Dynamical Systems For Biological Modeling An Introduction
Dynamical Systems For Biological Modeling An Introduction
FOR
DYNAMICAL SYSTEMS
BIOLOGICAL MODELING
DYNAMICAL SYSTEMS
FOR BIOLOGICAL
MODELING
Dynamical Systems for Biological Modeling: An Introduction pro-
vides both biology and mathematics students with the understanding
and techniques necessary to undertake basic modeling of biological
systems. It achieves this through the development and analysis of
dynamical systems. AN INTRODUCTION
The approach emphasizes qualitative ideas rather than explicit com-
putations. Some technical details are necessary, but a qualitative ap-
proach emphasizing ideas is essential for understanding. The model-
ing approach helps students focus on essentials rather than extensive
mathematical details, which is helpful for students whose primary in-
terests are in sciences other than mathematics.
The book discusses a variety of biological modeling topics, including
population biology, epidemiology, immunology, intraspecies competi-
tion, harvesting, predator–prey systems, structured populations, and
more.
The authors also include examples of problems with solutions and
some exercises that follow the examples quite closely. In addition,
problems are included that go beyond the examples, both in math- Fred Brauer
Brauer • Kribs
ematical analysis and in the development of mathematical models for
biological problems, in order to encourage deeper understanding and
an eagerness to use mathematics in learning about biology. Christopher Kribs
C664X
w w w. c rc p r e s s . c o m
Published Titles
Green’s Functions with Applications, Second Edition Dean G. Duffy
Introduction to Financial Mathematics Kevin J. Hastings
Linear and Integer Optimization: Theory and Practice, Third Edition
Gerard Sierksma and Yori Zwols
Markov Processes James R. Kirkwood
Pocket Book of Integrals and Mathematical Formulas, 5th Edition
Ronald J. Tallarida
Stochastic Partial Differential Equations, Second Edition Pao-Liu Chow
Dynamical Systems for Biological Modeling: An Introduction
Fred Brauer and Christopher Kribs
Advances in Applied Mathematics
DYNAMICAL SYSTEMS
FOR BIOLOGICAL
MODELING
AN INTRODUCTION
Fred Brauer
University of British Columbia
Vancouver, British Columbia, Canada
Christopher Kribs
University of Texas at Arlington
Arlington, Texas, USA
Back cover image credit: (Photo of healthcare workers) Cleopatra Adedeji, CDC PHIL public domain
collection.
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2016 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business
This book contains information obtained from authentic and highly regarded sources. Reasonable
efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and
publishers have attempted to trace the copyright holders of all material reproduced in this publication
and apologize to copyright holders if permission to publish in this form has not been obtained. If any
copyright material has not been acknowledged please write and let us know so we may rectify in any
future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information stor-
age or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, please access www.copy-
right.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222
Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that pro-
vides licenses and registration for a variety of users. For organizations that have been granted a photo-
copy license by the CCC, a separate system of payment has been arranged.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
to our children and grandchildren
This page intentionally left blank
Contents
Preface xi
Acknowledgments xiii
I Elementary Topics 1
1 Introduction to Biological Modeling 3
1.1 The nature and purposes of biological modeling . . . . . . . 3
1.2 The modeling process . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Types of mathematical models . . . . . . . . . . . . . . . . . 11
1.4 Assumptions, simplifications, and compromises . . . . . . . . 14
1.5 Scale, and choosing units . . . . . . . . . . . . . . . . . . . . 17
vii
viii Contents
4.3.3 Contact rate saturation and the “Pay It Forward” model 237
4.4 Parameter changes, thresholds, and bifurcations . . . . . . . 242
4.4.1 Hysteresis . . . . . . . . . . . . . . . . . . . . . . . . . 249
4.4.2 The spruce budworm . . . . . . . . . . . . . . . . . . . 251
4.5 Numerical analysis of differential equations . . . . . . . . . . 258
4.5.1 Approximation error . . . . . . . . . . . . . . . . . . . 259
4.5.2 Euler’s method . . . . . . . . . . . . . . . . . . . . . . 260
4.5.3 Other numerical methods . . . . . . . . . . . . . . . . 264
4.5.4 Eutrophication . . . . . . . . . . . . . . . . . . . . . . 270
Miscellaneous exercises . . . . . . . . . . . . . . . . . . . . . . . . 277
Bibliography 459
Index 467
Preface
xi
xii Preface
We thank Bob Stern at CRC Press for planting the initial idea for this book,
and both him and Bob Ross for their support and understanding in the face
of long delays during the preparation of the manuscript. (Although we do not
include delay equations in this book, there certainly were delays.) We also
thank everyone at Taylor & Francis who helped with the production of this
book, especially Shashi Kumar and Marcus Fontaine for critically useful help
involving LaTeX, Karen Simon, and Kevin Craig for graphic design assistance.
The figures and photos accompanying discussion of the various biologi-
cal systems studied and discussed play a crucial role in bringing the models
(and their motivations) to life for the reader, and many appear in this work
through the kind permission of others. We therefore acknowledge here all those
who generously permitted us to print their photos, or helped us obtain permis-
sion: Donna Anstey, Francine Bérubé, Daniel Bowen, John Calambokidis, Tom
Chrzanowski, Patricia Ernst, Carla Flores, Tim Gerrodette, Stefano Guerrieri,
Ray Hamblett, Alan M. Hughes, Duncan Jackson, Russell S. Karow, Carolyn
Kribs, Joel Michaelsen, Bernard E. Picton, Dave Powell, Francis Ratnieks,
Helen Sarakinos, Howard Swatland, Michael Tildesley, and S. Bradleigh Vin-
son.
We also thank all those photographers who contributed indirectly, in-
cluding photos in the public domain: Cleopatra Adedeji, Lennert B., David
Burdick, Janice Haney Carr, Mark Conlin, Karen Couch, Jan Derk, Gary
Fellers, Ryan Hagerty, William Chapman Hewitson, Steve Hillebrand, Tom
Hodge, John & Karen Hollingsworth, Al Mare, Maureen Metcalfe, Benjamin
Mills, A.J. Nicholson, Sergey Nivens, Zev Ross, Jeff Schmaltz, Greg Webster,
Gary Zahm, and the following organizations: the Centers for Disease Con-
trol, Florida Keys National Marine Sanctuary, Ken Gray Image Collection
at Oregon State University, MODIS/NASA, National Diabetes Information
Clearinghouse, NOAA, the Pennsylvania Dept. of Conservation and Natural
Resources, River Alliance of Wisconsin, the Sickle Cell Foundation of Georgia,
the Southwest Fisheries Science Center of the NOAA Fisheries Service, U.S.
Fish and Wildlife Service, U.S. Geological Survey, U.S. National Park Service,
and the Wisconsin Dept. of Natural Resources.
xiii
This page intentionally left blank
Part I
Elementary Topics
1
This page intentionally left blank
Chapter 1
Introduction to Biological Modeling
3
4 Dynamical Systems for Biological Modeling: An Introduction
tion: mathematical modeling in the practice of biologists and mathematicians, Science and
Education 6(5): 441–472.
2 G.E.P. Box (1976). Science and statistics, Journal of the American Statistical Associ-
evaluation
mate solutions numerically, and are useful for illustration purposes, or when a
model is too complicated to complete a qualitative analysis, or when we have
good estimates of the parameters involved. In this text we shall show how to
apply all three types of analysis, although qualitative analysis will receive the
most emphasis, as a way to prove how complicated models behave, regard-
less of the exact parameter values. Although the mathematical nature of such
models can generate convincingly real data, it is often important to remember
that, as gross oversimplifications of reality, mathematical models are just as
much rough sketches as the nonmathematical theories with which biologists
are familiar, and qualitative descriptions of behavior befit this notion.
(4) Interpretation Interpreting mathematical results in biological
terms requires a return to the original context of the problem, and is easier
with a well-articulated research question. That is, rather than simply asking,
“What happens if we model the following biological system?” one ought to
ask a specific question that can be addressed with a model — for example,
“What effects do seasonal temperature variations have on the size of this pop-
ulation?” or “How does behavior change based on observed infection levels
affect the eradicability of this disease?” (A general statistical analysis of an
experimental data set, however, may show up trends that shape subsequent
research directions.) Interpretation is also facilitated by a close correspon-
dence between each element of the mathematical model and the elements of
the biological system. The conclusion “this population will go extinct if beta
x is less than 1” is only useful if one can put “beta x” in biological terms —
what do beta and x mean? can we measure them (or at least their product)?
Which might we be able to control, and which are inherent properties of the
biological system?
The basis for an interpretation may be numerical, qualitative, or graphical.
When most parameters (fixed quantities) in a model can be estimated fairly
well, numerical information can provide ranges within which some varying
or controllable quantity causes the biological system to behave in a partic-
ular way, or critical values which cause that behavior to change. Identifying
thresholds between survival and extinction, or between settling down to an
equilibrium level and oscillating forever, or comparing the behavior of two
related but slightly different models, can answer qualitative questions such
as, “Which explanation for this phenomenon better accounts for our obser-
vations?” or “Does the system behave differently if we take into account this
additional factor?” The ready availability of numerous computational tools
also makes it possible to represent results graphically, and a graph can be
an especially compelling, concise statement of a complex result. Graphs can
illustrate both general trends and striking thresholds.
However, just as with algebraic expressions, graphical depictions of math-
ematical information can be so concise that the job of unpacking them into
nonmathematical terms can be significant. The ability of mathematics in all its
forms to represent complicated ideas compactly makes it both powerful and
challenging, and if one considers all the time invested in becoming familiar
Introduction to Biological Modeling 9
Rate
births deaths
Population size
FIGURE 1.2: A graph comparing birth and death rates vs. population size,
after Roth (2001).
• Note first that the horizontal axis in the graph is not time but population
size. That is, the graph shows how birth and death rates vary depending
on the population, rather than over time. Since many continuous graphs
depict how some quantity changes over time, one must readjust one’s
reading of the graph to reflect what it is that is varying.
• Note second that the vertical axis in the graph gives birth and death rates
(measured in individuals per day, week, or year), and not total births
and deaths. That is, the graph shows how fast births and deaths occur,
as population size grows (varies). The bending downward of the birth
rate curve, for example, as population size increases does not address
what has happened to previous births in the past of any population that
reaches that size. The distinction (and relationship) between population
4 W.-M. Roth (1998), Unspecified things, signs, and “natural objects”: towards a phe-
models which keep track only of the overall sizes of otherwise homogeneous
groups may allow the variables representing them to vary continuously, espe-
cially if the unit (see Section 1.5) makes this meaningful: biomass measured
in kilograms, or individuals measured in thousands or millions (where each
individual is such a small quantity that the difference between discrete and
continuous is minimal). In general, however, the choice between discrete and
continuous state variables is linked to a more fundamental decision: whether
to make the model deterministic or stochastic.
A deterministic model describes a system in terms of averages: every possi-
ble outcome — dying or not dying, reproducing or not reproducing, recovering
from a disease at any moment from just after infection to years afterward —
occurs, in the proportions corresponding to the relative probabilities of each
event happening. If the average lifetime for a population under study is 30
years, then the corresponding average (per capita) death rate is its reciprocal,
1 1
30 /yr ≈ 0.03/yr, and in a deterministic model precisely 1/30th, or 3 3 %, of the
population will die each year. In some sense, it is more accurate to say that
in a deterministic model, 1/30th of each individual dies per year. This type
of model is appropriate for large populations, where we call on the so-called
Law of Large Numbers, which says that as the number of trials of a random
event (such as whether or not a given individual dies) increases, the actual
experimental probability (here the proportion of the population that under-
goes a certain change) approaches the theoretical probability of the event.
Deterministic models also commonly use continuous state variables.
On the other hand, a stochastic, or probabilistic, model describes all events
in terms of probabilities, and essentially flips a coin or tosses a die for each
individual member of the population, to see whether that member dies or
not, reproduces or not, etc. Consequently, these models use discrete state
variables. A stochastic approach generally requires a little more mathematical
machinery — in particular, familiarity with probability distributions — but is
appropriate in cases where populations are small enough that the Law of Large
Numbers does not apply, or where for some other reason it is important to
couch the model in terms of probabilities: for example, to determine a possible
probability distribution for the outcome of a certain event (like how many
cases an outbreak of disease will cause). Recently, studies have been made of
the structure of social contact networks,8 and in some cases detailed data has
been recorded for contact networks (e.g., SARS outbreaks, and sexual contact
networks in small closed populations); studying how these networks affect
the spread of contact-driven phenomena such as infectious diseases requires
a stochastic approach. Likewise, studies of disease eradication must focus on
the last few infected individuals, where probabilistic effects are important. For
the large-scale study of gonorrhea mentioned earlier, however, a deterministic
8 See, for instance, S.H. Strogatz (2001). Exploring complex networks, Nature 410: 268–
276, and M.E.J. Newman (2003). The structure and function of complex networks, SIAM
Review 45: 167–256.
14 Dynamical Systems for Biological Modeling: An Introduction
approach is not only simpler but has provided significant enough insights to
inform disease control strategies.
In addition, statistical analyses can be especially important for some mod-
els. Two particular calculations of relevance here are sensitivity analysis and
uncertainty analysis. In a sensitivity analysis, a probability distribution of pa-
rameter values is chosen around the parameters’ estimated averages, in order
to determine how strongly the model results depend on each parameter; the
outcome is a ranking of parameters to which some particular model output —
say, the minimum influx of new individuals per year needed in order to sustain
a population which reproduces slowly — is sensitive. For example, one might
find this hypothetical minimum influx to be more sensitive to the existing
population’s death rate than to its reproductive rate. An uncertainty analysis
studies the impact of measurement errors (how far off the estimated param-
eter values are); the outcome is a probability distribution for some model
output, say, the spread of possible values for the required minimum influx
given estimated probability distributions for the birth and death rates.
No single text can hope to make a reader conversant with all of these
types of mathematical models. In this book we shall focus on deterministic
models, with continuous state variables, stratified discretely when necessary
to incorporate heterogeneity, the alternative being continuous variation us-
ing partial differential equations. Chapter 2 will consider models of biological
systems appropriate for discrete dynamical systems (recall that by discrete
we mean discrete in time), while the remaining chapters will consider mod-
els using continuous dynamical systems. In each case, we will begin our study
with simple (linear) single equations, proceed to more complicated (nonlinear)
single equations, and finally consider systems of interacting populations.
the way radioactive decay works (as opposed to, say, a fixed period of time);
although some biological processes do work this way, the real reason we make
this assumption is to make our models be differential equations rather than
something more complicated, like delay or integral equations. Second, at the
same time our willingness to consider reasonably complex models means giving
up exact solutions to the equations in favor of qualitative information about
the solutions: do they eventually taper off toward some equilibrium, or do
they oscillate forever, or do they crash to zero in finite time, or do they do
something stranger? In the case of a quantitative model, we give up exact
solutions and broad, analytical guarantees of possible behaviors in exchange
for quantitative approximations given specific parameter values.9
Third, typically we also make some kind of assumption of homogeneity
of individuals — that, within a given group, they all behave the same way
with regard to any events described in the model. Recall also the discussion of
handling heterogeneity in the previous section; stratification typically includes
this assumption of homogeneity within each class, in exchange for a simpler
type of model. This assumption, as well as the first one above, arises from the
motivation to minimize the number of state variables used (the dimension of
the model).
Fourth, we sometimes give up immediate biological interpretability in re-
turn for tractability — that is, after initially defining a model in biological
terms, we may further simplify it by redefining variables and parameters, in a
way that makes the model mathematically simpler (usually reducing the num-
ber of different variables and/or parameters) but may complicate our ability
to interpret the new quantities in biological terms. For example, one model
for foraging ants10 studied in Section 4.1 begins with five parameters:
αN K(βN − α) βK 2
a= , b= , c= .
σ σ σ
9 Computational schemes used in numerical analysis, however, are typically shown to
provide approximations that converge to the exact solution as the time-steps involved get
smaller. That is, one can make the approximate solutions as close as one likes to the exact
solution, if one is willing to wait longer for the computer to run its calculations.
10 D.J.T. Sumpter and S.C. Pratt (2003). A modelling framework for understanding social
The resulting model is simpler to analyze, but the new parameters are more
difficult to interpret in biological terms because the biological information
has been packed mathematically, and must be unpacked following analysis in
order to be interpreted more easily. This process is called rescaling or non-
dimensionalization and is discussed further in Section 1.5 below.
In articulating the simplifying assumptions involved in forming a model,
one should consider the consequences of introducing these “errors.” For ex-
ample, what do we lose from the results of a model if we assume homogeneous
mixing of sediments in a lake, or of a human population? The answers here,
of course, may be very different: in the first case, we lose the ability to dis-
tinguish the effects of sediment build-up as a function of depth, or distance
from the shore, including the resulting distribution of creatures in the lake
that are affected by those sediments (small animals may hide from predators
in the sediment, small animals and plants may even eat it; others, like the
predators, may be hindered by it). In the second case we lose our ability to
determine to what extent some individuals, who come into contact with more
people than most, such as a supermarket checker or a delivery person, may
be instrumental in the transmission of some contact-based process, such as an
infectious disease or the spreading of news and rumors by word of mouth.
There is, finally, another type of simplification which is sometimes made
without any intention of biological justification. In order to make inroads on
a tough problem, we may deliberately oversimplify our model — for instance,
by setting one or more quantities to zero — when the simpler model is easier
to analyze, with the purpose of bringing those results back to the original
problem to inform our intuition and expectations of how the full analysis will
turn out. For example, if there are two processes taking place and one proceeds
much more slowly than the other, we may assume temporarily that the “slow”
variables remain constant. We mention this practice (discussed in Chapter 6)
simply because it can be a useful tool for analyzing models at the edge of
our ability to handle, and because it might not occur to modelers focused
on justifying biologically all mathematical simplifications (the difference here
being that they are only temporary simplifications).
In any case, since models are imperfect descriptions of reality, unless our
main focus really is close fitting of data, our ultimate goal for conclusions
drawn from modeling should be a notion called robustness. This means that the
biological insights obtained from a mathematical model should be independent
of the particular mathematical structure and functions used, to the extent that
we are uncertain what the exact structure or function should be. For example,
as a population grows to fill its habitat, it runs up against the limitations
on the habitat’s resources, which begins to limit the population’s ability to
reproduce. This is true on scales anywhere from a Petri dish to the planet
Earth. This limitation on the birth rate can be described mathematically by
any function which tapers off eventually toward zero after an initial increase
(see, e.g., Figure 1.2). The most common function used for this purpose is
the quadratic that leads to the logistic equation, analysis of which shows that
Introduction to Biological Modeling 17
the population will eventually level off at a size determined by the resources.
However, other functions of similar shape (that described above) and similar
properties can be used instead, with the same results. Likewise, we might
hope that stratifying a population into three or four age or risk categories will
provide results qualitatively similar to those of a model which stratifies the
population into ten or twenty classes. (We certainly would not like to analyze
the latter model!) Another example is that of threshold quantities, mentioned
above and discussed in Section 4.4. In this way, we justify to a limited extent
the imperfection of our model.
Sometimes models make predictions that are surprising or even alarming.
Such predictions from simple models imply a need for further careful study
of the phenomenon, in order to see if more detailed models make similar
predictions. Predictions which are robust across multiple models are more
likely to be accurate (but no more accurate than the models) and should be
taken seriously.
than with others outside the collective. Recent research in mathematical bi-
ology has identified numerous situations in which collectives are important:
for example, individual households in studying infectious diseases. Within a
household, infections tend to spread faster among household members than
between a household member and an individual outside the household. This
is true of human respiratory diseases, as many families know, but also of pests
such as mice, which carry hantaviruses, and insects, which act as vectors for
diseases such as malaria (mosquitoes) and Chagas’ disease (triatomines or
“kissing bugs”). At the population level, it then also becomes reasonable to
make households the unit of measurement, and count infected houses rather
than infected individuals, with recovery coming only when no individuals in
a given household remain infected, or when treatment has eliminated all the
pests in the house.
Another issue related to scale and units is the rescaling procedure men-
tioned in the previous section. Sometimes we find it possible to reduce the
mathematical complexity of a model by reducing the number of variables
and/or parameters. Although the rescaled model can be more difficult to inter-
pret immediately, the resulting simplifications in our calculations are typically
worth it, especially when we can always undo the transformation afterward to
facilitate interpretation. Probably the most common type of rescaling is to re-
define variables relative to benchmark parameters. For example, in the model
of foraging ants mentioned in the previous section (and explored more fully
in Section 4.1), Sumpter and Pratt also rescaled both the state variable E(t),
representing the number of ants exploiting the food site, and the independent
variable t for time, as follows:
E t
x= , τ= ,
K K/σ
so that both x and τ are dimensionless. That is, x gives the number of ex-
ploring ants as a multiple of the benchmark value K, and τ gives the time
as a multiple of the benchmark value K/σ (which is how much time it takes,
on average, for K ants to retire). This is why the process is also called non-
dimensionalization. (The transformation to x in the article also involves elimi-
nation of another state variable for the number of ants not exploiting the food
source by assuming the total ant population constant, another simplifying
technique mentioned in the previous section.)
Introduction to Biological Modeling 19
Exercises
Consider the following biological systems from a modeling perspective. For
each, determine which type of mathematical model and analysis you believe
would be most appropriate with regard to the criteria developed in this chap-
ter:
• discrete or continuous independent variable(s) (time, age, space),
• discrete or continuous dependent variable (population or quantity),
• type of stratification or continuous distribution of traits, if any,
• deterministic or stochastic,
• quantitative (numerical) or qualitative analysis,
• scale and units.
Defend or justify each of your choices with a sentence. (This exercise is most
useful if discussed in a group setting after each person has written down
his/her choices.)
1. the number of salmon that return to a particular pool to spawn each
year
2. a population of pea aphids whose genetic resistance to infection by fungal
spores varies inversely with their resistance to attack by parasitic wasps
(studied by students in the paper by Smith, Haarer and Confrey)
3. the frequency of a particular allele within each generation of a population
4. the likelihood of extinction of a rare allele (for reasons other than ge-
netic) over many generations of a population
5. the geographic dispersal (movement) of house sparrows within the
United States, starting with the original eight pairs released in the spring
of 1851 in Brooklyn, New York
6. the spread of tall grasses around the shore of a pond from an isolated
cluster
7. seasonal competition for space among different species of trees in a rain-
forest
8. weekly harvesting (removal) of fish in a fish hatchery
9. the growth of a yeast culture in a Petri dish
10. the mixing and growth of bacteria and nutrients in a chemostat
11. the growth of a deer population subject to an annual hunting season
20 Dynamical Systems for Biological Modeling: An Introduction
26. In what important ways does Figure 1.4 below differ from Figure 1.2?
Consider what will happen to the size of the population if it begins in
each of the three regions marked (a), (b) and (c). (This graph exhibits
something called an Allee effect, which is explored in Section 2.2.)
Introduction to Biological Modeling 21
Rate
deaths
births
Population size
(a) (b) (c) Graph courtesy M. Tildesley
FIGURE 1.4: Graph for Exer- FIGURE 1.5: Graph for Exer-
cise 26. cise 27.
27. Figure 1.5 shows the distribution of farms in the U.K. during the 2001
outbreak of foot-and-mouth disease (FMD), in terms of the numbers of
cows and sheep on each farm. Color intensity indicates the number of
farms with the given numbers of animals.
(a) Where, on this graph, are the farms with large numbers of cows
but no sheep? Where are the farms with many sheep but no cows?
(b) What does the graph indicate about the distinct types of farms
that exist?
(c) Since there is no treatment for FMD, the epidemic was brought
under control by culling — slaughtering all susceptible animals on
farms within a certain radius of any animal found to be infected.
This wiped out numerous small farms altogether. How might the
small farmers use this graph to argue for a change in control policy?
28. Smith, Haarer and Confrey (1997) studied a group of graduate students
working together in a course on mathematical biology (see footnote 1).
They found that the goals of students from different fields, as well as the
course instructors, who had different backgrounds, differed at times. If
we consider three perspectives from which students using mathematical
models to represent biological systems might come, we might say that:
Those who came with the perspective of an experimental biologist wanted
models to incorporate all their data, and to fit and explain them.
Those who came with the perspective of a theoretical biologist wanted
to use the data obtained beforehand to suggest trends to explore in the
model, to use the model as a means of exploring the effects of particular
phenomena (like genetic variation) across particular systems, i.e., for
many different species and habitats. They saw the model as a virtual
22 Dynamical Systems for Biological Modeling: An Introduction
23
24 Dynamical Systems for Biological Modeling: An Introduction
Photo courtesy U.S. Fish and Wildlife Service, dls.fws.gov Photo courtesy Joel Michaelsen
FIGURE 2.1: The semelparous salmon (left) and the monocarpic century
plant (right) reproduce in distinct generations. Note the different stages of the
blooming spikes on the three century plants.
These equations are called difference equations, since they give the difference in
size between one generation and the next, and first-order difference equations1
have the form yk+1 = f (yk ), which can be read as saying that the size of the
next generation (generation number k + 1 of population y) is a function f of
the present generation (generation number k of population y). In the rest of
this section, we will discuss the simplest type of difference equation and what
its solutions look like. Following that, the next sections in this chapter develop
some tools, both analytical and computational, for studying more complicated
difference equations. Later sections will look at classes of biological systems
that can be studied using discrete dynamical systems, including population
genetics, competition, harvesting, and finally systems involving more than one
quantity or population.
where the symbol ≈ signifying approximate equality means that the error in
this approximation is small for small h, in the sense that this error divided by
1 First-order difference equations will be our primary focus in this chapter.
Difference Equations (Discrete Dynamical Systems) 25
Example 1.
Solve the difference equation yk+1 = 2yk , with y0 = 1.
Solution: From y1 = 2y0 (the difference equation with k = 0) we see that
y1 = 2y0 = 2. Then y2 = 2y1 = 4, y3 = 2y2 = 8, . . . We may then guess (and
prove by induction) that yk = 2k . This is the solution of the given problem.
Graphing will show the familiar exponential curve.
Example 2.
Verify that yk = 5(1.1)k is the solution of the difference equation yk+1 =
(1.1)yk , y0 = 5.
Solution: First verify the initial condition: The expression yk = 5(1.1)k with
k = 0 gives y0 = 5(1.1)0 = 5, as desired. Now substitute
the proposed solution
into the difference equation: yk+1 = (1.1) 5(1.1)k = 5(1.1)k+1 . Thus the
given yk satisfies the given difference equation and initial condition.
Equation (2.5) fits the general form yk+1 = f (yk ) of a first-order difference
equation; it also belongs to a more specific class of equations of the form
called linear difference equations because the right-hand side f (yk ) is a linear
function of yk . The special case (2.5) where b = 0 is known as the homogeneous
case. These simplest types of difference equation have special names for two
reasons: first, as we shall see below, they are relatively easy to solve outright
(in fact, they are the only kind of difference equation we shall attempt to solve
explicitly); and, second, as we shall see in a later section, they turn out to be
key to understanding the behavior of more complicated difference equations.
y1 = r y0 , y2 = r y1 = r 2 y0 , y3 = r y2 = r 3 y0 , . . .
yk = r k y0 , (2.7)
Example 3.
Find the solution of the difference equation yk+1 = −yk , y0 = 1.
Solution: From the formula (2.7) with r = −1 we have yk = (−1)k ; notice
that yk oscillates between positive and negative values.
Difference Equations (Discrete Dynamical Systems) 27
Example 4.
Find the solution of the non-homogeneous linear difference equation
yk+1 = −yk + 1, y0 = 1.
and then conjecture that yk alternates between 0 and 1. To verify the correct-
ness of this conjecture, we need only note that if yk = 1, then yk+1 = 0 and
if yk = 0, then yk+1 = 1. Thus yk does alternate between 0 and 1, which we
may express explicitly as yk = 12 [1 + (−1)k ].
y1 =r y0 + b
y2 =r y1 + b = r(r y0 + b) + b = r2 y0 + r b + b
y3 =r y2 + b = r(r2 y0 + r b + b) = r3 y0 + r2 b + r b + b, etc.
yk = rk y0 + b 1 + r + r2 + . . . + rk−1 .
(2.8)
1 − rk
1 + r + r2 + . . . + rk−1 =
1−r
for the sum of a geometric series (with r 6= 1), we may write this solution in
the form
1 − rk
b b
yk = r k y0 + b = y0 − rk + . (2.9)
1−r 1−r 1−r
We may also consider what happens to solutions of linear difference equa-
tions over long periods of time – in mathematical terms, as k → ∞. Consider-
ing the solution (2.9), we see that if r is large in size, i.e., r > 1 or r < −1, then
rk grows unbounded as k → ∞, and thus yk grows unbounded too. The only
b b
exception occurs when y0 = 1−r , in which case yk is a constant 1−r for all
b
k (cf. (2.9)) and thus approaches 1−r as k → ∞. Small differences, however,
from this equilibrium solution will be magnified for large r (cf. again (2.9)).
If r is instead small in size (between −1 and 1), then rk approaches zero
b
as k → ∞, and in view of (2.9) the solution yk approaches the limit 1−r as
k → ∞ regardless of the initial value y0 .
If r = −1, rk alternates between −1 and 1, and yk does not have a limit
28 Dynamical Systems for Biological Modeling: An Introduction
b
(unless y0 = 1−r ). If r = 1, the formula (2.9) is meaningless, but the formula
(2.8) becomes yk = y0 + rk; thus yk becomes unbounded as k → ∞.
The way the solution of the linear homogeneous difference equation (2.5)
depends on the value of the survival-and-growth term r will be important
when we study qualitative behavior of solutions of difference equations in
Section 2.3. From the formula (2.7) we see that yk → 0 as k → ∞ if |r| < 1
and yk grows unbounded as k → ∞ if |r| > 1. More precisely, if 0 ≤ r < 1,
yk decreases monotonically to zero and if −1 < r < 0, yk oscillates between
positive and negative values in approaching zero. If r > 1, yk increases to +∞
and if r < −1, yk oscillates unboundedly. The “boundary” cases are r = −1,
in which yk oscillates between ±y0 and does not approach a limit, and r = 1,
in which yk is the constant y0 . The essential property we shall need is the
following result.
Example 5.
Find for which values of a every solution of the difference equation yk+1 =
(1 + a)yk approaches zero as k → ∞.
Solution: Every solution approaches zero if and only if |1 + a| < 1, or −1 <
1 + a < 1, or −2 < a < 0.
ryk2
yk+1 = (2.11)
yk2 + A2
have been suggested as descriptions for populations whose growth rates satu-
rate for large population sizes. Here A is the population size at 50% saturation
2 P. F. Verhulst, Récherches mathématiques sur la loi d’accroissement de la population,
(plug in A for yk in (2.10) to see why). For small yk (relative to A), the denom-
inator of the right-hand side of (2.10) and (2.11) is essentially A, so that (2.10)
has yk ≈ Ar yk , similar to the unrestricted growth of the linear equation (2.5).
However, for large yk , A is relatively insignificant, making the right-hand sides
of (2.10) and (2.11) (and thus yk+1 ) approximately equal to r.
Another much-studied example is the logistic difference equation
yk
yk+1 = ryk 1 − , (2.12)
K
also introduced by Verhulst, with a growth rate which decreases to zero as yk
approaches the carrying capacity K and which becomes negative for yk > K.
The logistic difference equation should not be taken seriously as a model for
large population sizes as yk+1 is negative if yk > K. Other difference equations
which have been used as models to fit field data are
and (
ryk1−β , for yk > ǫ,
yk+1 = (2.14)
ryk , for yk < ǫ.
None of the difference equations (2.10), (2.11), (2.12), (2.13), (2.14) are
derived from actual population growth laws. Rather, they are attempts to
give quantitative expression to rough qualitative ideas about the biological
laws governing population growth. For this reason, we should be skeptical
of the biological significance of any deduction from a specific model which
depends on the precise formula for the solution of that model. Our goal should
be to formulate principles which are robust, that is, which are valid for a large
class of models embodying some set of qualitative hypotheses. Therefore, we
shall be more concerned with qualitative properties of solutions of difference
equations than with formulae for solutions. In fact, although we have seen
some examples of linear difference equations for which a solution formula is
available, the solution of nonlinear difference equations like (2.10), (2.11),
(2.12), (2.13), and (2.14) usually cannot be found in general as functions of
the equation parameters. In Section 2.3 we will develop tools for analyzing
the behavior of these more complicated equations.
The most that we can do quantitatively with many nonlinear difference
equations is to calculate solutions numerically by iteration for particular
choices of parameter values, as illustrated by the following two examples.
Example 6
Find the first four terms of the solution of the difference equation yk+1 =
yk (1 − yk ) with y0 = 12 .
Solution: We have y1 = 12 (1 − 12 ) = 14 = 0.25, y2 = 14 (1 − 41 ) = 16
3
= 0.1875,
3 3 39 39 39
y3 = 16 (1 − 16 ) = 256 = 0.152344, y4 = 256 (1 − 256 ) = 0.129135.
30 Dynamical Systems for Biological Modeling: An Introduction
0.5 1
0.4 0.8
0.3 0.6
0.2 0.4
0.1 0.2
0 1 2 3 4 0 5 10 15 20 25
Example 7.
1
Verify that the constant sequence yk = 2 (k = 0, 1, 2, . . .) is a solution of the
difference equation yk+1 = 2yk (1 − yk ).
Solution: If yk = 12 , then 2yk (1 − yk ) = 2( 12 )(1 − 21 ) = 12 . Thus yk = 1
2 satisfies
yk+1 = 2yk (1 − yk ).
Exercises
1. The bacteria E. coli doubles in a little over 20 minutes in good lab
conditions. (For purposes of this exercise, we will assume it takes exactly
20 minutes.)
(a) Write a difference equation that models the growth of an E. coli
Difference Equations (Discrete Dynamical Systems) 31
wn+1 = F wn + µ,
possible reasons and implications, Journal of Theoretical Biology 75: 349–371, 1978.
32 Dynamical Systems for Biological Modeling: An Introduction
yk
5. Find the first three terms of the solution to yk+1 = yk +1 , y0 = 1.
2
yk
6. Find the first three terms of the solution to yk+1 = 2 +1 ,
yk
y0 = 1.
yk+1 = r yk (2.15)
which we have already solved analytically in Section 2.1. The method is also
applicable to difference equations which cannot be solved analytically. It may
be carried out on a computer with the aid of a computer algebra system such
as Maple, Mathematica, or MATLAB. Some of the figures in this section were
produced using Maple, with a program given in Appendix A.
This method begins by drawing in the y − z plane the reproduction curve,
which is the graph of the function on the right-hand side of our difference
equation, and the line z = y, which we shall use for reflection purposes. For
equation (2.15), the reproduction curve is the line z = r y (taken from r yk ).
Next, we mark y0 on the y-axis, and go vertically to the reproduction curve
(meeting it at height ry0 ). The next step is to go horizontally to the line
z = y, meeting it at the point (y1 , y1 ). (Recall y1 = ry0 from (2.15).) Then
we repeat the process, going vertically to the reproduction curve (at height
ry1 ), then horizontally to the line z = y at the point (y2 , y2 ), where y2 = ry1 .
Continuing in the same manner, we reach successively the values y3 , y4 , y5 ,
. . .. The graphic portrayal will have four different cases — r > 1 (Figure 2.4),
0 ≤ r < 1 (Figure 2.5), −1 < r < 0 (Figure 2.6), r < −1 (Figure 2.7) —
corresponding to different relative positions of the reproduction curve and the
line z = y. In each case, the graphical solution illustrates the behavior already
z=ry
y =r y
1 0
y =r y
2 1
y =r y
2 1
y =r y
1 0
y y
y y y y
0 1 1 0
z z=y z z=y
y0 y0
y y
z=ry
z=ry
Example 1.
Apply the cobwebbing method to describe the solutions of the Verhulst equa-
tion
r yk
yk+1 = .
yk + A
ry
Solution: The reproduction curve is z = y+A and its slope is
dz rA
= .
dy (y + A)2
At y = 0 the slope is r/A. If r < A, this slope is less than 1, so the reproduction
curve lies below the line z = y, while if r > A this slope is greater than 1, so the
reproduction curve begins above the line z = y but comes down to intersect
it at y = r − A. If r > A, every solution, regardless of the initial value y0 ,
approaches the limit y∞ = r − A (Figure 2.8), and if r < A, every solution
approaches zero (Figure 2.9). In either case, the limit is an intersection of the
reproduction curve and the line z = y.
Difference Equations (Discrete Dynamical Systems) 35
z z
z=y z=y
ry
z=
y+A
ry
z=
y+A
y
y <r−A r−A y0>r−A y y
0 0
Example 2.
Apply the cobwebbing method to describe the solutions of the difference equa-
tion
r y2
yk+1 = 2 k .
yk + A
In Section 2.5 we shall see how we can identify limits of solutions (defined
in the next section as equilibria) graphically for different parameter values, by
this same superposition of two curves.
36 Dynamical Systems for Biological Modeling: An Introduction
z
z z
z=y
z=y z=y
RC
RC
RC
y y y y
y y
0 0 0
FIGURE
√ 2.10: r < FIGURE
√ 2.11: r > FIGURE
√ 2.12: r >
2 A in Example 2. 2 A, y0 < y− in Exam- 2 A, y0 > y− in Exam-
ple 2. ple 2.
Exercises
In Exercises 1-6, apply the cobwebbing method to describe the solutions
of the given difference equation.
yk+1 = 8yk3 (1 − yk ),
which provides another way to model growth limitations for large pop-
ulations (here large means closer to 1 than to 0, so consider the units
of y to be the maximum population size at which reproduction can
occur). Draw cobweb diagrams for three different initial conditions:
Difference Equations (Discrete Dynamical Systems) 37
FIGURE 2.13: Three views of the function y = x3 − 10x + 10 sin 10x, each
centered at the origin. As we restrict our view to smaller x-intervals around
0, the graph appears increasingly flat.
if there are any, are particularly easy to find. They are simply the solutions
of the equation g(y) = y. We define an equilibrium of the difference equation
(2.17) to be a solution y∞ of the equation g(y) = y. This is also referred to
as a fixed point of the map g. Geometrically, an equilibrium is an intersection
of the reproduction curve z = g(y) and the line z = y. In the nonlinear
difference equations used as examples in the previous section, we have observed
a tendency for solutions to approach a limit as k → ∞, and these limits are
equilibria. Go back and look at the cobwebbing diagrams in the previous
section to see this (note that in most cases, cobwebbing leads us to a point
where the reproduction curve meets the line z = y). It will turn out that
we can obtain a great deal of information about the solutions of a difference
equation by studying the nature of its equilibria.
An essential fact in studying the behavior of solutions of a difference equa-
tion near an equilibrium is that a difference equation may be approximated
by a linear difference equation near an equilibrium. If y∞ is an equilibrium
of the difference equation (2.17), so that g(y∞ ) = y∞ , we make the change of
variable
u k = yk − y∞
so that uk represents the deviation from the equilibrium. Substitution into
(2.17) gives
y∞ + uk+1 = g(y∞ + uk ).
By Taylor’s Theorem (see Appendix B), we may write
g ′′ (c) 2
g(y∞ + uk ) = g(y∞ ) + g ′ (y∞ )uk + u
2! k
for some value c between y∞ and y∞ + uk . Since g(y∞ ) = y∞ , the difference
equation (2.17) is equivalent to the difference equation
g ′′ (c) 2
uk+1 = g ′ (y∞ )uk + u (2.18)
2! k
Difference Equations (Discrete Dynamical Systems) 39
′′
for u. The term g 2!(c) u2k is not really a quadratic function of uk because the
intermediate value c may depend on uk , but it is small compared to uk in the
sense that this term divided by uk approaches zero as uk → 0. The lineariza-
tion of (2.17) at the equilibrium y∞ is defined to be the homogeneous linear
equation
uk+1 = g ′ (y∞ )uk (2.19)
obtained from (2.18) by neglecting this “higher order” term. Note that the
constant term drops out only because we are linearizing about an equilibrium,
where g(y∞ ) = y∞ .
Example 1.
Find the linearization of the logistic difference equation
yk
yk+1 = ryk 1 −
K
at each of its equilibria.
Solution: The equilibria are the solutions of the equation
y
y = ry 1 − ,
K
namely y = 0 and y = K(1 − 1r ) (the latter positive for r > 1). Here the
function g(y) is given by y
g(y) = ry 1 − ,
K
and
2ry
g ′ (y) = r − .
K
Thus g ′ (0)
= r, and the linearization at the equilibrium y = 0 is uk+1 = r uk .
Also, g ′ K(1 − 1r ) = r − 2r(1 − r1 ) = 2 − r, and so the linearization at
y = K(1 − 1r ) is uk+1 = (2 − r)uk .
The importance of the linearization lies in the fact that the behavior of
solutions to the linearization of a difference equation at an equilibrium de-
scribes the behavior of solutions of the original difference equation near the
equilibrium. In particular, what we want to know is, if a solution (the popula-
tion size) is close to a particular equilibrium, will it approach the equilibrium,
or will it go away from the equilibrium? This attractive-repulsive quality is
called stability, and because it is based on linearizations, which are only good
approximations of the original when we are very close to an equilibrium, the
stability information we get from them is called local stability, as it does not
(yet) tell us what happens when we start far away from an equilibrium.
An equilibrium of the difference equation (2.17) is said to be asymptot-
ically stable if every solution whose initial value is sufficiently close to the
equilibrium remains close to the equilibrium and approaches it as k → ∞.
An equilibrium is said to be unstable if there are solutions with initial values
40 Dynamical Systems for Biological Modeling: An Introduction
arbitrarily close to the equilibrium which fail to remain near the equilibrium.
For an example, consider the generic homogeneous linear difference equation
(2.15), yk+1 = r yk , which has the unique equilibrium y = 0. As seen in the
previous section, when −1 < r < 1 (the second and third cases, Figures 2.5
and 2.6), solutions approach the equilibrium at 0; in these cases, the equilib-
rium is asymptotically stable. However, when r > 1 or r < −1 (the first and
fourth cases, Figures 2.4 and 2.7), solutions grow unbounded away from 0; in
these cases, the equilibrium is unstable. This example actually provides the
basis for the equilibrium stability theorem below, as all linearizations have
this form. The two examples in the previous section can also be analyzed in
this way (see, for instance, Example 3 below).
In applications to biological or other sciences, a difference equation model
is only an approximation to reality. There are inevitably errors or approxima-
tions in the model, and there are experimental errors in the measurement of
data to determine the parameters of the model. If an equilibrium is asymp-
totically stable, these errors are unimportant in the long-time behavior of the
system, because a small movement away from the equilibrium will be wiped
out over time. On the other hand, an unstable equilibrium is not experimen-
tally observable because any small error will cause the solution to move away
from it. Thus asymptotic stability of an equilibrium is an essential property for
applications. (Sometimes errors or changes in parameter values can actually
affect the stability of an equilibrium; these changes in stability, which occur
for continuous-time systems also, are called bifurcations and are discussed in
Section 4.4.)
It is now possible to establish the following important result, which we shall
use often (we state it here without proof). It is valid not only for the first-
order difference equations we have been studying here but also for systems of
difference equations, and indeed for many other kinds of dynamical systems,
including differential equations, which we shall meet in Chapter 3, and systems
of differential equations, which we shall meet in Chapter 5.
Example 2.
Show that if 1 < r < 3, the equilibrium y = 0 of the logistic difference equation
yk
yk+1 = ryk 1 −
K
is unstable and the equilibrium y = K(1 − 1r ) is asymptotically stable. Show
also that if r > 3 there is no asymptotically stable equilibrium.
Difference Equations (Discrete Dynamical Systems) 41
We have used the linearization theorem and our knowledge of how linear
difference equations behave to develop the stability theorem, because this
approach extends naturally to systems of difference equations. For the first-
order difference equations we are studying here, however, there is also a simple
direct approach given in Appendix D which uses the Mean Value Theorem.
It is also possible to refine this theorem and show that the approach to
an asymptotically stable equilibrium y∞ is monotone if 0 < g ′ (y∞ ) < 1 and
oscillatory if −1 < g ′ (y∞ ) < 0. The truth of this is suggested by the cobweb-
bing method (cf. Figures 2.5 and 2.6). For instance, we saw in Example 2 that
the equilibrium y∞ = K(1 − 1r ) of the logistic difference equation is asymp-
totically stable if 1 < r < 3. As g ′ (y∞ ) = 2 − r, it follows that solutions of
the logistic difference equation approach y∞ monotonically if 1 < r < 2 and
with oscillations if 2 < r < 3. It is natural to conjecture that if r > 3, so
that there is no asymptotically stable equilibrium, all solutions oscillate but
without approaching an equilibrium.
This is a different sort of question than asking what happens near a partic-
ular equilibrium; it changes the scope of our study from local to global. Once
the local stability of each equilibrium of a model has been determined, one
can put the results together to form a picture of the overall model behavior.
42 Dynamical Systems for Biological Modeling: An Introduction
In some cases, there is only one stable equilibrium, and within the state space
(the set of possible values for our variable(s)) that equilibrium may then be
not only locally stable but globally stable, that is, all solutions approach it no
matter what the initial condition. For example, when 0 < r < 1, the logistic
difference equation has only one equilibrium, at 0, and it is locally stable. We
can also see directly by inspection that yk+1 < yk when r < 1, and in fact in
this case the equilibrium y = 0 is globally stable. In biological terms, the pop-
ulation’s maximum reproductive ratio r is so low that the population cannot
sustain itself. Then, when 1 < r < 3, the equilibrium y = 0 is unstable, but
the equilibrium y = K 1 − 1r exists and is locally stable. In this case, it can
be shown that the latter equilibrium is globally stable.
There are other, more complicated possible scenarios for a model’s global
behavior. One possibility, which we shall explore further at the end of this
section, is multiple locally stable equilibria. Another, which corresponds to
the logistic difference equation with r > 3, is no stable equilibria at all. We
shall return to this question in Section 2.6. For now, however, let us consider
one more abstract example of local stability analysis before applying this
technique to a biological problem.
Example 3.
Determine the asymptotic stability of each equilibrium of the Verhulst differ-
ence equation
r yk
yk+1 = .
yk + A
ry rA
Solution: We have g(y) = y+A , g ′ (y) = (y+A) 2 . For the equilibrium y = 0,
bridge, 1968.
Difference Equations (Discrete Dynamical Systems) 43
precisely. There are several factors which might influence the distribution of
alleles from one generation to the next, including: the number of alleles for
the gene under study; the effect of genotype (which alleles a single individual
has) on viability, the proportion6 of individuals who survive to reproduce; the
effect of genotype on mating preferences; the effect of genotype on fertility, the
number of offspring an average individual produces; the (potentially genotype-
dependent) mutation rate from one allele to another; and possible gender
asymmetries. Let us examine here the role of genotype-dependent viability
in shaping the distribution of alleles over many generations. To simplify our
discussion, we will then assume in what follows that
(1) the gene under study has two distinct alleles, A and a.
There are therefore three possible genotypes a given individual might have:
homozygous AA, heterozygous Aa, and homozygous aa. We further assume
that
(2) genotype affects viability (some genotypes make individuals
more likely to survive and reproduce than others), in the same
manner for each generation.
In order to examine only the effects of this particular factor, we will also
assume that the other potential influences do not play a role here; that is,
that
(3) fertility is genotype-independent;
(4) mating preferences are genotype-independent;
(5) the mutation rate for all alleles is zero;
(6) all effects are gender-independent.
Note also that we are implicitly assuming distinct generations, with repro-
duction occurring only within a given generation (i.e., no mating between
generations); this is appropriate only for some populations, although the gen-
eral conclusions we draw can be applied more broadly.
Following assumption (2) we will incorporate differential viability by defin-
ing three parameters, φAA , φAa , φaa , each of which measures the proportion
of individuals of the given genotype who survive to reproduce. φAA , φAa , φaa
will take on values between 0 and 1 and will be assumed not to vary from
one generation to the next. In order to measure the frequency of each allele,
we define our variable pn to be the proportion of alleles in the nth generation
which are A, and qn = 1−pn to be the complementary proportion, the relative
frequency of a alleles.
Now, in order to calculate the relative frequencies pn+1 and qn+1 in the
next generation, we observe that each new individual receives one allele from
6 Although we might more intuitively speak of the likelihood or probability of surviving
each parent. From assumption (4) alleles are mixed at random, that is, in
proportion to the relative frequencies of the gametes (eggs and sperm) pro-
duced (which are, in turn, the same as the relative frequencies of the alleles,
by assumption (3)). Therefore a proportion pn pn = p2n of the new individuals
receive A alleles from both parents, a proportion pn qn receive an A allele from
the mother and an a allele from the father, a proportion qn pn receive an a
allele from the mother and an A allele from the mother, and a proportion
qn qn = qn2 receive a alleles from both parents. (The reader can verify that
these four terms sum to (pn + qn )2 = 1, thereby accounting for the entire
next generation.) Note that the relative frequency of heterozygotes Aa in the
next generation is 2pn qn , since the result (the zygote) is the same regardless
of which allele came from which parent.
The genotype-dependent viability of assumption (2) means that of those
AA-homozygotes born into generation n + 1, only a proportion φAA of them
survive to reproduce, that is, a proportion φAA p2n of the entire generation
of newborns. Likewise, a proportion φAa 2pn qn become mature heterozygotes,
and a proportion φaa qn2 become mature aa-homozygotes. Since each AA indi-
vidual has two A alleles, and each Aa individual has one, the total frequency
of A alleles is now
fA = 2 φAA p2n + (φAa 2pn qn ) .
G
6
p = 0 stable
p = 0 stable p∞ unstable
p = 1 unstable p = 1 stable
p → 0, a wins
IC selects winner
1
p = 0 unstable p = 0 unstable
p∞ stable p = 1 stable
p = 1 unstable p → 1, A wins
p → p∞ , coexist
-F
0 1
We can now put these results together to form a complete picture of the
model’s behavior. There are four cases or regions in terms of F and G, as
depicted in Figure 2.14. When F < 1 and G > 1 (which we can also write as
φAA < φAa < φaa , so that AAs are less viable than Aas, which are in turn
less viable than aas), the equilibrium at 0 is unstable, and the equilibrium at
1 is (globally) stable, so that pn → 0 as n → ∞, that is, the proportion of
A alleles dwindles to zero from one generation to the next, until eventually
almost everyone is aa homozygous. Similarly, when F > 1 and G < 1, so
that φaa < φAa < φAA and the A allele contributes more to viability than
the a allele, then the equilibrium at 0 is unstable, the equilibrium at 1 is
(globally) stable, and pn → 1 as n → ∞. In this case, eventually nearly
everyone is AA homozygous. Note that in both of these two cases, one allele
clearly contributes more to viability than the other, so it is not surprising
that it comes to dominance. (Also note that in both cases, the coexistence
equilibrium p∞ is not present.)
When F < 1 and G < 1, both AA and aa genotypes are less viable than
the heterozygous Aa genotype. This is the case, for example, in Africa with
regard to the allele for sickle cell anemia. This allele provides some immunity
against malaria, but homozygotes with two copies of the allele develop sickle
cell anemia and often do not survive to maturity. Thus it is beneficial to have
precisely one copy of the allele. In this scenario, the equilibria at 0 and 1
are both unstable, but the coexistence equilibrium is (globally) stable. Here
Difference Equations (Discrete Dynamical Systems) 47
Exercises
In Exercises 1–6, find all positive equilibria of the given difference equation,
and for each equilibrium determine whether it is asymptotically stable or
unstable. If the difference equation involves one or more parameters, your
answer may depend on the range of values of the parameters.
48 Dynamical Systems for Biological Modeling: An Introduction
r y2
7. Find all positive equilibria of the difference equation yk+1 = y2 +Ak
from
k
Example 2 of Section 2.2, and for each equilibrium determine when (in
terms of r and A) it is locally asymptotically stable or unstable. Compare
your results to those given for the cobwebbing approach.
8. Solve the equilibrium condition (2.22) and show that the coexistence
G−1
equilibrium p∞ = F +G−2 can only exist when F and G are either both
greater than 1 or both less than 1.
9. Suppose that, in the genetics discussion in this section, survival is
genotype-independent, i.e., φAA = φAa = φaa . Show that the fre-
quency of each allele is then constant across generations, establishing
the Hardy-Weinberg law of genetics (so named for its independent dis-
coveries by G.H. Hardy and Wilhelm Weinberg in 1908 [33, 104]), as
follows: (a) Find the resulting values of the relative viabilities F and G.
Where does this place the scenario relative to Figure 2.14? (b) Simplify
equation (2.21).
10. Suppose that, in the genetics discussion in this section, we are studying a
gene for which aa homozygotes never reproduce (they are either infertile
or do not survive to maturity). The a allele may nevertheless survive
among heterozygotes. In this case φaa = 0, and thus G = 0. (Note also
in this case we must have p0 > 0, or else the entire population consists
of aa homozygotes, and there is no next generation.)
(a) How do equation (2.21) and its behavior change when G = 0?
(b) Interpret biologically the criterion developed in (a) for survival of
the a allele when G = 0.
(c) Show that the G = 0 model in (a) guarantees pn > 1/2 for n ≥ 1.
Explain biologically why this must be so.
(d) Sketch (by hand or by computer) a graph of the remaining equi-
libria p as a function of the remaining parameter F . This type of
graph is called a bifurcation diagram, and we shall discuss bifurca-
tion diagrams further in Section 2.6, and in later chapters.
∗
(e) Have a computer generate a three-dimensional bifurcation dia-
gram for the model (2.21) in terms of both parameters F and G.
(Segel8 , p. 50) If we reverse assumption (2) and suppose that all
8 Lee A. Segel, Modeling dynamic phenomena in molecular and cellular biology, Cam-
population size over time. Nicholson discovered9 that the blowfly population,
which emerged in distinct (or nearly distinct) generations, was controlled en-
tirely by the rate at which food was provided, so that (when this rate was held
constant) a small generation of blowflies, encountering no resource limitations,
produced a large next generation. The next generation, however, was so large
that each individual had a minuscule share of the food, and consequently
very few were able to develop eggs, resulting in a very small next generation.
These oscillations were sustained over a period of months (see Figure 2.15),
and others have since observed similar fluctuations in fly populations.
Contest competition, on the other hand, assumes that some individuals —
the most fit — are able to secure enough resources for themselves, while the less
fit may end up with nothing. Here we may imagine a ranking in which resources
are distributed to individuals in decreasing order of fitness, until resources are
exhausted. Under contest competition, some number of individuals (as many
as resources can support) will always be able to reproduce. For example, if we
consider the number of trees of a given species in a certain patch of woods,
often the tall trees spread their branches to catch as much sunlight as needed
for growth through photosynthesis, crowding out smaller trees nearby (which
then stop growing or may even die) until each tall tree’s foliage meets that
of its neighbors, allowing little sunlight to reach the ground. New trees then
grow only when a large tree falls, creating an opening.
The differences between these two assumptions will lead naturally to dif-
ferences in the mathematical models we develop to describe such populations.
In particular, scramble competition is modeled with reproduction functions
which rise to a maximum and then fall off, as excessive size reduces repro-
9 A.J. Nicholson, An outline of the dynamics of animal populations, Australian J. Zoology
3: 9–65, 1954.
52 Dynamical Systems for Biological Modeling: An Introduction
duction. In some cases reproduction may reach zero for some finite carrying
capacity, as with the logistic model (2.12), while in others it may taper off
gradually, as with Ricker’s model (first seen in the exercises at the end of
Section 2.2 and discussed in more detail below). Contest competition, mean-
while, is modeled with reproduction functions which rise monotonically and
flatten out as they reach the ceiling created by the resource limitations. The
Verhulst model (2.10) incorporates contest competition. Figure 2.16 sketches
the general form of these two types of functions.
We can consider the consequences of each type of intraspecies competition
via the properties described above of the corresponding reproduction func-
tions. In either case we will also require that the reproduction function g(y)
in the difference equation yk+1 = g(yk ) have the following properties:
• For small populations y, reproduction should be roughly linear in y,
corresponding to a constant (maximum) per capita ratio r called the
intrinsic growth rate, i.e.,
g(y)
lim = r > 0.
y→0 y
in the United Kingdom (M. Graham, ed.), Edward Arnold, London, 1956, pp. 372–441.
11 W.E. Ricker, Stock and recruitment, J. Fisheries Research Board Canada 11: 559–623,
1954.
Difference Equations (Discrete Dynamical Systems) 55
Photo by Bernard E. Picton and Christine C. Morrow Ken Gray Image Courtesy of Oregon State University
FIGURE 2.17: The bottom-feeding plaice (left), hard to see against the
sea floor except when they move, have high fertility and low survivorship to
maturity, leading to contest competition. The Mediterranean flour moth or
mill moth (right), seen in its characteristic resting pose, undergoes different
types of intraspecies competition during its life cycle as a result of parasitism.
Thus the equilibrium y∞ is asymptotically stable if and only if −1 < 1−log a <
1. This condition reduces to log a < 2, or a < e2 . The result suggests that fish
which cannibalize their young and have a high birth rate (a > e2 ) may have
an unstable equilibrium. We shall see in Section 2.6 that this may indicate
oscillations in population size. Oscillations in population size can lead to times
at which the population size is very small, so that a small perturbation could
lead to wiping out the population. Fish populations which satisfy a Ricker
model may be quite vulnerable to external influences.
In this and later sections, we will use the Ricker form g(y) = aye−by as a
ry
prototypical scramble recruitment function and the Verhulst form g(y) = y+A
as a prototypical contest competition recruitment function.
56 Dynamical Systems for Biological Modeling: An Introduction
283–295, 1975.
13 J. Maynard Smith and M. Slatkin, The stability of predator-prey systems, Ecology 54:
384–391, 1973.
14 S.D. Lane and N.J. Mills, Intraspecific competition and density dependence in an Eph-
Exercises
In Exercises 1–4 determine the values of r for which there is an asymptotically
stable equilibrium.
2
ryk
1. yk+1 = 2 +A
yk
2.5 Harvesting
Another mechanism besides competition which regulates population sizes
is the removal of individuals by another species, usually for food. Fish eat
plankton, and aquarium owners often buy algae-eating fish to keep their tanks
clean. Wolves and coyotes hunt deer and smaller mammals in the woods of
North America; lions hunt gazelles on the plains of Africa. Humans also cut
down trees to make paper and wood for construction. In areas where humans
have removed natural predators, the prey populations often grow unchecked
and can cause damage to the environment or become a nuisance. Deer popu-
lations, for instance, in areas where wolves have been shot by farmers, rise to
the point that they run onto roads, causing traffic accidents, and wander into
residential neighborhoods and eat flowers, bushes and gardens bare. Hunting
seasons have been developed and managed as a response to this overpopula-
tion, as a means of population control. Human colonization of wild areas has
also led to other, more accidental removals of natural predator or competi-
tor species, necessitating further population management programs based on
careful study. All these types of removal and control are known as harvest-
ing, and harvesting has become a major concern in wildlife and environmental
management. Rempel and Kaufman, for instance, studied16 the effects of both
spatial and temporal tree harvesting patterns in the Nakina Forest of north-
western Ontario, Canada, on the preservation of wildlife habitat for animals
such as caribou, moose, and martens, balancing commercial timber production
objectives with a concern not to fragment forest habitats.
