College of Natural and Computational Sciences

Department of Mathematics

Mathematical Modeling(Math445)
Module
Writers: Dereje Tigabu (M.sc)
Tesfaye Sama (M.sc)
Content Editor: Doyo Kereyu (M.sc)
Language Editor:
Pedagogy Editor:

July 4, 2019

Haramaya, Ethiopia
 TABLE OF CONTENTS

PREFACE v

LIST OF FIGURES vi

LIST OF TABLES x

MODULE INTRODUCTION xi

1 INTRODUCTION TO MODELING 1

1.1 Models and Reality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Principles of Mathematical Modeling . . . . . . . . . . . . . . . . . . . . . 5
1.3 Some Methods of Mathematical Modeling . . . . . . . . . . . . . . . . . . 8

1.3.1 Dimensional Homogeneity and Consistency . . . . . . . . . . . . . . 8

1.3.2 Abstraction and Scaling . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.3 Conservation and Balance Principles . . . . . . . . . . . . . . . . . 9

1.3.4 Constructing Linear Models . . . . . . . . . . . . . . . . . . . . . . 11

i
 TABLE OF CONTENTS

1.4 Building a Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.5 Examples of Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.5.1 Modeling with Difference Equations . . . . . . . . . . . . . . . . . . 13

1.5.2 Modeling with Ordinary Differential Equations . . . . . . . . . . . . 14
1.5.3 Modeling with Partial Differential Equation . . . . . . . . . . . . . 15

1.5.4 Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5.5 Modeling with Simulations . . . . . . . . . . . . . . . . . . . . . . . 16
1.5.6 Function Fitting: Data Modeling . . . . . . . . . . . . . . . . . . . 16

1.6 Why Study Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.7 Classifications of Mathematical Modeling . . . . . . . . . . . . . . . . . . . 17

2 ARGUMENTS FROM SCALE 20

2.1 Effects of Size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.2 Dimensional Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2.1 Dimensions and Units . . . . . . . . . . . . . . . . . . . . . . . . . 30

2.2.2 Dimensions of Common Physical Quantities . . . . . . . . . . . . . 31

2.2.3 Dimensional Homogeneity . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.4 Nondimensionalization of Equations . . . . . . . . . . . . . . . . . . 36

2.2.5 Dimensional Analysis and Similarity . . . . . . . . . . . . . . . . . 41

2.2.6 The Buckingham Pi-Theorem and Method of Repeating Variables . 46

3 GRAPHICAL METHODS 55

3.1 Using Graphs in Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.2 Comparative Statics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3 Stability Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

ii
 TABLE OF CONTENTS

3.4 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.5 Equilibrium Points and Stability Analysis . . . . . . . . . . . . . . . . . . 83
3.5.1 Equilibrium Points . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.5.2 Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

4 APPLICATION OF MATHEMATICAL MODELING 98

4.1 Continuous Population Model for Single Species . . . . . . . . . . . . . . . 99

4.1.1 Continuous Growth Models . . . . . . . . . . . . . . . . . . . . . . 99

4.1.2 The Logistic Population Model . . . . . . . . . . . . . . . . . . . . 103
4.1.3 Qualitative Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 109

4.2 Discrete Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

4.2.1 Newton’s Law of Cooling . . . . . . . . . . . . . . . . . . . . . . . . 118

4.2.2 Bank Account Problem . . . . . . . . . . . . . . . . . . . . . . . . 119

4.2.3 Drug Delivery Problem . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.2.4 Economic Model (Harrod Model) . . . . . . . . . . . . . . . . . . . 121

4.2.5 Arms Race Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

4.2.6 Discrete Population Models for a Single Species . . . . . . . . . . . 123
4.3 Continuous Models for Interacting Populations . . . . . . . . . . . . . . . . 132

4.3.1 Predator-Prey Models: Lotka-Volterra Systems . . . . . . . . . . . 132

4.3.2 Competition Models: Principle of Competitive Exclusion . . . . . . 137
4.3.3 Mutualism or Symbiosis . . . . . . . . . . . . . . . . . . . . . . . . 144

4.4 Mathematical Epidemiology . . . . . . . . . . . . . . . . . . . . . . . . . . 146

4.5 Mathematical Modeling of Infectious Diseases: Issues and Approaches . . . 147
4.5.1 Deterministic Epidemic Models:
Compartmental Approach . . . . . . . . . . . . . . . . . . . . . . . 152

iii
 TABLE OF CONTENTS

4.5.2 SI Model with out Demography . . . . . . . . . . . . . . . . . . . . 155

4.5.3 SI Model with Demography . . . . . . . . . . . . . . . . . . . . . . 156
4.5.4 SIR Model with out Demography . . . . . . . . . . . . . . . . . . . 157

4.5.5 SIR Model with Demography . . . . . . . . . . . . . . . . . . . . . 158

4.5.6 SIS Model with out Demography . . . . . . . . . . . . . . . . . . . 159

4.5.7 SIS Model with Demography . . . . . . . . . . . . . . . . . . . . . . 160

4.5.8 SEIR Model with out Demography . . . . . . . . . . . . . . . . . . 161
4.5.9 SEIR Model with Demography . . . . . . . . . . . . . . . . . . . . 161

REFERENCES 162

iv
 PREFACE

This module is designed for Mathematics undergraduate students taking the course math-
ematical modeling. The target groups are mainly mathematics students in the undergrad-
uate program at the department of Mathematics, Haramaya University. This material is
prepared so as to address the problem of shortage of reference and/or text books, to some
extent. It is believed that this module will be a valuable resource for introducing mathe-
matical modeling and motivating students to do research in applied mathematics in real
life problems. The module is organized into four chapters. There are exercises that may
help the learners to understand basic techniques in Modeling.

v
 LIST OF FIGURES

1.1 A first-order view of mathematical modeling that shows how the questions
asked in a principled approach to building a model relate to the develop-
ment of that model (inspired by Carson and Cobelli, 2001). . . . . . . . . . 6
1.2 A system boundary surrounding the object or system being modeled. The
influx qin (t), efflux qout (t), generation g(t), and consumption c(t), affect
the rate at which the property of interest, Q(t), accumulates within the
boundary (after Cha, Rosenberg, and Dym, 2000). . . . . . . . . . . . . . . 10

2.1 (a) Top view. (b) Cross section of center. I = length; b = beam; A =
cross-sectional area. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.2 Object falling in a vacuum. Vertical velocity is drawn positively, so w < 0

for a falling object. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.3 In a typical fluid flow problem, the scaling parameters usually include a
characteristic length L, a characteristic velocity V , and a reference pressure
difference P0 − P ∞. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.4 Kinematic similarity is achieved when, at all locations, the velocity in the
model flow is proportional to that at corresponding locations in the proto-
type flow, and points in the same direction. . . . . . . . . . . . . . . . . . 43

vi
 LIST OF FIGURES

2.5 The classical pendulum oscillating through angles θ due to gravitational

acceleration g. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3.1 Country 1 introduces ABMs. A = initial status (shaded area stable) ; B =

country I protects its missiles; C = country I protects its cities. Axes show
number of missiles. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2 Country 1 introduces MIRVs. Axes show number of missiles. . . . . . . . . 62

3.3 Migration .and extinction curves for islands. (a) Typical curves. (b) Effect
of distance. (c) Effect of size. . . . . . . . . . . . . . . . . . . . . . . . . . 66

3.4 Marginal cost and income curves. Axes show quantity produced per unit
time and dollars per unit time. . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.5 Supply and demand curves. Increased marginal costs shift supply curve
upward to dashed position. . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.6 The cobweb model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.7 Dynamics in the activity-friendliness plane. . . . . . . . . . . . . . . . . . . 76

3.8 Time series plot of dx/dt = x. . . . . . . . . . . . . . . . . . . . . . . . . . 86

3.9 Time series plot of dx/dt = 2 − x. . . . . . . . . . . . . . . . . . . . . . . . 86
3.10 Time series plot of dx/dt = x2 − 5x + 6. . . . . . . . . . . . . . . . . . . . 87

3.11 The solution curves of the planar system at equilibrium point (0, 0) has
stabe node. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

3.12 The solution curves of the planar system at equilibrium point (0, 0) has
unstable focus. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

4.1 Solution of the logistic equation. . . . . . . . . . . . . . . . . . . . . . . . . 106

4.2 Solution curves of the logistic equation. . . . . . . . . . . . . . . . . . . . . 110
4.3 Closed (u, v) phase plane trajectories, from (4.26) with various H, for the
LotkaVolterra system (4.24): H1 = 2.1, H2 = 2.4, H3 = 3.0, H4 = 4. The
arrows denote the direction of change with increasing time τ . . . . . . . . 134

vii
 LIST OF FIGURES

4.4 Periodic solutions for the prey u(τ ) and the predator v(τ ) for the Lotka-
Volterra system (4.24) with α = 1 and initial conditions u(0) = 1.25, v(0) =
0.66. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

4.5 The null clines for the competition model (4.33). f1 = 0 is u1 = 0 and
1 − u1 − a12 u2 = 0 with f2 = 0 being u2 = 0 and 1 − u2 − a21 u1 = 0. The
intersection of the two solid lines gives the positive steady state if it exists
as in (a) and (b): the relative sizes of a12 and a21 as compared with 1 for
it to exist are obvious from (a) to (d). . . . . . . . . . . . . . . . . . . . . 139
4.6 Schematic phase trajectories near the steady states for the dynamic be-
haviour of competing populations satisfying the model (4.33) for the various
cases. (a) a12 < 1, a21 < 1. Only the positive steady state S is stable and
all trajectories tend to it. (b) a12 > 1, a21 > 1. Here, (1, 0) and (0, 1) are
stable steady states, each of which has a domain of attraction separated by
a separatrix which passes through (u∗1 , u∗2 ). (c) a12 < 1, a21 > 1. Only one
stable steady state exists, u∗1 = 1, u∗2 = 0 with the whole positive quadrant
its domain of attraction. (d) a12 > 1, a21 < 1. The only stable steady state
is u∗1 = 0, u∗2 = 1 with the positive quadrant as its domain of attraction.
Cases (b) to (d) illustrate the competitive exclusion principle whereby 2
species competing for the same limited resource cannot in general coexist. . 142
4.7 Phase trajectories for the mutualism model for two species with limited
carrying capacities given by the dimensionless system (4.40). (a) a12 a21 >
1: unbounded growth occurs with u1 → ∞ and u2 → ∞ in the domain
bounded by the null clinesthe solid lines. (b) a12 a21 < 1: all trajectories
tend to a positive steady state S with u∗1 > 1, u∗ > 1 which shows the
initial benefit that accrues since the carrying capacity for each species is
greater than if no interaction were present. . . . . . . . . . . . . . . . . . . 146

4.8 Transfer diagram for an SIR compartment model . . . . . . . . . . . . . . 152

4.9 Flow chart of SI model with out birth and death rates. . . . . . . . . . . . 155
4.10 Flow chart of SI model with birth and death rates. . . . . . . . . . . . . . 156

4.11 Flow chart of SIR model with out birth and death rates. . . . . . . . . . . 157

viii
 LIST OF FIGURES

4.12 Flow chart of SIR model with birth and death rates. . . . . . . . . . . . . 158
4.13 Flow chart of SIS model with out birth and death rates. . . . . . . . . . . 159
4.14 Flow chart of SIS model with birth and death rates. . . . . . . . . . . . . . 160

4.15 Flow chart of SEIR model with out birth and death rates. . . . . . . . . . 161
4.16 Flow chart of SEIR model with birth and death rates. . . . . . . . . . . . . 162

ix
 LIST OF TABLES

2.1 Times of racing crews in four meets . . . . . . . . . . . . . . . . . . . . . . 26

2.2 Shell design parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.3 Derived physical quantities with their units . . . . . . . . . . . . . . . . . . 32
2.4 Some common physical quantites and their dimensions . . . . . . . . . . . 32

x
 MODULE INTRODUCTION

Dear summer students, this is your distance material for the course Mathematical Model-
ing which consists of four chapters. Chapter one, the introductory unit, consists of models
and reality, properties of models, building a model, examples of models, and why study
modeling. In the second chapter, you see arguments from scale which consists effects of
size and dimensional analysis. The third chapter deals with graphical methods of mathe-
matical modeling. The last chapter; chapter four deals with application of mathematical
modeling in real life and illustrative examples of mathematical modeling.

The following are the objectives that you should aspire to achieve in studying this module.

Objectives:

After completing this module, you will be able to:

¡ understand the importance of model,

¡ understand properties of models,

¡ give examples of models,

¡ understand dimensional analysis,

xi
 MODULE INTRODUCTION

¡ use graph in modeling,

¡ develop mathematical models representing a physical problem,

¡ test stability of the model,

¡ use models for forecasting.

To check your understandings, please try to practice exercises that are given at the end of
each chapter. We hope that after completing the module you will grasp better knowledge
about mathematical modeling.

We hope that you will enjoy studying this module and work hard to make the best out
of your study. Finally, we would like to advise you to spend sufficient time to study the
module.

xii
CHAPTER 1

INTRODUCTION TO MODELING

Objectives:

By the end of this chapter you will be able to:

• define and list the basic idea of mathematical modeling.

• list down and clarify steps to build mathematical modeling.

• formulate some problems using mathematical modeling.

• identify all procedure to solve problems through mathematical modeling.

• list classification of mathematical modeling.

• know characteristics of mathematical modeling.

Let’s begin with a dictionary definition of the word model:

Model (n): A miniature representation of something; a pattern of something to be

made; an example for imitation or emulation; a description or analogy used to help visu-
alize something (e.g., an atom) that cannot be directly observed; a system of postulates,

1
 1.1. MODELS AND REALITY

data and inferences presented as a mathematical description of an entity or state of affairs.

This definition suggests that modeling is a cognitive activity in which we think about and
make models to describe how devices or objects of interest behave.

There are many ways in which devices and behaviors can be described. We can use words,
drawings or sketches, physical models, computer programs, or mathematical formulas. In
other words, the modeling activity can be done in several languages, often simultaneously.
Since we are particularly interested in using the language of mathematics to make models,
the definition given above is refined as follows:

Definition 1.1. A mathematical model is an abstract, simplified, mathematical construct

related to a part of reality and created for a particular purpose.

1.1 Models and Reality

The theoretical and scientific study of a situation centers around a model,that is, some-
thing that mimics relevant features of the situation being studied. For example, a road
map, a geological map, and a plant collection are all models that mimic different aspects
of a portion of the earth’s surface. The ultimate test of a model is how well it performs
when it is applied to the problems it was designed to handle. (You cannot reasonably
criticize a geological map if a major highway is not marked on it; however, this would be
a serious deficiency in a road map.) When a model is used, it may lead to incorrect pre-
dictions. The model is often modified, frequently discarded, and sometimes used anyway
because it is better than nothing. This is the way science develops.
Mathematics and physical science each had important effects on the development of the
other. Mathematics is starting to play a greater role in the development of the life and
social sciences, and these sciences are starting to influence the development of mathemat-
ics. This sort of interaction is extremely important if the proper mathematical tools are
going to be developed for the various sciences. S. Bochner (1966) discusses the hand-
in hand development of mathematics and physical science. Some people feel that there

2
 1.1. MODELS AND REALITY

is something deeper going on than simply an interaction leading to the formulation of

appropriate mathematical and physical concepts.
As far as a model is concerned, the world can be divided into three parts :

1. Things whose effects are neglected.

2. Things that affect the model but the model is not designed to study its behavior.

3. Things that the model is designed to study the behavior of.

The model completely ignores item 1 . The constants, functions, and so on, that appear
in item 2 are external and are referred to as exogenous variables (also called pa-
rameters, input, or independent variables). The things the model seeks to explain
are endogenous variables (also called output or dependent variables). The exogenous-
endogenous terminology is used in some areas of modeling. The input-output termi-
nology is used in areas of modeling where the model is viewed as a box into which we
feed information and obtain information from. The parameter-independent -dependent
terminology is the standard mathematical usage.
Example

Suppose we are hired by a firm to determine what the level of production should be
to maximize profits. We would construct a model that enables us to express profits
(the dependent variable) in terms of the level of production, the market situation, and
whatever else we think is relevant (the independent variables). Next, we would measure all
the independent variables except the level of production and use the model to determine
which value of the level of production gives the greatest profit.

Now, let’s look at things from the point of view of an economist who is seeking to explain
the amount of goods firms produce. A two-part model could be constructed: Firms seek
to maximize profits, and profits can be determined as sketched in the previous paragraph.
In this model profits become an internal variable (of no interest except for the machina-
tions of the model), and level of production changes from an independent to a dependent
variable.

3
 1.1. MODELS AND REALITY

These three categories (neglected, input, and output) are important in modeling. If the
wrong things are neglected, the model will be not good. If too much is taken into consid-
eration, the resulting model will be hopelessly complex and probably require incredible
amounts of data. Sometimes, in desperation a modeler neglects things not because he
thinks they are unimportant, but because he cannot handle them and hopes that neglect-
ing them will not invalidate the conclusions.
Proper choice of dependent variables (i.e., output) is essential ; we must seek to explain
the things we can explain. Often this choice is relatively clear, as in the example involving
the economist who wished to explain the level of production of a firm. Sometimes we need
to be careful ; for example, we could explain profits in terms of level of production, but
not conversely as we might naively try to do, since we were asked to determine the best
level of production.
Since different models make different types of simplifying assumptions, there is usually no
single best model for describing a situation. R. Levins ( 1968, p . 7 ) observed that ” it is
not possible to maximize simultaneously generality, realism, and precision.” In the social
sciences one is often content with a statement that something will increase ; precision
has been sacrificed for realism and (hopefully) generality. Simulation models usually try
for precision and realism but sacrifice generality. These three trade-offs should become
clearer after you have studied some actual models.
Definitions of the variables and their interrelations constitute the assumptions of the
model. We then use the model to draw conclusions (i.e., to make predictions). This is a
deductive process : If the assumptions are true, the conclusions must also be true. Hence
a false prediction implies that the model is wrong in some respect. Unfortunately things
are usually not this clear-cut. We know our model is only an approximation, so we cannot
expect perfect predictions. How can we judge a model in this case?
A conclusion derived from a crude model is not very believable, especially if other models
make contrary predictions. A result is robust if it can be derived from a variety of different
models of the same situation, or from a rather general model. A prediction that depends
on very special assumptions for its validity infrangible. The cruder the model, the less
believable its fragile predictions.

4
 1.2. PRINCIPLES OF MATHEMATICAL MODELING

You may notice that we have talked about conclusions, not explanations. Can a model
provide explanations? This is a somewhat philosophical question, and different people
have different notions of what constitutes an explanation. Let us grant that, in some
sense, models can provide explanations. A decision about the validity of a model is
usually based on the accuracy of its predictions. Unfortunately, two different models may
make the same predictions but offer different explanations. How can this be ?

1.2 Principles of Mathematical Modeling

Mathematical modeling is a principled activity that has both principles behind it and
methods that can be successfully applied. The principles are over-arching or meta-
principles phrased as questions about the intentions and purposes of mathematical mod-
eling. These meta-principles are almost philosophical in nature.We will now outline the
principles, and in the next section we will briefly review some of the methods. visual por-
trayal of the basic philosophical approach is shown in Figure 1.1. These methodological
modeling principles are also captured in the following list of questions and answers:

• Why? What are we looking for? Identify the need for the model.

• Find? What do we want to know? List the data we are seeking.

• Given? What do we know? Identify the available relevant data.

• Assume? What can we assume? Identify the circumstances that apply.

• How? How should we look at this model? Identify the governing physical principles.

• Predict? What will our model predict? Identify the equations that will be used,
the calculations that will be made, and the answers that will result.

• Valid? Are the predictions valid? Identify tests that can be made to validate the
model,that is, is it consistent with its principles and assumptions?

• Verified? Are the predictions good? Identify tests that can be made to verify the
model, i.e., is it useful in terms of the initial reason it was done?

5
 1.2. PRINCIPLES OF MATHEMATICAL MODELING

• Improve? Can we improve the model? Identify parameter values that are not
adequately known, variables that should have been included, and/or assumption-
s/restrictions that could be lifted. Implement the iterative loop that we can call
”model-validate-verify-improve-predict.”

• Use? How will we exercise the model? What will we do with the model?

Figure 1.1: A first-order view of mathematical modeling that shows how the questions
asked in a principled approach to building a model relate to the development of that
model (inspired by Carson and Cobelli, 2001).

This list of questions and instructions is not an algorithm for building a good mathematical
model. However, the underlying ideas are key to mathematical modeling, as they are key

6
 1.2. PRINCIPLES OF MATHEMATICAL MODELING

to problem formulation generally. Thus, we should expect the individual questions to

recur often during the modeling process, and we should regard this list as a fairly general
approach to ways of thinking about mathematical modeling. Having a clear picture of why
the model is wanted or needed is of prime importance to the model-building enterprise.
Suppose we want to estimate how much power could be generated by a dam on a large
river, say a dam located at Benishangul-Gumuz region in Ethiopia. For a first estimate
of the available power, we wouldn’t need to model the dams thickness or the strength of
its foundation. Its height, on the other hand, would be an essential parameter of a power
model, as would some model and estimates of river flow quantities. If, on the other hand,
we want to design the actual dam, we would need a model that incorporates all of the
dams physical characteristics (e.g., dimensions, materials, foundations) and relates them
to the dam site and the river flow conditions. Thus, defining the task is the first essential
step in model formulation.
We then should list what we know for example, river flow quantities and desired power
levels as a basis for listing the variables or parameters that are as yet unknown. We
should also list any relevant assumptions. For example, levels of desired power may be
linked to demographic or economic data, so any assumptions made about population and
economic growth should be spelled out. Assumptions about the consistency of river flows
and the statistics of flooding should also be spelled out.
Which physical principles apply to this model? The mass of the river’s water must
be conserved, as must its momentum, as the river flows, and energy is both dissipated
and redirected as water is allowed to flow through turbines in the dam (and hopefully
not spill over the top!). And mass must be conserved, within some undefined system
boundary, because dams do accumulate water mass from flowing rivers. There are well-
known equations that correspond to these physical principles. They could be used to
develop an estimate of dam height as a function of power desired. We can validate the
model by ensuring that our equations and calculated results have the proper dimensions,
and we can exercise the model against data from existing hydroelectric dams to get
empirical data and validation. If we find that our model is inadequate or that it fails
in some way, we then enter an iterative loop in which we cycle back to an earlier stage
of the model building and re-examine our assumptions, our known parameter values, the

7
 1.3. SOME METHODS OF MATHEMATICAL MODELING

principles chosen, the equations used, the means of calculation, and so on. This iterative
process is essential because it is the only way that models can be improved, corrected,
and validated.

1.3 Some Methods of Mathematical Modeling

Now, we will review some of the mathematical techniques we can use to help answer the
philosophical questions posed in Section 1.2. These mathematical principles include:

• dimensional homogeneity,

• abstraction and scaling,

• conservation and balance principles, and

• constructing linear models

1.3.1 Dimensional Homogeneity and Consistency

There is a basic, yet very powerful idea that is central to mathematical modeling, namely,
that every equation we use must be dimensionally homogeneous or dimensionally consis-
tent. It is quite logical that every term in an energy equation has total dimensions of
energy, and that every term in a balance of mass should have the dimensions of mass.
This statement provides the basis for a technique called dimensional analysis that we will
discuss in greater detail in Chapter 2.

1.3.2 Abstraction and Scaling

An important decision in modeling is choosing an appropriate level of detail for the prob-
lem at hand, and thus knowing what level of detail is prescribed for the attendant model.
This process is called abstraction and it typically requires a thoughtful approach to identi-
fying those phenomena on which we want to focus, that is, to answering the fundamental
question about why a model is being sought or developed.

8
 1.3. SOME METHODS OF MATHEMATICAL MODELING

For example, a linear elastic spring can be used to model more than just the relation
between force and relative extension of a simple coiled spring, as in an old-fashioned
butchers scale or an automobile spring. It can also be used to model the static and
dynamic behavior of a tall building, perhaps to model wind loading, perhaps as part of
analyzing how the building would respond to an earthquake. In these examples, we can
use a very abstract model by subsuming various details within the parameters of that
model. We will explore these issues further in Chapter 3.

In addition, as we talk about finding the right level of abstraction or the right level of
detail, we are simultaneously talking about finding the right scale for the model we are
developing. For example, the spring can be used at a much smaller, micro scale to model
atomic bonds, in contrast with the macro level for buildings. The notion of scaling includes
several ideas, including the effects of geometry on scale, the relationship of function to
scale, and the role of size in determining limits all of which are needed to choose the right
scale for a model in relation to the reality we want to capture.

1.3.3 Conservation and Balance Principles

When we develop mathematical models, we often start with statements that indicate
that some property of an object or system is being conserved. For example, we could
analyze the motion of a body moving on an ideal, frictionless path by noting that its
energy is conserved. Sometimes, as when we model the population of an animal colony
or the volume of a river flow, we must balance quantities, of individual animals or water
volumes, that cross a defined boundary. We will apply balance or conservation principles
to assess the effect of maintaining or conserving levels of important physical properties.
Conservation and balance equations are related in fact, conservation laws are special cases
of balance laws.

The mathematics of balance and conservation laws are straightforward at this level of
abstraction. Denoting the physical property being monitored as Q(t) and the independent
variable time as t , we can write a balance law for the temporal or time rate of change of

9
 1.3. SOME METHODS OF MATHEMATICAL MODELING

that property within the system boundary depicted in Figure 1.2 as:

dQ(t)
= qin (t) + g(t) − qout (t) − c(t), (1.1)
dt

where qin (t) and qout (t) represent the flow rates of Q(t) into (the influx) and out of (the
efflux) the system boundary, g(t) is the rate at which Q is generated within the boundary,
and c(t) is the rate at which Q is consumed within that boundary. Note that eq. (1.1)
is also called a rate equation because each term has both the meaning and dimensions of
the rate of change with time of the quantity Q(t).

Figure 1.2: A system boundary surrounding the object or system being modeled. The
influx qin (t), efflux qout (t), generation g(t), and consumption c(t), affect the rate at which
the property of interest, Q(t), accumulates within the boundary (after Cha, Rosenberg,
and Dym, 2000).

In those cases where there is no generation and no consumption within the system bound-
ary (i.e., when g = c = 0), the balance law in eq. (1.1) becomes a conservation law:

dQ(t)
= qin (t) − qout (t) (1.2)
dt

Here, then, the rate at which Q(t) accumulates within the boundary is equal to the
difference between the influx, qin (t), and the efflux, qout (t) .

10
 1.4. BUILDING A MODEL

1.3.4 Constructing Linear Models

Linearity is one of the most important concepts in mathematical modeling. Models of

devices or systems are said to be linear when their basic equations whether algebraic,
differential, or integral are such that the magnitude of their behavior or response produced
is directly proportional to the excitation or input that drives them. Even when devices
like the pendulum are more fully described by nonlinear models, their behavior can often
be approximated by linearized or perturbed models, in which cases the mathematics of
linear systems can be successfully applied.
We apply linearity when we model the behavior of a device or system that is forced or
pushed by a complex set of inputs or excitations. We obtain the response of that device
or system to the sum of the individual inputs by adding or superposing the separate
responses of the system to each individual input. This important result is called the
principle of superposition. Engineers use this principle to predict the response of a system
to a complicated input by decomposing or breaking down that input into a set of simpler
inputs that produce known system responses or behaviors.

1.4 Building a Model

Model building involves imagination and skill. Giving rules for doing it is like listing rules
for being an artist ; at best this provides a framework around which to build skills and
develop imagination. It may be impossible to teach imagination. It won’t try, but I hope
this module provides an opportunity for your skills and imagination to grow. With these
warnings, the followings are outline of the modeling process.

1. Formulate the Problem; What is it that you wish to know ? The nature of the
model you choose depends very much on what you want it to do.

2. Outline the Model; At this stage you must separate the various parts of the
universe into unimportant, exogenous, and endogenous . The interrelations among
the variables must also be specified .

3. Is It Useful? Now stand back and look at what you have . Can you obtain the

11
 1.4. BUILDING A MODEL

needed data and then use it in the model to make the predictions you want? If the
answer is no, then you must reformulate the model (step 2) and perhaps even the
problem (step 1) . Note that ”useful” does not mean reasonable or accurate ; they
come in step 4. It means : the model fits the situation, will we be able to use it?

4. Test the Model; Use the model to make predictions that can be checked against
data or common sense . It is not advisable to rely entirely on common sense,
because it may well be wrong. Start out with easy predictions-don’t waste time
on involved calculations with a model that may be no good . If these predictions
are bad and there are no mathematical errors, return to step 2 or step 1 . If these
predictions are acceptable, they should give you some feeling for the accuracy and
range of applicability of the model. If they are less accurate than you anticipated,
it is a good idea to try to understand why , since this may uncover implicit or false
assumptions .

At this point the model is ready to be used. Don’t go too far ; it is dangerous to apply the
model blindly to problems that differ greatly from those on which it was tested. Every
application should be viewed as a test of the model. You may not be able to carry out
step 2 immediately, because it is not clear what factors can be neglected. Furthermore,
it may not be clear how accurately the exogenous variables need to be determined. A
common practice is to begin with a crude model and rough data estimates in order to
see which factors need to be considered in the model and how accurately the exogenous
variables must be determined. Some models may require no data. If a model makes
the same prediction regardless of the data, we are not getting something for nothing
because this prediction is based on the assumptions of the model. To some extent, the
distinction between data and assumptions is artificial. In an extreme case, a model may
be so specialized that its data are all built into the assumptions. Sometimes step 4 may
be practically impossible to carry out. For example, how can we test a model of nuclear
war ? What do we do if we have two models of a nuclear war and they make different
predictions ? This can easily happen in fields of study that lack the precisely formulated
laws found in the physical sciences. At this point, experience is essential-not experience in
mathematics but experience in the field being modeled. Even if predictions can be tested,
the testing may be expensive to carry out and may require training in a particular field

12
 1.5. EXAMPLES OF MODELING

of experimental science since the absence of experimental verification leaves the modeling
process incomplete.

1.5 Examples of Modeling

Here, we do a quick tour of several examples of the mathematical process. We present

the models as finished results as opposed to attempting to develop the models.

1.5.1 Modeling with Difference Equations

Consider the situation in which a variable changes in discrete time steps. If the current
value of the variable is an , then the predicted value of the variable will be an+1 . A
mathematical model for the evolution of the (still unspecified) quantity an could take the
form
an+1 = αan + β

In words, the new value is a scalar multiple of the old value offset by some constant β.
This model is common, e.g., it is used for modeling bank loans. One might amend the
model to make the dependence depend on more terms and to include the possibility that
every iteration the offset can change, thus,

an+1 = α1 an + α2 a2n + βn

This could correspond to, for example, a population model where the the migration levels
change every time step. In some instances, it is clear that information required to predict
a new value goes back further than the current value, e.g.,

an+1 = an + an−1

13
 1.5. EXAMPLES OF MODELING

Note now that two initial values are required to evolve this model. Finally, it may be that
the form of the difference equations are unknown and the model must be written as

an+1 = f (an , an−1 , an−M −1 )

Determining the nature of f and the step M is at the heart of model formulation with
difference equations. Often observed data can be employed to assist in this effort.

