An Introduction to

Structured Population
Dynamics
This page intentionally left blank
J. M. Gushing
University of Arizona
Tucson, Arizona

An Introduction to
Structured Population
Dynamics

siam..
SOCIETY FOR INDUSTRIAL AND APPLIED MATHEMATICS

PHILADELPHIA
Copyright ©1998 by the Society for Industrial and Applied Mathematics.

10987654321

All rights reserved. Printed in the United States of America. No part of this book may be
reproduced, stored, or transmitted in any manner without the written permission of the
publisher. For information, write to the Society for Industrial and Applied Mathematics,
3600 University City Science Center, Philadelphia, PA 19104-2688.

Library of Congress Cataloging-in-Publication Data

Cushing, J. M. (Jim M.), 1942-

An introduction to structured population dynamics / J.M. Cushing.
p. cm. — (CBMS-NSF regional conference series in applied
mathematics ; 71)
Outgrowth of a series of lectures given at a conference held at
North Carolina University, Raleigh, during June of 1997.
Includes bibliographical references (p. ) and index.
ISBN 0-89871-417-6 (pbk.)
1. Population biology-Mathematical models. I. Title.
II. Series.
QH352.C87 1998
577.8'8' 98-19033

51oJTL is a registered trademark.

To Alina and Lara
This page intentionally left blank
 Contents

Preface ix

Chapter 1. Discrete Models 1

1.1 Matrix models 1
1.1.1 Notation and preliminaries 3
1.1.2 Linear models 4
1.1.3 Nonlinear models 12
1.2 Autonomous single species models 15
1.2.1 The extinction equilibrium 16
1.2.2 Matrix equations with parameters 19
1.2.3 Positive equilibria 21
1.2.4 Positive equilibrium destabilization 24
1.2.5 The net reproductive number 28
1.3 Some applications 33
1.4 A case study 46
1.4.1 Model parameterization and validation 47
1.4.2 Bifurcating beetles 54
1.4.3 Periodic habitats 58
1.5 Multispecies interactions 64
1.5.1 Some equilibrium theory 65
1.5.2 Applications 68

Chapter 2. Continuous Models 77

2.1 Age-structured models 78
2.2 Autonomous age-structured models 81
2.2.1 The extinction equilibrium 82
2.2.2 Positive equilibria 85
2.2.3 Hopf bifurcation 90
2.3 Some applications 91
2.4 Multispecies interactions 97
2.5 Other structured models 100
vii
viii CONTENTS

Chapter 3. Population Level Dynamics 103

3.1 Ergodicity and nonlinear models 103
3.1.1 Discrete matrix models 103
103
3.1.2 Continuous age-structured models 120
120
3.2 The linear chain trick 123
3.3 Hierarchical models 133
3.3.1 Continuous age-structured models 134
134
3.3.2 Discrete matrix models 139
139
3.4 Total population size in age-structured models 142
142

Appendix A. Stability Theory for Maps 147

A.I Linear maps 147
147
A.2 Linearization of maps 153

Appendix B. Bifurcation Theorems 161 161

B.I A global bifurcation theorem 161
161
B.2 Local parameterization 163

Appendix C. Miscellaneous Proofs 167

Bibliography 171

Index 191
191
 Preface

Interest in the dynamics of biological populations is quite old, its roots being
traceable to the dawn of civilization. Many illustrious names are associated
with early mathematical theories of population growth, e.g., Fibonnaci, Euler,
Halley, Malthus [382]. The flowering of mathematical ecology and population
dynamics occurred, however, during the first half of the twentieth century [245],
[374]. Many of the "classical" equations (and their many variants) resulting from
this development, such as the famous logistic equations, Volterra predator-prey
equations, and the Lotka-Volterra competition equations, have had a tremen-
dous influence on both theoretical and applied ecology and population dynam-
ics. They stimulated the formulation of, and gave theoretical support to, many
(if not most) of the fundamental tenets held today. These include exponen-
tial growth, carrying capacity, competitive exclusion, ecological niche, limiting
similarity, r-K selectors, and predator-prey oscillations.
In order to gain a better understanding of the dynamics of biological popu-
lations, theoretical biologists and applied mathematicians have, over the course
of the century, modified classical models and modeling methodologies in many
ways. All mathematical models make simplifying assumptions, of course, and
there is a relentless trade-off between biological accuracy and mathematical
tractability. One way to view many of the simplifying assumptions made in
population models is with regard to various uniformities and homogeneities
that are either explicitly or implicitly postulated. For example, two common
simplifications concern homogeneities in space and time. In such models there
is no attempt to account for the spatial extent or movement of individuals or
populations, and environments are assumed constant in time. This is the case
in the classical models mentioned above. There exists now, however, a rather
large and growing body of literature on spatial diffusion models for biological
populations. There is also a growing body of literature on models that include
either deterministic or stochastic environmental fluctuations.
Another important modeling assumption that is commonly made concerns
the homogeneity of individuals within a population. Mathematical models of-
ten involve equations for total population "size" (total number or density of
individuals or their total biomass, dry weight, etc.) and in effect treat all indi-

ix
x PREFACE

viduals within the population as identical. This is true of the classical models
mentioned above and, indeed, of the majority of models studied today. However,
individuals in biological populations differ with regard to their physiological and
behavioral characteristics and therefore in the way they interact with their envi-
ronment. As a result, vital processes such as birth, death, growth, metabolism,
resource consumption, etc., vary among individuals. Birth rates of younger in-
dividuals are generally quite different from those of older individuals, mortality
rates of larger individuals are usually different from those of smaller individu-
als, and so on. These differences can be considerable, with variances sometimes
being larger within a population than between different populations.
The vital rates of individuals ultimately determine the dynamics of the entire
population and how those dynamics are affected by the physical and biological
environment. Accurate models of population level dynamics therefore require a
connection to individual level vital rates. One such connection is provided by
so-called "structured" population models.
The structured models considered in this monograph describe the distribution
of individuals throughout different classes or categories. The definition of these
classes is based upon individual differences that are important with regard to
individual vital rates. For example, the categorization of individuals can be
based upon chronological age, a measure of body size, life cycle stages, gender or
genetic differences, biochemical makeup, spatial location, behavioral activities,
etc. A structured model describes how individuals move in time among the
defined classes. The model thereby describes the dynamics of the population
class distribution and as a result the dynamics of the population as a whole.
To cite just a few examples, structured models are required for the study of
questions dealing with the effects of maturation and gestation delays; intraclass
competition (between, say, small and large individuals or between juveniles and
adults); intraclass predation (cannibalism); juvenile bottlenecks (in which the
individuals are subjected to heavy competition or predation before reaching
reproductive maturity); selective predation on prey of certain ages or sizes;
parasitization on specific life cycle stages of hosts; the relationship between
body size and interspecific competitive success; mixed types of interactions (in
which, for example, two species compete during one life cycle stage, but do not
compete or even bear an entirely different relationship, such as a predator-prey
relationship, at a different life cycle stage). See [55], [154], [323], [408], [435] for
these and many other examples.
Structured models have many advantages. By making a link between the
individual level and population level, they can account for dynamical behavior
that unstructured models cannot. Environmental influences are very likely to
affect different individuals differently. Therefore a structured model can more
accurately describe and predict the importance that specific environmental fac-
tors have on the population's dynamics, as well as the consequences of changes
in these factors. For example, an individual's movement through the structuring
classes can cause delays in response to environmental changes, which can have a
profound effect on the dynamics of the population as a whole. Another advan-
tage of structured models is that they are more likely than unstructured models
 PREFACE xi

to involve parameters with clear biological interpretations that are amenable

to measurement and thereby provide a greater opportunity for connection with
data.
Many different types of mathematical equations have been used to formulate
structured population models. One broad distinction between types of models
is whether the variables are discrete or continuous. For example, model time
can be continuous or it can be a discrete sequence of census times; individual
body size might be a continuous variable or it might be classified by discrete
size intervals; and so on. At one extreme all structuring variables, all state
variables, and time are discrete; at the other extreme they are all continuous.
In Chapter 1 models of the former type are considered; models of the latter type
are considered in Chapter 2. Structured models can also be of mixed types.
For example, so-called compartmental models describe in continuous time the
dynamics of discrete state variable classes. Examples of models discrete in time
but continuous in the state variables can be found in [267], [268]. We will not
consider such mixed types in this monograph.
Both discrete and continuous structured models have a long tradition of use.
Both have their advantages and disadvantages. Discrete models, for example,
are usually easy to construct from the life cycle history of the population. Gen-
erally, discrete models avoid many technical difficulties that continuous models
entail (e.g., the difficulties surrounding partial differential equations concerning
well posedness of initial value problems, numerical simulations, rigorous justifi-
cation of linearization procedures, etc.). Indeed, discrete models with arbitrarily
general structuring offer no particular difficulties. For continuous models of such
generality, however, severe difficulties arise with regard to even the fundamen-
tal questions of existence and uniqueness of solutions. Due to such difficulties
a complete and rigorous theory of continuous models has been worked out only
for restricted types of structuring (e.g., age structure). Furthermore, by the
recursive nature of the equations involved, discrete models are extraordinarily
easy to simulate on computers. Also, stochastic versions of discrete models are
generally easier to construct and analyze.
On the other hand, models with discrete time cannot account for the dynam-
ics between its census times. Unless the structuring classes are approximately
discrete in the biological population, the dynamics of a model with discrete
classes might be sensitive to how the classes are defined and measured. Of
course, in principle discrete models can be constructed that approximate a con-
tinuous classification arbitrarily closely (e.g., by shrinking the length of size or
age classes), but the model will become large in size and a continuous model
might be more tractable. (Interestingly, a complete and rigorous study of dis-
crete models as approximations to continuous models, or vice versa, is yet to
be made; see [409].) For more discussion of discrete and continuous models see
[408, Chapter 1].
The focus in this monograph is on the asymptotic dynamics of deterministic
models. Except for a brief appearance in section 1.4, stochastic models are not
considered. A general modeling theory for structured population dynamics is
presented. A general treatment of equilibria and stability is given from the point
xii PREFACE

of view of bifurcation theory. Bifurcation theory is particularly appropriate in

theoretical population dynamics since one of the fundamental expectations of a
mathematical model is a description of the circumstances under which a pop-
ulation has certain kinds of asymptotic dynamics and how these dynamics are
predicted to change if perturbations occur. Also included in this monograph is
a selection of applications. These applications were chosen to illustrate both the
mathematical theories and a selection of biological problems that have received
attention in the literature.
Attention will be focused on structuring variables related to physiological
characteristics. Structured models could also be constructed using classes based
upon spatial location and/or inhomogeneities in the physical habitat. Even
though spatial structure is extremely important in population dynamics and
there is a great deal of literature on spatial diffusion, migration, patchiness,
etc., models that include explicit spatial structure are not considered.
This monograph is restricted almost exclusively to autonomous models. How-
ever, much of the theory and mathematical results presented here have been ex-
tended to periodically forced model equations, which are appropriate for popu-
lations subjected to periodic oscillations in their vital rates and/or environments
(e.g., due to seasonality). See section 1.4.3.
The topics covered reflect the particular interests of the author and are not
meant to be comprehensive, either mathematically or biologically. For example,
there is a large literature on modeling the dynamics of cell growth which is not
touched on. This monograph focuses exclusively on population dynamics and
ecological interactions. No topics are included from many related disciplines,
e.g., epidemiology, genetics, evolutionary biology, renewable resource manage-
ment, or bioeconomics.
Chapter 1 contains a treatment of discrete models in discrete time which
allow for very general structuring of a population. A methodology is presented
for studying basic equilibrium and stability questions for such models from the
perspective of bifurcation theory. Nonequilibrium dynamics are also covered,
insofar as they arise from equilibrium destabilization and local bifurcations.
A similar tact is taken in Chapter 2 for continuous models. However, only
age-structured (and some simpler size-structured) models are treated, since the
mathematical theory of continuous models with more general structuring is
difficult and is not as complete. In Chapter 3 there appear some special types of
structured models, both discrete and continuous, for which dynamical equations
at the population level can be uncoupled from those at the individual level. Some
details of local stability theory for (not necessarily invertible) maps are given in
Appendix A. Other mathematical details appear in Appendices B and C.
This monograph is an outgrowth of a series of lectures given at a National
Science Foundation Regional Conference arranged by the Conference Board of
Mathematical Sciences and held at North Carolina State University, Raleigh,
during June of 1997. I would like to thank John Franke and Abdul-Aziz Yakubu
for organizing the conference and inviting me to give these lectures. The tremen-
dous success of the conference was in large part due to their efforts and those of
their supporting staff. Also contributing to this success were the many partic-
 PREFACE xiii

ipants at the conference, who came from a wide variety of disciplines and who
provided stimulating discussions and contributing talks. Special thanks are also
due to Jim Yorke, whose provocative and insightful lectures were certainly a
highlight of the meeting.
I would like to make a special acknowledgment to my colleagues R. F. Costan-
tino, Brian Dennis, and R. A. Desharnais who have shown me how exciting and
fruitful interdisciplinary collaborations can be.
I am very grateful to the National Science Foundation for its generous support
of my research over the years.
With the usual caveat that responsibility for all errors is mine, I thank Shan-
delle Henson, William Mueller, and Joseph Watkins for their aid in proofreading
this manuscript.

J. M. Gushing
University of Arizona
This page intentionally left blank
 CHAPTER 1
Discrete Models

This chapter deals with dynamical models for structured populations in which
both time and the structuring variables are discrete. These models involve
systems of difference equations (or maps) of a type called matrix population
models [55]. Matrix models are introduced in section 1.1. The asymptotic
dynamics of autonomous matrix models are studied in section 1.2 from the
point of view of bifurcation theory. Several applications are given in sections
1.3 and 1.4. These sections deal only with a single population. Matrix models
for the interaction of several structured species arc considered in section 1.5.

1.1 Matrix models

Suppose that the individuals of a population are categorized into a finite number
of classes (e.g., by chronological age or some measure of body size). Let Xi(t),
for 2 = 1 , 2 , . . . , TO, denote the number or density1 of individuals in the ith
class at time t = 0. 1, 2, . . . . Let
the fraction of j-class individuals expected to
survive and move to class i per unit of time.
Then at time t + 1 the density of individuals in class i who were alive at time t
is expected to be

Let
f the expected number of (surviving) i-cla
ffspring per j-class individual per unit
' Then at time t + 1 the number of i-class newborns is expected to be

Biomass. dry weight, or some other measure of population abundance of the individuals1
in the structuring classes could also be used.

