skeptical in science,
and not only of the work of others. One must also be relentlessly
skeptical of one’s own work. This was a question that demanded
rigorous thinking.
In order to see whether making the third equation of the LPA
model exact was forcing the experimental dynamics to fit the model
predictions, we consider the following thought experiment. Note that
the experimental protocol for manipulating µa and cpa depends only
on the form of the third equation. Thus, in the lab, you would carry out
the exact same experimental protocol in testing any larval-pupal-adult
model whose third equation was of the form

zt+1 = yt e−cpa zt + (1 − µa ) zt .

As an extreme example, we can imagine that one could have been

testing the following model (rather than the LPA model):

xt+1 = 0
yt+1 = 0
zt+1 = yt e−cpa zt + (1 − µa ) zt .

This ridiculous model predicts zero larvae and pupae for all t > 0, along
with a geometrically decreasing number of adults. Nevertheless, the
experimental results would have been the same as in Figure 6.4! Clearly,
“making the third equation exact” does not force the experimental
dynamics to follow the predictions of the model.
Can you think of a situation in which the experimental approach of
“making one equation exact” would not be legitimate? See Exercise 11.

6.11 EXERCISES
1. In this problem, you will carry out a stability analysis for the
extinction state of the LPA model (6.1).

a. Prove that the extinction state (0, 0, 0) is an equilibrium for

the LPA model.
 Flour Beetle Dynamics: A Case Study ■ 131

b. Find the linearization of the LPA model at the equilibrium

(0, 0, 0).
c. Find the characteristic equation of the linearization. Hint:
If you do not know how to find the determinant of a 3 ×
3 matrix, look up the procedure in an introductory linear
algebra text or calculus text.
d. Can you find conditions on the parameters that guarantee
the stability/instability of the extinction state (0, 0, 0)?
e. If you were able to complete part (1d), define the inherent
net reproductive number n by n = b (1 − µl ) /µa . This is
the expected number of adults arising from one adult in the
absence of cannibalism. Rewrite your conditions from part
(1d) in terms of n. Give a biological interpretation of these
conditions.

2. Consider a two-dimensional beetle model. Here, we will combine

the larvae and pupae into a juvenile stage Jt . The time step
is approximately 4 weeks, the sum of the durations of these
stages. Assume that (a) the egg stage can be ignored; (b) in
the absence of cannibalism, the per capita juvenile recruitment
rate is a constant b > 0 juveniles per adult per unit time;
(c) in the presence of cannibalism, the juvenile recruitment
rate is decreased by a Ricker factor of the form e−cej J due to
cannibalism of eggs by larvae; (d) in the presence of cannibalism,
the juvenile recruitment rate is further decreased by a Ricker
factor of the form e−cea A due to the cannibalism of eggs by adults;
(e) in the absence of cannibalism, there are no juvenile deaths;
(f) in the presence of cannibalism, the juvenile survivorship is
decreased by a Ricker factor of the form e−cja A due to the
cannibalism of juveniles by adults; (g) a fraction µa of adults
die per unit time, where 0 < µa < 1; (h) there is no measurement
error; (i) demographic stochasticity is the main source of noise
in the system; and (j) stochastic perturbations are uncorrelated
across life-cycle stages and uncorrelated in time. Let

Jt = Number of larvae and pupae at time t

At = Number of adults at time t.
132 ■ Mathematical Modeling in Biology

a. Draw the Leslie Diagram for the JA model.

b. Write the equations for the deterministic JA model. Always
state the time step when you pose a discrete-time model.
c. Write the equations for the stochastic JA model assuming
demographic noise.

3. Write equations for the “environmental noise LPAalt model” and

the “demographic noise LPAalt model.”

4. Derive equation (6.9) from equations (6.7) and (6.8).

5. In this problem, you will parameterize the two alternative models

presented in this chapter, the LPA and LPAalt models, using the
data in Data Set 6.1 as the estimation data.

a. Reproduce all the values in Tables 6.1–6.4 for the

environmental noise models.
b. Complete Tables 6.1–6.4 for the demographic noise models.

6. In this problem, you will use the AIC to select the best of the two
alternatives, the LPA and LPAalt models, under the assumption
of environmental noise.

a. Using equation (6.9), complete Table 6.5.

b. Select the best model and explain your selection.

7. In this problem, you will explore the dynamics of the param-

eterized LPA model and make testable predictions. Use the
parameters estimated from the environmental noise LPA model
(Tables 6.1–6.2).

a. Write a program to simulate the deterministic LPA model.

Produce a time series of length 40, given the initial condition
(250, 5, 100). Plot the time series for L, P, and A on the same
graph. Attach (i) your program, (ii) the time series (as a
matrix with 40 rows and 3 columns), and (iii) the graph.
 Flour Beetle Dynamics: A Case Study ■ 133

b. Write a program to simulate the stochastic LPA model with

environmental noise (6.5). Produce a typical stochastic time
series of length 40, given the initial condition (250, 5, 100).
Plot the time series for L, P, and A on the same graph.
Attach (i) your program, (ii) the time series (as a matrix
with 40 rows and 3 columns), and (iii) the graph.
c. Use the computer to graph the LPA model bifurcation
diagram for larval numbers, using µa as a bifurcation
parameter with range 0 ≤ µa ≤ 1.
d. The bifurcation diagram in (7c) is a prediction about the
behavior of the flour beetle dynamics as a function of the
adult death rate µa . What dynamic does the model predict
for µa = 0.1065? For µa = 0.50? For µa = 0.96? Try to
answer these questions simply by looking at the bifurcation
diagram.
e. Imagine planning a laboratory experiment to test the
predictions in part (7d). Your design would involve a
control (in which µa is not manipulated) and perhaps two
treatments in which µa is manipulated at the fixed values
µa = 0.5 and µa = 0.96, with replicates of the control and
each treatment. In order to further explore the predicted
outcomes:
i. Prepare time series graphs that show the deterministic
model predictions for each of your experimental treat-
ments and the control for 40 time steps using the initial
condition (250, 5, 100). Attach graphs of time series for
L and A.
ii. Prepare time series graphs that show typical stochastic
model simulations for each of your experimental
treatments and the control for 40 time steps using the
initial condition (250, 5, 100). Attach graphs of time
series for L and A.

8. In this problem, you will check to see if the LPA model fits Data
Set 6.2 without re-estimating the parameters. Use the parameters
estimated from the environmental noise LPA model (Tables 6.1–
6.2). Note that the experimental manipulations began at week 12
(Data Set 6.2).
134 ■ Mathematical Modeling in Biology

a. Fill in the R2 values for the estimation data (Data Set 6.1)
in Table 6.6.
b. Without re-estimating the model parameters, compute the
goodness of fit for the data in Data Set 6.2. That is, fill in
the R2 values for the validation data in Table 6.6. Hint: This
exercise will require some careful thought and programming.
There are several issues to keep in mind. First, look at Data
Set 6.2 and note that there is a “split” in each treatment.
Weeks 0–12 did not undergo manipulation. At week 12,
the experiment was “restarted” using the week 12 observed
values as initial conditions and applying the manipulations.
This is why week 12 is listed twice in Data Set 6.2 for
each treatment. Second, you will use the estimated value of
µa = 0.1065 not only for the unmanipulated control, but also
for the initial 12 weeks of each of the other two treatments.
For the remaining weeks of the manipulated treatments, you
will set µa = 0.5 or µa = 0.96, as appropriate. Third, when
creating residuals, it is important to correctly match up
predictions with observations with these “splits” in mind.
Fourth, note that Data Set 6.2 contains some values of zero,
which will cause problems with log residuals. When comput-
ing the log of the state variables, replace values of 0 with 0.5.
c. A comparison of the R2 values in Table 6.6 suggests that the
LPA model under the assumption of environmental noise is
not well validated for the L-stage. This is not surprising,
because the collection of the estimation and validation data
sets occurred over a decade apart in different laboratories
using different culture protocols and different strains of
Tribolium castaneum. In Dennis et al. (1997) the entire data
set on which Data Set 6.2 is based was randomly divided into
estimation data and validation data. The subsequently well-
validated model documented a transition between equilibria
(control), 2-cycles (µa = 0.5), and aperiodic cycles (µa =
0.96) for the treatments in Data Set 6.2 (Exercise 9). You
can see these transitions if you look at the time series in Data
Set 6.2. In this current exercise, however, continue to use
the parameters estimated from the environmental noise LPA
 Flour Beetle Dynamics: A Case Study ■ 135

model in Tables 6.1–6.2. Systematically perturb the parame-

ters and rerun the bifurcation diagram in (7c) to explore how
it changes. Can you find parameters “nearby” in parameter
space that produce a transition from an equilibrium to
2-cycles to aperiodic cycles as µa is “tuned” from 0 to 1?

9. Read Dennis et al. (1997). We will use the parameters

estimated for the SS strain in this problem: b = 7.483, cel =
0.01200, cea = 0.009170, µl = 0.2670, cpa = 0.004139,
µa = 0.003620.

a. Use the computer to graph the LPA model bifurcation

diagram for larval numbers, using µa as a bifurcation
parameter with range 0 ≤ µa ≤ 1.
b. Repeat the validation procedure in problem (8b) using these
parameters.

10. Read Costantino et al. (1997) and Dennis et al. (2001). Reproduce
the LPA bifurcation diagram in Figure 6.3 using the parameters
shown in the figure caption.

11. Suppose you are testing the predictions of a one-dimensional

Ricker model
xt+1 = bxt e−cxt
for fish in the laboratory, with time step of one year, and you
want to manipulate the parameters as follows. Suppose you want
to manipulate the parameter c to be c = 0.01 and you want to
set up experimental cultures with b = 0.5, b = 5, and b = 10.

a. Draw the bifurcation diagram, using b as a bifurcation

parameter for 0 ≤ b ≤ 12.
b. What are the model-predicted transitions as b is tuned from
0 to 12? What are the predicted dynamics for each of b = 0.5,
b = 5, and b = 10?
c. Would it be legitimate to test the Ricker predictions in
the laboratory by manipulating b and c through adding or
removing fish at each time step (after the observed census
was recorded) in order to “make the Ricker equation exact”?
136 ■ Mathematical Modeling in Biology

Stated negatively, does that experimental protocol force the

observed dynamics to correspond to the predicted Ricker
dynamics? Explain your reasoning.

BIBLIOGRAPHY
Costantino, R. F., Cushing, J. M., Dennis, B., and Desharnais, R. A.
1995. Experimentally induced transitions in the dynamic behavior
of insect populations. Nature 375:227–230. DOI: 10.1038/375227a0.
[“Announcement” paper with first appearance of the LPA model in
the literature. Documents model-predicted bifurcations from stable fixed
points to periodic cycles to aperiodic oscillations through experimental
manipulation of the adult mortality rate in the laboratory. Both theoretical
and empirical: models connected to data. The follow-up paper by Dennis
et al. (1997) gives the details.]

Costantino, R. F., Cushing, J. M., Dennis, B., Desharnais, R. A.,

and Henson, S. M. 1998. Resonant population cycles in temporally
fluctuating habitats. Bulletin of Mathematical Biology 60:247–273. DOI:
10.1006/bulm.1997.0017. [LPA model predicts, under certain circum-
stances, a larger average total population abundance when the habitat
volume periodically fluctuates than when the habitat volume is held
constant at the average volume. Tested this prediction empirically in
laboratory populations. Both theoretical and empirical: models connected
to data.]

Costantino, R. F. and Desharnais, R. A. 1991. Population Dynamics and the

Tribolium Model: Genetics and Demography. Springer-Verlag, New York.
[Provides a comprehensive history of the mathematics of Tribolium.]

Costantino, R. F., Desharnais, R. A., Cushing, J. M., and Dennis, B.

1997. Chaotic dynamics in an insect population. Science 275:389–391.
DOI: 10.1126/science.275.5298.389. [“Announcement” paper heralding the
documentation of chaos in a laboratory insect population. Both theoretical
and empirical: models connected to data. The follow-up paper by Dennis
et al. (2001) gives the details.]

Costantino, R. F., Desharnais, R. A., Cushing, J. M., Dennis, B., Henson,

S. M., and King, A. A. 2005. The flour beetle Tribolium as an effective
tool of discovery. Advances in Ecological Research 37:101–141. DOI:
10.1016/S0065-2504(04)37004-2. [Summary of the methodology of the
Beetle Team, emphasizing that both deterministic and stochastic forces
are important in ecological systems. Discusses how a number of nonlinear
phenomena have been documented in a laboratory system, including
 Flour Beetle Dynamics: A Case Study ■ 137

chaotic dynamics, population outbreaks, saddle nodes, phase switching,

and lattice effects. Also discusses the anatomy of chaos and mechanistic
models of stochasticity. Both theoretical and empirical: models connected
to data.]

Cushing, J. M., Costantino, R. F., Dennis, B., Desharnais, R. A., and Henson,
S. M. 1998. Nonlinear population dynamics: Models, experiments, and
data. Journal of Theoretical Biology 194:1–9. DOI: 10.1006/jtbi.1998.0736.
[“Perspectives” piece on the research paradigm of the Beetle Team. Both
theoretical and empirical: models connected to data.]

Cushing, J. M., Costantino, R. F., Dennis, B., Desharnais, R. A., and

Henson, S. M. 2003. Chaos in Ecology: Experimental Nonlinear Dynamics.
Academic Press, San Diego, CA. [Summarizes the large body of work
involved in the documentation of chaos in a laboratory population.
Includes several data sets in the appendices.]

Cushing, J. M., Dennis, B., Desharnais, R. A., and Costantino, R. F. 1996. An

interdisciplinary approach to understanding nonlinear ecological dynamics.
Ecological Modelling 92:111–119. DOI: 10.1016/0304-3800(95)00170-0.
[Describes a methodology for testing nonlinear population theory using
data, models, experiments, and statistics. Both theoretical and empirical:
models connected to data.]

Cushing, J. M., Dennis, B., Desharnais, R. A., and Costantino, R. F.

1998. Moving toward an unstable equilibrium: saddle nodes in population
systems. Journal of Animal Ecology 67:298–306. DOI: 10.1046/j.1365-
2656.1998.00194.x. [Empirical demonstration of a model-predicted “saddle
fly-by” in Tribolium data, in which the system approached an unstable
equilibrium along a two-dimensional stable manifold and departed along a
one-dimensional unstable manifold. Both theoretical and empirical: models
connected to data.]

Cushing, J. M., Henson, S. M., Desharnais, R. A., Dennis, B., Costantino,

R. F., and King, A. 2001. A chaotic attractor in ecology: Theory
and experimental data. Chaos, Solitons, and Fractals 12:219–234. DOI:
10.1016/S0960-0779(00)00109-0. [The LPA chaotic attractor and the
stochastic LPA model’s stationary distribution are compared in detail
to the experimental data in phase space and are showed to correspond
to an amazing degree. Model-predicted temporal patterns on the chaotic
attractor are demonstrated in the data, as well. Both theoretical and
empirical: models connected to data.]

Dennis, B., Desharnais, R. A., Cushing, J. M. and Costantino, R. F.

1995. Nonlinear demographic dynamics: mathematical models, statistical
138 ■ Mathematical Modeling in Biology

methods, and biological experiments. Ecological Monographs 65:261–281.

DOI: 10.2307/2937060. [Introduces the LPA model and analyzes the
Desharnais data, the controls of which appear in Data Set 6.1.]

Dennis, B., Desharnais, R. A., Cushing, J. M., and Costantino, R.

F. 1997. Transitions in population dynamics: equilibria to periodic
cycles to aperiodic cycles. Journal of Animal Ecology 66:704–729. DOI:
10.2307/5923. [Detailed follow-up to the “announcement” paper in Nature
(Costantino et al. 1997). From the abstract: “The rigorous statistical
verification of the predicted shifts in dynamical behaviour provides
convincing evidence for the relevance of nonlinear mathematics in
population biology.” Both theoretical and empirical: models connected to
data.]

Dennis, B., Desharnais, R. A., Cushing, J. M., Henson, S. M., and

Costantino, R. F. 2001. Estimating chaos and complex dynamics in an
insect population. Ecological Monographs 71:277–303. DOI: 10.1890/0012-
9615(2001)071[0277:ECACDI]2.0.CO;2. [Detailed follow-up to the “an-
nouncement” paper in Science (Costantino et al. 1997).]

Desharnais, R. A. and Costantino, R. F. 1980. Genetic analysis of a population

of Tribolium. VII. Stability: Response to genetic and demographic
perturbations. Canadian Journal of Genetics and Cytology 22:577–589.
DOI: 10.1139/g80-063. [First appearance in the literature of the data in
Data Set 6.1. Both theoretical and empirical: models connected to data.]

Desharnais, R. A., Costantino, R. F., Cushing, J. M., and Dennis, B. 1997.

Estimating chaos in an insect population. Science 276:1881–1882. DOI:
10.1126/science.276.5320.1881. [Part of a discussion in Science about
positive stochastic Lyapunov exponents and whether they indicate chaos
in noisy data, motivated by Costantino et al. (1997).]

Desharnais, R. A., Costantino, R. F., Cushing, J. M., Henson, S. M.,

Dennis, B., and King, A.A. 2006. Experimental support for the scaling
rule of demographic stochasticity. Ecology Letters 9:537–547. DOI:
10.1111/j.1461-0248.2006.00903.x. [Experimental test of the ecological
tenet that the magnitude of demographic-stochastic fluctuations in
populations scales inversely with the square root of population size.
We experimentally verified the scaling rule for populations displaying
noisy chaotic dynamics. Deterministic dynamics were clarified in larger
populations. Both theoretical and empirical: models connected to data.]

Desharnais, R. A., Dennis, B., Cushing, J. M., Henson, S. M., and Costantino,
R. F. 2001. Chaos and population control of insect outbreaks. Ecology
 Flour Beetle Dynamics: A Case Study ■ 139

Letters 4:229–235. DOI: 10.1046/j.1461-0248.2001.00223.x. [LPA-model-

guided laboratory protocol for applying small perturbations to Tribolium
populations in order to control chaotic dynamics. Both theoretical and
empirical: models connected to data.]

Dong, Q. and Polis, G. A. 1992. The dynamics of cannibalistic populations:

A foraging perspective. In M. A. Elgar & B. J. Crespi (Eds.) Cannibalism:
Ecology and Evolution among Diverse Taxa (pp. 13–37). Oxford University
Press, Oxford. [Classic reference on cannibalism.]

Elgar, M. A. and Crespi, B. J. 1992. Ecology and evolution of cannibalism.

In M. A. Elgar & B. J. Crespi (Eds.) Cannibalism: Ecology and Evolution
among Diverse Taxa (pp. 1–12). Oxford University Press, Oxford. [Classic
reference on cannibalism.]

Fox, L. R. 1975. Cannibalism in natural populations. Annual Review of Ecol-

ogy and Systematics 6:87–106. DOI: 10.1146/annurev.es.06.110175.000511.
[Classic reference on cannibalism.]

Henson, S. M. 2000. Multiple attractors and resonance in periodically-forced

population models. Physica D: Nonlinear Phenomena 140:33–49. DOI:
10.1016/S0167-2789(99)00231-6. [Mathematical treatment of Henson et al.
(1999). Both theoretical and empirical: models connected to data.]

Henson, S. M., Costantino, R. F., Cushing, J. M., Dennis, B., and

Desharnais, R. A. 1999. Multiple attractors and population dynamics in
periodic habitats. Bulletin of Mathematical Biology 61:1121–1149. DOI:
10.1006/bulm.1999.0136. [Populations that oscillate can develop multiple
attractors in the advent of periodic forcing. LPA-model predictions include
multiple attractors, resonance and attenuation phenomena, and saddle
fly-bys. Both theoretical and empirical: models connected to data.]

Henson, S. M., Costantino, R. F., Cushing, J. M., Desharnais, R.

A., Dennis, B., and King, A. A. 2001. Lattice effects observed in
chaotic dynamics of experimental populations. Science 294:602–605. DOI:
10.1126/science.1063358. [Chaotic attractors live in continuous phase
space, that is, they take on a continuum of values. Populations live on a
lattice in phase space; a bounded population must take on discrete values
on a finite lattice. The stochastic LPA model, when confined to a lattice in
phase space, shows fragments of signature cycles. These model-predicted
signatures are verified in the “Hunt for Chaos” empirical data set. Both
theoretical and empirical: models connected to data.]

Henson, S. M., Costantino, R. F., Desharnais, R. A., Cushing, J. M., and

Dennis, B. 2002. Basins of attraction: population dynamics with two
140 ■ Mathematical Modeling in Biology

stable 4-cycles. Oikos 98:17–24. DOI: 10.1034/j.1600-0706.2002.980102.x.

[Focuses on LPA-model-predicted multiple attractors and basins of
attraction, and verifies these phenomena in empirical population data.
Both theoretical and empirical: models connected to data.]

Henson, S. M. and Cushing, J. M. 1997. The effect of periodic habi-

tat fluctuations on a nonlinear insect population model. Journal of
Mathematical Biology 36:201–226. DOI: 10.1007/s002850050098. [LPA
model illustrates how a periodically fluctuating environment can increase
population numbers.]

Henson, S. M., Cushing, J. M., Costantino, R. F., Dennis, B., and Desharnais,
R. A. 1998. Phase switching in population cycles. Proceedings of the
Royal Society, London B 265:2229–2234. DOI: 10.1098/rspb.1998.0564.
[Oscillation phase switches in population cycles are explained as stochastic
jumps between basins of multiple attractors. Both theoretical and
empirical: models connected to data.]

King, A. A., Costantino, R. F., Cushing, J. M., Henson, S. M. Desharnais,

R. A., and Dennis, B. 2003. Anatomy of a chaotic attractor: Subtle model
predicted patterns revealed in population data Proceedings of the National
Academy of Sciences 408–413. DIO: 10.1073/pnas.2237266100. [From the
abstract: “By using data drawn from chaotic insect populations, we show
quantitatively that chaos manifests itself as a tapestry of identifiable and
predictable patterns woven together by stochasticity.” Both theoretical
and empirical: models connected to data.]

King, A. A., Desharnais, R. A., Henson, S. M., Costantino, R. F., Cushing,

J. M., and Dennis, B. 2002. Random perturbations and lattice effects in
chaotic population dynamics: Reply to Domokos and Scheuring. Science
297:2163a. [Part of a discussion in Science about lattice effects in noisy
population data, motivated by Henson et al. (2001).]

Polis, G. A. 1981. The evolution and dynamics of intraspecific predation.

Annual Review of Ecology and Systematics 12:225–251. DOI: 10.1146/an-
nurev.es.12.110181.001301. [Classic paper on cannibalism.]

Sokoloff, A. 1972. The Biology of Tribolium Vol 1: With Special Emphasis

on Genetic Aspects. Oxford University Press, Oxford. [Comprehensive
reference for the life history of flour beetles.]

Sokoloff, A. 1975. The Biology of Tribolium Vol 2: With Special Emphasis

on Genetic Aspects. Oxford University Press, Oxford. [Comprehensive
reference for the life history of flour beetles.]
 Flour Beetle Dynamics: A Case Study ■ 141

Sokoloff, A. 1978. The Biology of Tribolium Vol 3: With Special Emphasis

on Genetic Aspects. Oxford University Press, Oxford. [Comprehensive
reference for the life history of flour beetles.]

Tong, H. 1990. Non-Linear Time Series: A Dynamical System Approach.

Oxford University Press, Oxford. [Valuable introduction to the analysis
of nonlinear time series.]
 III
Continuous-Time Models

143
 CHAPTER 7

Introduction to
Differential Equations

7.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

Suppose x(t) is the state of a dynamical system at time t. If the system
involves mainly continuous-time processes, our model equations would
likely involve the continuous-time rate of change of x, which is the
derivative dx/dt. Equations involving derivatives are called differential
equations. Thus, continuous-time processes are often modeled with
differential equations.
Undergraduate mathematics majors typically take a course in
differential equations after the calculus sequence. In that class, they
learn to solve many types of differential equations. Many undergraduate
biology majors, however, cannot fit the course into their schedules
and proceed to graduate or professional school without having learned
how to solve differential equations. Although it would be best if all
biology majors took a differential equations course, it is important
for biologists, and particularly ecologists, to realize that they can
understand how to model with differential equations and they can
do much of the most important analysis without ever solving the
equations. We will focus on such techniques in Chapter 8.
The current chapter is an introduction to compartmental differ-
ential equations and is composed of three examples that illustrate
important ideas: (i) an example of how to set up a continuous-time
compartmental model using dimensional analysis; (ii) an example of

DOI: 10.1201/9781003265382-10 145

146 ■ Mathematical Modeling in Biology

how to study solutions of compartmental models without solving the

equations; and (iii) an example of using a powerful solution method to
solve a famous differential equation in ecology. We will briefly introduce
two techniques (partial fractions and separation of variables) that you
would see in second-semester calculus and differential equations classes,
but we will explain as we go for those who have not had those courses.

7.2 COMPARTMENTAL MODELS

For compartmental models, the main modeling technique is to look at
the net flux, which we can write as
dx
= Inflow rate − Outflow rate.
dt
Let’s illustrate this idea with three examples.

7.2.1 A Tank Problem

The first example is a generic “tank” model that illustrates inflow
and outflow rates as well as the usefulness of thinking in terms of
dimensional analysis. Suppose two 100 L tanks are initially full of pure
water. Salt solution of concentration 20 g/L flows into tank A at a rate
of 3 L/min. The well-mixed solution flows into tank B at a rate of 3
L/min, and tank B is allowed to overflow. Our goal is to model the
amount of salt in each tank (Figure 7.1).
Because we want to model the amount (in grams) of salt in each
tank, we let x (t) and y (t) be the number of grams of salt in tank A
and B, respectively, at time t. Time t is in minutes, so the derivatives
dx/dt and dy/dt will have units of g/min. Thus, each term in the
compartmental model—that is, each inflow rate and outflow rate—
must also have units of g/min. In order to obtain these fluxes in
units of g/min, we must multiply each flow rate of liquid (L/min)

Figure 7.1: Compartments and flows for a system of tanks.

 Introduction to Differential Equations ■ 147

by its salt concentration (g/L). An easy way to see this is that

(L/min)(g/L)=g/min because the liter units in each fraction cancel.
The initial amounts of salt in each tank are zero, so the initial
conditions are x(0) = 0 and y(0) = 0. Furthermore, the tanks always
remain full of 100 L of solution. However, the concentrations in both
tanks change over time. We know the concentration of the inflow into
tank A is a constant 20 g/L. The concentration of the outflow from
tank A (= the inflow to tank B) will be x grams of salt in tank A per
100 L of water in tank A, that is, x/100 with units of g/L. A similar
consideration yields the outflow from tank B as y/100 with units g/L
(Figure 7.1).
The compartmental differential equations are
   
dx L g L x g
= 3 20 − 3
dt min L min 100 L
   
dy L x g  L y g
= 3 − 3 .
dt min 100 L min 100 L
The initial value problem is therefore
dx 3x
= 60 −
dt 100
dy 3x 3y
= −
dt 100 100
x (0) = 0
y (0) = 0.

7.2.2 The SIR Model

The second example is a famous one from epidemiology, published
by Kermack and McKendrick (Kermack and McKendrick 1927). The
model is called an SIR model (for susceptibles/infectives/recovereds),
and it models a disease with rapid spread and permanent acquired
immunity. Many models in epidemiology are based on the SIR model.
Let S be the class of susceptibles (those who can catch, but not
spread the disease). Let I be the class of infectives (those who can
infect susceptibles). Let R be the class of recovereds (those who are
recovered with permanent immunity). Let’s assume that
(A) S + I + R is constant (so that the time derivatives S ′ + I ′ +
R′ = 0).
148 ■ Mathematical Modeling in Biology

(B) The per capita rate at which S are infected is proportional to

I.
(C) No new S enter the population.
(D) The per capita rate at which I recover is constant.
Given assumption (B), the per capita flow rate from compartment
S to compartment I is βI for some β > 0, and so the total flow rate
from S to I is (βI) (S). The per capita flow rate from I to R is constant
γ > 0, and so the total flow rate from I to R is γI (Figure 7.2). The
SIR model is

dS
= −βSI
dt
dI
= βSI − γI
dt
dR
= γI (7.1)
dt
S (0) = S0 > 0
I (0) = I0 > 0
R (0) = R0 > 0,

where β, γ > 0 are parameters.

Several interesting implications arise from these model equations.
For example, the number of susceptibles S(t) is decreasing as long as
there are both susceptibles and infectives in the population, since its
derivative is negative. It can be shown, in fact, that S(t) tends to zero
in the long run. This means that the derivative of I(t) must eventually
go negative, as well, and so I(t) must ultimately decrease. In fact, it can
be shown that I(t) also tends to zero in the long run. This means the
disease ultimately dies out. However, I (t) can initially increase, before
it begins decreasing. If this happens, we say there is an epidemic, or

Figure 7.2: Compartments and per capita flow rates for the SIR model.
 Introduction to Differential Equations ■ 149

Figure 7.3: If dI/dt is initially positive, we say there is an outbreak.

outbreak . That is, if at time t = 0, we have dI/dt > 0, then there

is an outbreak, whereas if dI/dt ≤ 0 at t = 0, there is no outbreak
(Figure 7.3).
Note that I(t) is non-increasing as long as

dI
= βSI − γI ≤ 0,
dt
that is, as long as
γ
S≤ .
β
Thus, there will be no outbreak if S0 ≤ γ/β, whereas there will be an
outbreak if S0 > γ/β.
What are the equilibria of the SIR model? The equilibria occur
where the derivatives are zero:
dS
= −βSI = 0
dt
dI
= βSI − γI = I (βS − γ) = 0
dt
dR
= γI = 0.
dt
150 ■ Mathematical Modeling in Biology

From the third equation, we see that the equilibrium value of I must be
zero, while S and R can be any non-negative value. So the equilibria
all have the form (Se , 0, Re ) , which includes the trivial equilibrium
(0, 0, 0). This means there is no endemic disease state (i.e., no level of
disease that can be maintained at a steady state in the host).
One of the interesting aspects of the SIR example is the amount of
information about the system we can obtain without actually solving
the differential equations. It is important for biologists to realize
that, in general, it is not necessary to actually solve a differential
equation in order to study its equilibria and stability. In Chapter
8, we will learn more techniques for analyzing differential equations
without solving them. But to finish the current chapter, we illustrate a
powerful solution method for solving a famous differential equation
from ecology. Hopefully, students who have not taken a course in
differential equations will become intrigued and be motivated to do
so.