1.5.2 Modeling with Ordinary Differential Equations

Although modeling with ordinary differential equations shares many of the ideas of model-
ing with the difference equations discussed above, there are many fundamental differences.
At the center of these, differences is the assumption that time is a continuous variable.

One of the simplest differential equations is also an extremely important model, i.e.,

dx
= αx
dt

In words, the rate of change of the quantity x depends on the amount of the quantity. If
α > 0, then we have exponential growth. If α < 0,then the situation is exponential decay.
Of course, additional terms can be added that fundamentally alter the evolution of x(t).
For example,
dx
= α 1 x + α 2 x2
dt
The model formulation again requires the development of the appropriate right hand side.
In the above model the value x on the right hand side is implicitly assumed to be evaluated
at the time t. It may be that there is evidence that the instantaneous rate of change at
time t is actually a function of a previous time, i.e.,

dx
= f (x(t)) + g(x(t − τ ))
dt

This is referred to as a delay differential equation.

14
 1.5. EXAMPLES OF MODELING

1.5.3 Modeling with Partial Differential Equation

In the previous sections on modeling the behaviour of a variable as a function of time we

assumed that there was only one independent variable. Many situations arise in practice
where the number of independent variables is more than two. For spatio-temporal models
we might have time and space.
Example,
∂f ∂ 2f
=α 2
∂t ∂x
or
∂ 2f ∂ 2f ∂ 2f
= + 2
∂t2 ∂x2 ∂y

1.5.4 Optimization

In many modeling problems the goal is to compute the ”best” solution. This may corre-
spond to maximizing profit in a company, or minimizing loss in a conflict. It is no surprise
that optimization techniques take a central seat in the mathematical modeling literature.
Now one may allow x ∈ Rn and require that

x∗ = argminf (x)

The quantity f (x) is referred to as the objective function while the vector x consists of
decision variables. Because x in Rn , the problem is referred to as unconstrained.

Alternatively, one might require that the solution x have all positive components. If we
refer to this set as S then the optimization problem is constrained

x∗ = arg |{z}
min f (x)
x∈S

If the objective function as well as the equations that define the constraint set are linear,
then the optimization problem is called a linear programming problem. Otherwise, the
problem is referred to as a nonlinear programming problem. As we shall see, solution
methods for linear and nonlinear programming problems are very different.

15
 1.6. WHY STUDY MODELING

1.5.5 Modeling with Simulations

Many problems may afford a mathematical formulation yet be analytically intractable.

In these situations a computer can implement the mathematics literally and repetitively
often times to extreme advantage.
Simulating Games

• What is the probability that you can win a game of solitaire?

• What is the best strategy for playing blackjack?

• Given a baseball team consisting of certain players, in what order should they hit?

On the other hand, computer simulations can be employed to model evolution equations.
Applications in the realm of fluid dynamics and weather prediction are well established.
A striking new example of such simulation modeling is attempting to model electrical
activity in the brain.

1.5.6 Function Fitting: Data Modeling

Often data is available from a process to assist in the modeling. How can functions be
computed that reflect the relationships between variables in the data. Produce a model

y = f (x; w)

and using the set of input output pairs compute the parameters w. In some cases the
form of f may be guessed. In other cases a model free approach can be used.

1.6 Why Study Modeling

Why not always deal with the real world instead of studying models ? We study modeling
for the following reasons (but not limited to):

1. To understand the behavior of the existing system.

16
 1.7. CLASSIFICATIONS OF MATHEMATICAL MODELING

2. To reduce the complexity of the system ( that is, reduce the need for costly, unde-
sirable, or impossible experiments with the real world).

3. To predict the effect of changes to the system.

1.7 Classifications of Mathematical Modeling

Mathematical models are usually composed of relationships and variables. Relationships

can be described by operators, such as algebraic operators, functions, differential op-
erators, etc. Variables are abstractions of system parameters of interest, that can be
quantified. Several classification criteria can be used for mathematical models according
to their structure:

• Linear vs. nonlinear: If all the operators in a mathematical model exhibit linear-
ity, the resulting mathematical model is defined as linear. A model is considered to
be nonlinear otherwise. The definition of linearity and nonlinearity is dependent on
context, and linear models may have nonlinear expressions in them. For example,
in a statistical linear model, it is assumed that a relationship is linear in the param-
eters, but it may be nonlinear in the predictor variables. Similarly, a differential
equation is said to be linear if it can be written with linear differential operators,
but it can still have nonlinear expressions in it. In a mathematical programming
model, if the objective functions and constraints are represented entirely by linear
equations, then the model is regarded as a linear model. If one or more of the
objective functions or constraints are represented with a nonlinear equation, then
the model is known as a nonlinear model.
Nonlinearity, even in fairly simple systems, is often associated with phenomena such
as chaos and irreversibility. Although there are exceptions, nonlinear systems and
models tend to be more difficult to study than linear ones. A common approach to
nonlinear problems is linearization, but this can be problematic if one is trying to
study aspects such as irreversibility, which are strongly tied to nonlinearity.

• Static vs. dynamic: A dynamic model accounts for time-dependent changes in

the state of the system, while a static (or steady-state) model calculates the system

17
 1.7. CLASSIFICATIONS OF MATHEMATICAL MODELING

in equilibrium, and thus is time-invariant. Dynamic models typically are represented

by differential equations or difference equations.

• Explicit vs. implicit: If all of the input parameters of the overall model are known,
and the output parameters can be calculated by a finite series of computations, the
model is said to be explicit. But sometimes it is the output parameters which are
known, and the corresponding inputs must be solved for by an iterative procedure,
such as Newton’s method (if the model is linear) or Broyden’s method (if non-linear).
In such a case the model is said to be implicit. For example, a jet engine’s physical
properties such as turbine and nozzle throat areas can be explicitly calculated given
a design thermodynamic cycle (air and fuel flow rates, pressures, and temperatures)
at a specific flight condition and power setting, but the engine’s operating cycles at
other flight conditions and power settings cannot be explicitly calculated from the
constant physical properties.

• Discrete vs. continuous: A discrete model treats objects as discrete, such as the
particles in a molecular model or the states in a statistical model; while a continuous
model represents the objects in a continuous manner, such as the velocity field of
fluid in pipe flows, temperatures and stresses in a solid, and electric field that applies
continuously over the entire model due to a point charge.

• Deterministic vs. probabilistic (stochastic): A deterministic model is one in

which every set of variable states is uniquely determined by parameters in the model
and by sets of previous states of these variables; therefore, a deterministic model
always performs the same way for a given set of initial conditions. Conversely,
in a stochastic model-usually called a ”statistical model”-randomness is present,
and variable states are not described by unique values, but rather by probability
distributions.

• Deductive, inductive, or floating: A deductive model is a logical structure based

on a theory. An inductive model arises from empirical findings and generalization
from them. The floating model rests on neither theory nor observation, but is
merely the invocation of expected structure. Application of mathematics in social
sciences outside of economics has been criticized for unfounded models. Application
of catastrophe theory in science has been characterized as a floating model.

18
 1.7. CLASSIFICATIONS OF MATHEMATICAL MODELING

Exercises
1. What is model?

2. What is modeling?

3. What is mathematical modeling?

4. List the steps to build or construct a model.

5. List some types of models.

6. What you understand from methods of mathematical modeling?

7. Why you study modeling?

8. Suppose people enter the elevators in a skyscraper at random during the morning
rush. The result will be several elevators stopping on each floor to discharge one or
two passengers each.

(a) Discuss schemes for improving the situation.

(b) How could improvement be measured ?
(c) How could you model the situation to decide what scheme to adopt ?

19
CHAPTER 2

ARGUMENTS FROM SCALE

Objectives:

By the end of this chapter you will be able to:

• know the effect of size on an object.

• identify endogenous and exogenous variables in the model.

• describe various techniques for reasoning about the scales.

• provide a lot of different examples on making models dimensionless with physically

correct scales.

• develop a better understanding of dimensions, units, and dimensional homogeneity

of equations.

• understand the numerous benefits of dimensional analysis.

• know how to use the method of repeating variables to identify nondimensional pa-
rameters from the model.

20
 2.1. EFFECTS OF SIZE

• apply Buckingham’s Pi-theorem to predict expected number of dimensionless prod-

ucts in the problem.

• understand the concept of dynamic similarity and how to apply it to experimental

modeling.

In this chapter, we consider arguments based on proportionality. For example, if you

make a scale model of an object with a scale of 1 : l, surface area will have a scale of 1 : l2
and the volume will have a scale of 1 : l3 . Models using this sort of idea are discussed
in the first section. The second section is based on the observation that physical laws
remain the same if the units of measurement are changed.

2.1 Effects of Size

Cost of Packaging
Consider a product like flour, detergent, or water, which is packaged in containers of
various sizes. You’ve probably noticed that larger packages of such products usually cost
less per kilogram. This is often attributed to savings in the cost of packaging and handling.
If this in fact the major cause or are there other important factors? We try to see where
this idea leads by constructing a simple model.
The cost of a product is the endogenous variable. We are interested in seeing how it varies
with the exogenous variable, size. Cost clearly depends on competition and the scale of
the business. We neglect these factors and concentrate on expenses due to materials and
handling. Since we are neglecting some important factors (name some), the resulting
predictions will be crude. In addition, there are various constants involved which we do
not even pretend to evaluate.
Let’s begin by studying the wholesale cost, that is, the price the retailer pays for the
product. This is a sum of several costs plus various profit markups by middlemen. Since
profit markups are usually in terms of percentages,we can absorb them in constants later
; for example, a 30% markup multiplies constants by 1.30. The main costs that enter the
wholesale price are :

21
 2.1. EFFECTS OF SIZE

1. Cost of producing the product, a.

2. Cost of packaging the product, b.

3. Cost of shipping the product, c.

4. Cost of the packaging material, d.

We will consider each of these in turn.

It is reasonable to assume that a is proportional to the amount of the good being produced.
We write this as a ∝ w, which is read ” a is proportional to the weight w.”
The costs of packaging depend on

• how long it takes to fill the package,

• how long it takes to close the package, and

• how long it takes to load the package into a box for shipping.

The first time is probably nearly proportional to the volume (hence the weight), while
the latter two times are probably about the same for all sizes of packages in a reasonable
range. Thus b ≈ f w + g for some positive constants f and g. (The symbol ≈ means ”
approximately equal to.”)

Shipping charges may depend on both weight and volume. Since volume is proportional
to weight for filled packages, we have c ∝ w.

The cost of the packaging material is more complicated. It depends on the costs the
package manufacturer must meet. Thus we must consider a, b, c, and d for the package
manufacturer. We neglect d; that is, we neglect the cost of the containers for the material
from which the final packages are made. From the analysis we have just completed,
the cost per package depends on the weight and volume of the package. If the range of
packages we are considering is not too large, it is reasonable to assume that the packaging
material is the same for all sizes of packages. Therefore the amount of material per
package (hence the weight of a package) is proportional to the area of the surface to be
covered. The volume per package is proportional to either the surface area or the volume

22
 2.1. EFFECTS OF SIZE

of the package, depending on whether the packaging is shipped collapsed (like cardboard)
or preformed (like glass). Therefore the expenses per package of the package supplier are
hw + kS + m , for constants h ≥ 0, k > 0 , and m > 0, where S is the surface area. Except
for a markup, this is the cost d to the packager.
We now use a scale argument to reduce everything to one independent variable, weight.
Let us assume that the various packages are roughly geometrically similar. The volume
is nearly proportional to the cube of a linear dimension, and the surface area is nearly
proportional to the square of a linear dimension . Thus, v ∝ l3 and S ∝ l2 implies
1 2
v ∝ Sl ⇒ v ∝ Sv 3 ⇒ S ∝ v 3
2 2
Hence S ∝ v 3 . Since v ∝ w, we have S ∝ w 3 . Thus the wholesale cost per kilogram is

cost a+b+c+d
= (2.1)
w w

Now, from the above discussion we have the followings relations.

a ∝ w ⇒ a = c1 w
b = fw + g
c ∝ w ⇒ c = c2 w
d = hw + kS + m
2 2
S ∝ w 3 ⇒ S = c3 w 3
Substitute this relationship in eq.(2.1). We get

cost a+b+c+d
=
w w
c1 w + f w + g + c2 w + hw + kS + m
=
w
2
c1 w + f w + g + c2 w + hw + kc3 w 3 + m
=
w
2
(c1 + f + c2 + h)w + kc3 w 3 + g + m
=
w
1 g+m
= (c1 + f + c2 + h) + kc3 w− 3 +
w
cost 1 q
= n + pw− 3 + (2.2)
w w

23
 2.1. EFFECTS OF SIZE

Where, n = c1 + f + c2 + h, p = kc3 and q = g + m

From this we see that the cost per kilogram decreases as the size of the package increases,
in agreement with the observation made at the start of this discussion.

Can we make any interesting predictions? Given three different costs and weights, we
could solve for n, p, and q in (2.2) and use the results to predict the prices for packages
of other sizes. Because of the crudity of our model,it is unlikely that our equation will
fit very well. We should not take the exact form of (2.2) too seriously. Another way
to fit a curve, which allows for inaccuracies, is the method of least squares. For this to
be a reasonable test of the model, we should have more data points than parameters.
Since (2.2) involves three constants, we should have more than three values for the cost
and weight of a single product. This is hard to obtain because of the limited number of
different-sized packages in which a particular product is available. Therefore we need a
different approach.

The cost per kilogram decreases at a rate

d(cost/w) p q
r=− = 4 + (2.3)
dw 3w 3 w2

This is a decreasing function of w. Thus the increase in the rate of savings per kilogram
is less when the package is larger. We can also compute the rate of total savings rw

p q
rw = 1 +
3w 3 w

It is also a decreasing function of w.

The consumer is not likely to understand this. We can make a statement like (2.3) in
simpler terms. In purchasing prepackaged products, doubling the size of the package
purchased tends to result in greater savings per kilogram when the packages are small
than when they are large.

This discussion concerned wholesale prices. What about retail prices ? The retailer’s
costs depend on wholesale prices and handling and storage costs. As above, the latter
two costs are of the form Hw + M . If the wholesaler sets his price at a fixed percentage
above his costs, then we again obtain an equation of the form (2.2). The conclusions we

24
 2.1. EFFECTS OF SIZE

reached above are therefore valid for retail prices too.

Speed of Racing Shells

In the college sport of crew racing the best times vary from class to class. Why? Can we
advise a coach how to adjust the shells so that he can pit his teams against each other on
an equal basis in practice ? This model is adapted from T. A. McMahon’s article (1971)
and deals with data for men only.
Racing shells are boats propelled by oarsmen in sporting contests. They hold one, two,
four, or eight oarsmen and are built to certain specifications. Figure2.1 is a rough diagram
of a racing shell. For an eight-man crew there is a lightweight category and a heavyweight
category.

Figure 2.1: (a) Top view. (b) Cross section of center. I = length; b = beam; A =
cross-sectional area.

Heavyweight oarsmen average about 86 kilograms, and lightweight oarsmen about 73

kilograms. This gives five classes. (There are others which we ignore because of a lack
of data.) McMahon observed that there is a rather consistent difference between the

25
 2.1. EFFECTS OF SIZE

Table 2.1: Times of racing crews in four meets

Number of men I II III IV
8 5.87 5.92 5.82 5.73
4 6.33 6.42 6.48 6.13
2 6.87 6.92 6.95 6.77
1 7.16 7.25 7.28 7.17

best times of the various classes. Table 2.1 lists the information he presented on best
times for 2000 meter races in four international competitions. The eight-man entry is the
heavyweight time.McMahon also states that the time of an eight-man heavyweight crew
is about 5% better than the time of an eight-man lightweight crew.

We want to explain all this.

Rather than present the underlying assumptions of the model in one ad hoc package, we
develop them as we proceed.
A shell is propelled by the power of the oarsmen and retarded by the drag of the water.
The balance of these two forces determines the speed of the shell, hence its time in the
race. We assume

1. The only drag force on the shell is due to skin friction and this force is proportional
to Sv 2 , where S is the wetted surface area and v is the velocity.
The expression for the skin friction drag given in the assumption is obtained from
hydrodynamics. The power P required to maintain velocity v is, by definition, equal
to the drag force times the velocity. Hence P ∝ Sv 3 , and so v ∝ (P/S)1/3 .

We assume

2. The oarsmen in the shell all have the same weight and the same constant power
output for the entire course of the race.

It follows that v is constant, except for the brief period when the shells are starting up.
Hence the course time t is proportional to v −1 , and so
 1/3
S
t∝ . (2.4)
P

26
 2.1. EFFECTS OF SIZE

We now consider the time difference between the heavyweight and lightweight eight-man
crews. We want to explain it and then see if we can find a way to redesign the shells so that
the two classes will be more nearly equal. The subscripts H and L denote heavyweight
and lightweight, respectively. From (2.4) we obtain
  13   13
tL SL PH
= . (2.5)
tH SH PL

We must say something about power output and wetted surface area if we are going to
explain the 5% edge of the heavyweight team. Unfortunately power output information
is not obtainable ; however, we know that the ratio of the weights of heavyweight and
86
lightweight oarsmen is about 73
kilograms = 1.18. Therefore we try to relate power and
weight.
Sustained power output depends on such factors as lung volume (actually lung surface
area, but this is proportional to volume because the lungs consist of small cells whose size
is independent of the size of the person) and muscle volume. For similarly proportioned
people, these are proportional to the total weight. Hence we can expect power output
to be proportional to the weight w of an oarsman times the number of oarsmen . Since
wH PH
wL
= 1.18 and both shells have eight oarsmen, PL
= 1.18. Combining this with (2.5),

 1/3  1/3
tL SL SL
= (1.18)1/3 = 1.06 . (2.6)
tH SH SH

If we make the rough assumption that SL = SH , then (2.6) comes close to the 5% observed
difference. Actually the surface area for a loaded heavyweight shell is slightly greater than
that for a lightweight shell. When this is taken into account, the 6% edge in (2.6) decreases
slightly. We haven’t predicted the edge precisely, but we have explained why it is in the
neighborhood of 5%.
How can the shells be redesigned to achieve equality ? For fixed power output we obtain
t ∝ S 1/3 from (2.4). To change the time we must change the wetted surface area of
the loaded shell. Let the subscripts p and r denote the present and redesigned shells,

27
 2.1. EFFECTS OF SIZE

respectively. Then  3
Sr tr
= .
Sp tp
The lightweight crews will have times about equal to those of the heavyweight crews if
tr
tp
= 0.95. By the above equation, SSpr = 0.86. In words, the wetted surface area of a loaded
lightweight shell should be decreased by about 14%, or we could slow the heavyweights
1
down by an increase of wetted area of about 16%( 0.86 = 1.16).
We now compare the times of various-sized shells by expressing the endogenous variable,
course time, in terms of the exogenous variable, team size. To do this we have to relate
S and P to the size of the team. If assumption 2 is extended to all oarsmen in all shells,
the power will be proportional to the number of oarsmen n. Hence (2.4) reduces to
 1/3
S
t∝ . (2.7)
n

We need some information about the relative sizes of the various shells so that we can
compute S. The information in Table 2.2 was presented by McMahon as evidence for the
assumption:
3. The shells are geometrically similar, and their loaded weights are proportional to n.
Furthermore, the submerged parts of the loaded shells are also geometrically similar.

Table 2.2: Shell design parameters

n l b l/b weight/n
8 18.28 0.610 30.0 14.7
4 11.75 0.574 21.0 18.1
2 9.76 0.356 27.4 13.6
1 7.93 0.293 27.0 16.3

Note : I = length; b = beam. The variation in the 00 l/b00 and 00 weight/n00 columns shows
that this is a rather crude assumption, but it is about the best we can do, since a table
of wetted surface areas is not available.
The volume of water displaced by a shell is proportional to its total weight. This volume
is also proportional to lA. By assumption 3, weight is proportional to the number of

28
 2.1. EFFECTS OF SIZE

oarsmen n, and A ∝ l2 . Thus

n ∝ lA ∝ l3 . (2.8)

The values of l and n listed in Table 2.2 do not satisfy n ∝ l3 . Therefore the similarity
assumption is wrong. What can we do about it ? For the sake of continuity, we postpone
discussing this problem.

The total submerged surface area is proportional to l times the submerged perimeter of
cross section A in Figure 2.1. By assumption 3, this perimeter is proportional to A1/2
which is in turn proportional to l. Thus S ∝ l2 . From (2.8) we obtain S ∝ n2/3 , and so
(2.7) becomes
t ∝ n−1/9 . (2.9)

This yields the prediction:

Times are proportional to the number of oarsmen raised to the power − 91 .

We can test this prediction by graphing t versus n in some fashion. It is much easier to
see if points are close to a straight line than it is to see if they are close to a curve. For
this reason relationships like (2.9) are usually plotted on what is called log-log paper. It
gives the effect of plotting log n against log t, which equals c - log n/9 if (2.9) is correct.
If you do this, you will discover that the points come close to lying on a straight line of
slope − 19 as predicted.
We are in an awkward situation : the prediction in (2.9) has been verified, but the
intermediate result in (2.8) is wrong. One possible explanation for this is that the central
portions of the shells (which displace most of the water, hence are the most important)
obey the similarity assumptions better than the ends of the shells. We do not have
the data to check this possibility. This central length λ and the cross section enter
into the calculations for volume and surface area. A reasonable rough approximation is
that volume and surface area are proportional to λ3 and λ2 , respectively. The previous
calculations can then be carried out with λ replacing l.
We can give a more robust argument that leads to (2.9). The volume of the submerged
portion of the shell is proportional to the weight of the loaded shell by Archimedes’ law.
The weight is very nearly proportional to n. Hence the volume is very nearly proportional
to n. Since the shells are all approximately the same shape, the surface area is nearly

29
 2.2. DIMENSIONAL ANALYSIS

proportional to the 32 power of the volume. Hence S is nearly proportional to n2/3 . By

 2/3 1/3
(2.7), t ∝ n n = n1/9 . The important point in this argument is that surface area
tends to remain proportional to the 32 power of the volume, even when the shape varies
somewhat from shell to shell. Thus we do not need the exact similarity assumption 3.

2.2 Dimensional Analysis

2.2.1 Dimensions and Units

A dimension is a measure of a physical quantity (without numerical values), while a

unit is a way to assign a number to that dimension. For example, length is a dimension
that is measured in units such as microns (µm), feet (f t), centimeters (cm), meters (m),
kilometers (km), etc. In the SI system, there are seven such primary dimensions (also
called fundamental or basic dimensions ) corresponding physical quantities: meter (m) for
length, kilogram (kg) for mass, second (s) for time, kelvin (K) for temperature, ampere
(A) for electric current, candela (cd) for luminous intensity, and mole (mol) for the amount
of substance.

We need some suitable mathematical notation to calculate with dimensions like length,
mass, time, and so forth. The dimension of length is written as [L], the dimension of
mass as [M ], the dimension of time as [T ], and the dimension of temperature as [θ]. The
dimension of a derived unit like velocity, which is distance (length) divided by time, then
becomes [LT −1 ] in this notation. The dimension of force, another derived unit, is the
same as the dimension of mass times acceleration, and hence the dimension of force is
[M LT −2 ].

Example: A mechanical system undergoing one-dimensional damped vibrations can be

modeled by the equation
mu00 + bu0 + ku = 0, (2.10)

where m is the mass of the system, b is some damping coefficient,k is a spring constant,
and u(t) is the displacement of the system. This is an equation expressing the balance of
three physical effects: mu00 (mass times acceleration), bu0 (damping force), and ku (spring

30
 2.2. DIMENSIONAL ANALYSIS

force). The different physical quantities, such as m, u(t), b and k, all have different
dimensions, measured in different units, but mu00 , bu0 , and ku must all have the same
dimension, otherwise it would not make sense to add them.

Let us find the dimensions of the terms in (2.10). A displacement u(t) has dimension
[L]. The derivative u0 (t) is change of displacement, which has dimension [L], divided by
a time interval, which has dimension [T ], implying that the dimension of u0 is [LT −1 ].
This result coincides with the interpretation of u0 as velocity and the fact that velocity is
defined as distance ([L]) per time ([T ]).
Looking at (2.10), and interpreting u(t) as displacement, we realize that the term mu00
(mass times acceleration) has dimension [M LT −2 ]. The term bu0 must have the same
dimension, and since u0 has dimension [LT −1 ], b must have dimension [M T −1 ]. Finally,
ku must also have dimension [M LT −2 ], implying that k is a parameter with dimension
[M T −2 ].
The unit of a physical quantity follows from the dimension expression. For example, since
velocity has dimension [LT −1 ] and length is measured in m while time is measured in
s, the unit for velocity becomes m/s. Similarly, force has dimension [M LT −2 ] and unit
kgm/s2 . The k parameter in (2.10) is measured in kgs−2 .
Dimension of Derivatives The easiest way to realize the dimension of a derivative, is
to express the derivative as a finite difference. For a function u(t) we have

du u(t + ∆t) − u(t)

≈
dt ∆t

where ∆t is a small time interval. If u denotes a velocity, its dimension is[LT −1 ] , and u(t+
∆t)−u(t) gets the same dimension.The time interval has dimension [T ], and consequently,
the finite difference gets the dimension [LT −2 ]. In general, the dimension of the derivative
du
dt
is the dimension of u divided by the dimension of t.

2.2.2 Dimensions of Common Physical Quantities

Many derived quantities are measured in derived units that have their own name. Force
is one example: Newton (N) is a derived unit for force, equal to kgm/s2 . Another derived

31
 2.2. DIMENSIONAL ANALYSIS

unit is Pascal (Pa) for pressure and stress, i.e., force per area. The unit of Pa then equals
N/m2 or kg/ms2 . In Table 2.3 there are more names for derived quantities, listed with
their units.

Table 2.3: Derived physical quantities with their units

Name Symbol Physical quantity Unit
radian rad angle 1
hertz Hz frequency s−1
newton N force,weight kgm/s2
pascal Pa pressure, stress N/m2
joule J energy,work,heat Nm
watt W power J/s

Some common physical quantities and their dimensions are listed in Table 2.4.

Table 2.4: Some common physical quantites and their dimensions

Quantity Relation Unit Dimension
stress force/area N/m2 = P a [M T −2 L−1 ]
pressure force/area N/m2 = P a [M T −2 L−1 ]
density mass/volume kg/m2 [M L−3 ]
strain displacement/length 1 [1]
Young’s modulus stress/strain N/m2 = P a [M T −2 L−1 ]
Poisson’s ratio transverse strain/axial strain 1 [1]
moment (of a force) distance × force Nm [M T −2 L2 ]
impulse force × time Ns [M LT −1 ]
linear momentum mass × velocity kgm/s [M LT −1 ]
angular momentum distance × mass × velocity 2
kgm /s [M L2 T −1 ]
work force × distance Nm = J [M L2 T −2 ]
energy work Nm = J [M L2 T −2 ]
power work/time N m/s = W [M L2 T −3 ]
heat work J [M L2 T −2 ]
−2
heat flux heat rate/area Wm [M T −3 ]
temperature base unit K [Θ]
heat capacity heat change/temperature change J/K [M L2 T −2 Θ−1 ]
−1 −1
specific heat capacity heat capacity/unit mass JK kg [L2 T −2 Θ−1 ]
−1 −1
thermal conductivity heat flux/temperature gradient Wm K [M LT −3 Θ−1 ]
−1 −1
dynamic viscosity shear stress/velocity gradient kgm s [M L−1 T −1 ]
kinematic viscosity dynamic viscosity/density m2 /s [L2 T −1 ]
surface tension energy/area J/m 2
[M T −2 ]

Prefixes for units; Units often have prefixes. For example, kilo (k) is a prefix for 1000,

32
 2.2. DIMENSIONAL ANALYSIS

so kg is 1000g. Similarly, GPa means giga pascal or 109 Pa.

2.2.3 Dimensional Homogeneity

A very fundamental assumption in physics is that physical principals can be expressed

in terms of mathematical equations that describe the underlying relation of different
physical quantities. Mathematical equations that reflect physical laws are referred to as
physical equations. For physical equations to remain valid for any chosen unit system,
each summand of such an equation must carry the same dimension. Equations which
satisfy this property are called dimensionally homogeneous.
In the context of physical considerations, the addition (or subtraction) of different entities
is not allowed, because it does not lead to physical equations. Such equations are not
dimensionally homogeneous since an alternate choice of units does not necessarily change
the numerical values of all summands by the same factor.

In engineering sciences sometimes empirical equations or numerical value equations are

used. When repeatedly applying the same equation, that contains constants and material
values, it is often perceived as practical to calculate the recurring numerical values.

The law of dimensional homogeneity, stated that : every additive term in an

equation must have the same dimensions.

As an example, we use the thermal state equation of an ideal gas law

pv = RT (2.11)

that determines the relation between pressure p, specific volume v and temperature T
of an ideal gas. In this equation, we have also the gas constant R, which takes different
values for each gas. Equation (2.11) is dimensionally homogeneous, and therefore valid
for any choice of unit system. However, if we are introducing the numerical value of
R = 287J/(kgK) for air, we obtain from Eq. (2.11)

pv = 287T, where [p] = M L−1 T −2 , [v] = M −1 L3 , and [T ] = K. (2.12)

33
 2.2. DIMENSIONAL ANALYSIS

This is a numerical value equation, which is valid only for the units meter, second and
Kelvin. Equation (2.12) may be regarded as an empirically derived equation, where the
experimenter has measured the ratio of p, v and T and found that this ratio is constant
for air. Although this equation can be quite useful for practical purposes, it is not dimen-
sionally homogeneous, and therefore does not constitute a physical equation. Equation
(2.12) may be converted back into a dimensionally homogeneous form by interpreting the
parameter 287 in Eq. (2.12) as a dimensional quantity with a numerical value of 1 and
the unit L2 T −1 K −1 .
This consideration can be generalized: any empirical equation can be converted into a
dimensionally homogeneous form by introducing appropriate dimensional constants. In
fact, the laws of physics are dimensionally homogeneous because they are brought in a
dimensionally homogeneous form by introducing corresponding dimensional constants.
Variables that change their value with alternate choice of base units are to be counted
as additional variables in dimensional analytical considerations. Therefore, in addition
to the variables in a strict mathematical sense one must also include material values and
dimensional constants.