1
2 CHAPTER 1

If only birth and death processes are allowed, we have

Using matrix notation these equations can be written in the compact form

where

is the class distribution vector at time t and the matrix

is the sum of the transition matrix

and the fertility matrix

The quantities /»j and tij are built from "submodels" based on class specific
hypotheses about these vital quantities. These submodels would take into con-
sideration relevant class specific rates of mortality, fertility, resource availability
and consumption, metabolism, body growth, etc. These rates, and hence t^ and
fij, might be related to population crowding or so-called density effects (e.g.,
due to competition), in which case they become functions of one or more of the
class densities Xi. Recursion equations of the form (1.1) are called matrix equa-
tions. (Rates not proportional to class densities are not included in this model.
For example, immigration and emigration rates might be of such a type.)
The matrix P is called the projection matrix.2 If P is constant, then the
matrix equation (1.1) is linear and autonomous. If P = P(t) depends explicitly
on time t, then (1.1) is linear and nonautonomous. If P = P(t, x(t)) depends on
x(t), then (1.1) is nonlinear. In any case, given an initial class distribution x(Q),
the recursion formula (1.1) defines a unique sequence x ( t ) , t = 0, 1, 2,... , called
a solution (or more precisely a forward solution) of (l.l). 3 Since all entries in
P are nonnegative it follows that Xi(0) > 0 implies that Xi(t) > 0 for all t = 0,
1,2,....
2
P is not a projection matrix in the geometric or functional analysis sense.
3
In general, P is not invertible and hence unique "backward" solutions are not denned.
See A.I.
 DISCRETE MODELS 3

1.1.1 Notation and preliminaries. The set of real numbers and the set of
nonnegative real numbers are denoted by Rl and R+, respectively. Let Rm =
R1 x • • - x R1 and 7?™ = R+ x • • • x R*_ denote m-fold Cartesian products. Forl
x = [xi] e Rm we define the vector norm

The transpose of x will be denoted by XT . Then the usual inner product is xry.
The boundary of R™ is denoted by dR™. For a matrix P — [pij] we use the
onerator norm

Then \Px\ < \\P\\ x . Let /(Qi.Qa) denote the set of integers i satisfying q\ <
i < <?2 (li can be -co and/or 2 can be +00). Let I [11,12) = 1(11,12) U {171},q
etc.
By a positive (nonnegative) vector x > 0 (x > 0) or matrix P > 0 (P > 0)
we mean all entries are positive (nonnegative). Eigenvalues r (real or complex)
arid right and left eigenvectors v and w (real or complex) of a matrix M satisfy
Mv = rv, v ^ 0. and wrM = rwT, w ^ 0. respectively. A positive, strictly
dominant eigenvalue r\ satisfies r\ > |r for all other eigenvalues r. The al-
gebraic multiplicity of an eigenvalue r is the multiplicity of r as a root of the
characteristic polynomial det (rl — M). An eigenvalue is algebraically simple if
its algebraic multiplicity is equal to 1. In this case, all right and left eigenvectors
have the form cv and cw, c 6 R1. The characteristic values of a matrix M are
the reciprocals of the nonzero real eigenvalues of M.
Of particular interest in population models are nounegative matrices P > 0.
A remarkable theorem of Perron states that a positive matrix necessarily has a
positive, (algebraically) simple, strictly dominant eigenvalue and to this eigen-
value there correspond positive right and left eigenvectors [343]. Matrices of
interest in population dynamics are not always positive, however. A famous
theorem of Froberiius adapts Perron's theorem to nonnegative matrices under
the assumption of irreducibility [176], [179]. A matrix is reducible if a permu-
tation of its rows and corresponding columns (which amounts to the reordering
of the classes in a structured model) results in a block triangular matrix, i.e.. a
matrix in the form

If no such permutation exists, the matrix is irreducible. Equivalently, a ma-

trix is irreducible if and only if its associated graph is strongly connected. The
Frobenius theorem states that a nonnegative, irreducible matrix has a posi-
tive, (algebraically) simple eigenvalue whose magnitude is exceeded by no other
eigenvalue and to this eigenvalue there correspond positive right and left eigen-
vectors. If, in addition, this maximal eigenvalue is strictly dominant, then the
4 CHAPTER 1

matrix is called primitive. It turns out that under these conditions there cannot
exist two independent nonnegative eigenvectors [179, p. 63]. A necessary and
sufficient condition that a nonnegative matrix P > 0 be irreducible and prim-
itive is that there exists a positive integer j such that PJ > 0. See [179] for
proofs of these facts.
THEOREM 1.1.1. (Perron/Frobenius). If a nonnegative matrix is irreducible
and primitive, then it has a positive, (algebraically) simple, strictly dominant
eigenvalue and to this eigenvalue there correspond positive right and left eigen-
vectors. Moreover, there cannot exist two independent nonnegative right (or
left) eigenvectors.

1.1.2 Linear models. Consider the linear autonomous matrix equation (1.1)
when P > 0 is a constant projection matrix. Matrix equations of this type were
first used to describe the dynamics of age-structured populations [281], [282],
[288]. In these so-called Leslie matrix models, a population is divided into age
(since birth) categories, all of which have the same length as the discrete census
time. For this reason the transition matrix T has the form

Since newborns necessarily lie in the first age class, the fertility matrix has the
form

and the projection matrix

is a so-called Leslie matrix. If there are fe juvenile classes, then u = fa =f

• • • = lk = 0. If /im > 0, then P is irreducible. When is P primitive? Letf
/i,mi, /i,m 2 ,-> /i,mj, /im, mj ^ m, be the nonzero fertilities appearing in the
first row of F listed in order. The Leslie matrix P is primitive if and only if
the greatest common divisor of the integers m — mj, mj — mj_i,..., m^ - m\ is
 DISCRETE MODELS 5

equal to 1 [179], [245]. In particular, P is primitive if there are two consecutive

fertile age categories.
A mathematical generalization of the Leslie matrix is given by the transition
matrix

This type of transition matrix has been utilized for populations whose individ-
uals are categorized by (increasing) size classes. The diagonal entry tu is the
fraction of individuals in size class i who survive and remain in class i after one
time unit, and t^+i^ is the fraction that survives and moves to the next largest
size category i + l. No individual can shrink in size or grow more than one class
in one unit of time. If all newborns lie in the smallest size class, the resulting
projection matrix

is called an Usher matrix [410], [411], [412] (or the standard size-classified matrix
[55]). An Usher matrix is irreducible if all i^i-i > 0 and f\m > 0. It is also
primitive if two successive size classes are fertile.
Most projection matrices used in applications are irreducible and primitive.
An example of an irreducible Leslie matrix that is not primitive is the juve-
nile/adult model

If adults are allowed to live more than one unit of time, the resulting modified
Leslie matrix

is irreducible and primitive.

Consider the general linear autonomous matrix equation
6 CHAPTER 1

and its solution

If P is nonnegative, irreducible, and primitive, then by Theorem 1.1.1 there is a

simple, positive, and strictly dominant eigenvalue r and positive right and left
eigenvectors v > 0 and w > 0. For simplicity, suppose P is diagonalizable and
let i>i, «2,... , m be a basis of right eigenvectors and wi, W2, • • • , wm a basisv
of left eigenvectors, where v\ — v and Wi = w. Write,

Since the inner products wTVi = 0 for i 7^ 1, it follows that c = wTx(0)/wrv > 0.
If x(0) ^ 0, then c > 0. Let 7-4, i € /[2,m], denote the remaining m — 1
eigenvalues of P. Let

denote the total population size. From

and r > \Ti for alH € J[2, m], we find that

The unit eigenvector v/ \v is called the stable distribution, and the limit (1.7)
is called the (strong) ergodic property [55], [64], [65], [66], [67], [245]. It remains
true for nondiagonalizable projection matrices P as well (see [245] for a proof).
The limit (1.7) describes the asymptotic behavior of the normalized distribu-
tion x(t)/p(t], regardless of the asymptotic dynamics of the class distribution
x(t) itself or that of the total population size p(t}. The equation

for the dynamics of p(t) can be obtained directly from p(t + 1) = \Px(t)\. This
is a scalar, nonautonomous difference equation for p(t) that is coupled with the
equation (1.6) for x(t). However, by (1.7) the coefficient of p(t) in (1.8) satisfies
 DISCRETE MODELS 7

This suggests that the asymptotic dynamics of p(t) are related to those of the
so-called lirnitinq equation

a particularly simple scalar recursion formula. The solution q(t) — r*g(0) tends,
of course, to 0 if r < 1 and grows exponentially if r > 1. Is this true of the
total population size p(t) satisfying (1.8)? The answer is "yes,'' a fact proved in
Appendix C. We summarize these results in the following theorem. (For further
theorems of this type see [55], [64], [65], [66], [67].)
THEOREM 1.1.2. (The Fundamental Theorem of Demography). Suppose that
the nonnegative matrix P > 0 is irreducible and primitive. Let r be the strictly
dominant eigenvalue of P and v > 0 be an associated eigenvector. Let x(i) be
the solution of the linear matrix equation x(t + 1) = Px(t), t e 7[0, +00), with
an initial state satisfying 0 < .x(0) 7^ 0, and letp(i) = x(t)\. Then
(a) (1.7) holds and
(b) \imt^+00p(t) = 0 if r < I and limt_> +00 7>(£) — +°° */r > 1-
The limiting equation (1.9) is an example of a dynamical equation for a pop-
ulation level statistic derived from a structured model. The individual level
parameters in the projection matrix are encapsulated in its dominant eigen-
value r and can therefore be related to population level dynamics.
The dominant eigenvalue r is the (inherent) growth rate of the population.
The magnitude of r determines whether a population with positive initial size
p(0) > 0 goes extinct (r < 1) or grows exponentially without bound (r >
1). Thus, a bifurcation occurs as r passes through the critical value 1. When
r = 1 (and for no other value of r), there exist equilibrium class distributions,
namely, any scalar multiple of the eigenvector v. The extinction equilibrium
x(t) = 0 (or p(t) = 0) always exists for the matrix equation x(t + 1) = Px(t).
If equilibria are viewed as functions of r, we see that there is an intersection
(or bifurcation) of two equilibrium branches, the extinction equilibria and the
nonzero (eigenvector) equilibria, which occurs at the bifurcation point (r, x) —
(1,0) (or (r,p) — (1,0)). This bifurcation is vertical because the spectrum for
the noncxtinction equilibria consists of the single point r = 1. See Fig. 1.1.
Another important quantity is the inherent net reproductive number n (some-
times called the net reproductive value or rate [184]). This quantity is defined
to be the expected number of offspring per individual per lifetime. For a Leslie
model a formula for n is obtained by summing the products of the expected
number of offspring fn from each age class and the probability £21^32 • • • ti,i-\
of reaching that age class. Thus,

where for notational convenience tw is defined to be 1. For a general projection

matrix P — T + F the inherent net reproductive number n can be defined as
follows.
8 CHAPTER 1

FlG. 1.1. A plot of the. inherent growth rate r against the total population size p = \x\ of
the equilibria x of the linear matrix equation (1.6) shows two branches intersecting at (r,p) =
(1,0). Namely, the branch of trivial equilibrium (r,0) intersects the branch of nontrivial
equilibria (l,p), p G Rl. The extinction equilibrium p = 0 is stable for r < 1 and is unstable
for r > 1.

DEFINITION 1. LetT and F be the transition and fertility matrices (1.2) and
(1.3). Ifl — T is invertible and F(I~T)~l has a positive, (algebraically) simple,
strictly dominant eigenvalue n and a nonnegative eigenvector u > 0, then n is
called the inherent net reproductive number for the projection matrix P = T+F.
Necessary for the invertibility of / — T is that at least one of the column sums
SS=i t%j °f T ig strictly less than 1. A sufficient condition is that all column sums
be strictly less than 1, i.e., X^i tij < 1 f°r eacn 3 e [l>ml- This biologically
reasonable condition means that over each time interval there is some loss to
every class of individuals (e.g., due to deaths). Since (/ — T)~l — X^o-^ > 0,
we see that F(I — T)"1 is a nonnegative matrix. Sufficient, but not necessary,
for the definition of n is that F(I — T)~l be irreducible and primitive.
Definition 1 is a mathematical one. What is the biological interpretation of
n? By Definition 1, n is given by [179]

From this formula we can obtain a biological interpretation for n as follows.

The i,j entry in (/ - T)"1 = / + T + T2 4- • • • is the expected amount of time
that an individual starting in class j will spend in class i over the course of its
lifetime. The i,j entry in the matrix F(I - T)"1 is the expected number of i
class offspring that an individual born into class j will produce over the course
of its lifetime. Thus, the total expected offspring of an individual born in class
j over its lifetime is the sum of elements in the jth column of F(I - T)~l.
Consider first the simplest and most common case in matrix models in which
 DISCRETE MODELS 9 9

there is only one newborn class. Then F has only one nonzero row which we
take, without loss of generality, to be the first row. and consequently only the
first row of F(I-T)~l is nonzero. This means, in this case, that F(I-T)~l has
0 as an m-1 repeated eigenvalue and the dominant eigenvalue (the inherent net
reproductive number, by definition) is the first row, first column entry. Since
this entry is the only nonzero entry in the first column, it is equal to the sum of
entries in the first column. It follows that for the case of a single newborn class
the inherent net reproductive number n is the expected number of offspring per
newborn over the course of its lifetime.
More generally, suppose that there are j > 1 newborn classes, which without
loss of generality we list first so that the last m — j rows of F consist of zeros
only. If we denote the ith row of F by Jl and the jth column of (/ — T}~1 by
Cj, i.e.,

then

Thus. 0 is an eigenvalue of multiplicity at least m — j. The dominant eigenvalue

of F(T — T)"1 is the dominant eigenvalue of the subrnatrix S-j from the upper
left-hand corner; thus,

For a class distribution of individuals x 6 R'" the expected distribution of

offspring over the course of their lifetimes is F(I — T)~1x. If we consider the
expected distribution of offspring from a group of newborn members of the
population, then x has nonzero entries in only its first j entries so that

where y e R+ is the distribution of those newborns. Then the distribution ofJ

expected offspring from this group is
10 CHAPTER 1

and the total expected offspring is |F(J —T) lx\ = \Sjy\. Prom (1.12) we
see that n is the maximum expected per capita lifetime number of offspring from
newborn individuals, where the maximum is taken over all possible newborn class
distributions. This maximum is attained by class distributions proportional to
the eigenvector u associated with n as the dominant eigenvalue of F(I — T)"1.
Note that any eigenvalue of F(I—T)" 1 is also an eigenvalue of (7—T)~ 1 F and
vice versa, since both matrices have the same characteristic polynomial. Thus,
the inherent net reproductive number n is also a strictly dominant positive
eigenvalue of the matrix (7 — T)~1F. It is easy to show that (/ — T)~lu > 0 is
a right eigenvector of (I — T)~1F associated with n. The following theorem is
proved in [115, Theorem 3 and Corollary 7].
THEOREM 1.1.3. Consider a nonnegative projection matrix 0 < P = T + F,
where the transition and fertility matrices T and F satisfy (1.2)-(1.3). Assume
that P has a positive, (algebraically) simple, strictly dominant eigenvalue r with
a positive right eigenvector v > 0. Assume further that T and F satisfy the
requirements of Definition 1 with (I -T)~~lu> Q.
Then r < I if and only if n < I and r > 1 if and only if n > 1. (Thus, r = 1
if and only if n — l.)
Moreover, n < 1 implies that n < r < 1 and n > 1 implies that 1 < r < n.
Thus, a population grows exponentially if its inherent net reproductive num-
ber is n > 1 and dies exponentially if n < 1 (not an unexpected result given the
biological interpretation of n).
Whereas formulas for the population growth rate r in terms of the entries in
the projection matrix P are not in general available, such formulas for n are
available for broad classes of projection matrices [115]. For example, in the
common case of models with a single newborn class we have from (1.12) that
the inherent net reproductive number is given by the inner product of the first
(fertility) row of F with the first column of (/ — T)~l, namely,

A formula for n is thus available when a formula for (I — T) 1 is available.