We shall consider here only one type of harvesting, that is constant-rate
or constant-yield harvesting, by which we mean removing a fixed number of
members just before the annual birth period. Thus a population described
in the absence of harvesting by a difference equation yk+1 = g(yk ) would
be described under harvesting by a difference equation yk+1 = g(yk ) − H,
where H is the number of individuals harvested per year. The questions of
interest to investigate involve the relationship between the harvesting rate and
the equilibrium population size. What is the maximum harvest rate that will
allow the population to persist? What is the minimum harvest rate necessary
to prevent excessive population growth? To what extent does harvesting make
a population more vulnerable to extinction from a catastrophic event such as
drought or disease? Can harvesting make a population less vulnerable to such
events? We begin by considering a simple example.
Example 1.
Consider a population governed by a logistic difference equation from which
16 R.S. Rempel and C.K. Kaufman, Spatial modeling of harvest constraints on wood sup-
ply versus wildlife habitat objectives, Environmental Management 32(5): 646–659, 2003.
doi: 10.1007/s00267-003-0056-8.
Difference Equations (Discrete Dynamical Systems) 59
if and only if
2
1 4HK
K2 1 − − ≥ 0,
r r
which we can rewrite as
(r − 1)2
H ≤ Hmax = K.
4r
This gives the range of values of H (for a given value of r) for which the
population will persist.
To determine the (local) stability of the two equilibria found in (2.30), we
recall the condition |g ′ (y∞ )| < 1 for the stability of an equilibrium y∞ of the
difference equation yk+1 = g(yk ). Here we find that for (2.28)
y 2y
g(y) = ry 1 − − H, g ′ (y) = r 1 − ,
K K
which reduces to
4rH
(r − 1)2 − < 4.
K
We can rewrite this in terms of H as
(r − 1)2 − 4 (r − 3)(r + 1)
H > Hmin = K= K,
4r 4r
which is satisfied for all H ≥ 0 when r < 3. Therefore, when r < 3, all
H < Hmax allow the population to persist at a (stable) positive equilibrium,
provided that the initial population is above the unstable (smaller) equilib-
rium. (Too small an initial population will go extinct from even a little har-
vesting.)
The perhaps surprising result comes from addressing the second half of
the original question. Recall (Section 2.3, Example 2) that for r > 3 the
logistic equation without harvesting has no stable equilibria. When r > 3,
however, (2.28) has a stable positive equilibrium for H in the interval Hmin <
H < Hmax . For H > Hmax , harvesting is too great, and the population is
wiped out in a finite number of generations (since there is no equilibrium). For
H < Hmin , harvesting is not strong enough to temper the high reproductive
rate, the equilibria remain unstable, and the population oscillates, as it does in
the absence of harvesting. Therefore, we see that for populations that exhibit
scramble competition (as in the logistic model) and a high fertility rate (r >
3, measured in appropriate units), it is actually possible to use harvesting
as a means of stabilizing the population, which leaves it less vulnerable to
fluctuations caused by other external or environmental factors.
FIGURE 2.18: At left, fishery biologists pumping eggs from the Tenor River,
Alaska, 1966. At right, a fishery biologist returns fish to their habitat. Fishery
biologists play a key role in restoring such important species as Atlantic and
Pacific salmon, American shad, striped bass, and Great Lakes trout.
given by
s 2
1 a−1 a−1 H
y∞ = −H ± −H −4 . (2.33)
2 b b b
With some algebra (see Exercise 4) √ one can show that these equilibria exist
and are positive for H ≤ Hc = ( a − 1)2 /b, a > 1. (For H = 0 these are
the equilibria 0 and a−1
b found in the previous section for the Beverton-Holt
model without harvesting.) Hc , which is the greatest harvesting yield that
allows the population to persist, is called the critical harvesting yield. Here
ay a
g(y) = 1+by − H and g ′ (y) = (1+by) ′
2 , so the criterion |g (y∞ )| < 1 for stability
√
becomes y∞ > a−1 b . With, again, some algebra, one can show (see Exercise 5)
√
a−1
that b lies between the two equilibria (2.33) when H ≤ Hc , a > 1, so
that only the larger equilibrium is stable. Thus, as with the logistic model
(Example 2.5), any amount of harvesting (H > 0) will drive a population
with low fertility (a < 1) to extinction in finite time (since there are no
equilibria). Populations with high fertility (a > 1) can survive harvesting
as long as the initial population is large enough (greater than the smaller,
unstable equilibrium) and there is no over-harvesting (i.e., as long as H < Hc ).
If we compare these results with those for the Beverton-Holt model without
harvesting derived in the previous section, we see that the effects of harvesting
on species exhibiting contest competition are threefold:
62 Dynamical Systems for Biological Modeling: An Introduction
1
z
ry
z=
z=y 1+by
z=y
0.8
0.6
z
ry
z= 0.4
1+by
0.2
y 0 0.2 0.4
y
0.6 0.8 1
ry
z=
1+by
0.6 z=y+H 0.6 0.6
z z z
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
y y y
sufficiently small H, there are two intersections, with one starting at 0 and
increasing and the other starting at a−1b and decreasing (Figure 2.21). At the
smaller intersection, the slope of the curve is greater than the slope of the line
(which is 1) and thus by the stability theorem of Section 2.4 this equilibrium
is unstable. At the larger intersection, the slope of the curve is positive but
less than 1 and so by the stability theorem this equilibrium is asymptotically
stable. This means that the limiting population size is the second equilibrium,
as long as the initial population size is greater than the first equilibrium
(otherwise the population will die out).
However, as H increases, it will reach a value where the two equilibria come
together and curve and line are tangent (Figure 2.22). If H increases beyond
this value, there will be no equilibrium (Figure 2.23). Thus the qualitative
behavior of the system will be that there is an asymptotically stable positive
equilibrium until H reaches this critical value, which we denote by Hc , and
then the equilibrium jumps to zero. A jump in equilibrium value is called a
catastrophe, because biologically it corresponds to the disaster of wiping out
the population.
√
It is also possible to use graphs to show that the critical harvest rate is
( a−1)2
and that the equilibrium population size at this critical harvest rate
√b
a−1 ay
is b . This may be shown by recognizing that, since the graphs of z = 1+by
and z = y + H are tangent when H = Hc (Figure 2.23), the equilibrium
population size at the critical harvest rate is the point on the reproduction
curve at which the slope is 1. Substituting this value of y into equation (2.32)
and then solving for H will give us the critical harvest rate Hc (see Exercise 3
below).
We now consider a particular application.
64 Dynamical Systems for Biological Modeling: An Introduction
Example 2.
As reported in the previous section, Reeve, Rhodes and Turchin (1998) found
that the southern pine beetle Dendroctonus frontalis experiences contest com-
petition early in its life cycle, during oviposition. Their study measured the
numbers of eggs deposited in galleries in each attack site on a tree, and, match-
ing their data to the flexible competition model (2.27) proposed by Maynard
Smith and Slatkin (1973), gave an estimate of the exponent n which (as also
described in the previous section) measures the type of competition exhibited.
The estimate n = 1.5 corresponds roughly to the Beverton-Holt model. They
also gave parameter estimates of a = 54 and b = 0.226/egg.
Use these estimates to find the expected average (equilibrium) number of
eggs per attack site, and the critical harvest rate Hc , the maximum number
of eggs which could be removed or parasitized per attack site without driving
the population to extinction.
Solution: In the absence of harvesting, the equilibrium value is (a − 1)/b as
long as a > 1 (which it is here). This gives (54 − 1)/0.226 ≈ 235 eggs per
attack site, with a critical harvest yield
√ √
( a − 1)2 ( 54 − 1)2
Hc = = ≈ 178 eggs.
b 0.226
The harvested Ricker model is
aye−by = y + H, (2.35)
the intersections of the graphs of the two expressions. Since this equation is
transcendental and cannot be solved explicitly, we will proceed with a graph-
ical analysis. The curve z = aye−by starts at the origin with slope a, increases
a
to a maximum value of be at y = 1b and then decreases, approaching a limit
of zero as y → ∞. If a < 1, the line z = y + H lies above this curve, except
for an intersection with the curve at y = 0 if H = 0, as was the case for
the Beverton-Holt model. Thus if a < 1 there is again no equilibrium under
harvesting, and the population dies out.
If a > 1, the analysis proceeds very similarly to the harvested Beverton-
Holt model. For H = 0 (no harvesting), the line z = y starts below the curve
at y = 0 and intersects it again at a positive value of y (Figure 2.24). Again,
increasing H from zero means moving the line z = y + H up, parallel to itself.
For sufficiently small H there are two intersections, with one starting at 0 and
increasing, and the other starting at logb a and decreasing (Figure 2.25). At the
smaller intersection, the slope of the curve is greater than the slope of the line
(which is 1) and thus by the stability theorem of Section 2.4 this equilibrium
is unstable. At the larger intersection, stability depends on the slope of the
Difference Equations (Discrete Dynamical Systems) 65
1
z z
0.8
z=y+H z=y+H
0.6
stable z=rye−by
z −by
z=rye
0.4
0.2
unstable
unstable
0 0.2 0.4 0.6 0.8 1 y y
y 1/b 1/b
curve. If the curve and line intersect to the left of y = 1/b (Figure 2.25a),
the slope of the curve is positive but less than 1 and by the stability theorem
this equilibrium is asymptotically stable. To the right of y = 1/b the curve’s
slope is negative (Figure 2.25b). If this slope is between 0 and −1 (which is
guaranteed if a < e2 ), the equilibrium will be asymptotically stable, but it is
possible for the curve’s slope to be less than −1 here, in which case neither
equilibrium is stable.
As H increases, it will reach a value Hc where the two equilibria come
together and curve and line are tangent (Figure 2.26). If H increases beyond
this value, there will be no equilibrium (Figure 2.27). Thus the qualitative
behavior of the system will be that for H near Hc there is an asymptoti-
cally stable positive equilibrium until H reaches this critical value, past which
the population size falls to zero. Again, we have a catastrophe, because this
situation corresponds biologically to the disaster of wiping out the population.
By matching the derivatives of the two graphs at the point of tangency
when H = Hc (Figure 2.26), we see that the equilibrium population size at
this point is the solution y∞ of the transcendental equation
ae−by∞ (1 − by∞ ) = 1
(the left-hand side is the derivative of the reproduction curve; the slope of the
line is 1). From the equilibrium condition (2.35) we also see that the critical
harvest rate is aye−by − y with this value of y∞ . However, it is not possible to
solve the equation (2.35) explicitly for y. In practice, one would use a numerical
approximation method to solve (2.35) and then one could calculate the critical
harvest rate. It would also be necessary to use a numerical approximation
method to find the equilibrium population size for a given harvest rate.
The stability analysis of the harvested Ricker equation is rather more com-
plicated, and we shall not attempt to carry it out. In fact, as with the logistic
66 Dynamical Systems for Biological Modeling: An Introduction
1 1
0.8 0.8
0.6 0.6
z z
0.4 0.4
0.2 0.2
Exercises
1. Reeve, Rhodes and Turchin (1998) also fit the number of pine beetle
oviposition (egg-laying) galleries per attack site on a tree to the Maynard
Smith and Slatkin model (2.27) and found that it too exhibits contest
competition, with an estimated n = 0.828 for naturally occurring attack
sites and n = 0.914 for field-experiment sites (recall n = 1 corresponds
to the Beverton-Holt model). For naturally occurring sites, the other
parameter estimates were a = 27.596 and b = 0.114/gallery; for the
field-experiment trees, they estimated a = 25.639 and b = 0.133/gallery.
Assume a Beverton-Holt model in both cases.
(a) Find the equilibrium gallery densities for both types of sites.
(b) Calculate the critical harvest rate for both types of sites.
Difference Equations (Discrete Dynamical Systems) 67
(c) Find the limiting numbers of galleries per attack if a certain num-
ber H of galleries are “harvested” (removed or prevented by ex-
perimenters) for H = 75 and 150. Which of the two types of site
appears more sensitive to harvesting?
2. Suppose we now apply the parameter estimates from the previous prob-
lem (a = 27.596 and b = 0.114/gallery for naturally occurring attack
sites) to a Ricker model with harvesting.
(a) Estimate the critical harvest rate numerically.
(b)∗ Estimate numerically the minimum harvest rate required to sta-
bilize the population (i.e., the positive equilibrium). [Hint: Plot
g ′ (y∞ (H)) as a function of H and see where it falls inside the in-
terval (−1, 1).]
ay
3. (a) Show that the point on the curve z = 1+by which has slope 1 is
√
a−1
y= b .
√
ay a−1
(b) Solve the equation 1+by = y + H with y = b for H.
(a) Show that this model has a zero equilibrium which is stable when
E − 1 < a < E + 1.
a
(b) Show that this model has a positive equilibrium E+1 − 1 /b
(E+1)2
which is stable when E < 1 or a < E−1 .
2
(c) What happens when a > (E+1)
E−1 ? when a < E − 1? It may help to
try some numerical examples or graph these criteria in the E − a
plane.
(d) Based on this analysis, summarize the effects of constant-effort har-
vesting on a population undergoing contest competition.
10∗ . Write the difference equation for the logistic model with constant-effort
harvesting, and derive the stability criterion for its zero equilibrium.
Difference Equations (Discrete Dynamical Systems) 69
1
0.6
0.5
0.8
0.4
0.6
y
0.3
0.4
0.2
0.2
0.1
0 2 4 6 8 10 12 14 16 18 20
k
0 0.2 0.4 0.6 0.8 1
may have plenty of resources, will produce a second generation which is too
great for the local resources to support. If the population exhibits scramble
competition as in the logistic model, then the second generation will then be
nearly unable to reproduce — that is, the third generation will be quite small,
as very few members of the second generation will have had children. The third
generation, however, being few in number, will have all the resources needed
to reproduce at their intrinsic capacity, and the fourth generation will again
be quite large. We observe an oscillation or alternation here between small and
large generations, based on three factors: discrete generations (reproduction
occurs simultaneously, rather than continually until environmental resources
are all in use), scramble competition, and a very high intrinsic reproduction
rate. In fact, we see this kind of oscillation in the logistic model when 2 <
r < 3 — for instance, in Figures 2.28 and 2.29, where r = 2.5 — but the
oscillation dies down, and the population level eventually approaches a steady
intermediate state. We might therefore conjecture that populations with an
extremely high intrinsic reproduction rate may cause such wild fluctuations
that the population size never settles down. Indeed, as we shall see, this is
precisely what happens in the logistic model with r > 3.
When r increases beyond 3, a solution of period 2 appears in the lo-
gistic model. By this we mean that there is a pair of values y + and y −
such that rg(y + ) = y − and rg(y − ) = y + . Thus the alternating sequence
{y + , y − , y + , y − , . . .} is a solution of the difference equation (2.36). Since
this solution repeats after two terms, we call it a solution of period 2, or a
2-cycle. We can verify this behavior and find the values of y + and y − either
numerically (choose values for the parameters, say K = 1 and r just beyond 3,
and then iterate the solution many times) or analytically. In order to find the
values analytically, we must solve the equation yk+2 = yk , i.e., rg[rg(y)] = y.
As the function g(y) is a quadratic polynomial, this is a fourth degree polyno-
mial equation. We already know two of the roots of this polynomial equation,
namely the equilibria y = 0 and y = K(1 − 1r ). By dividing out the factors
corresponding to these two roots, we may reduce the fourth-degree equation
to the quadratic equation
2 r+1 2 r+1
y −K y+K =0
r r2
0.8
0.6
0.6
y
0.4
0.4
0.2
0.2
0 2 4 6 8 10 12 14 16 18 20
k
0 0.2 0.4 0.6 0.8 1
We may find equilibria of such equations by looking for constant solutions; for
(2.38) the condition is y∞ = g2 (y∞ ). The linearization about each equilibrium
is obtained by replacing any nonlinear functions by the linear part of their
Taylor expansion and forming the corresponding linear difference equation; for
the second-order equation (2.38) we replace g2 (y) by g2′ (y∞ )u and obtain the
linearization uk+2 = g2′ (y∞ )uk . Finally, the stability criterion is then obtained
by recalling from Section 2.1 that solutions of linear difference equations have
the form uk = u0 λk for some number λ, and substituting u0 λk+2 for uk+2
and u0 λk for uk ; for (2.38) this gives λ2 = g2′ (y∞ ). The criterion |λ| < 1 for
stability then becomes |g2′ (y∞ )| < 1.
y
For the logistic function g(y) = y(1 − K ), these computations are com-
plicated. Exercises 3, 4 and 5 at the end of this section suggest one way to
∗ ∗
manage these calculations. It is possible to show that the equilibria y+ and y−
found above for (2.37) are asymptotically√stable (and hence the corresponding
2-cycle of (2.36) is stable) if 3 < r < 1 + 6 ≈ 3.4495. That is, for r between 3
and 3.4495, every solution of the logistic equation (2.36) approaches a solution
of period 2.
For r > 3.4495, the solution of period 2 is unstable, but it is possible to
show (in the same way) that a solution of period 4 appears, and that this
72 Dynamical Systems for Biological Modeling: An Introduction
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0 2 4 6 8 10 12 14 16 18 20
k
0 0.2 0.4 0.6 0.8 1
solution is asymptotically stable if 3.4495 < r < 3.544. Figure 2.32 shows the
graph of the solution, and Figure 2.33 shows the cobweb diagram. When the
solution of period 4 becomes unstable, a solution of period 8 appears, and this
solution is asymptotically stable for 3.544 < r < 3.564 (see Figures 2.34 and
2.35).
This period doubling phenomenon continues until r = 3.570, when periodic
solutions whose periods are not powers of 2 begin to appear. Figure 2.36
shows what appears to be close to a solution of period 5 for r = 3.75, and
Figure 2.37 shows the corresponding cobweb diagram. If we iterate more terms,
the cobweb diagram looks less like a periodic solution. (Figures 2.38 and 2.39
show a solution for r = 3.9.) For any positive integer p there is some value of
r > 3.828 for which the logistic equation has a periodic solution with period p,
but different initial values give solutions whose behaviors are quite different.
Such behavior is called chaotic.
These facts, whose proofs are difficult and require a close examination of
the properties of continuous functions and of the fixed points of iterates of
continuous functions, are not restricted to the logistic difference equation. It
is a remarkably robust fact that for every difference equation of the form
yk+1 = g(yk ), where g(y) is a function which increases from zero to a unique
maximum and then decreases and approaches zero as y → ∞ (possibly hitting
zero for finite y), the period doubling phenomenon and the onset of chaos
occur. The Ricker model introduced in Section 2.4 is another example of such
a g(y). (The Beverton-Holt model is not, because its g(y) never decreases.)
In fact, the period doubling begins to happen at the same rate: the intervals
of r values during which a given solution of period 2n is stable begin to decrease
in size by the same factor, as n increases. More specifically, if rn is the value
of r for which the asymptotically stable solution of period 2n appears, then
Difference Equations (Discrete Dynamical Systems) 73
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0 2 4 6 8 10 12 14 16 18 20
k
0 0.2 0.4 0.6 0.8 1
a number which is called the Feigenbaum constant. Usually, the limiting value
is approached very rapidly. This means that the period doubling values of r
occur closer and closer together.
One way to illustrate period doubling and chaotic behavior is by means of
a bifurcation diagram, which we shall discuss in greater detail in Section 4.4. A
bifurcation diagram in general plots equilibria as a function of the parameter
r. If there is an orbit of period 2 for a given value of r, then the bifurcation
diagram will have two points for that value of r, and similarly an orbit of any
period will show in the bifurcation diagram as a number of points equal to the
period of the orbit for that value of r. To draw a bifurcation diagram for, say,
the logistic difference equation, we iterate the logistic function for a sequence
of values of r and plot the results obtained starting after enough iterations
for the orbit to have approached the periodic orbit. This may be done using
74 Dynamical Systems for Biological Modeling: An Introduction
1 1
0.8 0.8
0.6 0.6
0.4
0.4
0.2
0.2
0 2 4 6 8 10 12 14 16 18 20
k
0 0.2 0.4 0.6 0.8 1
0.8 0.8
0.6 0.6
0.4
0.4
0.2
0.2
0 10 20 30 40
k
0 0.2 0.4 0.6 0.8 1
opens the door to chaos for such populations. We may wonder what other
biological mechanisms may produce chaotic behavior in populations with dis-
tinct generations. Although we cannot provide an exhaustive answer to this
question, in the remainder of this section we shall consider two other examples
of biological phenomena which have been associated with chaotic behavior.
2.6.1 Dispersal
The term dispersal refers to spatial redistribution of some or all of a popu-
lation, often in response to population pressures (intraspecies competition as
population density builds in a single area). While models for biological dis-
persal typically keep track of population sizes in several different habitats or
patches (as we shall see in the following section), some simple studies have
been made on the effects of dispersal in a single patch.
Since dispersal from a single source tends to reduce the population den-
sity in that location, we might suspect that it will also counter the chaotic
tendencies of populations with high reproduction rates. We can see this very
simply if we consider the usual form of dispersal, in which a given proportion
d of individuals leave the patch (habitat) under study following reproduction.
In this case, the model yk+1 = g(yk ) becomes
yk+1 = (1 − d)g(yk ),
0.8
0.6
FIGURE 2.41: An
eosinophil, a type
of granulocyte, a
specialized white
0.4 blood cell, surrounded
here by erythrocytes
(red blood cells).
Granulocytes fight
infection by phago-
0.2
cytosis; eosinophils
are identifiable by
their multi-lobed nu-
clei. Their production
exhibits regular oscil-
0 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 lation patterns but can
be affected by diseases
FIGURE 2.40: Bifurcation diagram for the logis- such as leukemia.
tic difference equation, with the parameter r on the
horizontal axis.
other nearby patches into this one. Since we are not modeling the population
densities in the other patches, we incorporate immigration as a constant c,
making the basic form
yk+1 = g(yk ) + c. (2.39)
The result of this immigration is to increase the equilibrium value y∞ above
what it would be without immigration, but the effect upon |g ′ (y∞ )| is difficult
to predict, and so in terms of simplicity of dynamics, constant migration may
either stabilize or destabilize a population’s dynamics. By the same token,
then, a constant dispersal (i.e., c < 0, as opposed to the proportional disper-
sal d discussed above) can also have either effect, in addition to reducing the
value of the positive equilibrium. Note that constant dispersal is equivalent to
constant-yield harvesting, so a high enough dispersal rate will drive a popula-
Difference Equations (Discrete Dynamical Systems) 77
tion to extinction in finite time (as seen in Section 2.5). Doebeli discusses17 a
way to use variable dispersal to stabilize a population’s dynamics by making
the magnitude of the dispersal dependent on population density.
We shall now look briefly at examples where immigration and dispersal
can either stabilize or destabilize a population. We shall follow Doebeli’s lead
in using the flexible model (2.27) proposed by Maynard Smith and Slatkin
(1973) to describe a population’s natural growth. Constant migration applied
to this model gives us the equation
ayk
yk+1 = + c. (2.40)
1 + (byk )n
Example 1.
Describe the behavior of the Maynard Smith and Slatkin model with constant
immigration (2.40) and parameter values a = 32 , b = 1/pop. unit, and n = 5,
for migration values of c = 0, c = 12 , and c = 1 pop. units.
Solution: The parameter values given correspond to scramble competition
(n = 5) with an intermediate intrinsic growth ratio(a = 3/2). We may then
expect the positive equilibrium to be stable, and it is. We calculate
1/5
(a − 1)1/n 1 1 − (a − 1)(n − 1) 2
y∞ = = ≈ 0.870551, g ′ (y∞ ) = =− ,
b 2 a 3
What happens here is that a little immigration increases the first genera-
tion just enough to make the second generation considerably smaller — that
is, the decrease in natural population size outweighs the boost given by immi-
gration — but a lot of immigration will end up dominating the population’s
natural growth, so that each new generation will consist almost entirely of
newly arrived immigrants, and very few descendants of the previous genera-
tion (which was so numerous it was practically unable to reproduce).
Example 2.
Describe the behavior of the Maynard Smith and Slatkin model with constant
dispersal (2.40) and parameter values a = 10, b = 1/pop. unit, and n = 2, for
dispersal values of c = 0 and c = −1 pop. units.
Solution: The parameter values correspond to intermediate or very weak
17 M. Doebeli, Dispersal and dynamics, Theoretical Population Biology 47: 82–106, 1995.
78 Dynamical Systems for Biological Modeling: An Introduction
(a − 1)1/n 1 − (a − 1)(n − 1) 4
y∞ = = 3, g ′ (y∞ ) = =− ,
b a 5
verifying stability for c = 0. With a dispersal of 1 pop. unit per generation,
however, we compute y∞ ≈ 2.47419, g ′ (y∞ ) ≈ −1.00983, which is just enough
to destabilize the equilibrium. As the dispersal amount increases, things get
worse; for c < −4.085, there are no positive equilibria at all, and the pop-
ulation must “crash” to zero in finite time. Here, in some sense, dispersal
magnifies the effect of the competition, so that instability occurs.
Science 197: 287–289, 1977; L. Glass and M.C. Mackey, Pathological conditions resulting
from instabilities in physiological control systems, Ann. N.Y. Acad. Science 316: 214–235,
1979.
Difference Equations (Discrete Dynamical Systems) 79
Exercises
1. Find the value of r for which a solution of period 2 appears for the
difference equation
yk+1 = ryk e1−yk .
2. Consider the Hassell model yk+1 = ayk (1 + nb yk )−n with a, b > 0, n > 1.
Show that this model does not exhibit period doubling if n < 2 and find
the value of a for which a solution of period 2 appears if n > 2. [Hint:
See Exercise 12, Section 2.4.]
3. Let {y+ , y− } be a solution of period 2 of the difference equation yk+1 =
g(yk ). Show that both y+ and y− are equilibria of the second-order
difference equation yk+2 = g[g(yk )].
4. ∗
Define a new index m = k2 for k even and define the iterated function
g2 (y) = g[g(y)]. Show that the solutions y+ , y− of period 2 from Exer-
cise 3 are equilibria of the first-order difference equation ym+1 = g2 (ym ).
[Remark: Exercise 4 and the stability theorem of Section 2.3 show that
an equilibrium y∗ of this second-order difference equation is asymptot-
ically stable if |g2′ (y∗)| < 1. Exercise 5 below gives another criterion
for the asymptotic stability of a solution of period 2 of the difference
equation yk+1 = g(yk ).]
∗
5. (a) Let {y+ , y− } be a solution of period 2 of the difference equation
yk+1 = g(yk ). Use the chain rule of calculus to show that if g2 (y) =
g[g(y)], then
d
g2′ (y+ ) = {g[g(y)]} = g ′ (y− )g ′ (y+ ).
dy y=y+
(b) Deduce from part (a) that the solution of period 2 is asymptotically
stable if |g ′ (y− )g ′ (y+ )| < 1.
∗
6. Doebeli (1995) writes regarding constant dispersal (or immigration)
applied to the Maynard Smith and Slatkin (1973) model (2.27), “one can
choose the parameters in [this equation] so that ... dispersal destabilizes
the system. For example, [the equation] can have a stable equilibrium
Difference Equations (Discrete Dynamical Systems) 81
9. Identify and discuss at least three modeling issues raised in the discus-
sion of the blood cell model in this section.
10. What practical difference is there between constant-yield harvesting and
constant-rate dispersal?
82 Dynamical Systems for Biological Modeling: An Introduction
yk+1 = f (yk , zk ),
(2.43)
zk+1 = g(yk , zk ),
where the growth of each new generation depends on the sizes of both popula-
tions in the previous generation. Algebraically, an equilibrium of this system
is a solution of the pair of equations
y = f (y, z),
(2.44)
z = g(y, z).
and denote the 2 × 2 matrix above as A(y∞ , z∞ ), and the vector [u0 v0 ]T as
~u0 , then we can write the equation in the form A~u0 = λ~u0 , or
(A − λI)~u0 = 0, (2.47)
where I is the 2 × 2 identity matrix. From linear algebra, we know that the
only way for (2.47) to have nonzero solutions ~u0 is for the matrix (A − λI) to
be singular. That is, this system has nontrivial solutions if λ satisfies
Our result, therefore, is that all solutions of the linearization (2.46) approach
zero if all roots λ of the characteristic equation (2.48) or (2.49) satisfy |λ| < 1.
In this matrix form, the stability theorem of Section 2.3 generalizes to two-
dimensional systems, and indeed to systems of any dimension. (Note that for
single difference equations, A and I are simply scalars, and the characteristic
equation for yk+1 = g(yk ) at y∞ has the unique solution λ = A = g ′ (y∞ ).)
We shall state this result first for a system of any dimension, and then more
specifically for two-dimensional systems.
and let A(~y∞ ) be the matrix of the linearization of the system at this
equilibrium. If all roots of the characteristic equation (2.48) satisfy |λ| <
1, then the equilibrium is asymptotically stable, and if the characteristic
equation has a root with |λ| > 1, the equilibrium is unstable.
There is a set of conditions, known as the Jury criterion, which gives nec-
essary and sufficient conditions that all roots of a polynomial equation satisfy
|λ| < 1. (In general, for an equation of degree n, the Jury criterion consists
of n different inequalities.) The characteristic equation at an equilibrium of a
two-dimensional system of difference equations is a quadratic equation. The
Jury criterion for two-dimensional systems is as follows: Both roots of the
quadratic equation
λ2 + a1 λ + a2 = 0
satisfy |λ| < 1 if and only if
Example 1.
For each equilibrium of the system yk+1 = 2zk , zk+1 = yk2 , determine whether
it is asymptotically stable.
Solution: Equilibria are solutions of y = 2z, z = y 2 . Substituting the first of
these equations into the second, we obtain z = 4z 2 , which has two solutions,
z = 0 and z = 14 . If z = 0, then y = 0, and if z = 14 , then y = 12 . Thus there
are two equilibria: (0,0) and ( 12 , 14 ). The matrix of the linearization at (y, z) is
0 2
.
2y 0
If y = 0, this matrix has trace 0 and determinant 0 and the conditions (2.50)
are satisfied. Thus the equilibrium (0,0) is asymptotically stable. If y = 12 , the
matrix has trace 0 and determinant −2, and the condition |tr A| < det A + 1
is violated. The equilibrium ( 21 , 14 ) is therefore unstable.
Example 2.
Determine the condition for asymptotic stability of an equilibrium y∞ of the
second-order difference equation yk+2 = g(yk ) by rewriting it as a first-order
system.
Solution: We rewrite the equation as the system yk+1 = zk , zk+1 = g(yk ). An
equilibrium is a solution of y = z = g(y), and the matrix of the linearization
at an equilibrium y is
0 1
.
g’(y) 0
86 Dynamical Systems for Biological Modeling: An Introduction
p
The characteristic equation is λ2 − g ′ (y) = 0, with roots ± g ′ (y). Thus
the condition for asymptotic stability is |g ′ (y)| < 1.
yk+1 = py yk + qz zk + yk B(yk ),
(2.51)
zk+1 = qy yk + pz zk .
The system (2.51) has two equilibria: extinction (0,0), which always exists,
19 E.M. Stock, Deterministic discrete-time epidemic models with applications to amphib-
FIGURE 2.42: Left: a Pacific tree frog. Right: a high sierran lake with
mountain yellow-legged frogs. Amphibians such as these frogs move between
lakes and the surrounding land; the lakes, which serve as breeding sites, also
act as focal points for viral and fungal infections and pesticide runoffs.
qy
and survival B −1 (m), 1−p z
B −1
(m) , which exists when B(0) > m, where m
is an average effective mortality ratio for the playa patch,
qy qz µz
m = (1 − py ) − = µy + qy
1 − pz qz + µz
(in every generation of reproductive adults yk , a proportion qy of them leave
for land; a proportion qzµ+µ
z
z
of those die before returning to the water).
The asymptotic stability of the extinction equilibrium can be analyzed via
the matrix for the linearization of (2.51) about (0,0),
py + B(0) qz
A= .
qy pz
Applying the Jury criterion, we see that the extinction equilibrium is stable
if (and only if)
If tr A > 0, then we drop the absolute value bars and rewrite the condition
(with some algebra20) as m+ B −1 (m)B ′ (B −1 (m)) < m, which is again always
true because B is decreasing. However, if tr A < 0, the condition can be
rewritten as
qy qz
m + B −1 (m)B ′ (B −1 (m)) > − (1 + py ), (2.53)
1 + pz
Example 3.
Show that the survival equilibrium of (2.51) is always stable when it exists, if
the fertility is described by the Beverton-Holt model.
ay a
Solution: We have yB(y) = 1+by , so that B(y) = 1+by . Thus B −1 (m) =
1 a
b m − 1 , and the matrix for the linearization of (2.51) about the survival
equilibrium is 2
py + ma qz
A= .
qy pz
m2
Here tr A = py + pz + a > 0, so the Jury criterion holds as noted above.
Example 4.
Determine the stability condition for the survival equilibrium of (2.51), if the
fertility is described by the Ricker model.
20 Move p + p to the right-hand side, move the last term on the right to the left, and
y z
divide through by 1 − pz . The right-hand side becomes m, and the given inequality is
obtained.
Difference Equations (Discrete Dynamical Systems) 89
1 a
Solution: We have yB(y) = aye−by , so that B(y) = ae−by . Now y∞ = b log m
exists for a > m, and the case tr A < 0 becomes
py + pz
a > m exp 1 + .
m
The general behavior of the models in the two examples above is similar to
that of the simpler Beverton-Holt and Ricker models analyzed in Section 2.4.
So in what way do these dispersal models give a better understanding of the
effects of dispersal on the stability of the underlying populations?
In order to answer this question, we should compare the behaviors of the
two two-patch dispersal models with the behaviors of the one-patch models
that most closely correspond to them: models with the given reproductive
functions, mortality ratios µ, and no dispersal (i.e., q = 0, so that p + µ = 1
and we consider the entire population as a whole, without separating those in
the lake or playa from those on land). For Beverton-Holt fertility, this is
ayk
yk+1 = + (1 − µ)yk ;
1 + byk
for Ricker fertility, it is
The analyses of these models are left as an exercise for the reader.
The one-
1 a
patch model with Beverton-Holt fertility has a survival equilibrium b µ − 1 ,
which exists and is stable when a/µ > 1. The corresponding two-patch model
replaces µ with m. The one-patch model with Ricker fertility has a survival
equilibrium 1b log µa which is stable for 1 < a/µ < e2/µ . The two-patch Ricker
model has a survival equilibrium with y∞ = 1b log m a
which is stable when
a > m and (2.54) holds. Therefore the effects of water-land dispersal for
90 Dynamical Systems for Biological Modeling: An Introduction
0 re−bz (1 − bz)
.
p 1−µ
the land patch before dying or returning to water is 1/(1 − pz ) in this model.
Difference Equations (Discrete Dynamical Systems) 91
0 r
,
p 1−µ
with trace 1 − µ and determinant −rp. The equilibrium is asymptotically
stable if 1 − µ < 1 − rp < 2. Thus the asymptotic stability condition is rp < µ,
which is just the condition that (0,0) is the only equilibrium. At the survival
equilibrium, the matrix is
0 µp (1 − bz)
,
p 1−µ
with trace 1−µ and determinant µ(bz −1). The asymptotic stability condition
is
1 − µ < µ(bz − 1) + 1 < 2.
The first inequality is satisfied automatically and the second condition is
µ(bz − 1) < 1. Using the equilibrium value of z, we reduce this to µ log rp
µ < 1.
rp 1/µ 1/µ
Taking exponentials, we have µ < e , or rp < µe . Since the existence
of this equilibrium requires rp > µ, we have an asymptotically stable survival
equilibrium if and only if
1
µ < rp < µe µ .
This result suggests that populations described by this model will tend to
reach a (more or less) constant level when the effective reproductive ratio rp
(the reproductive ratio multiplied by the proportion of larvae that survive to
reproductive age) outpaces natural mortality µ but not by too much. Popula-
tions undergoing intraspecific scramble competition which have an excessively
high growth ratio will tend to suffer fluctuations, just as observed for the
simple logistic model in Section 2.6.
In evaluating the utility of this model, we should consider what additional
detail it gives over the single-stage model with Ricker-function fertility and
proportional mortality described in the subsection on dispersal above. The
general behavior is the same; in the simpler model, the population tends to
the same equilibrium value (with a = rp) as long as the effective reproductive
ratio outpaces natural mortality µ, but not by too much. In the simpler model,
2
the ratio a/µ should fall in the interval (1, e µ ), while for the two-stage model
1
it should fall within the interval (1, e µ ) in order for the population density to
stabilize. This suggests that the added delay incurred by newborns requiring
a full generation of time to reach reproductive age makes it more difficult to
maintain stability; this conclusion is consistent with our previous observation
that delays (as represented by the discrete time steps implicit in difference
equations) make fluctuations more likely. The other difference, of course, is
that the two-stage model allows us to estimate the size of the larval population
relative to the adult population at any time.
Emmert and Allen used a similar model (with an additional term qyk in
the first equation denoting those juveniles who survive without having reached
92 Dynamical Systems for Biological Modeling: An Introduction
Exercises
In each of the following exercises, find all equilibria and determine which of
them are asymptotically stable.
22 K.E. Emmert and L.J.S. Allen, Population persistence and extinction in a discrete-
interaction, Insect Populations: in Theory and Practice (J.P. Dempster and I.F.G. McLean,
eds.), Chapman and Hall, London, 1998.
94 Dynamical Systems for Biological Modeling: An Introduction
bz ab b b
Because of the equilibrium conditions e−bz/y = e−ab , y2 = Be−ab , y = Be−ab ,
this matrix reduces to
ab −b
.
a 0
The trace of this matrix is ab and the determinant is also ab; thus the stability
condition is ab < ab + 1 < 2, or ab < 1. Thus the equilibrium is asymptotically
stable if and only if ab < 1 (note that ab is a dimensionless measure of the
amount by which the moth egg density reduces the ragwort biomass). Obser-
vations indicate both asymptotically stable equilibria and unstable equilibria
with oscillations are possible.
the host equation is yk+1 = ryk e−bzk . Since each parasitized host larva leads
to one adult parasitoid, the parasitoid equation is zk+1 = ryk (1 − e−bzk ). This
gives the Nicholson-Bailey model for host-parasitoid dynamics,24
re−bz −rbye−bz
.
r(1 − e−bz ) rbye−bz
with trace r and determinant zero. The stability conditions are r < 1 < 2,
or r < 1. Thus if r < 1, the extinction equilibrium is the only equilibrium
and it is asymptotically stable. At the survival equilibrium, the matrix of the
linearization is
1 −by∞
,
r − 1 by∞
with trace 1 + by∞ and determinant rby∞ . The stability conditions are 1 +
by∞ < 1 + rby∞ < 2 or by∞ < rby∞ < 1. The first of these conditions is
r
satisfied if and only if r > 1, and the second is satisfied if rby∞ = r−1 log r < 1.
However, this condition is not satisfied for any value of r > 1. Thus the survival
equilibrium of the Nicholson-Bailey model can not be asymptotically stable.
This means that either both host and parasitoid populations will be wiped
out (if r < 1) or both populations will oscillate. These oscillations may have
large amplitude and may bring one of the populations to a very low level from
which a small perturbation could wipe it out.
Thus this rather simple model supplies an explanation for outbreaks and
subsequent crashes of an insect population with a parasitoid. However, host-
parasitoid populations which persist for many generations have been observed,
24 A.J. Nicholson and V.A. Bailey, The balance of animal populations, Proc. Zoological
This model does allow for the possibility of an asymptotically stable survival
equilibrium, but the computations to show this are somewhat complicated
(see Exercise 5). The result is that most of the adult hosts must survive from
one generation to the next in order to overcome a high reproductive ratio r
and stabilize the equilibrium.
The biological systems described in this section share the property that the
equilibrium is unstable if the natural growth rate is too large. As discussed in
previous sections, many discrete models have this property. In the next chap-
ter, we shall consider continuous models, or ordinary differential equations,
which do not share this property. Discrete models have a built-in time-lag,
and this is what makes instability possible.
Exercises
In each of the following exercises, find all equilibria and determine which of
them are asymptotically stable.
1. Determine which of the systems in Exercises 1–6 of Section 2.7 represent
predator-prey systems, by considering yk and zk as two separate species
in each case, and examining the effect that each population has on the
other (via the sign of the associated coefficients).
2. In developing the model (2.57), van der Meijden, Crawley, and Nisbet
assumed that the number of new eggs laid is proportional only to the
biomass of ragwort, and not to the number of adult moths. What implicit
assumption is being made about the relevant factors? That is, why might
this be a reasonable assumption?
3. Suppose that the assumption discussed in the previous problem is not
valid, and that in fact the number of new eggs laid is proportional to
both the biomass of ragwort and the number of adult moths, i.e., zk+1 =
ayk zk . Find the equilibria of this system, and derive criteria for their
asymptotic stability.
4. Compare the behavior of the plant-herbivore model in the main text with
that of the modified system proposed in the previous problem. Explain
Difference Equations (Discrete Dynamical Systems) 97
(r + p)(1 − p) r
log < 1.
(r + p − 1) 1−p
Miscellaneous exercises
For each of the difference equations in Exercises 1–4, verify that the given
function is a solution of the given difference equation.
1. yk+1 = yk (1 − yk2 ), yk = −1
2. yk+1 = 1/yk , yk = c, (k odd), yk = 1/c (k even)
3. yk+1 = yk er(1−yk ) , yk = 1
4. yk+1 = yk /(1 + yk ), yk = 0
Solve each of the difference equations in Exercises 5–8 with initial value
y0 = c.
5. yk+1 = (k + 1)yk 7. yk+1 = yk /(k + 1)
k
6. yk+1 = 2 yk 8. yk+1 = yk + 1
For each of the difference equations in Exercises 9–12, find the equilibria, and
determine which equilibria are asymptotically stable.
98 Dynamical Systems for Biological Modeling: An Introduction
99
100 Dynamical Systems for Biological Modeling: An Introduction
Photo and illustration courtesy U.S. Fish and Wildlife Service, dls.fws.gov
our study to biological systems that can be modeled with first-order ordinary
differential equations.
We saw in the previous chapter that the simplest sort of change other
than a constant-valued derivative is one in which the rate of change is a
constant multiple of the quantity under study. This type of equation is called
linear, and, although many linear models are far too simple to capture the
complicated behavior of real biological systems, we also saw that the behavior
of more complex, nonlinear equations can be analyzed in terms of the behavior
of related linear equations. We thus begin our study of biological systems that
change continuously in time with a look at the simplest sort of first-order
ODE: the linear equation. Later sections will develop the tools necessary to
examine nonlinear models, and consider some of the types of models and
systems that have contributed most to our understanding of the biological
world, from chemical and neural functions within a single organism to the
growth and management of entire populations.
with the approximate equality (≈) signifying that the difference between the
two sides of the relation is small compared to h, in the sense that this difference
divided by h approaches zero as h → 0, i.e., that
y(t + h) − y(t) − ahy(t)
lim = 0. (3.3)
h→0 h
So far, this derivation is exactly the same as the one which began Section 2.1.
Now, however, instead of thinking of h as a fixed time interval, we allow h to
approach zero. It follows from (3.3) that
y(t + h) − y(t)
y ′ (t) = lim = ay(t).
h→0 h
Thus the population size y(t) satisfies the differential equation (3.1). If a frac-
tion b of the members reproduce by splitting and a fraction d of the members
die in unit time, then (3.2) would be true with a = b − d. The constant a may
be either positive or negative, depending on whether b > d or b < d. We will
use a to designate the constant of proportionality in (3.1); this constant may
be either positive or negative.
Although linear models may appear too simple to provide useful pre-
dictions, the hypotheses of constant per capita birth rates in the absence
of resource limitations, and constant per capita death rates, do have some
experimental support, even for relatively complex organisms. For instance,
Hutchinson cited2 the growth of the collared dove Streptopelia decaocto in
Great Britain, which it invaded in 1954 after spreading across much of west-
ern Europe. In the absence of initial competition among the first pioneer birds
to settle throughout Great Britain, the data (reproduced in Figure 3.2) show
a remarkably close fit to exponential growth (the figure is on a logarithmic
scale, making exponential curves appear linear) for several years, until pop-
ulation density slowed the growth. Hutchinson also observed a constant per
1 T.R. Malthus, An essay on the principle of population, Harmondsworth, Middlesex
Figure reprinted from G. Evelyn Hutchinson, Figure reprinted from G. Evelyn Hutchinson,
An introduction to population ecology, Yale An introduction to population ecology, Yale
University Press, 1978. Image copyright 1978 University Press, 1978. Image copyright 1978
Yale University Press, used with permission. Yale University Press, used with permission.
Example 1.
Suppose that a given population of protozoa develops according to a simple
growth law with a growth rate of 0.7944 per member per day, that there are
no deaths, and that on day zero the population consists of two members. Find
the population size after 6 days.
Solution: The population size satisfies the differential equation (3.1) with a =
0.7944, and is therefore given by y(t) = ce0.7944t . Since y(0) = 2, we substitute
t = 0, y = 2, and we obtain 2 = c. Thus the solution satisfying the given initial
condition is y(t) = 2 e0.7944t , and the population size after 6 days is y(6) =
2 e(0.7944)(6) = 235 (rounding the population size to the nearest integer).
Solution: The population size at time t satisfies y(t) = ceat and y(0) = 100,
y(100) = 150. Thus y(0) = 100 = ce0 , y(100) = ce100a = 150. It follows
that c = 100 and 150 = 100e100a . We obtain e100a = 1.5, a = log 1.5
100 =
−3
4.05465 × 10 . Finally, we obtain
−3
y(150) = 100 e150(4.05465×10 )
= 183.71.
Rounding off to the nearest integer, we obtain the population size 184 after
150 days.
First-Order Differential Equations (Continuous Dynamical Systems) 105
y ′ = −ky (3.7)
with k > 0.
Example 3.
The radioactive element strontium 90 has a decay constant 2.48 × 10−2
years−1. How long will it take for a quantity of strontium 90 to decrease
to half of its original mass?
Solution: The mass y(t) of strontium 90 at time t satisfies the differential
equation (3.7) with k = 2.48 × 10−2 . If we denote the mass at time t = 0 by
−2
y0 , then y(t) = y0 e−(2.48×10 )t . The value of t for which y(t) = y0 /2 is the
solution of
y0 −2
= y0 e−(2.48×10 )t .
2
If we divide both sides of this equation by y0 and then take natural logarithms,
we have
1
−(2.48 × 10−2 )t = log = − log 2
2
so that t = (log 2)/(2.48 × 10−2 ) = 27.9 years.
Example 4.
Radium 226 is known to have a half-life of 1620 years. Find the length of
time required for a sample of radium 226 to be reduced to three fourths of its
original size.
log 2
Solution: The decay constant for radium 226 is k = 1620 = 4.28×10−4 years−1 .
106 Dynamical Systems for Biological Modeling: An Introduction
3y0
= y0 e−kτ
4
3
or 4 = e−kτ . Taking natural logarithms we obtain −kτ = log 34 , which gives
log 43 1620(log 34 )
τ =− = = 672 years.
k log 2
Example 5.
Living tissue contains approximately 6 × 1010 atoms of carbon 14 per gram of
carbon. A wooden beam in an ancient Egyptian tomb from the First Dynasty
contained approximately 3.33 × 1010 atoms of carbon 14 per gram of carbon.
How old is the tomb?
Solution: The number of atoms of carbon 14 per gram of carbon, y(t), is given
by y(t) = y0 e−kt , with y0 = 6 × 1010 , k = 1.244 × 10−4, and y(t) = 3.33 × 1010
for this particular t value. Thus the age of the tomb is given by the solution
of the equation
−4 3.33 × 1010 3.33
e−(1.244×10 )t = = ,
6 × 1010 6
and if we take natural logarithms this reduces to
log 3.33 − log 6
t=− = 4733 years.
1.244 × 10−4
Exercises
In Exercises 1–6, assume the population size satisfies a simple growth law.
1. Suppose that the birth rate of a given population is 0.36 per member
per day with no deaths. If the population size on day zero is 50, what
is the population size 10 days later?
First-Order Differential Equations (Continuous Dynamical Systems) 107
2. Suppose that the birth rate of a given population of protozoa is 0.2 per
member per day with no deaths. If the population size on day zero is
10, find the population size 20 days later.
3. Suppose a population has 173 members at t = 0 and 262 members at
t = 10. Estimate the population size at t = 5.
4. Suppose that a population has 87 members at t = 0 and 125 members
at t = 4. Estimate the population size at t = 6.
5. Suppose that a population has 12 members at t = 3 and 5 members at
t = 10. What was the population size at t = 0?
6. Suppose that a population has 13 members at t = 4 and 20 members at
t = 6. What was the population size at t = 0?
7. If the half-life of a radioactive substance is 30 days, how long would it
take until 99 % of the substance decays?
8. How long does it take for a piece of carbon 14 to decrease to 20 % of its
original size?
9. In a sample of uranium 238, it is found that 0.0000154 % of the mass
disintegrates in 1000 years. Find the half-life of uranium 238.
10. How old is a fossil in which 85 % of the carbon 14 has disintegrated?
11∗ . Show that the solution of the initial value problem y ′ = ay, y(t0 ) = c is
y(t) = cea(t−t0 ) .
12∗ . Use the graph in Figure 3.2 to estimate (a) the per capita birth rate
of the collared turtledove, in units of per individual per year, and (b)
how long it would take for a population of this bird to double in an
environment with enough resources to prevent competition. Hint: What
form does log y have if dy/dt = ay? How can you read a from the graph?
13. Explain why the graphs of solutions to the linear differential equation
dy/dt = ry appear linear on a logarithmic scale. Calculate the slope of
such a line (supposing an initial condition y(0) = y0 ) and interpret it.
14. Suppose that a given population exhibits exponential growth when it is
small. How and why would you expect that to change when the popu-
lation is large? How would the graph of the population over time differ
from the graph of a purely exponential function?
108 Dynamical Systems for Biological Modeling: An Introduction
includes the constant solution y = K (with c = 0) but not the constant solution y = 0. The
existence and uniqueness theorem of Section 3.3 shows that, since we have now obtained a
solution corresponding to each possible initial condition, we have obtained all solutions of
the logistic differential equation.
110 Dynamical Systems for Biological Modeling: An Introduction
To find the solution which obeys the initial condition y(0) = y0 , we sub-
K
stitute t = 0, y = y0 into the form (3.12), obtaining 1+c = y0 which implies
K−y0
c = y0 as long as y0 6= 0 and gives the solution
K Ky0
y= = (3.13)
y0 + (K − y0 ) e−rt
K−y0
1+ y0 e−rt
to the initial value problem with y0 6= 0. Note that the denominator begins at
K (for t = 0) and moves toward y0 as t → ∞.
Now suppose that K represents some physical quantity such that K > 0.
One can see from the form (3.13) that if y0 > 0, then the solution y(t) exists
for all t > 0, and limt→∞ y(t) = K. If y0 < 0, then this solution does not
exist for all t > 0, because y(t) → −∞ wherethe denominator changes sign:
as y0 + (K − y0 )e−rt → 0, or t → log y0y−K 0
. If y0 = 0, the solution of the
initial value problem is not given by (3.13), but is the identically zero function
y = 0. If, instead, K < 0, as will occur in some examples presented in this
chapter, then the solution exists for all time if y0 < 0, but approaches positive
infinity in finite time if y0 > 0, for the same reason as above.
Note also that (for K > 0, y0 > 0) if r > 0 then y → K, while if r < 0 (as
also occurs for some examples in this chapter) then y → 0. In either case a
given solution approaches a limit monotonically, regardless of the magnitude
of r, in contrast to the behavior of the logistic difference equation of Chap-
ter 2, where more complicated behavior occurs for large values of the growth
constant r.
t (days) 8 9 10 11 12 13 14 15 16
y (observed) 50 76 69 51 57 70 53 59 57
y (logistic) 61 62 63 64 64 64 64 64 64
TABLE 3.2: Yeast population y vs. time t in hours, from Carlson (1913).
t 0 1 2 3 4 5 6 7 8 9
y(t) 9.6 18.3 29.0 47.2 71.1 119.1 174.6 257.3 350.7 441.0
t 10 11 12 13 14 15 16 17 18
y(t) 513.3 559.7 594.8 629.4 640.8 651.1 655.9 659.6 661.8
K − y0
z = log − rt,
y0
the graph of which (versus time) is a line with slope −r and y-intercept
log(K − y0 )/y0 . Thus a modified linear regression10 can provide estimates
7 T. Carlson, Über Geschwindigkeit und Grösse der Hefevermehrung in Würze, Bio-
in Mathematical Biology 11, Cambridge University Press, Cambridge, 1995, pp. 53–55.
10 In this case, the usual linear regression calculation must be modified slightly since the
data zn are functions of the parameter K (otherwise we must first fix a value for K and
then find the best r for that K). In searching for a minimum (with respect to K and r) of
112 Dynamical Systems for Biological Modeling: An Introduction
y[t] z[t]
4
600
500 2
400
t
2.5 5 7.5 10 12.5 15 17.5
300
200 -2
100
-4
t
2.5 5 7.5 10 12.5 15 17.5
FIGURE 3.4: Yeast culture growth FIGURE 3.5: Logarithmic (z) data
data from Carlson (1913) and its from Carlson (1913) and its best-fit
best-fit logistic curve. line.
for the parameters r and K (given y0 ). The best-fit values for the data in Ta-
ble 3.2 are K ≈ 664.5, r ≈ 0.539/hr. Graphs of y(t) and z(t) comparing data
and best-fit solution (similar to those in Renshaw, but using the parameter
values given here) are shown in Figures 3.4 and 3.5. The fit is clearly quite
good.
The logistic model has also been used with some success to fit data on
growth of larger, more complex organisms, as the following example illustrates.
Example 1.
Census figures (in millions) for the United States, P (t) in the table below, fit
the solution
265
y(t) =
1 + 69 e−0.03t
of the logistic differential equation, with t = 0 corresponding to the year 1790,
reasonably well. Predict the population in the years 1990 and 2000, and the
limiting population size.
substituting zn = log[(K − yn )/yn ] transforms the problem into one in which a computer
can search simultaneously for the best-fit K and r.
First-Order Differential Equations (Continuous Dynamical Systems) 113
Solution: The limiting population size as seen from the expression for y(t) is
265. The population size in 1990 (t = 200) would be predicted from P (t) as
265
= 226.3,
1 + 69 e−(0.03)(200)
and the prediction for 2000 (t = 210) would be
265
= 235.2.
1 + 69 e−(0.03)(210)
The actual population found in the 1990 census was approximately 250 million.
The actual population given in the 2000 census was about 281.4 million.
8000
2000
6000
1500
4000
1000
2000
500
year year
1860 1880 1900 1920 1940 1840 1860 1880 1900 1920
seems to be that the continuous logistic model often does a remarkably good
job of describing the rapid growth of even complex populations approaching
carrying capacity, but is of at best limited use in describing the growth of
populations at or near that capacity.
We finish this section with two extensions of the logistic model to multiple
populations, which provide a glimpse of the modeling of multi-population
biological systems undertaken in Chapters 5, 6, and 7.
Photo courtesy U.S. National Park Service, www.nps.gov Photo by Dave Powell, USDA Forestry Service
From our analysis earlier in this section, we know that (3.15) has the solution
m1
1− c1
p1 = ,
1+ Ce−(c1 −m1 )t
14 R. Levins, Some demographic and genetic consequences of environmental heterogeneity
for biological control, Bulletin of the Entomological Society of America 15: 237–240, 1969.
116 Dynamical Systems for Biological Modeling: An Introduction
where C = 1 − m c1 − p1 (0) /p1 (0), as long as p1 (0) 6= 0. Therefore, if c1 >
1
with intuition, in that the criterion for the species to survive is that it should
reproduce faster than it dies out.
Now suppose that there is another species, species 2, which competes with
species 1 for resources (and space) in the rainforest. Suppose that species 1
dominates species 2 in this competition, so that species 2 can only grow in
those sites not occupied by species 1. Then this inferior competitor colonizes
the proportion (1−p1 −p2 ) of sites occupied by neither the superior competitor
nor itself at a base rate of c2 p2 , and vacates sites due to natural mortality at
a rate m2 p2 , and due to displacement by the superior competitor at a rate
c1 p1 p2 (species 1 colonizes at a rate c1 p1 ; a proportion p2 of those sites are
occupied by species 2, causing displacement). These assumptions, together
with those on species 1, give us a system of equations
dp1
=c1 p1 (1 − p1 ) − m1 p1 ,
dt
dp2
=c2 p2 (1 − p1 − p2 ) − m2 p2 − c1 p1 p2 . (3.16)
dt
As species 1 remains unaffected by species 2, this system is said to decouple,
that is, we can analyze the growth of the two species separately. Species 1
will approach its carrying capacity p∗1 as discussed in the previous paragraph,
p1 → p∗1 , at which point (3.16) becomes
dp2
= c2 p2 (1 − p∗1 − p2 ) − m2 p2 − c1 p∗1 p2 ,
dt
which we can rewrite in explicit logistic form as
dp2 p2
= r2 p2 1 − ,
dt r2 /c2
where
c2 m
c1 − m2 − (c1 − m1 ), c1 > m1 ;
1
r2 = c2 − m2 − (c1 + c2 )p∗1 =
c2 − m 2 , c1 < m 1 .
The behavior of p2 thus again depends on the sign of the coefficient r2 , ap-
proaching r2 /c2 if r2 > 0 and 0 if r2 < 0. Under the (reasonable) assumption
that for each species i found in the rainforest, ci > mi , that is, that in the
absence of competition the species would persist, we take the corresponding
form for r2 and write the persistence condition for species 2 as
c1
c2 > (m2 + c1 − m1 ).
m1
This system can be extended to a complete hierarchy of n species, ranked
First-Order Differential Equations (Continuous Dynamical Systems) 117
Since the competition is completely ordered, each equation in the system can
be analyzed independently (beginning with species 1), rewritten as a logistic
equation, and shown to approach a unique equilibrium value, which is the
greater of 0 and
i−1
mi X ∗ cj
p∗i = 1 − − pj 1 + . (3.18)
ci j=1
ci
Note that the last term in the expression for p∗i represents the reduction in
distribution of species i due to displacement and exclusion by its superior
competitors. Tilman observed that depending on parameter values (i.e., a
species’s ability to colonize and survive), an inferior competitor may actually
occupy a greater proportion of sites than a superior competitor. For instance,
suppose species 1 has a high mortality rate, so that p∗1 is small. Then if species 2
has a low mortality rate,
m2 c1 m2
p∗2 = 1 − − p∗1 1 + ≈1− > p∗1 .
c2 c2 c2
This model can be extended to accommodate, for example, habitat de-
struction and fragmentation (see Exercises 8 and 9 below). Bampfylde, who
was interested in modeling competition in rainforests, observed that displace-
ment does not occur among rainforest species, so competition is purely for
sites opened through natural mortality. In this case, the displacement term in
(3.17) is omitted, and the colonization term is adjusted to exclude displace-
ment. ThusPthe colonization rate ci pi is reduced by the proportion of open
n
sites, (1 − j=1 pj ), rather than merely by the proportion of sites not occu-
Pi
pied by superior competitors, (1 − j=1 pj ). The resulting model does not
decouple, however, so we shall reserve its further study for Chapter 6.
epidemics, Part II, Proc. Royal Soc. London 138 (1932), 55–83.