Therefore, for dimensional analytical considerations the following is important: The meth-
ods of dimensional analysis can be only applied to dimensionally homogeneous equations,
and dimensional constants (for instance, to bring dimensionally non-homogeneous equa-
tions in a dimensionally homogeneous form) are to be counted as variables in the sense
of dimensional analysis.
Again, let us consider, the change in total energy of a simple compressible closed system
from one state and/or time (1) to another (2). The change in total energy of the system
(∆E) is given by
∆E = ∆U + ∆KE + ∆P E (2.13)

where E has three components: internal energy (U ), kinetic energy (KE), and poten-
tial energy (P E). These components can be written in terms of the system mass (m);
measurable quantities and thermodynamic properties at each of the two states, such as
speed (V ), elevation (z), and specific internal energy (u); and the known gravitational

34
 2.2. DIMENSIONAL ANALYSIS

acceleration constant (g),

1
∆U = m(u2 − u1 ) ∆KE = m(V22 − V12 ) ∆P E = mg(z2 − z1 ) (2.14)
2

It is straightforward to verify that the left side of Eq. (2.13) and all three additive terms
on the right side of Eq. (2.14) have the same dimensions energy. Using the definitions of
Eq. (2.14), we write the primary dimensions of each term,

[∆E] = [Energy] = [Force.Length] → [∆E] = M L2 /T 2

 
Energy
[∆U ] = Mass = [Energy] → [∆U ] = M L2 /T 2
Mass
Length2
 
[∆KE] = Mass 2 → [∆KE] = M L2 /T 2
Time
 
Length
[∆P E] = Mass Length → [∆P E] = M L2 /T 2
Time2
If at some stage of an analysis we find ourselves in a position in which two additive terms
in an equation have different dimensions, this would be a clear indication that we have
made an error at some earlier stage in the analysis. In addition to dimensional homo-
geneity, calculations are valid only when the units are also homogeneous in each additive
term. For example, units of energy in the above terms may be J, N.m, or kgm2 /s2 , all
of which are equivalent. Suppose, however, that kJ were used in place of J for one of the
terms. This term would be off by a factor of 1000 compared to the other terms. It is wise
to write out all units when performing mathematical calculations in order to avoid such
errors.

Example: Dimensional Homogeneity of the Bernoulli Equation

Probably the most well-known (and most misused) equation in fluid mechanics is the
Bernoulli equation. The standard form of the Bernoulli equation for in.compressible
irrotational fluid flow is
1
P + ρV 2 + ρgz = C (2.15)
2

35
 2.2. DIMENSIONAL ANALYSIS

1. Verify that each additive term in the Bernoulli equation has the same dimensions.

2. What are the dimensions of the constant C?

Solution: We are to verify that the primary dimensions of each additive term in Eq.
(2.15) are the same, and we are to determine the dimensions of constant C.
(1) Each term is written in terms of primary dimensions,
   
Force Length 1 M
[P ] = [Pressure] = = Mass. 2 . 2 =
Area Time Length LT 2
  "  2 #
1 2 Mass Length M
ρV = =
2 Volume Time LT 2
 
Mass Length M
[ρgz] = 2 Length =
Volume Time LT 2
Indeed, all three additive terms have the same dimensions.
(2) From the law of dimensional homogeneity, the constant must have the same dimensions
as the other additive terms in the equation. Thus,

M
[C] =
LT 2

Discussion If the dimensions of any of the terms were different from the others, it would
indicate that an error was made somewhere in the analysis.

2.2.4 Nondimensionalization of Equations

The law of dimensional homogeneity guarantees that every additive term in an equation
has the same dimensions. It follows that if we divide each term in the equation by a col-
lection of variables and constants whose product has the same dimensions, the equation is
rendered nondimensional. If, in addition, the nondimensional terms in the equation are of
order unity, the equation is called normalized. Normalization is thus more restrictive
than nondimensionalization, even though the two terms are sometimes (incorrectly) used
interchangeably.

36
 2.2. DIMENSIONAL ANALYSIS

Each term in a nondimensional equation is dimensionless.

In the process of nondimensionalizing an equation of motion, nondimensional parameters
often appear most of which are named after a notable scientist or engineer (e.g., the
Reynolds number and the Froude number). This process is referred to by some authors
as inspectional analysis.

As a simple example, consider the equation of motion describing the elevation z of an

object falling by gravity through a vacuum (no air drag), as in Fig. 2.2. The initial
location of the object is z0 and its initial velocity is w0 in the z-direction.

Figure 2.2: Object falling in a vacuum. Vertical velocity is drawn positively, so w < 0 for
a falling object.

From high school physics equation of motion is written as,

d2 z
= −g (2.16)
dt2

Dimensional variables are defined as dimensional quantities that change or vary in

the problem. For the simple differential equation given in Eq.(2.16), there are two dimen-

37
 2.2. DIMENSIONAL ANALYSIS

sional variables: z(dimension of length) and t (dimension of time). Nondimensional (or

dimensionless) variables are defined as quantities that change or vary in the problem,
but have no dimensions; an example is angle of rotation, measured in degrees or radi-
ans which are dimensionless units. Gravitational constant g, while dimensional, remains
constant and is called a dimensional constant. Two additional dimensional constants are
relevant to this particular problem, initial location z0 and initial vertical speed w0 . While
dimensional constants may change from problem to problem, they are fixed for a par-
ticular problem and are thus distinguished from dimensional variables. We use the term
parameters for the combined set of dimensional variables, nondimensional variables, and
dimensional constants in the problem.
Equation (2.16) is easily solved by integrating twice and applying the initial conditions.
The result is an expression for elevation z at any time t:

1
Dimensional result: z = z0 + w0 t − gt2 (2.17)
2

The constant 12 and the exponent 2 in Eq.(2.17) are dimensionless results of the inte-
gration. Such constants are called pure constants. Other common examples of pure
constants are π and e.

To nondimensionalize Eq. (2.16), we need to select scaling parameters, based on the

primary dimensions contained in the original equation. In fluid flow problems there are
typically at least three scaling parameters, e.g., L, V , and P0 − P∞ (see Fig. 2.3), since
there are at least three primary dimensions in the general problem (e.g., mass, length,
and time). In the case of the falling object being discussed here, there are only two
primary dimensions, length and time, and thus we are limited to selecting only two scaling
parameters. In a typical fluid flow problem, the scaling parameters usually include a
characteristic length L, a characteristic velocity V , and a reference pressure difference
P0 −P∞ . Other parameters and fluid properties such as density, viscosity, and gravitational
acceleration enter the problem as well.

We have some options in the selection of the scaling parameters since we have three
available dimensional constants g, z0 , and w0 . We choose z0 and w0 . You are invited to
repeat the analysis with g and z0 and/or with g and w0 . With these two chosen scaling

38
 2.2. DIMENSIONAL ANALYSIS

Figure 2.3: In a typical fluid flow problem, the scaling parameters usually include a
characteristic length L, a characteristic velocity V , and a reference pressure difference
P0 − P ∞.

parameters we nondimensionalize the dimensional variables z and t. The first step is to

list the primary dimensions of all dimensional variables and dimensional constants in the
problem,

Primary dimensions of all parameters:

[z] = L [t] = T [z0 ] = L [w0 ] = L/T [g] = L/T 2

The second step is to use our two scaling parameters to nondimensionalize z and t (by
inspection) into nondimensional variables z ∗ and t∗ ,

z wo t
Nondimensionalized variables: z ∗ = t∗ = (2.18)
z0 z0

Substitution of Eq. (2.18)into Eq. (2.16) gives

d2 z d2 (z0 z ∗ ) w02 d2 z ∗ w02 d2 z ∗

= = = −g → = −1 (2.19)
dt2 d(z0 t∗ /w0 )2 z0 dt∗2 gz0 dt∗2

which is the desired nondimensional equation. The grouping of dimensional constants

in Eq. (2.19) is the square of a well-known nondimensional parameter or dimensionless

39
 2.2. DIMENSIONAL ANALYSIS

group called the Froude number,

wo
Froude number: F r = √ (2.20)
gz0

The Froude number also appears as a nondimensional parameter in free surface flows, and
can be thought of as the ratio of inertial force to gravitational force. You should note
that in some older textbooks, F r is defined as the square of the parameter shown in Eq.
(2.20). Substitution of Eq. (2.20) into Eq. (2.19) yields

d2 z ∗ 1
Nondimensionalized equation of motion: ∗2
=− 2 (2.21)
dt Fr

In dimensionless form, only one parameter remains, namely the Froude number. Equation
(2.21) is easily solved by integrating twice and applying the initial conditions. The result
is an expression for dimensionless elevation z ∗ at any dimensionless time t∗ :

1 ∗2
Nondimensional result: z ∗ = 1 + t∗ − t (2.22)
2F r2

Comparison of Eqs. (2.17) and (2.22) reveals that they are equivalent. In fact, for practice,
substitute Eqs. (2.18) and (2.20) into Eq. (2.17) to verify Eq. (2.22).

It seems that we went through a lot of extra algebra to generate the same final result.
What then is the advantage of nondimensionalizing the equation? Before answering this
question, we note that the advantages are not so clear in this simple example because we
were able to analytically integrate the differential equation of motion. In more compli-
cated problems, the differential equation (or more generally the coupled set of differential
equations) cannot be integrated analytically, and engineers must either integrate the equa-
tions numerically, or design and conduct physical experiments to obtain the needed results,
both of which can incur considerable time and expense. In such cases, the nondimensional
parameters generated by nondimensionalizing the equations are extremely useful and can
save much effort and expense in the long run.
There are two key advantages of nondimensionalization. First, it increases our insight
about the relationships between key parameters. Equation (2.20) reveals, for example,
that doubling w0 has the same effect as decreasing z0 by a factor of 4. Second, it reduces

40
 2.2. DIMENSIONAL ANALYSIS

the number of parameters in the problem. For example, the original problem contains
one dependent variable, z; one independent variable, t; and three additional dimensional
constants, g, w0 , and z0 . The nondimensionalized problem contains one dependent pa-
rameter, z ∗ ; one independent parameter,t∗ ; and only one additional parameter, namely
the dimensionless Froude number, F r. The number of additional parameters has been
reduced from three to one.

2.2.5 Dimensional Analysis and Similarity

Non dimensionalizing a mathematical model is a constructive way to formulate the model

in terms of dimensionless quantities only. A big achievement is that dimensional analysis
yields insight in the scaling relations of the system without using knowledge of any gov-
erning equation. Non dimensionalization of an equation by inspectional analysis is useful
only when one knows the equation to begin with. However, in many cases in real-life
engineering, the equations are either not known or too difficult to solve; often times ex-
perimentation is the only method of obtaining reliable information. In most experiments,
to save time and money, tests are performed on a geometrically scaled model, rather
than on the full-scale prototype. In such cases, care must be taken to properly scale the
results. We introduce here a powerful technique called dimensional analysis. While typi-
cally taught in fluid mechanics, dimensional analysis is useful in all disciplines, especially
when it is necessary to design and conduct experiments. You are encouraged to use this
powerful tool in other subjects as well, not just in fluid mechanics. Dimensional analysis
is a method for reducing the number and complexity of experimental variables that affect
a given physical phenomena.
Dimensional analysis is very useful for planning, presentation, and interpretation of ex-
perimental data. As discussed previously, most practical fluid mechanics problems are
too complex to solve analytically and must be tested by experiment or approximated by
computational fluid dynamics (CFD). These data have much more generality if they are
expressed in compact, economic nondimensional form.
The three primary purposes of dimensional analysis are

• To generate nondimensional parameters that help in the design of experiments

41
 2.2. DIMENSIONAL ANALYSIS

(physical and/or numerical) and in the reporting of experimental results.

• To obtain scaling laws so that prototype performance can be predicted from model
performance.

• To (sometimes) predict trends in the relationship between parameters.

Before discussing the technique of dimensional analysis, we first explain the underlying
concept of dimensional analysis the principle of similarity. There are three necessary
conditions for complete similarity between a model and a prototype. The first condition
is geometric similarity the model must be the same shape as the prototype, but may
be scaled by some constant scale factor. The second condition is kinematic similarity,
which means that the velocity at any point in the model flow must be proportional (by
a constant scale factor) to the velocity at the corresponding point in the prototype flow
Figure 2.4. Specifically, for kinematic similarity the velocity at corresponding points
must scale in magnitude and must point in the same relative direction. You may think
of geometric similarity as length-scale equivalence and kinematic similarity as time-scale
equivalence. Geometric similarity is a prerequisite for kinematic similarity. Just as the
geometric scale factor can be less than, equal to, or greater than one, so can the velocity
scale factor.

In Figure 2.4, for example, the geometric scale factor is less than one (model smaller than
prototype), but the velocity scale is greater than one (velocities around the model are
greater than those around the prototype).

The third and most restrictive similarity condition is that of dynamic similarity. Dy-
namic similarity is achieved when all forces in the model flow scale by a constant factor
to corresponding forces in the prototype flow (force-scale equivalence). As with geometric
and kinematic similarity, the scale factor for forces can be less than, equal to, or greater
than one. In Figure 2.4 for example, the force-scale factor is less than one since the force
on the model building is less than that on the prototype. Kinematic similarity is a neces-
sary but insufficient condition for dynamic similarity. It is thus possible for a model flow
and a prototype flow to achieve both geometric and kinematic similarity, yet not dynamic
similarity. All three similarity conditions must exist for complete similarity to be ensured.

42
 2.2. DIMENSIONAL ANALYSIS

Figure 2.4: Kinematic similarity is achieved when, at all locations, the velocity in the
model flow is proportional to that at corresponding locations in the prototype flow, and
points in the same direction.

Advantages of dimensional analysis

1. Reduce the number of variables: Suppose the force F on a particular body shape
immersed in a stream of fluid depends only on the body length L, velocity V , fluid
density ρ and viscosity µ:
F = f (L, V, ρ, µ) (2.23)

In general, it takes about 10 points to define a curve. To find the effects of each
parameter on the force, we need to perform 10 × 10 × 10 × 10 = 104 tests. However,
using dimensional analysis, we can reduce the parameters to only one:
 
F ρV L
=g (2.24)
ρL2 V 2 µ

or
CF = g(Re)

43
 2.2. DIMENSIONAL ANALYSIS

That is, the non-dimensional force is a function of the dimensionless parameter

Reynolds number.
The function g is different mathematically from the original function f , but it con-
tains all the same information. Nothing is lost in a dimensional analysis. And think
of the savings: We can establish g by running the experiment for only 10 values of
the single variable called the Reynolds number. We do not have to vary L, V , ρ, or
µ separately but only the grouping ρV L/µ. This we do merely by varying velocity
V in, say, a wind tunnel or drop test or water channel, and there is no need to build
10 different bodies or find 100 different fluids with 10 densities and 10 viscosities.

2. Non-dimensional equations: that will provide insight on controlling parameters and

the nature of the problem. It suggests dimensionless ways of writing equations before
we waste money on computer time to find solutions. It suggests variables which can
be discarded; sometimes dimensional analysis will immediately reject variables, and
at other times it groups them off to the side, where a few simple tests will show
them to be unimportant. Finally, dimensional analysis will often give a great deal
of insight into the form of the physical relationship we are trying to study.

3. Scaling laws: that allows testing models instead of expensive large full-scale proto-
types. There are rules for finding scaling laws or conditions of similarity. In our
force example, if the similarity condition exists:

Rem = Rep

Then one can write:  2  2

Fp ρp Vp Lp
=
Fm ρm Vm Lm
where subscripts m and p indicate model and prototype, respectively. This equation
is the scaling law: if you measure the model force at the model Reynolds number,
the prototype force at the same Reynolds number equals the model force times the
density ratio times the velocity ratio squared times the length ratio squared.

Do you understand these introductory explanations? Be careful; learning dimensional

analysis is like learning to play tennis: There are levels of the game. We can establish

44
 2.2. DIMENSIONAL ANALYSIS

some ground rules and do some fairly good work in this brief chapter, but dimensional
analysis in the broad view has many subtleties and nuances which only time and practice
and maturity enable you to master. Although dimensional analysis has a firm physical
and mathematical foundation, considerable art and skill are needed to use it effectively.

Example A copepod is a water crustacean approximately 1 mm in diameter. We want to

know the drag force on the copepod when it moves slowly in fresh water. A scale model
100 times larger is made and tested in glycerin at V = 30cm/s. The measured drag on
the model is 1.3N . For similar conditions, what are the velocity and drag of the actual
copepod in water? Assume that Eq. (2.23) applies and the temperature is 200 C.

Solution: We have the following fluid properties

Water (prototype): µp = 0.001kg/(m.s) ρp = 998kg/m3

Glycerin (model): µm = 1.5kg/(m.s) ρm = 1263kg/m3

The length scales are Lm = 100mm and Lp = 1mm. We are given enough model data to
compute the Reynolds number and force coefficient

ρm Vm Lm (1263kg/m3 )(0.3m/s)(0.1m)
Rem = = = 25.3
µm 1.5kg/(m.s)

Fm 1.3N
CF m = 2 2
= = 1.14
ρm Vm Lm (1263kg/m )(0.3m/s)2 (0.1m)2
3

Both these numbers are dimensionless, as you can check. For conditions of similarity, the
prototype Reynolds number must be the same, and Eq. (2.24) then requires the prototype
force coefficient to be the same

998Vp (0.001)
Rep = Rem = 25.3 =
0.001

or
Vp = 0.0253m/s = 2.53cm/s

45
 2.2. DIMENSIONAL ANALYSIS

Fp
CF p = CF m = 1.14 =
998(0.0253)2 (0.001)2
or
Fp = 7.31 × 10−7 N

It would obviously be difficult to measure such a tiny drag force.

2.2.6 The Buckingham Pi-Theorem and Method of Repeating

Variables

The Buckingham Pi-Theorem states that:

If there exists a (physical proper) relation

f (R1 , R2 , ..., Rn ) = 0, (2.25)

between the quantities R1 , R2 , ..., Rn , there also exists an equivalent relation

φ(π1 , π2 , ..., πn−m ) = 0, (2.26)

where n is the number of the parameters on the problem and m is the number of funda-
mental (primary) dimensions.
Note that the theorem assumes that there is a relationship between R1 , R2 , ...; and Rn .
This has to be ensured, or at least assumed, before we apply the theorem. In fact, Buck-
ingham’s Pi-theorem may also prove that no such relation exists. Buckingham Pi-theorem
reduces the number of parameters, and if none of the dimensionless π-s are completely re-
dundant, n − m is also the least possible number of variables we have in our problem. The
variables R1 , R2 , ..., Rn are help to create the dimensionless combinations. These variables
are often called repeating (core) variables. Usually there are several possibilities for the
core (repeating) variables, and what is appropriate depends on the problem. Note that
the number of dimensionless variables will be the same regardless the choice of repeating
variables and combinations.

The Method of Repeating Variables

46
 2.2. DIMENSIONAL ANALYSIS

We have seen several examples of the usefulness and power of dimensional analysis. Now
we are ready to learn how to generate the nondimensional parameters, i.e., the π’s. There
are several methods that have been developed for this purpose, but the most popular (and
simplest) method is the method of repeating variables, popularized by Edgar Buckingham
(1867)-(1940). The method was first published by the Russian scientist Dimitri Riabouch-
insky (1882)-(1962) in 1911. We can think of this method as a step-by-step procedure or
”recipe” for obtaining nondimensional parameters. There are six steps that compromise
the method of repeating variables. These steps are explained in further detail as we work
through a number of example problems.
Step 1: List the parameters (dimensional variables, nondimensional variables, and di-
mensional constants) and count them. Let n be the total number of parameters in the
problem, including the dependent variable. Make sure that any listed independent pa-
rameter is indeed independent of the others, i.e., it cannot be expressed in terms of them.
(E.g., don’t include radius r and area A = πr2 , since r and A are not independent.)
Step 2: List the primary dimensions for each of the n parameters.
Step 3: Guess the reduction j. As a first guess, set j equal to the number of primary
dimensions (m) represented in the problem. The expected number of π’s (k) is equal
to n minus m, according to the Buckingham Pi theorem. If at this step or during any
subsequent step, the analysis does not work out, verify that you have included enough
parameters in step 1. Otherwise, go back and reduce j by one and try again.
Step 4: Choose j repeating parameters that will be used to construct each π. Since
the repeating parameters have the potential to appear in each π, be sure to choose them
wisely.
Step 5: Generate the π’s one at a time by grouping the j repeating parameters with
one of the remaining parameters, forcing the product to be dimensionless. In this way,
construct all k π’s. By convention the first π, designated as π1 , is the dependent π(the
one on the left side of the list). Manipulate the π’s as necessary to achieve established
dimensionless groups.
Step 6: Check that all the π’s are indeed dimensionless. Write the final functional
relationship in the form of the problem.
Let us illustrate the Buckingham’s Pi-theorem to find the relations between the variables.
That is to found dimensionless products.

47
 2.2. DIMENSIONAL ANALYSIS

For example: Let we will explore ”golden oldies” of physics, modeling the small angle,
free vibration of an ideal pendulum Figure 2.5. There are six variables to consider in this
problem, and they are listed along with their fundamental dimensions listed below. In
this case we have n = 6 and m = 3, so that we can expect three dimensionless groups.
Derived quantities Dimensions
Length (l) L
Gravitational acceleration (g) LT −2
Mass (m) M
Period T0 T
Angle θ 1
String tension (T ) M LT −2

Figure 2.5: The classical pendulum oscillating through angles θ due to gravitational
acceleration g.

Step 1
There are six parameters (dimensional variables, nondimensional variables, and dimen-
sional constants) in this problem; n = 6. They are listed in functional form, with the
dependent variable listed as a function of the independent variables and constants:

48
 2.2. DIMENSIONAL ANALYSIS

List of relevant parameters: l = f (m, g, θ, T0 , T ) n = 6

Step 2
The primary dimensions of each parameter are listed here. We recommend writing each
dimension with exponents since this helps with later algebra.

l g m θ T0 T
L LT −2 M 1 T M LT −2

Step 3

As a first guess, j = m is set equal to 3, the number of primary dimensions represented

in the problem (M, L, T ).

If this value of j is correct, the number of π’s predicted by the Buckingham-Pi theorem is
Number of expected π’s: k = n − j = 6 − 3 = 3
Step 4

We need to choose three repeating parameters since j = 3. Since this is often the hardest
(or at least the most mysterious) part of the method of repeating variables, several guide-
lines about choosing repeating parameters are listed above. Depending on the guidelines,
he wisest choice of three repeating parameters is g, m and T0 .
Step 5

Now we combine these repeating parameters into products with each of the remaining
parameters, one at a time, to create the πs. The first π is always the dependent π and is
formed with the dependent variable l.

[π1 ] = [lg a mb T0c ] (2.27)

where a, b and c are constant exponents that need to be determined. We apply the primary
dimensions of step 2 into Eq.(2.27) and force the π to be dimensionless by setting the
exponent of each primary dimension to zero:

[π1 ] = [M 0 L0 T 0 ] = [lg a mb T0c ] = [(L)(LT −2 )a (M )b (T )c ]

49
 2.2. DIMENSIONAL ANALYSIS

Since primary dimensions are by definition independent of each other, we equate the
exponents of each primary dimension independently to solve for exponents a, b and c.




 M0 = Mb ⇒ b = 0

0 = 1 + a ⇒ a = −1


= −2a + c ⇒ c = −2

0

Equation (2.27) thus becomes

π1 = lg −1 T0−2 . (2.28)

Now let’s modify π1 to be raised to power of − 21 . That is:

−1/2 T0 g 1/2 T0
π1 = lg −1 T0−2 = 1/2
=p . (2.29)
l l/g

The second dimensionless product π2 is θ, since it is dimensionless itself, which is angle

of rotation stands alone, that is, it is apparently not related to any of the other variables.
This follows from the assumption of small angles, which makes the problem linear, and
makes the magnitude of the angle of free vibration a quantity that cannot be determined.

π2 = θ. (2.30)

In similar fashion we create the second independent π(π3 ) by combining the repeating
parameters with independent variable T .

[π3 ] = [T g a mb T0c ] (2.31)

This implies that:

[π3 ] = [M 0 L0 T 0 ] = [T g a mb T0c ] = [(M LT −2 )(LT −2 )a (M )b (T )c ]

Now we solve for constants by equating the exponents of each primary dimension.

50
 2.2. DIMENSIONAL ANALYSIS




 0 = 1 + b ⇒ b = −1

0 = 1 + a ⇒ a = −1


0 = −2 − 2a + c ⇒ c = 0


Equation ((2.31))becomes
T
π3 = T g −1 m−1 = . (2.32)
mg
These groups show how the period depends on the pendulum length l and the gravitational
constant g , and the string tension T on the mass m and g

Exercises
1. The lift force F(MLT−2 ) on a missile depends on its length r(L), velocity v(LT −1 ),
diameter D(L), and initial angle θ(L0 M 0 T 0 ) with the horizon; it also depends on
the density ρ(M L−3 ), viscosity µ(M L−1 T −1 ), gravity g(LT −2 ) and speed of sound
s(T ) of the air. Find:

(a) All the dimensionless products by using dimensional analysis.

(b) The force F as a function of the other variable using π’s−Theorem.

2. The English physicist G. I. Taylor watched an amateur film of the first American
atomic bomb explosion in the Nevada desert, and measured the radius r(L) of the
fireball as a function of time t(T ). He argued that r, apart from t, should depend on
the energy E(M L2 T −2 ) that is released in the explosion and the density ρ(M L−3 )
of the air, since the flame-front needs to accelerate the mass of the surrounding air.
Find

(a) The dimensionless product.

(b) Using dimensional analysis find the relation for r and be sure that the relation
is dimensionally computable.

3. Since a right-angled triangle is completely determined by the length of the hy-

potenuse c(L) and the smallest angle α(M 0 L0 T 0 ), there must be a relationship

51
 2.2. DIMENSIONAL ANALYSIS

between surface area A(L2 ) of the triangle, the length of the hypotenuse and the
angle,

f (C, A, α) = 0.

Then,

(a) Find the dimensionless products from the relation.

(b) Find the relation for the area A using Buckingham’s theorem.
(c) Show that there is a relation

c 2 = a2 + b 2 ,

where a and b are lengths of the adjacent and opposite sides of a triangle respectively.

4. Mechanical stress has the same unit as pressure (Force per unit area). For a New-
tonian fluid (like water and air) flowing in the x-direction the so-called shear stress
on a plane parallel to the xy-plane is given by

∂u(x, y, z)
τ =µ ,
∂y

where u is the velocity in the x-direction at (x, y, z). What is the unit for the
constant µ, called the dynamic viscosity?

5. An open cylindrical tank with diameter, D, is filled to height, h, with a fluid of

density, ρ. The bottom has thickness, d, and an elasticity module, E (E is measured
in Pascal, like stress). Because of the weight of the fluid, the bottom will sink
somewhat, most at the center (No sinkage at the rims). Show that the sinking
(distance, δ) in the center of the bottom may be expressed as
 
δ h d E
=f , ,
D D D Dgρ

where g is the acceleration of gravity.

52
 2.2. DIMENSIONAL ANALYSIS

6. The various constants of physics often have physical dimensions (dimensional con-
stants) because their values depend on the system in which they are expressed.
For example, Newton’s law asserts that the attractive force between two bodies is
proportional to the product of their masses divided by the square of the distance
between them, or, symbolically,

Gm1 m2
F = ,
r2

where G is the gravitational constant. Find the dimension of G so that Newton’s

law is dimensionally compatible.

7. A skydiver in free fall with speed U experiences a drag (friction force ) from the
surrounding air. The drag may be written as
√ !
1 u A
Fd = ρair AU 2 f ,
2 ν

where ρair is the density of air, A is the cross-sectional area of the skydiver, and ν
the kinematic viscosity of the air.

(a) Show how this expression for Fd may be found by dimensional analysis.
(b) After a while the free fall jumper will be falling with constant speed. Find
an expression for this speed if we assume that f (x) = 1 (Hint: The force of
gravity, pulling the skydiver downwards, is Fd = mg, where m is the skydiver’s
mass and g the acceleration of gravity. Use that Fg is equal to Fd when the
speed is constant).
(c) Estimate the free fall speed in km/hour if we assume that f (x) = 1.

8. An industrial tank holding a chemical liquid has a hole near the bottom. The
chemical is flowing through the hole at an amount Q, measured in L3 T −1 . It is
reasonable to assume that Q depends on the diameter d of the hole and the pressure
difference ∆p in the fluid between the inner and outer sides of the hole. In addition,
we expect that the flow is governed by the fluid’s density ρ and dynamic viscosity µ.
Use dimensional analysis to show that the expression for Q under these assumptions

53
 2.2. DIMENSIONAL ANALYSIS

may be written
1 1 1
!
d2 ∆p 2 dρ 2 δp 2
Q= f ,
ρ 12 µ

where f is an unknown function of only one variable.

9. The necessary force F to keep a ship at a constant speed U depends on its shape,
primarily the length L, width W , and its depth into the water D. In addition, the
water density, ρ, the viscosity, ν, and the acceleration of gravity, g, play a part.
Use dimensional analysis to find an expression for the force which includes the two
most famous dimensionless numbers in ship design:

• Froude number: Fr = √U ,
Lg

• Reynolds number:Re = LU
ν
.

10. In cheking the dimensions of an equation, you should note that derivatives also
possess dimensions. For example, the dimension of ds
dt
is LT −1 and the dimension of
d2 s
dt2
is LT −2 , where s denotes distance and t denotes time. Determine whether the
equation
  2  
dE 2 dθ dθ
= mr 2
mgrsinθ
dt dt dt

for time rate change of total energy E(M L2 T −2 ) in a pendulum system with damp-
ing force is dimensionally compatible.

54
CHAPTER 3

GRAPHICAL METHODS

Objectives:

By the end of this chapter you will be able to:

• know the use of graphing in mathematical modeling.

• interpret various mathematical models graphically.

• know the use of graphical method in studying stability questions.

• apply graphical method in optimization problems.

• know how to determine the equilibrium points of linear and non linear autonomous
systems.

• understand how to determine the stability analysis of equilibrium points.

• interpret stability analysis mathematically as well as biologically.

55
 3.1. USING GRAPHS IN MODELING

3.1 Using Graphs in Modeling

Graphs can be very useful in modeling if you are aware of their uses and limitations. Since
many people expect either too much or too little from them, we discuss their uses and
limitations before going into specific models.
People can take in an entire picture rather quickly and then deduce consequences by using
their geometric intuition. It follows that graphs should be useful in conveying information.
Those wonderful analog computers people carry in their skulls can rapidly locate certain
patterns in visually presented data. One of the easiest to spot is a straight line. For this
reason a variety of forms of graph paper (rectangular, polar, log-log, normal probability,
etc.) are marketed so that plotted data will appear linear if the anticipated relationship
exists.

Graphs are most useful in conveying qualitative relationships or approximate data which
involve only a few variables. A graphical approach to a problem is most likely to be
useful when not much information is available or when it is given in a rather imprecise
form. Analytical methods are usually more appropriate when more precise information
is available. In complex simulation models, graphs are frequently used to illustrate the
qualitative behavior of several time varying endogenous variables simultaneously. This
helps one obtain a qualitative feel for the behavior of a complicated simulation model.

So far we have talked about graphs primarily as a way of presenting data. Now let’s
consider some major roles graphs play in model formulation.

Since our imagination is limited to three dimensions, graphical representations of the

interrelations of more than three variables are not directly useful. However, it is often
possible to graph a function with most variables held fixed and then determine how the
graph will change when one of the fixed variables is changed. This is the heart of the
geometric approach to comparative statics which is discussed in Section 3.2. The differen-
tial calculus approach parallels the geometric arguments and provides a firm foundation
for making statements when any number of variables is involved. The basic problem of
comparative statics can be stated as follows

• How does the equilibrium point of a system move when certain exogenous variables

56
 3.2. COMPARATIVE STATICS

are changed ?