7.2.3 The Continuous-Time Logistic Model

The third and final example in this chapter concerns a famous model
from population ecology called the logistic model (Verhulst 1838;
Verhulst 1845). This model can be posed as a compartmental model

dx
= bx − dx2 ,
dt

where b, d > 0 are parameters, x(t) is the population size or density

at time t, bx is the total population birth rate, and dx2 is the total
population death rate. We will examine the assumptions underlying
this model in Chapter 8. For now, we simply note that the logistic
model is traditionally written as

dx  x
= rx 1 − , (7.2)
dt K

where r = b and K = b/d (Exercise 4) (Pearl and Reed 1920). Our

goal here is to illustrate how to solve this type of differential equation.
The first step in solving equation (7.2) is to note that there are two
equilibrium solutions, namely xe = 0 and xe = K.
 Introduction to Differential Equations ■ 151

The next step is to note that equation (7.2) is multiplicatively

separable, which means it can be written in the form

dx
= f (x) g (t) . (7.3)
dt
(In the logistic model, we can take f (x) = x(1 − x/K) and g(t) = r.)
It turns out that one can legitimately treat dx/dt as a fraction of
differentials to multiplicatively “separate the variables” as

dx
= g (t) dt.
f (x)

One can then integrate each side with respect to its variable. It is
important to note, however, that dividing by f (x) assumes f (x) = ̸ 0,
which “drops” all equilibrium solutions of equation (7.3). In our current
example, we can write

dx
= rdt
x (1 − x/K)

as long as x ≠ 0 and x = ̸ K. So we see that we have “lost” the

equilibrium solutions xe = 0 and xe = K.
Then Z Z
dx
= rdt + C1 ,
x (1 − x/K)
for C1 ∈ R, and so
Z
dx
= rt + C1 .
x (1 − x/K)

The integrand on the left-hand side is a rational function with

linear factors in the denominator. To integrate such a function, we
use a technique from second-semester calculus called “partial fraction
decomposition.” The idea is to decompose the integrand as a sum with
unknown coefficients over the linear factors and then determine the
coefficients. We therefore decompose the integrand into two fractions
and then add them back together:

1 A B A (1 − x/K) + Bx
= + = .
x (1 − x/K) x (1 − x/K) x (1 − x/K)
152 ■ Mathematical Modeling in Biology

Then, ignoring for a moment the intermediate step, we set the

numerators equal to each other:

1 = A (1 − x/K) + Bx
= A + (B − A/K) x.

Setting “like coefficients” on x to be equal, we obtain

1 = A and B − A/K = 0,

which leads to A = 1 and B = 1/K. Thus, using the partial fraction

decomposition, we can write
Z  
1 1/K
+ dx = rt + C1 .
x (1 − x/K)

Integration yields

ln |x| − ln |1 − x/K| = rt + C1 .

By a property of logarithms, we have

x
ln = rt + C1 .
1 − x/K

To solve for x, we exponentiate both sides and then drop the absolute
values by introducing ± on the right:

x
= ert+C1 = eC1 ert
1 − x/K
x
= ±eC1 ert = Cert .
1 − x/K

Note that C = ±eC1 can be any number except 0. Multiplication of

both sides by the denominator yields

x = Cert (1 − x/K) ,

and solving for x yields

Cert
x= ̸ 0.
for C = (7.4)
1 + Cert /K
 Introduction to Differential Equations ■ 153

Equation (7.4) is (well, almost) the general solution, meaning the

set of all possible solutions as indexed by the constant C. But it is
NOT the set of all possible solutions, because our algebraic procedure
dropped the two solutions xe = 0 and xe = K. We must include these
in the general solution, so we write
Cert

 1+Ce rt /K for C ̸= 0
x(t) = 0 .

K
Note that if C were allowed to be 0, we would pick up the equilibrium
solution xe = 0 as part of our solution formula. So why bother with
keeping track of C = ̸ 0 and the lost equilibrium solutions throughout
the algebraic procedure if, in fact at the end, we allow C = 0? The
answer is that setting C = 0 does not necessarily absorb all the
equilibrium solutions into the formula. In this example, it is impossible
to obtain the solution xe = K from the solution formula no matter what
you choose for C (Exercise 5). Therefore, the true general solution,
denoted xg , can be written as
(
Cert
1+Cert /K for C ∈ R
xg (t) = . (7.5)
K

A solution such as xe = K that cannot be “fit” into the solution formula

is called a singular solution.
Let the initial condition be x (0) = x0 , where x0 ̸= K. Applying
this initial condition to equation (7.5) obtains a relationship between
x0 and C:
KC
x0 = . (7.6)
K +C
Solving for C yields (Exercise 6)
Kx0
C= . (7.7)
K − x0
Equations (7.5) and (7.7) together yield the particular solution,
meaning the solution that satisfies the initial condition. The particular
solution is (Exercise 7)
Kx0
x(t) = . (7.8)
x0 + (K − x0 ) e−rt
154 ■ Mathematical Modeling in Biology

In Chapter 8, we will study the behavior of these solutions of the

logistic equation as they depend on the initial condition x0 in the
context of ecology.

7.3 EXERCISES
1. Consider two 100 L tanks. Tank A is initially full of pure water,
and tank B initially contains 50 L of pure water. Seawater of
concentration 35 g/L flows into tank A at a rate of 2 L/min.
The well-mixed solution overflows into tank B, and when tank
B is full, it is allowed to overflow as well. Write the differential
equation initial value problem for the amount of salt in each tank.
Hint: It is okay to use piecewise functions.

2. Suppose a disease spreading through a school can be modeled

with the SIR model. How many of the susceptible children should
be vaccinated in order to avoid an epidemic?

3. Suppose a disease is spreading through a school, with S + I +

R constant (so that S ′ + I ′ + R′ = 0). Suppose that the per
capita rate at which susceptibles are infected is proportional to
the number of infectives, the per capita rate at which infectives
recover is constant, and recovereds become susceptible again at
a constant per capita rate. Draw a compartment diagram and
construct the differential equation model.

4. Show that the differential equation

dx
= bx − dx2
dt
can be written as
dx  x
= rx 1 − ,
dt K
where r = b and K = b/d.

5. Prove that there is no value of C for which the formula

Cert
x(t) =
1 + Cert /K
satisfies the initial condition x(0) = K > 0.
 Introduction to Differential Equations ■ 155

6. Solve equation 7.6 for C to obtain equation 7.7.

7. Show that equations (7.5) and (7.7) together yield the particular
solution (7.8). Be careful with the singular solution.

8. Consider the differential equation

dx
= t2 x2 .
dt
Note that this equation is separable and that there is one
equilibrium: xe = 0.

a. Find the general solution using the method of separation of

variables.
b. Find the particular solution given the initial condition
x(0) = 1/2.
c. Find the particular solution given the initial condition
x(0) = 0.

9. Consider the differential equation

dx
= t 1 − x2 .


dt

a. Find all equilibria xe .

b. Find the general solution using the method of separation of
variables. Hint: Write (1 − x2 ) as (1 − x) (1 + x).
c. Find the particular solution given the initial condition
x (0) = 2.
d. Find the particular solution given the initial condition
x (0) = −1.
e. Find the particular solution given the initial condition
x (0) = 1.

10. The discrete-time logistic map

xt+1 = rxt (1 − xt ) ,
156 ■ Mathematical Modeling in Biology

considered in Exercise 9 of Chapter 3, is so-called because its

quadratic nonlinearity reminds us of the well-known continuous-
time logistic model discussed in this chapter

dx/dt = rx (1 − x/K) .

However, the discrete-time logistic map is not a good discrete

analogue of the continuous-time logistic model. In fact, it turns
out that the Beverton-Holt map

bxt
xt+1 = ,
1 + cxt
considered in Exercise 8 of Chapter 3, is the discrete-time
analogue of the continuous-time logistic model. In this exercise
we will see why.

a. In the Beverton-Holt map, let K = (b − 1)/c. Prove that the

closed-form particular solution of the Beverton-Holt map
with initial condition x(0) = x0 is

Kx0
xt = . (7.9)
x0 + (K − x0 ) b−t

Hint: Prove that the proposed solution satisfies the

Beverton-Holt iterative map.
b. How do solutions (7.9) of the discrete-time Beverton-Holt
map compare to solutions (7.8) of the continuous-time
logistic differential equation?

BIBLIOGRAPHY
Kermack, W. O. and McKendrick, A. G. 1927. A contribution to the
mathematical theory of epidemics. Proceedings of the Royal Society of
London A. 115:700–721. DOI: 10.1098/rspa.1927.0118. [Historic paper
introducing the SIR model with random infectious encounters, similar to
the law of mass action in chemistry. Many models in epidemiology are
variations of the SIR model.]

Pearl, R. and Reed, L. J. 1920. On the rate of growth of the population of the
United States since 1790 and its mathematical representation. Proceedings
 Introduction to Differential Equations ■ 157

of the National Academy of Sciences of the United States of America

6:275–288. DOI:10.1073/pnas.6.6.275. [Pearl and Reed rediscovered the
Verhulst’s logistic growth model, sometimes called the Verhulst-Pearl
model.]

Verhulst, P.-F. 1838. Notice sur la loi que la population suit dans son
accroissement. Correspondance mathematique et physique 10:113–121.
[(Note on the law of population increase.) Verhulst introduces the logistic
differential equation model of population growth, which is a modification
of the Malthusian exponential growth model by the inclusion of crowding
effects.]

Verhulst, P.-F. 1845. Recherches mathematiques sur la loi d’accroissement de

la population. Nouveaux Memoires de l’Academie Royale des Sciences et
Belles-Lettres de Bruxelles 18:1–42. [(Mathematical researches into the law
of population growth increase.) Verhulst names the solution the “logistic
curve.”]
 CHAPTER 8

Scalar Differential
Equations

8.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

A scalar differential equation is a differential equation with one state
variable. In this chapter, we consider only autonomous differential
equations (ones which do not explicitly depend on time). Instead of
focusing on solution techniques, we focus on equilibria, stability, and
bifurcation techniques. It is important for biologists to realize that,
in general, it is not necessary to actually solve a nonlinear differential
equation in order to study its equilibria and stability. The main tools
come from first-semester calculus.
For the bifurcation approach to scalar differential equations given
here and the section on human population growth, the authors are
indebted to J. M. Cushing’s textbook Differential Equations: An
Applied Approach (Cushing 2004), from which co-author Henson spent
many pleasant years teaching differential equations.

8.2 LINEAR EQUATIONS

We first consider the linear equation
dx
= rx (8.1)
dt
x (0) = x0 ,

DOI: 10.1201/9781003265382-11 159

160 ■ Mathematical Modeling in Biology

where r is a real parameter. What are the equilibria of this model? The
system is at equilibrium if and only if the derivative dx/dt is zero for
all time t. Thus, the equilibrium equation (or fixed point equation) is

0 = rx.

To find the equilibria, we must solve the equilibrium equation for

constant x. The extinction state xe = 0 is the only equilibrium if r = ̸ 0.
If r = 0, then every constant number x is an equilibrium.
The general solution (i.e., the set of all solutions) of the differential
equation (8.1) is
xg (t) = cert ,

where c is any constant (Exercise 1). Now apply the initial condition
x(0) = x0 to obtain
x0 = cer0 ,

which implies c = x0 . Thus, the particular solution (i.e., the solution

that satisfies the initial condition) is

xp (t) = x0 ert .

From this solution formula, we can see that if r > 0, then the extinction
equilibrium xe = 0 is unstable, because ert is growing over time. If
r < 0, then the extinction state is asymptotically stable, because ert
is approaching zero over time. If r = 0, then every real number x0
is an equilibrium, and these equilibria are all neutrally stable. The
bifurcation diagram in Figure 8.1 summarizes these dynamics as a
function of the bifurcation parameter r. The bifurcation at r = 0 is
called a vertical bifurcation (Figure 8.1).
For a specific example, consider the initial value problem

x′ = 2x
x (0) = 5.

The general solution is x = ce2t . Plugging in the initial condition to

obtain c = 5, we find the particular solution

x (t) = 5e2t .
 Scalar Differential Equations ■ 161

Figure 8.1: Bifurcation diagram for linear scalar differential equation.

A vertical bifurcation occurs at r = 0.

8.2.1 Malthusian Growth

Now consider a population with constant per capita birth rate b >
0 and constant per capita death rate d > 0 and no emigration or
immigration. The total influx (inflow rate) is bx, and the total outflux
(outflow rate) is dx. Thus, we have

dx
= bx − dx = (b − d) x (8.2)
dt
x (0) = x0 .

The general solution is x = ce(b−d)t . Applying the initial condition

yields the particular solution

x (t) = x0 e(b−d)t .

Figure 8.2 gives time series graphs for the three cases b < d, b = d, and
b > d. If b < d, then all nonzero solutions are decaying exponentially
toward zero. If b = d, then the initial condition itself is an equilibrium
solution. If b > d, then all nonzero solutions are growing exponentially.
The exponential growth when b > d is called Malthusian growth,
after Thomas Malthus, an economist and English curate who wrote
the influential piece An Essay on the Principle of Population in 1798
(Malthus 1798).
Note that linear models have exponential solutions.
162 ■ Mathematical Modeling in Biology

(a)

(b)

(c)

Figure 8.2: Time series for model (8.2). (a) b < d, (b) b = d, (c) b > d.

8.3 NONLINEAR EQUATIONS

8.3.1 Logistic Growth
Consider again equation (8.2). When b and d are assumed to be
constant and b > d, the model exhibits Malthusian growth, which
of course cannot be sustained indefinitely because of crowding effects.
Sooner or later, the vital rates b and d will themselves become functions
of population density. In this case, the population model becomes
nonlinear and the population is said to have density dependence. For
example, suppose that a population grows with constant per capita
birth rate b > 0, but that the per capita death rate is proportional to
the population size. Assume there is no emigration or immigration. The
total population inflow rate is therefore bx, and the total population
 Scalar Differential Equations ■ 163

outflow rate has the form (dx) x = dx2 . Thus, we have the nonlinear
model
dx
= bx − dx2 . (8.3)
dt
This is the famous logistic model, but it is traditionally written in a
different form. Choose two new parameters r and K defined by r = b
and K = b/d. Then we can rewrite equation (8.3) as
dx  x
= rx 1 − . (8.4)
dt K
Equation (8.4) is the traditional form of the continuous-time logistic
model . What are the dynamics? First let’s compute the equilibria, that
is, the constant solutions. The solution x (t) is constant if and only if
dx/dt = 0 for all time t, so the equilibrium equation is
 x
rx 1 − = 0.
K
The equilibria are therefore xe = 0 and xe = K. Figure 8.3 illustrates
the behavior of all possible solutions as a function of t. Note from
Figure 8.3 that xe = 0 is an unstable equilibrium, whereas xe = K is
an asymptotically stable equilibrium.
The parameter r is called the intrinsic growth rate. It is the
exponential rate of growth for small population sizes, when crowding
effects are negligible. The parameter K is called the carrying capacity
of the environment.

Figure 8.3: Illustration of all possible solutions of the logistic model

(8.4).
164 ■ Mathematical Modeling in Biology

8.3.2 Allee Effects

The per capita growth rate of a population of density x (t) is defined
to be
1 dx
.
x dt
We tend to think of the cases in which density dependence causes
the per capita growth rate to decrease if population density increases.
Sometimes, however, it works the other way around so that an increase
in population density x results in an increase in the per capita growth
rate. This is called an Allee effect. Allee effects can happen when,
for example, a population requires a threshold density in order to
reproduce.
Consider again the logistic model
dx  x
= rx 1 − ,
dt K
where r, K > 0. The per capita growth rate is
 x
r 1− .
K
Here, we see that an increase in x causes a decrease in the per capita
growth rate (Figure 8.4). Thus, the logistic growth does not exhibit
Allee effects.
We can modify the logistic equation to include an Allee effect:
dx  x
= rx (x − a) 1 − , (8.5)
dt K

Figure 8.4: Per capita growth rate in the logistic model graphed
against population size x.
 Scalar Differential Equations ■ 165

Figure 8.5: Per capita growth rate in model (8.5) graphed against
population size x.

where r, a, K > 0 and a < K. The equilibria for this model are xe =
0, a, K, and the per capita growth rate is
 x
r (x − a) 1 − .
K

In Figure 8.5, we see that at low population densities (small values

of x), an increase in x causes an increase in the per capita growth
rate. Thus, the dynamics have an Allee effect at low population sizes.
Figure 8.6 illustrates the behavior of all possible solutions of model
(8.5) in time t. Note from Figure 8.6 that xe = 0 is asymptotically
stable, xe = a is unstable, and xe = K is asymptotically stable. The
equilibrium xe = a is called an Allee threshold. The initial population
size must be above this threshold in order for the population to survive
(Figure 8.6).

Figure 8.6: Illustration of all possible solutions of model (8.5).

166 ■ Mathematical Modeling in Biology

8.3.3 The “Doomsday Model” of Human Population Growth

Consider again the Malthusian growth model
dx
= rx
dt
x(0) = x0 > 0

with r > 0. The particular solution is

x (t) = x0 ert .

After a certain amount of time T, the population size will have doubled
to 2x0 . What is the doubling time T ? That is, for what value of T is

2x0 = x0 erT ?

Note that the initial population size x0 cancels if it is nonzero, and we

can solve for T to obtain a formula for the doubling time that does not
depend on the initial population size:
ln 2
T = .
r
Note that exponential growth is characterized by a constant doubling
time. Similarly, exponential decay is characterized by a constant half-
life.
An interesting problem in the differential equations textbook by J.
M. Cushing (Cushing 2004) took another look at human population
growth using census data for years 1900–1990 from the United Nations
(see bibliography entry for URL). In Table 8.1, we see that the doubling
time was not constant, but was decreasing. A decreasing doubling
time means that the human population was growing faster than any
exponential function.
Consider the population model
dx
= rxα (8.6)
dt
x (0) = x0

with constants r, α > 0. For an exponentially growing population,

α = 1 and doubling times are constant. If α > 1, the population is
 Scalar Differential Equations ■ 167

TABLE 8.1 World Population Size 1900–1990

Year World Population Approximate Doubling Time

(billions) (years)
1900 1.65 60.4
1910 1.75 50.8
1920 1.86 50.0
1930 2.07 40.6
1940 2.30 40.2
1950 2.52 30.7
1960 3.02
1970 3.70
1980 4.45
1990 5.30

growing faster than exponential and doubling times are decreasing. If

0 < α < 1, the population is growing, but it is growing more slowly
than exponential and doubling times are increasing. The decreasing
doubling times in Table 8.1 suggest that α > 1 for the years shown.
To solve equation (8.6), we separate the variables as illustrated in
Chapter 7:
dx
α
= rdt
Z x Z
dx
= rdt + C
xα
x1−α
= rt + C (8.7)
1−α
as long as α ̸= 1. Equation (8.7) is the general solution in implicit
form. To find the particular solution, we apply the initial condition
x(0) = 1.65, in which year 1900 is taken to be t = 0. This yields
1.651−α
C= . (8.8)
1−α
From equations (8.7) and (8.8), we see that the particular solution
satisfies
x1−α = rt (1 − α) + 1.651−α ,
and can be written
1/(1−α)
xp (t) = rt (1 − α) + 1.651−α

. (8.9)
168 ■ Mathematical Modeling in Biology

We need to estimate two parameters, r and α. In his textbook

(Cushing 2004), Cushing accomplished this by requiring that the
particular solution pass through the data points for years 1940 and
1990 (t = 40 and t = 90):

x(40) = 2.30
x(90) = 5.30.

Thus, we need to solve the system of two equations

1/(1−α)
2.30 = 40r (1 − α) + 1.651−α

(8.10)
1/(1−α)
5.30 = 90r (1 − α) + 1.651−α


for the two unknowns r and α.

The parameter estimates, here obtained from a computer, are
(Exercise 2)

r = 3.6022 × 10−3 (8.11)

α = 2.2633.

Substituting the values of α and r given in equation (8.11) into

equation (8.9), we find
1
x (t) = 0.7916 .
[(−4.551 × 10−3 ) t + 0.5312]
The strange and rather shocking fact from the point of view of
modeling is that this faster-than-exponentially-growing solution blows
up in finite time. That is, the solution has a vertical asymptote
when
−4.551 × 10−3 t + 0.5312 = 0,

that is, when t = 116.7. This time corresponds to the year 1900 +
116.7 = 2016.7, or about the year 2017. A similar model published in
the journal Science in 1960, called the “Doomsday Model,” predicted
a doomsday of November 13, 2026 (von Foerster, Mora, and Amiot
1960).
The Doomsday Model and these various dates for “doomsday”
are humorous, disturbing, and, for the modeling student, enlightening.
 Scalar Differential Equations ■ 169

TABLE 8.2 World Population Size

2000–2020

Year World Population (billions)

2000 6.14
2010 6.96
2020 7.79

They are humorous because it is obvious that, on a bounded planet,

a population size cannot blow up in finite time. And we laugh
because, at this writing, the year 2017 has already passed. Even
so, this exercise is disturbing because one realizes that the dynamic
mechanisms behind supra-exponential growth could deplete the earth’s
resources extremely quickly, leading to untold suffering. For the
modeling student, the Doomsday Model makes a very important point
about the interpretation of predictions by mathematical models. When
modeling dynamical systems, it is important to remember that model
predictions are based on model assumptions. Predictions change if
underlying assumptions are altered.
Let’s now consider the United Nations data for years 2000–2020
(Table 8.2) (see bibliography for URL).
The data for 2000–2020 show almost linear growth, which is sub-
exponential. Indeed, for model (8.6), applied to the data in Table 8.2,
α would be close to zero. Clearly, the dynamic mechanisms changed,
and a model for the data during 1900–2020 should be more like the
logistic model. At this point, our concern turns to the predicted value
of the carrying capacity K (Exercise 3). When world population levels
off, will there be enough resources for every human being to have a
decent standard of living while preserving the rest of the biodiversity
and beauty of the planet?
Scientists have a duty to make it clear to the public that model
predictions are always conditional projections. Predictions change if
underlying model assumptions are altered. As the ghost said about
Tiny Tim in Charles Dickens’ novel A Christmas Carol, “If these
shadows remain unaltered by the Future, none other of my race...will
find him here.”
170 ■ Mathematical Modeling in Biology

8.4 LINEARIZATION
Consider the differential equation

dx
= x (x − 1) (x − 2) .
dt

Let f (x) = x (x − 1) (x − 2). The equilibria are

xe = 0, 1, 2.

Figure 8.7a shows the graph of f (x) vs. x. The roots of f are
the equilibria. When the graph of f is above the x-axis, dx/dt > 0
and so x (t) is increasing. We denote this with a time arrow to the
right on the x-axis. When the graph of f is below the x-axis, we insert
an arrow to the left on the x-axis. If we extract the x-axis by itself
with the equilibria shown as circles and with the time arrows in each
interval between equilibria, we obtain the phase line portrait (Figure
8.7b). If the arrows point toward the equilibrium on each side, then
the equilibrium is a sink and is asymptotically stable. If the arrows on
each side point away from the equilibrium, then the equilibrium is a
source and is unstable. All possible time series for x vs. t are shown in
Figure 8.7c.
Note from the geometry in Figure 8.7a that if the graph of f is
increasing through the equilibrium as a function of x, then the arrows
point away from the equilibrium, and the equilibrium is unstable.
Similarly, if f is decreasing through the equilibrium as a function
of x, then the arrows point toward the equilibrium, which is then
asymptotically stable. In the current example, we have

df
(0) > 0 =⇒ xe = 0 unstable
dx
df
(1) < 0 =⇒ xe = 1 asymptotically stable
dx
df
(2) > 0 =⇒ xe = 2 unstable.
dx

Note: Be careful! df /dx is NOT the same as dx/dt.

These considerations lead to the following definition and theorem.
 Scalar Differential Equations ■ 171

(a)

(b)

(c)

Figure 8.7: (a) Graph of f (x) vs. x. The roots of f are the equilibria.
(b) Phase line portrait. (c) Time series for x(t) vs. t showing
equilibrium solutions (horizontal lines) and illustrating the behavior
of all other solutions.

Definition 8.1 Consider the differential equation

dx
= f (x) .
dt

df
An equilibrium xe is hyperbolic if and only if dx (xe ) ̸= 0.
172 ■ Mathematical Modeling in Biology

Theorem 8.1 (Linearization Theorem) Let xe be a hyperbolic equilib-

rium of
dx
= f (x)
dt
and suppose that the function f is continuously differentiable in x (that
is, df /dx exists and is continuous). Define

df
λ= (xe ) .
dx
Then

λ > 0 =⇒ xe is unstable
λ < 0 =⇒ xe is asymptotically stable.

As you will see in the next chapter, the number λ is a very important
number called the eigenvalue. Figure 8.8 summarizes the relationships
between sinks, sources, and eigenvalues. Note that for λ = 0 (the
nonhyperbolic case) you cannot conclude anything from linearization;
the linearization theorem does not apply. See Figure 8.9.
For an example, consider the logistic model
dx  x
= rx 1 − ,
dt K
where r, K > 0. The equilibria are xe = 0, K. Also, f (x) =
rx (1 − x/K). It is easiest to use the product rule to find df /dx =

Figure 8.8: Phase line portraits for hyperbolic equilibria. The

linearization theorem applies, and the sign of the eigenvalue λ
determines stability.
 Scalar Differential Equations ■ 173

Figure 8.9: Phase line portrait possibilities for nonhyperbolic equilib-

ria. The linearization theorem does not apply.

r (1 − x/K) − rx/K, because as you plug in each equilibrium, all terms

will zero out except one. At xe = 0, df /dx = r > 0, so the equilibrium
xe = 0 is unstable. At xe = K, df /dx = −r < 0, so the equilibrium
xe = K is asymptotically stable (Figure 8.10).
For another example, let’s consider

dx
= x2 (x − a) ,
dt

where a ≠ 0. The equilibria are xe = 0, a. Also, f (x) = x2 (x − a) , so,

using the product rule, we have df /dx = 2x (x − a) + x2 . At xe = 0,
df /dx = 0, and at xe = a, df /dx = a2 . Thus, from the linearization
theorem, the equilibrium xe = a is unstable and no conclusion can be
drawn about xe = 0. The geometric method (Figure 8.11) shows that
xe = 0 is a shunt (and thus unstable).
174 ■ Mathematical Modeling in Biology

Figure 8.10: Phase line portrait for the logistic model.

Figure 8.11: Phase line portrait for dx/dt = x2 (x − a), shown for
a > 0. For a < 0, xe = 0 becomes a shunt to the right.

For a final example, let’s consider

dx
= x (x − a) (10 − x) ,
dt

where a ∈ R. The equilibria are xe = 0, a, 10. Using the extended

product rule, we find that df /dx = (x − a) (10 − x) + x (10 − x) −
x (x − a) . At xe = 0, df /dx = −10a. At xe = a, df /dx = a (10 − a) .
At xe = 10, df /dx = −10 (10 − a) . The stabilities of the equilibria
depend on the value of a. For example, for xe = 0 we have df /dx >
0 for a < 0 and df /dx < 0 for a > 0. Thus, the equilibrium
xe = 0 is unstable for a < 0 and asymptotically stable for a > 0.
Similarly, the equilibrium xe = a is unstable for 0 < a < 10 and
asymptotically stable for a < 0 and a > 10. The equilibrium xe = 10 is
asymptotically stable for a < 10 and unstable for a > 10. The phase line
portraits and bifurcation diagram in Figures 8.12–8.13 summarize this
information.
 Scalar Differential Equations ■ 175

Figure 8.12: Possible phase line portraits for dx/dt =

x (x − a) (10 − x).

Figure 8.13: Bifurcation diagram for dx/dt = x (x − a) (10 − x), using

a as the bifurcation parameter.