First-Order Differential Equations (Continuous Dynamical Systems) 119
The differential equation (3.19) has the same form as the logistic equation
(3.8), with r replaced by βN − γ) and K replaced by βNβ−γ . It is important
to note that, as was true for the plant competition model studied earlier in
this section, in (3.19) it is possible for r = βN − γ to be either positive or
negative. Recalling the solution (3.12) of (3.8), we again observe that if r > 0
(and K > 0), then every solution y(t) with y(0) > 0 has limt→∞ y(t) = K,
while if r < 0, then every solution y(t) > 0 has limt→∞ y(t) = 0. If we translate
this result to (3.19), we see that if βN/γ < 1, then limt→∞ I(t) = 0, while if
βN/γ > 1, then limt→∞ I(t) = N − βγ > 0.
This result is the famous threshold theorem of Kermack and McKendrick:
If the quantity βN/γ, called the basic reproductive number, is less than 1,
then the number of infectives approaches zero, and the number of susceptibles
approaches N . In epidemiological terms this means that the infection dies
out. On the other hand, if the basic reproductive number βN/γ exceeds 1,
then the number of infectives remains positive and tends to a positive limit.
In epidemiological terms, this means that the infection remains endemic. The
quantity βN/γ represents the number of secondary infections caused by each
infective over the duration of the infection. Since βN is a number of contacts
per infective in unit time, the dimensions of βN are time−1 . Thus the basic
reproductive number βN/γ is dimensionless, and does not depend on the units
used in describing the model.
One infectious disease which has been studied mathematically by models
such as (3.19), as well as by considerably more detailed models, is gonorrhea.
For gonorrhea the average infective period (1/γ) is known to be about one
month. However, the transmission coefficient β must be estimated by some
indirect means. One way to estimate β is to use the fact that for small values
of I a solution of (3.19) grows exponentially with exponent r = (βN − γ).
Thus if an infection in some region begins with a small number of infectives,
the time t̂ for the number of infectives to double (ert̂ = 2) is approximately
log 2 log 2
= .
r βN − γ
In one reported gonorrhea outbreak this doubling time was approximately
1.7 months, which leads to the estimate βN/γ = 1.4, which indicates a basic
reproductive number of 1.4.
Example 2.
A new strain of influenza is introduced into a town with 1200 inhabitants by
two visitors. Assume that the average infective is in contact with 0.4 inhabi-
tants per day and that the average duration of the infective period is 6 days.
Will the infection die out or will the flu persist?
Solution: The number of potentially infective contacts per infective per day
is βN = 0.4 days−1 . Thus the basic reproductive number is βN/γ =
(0.4 days−1 )(6 days) = 2.4; since this exceeds 1, the flu will persist.
120 Dynamical Systems for Biological Modeling: An Introduction
There are other problems such as the spread of a rumor and rates of
chemical reactions which may also lead to a logistic differential equation model
(see Exercises 17 and 18 below, and Exercise 1 in Section 4.3).
Exercises
1∗ . Suppose that a population satisfies a logistic model (3.8). Show that its
growth rate y ′ is at a maximum when y = K 2 and that its maximum
rK ′
growth rate is 4 . Hint: Maximize y by setting
d y
(y ′ )′ = ry(1 − ) = 0.
dt K
2. The population of the United States in 1970 was 202.0 (millions) and the
per capita rate of growth was approximately 0.7% per year. Use these
data and an exponential growth model to predict the population in the
years 1990, 2000, and 2100.
(a) Show that in a population of fixed size K, such a disease will even-
tually spread to the entire population.
(b) Given the meaning in practical terms of “eventually,” what mod-
eling assumption made implicitly above accounts for the fact that
there always remain some individuals who never get infected, even
if the population remains closed and of fixed size, and infected in-
dividuals really never recover.
16∗ . Consider a disease spread by carriers who transmit the disease without
exhibiting symptoms themselves. Let C(t) be the number of carriers
and suppose that carriers are identified and isolated from contact with
others at a constant per capita rate α, so that C ′ = −αC. The rate
at which susceptibles become infected is proportional to the number of
carriers and to the number of susceptibles, so that S ′ = −βSC. Let C0
and S0 be the number of carriers and susceptibles, respectively, at the
time t = 0.
(a) Determine the number of carriers at time t from the first equation.
(b) Substitute the solution to part (a) into the second equation and
determine the number of susceptibles at time t.
(c) Find the number of members of the population who escape the
disease, limt→∞ S(t).
17∗ . Consider a population of fixed size K in which a rumor is being spread
by word of mouth. Let y(t) be the number of people who have heard
the rumor at time t and assume that everyone who has heard the rumor
passes it on to r others in unit time. Thus from time t to time t + h the
rumor is passed on rh y(t) times, but a fraction y(t)
K of the people
who
hear it have already heard it and thus there are only rh y(t) K−y(t)
K
people who hear the rumor for the first time. Use these assumptions to
obtain an expression for y(t + h) − y(t), divide by h and take the limit
as t → ∞ to obtain a differential equation satisfied by y(t).
18. At 9 AM, 1 person in a village of 100 inhabitants has heard a rumor.
Suppose that everyone who has heard the rumor tells one other person
per hour. Using the model of Exercise 17, determine how long it will
take until half the village has heard the rumor.
First-Order Differential Equations (Continuous Dynamical Systems) 123
3.3.1 Solutions
Recall that by a differential equation we mean simply a relation between
an unknown function and its derivatives. The general form of a first-order
differential equation is
dy
y′ = = f (t, y), (3.20)
dt
with f a given function of the two variables t and y. By a solution of the
differential equation (3.20) we mean a differentiable function y of t on some
t-interval I such that, for every t in the interval I,
In other words, differentiating the function y(t) results in the function f (t, y).
For example, as we saw in Example 1 of Section 3.1, the function y = 2eat is a
solution of the differential equation y ′ = ay on every t-interval. We see this by
differentiating 2eat to get a 2eat , which we can rewrite as ay. To verify whether
a given function is a solution of a given differential equation, we need only
substitute into the differential equation and check whether it then reduces to
an identity.
Example 1.
1
Show that the function y = t+1 is a solution of the differential equation
y ′ = −y 2 .
Solution: For the given function,
dy 1
=− = −y 2
dt (t + 1)2
and this shows that it is indeed a solution.
124 Dynamical Systems for Biological Modeling: An Introduction
In the same way we can verify that a family of functions satisfies a given
differential equation. By a family of functions we mean a function which in-
cludes an arbitrary constant, so that each value of the constant defines a
distinct function. √ The family ce5t , for instance, includes the functions e5t ,
5t 5t
−4e , 12e , and 3 e5t , among others. When we say that a family of func-
tions satisfies a differential equation, we mean that substitution of the family
(i.e., the general form) into the differential equation gives an identity satisfied
for every choice of the constant.
Example 2.
1
Show that for every c the function y = t+c is a solution of the differential
′ 2
equation y = −y . (This equation models a decaying population whose per
capita mortality rate is linear in population size, rather than constant.)
Solution: For the given function,
dy 1
=− = −y 2 ,
dt (t + c)2
and this shows that each member of the given family of functions is a solution.
Example 3.
Show that the family of functions
1 + cet
y=
1 − cet
is a solution of the differential equation
y2 − 1
y′ =
2
for every value of the constant c.
Solution: For the given function y,
y2 − 1 1 (1 + cet )2 − (1 − cet )2
= ,
2 2 (1 − cet )2
and algebraic simplification shows that these two expressions are equal.
through the point (t0 , y0 ) in the t-y plane. Physically, this corresponds to mea-
suring the state of a system at the time t0 and using the solution of the initial
value problem to predict the future behavior of the system.
Example 4.
1
Find the solution of the differential equation y ′ = −y 2 of the form y = t+c
which satisfies the initial condition y(0) = 1.
1
Solution: We substitute the values t = 0, y = 1 into the equation y = t+c ,
1
and we obtain a condition on c, namely 1 = c , whose solution is c = 1.
The required solution is the function in the given family with c = 1, namely
1
y = t+1 .
Example 5.
Find the solution of the differential equation y ′ = −y 2 which satisfies the
general initial condition y(0) = y0 , where y0 is arbitrary.
1
Solution: We substitute the values t = 0, y = y0 into the equation y = t+c and
1 1
solve the resulting equation y0 = c for c, obtaining c = y0 provided y0 6= 0.
Thus the solution of the initial value problem is
1
y=
(t + y10 )2
f is reasonably smooth. We will state this result and ask the reader to ac-
cept it without proof because the proof requires more advanced mathematical
knowledge than we have at present.
Even though the function f (t, y) may be well-behaved in the whole t-y
plane, there is no assurance that a solution will be defined for all t. As we
1
have seen in Example 1, the solution y = t+1 of y ′ = −y 2 , y(0) = 1 exists only
for −1 < t < ∞. As we have seen in Example 2, each solution of the family
1
of solutions y = t+c has a different interval of existence. In Example 5 we
have shown how to rewrite a family of solutions for a differential equation in
terms of an arbitrary initial condition — that is, as a solution of an initial value
problem for that differential equation. We have also seen how to identify those
initial conditions which cannot be satisfied by a member of the given family.
Often there are constant functions which are not members of the given family
but which are solutions and satisfy initial conditions that cannot be satisfied
by a member of the family. The existence and uniqueness theorem tells us that
if we can find a family of solutions, possibly supplemented by some additional
solutions, so that we can find this collection contains a solution corresponding
to each possible initial condition, then we have found the set of all solutions
of the differential equation.
There are computer programs, both self-contained and portions of more elab-
orate computational systems such as Mathematica, MATLAB and Maple,
which can generate direction fields for a differential equation and can also
sketch solution curves corresponding to these direction fields. We give some
examples here which have been produced by Maple (see Appendix A); the
reader with access to a facility which is capable of drawing direction fields is
urged to reproduce these examples before trying to produce other direction
fields.
Example 6.
Draw a direction field and some solutions of the differential equation y ′ = y.
(This is the prototypical linear homogeneous differential equation, with the
time unit rescaled so that the growth rate a is equal to 1; see Section 3.1 for
applications.)
Solution: See Figure 3.9. The direction field suggests exponential solutions,
which we know from Section 3.1 to be correct.
4 1
0.8
3
0.6
y(t) 2 y(t)
0.4
1
0.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1
t t
–0.2
–1
–0.4
–2
FIGURE 3.9: Direction field and FIGURE 3.10: Direction field and
solutions for y ′ = y. solutions for y ′ = −y 2 .
Example 7.
Draw a direction field and some solutions of the differential equation y ′ = −y 2 .
Solution: See Figure 3.10. The direction field indicates that solutions below
the t-axis become unbounded below while solutions above the t-axis tend to
zero as t → ∞, as could be seen from Example 2.
128 Dynamical Systems for Biological Modeling: An Introduction
Example 8.
Draw a direction field and some solutions of the differential equation y ′ =
y(1 − y). This is the prototypical logistic differential equation, in which both
time and population units have been rescaled so that r and K are both equal
to 1.
Solution: See Figure 3.11. The direction field indicates that solutions below
the t-axis are unbounded below while solutions above the t-axis tend to 1 as
t → ∞, consistent with what we have established for this logistic differential
equation.
1.5 1.5
1 1
y(t) y(t)
0.5 0.5
–0.5 –0.5
–1 –1
FIGURE 3.11: Direction field and FIGURE 3.12: Direction field and
solutions for y ′ = y(1 − y). solutions for y ′ = 1 − t2 − y 2 .
Example 9.
The differential equation y ′ = 1 − y 2 is a slight variation on the logistic
equation y ′ = y − y 2 ; the important difference between the two for purposes
of modeling biological systems is that when y = 0 the new model has y ′ = 1,
rather than y ′ = 0 as the logistic model has. This allows the population to
grow to its asymptotic value of 1 even when measurement errors in the initial
population size, which may be very small (given the rescaling of units so that
K = 1), yield a value of 0 or even a small negative number. Although the new
model is not quite biologically sound at y = 0, it may be used with the given
caveat.
Suppose now that we wish to incorporate an additional feature into the
model: namely, an eventual decline in population once it has reached its carry-
ing capacity (as mentioned in Section 3.2, such declines have been observed fol-
First-Order Differential Equations (Continuous Dynamical Systems) 129
lowing the initial growth period). A simple modification might involve adding
a negative term dependent on t, say −t2 , to y ′ . Draw a direction field and
some solutions of the differential equation y ′ = 1 − y 2 − t2 , to evaluate the
effects of this modification.
Solution: See Figure 3.12. It can be seen that regardless of initial conditions
(i.e., above or below the carrying capacity of 1), solutions eventually grow
negative and unbounded. The −t2 term does produce an eventual decline
from y = 1, but the decline is too strong; the term needs to be modified to
restrict its maximum size, so that solutions remain positive.
As the reader can see, direction fields provide a useful visual indication
of system behavior for a range of initial conditions. Although the solutions
obtained by connecting the segments are approximate rather than exact, and
approximate numerical solutions for specific initial conditions are found more
commonly via numerical analysis, direction fields are useful tools in suggesting
qualitative behavior that varies by initial condition in cases when it may be
difficult to produce a complete analysis using the usual qualitative tools (which
will be developed in Section 3.5).
The geometric view of differential equations presented by the direction field
will appear again when we examine some qualitative properties in Section 4.1,
and is also the basis of the numerical approximation of solutions in Section 4.5.
Exercises
1
1. Show that y = (1 − t2 )− 2 is a solution of the differential equation y ′ =
ty 3 .
2. Show that y = 3t is a solution of the differential equation y ′ = y/t.
1
3. Show that y = (c−t2 )− 2 is a solution of the differential equation y ′ = ty 3
for every choice of the constant c.
4. Show that y = ct is a solution of the differential equation y ′ = y/t for
every choice of the constant c.
1
5. Among the family of solutions y = (c − t2 )− 2 of y ′ = ty 3 , find the
solution such that y(0) = 1.
6. Find the solution of y ′ = ty 3 such that y(0) = 0.
7. Show that the solution of the initial value problem y ′ = ty 3 , y(0) = −1
1
is y = −(1 − t2 )− 2 .
t
1+ce ′ 1 2
8. Among the solutions y = 1−ce t of the differential equation y = 2 (y −1),
find the ones which satisfy the initial conditions y(0) = 1, y(0) = 0, and
y(2) = 0.
130 Dynamical Systems for Biological Modeling: An Introduction
9. Show that if ŷ is a constant such that f (t, ŷ) = 0 for all t, then y = ŷ is
a constant solution of y ′ = f (t, y).
10∗ . * If y(t) is a solution of y ′ = f (t, y) which has a local maximum at the
point (t̂, ŷ), show that f (t̂, ŷ) = 0.
11. Draw a direction field and some solutions of the differential equation
y ′ = ty.
12. Draw a direction field and some solutions of the differential equation
y ′ = t2 + y 2 .
13. Draw a direction field and some solutions of the differential equation
y′ = 1 + y2.
14. Draw a direction field and some solutions of the differential equation
y ′ = y − t.
p
15∗ . For the differential equation y ′ = −t + 2y + t2 ,
(a) Show that y = ct+ 12 c2 is a solution for every choice of the constant
c.
2
(b) Show that y = − t2 is a solution.
2 2
(c) Find the tangent line to the curve y = − t2 at the point (−c, − c2 ).
16∗ . Improve the modification to the model y ′ = 1 − y 2 presented in Example
9 so that it does not cause all solutions to grow unbounded and negative.
that is, if the changes in population size due to population density y can be
First-Order Differential Equations (Continuous Dynamical Systems) 131
factored out from those due to the time index t (such as seasonal factors). For
y
example, y ′ = 1+t ′
2 and y = y
2
are both separable, whereas y ′ = sin t − 2ty is
not separable. The examples discussed in Section 3.1 are also separable, and
the method of solution described for equation (3.1) of Section 3.1 is a special
case of the general method of solution to be developed for separable equations.
The reason for the name separable is that the differential equation can then
be written as
y′
= p(t),
q(y)
with all the dependence on y on the left and all the dependence on t on the
right-hand side of the equation, provided that q(y) 6= 0 (indeed, this is the
first step in the method). An important special case, which includes most of
the examples in this chapter, is when f is completely independent of t and is
a function of y only (i.e., p(t) = 1); in this case, the equation is said to be
autonomous. The method of separation of variables applies to many relatively
complex differential equations, and is worth learning for this reason.
Although the models and systems explored in this section are thus united
by a mathematical, rather than biological, feature, two classes of models which
allow separation of variables are the linear and logistic differential equations,
which form the basis for many models of biological systems. We shall begin
our discussion with some applications of these models, and continue with some
other, more specialized separable models.
Physiology 59: 734–766, 1972, and D.S. Jones and B.D. Sleeman, Differential equations and
mathematical biology, Boca Raton, FL: CRC Press, 2003.
132 Dynamical Systems for Biological Modeling: An Introduction
membranes open and close to change the electrical potential via an exchange
of particular ions.
Single autonomous differential equations typically do not lead to sustained
oscillations like a heartbeat, but we can obtain periodic behavior by intro-
ducing a control variable u that is sensitive to the cellular voltage levels,
and switches on and off when the voltage crosses set thresholds. In this way
we may think of the control as a regulating force like that of an artificial
pacemaker (although such devices’ controls are typically calibrated to times,
rather than voltages). We shall assume, then, that we know four voltage lev-
els: the control values uL and uH and the threshold values yL and yH , with
uL < yL < yH < uH . We shall also assume that the control signal u works as
follows: when the cellular voltage level is low, y(t) ≤ yL , the control switches
on, u = uH ; when the voltage level rises above the upper threshold, y(t) ≥ yH ,
the neuron fires and the control switches off, u = uL . While the voltage level is
rising or falling between the two thresholds, yL < y < yH , the control remains
at its present level. Finally, we shall assume that, apart from the influence
of the control signal, the pacemaker signal y(t) naturally decays at a rate
proportional to its present level.
These assumptions lead to a linear heterogeneous model for the pacemaker
signal voltage,
y ′ = au − ay = a(u − y), (3.22)
where a > 0 is a rate constant which determines how quickly the signal decays
from its previous values (such constants are sometimes said to measure the
“forgetfulness” of the system). Although the control signal u may change from
time to time as the voltage y crosses the upper or lower threshold, it is said to
be piecewise constant—that is, constant on short periods of time—and so we
will study how this system behaves on these short periods, later connecting
the solutions during each period to form a long-term picture.
Equation (3.22) is separable, and we shall solve it using the method of sep-
aration of variables, which we will explain more generally later in this section.
(Another method for solving linear differential equations will be presented
in Section 3.5.) To separate variables, we first divide both sides of (3.22) by
a(u − y), which is permissible as long as y 6= u (and remember that the model,
if successful, will constrain y within [yL , yH ], so that it never reaches uL or
uH ). This yields
1 dy
= 1.
a(u − y) dt
Integrating both sides of this equation with respect to t, considering y as a
function y(t) of t,
1 dy
Z Z
dt = 1 dt,
a(u − y) dt
we simplify to
dy
Z Z
= dt. (3.23)
a(u − y)
First-Order Differential Equations (Continuous Dynamical Systems) 133
log |y − u| = −a(t + c)
y = u + Ce−at . (3.24)
Note that although our separation of variables explicitly excluded the constant
solution y = u, this solution is contained in the family (3.24).
In order to find the solution of (3.22) which satisfies the initial condition
y(0) = y0 (3.25)
If we now set time to zero at the moment when the control switches on,
134 Dynamical Systems for Biological Modeling: An Introduction
sinoatrial
(SA) node y(t)
uH
yH
yL
uL
bundle t
of His 0 T1 T1+T
2
u(t)
uH
uL
Purkinje
fibers t
atrioventricular
(AV) node 0 T1 T1+T
2
Illustration by Carolyn Kribs
FIGURE 3.13: The cardiac electri- FIGURE 3.14: Pacemaker and con-
cal system. trol voltage over one period.
Thus the voltage rises from yL toward uH , until it crosses the upper threshold
yH , at which point the neuron fires and the control resets to uL . At this
moment we have y(T1 ) = yH , which we use as a new initial condition in (3.26)
to obtain a description of the voltage as it falls back toward uL ,
we can determine the remaining voltage parameter and the rate constant a.
Conversely, if we know the four voltage parameters and a, we can determine
T1 and T2 . These relationships come from the continuity of the voltage y at
the points where the control signal changes:
y(T1 ) = uH + (yL − uH )e−aT1 = yH (3.27)
and
y(T1 + T2 ) = uL + (yH − uL )e−a(T2 ) = yL . (3.28)
Experimental observations have shown that the voltage level ranges from
yL = −70 mV to yH = +30 mV during one period, which lasts about one
second. If observations show T1 ≈ 600 msec and T2 ≈ 400 msec, and we
suppose a lower control voltage of uL = −100 mV, then we can calculate
yH − u H yL − u L
e−aT1 = , e−aT2 = ,
yL − u H yH − u L
so that
yL − u H
aT1 = log , (3.29)
yH − u H
yH − u L
aT2 = log . (3.30)
yL − u L
To find the rate constant a, we substitute into (3.30):
30 mV − (−100 mV) 130
a(400 msec) = log = log ,
−70 mV − (−100 mV) 30
and find that a = 0.003666 msec−1 ≈ 0.004 msec−1 , or 4 sec−1 . To find uH , we
can then substitute into (3.29) and solve, to find uH = +42.5 mV ≈ +40 mV.
While this model uses an external control variable rather than incorporat-
ing details of the ion channel biochemistry to generate periodic behavior, the
exercise above does illustrate the utility of even simple models, as well as the
procedure of separation by variables. Before considering further applications,
we shall next give a general description of this procedure.
1 dy
= p(t),
q(y(t)) dt
on any interval on which q(y(t)) 6= 0, and this shows that y(t) satisfies the
differential equation (3.31). There may be constant solutions of (3.31) which
are not included in the family (3.33). A constant solution to the original
equation (3.31) corresponds to a solution of the equation q(y) = 0.
If we wish to solve an initial value problem and find the particular solution
of equation (3.31) which satisfies the initial condition y(t0 ) = y0 , then instead
we should make the indefinite integrals in equation (3.32) definite, integrating
forward from the initial condition (t0 , y0 ):
Z y Z t
du
= p(s) ds. (3.34)
y0 q(u) t0
Example 1.
Find all solutions of the differential equation y ′ = −y 2 .
Solution: We divide the equation by y 2 , permissible if y 6= 0, to give
1 dy
= −1.
y 2 dt
We then integrate both sides of this equation with respect to t. In order to
integrate the left side of the equation we must make the substitution y = y(t),
where y(t) is the (as yet unknown) solution. This substitution gives
1 dy dy 1
Z Z
2
dt = 2
=− +c
y dt y y
and we obtain
1
− = −t − c (3.35)
y
1
or (since Q(y) = −1/y is easily inverted here) y = t+c , with c a constant of
integration. Observe that no matter what value of c is chosen this solution
is never equal to zero. We began by dividing the equation by y 2 , which was
legitimate provided y 6= 0. If y = 0, we cannot divide, but the constant
function y = 0 is a solution. We now have all the solutions of the differential
equation, namely the family (3.35) together with the zero solution.
138 Dynamical Systems for Biological Modeling: An Introduction
Example 2.
Solve the initial value problem
y ′ = −y 2 , y(0) = 1.
1
Solution: The results of Example 1 give the family of solutions y = t+c for
the differential equation y ′ = −y 2 . When we substitute t = 0 and y = 1, we
1
get 1 = 0+c = 1c , which implies that c = 1. Thus the solution to the initial
1
value problem is y = t+1 , defined for all t ≥ 0.
Alternatively, separating variables and integrating as before yields
dy
Z Z
= −1 dt + c,
y2
or, in its definite form using the initial condition (t, y) = (0, 1),
Z y Z t
du
2
= −1 ds.
1 u 0
Example 3.
Solve the initial value problem
y ′ = −y 2 , y(0) = 0.
Example 4.
Solve the initial value problem
y ′ = ty 3 , y(0) = 1.
1 t2
− 2
= + c.
2y 2
The initial condition y(0) = 1 gives c = − 12 , which gives the solution implicitly
by
1
− 2 = t2 − 1.
y
In this example it is easy to find the solution explicitly. We have y 2 = (1 −
t2 )−1 . Since y(0) = 1, we must take the positive square root and thus we
obtain the solution y = (1 − t2 )−1/2 defined on the interval −1 < t < 1.
Example 5.
In Example 1 of Section 3.2 we saw that the population of the United States
can be described reasonably well by a logistic growth curve with intrinsic
(maximum) growth rate r = −0.03/yr and carrying capacity K = 265 (million
people). However, in the year 2000 the U.S. population was estimated at 281.4
million. Adapt the model to account for the observed population exceeding
the apparent carrying capacity.
Solution: One possible explanation is an increased carrying capacity due to
improved agricultural techniques (or the converting of forest land into farm-
land). In practice, such changes are continuous and ongoing, but making K a
continuously varying function of time keeps the model from being separable.
A simple way to incorporate an increase in K is to choose a single year in
which to consider the carrying capacity to have risen. Since, from the data
given in Example 1 of Section 3.2, 1970 was the first since 1910 in which the
real population surpassed the model prediction, let us choose 1970 as our new
First-Order Differential Equations (Continuous Dynamical Systems) 141
e−µ0 T u(0)
u(T ) = µ1 .
µ0 (1 − e−µ0 T )u(0) + 1
We now substitute the initial condition u(0) = ryk and define new constants
First-Order Differential Equations (Continuous Dynamical Systems) 143
of multiple processes on local abundance: a cohort approach for open populations, Ecology
Letters 2: 294–303, 1999.
19 J. Dumas and P. Prouzet, Variability of demographic parameters and population dy-
namics of Atlantic salmon Salmo salar L. in a south-west French river, ICES Journal of
Marine Science 60(2): 356–370, April 2003.
20 Sukgeun Jung and Edward D. Houde, Recruitment and spawning-stock biomass dis-
tribution of bay anchovy (Anchoa mitchilli) in Chesapeake Bay, Fishery Bulletin 102(1):
63–77, 2004.
144 Dynamical Systems for Biological Modeling: An Introduction
3.4.5 Allometry
Let x(t) and y(t) be the sizes of two different organs or parts of an indi-
vidual at time t. There is considerable empirical evidence to suggest that the
dy
relative growth rates dx
dt /x and dt /y are proportional. This means that there
is a constant k, depending on the nature of the organs, such that
1 dy 1 dx
=k . (3.41)
y dt x dt
The relation (3.41) is called the allometric law, and the identification of such
differential growth ratios is called allometry (the scientist Julian Huxley re-
ferred to it as heterogony). The single equation (3.41) does not provide enough
information to determine x and y as functions of t, but we can eliminate t and
obtain a relation between x and y. If we consider y as a function of x, then
according to the chain rule of calculus,
dy dy . dx
= . (3.42)
dx dt dt
Combining (3.41) and (3.42) we have
dy y
=k ,
dx x
R R
which can be solved by separation of variables dy/y = k dx/x to give
and finally
y = ek log x ea = xk ea = cxk . (3.44)
In order to determine the constant c, we need the values of x and y at some
starting time.
In experiments, one might measure both x and y at various times and
then use these measurements to plot log y against log x, that is, to plot the
experimental data on logarithmically scaled graph paper. According to the
relation (3.43), the graph should be a straight line with slope k and y-intercept
a = log c. Because of experimental error, the points may not line up perfectly,
but it should be possible to draw a line fitting the data well. One can then
measure the slope and y-intercept of the line to determine the constants k and
c in the relation (3.44).
First-Order Differential Equations (Continuous Dynamical Systems) 145
Example 6.
Thompson’s classic work On Growth and Form21 reports the following data
from Huxley for the weights of the claw and body of the fiddler crab Uca
pugnax:
Use these data to determine the constants c and k in the allometric law
y = cxk describing the relation between claw mass x and body mass y.
Solution: Transforming the data, we have:
log x 1.61 2.20 2.64 3.22 3.64 3.97 4.08 4.36 4.65 4.91 5.11 5.28
log y 4.06 4.38 4.69 5.05 5.30 5.47 5.60 5.70 5.87 6.04 6.15 6.28
5 400
4 300
3
200
2
100
1
x
1 2 3 4 5 6 50 100 150 200
FIGURE 3.15: Log-log plot of Hux- FIGURE 3.16: Linear plot of Hux-
ley’s allometric data, and the line of ley’s allometric data, and the corre-
best fit. sponding curve.
A caveat: Thompson warned that allometric laws hold only during periods
of rapid growth, and claimed that simpler relationships are likely to hold at
other times (see Exercise 24 for Thompson’s discussion of this with regard
to the fiddler crab). It is certainly true that all models have limitations; this
one is similar to those we observed in earlier sections for linear and logistic
models.
21 D’Arcy Wentworth Thompson, On growth and form, Vol. I, 2nd ed., Cambridge: Cam-
Exercises
In Exercises 1–8 below, find all solutions of the given differential equation. Do
not overlook constant solutions.
1. y ′ = t2 y 5. y ′ = −2ty
2. y ′ = 7y 3 cos t y2
6. y ′ = 1−t
t
3. y ′ = y 7. y ′ = t2 y − 4t2
y+1
4. y ′ = ty 2 8. y ′ = t
In Exercises 9–16 below, find the solution of the given differential equation
which satisfies the given initial condition.
9. y ′ = t2 y, y(5) = 1 13. y ′ = ty 2 , y(0) = −2
2
10. y ′ = 7y 3 cos t, y(1) = 0 y
14. y ′ = 1−t , y(−1) = log1 2
17. We can rewrite the pacemaker model (3.22) to emphasize how the pace-
maker signal y follows the control u: y ′ = −a(y−u). Since we are suppos-
ing the control signal u to be piecewise constant, then (y − u)′ = y ′ , and
we can make a substitution of variables, as is sometimes done in evalu-
ating integrals in calculus courses, to rewrite this equation in terms of
the difference z = y − u: (y − u)′ = −a(y − u) becomes
z ′ = −az. (3.45)
(b) Given the nature of the solution y(t), what will happen as yH ap-
proaches uH ? (Hint: Consider equation (3.29).)
(c) If the control signal becomes sufficiently desensitized, it may be-
come necessary to introduce an external pacemaker which drives
the control based explicitly on timing rather than on values of y.
That is, if the neuron does not fire within a certain amount of
time after last resetting, the artificial pacemaker forces the signal
to switch. This has the effect of controlling T1 and T2 directly.
If an external pacemaker of this type is applied to a heart with the
parameter values given in this section, and the maximum period
T1 allowed by the external pacemaker before causing the control
to switch off is 800 msec, then how high can the threshold yH rise
before the external pacemaker activates?
19. In the pacemaker model of this section, what happens to the firing times
T1 and T2 when uH − yH = yL − uL ?
20. In the logistic model, suppose that we have a population whose repro-
ductive rate undergoes seasonal fluctuations—many species of animal
reproduce only at certain times of year. This has the effect of making
the growth rate r a function of time. If we let t = 0 correspond to Jan-
uary and suppose that most reproduction occurs in June or July, then
we might model the growth rate as
tions. Fishery Investigations Series 2 (19). London: Great Britain Ministry of Agriculture.
148 Dynamical Systems for Biological Modeling: An Introduction
(a) Determine the constants in the allometric law that best fit these
additional data.
(b) Now try instead a simple linear fit, of the form y = mx + b.
(c) Which of the two models above (allometric or linear) appears to be
a better fit to these data?
(d) Repeat the analysis in parts (a)–(c) above, using the entire data set
(24 points). Which model is the better fit for the entire data set?
How would a model defined piecewise (allometric for part of the
data set, linear for the rest) compare to the two one-piece models?
25. As in Exercise 24, determine, evaluate and compare the allometric and
linear models for the following data on the stag beetle Lucanus cervus
from Huxley, republished by Thompson (p. 208).
Length of mandible (mm) 6.0 7.8 9.0 10.0 11.2 11.9 12.8 14.4
Total length (mm) 31.0 38.8 40.5 42.6 45.0 46.9 49.2 53.6
26. As in Exercise 24, determine, evaluate and compare the allometric and
linear models for the following data on the reindeer beetle Cyclommatus
tarandus from Huxley, republished by Thompson (p. 209).
Length of mandible (mm) 3.9 10.7 14.1 19.9 24.0 30.7 34.5
Total length (mm) 20.4 33.1 38.4 47.3 54.2 66.1 74.0
SYSTEM
inflow outflow
uniform mixing
- -
........
rate b rate ay
population y(t)
equations24 ; here we will introduce one which will also be of use when we
consider the more general first-order equation in Section 3.6, in which the
coefficients a and b may be functions of the independent variable t (that is,
they may change over time).
The technique of an integrating factor involves using the product rule for
derivatives in reverse, much like integration by parts. We begin by writing
(3.46) as
dy
+ ay = b
dt
and then look for a way to rewrite the left-hand side as an exact deriva-
tive: that is, the two-term derivative of a product of functions. If we multiply
through by a function g(t), then the left-hand side becomes gy ′ + agy. The
derivative of the product (gy) is gy ′ + g ′ y, Rso we have a match if g ′ = ag. Our
integrating factor g should therefore be e a dt (making g ′ = ag), which for
constant a reduces to eat . Now multiplying the original equation by g yields
dy
eat + aeat y = beat ,
dt
which by the above line of reasoning we can rewrite as
d at
e y = beat . (3.47)
dt
From this point we can apply either an indefinite or definite integral. In the
first case, we get Z
eat y = beat dt;
b at
for a and b constant, the right-hand side simplifies to ae plus an arbitrary
constant of integration C, which gives us the solution
b
y= + Ce−at . (3.48)
a
If we have the initial condition y(0) = y0 , then the definite integral of (3.47)
becomes Z t Z t
d aτ
(e y(τ )) dτ = beaτ dτ
0 dτ 0
b at
eat y(t) − e0 y(0) = e − e0 ;
a
that is,
b
y(t) = y0 e−at + 1 − e−at ,
(3.49)
a
24 Including separation of variables, cf. the cardiac pacemaker model of section 3.4.
152 Dynamical Systems for Biological Modeling: An Introduction
a time-weighted average of the initial value y0 and the final (asymptotic) value
b/a. (In terms of (3.48), then, C = y0 − ab .)
In the discussions of related models that follow, we shall take as given the
solution (3.48) to the differential equation (3.46), and the solution (3.49) to the
corresponding initial value problem. While we wish to point out the breadth
and variety of examples, our main goal is to encourage some understanding
of the modeling process in areas of interest to the reader.
3.5.1 Chemostats
The chemostat is a laboratory device proposed in 1950 by Novick and
Szilard25 and Monod26 to reproduce classic continuous linear diffusion in cul-
tures of bacteria, and has since become a regular fixture in many biology
laboratories. It consists of a large container, the chemostat reactor, to which
are connected inflow and outflow tubes (connected, respectively, to a nutrient
supply tank and an outflow collector), as well as other instruments which help
maintain constant conditions inside the reactor. The underlying assumption
governing chemostat populations is that growth is limited by one key nutri-
ent (the rest being present in abundant supply). The chemostat is a physical
model of a natural biological system in much the same way as a differential
equation such as (3.46) is a mathematical model of a biological system. In ad-
dition to its use as a constant supply of bacterial cultures in the lab, it is used
to approximate and study the workings of systems from populations of cells
to entire ecosystems, especially where field and in vivo studies are impossible.
Figure 3.18 shows a typical setup.
Many chemostat models are complex, incorporating multiple populations
or substances within the reactor (for example, nutrient supply and the or-
ganisms that consume them), multiple spatial layers within the reactor, etc.,
but here we shall consider the simplest possible model, in which we track a
single substance or population within a system. For example, Carman and
Woodburn27 used a chemostat to study the impact of low levels of antibiotics
used in food-producing animals on human intestinal flora: in particular, the
extent to which low levels of the common antibiotic ciprofloxacin may induce
resistance in intestinal bacteria. Such drugs are commonly used in food ani-
mals, but residues may remain in meat ingested by consumers in levels low
enough to produce resistance (to the same or related drugs) in gut bacteria.
Carman and Woodburn used a chemostat containing 500 mL of culture, with
a nutrient inflow of 35 mL/hr. If we assume a constant balancing outflow of
25 Aaron Novick and Leo Szilard, Description of the chemostat, Science 112: 715–716,
1950.
26 Jacques Monod, La technique de culture continue: théorie et applications, Ann. Inst.
chemostat model of the human colonic microflora, Regulatory Toxicology and Pharmacology
33: 276–284, 2001.
First-Order Differential Equations (Continuous Dynamical Systems) 153
FIGURE 3.18: Left: A chemostat, perhaps the most classical physical model
of a linear mixing process. In the two chemostats pictured, the containers
in front are the chemostat reactors, the large bottles above are the nutrient
reservoirs, and the metal boxes circulate water around the reactors to maintain
temperature. The outflow collectors are not in view. Right: Loch Linnhe in
western Scotland, the fjordic lake studied by Ross et al. (2001), in which the
nutrient dynamics behave like a chemostat.
For the higher inflow rates, we use 2150 or 21500 in place of 215 and obtain
time values of 0.336 hr and 0.0333 hr, respectively (about 20 min and 2 min).
We may observe that the primary adjustment necessary in applying (3.49) here
is the conversion between concentration and total amount; this is a common
issue with simple nonbiological mixing problems as well.
154 Dynamical Systems for Biological Modeling: An Introduction
28 Holger W. Jannasch, Steady state and the chemostat in ecology, Limnology and
to obtain
1 − e−3qT 1 − e−3qT
C3 = A −qT
, R3 = Ae−qT .
1−e 1 − e−qT
From this we may conjecture (and prove by induction) that for every positive
integer n,
1 − e−nqT −qT 1 − e
−nqT
Cn = A , Rn = Ae .
1 − e−qT 1 − e−qT
As n → ∞, e−nqT → 0 because e−qT < 1. Thus Cn and Rn approach limits
C and R, respectively, as n → ∞, where
A Ae−qT
C= −qT
, R= . (3.50)
1−e 1 − e−qT
It is easy to see that both the initial and residual concentration increase
with each dose but never exceed the limit values. For example, if a drug with an
elimination constant of 0.1 hours−1 is administered every 8 hours in a dosage
e−0.8
of 1 mg/L, the limiting residual concentration is 1−e −0.8 = 0.816 mg/L.
If the time interval T between doses is short, then the ratio of residual
concentration to dose,
R e−qT 1
= = qT ,
A 1 − e−qT e −1
is large (for small T , eqT − 1 is near 0), as each new dose builds upon the
previous one before it has had time to dissipate. On the other hand, if the
time interval between doses is long, each Rn is close to zero, and each Cn is
close to A; each individual dose dissipates almost completely before the next
dose is given.
Example 1.
Harms et al.30 studied the pharmacokinetics of oxytetracycline, an antibiotic
also used as a marker to label bone for age studies, in the loggerhead sea turtle
Caretta caretta (see Figure 3.19). They administered a single dose of the drug
and monitored its gradual elimination from the turtles’ blood plasma. They
reported results in terms of the elimination half-life, which is the amount of
time required to eliminate half of the drug from the blood. For a single dose
of 25 mg/kg administered intramuscularly to turtles with an average mass of
8 kg, the mean elimination half-life was 61.9 hours.
What is the minimum [intramuscular] dosage required in order to keep the
level of oxytetracycline in an 8 kg loggerhead’s blood plasma above 100 mg,
if doses are given once every day?31 Once every three days? Once a week?
Photo courtesy U.S. Fish and Wildlife Service, dls.fws.gov Image courtesy Benjamin Mills
FIGURE 3.19: Left: A loggerhead sea turtle fitted with a tracking device.
Right: The molecular structure of oxytetracycline, so named for the oxygen
(dark atoms along top and bottom) bonded to four (tetra) hydrocarbon rings.
120 600
100 500
80 400
60 300
40 200
20 100
time[days] time[days]
5 10 15 20 5 10 15 20
FIGURE 3.20: Left: Projected drug concentration with daily doses of 31 mg.
Right: Projected drug concentration with weekly doses of 556 mg.
1
Solution: The equation = e−qt1/2 for the elimination half-life t1/2 can be
2
rewritten for the elimination constant q:
1
q = − log /t1/2 .
2
where T0 is the temperature of the body when it is placed into the surround-
ings. Since (3.51) is the same as (3.46) with a and b replaced by k and kT ∗
respectively, its solution (from (3.49)) is
From this we see that the temperature of the body will approach the ambient
temperature as t → ∞. In practical applications, the constant of proportion-
ality k cannot be measured directly but must be calculated from temperature
readings at more than one time.
Example 2.
In an experiment published in 1937, James Hardy tried to determine a con-
stant of proportionality for heat loss from the human body.32 Immediately
there are many complicating factors which must be acknowledged, includ-
ing wind, body position (which affects the amount of surface area exposed),
and substances such as water or fabric (clothes) which affect the transfer of
heat from the skin to the surrounding air. The body’s own defenses, such as
the chill or shivering reflex, which attempts to create heat through expend-
ing [kinetic] energy, also interrupt the heat flow described by Newton. Hardy
therefore limited his primary investigations to the transfer of heat from dry
32 James D. Hardy, The physical laws of heat loss from the human body, PNAS 23: 631–
637, 1937.
First-Order Differential Equations (Continuous Dynamical Systems) 159
skin directly to still air, in the limited range of 23◦ C to 28◦ C, with his human
subjects lying perfectly still and straight. Hardy also wrote the heat exchange
rate coefficient as being directly proportional to surface area, k = KA, where
K is a constant rate per unit area and A is the surface area across which heat
is being exchanged.
Hardy used two human subjects, as well as an elliptical iron cylinder for
comparison purposes (the cylinder had dimensions roughly those of an adult
human torso, and afforded more detailed measurements). He reported heat
exchange rate coefficients of 5.2 cal/◦ C hr m2 for the human subjects and
6.75 cal/◦ C hr m2 for the cylinder. Hardy gave this latter figure as being
equivalent to 1.64 cal/◦ C hr overall (from which we may infer an effective
area of about 0.24 m2 , though Hardy’s figure is twice this), and gave the
cylinder’s mass as 24.98 kg. Using the common mass and surface area estimates
of 68 kg and 1.8m2 for an adult human male, determine the two coefficients
k for Newton’s law of cooling with the human subjects and the cylinder, and
the amount of time required for each to cool from 28◦ C to 23◦ C in a 20◦ C
environment. The specific heats of iron and the human body are 829 cal/kg◦ C
and 112 cal/kg◦C, respectively.
Solution: Hardy wrote in terms of heat exchange; we wish to write in terms
of cooling, or temperature change, so we need to use the specific heat of each
object to convert Hardy’s data (in calories) into purely temperature-based
terms. For the iron cylinder, we have
cal cal
k = 1.64 ◦ ÷ 112 ◦ ÷ 24.98kg = 0.000586/hr.
C hr kg C
For the human subjects, we multiply 5.2 cal/◦ C hr m2 by the estimated surface
area of 1.8m2 and then proceed similarly:
cal cal
k = 5.2 ◦ × 1.8m2 ÷ 829 ◦ ÷ 68kg = 0.000166/hr.
C hr m2 kg C
Example 3.
Newton’s law of cooling may also be applied forensically, to estimate the time
of death based on body temperature measurements after death. This appli-
cation is especially common at the beginning and end of hunting seasons, to
determine whether game animals such as deer were killed within the legal
season. Kienzler et al.33 conducted a study on the cooling rate for white-
tailed deer.34 They found a linear cooling rate constant of k = 0.087656/hr,
33 J.M. Kienzler, P.F. Dahm, W.A. Fuller, A.F. Ritter, Temperature-based estimation for
Example 4.
We can also sometimes use data to generate predictions directly without cal-
culating k explicitly. If a loaf of freshly baked bread cools from 100◦ C to 60◦ C
in the first 10 minutes it is on a cooling rack, when the surroundings are at
20◦ C, what will the loaf’s temperature be 30 minutes after it was placed on
the cooling rack?
Solution: From the second form of (3.52) with T0 = 100◦ C, T ∗ = 20◦ C, we
have T (10) = 20 + (100 − 20)e−10k = 60, and T (30) = 20 + (100 − 20)e−30k .
From the first of these, e−10k = 40/80 = 1/2, so that e−30k = (e−10k )3 = 1/8.
Thus T (30) = 20 + 80(1/8) = 30◦ C.
3.5.4 Migration
Another process which can be described in terms of mixing is the migra-
tion of populations, between or among groups in general, and into and out of
a single group in the context of one-equation models. From cells to organisms
and households, migration adds an important dimension to the dynamics of
populations that are not entirely closed. Cell migration is a necessary mecha-
nism for such processes as cell differentiation and the healing of wounds, and
the migration of tumor cells is metastasis, the defining characteristic of ma-
lignant cancers. Many animals such as insects and birds undertake periodic
seasonal migrations, and transnational immigration of humans has become a
major political issue in the early twenty-first century.
Although some behavioral and social scientists would distinguish between
seasonal or temporary movements and those movements of populations in-
tended to be permanent, we will include in our definition of migration any
movement which can be captured on the time scale of interest to us. If we
are dealing with migration over the course of years, seasonal, month-to-month
movements into and out of the zone under study will only manifest in our
model to the extent that they accumulate. On a time scale of hours, however,
every daily activity has the potential to affect the model. When a population
time since death; see Exercise 15. A linear regression of their data assuming proportional
dependence on time only would yield a different value for k than the one used here.
First-Order Differential Equations (Continuous Dynamical Systems) 161
TABLE 3.3: United States population and per capita demographic rates for
selected years.
Year Population Births Deaths Immigrations Emigrations
(millions) (/yr) (/yr) (/yr) (/yr)
1950 152 0.0241 0.0096 0.000681 0.000185
1960 180 0.0237 0.0095 0.001397 0.000236
1970 204 0.0184 0.0095 0.001628 0.000441
1980 226 0.0159 0.0088 0.001988 0.000520
1990 250 0.0167 0.0086 0.002935 0.000640
2000 281.4 0.0144 0.0085 0.003232 0.000831
TABLE 3.4: United States population predictions (in millions) and percent
errors for models (3.54) and (3.53), alongside the actual values.
Updated parameters 1950 parameters
Year Const. % Prop’l % Const. % Prop’l % True
immig. error immig. error immig. error immig. error value
1950 152 — 152 — 152 — 152 — 152
1960 176.5 −1.94% 176.6 −1.89% 176.5 −1.94% 176.6 −1.89% 180
1970 209.7 2.78% 209.9 2.89% 204.8 0.38% 205.2 0.57% 204
1980 225.5 −0.23% 225.7 −0.15% 237.4 5.05% 238.4 5.47% 226
1990 246.0 −1.59% 246.2 −1.51% 275.1 10.03% 276.9 10.77% 250
2000 277.0 −1.57% 277.4 −1.43% 318.5 13.19% 321.7 14.33% 281.4
to a single variable, the only alternative is to make this rate a constant (or,
conceivably, a function of time alone: see Exercise 16 below for an exploration
of this possibility). This assumption yields an alternative model,
dx
= bx − dx + I − ǫx, (3.54)
dt
where I is the total immigration rate (in, say, people per year).
One way to decide which assumption about the factors driving immigration
gives a better description of population growth is to compare the predictions
of both models with actual data. Both (3.53) and (3.54) are of the form (3.46),
so we may give their solutions, from (3.49), as
I
x(t) = x(0)e(b−d+ι−ǫ)t and x(t) = x(0)e(b−d−ǫ)t + e(b−d−ǫ)t − 1 ,
b−d−ǫ
respectively. Since both models make the simplifying assumption that the as-
sociated per capita rates are constant over time, as well as the assumption
that the underlying forces are not constrained by resource limitations (hence
the linear terms, rather than, say, logistic), the model should only be used
for limited ranges of time. Table 3.4 compares the predictions made by these
two models in two ways: first using the data for each given year to estimate
only the next decade’s population, and then using the true values for the
next decade to generate the next prediction; and then using the initial (1950)
data to project as far as 2000. One can see that the proportional immigration
model (3.53) always makes higher predictions, which is not surprising since
the immigration rate in (3.53) increases continuously with the population,
rather than remaining fixed as in (3.54). Thus the proportional immigration
model is slightly more accurate than the constant migration model when they
undershoot, and less accurate when they overshoot. In general, however, the
difference in predictions is less than one percent, which implies that the dif-
ference between these two assumptions on immigration rates is less important
for predicting population size than other factors.
As can be seen in Table 3.3, the per capita birth, death, and migration rates
are not really constant over time. In the following section we shall consider
First-Order Differential Equations (Continuous Dynamical Systems) 163
Exercises
1. At time t = 0, a tank contains 3 lb. of salt dissolved in 100 gal. of water.
A mixture containing 12 lb. of salt per gallon is pumped in at a rate
of 5 gal./min., and the mixture is pumped out at the same rate. Find
the weight of salt in the tank at time t. How much salt is there after
one hour? Does the weight of salt in the tank at time t have a limit as
t → ∞? [Note Exercises 1–3 involve nonzero initial populations.]
2. A tank contains 100 liters of water and 10 kg. of salt, thoroughly mixed.
Pure water is added at a rate of 5 liters/minute and the mixture is
poured off at the same rate. How much salt is left in the tank after 1
hour, assuming instantaneous and complete mixing?
3. A tank contains 30 lb. of a pollutant dissolved in 200 gallons of water.
Fresh water is poured into the tank at a rate of 10 gal./min. and the
solution is pumped out at the same rate, with immediate and complete
mixing. Find the concentration of pollutant as a function of time, and
the time required for the concentration to drop to 1 lb. per 100 gallons.
4. Water containing 1 pound of salt per gallon is poured into a 50 gallon
tank of water at a rate of 2 gallons per minute and the mixture runs out
at the same rate. After 20 minutes, this process is stopped, and fresh
water is pumped into the tank at a rate of 2 gallons per minute, with
the mixture running out at the same rate. Find the amount of salt in
the tank after 10 more minutes.
5. At time t = 0 cigarette smoke containing 4% carbon monoxide is intro-
duced into a previously carbon monoxide-free room with volume 1800
cubic feet, at a rate of 0.6 cubic feet per minute. The mixture leaves
the room at the same rate. After how long will the carbon monoxide in
the room reach a concentration of 10−4 (which is considered harmful)?
[Make the simplifying assumption that the gas has constant density.]
6. Use either the technique of separation of variables or an integrating
factor to show that the solution to the initial value problem C ′ (t) =
−qC(t), C(T ) = C2 , for T < t ≤ 2T (between the first and second drug
doses), is C(t) = C2 e−q(t−T ) .
7. Suppose that a dose of d milligrams of a drug is injected into the blood-
stream. Assume that the drug is eliminated from the blood at a rate
proportional to the amount of the drug present in the blood. Assume
in addition that half of the drug dose has been eliminated after 1 hour.
Find the time at which the amount of drug in the bloodstream is 20%
of the original dose.
164 Dynamical Systems for Biological Modeling: An Introduction
3.6.1 Superposition
In order to understand the solution methods to be presented in this section,
it is helpful to understand how solutions to a linear ODE can be written in
terms of the solutions to a special, simplified version of the equation. The idea
is that each solution is the sum of two functions, one of which is a solution to
this simpler equation. This idea is called the superposition principle. Although
it applies, properly speaking, only to linear differential equations, for reasons
shown below, we will also use some helpful substitutions to allow us to apply
it to some nonlinear models such as the logistic equation.
The first-order linear differential equation seen in the previous section can
be written in its most general form as
dy
= −a(t)y + b(t), (3.55)
dt
where the coefficients a(t) and b(t) are functions of t. (Recall that a linear
differential equation y ′ = f (t, y) is one in which the function f (t, y) is linear in
166 Dynamical Systems for Biological Modeling: An Introduction
y.) In the special case that a and b are constants, equation (3.55) is separable,
and we found in the previous section that its solution has the form
b
y= + ce−at (3.56)
a
for some constant c. This solution is a sum of two terms, the first a specific
solution y = ab of y ′ + ay = b and the second a family of solutions y = ce−at
of the corresponding homogeneous differential equation y ′ + ay = 0. It turns
out that an analogous result holds for the more general linear ODE (3.55) —
that is, it turns out that solutions to (3.55) can be written as a sum of two
terms, as done in equation (3.56), even if a and b are not constants. First let
us state this result more precisely; then we will see why it is true.
We will call a differential equation such as (3.55) homogeneous if the func-
tion b(t) on the right side is the identically zero function, so that (3.55) takes
the form
dy
= −a(t) y. (3.57)
dt
We will say that the differential equation (3.55) is non-homogeneous (or hetero-
geneous) if b(t) is not the identically zero function. For a non-homogeneous dif-
ferential equation (3.55), the differential equation (3.57) is said to be the corre-
sponding homogeneous differential equation. A homogeneous equation (3.57)
is separable, and by separation of variables has a family of solutions of the
form R
y = c yH (t) = c e− a(t) dt . (3.58)
Here, yH (t) is a solution of the homogeneous differential equation (3.57), and
every solution y of (3.57) is a constant multiple of yH (t) (because this is a
first-order equation35 ). The relation between the solutions of the general equa-
tion (3.55) and the homogeneous equation (3.57) is called the superposition
principle.
To establish the superposition principle, we can show that any two solu-
tions y1 (t) and y2 (t) of (3.57) differ by some constant multiple of yH . We first
observe that
y1′ (t) = −a(t)y1 (t) + b(t), y2′ (t) = −a(t)y2 (t) + b(t),
Example 1.
Find all solutions of the differential equation
y ′ + 2y = 2 et .
Integrating, we have
2 3t
ye2t = e +c
3
for arbitrary constant of integration c. Thus we have the family of solutions
2 t
y= e + c e−2t .
3
Example 2.
Find the solution of the initial value problem
y ′ + 2y = 2et , y(0) = 4.
Example 3.
Find all solutions of the differential equation
1
y′ + y = 1.
1+t
Example 4.
Solve the initial value problem
ty ′ + 2y = t3 , y(1) = 0.
170 Dynamical Systems for Biological Modeling: An Introduction
Solution: We first put the differential equation in standard form (where the
coefficient of y ′ is 1) by dividing both sides of the differential equation by t to
give the initial value problem
2
y ′ + y = t2 , y(1) = 0.
t
Next, we need to solve the corresponding homogeneous differential equation
2
y ′ + y = 0.
t
Separation of variables gives the solution
dy dt
Z Z
= −2 ,
y t
log |y| = −2 log |t| + k,
c
y = ±e−2 log t ek = 2 .
t
With yH = 1/t2 , we can make the substitution y = u/t2 , from which
u′ 2u
y′ = 2
− 3.
t t
Substituting for y ′ and y in the original equation gives us
′
u′
u 2u 2u 2
− + = t , or = t2 ,
t2 t3 t t2 t2
1 t5 t3
c
y(t) = 2 +c = + 2.
t 5 5 t
Example 5.
Solve the initial value problem
Example 7.
A culture tank contains 10 liters of water. A solution containing 1 kg/liter of
nutrient is added at a rate of 0.5 liters/hour, and the mixture is drained off at
a rate of 0.3 liters/hour. Find the concentration of the nutrient solution after
24 hours.
Solution: Let s(t) be the weight of nutrient in the tank at time t. Since the
volume of solution in the tank at time t is (10 + 0.2t) liters, the concentration
of nutrient at time t is s(t)/(10 + 0.2t). Thus the weight of nutrient being
removed at time t is 0.3s(t)/(10 + 0.2t) kg/hr. The weight of nutrient being
added is 0.5 kg/hr. Thus the rate of change of the nutrient weight is given by
ds 0.3s
= 0.5 − ,
dt 10 + 0.2t
which we write in standard form as
3
s′ + s = 0.5.
100 + 2t
To find the appropriate integrating factor, we solve the corresponding ho-
mogeneous differential equation by separation of variables,
ds 3 dt
Z Z
=− ,
s 100 + 2t
Since we need only one solution to obtain an integrating factor, we may omit
3
the constant of integration and write sH = (100 + 2t)− 2 . Then the integrating
3
factor is (100 + 2t) 2 , multiplication by which gives
3 1 3 3
(100 + 2t) 2 s′ + (100 + 2t) 2 s = [(100 + 2t) 2 s]′ = 0.5(100 + 2t) 2 .
Integration gives
Z
3 3 5
(100 + 2t) 2 s = 0.5(100 + 2t) 2 dt = 0.1(100 + 2t) 2 + c.
5
Substitution of the initial condition s(0) = 0 gives 0 = 0.1(100) 2 + c, c =
5
−0.1(100) 2 = −104 , leading to the solution
3 5
(100 + 2t) 2 s = 0.1(100 + 2t) 2 − 104 ,
3
s = 0.1(100 + 2t) − 104 (100 + 2t)− 2 .
3
When t = 24, s = 0.1(148) − 104(148)− 2 ≈ 9.25 kg. Also, when t = 24, the
First-Order Differential Equations (Continuous Dynamical Systems) 173
volume of water in the tank is 14.8 liters, and the concentration of nutrient is
9.25
14.8 ≈ 0.625 kg/liter. The concentration at time t, when the volume of water
is (10 + 0.2t) liters, is
3 − 52
10 + 0.2t − 104 (100 + 2t)− 2
10
= 1− kg/liter.
10 + 0.2t 10 + 0.2t
Example 8.
If we examine the per capita demographic rates for the United States given in
Table 3.3, we can see clear trends of change in each rate: decreasing birth and
death rates and increasing immigration and emigration rates. Graphs of these
rates show that the changes can be approximately described as linear in time.
Assume rates of this form and the hypothesis discussed in Section 3.5 that
immigration is independent of the U.S. population size, to derive the linear
model x′ = r(t)x + I(t), where r(t) = b − d − ǫ. Compare the resulting popu-
lation predictions with those of the constant-coefficient model of Section 3.5.
Does this refinement affect the model’s predictions significantly?