For example, how will the output of a firm be affected by a higher tax rate ?

Graphical methods are also useful in studying stability questions. The analytical treat-
ment of local and global stability theory is not easy. Therefore it is desirable to use
graphical methods whenever possible to suggest and perhaps prove results. Section 3.3
touches on this approach.
As a glance at the figures in this chapter shows, the intersections of curves are of major
importance in comparative statics. This is because they determine the equilibrium points.
A subtler observation is that slopes of curves play a central role in stability questions.
The slope of a curve is a rate, and rates play a crucial role in stability theory.

3.2 Comparative Statics

Comparative statics are rates of changes in equilibrium values of endogenous variables

in response to changes in exogenous variables. A large part of the mathematical modeling
we do in this regard is concerned with comparative statics, that is, the comparison of
different equilibrium states that are associated with different sets of values of parameters
and exogenous variables. To make such a comparison, we always start by assuming an
initial equilibrium state. We then allow some kind of disequilibrating change in the model
(through some change in a parameter or an exogenous variable). When this occurs, the
initial equilibrium will of course be upset, and so the endogenous variables will have to
adjust. The question we concern ourselves with in comparative statics is

• How will the new equilibrium position compare with the old?

NB: When we study comparative statics, we simply compare the initial (pre-change)
equilibrium position to the post-change equilibrium position. We cannot say anything
about the process of adjustment. Our comparative static analysis can be either quantita-
tive or qualitative. If our analysis is purely qualitative, this means that we will only be
able to talk about the direction of the change that occurs. If it is quantitative, we will
actually be able to talk about the magnitude of the change that has occurred. (Obviously,

57
 3.2. COMPARATIVE STATICS

if we know the magnitude, we will also know the direction of the change, so in effect, the
quantitative analysis involves a qualitative element as well).
The crux of all of this is that in doing comparative statics, we are looking for a rate of
change, namely the rate of change of the equilibrium value of an endogenous variable
with respect to a change in the particular parameter or exogenous variable. (How does
the endogenous variable change in response to a change in the exogenous variable, or a
change in the parameter) To get at this, we will make use of the concept of a derivative,
a concept that is concerned with rates of change.
Example 1: We analyse the comparative statics of the equilibrium of economic models.
This model is the familiar supply and demand framework in a market producing one
good. We will now use calculus to look at how changes in the intercept and slopes of
demand and supply functions affect equilibrium price and quantity The emphasis on the
word equilibrium is important here as it indicates that what we interested in is how
the solution to the demand and supply functions changes when you change one of the
parameters of the model. (i.e. comparative static analysis).

Qd = Qs
Qd = a − bP (a, b > 0)
Qs = −c + dP (c, d > 0)

We solve the equilibrium price and quantity.

a+c
P∗ =
b+d
ad − bc
Q∗ =
b+d

We are now interested in analysing the comparative statics of the model.

∂P ∗ 1
= >0
∂a b+d

58
 3.2. COMPARATIVE STATICS

This tells us P ∗ will increase (decrease) if a increases (decreases).

∂P ∗ −(c + a)
= <0
∂b (b + d)2

This tells us P ∗ will decrease (increase) if b increases (decreases).

∂P ∗ 1
= >0
∂c b+d

This tells us P ∗ will increase (decrease) if c increases (decreases).

∂P ∗ −(c + a)
= <0
∂d (b + d)2

This tells us P ∗ will decrease (increase) if d increases (decreases).

Find the partial derivatives of Q∗ and check your results using graphs.
Example 2: The Nuclear Missile Arms Race

The United States and the U.S.S.R. both feel that they require a certain minimum number
of intercontinental ballistic missiles (ICBMs) to avoid ” nuclear blackmail.” The idea is to
ensure that enough missiles will survive a sneak attack so that ” unacceptable damage ”
can be inflicted on the attacker. Given this philosophy, it is claimed by some and denied
by others that the introduction of antiballistic missiles (ABMs) and/or multiple warheads
on each missile (MIRVs) will cause both nations to increase their stock of missiles. Is
this true ? What about making missiles less vulnerable to attack by hardening silos or
building missile firing submarines? The wrong answers to these questions could have
drastic consequences. Who is right ?
Obviously we cannot hope to settle the debate. However, a simple graphical model can
shed some light on the problems involved and hopefully help lead to more intelligent
debate. The following discussion is adapted from T. L. Saaty (1968, pp. 22-25).
We deal with two countries which we call country 1 and country 2. Let x and y be the
number of missiles possessed by countries 1 and 2,respectively. We treat x and y as real
numbers. Of course they are actually integers ; but since they are large, the relative errors

59
 3.2. COMPARATIVE STATICS

introduced by treating them as real numbers will be small ; for example, the percentage
difference between 500 and 500.5 is quite small. For the time being we assume that all
missiles are the same and are equally protected. From the above discussion it follows that
there exist continuous, increasing functions f and g such that country 1 feels safe if and
only if x > f (y), and country 2 feels safe if and only if y > g(x). These functions are
plotted in Figure (3.1) . The shaded region is the area in which armaments are stable,
since both countries feel they have sufficient weapons to prevent a’sneak attack. We
consider questions such as : Does such a region actually exist ? What affect do such
things as ABMs, MIRVs, and so on, have on the point A = (xm , ym )? First we show

Figure 3.1: Country 1 introduces ABMs. A = initial status (shaded area stable) ; B =
country I protects its missiles; C = country I protects its cities. Axes show number of
missiles.

that the solid curves in Figure 3.1 are qualitatively correct. Let’s look at things from the
point of view of country 1. A certain number of missiles xo is needed to inflict what is
considered unacceptable damage on country 2. When country 2 has no missiles, country
1 requires xo .

60
 3.2. COMPARATIVE STATICS

We show that for any r > 0 the curve x = f (y) crosses the line y = rx. It suffices to show
that there is a function x(r) such that, whenever x ≥ x(r) and y = rx,country 1 believes
that it has enough missiles so that the number surviving a sneak attack by country 2 will
be able to inflict unacceptable damage on country 2. In other words, country 1 wants to
be practically certain of at least xo of its missiles surviving a sneak attack by country 2.
Suppose that y = rx. To destroy the most missiles, country 2 should aim about r missiles
at each of country 1 ’s missiles. Since a warhead may fail to reach and destroy its target,
there is some probability, p(r) > 0, that a given missile belonging to country i will survive
a sneak attack. Thus country 1 can expect xp(r) missiles to survive. For large enough
x = x(r), this will exceed xo by an amount large enough to allow for uncertainties. This
completes the proof that the curves intersect. Thus the curve x = f(y) starts at (xo , 0)
and curves upward with a slope increasing to 0 . By a symmetry argument, y = g(x) has
the form shown, with a slope decreasing to 0. Two such curves meet at exactly one point
which we call (xm , ym ), the minimum stable values for x and y.
This analysis applies to all the situations discussed below, so there is always a unique
minimum stable point. We want to know how its position compares with (xm , ym ).
Suppose the missiles of country 1 are made less vulnerable to sneak attack by the use of
hardened silos, ABM protection, or some other means. This increases p(r), the probability
that any given missile belonging to country 1 will survive a sneak attack. Hence the curve
f(y) moves to the left with the point xo fixed. The shape of the curve is altered somewhat
in the process. The new curve is shown dashed in Figure (3.1) . We can see that both
countries require fewer missiles for stability.
Suppose that country 1 protects its cities by some device such as ABMs. Country 2 now
requires more than yo missiles to inflict unacceptable destruction on country 1 . Thus
the curve g(x) moves upward as shown by the × − × − × curve in Figure (3.1). Both
countries require more missiles for stability.

What happens if multiple warheads are installed ? This situation is more complicated
than the previous two. Suppose country 1 replaces the single warheads on each of its
missiles with N warheads. It will then require that fewer of its missiles survive a sneak
x0
attack. (The number required is about N
.) Thus x = f(y) moves to the left as in Figure
(3.2). Country 2 will be faced with N times as many warheads in a sneak attack, so

61
 3.2. COMPARATIVE STATICS

from its point of view the scale of the x axis has changed by about a factor of N, as
shown in Figure (3.2). It appears that country 2 will require more missiles, and country 1
will require fewer ; however, this depends on the detailed shape of the curves. Therefore
probabilistic models should be used instead of, or in conjunction with, graphical ones.
This would require us to make more precise assumptions regarding the capabilities of the
missiles, so we do not go into it here. It seems unreasonable to assume that country 2 will

Figure 3.2: Country 1 introduces MIRVs. Axes show number of missiles.

not also develop and deploy multiple warheads if country 1 does. Therefore we should
analyze the situation in which both countries deploy multiple warheads. There are two
conflicting effects :

1. Since the axes measure missiles, the points [f (0), 0] and [0, g(0)] will move toward
the origin, tending to decrease (xm , ym ).

2. f(y) becomes more horizontal and g(x) becomes more vertical, tending to increase
(xm , ym ).

62
 3.2. COMPARATIVE STATICS

We cannot decide without further information which effect will dominate. T. L. Saaty
(1968, p. 24) presents an analytical model which leads to the conclusion that both coun-
tries will require many more missiles. In the above discussion, we assumed that all missiles
were the same. This is unrealistic. If we drop this assumption, each country will change
its strategy by aiming different numbers of missiles at the various enemy missiles. Of
these, some targeting makes the expected surviving firepower a minimum. This targeting
gives the curves for Figure (3.1), and the analysis proceeds as before.

You may be interested in the article by K. Tsipis (1975a) which contains a discussion of
the technology behind ultraaccurate MIRVs.

Example 3: Biogeography : Diversity of Species on Islands

The diversity of species varies considerably from place to place, even when the habitats
appear to be the same. Conservationists have argued that the size of a region is important
for diversity, and so they often favor a few large wilderness areas rather than many tiny
ones. The subject is far from understood. We study one corner of it briey.
The world is broken into patches of diering habitats. Often a habitat of a species finds
acceptable is surrounded by a large expanse of unacceptable territory. Examples are alpine
meadows, farm woodlots, lakes, game preserves, and islands. The following discussion is
confined to islands ; however, most of the ideas and results apply to other types of isolated
habitats. The material is adapted from R. H. MacArthur and E. O. Wilson (1967, Ch.
3) which treats the subject in much greater depth.
Studies have indicated that, the size of an island is an important factor in determining
the number of species the island is likely to contain. Also, islands closer to the mainland
tend to contain a greater variety of species than more isolated islands. It seems reasonable
that the effects of migration of species and extinction of species (on islands) can account
for this. We develop this idea and briefly consider some of its consequences.

A species can become established on an island only by migrating to it and prospering

there. An organism migrates by flying, being carried, drifting on currents, and so on.
Since a population on an island is relatively small, it can die out because of random
variations in the environment. As a result we expect the list of species present on an
island to change much faster than the list of species present on the mainland.

63
 3.2. COMPARATIVE STATICS

This is somewhat vague. Does a flock of migrating birds that stops on the island for a day
or a season become established and then die out ? Even if a species ” intends ” to stay on
the island, we are still faced with the problem of what we mean by ” become established
” if the island is too small to support a large population, the species will always be on
the verge of extinction. When is a species established in this case ? Since we are dealing
with a fairly crude model, we can afford to ignore these problems. A more refined model
would have to come to grips with them.

If we completely understood all the aspects of the situation (e.g., the biological, geo-
graphical, and meteorological), we could determine the probability of a particular species
composition being present on the island at a given future time. These would be tremen-
dously complex calculations involving vast quantities of data, and this approach would
be hopeless.
Let’s combine practically all the endogenous variables into one measure : the total number
of species present on the island. It seems reasonable to suppose that this should vary
around some average number of species in a steady state situation. We discuss this
average. For a discussion of transient behavior see R. H. MacArthur and E. O. Wilson
(1967, Ch. 3) or E. O. Wilson and W. H. Bossert (1971, Ch. 4).
When the number of species present on the island is in equilibrium, migration and ex-
tinction cancel out numerically; that is, the rate of migration of new species to the island
equals the rate of extinction of species already on the island. These rates depend in a
complicated way on the species present, the season, and many other factors. If we regard
a year as a short period of time, seasonality will present no problem. In this sort of crude
averaging over many species, which species are actually present probably doesn’t matter
much. Therefore it makes sense to talk about rates in a crude way independent of which
species are actually present on the island.
In Figure (3.3), the number of species N on the island versus the migration and extinction
rates are plotted. The two smaller graphs illustrate the effect of distance from the main-
land and the effect of island size. We discuss the reasons for the shapes and positions of
the curves.

64
 3.2. COMPARATIVE STATICS

Let’s consider the extinction rate curves. When more species are present on the island, the
chances that at least one species will become extinct in a given time are greater. Hence,
the extinction rate curves have a positive slope. Since extinction rates depend only on
the island and the species present, the extinction curve is not affected by the distance
from the mainland. However, we can expect that a species is more likely to die out on a
small island because the lack of space keeps the population lower. Thus, the extinction
rate curves shift upward as the islands become smaller.

Why do the migration rate curves have a negative slope ? The migration rate relates to
species not present on the island. The greater the diversity on the island, the smaller
the pool of potential migrating species on the mainland. Hence the chances of migration
decrease as the number of species on the island increases. Migration rates depend on the
distance of the island from the mainland and on the size of the island. The rates decrease
with distance, because any given organism is less likely to reach the island. The rates
increase with island size because;

1. an organism has a larger land area as a target and

2. an organism is more likely to be able to establish itself on a larger island.

It follows from Figure (3.3) that, the number of species present increases with island
size and decreases with distance from the mainland. This is not so surprising, since we
practically put these results in as initial assumptions.
We can say something about species turnover by looking at the graphs a bit more. Note
that the equilibrium extinction rate (which equals the equilibrium migration rate) is
greater for near islands than for far islands. Hence the species composition for two islands
of equal size should change more rapidly on the island closer to shore. If the effect of
island size on migration rate is not too great, we can similarly conclude that the species
composition changes faster on small islands than on large islands. Since small islands
have fewer species at equilibrium than large islands, this effect should be quite noticeable.

There is some data supporting the conclusions that species turnover is relatively and
absolutely more rapid on smaller islands. R. H. MacArthur and E. o. Wilson (1967, pp.
52-54) discuss the results of two botanical surveys of some small islands off the Florida

65
 3.2. COMPARATIVE STATICS

Figure 3.3: Migration .and extinction curves for islands. (a) Typical curves. (b) Effect of
distance. (c) Effect of size.

Keys. The first survey (1904) was conducted quickly and so may be incomplete. Since
the 1916 survey was quite complete, the species present in 1904 and absent in 1916 give
some measure of the turnover rate. Unfortunately, the data involve only six islands, two
of which are very small. .
R. H. MacArthur and E. O. Wilson (1967, pp. 55-60) also report on some results of R.
Patrick. She suspended glass slides in a spring in Pennsylvania and counted the number of
diatoms of various species that were present. The glass slides can be thought of as islands.
Four experiments were done two times each : A glass slide with an area of either 12 or 25
square millimeters was placed in the water for either 1 or 2 weeks. The slides submerged
for 1 week had more species present than those submerged for 2 weeks. We can explain
this apparent contradiction by observing that as a barren area becomes more populated
the interaction between species may cause extinction. This was not allowed for in our

66
 3.2. COMPARATIVE STATICS

model. Clearly care must be taken in modeling islands that are far from equilibrium.
Because of this, we do not consider the 1 week data further. We can check out two
predictions using the 2 week data :

1. Larger area implies more species : The smaller slides had 24 and 21 species, and
the larger had 29 and 28.

2. Smaller area implies a higher migration rate : The migration rate may be reflected
somewhat in the differences in the species composition of the slides. (Why ?) Seven
species appeared on one but not both of the smaller slides. For the larger slides the
number was one.

Example 4: Theory of the Firm

Suppose you are the manager of a firm which produces, among many other items, ”zowies.”
How can you decide on a level of production? The price of the main raw material for your
zowies is going to increase. Perhaps, you can pass some of the cost on to your customers.
How much? Can you pass on enough to make it worthwhile to continue manufacturing
zowies? Quantitative results are hard to obtain because data collection is extremely
difficult; however, we can obtain a qualitative picture of the situation fairly easily. In
the usual theory of the firm it is assumed that the manager of the firm has complete
information, that his decisions are carried out, and that he acts so as to maximize the
profits of the firm. There is an ongoing debate about the usefulness of these assumptions,
but we don’t want to get into that here. If you are interested in the subject, see R. M.
Cyert and 1. G. March (1963, pp. 5-16) for a discussion of both sides of the question.
In addition to the above assumptions, we generally assume (as is often done in economic
theory) that the functions with which we are dealing are well behaved; that is, they are
continuous and usually differentiable.
The theory of the firm is discussed in most textbooks on mathematical economics. There
also exist books devoted exclusively to the topic, such as K. J. Cohen and R. M. Cyert
(1965). Consult such sources if you wish to see the ideas in this example developed further.
For simplicity we assume that the firm produces only one product, so that we can speak
unambiguously of the level of production. It is measured in units per time period, where
a time period can be a day, a month, or any other convenient interval. We want to find a

67
 3.2. COMPARATIVE STATICS

way to determine the level of production, so that we can discuss the influence of changing
costs and prices on the production level.
Suppose the production of the firm is at some equilibrium level. Since profits are being
maximized, the additional cost that would be incurred in raising production slightly is
equal to the additional gross income that would be obtained by marketing these additional
units of the product. You should convince yourself that this is simply a restatement of
the calculus theorem that the function

Total gross income - Total cost

has maxima and minima where its derivative vanishes.

The additional cost required to produce one additional unit is called the marginal cost,
and the additional income is called the marginal income. In general, both marginal
cost and marginal income are functions of the level of production. We have shown that
marginal cost equals marginal income at equilibrium. This equality could imply that the
profit is a minimum instead of a maximum. How can we distinguish one from the other
? If we move away from the equilibrium, profits must decrease. Thus the marginal cost
curve must lie above the marginal income curve for higher production levels and below it
for lower production levels. This is shown in Figure 3.4 where the horizontal axis is the
quantity produced per unit time. This is the basic result with which we work.

Although marginal cost and marginal income may seem to be straightforward concepts,
they can be a bit fuzzy. During a short period of time (the short term), wages and the
cost of raw materials are fixed costs, because they have been contracted for ; consequently,
they do not enter into marginal calculations. From a slightly longer point of view, they
are both variable costs and so enter into the marginal costs. Equipment depreciation
is a fixed cost ; but maintenance, fuel, and replacement costs enter into the marginal
calculations. Since our marginal curves vary with how long a view we take, the optimum
level of output may depend on the length of time we want to consider. In the following
discussion we make the vague assumption that the manager is concerned with the firm’s
profits over a reasonably long time interval. As long as we don’t try to make any detailed
applications, we can afford to be vague.

68
 3.2. COMPARATIVE STATICS

Figure 3.4: Marginal cost and income curves. Axes show quantity produced per unit time
and dollars per unit time.

What effect will taxation have on production? If the firm is required to pay a lump
sum tax independent of production (e.g., a property tax), the marginal curves will not
be affected. Hence the production level will be unchanged. If the firm is required to
pay a tax that depends on the level of production (e.g., an income tax or a value-added
tax), the result will depend on whether or not the tax is passed on to the consumer. If
it is not passed on, the marginal cost curve will rise. We have shown that the marginal
cost curve intersects the marginal income curve from below at a maximum. It follows
that the new intersection will be to the left of the old one. Therefore the production
level will decrease.What will happen if the tax is passed on to the consumer ? In this
case both marginal curves will move upward by an amount equal to the tax per unit of
production, and the production level will be unchanged. The above result on taxation
can be generalized considerably, and we can state another result on the income side :
The optimum level of production moves in the opposite direction from the marginal cost
and moves in the same direction as the marginal income.

Convince yourself that this is true by giving a graphical argument. Suppose the price of
raw materials increases. This raises the marginal cost, so the production level tends to

69
 3.2. COMPARATIVE STATICS

decrease. Decreased production may cause consumers to drive up the cost (per unit) of
the product, thereby increasing the producer’s marginal income. Consequently the level
of production will rise. Since the product now costs more, the amount purchased by
consumers will probably be less. Thus the increase in cost will not be quite enough to
push production back to its original level. We discuss this in terms of supply and demand
curves.
In industries where the number of firms is large, it is reasonable to suppose that the price
per unit of product does not depend on the amount any single firm produces. In this
case the marginal income curve is horizontal. The marginal cost curve is then the supply
curve for the firm’s product, since at a price p the firm produces the quantity Q at which
the marginal cost equals p. Since the marginal income curve is horizontal, our earlier
discussion shows that the supply curve must have a positive slope to ensure stability.
This agrees with the intuitive notion that higher selling prices lead to greater production.

The demand curve is the amount of the product that will be purchased at a given price.
Usually demand falls as price increases. Figure (3.5) shows typical supply and demand
curves. At equilibrium, the quantity purchased must equal the quantity sold. Hence the
intersection of the supply and demand curves gives the equilibrium values of price and
quantity.
From Figure (3.5) we can see how much of the increased marginal costs will be passed on
to the consumer. The dashed curve shows the supply curve (marginal cost curve) after
the marginal costs have increased. The flatter the demand curve, the greater the fraction
of the increase the producer must absorb. What does a flat demand curve mean ? It
indicates that consumer buying patterns are very sensitive to price. Thus, if consumer
buying patterns are insensitive to price, you can pass most of your increased expenses on
to the consumer.
What about the theory of a firm that produces several products? It is better to study
such a situation using tools from calculus. However, our graphical analysis indicates the
sort of results we can expect to find in this case.

70
 3.3. STABILITY QUESTIONS

Figure 3.5: Supply and demand curves. Increased marginal costs shift supply curve
upward to dashed position.

3.3 Stability Questions

Cobweb Models in Economics

We consider the dynamics of supply and demand when there is a fairly constant time lag
in production as, for example, in agriculture. It has been observed that there are fairly
regular price fluctuations in such situations. This situation was studied by economists
in the 1920s and 1930s. The problem contrasts sharply with the theory of the firm in
Section 3.2, where we ignored time. The following discussion is adapted from M. Ezekiel
(1937/8).
When a commodity is marketed, the selling price is determined by the demand curve.
This price is one of the factors producers use in determining how to alter production. In
a ”pure” situation, they produce the amount on the supply curve that corresponds to the
present price. (Supply and demand curves are discussed in detail in the theory of the firm
model in Section 3.2. There were interested in the intersection point of the curves.) Thus
(see Figure (3.6) ), if the amount of potatoes produced in year 1 is q1 , the price per bushel
will be p1 . As a result, farmers will decide to produce the amount q2 in year 2, the market

71
 3.3. STABILITY QUESTIONS

will set a price p2 per bushel for this crop, and so on. Because of the picture(Figure 3.6),
this idea is referred to as the cobweb theorem. In practice one does not know the supply
and demand curves, but the above model predicts that the demand curve can be obtained
by plotting (qn , pn ) and the supply curve by plotting (qn , pn−1 ).

Figure 3.6: The cobweb model.

How realistic is this model? The existence of a supply curve assumes that producers can
control output perfectly. This is not true in the agricultural sector where weather is very
important, but it may be a reasonable approximation. If the supply and demand curves
move erratically, the model will be upset. Changes in prices for other goods the supplier
may produce, sudden changes in demand (e.g., the sale of wheat by the United States to
the U.S.S.R. in 1972), and sudden changes in supply (e.g., crop blights) may cause this to
happen. If the suppliers have some understanding of price fluctuations, they will not raise
production levels much in spite of higher prices. However, this does not wreck the model.
In this case the supply curve will be nearly independent of price near the equilibrium
price, but the model will still apply. It predicts small fluctuations in supply and a rapid
approach to stability.
Phase Planes
The previous model dealt with the stability of a difference equation. A similar procedure
is used for differential equations. This requires the notion of a phase plane. Suppose we

72
 3.3. STABILITY QUESTIONS

are dealing with the two equations

x0 = f (x, y), y 0 = g(x, y). (3.1)

At each point (x, y) in the xy plane we can plot a vector proportional to (x0 , y 0 ). This is
called the direction field of (3.1). To graph a solution of (3.1) we then start at an initial
point and follow a path parallel to the direction field. (Since the direction field varies from
point to point, the path is usually curved.) The speed is determined by the magnitude of
the vector tangent to the path at that point. If we start at a point with f = g = 0, we
will not move from it. Such points are called equilibrium points. Since we have only
crude information about f and g, our phase plane diagrams cannot be this detailed. To
answer stability questions, it is often sufficient to plot the two curves f = 0 and g = 0 and
indicate roughly the vectors (x0 , y 0 ) in the neighborhood of these curves. The intersections
of the curves are the equilibrium points of (3.1). The curve f = 0 divides space into two
regions such that x0 > 0 in one and x0 < 0 in the other. If you determine which region
is for f = 0, and likewise for g = 0, the rest will be easy. The vectors cross f = 0
vertically, and the direction will be upward if and only if g > 0. Similarly, they cross g
= 0 horizontally, and the direction will be rightward if and only if f > 0. In plotting
f = 0 and g = 0, it is helpful to determine the slopes of the curves. This can be done by
implicit differentiation: For f = 0,

dy ∂f /∂x
=− ,
dx ∂f /∂y

and similarly for g = 0. It is important to remember that the partial derivatives for the
slope of f = 0 are evaluated at values of x and y at which x is at equilibrium ; that is,
x0 = 0. (This is important in determining the sign of ∂f /∂x.) The partial derivatives also
help to decide which region is corresponds to f > 0 and which is to f < 0 : f > 0 to the
right of (or above) f = 0 if and only if ∂f /∂x > 0 (or ∂f /∂y > 0).
Small - Group Dynamics
Suppose you wish to set up a local committee to help elect a candidate to office. What
keeps a group together and working ? Does more work improve a task-oriented group or
harm it? Very little mathematical modeling has been done in this area and, unfortunately,

73
 3.3. STABILITY QUESTIONS

the following is rather crude and lacking in practical advice.

We want to study the stability and comparative statics of a group which has a required
activity imposed from the outside (a task). The model is taken from H. Simon (1952),
who based it on a nonmathematical model proposed by G. C. Homans (1950).
There are four basic functions of time :
I(t), the intensity of interaction among the group members.
F (t), the level of friendliness among the group members .
A(t), the amount of activity within the group.
E(t), the amount of activity imposed on the group by the external environment.
The variables can be treated as averages over all group members or as some overall
measure for the entire group. We regard I, F , and A as endogenous variables and E as
an exogenous variable which we generally treat as being constant.

To make the concepts more concrete, let’s consider an example. The imposed activity
E is the laying in firewood. The group may be engaged in this for wages, or they may
be friends preparing for winter. The various activities A include locating wood sources,
sawing logs, stacking logs and setting up a football pool. Note that some activities may
not be directed toward the externally imposed task. G. C. Homans (1950, p.101) says,
”By our definition interaction takes place when the action of one man sets off the action of
another.” ” Action” here refers to activity, so that activity is required for interaction, but
not conversely-a person can work alone. The many situations in our example that involve
interaction include discussing where to obtain wood, working opposite ends of a saw while
cutting logs, passing wood from one person to another in stacking, and conversing idly.
Some of the interaction is necessary, but a lot of it can be reduced considerably. The same
is true of activity, as any efficiency expert knows ; however, this may involve changes in
habit patterns and so require more time.
There are three relations on which the model is based:

1. I(t) depends on A(t) and F (t) in such a way that it increases if either A or F does.
The adjustment is practically instantaneous.

2. F (t) depends on I(t). It tends to increase when it is too low for the present level of

74
 3.3. STABILITY QUESTIONS

interaction and to decrease if there is not enough interaction to sustain its present
level. This adjustment requires time, and the rate of adjustment is greater when
the discrepancy between present and equilibrium levels is greater.

3. A(t) depends on F (t) and E(t). It tends to increase when it is too low for the present
level of F or E and to decrease when it is too high. This adjustment requires time,
and the rate of adjustment is greater when the discrepancy between present and
equilibrium levels is greater.

Criticize the assumptions.

These assumptions can be turned into equations :

∂r ∂r
I(t) = r(A, F ), > 0, > 0, (3.2a)
∂A ∂F

∂s ∂s
F 0 (t) = s(I, F ), > 0, < 0, (3.2b)
∂I ∂F

∂ψ ∂ψ ∂ψ
A0 (t) = ψ(A, F, E), < 0, > 0, > 0. (3.2c)
∂A ∂F ∂E

The reasoning behind ∂s/∂F < 0 and ∂ψ/∂A < 0 deserves an explanation. The same
idea applies to both cases. Let’s consider Φ. If A, F, and E are at some level, ψ = A0 will
be determined. If we now increase A, we will either reduce the pressure for A to increase
(if ψ > 0) or increase the pressure for A to decrease (if ψ < 0). In either case ∂ψ/∂A < 0.
By substituting (3.2a) into (3.2b) we obtain

∂φ ∂s ∂r ∂φ ∂s ∂s ∂r
F 0 (t) = φ(A, F ), = / >0⇒ = + / . (3.3)
∂A ∂I ∂A ∂F ∂F ∂I ∂F

This equation says that a high level of A tends to cause F to increase. The effect of a high
level of F is ambiguous : It may tend to cause F to increase or decrease. The statement
that ∂φ/∂F > 0 can be interpreted as : ”The greater the friendliness; the faster it tends
to increase (or the slower it tends to decrease, if it is decreasing).” While this may be true
at some points in the AF plane, it is unlikely to be true when F is large because of limits

75
 3.3. STABILITY QUESTIONS

on friendliness. We assume that ∂φ/∂F < 0 everywhere. The curves ψ = 0 and φ = 0 are
plotted in Figure 3.7. The slope of the curves is positive, since, for example, on the curve
ψ = 0, dF/dA = −(∂ψ/∂A)/(∂ψ/∂F > 0. The slope of the ψ = 0 curve is increasing,
because we assume a saturation effect : When A and F are both large and A0 = 0, a
fairly large increase in F is required to balance a small increase in A. In other words, the
group tends to resist increases in activity more when it is already quite active. Discuss
the curve φ = 0. Verify the general shape of the direction field shown in the figure. It

Figure 3.7: Dynamics in the activity-friendliness plane.

can be seen that the upper equilibrium point is stable and that the lower one is unstable.

We now consider the effect of changing E. We have

∂ψ ∂ψ ∂ψ
∆ψ ≈ ∆A + ∆F + ∆E.
∂A ∂F ∂E

Since ∆ψ = 0 on the curve ψ = 0, it follows from (3.2c) that, when ∆A = 0, ∆E and

∆F have opposite signs. Thus the ψ = 0 curve moves downward as E increases. Hence
The equilibrium levels of A and F are increasing functions of E.
If the ψ = 0 curve moves sufficiently far up, it will no longer intersect the φ = 0 curve,

76
 3.4. SCALING

and so there will be no equilibrium point. In this case the group will not continue to exist.
Consequently it is possible that a group will break up if externally imposed activity falls
below a certain level.