As an example, consider an irreducible and primitive Usher matrix (1.5) with
0 < tjj < 1. Then
 DISCRETE MODELS 11

The matrix

has dominant eigenvalue

(where for notational convenience we have defined £10 = 1) with a unit eigen-
vector

Also.

Thus, Theorem 1.1.3 applies.

As a second example, consider an Usher-type model with two classes of new-
borns (taken without loss of generality to be classes i = 1 and 2) in which only
the largest class is fertile. For this case the fertility matrix is

and (I — T) is again given by (1.14). From (1.11) we obtain the formula

12 CHAPTER 1

for the inherent net reproductive number. Also,

and

so that Theorem 1.1.3 applies.

1.1.3 Nonlinear models. Linear matrix models (1.1) with constant projec-
tion matrices P imply exponential dynamics. In the case of exponential growth,
linear models therefore cannot describe the long term (asymptotic) dynamics
of populations. It follows that models suitable for asymptotic dynamics and
population growth must involve projection matrices P whose entries (namely,
the fertility arid transition rates appearing in F and T] are not constant in time.
If P depends implicitly on time through a dependence on class densities (called
"density effects"), then the matrix model (1.1) is nonlinear. A projection matrix
may also depend explicitly on time t (deterministically and/or stochastically).
One way that nonlinear projection matrices arise is through fractional de-
creases in one or more entries in the transition matrix T and/or the fertility
matrix F that are dependent upon class densities. For example, suppose the
transition probability tij of a j class individual to class i is adversely affected
in some way by an encounter with (or simply the presence of) a k class indi-
vidual. If the probability of such an encounter is taken to be (approximately)
proportional to the length of time involved, then during an interval At of time
the probability of no encounter occurring between a j class individual and a
k class individual is approximately 1 — cjjj.fcAi, where w^t is the constant of
proportionality. If there are Xk(t) individuals of class k present at time t and
the encounters are independent, then the probability of no encounter during an
interval At of time is approximately (1 — w^^Ai) 1 *. During one full unit of
time the probability of no encounter is (assuming independence)
 DISCRETE MODELS 13

Then

Letting A.t —•> 0 we obtain

If we approximate ft Xk(s)ds « Xk(t), then the probability of no encoun-

ters with A; class individuals during one unit of time is approximately equal to
exp(—u>ij,fc£fc(£))- Finally, if ir^ is the probability that a j class individual sur-
vives one unit of time and moves to class i independent of the presence of k
class individuals, then

is the i,j entry in the transition matrix T at time t. If other classes also affect
tij (independently), then additional exponential factors appear and one obtains

where

is now a weighted total population size.

Exponential nonlinearities of the form exp(—qp, ? ) are often referred to as
Richer-type nonlinearities. (See [294], [391] for further treatments of exponential
types of nonlinearities in age-structured models.) Other nonlinear expressions
also appear frequently in applications. These may be derived from modeling
assumptions concerning the class specific interactions between individuals or
they may be ad hoc nonlinearities with desired qualitative features. Exam-
ples include the Bc.verton-Holt nonlinearity (1 + cpjj)" 1 or, more generally,
(1 + cpf,j}~1, and the modified exponential e x p ( — c p f ^ ) , a > 0. In general, if
population density effects are deleterious, then appropriate entries in T and
F are multiplied by fractions which are decreasing functions of weighted total
population sizes. Thus, tl} = ^ijfij(pij) and/orf2J— bijU'-^jipij)- where (pt].
ibi:i <G Cl(R\, [0,1]) are decreasing functions of their arguments.
Examples of nonlinear projection matrices with various types of nonlinearities
14 CHAPTER 1

taken from the literature are

with p = x\, and

The matrix (1.16) was introduced by Leslie [282]. It assumes that all transitions
and fertilities are decreased by the same Beverton-Holt-type fraction (1 + cp)~
as a function of total population size. In the matrix (1.17) only fecundity and/or
newborn survival rates are affected by total population size p. This matrix is
used in [283] with ip(p) = exp (—cp) and with other types of nonlinearities
in [160]. The matrix with two juvenile stages (1.19) was introduced in [132];
it utilizes two exponential nonlinearities involving two different weighted total
population sizes. The matrix (1.18) is from a size-structured model with a
juvenile stage and two adult classes [112]. In all of these examples, increased
population density is deleterious in the sense that survival and/or fertility rates
decrease with increased density.
Models in which increased population density is not deleterious, but is in fact
advantageous (an effect called "strict depensation" or the "Allee effect"), are of
biological interest [5], [6], [130]. Nonlinear factors that can be used to model
this effect include
 DISCRETE MODELS 15

An example is given by the projection matrix

from a model for spotted owl dynamics [209], [272], [275] in which the transition
rate £21 has an Allee effect.
For other examples of matrix equation applications to population dynamics
see [55], [71], [72], [73], [93], [99], [98], [101], [110], [112], [116], [151], [152], [153],
[154], [160], [178], [238], [283], [294], [299], [335], [342], [376], [386], [391], [408],
[410], [413], [415], [448] (and references cited therein).
Multispecies interactions among several species can be modeled by coupled
systems of matrix models in which the projection matrices of some species are
functions of the densities of other species. For example, the system

denotes a two species interaction in which the entries of the transition and/or
fertility matrices of each species are functions of the class densities of the other
species. Predator-prey, competition, and mutualistic interactions can be mod-
eled in this way. So can mixed-type interactions in which some stages might
compete while others have a predator prey relationship [435]. The vectors x ( t )
and y ( t ) , and hence the projection matrices PI and P%, need not be of the same
size; i.e., the number of structure classes (or their type) need not be the same
for both species.
Nonautonomous matrix equations

model situations in which transition or fertility rates depend explicitly on time

in some manner. For example, one or more of these rates might vary peri-
odically according to some environmental periodicity (e.g., seasonal or daily
fluctuations), or stochastically [74], [90], [107], [227], [232], [340].

1.2 Autonomous single species models

Consider the general nonlinear autonomous matrix equation

We will always assume that P(x) is nonnegative for nonnegative x and is con-
tinuously differentiable in x. Specifically,
16 CHAPTER 1

where fJ is an open set in Rm that contains the (closed ) nonnegative cone R™.
Clearly, a unique (forward) solution x(i) is denned for each initial condition
x(0) e ft. Moreover, x(0) > 0 implies that x(t) > 0 for all t e /[O,+00); i.e.,
R™ is forward invariant. Thus, matrix models have no mathematical difficul-
ties associated with the existence and uniqueness of solutions of initial value
problems, nor with the positivity of solutions.
The study of the asymptotic dynamics associated with (1.20) naturally begins
with a study of equilibrium solutions x ( t ) = x € -R™ (i.e., time independent
solutions). The equilibrium equation associated with (1.20) is the nonlinear
algebraic system x = P(x)x. We are interested only in nonnegative solutions
of this equation. Clearly, x = 0 is such a solution and it will be referred to
as the extinction equilibrium. With regard to nonzero, nonnegative equilibria
0 7^ x > 0, there are two basic problems: existence and stability.
An equilibrium x £ R™ of equation (1.20) is called (locally) stable if for
each £ > 0 there exists a 6 = 6(s) > 0 such that |x(0) — x\ < 6 implies that
\x(t) — x\ < e for all t € /[O, +00). If x is not stable, then it is called unstable.
The equilibrium x is an attractor if there exists a 6 > 0 such that |x(0) — x < 6
implies that lim t _ +00 \x(t) - x\ — 0. If x is a stable attractor then it is called
(locally) asymptotically stable.
The Jacobian of P(x)x of the right-hand side of equation (1.20) is important
with regard to the stability properties of an equilibrium x. This Jacobian is
given by the formula4

A matrix is hyperbolic if all its eigenvalues £ satisfy \C\ ^ 1. An equilibrium

x is called hyperbolic if the Jacobian J(x) is hyperbolic. The (local) stability
properties of a hyperbolic equilibrium can be determined from the eigenvalues of
J(z). While this fact is well known for (local) diffeomorphisms [156], [193], [436]
(i.e., if £ = 0 is not an eigenvalue of J ( x ) ) , the supporting theorems for a local
stability analysis without this restriction are not. Therefore these theorems and
their proofs are given in Appendix A.2. Among other things these theorems
imply that an equilibrium x is (locally) asymptotically stable if all eigenvalues
of the Jacobian J(x) satisfy |C| < 1 and is unstable if at least one eigenvalue
satisfies |£| > 1. First we turn our attention to the extinction equilibrium x = 0.

1.2.1 The extinction equilibrium. For the extinction equilibrium x = 0

the Jacobian J(0) equals the matrix P(0). which we call the inherent projection
matrix. The inherent projection matrix is the projection matrix when all density
effects are ignored; therefore it governs the dynamics at low population levels.
By assumption (1.21), the inherent projection matrix P(0) > 0 is nonnegative.
Assume that P(0) is also irreducible and primitive. Then by Theorem 1.1.1 there
is a positive strictly dominant simple eigenvalue r > 0 (with positive left and

*[(diP)x] is the matrix whose ith column is (diP(x))x.

 DISCRETE MODELS 17

right eigenvectors w > 0 and v > 0). If r < 1, then the extinction equilibrium
x — 0 is (locally asymptotically) stable. This means, of course, that populations
with sufficiently small initial total population sizes p(Q) = z(0)| will go extinct.
It may not be true, however, that all initial conditions lead to extinction.
Under additional conditions a stronger stability statement can be made when
r < I . The most common assumption in population models is that density ef-
fects are deleterious at all density levels. This means that the entries t^ = t,j(x)
in the transition matrix T are decreasing (or at least nonincreasing) functions
of the components of x and in particular 0 < x implies that 0 < T(x) < T(0);
the same is commonly true for the fertility matrix F = F(x).
Under the assumption that the projection matrix P(x) satisfies

it follows that 0 < x < y implies that 0 < P(x)x < P(0)x < P(0)y. Since
z(0) € R% implies that x(t + 1) e R% for all t € /[O, +00), we find that

If we let y(t) denote the solution of the initial value problem

a straightforward induction shows that 0 < x(t) < y ( t ) , t G /[(), +00). If r < 1,
then \y(t)\ tends exponentially to 0 as t —> +00 (cf. Theorem 1.1.2) and conse-
quently so does p(t) = x ( t ) \ . Thus, (1.22) and r < I imply global extinction.
(This result can be found in [8].)
If r > 1, then the extinction equilibrium is unstable. From the definition of
instability it does not necessarily follow that no population will go extinct. In
fact, if x — 0 is hyperbolic, we know from Theorem A.2.2 (see Appendix A.2)
that there may exist a stable manifold of initial conditions near x = 0 whose
solutions tend to 0 as t —> +oc. However, this stable manifold is tangent at
x = 0 to the stable manifold of the linearization. Since the Jacobian is P(0).
the stable manifold of the linearization cannot intersect the cone R'" (except
at x = 0). Therefore the stable manifold cannot locally intersect the cone H"'
(except at x = 0).
If. in addition to r > 1, all solutions with x(0) > 0 are bounded, a stronger
statement can often be made. An assumption that usually implies that solu-
tions of population models are bounded is that both transition probabilities
and fertilities decrease with increased population densities. In fact, a stronger
condition usually holds for population models, namely,
there exists a constant c > 0 so that for
(1.23) each z(0) e R'^ there is an integer t' > 0
such that p(i] < c for all t £ I [ t ' . +oc).

With this property (1.20) is called point dissipative [211]. Furthermore, the map
M : -R™ —» R™ defined by M(x] == A(x)xis continuous on the metric space
18 CHAPTER 1

R™, and therefore maps bounded sets to bounded (hence precompact) sets. In
the jargon of [211], M is completely continuous. By Lemma 2.3.1 of [211] M is
"asymptotically smooth," and it follows from Theorem 2.4.6 of [211] that there
exists a global (connected) attractor in .R™.5 Using this fact, it can be further
proved that no initial total population size will lead to extinction in the sense
of the following definition.
DEFINITION 2. The equation (1.20) is uniformly persistent with respect to
x = 0 if there exists a constant r\ > 0 such that x(0) £ B?+/ {0} implies that
lim inf^+oo p(t) > r/.
The proof of the following theorem appears in Appendix C.
THEOREM 1.2.1. Consider the matrix equation (1.20)-(1.21). Assume that
the inherent projection matrix P(0) is irreducible, primitive, and hyperbolic,
and let r > 0 be its strictly dominant eigenvalue.
(a) // r < 1, then x = 0 is (locally asymptotically) stable. If, in addition,
(1.22) holds, then lim t _ +00 p(<) = 0 for all x(0) e R%.
(b) If r > 1, then x = 0 is unstable. If, in addition, (1.20) is point dissi-
pative, i.e., (1.23) holds, then (1.20) is uniformly persistent with respect to the
extinction equilibrium x = 0.
For a "weak persistence" result see [8].
The following theorem gives some conditions under which (1.20) is point
dissipative.
THEOREM 1.2.2. Consider the matrix equation (1-20) with projection matrix
P(x) — (tij (x)} + [fij (x)}. Suppose that there exist constants 6 > 0 and / > 0
such that for all x € R™ and i,j 6 /[I, m]

and

Then (1.23) holds; i.e., (1.20) is point dissipative.