Solution: Let us define t as time in years since 1950, as was done in the
previous section. To make the demographic coefficients linear in time, we write
r(t) = r0 + r1 t and I(t) = I0 + I1 t, and substitute into the original equation
to obtain
x′ (t) = (r0 + r1 t)x + (I0 + I1 t), (3.62)
for which the corresponding homogeneous equation x′ (t) = (r0 + r1 t)x can be
solved by separation of variables:
dx
Z Z
= (r0 + r1 t)dt
x
t2
log |x| =r0 t + r1 + k
2
2
xH =cer0 t+r1 t /2
from which Z t
(I0 + I1 s)e−(r0 s+r1 s /2)
2
u(t) = x0 + ds,
0
since u(0) = x(0)/xH (0) = x0 /1. We can simplify part of the integral via the
substitution w = r0 s + r1 s2 /2: then
r0 I1 r0 I1 dw
I0 + I1 s = I0 − I1 + (r0 + r1 s) = I0 − I1 + ,
r1 r1 r1 r1 ds
174 Dynamical Systems for Biological Modeling: An Introduction
and
I1 I1 I1
Z Z
−(r0 s+r1 s2 /2)
(r0 + r1 s)e ds = e−w dw = − e−w + c
r1 r1 r1
I1 −(r0 s+r1 s2 /2)
=− e + c.
r1
Thus
Z t
r0 I1 −(r0 t+r1 t2 /2)
e−(r0 s+r1 s /2) ds −
2
u(t) = x0 + I0 − I1 e −1 ,
r1 0 r1
and
I1 2 I1
x(t) = xH u = x0 + er0 t+r1 t /2 −
r1 r1
Z t
r0
e−(r0 s+r1 s /2) ds.
2
r0 t+r1 t2 /2
+ I0 − I1 e
r1 0
FIGURE 3.21: Migration model predictions for the U.S. population using
fixed coefficients (short dashes) and coefficients linear in time (long dashes),
compared to census values (solid line).
Example 9.
Solve the initial value problem
1
y ′ − y = −ty 2 , y(0) = − .
2
Then we have
1 1
y= = .
z (t − 1) + ce−t
Substitution of the initial condition y(0) = 21 gives 12 = c−1
1
, or c = 3. Thus
the solution of the given initial value problem is y = [(t − 1) + 3e−t ]−1 .
model, Economica 69: 207–221, 2002. See their Appendix B for details.
First-Order Differential Equations (Continuous Dynamical Systems) 177
In order to find lim z(t), we may use L’Hôpital’s rule for the indeterminate
t→∞
form R t r ers r ert
0 K(s) ds K(t) 1
lim rt
= lim rt
= lim .
t→∞ e t→∞ re t→∞ K(t)
1
Thus if K(t) approaches a limit as t → ∞, then z(t) → limt→∞ K(t) , and thus
y(t) and K(t) have the same limit. If the carrying capacity does not approach
a limit (for instance, if it is seasonal), then the population will not do so, either
(since its “target level” is moving). However, the solution can be shown (see
Exercise 15 below) to be asymptotically bounded within the range of values
taken on by K(t).
Example 10.
In Section 3.2 we discussed the use of the logistic model with constant coeffi-
cients in accounting for Australian sheep populations, noting that the model
described the initial period of rapid growth well but failed to account for
the sustained significant (10% or more) fluctuations seen after the population
reached its carrying capacity. One hypothesis has been that the pasture land
required some time to recover from the overgrazing it experienced. Assume
that the pastures undergo cycles of overgrazing followed by recovery. To what
extent can this assumption account for the variations in the Tasmanian sheep
population seen in Figure 3.7?
Solution: From the figure, if the period 1846–1856 represents the period of
initial overgrazing (causing growth greater than expected for the amount of
system resources), then 4 cycles of fluctuation can be observed during the ap-
proximately 68 years which follow, with amplitude approximately 170 (thou-
sand sheep). This can be modeled most simply by adding a sine wave of
amplitude 170 and period 17(= 68 ÷ 4) to the estimated carrying capacity of
1670 (thousand sheep) given in the figure caption, making
2π
K(t) = 1670 + 170 sin (t − 1856),
17
where t is the year. If we assume the sheep’s intrinsic growth rate still constant
at r = 0.13125/yr, then we have the second special case described above, which
leads to (3.69). Taking our starting point as t0 = 1857, for which the original
solution y(t) = 1670/[1 + exp(240.81 − 0.13125 t)] has the value 1585, so that
z0 = 1/1585, we can then rewrite (3.69) as
Z t rs
re
z(t) = z0 e−r(t−t0 ) + e−rt ds.
t0 K(s)
Figure 3.22. We see that the population size does not approach the variable
carrying capacity asymptotically; rather, it increases for that part of the cycle
when y(t) < K(t), and decreases as long as y(t) > K(t) (see Figure 3.23). We
also see that the fluctuations appear to have a negative effect overall on the
population, in that the population predicted using a variable carrying capacity
is less than the average carrying capacity more than half of the time (that is,
it is below the line more often than above it). In general, we can see that this
model is more consistent with observed population values than the original,
though we should note that this is still only a rough approximation at best.
Exercises
In Exercises 1–12, solve the given initial value problem.
1
1. y ′ + 2y = 2, y(0) = 2 7. ty ′ + 2y = t2 , y(−2) = 5
y y3
13. Find the solution of y ′ + t = t2 passing through the point (6,3).
14. Find the solution of y ′ − y = ty 2 passing through the point (1, −1).
180 Dynamical Systems for Biological Modeling: An Introduction
15. Find the solution of the differential equation ty ′ = 2t2 y + y log y through
the point (1,1). [Hint: Make the change of dependent variable v = log y.]
16∗ . Discuss the behavior of solutions of the initial value problem y ′ = λy,
y(0) = y0 as t → ∞ for each of the cases λ > 0, λ = 0, and λ < 0. (We
will see in Section 4.1 that this is a key issue in the qualitative analysis
of differential equations.)
17∗ . Solve and then discuss the behavior of solutions as t → ∞ for each of
the following differential equations satisfying the generic initial condition
y(t0 ) = y0 , where t0 and y0 are given constants: (a) y ′ = −2y + e−t (b)
1
y ′ = −2y +et (c) y ′ = −2y +1 (d) y ′ = −2y + 1+t 2 [Hint: Use L’Hôpital’s
ty ′ + ay = f (t)
has a unique solution that is bounded as t → 0+, and find the limit of
this solution as t → 0+.
22. Verify that the function defined by equation (3.61) is the solution of
Example 6.
23. A tank contains 10 liters of water to which is added a nutrient solution
containing 0.3 kg of nutrient per liter. This nutrient solution is poured
in at the rate of 3 liters/min, is thoroughly mixed, and then the mixture
is drained off at the rate of 2 liters/min. How much nutrient is in the
tank after 5 minutes?
First-Order Differential Equations (Continuous Dynamical Systems) 181
24. A 500 gallon runoff containment tank contains 200 gallons of a solution
containing 15 lb. of pollutant, when heavy rains flood the system. Fresh
water pours into the tank at a rate of 12 gallons per minute, and the
mixture is pumped out for processing at a rate of 6 gallons per minute.
Find the amount of pollutant in the tank when the tank overflows.
25. A catfish pond has 1000 L of water containing 2 kg of dissolved salt.
Flooding causes seawater to enter the pond at 60 L/min. (Seawater has
an average salt content of 36 g/L.) At the same time, a drainage system
turns on and pumps the water out at a rate of 100 L/min. (a) Find the
amount of salt in the tank at any time t. (b) Find the concentration of
salt (in grams per liter) as a function of time. Chervinski found37 that
young catfish can survive salinity levels up to 1/4 that of seawater. How
long does the pond’s caretaker have to rescue the catfish? (c) Sketch the
graphs of the amount of salt and the concentration of salt against time,
and determine the absolute maximum of each quantity.
26. A 400 liter tank contains a mixture of water and chlorine with 0.05 grams
of chlorine per liter. In order to reduce the concentration of chlorine in
the tank, fresh water is pumped into the tank at a rate of 4 L/sec. After
thorough stirring, the mixture is pumped out at a rate of 10 L/sec. Find
the amount of chlorine in the tank as a function of t.
27. A very crude model of a population with migration is given by
y ′ + ay = M (t), y(0) = y0 ,
where a is the proportional rate of decrease without migration and M (t)
is the rate at which members are added to the population. Here, positive
values of M (t) correspond to immigration, and negative values of M (t)
correspond to emigration. Find the population size as a function of t if
M (t) is a positive constant M0 .
28. Find the population size y modeled by y ′ +ay = M0 +A sin ωt, y(0) = y0
where the sine term models fluctuations in migration. [Observe that if
|M0 | < |A|, then the migration rate takes on both positive and negative
values, indicating that there are periods of net immigration and periods
of net emigration.]
29. Solve the logistic equation with constant coefficients y ′ = ry(1 − y/K)
via the Bernouilli substitution.
30. Suppose that a population governed by the logistic model (3.65) has
a carrying capacity bounded by constants Kmin ≤ K(t) ≤ Kmax for
all time (and a constant intrinsic growth rate r). Show that the result-
ing population size y(t) is bounded by the two solutions of the form
37 J. Chervinski, Salinity tolerance of young catfish, Clarias lazera (Burchell), Journal of
Miscellaneous exercises
In Exercises 1–4, verify that the given family of functions is a solution of the
given differential equation for every value of the constant c.
sin t − 2 cos t ′
1. y = ce2t + , y − 2y = cos t
5
2
2. y = cet /2
, y ′ − ty = 0
e−1/t 2 ′
3. y = c , t y = y(1 − t)
t
y
4. y = ct − t log t, y ′ = −1 + t
In Exercises 5–8, draw a direction field for the given differential equation on
the given rectangle.
5. y ′ − 2y = cos t; 0 ≤ t ≤ 5, −2 ≤ y ≤ 2
6. y ′ − ty = 0; −2 ≤ t ≤ 2, −2 ≤ y ≤ 2
1
7. t2 y ′ = y(1 − t); 2 ≤ t ≤ 4, −1 ≤ y ≤ 1
8. y ′ = −1 + yt ; 1
2 ≤ t ≤ 4, 0 ≤ y ≤ 4
For each of the differential equations in Exercises 9–12, construct the direction
field and use it to sketch the solution passing through the origin.
9. y ′ = t + y
t
10. y ′ =
y−t
11. y ′ = (y + 1)(t + 1)
12. y ′ = y 3 − 1
In Exercises 13–26, find the solution of the given initial value problem.
13. y ′ = t2 y 2 , y(1) = 1
14. y ′ = ty log y, y(0) = e
2
15. y ′ = tey et , y(1) = 0
First-Order Differential Equations (Continuous Dynamical Systems) 183
34. To a tank containing 100 lb. salt dissolved in 100 gallons of water is
added a solution containing 2 lb. salt per gallon at a rate of 2 gallons
per minute. The mixture runs out at a rate of 4 gallons per minute. Find
the weight of salt and the concentration of the mixture after 10 minutes.
35. A body with initial temperature 100◦ C cools in air at 20◦ C, taking 5
minutes to cool to a temperature of 80◦ C. How long does it take to reach
a temperature of 60◦ C?
36. A hot body is cooled in air at 20◦ C, reaching temperatures of 45◦ C 10
minutes after it starts to cool and 40◦ C 20 minutes after it starts to
cool. What was its original temperature when cooling began?
p
37∗ . Explain why the initial value problem y ′ = 1 − y 2 , y(0) = 2 has no
solution.
38∗ . Show that the function y(t) defined as y = 0 for 0 ≤ t < 1 and as y = t3
for 1 ≤ t ≤ 2 is differentiable for 0 ≤ t ≤ 2 and solves the initial value
problem y ′ = 3y 2/3 , y(0) = 0.
ry
39. The differential equation y ′ = y+A − y may be thought of as a limiting
case of the Verhulst difference equation yk+1 = ykry+A
k
(think of yk+1 − yk
′
as an approximation to the derivative y ). Compare the behaviors of the
solution of the difference equation and the solution of the differential
equation, with A = 1 and various values of r. For what values of r do
the two equations behave similarly?
40. (a) Formulate a differential equation which is a limiting case of the
Beverton-Holt difference equation
ayk
yk+1 = .
1 + byk
(b) For what values of the parameters a and b do the solutions of the
difference equation and the differential equation derived in part (a)
behave similarly?
41∗ . As has been mentioned in the text, use of a continuous variable to rep-
resent a discrete population is an approximation, the error involved in
which we are usually willing to overlook if it is small relative to the
population size; that is, if a model predicts a population of 761.23 birds
at a given time, we can use the general size rounded to the nearest inte-
ger as an indication that the population will include something close to
761 birds. However, if we keep in mind the fact that deterministic mod-
els such as differential equations represent an average over all possible
outcomes, we can interpret non-integer (or indeed all) results in these
terms. Write probabilistic interpretations of the results in the first two
simple examples of Section 3.1.
Chapter 4
Nonlinear Differential Equations
between disordered and ordered foraging in Pharaoh’s ants, PNAS 98(17): 9703–9706, 14
August 2001; and David J. T. Sumpter and Madeleine Beekman, From nonlinearity to
optimality: pheromone foraging by ants, Animal Behaviour 66: 273–280, 2003.
185
186 Dynamical Systems for Biological Modeling: An Introduction
ants, these foragers lay pheromone trails connecting any food sources they find
back to their nests, in order to help nestmates find the food (see Figure 4.1).
These pheromone trails are volatile, lasting only about 10 minutes unless ants
still on the trail lay down more. Beekman, Sumpter, and Ratnieks studied
the effects of colony size, distance from food source to nest, and strength
of the pheromone trail on the number of ants using the trail. Their model
hypothesized, first, that ants begin to forage at a given food site either by
finding it independently, at a rate α inversely proportional to the food site’s
distance from the nest, or by being drawn to the pheromone trail, at a rate
dependent both on the number of ants laying the trail down and on the quality
of the food source, measured by a parameter β. If we let x(t) represent the
number of ants foraging on the trail at time t, and denote the colony size by
n, then the rate at which an ant begins using the trail is given by α + βx, the
number of ants not using the trail is n−x, and thus the total rate at which ants
join the trail is (α+βx)(n−x). Their model also hypothesized that ants lose or
stop using the pheromone trail at a rate dependent, in part, on the strength
of the trail. More specifically, they hypothesized that the rate at which an
individual ant loses the trail is inversely proportional to the number of ants x
renewing the trail, in such a way that the total rate at which ants leave the
trail never passes a maximum rate s. (Note the distinction here between the
rate for a single ant, or per capita rate, and the total rate.) This hypothesis
sx
gives a trail loss rate of x+K , where K is the number of ants on the trail at
which the total trail loss rate reaches s/2. Since the total rate never surpasses
s, this function is said to saturate; we shall see other examples of biological
processes that saturate in the sections ahead.
The model that results from the above hypotheses is
dx sx
= (α + βx)(n − x) − . (4.1)
dt x+K
Note that although equation (4.1) is separable, the technique of separation
of variables does not lead to a closed-form expression for x(t) because of the
algebraic complexity of the terms. We can nevertheless learn a lot about the
behavior this set of rules produces, results which Beekman, Sumpter, and
Ratnieks confirmed experimentally and interpreted to reveal a remarkable
shift in the efficiency of such behavior as colony size grows. In order to do so,
we now develop the tools we will need.
This is true because the derivative of a constant function is the zero func-
tion, so if y = y ∗ , then dy
dt = 0, and this function obeys the differential equation
dy ∗
dt = f (y) if and only if f (y ) = 0. Such a solution is called an equilibrium
or fixed point of the differential equation. Qualitative analysis of differential
equations begins with the study of the equations’ equilibria.
Because any solution which includes an equilibrium remains at that equi-
librium for all time, by virtue of the informal argument raised earlier, the
equilibria of a single differential equation separate the state space (the set of
all possible values of y(t)) into disjoint intervals: no solution can cross an equi-
librium (or else two solutions pass through that point, violating uniqueness),
so any solution must stay within a single interval. Furthermore, since f (y) is
continuous, and is zero only at equilibria, f (y) cannot change sign within an
interval (between equilibria): it will either be uniformly positive, in which case
all solutions in that interval will increase monotonically, or uniformly nega-
tive, in which case all solutions in that interval will decrease monotonically.
If we consider the graph of y(t) over time, these intervals become horizontal
bands on the graph.
Example 1.
Describe the bands for the logistic differential equation
dy y
= ry 1 − (4.3)
dt K
with r, K > 0, and find which solutions are increasing, and which decreasing.
Solution: The
bands are bounded by equilibria, which are the roots of f (y) =
y
ry 1 − K , namely y = 0 and y = K. There are therefore three bands to
Nonlinear Differential Equations 189
y
consider (Figure 4.2). If y > K (Band 1), then 1 − K < 0, and the function
y
f (y) = ry 1 − K is negative. Therefore, solutions y(t) of (4.3) with y(0) > K
are decreasing for all t. If 0 < y < K (Band 2), the function f (y) is positive,
and therefore solutions y(t) with 0 < y(0) < K are increasing for all t. If y < 0
(Band 3), then f (y) is negative, and therefore solutions y(t) with y(0) < 0 are
decreasing for all t. (Compare Figure 4.2 with the direction field in Figure 3.11,
where r = K = 1.)
3
y
Band 1 2.5
y(t) decreasing
2
K
Band 2 y(t)1.5
y(t) increasing
1
0 t
Band 3 0.5
y(t) decreasing
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
t
FIGURE 4.2: The t − y plane di- FIGURE 4.3: Some solutions of the
vided into bands by the equilibria of logistic equation (4.3) with r = 2,
the logistic equation. K = 2.
Note that if y(0) is above the largest equilibrium of (4.2), then the interval
or band containing the solution is unbounded, and if f (y) > 0 in this band,
then the solution y(t) may be (positively) unbounded and does not neces-
sarily exist for all t ≥ 0. Likewise, if y(0) is below the smallest equilibrium
of (4.2) and f (y) < 0 in the band containing the solution, then the solution
may be unbounded (negatively) and may fail to exist for all t ≥ 0. In the
example above of the logistic equation, solutions with y(0) < 0 (Band 3) grow
negatively without bound, but nevertheless exist for all t ≥ 0.
In practice, it is often easy to check for unbounded growth in solutions of a
single differential equation, by seeing whether the growth rate f (y) is bounded
away from zero for large y. A solution may become unbounded positively
(i.e., y(t) → +∞) if it is monotone increasing when y is large. Models of
most biological systems account for limitations on system resources, so that
190 Dynamical Systems for Biological Modeling: An Introduction
f (y) < 0 for very large y, and in such a case every solution remains bounded,
because solutions which become large and positive must be decreasing. If, as is
also frequently the case in applications, y = 0 is an equilibrium and only non-
negative solutions are of interest, then solutions cannot cross the line y = 0
or become negatively unbounded (i.e., y(t) → −∞). In many applications, y
stands for a quantity such as the number of members of a population, the
mass of a radioactive substance, the quantity of money in an account, or the
height of a particle above ground which cannot become negative. In such a
situation, only non-negative solutions are significant, and if y = 0 is not an
equilibrium but a solution reaches the value zero for some finite t, we will
consider the population system to have collapsed and the population to be
zero for all larger t. If this is the case, we need not be concerned with the
possibility of solutions becoming negatively unbounded even if y = 0 is not
an equilibrium.
In the above case of the logistic equation (4.3), we can see that every
non-negative solution remains bounded, by observing that f (y) < 0 (and thus
dy/dt < 0) when y > K (Band 1).
We can now combine the properties of boundedness and monotonicity to
argue that any solution of a differential equation (4.2) which does not grow
unbounded must instead approach a limit. From calculus, any function which
is either bounded above and monotone increasing, or bounded below and
monotone decreasing, must approach a limit. (Consider the analogy of a car
stuck in forward gear approaching a wall along a narrow lane. At some point—
the wall, if not sooner—it must approach a stopping point.) However, it can
be shown that if a differentiable, monotone function—a solution of (4.2)—
tends to a limit as t → ∞, then its derivative meanwhile must tend to zero.
That is, if y(t) → L, then dy d
dt → dt L = 0. At the same time, by continuity
dy
of f , f (y(t)) → f (L); since dt = f (y(t)), this suggests that f (L) = 0: that
is, that L is a root of f (y). By Property 1, this means that L must be an
equilibrium of the system. Thus every bounded solution of (4.2) approaches
an equilibrium.
In the case of the logistic equation (4.3), this means that since all non-
negative solutions are bounded, all non-negative solutions must approach an
equilibrium, either y = 0 or y = K. Note, however, that different solutions
to the same equation may approach different equilibria. Here it follows from
Property 3 and the arguments in Example 1 that (as seen in Section 3.2)
all non-negative solutions approach y = K, except for the constant solution
Nonlinear Differential Equations 191
y(t) = 0, which is already at the zero equilibrium. Figure 4.3 illustrates the
behavior with r = 2, K = 2.
Note also that these last two properties are carefully worded to apply
only to single differential equations, and not to systems of several differential
equations. We shall see in Chapter 5 that solutions to systems of two or more
differential equations need not be monotone, and consequently their bounded
solutions need not approach equilibria.
Before continuing on, we invite the reader to consider how these properties
apply to the other equation familiar to us from Chapter 3: the first-order linear
differential equation y ′ = −ay + b. We will discuss the results in an example
later in this section.
4.1.2 Stability
The question of whether or not solutions approach a given equilibrium
motivates us to study more formally the concept of stability, which has to
do with whether solutions approach (and remain near) a given equilibrium.
In applications, the initial condition usually comes from observations and is
subject to experimental error. For a model to be a plausible predictor of what
will actually occur, it is important that a small change in the initial value not
produce a large change in the solution. For example, the solution of the logistic
differential equation with initial value zero remains zero for all values of t, but
every solution with positive initial value, no matter how small, approaches
the limit K as t → ∞. Because of its extreme sensitivity to changes in the
initial value, we do not ascribe practical significance to the equilibrium zero
as a limiting value for solutions of the logistic equation.
An equilibrium y ∗ of (4.2) is said to be stable if every solution with initial
value close enough to y ∗ remains close to it for all time. The equilibrium
is said to be locally asymptotically stable when every solution with initial
value close enough to y ∗ actually approaches it as t → ∞. The equilibrium
is said to be globally asymptotically stable when every solution approaches it,
regardless of initial condition. If there are solutions which start arbitrarily
close to an equilibrium but move away from it, then the equilibrium is said to
be unstable. For the logistic differential equation (4.3) the equilibrium y = 0
is unstable, and we can see from examining Bands 1 and 2 in Example 1 that
the equilibrium y = K is locally asymptotically stable; in fact, it is globally
asymptotically stable in the positive (y > 0) state space. (In asserting the
global nature of the stability of y = K, we discount the initial condition
y = 0: in applications, unstable equilibria have no significance, because they
can be observed only if the initial condition is “just right.”)
Note that the process for determining global asymptotic stability, which
requires knowledge of other equilibria, the state space, and boundedness of so-
lutions, is distinct from that for determining local asymptotic stability, which
requires only information about what happens very close to one equilibrium
point. Therefore, the standard qualitative analysis process we shall now de-
192 Dynamical Systems for Biological Modeling: An Introduction
velop involves three steps: (1) identifying all equilibria, (2) performing a local
stability analysis on each equilibrium, and (3) assembling the pictures of local
behavior into a single (global) portrait of how solutions may behave in gen-
eral. From Property 1 we already have a methodical way to find equilibria of
a differential equation (4.2): they are the zeroes of the function f (y).
There is also a methodical way to see whether or not a given equilibrium is
locally asymptotically stable or not, beyond testing the sign of f (y) just above
and below the equilibrium (this latter method also does not extend gracefully
to systems of more than one equation, where there are many directions in
which one can be close to an equilibrium). As was the case with discrete-time
models, this way involves the derivative (slope) of the growth function f (y)
at the equilibrium.
Let us look at the piece of the graph of f (y) near an equilibrium y ∗ of
(4.2). By Property 1, f (y ∗ ) = 0; normally this means that the graph of f (y)
crosses the y-axis from one side to the other at y ∗ (in a few unusual cases, f
might turn around and go back the way it came, as in Figure 4.4(c)). We can
determine the direction of the crossing from the sign of the derivative at the
equilibrium, f ′ (y ∗ ). If f ′ (y ∗ ) < 0, then f (y) has a negative slope there, so that,
close to y ∗ , f (y) is positive if y < y ∗ and negative if y > y ∗ (Figure 4.4(a)).
In this case solutions above the equilibrium decrease toward the equilibrium,
while solutions below the equilibrium increase toward the equilibrium. This
shows that an equilibrium y ∗ with f ′ (y ∗ ) < 0 is locally asymptotically stable.
By a similar argument, if f ′ (y ∗ ) > 0, solutions above the equilibrium in-
crease and solutions below the equilibrium decrease, with both moving away
from the equilibrium (Figure 4.4(b)). In this case we see that the equilibrium
is unstable. The only case in which this approach fails to determine an equi-
librium’s stability is when f ′ (y ∗ ) = 0, one example of which is illustrated
in Figure 4.4(c). In such a case (which occurs relatively rarely in practice),
both stability and instability are possible, and we must investigate the sign of
f (y) above and below the equilibrium. If f (y) < 0 above the equilibrium and
f (y) > 0 below the equilibrium, then the equilibrium is locally asymptotically
stable.
y y y
y* y* y*
FIGURE 4.4: Equilibria y ∗ with (a) f ′ (y ∗ ) < 0, (b) f ′ (y ∗ ) > 0, (c) f ′ (y ∗ ) = 0.
Example 2.
For each equilibrium of the logistic differential equation (4.3), determine
whether it is locally asymptotically stable or unstable.
y
, f (y) = r − 2ry 2y
′
Solution: We have f (y) = ry 1 − K K = r 1 − K . Since
f ′ (K) = −r < 0, the equilibrium y = K is asymptotically stable, and since
f ′ (0) = r > 0, the equilibrium y = 0 is unstable.
Example 3.
Show that the equilibrium y = 0 of the differential equation y ′ = −y 3 is locally
asymptotically stable.
Solution: f ′ (0) = −3(0)2 = 0, so we cannot apply Property 4 here. However,
for y small and negative we see that −y 3 > 0, while for y small and positive
−y 3 < 0. Thus all solutions approach the limiting value zero, and the equilib-
rium y = 0 is locally asymptotically stable. The behavior of solutions is shown
in Figure 4.5.
1 1
0.8
0.6
y(t)0.5
y(t)
0.4
0.2 t
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
t
–0.2
–0.4 –0.5
–0.6
–0.8
–1
–1
FIGURE 4.5: Some solutions of the FIGURE 4.6: Some solutions of the
equation y ′ = −y 3 near 0. equation y ′ = −y 2 near 0.
194 Dynamical Systems for Biological Modeling: An Introduction
Example 4.
Show that the equilibrium y = 0 of the differential equation y ′ = −y 2 is
unstable.
Solution: The function −y 2 is negative for all y except y = 0. Thus solutions
of the differential equation y ′ = −y 2 are decreasing. Since solutions with
negative initial value move away from y = 0, the equilibrium y = 0 is unstable.
The behavior of solutions is shown in Figure 4.6. Note that for each of the
differential equations in Examples 3 and 4 we have f (0) = 0, f ′ (0) = 0;
no conclusion about stability can be drawn from f ′ (0) = 0 without further
examination.
We note that this theorem does not cover the case f ′ (y ∗ ) = 0. As we have
seen in Examples 3 and 4 above, a situation with f (y ∗ ) = f ′ (y ∗ ) = 0 must be
examined individually.
Let (y ∗ ) be an equilibrium of a differential equation (4.2), that is, a solution
of the equation f (y) = 0. We assume that the equilibrium is isolated, that is,
that there is an interval around y ∗ that does not contain any other equilibrium.
We shift the origin to the equilibrium by letting y = y ∗ + u, and then use
Taylor’s theorem to approximate f (y ∗ + u). Our approximation is
f (y ∗ + u) = f (y ∗ ) + f ′ (y ∗ )u + h1 = f ′ (y ∗ )u + h (4.4)
Nonlinear Differential Equations 195
-3 -2 -1 1 2
y −2 0 1 y
-1
-2
-3
To read a phase line portrait, we think of the phase line as the state space
within which the solution curve y(t) moves, thinking of t as a parameter.
Nonlinear Differential Equations 197
Thus the solution is described by motion along the line, in the direction given
by the arrows. The graph of Figure 4.7 describes a situation in which there
are asymptotically stable equilibria at y = −1 and y = 2, and an unstable
equilibrium at y = 0.
We illustrate this approach with two examples, the first the familiar linear
differential equation studied in Chapter 3.
Example 5.
dy
Describe the asymptotic behavior of solutions of the differential equation dt =
−ay + b, analytically and with a phase line.
Solution: Here f (y) = −ay + b, f ′ (y) = −a for all y, and the only equilibrium
is y = ab . If a > 0, then any equilibrium is asymptotically stable (since f ′ (y) =
−a < 0), and there are no unbounded solutions (since f (y) < 0 if y is large and
positive, f (y) > 0 if y is large and negative). This means that every solution
is bounded and approaches the limit ab .
If a < 0, however, the equilibrium is unstable, as now f ′ (y) = −a > 0 for
all y. Further, since f (y) > 0 above the equilibrium and f (y) < 0 below the
equilibrium, every solution grows unbounded, either positively or negatively.
Figure 4.8 shows the two corresponding phase lines.
(a) (b)
b/a b/a
dy
FIGURE 4.8: Phase lines for dt = −ay + b with (a) a > 0, (b) a < 0.
For our second example we consider a model which has been proposed as
an alternative to the logistic model. Recall that the logistic model was origi-
nally conceived to account for the effect of resource limitations on population
growth, by making the per capita growth rate a decreasing function of pop-
ulation size, positive for small y but negative for large y. Another way to do
so is with a per capita birth rate re−y and a constant per capita death rate
d, which yields the differential equation
dy
= y re−y − d .
(4.6)
dt
Since the precise form of the logistic model was derived empirically, and we
want model predictions that are robust in the sense of being tied more to
the qualitative assumptions than to the specific function(s) used, we can now
apply the analysis tools derived in this section to see whether this alternative
model predicts the same kind of population growth as the logistic model.
198 Dynamical Systems for Biological Modeling: An Introduction
Example 6.
Describe the asymptotic behavior of solutions with y(0) ≥ 0 of the differential
equation (4.6), with r, d > 0.
Solution: The equilibria of (4.6) are the solutions of y(re−y − d) = 0. Thus
there are two equilibria, namely y = 0 and the solution y ∗ of re−y = d, which is
y ∗ = log dr . If r < d, then y ∗ < 0, and only the equilibrium y = 0 is of interest.
In this case, f (y) < 0 for y > 0, and solutions with y(0) > 0 decrease to 0
(Figure 4.9). That is, the zero equilibrium is globally asymptotically stable in
the non-negative state space.
If instead r > d, we define K = log dr > 0, so that the positive equilibrium
is y = K, and e−K = d/r. We may now rewrite the differential equation (4.6)
as y ′ = ry e−y − e−K −y −K
. For the function f (y) = ry e − e , we have
f ′ (y) = r e−y − e−K − rye−y and f ′ (0) = r(1 − e−K ) > 0, implying that the
fHyL
fHyL
y y
(cf. equation 1 in Beekman et al.). Since f (0) > 0 and f (n) < 0, by continuity
of f there must be at least one equilibrium x∗ in between; we see from the
cubic equilibrium condition that there is in fact either one or three. There is
a cubic formula which can be used to find any equilibria, but the resulting
expressions have many terms and may be difficult to interpret.
Since the expressions for the equilibria are complicated, the first derivative
test given in the equilibrium stability theorem may be difficult to apply in
general. Instead, we shall use the phase line, together with the fact that the
one or three equilibria are simple zeroes of f , to determine their stability. If
there is only one equilibrium, then the phase line is split into two intervals, on
the lower of which f is increasing, and on the upper of which f is decreasing
200 Dynamical Systems for Biological Modeling: An Introduction
(a) (b)
0 E n 0 E1 E2 E3 n
FIGURE 4.11: Phase lines for (4.1) with (a) one equilibrium, (b) three
equilibria.
improvement in foraging organization for colony sizes above 600 ants. We can
also observe that if we expand the expression for ∆ as a polynomial in n,
the lead term is −β 2 (α − βK)2 n4 , implying that for sufficiently large colonies
∆ < 0 and the more complicated behavior occurs.
We can also address the hypothesis of more efficient foraging (i.e., a higher
proportion of ants using the trail) in larger colonies by dividing the equilib-
rium condition (4.7) by n3 in order to rewrite it in terms of the equilibrium
proportion y = x/n of ants using the trail:
α + βK s + αK α + βK αK
βy 3 + − β y2 + 2
− y − 2 = 0.
n n n n
Exercises
For each of the differential equations in Exercises 1–8, draw the phase line,
find all equilibria and describe the behavior of solutions as t → ∞.
1. y ′ = y 5. y ′ = y(1 − y)(2 − y)
2. y ′ = −y 6. y ′ = −y(1 − y)(2 − y)
3. y ′ = y 3 7. y ′ = y 2 (y + 1)
4. y ′ = sin y 8. y ′ = −c(y − 1)2
For each of the differential equations in Exercises 9–12, describe the behavior
of solutions with y(0) > 0.
9. y ′ = ry log K
y
202 Dynamical Systems for Biological Modeling: An Introduction
10. y ′ = ry −1 + Ky
ry(K−y)
11. y ′ = K+Ay
h i
y θ
12. y ′ = ry 1 −
K
∗
13. Consider a differential equation y ′ = g(y) with g(y) = yh(y), where
h(y) is a decreasing function of y, so that h′ (y) < 0, and h(y) > 0 for
0 < y < K, but h(y) < 0 for y > K. Show that the only equilibria are
y = 0 and y = K, and that the equilibrium y = 0 is unstable while the
equilibrium y = K is asymptotically stable.
∗
14. (a) Show that if y(t) is a solution of an autonomous differential equa-
tion y ′ = g(y), then y ′′ (t) = g ′ {y(t)} g{y(t)}. (b) Deduce from part (a)
that if the solution y(t) has an inflection point, this inflection point must
occur at a point (t, y) with g ′ (y) = 0.
15. In this section, we analyzed the long-term behavior of Beekman et al.’s
model for ants following a pheromone trail to a single food source. In
what sense is this behavior long-term, i.e., what is the time scale implicit
in this discussion? How does this time scale relate to the assumption that
the colony size is constant?
16. How would the ant foraging model (4.1) change if
(a) the ants did not use pheromone trails?
(b) ants using the trail never left it?
(c) ants using the trail left it at a constant per capita rate σ?
In each case, indicate how the resulting model’s behavior differs from
that of Beekman et al.
17. How would the interpretation of the rate at which ants leave the trail
be affected if it were modeled by
p
x
s ,
x+K
4.2 Harvesting
In Section 2.5 we studied harvesting, or population management, in
discrete-time models, where both reproduction and harvesting occur at spe-
cific discrete times. If the harvesting events are frequent enough, they can be
considered to be nearly continuous and ongoing, in the same way that we con-
sider ongoing reproduction and mortality for many biological systems. In this
section we will look at the management of populations from a continuous-time
perspective, including not only such activities as hunting, fishing and logging,
but also production of bacteria in laboratory settings and even the dispersal
of human populations.
In our previous study of population harvesting we considered primarily
constant-yield harvesting, a practice in which members of a population are
removed at a constant rate. Constant-yield harvesting models3 are appropriate
for systems in which there is a set quota or deliberate, controlled management
of the population. In other situations, however, such as the notorious bycatch
of dolphins in the tuna industry in the late 20th century that eventually led to
changes in fishing practices, harvesting may be either an incidental byproduct
of an activity not directly designed to control that population, or a process
with a duration or intensity set by factors other than the population size: for
instance, occasional fishing in a small lake or pond, where fishermen go a given
number of times, and the number of fish caught depends upon the availability
(population density) of fish in the pond. These situations are better modeled
by constant-effort harvesting, in which members of the population are removed
at a rate proportional to the population size.4 In this section we shall consider
biological systems of both types.
y ′ = f (y) − H, (4.8)
3A primary reference is Fred Brauer and David A. Sánchez, Constant rate population
harvesting: equilibrium and stability, Theor. Pop. Biol. 8(1): 12–30, August 1975.
4 A primary reference is M.B. Schaefer, Some aspects of the dynamics of populations
important to the management of commercial marine fisheries, Bull. Inter-Amer. Trop. Tuna
Comm. I: 25–56, 1954.
204 Dynamical Systems for Biological Modeling: An Introduction
where H > 0 is the constant harvest rate (yield per time) at which members are
removed from the population. (Note that the units of H here are population
per time, whereas in discrete-time harvesting models it had simply population
units.) We assume that the rate of removal is uniform (not seasonal as in many
types of hunting).
Although with a few reasonable assumptions we can prove results describ-
ing the behavior of any constant-yield harvesting model (4.8), for illustrative
purposes we will assume that the populations under study in this section obey
a logistic law, so that f (y) = ry(1 − y/K). Then the population size satisfies
a differential equation of the form
y
y ′ = ry 1 − − H, (4.9)
K
where r, K, and H are positive constants. While this equation can be solved
explicitly by separation of variables, the integration must be handled by ex-
amining three different cases depending on the values of the constants (see
Exercise 1). It is simpler (and probably more informative) to carry out a
qualitative analysis.
r 2
The equilibria of (4.9) are the solutions of ry − K y − H = 0, or
HK
y 2 − Ky + = 0. (4.10)
r
These are given by " r #
1 4HK
y= K ± K2 − . (4.11)
2 r
We must distinguish three cases:
4HK rK
(i) K 2 − r > 0, i.e., H < 4 ,
4HK rK
(ii) K 2 − r = 0, i.e., H = 4 ,
4HK rK
(iii) K 2 − r < 0, i.e., H > 4 .
The number and nature of the equilibria are different in the three cases.
In case (i) there are two distinct equilibria given by (4.11), and it is clear
that one, which we shall call y1 , is smaller than K 2 while the other, which
we shall call y2 , is larger than K 2 . If we let f (y) = ry − Kr 2
y − H, then
2r 2r K K
f (y) = r − K y = K 2 − y . Thus f (y) > 0 if y < 2 , and f ′ (y) < 0
′ ′
if y > K 2 . We now see, using the equilibrium stability theorem, that any
equilibrium y ∗ of (4.9) with y ∗ > K/2 is asymptotically stable, while any
equilibrium y ∗ with y ∗ < K/2 is unstable. More specifically, in case (i), we
see that y1 is unstable while y2 is asymptotically stable.
To consider the global picture for case (i), we note that since f (y) < 0 for
0 < y < y1 , any solution y(t) with 0 < y(0) < y1 is monotone decreasing and
reaches zero in finite time. When this happens, we consider the population to
Nonlinear Differential Equations 205
have been wiped out and the model to have collapsed. If instead y(0) > y1 ,
the solution y(t) approaches the limit y2 as t → ∞. This is illustrated in
Figure 4.12 with r = 2, K = 2, H = 95 .
In case (ii) there is a single equilibrium K
2 which is a double root of (4.10),
′ K
and f ( 2 ) = 0. The equilibrium stability theorem does not apply here, but if
we rewrite f (y) as
2
r 2 rK r 2 rK r K
f (y) = ry − y − =− y − ry + =− y− ,
K 4 K 4 K 2
0 0.5 1 1.5 2 2.5 3 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.5 1 1.5 2 2.5 3
t t t
If we now compare the three cases, we can understand the overall effect
of constant-yield harvesting on a population with logistic growth. Recall that
with no harvesting (H = 0) the population approaches the carrying capacity
K regardless of its initial size. As H increases from 0 to rK 4 , the limiting
population size y2 decreases monotonically from K to K/2, while at the same
time the minimum population size y1 required for survival increases from 0
toward K/2. That is, low harvest rates affect the population in two ways:
206 Dynamical Systems for Biological Modeling: An Introduction
(i) (ii) (iii)
0 y y K 0 K/2 K 0 K
1 2
FIGURE 4.15: Phase lines for (4.9) in the three cases given in the text. Note
the “semi-stable” equilibrium in case (ii).
by lowering its long-term size and by requiring a minimum size for survival.
As the harvest rate increases, the long-term size and minimum size approach
each other, narrowing the window for survival. When the harvest rate reaches
the critical value Hc = rK 4 , these two values meet at K/2, half the original
carrying capacity.
However, for harvest rates beyond the critical value (H > rK 4 ), the pop-
ulation is wiped out in a finite amount of time. Thus the limiting population
size decreases continuously (from K to K 2 ) as H increases from 0 to rK/4,
and then drops to zero as H passes through the critical value rK/4. Such a
discontinuity in the limiting behavior of a system is called a (mathematical)
catastrophe. The corresponding biological catastrophe is the extinction of the
population.
The implications of this model for the management of populations under
a quota system are entailed in the critical harvest rate rK/4. In cases such as
fishing and logging where harvesting is tied to industry, the economic goal of
maximizing the harvest rate must be tempered by an awareness of the fragility
created by harvesting. In this case, the harvest rate must be held sufficiently
far below the critical rate to ensure that occasional or seasonal fluctuations
in reproduction or mortality do not cause the population to drop below the
minimum survival size.
As suggested at the beginning of this section, the logistic growth model
is representative of a wide range of population growth functions. It turns out
that for any general population model of the form y ′ = f (y) whose per capita
growth rate is a monotone decreasing function of population size, and positive
for small y but negative for large y,5 the response to constant-yield harvesting
is similar. (The per capita growth rate in the logistic model, r(1 − y/K), fits
this description.) For such models, with initial conditions close to the natural
carrying capacity, we can show (see Exercise 3d for a special case) that the
limiting population size is positive when H is less than a critical harvest rate
Hc , but drops to zero as H passes Hc . We observed similar behavior for discrete
models in Section 2.5, but for discrete models the situation is complicated by
the possibility that all equilibria may be unstable.
Note that for the logistic model, the critical harvesting rate Hc = rK/4
is the maximum value of the function f (y) = ry(1 − y/K). This is not a
5 Note that these conditions are similar to our definition of scramble competition for
Photo courtesy Wisconsin Dept. of Natural Resources Graph courtesy Wisconsin Dept. of Natural Resources
Example 1.
Wisconsin is one of many states in the U.S. where human expansion has
removed most of the predators of the deer native to the area, with the result
that the deer population grows unchecked, until deer frequently appear on
roads and in yards, causing accidents and damage to gardens and shrubs.
As a result, the Wisconsin Department of Natural Resources (WDNR) issues
hunting permits for a fixed number of deer to be removed each hunting season,
in order to manage the burgeoning population.
The Wisconsin Department of Natural Resources (WDNR) first estab-
lished population goals for white-tailed deer (Figure 4.16) in 1962. As deer
range expanded and hunting interest increased, the goal has grown, to the
point that in the early 21st century it stands at about 700,000. As can be seen
in Figure 4.17, the deer population has continued to grow steadily nonethe-
less; beginning in the mid-1990s, winter populations have been about 50% over
goal. In the 2006–2007 season, just over 500,000 deer were hunted, but by the
fall of 2007, the population had reached about 1.7 million deer, in keeping
with the reproductive rate of about 0.5/yr which can be estimated from the
graph.
(a) What is the minimum equilibrium population value we can infer from
the continued growth shown in the graph together with these values for
H and r?
(b) If hunting half a million deer per year eventually keeps the population
around 3 million deer, how many deer would there be with no hunting?
(This question requires many unrealistic simplifying assumptions, such
as keeping the deer within a fixed habitat.)
208 Dynamical Systems for Biological Modeling: An Introduction
(c) Using this value of K, what are the critical harvesting rate and resulting
equilibrium level?
(d) Using this value of K, what harvesting rate H would meet the WDNR
goal of 700,000 deer?
Solution:
(a) From the graph, we can see that the population keeps growing even with
1/2 million deer harvested per year. Therefore, harvesting is presently
below the critical rate, that is, H < rK/4. Solving for K, we have K >
4H/r = 4(0.5)/0.5 = 4 million deer as a minimal carrying capacity. From
(4.11), the stable equilibrium must be greater than K/2 = 2 million deer.
(b) Solving the equilibrium condition (4.11) for K, we get (after some alge-
bra)
y ∗2
K= ∗ H.
y − r
Substituting r = 0.5/yr, H = 0.5M deer/yr, y ∗ = 3M deer yields a value
of K = 4.5M deer.
(c) Hc = rK/4 = 0.5625M deer/yr, and the equilibrium there is K/2 =
2.25M deer.
(d) As mentioned in the answer to (a), the constant-yield harvesting model
(4.9) cannot give a stable equilibrium value lower than K/2. Of course,
our estimates for r and y ∗ are very rough, but in general the true carry-
ing capacity must be less than twice the goal, in order for the goal to be
reachable through constant-yield harvesting. Since present pre-hunt pop-
ulations are more than twice the present population goal, the goal can
only be reached with variable yield harvesting: that is, harvesting that
depends upon the population size. WDNR continues to adjust their har-
vest quotas annually; we will consider one type of population-dependent
harvesting later in this section.
The great whales once numbered in the millions, but whaling emerged as
a major industry beginning in the eighteenth century, hunting whales of sev-
eral types for their oil and other products, and by the late twentieth century
many species had been reduced to just a few thousand members each (see Ta-
ble 4.16 ). As a result, in 1986 the International Whaling Commission (IWC)
implemented a moratorium on commercial whale hunting, and began a careful
study of the different species’ ability to recover, while setting guidelines for
6 Data from G. A. Knox, The key role of krill in the ecosystem of the southern ocean with
special reference to the convention on the conservation of Antarctic marine living resources,
Ocean Management 9: 113–156, 1984.
Nonlinear Differential Equations 209
TABLE 4.1: Population data (in thousands) for five species of whales prior
to the 1986 IWC moratorium, from Knox, 1984.
Species \ Year 1900 1910 1920 1930 1940 1950 1960 1970 1980
Fin 400 380 360 320 275 220 140 90 80
Blue 200 185 150 95 30 20 10 7 3
Humpback 100 95 70 50 35 20 10 5 3
Sei 95 90 87 83 80 76 70 50 20
Minke 25 25 25 30 34 38 65 96 100
Example 2.
The International Whaling Commission decided in 1965 to protect blue
whales, whose population decreased to fewer than 2000 before the moratorium
was enacted, in spite of an estimated carrying capacity of 200,000 whales and
intrinsic growth rate of 0.05/yr, as a result of annual catches averaging 13,000
whales per year over the period 1926–1940. (a) Assuming that blue whales
obey a logistic law, estimate the maximum number of whales that can be
caught per year without wiping out the population. (b) The IWC estimated a
population of 2,300 blue whales (excluding pygmy blues) in the southern hemi-
sphere in 1997–1998, and a growth rate of 0.082/yr. If we make the simplistic
assumption that the southern hemisphere’s carrying capacity for blue whales
is half that for the entire planet, how long will it take before the population
reaches the IWC’s goal of 54% of carrying capacity using their figures?
Solution: (a) The critical harvest rate is rK/4 = (0.05)200, 000/4 = 2500
whales/yr, less than one fifth the historical catch rate cited.
210 Dynamical Systems for Biological Modeling: An Introduction
Example 3.
Another species of whale hunted to near-extinction in the early twentieth cen-
tury is the right whale,whose post-moratorium fate in the two hemispheres
has been quite different. Southern right whales off the coast of Argentina
numbered about 7500 at the end of the twentieth century, according to the
IWC, and reproduction rates of approximately 0.07/yr have been cited. They
numbered about 100,000 a century earlier. However, the North Atlantic right
whale is now extinct in the eastern Atlantic, and numbered about 300 in the
western Atlantic as of the end of the twentieth century. In the first fifteen years
of the IWC moratorium, their mortality rate increased more than fivefold, al-
though they too numbered about 100,000 a century before. This population
has been observed to be declining at a rate of 0.024/yr.7 (a) If some southern
right whales were to be transported somehow to bolster the northern popu-
lation, what removal rate could the present southern population sustain and
still make progress toward carrying capacity? (b) If we suppose the northern
right whales to be in exponential decline at the given rate (a logistic law makes
no sense in this case) and add whales from the southern hemisphere at the
rate identified in (a), what will the expected equilibrium value be?
Solution: (a) In order for removal not to threaten the southern population,
the present population level must be above the lower equilibrium y1 in (4.11).
If we set y1 = 7500 whales, r = 0.07/yr, and K = 100, 000 whales and solve
for H, we get H = 485 whales/yr.
(b) Exponential decline with importation at a constant rate yields the
mixing equation y ′ = −ay + b studied in Section 3.5, the equilibrium value for
which is b/a. Taking b = 485 whales/yr and a = 0.024/yr yields b/a = 20, 200
whales, a large enough population that one might hope its decline could be
reversed.
The plan suggested in the above example is overly simplistic for many rea-
sons but illustrates the strategy of importing members of related populations
to save endangered populations that has been used successfully with some
7 Hal Caswell, Masami Fujiwara, and Solange Brault, Declining survival probability
threatens the north Atlantic right whale, Proceedings of the National Academy of Sciences
USA 96(6): 3308–3313, March 16, 1999.
Nonlinear Differential Equations 211
other species, especially birds. The failure of some species, like the northern
right whale, to recover following the cessation of decades of overharvesting is
not yet understood, and remains the subject of intense study. One possibility
is a lack of sufficient diversity (genetic and otherwise) in small populations,
but there are many other, behavioral aspects of these populations’ lives that
have been completely disrupted and are not well-known. Some dolphin species
are suffering a similar problem, as we shall see later in this section. A few other
whale populations are considered in the exercises at the end of this section.
Logging is another major industry, producing both wood and paper for uses
around the world. This type of harvesting has great impact on the world’s
forests. For example, clear-cutting—a technique in which all the trees in a
given area are cut down, leaving bare ground (see Figure 4.20)—has destroyed
many primary (old-growth) forest habitats. In population terms, clear-cutting
has the effect of reducing the area of wooded land and thus the carrying
capacity of the system. However, in some places lumber and paper companies
replant trees on the lands where they cut, so that, with many such areas
in a given region, each in a different stage of regrowth, the total wooded
area remains constant, even though the harvesting is ongoing. In the United
States, conservation efforts are holding the amount of forested land roughly
constant. About one third of the U.S., or 303 million hectares,8 is forested
(including Alaska). Of this, just over one third (104.2 million hectares) is
classified as primary forest, the most biodiverse form of forest. In general, the
world’s temperate forests (including those of the U.S. and Canada) are being
roughly conserved. However, the world’s tropical forests, almost all of them
in developing countries, are disappearing at a rate of 8% per decade, while
global wood-consumption is growing at 26% per decade.
Logging yields depend on the intensity of forest management. Intensive
management (used only in Europe as of the end of the twentieth century)
8A hectare is 10,000 m2 .
212 Dynamical Systems for Biological Modeling: An Introduction
MSY
6
intensive
5.5 production
5
4.5
4
3.5 consumption semi-intensive
production
t
2 4 6 8 10 12 14
Photo by Steve Hillebrand, courtesy U.S. FWS
FIGURE 4.21: Graphs of estimated
FIGURE 4.20: This photo of a global wood production vs. consump-
clear-cut hill in Oregon was taken in a tion, in 109 m3 /yr vs. years since 1984,
study of habitat loss for the northern in Example 4.
spotted owl. .
TABLE 4.2: Estimated maximum sustainable stemwood yield for the world’s
forests as of 1984, assuming intensive management and, for tropical forests, a
75-year rotation cycle.
Area Productivity MSY
Forest type (106 km2 ) (m3 /km2 /yr) (109 m3 /yr)
Cold temperate closed 10.56 180 1.91
Cold temperate open 3.09 27 0.09
Moderate temperate closed 3.88 280 1.09
Moderate temperate open 1.73 42 0.08
Tropical closed 11.9 207 2.46
Tropical open 7.3 31 0.23
Total 38.46 152 5.86
Example 4.
If we make the wildly unrealistic assumption that all the world’s forests began
intensive management in 1984, and take the conservation rates and consump-
tion growth rates given in the text, along with an estimated 1984 global wood
consumption rate of 3.67 × 109 m3 /yr, in what year would simple exponen-
tial models predict consumption to outpace production? What if we assume
intensive management of only temperate forests?
Solution: An exponential model for production under the intensive maximum
sustainable yield levels in Table 4.2 would give
where t is measured in years since 1984 and MSY in 109 m3 /yr; the first (con-
stant) component comes from temperate woods and the second (exponential)
from tropical woods, shrinking at the rate of 8% per year. Consumption, mean-
while, can be calculated in the same units as 3.67(1.26)t/10 (growth at the rate
of 26% per decade). Although we cannot set these two expressions equal and
solve for t in closed form, by graphing both quantities as functions of time we
can easily find the point at which they intersect, which occurs for t = 8.63,
that is, in 1992 or 1993.
If we instead assume intensive management of only temperate forests, pro-
duction is M SY (t) = 3.17 + 1.35(0.92)t. In this case, consumption outpaces
production at t = 4.65, i.e., in 1988 or 1989. Both production curves are
graphed vs. consumption in Figure 4.21. Of course, without intensive man-
agement consumption is already beyond production at t = 0.
Example 5.
Brazil is home to one of the largest tropical forests in the world. 57%—or about
477.7 million hectares (Mha)—of Brazil was forested as of 2005. Between 1990
and 2005, Brazil lost 8.1% of its forest cover, or around 42.33 million hectares.
This land was largely clear-cut for agricultural or other non-logging reasons, so
that the habitat is effectively lost. Suppose, however, that in 2005 developed
countries began providing subsidies to protect and restore the remaining forest
land. If individuals, meanwhile, continued clear-cutting at the same rate (by
1980, over 72% of Brazil’s forest clearing was due to ranching), what effective
inherent growth rate r would be necessary in order for replanting to save the
forests?
Solution: The average harvesting rate is H = 42.33Mha/15yr= 2.822 Mha/yr.
In order for the constant-yield harvesting model to have an equilibrium, we
must have H < rK/4, from which r > 4H/K = 4(2.822)/477.7 = 0.02363/yr.
At this rate, however, the equilibrium forest area is only half its 2005 size.
This solution also ignores issues of old (primary) vs. new growth.
214 Dynamical Systems for Biological Modeling: An Introduction
Example 6.
A less controversial example of harvesting is the use of laboratory bacteria
cultures to provide a source of bacteria for use in experiments. If a given
culture of bacteria doubles in 12 hours in a nutrient dish capable of sustaining
106 bacteria, what is the maximum sustained rate at which bacteria can be
removed from the dish? How would the answer change if we assume simple
exponential growth rather than logistic?
Solution: The inherent reproduction rate r can be written either as
(ln 2)/12hr = 0.058/hr or as (ln 2)/0.5day = 1.39/day, using the methods
of Section 3.1. The MSY under the logistic constant-yield harvesting model is
rK/4 = (1.39)106 /4 = 347, 000 bacteria/day.
y ′ = ry − H,
the idea of finding a certain number of animals, but the more of them there
are in his fields, the more often he will find them. The key distinction from
constant-yield harvesting is that here the process is not tied to a harvesting
quota; instead, the frequency of encounters is assumed to be proportional to
the population size (an idea central to the notion of contact processes, which we
shall investigate in the next section). For a population governed by the model
y ′ = f (y), constant-effort harvesting produces the equation y ′ = f (y) − Ey.
Once again we shall consider the specific case of a population governed by
a logistic law. In this case the model becomes
y
y ′ = ry 1 − − Ey. (4.12)
K
Clearly y = 0 is one equilibrium of this equation; the other we calculate to be
∗ E
yH = K 1 − . (4.13)
r
Note that this second equilibrium is only biologically meaningful (i.e., positive)
if E < r, that is, if harvesting events occur less frequently than reproduction.
If instead E > r, then the zero equilibrium is the only equilibrium.
To determine local asymptotic stability of the equilibria, we compute the
derivative of the right-hand side of (4.12) and use the criterion
2y ∗
r 1− − E < 0,
K
which simplifies to
1 E 1 ∗
y∗ > K 1− = yH .
2 r 2
If 0 ≤ E < r, then this criterion is clearly true for the positive equilibrium
∗ ∗
yH and false for the zero equilibrium. If, however, E > r, then yH < 0, so
that the stability criterion is true for the zero equilibrium (and false for the
∗
negative equilibrium yH ). In both cases, the local asymptotic stability extends
to global asymptotic stability in the non-negative state space, since there are
no unbounded solutions (f (y) is negative for large y). Figure 4.22 gives phase
line diagrams for both cases.
E<r E>r
0 y* K 0 K
H
FIGURE 4.22: Phase lines for (4.12) in the two cases given in the text.
it gradually drives the population extinct. The extinction in the latter case
is gradual rather than sudden (as with constant-yield harvesting) because as
the population becomes scarcer, its members become harder to find with the
same level of effort. (One might argue that when the population becomes very
small, its members may become too difficult to find to guarantee a constant
yield. Exercise 15a at the end of this section suggests a way to address this
scenario.)
We may wish to compare the maximum sustainable yield (MSY) for
constant-effort harvesting with that obtained for constant-yield harvesting
(rK/4 for a population with logistic growth). The maximum sustainable ef-
fort is a little below r, but recall that the actual harvest rate is Ey rather
∗
than E. We therefore find the maximum of EyH over all possible efforts:
E2
∗ E
M SY = max EyH = max E K 1 − = max K E − . (4.14)
E E r E r
To find the maximum of this last expression with respect to E, we set its
derivative to zero:
2E r
K 1− =0 ⇔ E= .
r 2
Since the MSY is a downward-facing quadratic in E (to see this, just graph
it vs. E), we know that this unique critical point is indeed a maximum. If we
substitute E = r/2 into (4.14), we obtain
(r/2)2
r r r rK
M SY = K − =K − = .
2 r 2 4 4
The reader who is surprised to find the MSY for constant-effort harvesting to
be the same as that for constant-yield harvesting should note that, given a
constant effort E, the population y will eventually settle down to the corre-
∗ ∗
sponding equilibrium yH , so that the product EyH which represents the yield
is asymptotically constant as well. The difference in practice is that constant-
yield harvesting is often managed to obtain the MSY, while constant-effort
harvesting typically describes scenarios in which the effort is fixed by other
considerations than the MSY.
We now consider two applications of constant-effort harvesting to ecology.
number of whales of all species killed commercially during the entire twenti-
eth century is only two million. The two primary species of dolphins affected
are the offshore pantropical spotted dolphin, Stenella attenuata (Figure 4.23),
and the eastern spinner dolphin, Stenella longirostris orientalis (Figure 4.24).
An informative and very readable reference on this issue is the report by Tim
Gerrodette,9 from which Figures 4.25–4.28 are taken.