3.4 Scaling

Scaling is an extremely useful technique in mathematical modeling and numerical sim-

ulation. After establishing a mathematical model in the form of an equation, it will be
necessary to introduce dimensionless variables. Usually it is not difficult to do this, but
it can be carried out in several ways, and it is not always easy to see what is most ap-
propriate way. However, there exists an intelligent way of doing this called scaling the
equations.
When the equations are scaled, it is easy to see which parts are important and which are
less important. It may be difficult to scale equations, and in any case this will depend on
the problem we are considering, even if the equation is basically the same all the time.
Somewhat simplified we can say that the scales force us to think about the situation, and
in this way we gain insight into what we are doing.

The purpose of the technique is three-fold:

1. Make independent and dependent variables dimensionless.

2. Make the size of independent and dependent variables about unity.

3. Reduce the number of independent physical parameters in the model.

The first two items implies that for any variable, denote it by u, we introduce a corre-
sponding dimensionless variable
u − u0
u∗ = ,
uc
where u0 is a reference value of u (u0 = 0 is a common choice) and uc is a characteristic
size of |u|, often referred to as ”a scale”. Since the numerator and denominator have the
same dimension, u∗ becomes a dimensionless number.

77
 3.4. SCALING

If uc is the maximum value of |u − u0 |, we see that 0 < |u∗ | < 1. How to find uc is
sometimes the big challenge of scaling. Examples will illustrate various approaches to
meet this challenge.

The many coming examples on scaling differential equations contain the following peda-
gogical ingredients to meet the desired learning outcomes.

• Learn the technic al steps of making a mathematical model, based on differential

equations, dimensionless.

• Describe various techniques for reasoning about the scales, i.e., finding the charac-
teristic sizes of quantities.

• Learn how to identify and interpret dimensionless numbers arising from the scaling
process.

• Provide a lot of different examples on making models dimensionless with physically

correct scales.

The Basic Model Problem

Processes undergoing exponential reduction can be modeled by the ODE problem

u0 (t) = −au(t), u(0) = I, (3.4)

where a, I > 0 are prescribed parameters, and u(t) is the unknown function. For the
particular model with a constant a, we can easily derive the exact solution, u(t) = Ie−at ,
which is helpful to have in mind during the scaling process.

Example 1: Population dynamics. The evolution of a population of humans, animals,

cells, etc., under unlimited access to resources, can be modeled by Eq. (3.4). Then u is the
number of individuals in the population, strictly speaking an integer, but well modeled
by a real number in large populations. The parameter a is the increase in the number of
individuals per time and per individual.
Example 2: Decay of pressure with altitude. The simple model (3.4) also governs
the pressure in the atmosphere (under many assumptions, such air is an ideal gas in

78
 3.4. SCALING

equilibrium). In this case u is the pressure, measured in N m−2 ; t is the height in meters;
and a = M/(R∗ T ), where M is the molar mass of the Earth’s air (0.029kg/mol), R∗ is
the universal gas constant (8.314N m/molK), and T is the temperature in Kelvin (K).
The temperature depends on the height so we have a = a(t).

The Technical Steps of the Scaling Procedure

Step 1: Identify independent and dependent variables. There is one independent

variable, t, and one dependent variable, u.
Step 2: Make independent and dependent variables dimensionless. We introduce
a new dimensionless t, called t∗ , defined by

t
t∗ = , (3.5)
tc

where tc is a characteristic value of t. Similarly, we introduce a dimensionless u, named

u∗ , according to
u
u∗ = , (3.6)
uc
where uc is a constant characteristic size of u. When u has a specific interpretation, say
when (3.4) models pressure in an atmospheric layer, uc would be referred to as character-
istic pressure. For a decaying population, uc may be a characteristic number of members
in the population, e.g., the initial population I.

Step 3: Derive the model involving only dimensionless variables. The next task
is to insert the new dimensionless variables in the governing mathematical model. That
is, we replace t by tc t∗ and u by uc u∗ in (3.4). The derivative with respect to t∗ is derived
through the chain rule as

du d(uc u∗ ) dt∗ du∗ 1 uc du∗

= = u c = .
dt dt∗ dt dt∗ tc tc dt∗

The model (3.4) now becomes

uc du∗
= −auc u∗ , uc u∗ (0) = I. (3.7)
tc dt∗

79
 3.4. SCALING

Step 4: Make each term dimensionless. Eq. (3.7) still has terms with dimensions.
To make each term dimensionless, we usually divide by the coefficient in front of the term
with the highest time derivative (but dividing by any coefficient in any term will do). The
result is
du∗
= −atc u∗ , u∗ (0) = u−1
c I. (3.8)
dt∗
Step 5: Estimate the scales. A characteristic quantity like tc reflects the time scale in
the problem. Estimating such a time scale is certainly the most challenging part of the
scaling procedure. There are different ways of reasoning. The first approach is to aim at
a size of u∗ and its derivatives that is of order unity. If uc is chosen such that |u∗ | is of size
unity, we see from (3.8) that du∗ /dt∗ is of the size of u∗ (i.e., unity) if we choose tc = 1/a.
Alternatively, we may look at a special case of the model where we have analytical insight
that can guide the choice of scales. In the present problem we are lucky to know the exact
solution for any value of the input data as long as a is a constant. For exponential decay,
u(t) ≈ e−at , it is common to define a characteristic time scale tc as the time it takes to
reduce the initial value of u by a factor of 1/e (also called the e-folding time):

1
e−atc = e−a.0 ⇒ e−atc = e−1 ,
e

from which it follows that tc = 1/a. Note that using an exact solution of the problem to
determine scales is not a requirement, just a useful help in the few cases where we actually
have access to an exact solution.
In this example, two different, yet common ways of reasoning, lead to the same value of tc .
However, instead of using the e-folding time we could use the half-time of the exponential
decay as characteristic time, which is also a very common measure of the time scale in
such processes. The half time is defined as the time it takes to halve u:

1
e−atc = e−a.0 ⇒ tc = a−1 ln 2.
2

There is a factor ln 2 = 0.69 difference from the other tc value. As long as the factor is
not an order of magnitude or more different, we do not pay attention factors like ln 2 and
skip them, simply to make formulas look nicer. Using tc = a−1 ln 2 as time scale leads to a

80
 3.4. SCALING

scaled differential equation u0 = −(ln 2)u, which is fine, but an unusual form. People tend
to prefer the simpler ODE u0 = −u, which arises from tc = 1/a, and we shall therefore
use this time scale.

Regarding uc , we may look at the initial condition and realize that the choice uc = I
makes u∗ (0) = 1. For t > 0, the differential equation expresses explicitly that u decreases,
so uc = I gives u∗ ∈ (0, 1]. Scaling a variable u such that |u∗ | ∈ [0, 1] is always the
ultimate goal, and this goal is in fact obtained here. Next best result is to ensure that
the magnitude of |u| is not ”big” or ”small”, in the sense that the size is neither as large
as 10 or 100, nor as small as 0.1 or 0.01. (In the present problem, where we are lucky to
have an exact solution u(t) = Ie−at , we may look at this to explicitly see that u ∈ (0, I]
such that uc = I gives u∗ ∈ (0, 1]).
With tc = 1/a and uc = I, we have the final dimensionless model

du∗
∗
= −u∗ , u∗ (0) = 1. (3.9)
dt

This is a remarkable result in the sense that all physical parameters (a and I) are removed
from the model! Or more precisely, there are no physical input parameters to assign before
using the model. In particular, numerical investigations of the original model (3.4) would
need experiments with different a and I values, while numerical investigations of (3.9)
can be limited to a single run!
It is very common to drop the stars when the scaled problem has been derived and work
further with (3.9) simply written as

du
= −u, u(0) = 1.
dt

81
 3.4. SCALING

Scaling a Cooling Problem with Constant Temperature in the

Surroundings

The heat exchange between a body at temperature T (t) and the surroundings at constant
temperature Ts can be modeled by Newton’s law of cooling:

T 0 (t) = −k(T − Ts ), T (0) = T0 , (3.10)

where k is a prescribed heat transfer coefficient.

Exact solution. An analytical solution is always handy to have as a control of the

choice of scales. The solution of (3.10) is by standard methods for ODEs found to be
T (t) = Ts + (T0 − Ts )e−kt .
Scaling. Physically, we expect the temperature to start at T0 and then to move toward
the temperature of the surroundings (Ts ). We therefore expect that T lies between T0
and Ts . This is mathematically demonstrated by the analytical solution as well. A proper
scaling is therefore to scale and translate T according to

T − T0
T∗ = . (3.11)
Ts − T0

Now, 0 ≤ T ∗ ≤ 1.
Scaling time by t∗ = t/tc and inserting T = T0 + (Ts − T0 )T ∗ and t = tc t∗ in the problem
(3.10) gives
dT ∗
∗
= −tc k(T ∗ − 1), T (0) = 0.
dt
A natural choice, as argued in other exponential decay problems, is to choose tc k = 1,
which leaves us with the scaled problem

dT ∗
∗
= −(T ∗ − 1), T ∗ (0) = 0. (3.12)
dt

No physical parameter enters this problem! Our scaling implies that T ∗ starts at 0 and
approaches 1 as t → ∞, also in the case Ts < T0 . The physical temperature is always

82
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

recovered as
T (t) = T0 + (Ts − T0 )T ∗ (kt∗ ). (3.13)

Therefore our scaled equation of the model, (3.12) is written as

dT
= −(T − 1), T (0) = 0.
dt

3.5 Equilibrium Points and Stability Analysis

In this section we will discuss some important concepts met when we are analyzing dy-
namical systems.
System dynamics is both a modeling and a simulation methodology. As a modeling tech-
nique it focuses on the complex structure of the aspect of reality being modeled. That
complexity is due to multiloop feedback substructures, nonlinearities and delays. The con-
sideration of these characteristics leads to mathematical models that are quite difficult to
deal with using analytical tools. In this way some dilemma is produced between math-
ematical analysis and computer modeling. Nevertheless, mathematical analysis yields
benefits for system dynamics that should not be disregarded as they can help to better
understand the behavior of a model. If a mathematical language is used, the consequences
of that usage ought to be fully maintained. A system dynamics model is a mathematical
object and, as such, it deserves mathematical analysis.

3.5.1 Equilibrium Points

One-dimensional Equations
Let x represents a given population. The one dimensional equations of an ODE is defined
as:
”The rate of change on the number of population x with respect to a given time t”
represented by
dx
= f (x). (3.14)
dt

83
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Normally, the function f is highly involved as it represents the whole model, but, at least
conceptually, a model can be thought as a mathematical object of the form (3.14).
Equilibrium points of one dimensional system
Definition: A point x∗ in R is said to be an equilibrium point of one dimensional equation
(3.14) if it satisfies the equation f (x∗ ) = 0. Equivalently, it can be defined as follows. The
equilibrium points (or simply the equilibria) of a model like (3.14) are defined as the
values x∗ of x for which if, at a specific time t0 , x = x∗ then x will remain unchanged for
all t > t0 . This implies that all the model variables will remain constant and, therefore
dx/dt = 0. This condition can be used to obtain the equilibrium points of a system.
An equilibrium point is also named as fixed point or critical point or stationary point.

As will be seen later, the number and values of the equilibria of a system play an important
role in the behavior of that system. An equilibrium point x∗ is said to be stable when
initial conditions close to that point produce trajectories (time evolution’s of x) which
approach the equilibrium. On the contrary, if these trajectories move away from x∗ , the
equilibrium is unstable.
Any dynamical system may have no, one or several equilibrium points, each of which may
either be stable or unstable. A plain mechanical example is a simple pendulum. If the
pendulum is at the downward position, the system will remain there forever, and thus,
the downward position is an equilibrium point. Obviously, this equilibrium is stable (if
friction is considered). Likewise, it is easy to see that the upward position is an unstable
equilibrium.
Example 1: Consider the one dimensional differential equations. Find the equilibrium
point(s)
a) dx
dt
= x. Here f (x) = x, and f (x) = x = 0 when x = 0. Hence x∗ = 0 is an equilibrium
point.
b) dx
dt
= 2 − x. Here f (x) = 2 − x, and f (x) = 2 − x = 0 when x = 2. Hence x∗ = 2 is an
equilibrium point.
dx
c) dt
= x2 − 5x + 6. Here f (x) = x2 − 5x + 6, and f (x) = x2 − 5x + 6 = 0 when x = 2 or
x = 3. Hence x∗ = 2 or x∗ = 3 are an equilibrium points.

84
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Now, after finding equilibrium points let us see about stability analysis.

3.5.2 Stability Analysis

Stability Analysis of Equilibrium Points on One Dimensional Equation.

Definition: Let x∗ be an equilibrium point of the one dimensional equation (3.14)

1. If f 0 (x∗ ) < 0, then the equilibrium point x∗ is stable.

2. If f 0 (x∗ ) > 0, then the equilibrium point x∗ is unstable.

3. If f 0 (x∗ ) = 0, nothing to say the stability of x∗ .

Physical Meaning of Stability Analysis

• An equilibrium point of a given differential equation is stable means ”all solution

curves of the equation attract towards the equilibrium point.”

• An equilibrium point of a given differential equation is unstable means ”all solution

curves of the equation go away from the equilibrium point.”

Example 2: Let us consider example 1, again.

Under (a) we have f (x) = x and x∗ = 0. Now f 0 (x) = 1 so that f 0 (x∗ ) = f 0 (0) = 1 > 0.
Hence the equilibrium point x∗ = 0 is unstable. That is all solution curves x(t) = cet of
the differential equation
dx
=x
dt
go away from the equilibrium point x∗ = 0.

Under (b) we have f (x) = 2 − x and x∗ = 2. Now f 0 (x) = −1 so that f 0 (x∗ ) = f 0 (2) =
−1 < 0. Hence, the equilibrium point x∗ = 2 is stable. That is all solution curves
x(t) = 2 − ce−t of the differential equation

dx
=2−x
dt

attract towards the equilibrium point x∗ = 2.

85
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Figure 3.8: Time series plot of dx/dt = x.

Figure 3.9: Time series plot of dx/dt = 2 − x.

Under (c) we have f (x) = x2 − 5x + 6 and x∗ = 2 or x∗ = 3. Now f 0 (x) = 2x − 5

so that f 0 (x∗ ) = f 0 (2) = −1 < 0. Hence the equilibrium point x∗ = 2 is stable. And
f 0 (x∗ ) = f 0 (3) = 1 > 0. Hence the equilibrium point x∗ = 3 is unstable. That is all

86
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

3−2cet
solution curves x(t) = 1−cet
of the differential equation

dx
= x2 − 5x + 6
dt

attract towards the equilibrium point x∗ = 2 and go away from the equilibrium point
x∗ = 3.

Figure 3.10: Time series plot of dx/dt = x2 − 5x + 6.

Note: If we study two or more differential equations at the same time then it is known
to be studying a dynamical system.

Planar linear systems (n-dimensional dynamical system)

We begin the study of systems of differential equations. A system of differential equations
is a collection of n interrelated differential equations of the form

x01 = f1 (x1 (t), x2 (t), x3 (t), ..., xn (t))

x02 = f2 (x1 (t), x2 (t), x3 (t), ..., xn (t))
x03 = f3 (x1 (t), x2 (t), x3 (t), ..., xn (t)) (3.15)
..
.
x0n = fn (x1 (t), x2 (t), x3 (t), ..., xn (t)).

87
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Here, the functions fj are real-valued functions of the n+1 variables x1 +, x2 +, ..., +xn ,
and t. Unless otherwise specified, we will always assume that the fj are C ∞ functions.
This means that the partial derivatives of all orders of the fj exist and are continuous.

To simplify notation, we will use vector notation:


x1
 . 
X= . 
 . .
xn

We often write the vector X as (x1 , ..., xn ) to save space.

Our system may then be written more concisely as

X 0 = F (t, X),

where  
f1 (t, x1 , ..., xn )
 .. 
F (t, X) = 
 . .

fn (t, x1 , ..., xn )

A solution of this system is then a function of the form X(t) = (x1 (t), ..., xn (t)) that
satisfies the equation, so that
X 0 (t) = F (t, X(t)),

where X 0 (t) = (x01 (t), ..., x0n (t)).

The system of equations is called autonomous if none of the fj depends on t, so the
system becomes X 0 = F (X). For most of the rest of this module we will be concerned
with autonomous systems.
In analogy with one-dimensional differential equations, a vector X ∗ for which F (X ∗ ) = 0
is called an equilibrium point for the system. An equilibrium point corresponds to a
constant solution X(t) ≡ X ∗ of the system as before.
Just to set some notation once and for all, we will always denote real variables by lowercase
letters such as x, y, x1 , x2 , t, and so forth. Real-valued functions will also be written in

88
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

lowercase such as f (x, y) or f1 (x1 , ..., xn , t). We will reserve capital letters for vectors such
as X = (x1 , ..., xn ), or for vector-valued functions such as

F (x, y) = (f (x, y), g(x, y)),

or  
h1 (x1 , ..., xn )
 .. 
H(x1 , ..., xn ) = 
 . .

hn (x1 , ..., xn )

We will denote n-dimensional Euclidean space by Rn , so that Rn consists of all vectors of

the form X = (x1 , ..., xn ).
Two-Dimensional Dynamical System

Many of the most important differential equations encountered in science and engineering
are second-order differential equations. We will deal with autonomous systems in R2 ,
which we will write in the form
x0 = f (x, y)
(3.16)
y 0 = g(x, y)
thus eliminating the annoying subscripts on the functions and variables. As above, we
often use the abbreviated notation X 0 = F (X) where X = (x, y) and F (X) = F (x, y) =
(f (x, y), g(x, y)).
Equilibrium Points of Two-Dimensional System

A point (x∗ , y ∗ ) is an equilibrium point of the two dimensional dynamical system (3.16)
if it satisfies
f (x∗ , y ∗ ) = (0, 0)

g(x∗ , y ∗ ) = (0, 0).

Remark: An equilibrium point of two dimensional dynamical system (3.16) is the inter-
section point of f (x, y) = 0 and g(x, y) = 0.

89
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Example 1: Find the equilibrium point of the two dimensional system given by

x0 = x + 1
y0 = y − 1

Solution: The equilibrium point is obtained by making

x0 = x + 1 = 0
y 0 = y − 1 = 0.

That is (x, y) = (−1, 1). Hence an equilibrium point of a system is (x∗ , y ∗ ) = (−1, 1).

Example 2: Find the equilibrium point of the two dimensional system given by

x0 = x2 − 1
y0 = y2 − 4

Solution: The equilibrium point is obtained by making

x0 = x2 − 1 = 0
y 0 = y 2 − 4 = 0.

This implies that x0 = (x−1)(x+1) = 0 and y 0 = (y −2)(y +2) = 0. Since the equilibrium
point of a system is the intersection of x0 = 0 and y 0 = 0. We have x = 1 or x = −1
and y = 2 or y = −2. Thus the equilibrium point of the system are (x∗ , y ∗ ) = (1, 2),
(x∗ , y ∗ ) = (1, −2), (x∗ , y ∗ ) = (−1, 2) and (x∗ , y ∗ ) = (−1, −2).
Example 3: Find the equilibrium point of the two dimensional system given by

x0 = 2x − xy
y 0 = xy − 3y

90
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Solution: The equilibrium point is obtained by making

x0 = 2x − xy = 0
y 0 = xy − 3y = 0.

This implies that x0 = x(2 − y) = 0 and y 0 = y(x − 3) = 0. We have x = 0 or 2 − y = 0

and y = 0 or x − 3 = 0. Therefore (x∗ , y ∗ ) = (0, 0), (x∗ , y ∗ ) = (3, 2) are equilibria points
of the system.
Stability Analysis of Equilibrium Points on Two Dimensional Dynamical Sys-
tem
Note: Basically there are six types of equilibrium points which are identified by stability
analysis, these are

1. Stable node

2. Unstable node

3. Saddle point

4. Stable focus

5. Unstable focus

6. Center

Technique to Determine Types of Equilibrium Points

Reconsider the planar system

x0 = f (x, y)
y 0 = g(x, y)

Assume that (x∗ , y ∗ ) be an equilibrium point of the above system. We define the Jacobian
matrix J(x, y) of the system as
!
∂f ∂f
∂x ∂y
J(x, y) = ∂g ∂g
. (3.17)
∂x ∂y

91
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Note that the Jacobian matrix determines whether the relation between the two species
is strong or week.
Question: What is the nature of the solution of the system (3.16) around the feasible
equilibrium points? To answer these question we have follow the following steps.
First, determine the eigenvalue (λ) of the Jacobian matrix J(x, y) at the equilibrium
point(s) (x∗ , y ∗ ). That is det(J(x∗ , y ∗ ) − λI2 ) = 0.

∂f ∂f
−λ
det(J(x∗ , y ∗ )) = ∂x
∂g ∂g
∂y
= 0.
∂x ∂y
−λ

   
2 ∂f ∂g ∂f ∂g ∂g ∂f
⇒λ − + λ+ − = 0.
∂x ∂y ∂x ∂y ∂x ∂y
p
2 T r(J) ± (T r(J))2 − 4det(J)
⇒ λ − T r(J)λ + det(J) = 0 ⇒ λ = .
2

Now let us discuss about the nature of the solution or stability analysis.
Note: Based on the algebraic sign of the eigenvalues of the Jacobian matrix λ1 and λ2
we can determine the type of the equilibrium points.

1. If both eigenvalues are real, distinct and negative, then the equilibrium point is
stable node.

2. If both eigenvalues are real, distinct and positive, then the equilibrium point is
unstable node.

3. If both eigenvalues are real, distinct and opposite signs, then the equilibrium point
is saddle point.

4. If the eigenvalues are complex with negative real parts, then the equilibrium point
is stable focus.

5. If the eigenvalues are complex conjugate with positive real part, then the equilibrium
point is unstable focus.

6. If the eigenvalues are pure imaginary, then the equilibrium point is a center.

92
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Example 1 [Stable Node]

Consider the two dimensional dynamical system

x0 = −x
y 0 = −4y

The equilibrium point of the system is the origin (0, 0). The Jacobian matrix of the system
at the point (x, y) is !
1 0
J(x, y) = .
0 −4

Now let us find the Jacobian matrix of the system at the equilibrium point (0, 0)
!
1 0
J(0, 0) = .
0 −4

Then the determinant of the Jacobian at equilibrium point is

−1 − λ 0
det (J(0,0)) = = 0 ⇒ (−1 − λ)(−4 − λ) = 0 ⇒ λ1 = −1, λ2 = −4.
0 −4 − λ

The eigenvalues are λ1 = −1 or λ2 = −4. Since both eigenvalues are real, distinct and
negative thus, the equilibrium point (0, 0) is stable node.
This means that all solution curves of the two dimensional dynamical system attract
towards the equilibrium point (0, 0).

Example 2 [Unstable focus]

Consider the two dimensional dynamical system

x0 = x + 2y
y 0 = −100x + y

The equilibrium point of the system is the origin (0, 0). The Jacobian matrix of the system

93
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Figure 3.11: The solution curves of the planar system at equilibrium point (0, 0) has stabe
node.

at the point (x, y) is !

1 2
J(x, y) = .
−100 1

Now let us find the Jacobian matrix of the system at the equilibrium point (0, 0)
!
1 2
J(0, 0) = .
−100 1

Then the determinant of the Jacobian at equilibrium point is

1−λ 2 √ √
det (J(0,0)) = = 0 ⇒ (1−λ)(1−λ)+200 = 0 ⇒ λ1 = 1+20 2i, λ2 = 1−20 2i.
−100 1 − λ
√ √
The eigenvalues are λ1 = 1+20 2i or λ2 = 1−20 2i. Since both eigenvalues are complex
conjugate with positive real part thus,the equilibrium point (0, 0) is unstable focus.
This means that all solution curves of the two dimensional dynamical system spirally go
away from the origin (0, 0).

94
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Figure 3.12: The solution curves of the planar system at equilibrium point (0, 0) has
unstable focus.

Since we have six (6) different types of equilibrium points which are determined by stability
analysis exists for the given two-dimensional dynamical system. We have discussed some
of them above but for the remaining cases:

• If an equilibrium point is unstable node, then all solution curves of the two dimen-
sional dynamical system go away from equilibrium point.

• If an equilibrium point is saddle point, then all solution curves of the two dimensional
dynamical system asymptotic to the axes.

• If an equilibrium point is stable focus, then all solution curves of the two dimensional
dynamical system spirally towards the equilibrium point.

• If an equilibrium point is center, then all solution curves of the two dimensional
dynamical system makes a circle around the equilibrium point.

95
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

Exercises
1. The following expressions have been proposed as the time scale for the function
u∗ (t∗ ) = A cos(2πf0 t∗ )
T = 1/f0 ,
T = 1/(2πf0 ),
(3.18)
T = 1/(πf0 ),
T = 500πf0−1
May all these be used as the time scale?

2. A common mathematical model for the size of a population y ∗ (t∗ ) as a function of

time t∗ is described by the logistic equation

dy ∗ y∗
 
∗
= ry 1 − . (3.19)
dt∗ K

Here r is called the growth rate and K is the sustainable capacity.

(a) Which scale is suitable for y ∗ ?

(b) Determine a time scale when y ∗ << K.
(c) Introduce these scales into the equation so that it becomes dimensionless (The
equation can easily be solved by inserting y = 1/u and solving for u).

3. Determine the equilibrium points and their stability for

du
= u(25 − u) (3.20)
dt

4. Find the equilibrium solutions of the following differential equations and determine
whether they are stable or unstable and sketch of the bifurcation diagrams indicating
stable and unstable equilibrium points.
du
(a) dt
= u2 (u2 − 1),
du
(b) dt
= (u − u2 )(u − µ), u ≥ 0, µ ≥ 0.

96
 3.5. EQUILIBRIUM POINTS AND STABILITY ANALYSIS

5. A certain population develops according to the equation of logistic growth. For

small populations, there is some possibility that individuals may die before they
have met a partner, and this can be modeled by modifying the logistic model as
follows,
dy
= y − y 2 − µy, µ ≥ 0. (3.21)
dt
(a) Determine the equilibrium solutions of Eq. (3.21) and determine whether they
are stable or unstable.
(b) For larger populations, this modification will be less important. The following
model has therefore been proposed:
 
dy y µ
=y 1− − , (3.22)
dt M 1+y

Here we assume that M > 1 and, as above, that µ > 0. Sketch the bifurcation
diagram for the modified model and determine stable and unstable equilibrium
solutions. Investigate what happens when the population is stable and positive,
(M +1)2
and µ grows slowly past the value 4M
.

6. Find the general solution and sketch phase diagrams for the following systems.
Characterize the type and the stability of the equilibria:

(a)
x0 = 4y,
(3.23)
y 0 = −9x.

(b)
x0 = −x + 2y,
(3.24)
y 0 = −100x − y.

(c)
x0 = −x − y,
(3.25)
y 0 = 4x + y.

97
CHAPTER 4

APPLICATION OF MATHEMATICAL MODELING

Objectives:

By the end of this chapter you will be able to:

• know the use of mathematics in biological sciences.

• identify Newton’s law of cooling.

• develop mathematical models representing a physical problem.

• develop better understanding of mathematics to the process of modeling in the

natural and social sciences.

• understand how to generate graphical representation to the solution of the model.

• understand the use of mathematics in Epidemiology.

98
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

4.1 Continuous Population Model for Single Species

The increasing study of realistic and practically useful mathematical models in popula-
tion biology, whether we are dealing with a human population with or without its age
distribution, population of an endangered species, bacterial or viral growth and so on, is
a reflection of their use in helping to understand the dynamic processes involved and in
making practical predictions.
The continually expanding list of applications is extensive as are the number of books on
various aspects of the field. There are also highly practical applications of single-species
models in the bio-medical sciences.

4.1.1 Continuous Growth Models

In this section we look at a population in which all individuals develop independently of

one another while living in an unrestricted environment where no form of competition is
possible. If the initial population size is small, then a stochastic model is more appropriate,
since the likelihood that the population becomes extinct due to chance must be considered.
Deterministic models often provide useful ways of gaining sufficient understanding about
the dynamics of populations whenever they are large enough. Furthermore, perturbations
to large populations at equilibrium often generate over short time scales independent in-
dividual responses, which may be appropriately modeled by deterministic models. For
example, the introduction of a single infected individual into a large disease-free popula-
tion leads to the generation of secondary cases of infection, propagating a disease. The
environment is free of interference competition, at least at the beginning of the outbreak,
when a large population of susceptibles provides a virtually unlimited supply of hosts.
The spread of disease in a large population of susceptibles may be thought of as an inva-
sion process generated by independent contacts between a huge pool of susceptibles and
a few infectious individuals.
The population density of a single species at time t will be denoted by N (t), where
it is assumed that N is everywhere differentiable, that is, N is a smooth function of
t. Although unrealistic since N (t) is an integer-valued function and thus not continu-

99
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

ous, for populations with a large number of members, the assumptions of continuity and
differentiability provide reasonable approximations. In many biological experiments the
population biomass, which one might expect to be more nearly described by a smooth
function than the population size, is often taken as the definition of N (t).
The rate of change of population density can be computed if the birth, death, and migra-
tion rates are known. A closed population has, by definition, no migration either into or
out of the population. In this case, the population size changes only through births and
deaths and the rate of change of population size is simply the birth rate minus the death
rate. The formulation of a specific model requires explicit assumptions on the birth and
death rates. Ideally, these assumptions are made with the goal of addressing specific bio-
logical questions such as under what conditions will interference competition (competition
for hosts) and pathogen virulence lead to host-pathogen long-term coexistence.
For microorganisms, which reproduce by splitting, it is reasonable to assume that the
rate of birth of new organisms is proportional to the number of organisms present. In
mathematical terms, this assumption may be expressed by saying that if the population
size at time t is N , then over a short time interval of duration h from time t to time
(t + h), the number of births is approximately bhN for some constant b, the per capita
birth rate. Similarly, we may assume that the number of deaths over the same time
interval is approximately dhN for some constant d, the per capita death rate. Hence, the
net change in population size from time t to time (t + h), which is N (t + h) − N (t), may
be approximated by [(bh − dh)]N (t). The duration h of the time interval must be short
to ensure that the population size does not change very much and thus the numbers of
births and deaths are approximately proportional to N (t). We obtain the approximate
equality
N (t + h) − N (t) ≈ (b − d)N (t)h. (4.1)

(The symbol ≈ is used to denote approximate equality in a sense that must be specified.)
Division by h gives,
N (t + h) − N (t)
≈ (b − d)N (t), (4.2)
h
and the limit as h → 0 gives
dN
= (b − d)N (t), (4.3)
dt

100
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

under the assumption that the function N (t) is differentiable. The approximate equality
in (4.1) means that the difference between the two sides of (4.1) is so small that the result
of dividing this difference by h gives a quantity that approaches zero as h → 0.
If the net growth rate is naturally defined as

r ≡ b − d,

then another way of looking at this model is to observe that if the population size at time
t is N (t), then in the next tiny time interval of length h the net increase in population size
due to a single organism will be rh. Since all individuals are independent (no competition
in an unrestricted environment), the net increase in population due to all N (t) organisms
will be rhN (t), and thus we arrive again at the differential equation

dN
= rN (t). (4.4)
dt

This differential equation has the infinite family of solutions given by the one-parameter
family of functions N (t) = kert ; hence, this one parameter family gives a solution of (4.4)
for every choice of the constant k. The most convenient way to impose a condition that
will describe the population dynamics of a specific population is by specifying the initial
population size at time t = 0 as,
N (0) = N0 . (4.5)

This choice selects the solution, N (t) = N0 ert . Condition (4.5) is called an initial condi-
tion, and the problem consisting of the differential equation (4.4) together with the initial
condition (4.5) is called an initial value problem.
As pointed out above, the above initial value problem has the unique solution

N (t) = N0 ert ,

where r > 0 (or equivalently b > d) implies that the population size will grow un-
boundedly as t → ∞, while r < 0 (or b < d) implies that the population size will
approach zero as t → ∞. In the absence of births (b = 0), the population is deplenished
by deaths at the rate d, and consequently, the average life-span of a member of this

101
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

population is 1/d. If b > 0, then the average number of offspring over the lifetime of
an average individual under Malthuss model would be b/d. If this ratio (usually referred
to as the basic reproductive number or ratio R0 ) is greater than one, then births exceed
deaths and the average number of offspring per person over their lifetime is greater than
one, that is, the population explodes; if this ratio is less than one, then deaths exceed
births, the average number of offspring per person is less than one, and the population
dies out.