8.5 BIFURCATIONS

In this section, we illustrate some canonical bifurcations for scalar

differential equations.
176 ■ Mathematical Modeling in Biology

8.5.1 Transcritical Bifurcation

Consider the equation
x′ = x (x − a),

where the “prime” denotes the time derivative dx/dt and a ∈ R is

a bifurcation parameter. The equilibria are xe = 0 and xe = a. In
the bifurcation diagram, these two equilibrium branches cross at a =
0 (Figure 8.14) at a transcritical bifurcation. In Exercise 6, you will
prove that for a < 0, the equilibrium xe = 0 is unstable and xe = a
is asymptotically stable, and for a > 0, the equilibrium xe = 0 is
asymptotically stable and xe = a is unstable. This means that a typical
exchange of stability occurs at the bifurcation point a = 0.
The correct interpretation of Figure 8.14 is important. For each
fixed value of the parameter a, one obtains a different differential
equation. The phase line portrait for that differential equation is the
vertical line through that value of a. We might say the dynamics for
any fixed a “live” on a vertical line. Thus, the bifurcation diagram
is the collection of all possible phase line portraits, slotted together
as vertical lines, as a function of parameter a. In Exercise 6, you will
determine the stabilities in two ways, first by using the linearization
theorem, and then by using the geometric method. The geometric

Figure 8.14: Transcritical bifurcation at a = 0 with typical exchange

of stability.
 Scalar Differential Equations ■ 177

method consists of drawing all possible phase line portraits (as a

function of a) with their stability arrows and then fitting them together
vertically to determine the stabilities of the equilibrium branches in the
bifurcation diagram.

8.5.2 Saddle-Node Bifurcation

Consider
x′ = x2 − a.
√
The equilibria are xe = ± a. In the bifurcation diagram, this is a
single branch of equilibria (Figure 8.15) with a saddle-node bifurcation
at a = 0. For a < 0, there is no equilibrium. In Exercise 7, you will
√
show that for a > 0, the upper part of the branch xe = a is unstable
√
and the lower part xe = − a is asymptotically stable.

8.5.3 Pitchfork Bifurcation

Now consider the equation

x′ = x x 2 − a .

√
The equilibria are xe = 0 and xe = ± a. In the bifurcation diagram,
these two branches cross at a = 0 in a pitchfork bifurcation (Figure
8.16). In Exercise 8, you will show that for a < 0, the equilibrium xe = 0
is unstable, and for a > 0, the equilibrium xe = 0 is asymptotically

Figure 8.15: Saddle-node bifurcation at a = 0.

178 ■ Mathematical Modeling in Biology

Figure 8.16: Pitchfork bifurcation at a = 0.

√
stable and xe = ± a are unstable. An exchange of stability occurs at
the bifurcation point a = 0.

8.5.4 Hysteresis
Finally, consider
x′ = x3 − x + a.
It is not so easy here to find the equilibria, but we note that the
bifurcation parameter a in f (x) = x3 − x + a vertically translates
the graph of f . First, graph f (x) for a = 0. The roots are −1, 0, and
1 (Figure 8.17). From the geometry we can include stability arrows

Figure 8.17: f (x) vs. x for various values of the parameter a. The
graph is shifted vertically upward when a > 0, which corresponds to
dropping the horizontal axis, etc.
 Scalar Differential Equations ■ 179

Figure 8.18: Hysteresis.

on the x-axis. If we “tune” a to small positive numbers, we raise the

graph of f slightly. The lower equilibrium moves to the left, and the
two upper equilibrium move toward each other. As we continue to
increase the parameter a, the two upper equilibria collide and wink out
of existence, while the lower equilibrium moves off toward −∞. If we
tune a from zero down to negative values, a similar thing happens. The
original graph of f (x) drops, the lower two equilibria move together
and annihilate each other, while the upper equilibrium moves off to
+∞ (Figure 8.17).
If we graph the path of the equilibria in Figure 8.17 as a function
of a, we obtain the bifurcation diagram in Figure 8.18. The stability
arrows in Figure 8.18 come from those in Figure 8.17 via the geometric
method. This bifurcation diagram is called a hysteresis and is composed
of two saddle-node bifurcations. In Exercise 9, you will use the
linearization theorem to verify the stabilities shown in Figure 8.18.
Exercise 10 illustrates a hysteresis loop, which occurs if the stabilities
of the segments of the branch are reversed from what they are in Figure
8.18.

8.6 EXERCISES
1. Use the method of separation of variables from Chapter 7 to
show that the general solution of dx/dt = rx is x (t) = cert for
c ∈ R. Keep track of lost equilibrium solutions and the values of
constants as we did in Chapter 7.
180 ■ Mathematical Modeling in Biology

2. Use the computer to estimate the parameters in equation (8.11)

from the system of equations (8.10). Hint: To numerically solve
g(x) = h(x), one can use the Nelder-Mead algorithm to minimize
2
the function (g(x) − h(x)) ; if the minimum is zero, you have
obtained a solution. Don’t forget that you may need to try
different initial guesses for the downhill method.

3. In this exercise, you will fit the logistic model

dx  x
= rx 1 −
dt K
to the world population data for the years 1900–2020 in Tables
8.1–8.2 in order to estimate the value at which the population is
expected to level off, that is, the value of the carrying capacity
K. One could approach this in several different ways, but here
we specify two such methods.

a. Try fitting the particular solution

Kx0
x(t) =
x0 + (K − x0 ) e−rt

of the logistic model (which we obtained in Chapter 7) to the

data using the method of nonlinear least squares under the
assumption of demographic noise. Here, x0 = 1.65, where
the units are in billions and t = 0 denotes year 1900. Hint:
You will need to use your skills from Chapter 2 to write a
program that minimizes the RSS and estimates the values
of r and K (the latter of which is in units of billions). You
should be able to modify one of your existing programs from
that chapter. Remember to try multiple starting values for
your parameters.
b. Now try fitting the general solution

Cert
x(t) =
1 + Cert /K

of the logistic model (which we obtained in Chapter 7)

to the data using the method of nonlinear least squares
under the assumption of demographic noise. Here, you are
 Scalar Differential Equations ■ 181

letting the initial condition “float,” meaning that you are

not forcing the curve to fit the data point at t = 0, and you
are estimating C as well as r and K.
c. Graph the scatter plot of data along with the fitted logistic
curves for both fitting methods above.
d. What are your thoughts about the logistic model as a
description of world population growth during the years
1900–2020? Can you propose a simple model that you think
might fit better?
e. What are your thoughts about world population growth, and
have they changed in any way after reading this chapter?
4. In this problem, you will do a complete equilibrium stability
analysis for the logistic model with Allee effects
dx  x
= rx (x − a) 1 −
dt K
r, K > 0.
a. Find all the equilibria.
b. In the context of the linearization theorem, what is f (x) in
this model? What is df /dx? Use the extended product rule
when computing df /dx.
c. Use the linearization theorem to determine conditions on
the parameter a that guarantee the stability/instability of
each equilibrium.
d. Assume r, K > 0 are fixed, and assume you can manipulate
the parameter a. On one graph, plot each equilibrium xe
versus a. Use the ranges −∞ < x < ∞ and −∞ < a < ∞.
This is a bifurcation diagram.
e. Indicate stabilities on your bifurcation diagram.
5. Consider the nonlinear differential equation
dx
= f (x) (8.12)
dt
with hyperbolic equilibrium xe . Recall that xe is an equilibrium if
and only if f (xe ) = 0. Before beginning this problem, you should
review Section 3.5 of Chapter 3.
182 ■ Mathematical Modeling in Biology

a. In calculus, you learned how to linearize a function about

a point by means of a tangent line. Use that procedure to
prove that
f (x) ≈ λ (x − xe )
df
for x ≈ xe , where λ = dx ̸ 0 in this
(xe ) . Explain why λ =
particular problem.
b. Write down a linear differential equation that approximates
the nonlinear one (8.12) for x ≈ xe .
c. Define the variation or displacement of the system from
equilibrium to be u(t) = x(t) − xe . How is du/dt related
to dx/dt?
d. If the system is near equilibrium, that is, if x ≈ xe , then
what can we say about the value of the variation u?
e. Use the change of variables u = x − xe to derive the
differential equation
du
= λu. (8.13)
dt
This is the variation equation. It is also called the
linearization of dx/dt = f (x) at the equilibrium xe .
f. The equilibrium of equation (8.12) is xe . The equilibrium
of equation (8.13) is u = 0. Explain why this makes sense
intuitively in terms of the meaning of u(t).
g. What is the general solution of equation (8.13)?
h. Using the general solution of equation (8.13), explain why

λ > 0 =⇒ ue = 0 unstable
λ < 0 =⇒ ue = 0 asymptotically stable.

i. Intuitively, how should the sign of λ and the resulting

stability/instability of ue = 0 be related to the stabil-
ity/instability of xe ? Explain. Compare your answer to the
linearization theorem.
j. How is the linearization theorem for one-dimensional differ-
ential equations different from the linearization theorem for
one-dimensional difference equations?
 Scalar Differential Equations ■ 183

6. In this problem, you will verify the stabilities of the equilibrium

branches in Figure 8.14.

a. Use the linearization theorem to verify the stabilities of the

equilibrium branches in Figure 8.14.
b. Use the geometric method to verify the stabilities of the
equilibrium branches in Figure 8.14. The geometric method
consists of drawing all possible phase line portraits (as a
function of a) with their stability arrows and then fitting
them together vertically to determine the stabilities of the
equilibrium branches in the bifurcation diagram.

7. Repeat Exercise 6 for Figure 8.15.

8. Repeat Exercise 6 for Figure 8.16.

9. Use the linearization theorem to verify the stabilities of the

equilibrium branches in Figure 8.18.

10. Consider
x′ = −x3 + x + a.

a. Draw the bifurcation diagram, using a as a bifurcation

parameter. Include stability arrows.
b. The bifurcation diagram contains a hysteresis composed of
two saddle-node bifurcations. Prove
√ that the saddle-node
bifurcations occur at a = ±2/(3 3).
c. Consider your bifurcation diagram. Imagine “tuning” the
parameter a from −∞ to +∞. For each fixed value of a,
we obtain a specific differential equation, and the system’s
phase line portrait is the vertical line through that value
of a. We might say that the dynamics always √ “live” on a
vertical line. If a is fixed with a < −2/(3 3), the system
approaches an equilibrium on the lower branch of equilibria,
which is √stable. As a is tuned through the bifurcation
√ value
a = 2/(3 3) and becomes slightly larger than 2/(3 3), the
lower branch disappears and the system must approach an
equilibrium on the upper branch, which is also stable. If
a is then tuned back down through the bifurcation value
184 ■ Mathematical Modeling in Biology
√
a = −2/(3 3), which branch does the system approach?
Using two large vertical arrows, draw the hysteresis loop on
the bifurcation diagram. This is the loop traced out by the
system’s √
final states as the parameter a is increased through
√
a = 2/(3 3) and then decreased through a = −2/(3 3).

11. This exercise uses the skills developed in Exercise 10. Consider
the logistic model for a fishery with an extra removal rate due to
fishing. Suppose the rate at which fish are removed due to fishing
is a constant h fish per unit time. The modified model is then
dx  x
= rx 1 − −h
dt K
with r, K > 0 and h ≥ 0.

a. Draw the bifurcation diagram using h as the bifurcation

parameter. Include stability arrows. Hint: The parameter h
translates the graph of f vertically.
b. Find the value of h for which the system has a saddle-node
bifurcation.
c. What happens in this fishery if h is increased slightly beyond
the bifurcation point? This unheralded collapse is sometimes
called a blue-sky bifurcation, because it comes “out of the
blue.”
d. If fishing is then reduced to levels below the bifurcation value
of h, does the fishery recover?
e. There is no hysteresis loop in this bifurcation diagram. What
must happen for the fishery to recover?

12. Draw the bifurcation diagrams for the following systems, using r
as a bifurcation parameter. Include stability arrows.

a. dx/dt = (r − 2) (x − r)
b. dx/dt = (x − 2) (x − r)
c. dx/dt = sin (r)
2
d. dx/dt = x (x − r)


e. dx/dt = x − r2 (r − x)
 Scalar Differential Equations ■ 185

BIBLIOGRAPHY
Cushing, J. M. 2004. Differential Equations: An Applied Approach. Prentice-
Hall, Upper Saddle River, NJ. [A good source for more phase line portrait
and bifurcation diagram exercises, as well as many interesting applied
projects using differential equations, including the “Doomsday Model.”]

Malthus, T. R. 1798. An Essay on the Principle of Population. J. Johnson,

London. [Malthus noted that there are not enough resources to keep up
with geometric (that is, exponential) population growth.]

United Nations Population Division. https://population.un.org.

von Foerster, H., Mora, P. M., and Amiot, L. W. 1960. Doomsday: Friday,
13 November, A.D. 2026. Science 132:1291–1295. DOI: 10.1126/sci-
ence.132.3436.1291. [As the authors wrote, “At this date human
population will approach infinity if it grows as it has grown in the last
two millenia.”]
 CHAPTER 9

Systems of Differential
Equations

9.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

In this chapter, we consider how to model continuous-time systems
whose descriptions require more than one state variable. We consider
only autonomous systems of differential equations (ones which do not
explicitly depend on time). Before beginning this chapter, you should
review basic matrix operations and introductory linear algebra by
working through the self-guided tutorial in Appendix A. A course
in linear algebra is not a prerequisite for this textbook; everything
you need to know from linear algebra for this and other chapters
is contained in Appendix A. You will need a computer program to
quickly graph phase plane portraits. The free Java version of PPLANE
(Castellanos and Polking) is very easy to use for this purpose.
In this chapter, you will need to know how to take partial
derivatives, a topic from third-semester calculus that we already
encountered in Chapter 5. As a reminder, they are easy to learn. If
you have a function of, say, two variables f (x, y), then the partial
derivative of f with respect to x, denoted ∂f /∂x, is the derivative
of f with respect to x while holding y constant. So, for example, if
f (x, y) = x2 + y 2 + x3 y 3 , then the partial derivatives are (showing
zeros to make a point):

∂f
= 2x + 0 + y 3 3x2


∂x

DOI: 10.1201/9781003265382-12 187

188 ■ Mathematical Modeling in Biology

∂f
= 0 + 2y + x3 3y 2 .


∂y
If this is your first experience with partial derivatives, work a few
problems in a calculus book to get comfortable with them.
The main ideas in this chapter are to learn how to solve and analyze
linear systems and then to apply those techniques to nonlinear systems
by means of linearization. You should keep track of how these methods
parallel those for discrete-time systems in Chapter 5. Thus, before
beginning this chapter you should reread the sections in Chapter 5 that
deal with solving linear discrete-time systems and the linearization of
nonlinear discrete-time systems.

9.2 LINEAR SYSTEMS OF ODES AND PHASE PLANE ANALYSIS

Consider the linear system of ordinary differential equations (ODEs)

x′ = ax + by (9.1)
′
y = cx + dy,

in which the “prime” indicates the derivative with respect to time t.

System (9.1) is linear because the terms are linear functions of the state
variables x and y and their derivatives. The equilibria are the solutions
of the algebraic system

0 = ax + by
0 = cx + dy.

The only equilibrium is (0, 0) as long as the determinant of the

coefficient matrix is nonzero, that is, as long as
a b
̸ 0.
=
c d
In what follows, we will always assume this holds.
There are six generic types of linear two-dimensional systems
(“generic” meaning those with distinct, nonzero eigenvalues). We use
an example to illustrate each type. Carefully work through the details
of the first example; the remaining five examples are presented without
showing all the work. Make sure you know how to work out the details
of each example by hand.
 Systems of Differential Equations ■ 189

9.2.1 Unstable Node

Consider
x′ = 2x + 2y (9.2)
1
y′ = x + 2y.
2
To solve this system of differential equations, we propose an Ansatz.
Remembering from previous chapters that the linear scalar differential
equation x′ = rx has exponential solutions of the form x = cert , we
propose an Ansatz of the form
x (t) = v1 eλt
y (t) = v2 eλt
and look for values of λ, v1 , and v2 , where v1 and v2 are not both zero.
Plugging the Ansatz into the system of differential equations (9.2), we
have
λv1 eλt = 2v1 eλt + 2v2 eλt
1
λv2 eλt = v1 eλt + 2v2 eλt ,
2
or
(2 − λ) v1 + 2v2 = 0 (9.3)
1
v1 + (2 − λ) v2 = 0.
2
This algebraic system can be visualized as two lines in the v1 -v2 plane,
both passing through the origin since (v1 , v2 ) = (0, 0) is a solution. In
order for two such lines to have a nontrivial solution (v1 , v2 ), they must
be coincident (i.e., the same line), which is true if and only if
2−λ 2
1 = 0.
2 2−λ
This holds if and only if
2
(2 − λ) − 1 = 0 (9.4)
2
(2 − λ) = 1
2 − λ = ±1
λ = 1, 3.
190 ■ Mathematical Modeling in Biology

Equation (9.4) is called the characteristic equation, and its solutions

λ = 1, 3 are the eigenvalues of the system.
Now we consider each eigenvalue in turn and look for the
corresponding v1 , v2 , not both zero. If λ = 1, then from system (9.3)
we obtain

v1 + 2v2 = 0
1
v1 + v2 = 0.
2
 
2
Thus, v1 = −2v2 , so an eigenvector belonging to λ = 1 is . The
−1
corresponding eigensolution is
   
x 2
= et .
y −1

If λ = 3, then

−v1 + 2v2 = 0
1
v1 − v2 = 0.
2
 
2
Thus, v1 = 2v2 , so an eigenvector belonging to λ = 3 is and the
1
corresponding eigensolution is
   
x 2
= e3t .
y 1

The general solution is a linear combination of the two independent

eigensolutions:
     
x 2 2
= c1 et + c2 e3t .
y −1 1

Here, you should think of the c1 et and c2 e3t as scalar multipliers on

the eigenvectors that stretch or shrink the eigenvectors.
In order to graph the general solution, that is, the set of all
solutions, we graph temporal orbits in the phase plane, that is, x-y
space. The temporal aspect is denoted by arrows indicating increasing
 Systems of Differential Equations ■ 191

time. First, we graph the orbits of the eigensolutions. These are

simply the lines in the phase plane passing through the origin that
are determined by the eigenvectors, with appropriate time arrows
(Figure 9.1). These lines are called manifolds. Both of the manifolds
in this example are called “unstable manifolds” because the system
moves away from the origin as time progresses (the scalars et and e3t
are increasing). For other solutions, as t → ∞, the scalar e3t dominates
the solution, which begins to roughly parallel the manifold determined
by eigenvalue (2, 1)⊤ . On the other hand, when t → −∞, the scalar
et begins to dominate, and, going backward in time, solutions roughly
parallel the manifold determined by (2, −1)⊤ . Figure 9.1 shows the
phase plane portrait for this system with the manifolds and a few other
orbits. This type of phase plane portrait for the equilibrium (here, the
origin) is called an unstable node.

9.2.2 Asymptotically Stable Node

Consider the system

x′ = −3x − y
y ′ = 2x.

Following the same procedure as in the previous section, we find the

eigenvalues by subtracting λ off the diagonal of the coefficient matrix

Figure 9.1: The equilibrium (0,0) is an unstable node.

192 ■ Mathematical Modeling in Biology

and setting the determinant equal to zero,

−3 − λ −1
= 0,
2 0−λ

so that

−λ (−3 − λ) + 2 = 0
λ2 + 3λ + 2 = 0
(λ + 1) (λ + 2) = 0.

The solutions of this characteristic equation are the eigenvalues λ =

−1, −2. Using the procedure from the previous section to find the
corresponding eigenvectors, we obtain
   
−1 1
for λ = −2 and for λ = −1.
1 −2

Thus, the general solution is

     
x 1 −t −1
= c1 e + c2 e−2t .
y −2 1

Here, both manifolds are stable (because the scalars e−t and e−2t are
approaching zero). The term with e−t dominates in forward time, and
the term with e−2t dominates in backward time, as shown in Figure 9.2.
This phase plane portrait is an asymptotically stable node (Figure 9.2).

9.2.3 Saddle
Consider the system

x′ = x
y ′ = −x − 2y.

The eigenvalues are the solutions of

1−λ 0
= 0,
−1 −2 − λ
 Systems of Differential Equations ■ 193

Figure 9.2: The equilibrium (0,0) is an asymptotically stable node.

that is, the solutions of the characteristic equation

(1 − λ) (−2 − λ) = 0.

The eigenvalues are λ = 1, −2, and the corresponding eigenvectors are

(make sure you can obtain these by hand)
   
0 −3
for λ = −2 and for λ = 1.
1 1
Thus, the general solution is
     
x 0 −2t −3
= c1 e + c2 et .
y 1 1

Here, we have a stable manifold and an unstable manifold, which gives

rise to a saddle (Figure 9.3).
Note that a saddle is unstable. In two dimensions, it has one
unstable direction and one stable direction. In higher dimensions, there
is at least one unstable direction and at least one stable direction.

9.2.4 Center
Consider the harmonic oscillator

x′ = y (9.5)
′
y = −x.
194 ■ Mathematical Modeling in Biology

Figure 9.3: The equilibrium (0,0) is a saddle. Saddles are unstable.

The characteristic equation for system (9.5) is

−λ 1
= 0,
−1 −λ

that is,
λ2 + 1 = 0.
The eigenvalues λ = ±i are complex numbers, in this case purely
imaginary numbers. For λ = i, we find the eigenvalues as before:

−iv1 + v2 = 0
−v1 − iv2 = 0.

If we take v1 = i, then v2 = −1. For λ = −i, we have

iv1 + v2 = 0
−v1 + iv2 = 0.

If we take v1 = i, then v2 = 1. Thus, we can express the general solution

over the complex numbers as
     
x i i
= c1 it
e + c2 e−it . (9.6)
y −1 1

We would like to obtain from this a general solution over R2 .

 Systems of Differential Equations ■ 195

From Euler’s formula eit = cos t + i sin t and the facts that cos t is
even and sin t is odd, one can obtain the complex solution (Exercise 1)
     
x sin t cos t
= (c2 − c1 ) + i (c1 + c2 ) . (9.7)
y cos t − sin t
A theorem from differential equations says that the real and imaginary
parts of this complex solution are two linearly independent real
solutions, in this case
   
sin t cos t
and .
cos t − sin t

The general solution over R2 is therefore

     
x sin t cos t
= c1 + c2 .
y cos t − sin t
It is straightforward to check that the component-wise form

x = c1 sin t + c2 cos t (9.8)

y = c1 cos t − c2 sin t

satisfies system (9.5) and that the solutions are closed curves in the
phase plane (Exercise 3). The direction of rotation can be determined
by choosing a test point, say x = 1, y = 0, and checking the signs of the
time derivatives at that point (x′ = 0, y ′ = −1). This tells us that y
is decreasing at that point, and so the rotation is clockwise in forward
time.
When eigenvalues are purely imaginary, as in this example, the
phase plane portrait is called a center (Figure 9.4). A center is neutrally
stable.

9.2.5 Unstable Spiral

Consider the system

x′ = x − y
y ′ = x + y.

Check that the eigenvalues are λ = 1 ± i. Note that e(1±i)t = et e±it =

et (cos t ± i sin t) . As in the previous example, the system oscillates,
196 ■ Mathematical Modeling in Biology

Figure 9.4: The equilibrium (0,0) of system (9.5) is a center with

concentric circles as orbits. Centers are neutrally stable.

but this time with increasing amplitude et . The equilibrium (0, 0) is

called an unstable spiral (Figure 9.5). The direction of rotation can be
determined with a test point as it was for the center. Check that in
this example, the spiral is counterclockwise.

9.2.6 Asymptotically Stable Spiral

Consider

x′ = −x + 2y
y ′ = −5x − 3y.

Figure 9.5: The equilibrium (0,0) is an unstable spiral.

 Systems of Differential Equations ■ 197

Figure 9.6: The equilibrium (0,0) is an asymptotically stable spiral.

Check that the eigenvalues are λ = −2 ± 3i. Note that e(−2±3i)t =

e−2t (cos 3t ± i sin 3t) . In this case, the system oscillates, but with
decreasing amplitude. The equilibrium (0, 0) is an asymptotically stable
spiral (Figure 9.6). Check that the direction of rotation is clockwise.

9.2.7 Summary: Eigenvalues Tell All

The six “generic” types of phase plane portraits for two-dimensional

linear systems are determined by the eigenvalues as summarized here
in Table 9.1 and the corresponding Figure 9.7. For the “nongeneric”
other cases, see Henson (2012).

TABLE 9.1 The Six Generic Phase Plane Portraits. (See Figure 9.7)

Figure 9.7 Eigenvalues Phase Portrait Type

Panel (a) λ1 < λ2 < 0 Asymptotically stable node
Panel (b) 0 < λ1 < λ2 Unstable node
Panel (c) λ1 < 0 < λ2 Saddle (unstable)
Panel (d) λ1,2 = ±bi Center (neutrally stable)
Panel (e) λ1,2 = a ± bi, a < 0 Asymptotically stable spiral
Panel (f) λ1,2 = a ± bi, a > 0 Unstable spiral
198 ■ Mathematical Modeling in Biology

(a) (d)

(b) (e)

(c) (f)

Figure 9.7: The six generic phase plane portraits. (a) Asymptotically
stable node. (b) Unstable node. (c) Saddle. (d) Center. (e) Asymptot-
ically stable spiral. (f) Unstable spiral. (See Table 9.1.)

9.3 NONLINEAR SYSTEMS OF ODES

Consider the nonlinear system

x′ = f (x, y)
y ′ = g (x, y)

with equilibrium (xe , ye ). As in Chapter 5, we approximate f and g

 Systems of Differential Equations ■ 199

with tangent planes near the equilibrium:

∂f ∂f
x′ = f (x, y) ≈ f (xe , ye ) + (xe , ye ) (x − xe ) + (xe , ye ) (y − ye )
∂x ∂y
∂g ∂g
y′ = g (x, y) ≈ g (xe , ye ) + (xe , ye ) (x − xe ) + (xe , ye ) (y − ye )
∂x ∂y
for x ≈ xe and y ≈ ye . Note that f (xe , ye ) = 0 and g(xe , ye ) = 0
because (xe , ye ) is an equilibrium. Define the variation or displacement
from equilibrium by the variables
u = x − xe
v = y − ye .
Note that u′ = x′ and v ′ = y ′ . Thus, near the equilibrium, the variation
from equilibrium can be approximated by the linear system
∂f ∂f
u′ ≈ (xe , ye ) u + (xe , ye ) v
∂x ∂y
∂g ∂g
v′ ≈ (xe , ye ) u + (xe , ye ) v,
∂x ∂y
for u ≈ 0 and v ≈ 0. This means that near the equilibrium, the system’s
behavior is determined by the eigenvalues of the coefficient matrix,
which is in this case the Jacobian
!
∂f ∂f
∂x (x e , y e ) ∂y (xe , ye )
∂g ∂g .
∂x (xe , ye ) ∂y (xe , ye )

Let’s apply this to an example. Consider the nonlinear system

x′ = x − xy
y ′ = 2y − 2xy.
First, we set up the equilibrium equations and factor:
x (1 − y) = 0
2y (1 − x) = 0.
From the first equation, x = 0 or y = 1. If x = 0, then from the
second equation y = 0. If, on the other hand, y = 1, then from the
200 ■ Mathematical Modeling in Biology

second equation x = 1. Thus, the two equilibria are (0, 0) and (1, 1) .
We linearize around each equilibrium in turn. The Jacobian is
 
1−y −x
J= .
−2y 2 − 2x

At the equilibrium (0, 0) , the Jacobian is

 
1 0
J (0, 0) = .
0 2

 
1
The eigenvalues are λ = 1 with eigenvector and λ = 2 with
0
 
0
eigenvector . Thus, the (0, 0) equilibrium is an unstable node
1
for the linear system, with unstable manifolds on the axes. Therefore,
as we shall see, the (0, 0) equilibrium of the nonlinear system is also an
unstable node, and the phase plane locally around (0, 0) is topologically
equivalent to that of the linear system.
At the equilibrium (1, 1) , the Jacobian is
 
0 −1
J (1, 1) = .
−2 0

√ √
 
1
√
The eigenvalues are λ = 2 with eigenvector and λ = − 2
− 2
 
√1
with eigenvector . Thus, the (0, 0) equilibrium of the linear
2
system is a saddle, with unstable and stable manifolds determined by
these eigenvectors. Therefore, as we shall see, the (1, 1) equilibrium
of the nonlinear system is also a saddle, and the phase plane locally
around (1, 1) is topologically equivalent to that of the linear system
around (0, 0).
Now we can piece the local information together to approximate
the phase plane portrait for the nonlinear system (Figure 9.8).
We make these ideas more precise in the next section.
 Systems of Differential Equations ■ 201

Figure 9.8: Nonlinear phase plane portrait.

9.3.1 Linearization
Definition 9.1 The linearization of a nonlinear system

x′ = f (x, y)
y ′ = g (x, y)

at an equilibrium (xe , ye ) is the linear system

∂f ∂f
u′ = (xe , ye ) u + (xe , ye ) v
∂x ∂y
∂g ∂g
v′ = (xe , ye ) u + (xe , ye ) v,
∂x ∂y
that is,
u′ = Ju,
 
u
where u = is the vector of state variables and J is the Jacobian
v
matrix evaluated at the equilibrium (xe , ye ) .

Definition 9.2 An equilibrium (xe , ye ) of a system

x′ = f (x, y)
y ′ = g (x, y)

is hyperbolic if and only if all of the eigenvalues of the Jacobian matrix

J (xe , ye ) have nonzero real parts.
202 ■ Mathematical Modeling in Biology

Theorem 9.1 (Linearization Theorem) If f (x, y) and g (x, y) are

continuously differentiable in both variables and (xe , ye ) is a hyperbolic
equilibrium of the system

x′ = f (x, y)
y ′ = g (x, y) ,

then
(1) if all the eigenvalues of the Jacobian J (xe , ye ) have negative real
parts, (xe , ye ) is asymptotically stable,
(2) if at least one of the eigenvalues of J (xe , ye ) has positive real part,
then (xe , ye ) is unstable.

The Hartman-Grobman theorem says that near hyperbolic

equilibria, the nonlinear phase plane portrait is (locally) topologically
equivalent to the phase plane portrait of the linearization.