Proof. Since \T(x)x < 6 x|, we have for each t € /[O, +00)

which by induction implies that for t € /[I, +00)

5
A global attractor is an invariant set A C R^ (i.e., a set for which M(A} — A) that
attracts each bounded set of R*? and is maximal (i.e., every compact invariant set of M
belongs to -4).
 DISCRETE MODELS 19

Since 6 < I , there exists an integer t' > 1 (depending on z(0)) such that

which implies (1.23) with c = If (I - 8 ) ~ 1 .

The condition (1.24) implies that at each time step, and for any class dis-
tribution x, there is loss of individuals during the transition from any class to
any other class (due to, say, mortality). Condition (1.25) implies that there is
an upper bound to fertility. For every i,j £ 7[l,m] it requires fij(x)xj to be
bounded for x <E /?"*, i.e., that there be intraclass density effects on fertility.
By Theorem 1.2.1 the extinction equilibrium x = 0 of (1.20) loses stability as
r increases through the critical value 1. Recall for linear matrix equations that
there exists an unbounded continuum of nonextinction equilibria that "bifur-
cates" at the point (r.x) — (1,0), but that the spectrum for these nonextinction
equilibria consists of the single point r — 1 (Fig. 1.1). The bifurcation of nonex-
tinction equilibria from this same point for the nonlinear matrix equation (1.20)
will be studied in section 1.2.3.
From Theorem 1.1.3 it follows that r can be replaced by the inherent ne repro-
ductive number n in the statement of Theorem 1.2.1, provided (/ — T}~lu> 0.
Theorem 1.2.1 can be applied to all of the examples (1.16), (1.17), (1.19), and
(1.18) above. In each case, r < 1 (or n < 1) implies global extinction and r > 1
(n > 1) implies uniform persistence with respect to x = 0.

1.2.2 Matrix equations with parameters. Theorem 1.2.1 implies that

if a model parameter A appearing in the projection matrix of a nonlinear matrix
equation (1.20) is changed in such a way as to cause the dominant eigenvalue
r of the inherent projection matrix to increase through the critical value 1.
then the extinction equilibrium x — 0 will lose its (local asymptotic) stability.
This suggests a possible bifurcation of nontrivial equilibria x =^ 0 from the
extinction equilibrium x = 0 at the critical value AQ of A, where r = 1. For
this scenario to happen, the model parameter A must be present (or related
to entries) in the inherent projection matrix P(0). Call such a parameter an
inherent parameter. In applications there are usually a considerable number of
inherent model parameters from which to choose. The choice can be dictated
by biological considerations related to the biological problem of interest or by
mathematical considerations as to which parameter is important with regard
to describing the dynamics of the specific model under consideration. One
parameter that can be used is the inherent net reproductive number n. While
this number generally does not appear explicitly in the projection matrix, it can
be introduced by scaling the class specific fertilities to n by writing fij = rup^.
The projection matrix can be written

where 3>(x) > 0 is the normalized fertility matrix. $ is normalized in the

sense that 1 is the dominant eigenvalue of T(0) + $(0) and hence the dominant
20 CHAPTER 1

eigenvalue of $(0) (/ - T(O))" 1 . (Note that 0 ^ $(0) > 0.) With this prototype
in mind we will consider matrix equations in which the parameter A appears
linearly in the inherent projection matrix.
Consider the matrix equation

Assume that

for some m x m matrices A and B.

DEFINITION 3. A real number AQ 6 R1 is called a critical value of A + \B
if I is an algebraically simple eigenvalue, of A + \oB and there exist left and
right eigenvectors w > 0 and v > 0, respectively, such that wTBv ^ 0. //, in
addition, I is a strictly dominant eigenvalue of A + \$B, then AQ is called a
strictly dominant critical value.
Note that since w is an eigenvector, w ^ 0 and it follows that WTV > 0.
Suppose that A0 is a critical value of P(A, 0) = A + \B. If c(r, A) denotes the
characteristic polynomial det (rl — A — \B) of P(A,0), then c(l,Ao) = 0 and
dc(l, \o)/dr ji 0 (since 1 is simple). The implicit function theorem implies that
there is a simple (infinitely differentiable) root r — r(A) of c(r. A) such that
r(Ao) = 1. Let v — v(\) denote the associated (infinitely differentiable) right
eigenvector. A differentiation of

leads, after evaluation at A = AQ, to

Since / — (A + XoB) is singular, it follows that WT (r'(Ao)i> — Bv) = 0, and

consequently

We see that r(A) is increasing or decreasing at AQ according to the sign o?u>TBv.

Using linearized stability theory at x = 0 (see Appendix A.2) together with
the fact that the Jacobian of P(X, x)x at x = 0 is equal to the inherent projection
matrix P(A, 0), we obtain the following result.
THEOREM 1.2.3. Assume that AQ is a strictly dominant critical value of A +
\B. If WTBv > 0, then the extinction equilibrium x = 0 of (1.27)-(1.28) loses
stability as A is increased through AQ. If uf Bv < 0, then the extinction equilib-
rium x = 0 loses stability as A is decreased through AQ.
For the case (1.26) wTBv = wT$(Q)v > 0. Theorems 1.1.3 and 1.2.3 imply
that x = 0 loses stability as n is increased through 1.
 DISCRETE MODELS 21

1.2.3 Positive equilibria. The equilibrium equation associated with (1.27)-

(1.28) is

A solution pair (\,x) of this equation will be called an equilibrium pair. An

equilibrium pair (A, 0) is called an extinction equilibrium pair. An equilibrium
pair (A,x) is positive if x > 0; it is nonnegative if x > 0. A nonextinction
equilibrium pair is an equilibrium pair (A, x) with x G -R"l/{0}- A boundary
equilibrium pair is an equilibrium pair (A, x) with x € dR™. Thus, for a nonex-
tinction boundary equilibrium pair (A, x) at least one component of x 6 /?™/{0}
is equal to 0. A continuum of equilibrium pairs is a closed and connected set in
Rl x Rm.
We can write the projection matrix in equation (1.27) as

where the entries in the matrix R ( X , x ) = P(\,x) - P(X.O) are continuous and
O(|o;|) near x — 0. The equilibrium equation becomes

where r(X.x) = R ( X , x ) x is continuous and |r(A,x)| = O(\x 2 ) near x = 0. If

/ — A is invcrtible, then the equilibrium equation can be written equivalently as

where h(X,x) is continuous and |/i(A,x)j — O(\x ) near x = 0. This equation

has the form (B.2) in Appendix B to which the theorems of Appendix B can be
applied. A bifurcation of nontrivial equilibria can only occur at characteristic
values of L. According to Theorem B.I.I (see Appendix B) a continuum of posi-
tive equilibria bifurcates from the extinction equilibrium at any (geometrically)
simple characteristic value of L associated with a positive characteristic vector.
LEMMA 1.2.1. Suppose that AQ is a critical value, of A + \B and w > 0 and
v > 0 are left and right eigenvectors of A + X^B. If I — A is mvertible, then
(a) AQ is a geometrically simple characteristic value of L = (I — A)~1B with
right characteristic vector v > 0;
(b) a, left characteristic vector q of L is given by qT = wT(I — A):
(c) t f v = wT(I - A)v = X0wTBv.
Proof. The equation (A + X0B)v = v is equivalent to X0Lv = v. Since the
null space of / — (A + X^B) is spanned by v, the same is true of the null space
of / — Aoi. This proves (a). From the definition of a critical value AO we hav
wT(A + AQ.B) = WT, and hence Xf,wTB= ^uT(I — A) — qT.From this follows
(c). To show (b) we note that qrX(>L = wT(I - A ) X 0 ( I - A)~1B = X0wTB
= <1T-
22 CHAPTER 1

FIG. 1.2. The three alternatives in Theorem 1.2.4 for the global bifurcating continuum C+
of positive equilibria are schematically represented by graphs of equilibrium total population
size p against X. In (c) p" is the total population size of a nonextinction boundary equilibrium
x

The following theorem follows immediately from this lemrna and Theorem
B.I.I. It states, roughly, that as A is varied a "branch" of positive equilibria
bifurcates from x = 0 at a critical value AQ and connects to the boundary of the
nonnegative cone R^f (oo is included in the boundary).
THEOREM 1.2.4. Suppose that k > 1 in (1.27)-(1.28), and suppose that I-A
is invertible. Suppose that AQ is a critical value of A + \B. Then there exists a
continuum C+ of equilibrium pairs such that (Ao,0) € C+ andC+ / {J?1 x dR™} /
0 contains only positive equilibria. Furthermore, one of the following alterna-
tives holds:
(a) C + /{(Ao,0)} is unbounded in R1 x R™ and contains only positive equi-
librium pairs;
(b) C+ contains a point (A*,0), where A* / AO is a characteristic value of
(I — A)~1B with a nonnegative characteristic vector;
(c) C+ contains a nonextinction boundary equilibrium (A*,x*) G Rl x dR™.
The three alternatives in Theorem 1.2.4 are illustrated graphically in Fig. 1.2.
In applications it is often the case that alternatives (b) and (c) can be ruled
out, with the result that the bifurcating branch is positive and unbounded as in
(a). For example, a case that often occurs is that L has only one characteristic
value AQ (or at least no other with a nonnegative characteristic vector). In this
case, alternative (b) is ruled out.
It also frequently occurs that the only nonnegative solution of the equilibrium
equation (1.29) is x = 0. In this case, alternative (c) is ruled out.
 DISCRETE MODELS 23

For example, a nonextiriction equilibrium x > 0 is an eigenvector of the pro-

jection matrix P(X,x) associated with eigenvalue 1. If the matrix P(X,x) is
irreducible and primitive for all (A, a:) 6 R1 x /££, then it can have no non-
negative eigenvectors other than the positive eigenvector associated with the
dominant eigenvalue. It follows that x > 0 (and 1 is the dominant eigenvalue).
Thus, for such projection matrices alternative (c) is ruled out. For an applica-
tion in which alternative (c) does occur, however, see the spotted owl model in
section 1.3.
Analytic formulas for positive equilibria of nonlinear matrix equations, ex-
pressed in terms of the model's parameters, are not in general available. How-
ever, general approximations can be obtained for the positive equilibrium pairs
on the continuum C+ near the bifurcation point (Ao,0). This can be done
by the classical procedure of parameterizing the branch of equilibrium pairs
(A, x) = (X(s),x(e)), (A(0),x(0)) = (Ao,0). in terms of a small parameter
e ss 0 (sometimes referred to as a Liapunov/Schmidt expansion). The details
of this procedure and a proof of its validity are given in Theorem B.2.1 (see
Appendix B.2). An application of that theorem to the matrix equation (1.27)-
(1.28) results in the theorem below. Here use is made of Lemma 1.2.1. Define
V± = {w 6 Rm WTV - 0}. and let G be the matrix such that x - Gf is the
unique solution of the linear algebraic equation x — A 0 (/ - A)~lBx + f lying
in V-.
THEOREM 1.2.5. Suppose thatk > 1 in (1.27)-(1.28), and suppose that I-A
is invertible. Suppose that AO is a critical value of A + XB and w7 Bv ^ 0. Then
in a sufficiently small neighborhood of the bifurcation point (Ao, 0) the continuum
C+ has the form

where, for some £0 > 0,

I f k > 2 , then

where d^ — (V x pij(A 0 ,0)) T <;.

The formula for AI in (1.31) allows for a determination of the direction of
bifurcation of C+, i.e., the sign of A — AO near the bifurcation point. If A > AQ
(A < AQ) for ( A , x ) <£ C+ near (Ao^O). we say that the bifurcation is to the
right (to the left). The direction of bifurcation is determined by the sign of
AH if AI > 0 (Ai < 0), then the bifurcation is to the right (to the left). The
coefficient z\ describes the effect the nonlinearities have on the equilibrium class
distribution x as a perturbation from the inherent (linearized) distribution v.
24 CHAPTER 1

The local stability of an equilibrium x of equation (1.27) is determined by the

eigenvalues of the JacobianJ(X.x)of the right-hand sideP ( X , x ) xevaluated
at x. Under the conditions of Theorem 1.2.5, the positive equilibrium pairs
near the bifurcation point (A 0 ,0) are parameterized by e as in (1.31). For these
equilibria the Jacobian J(A,x) = A + XB + J r ( X , x ) , Jr(X,x) = [djr^X, x)], and
hence its eigenvalues, are functions of the parameter e. A proof of the following
lemma appears in Appendix C.
LEMMA 1.2.2. Suppose that k > 2 in (1.27)-(1.28), and suppose that I — A
is invertible. Suppose that AQ is a critical value of A + XB and WTBv ^ 0. Let
C(e) be the eigenvalue of the Jacobian J(A, x) = A + XB + Jr(X, x) evaluated at
the nontrivial equilibria (1-31) such that £(0) = 1. Then

where X\ is given in (1.31).

If 1 is a strictly dominant eigenvalue of the Jacobian at (Ao,0), then the
stability of the positive equilibria (1.31) from the bifurcating continuum C+ in
Theorem 1.2.5 near the bifurcation point (Ao, 0) is determined by the eigenvalue
C(e) for small e; i.e., the positive equilibria are (locally asymptotically) stable
if (,(e) < 1 and unstable if £(e) > 1 for small £ > 0. Thus, local stability is
determined by the sign of C'(0). By Lemma 1.2.2 this sign is determined by the
sign of AI and hence (by Theorem 1.2.5) by the direction of bifurcation.
The local bifurcation described in Theorem 1.2.5 will be called a stable bi-
furcation if the positive equilibria (1.30) are (locally asymptotically) stable in a
sufficiently small neighborhood of the bifurcation point (Ao, 0) (i.e., for e > 0 suf-
ficiently small). If these positive equilibria are unstable, then the bifurcation will
be called unstable. The next theorem follows immediately from Lemma 1.2.2.
THEOREM 1.2.6. Suppose that k > 2 in (1.27)-(1.28), and suppose that I-A
is invertible. Suppose that AO is a strictly dominant critical value of A + XB.
Assume that AI ^ 0, where Aj is given by (1.31).
IfwrBv> 0,then the bifurcation described in Theorem 1.2.5 is stable if it is
to the right and unstable if it is to the left.
IfWTBv< 0, then the bifurcation described in Theorem 1.2.5 is stable if it is
to the left and unstable if it is to the right.
As a result of this theorem the stability or instability of the bifurcation at
(Ao,0) can be determined from the direction of bifurcation alone and an equi-
librium stability analysis is unnecessary. The four possibilities are illustrated in
Fig. 1.3.