In purse-seine fishing, a net is set around schools of tuna, often (about
half the time) associated with schools of dolphins (the reasons the two species
swim together are not well understood).10 Once the net is set around the dol-
phins and tuna (Figure 4.25), the net is closed, and a “backdown procedure”
initiated, in which one end of the net is held slightly underwater, so that the
dolphins may escape close to the surface (Figure 4.26). Most of the time, the
dolphins do escape (meaning that any given dolphin may be chased, netted
and freed many times during its lifetime), but sometimes things go wrong,
and the scale and frequency of these operations led over years to the millions
of dolphin deaths cited above, reducing the northeastern spotted and eastern
spinner dolphin stocks to 20% and 30%, respectively, of their population sizes
when the fishery began, as illustrated by the graphs in Figures 4.27 and 4.28.
Early 21st-century population estimates for these spotted and spinner dolphin
stocks are 640,000 and 450,000, respectively.
What is still not understood about this situation is the failure of these
two populations to recover following the halting of this massive harvesting. In
9 Tim Gerrodette, The tuna-dolphin issue, in W.F. Perrin, B. Würsig and
zones where low oxygen content in the water compresses the tuna habitat into a shallow
[oxygen-rich] region near the surface: M.D. Scott, S.J. Chivers, R.J. Olson, P.C. Fiedler,
K. Holland, Pelagic predator associations: tuna and dolphins in the eastern tropical Pacific
Ocean, Marine Ecology Progress Series 458: 283–302, 2012.
218 Dynamical Systems for Biological Modeling: An Introduction
the 1970s the United States passed laws that required American fishing ships
to reduce their dolphin bycatch to “insignificant levels approaching zero.” As
the fleets scaled down, however, tuna fleets from Latin American countries
grew, and so it was only through an international effort, including pressure
from U.S. consumers, who continued to buy most of the world’s commercially
processed tuna, that by the end of the century the annual dolphin mortality
had been reduced from onetime highs in the hundreds of thousands to under
3000. As can be seen in Figure 4.27, the two affected dolphin populations
stopped their steep decline in the mid-1970s, but even into the 21st century,
with relatively low mortality rates, the two populations persist at the reduced
levels to which purse-seine harvesting brought them. Median recovery rates of
1.7%/yr and 1.4%/yr have been calculated for northeastern offshore spotted
and eastern spinner dolphins, respectively.11
In studying the present populations, Gerrodette and colleagues have sug-
11 Two studies with preliminary data suggesting a stronger recovery in the early twenty-
first century are: Tim Gerrodette, George Watters, Wayne Perryman, Lisa Ballance, Esti-
mates of 2006 dolphin abundance in the Eastern Tropical Pacific, with revised estimated
from 1986-2003, NOAA Technical Memorandum NMFS-SWFSC-422, April 2008, available
at https://swfsc.noaa.gov/publications/TM/SWFSC/NOAA-TM-NMFS-SWFSC-422.pdf;
and International Dolphin Conservation Program Scientific Advisory Board,
7th meeting, Updated estimates of NM IN and stock mortality limits, Doc-
ument SAB-07-05, La Jolla, California, 30 October 2009, available at
http://www.iattc.org/PDFFiles2/SAB-07-05-Nmin-and-Stock-Mortality-Limits.pdf.
Nonlinear Differential Equations 219
Reprinted from Encyclopedia of Marine Mammals, 2nd edition, Tim Gerrodette, The tuna-dolphin issue,
pp. 1192–1195, Copyright 2009, with permission from Elsevier
ulations in the eastern tropical Pacific Ocean, Marine Ecology Progress Series 291: 1–21,
2005.
P.R. Wade, G.W. Watters, T. Gerrodette, and S.R. Reilly, Depletion of spotted and spin-
ner dolphins in the eastern tropical Pacific: modeling hypotheses for their lack of recovery,
Marine Ecology Progress Series 343: 1–14, 2007.
220 Dynamical Systems for Biological Modeling: An Introduction
Example 7.
Assume the two dolphin species are effectively at equilibrium. (a) If the ob-
served depression of the two dolphin species is to be explained by de facto
constant-effort harvesting, what is the effective frequency of these harvest-
ing events? (b) If instead harvesting “effort” is limited to the 3000 dolphins
per year cited in the discussion above, and the population depressions are
to be explained by changes to the baseline parameter r, what reproductive
rate would allow such limited harvesting to hold the population so far below
carrying capacity? (c) If harvesting effort is set as in (b) and the population
equilibrium is to be explained instead by changes to K, what is the resulting
carrying capacity from which the present populations are being depressed?
Solution: For the northeastern offshore spotted dolphins we have an estimated
present population of 640,000, original (pre-fishing) population of 3.2 million,
and r = 0.017/yr. For the eastern spinner dolphins we have an estimated
present population of 450,000, original (pre-fishing) population of 1.5 million,
and r = 0.014/yr.
(a) If we assume constant-effort harvesting, then the original population is
∗
the carrying capacity K and the present population is the equilibrium yH =
∗
K(1 − E/r), from which E = r(1 − yH /K). For spotted dolphins this gives
E = 0.017(1 − 0.64/3.2) = 0.0136/yr. For spinner dolphins this gives E =
0.014(1 − 0.45/1.5) = 0.0098/yr.
(b) We calculate effort from the given figures for harvesting, Ey ∗ , and the
present equilibrium, y ∗ . First, we arbitrarily divide the 3000 harvesting deaths
between the two species in proportion to their population sizes:
640, 000
3000 × = 1761
640, 000 + 450, 000
deaths/yr for the spotted dolphins, and the remaining 1239 deaths/yr for
the spinner dolphins. Then we divide: E = 1761/640, 000 = 0.00275/yr for
the spotted dolphins, and for the spinner dolphins E = 1239/450, 000 =
0.00275/yr. We now solve the equilibrium condition (4.13) for r:
E 0.00275
r= y∗ = = 0.00344/yr
1− K 1 − 0.2
for the spotted dolphins, and similarly we find r = 0.00393/yr for the spinner
dolphins. Both of these values are exceptionally low: that is, the reproductive
rates would have to be depressed by an order of magnitude from the usual
range of cetacean reproductive rates (a common estimate is 0.04/yr) in order
for changes in r to explain the depression. This would include both a reduced
birth rate and high calf mortality.
(c) We use the E = 0.00275/yr value calculated in (b) and the r and
equilibrium (y ∗ ) values from the literature to calculate K: for the spotted
dolphins
y∗ 640, 000
K= = ≈ 760, 000,
1 − Er 1 − 0.00275
0.017
Nonlinear Differential Equations 221
Reprinted from Encyclopedia of Marine Mammals, 2nd Photo by John and Karen Hollingsworth, courtesy
edition, Tim Gerrodette, The tuna-dolphin issue, U.S. Fish and Wildlife Service.
pp. 1192–1195, c 2009, with permission from Elsevier.
and for the spinner dolphins we likewise compute K = 560, 000 as the envi-
ronmentally reduced carrying capacity.
The sandhill crane (Grus canadensis, Figure 4.29) is the most populous of
the world’s cranes, with over 500,000 distributed across North America, from
the tip of Siberia to Cuba. There are six recognized subspecies, three of them
migratory and three nonmigratory subspecies restricted to very small habitats
(and populations) in Mississippi, Florida and Cuba. Most migratory sandhill
cranes pass through the Platte River in the northern U.S. where they form
breeding pairs that are typically lifelong.
Sandhill cranes have lost much of their original wetlands habitats to human
development, and this continues to be the main threat to their existence (espe-
cially loss of roosting grounds on the Platte). However, overhunting also poses
a major threat to some populations in the central U.S. and Canada, along with
collisions with vehicles, utility lines, and fences. Some farmers hunt the cranes
because of the damage they do to crops, in areas where their original habitat
has been drained for agricultural purposes. The rate at which these sandhill
cranes encounter farmers (and their crops), and other manmade artifacts, is
proportional to the population size, rather than occurring at a fixed rate as
with quota-managed populations. For this reason we can model these encoun-
ters as constant-effort harvesting (although see Exercise 11 for an application
of constant-yield harvesting).
222 Dynamical Systems for Biological Modeling: An Introduction
Example 8.
It has been estimated13 that of the 420,000 mid-continental sandhill cranes,
between 25,000 and 32,000 cranes are lost each year to hunting or collisions.
If we take the current population as an equilibrium level (this population is in
fact believed to be stable, although the present population size may not be a
true equilibrium), what is the average length of time before a crane has such
a fatal encounter, and what is the true carrying capacity K if we assume the
cranes’ intrinsic growth rate to be 0.098/yr?
Solution: If the current population is at equilibrium (y ∗ = 420000), then the
harvest rate is Ey ∗ . If we take an average of the two extremes given, this gives
Ey ∗ = 28500 cranes/yr, from which E = 28500/420000 = 0.0679/yr. The
average time before an encounter happens is the reciprocal of this figure, or
14.74 years. To find the carrying capacity, we solve the equilibrium condition
(4.13) for K, to get
∗
yH 420000
K= = = 1.37 million cranes.
1 − Er 1 − 0.0679
0.098
and conservation action plan. IUCN, Gland, Switzerland, and Cambridge, U.K. 294pp.
Northern Prairie Wildlife Research Center Online. http://www.npwrc.usgs.gov/resource/
birds/cranes/index.htm (Version 02MAR98).
Nonlinear Differential Equations 223
Example 9.
In Example 1, Section 3.2 we fitted the population of the United States to the
curve
265
y= , (4.15)
1 + 69 e−0.03t
with population size measured in millions and time measured in years from
1790. Suppose the population is governed by a logistic differential equation
for which this is the solution but that in addition there was an immigration
of one million people per year beginning in 1790. What would be the limiting
population size under these conditions?
Solution: For the solution y(t) given in (4.15), limt→∞ y(t) = 265, and thus
in the logistic model K = 265. From the form of the solution of the logistic
equation obtained in Section 3.3, comparison with (4.15) shows that r =
0.03/yr. Constantimmigration of M per year would change the logistic model
y r 2
to y ′ = ry 1 − K + M , with equilibria given by ry − K y + M = 0, whose
only positive solution is
" r # " r #
1 2
4KM K 4M
y= K+ K + = 1+ 1+ .
2 r 2 rK
Example 10.
An island with resources to support 1000 people becomes so attractive that
other people begin to hear about it from the inhabitants and come. On average,
each inhabitant’s messages convince inspire someone to immigrate there once
every fifteen years. If, as in the previous example, the normal reproduction
rate is r = 0.03/yr, what will the equilibrium size of the population be?
Solution: This situation corresponds to constant-effort harvesting, where the
harvesting (immigration) is proportional to the size of the population. Here the
immigration rate is Iy, where I = 1/(15yr) = 0.067/yr, and the equilibrium
size, adapting (4.13), is
I 0.067
y∗ = K 1 + = 1000 1 + ≈ 3222.
r 0.03
Once every fifteen years doesn’t sound like much, but in the end, despite the
impossibility of further natural growth (because of resource limitations), peo-
ple are arriving so fast (Iy ∗ ≈ 215 people/yr) that the immigration dominates
the system dynamics.
224 Dynamical Systems for Biological Modeling: An Introduction
4.2.4 Conclusions
In Section 2.5 we found that for discrete-time constant-yield harvesting of
logistic (and other scramble-competition) populations there are both maxi-
2
mum (Hmax = (r−1) 4r K) and minimum (Hmin =
(r−3)(r+1)
4r K) harvest rates
that allow stability since scramble populations can be naturally unstable. In
continuous-time harvesting we have only a maximum (Hmax = rK/4, quite
close to its discrete counterpart) since continuous-time scramble populations
do not have that inherent instability. A similar comparison can be made for
constant-effort harvesting: for example, discrete-time models have zero equi-
libria stable for r − 1 < E < r + 1 (see, e.g., Exercises 9 and 10 in Section 2.5),
while the continuous-time model (4.12) has a zero equilibrium stable for E > r.
We can also distinguish the results of constant-yield or quota harvesting,
where overharvesting leads to catastrophe, from those of constant-effort or
proportional harvesting, where overharvesting leads to a gradual decline. Re-
call, however, that the maximum sustainable harvesting rate turns out to be
the same for both models.
Exercises
1. Solve the logistic constant-yield harvesting equation (4.9) using separa-
tion of variables and the technique of partial fraction expansion. Note
that the quadratic expression in y involved in dy/dt factors differently
(or not at all) in each of the three cases discussed in the text.
∗
2. Solve the logistic constant-effort harvesting equation (4.12) using sep-
aration of variables and the technique of partial fraction expansion.
∗
3. Consider a general population model y ′ = f (y) for which the popula-
tion growth rate f (y) is differentiable, has f (0) = 0, rises to a unique
maximum fmax , and decreases thereafter, eventually becoming negative.
Under constant-yield harvesting, this population size is governed by a
differential equation y ′ = f (y) − H.
(a) Show graphically that if 0 ≤ H < fmax , there are two intersections
of the curve z = f (y) and the horizontal line z = H, corresponding
to equilibria of the differential equation.
(b) Show that the larger of the two equilibria is asymptotically stable
and the smaller of the two equilibria is unstable.
(c) Show that if H = fmax there is only one equilibrium.
(d) Show that if H > fmax there are no equilibria and every solution
reaches zero in finite time.
4. Over 40,000 deer-vehicle collisions were reported in 2007, when the deer
population rose as high as 1.7 million deer. From this data (and the
Nonlinear Differential Equations 225
were 120 cranes in the refuge. About how fast should the population be
growing at that point?
14. Suppose a small breeding population of 10 chipmunks is displaced into
a neighborhood capable of feeding 100 of them, but with an unfriendly
cat whom they encounter only very occasionally, about 5 times a year.
Assume the chipmunks’ reproduction satisfies a logistic model with
r = 0.3/yr, and that the feline encounters can be considered constant-
effort harvesting. How many chipmunks will there be after ten years? At
equilibrium? (Note: This is, of course, a facetious example, as at such
small sizes, stochastic effects dominate.)
15. In reality, the harvesting rate may depend both upon population size
and upon external (e.g., economic) considerations. Consider a scenario
in which harvesting activities are driven by a desired (constant-yield)
harvest rate H, but when harvesting begins to drive the population size
down, harvesting becomes more difficult, and is instead dependent upon
the (constant) effort expended to harvest them. In this case the effective
harvest rate becomes
Ey, y ≤ H/E;
R(y) =
H, y > H/E.
Since dA dC
dt + dt = 0, we can deduce that A + C must be a constant, call it KA ,
so that A(t) + C(t) = A(0) + C(0) = KA , and we can write A in terms of C:
A(t) = KA − C(t).
∗ 1n p o
C± = (KA + KB + Kd ) ± (KA + KB + Kd )2 − 4KA KB , (4.18)
2
where the ratio Kd = kk− +
of the reverse reaction rate to the forward reaction
rate is called the dissociation constant. (Note that the units for Kd are in
moles (amount), the same as KA and KB , because the units for k− and k+
are different.) A careful inspection of the expression inside the radical will show
that it is always positive, but smaller than (KA + KB + Kd ). Therefore both
equilibria are always positive. To determine their local asymptotic stability,
we calculate the derivative of dC/dt = f (C):
f ′ (C) = k+ (2C − KA − KB − Kd ).
We find that f ′ (C) < 0 if and only if C < 12 (KA + KB + Kd ). Since this is true
∗ ∗
for C− and false for C+ , the former must be locally asymptotically stable, and
the latter unstable (see the phase portrait in Figure 4.30).
0 *
C_ C*+ C
This appears to suggest that there is a region in phase space for which the
amount of chemical C grows without bound; however, this is illusory, as the
∗
equilibria C± are defined in terms of the (presumably known) initial condition
∗
C(0), and we can show that C(0) is well below the threshold level C+ :
∗ ∗
C− + C+ 1 1
= (KA + KB + Kd ) = C(0) + (A(0) + B(0) + Kd ) > C(0),
2 2 2
∗ ∗ ∗
that is, C(0) is below the midway point between C− and C+ . Therefore C−
is globally asymptotically stable, and the level of chemical C will approach
∗
C− (from which we can also calculate the equilibrium levels of A and B; cf.
Exercise 6).
230 Dynamical Systems for Biological Modeling: An Introduction
Example 1.
The dissociation reaction for an acid with chemical formula HA (where A
represents the ion that bonds with hydrogen to form the acid) is typically
written as
k+
H + + A− ⇋ HA,
k−
The water is commonly omitted since it does not affect the reaction (except
to bond with the dissociated hydronium ions H + ). In this context the disso-
ciation constant for acids is often denoted Ka rather than Kd .
Strong acids (distinguished by having Ka > 1), such as hydrochloric acid
(HCl) and sulfuric acid (H2 SO4 ), dissociate almost completely in water, leav-
ing little of the original acid left, while weak acids (Ka < 1) dissociate very
little. If the dissociation constant for sulfuric acid is 1000, find how much
sulfuric acid remains of 0.50 M (moles) when it is left to dissociate in water.
Solution: Here we begin with no reactants, A(0) = B(0) = 0, and an initial
condition of C(0) = C0 = 0.50 M. From (4.18) we have
1
q
∗ 2 2
C− = (2C0 + Ka ) − (2C0 + Ka ) − 4C0
2
1h p i
= 1001 − 10012 − 1 ≈ 0.00025M.
2
Example 2.
Acetic acid, CH3 COOH, is a weak acid, with a dissociation constant of 1.7 ×
10−5 M. How much acetic acid would remain if 0.50 M were left to dissociate
in water?
Solution: With the same initial conditions as in the previous example, we have
1
q
∗
C− = (2C0 + Ka ) − (2C0 + Ka )2 − 4C02
2
1h p i
= 1.000017 − 1.0000172 − 1 = 0.4999915M ≈ 0.50M.
2
In this case one can easily observe that almost none of the acid dissociates,
whereas in the previous example almost all of it did.
The law of mass action has formed the basis for describing contact pro-
cesses in many contexts within population biology, perhaps most notably
predator-prey relationships and the spread of infectious diseases. We shall
delay our own study of predator-prey systems until Chapter 6, in which we
Nonlinear Differential Equations 231
susceptible’s daily contacts c(N ), multiply by the proportion of infectious contacts I/N and
the number of susceptibles S.
Nonlinear Differential Equations 233
as N → 0 the contact rate remains steady. In reality, the contact rate should
be described as increasing more or less linearly with population size at first,
and then leveling off, possibly but not necessarily smoothly, as the population
size increases.
We can now take a qualitative approach to analyzing this more general
model. By inspection, one equilibrium of (4.20) is I = 0, called the disease-
free equilibrium. Factoring I out of the equilibrium condition, we are left with
N −I 1
c(N ) − = 0,
N τ
from which we find the endemic equilibrium
∗ 1
I =N 1− .
c(N )τ
I∗
′ 1
f (I) = c(N ) 1 − 2 − .
N τ
Left: photo by C.M. Kribs. Right: photo by Daniel Bowen/Public Transport Users Assoc., Melbourne, Australia
should die out. This kind of insight illustrates the reason why such mathemat-
ical models are useful. For this model, one might be able to make this same
argument without going through all the details of a qualitative analysis, but
not so for the more complicated models which arise in studying the problems
of interest to us. Mathematical results give rise to biological insights, and that
is our motivation.
So what difference does the form of the contact rate c(N ) make? In this
case, we have assumed the population size N to be constant, in order to be
able to study the system using a single differential equation. However, in many
more complicated systems, the population size may change during an outbreak
(especially for endemic situations). For populations whose size is on the order
of hundreds, or a few thousands, the mass-action assumption that the contact
rate is directly proportional to population size is reasonable: If you live in
a small town and the population doubles overnight, then you probably will
encounter double the usual number of people throughout the course of your
day: busier sidewalks and stores, fuller buses, etc.
However, at some point one’s ability to contact others in a way that might
transmit infection begins to saturate as the population grows. If you live in a
major metropolitan area like New York or Tokyo and the population doubles
overnight, you will probably not come into contact with twice as many people
on your daily routine, because the sidewalks, subways, and stores are already
all filled to capacity (compare the photos in Figure 4.32). In such a situation,
where the contact rate has saturated, we might instead assign c(N ) = a,
a constant value independent of changes in population size (as long as the
population remains at saturation levels). As discussed above, this assumption
produces the standard incidence model
dI SI 1
=α − I,
dt N τ
Nonlinear Differential Equations 235
Example 3.
Suppose that the infectious contact rate for the common cold saturates sharply
(like csw in Figure 4.33) at a level of α = 1.6/day for population sizes above
1 million, and increases linearly for smaller populations. (a) If the average
infectious period for the cold is 1 day, what is the critical population size
above which an outbreak of the cold will persist? (b) For large populations,
at what endemic level will the infection establish itself? (c) Despite model
predictions, outbreaks are observed in small populations, and outbreaks come
and go in large populations, rather than remain in a constant endemic state.
What limitations on the model are involved in these apparent contradictions?
Solution: (a) For large populations, we have the saturated R0 = ατ =
(1.6/day)(1day) = 1.6 > 1, so that the model predicts persistence of the
infection in populations of over 1 million. For smaller populations, we have
c(N ) = (1.6/day)(N/106) and R0 = (1.6/day)(N/106)(1day) = 1.6N/106, so
that R0 > 1 when N > 106 /1.6 = 625, 000. That is, the model predicts that
the infection will persist in populations of 625,000 people or more.
(b) The endemic equilibrium for populations of over 1 million has
I∗
1 1
= 1− = 1− = 0.375.
N R0 1.6
236 Dynamical Systems for Biological Modeling: An Introduction
That is, the model predicts that, eventually, 37.5% of the population will be
infected at any given time.
(c) The occurrence of outbreaks in small populations does not contradict
the model’s predictions, which simply state that when R0 < 1, eventually the
outbreak will die out. It is possible, however, that an outbreak might grow
within a small subset of the population which makes [potentially infectious]
contacts much more often than the general population—for instance, children
enrolled in school; this is a limitation of the model, which assumes that ev-
eryone in the population has roughly the same average contact rate. A more
detailed model could subdivide the population by contact rate, into high-risk
and low-risk groups.
The periodic disappearance and reappearance of outbreaks of the common
cold in large populations is most often attributed to seasonal factors which
are also not considered in this model, the most evident being weather.
Example 4.
In formulating the epidemic model (4.20), we assumed that the outbreak is
occurring on such a short time scale that demographic changes such as births
and deaths are negligible. How can the model be adapted to account for such
changes without contradicting the assumption of a constant total population
size?
Solution: As discussed in our initial attempts to model population growth,
established populations run up against the carrying capacity of their environ-
mental resources. For an established population, it is therefore reasonable to
assume that, on a time scale of a few years or less, the total birth rate roughly
evens out the total death rate, so that the population size remains constant
even though it is constantly gaining and losing members. We can incorporate
this assumption via a constant per capita mortality rate µ, which can be es-
timated by noting that 1/µ is then the average lifetime in the system. If we
assume that the disease is not transmitted vertically, so that all newborns (or
new recruits) enter the susceptible class, then our model becomes
1
S ′ (t) = µN + I − β(N )SI/N − µS,
τ
1
I ′ (t) = c(N )SI/N − I − µI.
τ
Adding the two equations yields dN/dt = dS/dt + dI/dt = µN − µS − µI = 0,
so N remains constant, and we can again write simply the equation for I,
substituting S = N − I.
Nonlinear Differential Equations 237
We leave the analysis of this model as an exercise but note that the basic
qualitative results remain unchanged from the original model (4.20), and the
quantitative results change only to incorporate µ into R0 .
where c(N ) is the per capita contact rate (making c(N )(P − N )/P the per
capita recruitment rate) and µ is the per capita rate at which individuals die or
drop out of the movement. We shall consider two ways to model the saturation
16 C.M. Kribs-Zaleta, To switch or taper off: the dynamics of saturation, Math. Biosci.
α csw
.
...
............. csm
. . ....
.... (a)
.. . .
..... 0 N
....
...
...... -N
0 A (b)
0 *
N_ N*+ N
FIGURE 4.33: Two different ways FIGURE 4.34: Phase portraits for
to model contact rate saturation for the “Pay It Forward” model (4.21).
large populations.
that occurs as the movement reaches a critical size (at which point the move-
ment is so well known that further growth does not make non-participants any
more familiar with it): smooth and sharp saturation corresponding to those
graphed in Figure 4.33. If we keep α as the maximum per capita contact rate
and A as the critical population level, then we can define the two contact rates
as
csw = α min(N/A, 1), csm = αN/(N + A).
A qualitative analysis of the model (4.21) using smooth saturation csm
reveals that, in addition to the zero equilibrium, a pair of positive equilibria
∗
N± 1 p
= (1 − m) + (1 − m)2 − 4ma
P 2
p
exists if and only if m ≤ 1 + 2a − 2 a(a + 1), where m = µ/α and a = A/P .
Stability analysis using the derivative
α (N + A)(2P N − 3N 2 ) − N 2 (P − N )
f ′ (N ) = −µ
P (N + A)2
shows that the zero equilibrium is locally asymptotically stable (f ′ (0) = −µ <
0), and when the positive equilibria exist, some algebra (which we omit here)
∗ ∗
can show that the lower one N− is always unstable and the upper one N+ is
always stable. This leaves us with a situation where, if individuals retire from
the movement faster than new ones can ever be recruited (m > 1), or if the
critical size at which the movement is well known to the public is large enough
(a > (1 − m)2 /4m), the movement is sure to die out (the zero equilibrium is
globally asymptotically stable, Figure 4.34(a)). If, on the other hand, the
maximum recruitment rate exceeds the retirement rate and that maximum
occurs at a small enough critical size, then survival of the movement depends
Nonlinear Differential Equations 239
∗
upon having a large enough initial core (N (0) > N− , see Figure 4.34(b)). In
this latter case, we have two locally asymptotically stable equilibria, 0 and
∗
N+ ; for this reason the equation is said to be bistable.
The analysis of this same model using the switching function csw to de-
scribe the contact rate is a little more complicated, but not more difficult.
It yields qualitatively similar results to those above (an indication that these
results are robust), and we leave the details as an exercise (see Exercise 12).
We may wonder how the phase portrait in Figure 4.34(a) can become the
one in Figure 4.34(b) simply by shifts in some parameters, or more generally
what consequences this change from a globally stable extinction equilibrium
to a bistable state implies for the future of the movement. These questions
can best be answered through the lens of bifurcations, first mentioned in
Section 2.6 as changes in the qualitative behavior of a dynamical system. In
the next section, we will develop techniques to study bifurcations, and study
or revisit several situations (including this one) where bifurcations identify
the important changes in a system’s dynamics.
Exercises
1. In a chemical reaction a substance S1 with initial concentration K is
transformed into a substance S2 . Let y(t) be the concentration of S2
at time t, so that K − y(t) is the concentration of S1 at time t. If the
reaction is autocatalytic (meaning that the reaction is stimulated by
S2 ), then dy
dt is proportional to y and K − y. Thus
dy
= αy(K − y)
dt
for some constant α. If the reaction is started at time t = 0 by intro-
ducing an initial concentration A of S2 , find the concentration of S2 as
a function of t.
2. In a second-order chemical reaction, a molecule of a substance S1 and a
molecule of a substance S2 interact to produce a molecule of a new sub-
stance S3 . Suppose the substances S1 and S2 have initial concentrations
a and b, respectively, and let y(t) be the concentration of S3 at time t.
Then the concentrations of S1 and S2 at time t are a − y(t) and b − y(t),
respectively. The rate at which the reaction occurs is described by the
differential equation
dy
= α(a − y)(b − y),
dt
where α is a positive constant (cf. (4.17)). Find the concentration y as
a function of t if y(0) = 0.
3. Find the concentration y(t) in Exercise 2 if a = b.
240 Dynamical Systems for Biological Modeling: An Introduction
4. Find the amount of each of the following acids that remains at equilib-
rium after 0.50 M of it is left to dissociate in water:
(a) boric acid, B(OH)3 (Ka = 5.8 × 10−10 M)
(b) iodic acid, HIO3 (Ka = 0.17 M)
(c) trichloroacetic acid, Cl3 COO2 H (Ka = 0.22 M)
5. Give the units for the reaction rate constants k+ and k− in (4.17).
∗
6. Show that the expression for C− in (4.18) is equivalent to the formula
used by chemists,
A∗ B ∗
Kd = .
C∗
with a backward bifurcation, J. Math. Biol. 40(6): 525–540, 2000, Example 4.1.
Nonlinear Differential Equations 241
given period or rate. It also involves using the Chain Rule from calculus to
dz dz dy
write a new differential equation: dx = dy dx . In this case, we need rescale only
the time variable, to eliminate the β; we define s = βt, rename y = I, and
define the bifurcation parameter p = βNβτ τ −1
= Rβτ
0 −1
; then, using the Chain
Rule,
dy dI dt 1
= = β[py − y 2 ] ,
ds dt ds β
so that the epidemic model (4.22) becomes
y ′ = py − y 2 . (4.23)
In this simplified form, we can better see the effects of the bifurcation param-
eter p.
Applying qualitative analysis techniques, we find two equilibria, y ∗ = 0
and y ∗ = p. We have f (y) = py − y 2 , f ′ (y) = p − 2y, so that f ′ (0) = p and
f ′ (p) = −p. Thus if p < 0, the equilibrium y = 0 is asymptotically stable and
the equilibrium y = p is unstable, while if p > 0, the equilibrium y = p is
asymptotically stable and the equilibrium y = 0 is unstable. (It is left as an
exercise to draw the two corresponding phase portraits and verify that there
are no non-negative unbounded solutions, so that local stability extends to
global.)
We can represent this information graphically in a bifurcation diagram by
graphing the equilibrium values y ∗ as functions of the bifurcation parameter
p. We draw solid lines or curves to represent stable equilibria and broken
or dashed lines to represent unstable equilibria. The bifurcation diagram for
equation (4.23) is shown in Figure 4.35. We see that as p increases through
zero, the two equilibria exchange stability; this type of exchange is called a
transcritical bifurcation. We can also now identify the bifurcation point via the
coordinates p = 0 and y = 0, since the two equilibria coincide at y = 0 when
p = 0.
We can also return to the original equation (4.22) and describe the bifur-
cation in terms of I ∗ , the equilibrium number of infectives, and R0 , the infec-
tion’s basic reproductive number (which has more biological meaning than p).
The endemic equilibrium y ∗ = p becomes I ∗ = (R0 − 1)/βτ , and its stability
criterion p > 0 becomes R0 > 1. Since we consider only a nonnegative state
space I ≥ 0 and by definition our epidemiological parameter R0 = βN τ ≥ 0,
we include in the bifurcation diagram (Figure 4.36) only the positive quad-
rant. In general, this bifurcation at R0 = 1, I ∗ = 0, which is so central to the
analysis of even the most complex epidemic models, is normally transcritical.
dN αN P −N
= N − µN. (4.24)
dt N +A P
244 Dynamical Systems for Biological Modeling: An Introduction
y* I*
R0
1
y′ = p − y2. (4.25)
√
This has no equilibria if p < 0, and the two equilibria y = ± p if p > 0.
√ √
Since f (y) = p − y 2 , f ′ (y) = −2y, and we have f ′ (± p) = ∓2 p. Thus
the positive equilibrium is asymptotically stable and the negative equilibrium
is unstable. Figure 4.38 shows the bifurcation diagram for (4.25). The term
bifurcation literally means “splitting,” and it is from situations such as this
that the phenomenon takes its name.
We may also observe that bifurcations are the reason why in some cases
we have needed to draw more than one phase portrait for a given differen-
tial equation. In fact, phase portraits can be seen as vertical “slices” of a
Nonlinear Differential Equations 245
N* y*
b
bc
Having now seen the three most common types of bifurcation that a single
differential equation may exhibit, we can now revisit some other models that
involve bifurcations, to see what type of bifurcation each model involves.
Example 1.
Classify any bifurcations exhibited by the constant-yield harvesting model
(4.9) of Section 4.2.
Solution: Since H is an externally controlled parameter (as opposed to r and
K, which are innate), it is the most sensible choice for bifurcation parame-
ter. As discussed in Section 4.2, the model’s only equilibria, given by equation
(4.11), exist if and only if H < rK/4. At H = rK/4 the equilibria coalesce into
a single point, and for H > rK/4 there are no equilibria. From this descrip-
tion, from their stabilities as determined earlier, and from the corresponding
bifurcation diagram graphing the equilibria as functions of H (see Exercise
8), we can identify the bifurcation at H = rK/4, y ∗ = K/2 as a saddle-node.
Example 2.
Classify any bifurcations exhibited by the constant-effort harvesting model
(4.12) of Section 4.2.
Solution: Here the effort E is the most reasonable choice of bifurcation param-
eter. As discussed in Section 4.2, the model’s two equilibria cross at E = r,
y ∗ = 0, with each one stable on a different side of the bifurcation point. From
this description and the bifurcation diagram (see Exercise 9), the bifurcation
can be identified as transcritical.
248 Dynamical Systems for Biological Modeling: An Introduction
y* y*
1-
---
v----
2
p r
1
--
4
1
----
-v----
2
Example 3.
Describe the behavior of solutions of the differential equation y ′ = ry +y 3 −y 5 .
Solution: Equilibria are y = 0 and the solutions of r + y 2 − y 4 = 0, which is
quadratic in y 2 ; solutions to the latter equation are given by
√
1 ± 1 + 4r
y2 =
2
if r ≥ − 14 . To determine stability, we substitute into f ′ (y) = r + 3y 2 − 5y 4
and find that f ′ (0) = r, so that the zero equilibrium is locally asymptotically
stable if and only if r < 0, while
√ √ √
1
f′
(1 ± 1 + 4r) = − 1 + 4r 1 + 4r ± 1 .
2
∗2
√
Thus for the two equilibria given√ by y√ = (1 + 1 + 4r)/2 which exist for
r ≥ −1/4 we have f ′ (y ∗ ) = − 1√+ 4r( 1 + 4r + 1), implying stability, while
for the two given√by y ∗2 = √ (1 − 1 + 4r)/2 which exist for −1/4 ≤ r ≤ 0 we
have f ′ (y ∗ ) = − 1 + 4r( 1 + 4r − 1) < 0, implying instability. We plot the
bifurcation curve (Figure 4.42), which appears to show a pitchfork bifurcation
√
at r = 0, y ∗ = 0 and two saddle-node bifurcations at r = − 41 , y ∗ = ±1/ 2.
4.4.1 Hysteresis
Hysteresis can be described mathematically by having two (or more) bi-
furcations at each of which a stable equilibrium either vanishes or becomes
unstable, without a direct “hand-off” of stability to an intersecting equilib-
rium. These multiple discontinuities in the limiting value for solutions give the
system what is thought of as “memory”: in the interval between the discon-
tinuities, there are multiple locally asymptotically stable equilibria, and the
one which a given solution approaches depends upon which of the two regions
beyond this interval the bifurcation parameter has most recently occupied.
In the case of Example 3 above, if −1/4 < r < 0, then the solution y(t) ap-
proaches the upper equilibrium if, prior to entering the interval [−1/4, 0], r
was most recently greater than 0, and approaches the zero equilibrium if r
was most recently less than −1/4. In this way a directed loop is formed.
Hysteresis loops have significant biological consequences as well. First, a
small change in environmental parameters may cause a sudden and very large
change in population size. Second, once the change occurs, it may be difficult
to undo: returning to the previous equilibrium requires a large change in the
environmental parameter(s), far beyond the value(s) held before the change.
For instance, in Example 3, if r increases past 0, in order to return the popu-
lation to the zero equilibrium one must get r below −1/4. These consequences
will be easier to understand in context, and so in the remainder of this sec-
tion we will look at the effects of hysteresis through the lens of two biological
systems which exhibit it.
The first such system is the ant foraging model (4.1) with which we began
this chapter,
dx sx
= (α + βx)(n − x) − .
dt x+K
In Section 4.1 we saw that this equation may have either one or three equilib-
ria. The appearance or disappearance of two equilibria signals a bifurcation.
Although the equilibrium condition (4.7) is cubic in the number x of ants
foraging at a given source, it is linear in the colony size n, and we can easily
solve for n in terms of x:
s
n= 1+ x,
(α + βx)(x + K)
which allows us to plot a bifurcation diagram of x∗ versus n. More generally,
when studying the relationship between a bifurcation parameter and some
property of a model (such as equilibria), we may often derive expressions which
are complicated in the state variable (here, x) but simple, even linear, in the
bifurcation parameter. Therefore, although we would really like to write an
250 Dynamical Systems for Biological Modeling: An Introduction
x* /K
5
n
35 40 45 50 55 60
x* /K
4
2
FIGURE 4.44: Foraging Pharaoh ants,
1 Monomorium pharaonis, are considered a
n
major nuisance in many places such as
10 20 30 40 50
hospitals. Their ability to send out new
colonies makes them difficult to eradi-
FIGURE 4.43: Bifurcation dia- cate.
grams for the ant foraging model
(4.1) with and without hysteresis.
expression for the state variable as a function of the parameter, we may instead
do the reverse, writing the parameter as a function of the state variable, in
order to graph it. This approach involves the notion of inverse functions, and
rather than graphing the independent variable on the horizontal axis and the
function on the vertical, we graph the variable (x, which we wish to interpret
as a function of the bifurcation parameter) on the vertical, and the inverse
function (which gives the parameter value) on the horizontal, thus giving us
the graph we wanted in the beginning. This approach may also be referred to as
defining the equilibrium value implicitly in terms of the bifurcation parameter.
Using the parameter estimates given by Beekman et al. and used in Sec-
tion 4.1, we see that the behavior varies with α, the rate at which ants find
the food source independently. The two possibilities are given in Figure 4.43:
either there are no bifurcations, and the stable equilibrium is always unique,
or else there are two saddle-node bifurcations, and a hysteresis loop exists in
the region where three equilibria exist. In the latter case, a growth in colony
size may result in a sudden leap in Pharaoh ants’ (Figure 4.44) collective for-
aging ability at a single site, whereas a drop in colony size below the lower
bifurcation point will suddenly wipe out the ants’ ability to make use of their
fellows’ pheromone trails. Either of these sudden changes will be difficult to
reverse, in that a much larger change in population size is necessary to undo
the change than was necessary to cause it in the first place.
A caveat is in order here: as can be seen from the two cases in Figure 4.43, n
is not the only bifurcation parameter. In fact, through rescaling we can show
(see Exercise 12 below) that our model has three fundamental bifurcation
Nonlinear Differential Equations 251
the spruce budworm and forest, Journal of Animal Ecology 47(1): 315–332, February 1978.
252 Dynamical Systems for Biological Modeling: An Introduction
Reproduced with permission from Natural Resources Reproduced with permission from Natural Resources
Canada, Canadian Forest Service, 2015. Canada, Canadian Forest Service, 2015.
FIGURE 4.45: Mature spruce bud- FIGURE 4.46: Spruce budworm in-
worm larva. duced whole tree mortality in a nat-
ural conifer stand.
pred./H
other prey
0.8 plentiful,
budworms
0.6 rare
budworms plentiful,
0.4
predation high
0.2
y/A
1 2 3 4
FIGURE 4.47: A second-order Hill function illustrates Holling type III pre-
dation, with opportunistic predators preying only on the most populous prey.
Nonlinear Differential Equations 253
These assumptions lead to the following model for spruce budworm growth:
y y2
y ′ = ry 1 − −H 2 . (4.29)
K y + A2
The model involves the four parameters r, K, H, and A. As discussed in Chap-
ter 1, it is often convenient to reduce the number of parameters by rescaling
the model: rescaling both population and time relative to certain benchmark
values will reduce the number of parameters by two (and with only two left, we
can more easily graph the different cases that arise). There are several options
for the rescaling benchmarks. Ludwig et al. used the half-saturation constant
A and the per-budworm-density predation rate H/A (see Exercise 13); here
we shall instead use the carrying capacity K as our population benchmark and
[the reciprocal of] the budworms’ natural growth rate r as our time bench-
mark, as did Robert May in his review of the work of Ludwig et al.21 The
qualitative results will be the same, and we encourage the interested reader
to compare our analysis below with that of Ludwig et al.
We therefore define the rescaled variables x = y/K (budworm larva den-
sity relative to the forest’s carrying capacity for them) and, if we call the
original time variable τ , the new time scale t = rτ , measured in budworm
“generations” (t = τ / 1r , where r1 is the average time for a budworm to re-
produce). We will also find it convenient to define the rescaled parameters
a = A/K, which gives the threshold density for predation relative to the for-
est’s carrying capacity, and h = H/Hc = H/ rK 4 , the predation rate expressed
as a proportion of the critical harvesting rate Hc (identified in Section 4.2).
Invoking the Chain Rule twice,
dx dy dx dτ
= ,
dt dτ dy dt
with dx/dy = 1/K, dτ /dt = 1/r, we obtain the rescaled model
dx h x2
= x(1 − x) − . (4.30)
dt 4 x2 + a2
∗
Looking for equilibria, we find that either x∗ = 0 or (1−x∗ )− h4 x∗2x+a2 = 0.
In the latter case we can multiply through by x∗2 + a2 to make the equation
polynomial:
h
G(x∗ ) ≡ (1 − x∗ )(x∗2 + a2 ) − x∗ = 0.
4
The function G is cubic with G(0) = a > 0 and G(1) = − h4 < 0, so it
2
Solving for h, we get h = 8x(1 − x)2 , from which a2 = x2√ (1 − 2x). Since by
definition a > 0, we take the positive square root: a = x 1 − 2x. This last
expression gives us the necessary bounds on the parameter x: 0 ≤ x ≤ 1/2;
that is, different combinations of a and h produce bifurcations at x values any-
where from 0 to 1/2, and letting x vary from 0 to 1/2 will allow us to trace out
on the a–h plane precisely the combinations that produce those bifurcations.
(We can refer to this approach as reverse parametrization.) Modern scientific
computing software such as Mathematica, Maple and MATLAB can create
such parametric plots, and the result is given in Figure 4.49.
Thus for most values of a and h there are only two equilibria (the un-
stable zero equilibrium and a globally stable positive one), but for small a
Nonlinear Differential Equations 255
h
1.2
x*
1
0.8 capacity
4 2 outstrips
0.6 predation
0.4
0.2 predation
dominates
a S
0.05 0.1 0.15 0.2 0.25 S1 S2
and small enough h—that is, when predation saturates (“turns on”) for even
relatively low budworm densities, and is too weak to keep up with budworm
reproduction—there are four (the unstable zero equilibrium, and two stable
positive equilibria separated by an unstable one).
We now return to the context in which we developed the model: the bud-
worm population developing on a fast timescale, and the forest on a slow
timescale. In order to see what effect the forest’s growth and decay has on the
budworm population, we need to establish what relationship exists between
the two. Of the model’s four original parameters, the carrying capacity K and
predation saturation threshold A were both found to be directly proportional
to the state of the forest (in particular, to the leaf surface area density, mea-
sured in branches per acre). If we consider our rescaled parameters a = A/K
and h = H/ rK 4 , we find a to be independent of forest development as the two
linear factors—say A = αS and K = κS for foliage density S—cancel each
other out, while h = 4H/rκS varies inversely with foliage density. Since the
equilibrium condition G(x) = 0 is linear in h, it is a simple matter to solve it
4H
for h in terms of x, and then to substitute into the expression S = rκh , to get
∗
the equilibrium budworm density x implicitly as a function of S:
H x
S= .
rκ (1 − x)(x2 + a2 )
TABLE 4.3: Empirical parameter estimates for (4.30), from Ludwig et al.
(1978).
Parameter Value Units
r 1.52 /yr
α 1.11 larvae/branch
κ 355 larvae/branch
H 43 200 larvae/acre/yr
Looking at the bifurcation diagram, we see that in young forests (or those
largely degraded by infestation), with S small, the budworm density is kept
relatively low, first because the carrying capacity K = κS is low, and second
because the predation threshold A = αS is also low, so that predators choose
the budworms as primary prey. Here A is low because there are as yet so
few [healthy] branches that it does not take predators long to find budworms.
In this way, predation keeps the budworm population in check. As the for-
est grows, both the carrying capacity and the predation threshold increase,
leading to a gradual increase in budworm density. When S reaches the upper
bifurcation point, marked S2 in Figure 4.50, the food supply has become so
great that predation can no longer contain the budworm reproduction, and
the budworm density rises exponentially quickly toward the upper equilib-
rium, where budworm growth is limited only by the carrying capacity of the
food supply. At this upper equilibrium, the budworm population begins to de-
grade the forest, and S begins to decrease. If the deforestation caused by the
budworms causes it to decrease below the lower bifurcation point, marked S1
in Figure 4.50, then the food supply—and cover for the budworms—disappear,
leaving the budworms again dominated by predation. The upper equilibrium
disappears, and the population crashes back down to the lower equilibrium,
where they remain while the forest begins to rebuild, thus completing the
hysteresis loop. Since the forest growth takes decades, these cycles will also,
explaining the roughly 40-year budworm irruptions that were long observed.
The remaining question, given that hysteresis only occurs in our model
for a certain range of parameter values, is whether in fact observations yield
parameter estimates within this range. Ludwig et al. report two sets of pa-
rameter estimates, one based upon general principles and the other based
upon empirical observations. In most cases the two estimates are fairly close.
We reproduce in Table 4.3 the relevant parameters from their Table 1. From
them, we can calculate
√ a = α/κ = 1.11/355 ≈ 0.00313, which is well within
the range a < 1/3 3 ≈ 0.192 where hysteresis occurs.
The phenomenon of hysteresis has been observed in the outbreak and sub-
sequent crash of many different insect populations. There is, however, another
possible behavior for the system which has been observed in practice. This in-
volves human intervention to keep the insect population down by spraying,
in order to avoid the collapse of the forest (of course, this is not incorporated
in our model). In this case, the insect population remains at a high equilib-
Nonlinear Differential Equations 257
rium level, held in check only by ongoing intervention. Such behavior is called
“perpetual outbreak.”
Exercises
In Exercises 1–7, describe the behavior of solutions of the given differential
equation as the parameter r is varied.
1. y ′ = r + y 2 5. y ′ = y + ry 3
2. y ′ = ry − yey y
6. y ′ = ry(1 − y) − y+1
3. y ′ = ry + y 2
y2
4. y ′ = r − y + y 2 +1 7. y ′ = ry + 1 + y 2
14. We can identify the range of values of the parameter a for which the
rescaled budworm model (4.30) undergoes two bifurcations √ as h (or S)
varies, by using the fact that the parametric curve a = x 1 − 2x, h =
8x(1 − x)2 (0 ≤ x ≤ 1/2) identified in the text “turns around” at the
cusp point. That is, as x increases from 0, the curve moves up and to
the right from the origin (that is, da/dx > 0 and dh/dx > 0). At the
cusp point, however, the curve begins to move down and to the left.
Therefore, at the cusp point da/dx = dh/dx = 0. Use this fact to find
the values of a and h at which this occurs.
the system’s behavior for those particular parameter values; the system’s be-
havior for other parameter values may be quite different, even if the differences
in values are slight. Another reason is that all approximations involve errors,
as we will discuss in more detail below.
The basic problem we would like to solve is to estimate the value of the
population y(t) at some particular moment in time T , given an initial popula-
tion size y0 and a description (a differential equation) of how that population
changes over time. In mathematical terms, we want to solve the initial value
problem
y ′ = f (t, y), y(t0 ) = y0 . (4.31)
Numerical analysis works by converting this continuous-time problem into a
discrete-time problem (i.e., a difference equation yk+1 = g(k, yk )), which can
be solved one time-step at a time, just as we did when we first encountered
nonlinear difference equations in Section 2.1. In order to accomplish this con-
version, we first subdivide the time interval of interest, t0 ≤ t ≤ T , into N
subintervals of length h, by defining t1 = t0 + h, t2 = t1 + h = t0 + 2h, . . .,
tN = T = t0 + N h. The number of subintervals N and the length of each
subinterval h are related by T − t0 = N h; that is, once the time interval is
known, choosing the time-step h is equivalent to choosing the number of time-
steps N = (T − t0 )/h. The fundamental idea is to proceed to y(T ) by first
approximating y(t1 ), then y(t2 ), then y(t3 ), etc. Each calculation is based on
an approximate value for y ′ (t) at that step, obtained using the value of the
solution at one or more points (but not involving any limiting process). In
other words, at each step we try to make the approximation yk as close as
possible to the real solution at that point, y(tk ), so that the final value yN is
as close as possible to the real solution y(tN ) = y(T ). A numerical method
must then have a way to use the function f (t, y) from the original differential
equation to derive a function g(k, y) that will produce faithful estimates.
However, a smaller time-step also means more calculations to reach the final
time T , which in turn increases the cumulative round-off error.
In practice, therefore, the choice of step size h is a challenge: If the step size
is too large, the discretization error at each stage may be large, and these errors
may accumulate catastrophically. On the other hand, if the step size is too
small the computation may require a long time, and the large number of steps
may produce large accumulated round-off errors. One common procedure is to
vary the step size, selecting an optimal h value at each stage to minimize the
total error of both kinds, but one may instead simply repeat the calculation
with successively smaller values of h until further reduction of h makes no
significant change in the approximate value of y(T ) for the chosen time T .
In practice, computer software often carries out the numerical approximation
automatically without asking the user to provide the step size and without
indicating what method is being used.
We should also be concerned that the qualitative behavior of the original
(continuous-time) system and its (discrete-time) approximation agree: for in-
stance, that any equilibrium of one is also an equilibrium of the other, and
that each equilibrium is stable for the same range of parameter values in both
models. Since we have seen that even with a simple logistic model can behave
very differently in continuous and discrete time, we must verify each of these
properties for any given numerical method.
y(t + h) − y(t)
y ′ (t) ≈
h
Nonlinear Differential Equations 261
Euler’s method therefore replaces the differential equation (6.18) with the
difference equation
y(t + h) = y(t) + h f (t, y), (4.32)
or in discrete terms (tk = t0 + kh)
to y(t1 ). We now take t = t1 and use the approximation y1 for y(t1 ) in (4.32)
to obtain an approximation
y2 = y1 + hf (t1 , y1 )
Example 1.
Use Euler’s method with a step size of 1 to approximate the solution of the
logistic equation y ′ = ry(1 − y/K) on the interval 0 ≤ t ≤ 10 for r = 1,
K = 10, and y0 = 1.
Solution: Following (4.33), we compute
and so forth, up to y10 ≈ 10. The graphs of the true solution (given by
equation (2.13)), the approximation u(t) based on the above values for yk ,
and a coarser approximation based on a step size of h = 1.5 are shown in
Figure 4.51.
In the example above and in Figure 4.51, we use impractically large step
262 Dynamical Systems for Biological Modeling: An Introduction
y y
10
10
8
8
6
6
4 4
2 2
t t
2 4 6 8 10 2 4 6 8 10
is made up of straight line segments joining the points (tk , yk ) and (tk+1 , yk+1 ).
Each straight line segment has the same direction as the element of the direc-
tion field (Section 3.2) at the point (tk , yk ). A good approximation is one for
which the graph of the function u(t) is close to the graph of the true solution
y(t).
We can also use this simple example to look at errors in more detail. Since
we used large step sizes in Example 1, the approximation error is due primarily
Nonlinear Differential Equations 263
Example 2.
Use the Euler method to estimate the value of e as y(1), where y(t) is the
solution of the initial value problem y ′ = y, y(0) = 1, and determine the
discretization error.
Solution: The Euler method gives
y(t + h) − y(t − h)
= f (t, y(t))
2h
or
y(t + h) = y(t − h) + 2hf (t, y(t)).
For a given time tk , we have tk + h = tk+1 and tk − h = tk−1 , so that we can
rewrite the difference equation as
If the initial value y0 = y(t0 ) and an approximation y1 for y(t1 ) are known,
Nonlinear Differential Equations 265
y2 = y0 + 2hf (t1 , y1 ).
Example 3.
Use the modified Euler method with a step size of 1 to approximate the
solution of the logistic equation y ′ = ry(1 − y/K) on the interval 0 ≤ t ≤ 10
for r = 1, K = 10, and y0 = 1. Compare the accuracy to that of Euler’s
method using the same step size.
Solution: As noted above, before we can apply the modified Euler method we
must first obtain an estimate y1 for y(t1 ). To keep things simple, let us use
the estimate y1 = 1.9 given by the original Euler method (with h = 1) in
Example 1 (a smaller h would give a better y1 ). Then, following (4.34), we
compute
etc., up to y10 ≈ 10. If we compare the results with those given by the simpler
Euler method, we find significant differences immediately: for y(2) ≈ 4.50853,
the modified Euler method gives y2 = 4.078, an error of less than 10%, rather
than the Euler method’s y2 = 3.439 which has an error of 23.7%. At the next
step, for y(3) ≈ 6.90568, the modified Euler method has an error of 2.5%
compared to the Euler method’s 17.5% error. However, in this case the equi-
librium y∞ = K of the difference equation (4.34) corresponding to the logistic
differential equation is not stable, unlike both the true solution and the Euler
method approximation with h = 1. Figure 4.52 graphs the approximations
u(t) based on the Euler method and modified Euler method along with the
true solution y(t) for the given interval; one can see the instability evidenced
by the oscillation as the modified Euler method solution approaches K. There-
fore, although the modified Euler method gives a much better approximation
here during the initial period of growth (say 0 ≤ t ≤ 5) than the simple Euler
266 Dynamical Systems for Biological Modeling: An Introduction
The final numerical method we shall mention in this section is perhaps the
most well known for this purpose: the fourth-order Runge-Kutta method (also
called RK4), which uses the approximation
K1 + 2K2 + 2K3 + K4
yk+1 = yk + h ,
6
where
h h
K1 = f (tk , yk ), K2 = f tk + , yk + K1 ,
2 2
h h
K3 = f tk + , yk + K2 , K4 = f (tk + h, yk + h K3 ).
2 2
An interpretation of these terms is that K1 is the same approximation to y ′ (tk )
given by the Euler method, K2 is a first approximation to y ′ (tk + h2 ) (the slope
halfway between tk and tk+1 ) based on K1 , K3 is a second approximation to
y ′ (tk + h2 ) based on K2 , and K4 is an approximation to y ′ (tk+1 ) (the slope at
the end of the interval) based on K3 . yk+1 then uses a weighted average of
these estimates of the slope y ′ (t) over the time interval [tk , tk+1 ].
It is possible to show that the discretization error at each step is no greater
than a constant multiple of h5 for this method, so that the accumulated dis-
cretization error is no greater than a constant multiple of h4 . This considerable
improvement is evident in the following example.
Example 4.
Use the fourth-order Runge-Kutta method with a step size of 1 to approximate
the solution of the logistic equation y ′ = ry(1−y/K) on the interval 0 ≤ t ≤ 10
for r = 1, K = 10, and y0 = 1. Compare the accuracy to that of Euler’s
method and the modified Euler method using the same step size.
Solution: Here we must first calculate the four Ki prior to finding each yk . We
therefore proceed as follows: To calculate y1 , we find
K1 = ry0 (1 − y0 /K) = (1)1(1 − 1/10) = 0.9,
h
K1 1 − y0 + h2 K1 /K
` ´ˆ ` ´ ˜
K2 = r y0 + 2
= (1)1.45(1 − 1.45/10) = 1.23975,
h
K2 1 − y0 + h2 K2 /K
` ´ˆ ` ´ ˜
K3 = r y0 + 2
≈ (1)1.62(1 − 1.62/10) ≈ 1.35748,
K4 = r(y0 + K3 )(1 − (y0 + K3 )/10) ≈ (1)1.68(1 − 1.68/10) ≈ 1.80171,
of 18.09% for Euler’s method (and the modified Euler method) but only 0.16%
for RK4. This dramatic (by a factor of 100) improvement in accuracy continues
throughout the interval of interest, with the error in later time-steps never even
reaching 1/5 of 1% for RK4, whereas the average magnitude of error is 7%
for Euler’s method and almost 6% for the modified Euler method over the
interval [0,10]. Table 4.4 gives the true solution and the approximations for
all three methods at each step on this interval, rounded to 4 decimal places.
in file MAIN.M:
in file LOGISTIC.M:
t0 = 0;
function yout = logistic(t,y)
tf = 10;
y0 = 1;
r = 1;
[t,y] = ode45(’logistic’,[t0
K = 10;
tf],y0);
yout = r*y*(1-y/K);
ans = interp1(t,y,2)
and MATLAB, respectively. Note that in each case the precise step size and
method are left implicit, although documentation accompanying each software
package describes the methods used. Here we solve the logistic differential
equation used in the foregoing examples.
We may now consider some examples for which the point is not the me-
chanics of the approximation method, but rather the insights that it produces.
Example 5.
In the budworm model (4.29) of Section 4.4, with the parameter values as
given in Table 4.3 (and A = αS, K = κS as discussed therein), suppose that
an initial budworm density is measured of approximately 30,000 larvae/acre
for a given forest.
(a) Find the range of forest densities [S1 , S2 ] (cf. Figure 4.50) for which a
hysteresis loop exists.
(b) If the forest density is 500 branches/acre, what will the budworm
density be in 1 year? 2 years? How long will it take for the budworm density
to be within 10% of its equilibrium value?
(c) If the forest density is 5000 branches/acre, what will the budworm
density be in 1 year? 2 years? How long will it take for the budworm density
to be within 10% of its equilibrium value?
(d) Interpret and explain the differences in the answers to (b) and (c).
Solution: (a) Using the parameter values, we can use a computer or graph-
ing calculator to graph S as a function of the equilibrium rescaled budworm
Nonlinear Differential Equations 269
Log10 [y[t]]
6
5
4
3
2
1
t
2 4 6 8 10
FIGURE 4.56: Solutions to the differential equation (4.29) showing the log
budworm densities of Example 5 over time, for forest densities of 500 (solid
curve) and 5000 (dashed curve) branches/acre.
density x, as described in Section 4.4, and find the two places where the
graph has a horizontal tangent. These prove to be (x, S) = (0.0031366, 12842)
and (0.49998, 320.22), so that S1 = 320.22 branches/acre and S2 = 12842
branches/acre.
(b) Using a computer to solve the differential equation numerically, we find
that y(1) ≈ 20, 644 larvae/acre, y(2) ≈ 39 larvae/acre, and it will take about
3.36 years for the population density to come within 10% of its equilibrium
value of 10.84 larvae/acre.
(c) Numerical solution with S = 5000 branches/acre indicates that y(1) ≈
36, 620 larvae/acre, y(2) ≈ 63, 613 larvae/acre, and it will take about 6.12
years for the population density to come within 10% of its equilibrium value
of 1.75 million larvae/acre.
(d) In the smaller forest (S = 500 branches/acre), the budworm density
drops by a factor of 1,000 in just 2 or 3 years, while in the larger forest
(S = 5000 branches/acre), the same initial density leads to a slower growth,
by a factor of nearly 100. Both forest densities fall within the hysteresis range,
but in one case the initial density given falls below the separatrix (the unstable
equilibrium), causing the density to crash, while in the other that same initial
density lies above the separatrix, leading the density to skyrocket to a major
infestation. In ecological terms, the tenfold increase in forest density allows the
given budworms to reproduce beyond the predators’ ability to keep them in
check, whereas the lower forest density does not provide an abundant enough
food supply for reproduction to overcome predation. Figure 4.56 shows the
two populations over the first ten years, graphed on a logarithmic scale since
the equilibria are orders of magnitude apart.