The prediction that population size will grow exponentially under these conditions was
first stated by Malthus (1798). Malthus predicted disaster, because food supplies could
not possibly be increased to keep pace with population growth at a constant positive per
capita growth rate. Populations that grow exponentially at first are commonly observed in
nature. However, their growth rates usually tend to decrease as population size increases.
In fact, exponential growth or decay may be considered typical temporary local behavior.
In other words, populations dynamics can usually be approximated by this simple model
only for short periods of time; that is, the dynamics of a population may be handled well
locally with linear models. The assumption that the rate of growth of a population is
proportional to its size (linear assumption) is usually unrealistic on longer time scales.
The next section considers nonlinear assumptions on the rate of population growth rates,
which lead to quite different qualitative predictions.

Exercises

In Exercises 1 through 8, assume that the rate of change of population size is proportional
to population size.

1. Suppose the growth rate of a population is 0.7944 per member per day. Let the
population have 2 members on day zero. Find the population size at the end of 5
days.

2. Suppose a population has 100 members at t = 0 and 150 members at the end of 100
days. Find the population at the end of 150 days.

3. Suppose a population has 39 members at t = 8, and 60 members at t = 12. What

102
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

was the population size at t = 0?

4. The population of the earth was about 5 × 109 in 1986. Use an exponential growth
model with the rate of population increase of 2 percent per year observed in 1986
to predict the population of the earth in the year 2000.

5. Bacteria are inoculated in a petri dish at a density of 10/ml. The bacterial density
doubles in 20 hours. Assume that this situation is described by the differential
equation
dN
= CN,
dt
where N is the bacterial density and C is a constant.

(a) Integrate this equation giving N as a function of time.

(b) Find the value of C.
(c) How long does it take for the density to increase to 8 times its original value?
To 10 times?

6. Suppose that a population has a constant growth rate r per member per unit time
and that the population size at time t = t0 is N0 . Show that the population size at
time t is N0 er(t−t0 ) .

7. Suppose that a population has a growth rate of r per member in unit time, with
r > 0. Show that the time required for the population to double its initial size
(called the doubling time) is (log 2)/r. [Note: We will always use ”log ” to denote
the natural logarithm.]

8. Suppose that a population has a growth rate of r per member per unit time with
r < 0. Show that the time required for the population to decrease to half its initial
size (called the half-life) is −(log 2)/r.

4.1.2 The Logistic Population Model

As before, N (t) denotes the size of a population at time t, and dN/dt or N 0 (t) is the
rate of change of population size. We shall continue to assume that the population

103
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

growth rate depends only on the population’s size. Such an assumption appears to be
reasonable for simple organisms such as micro-organisms. For more complicated organisms
like plants, animals or human beings this is obviously an oversimplification since it ignores
intra-species competition for resources as well as other significant factors, including age
structure (the mortality rate may depend on age rather than on population density, while
the birth rate may depend on the adult population size rather than on total population
size). Furthermore, the possibility that birth or death rates may be influenced by the
size of populations that interact with the population under study must also be considered
(competition, predation, mutualism). We shall consider the effects of some of these factors
in later sections.
Here,we study models in which the growth rate depends only on population size, because
in spite of their shortcomings, these models do predict the qualitative behavior of many
real populations. The per capita growth rate, or rate of growth per member(individuals),
is given by N 0 (t)/N (t), which we are assuming as a function of N (t). In the previous
section it was assumed that the total growth rate was proportional to population size (a
linear model), or equivalently, we took a constant per capita growth rate. In this section
total growth rates that decrease as population size increases are considered.

The simplest population model in which the per capita growth rate is a decreasing function
of population size is λ − aN . This assumption leads to the logistic differential equation

dN
= N 0 = N (λ − aN ),
dt

first introduced by Verhulst (1838) and later studied further by R. Pearl and L. J. Reed
(1920). This equation is commonly written in the form,
 
dN N
= N 0 = rN 1− , (4.6)
dt K

with parameters r = λ, K = λ/a. The parameters r and K, assumed positive, may then
be given biological significance. It is observed that N 0 ≈ rN when N is small, and that
N 0 = 0 when N is near K. In other words, when N is small the population experiences
exponential growth, while when N is near K the population hardly changes.

104
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

Separation of variables allows us to rewrite equation (4.6) as,

Z Z
dN r
= dt,
N (K − N ) K

while using partial fractions, giving

 
1 1 1 1
= + ,
N (K − N ) K N K −N

allows us to integrate it:

Z Z Z 
r dN 1 dN dN 1
t+c= = + = (log N − log(K − N )),
K N (K − N ) K N K −N K

where c is the constant of integration.

If the population size at time t = 0 is N0 , substitution of the initial condition N (0) = N0
gives,
1
c= (log N0 − log(K − N0 )).
K
We now have

1 r 1
(log N − log(K − N )) = t + (log N0 − log(K − N0 )),
K K K
 
N N0
log = rt + log ,
K −N K − N0
N (K − N0 ) N (K − N0 )
log = rt ⇒ = ert .
N0 (K − N ) N0 (K − N )

Further algebraic simplification gives,

N (K − N0 ) = N0 (K − N )ert = KN0 ert − N N0 ert ,

N (K − N0 + N0 ert ) = KN0 ert ,

105
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

and finally
KN0 ert KN0
N (t) = = . (4.7)
K − N0 + N0 ert N0 + (K − N0 )e−rt
The above solution is valid only if 0 < N0 < K, so that the logarithms obtained in the
integration are defined. To obtain the solution without this restriction, our integration
should have given logarithms of absolute values. Nevertheless, the formula (4.7) for the
solution of the logistic equation is valid for all N0 , as could be verified by a more careful
analysis.

The expression (4.7) for the solution of the logistic initial value problem shows that the
population size N (t) approaches the limit K as t → ∞ if N0 > 0. The value K is called
the carrying capacity of the population, because it represents the population size that
available resources can continue to support. The value r is called the intrinsic growth
rate, because it represents the per capita growth rate achieved if the population size were
small enough to ensure negligible resource limitations. The logistic model predicts rapid
initial growth for 0 < N0 < K, then a decrease in growth rate as time passes so that
the size of the population approaches a limit (Figure 4.1). This behavior is in agreement
with the observed behavior of many populations, and for this reason, the logistic model
is often used as a means of describing population size.

Figure 4.1: Solution of the logistic equation.

106
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

Example;. Census data for the population of the United States in millions fit the function

265
N (t) =
1 + 69e−0.03t

reasonably well, with the year 1790 taken as t = 0. We may compare this with (4.7) in
the form
K
N (t) =  
K−N0
1+ N0
e−rt

to see that this expression is a solution of the logistic model with K = 265, r = 0.03. If we
use this logistic model to describe the population of the United States we would predict
a carrying capacity of 265, 000, 000, and a 1990 census total of 226, 300, 000. (In fact, the
population size found in the 1990 census was approximately 250, 000, 000.)
It is possible to give a derivation of the logistic model based on a specific assumption
about the resources on which a population depends. Let C denote the concentration of
nutrients and assume that the per capita growth rate r is proportional to C, that is,
r = aC for some constant a. Assume also that we begin with a fixed concentration C(0)
of nutrients and that the consumption of a unit of nutrient produces b units of population
size. Then population size is governed by

N 0 = aCN (4.8)

with
1 0
N = −C 0 . (4.9)
b
Integration of (4.9) with respect to t gives,

1
C = − N + K,
b

and substitution of N = N0 , C = C(0) for t = 0 enables us to calculate the constant of

integration K as
N0
K = C(0) + .
b

107
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

Now (4.8) becomes,

   
0 1 N
N = aN K − N = aKN 1 − ,
b bK

and this is the logistic differential equation with intrinsic growth rate aK and carrying
capacity bK. The reader should observe that the intrinsic growth rate depends on the
initial concentration of nutrients, but the carrying capacity depends only on the rate of
conversion of nutrients into population (this derivation was taken from Edelstein-Keshet
(1988)).

Exercises
1. Suppose a population satisfies a logistic model with r = 0.4, K = 100, N (0) = 5.
Find the population size for t = 10.

2. Suppose a population satisfies a logistic model with carrying capacity 100 and that
the population size is 10 when t = 0 and 20 when t = 1. Find the intrinsic growth
rate.

3. The Pacific halibut fishery is modeled by the logistic equation with carrying capacity
80.5 × 106 , measured in kilograms, and intrinsic growth rate 0.71 per year. If the
initial biomass is one-fourth the carrying capacity, find the biomass one year later
and the time required for the biomass to grow to half the carrying capacity.

4. Use a logistic model with an assumed carrying capacity of 100 × 109 , an observed
population of 5 × 109 in 1986, and an observed rate of growth of 2 percent per year
when population size is 5 × 109 to predict the population of the earth in the year
2000.

5. Show that for every choice of the constant c, the function

K
N (t) =
1 + ce−rt

is a solution of the logistic differential equation.

108
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

6. One of the simplest models of population growth is the logistic equation dN/dt =
rN (1 − N/K).

(a) Interpret r and K. Discuss the model.

(b) Suppose you were given census data for a population (i.e., a table of date versus
population size). How could you test the fit of the logistic model to the data
? Remember that r and K are not given.

4.1.3 Qualitative Analysis

When using the logistic model in practice, one normally assumes that a population is
indeed described by a logistic model and then attempts to choose the parameters r and
K and the initial population size N0 to give the best fit with experimental data. It is
important to remember that the values of r, K, and N0 are therefore subject to error.
However, errors in r and N0 do not affect our prediction of the ultimate population size
K. The property that a small change in the initial size N0 of a solution has only a small
effect on the behavior of the solution as t → ∞ is called stability of the solution. We
will require stability of any solution to which we ascribe biological significance; if a small
disturbance can cause a large change in the solution it is unreasonable to consider the
solution meaningful.
It is also important to remember that the logistic model is an assumed form, not a
consequence of a fundamental law. We will want to consider larger classes of models and
examine properties that are valid for these larger classes rather than those properties that
depend on the specifics of the logistic model. A property that holds for a large class
of models is said to be robust, to indicate that it is more likely,in some sense, to have
biological significance.
The information that we derived about the behavior of solutions of the logistic model
was obtained from the explicit solution by separation of variables. If we wish to search
for robust properties we must learn to deduce properties of solutions from the differential
equation directly, without depending on analytic expressions for solutions.

The derivative of a solution N (t) at a point (t, N (t)) is the right side of the logistic

109
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

equation  
dN N
= rN 1− .
dt K
This is positive if 0 < N < K, zero if N = 0 or N = K, and negative if N < 0 or N > K.
Thus, a solution N (t) is an increasing function of t when 0 < N (t) < K and a decreasing
function of t when N (t) > K. (We ignore the case N (t) < 0 because it has no biological
significance.) The constants x ≡ 0 and x ≡ K are solutions. Differentiation of the logistic
equation with respect to t gives

d2 N
       
d N dN 2N dN 2 2N N
2
= rN 1− =r 1− =r N 1− 1− .
dt dN K dt K dt K K

From this we deduce that (d2 N )/(dt2 ) changes sign as N crosses the horizontal line N =
K/2 and thus a solution that crosses this line has an inflection point at the crossing. The
solution curves can be as shown in Figure 4.2.

Figure 4.2: Solution curves of the logistic equation.

If N (t) < K for some t, then the graph of N (t) cannot cross the line N = K (to see this,
we must use the uniqueness of solutions; if the graph did cross the line N = K there
would be two solutions passing through the point of crossing, and this is impossible) and

110
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

is increasing for all t. Thus, N (t) tends to a limit as t → ∞, but the only possible limits
are values of N for which the right side of the differential equation is zero, namely N = 0
or N = K. Since solutions near N = 0 tend away from N = 0, they cannot approach zero,
but must tend to K as t → ∞. A solution N (t) that is above K decreases for all t and
by a similar argument must tend to K as t → ∞. Thus, we see that every nonnegative
solution except the constant solution N ≡ 0 tends to K as t → ∞, and we have obtained
this information without explicitly solving the differential equation. The method we have
used may be adapted to more general first-order differential equations.
We will consider autonomous first order differential equations, that is, equations of the
form
N 0 = f (N ), (4.10)

in which the right side does not contain the independent variable t. Sometimes we will
write the equation in the form
N 0 = N r(N ), (4.11)

with r(N ) representing the per capita growth rate. We define an equilibrium of the
differential equation (4.10) to be a value N∞ such that f (N∞ ) = 0. An equilibrium
corresponds to a constant solution N (t) ≡ N∞ of the differential equation.
If N (t) is a solution of a differential equation N 0 = f (N ) that tends to a limit as t → ∞
then it is not difficult to show that its limiting value must be an equilibrium. In fact,
for a first-order differential equation, every solution must either tend to an equilibrium
as t → ∞ or be unbounded. However, not every equilibrium is a limit of nonconstant
solutions. For example, the only solution of the logistic equation that tends to zero as
t → ∞ is the identically zero solution.

In order to describe the behavior of solutions near an equilibrium, we introduce the process
of linearization. If N∞ is an equilibrium of the differential equation N 0 = f (N ), so that
f (N∞ ) = 0, we make the change of variable u(t) = N (t) − N∞ , representing deviation of
the solution from the equilibrium value. Substitution gives

u0 (t) = f (N∞ + u(t)),

111
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

and application of Taylor’s theorem gives

f 0 (c)
u0 (t) = f (N∞ ) + f 0 (N∞ )u(t) + (u(t))2
2!

for some c between N∞ and N∞ + u(t). We use f (N∞ ) = 0 and write h(u) = (f 0 (c)/2!)u2 .
Then we may rewrite the differential equation N 0 = f (N ) in the equivalent form

u0 = f 0 (N∞ )u + h(u).

The function h(u) is small for small |u| in the sense that h(u)/u → 0 as u → 0; more
precisely, for every  > 0 there exists δ > 0 such that |h(u)| < |u| whenever |u| < δ.
The linearization of the differential equation at the equilibrium N∞ is defined to be the
linear homogeneous differential equation

v 0 = f 0 (N∞ )v, (4.12)

obtained by neglecting the higher-order term h(u) in u0 = f 0 (N∞ )u+h(u). The importance
of the linearization lies in the fact that the behavior of its solutions is easy to analyze,
and this behavior also describes the behavior of solutions of the original equation (4.10)
near the equilibrium.
Theorem 4.1.1. If all solutions of the linearization(4.12) at an equilibrium N∞ tend to
zero as t → ∞, then all solutions of (4.10) with N (0) sufficiently close to N∞ tend to the
equilibrium N∞ as t → ∞.

We have stated the theorem in a form that generalizes readily to results for systems of
differential equations, as well as to results to be given later, when equations with delay
are discussed. In the specific situation covered in Theorem 4.1.1 the condition that all
solutions of the linearization tend to zero is f 0 (N∞ ) < 0. For an equilibrium N∞ with
f 0 (N∞ ) < 0 we must have f (N ) > 0 for N < N∞ and f (N ) < 0 for N > N∞ if N is
sufficiently close to N∞ .
A solution with N (0) > N∞ is monotone decreasing but bounded below by N∞ , and
therefore tends to a limit as t → ∞. Since the only possible limits of solutions are
equilibria, such a solution must tend to N∞ (if there are no other equilibria between N (0)

112
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

and N∞ ). By a similar argument a solution with N (0) < N∞ increases monotonically

to N∞ if there is no other equilibrium between N (0) and N∞ . Thus, all solutions with
N (0) sufficiently close to N∞ approach N∞ as t → ∞. Indeed, for first-order differential
equations we can be more precise: If N∞ is an equilibrium with f 0 (N∞ ) < 0, then every
solution whose initial value N (0) is between N∞ and the next equilibrium, in either
direction, must tend to N∞ as t → ∞.
An equilibrium N∞ is said to be stable if for every  > 0 there exists δ > 0 such that
|N (0) − N∞ | < δ implies |N (t) − N∞ | <  for all t > 0. It is implicit in this definition
that the existence of the solution N (t) is required for 0 ≤ t < ∞. An equilibrium N∞ is
said to be asymptotically stable if it is stable and if in addition, |N (0) − N∞ | < δ implies

lim N (t) = N∞ .
t→∞

Thus, stability means roughly that a small change in initial value produces only a small
effect on the solution, and this condition is a natural requirement for an equilibrium to be
biologically meaningful. It is possible for systems to have equilibria for which all solutions
starting near the equilibrium tend toward the equilibrium but only after traveling away
from the equilibrium. Such an equilibrium would not be stable, but our definition of
asymptotic stability requires stability in order to exclude this possibility. In biological
applications we will ordinarily require asymptotic stability rather than stability, both
because asymptotic stability can be determined from the linearization, while stability
cannot, and because an asymptotically stable equilibrium is not disturbed greatly by a
perturbation of the differential equation.
In terms of asymptotic stability we may restate Theorem 4.1.1 and a corresponding in-
stability result proved in the same way as follows:
Corollary 1. An equilibrium N∞ of (4.10) with f 0 (N∞ ) < 0 is asymptotically stable,
while an equilibrium N∞ with f 0 (N∞ ) > 0 is unstable.
We have already mentioned the logistic model, with r(N ) = r(1 − N/K), as an example.
Other examples that have been used in population models include

K
r(N ) = r log , [Gompertz (1825)]
N

113
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

r(K − N )
r(N ) = , [F. Smith (1963))]
K + aN
 θ !
N
r(N ) = r 1 − , [Ayala, Gilpin, and Ehrenfeld (1973)]
K

r(N ) = re1−N/K − d. [Nisbet and Gurney (1982)]

In using a model of this type to study a population problem one would assume a particular
form for r(N ); conduct experiments fit the resulting data to this form to estimate the
parameters of the model and then compare other observations with the predictions of the
model to judge its validity.

Every autonomous differential equation of the form N 0 = f (N ) or N 0 = N r(N ) has sep-

arable variables and thus can in principle be solved by integration. The reader should
observe that for each of the above examples, the necessary integration is sufficiently com-
plicated to make a qualitative approach attractive.
If in the model
N 0 = N r(N ) (4.13)

the function r(N ) is nonnegative and decreasing for 0 ≤ N ≤ K, then it is said to be

a compensation model. If the per capita growth rate r(N ) is increasing for small N ,the
model is said to be a depensation model. If the per capita growth rate is actually negative
for small N , then the model is said to be a critical depensation model. While compensation
models are the ones most commonly examined, both depensation and critical depensation
models arise in fishery studies.
A compensation model is characterized by the conditions

r(N ) ≥ 0, r0 (N ) ≤ 0 for 0 ≤ N ≤ K.

For a depensation model we assume

r(N ) ≥ 0, r00 (N ) ≤ 0 for 0 ≤ N ≤ K,

114
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

r0 (N ) > 0 for 0 < N < K ∗ ,

r0 (N ) < 0 for K ∗ < N < K.

Thus, r(N ) achieves a maximum at K ∗ , and since

lim r(N ) = lim f (N )/N = f 0 (0),

N →0 N →0

it follows that f 0 (0) < r(K ∗ ). Thus, the line joining the origin to the point (K ∗ , f (K ∗ )) on
the growth curve y = f (N ), which has slope r(K ∗ ), lies above the tangent to the growth
curve at the origin. Further, since

f 0 (N ) = N r0 (N ) + r(N ), f 00 (N ) = N r00 (N ) + 2r0 (N ),

we have f 00 (0) = 2r0 (0) ≥ 0 and

f 00 (K ∗ ) = K ∗ r00 (K ∗ ) + 2r0 (K ∗ ) = K ∗ r00 (K ∗ ) < 0.

This shows that the growth curve has an inflection point to the left of K ∗ .
For a critical depensation model we assume

f (N ) < 0 for 0 < N < K0 ,

f (N ) ≥ 0 for K0 ≤ N ≤ K.

Under this assumption it is not difficult to show that the equation N 0 = f (N ) has three
equilibria-an unstable equilibrium at K0 and asymptotically stable equilibria at 0 and K.
Then if the initial population size is below K0 , the population will die out. As we will
see in the next section, in the case of critical depensation hunting may drive a population
to extinction by bringing the population size below the critical level K0 , and this trend
to extinction will not be reversed if hunting ceases. The extinction through hunting of
the passenger pigeon in the nineteenth and early twentieth centuries from an original
population of 7 billion may have been an example of critical depensation. This property
is sometimes called the Allee effect [Allee(1931)].

115
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

In all of these models we have been assuming tacitly that the function f (N ) on the right
side of the differential equation is exact. Any realistic study would recognize that the
model can be at best an approximation, and that instead of N 0 = f (N ) we should really
be studying a differential equation of the form

dy
= f (y) + h(y),
dt

in which the term h(y) represents the error made in the assumption of the specific form
f (y). It is in the nature of h(y) that it cannot be known explicitly. Thus, instead
of looking for explicit formulas for solutions of y 0 = f (y) + h(y), we must be satisfied
with qualitative information about the solutions for a given class of functions h(y). It
is possible to establish the following result, which justifies our interest in asymptotically
stable equilibria of the model N 0 = f (N ).

Theorem 4.1.2. Let N∞ be an asymptotically stable equilibrium of N 0 = f (N ) with

f 0 (N∞ ) < 0. Then
(i) If h(y)/(y − N∞ ) → 0 as y → N∞ , and if |y(0) − N∞ | is sufficiently small, the solution
y(t) of y 0 = f (y) + h(y) tends to N∞ as t → ∞, i.e., N∞ is an asymptotically stable
equilibrium of y 0 = f (y) + h(y).
(ii) If |h(y)| ≤ A for all y and A is sufficiently small, and if |y(0) − N (0)| is sufficiently
small, then |y(t) − N (t)| ≤ (KA)/|f 0 (N∞ )| for some constant K, i.e., solutions of y 0 =
f (y) + h(y) are close to N∞ for all large t.

The essential content of Theorem 4.1.2 is that a perturbation h(y) that tends to zero
more rapidly than (y − N∞ ) as y → N∞ has no effect on the existence and asymptotic
stability of an equilibrium, while a bounded perturbation has at worst a bounded effect
on solutions. Thus conclusions drawn from analysis of the model N 0 = f (N ) are valid in
a sense for a large class of more refined models.

The description of interacting species will require systems of differential equations. The
use of vectormatrix notation and the methods of linear algebra will make it possible to
develop the theory of equilibria and asymptotic stability in a form analogous to what we
have described here. The central result is the analogue of Theorem 4.1.1 in its form stated
originally that if all solutions of the linearization at an equilibrium tend to zero, then the

116
 4.1. CONTINUOUS POPULATION MODEL FOR SINGLE SPECIES

equilibrium is asymptotically stable.

Exercises

In each of question 1 through 5, find all equilibria and determine which are asymptotically
stable.

1. N 0 = rN (1 − N/K).

2. N 0 = rN log K
N
.
rN (K−N )
3. N 0 = K+aN
.
 
N θ
4. N 0 = rN 1 −


K
(0 < θ < 1).

5. N 0 = N (re1−N/K − d).

6. (a) A population is governed by the differential equation

N 0 = N (e3−N − 1).

Find all equilibria and determine their stability.

(b) A fraction p(0 < p < 1) of the population in part (a) is removed in unit time,
so that the population size is governed by the differential equation

N 0 = N (e3−N − 1) − pN.

For what values of p is there an asymptotically stable positive equilibrium?

7. For which initial values y(0) does the solution y(t) of the differential equation

y 0 = y(2y − e−y )

approach zero as t → ∞?

117
 4.2. DISCRETE MODELS

4.2 Discrete Models

In discrete models, the state variables change only at a countable number of points in
time. These points in time are the ones at which the event occurs/change in state. Thus,
in discrete time modeling, there is a state transition function which computes the state at
the next time instant given the current state and input. The changes are really discrete
in many situations which occur at well defined time intervals. Moreover, in many cases,
the data are usually discrete rather than continuous. Hence, due to the limitations of the
available data, we may be compelled to work with the discrete model, even though the
underlying model is continuous.

4.2.1 Newton’s Law of Cooling

Suppose a cup of coffee, initially at a temperature of 1900 F is placed in a room, which is

held at a constant temperature of 900 F . After 1 minute, the coffee has cooled to 1800 F .
If we need to find the temperature of the coffee after 15 minutes, we will use Newton’s
law of cooling, which states that the rate of change of the temperature of an object is
proportional to the difference between its own temperature and the ambient temperature
(i.e. the temperature of its surroundings). Mathematically, this means,

tn+1 − tn = k(S − tn ).

Where tn is the temperature of the coffee after n minutes, S is the temperature of the
room and k is the constant of proportionality.
We first make use of the information given about the change in the temperature of the
coffee during the first minute to determine the value of the constant of proportionality k.
Then
1
t1 − t0 = k(S − t0 ) ⇒ 180 − 190 = k(70 − 190) ⇒ k = .
12
1 11 70
tn+1 − tn = (70 − tn ) ⇒ tn+1 = tn + .
12 12 12

118
 4.2. DISCRETE MODELS

The solution is given by

 n   n 
11 11
tn = 190 + 70 1 −
12 12
 n
11
tn = 70 + 120 , for n = 1, 2, 3, ....
12
For n = 15,  15
11
t15 = 70 + 120 = 102.54
12
11 n


Hence, after 15 minutes, the coffee has cooled to just under 1030 F . Also, limn→∞ 12
=
0, which implies that the temperature of the coffee will approach the equilibrium temper-
ature of 700 F , the room temperature as n increases.

4.2.2 Bank Account Problem

Suppose a savings account is opened that pays 4% interest compounded yearly with initial
deposit of Birr 10000.00 and a deposit of Birr 5000.00 is made at the end of each year.
For a savings account that is compounded yearly, the interest is added to the principal at
the end of each year. If is the amount at the end of year n(n = 0, 1, 2, 3, ...), then

a1 = a0 + ra0 = (1 + r)a0 ,
a2 = a1 + ra1 = (1 + r)a1 ,
.. (4.14)
.
an+1 = an + ran = (1 + r)an .

where r is the rate of interest. Now, if a deposit of Birr 5000.00 is made at the end of
each year, then the dynamic model which describes this scenario is given by

an+1 = (1 + r)an + 5000 = (1 + 0.04)an + 5000 = 1.04an + 5000

119
 4.2. DISCRETE MODELS

Thus the amount for three consecutive years will be

a1 = 1.04a0 + 5000 = 1.04 × 10000 + 5000 = 15400,

a2 = 1.04a1 + 5000 = 1.04 × 15400 + 5000 = 21016,

a3 = 1.04a2 + 5000 = 1.04 × 21016 + 5000 = 26856.64.

Let us now consider a different scenario, where no deposits are made but Birr 2000.00 is
withdrawn at the end of each year. We want to find out how much money be deposited,
so that we never run out of cash. The model for this scenario is

an+1 = 1.04an − 2000,

where we assume that the money is withdrawn after the interest from previous years has
been added and we are not penalized for withdrawing money each year. The equilibrium
value for this is given by

an+1 = an = a∗n ⇒ 1.04a∗n − 2000 = a∗n

2000
⇒ a∗n = = 50000.00
0.04
Thus, if the initial deposit (a0 ) in the account is Birr 50000.00 and we withdraw Birr
2000 each year, then the account will always have the same amount at the end of each
year. An obvious question is what happens if a0 < 50000 or a0 > 50000. If a0 is less that
50000, the amount in the account decreases to zero and the amount grows without bound
if is greater than 50000. Thus the system approaches zero or increases without bound if
a0 6= 50000 and therefore this equilibrium value is unstable.

4.2.3 Drug Delivery Problem

Suppose a patient is given a drug to treat some infection. The amount of drug in the
patients bloodstream decreases at the rate of 50% per hour. To sustain the drug to a
certain level, an injection is given at the end of each hour that increases the amount of

120
 4.2. DISCRETE MODELS

drug in the bloodstream by 0.2 units. The dynamic model which describes this scenario
is given by
an+1 = 0.5an + 0.2.

Where an is the amount of drug in the blood at the end of each hour. The equilibrium
solution of this model is given by

an+1 = an = a∗n ⇒ 0.5a∗n + 0.2 = 0 ⇒ a∗n = 0.4

The long-term behavior of the system will depend on the initial value a0 . No matter what
is the value of a0 , the system always approaches the value of 0.4, implying that 0.4 is a
stable equilibrium.

4.2.4 Economic Model (Harrod Model)

The Harrod model, which was developed in the 1930s, gives some insight into the dynamics
of economic growth. The model aims to determine an equilibrium growth rate for the
economy. Let Gn be the Gross Domestic Product (GDP) on national income, which is one
of the primary indicators to determine a countrys economy and Sn and In be the savings
and investment of the people. The Harrod model assumed that in a country peoples
savings depend on GDP or national income, that is, savings is a constant proportion of
current income, which implies
Sn = kGn , k>0 (4.15)

Harrod further assumed that the investment made by the people depends on the difference
between the GDP of the current year and the last year, that is,

In = a(Gn − Gn−1 ), a>k (4.16)

Finally, the Harrod model assumed that all the savings made by the people are invested,
that is,
Sn = In (4.17)

121
 4.2. DISCRETE MODELS

From Eq.(4.15), (4.16) and (4.17), we obtain

aGn−1
a(Gn − Gn−1 ) = Sn = kGn ⇒ Gn = ,
a−k

whose solution is  n
a
Gn = G(0)
a−k
Thus, Harrods model concludes that GDP or national income increases geometrically with
time.