9.4 LIMIT CYCLES, CYCLE CHAINS, AND BIFURCATIONS

If a nonlinear two-dimensional system is autonomous (meaning there
is no explicit time dependence in the model equations), it can have
a limited number of other phase plane configurations that are not
possible in autonomous linear systems.
Periodic solutions of two-dimensional systems are associated with
closed-loop orbits, called cycles, in the phase plane. A famous example
that has concentric cycles in the phase plane (that is, a nonlinear cen-
ter) is the Lotka-Volterra predator-prey model, which we will discuss
in the next section. A cycle always surrounds at least one equilibrium.
If it attracts nearby orbits, it is called a limit cycle (Figure 9.9a). We
will see an example of a limit cycle in the Van der Pol oscillator (9.9)
below.
Another type of configuration in the phase plane is the cycle
chain (also known as a heteroclinic cycle). This occurs when multiple
equilibria are connected by heteroclinic orbits (orbits that connect
two different equilibria) in a loop structure (Figure 9.9b). Oscillatory
solutions may approach a cycle chain, misleading the researcher to
think they are approaching a limit cycle. The cycle chain, however, is
 Systems of Differential Equations ■ 203

(a)

(b)

Figure 9.9: (a) Limit cycle of the Van der Pol oscillator (equation 9.9).
(b) Cycle chain.

not itself an orbit (although it is an invariant set). You will analyze

systems with cycle chains in Exercises 8 and 9.
A bifurcation is an abrupt change in phase portrait type that occurs
as a parameter is tuned through a critical value. The following example
204 ■ Mathematical Modeling in Biology

illustrates both limit cycles and bifurcations. Consider the Van der Pol
oscillator

x′ = y − x3 + ax (9.9)
′
y = −x,

in which a ≈ 0 (a is close to zero). You will show in Exercise 6 that

if a < 0, the equilibrium is hyperbolic and the nonlinear system has
an asymptotically stable clockwise spiral at the origin. If a is “tuned”
up to a = 0, the real parts of the eigenvalues become zero, so the
equilibrium is now non-hyperbolic and the linearization theorem does
not apply. Although the linearized system has a center at the origin,
you can see from simulations in PPLANE that the nonlinear system
still has an asymptotically stable spiral at the origin. If a is further
tuned up to a > 0, however, the equilibrium again becomes hyperbolic
and the nonlinear system has an unstable clockwise spiral at the origin.
Thus, the Van der Pol oscillator has a bifurcation at a = 0.
For small a > 0 in equation (9.9), local trajectories spiral out, away
from the origin, yet trajectories farther away from the origin are still
spiraling in toward the origin. These trajectories meet in a limit cycle
(Figure 9.9a). Thus, as a is tuned from negative to positive values, the
limit cycle is born at the origin when a = 0 and grows in radius as a
increases.
A theorem known as the Poincare-Bendixson theorem guaran-
tees that bounded orbits in the phase plane must approach equilibria,
cycles, or sets of equilibria connected by heteroclinic and homoclinic
orbits.

9.5 LOTKA-VOLTERRA MODELS AND NULLCLINE ANALYSIS

We cannot end this chapter without discussing the famous Lotka-
Volterra models of ecology. They usually are studied by means of
nullcline (or isocline) analysis. When it is complicated to solve for
equilibria and eigenvalues in a nonlinear system, one can determine
the direction of the vector field in the phase plane by means of curves
of zero growth rate. A nullcline for state variable x is a curve in the
phase plane along which dx/dt = 0. The intersections of the nullclines
for x and y are exactly the equilibria.
 Systems of Differential Equations ■ 205

9.5.1 Lotka-Volterra Competition

The well-known Lotka-Volterra competition model endowed ecology
with the theoretical idea of competitive exclusion. If x and y are two
competing species, then both have negative interactions with the other.
If we assume that each species grows logistically in the absence of the
other species and that the two species interact through mass action,
the model is

x′ = x (r1 − a11 x − a12 y) (9.10)

′
y = y (r2 − a21 x − a22 y)

with positive coefficients.

We find the nullclines for x by setting

0 = x (r1 − a11 x − a12 y) ,

which occurs for x = 0 and y = r1 /a12 − a11 x/a12 . Along these

two lines, the rate of change of x is zero; that is, the vector field
is vertical. Similarly, the nullclines for y are the lines y = 0 and
y = r2 /a22 − a21 x/a22 . Along these lines, the rate of change of y is
zero, and the vector field is horizontal. These two sets of nullclines have
four possible configurations in phase space (Figure 9.10). In two of the
cases, one species always wins (Figure 9.10a and b). In the third case,
there is a coexistence equilibrium, but it is a saddle and the winner
depends on the initial condition (Figure 9.10c). In the fourth case,
the coexistence equilibrium is a stable node and the species coexist
(Figure 9.10d).
Coexistence happens when the competition within the species is
stronger than the competition between the species (Exercise 10).

9.5.2 Lotka-Volterra Cooperation

In cooperation, both species benefit from the other. If we assume that
each species grows logistically in the absence of the other, with a mass
action benefit when together, the model is

x′ = x (r1 − a11 x + a12 y) (9.11)

′
y = y (r2 + a21 x − a22 y)
206 ■ Mathematical Modeling in Biology

(a) (c)

(b) (d)

Figure 9.10: Possible phase plane portraits for Lotka-Volterra com-

petition model (9.10). (a) Competitive exclusion: Species x wins. (b)
Competitive exclusion: Species y wins. (c) Competitive exclusion:
Winner depends on initial condition. (d) Stable coexistence.

with positive coefficients. The nullclines for x are the lines x = 0

and y = −r1 /a12 + a11 x/a12 , and the nullclines for y are the lines
y = 0 and y = r2 /a22 + a21 x/a22 . Figure 9.11 shows the two possible
configurations in the phase plane. In the first case (Figure 9.11a), the
cooperators coexist. In the second case, however, the mutual benefit
is so strong that it overwhelms the logistic limitations to growth and
both species grow without bound (Figure 9.11b)! Robert May famously
referred to the second case as an “orgy of mutual benefaction.” See
Exercise 11.

9.5.3 Lotka-Volterra Predator-Prey

The traditional Lotka-Volterra predator-prey model is

x′ = x (r1 − a12 y) (9.12)

′
y = y (−r2 + a21 x)

with positive coefficients, where x is the prey density and y is the

predator density. It is important to note that this model is quite
 Systems of Differential Equations ■ 207

(a)

(b)

Figure 9.11: Phase plane portraits for Lotka-Volterra cooperation

model (9.11). (a) Coexistence; (b) Robert May’s “orgy of mutual
benefaction.”

different from the previous two in that it assumes the prey population
grows exponentially in the absence of predators and that the predator
population declines exponentially in the absence of prey. Neither of
these assumptions is very realistic, because one would expect that
(i) both species would have a self-regulatory density-dependent term
such as the quadratic nonlinearity in the logistic model, and (ii) the
predators would switch to other prey in the absence of species x.
The nullclines for x are the lines x = 0 and y = r1 /a12 , and the
nullclines for y are the lines y = 0 and x = r2 /a21 . It is not clear
from the nullclines (Figure 9.12) whether the coexistence equilibrium
is a spiral or a nonlinear center. It is possible to prove, however, that
208 ■ Mathematical Modeling in Biology

Figure 9.12: Nullclines for the Lotka-Volterra predator-prey system

(9.12). It is possible to prove that the orbits are closed curves
surrounding the equilibrium, forming a nonlinear center.

the orbits are closed curves and hence surround a nonlinear center.
When the periodic solutions are graphed as time series, they exhibit
the well-known predator-prey oscillations.
One might consider a more realistic version of the Lotka-Volterra
predator-prey model. For example, the Seabird Ecology Team used

x′ = rx − rx2 /K − axy
y ′ = sy − sy 2 /C + bxy

to model eagle-gull predator-prey dynamics on Protection Island from

1980–2016, where x is the number of gulls and y is the number of
eagles (Henson et al. 2019). Here, both populations are assumed to
grow logistically in the absence of the other. We showed that the gull
population dynamics could be explained by the number of occupied
bald eagle territories in Washington with generalized R2 = 0.82. This
supported the hypothesis that the rise and decline in gull numbers
observed on Protection Island has been due largely to the decline and
recovery of the bald eagle population. See Chapter 10, Exercise 9.

9.6 EXERCISES
You will need a computer program to quickly graph phase planes. The
free PPLANE Java version (Castellanos and Polking) is very easy to
use for this purpose.
 Systems of Differential Equations ■ 209

1. Obtain equation (9.7) from equation (9.6).

2. Consider the linear system

x′ = Ax. (9.13)

For each of the six matrices A below, do the following calculations

by hand:

a. Find the eigenvalues. If the eigenvalues are real:

i. Find eigenvectors.
ii. Give two independent eigensolutions of equation (9.13).
iii. Give the general solution of equation (9.13).
b. Draw the phase plane portrait. If the eigenvalues are real,
include the manifolds (orbits of the eigensolutions) in your
sketch, along with a few other representative orbits. If the
eigenvalues are complex, show the direction of rotation.
c. Classify the (0, 0) equilibrium type and give its stability
(unstable node, asymptotically stable node, unstable spiral,
asymptotically stable spiral, center (neutrally stable), or
saddle (unstable)).
d. Check each of your phase portraits using online software
such as PPLANE to graph the vector fields.
 
3 −2
A =
4 −3
 
−1 −2
A =
2 −1
 
−3 −2
A =
0 −2
 
0 2
A =
−2 0
 
3 2
A =
1 2
 
3 2
A =
−2 3
210 ■ Mathematical Modeling in Biology

3. Consider the harmonic oscillator system (9.5).

a. Prove that equations (9.8) are solutions of system (9.5) for

all c1 , c2 ∈ R.
b. Prove that solutions (9.8) describe closed curves in the phase
plane.
p In particular, they describe concentric circles of radius
c1 + c22 centered on the origin.
2

4. Consider the nonlinear system

x′ = 2x + y 2
y ′ = x − y.

a. Find all equilibria.

b. Find the Jacobian at each equilibrium.
c. Which of the equilibria are hyperbolic?
d. For each hyperbolic equilibrium, determine the phase
portrait/stability type using the linearization theorem.
e. Use PPLANE to draw the complete phase portrait for this
nonlinear system.

5. Consider the nonlinear system

x′ = y − x3
y ′ = −x.

a. Show that (0, 0) is the only equilibrium.

b. Find the Jacobian at (0, 0).
c. Show that (0, 0) is nonhyperbolic and that the linearized
equation has a center at (0, 0).
d. Use PPLANE to see that (0, 0) is actually an asymptotically
stable spiral for the nonlinear equation.

6. Consider the Van der Pol equation from circuits

x′ = y − x3 + ax
y ′ = −x,
where a ≈ 0.
 Systems of Differential Equations ■ 211

a. Show that the only equilibrium is (0, 0).

b. Find the linearization at (0, 0). (Your answer should be a
linear system of DEs.)
c. Find the eigenvalues of the Jacobian matrix at (0, 0).
d. If a ⪆ 0, what is the phase plane portrait? Note: Let a ⪆ 0
denote positive values of a that are close to zero.
e. If a ⪅ 0, what is the phase plane portrait?
f. If a = 0, what (if anything) can you conclude from the
linearization theorem?
g. Use PPLANE to investigate how the phase plane portrait
changes as a ranges from a = −0.1 to a = 0.1. Use a window
size of −2 < x < 2 and −2 < y < 2. Note the birth of the
stable limit cycle (closed-loop solution) as a passes through
the critical value acr = 0. This is similar to what’s called a
Hopf bifurcation.
h. For small a > 0, can you form a conjecture about the radius
of the limit cycle as a function of a?

7. Carry out an isocline analysis of the Van der Pol equation in

problem (6). Do a separate analysis for each of a = 0, a ⪅ 0, and
a ⪆ 0.

8. This problem, designed by J. M. Cushing (Cushing 2004),

illustrates a cycle chain (heteroclinic cycle). Consider the system

x′ = x2 − 1 y


 
′ 2

3
y = 1−y x+ y .
10

a. There are seven equilibria. Find them and graph them on

the phase plane.
b. Find the Jacobian matrix as a function of x and y.
c. For each of the seven equilibria in turn, find the eigenvalues,
linearized phase plane portrait, and stability.
d. Along each of the two vertical lines x = ±1 in the phase
plane, what are the dynamics of y?
212 ■ Mathematical Modeling in Biology

e. Along each of the two horizontal lines y = ±1 in the phase

plane, what are the dynamics of x?
f. Graph the nonlinear phase plane portrait. Check your work
with PPLANE.

9. This problem illustrates a cycle chain. Part (d) requires

integration by separation of variables. If you haven’t had second-
semester calculus or a course in differential equations, you should
review the explanation of this technique in Chapter 7. Consider
the system

x′ = y
y ′ = −x 1 − x2 .

a. There are three equilibria. Find them and place them on the
phase plane in a graph.
b. Find the Jacobian matrix as a function of x and y.
c. Show that two of the equilibria are hyperbolic saddles and
the third is nonhyperbolic (so the linearization theorem does
not apply to it).
d. Prove that all solutions lie on curves of the form x2 + y 2 =
1 4
2 x + C. Hint: Compute dy/dx by recalling that dy/dt =
dy/dx · dx/dt. Use the technique of “separation of variables”
to integrate dy/dx.
e. Prove that the nonhyperbolic equilibrium is (locally) a
nonlinear center. Hint: Prove that if C ∈ (0, 1/2], then the
equations x2 + y 2 = 21 x4 + C describe a family of closed
curves containing the origin for −1 ≤ x ≤ 1.
f. Prove that two heteroclinic orbits connect the two saddle
equilibria. Hint: Prove that the two saddle equilibria lie on
the closed curve associated with C = 1/2.
g. Determine whether the trajectories on the closed curves for
C ∈ (0, 1/2] are clockwise or counterclockwise in forward
time.
h. Graph the nonlinear phase plane portrait and check your
work with PPLANE.
 Systems of Differential Equations ■ 213

10. For the Lotka-Volterra competition model (9.10), find conditions

on the parameters that ensure the possibility of stable coexis-
tence. Give a biological interpretation: Under what conditions is
competitive coexistence possible?

11. For the Lotka-Volterra cooperation model (9.11), find conditions

on the parameters that lead to unbounded growth (Figure 9.11b).
Give a biological interpretation of these conditions.

BIBLIOGRAPHY
Castellanos, J. and Polking, J. C. PPLANE, Java version.
https://www.cs.unm.edu/˜joel/dfield/. [Freeware for graphing phase
plane portraits.]

Cushing, J. M. 2004. Differential Equations: An Applied Approach. Prentice-

Hall, Upper Saddle River, NJ, p. 308. [A good source for more differential
equation exercises and introductory theory, as well as many interesting
applied modeling projects using differential equations.]

Henson, S. M. 2012. Phase plane analysis. In A. Hastings and L. Gross (Eds.)

Encyclopedia of Theoretical Ecology (pp. 538–545). University of California
Press, Berkeley. [Includes the “nongeneric” phase plane portraits not
covered in this chapter.]

Henson, S. M., Desharnais, R. A., Funasaki, E. T., Galusha, J. G., Watson,

J. W., and Hayward, J. L. 2019. Predator-prey dynamics of bald eagles
and glaucous-winged gulls at Protection Island, Washington, DC. Ecology
and Evolution 9:3850–3867. DOI:10.1002/ece3.5011. [Lotka–Volterra-type
predator–prey model fitted to gull nest count data and Washington State
eagle territory data 1980–2016 with R2 = 0.82, suggesting that gull
dynamics were due largely to eagle recovery dynamics. Both theoretical
and empirical: models connected to data.]
 CHAPTER 10

Seabird Behavior: A
Case Study

10.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

This chapter is based on the initial paper (Henson et al. 2004) of a large
body of research carried out by the Seabird Ecology Team, a National
Science Foundation-funded research group co-directed by the authors
of this textbook. The paper, which was published in The Auk, was co-
authored with animal behaviorist Joe Galusha and two undergraduate
research students. That paper, along with others by the team, is cited
in the annotated bibliography for this chapter and can be referred to for
more details on this particular project, as well as further information
on the team’s larger body of work.
The continuous-time modeling techniques in this chapter assume an
understanding of the methods of connecting models to data (Chapter
2), including the use of one-step predictions (Chapter 6), as well as
familiarity with ordinary differential equations (Chapters 7–9). The
chapter also requires the relevant skills in coding (Appendices B–C).

10.2 THE SCIENTIFIC PROBLEM

Wildlife managers and ecologists are interested in temporal changes
in numbers of organisms. Numbers of organisms provide baseline
information about populations—a starting point for understanding
how any ecosystem works.

DOI: 10.1201/9781003265382-13 215

216 ■ Mathematical Modeling in Biology

In the Pacific Northwest region of North America, glaucous-winged

gulls (Larus glaucescens) are important environmental indicators for
the quality of the marine environment, much like the death of canaries
indicated the presence of carbon monoxide in coal mines more than
a century ago. Gulls are ideal for field study because they are large,
abundant, and easily observed. Moreover, they engage in behaviors that
are reasonably complex, but not so complex as to make monitoring and
analysis difficult.
Recall that Chapter 6 explored the use of mathematical models to
understand the dynamics of laboratory populations of flour beetles.
A major feature of the beetle work was the rigorous connection of
models with actual data, an important step beyond the construction
of theoretical models. One of us, Shandelle Henson, was a member of
the “Beetle Team.” Meanwhile, the other one of us, Jim Hayward,
had for several years been collecting time series behavior data in
a breeding colony of gulls on Protection Island National Wildlife
Refuge, Strait of Juan de Fuca, Washington, the USA (48◦ 07′ 40′′ N,
122◦ 55′ 3′′ W). When we compared notes, we wondered: If mathematical
models could work so well for population dynamics of laboratory
insect populations, could they also describe and predict behavioral
dynamics of free-living seabirds on Protection Island? We decided to
find out.
For a pilot study, we chose a simple system. We decided to see if we
could successfully model and predict numbers of glaucous-winged gulls
loafing (standing, sitting, preening, resting, sleeping) on a pier adjacent
to the Protection Island breeding colony. Hayward had a large historical
data set spanning four field seasons in which he had monitored hourly
numbers of gulls in various habitats, including those loafing on the pier.
In the autumn of 2001, we used these data to parameterize and validate
an ordinary differential equation model that described fluctuations in
gull counts in relation to three environmental variables: day of the year,
solar elevation, and tide height. We then used the parameterized model
to predict what the counts would be for specific hours on days during
the (future) 2002 field season. Along with two undergraduate students,
we traveled to Protection Island in the spring of 2002 to test the a
priori model predictions.
This chapter is the story of how the study was carried out
and provides exercises that will allow you to work through the
 Seabird Behavior: A Case Study ■ 217

modeling processes we used. Subsequent team papers that use the same
methodology are found in the bibliography.

10.3 HISTORICAL DATA

10.3.1 Count Data
The Protection Island gull colony contained more than 3,000 gulls
during this study. The gulls move from habitat to habitat during the
day. For example, a breeding gull begins the day on its nesting territory.
It may fly to the nearby beach and wade into the water for a drink.
Then it might fly off to some remote location to feed. When it returns to
the island, instead of going directly to its territory, it may land in a com-
munal loafing area to preen and rest. One popular site for this activity
is the pier located in a small marina located adjacent to the colony.
One day each week Hayward had made an hourly count of the
gulls on the pier during every daylight hour (0600–2100 PDT), May
to August, 1997, 1998, 1999, and 2001. Counts were made hourly so
that they could be compared to fluctuations in environmental variables
such as tide height and solar elevation.

10.3.2 Dividing the Data

The historical count data were assigned randomly to two bins, the
first for parameter estimation and the second for independent model
validation. In order to preserve variability along the entire tidal cycle,
we assigned daily data sets to bins using stratified random sampling:
We first divided the 14-day tidal periods into roughly four quarters. We
then randomly selected half the daily data sets from each quarter for
the parameter estimation bin; the remaining data were placed in the
validation bin. The binned historical data are given in Data Set 10.1.

10.3.3 Tide and Solar Elevation Data

Tides in the Strait of Juan de Fuca are “semi-diurnal,” meaning there
are two high tides and two low tides during each 24-hour period;
moreover, one of the low tides is higher than the other low tide. Tidal
amplitude changes over a 14-day interval, with the lowest amplitude
tide at what we called a “node.” (See arrows, Figures 10.2 and 10.4.)
218 ■ Mathematical Modeling in Biology

We downloaded historical hourly tide height observations in meters

from the National Oceanic and Atmospheric Administration (NOAA)
(see bibliography entry for URL) for the Port Townsend, Washington
station 9444900, for the days corresponding to the historical pier
counts. The tide heights in meters were obtained by multiplying each
NOAA observation by 0.93, which is the correction factor for nearby
Protection Island. We normalized these corrected hourly tide data by
first subtracting the minimal tide value from each tide height in order
to set the minimum equal to zero, then dividing each by the resulting
maximum to set the new maximum equal to one, and finally by adding
one so that the result was between one and two:
tide − min (tide)
T (t) = + 1. (10.1)
max (tide − min (tide))

This resulted in a non-dimensional tide height function with 1 ≤ T (t) ≤

2.
Similarly, we obtained hourly solar elevations from NOAA (see
bibliographic entry for URL). We set negative solar elevations to zero
and then performed a normalization equivalent to equation (10.1)
in order to produce a non-dimensional solar elevation function with
1 ≤ S(t) ≤ 2.

10.4 GENERAL MODEL

We used an ordinary differential equation compartmental model to
track the dynamics on the pier. The net rate of change in the number
of gulls N loafing on the pier is the rate at which birds land on the
pier (inflow rate) minus the rate at which birds leave the pier (outflow
rate):

dN
= (Inflow rate) − (Outflow rate).
dt
We assumed that:

(H1) Fluctuations in numbers of gulls on the pier occurred in

response to an environmental variable E(t).

(H2) Numbers of gulls on the pier during daylight hours could

be described using a two-compartment model: one compartment
 Seabird Behavior: A Case Study ■ 219

consisting of the loafing area and the other compartment

consisting of all other locations.

(H3) Gulls landed on the pier at a per capita rate proportional

to E(t), and they left the pier at a per capita rate inversely
proportional to E(t).

(H4) The total number of gulls in the two compartments depended

on the time of year and was proportional to the weekly maximal
number of gulls landing on the pier.

We let N (t) be the number of birds on the pier at time t, where t

is the day of the year plus the decimal fraction of the time within that
day. We took βKp (t) to be the total number of gulls in the system, both
on the pier and everywhere else, at time t, where Kp (t) estimates the
maximum number of gulls on the pier anytime during the year. With
these definitions, the expression βKp (t)−N (t) represents the number of
all gulls in the system that are not on the pier at time t. The inflow rate
is the per capita inflow rate multiplied by the number of birds in the
system at other locations, that is, αE(t) (βKp (t)−N ). The outflow rate
is the per capita outflow rate multiplied by the number of birds on the
pier, that is, N/ (αE (t)). The constants of proportionality α, β > 0 are
the parameters that we needed to estimate from the historical data. The
maximum number of gulls in the adjacent colony was approximately
60 times the maximum value of Kp (t); consequently, we selected 80 as
a generous upper limit for β. The model is therefore
dN N
= αE(t) (βKp (t) − N ) − , (10.2)
dt αE (t)

where 0 < α < ∞ and 1 ≤ β ≤ 80.

We estimated Kp (t) from pier count data taken throughout January
1 to March 21 and May 23 to December 31 during 1997, 1998, 1999, and
2001. Using these occupancy data, we estimated the seasonal maximum
number of gulls for the fluctuating pier counts by fitting a modified log-
normal curve to the means of the maximal pier counts for every week
of the year:
" #
2
[ln (40.29 − t/7) − 2.504]
Kp (t) = 76.36 exp .
−0.7225
220 ■ Mathematical Modeling in Biology

After day 275 and before day 65 of the following year, the mean
maximum pier counts were zeros or ones; consequently, we set Kp (t) =
0 for those intervals. See Figure 10.1.

10.5 ALTERNATIVE MODELS

We tested three alternative hypotheses for (H1) on the historical data:
(H1a) E(t) = T (t)

(H1b) E(t) = 1/S(t)

(H1c) E(t) = T (t)/S(t)

The three alternative hypotheses substituted into equation (10.2)
produce three alternative models: the “tidal model,” the “solar model,”
and the “tidal-solar model.”

10.6 MODEL PARAMETERIZATION

Using the method of conditioned least squares (CLS), we estimated
parameters for each of the three alternative models. For each hour,
a one-step prediction of the next observation was conditioned on the
previous observation. Specifically, for a set of n + 1 successive hourly

Figure 10.1: Maximal historical counts. The height of each bar is the
mean historical maximal count for the pier for that week, averaged over
the years 1997–1999 and 2001. The curve is the fitted function Kp (t).
(Originally published in The Auk 121:382. Used with permission of
Oxford University Press.)
 Seabird Behavior: A Case Study ■ 221

observations {x0 , x1 , . . . , xn }, the model was integrated numerically

to produce n hourly one-step predictions {y1 , y2 , . . . , yn }, where yi+1
denotes the model prediction at time i + 1, given the observation
xi at time i as the initial condition for the model. The differences
xi − yi between the observed and predicted values at time i
yielded the conditioned one-step residual errors. The residual sum of
squares
Xn
2
RSS(θ) = (xi − yi )
i=1

was minimized as a function of the vector θ of model parameters. The

vector of CLS parameter estimates for the model was the minimizer
∧
θ. We produced the one-step predictions using the MATLAB ode45
integrator, and we minimized RSS(θ) with the Nelder-Mead algorithm
fminsearch in MATLAB.
We used the generalized R2 value given by

RSS(θb)
R2 = 1 − P
n
2
(xi − x)
i=1

as a measure of goodness of fit, where x denotes the sample mean of the

observations {x1 , x2 , ..., xn }. The value R2 estimates the proportion of
the observed variability explained by the model, thereby providing a
measure of model prediction accuracy.
Computational note: In order to numerically integrate equation
(10.2), which depends on the environmental functions T and/or S, one
must spline hourly tides and solar elevations in order to produce smooth
curves that give between-hour values for T and S. See Exercise 3.
CLS parameter estimates and R2 values for the three alternative
models as fitted to the estimation data are shown in Table 10.1. You
will reproduce these values in Exercise 7.

10.7 MODEL SELECTION

In this study, it was not necessary to use the AIC to select the best
model, because all three alternative models had the same number
222 ■ Mathematical Modeling in Biology

TABLE 10.1 CLS Parameter Estimates and R2 Values

for the Three Alternative Models as Fitted to the
Estimation Data

Model α
b β
b b2
σ R2
H1a 0.31 2.0 244 0.38
H1b 0.10 80 225 0.43
H1c 0.35 2.8 166 0.58
Note that for (H1b), the local minimizer for RSS occurs
on the boundary of allowed parameter space with β = 80.

of parameters. In this situation, the model with the lowest AIC

corresponds to the model with the highest R2 (Exercise 4).
Based on the R2 values in Table 10.1, we eliminated the tidal
and solar models in favor of the tidal-solar model. Tidal-solar model
simulations are displayed with the estimation data in Figure 10.2.

10.8 MODEL VALIDATION

Using the reserved historical validation data in Data Set 10.1, we
independently evaluated the performance of the best model without
re-estimating its parameters (Exercise 8). The results are in Table 10.2.
The R2 = 0.60 for the validation data is comparable to R2 = 0.58
for the estimation data. This similarity supported confidence in the
model validation outcome. Figure 10.2 displays the model predictions
for the validation data (without refitting).

10.9 TEST OF A PRIORI PREDICTIONS

The analysis described above occurred during the autumn of 2001. We
decided to use our parameterized and validated model to see if we
could predict what would happen in the gull colony at a future date.

TABLE 10.2 Model (H1c) Validation

Results
Estimation Data Validation Data
2
R 0.58 0.60
 Seabird Behavior: A Case Study ■ 223

Figure 10.2: Model prediction (lower curve), hourly historical obser-

vations (circles), and tide height (upper curve) for the estimation
data and the reserved validation data. Each panel represents one day,
with a horizontal scale of 5–20 hours PST, which is 6–21 hours PDT.
A typical tidal curve for Protection Island is shown at the bottom.
Tidal nodes are indicated with arrows. Data from days occurring
during the same quarter of the tidal sequence are stacked vertically.
(Originally published in The Auk 121:383. Used with permission of
Oxford University Press.)