270 Dynamical Systems for Biological Modeling: An Introduction
Photo courtesy River Alliance of Wisconsin Photo by Jeff Schmaltz, MODIS/Courtesy NASA
FIGURE 4.57: The surface of this lake FIGURE 4.58: This satellite
at Mounds Dam, Wisconsin is covered photo from June 2008 shows phy-
with algae blooms caused by eutroph- toplankton blooms from exten-
ication. Excess nitrogen and phospho- sive eutrophication in the Sea of
rus runoff from fertilizers make microbe Azov (upper right) and where the
populations mushroom, removing oxygen Danube River empties into the
from the water and creating dead zones Black Sea (upper left). The pale
where neither plants nor animals can live. swirls in the Black Sea reflect sed-
imentation.
4.5.4 Eutrophication
Aquatic ecosystems are affected strongly by the concentrations of many
different substances in the water, which determine the ability of plant and an-
imal species to survive. Some of the substances found in bodies of water serve
to feed the organisms in the ecosystem, in particular oxygen and various nutri-
ents, while others may be toxic. The extent to which sunlight is blocked may
also affect plants that use photosynthesis. The level of nutrients in the water
is characterized on a spectrum between oligotrophic, a state with relatively
low levels of nutrients and plant production, and consequently clear water,
and eutrophic, a state with high levels of nutrients and plant production, and
consequently murky water. Eutrophic lakes actually suffer from their over-
abundance of nutrients, as phytoplankton (microscopic photosynthetic plants
that live near the water’s surface) and other microbes use the excess nutrients
to reproduce exponentially, creating so-called blooms which deplete the oxy-
gen in the water that larger plants and animals need in order to survive. Left
unchecked, the eutrophication process can ultimately wipe out an ecosystem.
Eutrophication of lakes and even larger bodies of water has become a
major problem in the past century because of the rise of chemical fertilizers
in agriculture and urban lands, which pump high concentrations of nitrogen
and phosphorus into the soil, much of which runs off into the local water
table, streams, rivers, and lakes. In the twenty-first century, eutrophication
has been observed not only in ponds and lakes (see Figure 4.57) but in the
world’s seas and oceans (see Figure 4.58), including the Sea of Azor and the
Nonlinear Differential Equations 271
ject to potentially irreversible change, Ecological Applications 9(3): 751–771, Aug. 1999.
272 Dynamical Systems for Biological Modeling: An Introduction
fi [p]
L+r
L P
0
P
oligotrophic m eutrophic 0 P
P P P
higher than the maximum slope (derivative) of the recycling function,23 then
the line can only intersect the sigmoid curve once, and for low enough L the
only equilibrium will be oligotrophic (Figure 4.61(a)). Lakes such as these
are called reversible, since lowering the phosphorus runoff rate L will reverse
eutrophication. Reversible lakes tend to be deep and cool (a notable example
is Lake Washington, in the U.S. state of Washington).
If the lake’s phosphorus clearance rate s is not higher than the recycling
rate’s maximum slope, then the potential for reversing eutrophication depends
upon the extent to which the runoff rate L can be reduced. It is not possible
in practice to eliminate all phosphorus runoff, and if the minimum runoff
rate that can be achieved, Lmin , is too high, then there may never be an
oligotrophic equilibrium (Figure 4.61(a)). Such lakes are called irreversible,
since eutrophication is permanent even under the strongest restrictions on
runoff. These lakes are often shallow and experience high phosphorus runoff
rates (a well-studied example is Shagawa Lake in the U.S. state of Minnesota).
Finally, lakes whose clearance rate s falls below the critical level for re-
versibility, but whose minimum phosphorus runoff rate Lmin is low enough
to permit an oligotrophic equilibrium (Figure 4.61(b)), are called hysteretic,
as they undergo hysteresis as a function of runoff control. Such lakes, typ-
ically small and shallow with high recycling rates, can have eutrophication
permanently reversed through a temporary reduction in runoff rates.
It is possible to carry out further bifurcation analysis (see Exercise 14), but
not to find explicit expressions for the equilibrium values, nor conditions under
which three equilibria exist (hysteresis), so we turn to numerical analysis for
further insights.
Carpenter et al. studied the particular case of Lake Mendota, which is
immediately adjacent to the campus of the University of Wisconsin,24 in the
center of the city of Madison. They measured parameter values of Lmin =
3800 kg/yr, s = 0.817/yr, r = 731, 000 kg/yr, m = 116, 000 kg, q = 7.88,
and in a related study, Lathrop et al. (1998)25 published records of the annual
23 See Exercise 13 for the derivation of this maximum slope.
24 where both they and the authors of this book worked
25 Richard C. Lathrop, Stephen R. Carpenter, Craig A. Stow, Patricia A. Soranno, and
274 Dynamical Systems for Biological Modeling: An Introduction
phosphorus input into Lake Mendota during the period 1976–1996, which had
an average input rate of L = 34, 000 kg/yr, a range of 15,000–67,000 kg/yr,
and a mean phosphorus load of 57,600 kg.
If we use these parameter values in our model (4.35), we find Lake Men-
dota to be in a hysteretic state, with equilibria at 41,900 kg (oligotrophic),
77,500 kg (unstable), and 936,000 kg (eutrophic). However, during 14 of the
21 years observations were made, the estimated phosphorus load was in the
range 44,000–70,000 kg. Therefore, it is likely that the lake is closer to an
equilibrium value like the reported mean of 57,600 kg. Indeed, Carpenter et
al. estimated as much as a tenfold uncertainty in the value of the maximum
recycling rate r (that is, the true value of r might be up to ten times as
much as their estimate, or as little as one tenth), and a 100-fold uncertainty
in the value of the eutrophication threshold m. There is likewise some un-
certainty in L, since Lathrop et al. measured phosphorus runoff coming from
the principal streams feeding the lake, but could not measure that coming
directly from some urban areas next to the lake shore. Changes in the param-
eters r and m are not likely to affect the value of the oligotrophic equilibrium
much (cf. Figure 4.61(b)), but we can easily estimate the value of L which
would correspond to an equilibrium at the observed level: At equilibrium, we
q
have L = sP − r mqP+P q = 43, 873 kg/yr, which also generates equilibria at
69,900 kg (unstable) and 948,000 kg (eutrophic). This suggests that the lake
is currently in a largely oligotrophic state, although a one-time influx of as
little as 69, 900 − 57, 600 = 12, 300 kg would be enough to cause the lake to
experience eutrophication toward the upper equilibrium. Numerical analysis
reveals that this change would probably not be detected immediately simply
from data: in the first five years, the phosphorus level would only increase by
1000 kg, after seven years it would be up by about 5500 kg (still less than
10%), but after eleven years it would be within 10% of the 948,000 kg eu-
trophic equilibrium. (We invite the reader to duplicate these results using any
standard numerical software and plot the graph of P (t).)
In terms of management, we can also see that aggressive reduction of
phosphorus runoff (L = Lmin ) could reduce the oligotrophic equilibrium by
more than an order of magnitude, down to about 4650 kg. In such a situation,
beginning at the present level of 57,600 kg, we can use numerical analysis on
our model (4.35) to find that it would then take about 2.81 years to reduce
the phosphorus content in the lake to below 10,000 kg.
Numerical analysis allows us to quantify the rates of change in this study
in ways that help us recognize it in the data we collect. Carpenter et al. go
further to study the economics of managing the phosphorus content in the lake:
certain economic benefits that result from having an oligotrophic lake (such as
recreational activities and the health benefits associated with water quality)
compete with other economic benefits that generate much of the phosphorus
John C. Panuska, Phosphorus loading reductions needed to control blue-green algal blooms
in Lake Mendota, Canadian Journal of Fisheries and Aquatic Sciences 55: 1169–1178, 1998.
Nonlinear Differential Equations 275
Exercises
1. Estimate e by using the modified Euler method to approximate y(1)
where y(t) is the solution of the initial value problem y ′ = y, y(0) = 1.
Compare your results with those of Example 2 using h = 0.1.
In Exercises 2–7, use the technology available to you to approximate the so-
lution of the initial value problem for the given interval.
2. y ′ = −y, y(0) = 1; 0 ≤ t ≤ 5 5. y ′ + 2ty = sin t, y(0) = 1; 0 ≤ t ≤ 1
1
3. y ′ = sin y, y(0) = 1; 0 ≤ t ≤ 1 6. y ′ = |y| 2 , y(0) = 1; 0 ≤ t ≤ 1
1
4. y ′ = sin y, y(0) = 0; 0 ≤ t ≤ 1 7. y ′ = |y| 2 , y(0) = 0; 0 ≤ t ≤ 1
8. What is the maximum allowable step size h that would ensure that the
approximations generated by Euler’s method for solutions of the logistic
equation with parameters as given in Example 1 approach the (correct)
equilibrium?
9. Another relatively simple numerical method which, like the modified
Euler method, uses the idea of the derivative at a subinterval’s mid-
point being a good estimate for the derivative throughout the entire
subinterval, is the midpoint method, which approximates the differential
equation dy/dt = f (t, y) with the difference equation
h h
yk+1 = yk + hf tk + , yk + f (tk , yk .
2 2
10. Use numerical methods to find how long it would take an ant colony
of size 1000 ants governed by equation (4.1) to reach 90% engagement
in foraging, if initially only 25 ants find the source and establish the
trail. Use Beekman et al.’s parameter values of α = 0.0052/min., β =
0.00125/ant-min., s = 1 ant/min., and K = 10 ants.
11. Repeat the previous exercise with a colony size of 500 ants, and conjec-
ture a biological explanation for the difference in times.
276 Dynamical Systems for Biological Modeling: An Introduction
12. In the foraging ant model (4.1) using the parameter values suggested
in Section 4.1 carry out simulations for values of K close to the criti-
cal threshold value determined in Section 4.1 to exhibit the change in
behavior at the threshold.
13. For the eutrophication model of Carpenter et al., find the maximum
slope of the phosphorus inflow rate f1 (P ) by first finding the inflection
point P̄ at which f1′′ (P̄ ) = 0, and then calculating f1′ (P̄ ). (Your answer
should be r/m times a function of q.) Phosphorus clearance rates s
greater than this critical value indicate eutrophication-reversible lakes.
∗
14. For the eutrophication model of Carpenter et al., use rescaling and
the reverse parametrization method applied to the budworm model in
Section 4.4 to identify the parameter values for which there exist three
equilibria:
(a) Rescale both the state variable and the time variable to reduce
the number of model parameters to two. (Think about meaningful
benchmarks by which to rescale each variable.) Write the rescaled
differential equation using the new parameters a (replacing L) and
b (replacing s).
(b) Use the fact that at a bifurcation point not only dp/dt = f (p) = 0
but also f ′ (p) = 0, to write two equations which are linear in the
two new parameters a and b. Solve these equations to find expres-
sions for a and b as functions of p (this is the reverse parametriza-
tion).
(c) Show that (a, b) moves from (−1, 0) to (q, 0) for 0 ≤ p < ∞, with
a increasing in p while b hits a (positive) maximum followed by a
(negative) minimum.
(d) Using the known zeroes of b(p), show that the reverse-parametrized
curve (a(p), b(p)) passes through the positive quadrant if and only
if q > 2. What is the significance of this result?
Miscellaneous exercises
For each of the differential equations in Exercises 1–4, draw the phase line,
find the equilibria, and describe the asymptotic behavior of solutions.
2
1. y ′ = y 2 (y − 1)
y
2. y ′ = 1−y
3. y ′ = ye−y
4. y ′ = y sin y
5. (a) Consider the logistic differential equation
y
y ′ = ry 1 − , y(0) = A
K
with K = 100, A = 25 and several different values of r including
r = 1.5, r = 2.5, r = 3.5, and r = 4.5. Describe the behavior of the
solution for each value of r qualitatively.
(b) Consider the discrete logistic equation
yk
yk+1 = ryk 1 − , y0 = A
K
with K = 100, A = 25 and several different values of r including
r = 1.5, r = 2.5, r = 3.5, and r = 4.5. Describe the behavior of the
solution for each value of r qualitatively.
(c) Consider the alternative discrete version
yk
yk+1 = ryk e(1− K ) , y0 = A
(b) Use the cubic formula to factor B(x); depending upon parameter
values, it will have either one or three real roots. Next, apply the
method of partial fraction expansion to A(x)/B(x). In the case
where B(x) has three real roots, you can rewrite A(x)/B(x) as the
sum of three terms of the form Ai /(x−ri ) (where i = 1, 2, 3). In the
case where B(x) has only one real root, you can rewrite A(x)/B(x)
A1
as the sum x−r + x2A+B
2 x+A3
1 x+B2
. (You may have to look up the details
for both of these techniques.)
(c) Carry out the integration.
(d) Exponentiate both sides of the equation, and use the initial condi-
tion x(0) = x0 to rewrite any constant of integration.
7. Explain why no two consecutive equilibria can both be (locally) asymp-
totically stable. (Consider the behavior of solutions on the interval be-
tween them.)
8. Suppose a population of whales grows according to an exponential
growth law but with constant-effort harvesting. Then the population
size satisfies a differential equation y ′ = ry − Ey. (a) If r = 0.04/yr,
what catch per year will maintain the population at a constant size of
8000 whales? (Note the catch is not E but Ey.) (b) If harvesting takes
place for ten years with E = 0.06, how long will it take after harvesting
ceases until the population returns to its original size of 8000 whales?
9. A forest of 10,000 acres suffers from forest fires which destroy an average
of 1% of its area annually. If the natural growth rate is r = 0.05/yr, find
the equilibrium forest size.
∗
10. (a) Show that the minimum survival threshold y1 in the constant-
yield harvesting model has y1 > H/r for all values of K. (Hint: Work
backward.)
(b) Use calculus to show that y1 → H/r as K → ∞: that is,
" r #
1 4HK H
lim K − K2 − = .
K→∞ 2 r r
11. In the mid-1990s, the Canadian harp seal population was estimated at
4.8 million seals. In 1997 an estimated 344,000 harp seals were killed.
The harp seal’s reproductive rate has been estimated at 7%/year. (a)
Under a constant-yield scenario, what is the minimum overall carrying
capacity, if the observed harvest is sustainable? (b) Use the result of
Exercise 10 to evaluate the sustainability of this hunt.
12. In formulating the epidemic model (4.20), we assumed that individuals
mix randomly in making contacts with others, so that a proportion I/N
Nonlinear Differential Equations 279
281
This page intentionally left blank
Chapter 5
Systems of Differential Equations
283
284 Dynamical Systems for Biological Modeling: An Introduction
Photo by Mark Conlin, Southwest Fisheries Science Ctr, NOAA Fisheries Svc Photo by Stefano Guerrieri
FIGURE 5.1: The most common sharks and fish in the Adriatic Sea, the
subject of Volterra’s landmark study, are the blue shark, Prionace glauca (left),
and members of the goby family (pictured at right is one example, Didogobius
schlieweni).
ally consistent with observations, it gave great insight into the modeling of
predator-prey relationships and led to many important developments.
Let y(t) denote the number of fish and z(t) the number of sharks at time
t. We make the quite unreasonable simplifying assumption that all fish are of
the same kind and that all sharks are identical.1 We assume that the plankton
on which the fish feed is available in unlimited quantities, and thus that the
fish population would grow exponentially in the absence of sharks. Then, if
there were no sharks, the fish population would satisfy an exponential growth
equation
y ′ = λy.
The sharks, on the other hand, are assumed to depend on the fish as their food
supply. We assume that in the absence of fish the shark population would die
out exponentially, and thus that the shark population satisfies an exponential
decay equation
z ′ = −µz
if there are no fish. To describe the interaction between sharks and fish, we
assume that the presence of fish produces a linear increase in the per capita
shark growth rate, and the presence of sharks produces a linear decrease in
the per capita fish growth rate. That is, the rate of encounters between fish
y and sharks z is proportional to the size of each group, making it a multiple
of yz. Thus we assume per capita growth rates of λ − bz for fish and −µ + cy
for sharks, where b and c are constants which describe how sharply the shark
population affects the fish reproduction, and vice versa. This gives us a system
1 This assumption will allow us to write a reasonably simple model and investigate
whether the basic idea of interdependence can account for the observed fluctuations.
Systems of Differential Equations 285
from calculus (an application of the Chain Rule) to obtain a first-order differ-
ential equation for z as a function of y,
dz z(−µ + cy)
= . (5.2)
dy y(λ − bz)
The differential equation (5.2) has variables separable, and may be solved by
the methods of Section 3.4 (which we will now do).
Separation of variables in (5.2) gives
−µ + cy λ − bz
Z Z
dy = dz,
y z
and integration gives
V (y, z) = h. (5.4)
That is, as y and z change (over time, but we have made that implicit here),
they will follow the curve given by (5.4), so that the value of V (y, z) remains
constant, at its initial value h.
If we return briefly to (5.1), by setting both equations equal to zero we
can find that the system has two equilibria: y = z = 0 and y = µ/c, z = λ/b.
V (y, z) is undefined at (0,0) — it approaches ∞ as we approach the origin
286 Dynamical Systems for Biological Modeling: An Introduction
or either axis — but defined everywhere in the interior of the first quadrant;
since we saw above that V (y, z) must remain constant, we can conclude that
(y, z) can only approach (0,0) if we begin on one of the axes: in particular,
the z-axis. That is, the fish and shark populations will die out entirely only
in the uninteresting case where there are no fish, and the sharks die out in
the absence of their food supply. (In the remainder of our discussion below we
will assume that both fish and sharks are present.)
At the other equilibrium, however, we find that
µ λ
V (y, z) = h0 ≡ −µ log + µ − λ log + λ.
c b
In fact, we can show that h0 is the minimum value V (y, z) can ever take on,
by examining the first and second derivatives of V , just as with a function of
one variable. In particular, if we take the partial derivatives of V with respect
to y and z, we find2 that
∂V µ ∂V λ ∂2V µ ∂2V λ
= − + c, = − + b, = > 0, = 2 > 0,
∂y y ∂z z ∂y 2 y2 ∂z 2 z
so the equilibrium y = µ/c, z = λ/b is the only place where both ∂V ∂y = 0 and
∂V
∂z = 0. Furthermore, we see from the second derivatives that the function
is concave up in the interior of the first quadrant, so this critical point must
be a minimum.3 (Here we begin to see our first hint of new and interesting
behavior: since the value of V (y, z) must remain constant, and V (µ/c, λ/b)
is different (less) than the values of V for all the points surrounding it, no
trajectories lead toward this equilibrium, either.) Therefore every solution of
the system (5.1) is described by the relation (5.4) with some choice of h ≥ h0 ,
determined by initial conditions.
Recall now that our purpose in writing this model was to see if the in-
terdependence between the fish and shark populations could account for the
fluctuations observed. In order to see that we do have periodic solutions here,
we will extend the notion of a phase line introduced in Section 4.1. Since we
now have a system of two dimensions rather than one, we will speak of the
phase plane, i.e., the y − z plane. For each value of h ≥ h0 , the relation (5.4)
defines implicitly a curve in this plane. More directly, it gives the relation
between the two population sizes from the differential equation (5.2). It does
2 Note to the reader not accustomed to partial derivatives: the process is just like normal
differentiation with respect to the variable in question, while treating all other independent
variables as constants. The notation ∂V /∂y is used above, for instance, instead of dV /dy, in
order to show that V is a function of more than one variable. As a short example, consider
A(y, z) = y 2 sin z + 4. Then ∂A/∂y = 2y sin z, while ∂A/∂z = y 2 cos z. In this case we have
three different second-order partial derivatives: ∂ 2 A/∂y 2 = 2 sin z, ∂ 2 A/∂z 2 = −y 2 sin z,
and ∂ 2 A/∂y∂z = ∂ 2 A/∂z∂y = 2y cos z.
2
3 We can also note that ∂ V = 0, but we are borrowing just enough from multivariable
∂y∂z
calculus here to make our point.
Systems of Differential Equations 287
not, however, give the populations as functions of the time t. Such a curve is
called an orbit of the system (5.1).
To show that solutions to (5.1) are periodic, we must show that the curves
given by V (y, z) = h are closed. We can, of course, use computer programs
such as Maple, Mathematica or MATLAB to generate numerical solutions
and plot them, and Figure 5.2 below shows such a plot. An analytical way to
show that orbits of (5.1) are closed involves using a polynomial approximation
to the logarithms involved in the formula for V , similar to the linearization
process introduced in Section 4.1 to analyze stability of equilibria. Since it is
often important to be able to show results analytically — here, for example,
to distinguish genuine periodic orbits from spirals which grow or decay very
slowly — we will now consider this approach.
We begin by making the change of variables
µ λ
y= + u, z = + v
c b
in (5.4), to center our system (u, v) around the equilibrium ( µc , λb ). This gives
us
µ λ µ µ
V + u, + v = −µ log +u +c +u
c b c c
λ λ
−λ log +v +b + v = h. (5.5)
b b
is the same approximation process we would use in linearizing about the equi-
librium, as introduced in Section 4.1, but here we are keeping the quadratic
as well as the linear terms.) When we substitute these approximations into
(5.5) we obtain the curve
c2 2 b2 2
µ
λ
−µ log − cu + u + µ + cu − λ log − bv + v + λ + bv = h
c 2µ b 2λ
288 Dynamical Systems for Biological Modeling: An Introduction
or
c2 2 b 2 2 µ λ
u + v = h + µ log − µ + λ log − λ = h − h0
2µ 2λ c b
as an approximation to the curve (5.5). This curve, whose equation in the
y − z plane is
2
c2 µ 2 b2
λ
y− + z− = h − h0 , (5.6)
2µ c 2λ b
represents an ellipse with center at ( µc , λb ) as long as h > h0 .
This shows that the curve (5.6) is a closed curve around the equilibrium
( µc , λb ) if h − h0 is small and positive. Since the solution runs around a closed
curve, it must repeat itself and is therefore periodic. Thus the Lotka-Volterra
model predicts periodic fluctuations as had been observed in the shark and
fish populations. It is possible to prove that the√period of oscillation, or time
for a cycle to repeat itself, is approximately 2π/ λµ. The phase portrait (Fig-
ure 5.2), or sketch of orbits in the phase plane, shows that, because the orbit
is traversed in a counterclockwise direction (see Exercise 7 below), the maxi-
mum prey population (reached at the rightmost end of each orbit) occurs one
quarter of a cycle before the maximum predator population (reached at the
top of each orbit).
2
1.8
1.6
1.4
1.2
z 1
0.8
0.6
0.4
0.2
Now that we have seen an example, let us restate more generally the termi-
nology we shall use in this chapter. The Lotka-Volterra system is an example
Systems of Differential Equations 289
Example 1.
Describe the orbits of the system
y ′ = z, z ′ = −y.
2 2
and integration gives z2 = − y2 + c. Thus every orbit is a circle y 2 + z 2 = 2c
with center at the origin, and every solution is periodic.
290 Dynamical Systems for Biological Modeling: An Introduction
2 2
1.5 1.5
z z
1 1
0.5 0.5
FIGURE 5.3: Direction field and FIGURE 5.4: Direction field and
phase portrait for the Lotka-Volterra nullclines for the Lotka-Volterra
model. model (the coordinate axes are also
nullclines).
Exercises
In each of Exercises 1–6, describe the orbits of the given system.
1. y ′ = yz 2 , z ′ = zy 2 3. y ′ = cos z, z ′ = y
2. y ′ = ez , z ′ = e−y 4. y ′ = yez , z ′ = yzey
We will assume that the equilibrium is isolated, that is, that there is a circle
centered around (y∞ , z∞ ) which does not contain any other equilibrium. We
shift the origin to the equilibrium by letting y = y∞ + u, z = z∞ + v, and
then use Taylor’s theorem for two variables (see Appendix B) to approximate
F (y∞ + u, z∞ + v) and G(y∞ + u, z∞ + v). The difference here between a
one-dimensional system and a two-dimensional one is that the approximation
via Taylor’s theorem in two or more dimensions uses partial derivatives. Our
approximations are
where h1 and h2 are functions which are “quadratic” in u and v, in the sense
that they are negligible relative to the linear terms in (5.10) when u and v are
small (i.e., close to the equilibrium).
The linearization of the system (5.8) at the equilibrium (y∞ , z∞ ) is defined
to be the linear system with constant coefficients
into (5.8). By (5.9) the constant terms are zero, and for the linearization we
neglect the higher-order terms h1 and h2 . The coefficient matrix of the linear
system (5.11) is the matrix of constants
Fy (y∞ , z∞ ) Fz (y∞ , z∞ )
.
Gy (y∞ , z∞ ) Gz (y∞ , z∞ )
Example 1.
Find the linearization at each equilibrium of the Lotka-Volterra system
Solution: As noted in the previous section, the equilibria are the solutions of
y(λ − bz) = 0, z(−µ + cy) = 0. One solution is the origin, y = 0, z = 0,
representing extinction of both fish and sharks, and a second is y = µc , z = λb ,
representing coexistence. Because the partial derivatives of the functions on
the right side of the system are, respectively,
∂ ∂
[y(λ − bz)] = λ − bz, [y(λ − bz)] = −by,
∂y ∂z
∂ ∂
[z(−µ + cy)] = cz, [z(−µ + cy)] = −µ + cy,
∂y ∂z
the linearization at an equilibrium (y∞ ,z∞ ) is
u′ = (λ − bz∞ )u −by∞ v,
′
v = cz∞ u +(−µ + cy∞ )v.
u′ = λu, v ′ = −µv,
bµ cλ
u′ = − v, v ′ = u.
c b
(What these linearized systems tell us about the predator-prey system’s be-
havior is investigated later in this section, cf. Example 3).
Example 2.
Find the linearization at each equilibrium of the system
u′ = −βz∞ u −βy∞ v,
′
v = βz∞ u +(βy∞ − γ)v.
The equilibria are the points (y∞ ,0) with arbitrary y∞ , and the corresponding
linearization is
u′ = −βy∞ v, v ′ = (βy∞ − γ)v.
If we examine the second of these equations, we see that v(t) experiences
either exponential growth or exponential decay, depending on the sign of the
coefficient (βy∞ − γ). If y∞ > γ/β, this coefficient is positive and solutions
to the linearized system exhibit exponential growth in v, which corresponds
to growth in z, the number of infected individuals. Only if y∞ < γ/β does v
decay, corresponding to a drop in the number of infectives. As we will discover
in Section 6.1, this comes from the modeling assumption that infectious con-
tacts, like the predator-prey contacts in the Lotka-Volterra model, occur at
a rate proportional to the product of both populations, yz. Infection control
might therefore concentrate on reducing y, the number of susceptibles who
make potentially infectious contacts; two common ways to do so, illustrated
in Figure 5.5, are vaccination and protection from infectious contact.
An equilibrium of the system (5.8) with the property that every orbit
with initial value sufficiently close to the equilibrium remains close to the
equilibrium for all t ≥ 0, and approaches the equilibrium as t → ∞, is said to
be asymptotically stable. An equilibrium of (5.8) with the property that some
solutions starting arbitrarily close to the equilibrium move away from it is said
to be unstable. These definitions are completely analogous to those given in
Section 2.4 for difference equations and Section 4.1 for first-order differential
equations.
The fundamental property of the linearization which we will use to study
stability of equilibria is the following result, which we state without proof. The
proof may be found in any text which covers the qualitative study of nonlinear
294 Dynamical Systems for Biological Modeling: An Introduction
FIGURE 5.5: Two common strategies for reducing the number of individuals
at risk of infection in an epidemic are vaccination and protection from contact.
At left, a man vaccinates a child (in his mother’s arms) against smallpox in
the Republic of Chad during a worldwide vaccination program that began in
1967 and lasted over a decade. At right, riders on a Mexico City subway train
wear breathing masks during the 2009 H1N1 influenza epidemic.
that this case may indeed occur, recall the Lotka-Volterra model we studied
in the previous section, where periodic solutions arose.
Example 3.
Show that the equilibrium (0,0) of the Lotka-Volterra system
is unstable.
Solution: As we have seen in Example 1, the linearization of the system at the
equilibrium (0,0) is u′ = λu, v ′ = −µv. These two equations can be solved sep-
arately by separation of variables, and every solution of the linearization has
the form u = c1 eλt , v = c2 e−µt . As every solution with c1 6= 0 is unbounded,
the linearization theorem shows that the equilibrium (0,0) is unstable.
If we try to apply the theorem to the interior equilibrium µc , λb , we
must consider the linearization u′ = − bµ ′ cλ
c v, v = b u. Although we have
not yet discussed any way to analyze such linearizations in general, methods
we will discuss later in this chapter will enable us to show that all√solutions
√ linearization are periodic — namely, combinations of sin λµ t and
of this
cos λµ t. Since the linearization’s solutions neither approach zero nor be-
come unbounded, the linearization theorem above does not allow us to draw
any conclusions about the stability of this equilibrium; instead, we must turn
to other methods, such as the phase plane approach we took in the last section.
Example 4.
For each equilibrium of the system
y ′ = z, z ′ = −2(y 2 − 1)z − y,
∂ ∂
[−2(y 2 − 1)z − y] = −4yz − 1, [−2(y 2 − 1)z − y] = −2(y 2 − 1),
∂y ∂z
the linearization at (0,0) is u′ = 0u + 1v = v, v ′ = −u + 2v. We can actually
solve this system of equations outright by using a clever trick4 to reduce it
4 It has been said that the best way to solve a differential equation is by correctly guessing
the solution. It has also been said that a technique is a trick that can be used more than once.
Many techniques (and tricks) that have been developed to solve differential equations take
advantage of some peculiarity of the particular equations under consideration. The trick used
above certainly falls into that category. However, as our approach in this text concentrates
on qualitative methods, we shall not attempt to provide a list of such tricks, referring
the interested reader instead to a more general introductory text on solving differential
equations.
296 Dynamical Systems for Biological Modeling: An Introduction
to a single equation. We subtract the first equation from the second to give
(v − u)′ = (v − u). This is a first-order differential equation for (v − u),
whose solution is v − u = c1 et . This partial result is already enough to tell
us that the equilibrium is unstable, as since the difference between u and v
grows exponentially, at least one of them must therefore grow exponentially
as well. However, to give a more explicit conclusion we shall complete the
solution: Substitution of v = u + c1 et into u′ = v gives another first-order
linear equation u′ − u = c1 et . The solution of this equation, obtained by the
method of Section 3.5, is u = (c1 t + c2 )et , and therefore v = (c1 t + c1 + c2 )et .
As this has unbounded solutions, we (again) conclude that the equilibrium
(0,0) is unstable.
Exercises
In Exercises 1–8, find the linearization of the given system at each equilibrium.
1. y ′ = y + z − 2, z ′ = z − y 6. y ′ = z, z ′ = sin y
′ ′
2. y = y + z − 1, z = y 7. y ′ = y(λ − ay − bz),
′ 2 ′
3. y = y + z , z = y + 1 z ′ = z(µ − cy − dz), a, b, c, d > 0
4. y ′ = z − 2, z ′ = y 2 − 8z 8. y ′ = y(λ − ay + bz),
z ′ = z(µ + cy − dz), a, b, c, d > 0
5. y ′ = ez , z ′ = e−y
In each of Exercises 9–12, for each equilibrium of the given system determine
whether the equilibrium is asymptotically stable or unstable.
9. y ′ = −2y, z ′ = −z 11. y ′ = −y, z ′ = y 2 − z
10. y ′ = y, z ′ = −z 12. y ′ = y + z, z ′ = z − 1
13. Show that the equilibrium (0,0) of the system y ′ = y(λ − ay − bz),
z ′ = z(µ − cy − dz) is unstable (see Exercise 7).
14. Show that the equilibrium (0,0) of the system y ′ = y(λ − ay + bz),
z ′ = z(µ + cy − dz) is unstable (see Exercise 8).
15. How would the epidemic model of Example 2 have to change if
(a) the whole population were partially protected from infection?
(b) part of the population were born [permanently] partially protected
from infection?
(c) part of the population were initially temporarily protected?
(d) the outbreak were extended enough to warrant including demo-
graphic renewal (births and deaths) in the model?
In each case write the new model.
Systems of Differential Equations 297
y ′ = ay + bz, (5.12)
z ′ = cy + dz,
where λ, Y , and Z are constants to be determined, with Y and Z not both zero.
When we substitute the form (5.13) into the system (5.12), using y ′ = λY eλt ,
z ′ = λZeλt , we obtain two conditions
which must be satisfied for all t. Because eλt 6= 0 for all t, we may divide these
equations by eλt to obtain a system of two equations which do not depend on
t, namely
λY = aY + bZ,
λZ = cY + dZ
or
(a − λ)Y + bZ = 0, (5.14)
cY + (d − λ)Z = 0.
for some constants K1 and K2 . The form (5.16) with two arbitrary constants
Systems of Differential Equations 299
K1 and K2 is called the general solution of the system (5.12). If initial values
y(0) and z(0) are specified, these two initial values may be used to determine
values for the constants K1 and K2 , and thus to obtain a particular solution
in the family (5.16).
Example 1.
Find the general solution of the system
y ′ = −y − 2z, z ′ = y − 4z
If the characteristic equation (5.15) has a double root, this method gives
only one solution of the system (5.12), and we need to find a second solution
in order to form the general solution. The characteristic equation has a double
root if the discriminant (a+d)2 −4(ad−bc) = (a−d)2 +4bc = 0. It is possible to
show (and the reader can verify) that if λ is a double root of (5.15), λ = a+d
2 ,
then in addition to the solution y = Y1 eλt , z = Z1 eλt of (5.12) there is a
second solution, of the form
(a − λ)Y2 + bZ2 = Y1 ,
cY2 + (d − λ)Z2 = Z1 .
Thus the general solution of the system (5.12) in the case of equal roots is
Note that if (5.15) has a single root and b = 0, then λ = a = d, and we have
Y1 = 0, Z1 = cY2 , so that the general solution becomes
Example 2.
Find the general solution of the system
y ′ = z, z ′ = −y + 2z,
Example 3.
Find the general solution of the system
y ′ = −2z, z ′ = y + 2z,
y = −2K1 et sin t − 2K2 et cos t, z = (K1 − K2 )et sin t + (K1 + K2 )et cos t.
FIGURE 5.6: Direction field and phase portrait showing a stable node.
FIGURE 5.7: Direction field and phase portrait showing a saddle point.
are shown in Figure 5.8. Every orbit approaches the origin, but the approach
is by an inward spiral. Such an equilibrium is called a spiral point. Here the
roots of the characteristic equation are complex.
The direction field and phase portrait for the system y ′ = z, z ′ = −y are
shown in Figure 5.9. Every orbit is a closed orbit around the origin, and such
an equilibrium is called a center. In this case the roots of the characteristic
equation are purely imaginary.
In the above examples, every orbit approaches the origin for the node
and spiral point, but we could also give examples in which the directions
Systems of Differential Equations 303
FIGURE 5.8: Direction field and phase portrait showing a stable spiral
point.
are reversed and all orbits are unbounded. For a center, no orbits approach
the origin, and for a saddle point there are orbits which approach the origin
and also unbounded orbits. Every equilibrium point for a linear autonomous
system is of one of these four types.
As noted above, the nature of the origin as an equilibrium of the linear sys-
tem (5.12) depends on the roots of the characteristic equation (5.15). If both
roots of (5.15) are real and of the same sign, then the solutions of (5.12) are
combinations of either positive exponentials or negative exponentials. This
304 Dynamical Systems for Biological Modeling: An Introduction
′
implies that the slope of an orbit, which is yz ′ (t)
(t)
, must approach a limit as
t → ∞, and thus that the origin is a node. If the roots of (5.15) are real and
of opposite sign, then there are solutions which are positive exponentials and
solutions which are negative exponentials. Thus there are solutions approach-
ing the origin and solutions moving away from the origin, and the origin is a
saddle point. If the roots of (5.15) are complex, the solutions contain trigono-
metric functions and the orbits oscillate. If the real parts of the roots are zero,
the orbits are periodic, and the origin is a center. If the real parts are different
from zero, the orbits spiral, and the origin is a spiral point.
Example 4.
A tank contains 100 liters of water and 10 kg. of salt, thoroughly mixed. Pure
water is added at the rate of 5 liters/minute, and the mixture is poured off at
the rate of 5 liters/minute into a second tank which initially contains 80 liters
of water. The mixture in the second tank is then poured out as waste, at the
same rate. Formulate and solve a model to describe the weight of salt in each
tank as a function of time.
Solution: Let y(t) denote the weight of salt in the first tank and z(t) the
weight of salt in the second tank at time t. Then y(0) = 10 and z(0) = 0. The
y
concentration of salt in the first tank is 100 kg./liter, and the weight of salt
y y
poured into the second tank per minute is 5 100 = 20 kg. The concentration
of salt in the second tank is z/80 kg./liter, and the weight of salt poured out
z z
is 5 80 = 16 kg. Thus
TABLE 5.1: Data from Feldmann and Schneider [24] showing the concen-
tration q1 of BSP in the blood as a function of time t since introduction.
t (min) 3 6 9 12 15 20 30 40 60
q1 (t) (µg/L) 49 20 14 5 4 3 2 2 1
and models for the pharmacodynamics, in J. Berger et al. (eds.), Mathematical models
in medicine: workshop, Mainz, March 1976 (Berlin/New York: Springer-Verlag), Lecture
Notes in Biomathematics 11: 243–277, 1976.
308 Dynamical Systems for Biological Modeling: An Introduction
As always, real data contains “noise,” so will not fit any theoretical form
(i.e., (5.20)) perfectly. There are many sophisticated numerical and statistical
methods for deriving the best possible estimate for the coefficients in the
equation, but for the sake of simplicity here we will use only some of these
data and derive simple estimates (which nevertheless compare well with those
Jolivet obtained, q.v.). First, we assume that the difference in decay rates
between λ1 and λ2 is great enough that after enough time, most of the fast-
decaying eλ2 t component has dwindled to a negligible level, leaving primarily
the slower-decaying eλ1 t component: q1 (t) ≈ a1 eλ1 t . If we assume that this is
true for the last half of the data, then we can use the concentrations after 20
and 60 minutes to calculate the coefficients a1 and λ1 :
q1 (20) = 3 ≈ a1 e20λ1 , q1 (60) = 1 ≈ a1 e60λ1 ,
log 3 ≈ log a1 + 20λ1 , log 1 = 0 ≈ log a1 + 60λ1 ,
from which log a1 ≈ −60λ1 , and, substituting, log 3 ≈ −40λ1 , making λ1 ≈
(log 3)/40min = −0.0275/min and thus a1 ≈ e−60λ1 = 5.2µg/L (note all
logarithms here are natural, not base ten).
To recover a2 and λ2 we return to the earliest data points (3 and 6 minutes
after injection), where we assume the fast-decaying component can still be
detected:
q1 (3) = 49 = a1 e3λ1 + a2 e3λ2 , q1 (6) = 20 = a1 e6λ1 + a2 e6λ2 ,
49 = 5.2e3(−0.0275) + a2 e3λ2 , 20 = 5.2e6(−0.0275) + a2 e6λ2 ,
49 = 4.788 + a2 e3λ2 , 20 = 4.41 + a2 e6λ2 ,
log 44.212 = 3.789 = log a2 + 3λ2 , log 15.59 = 2.747 = log a2 + 6λ2 ,
so log a2 = 2.747 − 6λ2, making 3.789 = 2.747 − 3λ2, and thus λ2 = −(3.789 −
2.747)/3min = −0.347/min. Then, finally, a2 = e2.747−6(−0.347) ≈ 125µg/L.
This gives the concentration of BSP in the blood of
q1 (t) = 5.2e−0.0275t + 125e−0.347t . (5.22)
From this we see that most of the BSP is removed fairly quickly from the
blood (after an average of 1/λ2 = 2.88min), while the rest takes longer to be
eliminated (an average of 1/λ1 = 36.4min). Figure 5.12 superimposes plots
of the original data, the solution, and the slow-decaying component of the
solution.
To recover the original exchange rates we use a somewhat ad hoc process,
first observing that q1 (0) = 5.2 + 125 ≈ 130µg/L, q2 (0) = 0 so that (from
(5.18)
dq1
(0) = −k21 q1 (0) + k12 q2 (0) = −(130µg/L)k21 ,
dt
while at the same time (from (5.22))
dq1
(0) = 5.2(−0.0275) + 125(−0.347) = −43.518(µg/L)/min,
dt
Systems of Differential Equations 309
Exercises
In each of Exercises 1–14, find the general solution of the given system by
analytic solution, use a computer algebra system to examine the behavior
of solutions, and classify the origin as a node, saddle point, center, or spiral
point.
1. y ′ = y + 5z, z ′ = y − 3z
2. y ′ = y − z, z ′ = z
3. y ′ = 2y + z, z ′ = z
4. y ′ = −y, z ′ = y − z
5. y ′ = 4y, z ′ = 2y + 4z
6. y ′ = z, z ′ = −y + 2z
7. y ′ = z, z ′ = −2y − 3z
8. y ′ = y − z, z ′ = 4y − 3z
9. y ′ = y + 2z, z ′ = −3y + 6z
10. y ′ = 3y − 4z, z ′ = y − 2z
11. y ′ = 3y + 5z, z ′ = −5y + 3z
12. y ′ = z, z ′ = −y
13. y ′ = y + z, z ′ = z
14. y ′ = 5y + z, z ′ = 5z
15. A tank contains 1000 liters of water and a salt solution containing 10
kg./liter is pumped into it at a rate of 200 liters /minute. The mixture
is led to a second tank containing 1000 liters of water at a rate of 200
liters/minute, and the mixture is pumped out of the second tank at the
same rate. What is the concentration of salt in the second tank after
one hour?
310 Dynamical Systems for Biological Modeling: An Introduction
16. Two tanks begin with 10 kg. of salt dissolved in 100 liters of water.
Water is pumped into the first tank at a rate of 10 liters/minute, the
mixture is pumped from the first tank to the second tank at a rate of
10 liters/minute, and the mixture is pumped out of the second tank at
a rate of 10 liters/minute. What is the amount of salt in each tank after
one hour, and what is the amount of salt in each tank after a very long
time?
17. Obtain the general solution of the system y ′ = ay, z ′ = cy + dz if a 6= d
by solving y ′ = ay, substituting the result into z ′ = cy + dz and solving.
y ′ = y − z,
z ′ = 3y − 3z,
y ′ = ay + bz, (5.23)
z ′ = cy + dz
If the roots λ1 and λ2 of (5.24) are real, then the solutions of (5.23) are made
up of terms eλ1 t and eλ2 t , or eλ1 t and teλ1 t if the roots are equal. In order that
all solutions of (5.23) approach zero, we require λ1 < 0 and λ2 < 0, so that
the terms will be negative exponentials. If the roots are complex conjugates,
λ = α ± iβ, then in order that all solutions of (5.23) approach zero, we require
α < 0. Thus if the roots of the characteristic equation have negative real part,
all solutions of the system (5.23) approach zero as t→ ∞. In a similar manner,
we may see that if a root of the characteristic equation has positive real part,
then (5.23) has unbounded solutions.
It turns out, however, that it is not necessary to solve the characteristic
equation in order to determine whether all solutions of (5.23) approach zero,
as there is a useful criterion in terms of the coefficients of the characteristic
equation. The basic result, whose proof may be found in Appendix C, is that
the roots of a quadratic equation λ2 + a1 λ + a2 = 0 have negative real part
if and only if a1 > 0 and a2 > 0. Applying this to the characteristic equation
(5.24) and the system (5.23), we obtain the following result for linear systems
with constant coefficients:
y ′ = ay + bz, z ′ = cy + dz
Example 1.
Determine whether all solutions tend to zero or whether there are unbounded
solutions for each of the following systems:
(i) u′ = −u − 2v, v ′ = u − 4v
(ii) u′ = v, v ′ = −u − 2v
(iii) u′ = −2v, v ′ = u + 2v
of a system
y ′ = F (y, z), z ′ = G(y, z) (5.26)
at an equilibrium (y∞ , z∞ ), we obtain the following result.
or
Fy (y∞ , z∞ )Gz (y∞ , z∞ ) − Fz (y∞ , z∞ )Gy (y∞ , z∞ ) < 0
the equilibrium (y∞ , z∞ ) is unstable.
Systems of Differential Equations 313
Example 2.
Determine whether each equilibrium of the system
y ′ = z, z ′ = 2(y 2 − 1)z − y
Example 3.
Determine whether each equilibrium of the system
y ′ = y(1 − 2y − z)
z ′ = z(1 − y − 2z)
At (0,0), this matrix has trace 1 and determinant 1, and thus the equilibrium is
314 Dynamical Systems for Biological Modeling: An Introduction
unstable. At (0, 21 ), this matrix has trace − 12 and determinant − 21 , and thus the
equilibrium is unstable. At ( 12 ,0), this matrix has trace − 21 and determinant
− 12 , and thus the equilibrium is unstable. At ( 13 , 13 ), this matrix has trace − 34
and determinant 13 , and thus this equilibrium is asymptotically stable. Thus
the model suggests that evenly matched competitors can coexist.
The careful reader will have noticed that, like the equilibrium stability
theorem of Section 4.1, the equilibrium stability theorem given above has a
hole of sorts in its result, in that the theorem says nothing about the stabil-
ity of equilibria for which the trace and determinant lie on the boundary of
conditions (5.27) and (5.28) — in other words, for which the linearization has
solutions which do not approach zero as t → ∞ but stay bounded. The reason
for this “hole” is that in such cases, the linearization does not give enough
information to determine stability. The following example recalls such a case,
from the application with which we began this chapter.
Systems of Differential Equations 315
Example 4.
Determine the asymptotic stability or instability of each equilibrium of the
Lotka-Volterra system
y ′ = y(λ − bz),
z ′ = z(−µ + cy).
0 − bµ
c .
cλ
b 0
This matrix has a positive determinant, but the trace is zero. In this case,
the stability theorem does not give any information. However, as we saw in
Section 5.1, the orbits of the system neither tend to the equilibrium nor move
away from the equilibrium. Thus the equilibrium is neither asymptotically
stable nor unstable, but behaves like a center (cf. Figures 5.2, 5.9).
y ′ = y(1 − y 2 − z 2 ) − z,
z ′ = z(1 − y 2 − z 2 ) + y,
which has its only equilibrium at the origin. The equilibrium is unstable, and
Figure 5.14 illustrates the fact that all orbits not beginning at the origin spiral
counterclockwise [in or out] toward the unit circle y 2 + z 2 = 1.
-2 -1 1 2
-1
-2
FIGURE 5.14: Trajectories approaching a limit cycle.
In many applications the functions y(t) and z(t) are restricted by the
nature of the problem to non-negative values. For example, this is the case
if y(t) and z(t) are population sizes. In such a case, only the first quadrant
y ≥ 0, z ≥ 0 of the phase plane is of interest. For a system
Example 5.
Show that every orbit in the region y > 0, z > 0 of the system
yz
y′ = y(2 − y) − ,
y+1
yz
z′ = 4 −z
y+1
approaches a periodic orbit as t → ∞.
Solution: We have y ′ = 0 when y = 0, and z ′ = 0 when z = 0, so orbits
starting in the first quadrant remain in the first quadrant. Equilibria are the
z 4y
solutions of either y = 0 or 2 − y = y+1 , and either z = 0 or y+1 = 1. If y = 0,
z
we must also have z = 0. If 2 − y = y+1 , we could have z = 0, which implies
y = 2, or y = 31 , which implies z = 20 9 . Thus there are three equilibria, namely
(0,0), (2,0), and ( 13 , 20
9 ). By checking the values of the trace and determinant
of the community matrix, which is
y∞
" #
z∞
2 − 2y∞ − (1+y ∞)
2 − y∞ +1
4z∞ 4y∞ ,
(y∞ +1)2 y∞ +1 − 1
we may see that each of the three equilibria is unstable. In order to apply the
Poincaré-Bendixson theorem, we must show that every orbit starting in the
first quadrant of the phase plane is bounded.
To show this, we might like to show that y ′ and z ′ are negative when
y and/or z are sufficiently large, but a glance at the equations tells us this
isn’t necessarily so. Therefore we instead consider some positive combination
of y and z whose time derivative does become negative far enough from the
origin. In particular, consider the function V (y, z) = 4y + z. If an orbit is
unbounded, then along this orbit the function V (y, z) must also be unbounded.
The derivative of V (y, z) along an orbit is
d
V [y(t), z(t)] = 4y ′ (t) + z ′ (t) = 4y(2 − y) − z.
dt
This is negative except in the bounded region defined by the inequality z <
4y(2 − y). Therefore the function V (y, z) cannot become unbounded, because
it is decreasing (dV /dt < 0) whenever it becomes large (z > 4y(2−y), which is
true, for example, whenever V > 9). This proves that all orbits of the system
318 Dynamical Systems for Biological Modeling: An Introduction
are bounded. Now we may apply the Poincaré-Bendixson theorem to see that
every orbit approaches a limit cycle.
Exercises
In Exercises 1–6, for each equilibrium of the given system determine whether
the equilibrium is asymptotically stable or unstable.
y ′ = y(1 − y − 2z),
z ′ = z(1 − 2y − z).
y ′ = z, z ′ = −y − z 3 .
r′ = r(1 − r), θ′ = 1.
Miscellaneous exercises
In each of Exercises 1–2, describe the orbits of the given system.
1. y ′ = y 2 z, z ′ = zy 2
2. y ′ = e−z , z ′ = ey
320 Dynamical Systems for Biological Modeling: An Introduction
In each of Exercises 3–6, find the linearization of the given system at each
equilibrium.
3. y ′ = y + z − 4, z ′ = y − z 5. y ′ = y 2 z, z ′ = zy 2
4. y ′ = z − 1, z ′ = y 6. y ′ = e−z , z ′ = ey
In each of Exercises 7–8, for each equilibrium of the given system determine
whether the equilibrium is asymptotically stable or unstable.
7. y ′ = y 2 z, z ′ = zy 2 8. y ′ = e−z , z ′ = ey
In each of Exercises 9–16, find the general solution of the given system by
analytic solution, use a computer algebra system to examine the behavior
of solutions, and classify the origin as a node, saddle point, center, or spiral
point.
9. y ′ = z − 3y, z ′ = 5y + z 13. y ′ = −z, z ′ = y
10. y ′ = y, z ′ = y + z 14. y ′ = y − z, z ′ = y + z
11. y ′ = 2y + z, z ′ = z 15. y ′ = y + z, z ′ = y + 2z
12. y ′ = −y, z ′ = y − z 16. y ′ = y + z, z ′ = y + z
17. A tank contains 1000 liters of water and a salt solution containing 20
kg./liter is pumped into it at a rate of 100 liters /minute. The mixture
is led to a second tank containing 1000 liters of water at a rate of 100
liters/minute, and the mixture is pumped out of the second tank at the
same rate. What is the concentration of salt in the second tank after
one hour?
18. Two tanks begin with 10 kg. of salt dissolved in 100 liters of water.
Water is pumped into the first tank at a rate of 20 liters/minute, the
mixture is pumped from the first tank to the second tank at a rate of
20 liters/minute, and the mixture is pumped out of the second tank at
a rate of 20 liters/minute. What is the amount of salt in each tank after
one hour, and what is the amount of salt in each tank after a very long
time?
In Exercises 19–22, for each equilibrium of the given system determine whether
the equilibrium is asymptotically stable or unstable.
19. y ′ = y + z − 4, z ′ = y − z 21. y ′ = y 2 z, z ′ = zy 2
20. y ′ = z − 1, z ′ = y 22. y ′ = e−z , z ′ = ey
Chapter 6
Topics in Modeling Systems of
Populations
There are many questions involving the interaction of two different popula-
tions. These different populations may be members of a single population but
distinguished by gender, age, or their infection status with respect to a dis-
ease present in the population, or they may be members of two quite different
species, cooperating, competing for a common resource, or in a predator-prey
relationship. The modeling of such questions leads naturally to systems of
differential equations, typically with each differential equation describing one
of the interacting populations. This chapter is devoted to some examples, de-
scribing applications of the general theory of systems of differential equations
developed in the previous chapter.
321
322 Dynamical Systems for Biological Modeling: An Introduction
three classes, but in each case one of the equations may be eliminated since
we can use the above relation to find S, I or R in terms of the other two. Thus
we will obtain a system of two differential equations to describe the spread of
diseases for which there is a removed class.
A model was proposed by W. O. Kermack and A. G. McKendrick1 to
explain the rapid rise and fall of cases frequently observed in epidemics, in-
cluding the Great Plague of 1665–66 in England, cholera in London in 1865,
and plague in Bombay in 1906. This model is
S ′ = −βSI,
I ′ = βSI − γI, (6.1)
R′ = γI.
The only difference from the model of Section 3.2.3 is that the term γI now
represents a rate of transition from the class I to the class R, instead of a rate
of return to the class S. The rate of recoveries in unit time is γI, and the rate
of transmission of infection from infectives to susceptibles is βSI. Note that
this model is only appropriate if the duration of an outbreak is short enough
that demographics (natural births and deaths) can be ignored.
We consider the model as a system of two equations, viewing R as deter-
mined by S and I, R = K − S − I, since the first two equations do not involve
R:
S ′ = −βSI, (6.2)
′
I = βSI − γI.
The equilibria of the system (6.2) (seen in slightly different form in Section 5.2,
Example 2) are the solutions of the pair of equations βSI = 0, βSI − γI = 0.
The first of these implies that either S = 0 or I = 0. If S = 0, the second
equation is satisfied only if I = 0, while if I = 0 the second equation is satisfied
for every S. Thus there is a line of equilibria (S∞ ,0) with S∞ arbitrary, 0 ≤
S∞ ≤ K. If we compute the linearization of the system (6.2) at an equilibrium
(S∞ ,0), we obtain
u′ = −βS∞ v,
′
v = (βS∞ − γ) v,
However, if S(0) > βγ , then I(t) increases so long as βS > γ, and thus I(t)
increases initially before decreasing to zero. We think of introducing a small
number of infectives into a susceptible population so that I(0) = ǫ > 0,
S(0) = K − ǫ. Then if βK/γ < 1, I(t) decreases monotonically to zero and
the infection dies out. On the other hand, if βK/γ > 1, an epidemic occurs,
as S(0) > βγ (for ǫ small), so I(t) increases to a maximum and then decreases
to zero. This is another threshold theorem of Kermack and McKendrick with
the threshold quantity βK/γ. This threshold quantity distinguishes between
two possible behaviors just like the threshold quantity in Section 3.2.3, but
the possible behaviors are not the same as in Section 3.2.3.
One might suppose that the reason for the eventual disappearance of the
infection in the epidemic case is that all susceptibles become infected, but
observations of epidemics indicate that this is not the case. The model (6.2)
agrees with observation in that it implies that the limiting value S(∞) =
limt→∞ S(t) of every solution of the system (6.2) obeys S(∞) > 0. We may
see this by calculating
γ S ′ (t)
d γ
S(t) + I(t) − log S(t) = S ′ (t) + I ′ (t) −
dt β β S(t)
γ
= −βSI + [βSI − γI] − (−βI) = 0
β
and therefore h i
S(0)
log S(∞)
β/γ = . (6.3)
S(0) − S(∞)
The quantity β/γ is known as the contact number. Not only does (6.3) imply
S(∞) > 0 (because if S∞ were zero, the right side of (6.3) would be infinite but
the left side is finite), but it also gives a means of estimating the contact rate β,
which generally cannot be measured directly. By making a serological survey
(testing for immune responses in the blood) in the population before and after
an epidemic, one may estimate S(0) and S(∞), and then (6.3) gives β/γ. If
the mean infective period 1/γ is known as well, then β can be calculated. The
contact rate β depends on the disease as well as on other factors such as the
rate of mixing in the population.
324 Dynamical Systems for Biological Modeling: An Introduction
30
25
20
l(t)
15
10
FIGURE 6.1: During the Great Plague of FIGURE 6.2: A phase por-
1665–66, the village of Eyam in England vol- trait for model (6.2).
untarily quarantined itself in hopes of pre-
venting the plague from spreading to neigh-
boring villages. Inhabitants of the neighbor-
ing populations left food and other supplies
for Eyam residents at the Boundary Stone
(pictured) just outside the village.
log 254
β/γ = 83
= 6.54 × 10−3 .
254 − 83
The infective period was 11 days, or 0.3667 months. Using a month as the
unit of time we obtain the estimate β = 0.0178. This data, with 7 initial
infectives, gives the phase portrait of Figure 6.2, traversed from right to left
as time progressed and the number of susceptibles decreased. Note that in this
case infected individuals were removed to the R class through death for the
most part, rather than recovery with immunity. Our simple model (6.2) still
describes this process, however different the interpretation may be, as the R
class of the model simply includes individuals no longer involved in the spread
of the disease.
The criterion βK/γ > 1 for the establishment of a disease can also be
expressed as the requirement that the susceptible population density exceeds
a certain critical value βγ . For fox rabies in Europe, observations indicate a
Topics in Modeling Systems of Populations 325
Example 1.
A survey of freshman students at Yale University2 found that 25% were sus-
ceptible to rubella at the beginning of the year and 9.65% were susceptible at
the end of the year. What fraction would have had to be immunized to avoid
the spread of rubella?
Solution: Using S(0) = 0.25, S(∞) = 0.0965 and substituting in (6.2), we
obtain
0.25
log 0.0965
β/γ = = 6.20.
0.25 − 0.0965
γ
In order to avoid the spread of rubella, the requirement is S(0) < β = 0.16.
2 A.S. Evans, Viral Infections of Humans, 2nd ed., Plenum Press, New York (1982),
Photo by AlMare
FIGURE 6.3: The notion of herd immunity comes from the fact that mem-
bers of a homogeneously mixing population (a “herd”) making potentially
infectious contacts with neighbors can be protected from infection if enough
of those neighbors are immune. Imagine here that the individuals in the car
in the background are in close contact with the neighboring members of the
herd surrounding them (here, a herd of goats on a road in Greece). If enough
of those neighbors are vaccinated against infection, then by the time the in-
dex individual comes into contact with an unvaccinated individual, the latter
may no longer be infected. With the infection rate reduced so drastically, the
infection dies out in the population.
Example 2.
In Example 2, Section 3.2.3, a disease was described spreading in a population
1
of 1200 members. Suppose the disease, with β = 3000 , 1/γ = 6 days, had
conferred immunity on recovered infectives. How many members would have
had to have been immunized to avoid an epidemic?
1
Solution: The basic reproductive number is βK/γ = 3000 × 6 × 1200 = 2.4.
γ
Thus herd immunity, p > 1 − βK , would require immunization of a fraction
1
1 − 2.4 = 0.5833, or 700 members.
γ γ
S(t∗ ) + I(t∗ ) − log S(t∗ ) = S(0) + I(0) − log S(0)
β β
and
γ γ γ γ
+ I(t∗ ) − log = S(0) − log S(0).