1.2.4 Positive equilibrium destabilization. The stability property of the

bifurcating continuum C+ of equilibria in Theorem 1.2.6 may not persist outside
of a neighborhood of the bifurcation point (Ao, 0). The stable positive equilibria
arising from a stable bifurcation at (Ao, 0) may lose their stability as one "moves
away" from (Ao,0) along C+. Or, conversely, the unstable positive equilibria
arising from an unstable bifurcation may gain stability.
 DISCRETE MODELS 25

FIG. 1.3. The. local bifurcation alternatives for the branch of positive equilibria in Theorem
1.2.6 are schematically represented by graphs of total population size p against the parameter
A. The letter "s" denotes "stable" and "u" denotes "unstable."

Equilibrium destabilization will occur as one moves along C+ if the magni-

tude of an eigenvalue of the Jacobian becomes greater than 1; i.e., an eigenvalue
moves from the inside to the outside of the unit circle in the complex number
plane. When this happens, another attractor usually appears and a "bifurca-
tion" is said to occur. What kind of new attractor arises depends crucially on
the point at which the eigenvalue leaves the unit complex circle. General the-
orems describing the bifurcation possibilities for general maps appear in most
textbooks on discrete dynamical systems (e.g.. see [193], [436]). Nothing excep-
tional can be said about matrix population models (1.27) as a specialized class
of higher order maps. We will therefore only describe, for future reference, the
most common general local bifurcations that occur for higher order maps.
Stability is lost at an equilibrium xcr as the parameter A is increased (or
decreased) through a critical value A cr because an eigenvalue £ of the Jacobian
at x,.r moves out of the complex unit circle. That is to say, [£| = 1 for A = A cr
while, in a neighborhood of \cr, [£| < 1 for A < Acr and JC| > 1 for A > \cr (or
vice versa, |£| < 1 for A > \cr and |C| > 1 for A < A c r ). Moreover, it is usually
the case that no other eigenvalue leaves the unit circle at A = A,;r if £ is real; or
only C and its complex conjugate leave at A — A cr if (" is complex. If £ is real,
then either £ = 1 or — 1 at A — A e r .
When C = 1 at (\,x) = (\cr,xcr). the bifurcation typically involves equilibria
only. There are a variety of possibilities, the most fundamental of which are
the traw,smtical, saddle-node, and pitchfork bifurcations. A transcritical bifur-
cation is the crossing of two distinct branches of equilibria at the bifurcation
26 CHAPTER 1

FIG. 1.4. A transcritical bifurcation occurs at the point (A,p) = (Ao,0) where the ex-
tinction equilibrium p = 0 loses stability. A saddle-node bifurcation occurs at the point
(A,p) = (AenPcr)- The letter "s" denotes "stable" and "u" denotes "unstable."

point in such a way that equilibria from both branches exist for A values on
both sides of A cr . A transcritical bifurcation usually involves, at the bifurcation
point, an exchange of stability from the equilibria on one branch to those on
the other. The bifurcation of positive equilibria from the extinction equilibrium
(Xcr,Xcr) = (Ao,0) as described in Theorem 1.2.5 is a transcritical bifurcation
(the other "half" of the branch corresponding to £ < 0 is ignored because it
consists of negative equilibrium pairs and hence is not biologically relevant).
Transcritical bifurcations at positive equilibria xcr > 0 do not typically occur
in population models.
Saddle-node bifurcations at positive equilibria, on the other hand, do occur
in population models. A saddle-node bifurcation is one in which, at least in a
neighborhood of the bifurcation point (A cr , xcr), two positive equilibria exist for
A on one side of \cr while no positive equilibria exist on the other side of \CT.
Thus, at a saddle-node bifurcation point (Acr,a;cr.) the continuum C+ "turns
around." Usually (but not always) the two positive equilibria have opposite
stability properties, one being a saddle and the other a stable node. One cir-
cumstance in which a saddle-node bifurcation occurs in population models is
when the bifurcation at (A,x) = (A 0 ,0) is unstable (due, for example, to an
Allee effect), but the continuum C+ "turns around" at a saddle-node bifurca-
tion to create stable equilibria. See Fig. 1.4. The spotted owl model in section
1.3 is a specific example.
Finally, in a neighborhood of pitchfork bifurcation point (Xcr,xcr). three equi-
libria occur on one side of Acr while only one equilibrium occurs on the other
side. Pitchfork bifurcations have rarely occurred in population models.
 DISCRETE MODELS 27

FIG. 1.5. -4s A is increased through \CT the positive equilibrium is destabilized and a stable
"2-cycle emerges as shown in the bifurcation diagram and time series plots in graphs (a) and
(b). In phase space the attractor changes from an equilibrium point to a pair of points on the
2-cycle (shown in graph (c) for m — 2 dimensions).

The second possibility £ = — 1 at (A, a;) = ( X c r . x c r ) usually implies the bifur-

cation of a branch of 2-cycles, i.e., periodic solutions of period 2. This "period
doubling" bifurcation, familiar for one-dimensional maps, also occurs for higher
dimensional maps. These 2-cycles "grow" or "pop" out of the equilibrium at
the bifurcation point, starting with a small amplitude that increases as A moves
further from Xcr. The 2-cycles exist for A on one side of \cr, but not the other,
and at such a value of A the 2-cycle has the opposite stability of the equilibrium.
If the 2-cycles are stable (i.e.. exist on the same side of AQ as do the unstable
equilibria), the bifurcation is called stable or supercritical. In the opposite case
when the 2-cycles are unstable (i.e., exist on the same side of AQ as do the
stable equilibria), the bifurcation is unstable or subcritical. See Fig. 1.5 for an
illustration of a stable 2-cycle bifurcation.
The final possibility C = exp(±z#), 9 ^ 0 or TT, at (X,x) — (\cr.,xcr) can only
occur for maps of dimension m = 2 or higher and therefore are particularly
relevant for structured population models. In this case a so-called invariant
loop (usually) bifurcates from the equilibrium. This loop is a one-dimensional,
closed curve in Rm (roughly elliptical near the bifurcation point). It is not itself
a solution or orbit, but it is an invariant set of points. The orbits lying on the
invariant loop may either be "quasi periodic" or "period locked"; that is to say,
they may either move around the loop without being mathematically periodic
(and hence never hitting a point twice), or they may be periodic in their motion
around the loop. Invariant loops are best seen in phase space Rm rather than
28 CHAPTER 1

FIG. 1.6. As A is increased through Acr the positive equilibrium is destabilized and quasi-
periodic oscillations emerge as shown in the bifurcation diagram and time series plots in
graphs (a) and (b). In phase space the attractor changes from an equilibrium point to a
closed invariant loop (shown in graph (c) for m = 2 dimensions).

m time series plots of components of x against time t. See Fig. 1.6. The in-
variant loop bifurcation, sometimes called a discrete Hopf (or Naimark/Sacker)
bifurcation, requires the unexpected technical assumption that £ not be equal to
any of the first four roots of unity; at such points the nature of the bifurcation
is apparently not yet fully understood. The invariant loop bifurcation may be
either stable or unstable; in other words, orbits starting near the loop may tend
towards or away from the loop in forward time.
The analytical determination of the type and stability properties of a bifurca-
tion is usually difficult and intractable in applications, although it is in principle
possible by means of known procedures [306].

1.2.5 The net reproductive number. In this section we take a closer

look at using the inherent net reproductive number n as the bifurcation param-
eter A.
Consider the general projection matrix P(x) = T(x) + F(x) > 0. Assume
that T(x) + F(x} is irreducible and primitive and I - T(x) is nonsingular for all
x e R™. If F(x)(I — T(x))~1 has a positive, strictly dominant, simple eigenvaluea
n — n(x) with a nonnegative eigenvector v = v(x) > 0, then n(x) is called the
net reproductive number at x. The quantity n(x) is not to be confused with the
inherent net reproductive number n (which is, in fact, n = n(0)). Suppose that
x > 0 is a positive equilibrium so that x solves the equation x = (T(x) + F(x))x.
Since x > 0 and the matrix T(x) + F(x) is irreducible and primitive, it follows
 DISCRETE MODELS 29

that 1 is the dominant eigenvalue of T(x) + F ( x ) . From Theorem 1.1.3 we

conclude

The biological interpretation of this statement is, roughly, that at equilibrium

each individual exactly replaces itself over the course of its lifetime.
If the fertilities jij are scaled to the inherent net reproductive number n,
fij = n(Pij> tnen we can write F(x) = n$(x), where $(x) > 0 (^ 0) is the
normalized fertility matrix, and the matrix equation (1.27) becomes

We consider this equation under the following assumptions. For some integer
k e J[0, + oo)

(1.34) (d) T(x) + n$(x) is irreducible and primitive

for all x £ R%, n > 0;

(e) for all x 6 #™, &(x)(I - T(x})~1 has a

positive, strictly dominant, simple eigenvalue
value v(x), i>(0) = 1, with a nonnegative
eigenvector vector u(x) > 0, such that
(I -T(x))~lu(x) >0.

The last assumption implies that n(x) = rw(x), where n is the inherent net
reproductive number. From (1.32)

(1.35) nv(x) = 1 for all equilibria x > 0.

By Theorems 1.1.3 and 1.2.1, x — 0 loses stability as n is increased through

the critical value 1. From Theorems 1.2.5 and 1.2.6 the bifurcation of positive
equilibria at n = 1 is stable if the bifurcation is to the right and unstable if it is
to the left. Define the spectrum and range of the continuum C+ from Theorem
1.2.4 as

respectively. The spectrum and the range are connected sets. For linear models
s(C+) = {1} and p(C+) = {cu \ c 6 #"}.
30 CHAPTER 1

THEOREM 1.2.7. Consider the matrix equation (1.33) under the conditions
(1.34). Then
(i) p(C+} is unbounded and p(C+)/ {0} contains only positive equilibria;
(ii) <j(C+) contains only positive n > 0.
Proof, (i) First we rule out alternatives (b) and (c) in Theorem 1.2.4.
Suppose that (n\, 0) e C+, where n\ ^ 1 is a characteristic value of L = (I —
T(0))~1*(0) (and hence of $(0)(/ - T(O))" 1 ) with a nonnegative characteristic
vector £ > 0. Then 1 is an eigenvalue and £ is an associated eigenvector of
the matrix T(0) + ni$(0). Since T(0) + ni$(0) is irreducible and primitive, a
nonnegative eigenvector can be associated only with the dominant eigenvalue.
Therefore 1 is the dominant eigenvalue of this matrix. By Theorem 1.1.3 it
follows that the dominant eigenvalue of ni$(0)(/-T(0))' 1 is equal to 1, which
by (1.34(e)) implies the contradiction n\ = 1. This rules out alternative (b).
By (1.34(d)) the projection matrix T(x) + n$(x) can have no nonnegative
eigenvector other than the positive eigenvector associated with the dominant
eigenvalue. Since an equilibrium is an eigenvector of the projection matrix
associated with eigenvalue 1. it follows that a nonnegative equilibrium must in
fact be positive. This rules out alternative (c).
By alternative (a) of Theorem 1.2.4 p(C+) contains only positive equilibria
and either it or the spectrum is unbounded. If the range p(C+) were bounded,
then v(x] > 0 would be bounded away from 0 for all x 6 p(C+) and (1.35)
would imply that the spectrum o~(C+) would also be bounded, a contradiction.
This proves (i).
(ii) Suppose that there exists a pair (n,x) 6 C+ for which n < 0. Because
+
C is a continuum, it follows that there must exist an equilibrium pair (0, x) €
C+. By (i), x > 0. From the equilibrium equation it follows that x = T(x)x
which contradicts (1.34(b)). Thus, n > 0 for all (n,x) € C+.
The next theorem contains facts about the spectrum obtainable from prop-
erties of v(x) (or equivalently n(x}).
THEOREM 1.2.8. Consider the matrix equation (1.33) under the conditions
(1.34).
(i) // lini|a.|_++00v X^R™ v(x) = 0, then there, exists at least one positive equi-
librium for each n > 1.
(ii) If v(x) < I for x > 0, x « 0, then the bifurcation is to the right and stable.
If v(x) > 1 for x > 0, x « 0, then the bifurcation is to the left and unstable.
(iu} If v(x) < 1 for all x > 0, then there exists no positive equilibrium for
n < 1.
Proof. If lini|x|^+00! x6,Rm v(x) — 0, the unboundedness of the range p(C+)
and (1.35) imply that the spectrum o~(C+) is unbounded; i.e., o-(C+) is an
unbounded interval of positive real numbers that contains 1 in its closure. This
implies (i). (Note that it is sufficient that v(x) tends to 0 for x e p(C+).) Both
(ii) and (iii) follow immediately from (1.35).
In applications v(x) —> 0 as |a;| —> +00 is usually a consequence of the mod-
eling assumption that fertility and/or transition (e.g., survival) rates tend to 0
as \x increases without bound.
 DISCRETE MODELS 31

In applications, the inequality i/(x) < 1, x > 0. is usually the result of the
modeling assumption that all density effects are deleterious, i.e.. that fertility
and/or transition rates are, in the presence of other individuals, less than inher-
ent values. This condition will be fulfilled, for example, if for all 0 < x 6 .R™

To see this, note that from 0 < T(x)y < T(0)j/, 0 ^ y 6 R*?, we obtain
0 < Ti(x)y < Tl(0)y and hence

Then 0 < $(x) (/ - T(x))"1 y < $(0) (/ - T(O))"1 y, and since the unit sphere
is compact in Rm,

Often in applications density effects are modeled by a dependence of the

fertility matrix F = F(p) and/or the transition matrix T = T(p) on a weighted
total population size