4.2.5 Arms Race Model

We consider two countries engaged in an arms race. We assume that the two countries
have similar economic strength and the same level of distrust for each other. Let Mn be
the total amount of money spent by the two countries on arms. Let g > 0 measure the
restraint of growth due to economic strength (or weakness) of the countries and d > 0, the
level of distrust between the two countries. Both countries also spent a constant amount
(say, k ) of money for buying arms irrespective of involving in an arms race. Then, the
dynamic discrete model for the total amount of money Tn spent on arms by each country
after years is given by

Tn = (1 − g)Tn−1 + dTn−1 + k ⇒ Tn = (1 − g + d)Tn−1 + k

1 − (1 − g + d)n
 
n
Tn = (1 − g + d) T0 + k
1 − (1 − g + d)
1 − (1 − g + d)n
 
n
Tn = (1 − g + d) T0 + k
g−d
The equilibrium solution is

k
Tn = Tn−1 = T ∗ ⇒ (1 − g + d)Tn−1 + k = Tn−1 ⇒ T ∗ = , g > d.
g−d

Thus, as time increases, the total amount of money spent on arms reaches a steady state
(equilibrium point) and both the countries have a ”stable” arms race.

122
 4.2. DISCRETE MODELS

4.2.6 Discrete Population Models for a Single Species

In this section we shall consider populations with a fixed interval between generations or
possibly a fixed interval between measurements. Thus, we shall describe population size
by a sequence {xn }, with x0 denoting the initial population size, x1 the population size at
the next generation (at time t1 ), x2 the population size at the second generation (at time
t2 ), and so on. The underlying assumption will always be that population size at each
stage is determined by the population sizes in past generations, but that intermediate
population sizes between generations are not needed. Usually the time interval between
generations is taken to be a constant.

For example, suppose the population changes only through births and deaths, so that
xn+1 − xn is the number of births minus the number of deaths over the time interval
from tn to tn+1 . Suppose further that the birth and death rates are constants b and d,
respectively (that is, if the population size is x then there are bx births and dx deaths in
that generation). Then
xn+1 − xn = (b − d)xn ,

or
xn+1 = xn + (b − d)xn = (1 + b − d)xn .

We let r = 1 − b + d and obtain the linear homogeneous difference equation

xn+1 = rxn .

This together with the prescribed initial population size x0 determines the population
size in each generation. By a solution of the difference equation xn+1 = rxn with initial
value x0 we mean a sequence {xn } such that xn+1 = rxn for n = 0, 1, 2, ..., with x0 as
prescribed.

It is easy to solve the difference equation xn+1 = rxn algebraically. We begin by observing
that x1 = rx0 , x2 = rx1 = r2 x0 , x3 = rx2 = r3 x0 , and then we guess (and prove by
induction) that the unique solution is xn = rn x0 (n = 0, 1, 2, ...). It follows that if |r| < 1,
then xn → 0 as n → ∞, while if |r| > 1, then xn grows unboundedly as n → ∞. More
precisely, if 0 ≤ r < 1, xn decreases monotonically to zero; if 1 < r < 0, xn oscillates,

123
 4.2. DISCRETE MODELS

alternating between positive and negative values, but tends to zero; if r > 1, xn increases
to +∞; if r < −1, xn oscillates unboundedly. Negative values of xn for this difference
equation have no biological meaning, but we soon will consider difference equations in
which the unknown is a deviation from equilibrium (which may be either positive or
negative) rather than a population size. For this reason we have used the difference
equation xn+1 = rxn as our first example, even though a more plausible model for a real
population might be (
rxn for xn > 0,
xn+1 =
0 for xn ≤ 0,
which says that the population becomes extinct once it becomes zero in any generation.
This will occur if and only if r ≤ 0. The model xn+1 = rxn also arises under the
assumption that all members of each generation die, but there is a constant birth rate b
to form the next generation. In this case d = 1, so that r = b. We may form a different
model by allowing migration and assuming a constant migration rate β per generation,
with positive β denoting immigration and negative β denoting emigration. This leads to
the linear inhomogeneous difference equation

xn+1 = rxn + β,

which may also be solved iteratively,

x1 = rx0 + β,
x2 = rx1 + β = r(rx0 + β) + β = r2 x0 + rβ + β,
x3 = rx2 + β = r(r2 x0 + rβ + β) + β = r3 x0 + r2 β + rβ + β,
..
.

Again we may guess, and then prove by induction, that

xn = rn x0 + β(rn−1 + rn−2 + ... + r + 1)

β(1 − rn )
 
n β β
= r x0 + = x0 − rn + .
1−r 1−r 1−r

124
 4.2. DISCRETE MODELS

If r > 1, then xn grows unboundedly for β > −(r − 1)x0 , but xn reaches zero if β <
−(r − 1)x0 ; thus sufficiently large emigration will wipe out a population that would
otherwise grow unboundedly. If 0 < r < 1, then xn tends to the limit β/(1 − r) > 0 for
β > 0, while xn reaches zero for β < 0. Thus, immigration may produce survival of a
population that would otherwise become extinct.
The assumption of a constant growth rate independent of population size is unlikely to be
reasonable for real populations except possibly while the population size is small enough
not to be subject to the effects of overcrowding. Various nonlinear difference equation
models have been proposed as more realistic. For example, the difference equations

rxn
xn+1 = [Verhulst (1845)]
xn + A

and
rx2n
xn+1 = ,
x2n + A
have been suggested as descriptions of populations that die out completely in each genera-
tion and have birth rates that saturate for large population sizes. The difference equations
 xn   xn 
xn+1 = xn + rxn 1 − and xn+1 = rxn 1 − ,
K K

both called the logistic difference equation, and essentially equivalent, describe popula-
tions with growth rates that decrease to zero as the population grows large. Neither
should be taken too seriously for large population sizes since xn+1 becomes negative if xn
is too large. Another form, which could with some justification also be called the logistic
equation, is
xn+1 = xn er(1−xn /K) .

Here the growth rate decreases to zero as xn → ∞, but xn+1 cannot become negative.
Other difference equations, which have in fact been used as models to try to fit field data,
are
xn+1 = rxn (1 + αxn )−β [Hassell (1975)]

125
 4.2. DISCRETE MODELS

and (
rβ x1−β
n for xn > ,
xn+1 =
rxn for xn < .
It should be recognized that none of these models is derived from actual population growth
laws. Rather, they are attempts to give quantitative expression to rough qualitative ideas
about the biological laws governing the population. For this reason, we should be skeptical
of the biological significance of any deduction from a specific model that holds only for
that model. Our goal should be to formulate principles that are robust, that is, valid
for a large class of models (ideally for all models that embody some set of qualitative
hypotheses).

Exercises
1. Find the solution of the difference equation xn+1 = 12 xn , x0 = 2.

2. Find the solution of the difference equation xn+1 = 12 xn + 1, x0 = 2.

3. Find by calculating recursively the solution of the second-order difference equation

xn+1 = 21 xn + 1, x0 = 1, x1 = −1.

4. Consider the second order difference equation

xn+2 − 3xn+1 + 2xn = 0.

(a) Show that the general solution to the equation is of the form

xn = A 1 + 2 n A 2 ,

where A1 and A2 are any constants.

(b) Suppose that x0 and x1 are given. Then A1 and A2 must satisfy the system of
equations

A1 + A2 = x0 ,
A1 + 2A2 = x1 .

126
 4.2. DISCRETE MODELS

(c) From the general solution, solve for the specific solution with initial conditions
x0 = 10 and x1 = 20.

5. Find by calculating recursively the solution of the second-order difference equation

xn+2 = rxn , x0 = 1, x1 = 1.

6. Find the general form of the solution of the difference equation

xn+1 = c − xn

with c arbitrary for an arbitrary initial value x0 = a.

7. Consider the model  xn 

xn+1 = rxn 1− , r > 0.
K
(a) Show that xn+1 < 0 if and only if xn > K.
(b) Show that xn+1 > K is possible with 0 < xn < K only for r > 4.
(c) What conditions on x0 are necessary and sufficient to guarantee xn > 0 for
n = 1, 2, 3, ...?

8. Find the general form of the solution of the difference equation

xn+1 = 1 − xn

for an arbitrary initial value x0 = a.

9. The solution of the difference equation xn+2 = xn + xn+1 , x0 = 0, x1 = 1 is called

the Fibonacci sequence (originally formulated by Leonardo Fibonacci (1202) to
describe the number of pairs of rabbits under the hypothesis that each pair of rabbits
reproduces only at age one month and age two months and produces exactly one
pair of offspring on each of these two occasions, with all rabbits living exactly two
months).

(a) Calculate the first eight terms of the Fibonacci sequence.

(b) Suppose it can be shown that the ratio of successive terms xn+1 /xn of the
1
Fibonacci sequence tends to a limit τ as n → ∞. Show that τ = τ
+ 1.

127
 4.2. DISCRETE MODELS

√
1+ 5
(c) Deduce that τ = 2
.

Equilibrium Analysis

Consider the difference equation,

xn+1 = f (xn ). (4.18)

Definition 4.1. An equilibrium of a difference equation (4.18) is a value x∞ such that

x∞ = f (x∞ ), so that xn = x∞ (n = 0, 1, 2, ...) is a constant solution of the difference
equation.

In order to describe the behavior of solutions near an equilibrium, we introduce the process
of linearization just as we did in Section 4.1.3 for first-order differential equations. If x∞
is an equilibrium of the difference equation xn+1 = f (xn ), so that x∞ = f (x∞ ), we make
the change of variable un = xn − x∞ (n = 0, 1, 2, ...). Thus un represents deviation from
the equilibrium value. Substitution gives

x∞ + un+1 = f (x∞ + un ),

and application of Taylor’s theorem gives

f 00 (cn ) 2
x∞ + un+1 = f (x∞ + un ) = f (x∞ ) + f 0 (x∞ )un + un
2!

for some cn between x∞ and x∞ + un . We write h(un ) = f 00 (cn )u2n /2! and use the relation
x∞ = f (x∞ ) to form the difference equation equivalent to the original difference equation
(4.18),
un+1 = f 0 (x∞ )un + h(un ). (4.19)

The function h(u) is small for u, small in the sense that |h(u)/u| → 0 as |u| → 0; more
precisely, for every  > 0 there exists δ > 0 such that |h(u)| < |u| whenever |u| < δ. The
linearization of the difference equation xn+1 = f (xn ) at the equilibrium x∞ is defined to

128
 4.2. DISCRETE MODELS

be the linear homogeneous difference equation

vn+1 = f 0 (x∞ )vn , (4.20)

obtained by neglecting the higher-order term h(un ) in (4.19). The importance of the
linearization lies in the fact that the behavior of its solutions describes the behavior of
solutions of the original equation (4.18) near the equilibrium. The following result explains
the significance of the linearization at an equilibrium.

Theorem 4.2.1. If all solutions of the linearization (4.20) at an equilibrium x∞ tend to

zero as n → ∞, then all solutions of (4.18) with x0 sufficiently close to x∞ tend to the
equilibrium x∞ as n → ∞.

Proof. For convenience we write ρ = |f 0 (x∞ )|. The assumption that all solutions of the
linearization tend to zero is equivalent to ρ < 1. Now choose  > 0 so that ρ +  < 1. The
difference equation xn+1 = f (xn ) is equivalent to un+1 = f 0 (x∞ )un + h(un ). Then

|un+1 | ≤ |f 0 (x∞ )||un | + |h(un )| < ρ|un | + |un |

provided |un | < δ, where δ is determined by the condition that |h(u)| < |u| for |u| < δ.
Thus, |un+1 | ≤ (ρ + )|un | provided |un | < δ. If |u0 | < δ, it is easy to show by induction
that |un+1 | < δ for n = 0, 1, 2, .... This establishes |un+1 | ≤ (ρ + )|un | for n = 0, 1, 2, ....
Now it is easy to show, again by induction, that

|un | ≤ (ρ + )n |u0 |, n = 0, 1, 2, ....

Since ρ +  < 1, it follows that un → 0, and thus that xn → x∞ as n → ∞.

The content of Theorem 4.2.1 is that an equilibrium x∞ with |f 0 (x∞ )| < 1 has the property
that every solution with x0 close enough to x∞ remains close to x∞ and tends to x∞ as
n∞ . This property is called asymptotic stability of the equilibrium x∞ . The condition
f 0 (x∞ ) < 1 means that the curve y = f (x) crosses the line y = x from above to below
as x increases, while the condition f 0 (x∞ ) > −1 means that the curve y = f (x) cannot
be too steep at the crossing. If |f 0 (x∞ )| > 1, it is not difficult to show that except

129
 4.2. DISCRETE MODELS

for the constant solution xn = x∞ (n = 0, 1, 2, ...), solutions cannot remain close to x∞ .

This property is called instability of the equilibrium x∞ . An unstable equilibrium has
no biological significance, since any deviation, however small, is enough to force solutions
away.
Example 1; For the logistic difference equation
 xn 
xn+1 = xn + rxn 1 − ,
K

with f (x) = (1 + r)x − rx2 /K and f 0 (x) = (1 + r) − 2rx/K, it is easy to find equilibria
by solving the quadratic equation x = x + rx(1 − x/K) and obtaining the roots x = 0
and x = K. Since f 0 (0) = 1 + r, the equilibrium x = 0 is asymptotically stable if −1 <
1 + r < 1, or −2 < r < 0. Since r > 0 in applications, this means that the equilibrium
x = 0 is unstable. Since f 0 (K) = 1 − r, the equilibrium x = K is asymptotically stable
if 0 < r < 2. It is not difficult to show that for 0 < r < 2, every solution tends to the
equilibrium K. If r > 2, the equilibrium x = K is unstable and there is no asymptotically
stable equilibrium to which solutions can tend.
The logistic difference equation is sometimes presented in the form
 xn 
xn+1 = rxn 1− .
K

The study of the equation in this form is quite similar to the previous discussion; there
is an equilibrium at x = 0 that is asymptotically stable if r < 1, in which case every
solution tends to zero, and an equilibrium at x = K(1 − 1/r) that is asymptotically stable
if 1 < r < 3, in which case every solution tends to K(1 − 1/r), and if r > 3 there is no
asymptotically stable equilibrium.

Example 2; For the Verhulst equation

rxn
xn+1 = ,
xn + A

we have f (x) = rx/(x + A); f 0 (x) = rA/(x + A)2 . The solution of x = rx/(x + A) gives
two roots, x = 0 and x = r − A. Thus, if r < A the only equilibrium corresponding
to a non-negative population size is x = 0. Since f 0 (0) = r/A < 1, this equilibrium is

130
 4.2. DISCRETE MODELS

asymptotically stable and every solution tends to zero. If r > A there are two equilibria,
x = 0, and x = x∞ = r − A. Since f 0 (0) = r/A > 1, the equilibrium at x = 0 is unstable.
Since f 0 (x∞ ) = A/r < 1, the equilibrium x∞ is asymptotically stable.

Exercises

In question 1 through 5 find each equilibrium of the given difference equation and deter-
mine whether it is asymptotically stable or unstable.

rx2n
1. xn+1 = x2n +A
(r and A are nonnegative).

2. xn+1 = xn er(1−xn /K) .

3. xn+1 = rxn (1 + αxn )−β .

2xn
4. xn+1 = 1+xn
.
λxn
5. xn+1 = (1+axn )b
.

6. (a) A population is governed by the difference equation

xn+1 = xn e3−xn .

Show that all equilibria are unstable.

(b) The population of part (a) is to be stabilized by removing a fraction p(0 < p <
1) of the population in each time period after all births and deaths have taken
place, to give the model

xn+1 = (1 − p)xn e3−(1−p)xn .

For what values of p does the population have an asymptotically stable positive
equilibrium?

131
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

4.3 Continuous Models for Interacting Populations

When species interact the population dynamics of each species is affected. In general,
there is a whole web of interacting species, sometimes called a trophic web, which makes
for structurally complex communities. We consider here systems involving two or more
species, concentrating particularly on two-species systems. There are three main types of
interaction.

i If the growth rate of one population is decreased and the other increased, then the
populations are in a predator-prey situation.

ii If the growth rate of each population is decreased, then it is competition.

iii If each population’s growth rate is enhanced, then it is called mutualism or symbiosis.

4.3.1 Predator-Prey Models: Lotka-Volterra Systems

Volterra (1926) first proposed a simple model for the predation of one species by another
to explain the oscillatory levels of certain fish catches in the Adriatic. If N (t) is the prey
population and P (t) is that of the predator at time t, then Volterra’s model is;

dN
= N (a − bP ), (4.21)
dt

dP
= P (cN − d), (4.22)
dt
where a, b, c and d are positive constants.
The assumptions in the model are:

1. The prey in the absence of any predation grows unboundedly in a Malthusian way;
this is the aN term in (4.21).

2. The effect of the predation is to reduce the prey’s per capita growth rate by a term
proportional to the prey and predator populations; this is the −bN P term.

132
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

3. In the absence of any prey for sustenance the predator’s death rate results in expo-
nential decay, that is, the −dP term in (4.22).

4. The prey’s contribution to the predators’ growth rate is cN P ; that is, it is propor-
tional to the available prey as well as to the size of the predator population.

The N P terms can be thought of as representing the conversion of energy from one source
to another: bN P is taken from the prey and cN P accrues to the predators. We shall see
that this model has serious drawbacks. Nevertheless it has been of considerable value in
posing highly relevant questions and is a jumping-off place for more realistic models; this
is the main motivation for studying it here.

The model (4.21) and (4.22) is known as the Lotka-Volterra model since the same equations
were also derived by Lotka (1920; see also 1925) from a hypothetical chemical reaction
which he said it could exhibit periodic behaviour in the chemical concentrations.
As a first step in analysing the Lotka-Volterra model, we nondimensionalise (or scale) the
system by writing

cN (t) bP (t)
u(τ ) = , v(τ ) = , τ = at, α = d/a. (4.23)
d a

Substituting these values in the Lotka-Volterra model (4.21) and (4.22) which has four
parameters a, b, c and d, we obtain,

d du d a
= u(a − b v)
c dτ /a c b

a dv a d
= v(c u − d).
b dτ /a b c
Simplification gives the scaled Predator-Prey (Lotka-Volterra model)as:

du dv
= u(1 − v), = αv(u − 1). (4.24)
dτ dτ

Observe here that the dimensionless system has only one parameter α.

133
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

In the u, v phase plane these give,

dv αv(u − 1)
= , (4.25)
du u(1 − v)

which has singular points at u = v = 0 and u = v = 1. We can integrate (4.25) exactly

to get the phase trajectories
αu + v − ln uα v = H, (4.26)

where H > Hmin is a constant: Hmin = 1 + α is the minimum of H over all (u, v) and it
occurs at u = v = 1. For a given H > 1 + α, the trajectories (4.26) in the phase plane
are closed as illustrated in Figure 4.3.

Figure 4.3: Closed (u, v) phase plane trajectories, from (4.26) with various H, for the
LotkaVolterra system (4.24): H1 = 2.1, H2 = 2.4, H3 = 3.0, H4 = 4. The arrows denote
the direction of change with increasing time τ .

A closed trajectory in the u, v plane implies periodic solutions in τ for u and v in (4.24).
The initial conditions, u(0) and v(0), determine the constant H in (4.26) and hence the
phase trajectory in Figure 4.3. Typical periodic solutions u(τ ) and v(τ ) are illustrated in
Figure 4.4. From (4.24) we can see immediately that u has a turning point when v = 1
and v has one when u = 1.

134
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

A major inadequacy of the Lotka-Volterra model is clear from Figure 4.3, the solutions
are not structurally stable. Suppose, for example, u(0) and v(0) are such that u and v for
τ > 0 are on the trajectory H4 which passes close to the u and v axes. Then any small
perturbation will move the solution onto another trajectory which does not lie everywhere
close to the original one H4 . Thus a small perturbation can have a very marked effect,
at the very least on the amplitude of the oscillation. This is a problem with any system
which has a first integral, like (4.26) which is a closed trajectory in the phase plane. They
are called conservative systems; here (4.26) is the associated ”conservation law”. They
are usually of little use as models for real interacting populations (see one interesting and
amusing attempt to do so below). However, the method of analysis of the steady states
is typical.

Returning to the form (4.24), a linearisation about the singular points determines the type
of singularity and the stability of the steady states. A similar linear stability analysis has
to be carried out on equivalent systems with any number of equations.

Figure 4.4: Periodic solutions for the prey u(τ ) and the predator v(τ ) for the LotkaVolterra
system (4.24) with α = 1 and initial conditions u(0) = 1.25, v(0) = 0.66.

We first consider the steady state (u, v) = (0, 0). Let x and y be small perturbations
about (0, 0). If we keep only linear terms, (4.24) becomes
! ! ! !
dx
dτ
1 0 x x
dy
≈ =A . (4.27)
dτ
0 −α y y

135
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

The solution is of the form !

x(τ )
= Beλτ ,
y(τ )
where B is an arbitrary constant column vector and the eigenvalues λ are given by the
characteristic polynomial of the matrix A and thus are solutions of

1−λ 0
|A − λI| = =0 ⇒ λ1 = 1, λ2 = −α.
0 −α − λ

Since at least one eigenvalue, λ1 > 0, x(τ ) and y(τ ) grow exponentially and so u = 0 = v
is linearly unstable. Since λ1 > 0 and λ2 < 0 this is a saddle point singularity.

Linearising about the steady state u = v = 1 by setting u = 1 + x, v = 1 + y with |x| and

|y| small, (4.24) becomes
! ! !
dx
dτ
x 0 −1
dy
=A , A= , (4.28)
dτ
y α 0

with eigenvalues λ given by

−λ −1 √
=0 ⇒ λ1 , λ2 = ±i α. (4.29)
α −λ

Thus u = v = 1 is a center singularity since the eigenvalues are purely imaginary. Since
Re λ = 0 the steady state is neutrally stable. The solution of (4.28) is of the form
!
x(τ ) √ √
= lei ατ
+ me−i ατ
,
y(τ )

where l and m are eigen vectors. So, the solutions in the neighbourhood of the singular
√
point u = v = 1 are periodic in τ with period 2π/ α. In dimensional terms from (4.23)
this period is T = 2π(a/d)1/2 ; that is, the period is proportional to the square root of the
ratio of the linear growth rate, a, of the prey to the death rate, d, of the predators. Even
though we are only dealing with small perturbations about the steady state u = v = 1
we see how the period depends on the intrinsic growth and death rates. For example, an

136
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

increase in the growth rate of the prey will increase the period; a decrease in the predator
death rate does the same thing. Is this what you would expect intuitively?
In this ecological context the matrix A in the linear equations (4.27) and (4.28) is called
the community matrix, and its eigenvalues λ determine the stability of the steady states.
If Re λ > 0 then the steady state is unstable while if both Re λ < 0 it is stable.

The critical case Re λ = 0 is termed neutral stability (center).

There have been many attempts to apply the Lotka-Volterra model to real-world oscilla-
tory phenomena. In view of the system’s structural instability, they must essentially all
fail to be of quantitative practical use. As we mentioned, however, they can be important
as vehicles for suggesting relevant questions that should be asked.

4.3.2 Competition Models: Principle of Competitive Exclusion

Here two or more species compete for the same limited food source or in some way inhibit
each other’s growth. For example, competition may be for territory which is directly
related to food resources. Some interesting phenomena have been found from the study
of practical competition models; see, for example, Hsu et al. (1979). Here we discuss
a very simple competition model which demonstrates a fairly general principle which is
observed to hold in Nature, namely, that when two species compete for the same limited
resources one of the species usually becomes extinct.
Consider the basic 2-species Lotka-Volterra competition model with each species N1 and
N2 having logistic growth in the absence of the other. Inclusion of logistic growth in the
Lotka-Volterra systems makes them much more realistic, but to highlight the principle
we consider the simpler model which nevertheless reflects many of the properties of more
complicated models, particularly as regards stability. We thus consider
 
dN1 N1 N2
= r1 N1 1 − − b12 , (4.30)
dt K1 K1
 
dN2 N2 N1
= r2 N2 1 − − b21 , (4.31)
dt K2 K2

137
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

where

• r1 , K1 , r2 , K2 , b12 and b21 are all positive constants.

• r1 and r2 are the linear birth rates of N1 and N2 respectively.

• K1 and K2 are the carrying capacities of N1 and N2 respectively.

• b12 and b21 measure the competitive effect of N2 on N1 and N1 on N2 respectively:

they are generally not equal.

Note that the competition model (4.30) and (4.31) is not a conservative system like its
Lotka-Volterra predator-prey counterpart.
If we non-dimensionalise this model by writing

N1 N2 r2
u1 = , u2 = , τ = r1 t, ρ= ,
K1 K2 r1
(4.32)
K2 K1
a12 = b12 , a21 = b21
K1 K2

(4.30) and (4.31) become

du1
= u1 (1 − u1 − a12 u2 ) = f1 (u1 , u2 ),
dτ (4.33)
du2
= ρu2 (1 − u2 − a21 u1 ) = f2 (u1 , u2 ).
dτ

The steady states, and phase plane singularities, u∗1 , u∗2 , are solutions of f1 (u1 , u2 ) =
f2 (u1 , u2 ) = 0 which, from (4.33), are

u∗1 = 0, u∗2 = 0; u∗1 = 1, u∗2 = 0; u∗1 = 0, u∗2 = 1;

1 − a12 1 − a21 (4.34)
u∗1 = , u∗2 = .
1 − a12 a21 1 − a12 a21

The last of these is only of relevance if u∗1 ≥ 0 and u∗2 ≥ 0 are finite, in which case
a12 a21 6= 1. The four possibilities are seen immediately on drawing the null clines f1 = 0
and f2 = 0 in the u1 , u2 phase plane as shown in Figure 4.5. The crucial part of the null

138
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

clines are, from (4.33), the straight lines

1 − u1 − a12 u2 = 0, 1 − u2 − a21 u1 = 0.

The first of these together with the u2 -axis is f1 = 0, while the second, together with the
u1 -axis is f2 = 0.

Figure 4.5: The null clines for the competition model (4.33). f1 = 0 is u1 = 0 and
1 − u1 − a12 u2 = 0 with f2 = 0 being u2 = 0 and 1 − u2 − a21 u1 = 0. The intersection of
the two solid lines gives the positive steady state if it exists as in (a) and (b): the relative
sizes of a12 and a21 as compared with 1 for it to exist are obvious from (a) to (d).

The stability of the steady states is again determined by the community matrix which,

139
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

for (4.33), is
! !
∂f1 ∂f1
∂u1 ∂u2
1 − 2u1 − a12 u2 −a12 u1
A= ∂f2 ∂f2
= . (4.35)
∂u1 ∂u2 u∗1 ,u∗2
−ρa21 u2 ρ(1 − 2u2 − a21 u1 ) u∗1 ,u∗2

The first steady state in (4.34), that is, (0, 0), is unstable since the eigenvalues λ of its
community matrix, given from (4.35) by

1−λ 0
|A − λI| = =0 ⇒ λ1 = 1, λ2 = ρ,
0 ρ−λ

are positive. For the second of (4.34), namely, (1, 0), (4.35) gives

−1 − λ −a12
|A − λI| = =0 ⇒ λ1 = −1, λ2 = ρ(1 − a21 ),
0 ρ(1 − a21 ) − λ

and so  
stable a >1
21
u∗1 = 1, u∗2 = 0 is if (4.36)
unstable a < 1.
21

Similarly, for the third steady state, (0, 1), the eigenvalues are λ1 = ρ, λ2 = (1 − a12 ) and
so  
stable a >1
12
u∗1 = 0, u∗2 = 1 is if (4.37)
unstable a < 1.
12

Finally for the last steady state in (4.34) , when it exists in the positive quadrant, the
matrix A from (4.35) is
!
a12 − 1 a12 (a21 − 1)
A = (1 − a12 a21 )−1
ρa21 (a21 − 1) ρ(a21 − 1)

which has eigenvalues

λ1 , λ2 = [2(1 − a12 a21 )]−1 [(a12 − 1) + ρ(a21 − 1)

1/2
(4.38)
± [(a12 − 1) + ρ(a21 − 1)]2 − 4ρ(1 − a12 a21 )(a12 − 1)(a21 − 1)

].

140
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

The sign of λ, or Re λ if complex, and hence the stability of the steady state, depends on
the size of ρ, a12 and a21 . There are several cases we have to consider, all of which have
ecological implications which we come to below.

Before discussing the various cases note that there is a confined set on the boundary of
which the vector of the derivatives, (du1 /dτ, du2 /dτ ), points along it or inwards: here it
is a rectangular box in the (u1 , u2 ) plane. From (4.33) this condition holds on the u1 -
and u2 -axes. Outer edges of the rectangle are, for example, the lines u1 = U1 where
1 − U1 − a12 u2 < 0 and u2 = U2 where 1 − U2 − a21 u1 < 0. Any U1 > 1, U2 > 1 suffice. So
the system is always globally stable.

The various cases are: (i) a12 < 1, a21 < 1, (ii) a12 > 1, a21 > 1, (iii) a12 < 1, a21 > 1,
(iv) a12 > 1, a21 < 1. All of these are analyzed in a similar way. Figures 4.5(a) to (d)
and Figures 4.6(a) to (d) relate to these cases (i) to (iv) respectively. By way of example,
we consider just one of them, namely, (ii). The analysis of the other cases is left as an
exercise. The results are encapsulated in Figure 4.6.
The arrows indicate the direction of the phase trajectories. The qualitative behaviour of
the phase trajectories is given by the signs of du1 /dτ , namely, f1 (u1 , u2 ), and du2 /dτ
which is f2 (u1 , u2 ), on either side of the null clines.
Case a12 > 1, a21 > 1. This corresponds to Figure 4.5(b). From (4.36) and (4.37), (1, 0)
and (0, 1) are stable. Since 1 − a12 a21 < 0, (u∗1 , u∗2 ), the fourth steady state in (4.34),
lies in the positive quadrant and from (4.38) its eigenvalues are such that λ2 < 0 < λ1
and so it is unstable to small perturbations: it is a saddle point. In this case, then,
the phase trajectories can tend to either one of the two steady states, as illustrated in
Figure 4.6(b). Each steady state has a domain of attraction. There is a line, a separatrix,
which divides the positive quadrant into 2 non overlapping regions I and II as in Figure
4.6(b). The separatrix passes through the steady state (u∗1 , u∗2 ): it is one of the saddle
point trajectories in fact.
Now consider some of the ecological implications of these results. In case (i) where a12 < 1
and a21 < 1 there is a stable steady state where both species can exist as in Figure
4.5(a). In terms of the original parameters from (4.32) this corresponds to b12 K2 /K1 < 1
and b21 K1 /K2 < 1. For example, if K1 and K2 are approximately the same and the

141
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

Figure 4.6: Schematic phase trajectories near the steady states for the dynamic behaviour
of competing populations satisfying the model (4.33) for the various cases. (a) a12 <
1, a21 < 1. Only the positive steady state S is stable and all trajectories tend to it.
(b) a12 > 1, a21 > 1. Here, (1, 0) and (0, 1) are stable steady states, each of which
has a domain of attraction separated by a separatrix which passes through (u∗1 , u∗2 ). (c)
a12 < 1, a21 > 1. Only one stable steady state exists, u∗1 = 1, u∗2 = 0 with the whole
positive quadrant its domain of attraction. (d) a12 > 1, a21 < 1. The only stable steady
state is u∗1 = 0, u∗2 = 1 with the positive quadrant as its domain of attraction. Cases (b)
to (d) illustrate the competitive exclusion principle whereby 2 species competing for the
same limited resource cannot in general coexist.

interspecific competition, as measured by b12 and b21 , is not too strong, these conditions
say that the two species simply adjust to a lower population size than if there were no
competition. In other words, the competition is not aggressive. On the other hand if the
b12 and b21 are about the same and the K1 and K2 are different, it is not easy to tell what
will happen until we form and compare the dimensionless groupings a12 and a21 .