We chose the upcoming breeding season: May and June of 2002. In

particular, we used the tidal-solar model with parameters in Table 10.1
to generate hourly predictions for numbers of gulls loafing on the pier
during daylight hours for May and early June 2002 (Figures 10.3 and
10.4). The model predictions showed periodicities at three temporal
scales: high-frequency daily oscillations (Figure 10.4c), the expected
224 ■ Mathematical Modeling in Biology

Figure 10.3: Model (H1c) prediction (lower curve), hourly data from
2002 (circles), and tide height (upper curve). Each daily panel is
identified with the day of the year and has a horizontal scale of 5–
20 hours PST, which is 6–21 hours PDT. Tide height is graphed on a
vertical scale of −1 to 3 m. (Originally published in The Auk 121:385.
Used with permission of Oxford University Press.)

low-frequency seasonal fluctuation (Figure 10.4a), and an unexpected

medium-frequency oscillation (Figure 10.4c) that coincided with a
biweekly tidal pattern (Figure 10.4e, arrows). The form and minimal
values of the predicted daily fluctuations depended in an unexpected
way on the time within the tidal cycle relative to the solar cycle
(Figures 10.3 and 10.4c). On some days, for example, model predictions
oscillated out of phase with the tidal cycle (Figure 10.3, days 142
and 155), an unexpected result in light of previous studies in the
literature.
We began collecting data to test the model predictions at 0900 on
May 9, 2002. Thereafter, data collection continued at hourly intervals
from 0600 to 2100 PDT (dawn to dusk) until 2100 on June 6, 2002. Our
counts of gulls on the pier and associated structures were made from
a 30-meter-high bluff located approximately 100 m from the pier. If
there was a human disturbance on the pier during the count or within
30 minutes before the count, that count was eliminated.
The 2002 data, given in Data Set 10.2, contained all three
predicted periodicities (Figure 10.4b and d). Model predictions closely
approximated the count data (Figure 10.3) both qualitatively and
quantitatively, with a goodness of fit of R2 = 0.66. During the
 Seabird Behavior: A Case Study ■ 225

(a) (b)

(c)

(d)

(e)

Figure 10.4: Model (H1c) predicted oscillations on three different time

scales. (a and c) Model predictions for the spring of 2002, shown
with the seasonal pier envelope (dotted curve). Oscillations are present
on daily, biweekly, and yearly scales. (b and d) Data observations
corresponding to the predictions in (a and c). (e) Tidal oscillation for
the data collection time period in 2002. The tidal nodes are indicated
with arrows. (Originally published in The Auk 121:386. Used with
permission of Oxford University Press.)

first several days of each internodal period, daily minimal counts

lagged morning low-low tides (e.g., Figure 10.3, day 130), in mid-
period shifted to a point below midday low-low tides (e.g., Figure 10.3
day 137), later in the period preceded afternoon low-low tides (e.g.,
Figure 10.3, day 140), and at the node coincided with high tide
(e.g., Figure 10.3, day 142). After the node, daily maximal counts
increased and then decreased toward the next node in a reflection of
226 ■ Mathematical Modeling in Biology

the biweekly oscillation predicted by the model (Figure 10.4b and d).
Finally, from the first to the second biweekly oscillation period, the
average daily counts increased, which reflected the predicted increase
in counts during the spring (Figure 10.4b and d).

10.10 STEADY-STATE MODEL

The prediction of a differential equation model typically is determined
by integration over the past. But the gull-pier system recovers quickly
(empirically, in less than one hour) following a disturbance, and it
can be demonstrated that the state of the system from hour to hour
depends more on the current environmental conditions at time t than
on the past history of the system (Henson, Hayward, and Damania
2006). Said another way, the parameterized tidal-solar model exhibits
rapid transient dynamics (rapid in comparison with its steady-state
dynamics).
For such systems with fast transient recovery times, it is possible to
use a technique called multiple time scale analysis, which produces two
algebraic models: one for the transient dynamic and one for the steady-
state dynamic. A classic reference for multiple time scale analysis is the
book by Lin and Segel (1988). Here is an idea of how it works. We begin
with the differential equation
dN N (t)
= αE(t) (βKp (t) − N (t)) − ,
dt αE (t)
and we consider the fast time scale τ = t/ε, where ε > 0 is a small
number. Then using chain rule, we obtain
dN dt dN N (τ )
= = αE(τ ) (βKp (τ ) − N (τ )) − ,
dt dτ dτ αE (τ )
and so
dN N (τ )
ε = αE(τ ) (βKp (τ ) − N (τ )) − .
dt αE (τ )
For small enough ε, we have
N (τ )
0 ≈ αE(τ ) (βKp (τ ) − N (τ )) − . (10.3)
αE (τ )
Using E(τ ) = T (τ )/S(τ ), solving for N (τ ), and renaming τ as t, we
obtain the fast time scale (steady state) solution N0 (t), which is the
algebraic model (Exercise 5)
 Seabird Behavior: A Case Study ■ 227

βKp (t)
N0 (t) =  2 . (10.4)
1 S(t)
1+ α2 T (t)

The ability to make model predictions of the steady-state dynamics

of a system using this approach can be useful in some applications,
including resource management. Although the algebraic equation
(10.4) is simpler to use than the differential equation, it is unable
to account for transient dynamics following disturbances and cannot
make one-step predictions for model fitting. In Exercise 6, you will
compute the R2 for the steady-state algebraic model (10.4) on the
2002 data (Data Set 10.2), using the tidal-solar model parameters from
Table 10.1.

10.11 DISCUSSION
10.11.1 Importance of Scale
So, is animal behavior largely deterministic or largely stochastic? The
answer is: It depends on the scale.
Fluctuation in animal numbers over the short term indicates a
variety of competing functional needs. For example, during a single
day an individual bird may loaf, feed, and defend its territory against
intruders. To carry out all these behaviors, the bird must move
from habitat to habitat. Physiological limitations and environmental
variables are important in opening and closing opportunities for these
behaviors. It is not too surprising, then, that at the group level, some
behaviors are fairly deterministic, as we have seen in this chapter.
Nevertheless, animals such as gulls display high levels of individual
variation in behavior. An individual gull decides to move from one
habitat to another within a unique set of historical contingencies. While
it is true that a decision to change habitats may be influenced by what
other gulls do, the decision may also made without this influence.
We saw no evidence of coaction or social facilitation in arrivals and
departures from the pier, except in cases when the pier was disturbed
by human activity. Instead, gulls landed on the pier and left the
pier individually throughout the day, suggesting that the decision by
individual gulls to loaf may be more or less independent of the behavior
of the other gulls. The results in this chapter therefore suggest that
228 ■ Mathematical Modeling in Biology

deterministic forces at some group scales can be more important than

stochastic individual variability in a system.
The 1989 Robert H. MacArthur Award Lecture, presented by
Simon Levin in Toronto, Canada, and appearing in print in the journal
Ecology as the paper “The Problem of Pattern and Scale in Ecology,” is
a classic discussion of scale (Levin 1992). It should be read by everyone
who is interested in the dynamics of ecological systems.

10.11.2 Resource Management

Dynamic models of ecological systems are needed for proper un-
derstanding and management of natural resources. The successful
prediction of the dynamics of organisms in their natural environments
can lead to an understanding of food production, conservation, the
spread of disease, and resource management. Mathematical models
can generate testable hypotheses about natural systems. But this
can happen only if data are rigorously connected to models through
parsimonious modeling assumptions and model parameterization
using appropriate statistical methodologies, model selection from
a set of alternative models, and model validation. Data sampling
must be sufficiently dense to allow for model parameterization and
validation. Field sampling must occur at intervals smaller than
relevant environmental periodicities. It is important that models, once
parameterized, be validated on independent data sets to avoid simple
curve fitting, which has little explanatory power. Finally, the most
useful mathematical models are those that make successful a priori
predictions of future system dynamics, especially if the predictions are
unexpected.
The Seabird Ecology Team has successfully applied the methodol-
ogy described in this chapter to a number of other systems. See the
annotated bibliography for this chapter.

10.12 EXERCISES
1. Obtain from NOAA the Port Townsend hourly tidal observations
in meters for hours 0600–2100 PDT on July 2, 1997. Correct these
values for Protection Island with a factor of 0.93 as explained in
this chapter. Do your values match those in Data Set 10.1?
 Seabird Behavior: A Case Study ■ 229

2. Using all of the tidal observations and solar elevations in Data

Set 10.1, compute hourly normalized values of T (t) and S(t) with
equation (10.1). Do your values match those in Data Set 10.1?

3. Consider the vector of hours t = (0, 1, 2, 3, 4, 5) and a correspond-

ing vector of function values T (t) = (1.0, 0.54, −0.42, −0.99,
−0.65, 0.28). Use a cubic spline interpolator to estimate T (2.1)
and T (3.8). Hint: If you are using MATLAB, you can use the
cubic spline interpolator spline(t, T, τ ) in which t is the vector
of times at which you already know the function value, T is the
vector of function values at those times, and τ is the time (or
vector of times) at which you want to estimate the interpolated
value of the function.

4. Prove: If each model in a set of alternative one-dimensional

models has the same number of parameters, and if the residuals
are independent and Gaussian with mean zero and constant
variance σ 2 , then choosing the model with the smallest AIC is
equivalent to choosing the one with the largest R2 . Hint: See
Exercise 18 in Chapter 2.

5. In equation (10.3), replace ≈ with = and solve for N to obtain

equation (10.4).

6. How well does the steady-state algebraic model (10.4) with α =

0.35 and β = 2.8 (H1c; Table 10.1) fit the data in Data Set 10.2?
Compute the R2 . Show the fit visually with appropriate graphs
of daily data together with predictions. Attach your programs,
input files, and output.

7. Write the programs to estimate parameters for model (10.2) on

the estimation data. Reproduce the parameter estimates and
R2 values in Table 10.1. Attach your programs, input files,
and output. Hint: Use (0.1, 10)⊤ as the initial guess for the
parameter vector. Print the parameters to the screen as they
iterate. Answers may differ slightly depending on the software
you use.

8. Reproduce the R2 values in Table 10.2. Attach your programs,

input files, and output.
230 ■ Mathematical Modeling in Biology

9. Two-dimensional ODE parameterization project: This problem

concerns an eagle-gull predator-prey study (Henson et al. 2019).
In that paper, the Seabird Ecology Team fitted a Lotka-Volterra-
type predator-prey model to gull nest count data and Washington
State eagle territory data 1980–2016 with R2 = 0.82, suggesting
that gull dynamics were due largely to eagle recovery dynamics.
The paper is open access and contains the observed data.

a. Read the paper (Henson et al. 2019).

b. Reproduce the equilibrium and stability analysis given in
Appendix A of Henson et al. (2019).
c. Following the methods in that paper, reproduce the
parameter estimates in Table 2 of the paper and the R2
values.

BIBLIOGRAPHY
Cowles, J. D., Henson, S. M., Hayward, J. L., and Chacko, M. W. 2013. A
method for predicting harbor seal (Phoca vitulina) haulout and monitoring
long-term population trends without telemetry. Natural Resource Modeling
26:605–627. DOI: 10.1111/nrm.12015. [Applies the differential equation
approach and time scale analysis methods of this chapter to harbor seal
haul-out behavior. Both theoretical and empirical: models connected to
data.]

Damania, S. P., Phillips, K. W., Henson, S. M., and Hayward, J. L.

2005. Habitat patch occupancy dynamics of glaucous-winged gulls (Larus
glaucescens) II: A continuous-time model. Natural Resource Modeling
18:469–499. [Applies the differential equation approach and time scale
analysis methods of this chapter to the movement of gulls within a
system of three habitat patches dedicated to loafing. Both theoretical and
empirical: models connected to data.]

Hayward, J. L., Henson, S. M., Logan, C. J., Parris, C. R., Meyer, M. W,

and Dennis, B. 2005. Predicting numbers of hauled-out harbour seals:
a mathematical model. Journal of Applied Ecology 42:108–117. [Applies
the differential equation approach and time scale analysis methods of this
chapter to harbor seal haul-out behavior. Both theoretical and empirical:
models connected to data.]

Hayward, J. L., Henson, S. M., Tkachuck, R., Tkachuck, C., Payne, B.

G., and Boothby, C. K. 2009. Predicting gull/human conflicts with
 Seabird Behavior: A Case Study ■ 231

mathematical models: a tool for management. Natural Resource Modeling

22:544–563. [Empirical test of this chapter’s loafing model for glaucous-
winged gulls loafing at different locations during various stages of the
breeding season in different years, and for herring (Larus argentatus) and
great black-backed gulls (L. marinus) loafing on roof tops on Appledore
Island, Maine, the USA. Both theoretical and empirical: models connected
to data.]

Henson, S. M., Dennis, B., Hayward, J. L., Cushing, J. M., and Galusha,
J. G. 2007. Predicting the dynamics of animal behaviour in field popula-
tions. Animal Behaviour 74:103–110. DOI: 10.1016/j.anbehav.2006.11.015.
[Presents a general mathematical framework for modeling animal
behaviors and habitat patch occupancies with compartmental differ-
ential equations. Both theoretical and empirical: models connected
to data.]

Henson, S. M., Desharnais, R. A., Funasaki, E. T., Galusha, J. G., Watson,

J. W., and Hayward, J. L. 2019. Predator-prey dynamics of bald eagles
and glaucous-winged gulls at Protection Island, Washington, USA. Ecology
and Evolution 9:3850–3867. DOI:10.1002/ece3.5011. [Lotka-Volterra-type
predator-prey model fitted to gull nest count data and Washington State
eagle territory data 1980–2016 with R2 = 0.82, suggesting that gull
dynamics were due largely to eagle recovery dynamics. Both theoretical
and empirical: models connected to data.]

Henson, S. M., Galusha, J. G., Hayward, J. L., and Cushing, J. M. 2007.

Modeling territory attendance and preening behavior in a seabird colony
as functions of environmental conditions. Journal of Biological Dynam-
ics 1:95–107. DOI: 10.1080/17513750601032679 [Applies compartmental
modeling methods of this chapter to behaviors of territory attendance
and preening in a gull colony, where each behavior is considered a
compartment and the flow rates between compartments are functions of
environmental variables. Both theoretical and empirical: models connected
to data.]

Henson, S. M. and Hayward, J. L. 2010. The mathematics of animal behavior:

an interdisciplinary dialogue. Notices of the American Mathematical
Society 57:1248–1258. [Perspective piece on the integration of biology
and mathematics and the Seabird Team’s approach to modeling animal
behavior.]

Henson, S. M., Hayward, J. L., Burden, C. M., Logan, C. J., and

Galusha, J. G. 2004. Predicting dynamics of aggregate loafing behavior
232 ■ Mathematical Modeling in Biology

in gulls at a Washington colony. Auk 121:380–390. DOI: 10.1642/0004-

8038(2004)121[0380:PDOALB]2.0.CO;2. [The paper on which this chapter
is based. Both theoretical and empirical: models connected to data.]

Henson, S. M., Hayward, J. L., and Damania, S. P. 2006. Identifying

environmental determinants of diurnal distribution in marine birds
and mammals. Bulletin of Mathematical Biology 68:467–482. DOI:
10.1007/s11538-005-9009-0. [Describes the time scale technique mentioned
in this chapter. Both theoretical and empirical: models connected to data.]

Henson, S. M., Weldon, L. M., Hayward, J. L., Greene, D. J., Megna, L. C.,
and Serem, M. C. 2012. Coping behaviour as an adaptation to stress: Post-
disturbance preening in colonial seabirds. Journal of Biological Dynam-
ics 6:17–37. DOI: 10.1080/17513758.2011.605913. [Differential equation
approach incorporating logistic regression and Darwinian dynamics to
investigate theoretically how a behavior with crucial physiological function
might evolve into a comfort behavior. Both theoretical and empirical:
models connected to data.]

Levin, S. A. 1992. The problem of pattern and scale in ecology. Ecology

73:1943–1967. DOI: 10.2307/1941447. [Classic paper on scale in ecology,
presented as the Robert H. MacArthur Award Lecture in 1989 in Toronto,
Canada. Every student of ecology and mathematical biology should read
this paper.]

Lin, C. C. and Segel, L. A. 1988. Mathematics Applied to Deterministic

Problems in the Natural Sciences. SIAM, Philadelphia. [Classic text
in applied mathematics and mathematical biology. Belongs on every
mathematical biologist’s bookshelf.]

Moore, A. L., Damania, S. P., Henson, S. M., and Hayward, J. L. 2008.

Modeling the daily activities of breeding colonial seabirds: Dynamic occu-
pancy patterns in multiple habitat patches. Mathematical Biosciences and
Engineering 5:831–842. DOI: 10.3934/mbe.2008.5.831. [Applies methods
of this chapter to multiple habitat patches. Both theoretical and empirical:
models connected to data.]

National Oceanic and Atmospheric Administration. Solar Position Calendar.

https://gml.noaa.gov/grad/solcalc/azel.html.

National Oceanic and Atmospheric Administration. Tides and Currents.

https://tidesandcurrents.noaa.gov/waterlevels.html?id=9444900.

Payne, B. G., Henson, S. M., Hayward, J. L., Megna, L. C., and Velastegui
Chavez, S. R. 2015. Environmental constraints on haul-out and foraging
 Seabird Behavior: A Case Study ■ 233

dynamics in Galpágos marine iguanas. Journal of Coupled Systems and

Multiscale Dynamics 3:208–218. DOI: 10.1166/jcsmd.2015.1077. [Applies
the methods in this chapter to haulout in marine iguanas. Both theoretical
and empirical: models connected to data.]
 IV
Regression Models

235
 CHAPTER 11

Introduction to
Regression

11.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

This chapter briefly introduces basic ideas from regression theory. Re-
gression models are statistical models, but they are also mathematical
models that we parameterize with data, hence their place in this book.
We omit many useful tools in regression theory, but we hope this
brief, basic treatment will give the student a conceptual foundation
for further investigation.

11.2 LINEAR REGRESSION

We begin with linear regression, which students may encounter in high
school mathematics courses. The idea is to put a “best-fit line” through
data points of the form (x, y) that have an approximately linear trend
when y is graphed against x. This is called “regressing y against x.”
For example, we might regress crop yield against rainfall. Here, crop
yield is the outcome, which is the dependent variable. The dependent
variable is also called the response variable in regression. Rainfall is the
factor , which is the independent variable. The independent variable is
also called a predictor variable or explanatory variable in regression.
Note the order of words in the terminology “regress y against x,” and
“the response of y to x.” If we are thinking in terms of graphs, this is
equivalent to the way mathematicians say “graph y against x.” Always

DOI: 10.1201/9781003265382-15 237

238 ■ Mathematical Modeling in Biology

say “vertical against horizontal.” Many people mistakenly reverse the

order.

11.2.1 Simple Linear Regression (Single Factor)

Suppose we wish to know how a dependent variable y depends on a
factor x, under the assumption that x and y have a linear relationship.
The mathematical statement of this assumption is
y = β0 + β1 x,
where the coefficients β0 and β1 are the model parameters. Here, β1 is
the slope and β0 is the intercept.
The intercept β0 quantifies the value of y when x = 0. Let us
consider the interpretation of the slope β1 . Suppose factor x changes
by c units. Then the resulting change in y is
∆y = ynew − yold
= (β0 + β1 (x + c)) − (β0 + β1 x)
= β1 c.
That is, if x changes by c units, then y changes by β1 c units. In
particular, if x changes by one unit, then y changes by β1 units. Thus,
β1 is the rate of change of y with respect to the factor x. This is, of
course, consistent with our understanding from calculus that slope is
equivalent to the derivative, which is the rate of change. Note that if
β1 is positive, then y increases as x increases, whereas if β1 is negative,
then y decreases as x increases.

11.2.2 Multiple Linear Regression (Multiple Factors)

Suppose now that y depends linearly on n factors x1 , x2 , . . . , xn . The
multiple regression (linear regression with multiple factors) model is
y = β0 + β1 x1 + β2 x2 + . . . + βn xn (11.1)
Xn
= β0 + βj xj .
j=1

If factor xi changes by c units, with all other factors held constant,

then it is straightforward to check that the resulting change in y is
(Exercise 1)
 Introduction to Regression ■ 239

∆y = βi c.
Thus, coefficient βi is the rate of change of y with respect to factor xi
when all other factors are held constant. The sign of the coefficient
βi determines whether y is positively or negatively correlated with
factor xi .

11.2.3 Stochastic Model and Parameter Estimation

A stochastic version of model (11.1) in which Gaussian noise is additive
is
Xn
Y = β0 + βj xj + σε,
j=1

where σ > 0 is a parameter representing the standard deviation of

the noise and ε is a standard normal random variable (mean zero and
standard deviation one). The residuals are
 
Xn
res = y − β0 + βj xj  ,
j=1

where y is an observed value (realization of the random variable Y )

associated with factors having values (x1 , x2 , . . . , xn ). The residuals
are realizations of the random variable σε. That is, this model assumes
that the residuals are normally distributed about zero with standard
deviation σ.
The likelihood that an observed data point is a realization of the
stochastic model, that is, the likelihood that the residual comes from
the hypothesized distribution of noise, is
1 1 res 2
√ e− 2 ( σ ) .
σ 2π
Assuming the observations are independent, the likelihood L that all of
the residuals generated from the data set come from the hypothesized
distribution is
Y 1 1 res 2
L= √ e− 2 ( σ ) .
data
σ 2π
The ML parameter estimates are those that maximize the likelihood L.
240 ■ Mathematical Modeling in Biology

Parameter estimation routines for regression equations are available

in the libraries of scientific computing programs. In MATLAB, for
example, one can use the function glmfit.

11.2.4 Confidence Intervals for Regression Coefficients

Suppose a regression slope coefficient is estimated to be βb, with
standard error SE. The standard error of a parameter estimate is
the standard deviation of its sampling distribution. That is, if you
resampled the data many times and considered the distribution of the
estimated values βb, the standard deviation of that distribution would
be SE. The endpoints of the 95% confidence interval for β are
βb ± 1.96 × SE.

11.3 LOGISTIC REGRESSION

Logistic regression (Hosmer and Lemeshow 2000) is a technique
commonly used to quantify the effect of factors on a binary variable
(an event occurs, or it does not occur). Let P be the probability of
the event. Then P takes on values between zero and one, making
linear regression inappropriate for predicting P because regression lines
predict values from negative infinity to positive infinity. Thus, the
dependent variable P must be transformed so that its range is R.
The odds of the event are defined to be
P
.
1−P
Note that, whereas P takes on values between zero and one, the
associated odds take on values between zero and infinity. The log odds
of the occurrence of the event are given by
 
P
ln ,
1−P
which takes on values between negative infinity and positive infinity.
In logistic regression, the log odds are regressed on a vector x =
(x1 , x2 , . . . , xn ) of factors:
  n
P X
ln = β0 + βj x j , (11.2)
1−P j=1
 Introduction to Regression ■ 241

where P is the probability of the event. The intercept β0 calibrates

the baseline log odds of the event when all factors are zero, and the
regression coefficients β = (β1 , . . . , βn ) quantify the response of the log
odds to changes in the factors.

11.3.1 Odds Ratios (ORs)

The coefficients β have a convenient interpretation. If factor xi
increases by c units while all other factors remain constant, then the
log odds changes by βi c units:

   
P2 P1
ln − ln
1 − P2 1 − P1
 
 
n
X n
X
 
= 
β0 + βi (xi + c) +  − β0 +
βj xj  βj xj 

j=1 j=1
̸
j=i
= βi c,

where P1 and P2 are the probabilities of the event before and after the
change, respectively. Thus, by the laws of logarithms, the odds ratio,
denoted OR, is

P2 / (1 − P2 )
OR = = eβi c .
P1 / (1 − P1 )
This means that, given an increase in factor xi by c > 0 units, with all
other factors held constant, the odds of the event are eβi c times what
they were before.
If βi > 0, then the odds ratio is greater than one (OR > 1), meaning
the odds of the event have increased. If βi < 0, the odds ratio is less
than one (OR < 1), and the odds of the event have decreased. For
example, if OR = 1.25, the interpretation is that the odds of the event
increase 25% with a c-unit increase in xi . If OR = 0.85, the odds
decrease 15% with a c-unit increase in xi . The researcher chooses a
convenient value of the reference c so that the results are intuitively
clear. Note that these statements of association do not necessarily
imply causation.
242 ■ Mathematical Modeling in Biology

11.3.2 OR Confidence Intervals

Odds ratios typically are presented with confidence intervals. The 95%
confidence interval for OR = eβc is
 
ec(β−1.96×SE) , ec(β+1.96×SE) ,

where SE is the standard error of the regression coefficient β. If the

confidence interval includes the value of 1, this indicates that we are
not confident whether OR < 1 or OR > 1, and hence, we cannot claim
that the odds of the event significantly decrease or increase with an
increase in the associated factor.
Many other useful topics, such as how to rank factors according
to their importance, are found in the definitive books by Burnham
and Anderson (Burnham and Anderson 2002) and by Hosmer and
Lemeshow (Hosmer and Lemeshow 2000).

11.4 GENERALIZED LINEAR MODELS (GLMs)

As we saw above, logistic regression is a form of linear regression in
which a log-odds transformation of the dependent variable is regressed
against factors. In general, one can regress a transformed dependent
variable g (y) against the factors x = (x1 , x2 , . . . , xn ):
n
X
g (y) = β0 + βj xj . (11.3)
j=1

The transformation g is called the link function, and the model is called
a generalized linear model (GLM). Common examples of GLMs include
logistic regression for binary data (0 or 1 outcomes), Poisson regression
for count data, gamma regression for exponential response data, and
linear regression for linear response data. The link functions for these
types of regression are shown in Table 11.1.
For more information on GLMs, see the definitive book by
McCullagh and Nelder (1989).

11.5 INTERACTION TERMS

The right-hand side of equation (11.3) can also include interaction
terms of the form βij xi xj . These terms are necessary if the effect of
 Introduction to Regression ■ 243

TABLE 11.1 Link Functions for Regression

Regression Link Function Type of Data (Dependent Variable)
Type
Gamma g(y) = −y −1 Exponential response data, (0, ∞)
Linear g (y) = y Linear response data (−∞, ∞)
Logistic g(y) = ln (y/ (1 − y)) Binary data (yes/no), {0, 1}
Poisson g(y) = ln y Count data, {0, 1, 2, . . .}

one factor depends on the value of another factor. In such a case,

the meaning of the coefficients βi and the formulas for the confidence
intervals change. For example, consider the logistic regression model

ln (odds) = β0 + β1 x1 + β2 x2 + β12 x1 x2 .

If factor x1 is increased by c units with all other factors held constant,

then the change in log odds is

∆ ln (odds) = (β0 + β1 (x1 + c) + β2 x2 + β12 (x1 + c) x2 )

− (β0 + β1 x1 + β2 x2 + β12 x1 x2 )
= β1 c + β12 cx2 .

Thus, the odds ratio

OR = ec(β1 +β12 x2 )
depends on the value of factor x2 .

11.6 EXERCISES
1. Prove that if factor xi increases by c units in equation (11.1), and
all other factors are held constant, then the resulting change in
y is βi c.
2. Consider the Poisson regression model

ln y = β0 + β1 x1 + β2 x2 + β3 x3 .

If factor x1 increases by c units, and all other factors are held

constant, then how does y change?
3. Consider the GLM

g (y) = β0 + β1 x1 + β2 x2 ,
244 ■ Mathematical Modeling in Biology

where g is the link function for gamma, logistic, or Poisson

regression (Table 11.1). For each of these three types of models,
solve the regression equation for y.

4. Consider the following multiple regression model that predicts

VO2max from AGE in years, WEIGHT in kilograms, and
heart rate (HRT) in beats/minute in men (adapted from
https://statistics.laerd.com/stata-tutorials/multiple-regression-
using-stata.php):

VO2max = 101.04 − 0.17AGE − 0.39WEIGHT − 0.12HRT.

a. What predicted change in VO2max results from adding 2

extra kg of weight, with all other factors held constant?
b. What is the predicted change in VO2max if the heart rate
goes up by 3 beats/min, with all other factors held constant?

5. Consider the logistic regression model

ln (odds) = β0 + β1 x1 + β2 x2 + β3 x3 .

a. If the slope coefficient for factor x3 is β3 = −0.1 and you

increase x3 by c = 5 units, with all other factors held
constant, what is the odds ratio (OR)?
b. If x3 increases by five units, with all other factors held
constant, then the odds of the outcome decrease by what
percent?

6. Second-hand smoke increases the risk of lung cancer in

nonsmokers and is a serious health risk. Suppose the odds ratio
for lung cancer for a nonsmoker living with a smoker versus not
living with a smoker is about OR = 1.28, all other factors held
constant. Living with a smoker (vs. not living with a smoker)
increases a nonsmoker’s odds of developing lung cancer by what
percent?

7. Consider the outcome of passing a class (0 or 1) as a function of

the average number of study hours per week (STUDY) and the
average number of hours of sleep per night (SLEEP).
 Introduction to Regression ■ 245

a. Use the following fictional data set to parameterize the

logistic regression model

ln (odds of passing) = β0 + β1 x1 + β2 x2 .

Use the MATLAB library function glmfit(X,Y,‘binomial’) or

another GLM fitting function in the programming language
you are using.
PASS STUDY (x1 ) SLEEP (x2 )
0 6 4
1 5 8
1 3 4
0 1 5
1 6 5
0 0 4
1 6 6
1 4 7
1 5 8
0 1 7
0 1 9
1 7 7

b. Fill in the missing values in the results table using the given
reference value c.
Factor bi c SE OR 95% CI
STUDY 0.5 hours
SLEEP 0.3391 2 hours 0.4960 1.970 (0.2819, 13.77)

c. Interpret the results in words, in terms of the reference value

c, the odds ratio OR, and significance. For example, the
interpretation for SLEEP would be: “For each extra 2 hours
of sleep per night, the odds of passing the class increased by
97%. However, this trend was not significant given that the
OR confidence interval included numbers both above and
below one, due to the large standard error.”
246 ■ Mathematical Modeling in Biology

BIBLIOGRAPHY
Burnham, K. P. and Anderson, D. R. 2002. Model Selection and Multi-Model
Inference: A Practical Information-Theoretic Approach, 2nd ed. Springer-
Verlag, New York. [Comprehensive and user friendly text on methods of
model selection.]