β β β β
From this we conclude that I(t∗ ), the maximum number of infectives, is given
by
γ γ γ γ γ γ log βS(0)
I(t∗ ) = S(0) − log S(0) + log − = S(0) − − . (6.4)
β β β β β β γ
For the Great Plague in Eyam, this gives a maximum infective population of
30.4, confirmed by the phase portrait of Figure 6.2.
Example 3.
What is the maximum number of infectives in the rubella epidemic of Exam-
ple 1?
Solution: In Example 1 we were given S(0) = 0.25, and we calculated β/γ =
6.20. Then (6.4) gives
1 1
I(t∗ ) = 0.25 − − log(6.20)(0.25) = 0.018.
6.20 6.20
Thus at most 1.8% of the population is infected at any one time.
S′ = µK − βSI − µS (6.5)
I′ = βSI − (γ + µ)I.
and it is easy to see that this has negative eigenvalues if and only if R0 < 1.
Thus the disease-free equilibrium is asymptotically stable if and only if R0 < 1.
At the endemic equilibrium (S, I) the matrix is
−(βI + µ) −βS
.
βI 0
Since this matrix has negative trace and positive determinant, the endemic
equilibrium is always asymptotically stable if it exists. We may summarize the
analysis by saying that there is always a single asymptotically stable equilib-
rium, the disease-free equilibrium if R0 < 1 and the endemic equilibrium if
R0 > 1.
For most human diseases, the mean infective period 1/γ is less than one
month, much smaller than the mean life span 1/µ, which is onthe order of 70
years. Thus an endemic equilibrium, with
µ 1
I= 1− K,
γ +µ R0
the number of infectives is a tiny fraction of the carrying capacity. In a pop-
ulation of moderate size, say 1000, the number of infectives at an endemic
Topics in Modeling Systems of Populations 329
equilibrium might be less than 1, and small random effects might be enough
to wipe out the infective population. A numerical simulation of an epidemic
model (6.2) and an endemic model (6.5) would appear indistinguishable. How-
ever, in a large population of, say 1, 000, 000, there might be 1, 000 infectives
at an endemic equilibrium, and this is a number of disease cases large enough
to be significant.
For animal diseases, where the life span may be much shorter and the dis-
ease infective period may be much longer, the differences between an epidemic
and an endemic situation may be much more readily noticed. An extreme
example would be a disease with no recovery, such as rinderpest (a cattle
disease of ancient origin which has quite recently become the second disease,
after smallpox, to be eliminated). Such a disease would be described by an SI
model, like (6.5) but with γ = 0,
having
βK
R0 = ,
µ
and an endemic equilibrium
µ 1
S= , I= 1− K.
β R0
Exercises
1. The same survey of Yale students described in Example 1 reported that
91.1% were susceptible to influenza at the beginning of the year, and
51.4% were susceptible at the end of the year. Estimate the contact
number β/γ and decide whether there was an epidemic.
2. An influenza epidemic was reported at an English boarding school in
1970 which spread to 512 of the 673 students. Estimate the contact
number β/γ.
3. What fraction of the Yale students of Exercise 1 would have had to be
immunized to prevent an epidemic?
4. What fraction of the boarding school students of Exercise 2 would have
had to be immunized to prevent an epidemic?
5. What was the maximum number of Yale students of Exercises 1 and 3
missing classes because of influenza at any given time?
6. What was the maximum number of boarding school students of Exercises
2 and 4 suffering from influenza at any given time?
330 Dynamical Systems for Biological Modeling: An Introduction
Let us consider two species whose population sizes at time t are y(t) and
z(t), respectively. Suppose that each species would grow according to a logistic
law if there were no interaction with the other species. Suppose also that the
two species are competing for resources, and that the effect of this competi-
tion is to decrease the per capita growth rate of each species by an amount
proportional to the population size of the other species. These assumptions
lead to a model of the form
y ′ = y(λ − ay − bz) (6.7)
z ′ = z(µ − cy − dz)
with λ, µ, a, b, c, and d positive constants. The carrying capacities of the two
species are, respectively, λa and µd . It is easy to see that (0,0), ( λa ,0), and (0, µd )
are equilibria of the system (6.7). In addition, there may be an equilibrium
which we will call (y∞ ,z∞ ) with y∞ > 0, z∞ > 0, if the lines ay + bz = λ and
cy + dz = µ intersect in the first quadrant. Their point of intersection,
dλ − bµ aµ − cλ
y∞ = , z∞ = , (6.8)
ad − bc ad − bc
which exists as long as ad 6= bc, lies in the first quadrant when the numerators
and denominator in (6.8) are either all positive or all negative. Some algebra
shows that this happens when the relative growth ratio λ/µ (of y to z) lies
between a/c and b/d.
Because F (y, z) = λy − ay 2 − byz and G(y, z) = µz − cyz − dz 2 , the
community matrix at an equilibrium (y, z) is
λ − 2ay − bz −by
.
−cz µ − cy − 2dz
332 Dynamical Systems for Biological Modeling: An Introduction
Thus at the equilibrium (0,0) this matrix has trace λ + µ > 0 and determinant
λµ > 0, and the equilibrium is unstable. At the equilibrium ( λa ,0), the matrix
has trace −λ+ aµ−cλ b and determinant −λ( aµ−cλ
a ), so the trace is negative and
the determinant positive if aµ − cλ < 0. A similar argument shows that the
equilibrium (0, µd ) is asymptotically stable if dλ − bµ < 0. For the equilibrium
(y∞ ,z∞ ), if there is one, the community matrix is
−ay∞ −by∞
.
−cz∞ −dz∞
Thus the trace is negative, and the determinant is positive if and only if
ad − bc > 0 (or a/c > b/d).
Therefore the equilibrium (0,0) is always unstable, and we may summarize
the results for the other three equilibria as follows:
I. If b/d < λ/µ < a/c, there is an equilibrium (y∞ ,z∞ ) which is asymptot-
ically stable, and the other three equilibria are unstable.
II. If a/c < λ/µ < b/d, there are an unstable equilibrium (y∞ ,z∞ ) and two
locally asymptotically stable equilibria ( λa ,0) and (0, µd ).
III. If λ/µ < a/c, b/d, the equilibrium ( λa ,0) is unstable and the equilibrium
(0, µd ) is asymptotically stable.
IV. If λ/µ > a/c, b/d, the equilibrium ( λa ,0) is asymptotically stable and the
equilibrium (0, µd ) is unstable.
FIGURE 6.6: Case III: z survives. FIGURE 6.7: Case IV: y survives.
In case III, the relative growth ratio for species y (relative to z) is dom-
inated by both of the limiting ratios, λ/µ < a/c, b/d, while the relative
growth ratio for species z (relative to y) exceeds the two limiting ratios,
µ/λ > c/a, d/b. Therefore the more robust z-species survives, and the y-
species is wiped out.
In case IV, the tables are turned: the relative growth ratio for y (to z)
exceeds the two limiting ratios, λ/µ > a/c, b/d, while the relative growth
ratio for z (to y) is less than the two limiting ratios, µ/λ < c/a, d/b. Thus the
y-species survives, and the z-species is wiped out.
The four cases may also be distinguished by the locations of the isoclines,
which are all straight lines, as shown in Figures 6.4–6.7 for the four cases,
respectively. In these figures, each asymptotically stable equilibrium has been
marked @. By drawing the two nullclines for any competitive system and not-
ing their relative positions, we may identify which of the four cases describes
the system.
334 Dynamical Systems for Biological Modeling: An Introduction
Suppose now that we have two species of moth competing for the same
food supply, with respective initial per capita reproduction rates of 100 and
60 (in per capita per time units), and carrying capacities on a given patch
of habitat (such as a tree or field) of 25 and 30. (Figure 6.8 illustrates one
such competition, between the gypsy moth and the northern tiger swallowtail
butterfly.) In the next two examples we shall see that the extent to which
competition adversely affects each species determines whether the two can
coexist, or, if not, which one survives.
Example 1.
Determine the outcome of a competition modeled by the system
Example 2.
Now suppose instead that species y is only affected one quarter as much by
the presence of species z — that is, that each z individual reduces species y’s
per capita growth rate by one instead of four. Determine the outcome of the
competition modeled by the resultant system
with negative trace and positive determinant. Thus the coexistence equilib-
rium is asymptotically stable, and every orbit approaches it. Alternatively, we
note that λ/µ now lies between the other two ratios, b/d = 1 < λ/µ = 5/3 <
a/c = 4, as in case I. In this case, the level of competition (coefficients b and
c) is small enough for both species that they can coexist.
Finally, we revisit the model for plant competition first explored in Sec-
tion 3.2.2. Equation (3.17) gave the general equation for a ranked system
of competitors, in which the proportion pi of habitat occupied by species i
changes according to four processes: colonizing uninhabited patches, natural
mortality, displacing inferior competitors (with index j > i) and being dis-
placed by superior competitors (with index j < i). Analysis showed that each
species persisted at a positive equilibrium p∗i if and only if its colonization rate
ci exceeded its own mortality rate Pm i and the superior competitors left some
i−1
space available for colonization, j=1 pj < 1. However, Bampfylde observed
that displacement does not occur in rainforests;
Pn a competition model includ-
ing only colonization of open spaces (1 − j=1 pj for n species) and natural
mortality, but no displacement, becomes (cf. (2.18))
n
dpi X
= ci pi 1 − p j − mi p i (6.9)
dt j=1
for i = 1, 2, ..., n.
We analyze system (6.9) through its equilibria,
Pn found by setting dpi /dt = 0.
For each i this yields either p∗i = 0 or 1 − j=1 p∗j = mi /ci . In general the
values of mi and ci are independent from one species to another, so it is
unlikely to have mi /ci = mj /cj for
Pnany i 6= j; this then makes it impossible for
the proportion of free sites (1 − j=1 pj ) to match simultaneously more than
one mi /ci as in the equilibrium conditions. Thus no more than one species
336 Dynamical Systems for Biological Modeling: An Introduction
m3
persists], and (0, 0, 1 − c3 ) [only species 3 persists]. We can determine when
each equilibrium is asymptotically stable using the community matrix, which
calculations show to be
c1 (g − p1 ) − m1 −c1 p1 −c1 p1
−c2 p2 c2 (g − p2 ) − m2 −c2 p2 ,
−c3 p3 −c3 p3 c3 (g − p3 ) − m3
which readers familiar with linear algebra will observe has the eigenvalues
(solutions to the characteristic equation) c1 − m1 , c2 − m2 and c3 − m3 . Thus
the extinction equilibrium is asymptotically stable if and only if ci < mi for all
3 species (that is, each species’ mortality rate outstrips its ability to colonize
new territory), just as for the displacement model of Chapter 3.
At the second equilibrium, in which species 1 wins the competition, the
community matrix simplifies to
once again the eigenvalues are the diagonal entries, and this equilibrium is
asymptotically stable if they are all negative, i.e.,
c1 c2 c3
> max 1, , .
m1 m2 m3
We can extrapolate similar conditions for the asymptotic stability of the other
single-survivor equilibria, so that in the end the species with the highest ci /mi
Topics in Modeling Systems of Populations 337
ratio wins the competition. (We omit here the discussion of issues related to
global stability which would be required for a rigorous proof.)
This result suggesting competitive exclusion may appear counter-intuitive
when the dominant competitor will not (at equilibrium) occupy the entire
rainforest: the model which allows displacement predicts coexistence as long
as there is enough habitat for all, and without displacement the inferior com-
petitors would appear to be at an advantage relative to the same scenario with
displacement. However, what constitutes a superior competitor is quite differ-
ent in this new model (the species with the greatest ci /mi ratio) than in the
one studied in Chapter 3 (where species are ranked entirely independently of
their ci /mi ratios), and a little algebra shows that any “inferior competitor” in
this new model (say c2 /m2 < c1 /m1 ) would also approach a zero equilibrium
value in the prior model.
In the boundary case, mentioned as unlikely above, of a “tie” for highest
ci /mi ratio, there are an infinite number of [non-isolated] equilibria such that
the “tied” species’ proportions sum to their common mi /ci ratio.
Example 3.
Determine the behavior of a predator-prey system (for example, moths and
birds) modeled by the system
y ′ = y(180 − y − z), z ′ = z(−500 + 10y).
In the model (6.11) the term −byz in the equation for y ′ represents the
rate at which predators consume prey. Thus it is assumed that each predator’s
consumption is proportional to the prey population size. Biologically, it is
more plausible to assume that the rate of prey consumption per predator
increases with prey population size but is bounded as the prey population
becomes unbounded, that is, that there is a maximum rate of consumption per
predator no matter how plentiful the food supply. Beyond a certain point, the
prey population no longer limits the resources of the predators. For example,
we may assume that the rate of consumption of prey per predator has the
qy
form y+A where q and A are positive constants.3 Instead of the term cyz in
yz
the equation for z ′ , we incorporate a term proportional to z+A , representing
the conversion of food (prey) into predator biomass. This leads us to a model
of the form
y ayz
y ′ = ry 1 − − , (6.12)
K y+A
y J
z ′ = sz − .
y+A J +A
3 See Section 2.1.3 for an interpretation of this Verhulst-type expression, including the
significance of A.
Topics in Modeling Systems of Populations 339
FIGURE 6.9: A barracuda eats another fish off the Florida keys.
Here, we have also changed the names of some of the parameters in (6.11),
r sJ qyz
replacing λ by r, a by K , and µ by J+A . The term y+A (replacing byz)
in the first equation of (6.12) is called the predator functional response, and
syz
the term y+A (replacing cyz) in the second equation of (6.12) is called the
predator numerical response; the constant qs is the conversion efficiency of prey
into predators. The model (6.12) assumes that the prey population would obey
a logistic law in the absence of predators, and that the predator population
would die out exponentially in the absence of prey.
Here we have rewritten the natural decay rate of the predator population
in terms of J, which we can see from (6.12) is the minimum prey population
required to sustain the predator population (z ′ ≥ 0). As J decreases, so does
the rate at which the predator population would die out in the absence of the
prey. In the following two examples, we shall see that the parameter J (or,
equivalently, µ) is capable of changing the nature of the system’s behavior.
Example 4.
Determine the qualitative behavior of a predator-prey system (imagine this
time large and small fish in a pond, cf. Figure 6.9) modeled by the differential
equations
y yz
y′ = y 1 − − ,
30 y + 10
y 3
z′ = z − .
y + 10 5
340 Dynamical Systems for Biological Modeling: An Introduction
and thus the equilibrium is unstable. At the equilibrium (30,0), the community
matrix is
−1 − 43
3 ,
0 20
3
and since its determinant is − 20 < 0 this equilibrium is also unstable. At the
equilibrium (15,12.5), the community matrix is
1
− 5 − 53
1 ,
5 0
asymptotically stable. It is possible to show that every orbit with y(0) > 0
and z(0) > 0 — not just those which start close to (15,12.5) — approaches
this equilibrium. Thus predator and prey coexist here, with prey at half their
natural carrying capacity.
Example 5.
Suppose now that the environment of the pond in Example 4 improves in such
a way that the predator fish tend to live longer (or die off more slowly), so
that the natural per capita mortality rate drops from 53 (in per time units)
to 13 . Determine the qualitative behavior of this new predator-prey system,
modeled by the differential equations
y yz
y′ = y 1 − − ,
30 y + 10
y 1
z′ = z − .
y + 10 3
Since this matrix has positive trace, the equilibrium (5,12.5) is also unstable,
and the system has no asymptotically stable equilibrium. In order to show that
all orbits in the first quadrant are bounded, we apply the technique introduced
in Example 5, Section 5.4, adding the two equations of the model to obtain
y z
(y + z)′ = y 1 − − .
30 3
Thus y + z is decreasing except in the bounded region defined by z3 <
y
y 1 − 30 . In order for an orbit to be unbounded, y + z must be unbounded,
and, as in Example 5, Section 5.4, this is impossible since y + z is decreasing
whenever y + z is large. Thus all orbits in the first quadrant are bounded,
and the Poincaré-Bendixson theorem may be applied to show that there must
be a limit cycle (with the equilibrium (5,12.5) in its interior) to which every
orbit tends. Thus the two species co-exist, but their population sizes fluctuate
periodically. Some orbits are shown in Figure 6.10.
36
34
32
30
28
26
24
22
20
z18
16
14
12
10
8
6
4
2
5 10 15 20 25
y
Examples 4 and 5 show that for a model of the form (6.12) there may be
periodic orbits, periodic orbit] or every orbit may approach an equilibrium.
Which behavior occurs depends on the values of the parameters in the model
rather than on the form of the model. In fact, there is a class of models consid-
erably more general than (6.12) exhibiting the same two possible behaviors.
We shall now consider models of the general form
y ′ = yf (y) − yzφ(y), (6.14)
z ′ = z[syφ(y) − c].
Topics in Modeling Systems of Populations 343
In this model, as before, y(t) is the prey population size and z(t) is the predator
population size. Here the term yf (y) represents the prey population growth
rate in the absence of predators. We assume that the per capita growth rate
of the prey decreases as prey population size increases, and that there is a
prey carrying capacity K, so that
The term yzφ(y) represents the predator functional response, with consump-
tion of yφ(y) prey per predator in unit time. We assume that consumption of
prey is positive, and that the prey consumption per predator yφ(y) increases
with prey population size, but that the fraction φ(y) of prey population con-
sumed per predator decreases:
The term syzφ(y) is the predator numerical response, and cz is predator mor-
tality. Because of (6.16), the predator per capita growth rate syφ(y) − c in-
creases with prey population size and is positive for prey population size above
some minimum J, sJφ(J) = c. Normally, it is assumed that the minimum prey
population size for predator survival J is less than the prey carrying capacity
K, as otherwise it is clear that the predator population cannot survive.
Let us try to analyze the behavior of solutions of a system (6.14) under
the assumptions (6.15) and (6.16). The equilibria of (6.14) are the solutions
of
f (J)
y∞ = J, z∞ = .
φ(J)
Since f (0) > 0 and −c < 0, this equilibrium is unstable (by applying the
equilibrium stability theorem of the previous section).
The community matrix at the equilibrium (K, 0) is
Kf ′ (K)
−Kφ(K)
.
0 sKφ(K) − c
Because f ′ (K) < 0 and sKφ(K)−c > 0 (if K > J), we see that the equilibrium
(K, 0) is unstable if K > J. If instead K < J, then sKφ(K) − c < 0, and in
this case the equilibrium (K, 0) is asymptotically stable.
If K > J, there is a third equilibrium (y∞ ,z∞ ) with community matrix
cc[yf (y)]′y∞ − z∞ [yφ(y)]′y∞ −y∞ φ(y∞ )
.
sz∞ [yφ(y)]′y∞ 0
The conditions for asymptotic stability are
The equilibrium (y∞ , z∞ ) of the system (6.14) with y∞ > 0, z∞ > 0 under
the hypotheses (6.15) and (6.16) is asymptotically stable if and only if the
prey nullcline has negative slope at the equilibrium.
Topics in Modeling Systems of Populations 345
Because of the hypotheses (6.15) and (6.16), the slope of the prey nullcline
′
at its y-intercept (K, 0) is fφ(K)
(K)
< 0. The slope of the prey nullcline at y = 0
may be either positive or negative, depending on the functions f (y) and φ(y)
y q
and the parameter values. For example, if f (y) = r 1 − K and φ(y) = y+A ,
the prey nullcline has positive slope at y = 0 if A < K. Thus there are cases
in which the prey nullcline increases to a maximum and then decreases to zero
as y increases. In such cases the equilibrium (y∞ , z∞ ) is unstable if it is on
the increasing portion of the prey nullcline, and asymptotically stable if it is
on the decreasing portion of the prey nullcline.
Increasing the carrying capacity of the prey species amounts to increasing
the food supply for the predators. Geometrically, this would have the effect
of moving the prey nullcline upward and to the right. This could move the
equilibrium from the decreasing portion of the prey nullcline to the increasing
portion of the prey nullcline and thus change it from an asymptotically stable
equilibrium to an unstable equilibrium. This possibility, that increasing the
food supply for the predators could destabilize the population system, has
been called the “paradox of enrichment.”
If the equilibrium (y∞ , z∞ ) is unstable, the system (6.14) has no asymp-
totically stable equilibrium. It is possible to show that every solution of the
system (6.14) in the first quadrant is bounded as t → ∞ (see Exercise 13
below). Then by the Poincaré-Bendixson theorem (Section 5.4), every orbit
must approach a limit cycle, and the prey and predator populations oscillate
as in Example 5. It is possible that an orbit could come very close to one of
the axes, where a small perturbing force, say an environmental change, could
wipe out one of the populations and lead to the collapse of the population
system. Thus the oscillations caused by enrichment could turn out to be very
harmful to the predator species.
Example 6.
Describe the long-term behavior of a predator-prey system modeled by the
system
y ′ = y(10 − y) − yz, z ′ = z(y − 5).
Solution: This system is of the form (6.14) with f (y) = 10−y, so that K = 10,
and φ(y) = y, so that J = 5. The hypotheses (6.15) and (6.16) are satisfied.
The prey nullcline is the line z = 10 − y with negative slope, and thus the
coexistence equilibrium given by y = 5, z = 10 − y, or z = 5, is asymptotically
stable. Every orbit approaches this equilibrium, so the species will coexist.
6.2.3 Symbiosis
There are situations in which the interaction of two species is mutually
beneficial, for example, plant-pollinator systems (Figure 6.11 gives another
example). Such an interaction is called mutualistic or symbiotic. The interac-
tion may be facultative, meaning that the two species could survive separately,
346 Dynamical Systems for Biological Modeling: An Introduction
FIGURE 6.11: Two goby fish and a shrimp in a symbiotic relationship: the
shrimp maintains a burrow in the sand on the sea floor which also serves as a
safe refuge for the goby and their eggs, while the goby alert the shrimp (which
has poor eyesight) if danger approaches.
or obligatory, meaning that each species would become extinct without the as-
sistance of the other.
If we model a symbiotic system by a pair of differential equations with
linear per capita growth rates
the mutualism of the interaction is modeled by the positive nature of the in-
teraction terms cy and bz. In a facultative interaction, the constants λ and µ
are positive, while in an obligatory relation the constants λ and µ are nega-
tive. In each type of interaction there are two possibilities, depending on the
relationship between the slope a/b of the y isocline and the slope c/d of the
z-isocline. If ad > bc, the mutualistic effects are smaller than the self-limiting
terms in the per capita growth rates, and the slope of the y-isocline is greater
than the slope of the z-isocline.
In both facultative and obligatory interactions, if ad < bc there is a region
of the phase plane in which solutions become unbounded, and this suggests
that either we must restrict models of this form by requiring ad > bc, or we
must consider models with nonlinear per capita growth rates.
For models with linear per capita growth rates and ad > bc, the only
asymptotically stable equilibrium in the facultative case is the intersection
(y∞ , z∞ ) of the lines ay − bz = λ, −cy + dz = µ with y∞ > 0, z∞ > 0, and
Topics in Modeling Systems of Populations 347
every orbit tends to this equilibrium. To see this, we calculate the equilibrium
dλ + bµ cλ + aµ
y∞ = > 0, z∞ = > 0.
ad − bc ad − bc
The community matrix at (y∞ , z∞ ) is
λ − 2ay∞ + bz∞ by∞ −ay∞ by∞
=
cz∞ µ − 2dz∞ + cy∞ cz∞ −dz∞
and since this matrix has negative trace and positive determinant the equi-
librium (y∞ , z∞ ) is asymptotically stable. It is easy to verify that the other
equilibria (0, 0), (λ/a, 0), (0, µ/d) are unstable.
In the obligatory case with ad > bc, since λ < 0, µ < 0 the only equilib-
rium in the first quadrant is (0, 0), asymptotically stable since the community
matrix is
λ 0
,
0 µ
and asymptotic stability follows from λ < 0, µ < 0. Thus the only asymptoti-
cally stable equilibrium is the origin, and every orbit tends to the origin. In the
obligatory case neither species survives. While our model may be acceptable
in the facultative case, it is clear that the possibility of obligatory mutualism
is not described by this model. If we consider the obligatory case with ad < bc,
there is an equilibrium (x∞ , y∞ ) with x∞ > 0, y∞ > 0, which may be shown
to be a saddle point whose stable separatrices separate the phase plane into
a region of mutual extinction and a region of unbounded growth. Such a sep-
aration is plausible biologically, but it would be necessary to alter the model
so as to rule out the possibility of unbounded growth in order to give a more
realistic model.
Exercises
In each of Exercises 1–4, determine the outcome of the competition modeled
by the given system.
19. * Models for cell growth often assume that a cell can be in various
states, with switching (transfer) between one state an another.4 Assume
that there are two states, in only one of which there is proliferation
(cell division). If P (t) and Q(t) represent the concentrations of cells in
the two states, the following equations can be assumed, with all Greek
letters representing positive constants
u′ = −γv, v ′ = αu − qv,
A′ = αA − a1 A3 − a2 aB 2 , B ′ = βB − b1 BA2 − b2 B 3 .
6 This problem is taken from Lee A. Segel, Modeling dynamic phenomena in molecular
As noted in Section 4.5, it can be shown via Taylor approximations that the
truncation error of the Euler method at each step is no greater than a constant
multiple of h2 , and that the accumulated truncation error is no more than a
constant multiple of h.
The modified Euler method is given by
As in the first-order case, it requires the use of some starting method to give
y1 and z1 . It can be shown that the truncation error of the modified Euler
method in each step is no more than a constant multiple of h3 , and that the
accumulated truncation error is no more than a constant multiple of h2 . This
represents a considerable improvement over the Euler method.
The Runge-Kutta method is still more accurate, having a truncation error
in each step of no more than a constant multiple of h5 and an accumulated
truncation error of no more than a constant multiple of h4 . It is given by
where
K1 = F (yk , zk ), L1 = G(yk , zk ),
The warnings in Section 4.5 about the proper choice of step size h are also
relevant here. Some commercial software uses a variable step size to minimize
problems.
are also independent of population density, then we can write the system
cf f (1 − x)2 − µf cm f x2 cf F ∗
J= 2 2 , where x = ,
cf m(1 − x) cm mx − µm cf F + cm M ∗
∗
but the quantity x can take on any value from 0 to 1 as (F, M ) → (0, 0),
depending upon the path taken to approach the origin (as students of multi-
variable calculus will recall). [The equilibrium stability theorem of Chapter 5
does not apply here since the growth rates are not differentiable at (0,0).]
Therefore, in order to explore what determines the long-term behavior of
this system, we turn to numerical analysis. Computer systems such as Math-
ematica, Maple, and MATLAB can perform symbolic calculations in some
cases (such as computing the Jacobian matrix above), but here we illustrate
their ability to compute and graph numerical solutions (approximations) for
systems. Since exploration is inherently ad hoc rather than formulaic, we here
present a sample of commands and results representative of the problem solv-
ing process. Using Mathematica syntax, we may first define the equations and
354 Dynamical Systems for Biological Modeling: An Introduction
0.6
0.5
0.4
0.3
0.2
0.1
t
Out[7]= 2 4 6 8 10
It should not be surprising that F (t) and M (t) are equal (their graphs su-
perimposed) since we have given symmetric parameter values. However, we
see that for the given values, both populations die out quickly. Perhaps larger
initial population sizes would make a difference?
Topics in Modeling Systems of Populations 355
t
Out[10]= 2 4 6 8 10
This graph is identical to the previous one except for scale (the numbers are 10
times as great). So initial conditions may not matter. A partial phase portrait
may make this clearer, showing that trajectories from the two different sets
of initial conditions lead to the same equilibrium: the origin.
In[11]:= pp1=ParametricPlot[{F[t]/.set1, M[t]/.set1}, {t,0,tf},
AxesLabel→{”F”, ”M”}, PlotStyle→Thickness[0.01], PlotRange→All];
In[12]:= pp2=ParametricPlot[{F[t]/.set2, M[t]/.set2}, {t,0,tf},
AxesLabel→{”F”, ”M”}, PlotStyle→Dashing[0.05]];
In[13]:= Show[pp1,pp2]
M
10
Out[13]= 2 4 6 8 10
F
(The trajectories are superimposed since they both fall on the line F = M .)
Now what if we increase the reproduction rates, say by an order of mag-
nitude?
In[14]:= cf=10; cm=10;
356 Dynamical Systems for Biological Modeling: An Introduction
In[15]:= set3=twosexsoln;
In[16]:= plot3 = PopPlot[set3]
PHtL
6 ´ 107
5 ´ 107
4 ´ 107
3 ´ 107
2 ´ 107
1 ´ 107
t
Out[16]= 2 4 6 8 10
The outcome has now changed qualitatively: the population grows exponen-
tially instead of dying out exponentially. What if we now reduce one of the
reproduction rates, creating an asymmetry?
In[17]:= cf=3;
In[18]:= set4=twosexsoln;
In[19]:= plot4 = PopPlot[set4]
PHtL
450
400
350
300
250
200
150
t
Out[19]= 2 4 6 8 10
The overall growth rate is certainly slowed, but the population still grows,
and both populations remain matched. What if we force an asymmetry in the
population values by changing one of the initial conditions?
In[20]:= F0=50;
In[21]:= set5=twosexsoln;
In[22]:= plot5 = PopPlot[set5]
Topics in Modeling Systems of Populations 357
PHtL
250
200
150
100
t
Out[22]= 2 4 6 8 10
Despite the initial disparity in sizes, the two sexes’ populations quickly return
to even before growing, because all new births are evenly divided between
females and males.
Here we have used arbitrary values rather than realistic estimates because
the goal is to uncover qualitative insights about the model’s behavior. The
original question of the precise threshold condition between the two possible
behaviors (growth and decay) has not yet been answered precisely, although a
few data points have been established. A more methodical approach would fix
all but one or two key parameters (say cf and cm ), vary those independently
over a regular grid (say each taking on values between 1 and 10), then mark
on the grid which combinations led to extinction and which to growth, and
finally abstract from the results an overall pattern describing the influence of
these two parameters. We invite the reader to try such an experiment.
As one final note in this example, we observe that the unbounded growth of
the model above can be eliminated by introducing a logistic term in each equa-
tion, that is, by multiplying each sex’s growth rate by a factor of 1 − F +M K
for some carrying capacity K. While the resulting model is more complicated
to analyze (and still has the problem of not being linearizable at the origin),
even a partial analysis (see Exercise 6) sheds some light on the threshold
between survival and extinction for both that model and the one explored
above.
Exercises
Use a computer and available software to approximate the solution of each of
the following initial value problems on the given interval.
of Example 4, Section 6.2, with initial conditions y(0) = 40, z(0) = 10,
over the time range 0 ≤ t ≤ 10.
y yz y
3. The [pond] predator-prey system y ′ = y(1 − 30 )− y+10 , z ′ = z( y+10 − 13 )
of Example 5, Section 6.2, with initial conditions y(0) = 40, z(0) = 10,
over the time range 0 ≤ t ≤ 10.
1 1
4. The SIR disease model S ′ = − 200 SI, I ′ = 200 SI − 16 I of Section 6.1,
with initial conditions S(0) = 1190, I(0) = 10, over the time range
0 ≤ t ≤ 20.
In each case, compare the numerical approximation with the qualitative in-
sights obtained in Sections 6.1 and 6.2 where these systems are studied as
examples.
359
360 Dynamical Systems for Biological Modeling: An Introduction
w′ = ǫ(v − γw),
as this matrix has negative trace and positive determinant, the equilibrium
3 A.L. Hodgkin and A.F. Huxley, A quantitative description of membrane current and
its application to conduction and excitation in nerve, J. Physiology 117 (1952), 500–544.
362 Dynamical Systems for Biological Modeling: An Introduction
0.2 0.3
w 0.1 0.2
0.1
–0.2 0.2 0.4 0.6 0.8
v
Another experiment gives the cell a constant input of positive ions instead
of a single pulse. If we apply a constant current J, we add J to the rate of
change of potential (assuming that 1 unit of current raises the potential by 1
unit in unit time), so that the modification of (7.1) is
This moves the equilibrium (0,0) into the first quadrant. For small values of
J this equilibrium is asymptotically stable (Figure 7.2). For larger inputs it
becomes unstable and a periodic orbit is set up (Figure 7.3).
Thus a steady input leads to a periodic solution. Examination of v as
a function of t (Figure 7.4) shows a “bursting” behavior, similar to what is
observed in real neurons, with the potential rising close to 1 and then dropping
below zero.
Systems with Sustained Oscillations and Singularities 363
0.6
0.4
w v
0.5
0.2
FIGURE 7.5: Aminergic cell activity is high while cats are awake but low
during REM sleep.
4 R.W. McCarley and J.A. Hobson, Neuronal excitability modulation over the sleep cycle:
x′ = ax − bxy, (7.5)
y′ = −cy + dxy,
These imply that there are two equilibria: (0, 0) and (c/d, a/b). The matrix of
the linearization at an equilibrium (x, y) is
a − by −bx
.
dy dx − c
and it is clear that since the eigenvalues of this matrix are a, −c, this equilib-
rium is unstable.
At the equilibrium (c/d, a/b) the matrix of the linearization is
0 −bc/d
.
ad/b 0
Since the√ eigenvalues of this matrix are complex conjugates with real part
zero, ±i ac, the linearization at this equilibrium has periodic solutions. In
fact, the model (7.5) is the same as the model (5.1) for the Lotka-Volterra
system studied in Section 5.1, where it is shown that all solutions are given by
periodic orbits around the equilibrium. Thus the simple model (7.5) predicts
regular oscillations as observed.
Exercises
In each of Exercises 1–9, use a computer algebra system to display the behavior
of solutions of the Fitzhugh-Nagumo system (7.4) with the given values of the
parameters.
12. (a) Use a phase line or linearization to analyze the behavior of the solo
equation v ′ = −v(v − a)(v − 1), which represents neuron voltage in the
absence of a blocking mechanism.
*(b) Analyze the behavior of the following model, in which the blocking
mechanism w exists but fails to act on the voltage: v ′ = −v(v −a)(v −1),
w′ = ǫ(v − γw).
366 Dynamical Systems for Biological Modeling: An Introduction
with solution (y(t, ǫ), z(t, ǫ)). There is a corresponding reduced system ob-
tained by setting ǫ = 0,
f (y, z, 0) = 0 (7.7)
′
z = g(y, z, 0), z(0) = z0
called the boundary layer system, to give an approximation, called the inner
solution to the behavior of the system near t = 0. Away from t = 0 we
hope that the solution of the full problem is approximated well by the outer
solution.
Sometimes a problem is given in the form (7.9) from the start. The under-
lying idea in a singular perturbation problem is that there are two different
time scales inherent in the problem, and this makes it possible to analyze the
problem separately on each time scale. The reduction in dimension because of
this separation simplifies the analysis.
The mathematical treatment of singular perturbations began in the 1940’s
from the perspective of asymptotic expansions. A few years later the qualita-
tive result which justifies the use of the reduced system as an approximation
to the full system was obtained independently in the U.S.A. and the Soviet
Union:5
equations, Acta Math. 82 (1950), 71–106; A.N. Tihonov, On the dependence of the solutions
of differential equations on a small parameter, Mat. Sbornik NS 22 (1948), 193–204.
6 F.C. Hoppensteadt, Singular perturbations on the infinite interval, Trans. Amer. Math.
Example 1.
Describe the solution of the first-order differential equation
ǫy ′ = −y, y(0) = 1.
Solution: The solution of this initial value problem may be obtained easily by
separation of variables and is
y(t, ǫ) = e−t/ǫ .
We may calculate (
1 for t = 0
lim y(t, ǫ) =
ǫ→0 0 for t > 0.
This limit is discontinuous at t = 0. When ǫ = 0 the problem is no longer
an initial value problem but is just the relation y = 0 together with the
(incompatible) initial condition y(0) = 1. This indicates that the solution for
ǫ > 0 begins with the value 1 at t = 0 and then decreases rapidly to 0. A
graph of the solution for a small value of ǫ would indicate this boundary layer
(see Exercise 3 below).
Another way to approach this problem is to change the time scale by
making the change of independent variable t = ǫs to transform the problem
to
dy
= −y, y(0) = 1,
ds
whose solution is y(s) = e−s = e−t/ǫ .
Example 2.
Describe the behavior of solutions of the initial value problem
ǫy ′ = −y, y(0) = 1, (7.11)
′
z = −yz, z(0) = 1.
Solution: We let (y(t, ǫ), z(t, ǫ)) be the solution of this problem and we let
(y0 (t), z0 (t)) be the solution of the reduced problem
y = 0, (7.12)
z′ = −yz, z(0) = 1,
370 Dynamical Systems for Biological Modeling: An Introduction
(see Exercise 1 below). Since limǫ→0 z(t, ǫ) = 1 for all t ≥ 0 (see Exercise 2
below) we see that z(t, ǫ) → y0 (t) as ǫ → 0 for all t ≥ 0. However, as we have
seen in Example 1, (
1 for t = 0,
lim y(t, ǫ) =
ǫ→0 0 for t > 0,
and
lim y(t, ǫ) 6= y0 (t) ≡ 0
ǫ→0
0.1
B
w 0.1
B
w
v
A
v
A
seen in Figure 7.6 as the branches to the left of the point A, between the
points A and B, and to the right of the point B, respectively. The fast-time
dynamics in v make solutions move (horizontally) toward [the nearest stable
branch of] the QSS, and then the slow-time dynamics in w send solutions up
or down the QSS curve toward the w-coordinate v/γ. To see this, one can first
analyze the behavior of the “fast” dynamics in v (for fixed w, see Exercise 11)
and then the behavior of the “slow” dynamics in w (see Exercise 12). Analysis
of the v equation shows that the branch v2 is unstable while v1 and v3 are
locally asymptotically stable (in the v direction), so if we take w(0) = 0, then
v = a, where a is the middle intersection of the quasi-steady-state curve (the
unstable v2 ) with the v axis, separates solution behaviors into two distinct
groups, depending on the initial value of v. If v(0) < a, then v goes directly
to the equilibrium (0, 0). If v(0) > a, then v goes rapidly to the right branch
v = v3 (w) while w remains constant. Then, in the slow time, the orbit follows
the curve v = v3 (w) until it comes near the point B; this is the excited phase
of the motion. If γ is small enough, say γ < 4.5 (this seems arbitrary but it’s
not, see Exercise 13), then near B we have dw/dt = v − γw > 0, indicating
a continued upward trajectory, but at B the graph of the quasi-steady-state
w = −f (v) turns downward. The orbit thus cannot continue to follow the
quasi-steady-state curve past B; it is then governed by a fast system
ǫv ′ = −f (v) − w, w′ = 0
and moves rapidly horizontally until it reaches the branch v = v1 (w). It then
follows this branch of the quasi-steady-state curve to the equilibrium (0, 0).
372 Dynamical Systems for Biological Modeling: An Introduction
This behavior is illustrated in Figure 7.7, which shows an orbit for the system
(7.13) with parameter values a = 0.3, γ = 1, ǫ = 0.002 along with the quasi-
steady-state curve.
Now suppose a constant current is applied, changing the model to
dv
ǫ = −f (v) − w + J, (7.14)
dt
dw
= v − γw.
dt
Then the effect is to move the quasi-steady-state curve upward (it is now
w = −f (v) + J) and the equilibrium into the first quadrant (since the line
v = γw where dw/dt = 0 only passes through the first and third quadrants).
It is possible to show that when J is large enough for the equilibrium to reach
the portion v = v2 (w) of the quasi-steady-state curve (at the point labeled A)
the equilibrium becomes unstable and the orbit becomes periodic, oscillating
between the branches v = v1 (w) and v = v3 (w).
7.2.1 Bursting
In the Fitzhugh-Nagumo model, and also in the four-dimensional Hodgkin-
Huxley model, which models the behavior of a neuron somewhat more closely,
neurons may fire periodically if stimulated by a constant applied current.
Many cells exhibit a more complicated behavior called bursting, in which
periods during which the potential changes slowly alternate with periods of
rapid oscillation. Such behavior is observed, for example, in groups of pan-
creatic β-cells. Action potential bursting in these cells, clustered together in
the pancreas in groups called the islets of Langerhans (see Figure 7.8), plays
a critical role in secreting the hormone insulin, used for maintaining glucose
levels in the body. (Type II diabetes, hyperglycemia, develops when this burst-
ing behavior does not compensate correctly in response to glucose levels in
blood plasma.) Here we will not explore the rather complicated models which
have been formulated to attempt to explain bursting phenomena,7 but we
will sketch out a way to use multiple timescales and singular perturbations to
design a model that exhibits bursting.
The basic idea is to build a model which can exhibit either oscillation or
quiescence (a stable equilibrium), and then use a slow timescale to switch back
and forth between the two behaviors, to replicate bursting. In order to exhibit
oscillation within the fast timescale, we will need at least two dimensions
there, so we suggest a three-dimensional model with two fast variables and
7 One review of models for pancreatic β-cell bursting is Arthur Sherman and Richard
ǫv ′ = f1 (v, w, z),
ǫw′ = f2 (v, w, z), (7.15)
′
z = g(v, w, z).
We would like to arrange the fast variable system to have three equilibria:
an asymptotically stable node, a saddle point, and an unstable node with an
asymptotically stable limit cycle around it. In this way the unstable equilib-
rium (saddle point) serves to separate the fast variable state space into two
different regions, one where solutions approach the stable equilibrium and
one where solutions approach the limit cycle; the dividing curve (or surface)
that extends from the unstable equilibrium to do this is called a separatrix,
formed by the stable manifold of the saddle point, separating the domain of
attraction of the asymptotically stable node from the domain of attraction
of the limit cycle. (The unstable equilibrium branch v2 in the fast dynamics
of the Fitzhugh-Nagumo example serves similarly to separate the domains of
attraction for the stable branches v1 and v3 .)
The slow variable z then follows the quasi-steady state curve
f1 (v, w, z) = 0, f2 (v, w, z) = 0.
We would like to arrange for the slow variable to move solutions back and
forth across the separatrix. A move into the domain of attraction of the limit
cycle will trigger a rapid oscillation, and then a move back into the domain of
attraction of the asymptotically stable node will lead to a quiescent period.
This cycle will be repeated to give bursting.
374 Dynamical Systems for Biological Modeling: An Introduction
v′ = w − v 3 + 3v 2 , (7.16)
w′ = 1 − 5v 2 − w
coupled with a slow variable z and an applied voltage J to give the system
v′ = w − v 3 + 3v 2 + J − z,
w′ = 1 − 5v 2 − w, (7.17)
′
z = ǫ[s(v − v1 ) − z]
√
where v1 = ( 5 − 1)/2. In the slow dynamics, when v is large, z increases
and becomes large, but then v ′ is decreased by the larger z, until eventually v
decreases, and then so does z. With the parameter values J = 2, ǫ = 0.002, s =
4 the graph of v for this system is shown in Figure 7.9.
2
v 1
–1
–2
The phenomenon of bursting arises in many real cell situations, and the
above example is an extremely simplified version of one of the types of model
that displays bursting. Since study of the behavior of such models requires
very detailed examination of the phase portrait of the fast subsystem, we will
not go further into this subject.
s′ = −k1 se + k−1 c,
e′ = −k1 se + (k−1 + k2 )c, (7.18)
′
c = k1 se − (k−1 + k2 )c,
p′ = k2 c,
333–369.
376 Dynamical Systems for Biological Modeling: An Introduction
dy
k1 e20 = k1 e0 (1 − y)s0 z − (k−1 + k2 )e0 y, (7.22)
dτ
dz
k1 e0 s0 = −k1 e0 (1 − y)s0 z + k−1 e0 y.
dτ
We now define the parameters
k2 k−1 + k2 e0
λ= , K= , ǫ=
k1 s0 k1 s0 s0
so that K − λ > 0. Then the system (7.22) becomes
dy
ǫ = z(1 − y) + Ky, y(0) = 0, (7.23)
dτ
dz
= −z(1 − y) + (K − λ)y, z(0) = 1.
dτ
In many enzyme reactions the enzymes are very effective catalysts and
the concentration of enzyme needed is very small compared to the substrate
concentration. This means that e0 is much smaller than s0 and thus that ǫ is
very small. Often ǫ is between 10−2 and 10−7 . Thus we may view (7.23) as a
singular perturbation problem of the form (7.6) with
This means that except in the boundary layer very close to t = 0, where c may
change rapidly, we may approximate the solution of (7.19) by the solution of
the reduced problem
This gives
k1 e0 s
c= ,
k1 s + k−1 + k2
and then (7.20) reduces to the single differential equation
k−1 +k2
If we let Km = k1 ,Q = k2 e0 , this becomes
Qs
s′ = − . (7.25)
s + Km
This substrate uptake rate is called a Michaelis-Menten uptake, and reaction
rates of this form occur in other problems. For example, the predator func-
tional response in predator-prey models is often assumed to be of this form.
We recall that in the original model (7.18) the rate of formation of product is
given by p′ = k2 c. For the reduced problem we may replace c in this expression
by (7.20), and then we obtain p′ = −s′ . Thus the rate of reaction is given by
(7.25).
Exercises
−t/ǫ)
1. Show that the solution of z ′ = −e−t/ǫ z, z(0) = 1 is z(t, ǫ) = e−ǫ(1−e .
−t/ǫ)
2. Show that limǫ→0 e−ǫ(1−e = 1 for all t ≥ 0. [Hint: Show that
limǫ→0 ǫ(1 − e−t/ǫ ) = 0.]
3. Graph the solution of the initial value problem in Example 1 with the
values ǫ = 0.2, ǫ = 0.1, ǫ = 0.01.
4. Find the solution of the stretched version of the initial value problem of
Example 2.
5. Solve the initial value problem ǫy ′ = y − y 3 , y(0) = 1/2 and compare the
solution with the solution of the corresponding reduced problem.
6. Solve the initial value problem
7. Find the equilibria and analyze their stability for the system (7.16).
378 Dynamical Systems for Biological Modeling: An Introduction
à
æ à à2.
1 1 1 1 3 5 7 9 11
12 6 4
years
Photo courtesy CDC/Maureen Metcalfe, Tom
Hodge
FIGURE 7.11: This graph, adapted from
FIGURE 7.10: HIV-1 viri- widely circulated ones in Pantaleo et al.
ons. (1993) and Fauci et al. (1996), shows the pro-
gression of within-host viral load (circles) and
corresponding CD4+ T cell count (squares)
[vs. time since infection] in a typical HIV pa-
tient, over the initial acute phase, the clini-
cally latent phase, and eventual progression
to AIDS. Note the change in timescale after
3 months.
In modeling the interaction between T cells and HIV, one must make deci-
sions about which processes and approaches to include. Must all the immune
responses be modeled separately? Infected T cells spawn HIV virions inter-
nally during a quiescent or latent period, followed by lysis and the sudden
release of free virions into the system; since much of the viral load comes
from other types of infected cells, need each of these stages (or indeed the
infected T cells at all) be represented explicitly? Is the initial acute phase im-
portant in determining the long-term outcome? Do viral and immune system
dynamics take place on different enough timescales that a singular pertur-
bation approach would help? Some answers are offered by the specificity of
one’s hypothesis (which often claims one mechanism to be most important in
explaining observed behavior), while others can only be determined by trying
several models and/or analysis approaches, and then selecting the simplest
one which explains observations well. We shall now employ both ways, via
multiple tries beginning with a simple model for a healthy immune system.9
9 Much of the pioneering work in modeling HIV immunology builds on work by Perelsen
and colleagues; one set of models whose structure parallels that developed in this section
is discussed in Alan S. Perelson and Patrick W. Nelson, Mathematical analysis of HIV-1
dynamics in vivo, SIAM Review 41(1): 3–44, March 1999. Another helpful reference for this
topic is M.A. Nowak and R.M. May, Virus dynamics: mathematical principles of immunol-
ogy and virology, Princeton University Press, 2003.
Systems with Sustained Oscillations and Singularities 381
−µ − sa2 − k1 T0
g ,
0 b − k2 T0
V V
FIGURE 7.12: Nullclines of the sys- FIGURE 7.13: Nullclines of the sys-
tem (7.26) in the V -T plane, in the tem (7.26) in the V -T plane, in the
case where kg2 > bT0 and kg2 < s1k−s
1
2
. case where kg2 > bT0 , kg2 > s1k−s
1
2
and
the nullclines do not cross.
are to the right (Figure 7.13). Thus solutions will enter the region between the
curves, either from above or from below; any solution lying between the two
nullclines must remain between them for all further time (moving down and
to the right). Thus T will decrease and V will increase for all t, with V → ∞
and T → 0, indicating progression to AIDS.
Given the relatively complicated analysis, a simplification technique may
help us better understand the model system’s behavior. Since the changes in
V take place on a much faster time scale (a few weeks) than the changes in
T (several years), it is reasonable to consider rewriting (7.26) as a singular
perturbation problem of the form
′ s2 V
T = cT s1 − − µT − k1 T V ,
a+V
′ gV
ǫV = cV − k2 T V ,
b+V
with proportionality coefficients cT and cV , allowing us to rescale the time
variable to study either fast (inner system) or slow (outer system) time. Let
us suppose that the values of the coefficients in (7.26) support this notion.
Then our analysis of (7.26) can be decomposed into the analysis of the inner
system with T constant, T = T0 :
dV gV
= − k2 T0 V, V (0) = V0 ≥ 0, (7.27)
dτ b+V
in the boundary layer near t = 0 (where τ is the rescaled, “fast” time), and
an outer system for T (with V given in terms of T ).
For the inner system, the first-order equation (7.27) can easily be shown
384 Dynamical Systems for Biological Modeling: An Introduction
with g > k2 bT0 and, since the virus terms should make F (T ) < s1 − µT
(and thus T (t) ≤ T0 ), T (t) < g/k2 b more broadly. Thus in analyzing the
outer system we restrict our attention to the interval [0, g/k2 b]. Given the
complexity of F , we proceed using basic properties of F rather than finding
its roots. We observe
s1 − s2 g g
F (0) = k1 − , F (g/k2 b) = µ T0 − < 0,
k1 k2 k2 b
s2 agk2
and calculate F ′ (T ) = k1 b − µ + ,
[g + (a − b)k2 T ]2
s2 ak2 s2 k2 b2
F ′ (0) = k1 b − µ + , F ′ (g/k2 b) = k1 b − µ + .
g ga
g
Both F and F ′ have a pole (i.e., a vertical asymptote) at T = k2 (b−a) , but
it does not fall within [0, g/k2 b]: if a > b the pole is negative and F ′ (which
Systems with Sustained Oscillations and Singularities 385
decreases away from the pole) is then monotone decreasing on [0, g/k2 b], while
if a < b some algebra shows that the pole is greater than g/k2 b, making F ′
monotone increasing on [0, g/k2 b]. In either case F ′ is monotone, so F ′ has at
most one root (F ′ = 0) in [0, g/k2 b], meaning F can “turn around” at most
once.
Thus if F (0) > 0 there is a unique equilibrium in (0, g/k2 b) (since
F (g/k2 b) < 0), and since F crosses from positive to negative there, F ′ < 0
there, making the equilibrium asymptotically stable (locally, and also globally
since there are no other stable equilibria). This indicates an endemic state.
If a > b (making F ′ monotone decreasing and thus F concave down) and
′
F does have a root, Tc , in [0, g/k2 b], then Tc is a local maximum for F , and
it is possible to have F (0) < 0 but F (Tc ) > 0. In this case there are two
equilibria: one in (0, Tc ) (call it E0 ) which is unstable (since F crosses from
negative to positive there, making F ′ > 0), and one in (Tc , g/k2 b) (call it E1 ),
asymptotically stable (since F crosses from positive to negative there, making
F ′ < 0 as before). Here the dynamics are more complex, as the unstable E0
acts as a separatrix: If T0 < E0 , then T → 0 (and then V → ∞) and the
infection progresses to AIDS, but if T0 > E0 , then T → E1 , and the infection
persists in an endemic state.
If neither of the two sets of conditions above hold (i.e., F (0) < 0 and either
Tc does not exist in [0, g/k2 b] or F (Tc ) < 0), then F < 0 on all of [0, g/k2 b],
making T ′ < 0 so that T → 0, V → ∞ and the infection progresses to AIDS
regardless of initial conditions.
The condition F (0) > 0 (for a single, globally stable endemic equilibrium)
is equivalent to
g s1 − s2
< , (7.32)
k2 k1
seen in the analysis of the original (full) system (7.26). The more complicated
conditions that lead to two equilibria can be condensed (see Exercise 3) to
s r s s
b k2 µ − k1 b k2 k1 b s1 − s2 k2
< < − 1− 1− . (7.33)
a g s2 a g s2 a k1 g
0.0008
0.0006
V V
2
0.0004
1
0.0002
1 2 3 4 5 1 2 3 4 5
t t
800
60
600
T V
40
400
20
200
500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000
t t
12000
800
10000
600 8000
T V
6000
400
4000
200
2000
500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000
t t
values, this possibility may not arise), and under most other conditions there
is progression to AIDS.
For numerical simulations with given sets of parameter values, the rapid
changes in V near t = 0 may cause difficulties in the approximation. It is
Systems with Sustained Oscillations and Singularities 387
advisable to use the equilibrium of the inner system as an initial value. This is
essentially equivalent to finding a solution of the outer problem and matching
it to the solution of the inner problem, thus using the singular perturbation
approach at least implicitly. We will give the results of some numerical simu-
lations, using the parameter values
cells cells virus virus
s1 = 10 day, s2 = 7 day, a = 12 , b=8 , (7.34)
mm3 mm3 mm3 mm3
−1 −1
virus cells
k1 = 2.5 × 10−4 day , k2 = 0.01 day , µ = 0.01day −1 .
mm3 mm3
The values for s1 , s2 , µ, k1 , k2 have been derived from experimental data. The
graphs are very sensitive to changes in the parameters a and b. We use different
values of g to illustrate the possibilities of eradication of the virus, an infected
steady state, and progression to AIDS. Figures 7.14 and 7.15 show the short
term behavior with g = 1 and g = 125, respectively (units for g in all cases are
virus
mm3 day but will hereinafter be omitted for space constraints). Figures 7.16
and 7.17 show the long term behavior of T and V for g = 100, illustrating an
asymptotically stable infected steady state, while Figures 7.18 and 7.19 show
the long term behavior for g = 125, illustrating progression to AIDS.
− sa2 − k1 T0
−µ 0
0 −δ k1 T0 ,
g
0 δN b − k T
2 0
analyzing the full three-dimensional model (7.35), and thus this equilibrium
is asymptotically stable if and only (7.36) is satisfied. It is not difficult to
show that the system (7.37) has another equilibrium, which is positive and
asymptotically stable, if (7.36) is not satisfied but 0 < k2 − N k1 < Tg0 b , given
by
k1 T0 g g
I= −b , V = − b.
δ (k2 − N k1 )T0 (k2 − N k1 )T0
If, finally, k2 < N k1 , so that the virus clearance rate is outweighed by the
indirect virus creation rate from infected T cells alone, then one can show
(Exercise 6) that I, V → ∞, corresponding to AIDS.
The outer problem is
s2 V
T ′ = s1 − − µT − k1 T V
a+V
with V the appropriate limiting value of the inner system. If k2 − N k1 >
g/T0 b (so V = 0), which is the case of virus eradication, then the outer
system reduces to the healthy T cell equation and T → T0 . If k2 < N k1 (so
V → ∞), then T → 0 and the infection progresses to AIDS and death. If
0 < k2 − N k1 < g/T0 b, so that
g
V (T ) = − b,
(k2 − N k1 )T
then we may write the outer problem in the form (7.31) with k2 replaced by
k2 − N k1 in the expression for F (T ). The argument used in analyzing (7.31)
shows there is a positive asymptotically stable equilibrium if and only if
g s1 − s2
< (7.38)
k2 − N k1 k1
g s1 −s2
(analogous to (7.32)). For larger values of g ( k2 −N k1 > k1 ) two endemic
equilibria may exist if (analogous to (7.33))
r r r s
b k2 −N k1 µ−k1 b k2 −N k1 k1 b s1 −s2 k2 −N k1
< < − 1− 1− .
a g s2 a g s2 a k1 g
Otherwise f (T ) < 0 for all T , and once again the solutions of the outer
problem are monotone decreasing, and without a positive equilibrium every
solution must approach zero, so that we have progression to AIDS.
For numerical simulations, it is reasonable to use the parameter values of
(7.34) together with δ = 0.5/day, N = 10 virions per infected T cell. However,
the value of N is not well determined experimentally, and different values may
give quite different behavior.
390 Dynamical Systems for Biological Modeling: An Introduction
Exercises
Use the fact that the endemic equilibrium condition is quadratic to limit
the number of crossings of the two nullclines (two of the four cases above
will nevertheless have two possible portraits each).
3. Derive the inequality (7.33) from the set of conditions that lead to two
endemic equilibria for the system (7.31), as follows:
(a) Find Tc such that F ′ (Tc ) = 0. (The equation has 2 roots, but one
is always outside the interval [0, g/k2 b].) Find the condition that
Tc exist (be a real number.)
(b) Find conditions that 0 < Tc < g/k2 b when a > b. Verify that these
conditions imply that a > b and Tc is real. Show also that these
conditions are equivalent to F ′ (g/k2 b) < 0 < F ′ (0).
√ q 2
(c) Show that F (Tc ) = F (0) + s2 a − (µ − k1 b) kg2 (a − b).
(d) Find conditions that F (0) < 0 < F (Tc ) by solving F (Tc ) > 0 for
µ−k1 b
s2 a . Simplify to (7.33), and verify that it implies all the previous
conditions (i.e., that F (0) < 0, 0 < Tc < kg2 b , a > b, Tc is real).
A
H
a
p /2 p
H R
a 1−P
R
a
y1 + y2 + y3 = N.
Suppose that the density-dependent birth rate Λ(N ) is the same for all geno-
types, while proportional death rates d1 , d2 , d3 for genotypes AA, Aa, aa, re-
spectively, may differ. Initially, we will assume that these death rates are
equal, d1 = d2 = d3 = d. We also make some assumptions on the form of
Λ(N ): namely, that Λ(0) = 0 and that Λ(N ) is differentiable, with
Λ(N )
Λ′ (N ) ≤ for N > 0. (7.40)
N
That is, Λ increases slower than linearly as N increases. This bound prevents
unbounded growth (since the total death rate does increase linearly in N ) and
is satisfied by most common growth functions, including logistic.
To track the quantities of interest, we define the genotype frequencies D
(for dominant), H (for heterozygous) and R (for recessive):
y1 y2 y3
D= , H= , R= ,
N N N
so that D + H + R = 1, and the allele frequencies P for A and 1 − P for a.
Note that P = D + H2 since all of the alleles of the dominant (AA) genotype
and half those of the heterozygous (Aa) genotype are As. Correspondingly
1−P =R+ H 2.