142
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

In case (ii), where a12 > 1 and a21 > 1, if the K’s are about equal, then the b12 and b21 are
not small. The analysis then says that the competition is such that all three nontrivial
steady states can exist, but, from (4.36) to (4.38), only (1, 0) and (0, 1) are stable, as in
Figure 4.6(b). It can be a delicate matter which ultimately wins out. It depends crucially
on the starting advantage each species has. If the initial conditions lie in domain I then
eventually species 2 will die out, u2 → 0 and u1 → 1; that is, N1 → K1 the carrying
capacity of the environment for N1 . Thus competition here has eliminated N2 . On the
other hand if N2 has an initial size advantage so that u1 and u2 start in region II then
u1 → 0 and u2 → 1 in which case the N1 -species becomes extinct and N2 → K2 , its
environmental carrying capacity. We expect extinction of one species even if the initial
populations are close to the separatrix and in fact if they lie on it, since the ever present
random fluctuations will inevitably cause one of ui , i = 1, 2 to tend to zero.
Cases (iii) and (iv) in which the interspecific competition of one species is much stronger
than the other, or the carrying capacities are sufficiently different so that a12 = b12 K2 /K1 <
1 and a21 = b21 K1 /K2 > 1 or alternatively a12 > 1 and a21 < 1, are quite definite in the
ultimate result. In case (iii), as in Figure 4.6(c), the stronger dimensionless interspecific
competition of the u1 -species dominates and the other species, u2 , dies out. In case (iv)
it is the other way round and species u1 becomes extinct.
Although all cases do not result in species elimination, those in (iii) and (iv) always do
and in (ii) it is inevitable due to natural fluctuations in the population levels. This work
led to the principle of competitive exclusion which was mentioned above. Note that the
conditions for this to hold depend on the dimensionless parameter groupings a12 and
a21 : the growth rate ratio parameter ρ does not affect the gross stability results, just
the dynamics of the system. Since a12 = b12 K2 /K1 , a21 = b21 K1 /K2 the conditions for
competitive exclusion depend critically on the interplay between competition and the
carrying capacities as well as the initial conditions in case (ii).
Suppose, for example, we have 2 species comprised of large animals and small animals,
with both competing for the same grass in a fixed area. Suppose also that they are equally
competitive with b12 = b21 . With N1 the large animals and N2 the small, K1 < K2 and so
a12 = b12 K2 /K1 < b21 K1 /K2 = a21 . As an example if b12 = 1 = b21 , a12 < 1 and a21 > 1
then in this case N1 → 0 and N2 → K2 ; that is, the large animals become extinct.

143
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

The situation in which a12 = 1 = a21 is special and, with the usual stochastic variability in
nature, is unlikely in the real world to hold exactly. In this case the competitive exclusion
of one or the other of the species also occurs.

4.3.3 Mutualism or Symbiosis

There are many examples where the interaction of two or more species is to the advan-
tage of all. Mutualism or symbiosis often plays the crucial role in promoting and even
maintaining such species: plant and seed dispersal is one example. Even if survival is
not at stake the mutual advantage of mutualism or symbiosis can be very important.
As a topic of theoretical ecology, even for two species, this area has not been as widely
studied as the others even though its importance is comparable to that of predator-prey
and competition interactions. This is in part due to the fact that simple models in the
Lotka-Volterra vein give silly results. The simplest mutualism model equivalent to the
classical Lotka-Volterra predator-prey one is

dN1 dN2
= r1 N1 + a1 N1 N2 , = r2 N2 + a2 N2 N1 ,
dt dt

where r1 , r2 , a1 and a2 are all positive constants. Since dN1 /dt > 0 and dN2 /dt > 0, N1
and N2 simply grow unboundedly in, as May (1981) so aptly put it, ’an orgy of mutual
benefaction.’

Realistic models must at least show a mutual benefit to both species, or as many as are
involved, and have some positive steady state or limit cycle type oscillation.

As a first step in producing a reasonable 2-species model we incorporate limited carrying

capacities for both species and consider
 
dN1 N1 N2
= r1 N1 1 − + b12
dt K1 K1
  (4.39)
dN2 N2 N1
= r2 N2 1 − + b21 ,
dt K2 K2

where

144
 4.3. CONTINUOUS MODELS FOR INTERACTING POPULATIONS

• r1 , r2 , K1 , K2 , b12 and b21 are all positive constants.

• r1 and r2 are birth rates of N1 and N2 respectively.

• K1 and K2 are carrying capacity of N1 and N2 respectively.

• b12 and b21 measure the positive effect of N2 on N1 and N1 on N2 respectively.

If we use the same non-dimensionalisation as in the competition model (the signs preceding
the b’s are negative there), namely, (4.32), we get

du1
= u1 (1 − u1 − a12 u2 ) = f1 (u1 , u2 ),
dτ (4.40)
du2
= ρu2 (1 − u2 − a21 u1 ) = f2 (u1 , u2 ),
dτ

where
N1 N2 r2
u1 = , u2 = , τ = r1 t, ρ = ,
K1 K2 r1
(4.41)
K2 K1
a12 = b12 , a21 = b21 .
K1 K2
Analysing the model in the usual way we start with the steady states (u∗1 , u∗2 ) which from
(4.40) are
(0, 0), (1, 0), (0, 1),

1 + a12 1 + a21
 (4.42)
, , positive if δ = 1 − a12 a21 > 0.
δ δ
After calculating the community matrix for (4.40) and evaluating the eigenvalues λ for
each of (4.42) it is straightforward to show that (0, 0), (1, 0) and (0, 1) are all unstable:
(0, 0) is an unstable node and (1, 0) and (0, 1) are saddle point equilibria. If 1 − a12 a21 < 0
there are only three steady states, the first three in (4.42), and so the populations become
unbounded. We see this by drawing the null clines in the phase plane for (4.40), namely,
f1 = 0, f2 = 0, and noting that the phase trajectories move off to infinity in a domain in
which u1 → ∞ and u2 → ∞ as in Figure 4.7(a).
When 1 − a12 a21 > 0, the fourth steady state in (4.42) exists in the positive quadrant.
Evaluation of the eigenvalues of the community matrix shows it to be a stable equilibrium:
it is a node singularity in the phase plane. This case is illustrated in Figure 4.7(b). Here all

145
 4.4. MATHEMATICAL EPIDEMIOLOGY

Figure 4.7: Phase trajectories for the mutualism model for two species with limited car-
rying capacities given by the dimensionless system (4.40). (a) a12 a21 > 1: unbounded
growth occurs with u1 → ∞ and u2 → ∞ in the domain bounded by the null clines-
the solid lines. (b) a12 a21 < 1: all trajectories tend to a positive steady state S with
u∗1 > 1, u∗ > 1 which shows the initial benefit that accrues since the carrying capacity for
each species is greater than if no interaction were present.

the trajectories in the positive quadrant tend to u∗1 > 1 and u∗2 > 1; that is, N1 > K1 and
N2 > K2 and so each species has increased its steady state population from its maximum
value in isolation.
This model has certain drawbacks. One is the sensitivity between unbounded growth and
a finite positive steady state. It depends on the inequality a12 a21 < 1, which from (4.41)
in terms of the original parameters in (4.39) is b12 b21 < 1: the b’s are dimensionless. So if
symbiosis of either species is too large, this last condition is violated and both populations
grow unboundedly.

4.4 Mathematical Epidemiology

Introduction

Epidemiology is the study of infectious disease, the causes of their occurrence and their
spread in space and time.

146
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

Mathematical epidemiology is about obtaining understanding of epidemiological phe-

nomena, translating assumptions regarding epidemiological features to mathematical lan-
guage, finding solutions of mathematical problems and last but certainly not least trans-
lating the result in to epidemiology.
The main use of mathematics in epidemiology is to gain insight on epidemics to see how
the dynamics of an infectious disease depends on the basic parameters that characterize
it. Infectious disease concerns about a communicable disease which comes from direct
or indirect transmission. Direct transmission is a transfer of infectious agent from the
infected individual directly to the host just like touching, sexual intercourse for example
HIV/AIDS. Indirect transmission is the transfer of an infectious agent by contaminated
object like by an environmental water contamination or food contamination for example
Malaria.

4.5 Mathematical Modeling of Infectious Diseases:

Issues and Approaches

A disease is infectious if the causative agent, whether a virus, bacterium, protozoa, or

toxin, can be passed from one host to another through modes of transmission such as
direct physical contact, airborne droplets, water or food, disease vectors, or mother to
newborn.
The objective of a mathematical model of an infectious disease is to describe the trans-
mission process of the disease, which can be defined generally as follows: when infectious
individuals are introduced into a population of susceptibles, the disease is passed to other
individuals through its modes of transmission thus spreading in the population. An in-
fected individual may remain asymptomatic at the early stage of infection, only later
developing clinical symptoms and being diagnosed as a disease case. If the number of
cases rises above the usual average within a short period of time, a disease outbreak
occurs. When the disease spreads quickly to many people, it is an epidemic. Infected
individuals recover from infection, either through treatment or due to the action of the
immune system, and gain various degrees of acquired immunity against reinfection. When

147
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

the pool of susceptible individuals is sufficiently depleted, new infections will cease and
the epidemic slows down and stops. If fresh susceptibles are added to the population,
either from birth or migration, or if reinfection occurs easily, the epidemic may persist
and the infection may remain in the population over a long period of time. In this case,
the disease is said to be endemic in the population. If the disease spreads spatially on
a global scale to many countries and continents, a pandemic occurs. The 1918 Spanish
flu that spread to all continents and killed over 50 million people is a classic example of
a global pandemic. With modern air travel, many emerging and reemerging infectious
diseases have an increasing potential to cause a global pandemic.
Facing an imminent epidemic, public health authorities will be looking for answers to the
following important questions:

1. How severe will the epidemic be? The severity can be measured in two different
ways:

(a) Total number of infected people who may require medical care.
(b) Maximum number of infected people at any given time

2. How long will it last? When will it peak? What will be its time course?

3. How effective will quarantine or vaccination be?

4. What quantity of vaccine or anti-viral drugs should be stockpiled?

5. What are effective measures to contain, control, and eradicate an endemic disease?

Partial answers may be obtained using a variety of approaches. Mathematical modeling

has proven to be an important tool in assisting public health authorities to make informed
decisions.

This brings us to the obvious question: why is mathematical modeling of infectious dis-
eases useful? Part of the answer is that traditional methods using experimental and
statistical approaches may not be adequate for various reasons:

1. Infectious diseases often affect a large population of individuals over a large geo-
graphic area. Experiments conducted in laboratories are often inadequate simply
because of the huge difference in scale.

148
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

2. For infectious diseases in humans, large-scale experiments may be impossible or

unethical.

3. Existing data sets about the disease may not be complete or accurate enough for
the statistical analysis to be reliable. For instance, there will be little or no disease
surveillance data available on infected people who are asymptomatic.

4. Since repeated experiments of disease outbreaks in a population are rarely possible,

disease surveillance data often represents a single outbreak (sample) with a large
amount of information. Statistical analysis of high-dimensional data from small
samples is a challenge since statistics theory relies on large samples.

Mathematical modeling can provide an understanding of the underlying mechanisms of

disease transmission and spread, help to pinpoint key factors in the disease transmission
process, suggest effective control and preventive measures, and provide an estimate for
the severity and potential scale of the epidemic. Put it simply, mathematical modeling
should become part of the toolbox of public health research and decision-making.

The next question we would like to answer is: what is the general process of mathematical
modeling? Generally speaking, the modeling process involves the following six stages:

1. Make assumptions about the disease transmission process based on the best available
biological knowledge on the pathogenesis of the infection and epidemiology of the
disease.

2. Set up mathematical models for the transmission process based on these assump-
tions. This usually starts from drawing the transfer diagram and then deriving the
mathematical equations.

3. Perform mathematical analysis on the model to understand all possible qualitatively

distinct model outcomes. This is typically done by applying existing mathematical
theories on stability and bifurcations in conjunction with numerical simulations.

4. Interpret the mathematical findings within the modeling context. These interpre-
tations form our understanding of the disease transmission process entailed by the
set of assumptions made in Step (1).

149
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

5. Collect available disease data from public health agencies and from research publi-
cations. Validate the model using data.

6. Improve the model by modifying the earlier assumptions in Step (1), and produce
a more accurate understanding of the disease process.

It is important to keep in mind that a mathematical model is an approximation of the

real disease process. It is a mathematical translation of our hypotheses about disease
transmission. Another important role for mathematical models is hypothesis testing: by
comparing the model outcomes with existing knowledge or data of the disease, we can
use the model to test various hypotheses about the disease. Compared to experimental
approaches, the modeling approach has the advantage of saving enormous amounts of
time and resources.
We should caution that a mathematical model is not a magic bullet. There are often many
difficulties associated with mathematical modeling. The following is a list of important
issues involved in the mathematical modeling process:

1. Due to our limited knowledge about an infectious disease, realistic assumptions

about its transmission process are not always possible. Various degrees of simplifi-
cation need to be made. Very often, our assumptions are simply hypotheses. When
interpreting findings from the model, we always need to keep these limitations in
mind.

2. Model validation using disease data is important because it provides a test of our
modeling hypotheses. This may not always be possible or may be difficult to do
depending on the availability and quality of data.

3. Mathematical analysis of the model may be limited by existing mathematical theory.

There is always a trade-off in mathematical modeling between more realistic and therefore
more complex models and our ability to analyze the model mathematically and obtain
useful information for interpretation. Advancement in mathematical theory and method-
ology often allows us to successfully use more realistic models. When using mathematical
models to analyze or interpret disease data, it is not always true that a more realis-
tic model will do better. Part of the reason is that more realistic models incorporate

150
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

a greater degree of biological complexity and hence introduce more model parameters.
With the same data set, it may be more difficult to estimate all the parameter values
for a more complex model compared to a simpler model. The result might be a greater
degree of uncertainty in model outcomes.

There are three general approaches to mathematical modeling of infectious diseases:

1. Statistical models: These models are very data oriented and are constructed to deal
with a specific set of data.

• Advantages: statistical models are widely used in epidemiology and public

health research.
• Drawbacks: statistical models require large samples of data.

2. Deterministic models: These are typically models using differential and difference
equations of various forms. The key assumption is that the sizes of the susceptible
and infectious populations are continuous functions of time. The models describe
the dynamic interrelations among the rates of change and population sizes.

• Advantages: mathematical theories for this type of model are more mature
in comparison to stochastic models; the derivation of mathematical models
are less data dependent in comparison to statistical models; and mathematical
models are suited for making predictions.
• Drawbacks: these models are not expected to be valid if the population sizes
are very small, in which case stochastic disturbances become non-negligible.

3. Stochastic models: In this type of model, disease infection is treated as a stochastic

process. The models describe dynamic interrelations of its probability distributions.

• Advantages: stochastic models are suitable to deal with small population

groups such as a small community or a single hospital, or where a few infected
individuals are highly active and have a high number of infectious contacts.
• Drawbacks: mathematical analysis of stochastic models is difficult due to a
lack of mathematical machinery; model analysis largely relies on observations
from a huge number of numerical simulations.

151
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

4.5.1 Deterministic Epidemic Models:

Compartmental Approach

In this subsection, we explain how to set up a mathematical model for the transmission
process of an infectious disease using a compartmental approach. We first partition the
host population into mutually exclusive groups compartments according to the natural
history of the disease. For a simple infectious disease, possible compartments may be:
S: susceptible hosts, I: infectious hosts, R: recovered hosts.

Then, we illustrate the transmission process schematically in a carton, called a trans-

fer diagram, shown in Figure 4.8. In the transfer diagram, the arrows indicate movements
of individuals among compartments. The term ”removal” includes loss of individuals
through death or out-migration. The goal of modeling is to track the number of hosts

Figure 4.8: Transfer diagram for an SIR compartment model

in each of the three compartments at any given time t, and we denote these numbers by
S(t), I(t), and R(t) accordingly. To set up the compartmental model, we consider a small
time interval [t, t + ∆t] and the net change in the number of individuals in each compart-
ment. In the transfer diagram, the arrows indicate the direction of individual movement.
The net change of the number of hosts in a compartment is the number coming into
the compartment minus the number leaving the compartment during the time interval.

152
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

Applying this principle to each compartment, we arrive at the following equations:

∆S(t) = new susceptibles + transfer from R - new infections

−removal from S
∆I(t) = new infections - transfer into R - removal from I (4.43)
∆R(t) = transfer from I - transfer into S - removal from R

If we divide both sides of these equations by ∆t and let ∆t → 0, then the left-hand side
will be the derivatives S 0 (t), I 0 (t), and R0 (t), since

∆S(t) S(t + ∆t) − S(t)

= → S 0 (t)
∆t ∆t

as ∆t → 0, and similar relations hold for I 0 (t) and R0 (t). The terms on the right-hand
side will become instantaneous rates of incidence, recovery, and removal. We thus have
the following differential equations:

S 0 (t) = influx of new susceptibles + transfer rate from R - incidence rate

− removal rate from S
I 0 (t) = incidence rate - transfer rate into R - removal rate from I (4.44)
R0 (t) = transfer rate from I - transfer rate into S - removal rate from R .

If we express all the terms on the right-hand side as functions of S(t), I(t), and R(t), we
will obtain a system of differential equations for S(t), I(t), and R(t), which will form our
mathematical model. It is important to note that how these terms are expressed as func-
tions of S(t), I(t), and R(t) is based on our hypotheses regarding the biological processes
of disease transmission and population transfer among compartments. Therefore, differ-
ent hypotheses will give rise to different forms of the model, and may lead to different
model outcomes. If data is available to verify our model outcomes, then the model can
be used to test the validity of our hypotheses about the disease transmission process.
Deterministic compartmental modeling is by far the most common approach to theoret-
ical population biology. These models are simple to develop and they enable the use of

153
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

a powerful set of analytic tools such as elegant algebraic relationships. The determin-
istic compartmental models are often adequate and are probably the most appropriate
approach in most circumstances of established epidemics.

However, these models have their limitations and inappropriate assumptions . They are
certainly not appropriate for addressing circumstances for which there are only a handful
of individuals or when individual level intervention strategies, such as contact tracing, is
the focal question for investigation. There are times in which the stochastic elements are
essential to be modeled explicitly. But if there are large populations and the epidemic is
well-established, the stochastic models do not provide any more insight than deterministic
models, except they provide more variability in the outcome trajectories. There are several
deterministic models of which the most notable are:

• Susceptible- Infectious model(SI- model)

• Susceptible- Infectious- removed model(SIR- model)

• Susceptible- Infectious-Susceptible model(SIS- model)

• Susceptible-Exposed- Infectious- Removed model(SEIR- model)

Susceptible[S(t)]: is used to represent the number of individuals not yet infected with
the disease at a time t or those Susceptible for the disease, that can be cached by the
disease.

Infective [I(t)]: is used to represent the number of individuals who have been infected
by the disease and are capable of spreading the disease to those in the susceptible category
at a time.

Recovered [R(t)]: is the compartment used for those individuals who have been infected
and then removed from the disease either due to immunization or due to death . Those
in this category are not able to infected again or to transmit the infection to others.
Remark: The choice of which compartment to use in developing a model depends on
the nature of the disease and objective of the modeller. We will define and explain the
nature of those models as follows:

154
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

4.5.2 SI Model with out Demography

Differential equations describe the rate of change of a given quantity relative to something
else. The rate of change in the number of individuals in a given category over time is just
given by the difference between the number of individuals entering the category and the
number of individuals leaving the category per unit time. That is,
(The number of individuals entering the category per unit time)−(The number of indi-
viduals leaving the category per unit time).
The first step of a mathematical model is to divide the population, consisting of N in-
dividuals, into a set of compartments or categories. Susceptibles (S) are the number of
individuals who can be infected but have not yet contracted the disease. Infectives (I)
are the number of individuals who are infected and are able to transmit the disease. For
example let us consider the simple SI model. From the above flow chart we see that no

Figure 4.9: Flow chart of SI model with out birth and death rates.

one enters the Susceptible compartment and newly infected individuals leaving the com-
partment hence the rate of change in the number of susceptible individuals is given by
the expression:
dS
dt
= − number of individuals who are newly infected per unit time.

Using mathematical notation, the above equation would be written as:

dS
= −λ(t)S(t)
dt

where λ(t) is the force of infection and it can be replaced by the equation

λ(t) = β(t)I(t)

155
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

where β(t) is the rate at which two specific individuals come in to effective contact per
unit time and I(t) is the number of infectious individuals at time t. Similarly from the
flow chart we see that newly infected individuals enter in to this compartment. Then, the
corresponding dynamical system is given by

dS
= −λSI (4.45)
dt
dI
= λSI (4.46)
dt

S(0) = S0 ≥ 0; I(0) = Io ≥ 0

In this model we assume that the population is constant, i.e N = S(t) + I(t) and mixing
of a population is homogeneous. Infection rate is proportional to the number of infective
that is λI.

4.5.3 SI Model with Demography

This model considers the natural birth rate of the susceptible group and the natural death
rate of both groups and its flow chart is given by

Figure 4.10: Flow chart of SI model with birth and death rates.

The corresponding SI mathematical model represented by the following dynamical sys-

156
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

tems

dS SI
= µN − λ − µS (4.47)
dt N
dI SI
= λ − µI (4.48)
dt N

S(0) = S0 ≥ 0; I(0) = Io ≥ 0

Since we have assumed that the total population is constant and given by N = S(t)+I(t),
dS dI
therefore we do have dt
+ dt
= 0.

4.5.4 SIR Model with out Demography

This model consists of three compartments i.e Susceptible(S), Infected(I) and Removed(R)
Susceptibles [S(t)]: are the number of individuals who can be infected but have not
yet contracted the disease.

Infectives [I(t)]: are the number of individuals who are infected and are able to transmit
the disease.

Removed[R(t)]: are the number of individuals who have been infected and then re-
covered from the disease either due to immunization( E.g measle) or due to death(E.g
Rabies). Those in this group are not able to be infected again or to transmit the infection
to others.
Example: The dynamics of an epidemic like flu are much faster than the dynamic of
birth and death for such cases we used simple SIR model( SIR model with out birth rate
and death rate)
Its flow chart is given by

Figure 4.11: Flow chart of SIR model with out birth and death rates.

157
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

The corresponding simple SIR mathematical model represented by the following dynam-
ical systems,

dS
= −βSI (4.49)
dt
dI
= βSI − αI (4.50)
dt
dR
= αI (4.51)
dt
(4.52)

S(0) = So ≥ 0; I(0) = Io ≥ 0; R(0) = Ro ≥ 0

Since we have assumed that the total population is constant and given by N = S(t) +
dS dI dR
I(t) + R(t), therefore we do have dt
+ dt
+ dt
= 0.

4.5.5 SIR Model with Demography

In this subsection we consider birth rate and death rate because there are diseases that
stay and become endemic for long time so for such cases we need to include birth and
death. for example malaria.

Schematic representation of this model is give by:

Figure 4.12: Flow chart of SIR model with birth and death rates.

The corresponding simple SIR mathematical model represented by the following dynam-

158
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

ical systems,

dS
= Qo − βSI − µS (4.53)
dt
dI
= µS − αI − µI (4.54)
dt
dR
= αI − µR (4.55)
dt

S(0) = So ≥ 0; I(0) = Io ≥ 0; R(0) = Ro ≥ 0

Since we have assumed that the total population is constant and given by N = S(t) +
I(t) + R(t), therefore we do have dS
dt
+ dI
dt
+ dR
dt
= 0.

4.5.6 SIS Model with out Demography

SIR model capture the dynamics of acute infections that either kill or confer life long
immunity once recovered. However , numerous infectious disease confer no long lasting
immunity like sexual transmission disease, for these disease, individuals can be infected
multiples times through out their lives with no apparent immunity for example Gonorrhea
for such cases we consider a model called SIS model. In this case recovery from infection
is followed by an instant return to susceptible compartment. Schematic representation of
this model is give by

Figure 4.13: Flow chart of SIS model with out birth and death rates.

The corresponding simple SIS mathematical model represented by the following dynamical

159
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

systems,

dS
= γI − βSI (4.56)
dt
dI
= βSI − γI (4.57)
dt

S(0) = So ≥ 0; I(0) = Io ≥ 0

4.5.7 SIS Model with Demography

Schematic representation of this model is give by:

Figure 4.14: Flow chart of SIS model with birth and death rates.

The corresponding simple SIS model with birth rate and death rate represented by the
following dynamical systems,

dS
= Qo + γI − µS − βSI (4.58)
dt
dI
= βSI − γI − µI (4.59)
dt

S(0) = So ≥ 0; I(0) = Io ≥ 0

160
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

4.5.8 SEIR Model with out Demography

The SI, SIR and SIS model discussed above takes in to account only those disease which
cause an individual to be able to infect others immediately up on their infection, but many
diseases have what is termed as a latent or exposed phase , during which the individual
is said to be infected but not infectious.
In this model the SEIR contains one more compartment. The Exposed(E) compartment
contains those people who are infected but the symptom of the disease are not yet visible
or they may not communicate the disease. For some disease, it takes certain time for an
infective agent to multiple inside the host up to the critical level so that the disease actually
manifest it self in the body of the host(i.e a disease in incubation period) Schematic
representation of this model is give by

Figure 4.15: Flow chart of SEIR model with out birth and death rates.

The corresponding simple SEIR model with out birth rate and death rate represented by
the following dynamical systems,

dS
= −βSE (4.60)
dt
dE
= βSE − αE (4.61)
dt
dI
= αE − γI (4.62)
dt
dR
= γI (4.63)
dt

S(0) = So ≥ 0; E(0) = Eo ≥ 0; I(0) = Io ≥ 0.

4.5.9 SEIR Model with Demography

The flow chart of this model is given by:

161
 4.5. MATHEMATICAL MODELING OF INFECTIOUS DISEASES:
ISSUES AND APPROACHES

Figure 4.16: Flow chart of SEIR model with birth and death rates.

The corresponding SEIR model with birth rate and death rate represented by the following
dynamical systems,

dS
= Qo − βSE − µS (4.64)
dt
dE
= βSE − αE − µE (4.65)
dt
dI
= αE − γI − µI (4.66)
dt
dR
= γI − µR (4.67)
dt

S(0) = So ≥ 0; E(0) = Eo ≥ 0; I(0) = Io ≥ 0.

162
 REFERENCES

[1] Allee, W.C. (1931) The Social Life of Animals, Heinemann, London.

[2] Ayala, F.J., M.E. Gilpin, and J.G. Ehrenfeld (1973) Competition between species:
Theorical models and experimental tests, Theoretical Pop. Biol., 4:331-356.

[3] Bender, Edward A. (1942). An introduction to mathematical modeling.”A Wiley In-

terscience publication.”

[4] Bochner, S. (1966) . The Role of Mathematics in the Rise of Science. Princeton Uni-
versity Press.

[5] Clive L. Dym and Elizabeth Ivey (1980). Principles of Mathematical Modeling, Ist
Edition,, Academic Press.

[6] Cohen , K. J. and R. M. Cyert (1965) . Theory of the Firm : Resource Allocation in
a Market Economy. Prentice-Hall .

[7] Cyert, R. M. and J. G. March (1963). A Behavioral Theory of the Firm.Prentice-Hall

.

[8] E. Carson and C. Cobelli (2001.),Modelling Methodology for Physiology and Medicine,
Academic Press, San Diego, CA.

[9] Edelstein-Keshet, L. (1988) Mathematical Models in Biology, Random House, New

York.[reprinted as SIAM Classics in Applied Mathematics 46 (2005)].

163
 REFERENCES

[10] Ezekiel, M. (1937-1 938). The cobweb theorem . Quart. J. Econom ics 52 :255-2 80.
Reprinted in G. Haberler, ed . (1944) Readings in Business Cycle Theory. Irwin .

[11] Fibonacci, L. (1202) Tipographia, delle Scienze Mathematica, Roma.

[12] Gompertz, B. (1825) On the nature of the function expressing the law of human
mortality, Phil. Trans., 115:513-585

[13] Hassell, M.P. (1975) Density dependence in single-spacies populations, J. Animal.

Ecol., 44:283-295.

[14] Homans, G. (1950). The Human Group. Harcourt Brace .

[15] Levins, R. (1968) . Evolution in Changing Environmen ts . Princeton University Press

.

[16] Lotka, A.J. (1925) Elements of Physical Biology, Williams & Wilkins, Baltimore;
Reissued as Elements of Mathematical Biology, Dover, New York (1956).

[17] MacArthur , R. H. and E. O. Wilson (1967) . The Theory of Island Biogeography.

Princeton University Press.

[18] Malthus, T.R. (1798) An Essay on the Principle of Population, 1a Ed, J Johnson in
St Pauls Churchyard, London.

[19] May R. (1981) Theoretical ecology, principles and applications, Second Edition, Sin-
auer Associates, Massachusetts.

[20] McMahon, T. A. (1971). Rowing : A similarity analysis. Science 173 : 349-351.

[21] McMahon, T. A. (1973) . Size and shape in biology. Science 179 : 1201-1204 .

[22] Nisbet, R.M. and W.S.C. Gurney (1982) Modeling Fluctuating Populations, Wiley-
Interscience, Chichester.

[23] P. D. Cha, J. J. Rosenberg, and C. L. Dym(2000), Fundamentals of Modeling and

Analyzing Engineering Systems, Cambridge University Press,New York.

[24] Pearl, R. and L.J. Reed (1920) On the rate of growth of the population of the United
States since 1790 and its mathematical representation, Proc. Nat. Acad. Sci., 6:275-
288.

164
 REFERENCES

[25] Smith, F.E. (1963) Population dynamics in Daphnia magna and a new model for
population growth, Ecology, 44:651-663.

[26] Tsipis, K. (1975a). The accuracy of strategic missiles. Sci. Amer. 233(1) :14–2 3.

[27] Verhulst, P.F. (1838) Notice sur la loi que la population suit dans son accroissement,
Corr. Math. et Phys., 10:113-121.

[28] Verhulst, P.F. (1845) Recherches mathematiques sur la loi daccroissement de la pop-
ulation, Mem Acad Roy, Brussels, 18:1-38.

[29] Volterra, V. (1926) Variazioni e fluttazioni del numero dindividui in specie animali
conviventi, Mem. Acad. Sci. Lincei, 2:31-13.

[30] Wilson , E. O. and W. H. Bossert (1971). A Primer of Population Biology .Sinauer

Associates.

165