Henson, S. M., Weldon, L. M., Hayward, J. L., Greene, D. J., Megna,

L. C., and Serem M. C. 2012. Coping behaviour as an adaptation to
stress: Post-disturbance preening in colonial seabirds. Journal of Biological
Dynamics 6:17–37. DOI: 10.1080/17513758.2011.605913. [Incorporates
regression models into differential equation models for animal behavior.]

Hosmer, D. W. and Lemeshow, S. 2000. Applied Logistic Regression, 2nd ed.

John Wiley & Sons, New York. [The classic logistic regression text.]

McCullagh, P. and Nelder, J. A. 1989. Generalized Linear Models, 2nd ed.

Chapman & Hall/CRC Press, Boca Raton, FL. [Definitive text on GLMs.]
 CHAPTER 12

Climate Change and

Seabird Cannibalism: A
Case Study

12.1 WHAT YOU SHOULD KNOW ABOUT THIS CHAPTER

This chapter is based on a large study conducted by the Seabird Ecol-
ogy Team. The team collaborated with mathematician Lynelle Weldon
to examine the relationship between a climate-driven environmental
variable (sea surface temperature) and egg cannibalism in gulls. Results
from this study were first published in The Condor: Ornithological
Applications (Hayward et al. 2014). The team produced several
additional related publications, which are listed in the annotated
bibliography for this chapter.
The methodology in this chapter employs basic logistic regression
(Chapter 11) and coding skills (Appendix B).
The Condor study (Hayward et al. 2014) was comprehensive and
used a number of techniques we will not explain here, except perhaps in
a cursory fashion, such as Akaike weights, model-averaged parameter
estimation (Burnham and Anderson 2002), design variables, measures
of overdispersion (Hosmer and Lemeshow 2000), and advances in
techniques for validating logistic regression models (Giancristofaro and
Salmaso 2003). In the exercises, you will use basic logistic regression
without these supporting techniques, and the parameters you obtain
will therefore differ somewhat from those obtained in The Condor
paper (Hayward et al. 2014). That paper, however, might serve as

DOI: 10.1201/9781003265382-16 247

248 ■ Mathematical Modeling in Biology

a useful resource for approaching a comprehensive logistic regression

analysis in the reader’s own research projects.

12.2 THE SCIENTIFIC PROBLEM

Over the course of many field seasons at a colony of several
thousand glaucous-winged gulls (Larus glaucescens) at Protection
Island National Wildlife Refuge, Washington, the USA, we noticed
large accumulations of broken eggshell littering a few of the nesting
territories. We watched the owners of these territories invade the
territories of fellow residents, grab an egg, fly the egg back to their own
territories, and eat the contents. Eggshell fragments from the stolen
eggs accumulated on the territories of these egg cannibals (Figure 12.1).
A cannibal is any animal that, like these gulls, kills and eats
members of its own species, and their victims can be at any stage of the
life cycle, including the egg stage. Members of at least 1,300 species are
known to engage in cannibalism (Polis 1981), although the real number
is certainly much higher. Cannibals are found among zooplankton,
arthropods, fish, reptiles, birds, mammals, and other animal groups.
Once thought to be an aberration, cannibalism is now considered a nor-
mal life history trait of many species (Elgar and Crespi 1992; Fox 1975).
Cannibalism exerts important influences on ecological and evo-
lutionary features of animals. It can, for example, influence energy
relations among members of ecological communities, reduce their
reproductive success, change the sizes of their populations, modify their
social behaviors, promote kin selection, and even result in complex

(a) (b)

Figure 12.1: (a) Male cannibal holding egg in bill. (b) Accumulation
of eggshell fragments on the territory of an egg cannibal. (Photos by
James Hayward.)
 Climate Change and Seabird Cannibalism: A Case Study ■ 249

nonlinear population dynamics such as chaos (Cushing, Henson and

Hayward 2015; Elgar and Crespi 1992; Polis 1981).
Crowding, odd behavior patterns by subdominant individuals,
psychological and physiological stress, and the availability of victims
can elicit cannibalism (Fox 1975). The size, age, sex, habitat, and
developmental stage of potential victims also may play roles. Moreover,
some individuals are more genetically predisposed than others to
engage in cannibalistic behavior. Most commonly, however, lack of food
or poor food quality leads to cannibalism (Dong and Polis 1992).
Ornithologists have long known that gulls cannibalize their
neighbors’ eggs, a behavior reported for at least 19 gull species (Polski
et al. 2021). The behavior, however, has not been well studied, and most
such reports have come from incidental observations during research
into other aspects of gulls’ lives. We wondered about the extent and
impact of this behavior on a breeding colony, so we decided to examine
this phenomenon at Protection Island. One of the papers that resulted
from this project provides the first comprehensive assessment of the
behavioral ecology of cannibalism by gulls (Polski et al. 2021).
Only males on the Protection Island colony steal eggs. A
cannibalistic male locates an egg to steal in one of two ways. He
may fly slowly over an undisturbed colony looking for a momentarily
unprotected nest, or he may wait for an eagle to fly over, disturbing
residents in large areas of the colony so that they fly up from their nests
and leave their eggs exposed. In either case, the cannibal capitalizes
on the opportunity, quickly lands by the exposed nest, and grabs an
egg. The cannibal then flies the stolen egg back to his territory, gently
holding the egg in his bill or swallowing it into his crop whole. Once
back on his territory, the cannibal drops the egg or regurgitates it
from his crop, pecks it open, and devours the slimy contents. If his
mate is present, she typically begs to be allowed to participate in the
feeding. Most of the time, the male shares his loot with his mate, but
occasionally, he pecks at her and refuses to share.
Eating two eggs per day just about fulfills the energy requirements
of an adult gull (Hayward et al. 2014), but some cannibals in our study
stole more than two eggs per day. In June 2014, for example, one male
stole 81 eggs in 30 days, and another took 75 eggs during the same
period (Polski et al. 2021). Unlike most colony residents, super egg
predators like these spent little time looking for food off the colony.
250 ■ Mathematical Modeling in Biology

We compared the contents of regurgitated food pellets of non-

cannibals and egg cannibals. (Gulls regurgitate pellets of undigestible
material just like owls.) Not surprisingly, pellets from cannibals were
significantly more likely to contain fragmented eggshell than pellets
from non-cannibals, but pellets from cannibals were also significantly
less likely to contain fish remains than those of non-cannibals. This
suggested that cannibals were substituting eggs for fish as their primary
protein source.
Only about 1 in a 100 of the territories on the Protection
Island gull colony were defended by egg cannibals, yet these few
cannibals stole 1 out of every 9 eggs produced in the entire colony
in both 2014 and 2015. Despite the fact they are few in number,
egg cannibals impact the colony in disproportionate ways, and several
birds at Protection Island exhibited this behavior over multiple
years.
How did the reproductive output of egg cannibals compare with
those of fellow colony residents? Egg cannibals produced significantly
fewer eggs than non-cannibals. Moreover, the proportion of eggs taken
from the nests of cannibals by other cannibals was significantly higher
than the proportion of eggs taken from the nests of non-cannibals. So,
egg cannibalism appears to be a feeding tactic used by less attentive
and less successful breeders.
Marine birds associated with the Pacific Ocean experience food
shortages every few years as a result of El Nino-Southern Oscillation
(ENSO) events. These events raise the sea surface temperature
(SST), lower the thermocline (the depth at which the temperature of
water suddenly becomes cooler), and weaken upwellings (which bring
nutrients up to higher levels in the water column). As a result of these
changes, plankton become less abundant toward the sea surface, and
this drives forage fish to lower levels in the water column. Gulls, too
buoyant to dive, have fewer fish to eat during these times and must
travel longer distances to find enough food for themselves and their
young. All this leads to lower reproductive success, which may lead to
population declines.
We wondered whether ENSO-related rises in SST impacted the
rate of egg cannibalism in a breeding colony of glaucous-winged
gulls.
 Climate Change and Seabird Cannibalism: A Case Study ■ 251

12.3 DATA

We collected data from late May to mid-July, 2006–2011, at the

glaucous-winged gull colony at Protection Island National Wildlife
Refuge, Washington, the USA. Protection Island (48◦ 07′ 40′′ N,
122◦ 55′ 3′′ W) is located in the southeast corner of the Strait of Juan de
Fuca. The gull colony contained more than 2,400 breeding pairs during
this study.
We selected five rectangular sample plots (Figure 12.2, Plots A–
E), with a combined area of 4,205 m2 , and with a combined total of
199–267 sample nests per year. Each plot was checked every day in
the late afternoon. When the first egg was laid in a nest, a numbered
stake was placed near the nest and the egg was labeled as the A-egg.
Thereafter, the nest was monitored each afternoon. Each subsequent
egg was labeled and recorded as well. The ultimate fate of each egg was
determined and recorded as cannibalized, eagle predated, addled (dead
during incubation), died during hatching, hatched, or other (punctured,
nest flooded, or rolled out of nest). We distinguished cannibalized eggs
from eagle-predated eggs by the fact that cannibals steal eggs and fly

Figure 12.2: Violet point on Protection Island, showing five sample

areas A–E in the gull colony.
252 ■ Mathematical Modeling in Biology

them back to their own territories, whereas eagles eat eggs where they
find them, leaving distinctive patterns of broken eggshell behind.
We measured the distance from the center of each sample nest to
the center of the nest of its nearest neighbor. We recorded the nest as
residing in one of four habitats: (i) short or sparse vegetation (SV),
(ii) tall dune grass (TG), (iii) beside a shrub or log (SL), or (iv) beach
(BC). In 2007–2011, we also determined the mass of each egg on the
day it was laid.
The average SST from September to May prior to each breeding
season was determined using data from the Port Townsend, Wash-
ington buoy (PTWW1), a floating National Oceanic and Atmospheric
Administration (NOAA) structure located 12 km east of Protection
Island. The interval from September to May was chosen for two
reasons: (i) Breeding success in gulls is determined in part on resource
availability before egg production, and (ii) the effects of changes in the
physical environment on seabird populations are not immediate (Smith
et al. 2017). Warmer-than-average SSTs preceded the 2007 and 2010
breeding seasons due to El Nino conditions.
The data for 2006–2011 are given in Data Set 12.1.

12.4 LOGISTIC REGRESSION ANALYSIS

The fates of 2,932 eggs monitored over the five breeding seasons from
2007 to 2011 were assessed by logistic regression in terms of whether
eggs were cannibalized or not (1 or 0). Cannibalism was considered
as a function of SST, egg mass (MASS), nearest neighbor distance
(NN), number of days before or after the mean laying date for the
season (DAYS), sample area (PLOT; five values, A–E), habitat type
(HAB; four values, SV, SL, TG, and BC), egg order as laid in the nest
(ORDER), and the total number of eggs laid in the clutch (CSIZE).
The log odds of cannibalism were regressed on these eight factors and
on 17 interaction terms between PLOT and HAB (three of the 20
possible combinations did not occur) with an intercept term.
A suite of alternative models was obtained from the global model
by taking all submodels, that is, all possible linear combinations of
the eight factors, with intercepts. For models with both PLOT and
HAB variables, the interaction terms were included. We determined
 Climate Change and Seabird Cannibalism: A Case Study ■ 253

parameter estimates and the Akaike information criterion (AIC) for

the global model and all submodels.
The best model, that is, the one with the smallest AIC, included all
of the factors except ORDER, NN, and MASS. Because the best model
did not include MASS, which was the only measurement not taken in
2006, in the exercises you will analyze the entire 6-year (2006–2011)
data set (Data Set 12.1).

12.5 MODEL VALIDATION

We then validated the selected model. Because regression models
measure trends but are not mechanistically based, the values of R2
(both fitted and validation) are typically quite low. Thus, for regression
modeling the goodness-of-fit measure R2 used throughout the rest of
this text is often uninformative. Indeed, model validation techniques
for logistic regression models are an ongoing subject of research.
In this study (Hayward et al. 2014), we used a validation technique
by Giancristofaro and Salmaso (Giancristofaro and Salmaso 2003),
in which the egg data were split into two random samples: one for
parameter estimation (75% of the data) and one for validation (25%
of the data). The selected model was fitted to the estimation data,
and the model performance was then measured on both samples. This
process was repeated 100 times. Details are found in Giancristofaro
and Salmaso (2003) and Hayward et al. (2014).

12.6 OUTCOMES
The results of the analysis described above are summarized in Table
12.1, which shows only those variables found to be significant. The
parameters used to compute the ORs in Table 12.1 were not the
fitted parameters for the best model, but rather were the “model-
averaged” parameter estimates. A model-averaged parameter estimate
is the average over all models (in the suite of alternative models) that
contain that parameter, weighted by its Akaike weight. We do not
discuss Akaike weights or model-averaged parameter estimates here;
details can be found in Burnham and Anderson (2002).
The most important thing to understand in this section is how
to interpret the results in Table 12.1. Make sure you can derive the
following interpretations from Table 12.1.
254 ■ Mathematical Modeling in Biology

TABLE 12.1 Cannibalism Odds Ratios (ORs) with 95% Confidence

Intervals (CI) Associated with a c Unit Increase in the Factor or
Relative to the Given Reference Variable
Factor c or Reference Value OR 95% CI
SST 0.1 deg 1.10 (1.06, 1.13)
DAYS 1 day 1.09 (1.06, 1.11)
CSIZE 1 egg
2 eggs 0.13 (0.07, 0.21)
>2 eggs 0.09 (0.05, 0.15)
PLOT B HAB SV
HAB SL 0.28 (0.23, 0.33)
HAB TG 0.06 (0.04, 0.09)
PLOT C HAB SV
HAB SL 0.54 (0.45, 0.63)
HAB BC 0.68 (0.62, 0.75)
PLOT D HAB SV
HAB SL 0.28 (0.26, 0.32)
HAB BC 0.40 (0.34, 0.48)
PLOT E HAB SV
HAB SL 0.39 (0.35, 0.42)
HAB BC 0.68 (0.54, 0.87)
Habitats are short or sparse vegetation (SV), beside shrub or log (SL), beach (BC),
and beside or in tall grass (TG). Only significant variables and interactions are
listed here. For the complete results, see Hayward et al. (2014).

1. The odds of cannibalism increased 10% for each 0.1◦ C rise in

SST, with all other factors held constant.

2. The odds of cannibalism for an egg increased by 9% for each day

it was laid away from the mean laying date, with all other factors
held constant.

3. An egg in a three-egg nest was 91% less likely to be cannibalized

than an egg from a one-egg nest, and an egg from a two-egg nest
was 87% less likely to be cannibalized than an egg from a one-egg
nest (all other factors held constant).

4. Compared to an egg in a nest in short or sparse vegetation (SV),

an egg in a nest beside a shrub or log (SL) was the least likely
to be cannibalized, except in Plot B, in which eggs in tall grass
 Climate Change and Seabird Cannibalism: A Case Study ■ 255

(TG) were even less likely to be cannibalized. In Plots C, D,

and E, eggs in nests on the beach (BC) also were less likely to
be cannibalized than those in short or sparse vegetation (SV),
but they were more likely to be cannibalized than those beside a
shrub or log (SL), all other factors held constant.

12.7 CLIMATE CHANGE, CANNIBALISM, AND REPRODUCTIVE

SYNCHRONY

The fitness of individuals and the dynamics of populations are

directly affected by cannibalism (Dong and Polis 1992). Cannibalism of
juveniles by adults may allow population survival when resource levels
are low, followed by a redirection of reproductive effort to times when
resources are more available (Elgar and Crespi 1992; Henson 1997).
Cannibalism is especially common among adult gulls when food
supplies are low, although ours was the first study to directly link egg
cannibalism with environmental conditions associated with low food
supply.
During our research, egg cannibalism became more common when
the SST increased. In particular, a 0.1◦ C rise in SST was associated
with a 10% increase in the odds that an egg was cannibalized. When
SST is high, as during El Nino events, plankton and fish, primary
food items for gulls, drop to lower, cooler levels in the water column.
Unlike many seabirds, gulls cannot dive, so food is more difficult to
obtain during these times; eggs, nutritious and conveniently abundant,
become a more frequent food source for colony residents.
The highest degrees of egg cannibalism during our study occurred
during the 2007 and 2010 breeding seasons, both of which were
preceded by higher than average SSTs in months prior to the breeding
seasons. Egg cannibalism during these times provided relatively low-
cost, locally available food for hungry breeding birds. Indeed, as already
noted, some gulls cannibalize more than two eggs per day during the
incubation period, probably most of the caloric intake by these birds
during this time (Polski et al. 2021).
Sea surface temperatures in the Strait of Juan de Fuca surrounding
Protection Island increased by approximately 1◦ C from 1950 to 1998,
and this warming trend likely will continue in the face of global
256 ■ Mathematical Modeling in Biology

warming. Our data suggest this warming trend may increase the rate
of egg cannibalism.
It seems, however, that gulls have evolved a way to reduce the
impact of cannibalism on their reproductive efforts. During the years of
low SST and low egg cannibalism, egg-laying occurs during a relatively
short period, which overwhelms would-be egg and chick predators from
outside the colony with a glut of food; the chance that a given egg
is predated by bald eagles is reduced (the “Fraser Darling effect”
(Darling 1938)). By contrast, during the years of high SST and high
egg cannibalism, the egg-laying period lengthens and females nesting
close together tend to synchronize their every-other-day egg laying until
they achieve the typical three-egg clutch; by laying their eggs on the
same day as their neighbors, the chance that a given egg is predated
by a cannibal neighbor is reduced (Henson et al. 2010). In other words,
gulls switch between two breeding tactics, depending on whether the
threats to their offspring are from outside (eagles) or inside (cannibals)
the colony (Weir et al. 2020).
Under the lead of one of our colleagues, Gordon Atkins, we found
that females tended to avoid copulation on egg-laying days, but
welcomed copulation on the intervening days. The very loud, pulsating
copulation call of the males functioned as the synchronizing signal
(Atkins, Hayward, and Henson, 2021; Atkins et al. 2017).
Our studies of cannibalism raise many unanswered questions.
For example, are some gulls genetically predisposed to cannibalize
their neighbors’ eggs, or is this primarily a learned behavior? Are
eggs with certain colors and pigmentation patterns more vulnerable
to cannibalism than others? Do egg cannibals exhibit other forms
of antisocial behavior outside the breeding season? Will increasing
SSTs associated with climate change increase the incidence of egg
cannibalism on gull colonies? Hopefully, these and other questions will
be answered by future research.
Of the Seabird Team papers in the annotated bibliography below,
some are theorem-proof mathematics papers and some are empirical
papers. In terms of topics, they address various aspects of cannibalism
and egg-laying synchrony in the context of climate change. The
annotations will help the reader know how each paper relates to this
chapter.
 Climate Change and Seabird Cannibalism: A Case Study ■ 257

12.8 EXERCISES
1. Fit model

log (odds) = b0 + b1 SST + b2 DAYS + b3 CSIZE

(no interaction terms) to Data Set 12.1 using logistic regression,

where odds is the odds an egg is cannibalized. In this exercise,
let CSIZE be the actual number of eggs in the clutch, which
is slightly different from the definition of CSIZE in Hayward et
al. (2014). Also compute odds ratios (ORs) and 95% confidence
intervals for the ORs. Use the glmfit(x,y,‘binomial’) function in
MATLAB or other software for logistic regression.

a. Fill in the missing entries in this results table, given the

reference c.
Factor b c SE OR 95% CI
SST 0.1 deg
DAYS 1 day
CSIZE −0.5671 1 egg 0.0686 0.5672 (0.4958, 0.6489)

b. Interpret the results in words, in terms of the reference

value c, the odds ratio OR, and significance. For example,
the interpretation for CSIZE could be: “For each extra 1
egg in the clutch, the odds that a given egg is cannibalized
decreases by 43%. This decreasing relationship is significant
because the entire 95% confidence interval for OR is less
than one.”
c. Give a short biological discussion of the results. For example,
for CSIZE you might say: “Eggs from a larger clutch are less
likely to be cannibalized. There are at least two possible
reasons for this: (i) with more eggs in a clutch, each egg
is less likely to be taken during a predation event, and (ii)
parental investment in a full clutch is higher, leading to more
careful guarding of the nest by parents.”
d. Compare your results to those in Table 12.1, which are taken
from Hayward et al. (2014) and are based on model-averaged
parameter estimates.
258 ■ Mathematical Modeling in Biology

2. Fit model
log (odds) = b0 + b1 SST
to Data Set 12.1 using logistic regression, where odds is the odds
an egg is cannibalized. Use the glmfit(x,y,‘binomial’) function in
MATLAB or other software for logistic regression.

a. Fill in the following table:

Factor b c SE OR 95% CI
SST 0.1 deg

b. Interpret the results in words, in terms of the reference value

c, the odds ratio OR, and significance.
c. Compare your results to those in Table 12.1, which are taken
from Hayward et al. (2014).

3. Fit model
log (odds) = b0 + b1 DAYS
to Data Set 12.1 using logistic regression, where odds is the odds
an egg is cannibalized. Use the glmfit(x,y,‘binomial’) function in
MATLAB or other software for logistic regression.

a. Fill in the following table:

Factor b c SE OR 95% CI
DAYS 1 day

b. Interpret the results in words, in terms of the reference value

c, the odds ratio OR, and significance.
c. Compare your results to those in Table 12.1, which are taken
from Hayward et al. (2014).

4. Complete Exercises 1–3. Suppose, in general, that there are two

important factors X1 and X2 . Discuss in depth the difference
between regressing on each factor individually and regressing on
both factors at once.
 Climate Change and Seabird Cannibalism: A Case Study ■ 259

BIBLIOGRAPHY
Atkins, G. J., Hayward, J. L., and Henson, S. M. 2021. How do gulls
synchronize every-other-day egg laying? Wilson Journal of Ornithology
133:226–235. DOI: 10.1676/20-00019. [The connection between socially-
facilitated mounting and egg-laying synchronization. Both theoretical and
empirical.]

Atkins, G. J., Reichert, A. A., Henson, S. M., and Hayward, J. L.

2017. Copulation call coordinates timing of head-tossing and mounting
behaviors in neighboring glaucous-winged gulls (Larus glaucescens).
Wilson Journal of Ornithology 129:560–567. DOI: 10.1676/16-004.1.
[Mounting and copulation calls spread through the colony via social
facilitation. Empirical.]

Atkins, G. J., Sandler, A. G., McLarty, M., Henson, S. M., and Hayward, J. L.
2015. Oviposition behavior in glaucous-winged gulls (Larus glaucescens).
Wilson Journal of Ornithology 127:486–493. DOI: 10.1676/14-151.1. [First
detailed description of egg-laying behavior in gulls. Empirical.]

Burnham, K. P. and Anderson, D. R. 2002. Model Selection and Multi-Model

Inference: A Practical Information-Theoretic Approach, 2nd ed. Springer-
Verlag, New York. [Comprehensive and user friendly text on information-
theoretic methods of model selection.]

Burton, D. and Henson, S. M. 2014. A note on the onset of synchrony in avian

ovulation cycles. Journal of Difference Equations and Applications 20:664–
668. DOI: 10.1080/10236198.2013.870564. [Mathematical treatment of egg-
laying synchrony as an adaptive response to egg cannibalism in seabirds.
Based on the master’s thesis of Danielle Burton. Theoretical.]

Cushing, J. M. and Henson, S. M. 2018. Periodic matrix models for

seasonal dynamics of structured populations with application to a
seabird population. Journal of Mathematical Biology 77:1689–1720. DOI:
10.1007/s00285-018-1211-4. [Mathematical analysis of a general class
of discrete-time matrix models that account for changes in behavioral
tactics within the breeding season and their dynamic consequences at the
population level across breeding seasons. Application to cannibalism and
reproductive synchrony in gulls. Theoretical.]

Cushing, J. M., Henson, S. M., and Hayward, J. L. 2015. An evolutionary

game theoretic model of cannibalism. Natural Resource Modeling 28:497–
521. DOI: 10.1111/nrm.12079. [Uses matrix models and bifurcation theory
to investigate population and evolutionary dynamic consequences of adult-
on-juvenile cannibalism. In the presence of cannibalism, a population
260 ■ Mathematical Modeling in Biology

can survive under circumstances of low resource availability which, in

the absence of cannibalism, lead to extinction. The evolutionary version
of the model shows that cannibalism can be an evolutionarily stable
strategy (ESS). Presents a gull egg cannibalism model as an application.
Theoretical.]
Darling, F. 1938. Bird Flocks and the Breeding Cycle. Cambridge University
Press, Cambridge. [Presents the “Fraser Darling Effect” hypothesis as to
why large colonies of birds breed simultaneously during a short breeding
season.]
Dong, Q. and Polis, G. A. 1992. The dynamics of cannibalistic populations:
A foraging perspective. In M. A. Elgar & B. J. Crespi (Eds.) Cannibalism:
Ecology and Evolution Among Diverse Taxa (pp. 13–37). Oxford
University Press, Oxford. [Classic reference on cannibalism.]
Elgar, M. A. and Crespi, B. J. 1992. Ecology and evolution of cannibalism.
Journal of Evolutionary Biology 7:1–12. [Classic reference on cannibalism.]
Fox, L. R. 1975. Cannibalism in natural populations. Annual Re-
view of Ecology and Systematics. 6:87–106. DOI: 10.1146/annurev.es.
06.110175.000511. [Classic reference on cannibalism.]
Gallos, D., Gallos, C., Watson, W., and Henson, S. M. 2018. A
note on synchronous egg laying in a seabird behavior model. Jour-
nal of Difference Equations and Applications 24:1953–1966. DOI:
10.1080/10236198.2018.1544633. [Mathematical treatment of egg-laying
synchrony as an adaptive response to egg cannibalism in seabirds. Based
on the undergraduate Honors theses of Christiana Gallos and Dorothea
Gallos. Theoretical.]
Giancristofaro, R. A. and Salmaso, L. 2003. Model performance analysis
and model validation in logistic regression. Statistica 63:375–396. DOI:
10.6092/issn.1973-2201/358. [Presents a validation method for logistic
regression models.]
Hayward, J. L., Weldon, L. M., Henson, S. M., Megna, L. C., Payne, B. G.,
and Moncrieff, A. E. 2014. Egg cannibalism in a gull colony increases with
sea surface temperature. The Condor: Ornithological Applications 116:62–
73. DOI: 10.1650/CONDOR-13-016-R1.1. [Paper on which this chapter
is based, showing that a 0.1 degree rise in sea surface temperature is
associated with a 10% increase in the odds that an egg is cannibalized.
Both theoretical and empirical: models connected to data.]
Henson, S. M. 1997. Cannibalism can be beneficial even when its mean
yield is less than one. Theoretical Population Biology 51:109–117.
 Climate Change and Seabird Cannibalism: A Case Study ■ 261

DOI: 10.1006/tpbi.1997.1303. [Mathematical treatment of benefits of

cannibalism. Theoretical.]

Henson, S. M., Cushing, J. M., and Hayward, J. L. 2011. Socially-induced

ovulation synchrony and its effect on seabird population dynamics. Journal
of Biological Dynamics 5:495–516. DOI: 10.1080/17513758.2010.529168.
[Poses a discrete-time model of socially-stimulated ovulation synchrony.
Shows mathematically that synchrony can increase total population size
and allow the population to persist at lower birth rates. Both theoretical
and empirical: models connected to data.]

Henson, S. M., Hayward, J. L., Cushing, J. M., and Galusha, J. G. 2010.

Socially induced synchronization of every-other-day egg laying in a seabird
colony. Auk 127:571–580. DOI: 10.1525/auk.2010.09202. [Announcement
of first demonstration of egg-laying synchrony. Both theoretical and
empirical: models connected to data.]

Hosmer, D. W. and Lemeshow, S. 2000. Applied Logistic Regression, 2nd ed.

John Wiley & Sons, New York. [The classic logistic regression text.]

McWilliams, K. M., Sandler, A. G., Atkins, G. J., Henson, S. M., and

Hayward, J. L. 2018. Courtship and copulation in glaucous-winged gulls,
Larus glaucescens, and the influence of environmental variables. Wilson
Journal of Ornithology 130:270–285. DOI: 10.1676/16-151.1. [Detailed
description of courtship and copulation behavior based on the master’s
thesis of Kelly McWilliams. Empirical.]

Nurhan, Y. I. and Henson, S. M. 2021. Cannibalism and synchrony in

seabird egg-laying behavior. Natural Resource Modeling 34:e12325. DOI:
10.1111/nrm.12325. [Mathematical treatment of egg-laying synchrony as
an adaptive response to egg cannibalism in seabirds. Based on the
undergraduate Honors thesis of Yosia Nurhan. Theoretical.]

Polis G. A. 1981. The evolution and dynamics of intraspecific predation.

Annual Review of Ecology and Systematics. 12:225–251. DOI: 10.1146/an-
nurev.es.12.110181.001301. [Classic reference on cannibalism.]

Polski, A. A., Osborn, K. J., Hayward, J. L., Joo, E., Mitchell, A. T., Sandler,
A. G., and Henson, S. M. 2021. Egg cannibalism as a foraging tactic
by less fit glaucous-winged gulls (Larus glaucescens). Wilson Journal of
Ornithology 133:552–567. DOI: 10.1676/20-00072. [Comprehensive study
of the behavioral ecology of egg cannibalism in glaucous-winged gulls.
Based in part on the undergraduate Honors theses of Ashley Polski and
Karen Osborn. Empirical.]
262 ■ Mathematical Modeling in Biology

Sandler, A. G., Megna, L. C., Hayward, J. L., Henson, S. M., Tkachuck, C.