Then v = v(p) is a scalar function of a scalar variable and (1.35) describes a

curve in the (n,p) plane. A plot of this planar curve is one way of geometrically
describing the bifurcation of the continuum C+. For example, from (1.35) we
see that ^'(O) < 0 implies a right and stable bifurcation and f'(0) > 0 implies a
left and stable bifurcation.
THEOREM 1.2.9. In (1.33)-(1.34) suppose that F = F(p) and T = T(p) are
functions of a weighted total population size p — Y^=i ^iXi, u>i > 0, /^^Li ^ ¥"
0. // v'(p) < 0 for all p > 0, then there exists no positive, equilibrium for n < 1
and &(C+) ~ [1. l/i>oo), where vx — limp^+00 v(p) > 0 (replace 1/f oc by +oc
*/ ^oo = Oj. Moreover, for each n £ ff(C+) there exists exactly one positive
equilibrium x.
Proof. All that needs to be proved is the uniqueness statement; the rest of
the theorem follows from (1.35). Suppose that x > 0 and x* > 0 are equilibria
associated with the same value of n. By (1.35) x and x* have the same weighted
total population size p and hence both are positive eigenvectors, associated with
eigenvalue 1, of the same matrix T(p) + F(p] = T(p) + n<&(p). By assumption
(1.34(d)) this matrix is irreducible and primitive and therefore has a only one
independent positive eigenvector. Thus, x and x* are dependent. In fact, they
must be identical since they have the same weighted total population size.
32 CHAPTER 1

A similar argument shows that if f'(p) > 0 for all p > 0, then there exists
no positive equilibrium for n > 1 and a(C+) = [l/Voo> !)• Moreover, for each
n < 1, n 6 a (C + ), there exists exactly one positive equilibrium x.
A positive derivative v'(p) > 0 can arise from an "Allee" effect in which some
fertility and/or transition rates increase with increased total population size.
Allee effects are usually postulated to hold at low population densities, while
the usual deleterious density effects are assumed to hold at high population
densities. In such a model, v'(p) > 0 for small p, say 0 < p < pcr, and v'(p) < 0
for large p > pcr. Using (1.35) we can deduce a bifurcation diagram as in Fig.
1.4. Thus, in this case there exist two positive equilibria for n satisfying ncr —
\/v(pcr) < n < I and one positive equilibrium for n > 1.
In the results above, the sign of the derivative v'(p), or equivalently n'(p), for
small p > 0 determines the stability of the positive equilibrium pairs (n, x) near
the bifurcation point (1,0). A natural question to ask is whether the derivative
n'{p) can tell us anything about the stability of positive equilibria in general.
Return, for the moment, to the general case P(x) = T(x) + F ( x ) . By defini-
tion n(x) is the dominant eigenvalue of F(x)(I — T(x))~l with a nonnegative
eigenvector u(x) > 0. n(x] is also an eigenvalue of the matrix (I — T(x))~lF(x)
with eigenvector z(x] = (I — T(x))~lu(x) > 0. From the equation F ( x ) z ( x ) =
n(x)(I — T ( x ) } z ( x ) one can derive a formula for the partial derivative c^n. If this
formula is evaluated at an equilibrium x > 0 (where n(x) — I and z(x) = x),
one obtains

If both sides are multiplied by the left eigenvector wr(x) of P(x) at equilibrium
(for which wr(x) = wT(x}P(x)} we get the formula

Return now to the case where density dependence is through a dependence

on a single weighted total population size p

Then n — n(p), din(x) = Wjn'(p), and diP(x) = o^P'(p) in (1.37). Choosing an

i for which Wj ^ 0, we obtain from (1.37)

Can this derivative be related to the eigenvalues of the Jacobian J(p)7

As an example, consider the general nonlinear Usher matrix
 DISCRETE MODELS 33

for which (see (1.15))

A straightforward (but tedious) calculation shows (at equilibrium)

If n'(p) > 0, then the characteristic polynomial det (£/ — J(p)) of the Jacobian
is negative when evaluated at C, = 1. Since this polynomial tends to +00 for
real ( —> +oc. it follows that it must have a real root ( > 1. Thus, n'(p) >
0 implies that the positive equilibrium is unstable and n'(p) < 0 becomes a
necessary condition for equilibrium stability. This necessary condition has been
established for general matrix models of the form (1.38). The theorem below
can be found in [442]. Let ('.^(p) be the cofactor of the ilh diagonal element
in the matrix / — P(p)- (The strong dissipative condition X/i^i^X 3 -) — 1 'ls
required in [442], but it is only needed to guarantee the existence of the net
reproductive number. Conditions (1.34) are sufficient for this purpose and the
proof in [442] is unchanged.)
THEOREM 1.2.10. Consider the. matrix equation (1.33) with a projection ma-
trix of the form (1.38) under assumptions (1.34). Let x > 0 be. a positive equi-
librium with weighted total population size p = x\ > 0. Assume that ca(p) > 0
for all i = 1. 2 , . . . . m. Then n'(p) > 0 implies that x is unstable.
The condition that ca(p) > 0 for all z = 1, 2 . . . . . m can be shown to hold for
general nonlinear Usher models (and hence for general nonlinear Leslie models).
It can also be shown to hold for general matrix models of size m — 2 and
3; whether it holds for general matrix models of size TO > 4 remains an open
question [442].
While Theorem 1.2.10 shows that n'(p) < 0 is necessary for the stability of a
positive equilibrium, neither this condition, nor the stronger condition n'(p) <
0, is sufficient to guarantee stability. The famous m, = I dimensional Ricker
map x(t + 1) = bexp(-cx(t))x(t). for which n(x) = bexp(-cx), serves as a
counterexample.
The geometric interpretation of Theorem 1.2.10 is that equilibria whose weight-
ed total population size p lies on a decreasing branch of the graph of (1.35) in
the (n.p) plane are unstable. See Fig. 1.7.
Finally, we point out that while we have used (1.32) to deduce properties of the
bifurcating branch of equilibrium pairs when using the inherent net reproductive
number n as the bifurcation parameter A, this identity can also be used for the
same purpose when other bifurcation parameters are used.

1.3 Some applications

The applications below illustrate the use of the results from the previous sec-
tions. The first application provides a general treatment of a large class of
34 CHAPTER 1

FIG. 1.7. Under the conditions of Theorem 1.2.10 the decreasing portions of the equilib-
rium branch are unstable.

nonlinear matrix models. The second application is to a specific model for the
population dynamics of flour beetles and it illustrates a stable bifurcation from
the extinction state (followed by either 2-cycle or discrete Hopf bifurcations).
This matrix model is also used to illustrate how different choices of inherent
parameters can be made for use as a bifurcation parameter. The third appli-
cation involves a matrix model that has been used to study the dynamics of
the controversial spotted owl of the Pacific northwest. This model illustrates
an unstable bifurcation from the extinction state due to an Allee effect. The
final example involves a size-structured model that has been used to conduct a
theoretical study of the effects of competition on juvenile maturation size. This
example illustrates a "nongeneric" bifurcation from the extinction state.

A general nonlinear Usher model. Consider a nonlinear Usher projection

matrix (1.40) in which the density dependence is through a dependence on a
weighted total population size p = J2^Ll WjXj, &i > 0, X)Hi w« ^ 0- The net
reproductive number is (see (1.15))

Setting fu(p) = n<pn(p), where the <fu are normalized so that

 DISCRETE MODELS 35

the number n is the inherent net reproductive number and the nonlinear Usher
projection matrix has the form P(p) — T(p] + n$(p), where

If we assume, for an open interval A C Rl containing the half line [0, +00), the
smoothness conditions

** 6 C fc+1 (A, [0,1)), ti,i_i eC f c + 1 (A, (0,1]), ^-eC^A,^)

for some integer fe e 7[0, +00) and in addition the inequalities

then the nonlinear Usher matrix satisfies the first four conditions (1.34(a)-(d))
in section 1.2.5. From

we find

whose dominant eigenvalue and eigenvector are

36 CHAPTER 1

(For notational convenience we have defined ^io(p) = 1-) Then for all p € R\

and the last condition (1.34(e)) in section 1.2.5 is satisfied. Thus, all results in
sections 1.2 and 1.2.5 hold for the general nonlinear Usher model.
As an illustrative example, consider a nonlinear Usher model with Ricker-
type, exponential nonlinearities for fertilities and transition probabilities; specif-
ically

The normalized inherent fertilities j3i satisfy the normalization (1.41), namely,

One interpretation of these submodels for the transition probabilities is that

TTi is the probability that an individual in class i survives one unit of time and
exp (—dip) is the probability that such a survivor moves into (e.g., "grows" into)
the next class i + 1. The number of newborns per individual of class i is 0i and
the exponential exp(—Cip) is the probability that a newborn from a parent of
class i will survive to the next census time.
For this example, it is not difficult to see from (1.42) that v'(p) < 0 for all
p > 0 and limp_,+00 v(p) = 0. Thus, from the results in section 1.2.5 we find
that x = 0 loses stability as n increases through 1. For each n> 1 the nonlinear
Usher matrix model is uniformly persistent with respect to the extinction equi-
librium x = 0 (by Theorem 1.2.2) and there exists a unique positive equilibrium
these positive equilibria are (locally asymptotically) stable for n w 1 and are
unbounded as n —> +00. Moreover, (1.22) holds and consequently n < I implies
global extinction (i.e., limt_+00 \x(t)\ = 0 for all x(0) > 0).
In the preceding example only the facts that the submodels for the fertili-
ties and transition probabilities tend monotonically to 0 as p increases without
bound were used. Thus, other nonlinearities (or a mix of nonlinearities). such
as (1 + Cipa)~l, could be used with the same results.
 DISCRETE MODELS 37

The LPA flour beetle model. Flour beetles (Tribolium) live almost exclu-
sively in stored grain products and have apparently done so for nearly as long
as humans have stored and milled grains (e.g.. evidence of their presence have
been found in ancient Egyptian urns). Besides being a major agricultural pest,
flour beetles are an excellent animal for laboratory experiments in population
biology. For these reasons, flour beetles have been intensely studied for many
decades and their physiological, behavioral, genetic, and life-cycle characteris-
tics are very well understood. In [72], [73], [74], [116], [117], [131], [132], [133],
[232] the following matrix model (called the LPA model)

is used to describe the population dynamics of laboratory cultures of flour bee-

tles. In this model L(t) is the number of larvae, P(t) is the number of pupae
(but also including rionfeeding larvae and young nonreproducing adults), and
A(t) is the number of reproducing adults at time t. The time unit (census in-
terval) is two weeks. The nonlinear interactions in this model, as described by
the exponential terms, are attributed to cannibalism between certain life-cycle
stages, and the parameters cei > 0, cea > 0. and cpa > 0 are called the "can-
nibalism coefficients." Thus, the inherent number of larvae produced per adult
per unit of time, b > 0, is reduced by the fraction cxp( —c e iL — ceaA) due to larval
and adult cannibalism of eggs. Pupal mortality is due only to cannibalism by
adults. Larvae and adults have constant death rates 0 < fit < 1 and 0 < na < 1.
respectively. Cannibalism in flour beetles occurs from random encounters and
is therefore well modeled by exponential type nonlinearities (see section 1.1.3).
The LPA model (1.43) has the form of a nonlinear Leslie matrix model with

and inherent net reproductive number

It can be put in the form (1.33) with transition matrix

and normalized fertility matrix

38 CHAPTER 1

Note that

so that

Thus, all conditions in (1.34) are satisfied (for all k > 0).
Since P(x) < P(0) for all x € R3+, we find from Theorems 1.1.3 and 1.2.1
that n < 1 implies that the origin x = 0 is globally attracting in R^_ and
the model predicts asymptotic extinction. From these same theorems we have
that the origin is unstable for n > 1. Unfortunately, the transition matrix
associated with P(x] does not satisfy the condition (1.24) and therefore we
cannot apply Theorem 1.2.2 directly to show that the LPA model (1.43) is
dissipative. However, this can be shown as follows. From the first and second
equations in (1.43) we have the inequalities

which show that after two time steps the L and P components of all solutions
are bounded by common bounds. From the third equation in (1.43) we have

from which we conclude, using an induction argument as in the proof of Theorem

1.2.2.

for sufficiently large t > 0; i.e., (1.43) is dissipative. Uniform persistence (with
respect to the origin) follows from Theorem 1.2.1 for n > 1. Theorems 1.2.7 and
1.2.6 guarantee that the bifurcation, from the origin at n = 1, of a global contin-
uum of positive equilibrium pairs which has a positive spectrum, is unbounded
 DISCRETE MODELS 39

FIG. 1.8. Two bifurcation, diagrams for the flour beetle, model (1.43) show the. attractor,
represented by total population size p = L + P + A, plotte.d against the parameter b. In (a).
cea = cel = 0.01, cpa = 0.05, /Kj = 0.2, and /j,a = 0.9, and in (b), cea = cei = cpa = 0.01,
jij = 0.5. and /i,a = 0.20. To eliminate transients, 1000 iterations were performed before 100
values of p were plotted.

in B,\ x R^_ and is (locally asymptotically) stable near the bifurcation point.
The inequalities above (which are valid for all solutions, including equilibrium
solutions) show that the magnitude of equilibria are bounded above by a mul-
tiple of n, and from this it follows that the spectrum of the continuum cannot
be bounded (for if it were, then the global continuum would be bounded). It is
not difficult to show that the positive equilibrium for n > I is unique.
The equilibrium stability for n near 1 does not in general persist for large n.
The bifurcation diagrams in Fig. 1.8 show that both 2-cycle bifurcations and
discrete Hopf (invariant loop) bifurcations are possible for the LPA model (1.43),
depending on parameter values. Fig. 1.8(a) shows a stable 2-cycle bifurcation
occurring at a critical value of n > I , where the positive equilibrium loses
stability (because an eigenvalue of the Jacobian decreases through — 1 as n
increases), and Fig. 1.8(b) shows the bifurcation of an invariant loop (because
a complex conjugate pair of eigenvalues leaves the complex unit circle).
In the experiments described in [72], [133] the adults' death rate /;,a was
manipulated. We can also study (1.43) using the inherent death rate A = p,a as
a bifurcation parameter. To do this we write the inherent projection
40 CHAPTER 1

as P(0) = A + XB, where

The only critical value of A + XB is AQ = (1 — nt) b and it is simple and strictly

dominant. The right and left eigenvectors of A + X0B associated with the
eigenvalue 1 are

from which we calculate wrBv = — 1 < 0.

Theorem 1.2.3 implies that the extinction equilibrium x = 0 loses stability
as A = ^a decreases through AQ, and Theorem 1.2.4 implies the existence of
a bifurcating continuum C+ of equilibrium pairs that satisfies the alternatives
of that theorem. However, the uniqueness of the critical value AQ rules out
alternative (b). Moreover, it is straightforward to show, from the equilibrium
equations, that a nonnegative equilibrium must be positive; this rules out al-
ternative (c). The remaining alternative, (a), implies that the continuum C+ is
unbounded in Rl x R+ and consists of positive equilibrium pairs (A, x) (except
for the bifurcation point (Ao,0)). The identity (1.32), which must hold for all
equilibrium pairs from C+, is

Thus, the spectrum cr(C+) cannot be unbounded. The unboundedness of C+

then implies that the range p(C+) is unbounded. By (1.44) A = 0 is in the
closure of a(C+), Since (1.44) implies that A < AO, we deduce that a(C+) is
identical with the interval 0 < A < AQ and the bifurcation is to the left.
Furthermore, the formulas in Theorem 1.2.5 yield

and hence

Theorems 1.2.5 and 1.2.6 imply that the bifurcation is stable.