We retain the assumption from Section 2.3 that mating is random with
respect to these alleles and their associated genotypes. As noted above, we
also retain the assumption of genotype-independent fecundity, so that the
birth rate Λ(N ) applies to all genotypes. Then Figure 7.21 shows the possible
offspring for all combination of parent genotypes, and we can write expressions
for the proportions pD , pH , pR of offspring of each genotype, in terms of the
parent frequencies D, H, R. Note that in the figure, y1 , y2 , y3 are deliberately
shown as different in size to emphasize that variations in area within the square
correspond to proportions of offspring with different genetic compositions.
The dark shading shows how the proportion pD of homozygous dominant (D)
offspring can be computed as the number of offspring from two D parents,
half the offspring from a D parent and an H parent, and one quarter of the
offspring of two H parents, that is,
2
1 1 H
pD = D2 + 2 · DH + H 2 = D+ = P 2. (7.41)
2 4 2
Likewise the light shading in Figure 7.21, which represents homozygous reces-
394 Dynamical Systems for Biological Modeling: An Introduction
Finally, the two unshaded areas, both equal in proportion, representing het-
erozygous offspring, yield
H H
pH = 2 D + R+ = 2P (1 − P ). (7.43)
2 2
2
(Note that thus pD + pH + pR = D + H2 + H2 + R = P 2 + 2P (1 − P ) +
2
(1 − P ) = 1, as it should.)
The dynamics of the frequencies D, H, R (and P ) follow from those of
the populations with each genotype. Under the assumptions that mating and
fecundity are genotype-independent, we have
y1′ = pD Λ(N ) − d1 y1 ,
y2′ = pH Λ(N ) − d2 y2 , (7.44)
y3′ = pR Λ(N ) − d3 y3 .
so that
Λ(N ) 2
D′ = (P − D). (7.46)
N
Similarly y2 = HN , so y2′ = H ′ N + HN ′ and H ′ N = y2′ − HN ′ , from which,
by (7.43),(7.44),(7.45),
so that
Λ(N )
H′ = (2P (1 − P ) − H). (7.47)
N
Systems with Sustained Oscillations and Singularities 395
H′ Λ(N ) 2 Λ(N )
P ′ = D′ + = (P − D) + (P (1 − P ) − H/2)
2 N N
Λ(N ) H Λ(N ) 2
= P 2 + P (1 − P ) − D + = (P + P − P 2 − P ) = 0,
N 2 N
from which we see that the frequency of allele A remains constant over time
(P ′ = 0), and thus so does the frequency (1 − P ) of allele a. This makes sense
given that all genotypes reproduce at the same rate and die at the same rate:
each member of y1 reproduces pairs of A’s at the same rate as each member
of y3 reproduces pairs of a’s and each member of y2 reproduces sets of A and
a.
The result that P ′ = 0 answers one of the original questions: continuous,
density-dependent population dynamics does not disturb the constancy of the
allele frequencies. To track the genotype frequencies over time, we need only
adjoin equations (7.45),(7.46), since H can be found from D and P , and then
R from D and H (or directly as R = 1 + D − 2P ). The system for both
frequency types can thus be summarized as
Λ(N ) 2
N ′ = Λ(N ) − dN, D′ = (P − D), P ′ = 0. (7.48)
N
Since the first equation, (7.45), is independent of the other two, the pop-
ulation dynamics can be analyzed separately. The assumptions made on the
form of Λ can be used (see Exercise 1) to show that (7.45) has a unique,
globally stable positive equilibrium if and only if Λ′ (0) > d; otherwise the
extinction equilibrium N ∗ = 0 is globally stable. In what follows, therefore,
we assume that Λ′ (0) > d, so that N (t) approaches a stable positive limit, say
K.
If we consider the two-dimensional system in N and D in (7.48) with
P = P0 , it has the unique positive equilibrium (K, P02 ). Linearization produces
the Jacobian matrix
′
Λ (N ) − d 0
J(N, D) = Λ (N ) Λ(N )
′
) ,
N − N2 − Λ(N
N
since from (7.40) Λ′ (K) < Λ(K)/K = dK/K = d, making both factors of
the determinant negative. Thus, by the equilibrium stability theorem of Sec-
tion 5.4, the equilibrium (K, P02 ) is locally asymptotically stable. In addition,
since trJ(N, D) < 0 for all N, D > 0, a theorem called Bendixson’s Crite-
rion implies that the system (7.45),(7.46) does not have a periodic orbit in
the first quadrant. Since 0 ≤ N ≤ K, 0 ≤ D ≤ 1, solutions of (7.45),(7.46)
are bounded, and now by the Poincaré-Bendixson theorem the equilibrium
(K, P02 ) is globally asymptotically stable (see Exercise 2 for discussion of
N ∗ = 0). Since H = P − D and R = 1 + D − 2P the genotype frequen-
cies D, H, and R approach P02 , P0 (1 − P0 ), and (1 − P0 )2 , respectively.
In the discrete generation model of Section 2.3 the genotype frequencies
were constant. With nonlinear population dynamics included this is not neces-
sarily true (unless the population starts out already at this equilibrium), but
still the Hardy-Weinberg distribution is approached in the limit. This result
serves as a baseline for studying the effects of varying genotype fitness.
Note that, for small ǫ, P is a slow variable and N and D are fast variables;
in particular, the inner system (ǫ = 0) is given by (7.48), the system with
genotype-independent death(!). As seen in the previous subsection, the inner
solution approaches the Hardy-Weinberg proportions, D → P 2 (with P a
constant, say P0 ). The outer system, which governs the long-term evolution
of the slow variable P , is obtained using the slow-time variable10 s = ǫt, so
that dP 1 dP 2
ds = ǫ dt , and substituting D = P in (7.54) (see Exercise 3), to get
dP
= f (P ) = (∆D + ∆R )P (1 − P )(P ∗ − P ), (7.56)
ds
where
∆R
P∗ = .
∆D + ∆R
The equation (7.56) has three equilibria, P = 0, P = P ∗ , P = 1; the equilib-
rium P ∗ is biologically meaningful only if 0 ≤ P ∗ ≤ 1. The question now is
which of the meaningful equilibria are asymptotically stable, and this depends
on the signs of ∆D and ∆R .
We consider four separate cases depending on these signs, cf. Figure 7.22.
We sketch out the results in each case and leave the details for the reader to
work through (Exercises 4–7):
10 The variable s is here used to represent slow time relative to t, in contrast to its use in
but some studies suggest that it instead conferred protection against smallpox.
Systems with Sustained Oscillations and Singularities 399
∆
R
Case I Case III
D fittest H fittest
A dominates overdominance
0 ∆D
Case IV Case II
H least fit R fittest
underdominance a dominates
FIGURE 7.22: The slow dynamics of equation (7.56) are determined by the
signs of the differential mortality rates ∆D and ∆R .
described in this section is one where both the short-term and long-term
behavior are of interest.
Exercises
Λ′ (0) > d but the graph of y = Λ(x) has already crossed below the
line y = d · x).
(c) The two previous parts prove that a unique positive equilibrium
exists iff Λ′ (0) > d. The extinction equilibrium N ∗ = 0 of (7.45)
always exists, by inspection. Use linearization to determine stabil-
ity conditions for each equilibrium and complete the proof of the
desired result.
2. In this section we analyzed the stability of the equilibrium (K, P02 ) of sys-
tem (7.45),(7.46). Although the right-hand side of (7.46) is undefined
at N = 0, we can extend the system to this point by observing that
limN →0 Λ(N )/N = Λ′ (0) from the definition of the derivative, and sup-
posing that Λ′ (0) is positive (and finite). Then let D′ = h(N )(P 2 − D),
where h(N ) = Λ(N N
)
if N > 0 and h(0) = Λ′ (0). Then (0, P02 ) is an
equilibrium of this system. Determine the condition for its stability by
taking the limit of the expressions related to the Jacobian in the text,
as N → 0. How does this fit with the stability of (K, P02 )?
3. Derive the differential equation (7.56) from (7.54) using the change of
independent variable s = ǫt and the slow-time equilibrium condition
D = P 2.
4. Complete the analysis of Case I for equation (7.56), in which ∆D <
0, ∆R > 0. You will need to consider two subcases, depending on the
sign of ∆D + ∆R , but the result should be the same for both.
5. Complete the analysis of Case II for equation (7.56), in which ∆D >
0, ∆R < 0.
6. Complete the analysis of Case III for equation (7.56), in which ∆D >
0, ∆R > 0.
7. Complete the analysis of Case IV for equation (7.56), in which ∆D <
0, ∆R < 0.
8. Determine the asymptotic behavior of the system (7.56) in the case
∆D = 0, and interpret the results in terms of the genotype distribution.
9. Determine the asymptotic behavior of the system (7.56) in the case
∆R = 0, and interpret the results in terms of the genotype distribution.
10. * Rewrite system (7.48) to incorporate genotype-dependent reproduc-
tion (while maintaining genotype-independent mating preferences and
death): Let AA homozygous individuals have 1 + ǫ∆D times as many
offspring as heterozygous individuals, and aa homozygous individuals
1 + ǫ∆R times as many. Derive new differential equations for N, D, P .
402 Dynamical Systems for Biological Modeling: An Introduction
F = ma = mx′′ , which produce the equation mx′′ = −kx equivalent to the equation given
above.
Systems with Sustained Oscillations and Singularities 403
y ′ = z, (7.60)
z ′ = −qy − pz,
Example 1.
Find the general solution of the differential equation y ′′ + 4y ′ + 3y = 0, and
the particular solution satisfying the initial conditions y(0) = 1, y ′ (0) = 0.
Solution: The characteristic equation is λ2 +4λ+3 = 0, with roots −3 and −1.
Thus the general solution of the differential equation is y = K1 e−3t + K2 e−t .
To satisfy the given initial conditions, we differentiate this general solution,
obtaining y ′ = −3K1 e−3t − K2 e−t and then substitute t = 0, y = 1, y ′ = 0.
This gives the conditions 1 = K1 + K2 , 0 = −3K1 − K2 , whose solution is
K1 = − 21 , K2 = 23 . Thus the solution of the initial value problem is y =
− 21 e−3t + 32 e−t .
Example 2.
Find the solution of the initial value problem
y ′′ + 2y ′ + y = 0, y(0) = 0, y ′ (0) = 1.
Example 3.
Find the general solution of the differential equation y ′′ + ω 2 y = 0.
Solution: The characteristic equation λ2 + ω 2 = 0 has complex conjugate
roots ±iω, and thus the general solution of the differential equation is y =
K1 cos ωt + K2 sin ωt.
Example 4.
We can consider the motion of the human diaphragm in breathing as that of a
massive spring being moved up and down by actuating forces: muscles which
pull it down during inhalation (and, to some extent, up during exhalation),
the natural springlike restoring force of the diaphragm itself, and a frictionlike
resistance proportional to (and opposite) the movement. If we define y to be
the downward displacement of the diaphragm (y = 0 at rest), we can write a
simple model for the net force F acting on the diaphragm:
F = f − ky − ry ′ ,
my ′′ + ry ′ + ky = f.
y ′ = z,
k r f
z′ = − y− z+ .
m m m
Finally, the characteristic equation (which can be read from the single second-
order equation with f = 0, cf. (7.59), (7.61)) is mλ2 + rλ + k = 0.
Note, however, that because of the external forcing f , this model is not ho-
mogeneous, so finding the solution is somewhat more complicated than finding
the solution to a homogeneous equation, as done in the previous examples of
this section.
Systems with Sustained Oscillations and Singularities 405
Example 5.
Write the van der Pol equation as a system.
Solution: We let z = y ′ , and then z ′ = y ′′ = −µ(y 2 −1)y ′ −y = −µ(y 2 −1)z −y,
and we have the system
y ′ = z, (7.62)
′ 2
z = −µ(y − 1)z − y.
We note that because of the special form of the converted system (7.58),
any equilibrium of such a system satisfies z = 0, and thus may be found
simply by solving the equation F (y, 0) = 0 for y. The linearization of the
system (7.58) at an equilibrium (y∞ ,0) is the system
u′ = v,
v ′ = Fy (y∞ , 0)u + Fz (y∞ , 0)v,
Example 6.
Find the equilibria of the van der Pol equation, the linearization at each
equilibrium, and the stability of each equilibrium.
Solution: An equilibrium of (7.62) is found by solving z = 0, −µ(y 2 −1)z −y =
13 B. van der Pol and J. van der Mark, The heartbeat considered as a relaxation oscillation,
or
u′′ − µu′ + u = 0,
since the linear approximation to y 2 − 1 at y = 0 isp−1. The roots of the
characteristic equation r2 − µr + 1 = 0 are r = 21 µ ± µ2 − 4 . Since µ > 0,
these roots are both positive if µ > 2, and complex conjugates with positive
real part if µ < 2; in either case, the origin is an unstable equilibrium.
It is possible to show, although not without some difficulty, that every orbit
of the system (7.62) is bounded, and this implies by the Poincaré-Bendixson
theorem that there is a limit cycle. In fact, there is a unique limit cycle ap-
proached by every orbit, as suggested by the phase portrait in Figure 7.24
with µ = 1.5.
z(t)
2
–2 –1 1 2
y(t)
–2
–4
The van der Pol equation is usually treated as a system, but more infor-
mation can be derived by using a different system from (7.62) which allows
us to see more clearly the role of the parameter µ in determining the shape
of solutions (recall µ determines the strength of the nonlinear term). We first
define the function f (y) = y 2 − 1, so that the van der Pol equation may be
written as y ′′ + µf (y)y ′ + y = 0. We then define
Z y
y3
F (y) = f (u) du = − y, w = y ′ + µF (y).
0 3
Then by the Chain Rule
y ′ = w − µF (y),
w′ = −y.
y
z 1
1.5
1.0
10 20 30 40 50
0.5 t
y
-2 -1 1 2
–1
-0.5
-1.0
–2
-1.5
independent of axial height z and angle θ (around the axis, see Figure 7.27).
Assuming these two symmetries allows us to consider only the matter of how
oxygen is transported from the boundary of the fiber (at r = a) inward to the
center (r = 0).
z
θ
r=0 r=a
where the constant K1 describes the rate of diffusion. Since (under our as-
sumptions of symmetry) the diffusion at the center is even, the rate of change
of the concentration should level off as one approaches the center; this gives
us one boundary condition, dy dr (0) = 0. The concentration of oxygen entering
the fiber from outside (at r = a) is known, giving us our second boundary
condition, y(a) = Ya . (Since the diffusion begins outside the fiber boundaries,
we cannot know that dy dr (a) = 0.)
By completing the differentiation in (7.64) we can obtain a second-order
differential equation, which can be rewritten as a first-order system and ana-
lyzed using the methods illustrated in this section. Exercises 19–22 below out-
line some extensions and analysis of model (7.64). The simplest model gives
a quadratic distribution for y(r), reflecting the higher concentration closer to
the source (r = a). Mathematically the most interesting feature of the model
and its extensions is that the nature of the first boundary condition, specifying
dy/dr rather than r at 0, means there are no boundary layers for y: a slight
correction (corresponding to a fast-time or inner solution) is needed only for
dy/dr(0), not for y(0).
A more realistic model would consider the specific locations (in terms of
angle θ) of the surface capillaries from which diffusion originates, and might
further consider what happens when the muscle contracts, increasing the ra-
dius and changing the locations of the capillaries.
Exercises
In each of Exercises 1–12, find the general solution and the solution satisfying
the initial conditions y(0) = 1, y ′ (0) = −1.
1. y ′′ = 0 7. y ′′ + 5y ′ + 10y = 0
2. y ′′ − 9y = 0 8. 4y ′′ + 4y ′ + 13y = 0
3. y ′′ − 5y ′ + 6y = 0 9. y ′′ − 4y ′ + 13y = 0
4. y ′′ + 9y = 0 10. y ′′ − 4y ′ + y = 0
5. y ′′ + 10y ′ + 25y = 0 11. y ′′ − 2y ′ + 2y = 0
6. 4y ′′ − y = 0 12. ǫy ′′ + 2y ′ + y = 0 (0< ǫ < 1)
16. y ′′ + g(y) = 0, where g(y) is a smooth function with yg(y) > 0 for y 6= 0.
17. y ′′ + f (y)y ′ + g(y) = 0, where f (y) is a smooth function with f (y) > 0
for all y and g(y) is a smooth function with yg(y) > 0 for y 6= 0.
18. * To illustrate the complexity of the structure when the origin is not an
isolated equilibrium, consider the differential equation y ′′ + y ′ = 0.
(a) Show that there is a line of equilibria.
(b) Sketch the phase portrait in the y-y ′ phase plane.
19. In the model (7.64), treating the conversion rate f as a constant, com-
plete the differentiation in the above model, and write it first as a second-
order differential equation, and then as a system of two first-order equa-
tions. Can you solve the resulting equation for y ′ ?
20. * Given that oxymyoglobin is not consumed within the muscle fiber (it
acts to transport the oxygen), but that the oxygenization process men-
tioned above increases the amount of oxymyoglobin present, and also
that the oxymyoglobin does not leave the fiber (i.e., diffusion is limited
to 0 ≤ r ≤ a), write an equation modeling the radial diffusion of oxymyo-
globin in the muscle fiber, analogous to the one given above for oxygen,
using w(r) to represent the [radial] concentration of oxymyoglobin.
21. * In fact, f is not a constant, but depends on the amounts of reactants
(oxygen, myoglobin and oxymyoglobin) present, as well as on some con-
stants of proportionality. Let us take f = K3 xy − K4 w, where w is the
concentration of oxymyoglobin, x and y are the concentrations of myo-
globin and oxygen, respectively, and K3 and K4 are the constants of
proportionality. Rewrite the second-order equations for y and w using
this expression for f , and transform them into a system of four first-order
differential equations.
22. * Approximate values for the parameters in this model are as follows:
a = 2.5 × 10−3 cm, Ya = 3.5 × 10−8 mol/cm3 , g = 5 × 10−8 mol/cm3 sec,
K1 = 10−5 cm2 /sec, K2 = 5 × 10−7 cm2 /sec (diffusion rate constant
for oxymyoglobin), K3 = 2.4 × 1010 cm3 /mol sec, K4 = 65/sec, x =
2.8 × 10−7 mol/cm3 (the concentration of myoglobin in muscle fibers).14
Use a computer to find the distributions the two models predict for
oxygen and myoglobin as functions of r.
p. 42.
This page intentionally left blank
Part III
Appendices
413
This page intentionally left blank
Appendix A
An Introduction to the Use of
MapleTM
Maple is a computer algebra system which can assist you in solving math-
ematical problems which would be quite inaccessible without its help. This
introduction is not intended to make you an expert on Maple but it should be
a reasonably self-contained start which will enable you to handle many prob-
lems in the analysis of ordinary differential equations and difference equations
which are found in this book. Maple is intended as an aid, not as a replacement
for thinking. Ideally, you will begin by working on a problem to understand
what is required and will turn to Maple to provide graphic output where ap-
propriate. While Maple has considerable capacity for calculation (good enough
to pass most calculus courses), this is an aspect which we shall try not to em-
phasize. Our use of Maple will be concentrated on
(i) plotting graphs of functions or pairs of functions and thus finding inter-
sections of two graphs,
(ii) graphical solution of ordinary differential equations, including first-order
equations and systems of first-order differential equations,
(iii) difference equations, including the cobweb method for first-order differ-
ence equations and the graphing of solutions of difference equations of
first or higher order by iterative calculations.
Maple can be run on many different kinds of computers, including Win-
dows, Macintosh, and many UNIX systems. On many university campuses it is
installed on computers in student labs. In addition, for students planning on a
career which will use mathematics, the “academic version” of Maple available
for sale at a moderate price in many universities is an excellent investment.
Maple can be used for many “word-processor” functions to dress up your
worksheets and convert them into more formal reports. You can toggle between
Maple ([>) and text (T) input on the toolbar in order to add text. Inside a
Maple computation you can add comments; any Maple input beginning with
# is treated as a comment and is ignored as far as calculation is concerned.
Note that every time you hit the “Enter” key you begin a new input and if
you want a comment to extend for more than a paragraph you need to begin
each paragraph with “#”.
415
416 Dynamical Systems for Biological Modeling: An Introduction
Here are a few general instructions for using Maple. They are not meant
to be complete, and a manual on the use of Maple or the online Help available
in Maple contains much more useful information.
Every input for which you want Maple to do something must end with a
semicolon. If you want Maple to act but not display the output you would use
a colon.
The command restart at the beginning of a worksheet tells Maple to
forget any previous instructions. If you have been working on one problem
and defined a symbol and then begin working on another problem which uses
the same symbol, this instruction will avoid confusion.
The command with tells Maple to read in a package containing some
additional definitions. For example, when we solve differential equations we
will always use the package DEtools and thus will begin the worksheet with
the command with(DEtools):. If we use a semicolon rather than a colon at
the end of the command, Maple will print out a list of the commands included
in the package.
The syntax for defining a quantity for later use is : =. Thus the command
a: = 1 assigns the value 1 to the quantity a. This syntax is also used to define
functions.
0.4
0.3
0.2
0.1
0.5
y
0 0 1 2 3 4 5
1 2 x 3 4 5 x
> plot([f(x),g(x)],x=0..5);
(see Figure A.2)
> plot([f(x),g(x)],x=0..5,y=0..0.5);
(see Figure A.3)
> plot([f(x),g(x)],x=0..5,y=0..0.5,scaling=constrained):
(see Figure A.4)
1.4 1.4
1.2 1.2
1 1
y(t)0.8 y(t)0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 1 2t 3 4 0 1 2t 3 4
–0.2 –0.2
–0.4 –0.4
–0.6 –0.6
–0.8 –0.8
–1 –1
1.4
1.2
1
y(t)0.8
0.6
0.4
0.2
0 1 2t 3 4
–0.2
–0.4
–0.6
–0.8
–1
x=0..50,y=0..40,scene=[x,y],stepsize=0.05,linecolour=black,
thickness=0,arrows=none);
(see Figure A.8)
40
30
y
20
10
0 10 20 30 40 50
x
> DEplot({diff(x(t),t)=1*x*(1-x/40)
-x*y/(x+10),diff(y(t),t)=1*y*10*(x-20)/(30*(x+10))-0.4},
{x(t),y(t)},t=0..25,[[x(0)=50,y(0)=10],[x(0)=50,y(0)=20],
[x(0)=50,y(0)=30],[x(0)=50,y(0)=15],[x(0)=50,y(0)=5]],
x=0..50,y=0..40,scene=[t,x],stepsize=0.05,linecolour=black,
thickness=0,arrows=none);
(see Figure A.9)
> DEplot({diff(x(t),t)=1*x*(1-x/40)
-x*y/(x+10),diff(y(t),t)=1*y*10*(x-20)/(30*(x+10))0.4},
{x(t),y(t)},t=0..25,[[x(0)=50,y(0)=10],[x(0)=50,y(0)=20],
[x(0)=50,y(0)=30],[x(0)=50,y(0)=15],[x(0)=50,y(0)=5]],
x=0..50,y=0..40,scene=[t,y],stepsize=0.05,linecolour=black,
thickness=0,arrows=none);
(see Figure A.10)
50 40
40
30
30
x y
20
20
10
10
0 2 4 6 8 10 12 t 14 16 18 20 22 24 26 0 2 4 6 8 10 12 t 14 16 18 20 22 24 26
0.8
0.6
0.4
0.2
0.8
0.7
0.6
0.5
0.4
0 2 4 6 8 10 12 14 16 18 20 22 24
10 8
8
6
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
FIGURE A.13: Solution plot for FIGURE A.14: Solution plot for
yn in system. zn in system.
424 Dynamical Systems for Biological Modeling: An Introduction
0.8
0.6
0.4
0.2
′
Likewise, we may write g (t) as the integral of its derivative,
Z t
g ′ (t) = g ′ (y∞ ) + g ′′ (s)ds. (B.2)
y∞
425
426 Dynamical Systems for Biological Modeling: An Introduction
2
The error term can be expressed as g ′′ (c) u2 , with c some (unknown) point
between y∞ and y∞ + u. For our purposes, the factor u2 , which for small u
makes the error term negligible compared to the linear term, is what matters.
There is a two-dimensional version of Taylor’s theorem, which is needed
for linearization of two-dimensional systems (Sections 2.8, 5.4). We wish to
expand a function of two variables, g(y∞ + u, z∞ + v) in powers of u and v
with coefficients which depend on the value of the function g and its partial
derivatives at the base point (y∞ , z∞ ). In order to reduce this problem to
Taylor’s theorem in one variable, we define a function of one variable t,
Then G(0) = g(y∞ , z∞ ) and G(1) = g(y∞ + u, z∞ + v). Taylor’s theorem for
one variable says that
where R is the error term. We may use the chain rule for partial derivatives
to calculate G′ (t) and G′′ (t):
G′′ (t) = u2 gyy (y∞ +tu, z∞ +tv)+2uvgyz (y∞ +tu, z∞ +tv)+v 2 gzz (y∞ +tu, z∞ +tv)
Substitution of t = 0 gives
and it is possible to check that the error term R is quadratic in u and v, that
is, it is a sum of terms which are multiples of u2 , uv, and v 2 . This is the
expression which is needed for linearization of a two-dimensional system at an
equilibrium (y∞ , z∞ ).
Appendix C
Location of Roots of Polynomial
Equations
f (λ) = λ2 + a1 λ + a2 = 0. (C.1)
427
428 Dynamical Systems for Biological Modeling: An Introduction
If the roots are real, in order for both to be negative their product must be
positive and their sum must be negative. On the other hand, the conditions
a1 > 0, a2 > 0 are sufficient. To see this, we note that if a21 − 4a2 ≥ 0, so that
the roots are real, a2 > 0 implies that they have the same sign and a1 > 0
implies that both are negative. If a21 − 4a2 < 0, so that the roots are complex
conjugates, a1 > 0 implies that they have negative real part. To sum up, we
have obtained the following result:
or −(a2 + 1) < a1 < (a2 + 1), and this is equivalent to |a1 | < a2 + 1. The
condition |a2 | < 1 may be written −1 < a2 < 1, or 0 < a2 + 1 < 2, and we
may combine our two necessary conditions into the double inequality
These conditions are also sufficient. If (C.2) is satisfied, then f (−1) > 0
and f (1) > 0. If the roots are real, this implies that either both are less than
−1 or both between −1 and +1, or both greater than 1. However, if both
are less than −1 or greater than 1, their product is greater than 1, which
contradicts a2 < 1. Thus if the roots are real they must be between −1 and
1. If the roots are conjugate complex, their product a2 is the product of their
absolute values and thus less than 1. Since both roots have the same absolute
value, they must both have absolute value less than 1.
To sum up, we have obtained the following result:
429
430 Dynamical Systems for Biological Modeling: An Introduction
for k = 1, 2, . . .. This shows that each of the terms yk is in the interval I, that
the successive terms come closer to y∞ , and that yk approaches the limit y∞
as k → ∞. Thus every solution starting close enough to y∞ remains close to
y∞ and approaches y∞ as k → ∞, and this establishes asymptotic stability
of y∞ .
If |g ′ (y∞ )| > 1, we may pick a number ρ > 1 and an interval centered at
y∞ on which |g ′ (y)| > ρ and use the Mean Value Theorem in much the same
way but with the inequalities reversed to give
From this we see that the terms yk move away from y∞ and thus that the
equilibrium y∞ is unstable.
Geometrically, this means drawing the tangent line to the curve z = f (y)
at the point y0 and taking y1 to be the point where this tangent line meets
the y-axis (i.e., taking the tangent line as an approximation to the function
near y0 and finding where this approximation is zero). This process is then
repeated.
In order to view Newton’s method as a stability of equilibrium question,
we define
f (y)
g(y) = y − ′ .
f (y)
Then an equilibrium y∞ of the difference equation yk+1 = g(yk ) is a solution
of the equation f (y) = 0 (provided f ′ (y) 6= 0). Since
Chapter 2
Section 2.1
2. a.
µ µ
wn = w0 − Fn +
1−F 1−F
µ
b. If F < 1, then eventually wn → 1−F . This means that if the
weed is less fit than its competitor, then it will approach the given
equilibrium density thanks to the constant growth due to mutation.
c. If F > 1, meaning the weed is more fit than its competitor, then wn
grows without bound, i.e., weed density becomes arbitrarily high.
d. Unbounded growth in a single patch is unrealistic, so the model’s
operating range is limited (in w) when F > 1.
3. y1 = 3/16, y2 = 39/256, y3 = 8463/65536.
5. y1 = 1/2, y2 = 1/3, y3 = 1/4.
11. yk = (1.1)k .
13. yk = 1/2k .
15. yk = 1.
17. yk = (1/2)k − 1.
19. yk → 0.
21. yk → 5/4.
433
434 Answers to Selected Exercises
Section 2.2
8. yk → 0.
Section 2.3
1. Equilibrium y = 2, unstable.
3. Equilibrium y = β/(1 − α), unstable.
5. Equilibrium y = 1 + r if r > 1/e, asymptotically stable for r < e.
10. a. (2.21) becomes
F pn + (1 − pn )
pn+1 = .
F pn + 2(1 − pn )
The p = 0 equilibrium vanishes, and the coexistence eqm becomes
1/(2 − F ). The equilibrium p = 1 is again stable for F > 1, and
1/(2 − F ) is for F < 1.
Answers to Selected Exercises 435
b. Coexistence occurs only for F < 1, i.e., the a allele only survives if
the Aa zygote is more viable than the AA zygote.
c. We can rewrite the G = 0 model as pn+1 = K+1 K+2 , where K =
F pn /(1 − pn ) > 0. Since K+1
K+2 > 1
2 for K > 0, this guarantees A
alleles will always be in the majority. This makes sense because all
individuals who reproduce must have at least one A allele, so at
least half the alleles contributed to the next generation are As.
e.
13. a.
F αn + 14 (1 − αn − βn )
αn+1 =
F αn + (1 − αn − βn ) + Gβn
Gβn + 41 (1 − αn − βn )
βn+1 =
F αn + (1 − αn − βn ) + Gβn
Section 2.4
1. Equilibrium y = 0 is asymptotically stable for all r.
3. r < e.
√ √
5. H < ( r − A)2 .
7. H < 1/10.
ayk , byk < 1;
a
14. yk+1 = yk , byk = 1;
2
0, byk > 1.
Section 2.5
1. a. 233.3 (naturally occurring), 182.56 (field experiment)
b. 158.68 (naturally occurring), 124.15 (field experiment)
c. 154, 62.1 (naturally occurring), 104.9, 0 (field experiment)
2. a. 80.687
b. 78.92
4. √ √
( a − 1)2 a−1 ( a + 1)2
< <
b b b
5. Multiply the desired double inequality by 2, subtract ( a−1
b − H), and
rewrite as a single inequality
√
s
2
( a − 1)2
a−1
H− < − H − 4H/b.
b b
Section 2.6
1. r = e.
8. a. yk+1 = g((1 − d)yk )
b. Both are the same, r(1 − d).
Answers to Selected Exercises 437
Section 2.7
1. y = 0, z = 0, unstable.
3. y = 0, z = 4, unstable.
5. y = 0, z = 0, unstable.
a−µ
7. y = 0, asymptotically stable if 0 < µ − a < 2 and y = bµ , asymptoti-
cally stable if µ < a, µa (a − µ) < 2.
Section 2.8
2. The ragwort is the factor limiting the moth population size and/or repro-
ductive ability; a certain amount of ragwort is needed to provide energy
to produce each egg, and the amount available is less than the total
reproductive capacity of the moths in the absence of any limitations.
3. The equilibrium (B, 0) is stable if and only if aB √ < 1; the equilibrium
( a1 , ab
1
log aB) is stable if and only if 1 < aB < e.
5c.
r
10
p
0.2 0.4 0.6 0.8 1
438 Answers to Selected Exercises
Chapter 3
Section 3.1
1. 1830.
3. 211.66
5. 17.46.
7. 199 days.
9. 4.5 billion years.
Section 3.2
3. (a) r = 1.0668 × 10−5 , (b)205(1990), 206(2000), 208(2010).
5. a. World War I and the ”Spanish” flu epidemic of 1918-19
b. The model is pretty good through 1970. Changing immigration pat-
terns, better census counts, and better food production technology
might explain changes, and changing some parameter values might
give a better fit.
7. Biomass after i year is 32.5 × 106 kg. It takes 1.55 years for biomass to
reach K
2 .
Section 3.3
5. y = (1 − t2 )−1/2 .
1−et 1−et−2
8. y = 1, y = 1+et , y = 1+et−2
Answers to Selected Exercises 439
Section 3.4
2
1. y = cet .
3. y = (c + 2t)1/2 .
2
5. y = ce−t .
3
7. y = 4 + cet /3
.
3
9. y = e(t −125)/3
.
2
11. y = y0 e−t .
2
13. y = 1−t2 .
17. a. The pacemaker signal y follows the control u, with the difference
y − u between them decaying over time at a rate proportional to
that difference.
b. y(t) − u(t) = z(t) = z(0)e−at so y(t) = u(t) + [y(0) − u(0)]e−at .
18. a. The heart beats more slowly (or at least firing rate decreases).
b. Firing rate decreases toward zero: no beating!
c. Solve equation (3.29) for yH and substitute.
19. T1 = T2
20. The solution changes only in the exponent of the exponential in the
r1
denominator of (3.37), which goes from −rt to −r0 t + 2π cos 2πt. As
r1
t → ∞ the oscillation dies out since r0 t >> 2π , so y → K anyway.
21. a. u′ = −µ1 u leads to u(t) = u(0)e−µ1 t and u(tc ) = ryk e−µ1 tc .
440 Answers to Selected Exercises
b. u′ = −µ2 u leads to u(t) = u(tc )e−µ2 (t−tc ) = ryk e−µ1 tc e−µ2 (t−tc )
and u(T ) = ryk e−(µ1 −µ2 )tc e−µ2 T
27. 99.
31. f (t) ≡ 0,and f (t) = t/2.
32. f (t) = 0(t < c), f (t) = (t − c)/2(t > c) for every c ≥ 0.
Section 3.5
1. Let Q(t) denote the weight of salt in the tank at time t, so that Q(0) = 3.
The rate at which salt is flowing
in is 2.5 lb./minute, and the rate at
Q
which salt is flowing out is 5 100 , since the concentration of salt at
Q
time t is 100 lb./gal. and the solution is flowing out at 5 gal./minute.
This leads to the initial value problem
Q
Q′ = 2.5 − , Q(0) = 3,
20
which is of the form (3.46) with k = 0.05, b = 2.5, and has solution
Q = 50 − 47 e−0.05t.
log 5
t= = 2.32 hours.
log 2
11. 40 minutes.
13. 58.3 minutes.
Answers to Selected Exercises 441
Section 3.6
1. y = 1 − e−2t /2.
3. Y = e−2t + 4e−3t .
t2 +1
5. y = 2t .
7. y = t2 /4 + 4/t2 .
et +1−e
9. y = t .
11. y = t3 − 23 t2 + 32 t − 43 [1 − e−2t ].
p
13. y = 3t/2.
15. Y = e2t(t−1) .
19. y = (a + 1)et − 1.
100
24. x(t) = 500/(t + 3 ), x(50) = 6lb.
25. a.
272
s(t) = 1440(25 − t) − (25 − t)5/2
25
b.
s(t)
V (t) = 1000 − 40t = 40(25 − t) so = 36 − 0.272(25 − t)3/2
V (t)
The poor fish only have 3.56 minutes in which to be rescued.
29. The substitution y = 1/z, y ′ = −z ′ /z 2 gives
1 ′ 1 r 1
− z =r − ,
z2 z K z2
r
z ′ + rz = .
K
Using the integrating factor ert we obtain
r rt
ert z ′ + rert z = (ert z)′ = e ,
K
and integration gives
1
ert z = ert + c,
K
1
z= + ce −rt .
K
Returning to the original dependent variable y, we find the family of
solutions
1 K
y= 1 −rt
= −rt
.
K + ce 1 + cKe
This is the same family that we found before, in Section 3.3, except
that the arbitrary constant c in Section 3.3 is replaced by the arbitrary
constant cK here.
442 Answers to Selected Exercises
Chapter 4
Section 4.1
0
FIGURE S.6: Solution of Exercise 1
0
FIGURE S.7: Solution of Exercise 3
0 1 2
FIGURE S.8: Solution of Exercise 5
−1 0
FIGURE S.9: Solution of Exercise 7
Section 4.2
r
1. In case (i), where H < rK/4, we can factor dy/dt = − K (y − y1 )(y − y2 )
which leads to
y(t) − y2 y(0) − y2 y2 − y1
= exp −rt ,
y(t) − y1 y(0) − y1 K
from which
y2 − y1 Q exp −rt y2K
−y1
y(0) − y2
y(t) = , where Q = .
1 − Q exp −rt y2K−y1
y(0) − y1
r
In case (ii), where H = rK/4, we have dy/dt = − K (y − K/2)2 , the
solution to which is
K y(0) − K
2
y(t) = + rt K
.
2 1+ K y(0) − 2
Finally, in case (iii), where H > rK/4, there are no equilibria, so the
quadratic cannot be factored over the reals. Instead we complete the
square: " 2 #
dy r K HK K2
=− y− + − ,
dt K 2 r 4
so that after integration and solving for y(t) we have
" r #
K2
K −1 K rt HK
y(t) = + tan tan y(0) − − − .
2 2 K r 4
444 Answers to Selected Exercises
∗
2. Factoring dy/dt = (r/K)y(yH − y) and the partial fraction expansion
1 1 1 1
∗ − y) = ∗ − ∗ )
y(yH yH y (y − yH
Section 4.3
αKA
1. y = A+(K−A)e−αt .
a
3. y = a − αat−1 .
446 Answers to Selected Exercises
6. First solve (4.18) for Kd to get Kd = (KA − C ∗ )(KB − C ∗ )/C ∗ and then
substitute KA = A∗ + C ∗ , KB = B ∗ + C ∗ .
8.
β(N ) I ∗ 1
R0 = 1, N = 1− R ,
µ+ τ 0
and the disease persists if and only if R0 > 1, that is, the disease-
free equilibrium is globally asymptotically stable if R0 < 1, while the
endemic equilibrium is if R0 > 1.
9. The model formulation would only change inasmuch as β(N ) becomes
β(I), but the equilibrium analysis would be complicated considerably
with an additional function of I in the nonlinear term. If people change
their behavior in response to observed prevalence, β(I) should be a non-
increasing (perhaps even sharply decreasing) function of I. If the health-
care system breaks down beyond some critical level Ic , then β(I) should
be constant below Ic and increasing above it (perhaps up to some max-
imum).
10. The new effective population size for contacts is S + σI, so we get
dI σI 1
= β(S + σI)S − I.
dt S + σI τ
We reduce to one variable as before with S = N −I, S+σI = N −(1−σ)I.
1
11. a. R0 = σβ1 N τ , and I ∗ /N = 1 − R0 as before.
∗
b. R0 = σβ0 τ , and I /N = (R0 − 1)/(R0 − 1 + σ).
12. a. The recovery term disappears and R0 → ∞.
b. The proportion of infectives at the endemic equilibrium approaches
1: everyone is infected. This is unrealistic primarily because it ig-
nores timescale issues, such as demographic renewal.
13. Of course, dI/dt gets a new term −δI, but more significantly N is no
longer constant, so we cannot represent the system with a single equa-
tion. We must include an equation for either dS/dt or dN/dt.
14. a. In addition to the DFE, endemic equilibria
∗
r
I±
1 4
= 1± 1−
N 2 βτ N 2
since 0 < 1 − βτ4N 2 < 1, the square root is larger than the radi-
cand, and we conclude that f ′ (I−∗
) > 0 while f ′ (I+∗
) < 0, so that the
upper endemic equilibrium is LAS and the lower one is unstable.
This model behaves differently than the basic mass-action epidemic
model in two important ways: first, the DFE is always LAS, which
implies that R0 = 0 (that is, a single infective introduced into a
population of susceptibles never causes an outbreak, because infec-
tion requires many infectives (I 2 )); and second, for βτ N 2 > 4, the
persistence of the infection depends on the initial number of infec-
∗
tives: if it is high enough (I(0) > I− ), it will establish an endemic
state; otherwise it will die out quickly.
b. This model behaves
√ mostly like the usual mass-action incidence
model, with R0 = βτ .
16. a. dy/dt = ry(1 − y/K) − kxy
b. dx/dt = cxy − µx
c. With y constant, the model becomes linear: dy/dt = (cY − µ)x,
and the predators either grow unchecked (if cY > µ) or die out (if
cY < µ).
d. With x constant, the model becomes logistic, dx/dt = r2 y(1 −
y/K2 ), where r2 = r − kX and K2 = K(r − kX)/r, so that the
prey population x approaches a smaller carrying capacity K2 as
long as r > kX, and dies out if r < kX.
Section 4.4
1. If r > 0, no equilibrium. If r = 0,y = 0 is unstable. If r < 0, there are
two equilibria, one unstable and one asymptotically stable.
3. y = 0 is asymptotically stable if and only if r < 0. y = −r is asymptot-
ically stable if and only if r > 0.
p
5. If r ≥ 0, y = 0 is unstable. If r < 0, y = 0 is unstable and y = ± −1/r
are asymptotically stable.
7. If −2 < r < 2 there are no equilibria. If r < −2 or r > 2, there are two
equilibria, one unstable and one asymptotically stable.
y
14. Let y = x/K and τ = ts/K. Then dy/dτ = a + by − cy 2 − y+1 , where
a = αn/s, b = (βn − α)K/s, and c = βK 2 /s.
15. a. We use A as the unit of population size and make the change of
y
dependent variable u = A . This transforms the equation (4.29) to
u2
′ Au
Au = rAu 1 − −H
K u2 + 1
448 Answers to Selected Exercises
or
u2
A ′ A Au
u =r u 1− − 2 .
H H K u +1
Ar K
We define the new parameters R and Q by R = H , Q= A and
then we have the model
u2
A ′ u
u = Ru 1 − − 2 . (S.3)
H Q u +1
H
We could also make the change of independent variable s = A t,
which would replace (S.3) by
u2
du u
= Ru 1 − − 2 ,
ds Q u +1
but we need not do this in order to carry out the equilibrium analy-
sis. In terms of the original independent variable t we have the same
equilibria and stability; the change of dependent variable changes
only the exponential rate of approach to equilibrium. Effectively, we
have replaced the original four parameters by the two parameters
R and Q. An examination of the forest changes indicates that un-
der change of the forest, the parameter Q remains almost constant,
but the parameter R grows with the development of the forest. The
process of change in the parameter in many applications is the rea-
son for our study of bifurcations. In this case we will consider R to
be our bifurcation parameter.
b. Equilibria
of (S.3) are u = 0 and the intersections of the line v =
u u
R 1 − Q and the curve v = 1+u 2 , solutions of the equation
u u
R 1− = . (S.4)
Q 1 + u2
8 food
supply B
limitations
(M)
u
4
(L)
0.5 1 1.5 2
R
r
will be stable which have f ′ (u) < 0 there, i.e., where the slope of
the line is less than the slope of the curve. From the graph we can
see that this occurs at K and M , but not at 0 or L. Therefore
we have an unstable equilibrium at u = 0 and two asymptotically
stable equilibria K and M separated by an unstable equilibrium L.
The domain of attraction of the small equilibrium K is the interval
[0, L), and the domain of attraction of the large equilibrium M is
the interval (L, ∞).
If R (the slope of the line) increases until K and L coalesce, as in
Figure S.12, then only the equilibrium M remains. Thus if a system
is in equilibrium at K and R increases, there would be a jump from
the small equilibrium K to the large equilibrium M. Conversely, if
R is decreased until L and M coalesce, as in Figure S.13, then
only the equilibrium K would remain, and a system which is in
equilibrium at the large equilibrium M would crash to the small
equilibrium K. These two situations therefore involve bifurcations
like the saddle-node bifurcation seen earlier in this section.
c. We think of the parameter R as representing resources of the forest
as food supply for the budworms, and the equilibria K and M as
corresponding to budworm limitation by predators and food sup-
ply, respectively. In a young forest, R is small with only the small
equilibrium K (limited by predators). As the forest develops, the
food supply becomes so great that budworm growth exceeds con-
trol by predators, and R increases to the stage where there is only
the large equilibrium M , signifying an outbreak of the budworm
population. If this outbreak destroys the forest, the predators may
regain control as R decreases, and the population may crash to the
small equilibrium K. The model exhibits hysteresis; the outbreak
and crash will generally occur at different population levels. This
450 Answers to Selected Exercises
1.6 1.6 1.6
1 1 1
M M
0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12
u u u
may also be seen from the model (S.3) by drawing the graph of
the equilibria as a function of the bifurcation parameter R. The
equation (S.4) is easily solved for R as a function of y, and the
bifurcation curve is just the (multi-valued) inverse function (Fig-
ure S.10). When R grows from a small value, the equilibrium moves
along the bifurcation curve until it reaches the point A, and then
jumps to the higher point B and continues upwards. When R de-
creases from a high value, the equilibrium moves down the bifurca-
tion curve past B to C before crashing to the lower value D. Thus
the discontinuity in equilibrium on the way up is at a higher value
of R than the discontinuity in equilibrium on the way down.
Section 4.5
1. The modified Euler method gives 2.70796 instead of 2.5937, an error of
0.38% rather than 4.6%.
3. The Runge-Kutta approximation gives
8. h < 2.
10. 4.6 minutes
11. 8.0 minutes; there are fewer ants wandering around, so it takes longer
for the pheromone trail to become established enough to attract many
ants
13.
rmq qP q−1
f1′ (P ) = 2,
(mq + P q )
rqmq P q−2 [(q − 1)mq − (q + 1)P q ]
f1′′ (P ) = 3
(P q + mq )
so f1′′ (P ) = 0 when P = 0 or
1/q
q−1
P = P̄ = m .
q+1
Now
q−1
q
q−1
r q q+1
′
f1 (P̄ ) = 2 .
m
q−1
1 + q+1
Chapter 5
Section 5.1
1. Orbits are circles with center at the origin.
3. Orbits are curves sin z = y 2 /2 + c.
5. Orbits are curves z = ceβy/α .
Section 5.2
1. Linearization at (1, 1) is u′ = u + v, v ′ = −u + v.
3. Linearization at (−1, 1) is u′ = u + 2v, v ′ = u. Linearization at (−1, −1)
is u′ = u − 2v, v ′ = u.
5. System has no equilibria.
9. Equilibrium (0, 0) is asymptotically stable.
11. Equilibrium (0, 0) is asymptotically stable.
Section 5.3
1. y = 5K1 e2t + K2 e−4t , z = K1 e2t − K2 e−4t .
3. y = K2 et , z = K1 e2t − K2 et .
5. y = K1 e4t , z = 2K1 te4t , z = 2K1 te4t + K2 e4t .
7. y = K1 e−t + K2 e−2t , z = −K1 e−t − 2K2 e−2t .
Answers to Selected Exercises 453
9. y = K1 tet + K2 et , z = K1 et .
11. y = K1 e3t sin 5t + K2 e3t cos 5t, y = K2 e3t sin 5t + K1 e3t cos 5t.
13. y = K1 tet + K2 et , z = K1 et .
cK1 at
17. y = K1 eat , z = a−d e + K2 eat , (a 6= d).
Section 5.4
1. (1, 1) is unstable.
3. Both (−1, 1) and (−1, −1) are unstable.
5. System has no equilibria.
7. (0, 0) and (1/3, 1/3) are unstable; (0, 1) and (1, 0) are asymptotically
stable.
9. The origin is unstable (saddle point). The equilibrium (a/λ, 0) is asymp-
totically stable if and only if λc < aµ. If λc > aµ, there is an asymptot-
ically stable equilibrium with y > 0, z > 0.
Chapter 6
Section 6.1
1. 1.44.
4. 17.56 %.
5. 2.83 %
9. a. New model is
b. Reduced model is
Section 6.2
1. Equilibrium (60, 20) is asymptotically stable; species coexist.
3. Equilibrium (0, 45) asymptotically stable; y species goes extinct and z
species wins competition.
5. All orbits approach the asymptotically stable equilibrium (10, 10).
7. All orbits approach a limit cycle about the unstable equilibrium
(20, 10.8).
9. All orbits approach a limit cycle about the unstable equilibrium (30, 25).
12. There are three equilibria (0, 0), (0, M ), and (K, 0). They are all un-
stable. If ab≥ 1, orbits areunbounded. If ab < 1, a stable equilibrium
K+aM Mb+K
(y∞ , z∞ ) = 1−ab , 1−ab appears and all trajectories approach this
equilibrium.
15. (0, 0) is asymptotically stable and (70, 120) is unstable.
17. (20, 10 is asymptotically stable, and (0, 0), (20, 0), (0, 10) are unstable.
22.
The system has two equilibria: extinction (0,0) and persistence
d γ 1
γ+ǫd , γ+ǫd × 1 − Rd K; the extinction equilibrium is LAS iff Rd <
1, and the persistence equilibrium exists and is LAS iff Rd > 1. Here
γ
Rd = db µ+γ gives the birth-to-adult-death ratio, multiplied by the pro-
portion of juveniles who survive to maturity. Stability follows from stan-
dard methods using the Jacobian matrix and Routh-Hurwitz criteria.
Section 6.3
1. The Runge-Kutta approximation gives
t 0 2 4 6 8 10
y 40 22.99 17.27 13.71 11.92 11.94
z 10 13.25 15.07 15.21 14.00 12.44
Answers to Selected Exercises 455
Chapter 7
Section 7.1
In Exercises 1–9, only Exercise 4 exhibits a limit cycle and bursting be-
havior. Bursting behavior requires J > 0 and ǫ and γ small.
Section 7.2
q √ t/ǫ
5. 2y 1+y
1−y = 3e . Solution of reduced problem is y = 0(t = 0), y =
1/2(t > 0).
7. Equation (6.16) has a unique equilibrium (v, w) with 1/2 < v < 2/3, w <
0, and this equilibrium is unstable.
11. For dv/dt = −f (v) − w, an equilibrium v ∗ is LAS iff −f ′ (v ∗ ) < 0. But
this occurs precisely outside the interval [A, B]. We can find A and B
by setting f ′ (v) = 0, which yields
1h p i
A, B = (a + 1) ± a2 − a + 1 .
3
456 Answers to Selected Exercises
0.6
0.4
0.2
v
0.5 1.0
15. Fast: inner solution, boundary layer. Slow: outer solution, reduced prob-
lem, quasi-steady state.
Section 7.3
1. V → 0 if g < bK2 T0 ; V → g/K2 T0 − b if g > bK2 T0 .
2. (a) Solutions approach either the virus-free equilibrium (VFE) or pro-
gression to AIDS, depending on initial conditions (an unstable endemic
equilibrium provides a separatrix). (b) A unique endemic equilibrium
(EE) is globally asymptotically stable (GAS). (c) Two possible por-
traits: the VFE is GAS if the nullcline curves do not cross, but if they
do cross [twice], the VFE and an EE are both LAS (an unstable EE
again providing a separatrix). (d) Two possible portraits: progression to
AIDS if the nullcline curves do not cross (Figure 7.13), while if they do
cross [twice] there is also a LAS EE (with an unstable EE providing a
separatrix).
Answers to Selected Exercises 457
r
s2 a g g
3. (a) Tc = − (a − b), requires µ > k1 b
µ − k1 b k2 k2
s2 a b 2 g s2 a
(b) < <
µ − k1 b a2 k2 µ − k1 b
4. Behavior would change to progression to AIDS for g = 1200.
6. V ′ > VgV ′ ′ gV
+b − N I , so (V + N I) > V +b and thus increases without bound
in both systems. In the system with T , the growth of V pulls down
T ′ < 0 so eventually T → 0.
7. The virus-free equilibrium is asymptotically stable if g < bK2 s1 /µ.
Section 7.4
1. (a) The condition (7.40) provides an upper bound on the slope of
Λ(N ) which decreases as N increases: for any given N0 , Λ′ (N0 ) <
Λ(N0 )/N0 , the slope of the line segment connecting (0,0) with
(N, Λ(N )). Thus for any N1 > N0 , Λ(N1 ) < (Λ(N0 )/N0 )N1 , which
implies that the bound on Λ′ (N1 ) is even lower:
Section 7.5
1. y = K1 t + K2 , y = −t + 1.
3. y = K1 e3t + K2 e2t , y = −3e3t + 4e2t .
5. y = K1 e5t + K2 te5t , y = e5t − 6tE 5t .
√ √ √
y = K1 e−5t/2 cos√ 15t/2 + K2 e−5t/2 sin 15t/2, y = e−5t/2 cos 15/2 +
7. √
15/2e−5t/2 sin 15t/2.
9. y = K1 e2t cos 3t + K2 e2t sin 3t, y = e2t cos 3t − e2t sin 3t.
11. y = K1 et cos t + K2 et sin t, y = et cos t − 2et sin t.
√
13. Equilibria are (0, 0) and (0, ± −b) if b < 0.
15. Linearization at (0, 0) is u′ = v, v ′ = −v.
17. Linearization at (0, 0) is u′ = v, v ′ = −g ′ (0) − f (0)v.
Bibliography
459
460 Bibliography
H.W. Hethcote and J.A. Yorke, Gonorrhea Transmission Dynamics and Con-
trol. Lecture Notes in Biomathematics 56, Springer-Verlag, New York
(1984).
A. L. Hodgkin and A. F. Huxley, A quantitative description of membrane
current and its application to conduction and excitation in nerve, J.
Physiology, 117: 500–544 (1952).
F.C. Hoppensteadt, Singular perturbations on the infinite interval, Trans.
Amer. Math. Soc., 123: 521–535 (1966).
G. Evelyn Hutchinson, An introduction to population ecology, Yale University
Press, New Haven (1978).
Catherine Ryan Hyde, Pay it forward, Simon and Schuster, New York (1999).
International Dolphin Conservation Program Scientific Advisory Board, 7th
meeting, Updated estimates of NMIN and stock mortality limits, Doc-
ument SAB-07-05, La Jolla, California, 30 October 2009, available
at http://www.iattc.org/PDFFiles2/SAB-07-05-Nmin-and-Stock-
Mortality-Limits.pdf.
Holger W. Jannasch, Steady state and the chemostat in ecology, Limnology
and Oceanography, 19: 716–720 (1974).
Emmanuel Jolivet, Introduction aux modèles mathématiques en biologie,
INRA/Masson, Paris (1983).
D.S. Jones and B.D. Sleeman, Differential Equations and Mathematical Bi-
ology, CRC Press, Boca Raton, FL (2003).
Sukgeun Jung and Edward D. Houde, Recruitment and spawning-stock
biomass distribution of bay anchovy (Anchoa mitchilli) in Chesapeake
Bay, Fishery Bulletin, 102: 63–77 (2004).
J. Keener and J. Sneyd, Mathematical Physiology, Springer-Verlag, New York
(1998).
W. O. Kermack and A. G. McKendrick, A contribution to the mathematical
theory of epidemics, Proceedings of the Royal Society of London, 115:
700–721 (1927).
W.O. Kermack and A.G. McKendrick, Contributions to the mathematical
theory of epidemics, Part II, Proceedings of the Royal Society of London,
138: 55–83 (1932).
J.M. Kienzler, P.F. Dahm, W.A. Fuller, A.F. Ritter, Temperature-based es-
timation for the time of death in white-tailed deer, Biometrics, 40: 849–
854 (1984).
Bibliography 463
Curt D. Meine and George W. Archibald (Eds)., The cranes: Status survey
and conservation action plan. IUCN, Gland, Switzerland, and Cam-
bridge, U.K. 294pp. Northern Prairie Wildlife Research Center Online.
http://www.npwrc.usgs.gov/resource/birds/cranes/index.htm
(Version 02MAR98) (1996).
L. Michaelis and M.I. Menten, Die Kinetik der Invertinwirkung, Biochem.
Z., 49: 333–369 (1913).
R.S. Miller and D.B. Botkin, Endangered species: models and predictions,
American Scientist, 62: 172–181 (1974).
Jacques Monod, La technique de culture continue: théorie et applications,
Annales de l’Institut Pasteur, 79: 390–410 (1950).
J.S. Nagumo, S. Arimoto, and S. Yoshizawa, An active pulse transmission
line simulating nerve axon, Proc. Inst. Radio Engineers, 50: 2061–2071
(1962).
M.E.J. Newman, The structure and function of complex networks, SIAM
Review, 45: 167–256 (2003).
A.J. Nicholson, An outline of the dynamics of animal populations, Australian
Journal of Zoology, 3: 9–65 (1954).
A.J. Nicholson and V.A. Bailey, The balance of animal populations, Proc.
Zoological Society of London, 3: 551–598 (1935).
Aaron Novick and Leo Szilard, Description of the chemostat, Science, 112:
715–716 (1950).
M.A. Nowak and R.M. May, Virus dynamics: Mathematical principles of
immunology and virology, Princeton University Press, 2003.
R. Pearl, Introduction of medical biometry and statistics, Saunders, Philadel-
phia (1930).
Alan S. Perelson and P.W. Nelson, Mathematical analysis of HIV-1 dynamics
in vivo, SIAM Review 41: 3–44 (1999).
J.D. Reeve, D.J. Rhodes and P. Turchin, Scramble competition in the south-
ern pine beetle, Dendroctonus frontalis, Ecological Entomology, 23: 433–
443 (1998).
R.S. Rempel and C.K. Kaufman, Spatial modeling of harvest constraints on
wood supply versus wildlife habitat objectives, Environmental Manage-
ment, 32: 646–659 (2003). doi: 10.1007/s00267-003-0056-8.
Eric Renshaw, Modelling biological populations in space and time, Cambridge
Studies in Mathematical Biology 11, Cambridge University Press, Cam-
bridge (1995).
Bibliography 465
FOR
DYNAMICAL SYSTEMS
BIOLOGICAL MODELING
DYNAMICAL SYSTEMS
FOR BIOLOGICAL
MODELING
Dynamical Systems for Biological Modeling: An Introduction pro-
vides both biology and mathematics students with the understanding
and techniques necessary to undertake basic modeling of biological
systems. It achieves this through the development and analysis of
dynamical systems. AN INTRODUCTION
The approach emphasizes qualitative ideas rather than explicit com-
putations. Some technical details are necessary, but a qualitative ap-
proach emphasizing ideas is essential for understanding. The model-
ing approach helps students focus on essentials rather than extensive
mathematical details, which is helpful for students whose primary in-
terests are in sciences other than mathematics.
The book discusses a variety of biological modeling topics, including
population biology, epidemiology, immunology, intraspecies competi-
tion, harvesting, predator–prey systems, structured populations, and
more.
The authors also include examples of problems with solutions and
some exercises that follow the examples quite closely. In addition,
problems are included that go beyond the examples, both in math-
ematical analysis and in the development of mathematical models for
Brauer • Kribs Fred Brauer
biological problems, in order to encourage deeper understanding and
an eagerness to use mathematics in learning about biology. Christopher Kribs
C664X
w w w. c rc p r e s s . c o m