M., and Tkachuck, R. D. 2016. Every-other-day clutch-initiation synchrony
in ring-billed gulls (Larus delawarensis). Wilson Journal of Ornithology
128:760–765. DOI: 10.1676/15-121.1. [Shows that egg-laying synchrony
occurs in a second species of gulls. Both theoretical and empirical: models
connected to data.]

Smith, R. S., Weldon, L. M., Hayward, J. L., and Henson, S. M. 2017.

Time lags associated with effects of oceanic conditions on seabird breeding
in the Salish Sea region of the northern California Current system.
Marine Ornithology 45:39-42. [Uses logistic regression and model-selection
techniques to determine the time of year SST should be measured in
order to best explain reproductive success the following breeding season
on Protection Island. Based in part on the undergraduate Honors thesis
of Rashida Smith. Both theoretical and empirical: models connected to
data.]

Weir, S. K., Henson, S. M., Hayward, J. L., Atkins, G. J., Polski, A. A.,
Watson, W., and Sandler, A. G. 2020. Every-other-day clutch-initiation
synchrony as an adaptive response to egg cannibalism in glaucous-
winged gulls (Larus glaucescens). Wilson Journal of Ornithology 132:575–
586. DOI: 10.1676/19-82. [Demonstration that egg-laying synchrony is
adaptive in the presence of egg cannibalism in gulls. Based in part on
the undergraduate Honors thesis of Sumiko Weir. Both theoretical and
empirical: models connected to data.]
 V
Appendix

263
 APPENDIX A

Linear Algebra Basics

This appendix is a self-guided tutorial through the basics of matrices

and linear algebra. The student who works through the examples and
exercises in this chapter will be prepared for the linear algebra required
in this textbook. These exercises should be worked “by hand.”

A.1 MATRIX OPERATIONS

A n × m matrix is an array of numbers arranged in n rows and m
columns. For example, this is a 2 × 3 matrix:
 
1 −2 4
.
−1 0 3

A.1.1 Matrix Addition

Matrices of the same dimensions can be added. Matrix addition is
entry-wise, for example
       
1 −2 2 −1 1+2 −2 + (−1) 3 −3
+ = = .
−1 0 1 −3 −1 + 1 0 + (−3) 0 −3

A.1.2 Scalar Multiplication

A matrix can be multiplied by a scalar (real number). Scalar
multiplication is entry-wise, for example
     
1 −2 4(1) 4(−2) 4 −8
4 = = .
−1 0 4(−1) 4(0) −4 0

265
266 ■ Appendix A

A.1.3 Matrix Subtraction

Matrix addition and scalar multiplication together allow us to subtract
matrices of equal dimension. For example,
       
1 −2 2 −1 1 −2 2 −1
− = + (−1)
−1 0 1 −3 −1 0 1 −3
   
1 −2 −2 1
= +
−1 0 −1 3
 
−1 −1
= .
−2 3

This is, of course, just entry-wise subtraction; the intermediate steps

are henceforth unnecessary.

A.1.4 Matrix Multiplication

The product of a row vector (a 1 × n matrix) and a column vector (an
n × 1 matrix) is computed in the following way:

 

d
a b c  e  = ad + be + cf.
f

For example,
 

2
1 4 = 1(2) + 4(−3) = −10.
−3

This is called the dot product of the two vectors. Note that the dot
product gives a scalar, that is, a 1 × 1 matrix.
In general, two matrices can be multiplied together if the first has
dimension n × m and the second has dimension m × p. The result is
an n × p matrix in which the (i, j)th entry, meaning the entry in the
ith row and jth column, is the dot product of the ith row of the first
matrix and the jth column of the second matrix. This sounds more
difficult than it is; matrix multiplication is easy to carry out once you
 Appendix A ■ 267

see the pattern. Here are several examples:

 

4 7

1 2 3  5 8  = 1(4) + 2(5) + 3(6) 1(7) + 2(8) + 3(9)
6 9


= 32 50 ;

    
1 −2 2 −1 1(2) − 2(1) 1(−1) − 2(−3)
=
−1 0 1 −3 −1(2) + 0(1) −1(−1) + 0(−3)
 
0 5
= ;
−2 1
      
1 −2 2 1(2) − 2(−3) 8
= = .
−1 0 −3 −1(2) + 0(−3) −2
Note that matrix multiplication is not in general commutative. That
is, the order in which matrices are multiplied matters.

A.1.5 Determinants of Square Matrices

The determinant of a 2 × 2 matrix is the scalar number

a b
= ad − bc.
c d
This textbook mostly requires computing determinants for 2 × 2
matrices. If you need to compute the determinant of a higher-
dimensional matrix, simply look up the procedure in a more
comprehensive discussion of matrices.
Work all of the following exercises, which are designed to lead
you step by step through the major concepts you will need from
linear algebra. A list of solutions follows the exercises. You will find
a summary of the concepts at the end.

A.2 EXERCISES
  
2 −1 1 −2
1. =
1 −3 −1 0
268 ■ Appendix A
  
0 2 1 1
2. =
3 −2 1 −3
  
1 −2 1 0
3. =
−1 0 0 1
  
1 0 2 −1
4. =
0 1 1 −3
  
1 −2 x
5. =
−1 0 y
  
0 2t a
6. =
3t3 −2 b
   
1 −2 2 −1
7. − =
−1 0 1 −3
   
1 −2 2 −1
8. 3 +α =
−1 0 1 −3
    
1 −2 x 2t
9. + =
−1 0 y −t3

10. Compute A − λI,

 
1 −2
where λ is a scalar and A = . Note: The matrix
−1 0
 
1 0
I= is called the identity matrix.
0 1

1 −2
11. Compute the determinant. =
−1 0

5 −1
12. =
3 6

(2 − λ) 1
13. =
−1 (−1 − λ)

14. Find the determinant |A − λI| ,

 Appendix A ■ 269
 
5 −1
where λ is a scalar and A = . This is called the
3 6
characteristic polynomial of A.

15. Solve the following equation for λ:

|A − λI| = 0,

 
5 −1
where A = . The two λ-solutions are called the
0 6
eigenvalues of the matrix A.

16. Solve the following equation for λ:

|A − λI| = 0,

 
1 2
where A = . That is, find the eigenvalues of A. Hint:
−2 1
The eigenvalues in this case are complex numbers.
 
−3 2
17. Find the eigenvalues of A = .
−2 2

18. 
For eac
h of the eigenvalues in Problem 17, find a vector v =
v1
such that Av = λv; that is, find a vector v such that
v2
 
0
v satisfies the vector equation (A − λI)v = 0, where 0 =
0
is the zero vector. These vectors v are called eigenvectors. You
may wish to follow the complete solution given below as you work
through this procedure for the first time.

19. Find the eigen

values and associated eigenvectors for the matrix
1 3
A= .
0 2

20. Find the eigenvalues

 and associated eigenvectors for the matrix
2 3
A= .
−1 −2
270 ■ Appendix A

A.3 SOLUTIONS
Some of the following solutions omit intermediate steps.
    
2 −1 1 −2 3 −4
1. =
1 −3 −1 0 4 −2
    
0 2 1 1 2 −6
2. =
3 −2 1 −3 1 9
    
1 −2 1 0 1 −2
3. =
−1 0 0 1 −1 0
    
1 0 2 −1 2 −1
4. =
0 1 1 −3 1 −3
    
1 −2 x x − 2y
5. =
−1 0 y −x
    
0 2t a 2bt
6. =
3t3 −2 b 3at3 − 2b
     
1 −2 2 −1 −1 −1
7. − =
−1 0 1 −3 −2 3
     
1 −2 2 −1 3 + 2α −6 − α
8. 3 +α =
−1 0 1 −3 −3 + α −3α
      
1 −2 x 2t 2t + x − 2y
9. + =
−1 0 y −t3 −x − t3
     
1 −2 1 0 (1 − λ) −2
10. A − λI = −λ =
−1 0 0 1 −1 −λ

1 −2
11. = −2
−1 0

5 −1
12. = 33
3 6

(2 − λ) 1
13. = λ2 − λ − 1
−1 (−1 − λ)
 Appendix A ■ 271

14. The characteristic polynomial of A is

     
5 −1 1 0 5 − λ −1
−λ =
3 6 0 1 3 6−λ
= λ2 − 11λ + 33
15. |A − λI| = 0 reduces to (λ − 5) (λ − 6) = 0, which holds if and
only if λ = 5, 6. These two numbers are the eigenvalues of A.
16. The quadratic formula leads to the eigenvalues λ = 1 ± 2i.
17. The eigenvalues of A are λ = 1, −2.
18. The eigenvalues are λ = 1, −2. Also,
(A − λI)v = 0
is equivalent to
    
−3 − λ 2 v1 0
= ,
−2 2−λ v2 0
which reduces to

(−3 − λ) v1 + 2v2 = 0
. (A.1)
−2v1 + (2 − λ) v2 = 0
For λ = 1, Equation (A.1) becomes

−4v1 + 2v2 = 0
.
−2v1 + v2 = 0
Thus, we have v2 = 2v1 , so choose v1 = 1 and then v2 = 2. Note
that any v1 and v2 that satisfy the relationship v2 = 2v1 will
work.
For λ = −2, equation (A.1) becomes

−v1 + 2v2 = 0
.
−2v1 + 4v2 = 0

Thus, we have v2 = 12 v1 , so choose v1 = 2 and then v2 = 1.

 Therefore,
 an eigenvector belonging to eigenvalue λ = 1 is v =
1
, and an eigenvector belonging to eigenvalue λ = −2 is v =
2
 
2
.
1
272 ■ Appendix A

19. The eigen

 values
 are λ = 1, 2. An eigenvector belonging
 toλ=1
1 3
is v = . An eigenvalue belonging to λ = 2 is .
0 1

valuesare λ = 1, −1. An eigenvector belonging to λ = 1

20. The eigen
−3
is v = . An eigenvalue belonging to λ = −1 is v =
1
 
−1
.
1

A.4 SUMMARY OF LINEAR ALGEBRA CONCEPTS

Here is a summary of some terms and a note about eigenvectors.

1. The λ-polynomial |A − λI| is called the characteristic polynomial

of A.

2. The λ-equation |A − λI| = 0 is called the characteristic equation

for A.

3. The λ-solutions of the characteristic equation |A − λI| = 0 are

called eigenvalues of A.

4. The vector equation (A − λI)v = 0 is called the eigenvalue

problem for A.

5. Vector solutions v of the eigenvalue problem (A − λI)v = 0 are

called eigenvectors belonging to λ.

6. The eigenvector belonging to a  particular eigenvalue is not

−1
unique. For example, if v = is an eigenvector belonging
1
   
−2 5
to λ, then so is v = and so is v = . In fact, any
2 −5
 
−1
scalar multiple, that is, any vector of the form c , is an
1
eigenvector.
 APPENDIX B

MATLAB: The Basics

This MATLAB primer is a self-guided tutorial designed to be read

while sitting at a computer with MATLAB. The student who reads
and successfully works through all the examples in this chapter will
obtain the basic working knowledge of MATLAB syntax needed for
this textbook.

B.1 PRELIMINARIES
Set the defaults on computer so that file extensions are visible. It is
best to leave your computer this way. In Windows 10, right-click on
Start > File Explorer > View. Then enable the “File name extensions”
by checking the box.
Keep shortcuts for Notepad, Excel, and MATLAB on your desktop
or taskbar for easy access. You can find Notepad in Windows 10 by
left-clicking on Start.

B.2 SYNTAX AND PROGRAMMING

B.2.1 Command Line
The MATLAB command line prompt is >>. This tutorial will use a
single >. You can do simple programming directly at the command
line.

Example B.1 Type the following commands at the prompt. After

typing each command, press Enter. This performs the given command
and introduces a new prompt below it.

273
274 ■ Appendix B

> x = 5
> y = 6

B.2.2 Case-Sensitivity
MATLAB is case-sensitive. This includes variables and file names.

B.2.3 Displaying the Current Value of a Variable

Simply type the variable name at the command prompt if you wish to
know its current value.

Example B.2 Type the following commands at the prompt. Here, we

assign values to x and y, and then take z to be their sum. If we type z,
or x, or zˆ2 at the prompt, MATLAB prints the value to the screen.

> x = 2
> y = 3
> z = x+y
> z
> x
> z^2

B.2.4 Clearing All Variables

At any point, one can clear the memory of all variables by typing
> clear

B.2.5 Closing MATLAB

To close MATLAB, type
> exit
or
> quit
at the command prompt.

B.2.6 Variables and Arithmetic Operators

The equal sign in programming does not mean quite the same thing as
it does in mathematics. When you type
 Appendix B ■ 275

>x=5
at the command line in MATLAB, you are saying:
> variable = known numerical value
That is, you are saying that x is the name of a variable and you are
assigning the value 5 to it. The right-hand side of an equal sign must be
either a number, or a combination of variables that have already been
assigned numerical values. For example, in programming, the command
> x = x+1 is perfectly legitimate if x has already been assigned a value,
whereas in algebra, the statement x = x + 1 is always a contradiction.

Example B.3 At the command prompt, type the following commands.

In the first line, you are storing the value 5 in the variable x. In the
second line, you are storing the value of x (which is now 5 from the
previous line) in the variable y, so now y is 5. In the third line, you
are storing the value x+1 (which is 6) back into the variable x. Now x
is 6. Finally, in the last line, you are storing the value of x+y (which
is now 11) in the variable z.

> x=5
> y=x
> x=x+1
> z=x+y

The syntaxes for the five arithmetic operations are

+ − ∗ / ˆ
The symbol ˆ means “raised to the power.”The order of operations
is the same as it is in arithmetic: compute parentheses, then exponents,
then multiplication and division, and then addition and subtraction.
When everything is at the same level of operators, the calculation
proceeds left to right.

Example B.4 Type the following commands at the prompt. In the

third line, the operation in parentheses is computed first (6+2 is 8), and
then that value is squared, which is 64. Then the division is computed
(64/2 is 32), and finally, the x is added, giving 32+6, which is 38. The
fourth line gives z = 518, whereas the fifth line gives z = 32768. Make
sure you understand the order of operations in each case.
276 ■ Appendix B

> x=6
> y=2
> z = (x + y)ˆ2/y + x
> z = (x + y)ˆ(6/y) + x
> z = (x + y)ˆ6/(y + x)

If in doubt about the order of operations, you can use extra

parentheses to clarify your command, but it is generally bad form.
Learn the order of operations.

B.2.7 Programs
Typing a long sequence of complicated commands at the prompt is
cumbersome, so we write such sequences of commands in a separate
file called a program (or script, or code). When you run the program
at the MATLAB prompt, each line within the program is executed
sequentially. Here are the general steps for writing and running a
program.
1. To write a program using the MATLAB editor, go to “File” in
MATLAB; choose “New”and then “Script.”
2. Type code (lines of commands) into the file. Do not type > in
front of your commands.
3. Save the program file to the folder where you are keeping
your MATLAB work. It should be saved as programname.m. Here,
“programname” is a placeholder for whatever name you want to call
the program. The extension must be *.m for MATLAB.
4. In MATLAB, you need to “set the path,”which means directing
MATLAB to the working folder in which you store your programs.
The exact way in which you set the path depends on the MATLAB
version you are using. You may find “Set Path” under a “File” or
“Environment” tab, and many versions have a “browse for folder”
button at the top of the Editor window with which you can choose
your working folder.
5. To run the program called programname.m, go to the MATLAB
command prompt and type the program name without the extension
(no “.m”).
> programname
 Appendix B ■ 277

Example B.5 Type the following three lines in a script file, and save
the file to your working folder as hey.m. (The horizontal lines shown
here are not part of the program.) To run the program hey.m, simply
type > hey at the command prompt. The output to the screen will be
the values of x, y, and z.

x=6
y=2
z = (x + y)ˆ6/y + x

B.2.8 Comment Lines

Within a program, the symbol % is used to prevent the execution of a
particular line. Commenting out lines is useful for documenting your
program and for debugging. For example,
%This line is "commented out" and will not execute

B.2.9 Printing to the Screen

A semicolon is used at the end of a command if you don’t want the
value to print to the screen.

Example B.6 Type the following commands at the command prompt

and notice the difference. After you enter the first line and press Enter,
the value 5 prints to the screen. In the second line, the value does not
print to the screen, although it is still stored in the variable y, as you
see after executing the third line.

>x = 5
> y = 6;
>y

The semicolon at the end of lines is particularly useful in large

programs because printing a lot of output to the screen can drastically
slow down the program. Removing the semicolon from selective lines
in programs is useful for debugging because it allows you to see the
value of a certain variable at a certain point as you run the program.
278 ■ Appendix B

B.2.10 Numerical Format

The default numerical format for MATLAB is floating-point decimal
format. It will report numbers to four decimal places unless the format
is changed to “long” in which case it will report to fifteen decimal
places. If the format is changed to “rat” the numbers will be reported
as ratios (or fractions). One can change the format as follows.
>format long
>format short
>format rat
Scientific notation is denoted by e+###. For example, Avogadro’s
number could be entered as 6.022e + 23 or 6.022 × 10ˆ23. Either way,
it would appear in the short format setting as 6.0220e + 23.

B.2.11 Loops
Loops are used to repeat sections of the code (program) a specified
number of times. The syntax is:
for countername = startvalue : increment : endvalue
(various commands)
end
Note that we indent the commands that are inside the loop. This
helps highlight the logical structure of the code.

Example B.7 Write the following code (program). Save it as

tschuss.m. Note the comment lines that document the code. Note also
that MATLAB does not notice the extra vertical and horizontal spacing.
Use such spacing liberally to make your code easy to read. To run the
program, type > tschuss at the command prompt.

%Program tschuss.m
%Initialize x
x = 2

%Loop
for i = 1 : 1 : 6
x = x * x + i
end
 Appendix B ■ 279

Note that the “end” statement closes the loop.

B.2.12 Crashing a Program on Purpose

If you want to stop the execution of a long program, type control-c.
Example B.8 Create and run the following program. Call it wiege-
hts.m. The goal is to print the final value of x to the screen. Suppose that
you forget to put a semicolon after the statement x = i+100. All the
intermediate values of x will write to the screen, and the program will
take a long time to run because of this. While the program is running,
crash it using control-c. Now put a semicolon at the end of statement x
= i+100 to suppress the intermediate screen output. Save the program
file, and rerun the program. Note the speed at which the program runs
when it doesn’t have to print the intermediate values of x to the screen.
Also note that whenever you modify a program file, you must save it
before rerunning it.

%Program wiegehts.m
for i = 1 : 1 : 100000
x = i+100
end
x

B.2.13 Logical Statements (If-Then-Else)

The syntax for logical operators is
& and
| or
> greater than
>= greater than or equal to
< less than
<= less than or equal to
== equal to
Example B.9 Write and run the following code called hola.m. Try
various initial values of x. Don’t forget to save the program file after
each change before you rerun the program.
280 ■ Appendix B

%Program hola.m
%Initialize x
x = 7

%Set the variable "check" according to the value of x

if x < 5
check = 0
elseif x >= 5 & x < 10
check = 1
else
check = 2
end

Here, the “end” statement closes the logical decision tree.

B.2.14 Input and Output Files

Input files are text (*.txt) files created with Notepad. They are used to
import vectors or matrices of numerical data into MATLAB. The input
file should be in the working folder where you keep your programs. The
command to load the input file is
> load inputfilename.txt
The input matrix is saved in a variable called “inputfilename”.

Example B.10 Open a new Notepad file and type the first 10 positive
integers in a column vector. Save this input file as ciao.txt. Go to the
command prompt and type the following commands. After you load
ciao.txt, the vector of numbers from the input file is stored in a variable
named “ciao”. The second command renames the vector of numbers,
calling it “data”. The third command adds one to each entry in the
vector.

> load ciao.txt

> data = ciao
> data = data + 1
 Appendix B ■ 281

Example B.11 Suppose that, after typing the commands in the

previous example, you want to save the new values stored in “data” to
an output file called “bye.txt”. Type the following command. MATLAB
will create the file bye.txt in your working folder. Open bye.txt with
Notepad and see what it looks like.

>save bye.txt data -ascii

In general, the syntax for saving a variable to an output file is

>save outputfilename.txt variablename -ascii

B.2.15 Creating Functions

Here are the general steps for creating a function in MATLAB.
1. Open a new script file and save it as functionname.m. Here,
“functionname” is a placeholder for whatever name you want to give
the function.
2. The first executable line in the script should declare that this is
a function. The declaration line should look like
function y = functionname(inputvariable)
3. Define the function in the next lines.
4. Close the function definition with end.
5. At the command prompt, evaluate the function at a particular
value by typing
> functionname(value)

Example B.12 Create a script file for f(x) = x2 using the following
code. Save the program as “square.m”. Note that the name of the
program file should be the same as in the function declaration line.

%Create the squaring function square.m

function y = square(x)

%Define y as a function of x
y = x^2;

end
282 ■ Appendix B

After saving the above program to your working folder, type the
following commands at the prompt.
> square(5)
> square(10)
The screen output should be 25 and 100.

B.2.16 Subroutines
A subroutine is a subsidiary program that is called within a main
program. It is often easier to see the structure of the main program
when certain groups of commands are relegated to subroutines.

Example B.13 Create a subroutine called sub.m that is called by a

program called main.m. You will write two different programs as shown
below. Declaring variables as global means they are shared between the
two programs. With pencil and paper, work through the first several
loops of the program and make sure you understand what the program
does at each step. Then run the main program at the command prompt
by typing > main.

%Program main.m
%Calls subroutine sub.m
%Declare global variables
global x check

%Initialize x
x = 0;

%Loop
for i = 1 : 0.5 : 20
x = x + i;
%Call the subroutine sub.m
sub;
%Print out value of the variable check
check
end
 Appendix B ■ 283

%Subroutine sub.m
%Called by main.m
%Declare global variables
global x check

%Logical decision used to set value of check

if x < 5
check = 0;
elseif x >= 5 & x < 10
check = 1;
else
check = 2;
end

B.2.17 Vectors, Matrices, and Arrays

Example B.14 Type the following commands at the command prompt
and note the output. Make sure you understand each operation. The
parenthetical explanatory remarks are not part of the commands.

> x = [1 -3 5 9]
> y = [2 0 3 -1]
> a = 0 : 0.2 : 3
> a = a’ (transpose operation)
> n = length(y) (finds the “length” of the vector = number of
entries)
> z = x + y (addition of vectors/matrices)
> zz = sum(y) (this is the sum of the entries of the vector y)
> x = x’
> w = [1 -3 4]’
> w(1) (returns the first entry of the vector w)
> w(3) (returns the third entry of the vector w)
> y = [1 3 5; 2 4 -3; 7 -2 0] (semicolons denote different
rows)
> y = y’
> x = [1 2 5
3 7 1
-2 0 0] (line breaks denote different rows)
284 ■ Appendix B

> z = x + y
> z = x*y (multiplication of matrices)
> z = x.*y (entry-wise multiplication of matrices)
> z = 2*x (scalar multiplication of matrices)
> z = x/2 (scalar division of matrices)
> z = x.^2 (entry-wise squaring of matrix entries)
> z = x + 1

The syntax for locating an entry in a matrix (or an array) is

> arrayname(row,column)

Example B.15 Type the following commands at the command

prompt.

> x = [1 2 5
3 7 1
-2 0 0]
> z = x(2,3)
> z = x(1,2)

A colon can be used to denote “through”. When standing alone,

the colon denotes all rows or columns.

Example B.16 Using the same matrix x as in the previous example,

type the following commands.

> z = x(2:3,1) (second through third row, first column)

> z = x(:,3) (all rows, third column)
> z = x(1,:) (first row, all columns)

One can “stack” vectors or matrices of compatible dimensions, and

this can be done horizontally or vertically, as you see in the next
example.

Example B.17 Try the following commands and note the result of
each type of syntax.

> y = [1 2]
> v = [y 3]
> w = [y [3 4]]
> z = [y; [3 4]]
 Appendix B ■ 285

Example B.18 There are three commands for building special vec-
tors/matrices. Try these examples at the command prompt.

> z = ones(2,5) (matrix with 2 rows and 5 columns of ones)

> z = zeros(3,1) (matrix with 3 rows and 1 column of zeros)
> z = [] (empty matrix/vector)

Loops can be used to build vectors, as well.

Example B.19 Try the following two sample programs.

%Initialize x as a column vector of 10 zeros

x = zeros(10,1);

%Loop
for i = 1 : 1 : 10
x(i) = 3*i;
end
x

An alternative, but slower way to write the above program is

%Initialize x as empty vector

x = [];

for i = 1 : 1 : 10
x = [x; 3*i];
end
x

B.2.18 Functions in the MATLAB Library

Many functions already exist in the MATLAB library. The following
exercise contains the syntax for some of the most common functions.
The student can look up the syntax for other functions in the MATLAB
help files as needed.
286 ■ Appendix B

Example B.20 Type the following commands at the prompt. Note

what each command accomplishes. The parenthetical remarks are not
part of the commands.
> x = 2
> y = [1 2 3 4]
> log10(x) (this is log base 10)
> log10(y)
> log(y) (this is log base e, the natural log)
> sin(x)
> cos(y)
> tan(y)
> atan(y) (this is the arctangent of y)
> sec(x)
> csc(x)
> cot(y)
> exp(x) (this is e raised to the x power)
> exp(y)
> sqrt(x)

B.2.19 Plotting
To plot vector y against vector x, use the command
> plot(x,y)
Example B.21 Type the following commands at the prompt. Note
what each command accomplishes.
> x = 1 : 0.1 : 8
> y = sin(x)
> plot(x,y)

B.2.20 Simulating Discrete-Time Models

Discrete-time models, which are simply iterative rules, run quickly on
computers.
Example B.22 Write a program to simulate the Ricker model
xt+1 = bxt e−cxt
x0 = 1.
 Appendix B ■ 287

Set b = 8 and c = 0.01. Find x50 .

One way to write this program is to iteratively update the value of

x, as in the following code.

%Program Ricker.m
%Initialize x and set parameters b and c
x = 1;
c = 0.01;
b = 8;

%Loop
for i = 1 : 1 : 50
x = b*x*exp(-c*x);
end

%Output value to screen

x

A better way to update the value of x is the following.

%Program Ricker.m
%Initialize x and set parameters b and c
x = 1;
c = 0.01;
b = 8;

%Loop
for i = 1 : 1 : 50
xnew = b*x*exp(-c*x);
x = xnew;
end

%Output value to screen

x
288 ■ Appendix B

An even better way to write this program is to treat x as a vector

and define each entry.

%Program Ricker.m
%Initialize x and set parameters b and c
x = 1;
c = 0.01;
b = 8;

%Loop
for i = 1 : 1 : 50
x(i+1) = b*x(i)*exp(-c*x(i));
end

%Output value to screen

x(51)

%plot x vs. t for whole trajectory

time = 1 : 1 : 51;
plot(time,x)

In the last program above, make sure you understand why x(51) is
x50 . Hint: In MATLAB, a vector cannot have a zeroth entry.

Example B.23 Write a program to simulate this two-dimensional

Ricker model.

xt+1 = byt e−cyt

yt+1 = xt + (1 − a)yt
x0 = 1
y0 = 2.

Set a = 0.2, b = 8, and c = 0.01. Find x50 and y50 .

%Program Ricker2D.m
%Set parameters
 Appendix B ■ 289

c = 0.01;
b = 8;
a = 0.2;

%Initialize x and y
x = 1;
y = 2;

%Loop
for i = 1 : 1 : 50
xnew = b*y*exp(-c*y);
ynew = x + (1-a)*y;
x = xnew;
y = ynew;
end

%Output values to screen

x
y

Note that in the two-dimensional Ricker program above, x and y

cannot be updated with the code

x = b*y*exp(-c*y);
y = x + (1-a)*y;

because the second line would use the updated value of x on its right-
hand side instead of the previous value of x. It is very important to
understand this.
Often we want to plot the whole trajectory of x and y. Here is a
better way to code the previous example.

%Program Ricker2D.m
%Set parameters
c = 0.01;
b = 8;
a = 0.2;
290 ■ Appendix B

%Initialize x and y
x = 1;
y = 2;

%Loop
for i = 1 : 1 : 50
x(i+1) = b*y(i)*exp(-c*y(i));
y(i+1) = x(i) + (1-a)*y(i);
end

%Output values to screen

x(51)
y(51)

%Plot x and y on the same axes

time = 1 : 1 : 51;
plot(time,x,time,y)

B.2.21 Simulating Ordinary Differential Equations (ODEs)

Simulations of continuous-time models such as ODEs tend to run much
more slowly than those of discrete-time models. The syntax for the
MATLAB ode45 integrator is
> ode45(’RHS’, [starttime endtime], initialcondition)
where RHS refers to a program that computes the right-hand side of
the differential equation.

Example B.24 This example concerns a famous model from popula-

tion dynamics called the logistic model. Integrate the logistic model
 
dN N
= 2N 1 −
dt 100
N (0) = 5

from t = 0 to t = 42. Here, N (t) is the population size at time t. The

statement N (0) = 5 is called the initial condition. Graph the population
size N vs. time t. What is N (42)?
 Appendix B ■ 291

To do this example, first write a function program called

calcderiv.m that computes the RHS of the logistic ODE.