As we have noted above, the spectrum a values associated
of A =with fj,
positive equilibria from C+ is the interval 0 < A < AQ. Thus, either AQ > 1
and the extinction equilibrium is unstable for all A = jua of biological interest
(0 < fia < 1) or AQ < 1 and the bifurcation occurs at a biologically meaningful
value of the adult death rate //0. See Fig. 1.9.
 DISCRETE MODELS 41

FlG. 1.9. The two possible bifurcation diagrams for the beetle model (1.43) using the adult
death rate X — na as the bifurcation parameter are shown. On the. left AQ = (1 — fi{) b < 1,
while on the right AQ > 1. The local stability of the positive equilibria near the bifurcation
point (Ao,0) may not persist for all values of aA <=1.fj.

A spotted owl model. The northern spotted owl is a long-lived, territorial

predator that requires large tracts of old growth forest in which to live. Interest
in the population dynamics of the spotted owl was intensified during the 1980s
by an environmental controversy concerning the logging of old growth forests in
the Pacific northwest. A matrix model for these dynamics was introduced and
studied in [272], [275] (also see [209]). This model describes the dynamics of the
m = 2 variables
S ( t ) = number of single male owls at time t,
C(i) = number of paired adults (reproductive couples) at time t
by a matrix equation with the projection matrix

The unit of time is one year. This model is based upon the life cycle of the
spotted owl. After a year as juveniles, males establish territories and attract
breeding females. Males are either single or paired with a female (monoga-
mously). A juvenile must find a territory by the beginning of the second year
(or it will die or leave the area). Timber harvesting fragments the habitat
into small patches, which influences the population dynamics through effects on
territory availability (nesting sites) and mating success, dispersal, etc. In the
projection matrix above
probability of a female finding a mate.

probability a juvenile survives dispersal.

42 CHAPTER 1

The model parameters (taken here to be constants) are

fraction of single males surviving one year,
fraction of juveniles surviving one year (to become single adults),
number of offspring per adult pair per year,
probability that only the female of a pair dies in a year,
probability both individuals in a pair survive one year,
unoccupied site search efficiency by single males (number of
search attempts in a year before dying),
unmated female search efficiency (number of search
attempts in a year before dying or leaving the area),
total number of sites in the system,
number of suitable sites in the system.
The inherent projection matrix

has eigenvalues ss < 1 and pa < 1, and therefore the extinction equilibrium
is (locally asymptotically) stable for all model parameter values. Thus, in this
model small populations will go extinct. Does the model predict that all popu-
lations will go extinct? Or are there stable positive equilibria?
If we want to study this question using the bifurcation theory from the pre-
ceding sections, we need an inherent parameter which, when varied, can lead to
destabilization of the extinction equilibrium x = 0. The only choices are ss or ps.
If, for example, ps is mathematically allowed to increase through 1, x = 0 will
become unstable. Of course, values of ps > I are not biologically meaningful in
this model, so ultimately we must restrict attention to ps < 1.
Let A = ps, and write P(0) = A + \B, where

The only critical value of A + XB is AQ = 1. From the eigenvectors

we find wTBv = 2(1 — ss) > 0. Theorem 1.2.4 implies the existence of a bifur-
cating continuum C+ of equilibrium pairs that satisfies the alternatives of that
theorem. However, the uniqueness of the critical value AQ rules out alternative
(b). We will rule out alternative (a) below. Before doing that, however, we note
that the bifurcation is to the left and unstable. This follows from a calculation
that, shnws
 DISCRETE MODELS 43

Theorems 1.2.5 and 1.2.6 imply that the bifurcation is to the left and unstable.
The equilibrium equations are

Note that the second equation implies that A < 1 for any positive equilibrium.
Also, from these equations one can deduce that the only nonextinctiori equi-
librium pair (A*,x*) € R1 x dR2+, 0 < x* =£ 0, available for alternative (c) in
Theorem 1.2.4, is

A study of the equilibrium equations will show that alternative (a) of Theorem
1.2.4 (that C+ is unbounded in the positive quadrant) is ruled out. This can
be done by a contradiction argument, ruling out the possibility of the unbound-
edness of the spectrum and of the components S and C. We conclude from
Theorem 1.2.4(c) that the continuum C+ contains the point (A*.C*) given by
(1.45a).
We have shown that there exists a continuum of positive equilibrium pairs
(A,x) with A < 1 which connects the equilibrium pairs (1,0) and (l,x*). It
follows that the continuum must "turn around" at some point \cr < I, where a
saddle-node bifurcation occurs, and there exist at least two positive equilibria
for A on the interval Acr < A < 1; see Fig. 1.10. Although we have no proof of
this (except for the "lower" branch near A = 1), we suspect that the "smaller"
equilibria on the "lower" branch in Fig. 1.10 are unstable while the "larger"
equilibria on the upper branch are stable. Computer simulations bear this out.
The biological conclusion is that while sufficiently large populations of the
spotted owl are viable, small populations are in danger of extinction. This
threshold phenomenon (its estimation, its dependence on model parameters,
etc.) is the focus of the study in [272].

A juvenile growth model. Intraspecific competition between juveniles and

adults has received considerable attention in the literature. One question of in-
terest concerns the "stabilizing" or "destabilizing" effects of such an interaction.
Generally, competition between juveniles and adults is considered destabilizing,
although there can be exceptions and such an assertion depends on exactly what
is meant by "destabilization." Studies of this question using discrete models can
be found in [97], [112], [151], [152], [153], [299], and [110]. (Studies using con-
tinuous models can be found in [97], [113], [312], and [402].)
44 CHAPTER 1

FIG. 1.10. The unstable bifurcation in the spotted owl model at A = ps = 1 has a saddle-
node bifurcation at A C r. The dotted lines at (A,p) = (1,C"") represent equilibria that have left
the positive quadrant.

The nature of juvenile versus adult competition can be quite diverse, depend-
ing on details of the life cycle of the species. Most mathematical models in the
literature utilize age structure. However, it is often pointed out that body size,
rather than chronological age, is more often the determining factor in an indi-
vidual's interaction with its physical and biological environment and hence in
the effects that competition has on vital rates; see [154], [435]. A size-structured
model for juvenile versus adult competition is studied in [153]; this model is not
analytically tractable, however, and its study relied on computer simulations.
Here we will consider a simpler, but more tractable, size-structured model in-
troduced in [112]. This model focuses on how competition with adults effects
the size of a juvenile at maturation and hence its adult fertility.
The fertility and transition matrices

describe a population with a stage class distribution vector

where J is the number (or density) of a juvenile class and A\ and A% are the
numbers (or density) of individuals in two adult classes whose fertility rates are
ri/i and n/2, respectively. A fraction s of surviving juveniles become smaller
adults after one unit of time, the remaining fraction 1 — s become larger adults.
 DISCRETE MODELS 45

Competition for resources is assumed to affect a juvenile's size at maturation.

This is modeled by assuming s = s(p) is a function of the weighted total popu-
lation size

The weights (il measure the relative strength of the competition from the two
adult classes as experienced by juveniles. Increased competition is assumed to
result in less juvenile growth and hence smaller adults, so s(p) is assumed to be
an increasing function of p, as follows:

the last condition implying that as population density increases without bound
the fraction of juveniles growing to become large adults drops to 0. The coeffi-
cients /3j satisfy the normalization

and <T measures the strength of adult competition experienced by a juvenile

relative to competition from other juveniles.
The two adult classes are distinguished by the size of the individuals in them,
the individuals in the first class AI being the smaller. The larger adults are
assumed to be more fertile, and therefore 0 < f\ < /2- The coefficients f\ and
/2 satisfy the normalization

so that n is the inherent net reproductive number. Adults survive only one time
unit (or at least remain fertile only one time unit), and therefore this model is
appropriate only for semelparous populations.
From

we obtain the net reproductive number n(p) = nv(p). where

We can use the inherent net reproductive number n as a bifurcation parameter

and apply Theorems 1.2.4 and 1.2.5 to obtain a continuum C+ of equilibrium
pairs that bifurcates from the extinction equilibrium. From nv(p) — 1 and
y
'(p] — (/i - /2)s'(p) < 0 we deduce that the bifurcation is to the right. We
would expect, from Theorem 1.2.6, that the bifurcation is stable. This theorem,
however, cannot be applied since n = 1 is not a strictly dominant critical value
(because the inherent projection matrix has eigenvalues ±1 when n = I).
46 CHAPTER 1

FIG. 1.11. In the upper two graphs, bifurcation diagrams for the juvenile growth model
(1.46)-(1.47) with s(p) = 1 — exp(— p) are shown for two different values of the competition
coefficient rr, namely, a = 0.5 in (a) and a = 5.0 in (b). The values of p on the attractors
are plotted as a function of the inherent net reproductive number n. Other parameters are
f\ = 0.1, /2 = 1.0. /?! = 0.25, /?2 = 1-0. In (a) the bifurcating equilibria are stable while in
(b) the bifurcating 1-cycles are stable. The times series plots of these attractors are shown
for n — 4 in the lower graphs. In (b) note that the resulting synchronous 1-cycle is such that
juveniles and adults do not appear at the same time.

In fact, it is shown in [112] that an additional continuum consisting of 2-

cycles also bifurcates to the right from the extinction equilibrium at n = 1.
(The existence of this continuum is suggested by the eigenvalue — 1 at n = 1.)
These 2-cycles are "synchronous" in the sense that juvenile and adult numbers
periodically alternate between a positive value and 0 in an out-of-phase manner.
This means that in these 2-cycles adults and juveniles are never present at the
same time (thus avoiding direct competition!). It is shown in [112] that only
one of the two bifurcating branches can be stable. Specifically, the positive
equilibrium branch is stable near n = 1 if <r < 1 and the synchronous 2-cycles
are stable near n = 1 if a > 1. See Fig. 1.11. For large values of n the situation
is more complicated, however; see [112].

1.4 A case study

Can mathematical models be used to describe and predict the dynamics of a real
biological population? Can model predicted dynamics be observed in population
data? Can predicted bifurcations in dynamics be observed in a population when
 DISCRETE MODELS 47

physical or biological parameters are changed? Given the complexities of the

biological world it is reasonable to begin by asking these questions in a controlled
laboratory setting. In this section we give a brief description of a project that
addresses these questions by means of controlled laboratory experiments and a
relatively simple matrix model.
Mathematical models are always simplifications, of course, so that an impor-
tant first step in building a model is to determine the dominant, biological and
physiological mechanisms that account for the dynamics of the population of
interest under the relevant environmental circumstances. These mechanisms are
to be incorporated into the model while other mechanisms considered to be of
less importance are ignored. The next crucial issue is that of connecting the
resulting mathematical model with data. That is to say. how are model param-
eters to be assigned numerical values (i.e., how is the model "calibrated")? How
is the resulting model "validated"; i.e., how are the model and the data used in
the parameterization to be compared and a decision made as to how well the
model "fits" the data? How are the model and data not used in the parame-
terization to be compared and the model's predictive capability to be judged?
A thorough study of these issues fall outside the scope of this monograph. In
section 1.4.1 a very brief introduction to one approach to these problems is
given.
In section 1.4.2 a project involving controlled laboratory experiments is de-
scribed in which a matrix model and its predicted bifurcations constitute the
foundation on which the experiment protocol is based. This case study shows
convincingly that a matrix model can describe and predict the asymptotic dy-
namics of a biological population (in this case, flour beetles). The model used
in this study has been successful in other contexts as well. For example, model
predicted unstable equilibria, together with their stable manifolds, accurately
describe and explain certain transient behavior observed in laboratory data
[117]. In section 1.4.3 a modification of the model is used to explain an unex-
pected "resonance" phenomenon observed in experiments involving periodically
fluctuating habitats (and to predict some unintuitive dynamics that were sub-
sequently confirmed by means of further experiments) [262].
1.4.1 Model parameterization and validation. One way to obtain a nu-
merical value for a model parameter is directly from experimental measurements
for that parameter. For example, if a constant survival fraction tlt appears in
a matrix model arid if one has census counts of both living and dead i class
individuals after one time unit, then one can estimate this fraction from this
census data,. Procedures for obtaining parameter estimates in this way, from
"life history" data, are discussed for linear matrix models in [55].
Another way to obtain parameter values in a model is to control that param-
eter by environmental or physical manipulations, assuming this can be done.
Often (if not usually) neither of these options is available for some or all
of the model parameters. In this case, one can attempt to obtain parameter
estimates from observed time series data, which we assume are observed values
of the state variables Xi(t) in a matrix model at a finite number of censu
48 CHAPTER 1

times t £ /[O,^]. Given parameter values the matrix equation provides, at each
time t, a (deterministic) prediction for the data at the subsequent time t + 1.
One is interested, of course, in the differences between the model prediction
and the actual observed data at the subsequent time t + 1. These "residuals"
can be used both to obtain parameter estimates (by choosing parameter values
that minimize the residuals) and for model performance (by a study of their
statistical properties).
However, biological data tend to be "noisy," i.e., to have "random" fluctua-
tions not accounted for in the deterministic model. In order to relate data to
a model adequately one needs to account for this stochasticity. This requires a
"stochastic model" in which appropriate random variables have been included
which reflect the type of noise present and its effect on the components of the
model. The resulting stochastic model (i.e., the deterministic model "skeleton"
and the hypothesized noise structure) provides a "likelihood function" for one
step predictions. The stochastic model can be viewed as a "testable hypothesis"
and its one-step residuals provide a means for its parameterization and model
validation.
Although this monograph deals only with deterministic models, we will illus-
trate these general remarks with some simple examples in order to provide an
understanding of the case study in section 1.4.2. For more details of these issues
see [132], [116].
Consider a one-dimensional (m = I) map

with p parameters 6^. For given values of (9, this map makes the deterministic
prediction that, given a population of size x ( t ) at time t, the total population size
at time t + 1 will be f ( x ( t ) , 9\,... , 6P). In general, the actual total population
size at t + 1 will deviate from this prediction in an unpredictable way and we
wish to modify this map to reflect this stochasticity.
Ecologists draw a distinction between two broad types of stochasticity. "En-
vironmental" stochasticity affect many individuals in the population and are
usually due to extrinsic factors. "Demographic" stochasticity involves chance
variations among individuals. Environmental stochasticity, which typically pre-
dominates at high population levels, is additive on the logarithmic scale [131].
As an example, consider noise added on the logarithmic scale to the following
model (1.48):