%Program calcderiv.m
function Nprime = calcderiv(t,N)

%Define RHS of ODE

Nprime = 2*N*(1-N/100);
end

Then, in MATLAB, go to the command prompt and type the

following commands.
> [T pop] = ode45(’calcderiv’,[0 42], 5)
> plot(T, pop)
> n = length(T)
> pop(n)

Or, write a program called logistic.m that executes these same lines:

%Program logistic.m
[T pop]=ode45(’calcderiv’, [0 42], 5);
plot(T, pop)
n = length(T);
pop(n)

and then call the program from MATLAB:

> logistic
Note that in the example above, T is a vector of times along
the trajectory and pop is the vector of corresponding N values. The
command > n = length(T) sets n to be the number of entries in the
vector T . The command > pop(n) picks off the nth value of the vector
pop, which is the very last value, N (42), which is the population size
at time 42. Make sure you understand these variables by typing them
at the command prompt to see their values.
>T
> pop
292 ■ Appendix B

>n
> pop(1)
> pop(5)
> pop(n)
The last three commands above are not function evaluations. They
return the 1st, 5th, and nth entries, respectively, of the vector pop.

B.2.22 The Downhill Nelder-Mead Algorithm

The Nelder-Mead algorithm is a “downhill method” of numerically
finding minima of functions. MATLAB already has this algorithm in
its library of functions. The syntax is
> [x,fval] = fminsearch(’functionname’,x0)
Here, x0 is the initial “guess” from which the algorithm starts
downhill, x is the value of the independent variable that minimizes
the function, and f val is the function value at the minimum. The
independent variable x can be a scalar or a vector.

Example B.25 Write a program that finds a “minimizer” of the

2
multivariate function g(a, b) = (a − 2) + (b − 3)2 + 5. That is, write
a program that finds values of a and b that minimize the value of
2
(a − 2) + (b − 3)2 + 5. You will need to write two programs: a main
program and a subroutine that computes the value of g(a, b).

%Program findmin.m
%Calls fminsearch.m and gfun.m

%Set initial "guess" for a and b

theta = [0.1 0.03]’;
[x fval] = fminsearch(’gfun’,theta)

%Program gfun.m
function y = gfun(theta)

a = theta(1);
b = theta(2);
2 2
y = (a − 2) + (b − 3) + 5;
end
 Appendix B ■ 293

To run the above program, type

> findmin
at the command prompt. Your output to the screen should be x =
[2, 3]′ , and f val = 5. Here, x is the minimizer and f val is the value of
the function g(a, b) at the minimizer.

B.3 EXERCISES
1. Write a program that finds minimizers of f (x) = sin x. As you
know, sin x has infinitely many minima. Note how your answer
depends on your initial “guess” x0. Attach your programs and
screen output for several values of x0.

2. Use the Nelder-Mead algorithm to solve numerically the

2
transcendental equation x = e−x . Hint: Let g(x) = (x − e−x )
and find the minimizer for g. Attach your programs and screen
output.

3. Write a program to simulate the discrete-time system

2yt
xt+1 = + 0.25xt
1 + yt
yt+1 = 0.50xt + 0.7yt
x0 = 1
y0 = 1

from t = 0 to t = 100. Plot x vs. t and y vs. t on the same axes.

4. Consider the following model of a population of bacteria, where

time t is in hours and N (t) is the number of bacteria at time t.
 
dN N
= 0.03 N (N − 15) 1 −
dt 100
N (0) = 17.

Numerically integrate this model from t = 0 to t = 1. What is

the value of N (1)? Attach your programs and your output.

5. Consider the model in Exercise 4. Write a program that makes a

list of the model predictions at the top of each hour. In particular,
294 ■ Appendix B

use a loop and ode45 to build a vertical vector x having 20 entries,

such that x(1) = 17, x(2) = N (1), x(3) = N (2), etc. Save the
vector to an output file. Attach your programs and your output
file.
 APPENDIX C

Connecting Models to
Data: A Brief Summary
with Sample Codes

This appendix outlines some methods of connecting models to data and

gives samples of MATLAB code that will help students implement the
methods on the computer. Students unfamiliar with MATLAB should
first work through the self-guided MATLAB tutorial in Appendix B.

C.1 PARAMETERIZATION
The main ideas are:

1. A model, say N (t) = at2 + b, has dependent variables (in this

case, N ), independent variables (in this case, t), and parameters
(in this case, a and b).

2. Imagine graphing the model curve N vs. t as well as observed

data points (t, n) as a scatter plot on the same axes. The idea is
to make the curve “fit” the data points.

3. We adjust the shape of the curve by “tuning” the values of the

parameters a and b.

4. We want to find the values of a and b that make the curve “best”
fit the data. Such values b
a and bb are the parameter estimates.
There are various methods for obtaining parameter estimates.

295
296 ■ Appendix C

The two methods discussed in this book are maximum likelihood

(ML) and least squares (LS).

C.2 RESIDUAL ERRORS (RESIDUALS)

Let’s call the model prediction pred and the associated data point
observation obs.
1. The residual error is the observed minus the predicted: res =
obs − pred.
2. In general, we apply a variance-stabilizing transformation ϕ to
make noise approximately additive on the ϕ-scale. Thus, the residual
on the ϕ-scale is res = ϕ (obs) − ϕ (pred).
√
3. For demographic noise ϕ(·) = ·.
4. For environmental noise ϕ(·) = ln(·).
5. For dynamic models such as differential equations and difference
equations, it is sometimes best to compute “one-step residuals,” in
which the model prediction pred is conditioned on the observation at
the previous time step.

C.3 RSS AND R2

Suppose noise is approximately additive on the ϕ-scale.
1. The sum of squared residuals (residual sum of squares, or RSS)
is
X 2
RSS = (ϕ (obs) − ϕ (pred)) .
data

2. A generalized goodness of fit is

P 2
(ϕ (obs) − ϕ (pred))
R2 = 1 − data
P  2 ,
ϕ (obs) − ϕ (obs)
data

where ϕ (obs) denotes the sample mean of the transformed observa-

tions.
3. R2 gives the proportion of variability in the data that is explained
by the model, relative to using the mean of the data as a predictor.
 Appendix C ■ 297

C.4 MAXIMUM LIKELIHOOD (ML) PARAMETERS

In this section, we assume data and predictions have already been
transformed by ϕ.
1. The residuals depend on the values of the model predictions,
which depend on the values of the parameters. So if we “tune” the
parameters, we change the values of the residuals.
2. Suppose noise is Gaussian with mean zero and constant variance
σ 2 and that stochastic perturbations are uncorrelated in time. The
likelihood that a given residual res comes from the normal distribution
with mean zero and standard deviation σ is
1 1 res 2
√ e− 2 ( σ ) .
σ 2π
3. For a model with one dependent variable, the likelihood that
ALL the residuals come from the distribution is the product
1 1 res 2
√ e− 2 ( σ ) .
Y
L=
data
σ 2π

4. The likelihood function L is a function of the model parameters.

5. The ML parameter estimates are those parameter values that
maximize L, and these are equivalent to those that maximize the log-
likelihood ln L.
6. If there are q residuals, the log-likelihood is
q 1 X 2
ln L = −q ln σ − ln (2π) − 2 (res)
2 2σ
data
q 1
= −q ln σ − ln (2π) − 2 RSS.
2 2σ
6. Maximizing the log-likelihood in this context is equivalent to
minimizing the RSS.

C.5 LEAST SQUARES (LS) PARAMETERS

Note that:

1. The residuals, and hence RSS, are functions of the parameter

values.
298 ■ Appendix C

2. One can use the Nelder-Mead algorithm, a downhill search

method, to find the parameter values that minimize the RSS
function.

3. The minimizing parameters are the LS parameter estimates.

4. It is helpful to note that the LS method relaxes the restrictive

assumptions on the residuals; LS parameter estimates converge to
the true values even if the noise is non-normal and autocorrelated,
as long as the noise has a stationary distribution.

C.6 IMPLEMENTATION IN CODE

C.6.1 Basic Structure of Program
A code to estimate parameters requires three basic parts:

1. The main program, which sets the initial parameter vector

“guess,”then passes this initial vector and the name of the
RSS function to the downhill search algorithm (which attempts
to converge on the minimizers using this initial “guess”) and,
finally, recovers the parameter estimate vector from the downhill
algorithm.

2. The downhill search program. This is not something you need

to code yourself; it is likely in the library of functions of the
language you are using. For example, in MATLAB the downhill
search function is called fminsearch. When the main code passes
the initial parameter vector and the name of the RSS routine to
this function, it uses those initial parameters to compute RSS.
Then it “walks downhill” until it finds a local minimum value
of RSS. Finally, it returns the minimizer vector of parameters to
the main code.

3. The subroutine that computes the RSS as a function of the

parameters. This is where the model is coded, predictions and
residuals are computed, and the RSS value is computed for that
parameter vector.

Note: Most downhill search functions search for negative as well

as positive minimizers. In population models, however, parameters are
 Appendix C ■ 299

generally positive and we do not want to search negative parameter

space. One way to address this is to pass the initial parameter vector
v to the downhill search routine as ln(v) so that it can search all of
parameter space. In the RSS routine, the ln(v) must be converted back
via exponentiation before the parameters are used to compute model
predictions. Also, when the final parameter estimates are returned to
the main program, they must be exponentiated to recover the best
parameters on the correct scale.

C.6.2 Constructing Input Files

To construct input files:
1. Prepare an MS Excel data file with the appropriate columns.

2. Copy the contents of the Excel file and paste them into a Notepad
file. If you keep the column headings, the headings row must be
commented out with an initial %.

3. Save the Notepad file in your program folder as inputname.txt.

C.6.3 Example: Algebraic Model Using Vectors

The goal in this example is to fit the Gompertz model

f (x) = K exp −ea−bt

to the humerus estimation data in Data Set 2.4 (see Chapter 2) using
Kb = 11.96923077. Assume environmental noise. We want to find the
best-fit parameters a and b.

%Program gompertz.m
%Parameterizes the Gompertz model
global age hum

%Load data
load estimation.txt;
data = estimation;
age = data(:,1);
hum = data(:,2);
300 ■ Appendix C

%Initialize parameters
theta = log([0.01 0.03]’);

%Run Nelder-Mead algorithm and output best parameters

to screen
[output RSSbest] = fminsearch(’RSSgom’,theta);
params = exp(output);
a = params(1)
b = params(2)

%Program RSSgom.m
function RSS = RSSgom(theta)
global age hum

%Compute vector of predictions

a = exp(theta(1));
b = exp(theta(2));
K = 11.96923077;
pred = K*exp(-exp(a-b*age));

%Compute RSS on log scale for environmental noise

residuals = log(hum) - log(pred);
squareresiduals = residuals.^2;
RSS = sum(squareresiduals);
end

C.6.4 Example: Algebraic Model Using Loop

The goal in this example is to fit the piecewise-linear model

a (x − b) + K x < b
f (x) =
K x≥b

to the humerus estimation data in Data Set 2.4 (see Chapter 2) using
Kb = 11.96923077. Assume demographic noise. We want to find the
best-fit parameters a and b.
 Appendix C ■ 301

%Program lin.m
%Parameterizes the linear model
global age hum

%Load data
load estimation.txt;
data = estimation;
age = data(:,1);
hum = data(:,2);

%Initialize parameters
theta = log([0.1 15]’);

%Run Nelder-Mead algorithm and output best parameters

to screen
[output RSSbest] = fminsearch(’RSSlin’,theta);
params = exp(output);
a = params(1)
b = params(2)

%Program RSSlin.m
function RSS = RSSlin(theta)
global age hum

%Compute vector of predictions

a = exp(theta(1));
b = exp(theta(2));
K = 11.96923077;
n = length(age);

for i = 1 : n
if age(i) < b
pred(i) = a*(age(i)-b) + K;
else
302 ■ Appendix C

pred(i) = K;
end
end
pred = pred’;

%Compute RSS on sqrt scale for demographic noise

residuals = sqrt(hum) - sqrt(pred);
squareresiduals = residuals.^2;
RSS = sum(squareresiduals);

end

C.6.5 Example: Scalar Map with One-Step Predictions

The goal is to fit the Ricker model

Nt+1 = aNt e−bNt

to the data in Data Set C.1. That is, we want to find the best-fit
parameters a and b.

%Program ricker.m
%Parameterizes the Ricker model
global obst obsN

%load data
load RickerInput.txt;
data = RickerInput;
obst = data(:,1);
obsN = data(:,2);

%Initialize a and b
theta = log([2 0.03]’);
 Appendix C ■ 303

%Run Nelder-Mead algorithm and output best parameters

to screen
[output RSSbest] = fminsearch(’RSSfun’,theta);
params = exp(output);
a = params(1)
b = params(2)

%Function RSSfun.m
function RSS = RSSfun(theta)
global obst obsN

%Compute vector of one-step predictions

a = exp(theta(1));
b = exp(theta(2));
predN = a*obsN .*exp(−b*obsN);

%Compute RSS
numobs = length(obsN);
obsNahead = obsN(2:numobs);
predNlag = predN(1:numobs-1);
residuals = obsNahead - predNlag;
squareresiduals = residuals.^2;
RSS = sum(squareresiduals);

end

C.6.6 Example: Higher-Dimensional Discrete-Time Model with

One-Step Predictions
Here we fit the environmental noise LPA model to the data
in Data Set 6.1.

%Program LPA.m
%Parameterizes the LPA model
global Ldata Pdata Adata
304 ■ Appendix C

%Load data: user sets this block

load DataSet61.txt;
data = DataSet61;
reps = 4;

%Loop through replicates and construct separate L, P, A

data matrices
Ldata = [];
Pdata = [];
Adata = [];

for i = 1 : reps
Ldata = [Ldata data(:,(i-1)*3+1)];
Pdata = [Pdata data(:,(i-1)*3+2)];
Adata = [Adata data(:,(i-1)*3+3)];
end

%Initialize parameters
theta = log([
8
0.01
0.01
0.2
0.01
0.2
]);

[output RSSbest] = fminsearch(’RSSlpa’,theta);

params = exp(output)

%Program RSSlpa.m
function RSS = RSSlpa(theta)
global Ldata Pdata Adata

%Set parameters
par = exp(theta);
b = par(1);
cel = par(2);
cea = par(3);
 Appendix C ■ 305

mul = par(4);
cpa = par(5);
mua = par(6);

n = length(Ldata(:,1));

%Compute one-step prediction and residual matrices

Lpred = b*Adata.*exp(-cel*Ldata-cea*Adata);
Ppred = (1-mul)*Ldata;
Apred = Pdata.*exp(-cpa*Adata) + (1-mua)*Adata;

Lp = Lpred(1:n-1,:);
Pp = Ppred(1:n-1,:);
Ap = Apred(1:n-1,:);

Lo = Ldata(2:n,:);
Po = Pdata(2:n,:);
Ao = Adata(2:n,:);

Lres = log(Lo) - log(Lp);

Pres = log(Po) - log(Pp);
Ares = log(Ao) - log(Ap);

RSSL = sum(sum(Lres.^2));
RSSP = sum(sum(Pres.^2));
RSSA = sum(sum(Ares.^2));

RSS = RSSL + RSSP + RSSA;

end
Index

abiotic factor, 11 biology, 9

Akaike Information Criterion, birth rate
8, 43, 122, 221, 253 per capita, 57
Akaike weight, 253
Allee effect, 164, 181 cannibalism, 113, 247, 249, 256
Allee threshold, 165 flour beetles, 113
Ansatz, 93, 189 gulls, 247
art of modeling, 6, 114 carrying capacity, 163
assumptions Caswell, H., 86, 109
deterministic, 33 causation
model, 32 proximate, 12
stochastic, 33 ultimate, 12
Atkins, G. J., 256 vs. correlation, 12, 241
attractor, 77, 79 center, 193, 195
autonomous, 159, 187, 202 chaos, 79, 125
axiom, 20 characteristic equation, 97, 98,
190, 272
Beetle Team, 111, 125, 216 characteristic polynomial,
Beverton-Holt model, 73, 269, 272
82, 156 climate change, 247, 256
bifurcation, 78, 202, 203 closed-form solution, 58
blue-sky, 184 community, 11
diagram, 61, 76, 79, compartmental model, 56, 146
81, 160 continuous-time, 146
Hopf, 211 discrete-time, 56
hysteresis, 178, 179 competition, 86, 205
hysteresis loop, 179 computer
parameter, 61 code, 41
pitchfork, 177 program, 41
point, 61 conditioned least squares, 119,
saddle-node, 177 120, 220
transcritical, 176 confidence interval, 240, 242
vertical, 160 consistent
binning, 36 logically, 20

307
308 ■ Index

continuously differentiable, 67, nonlinear systems, 99

103, 172, 202 differential equations, 4, 145,
contradiction 159, 187
logical, 20, 27 linear scalar, 159
cooperation, 86, 205 linear systems, 188
correlation nonlinear scalar, 162
vs. causation, 12, 241 nonlinear systems, 198
Costantino, R. F., 111, 124, separable, 151
125, 127, 128 dimension, 86
cross-sectional study, 31 displacement from equilibrium,
crowding effects, 60, 162 66, 101, 182, 199
Cushing, J. M., 111, 114, 124, distribution, 17
159, 166, 168, 211 normal, 17
cycle chain, 202 stable age, 99, 106
Doomsday Model, 166
Darling, Fraser, 256
dot product, 266
data
doubling time, 71, 166
estimation, 36
downhill method, 40
validation, 36
dynamical system, 4
day of year, 216
dynamics
deduction, 19
steady state, 226
Dennis, Brian, 111
transient, 226
density dependence, 162
derivative, 13
partial, 85, 187 ecology, 10
Desharnais, R. A., 111, 125 behavioral, 12
determinant, 267 community, 11
2 × 2, 93 ecosystem, 11
n × n, 98 mathematical, 12
deterministic, 13, 79 physiological, 11
skeleton, 34, 117 population, 11
unpredictability, 80 ecosystem, 11
diagram eigensolution, 94, 190
bifurcation, 61, 76, 79, 160 eigenvalue, 59, 65, 66, 94, 97,
Leslie, 87, 90, 116 98, 172, 190, 269,
Dickens, Charles, 169 271, 272
differential equation, 4, 55, 85 complex, 194
linear scalar, 57 dominant, 99, 105
linear systems, 90 imaginary, 194
nonlinear scalar, 60 eigenvalue problem, 274
 Index ■ 309

eigenvector, 94, 97, 98, 190, equation, 14

269, 272 flour beetle, 23, 86, 99,
eight-cycle, 78 111, 112
El Nino-Southern four-cycle, 78
Oscillation, 250 fractal, 83
emergent properties, 10 Fraser Darling effect, 256
endemic disease, 150 function, 13
epidemic, 148 probability density, 17
epidemiology, 9 Fundamental Theorem of
epistemology, 22 Demography, 98
equation
autonomous, 159, 187, 202 Galusha, J. G., 215
difference, 4 Gaussian
differential, 4, 145 noise, 117
equilibrium, 14, 92 variable, 40, 44, 50, 229,
fixed-point, 14, 92 239, 297
Van der Pol, 204 general solution, 153
equilibrium, 14, 15, 59, 64, 92 generalized linear model, 242
asymptotically stable, 15 geometric growth, 58
hyperbolic, 67, 103, glaucous-winged gull, 30,
171, 201 216, 248
neutrally stable, 15 Gleick, James, 75
stable, 15 Gompertz model, 299
unstable, 15 goodness-of-fit, 8, 45, 124, 296
error growth, 30
measurement, 16, 32 bone, 29
observational, 16 exponential, 24
process, 16, 32 geometric, 58
residual, 35 Malthusian, 24, 57, 58, 161
Euler’s formula, 195 growth rate
exchange of stability, 176 intrinsic, 59
expected value, 18 per capita, 16, 164
exponential growth, 24 population, 16
gull
factor, 237 glaucous-winged, 30,
abiotic, 11 216, 248
fecundity, 57, 90
final state, 79 Haldane, J. B. S., 10
Fisher, Ronald, 10 half-saturation constant, 47
fixed point, 14, 92 Hardy, G. H., 10
310 ■ Index

harmonic oscillator, 193 King, A. A., 111

Hartman-Grobman KISS principle, 6, 114
Theorem, 202
Hastings, A., 124 Larus glaucescens, 30, 216, 248
heteroclinic cycle, 202 least squares, 297
histogram, 19 conditioned, 120, 121
Holling Type III, 47 nonlinear, 39
Hopf bifurcation, 211 Leslie
humanities, 21 diagram, 87, 90, 116
Hunt for Chaos matrix, 89, 91
experiments, 125 model, 86
hyperbolic, 67, 103, 171, 201 model, nonlinear, 99, 116
hypotheses Leslie, P. H., 86
alternative, 7, 220 Levin, S. A., 28, 228
hypothesis, 7 life-cycle stage, 86
hysteresis, 178, 179 likelihood function, 39, 123,
hysteresis loop, 179 297
limit cycle, 202
induction, 19 linearization
mathematical, 20, 26 importance of, 63
infectives, 147 of functions, 62, 63
information theory, 8, 43 of scalar difference
initial condition, 58, 153 equations, 64, 66
sensitivity to, 79 of scalar differential
initial value, 58 equations, 170
initial value problem, 58 of systems of difference
interaction equations, 100, 101
competition, 86 of systems of differential
cooperation, 86 equations, 201
interspecific, 86 Linearization Theorem, 67, 76,
intraspecific, 86 103, 172, 202
mutualism, 86 link function, 242
predator-prey, 86 loafing, 216
interaction terms log odds, 240
(regression), 242 log scale, 35
intrinsic growth rate, 59, 163 log-likelihood, 39
isocline, 204 log-likelihood function, 297
logistic map, 73, 82
Jacobian matrix, 102, 199 logistic model, 14, 24, 150, 156,
Janoschek curve, 34 163, 172, 291
 Index ■ 311

logistic model, (cont.) display current value of

discrete-time, 155 variable, 274
logistic regression, 247 fminsearch, 221, 292
longitudinal study, 31 functions, 281, 285
Lotka, Alfred J., 23 glmfit, 240, 257, 258
Lotka-Volterra model, 204 input files, 280
competition, 205 logical statements, 279
cooperation, 205 logistic model, 290
predator-prey, 206 loops, 278
LPA model, 23, 99, 116, 128 matrices, 283
demographic noise, 117 minimizer, 292
environmental noise, 117 Nelder-Mead
algorithm, 292
numerical format, 278
MacArthur, Robert H., 228
ode45, 221, 290
Malthus, Thomas, 161
ODEs, 290
Malthusian growth, 24, 57,
output files, 280
58, 161
plot command, 286
manifold, 191
printing to screen, 277
stable, 192
programs, 276
unstable, 191
Ricker model, 286
map, 4, 55
sample codes, 295
linear, 57 script, 276
nonlinear, 60 spline, 229
linearization, 64 subroutines, 282
logistic, 73, 82, 155 variables, 274
Ricker, 67, 75 vectors, 283
mathematical induction, 20, 26 matrices
mathematics noncommutativity of, 267
pure, 20 matrix, 265
Matlab addition, 265
arithmetic operations, 275 determinant, 267
case sensitivity, 274 identity, 268
clearing variables, 274 Jacobian, 102, 199
closing, 274 Leslie, 89, 91
code, 276 multiplication, 266
command line, 273 multiplication by
comment lines, 277 scalar, 265
crashing a program, 279 projection, 89
discrete-time model, 286 subtraction, 266
312 ■ Index

maximizer, 39 Janoschek, 34
maximum likelihood, 38, 296 Leslie, 86
parameters, 39 linear, 89
May’s Hypothesis, 81 logistic, 14, 24, 150, 156,
May, Robert, 81, 125, 206 163, 172, 291
mean, 19 Lotka-Volterra, 204
mechanism LPA, 23, 99, 116, 128
deterministic, 33 mathematical, 4
stochastic, 33 mechanistic, 6
metamorphosis, 86, 112 NLAR, 117
minimizer, 40 nonlinear, 89
model nonlinear Leslie, 99
fitting, 7, 8, 36, 295 regression, 237
parameterization, 119, 220 Ricker, 24, 61, 75, 82,
predictions, 8, 222 83, 135
selection, 8, 42, 122, 221 SIR, 147
validation, 8, 36, 45, 124, stage-structured, 86
222 stochastic, 34, 35
validation in logistic model-averaged parameter
regression, 253 estimates, 253
model modeling
age-structured, 86 art of, 6, 114
algebraic, 4 ecological, 10
assumptions, 6 multiple time scale
Beverton-Holt, 73, 82, 156 analysis, 226
compartmental, 56, 146, mutualism, 86
150, 218
continuous-time, 4 natural selection, 12
demographic noise Nelder-Mead algorithm, 40,
LPA, 119 300
deterministic, 33 net reproductive number, 107
difference equation, 4 inherent, 131
differential equation, 4 neutrally stable, 15, 195
discrete logistic, 82 NLAR model, 117
discrete-time, 4, 85 NOAA, 218
discrete-time logistic, 155 node
environmental noise asymptotically stable,
LPA, 118 191, 192
generalized linear, 242 unstable, 189, 191
Gompertz, 299 noise, 6, 16, 114
 Index ■ 313

noise, (cont.) phase plane, 190

demographic, 16, 296 phase plane portrait
environmental, 17, 296 asymptotically stable
Gaussian, 117, 239, 297 node, 191
nonlinear autoregressive asymptotically stable
model, 117 spiral, 196
nonlinear least squares, 39 center, 193
normal distribution, 17 saddle, 192
nullcline, 204 six generic, 197
unstable node, 189
Occam’s Razor, 27 unstable spiral, 195
odds, 240 phase space, 13
odds ratio, 241 Poincare-Bendixson
ODE Theorem, 204
linear system, 188 population, 11
nonlinear system, 198 predator-prey, 86, 206
orbit, 13 predictability, 80
organism, 10 prediction, 21
bones, 30 one-step, 220, 227
development, 30 probability density function, 17
growth, 30 projection matrix, 89
organs, 30 proportional to, 15
shape, 30 inversely, 15
size, 30 Protection Island, 30, 208,
outbreak, 149 216, 252
proximate cause, 12
paradox, 27
parameter, 6, 295 R-squared, 45, 124, 221, 296
bifurcation, 61 random variable
estimation, 7, 36, 39, 295 continuous, 19
least squares, 39, 296, 297 discrete, 19
maximum likelihood, 39, realization, 35
296, 297 standard normal, 35,
model-averaged, 253 117, 239
space, 40 rate
parameterization, 7, 36, 295 vital, 86
partial fractions, 151 rate of change, 13
per capita growth rate, 164 realization, 17
period-doubling cascade, 79 recovereds, 147
phase line portrait, 170 recruitment, 60
314 ■ Index

recursion formula, 4, 72 signal, 16

regression sink, 170
coefficients, 238 SIR model, 147
gamma, 242 solar elevation, 216, 217
interaction terms, 242 solution
intercept, 238 closed-form, 58
linear, 237, 242 general, 95, 97, 98, 153,
logistic, 240, 242, 247, 253 160
multiple, 238 particular, 95, 97, 98, 153,
Poisson, 242 160
slope, 238 singular, 153
residual, 35, 119, 221, 296 source, 170
log, 119 species, 11
one-step, 119, 221 social, 10
square-root, 35, 119, 120 solitary, 10
residual sum of squares, 39, spiral
120, 221, 296 asymptotically stable,
Ricker model, 24, 61, 67, 75, 196, 197
82, 83, 135, 287 unstable, 195, 196
Ricker nonlinearity, 70, 89, 131
spline, 221
RSS, 39, 120, 221, 296
square root scale, 35
saddle, 192, 193 stable, 15
sampling asymptotically, 15
stratified random, 36 neutrally, 15
saturation, 33 stable age distribution, 99
scale, 9, 10, 227, 228 stage-structured model, 86
community, 11 standard deviation, 18
ecosystem, 11 standard error, 240
organism, 10 standard normal random
population, 11 variable, 35
species, 11 state space, 13
science, 20 steady state, 226
sea surface temperature, 250 stochasticity, 13, 16, 80
seabird behavior, 215 demographic, 16, 17,
Seabird Ecology Team, 215, 35, 117
228, 247 environmental, 16, 17,
sensitivity to initial 35, 117
conditions, 79 Gaussian, 117
sigmoidal growth, 33 stratified random sampling, 36
 Index ■ 315

Strong Ergodic Property, Van der Pol equation, 204, 210

99, 105 variable
survivorship, 60 binary, 240
susceptibles, 147 continuous random, 19
synchrony, 256 coupled, 13
dependent, 5, 295
tangent plane, 100, 199 discrete random, 19
Tenebrionidae, 111 explanatory, 237
thermocline, 250 extrinsic, 6
tide, 217 independent, 5, 295
semi-diurnal, 217 intrinsic, 6
height, 216 predictor, 237
time series, 13, 76 random, 17
transformation realization of random, 17
variance-stabilizing, 35, response, 237
117, 296 standard normal random,
transient, 82, 227 117, 239
trend state, 5
deterministic, 16 variance, 19, 40
Tribolium, 23, 86, 111 variance-stabilizing
adult, 112 transformation, 35,
castaneum, 99, 111 117, 296
egg, 112 variation equation, 66, 182
larva, 112 vital rates, 86
pupa, 112
two-cycle, 78 Weldon, L. M., 247
Wright, Sewell, 10
ultimate cause, 12
unpredictability, 80
unstable, 15
 Taylor & Francis eBooks
www.taylorfrancis.com

A single destination for eBooks from Taylor & Francis

with increased functionality and an improved user
experience to meet the needs of our customers.

90,000+ eBooks of award-winning academic content in

Humanities, Social Science, Science, Technology, Engineering,
and Medical written by a global network of editors and authors.

TAYLOR & FRANCIS EBOOKS OFFERS:

Improved
A streamlined A single point search and
experience for of discovery discovery of
our library for all of our content at both
customers eBook content book and
chapter level

REQUEST A FREE TRIAL

[email protected]