Here Et is a normal random variable with mean 0 and variance v. We assume

that EQ, E j _ , . . . are uncorrelated. A stochastic model of this form is a type of
multivariate, nonlinear, autoregressive model [403]. The nonlinear deterministic
skeleton (1.48) is preserved on the logarithmic scale for the expectations of the
random variable x(t] in the sense that
 DISCRETE MODELS 49

Suppose that a time series of q + 1 data points

is given. A "likelihood function" L gives the probability that the observed

data would result from the proposed stochastic mechanism relative to all other
possible outcomes [132]. The data yt is a realization of the random variable
x ( t ) . On the log scale, Wt = Iny^ is a realization of the random variable \n.x(i).
The likelihood function L is

where p(wt \ wt-\} is the joint probability distribution function (pdf) that wt oc-
curs given that wt-i occurs. This is a normal pdf with mean \nf(yt-\,9i,... ,9P)
and variance v. Thus.

and

The maximum likelihood parameter estimates are those values of the parame-
ters ( 9 i , . . . ,9p,v that maximize L ( 9 \ < . . . , 9P, v), or equivalently that maximize
1 ( 9 1 , . . . , 9 p , v ] = l n ( L ( # i , . . . , 9 p , v } } . A calculation shows

where

are the "log-residuals." The critical points (9\,... , 9 p , v ] of / are zeros of the
derivatives

i.e.. are roots of the uncoupled equations

50 CHAPTER 1

The first equation represents p equations for the p parameters ( ? i , . . . , Op whose

solution gives the maximum likelihood estimates for 0\,... ,0P. The second
equation then gives the maximum likelihood estimate for the variance v in the
stochastic version of the model. In general, the equations (1.50) must be solved
numerically.
As an illustration consider the following stochastic version of the Ricker equa-
tion:

Here there are two parameters, #1 = b and 62 = c, in the deterministic skeleton.

For this model the equations (1.50), namely,

are linear in In 6 and c with solutions given by

where

From (1.51)

where

As an illustration of these formulas, we generate "simulation data" yt from

the deterministic (v = 0) Ricker map with b = 10, c = 1/100 and a;(0) = 230
and q = 60. Using the resulting x ( t ) for the data yt in the formulas above
 DISCRETE MODELS 51

FIG. 1.12. Adult flour beetle numbers in 20g of flour, shown plotted against time, yield
the parameter estimates b = 1.631745, c = 1.338939 x 10~3, v = 6.658133 X 10~3 for the
stochastic Richer model (1.52). The log-residuals (1.53), computed from these, data and pa-
rameter estimates, yield the histogram on the right. The. log-residuals have mean 0.000000
and variance 6.770980 x 10~3.

we obtain (to seven significant digits) the numbers s\ = 1.390557 x 104, 53 =

3.911992xl06, s3 = -9.005503xlCT1, s4 = -7.101169xl03, s5 = 6.893810X101
and b = e2-3025851 = 1.000000 x 101, c = 1.000000 x 10~2, v = 0.000000. We
find that the maximum likelihood parameter estimates accurately recover the
correct parameters (seven significant digits).
More interestingly we can apply the parameter estimation formulas above
using population census data and see how well the stochastic Ricker model (1.52)
accounts for that data. In Fig. 1.12 we show a time series plot of adult flour
beetle numbers, one of several replicates in a laboratory experiment in which
each replicate was grown in a separate bottle containing 20g of flour. The adult
beetles were counted at two week intervals for 120 weeks. Thus, q = 60 and there
are 61 data points starting with 100 adults at time t = 0. Using this data in the
formulas above, we obtain the maximum likelihood parameter estimates for b,
c, and v given in Fig. 1.12 for the stochastic Ricker model (1.52). How well does
this parameterized model do in accounting for the data? The parameterized
Ricker model produces, from each data point, a one-step prediction for the next
data point. The stochastic model, as a hypothesis to be tested, asserts that
the differences between these one-step predictions and the actual data points
(the "residuals'') have certain statistical properties. Statistical tests can be
performed to quantify whether the residuals do indeed have these properties
(to a certain level of confidence). The stochastic Ricker model (1.52) predicts
that on the log scale the residuals will be normally distributed, i.e., that the
52 CHAPTER 1

FIG. 1.13. Beetle numbers from a replicate culture taken from the same experiment that
produced the beetle numbers in Fig. 1.12 are shown. The histogram is that of the log-residuals
computed from these new data and the parameter estimates calculated from the data in
Fig. 1.12. The log-residuals have mean -3.508827 x 10~2 and variance 5.715805 x 10~3.

log-residuals

will be normally distributed with mean 0 and variance v. For the data and
the parameter values in Fig. 1.12 a histogram of these residuals is also plotted.
One test of the success of this parameterized model would be to determine
whether these residuals are normally distributed with mean 0 and variance v —
6.658133 x 10~3. Normality tests can be performed to see how well the residuals
fit this description. Other statistical tests could be carried out as well. (For
example, the stochastic Ricker model was built under the assumption that the
random variables Et were uncorrelated in time and hence autocorrelation tests
can be performed, on the logarithmic scale, to test this hypothesis.) In this
example we will be content with only a visual inspection histogram of log-
residuals as shown in Fig. 1.12, which we note has an accurate mean and variance
and does have a strong appearance of normality, although there is some skewness
to the left.
Suppose we accept that the stochastic Ricker model successfully describes
data in Fig. 1.12. A further, and perhaps stronger, test of the model would be
to ask that it also describe other data gathered under the same experimental
conditions. That is, how well does the parameterized model "predict" data that
were not used in its parameterization? A plot, of data from a replicate culture
from the same experiment that yielded the data in Fig. 1.12 appears in Fig.
1.13. Using the parameter estimates obtained from the data in Fig. 1.12, we can
calculate log-residuals (1.53) between the Ricker model one-step predictions and
the data in Fig. 1.13. The results appear in the histogram in Fig. 1.13. Again
 DISCRETE MODELS 53

FIG. 1.14. For the data plotted on the left the maximum likelihood estimates for the
stochastic Richer model (1.52) are b = 4.815900, c = 2.695051 x 10~2, and v = 5.362000 x
10"1. The log-residuals, whose, histogram is shown on the right, have mean 0.000026 and
variance 5.499105 x 10"1.

we find a normal looking distribution of log-residuals (with some skewness to

the left) with accurately predicted mean and variance.
If the "validation" and "prediction" fits in Figs. 1.12 and 1.13 are judged
acceptable, then the stochastic Ricker model (1.52), with parameter estimates
in Fig. 1.12, is judged adequate to describe the dynamics of the populations in
this experiment. Since the parameter estimates in Fig. 1.12 for the deterministic
Ricker model yield a stable equilibrium, we would conclude that the populations
in this experiment have a "noisy" equilibrium.
In Figs. 1.14 and 1.15 the procedure above is repeated for two more replicate
data sets taken from a different experiment. The data arc considerably more
oscillatory than those in Figs. 1.12 and 1.13. Visually, in this case, the log-
residuals are perhaps less convincingly normally distributed. If, however, these
validation and prediction fits were judged to be acceptable for these data, then
the conclusion would be that the populations in this experiment are adequately
described by the stochastic Ricker model (1.52). Since the parameter estimates
in Fig. 1.13 for the deterministic Ricker model yield a stable equilibrium, we
would conclude that the populations in this experiment also have a "noisy"
equilibrium.
A third data set appears in Fig. 1.16 together with the parameter estimates
for the stochastic Ricker model and a histogram of the resulting log-residuals.
Here the distribution of log-residuals clearly is not normally distributed, and
we reject the fit of this model to the data (which asserts that this data is also
a very noisy equilibrium). This means either the deterministic Ricker model or
the type of noise added to it in (1.52) are inadequate to account for the data in
this experiment.
54 CHAPTER 1

FlG. 1.15. Beetle numbers from a replicate culture taken from the same experiment that
produced the beetle numbers in Fig. 1.14 are shown plotted on the left. The histogram is that
of log-residuals computed from these data and the parameter estimates calculated from the
data in Fig. 1.14. The log-residuals have mean 5.069767 x 10~2 and variance 6.178831 x 10"1.

In the discussion above, the one-dimensional Ricker model was used. The
procedures outlined can be extended to the case of multivariate data sets and
higher dimensional models [132]. In the multivariate case there can be covari-
ances among the state variables as well as variances. The likelihood function is
more complicated in that case, of course, and its maximization must be done
numerically with the aid of a computer in order to obtain parameter estimates.
It is interesting to note that in the one-dimensional case treated above maximiz-
ing the likelihood function is equivalent to minimizing the residuals in the least
squares sense; see (1.49). This is not in general the case for the multivariate
data (unless all covariances are 0) and parameter estimates calculated by least
squares are in general different. Least square estimates relax the distributional
assumptions concerning the noise in the model, but should give similar results
if the log-normal assumption in the model is valid. In this way, least square es-
timates provide a cross check on the validity of the stochastic model. However,
the statistical theory for multivariate least squares estimates does not seem to
have been thoroughly worked out yet.

1.4.2 Bifurcating beetles. The data used in Figs. 1.12-1.16 were taken
from laboratory experiments using flour beetles (Tribolium). In these experi-
ments, beetles were grown in 20g of flour at a constant incubator temperature
and humidity. Every two weeks the flour was shifted and the counted animals
returned to a fresh 20g of flour. The data shown in Figs. 1.12-1.16 are numbers
of adult beetles taken from three different experiments (in which adult beetle
recruitment and morta, !,y were different). While the stochastic Ricker model
seemed to describe well some of these data sets, it did less well on others and,
in particular, did a very poor job on the data in Fig. 1.16. In order to construct
 DISCRETE MODELS 55

FIG. 1.16. Beetle numbers from a third experiment ore shown plotted on the left. On
the right appears a histogram of the log-residuals computed using the parameter estimates
b = 2.424700, c = 2.481400 x 10~2, and v = 8.900700 obtained from the data. The log-
residuals have mean 4.147428 X 10~2 and variance 9.055077.

a better model we should take into account what are considered the important
biological mechanisms that effect the dynamics of these beetles. Under the con-
ditions of the experiments (no resource limitation, constant habitat, no other
species present, etc.) these mechanisms, it turns out, involve interactions be-
tween the individuals of different life cycle stages. The species of flour beetles
used in the experiments have a life cycle that includes larval arid pupal periods
lasting approximately two weeks each (under the conditions of the experiment).
Specifically, adult and larvae are cannibalistic, both consuming eggs (affecting
larval recruitment). Another significant interaction is adult cannibalism of pu-
pae. Therefore a structured population model is called for that includes at least
larval, pupal, and adult stages. (Larvae and pupae were, in fact, also counted
in the experiment.)
Since the census period is two weeks, we choose the unit of time to be two
weeks and let L(t). P(t), and A(t) denote the numbers of larvae, pupae, and
adults at time £, respectively. Given that the larval and pupal periods are both
approximately two weeks in duration, a Leslie matrix

could serve as a projection matrix in the absence of any nonlinear (density)

effects. Here b is the larval recruitment rate per adult per unit time (two weeks)
and the ^/,'s denote the death rates for the three life-cycle stages. It turns out
that for this experiment pupal mortality (in the absence of adult cannibalism)
56 CHAPTER 1

is negligibly small so we set np = 0. Cannibalistic acts are random occurrences

as the organisms move through the container of flour. Therefore we model the
effects of cannibalism by means of exponential discounting terms of Ricker type
(see section 1.1.3). This results in a nonlinear projection matrix of the form

with

This is the projection matrix for the LPA model equations (1.43), i.e.,

The coefficients cei, cea, and cpa are referred to as the "cannibalism coefficients."
The goal is first to parameterize and validate this deterministic model, using
a stochastic version of the model, and then to carry out laboratory experiments
based on the dynamics predicted by this model if one or more of its parameters is
changed (i.e., manipulated by the experimenter). In particular, it is attempted
to document by means of the experiments any bifurcations in the asymptotic
dynamics predicted by this deterministic model. The data from the experiments
can be analyzed by means of a "prediction fit" (i.e., utilizing the a priori param-
eter estimates used to design the experiment, not parameter estimates obtained
from the resulting experimental data) or, since the experiments include replicate
cultures, the data can be divided and different sets used for validation and pre-
diction fits of the model (as in the example of section 1.4.1). We briefly describe
the results of such experiments as reported in [132], [72], [73], [116], [133].
To carry out the program described above it is necessary to understand as
much as possible about the asymptotic dynamics of the LPA model (1.54). In
section 1.3 it was shown that for n < 1 the origin is globally attracting in R^_
and for n > 1 the model system is uniformly persistent and there exists a unique
positive equilibrium which is (locally) stable for n sufficiently close to 1. (For a
global stability result for the positive equilibrium see [271].) It was pointed out
in section 1.3 that the positive equilibria of (1.54) are not stable for all parameter
values, however. While there are no general results available that locate stability
boundaries for the model parameters in (1.54), numerical simulations, such as
those in section 1.3, show that destabilization and bifurcation to nonequilibrium
attractors can occur as parameters are varied.
One special case has been thoroughly studied analytically, however. In [132]
the destabilization boundaries are analytically found for the case when larval
cannibalism of eggs is ignored ce; = 0. It is shown that positive equilibria can
destabilize as a parameter varies either because an eigenvalue of the Jacobian
passes through —1 or because a complex conjugate pair of eigenvalues passes
 DISCRETE MODELS 57

out of the complex unit circle. In the first case, a familiar period doubling
bifurcation to a 2-cycle occurs while in the second case, a bifurcation to an
invariant loop in Ft+ (and quasi-periodic oscillations) occurs. Which type of
destabilization and bifurcation occurs depends on the values of the other model
parameters. Extensive numerical studies show this is also the case when cei ^ 0.
Moreover, these studies show that further destabilizations and bifurcations often
occur and chaotic attractors can arise [72], [73], [133].
The richness of dynamics associated with the LPA model (1-54) and the ready
availability of flour beetles for laboratory experiments present an opportunity to
coordinate theoretical modeling and predictions with experimental data. More-
over, this system offers an opportunity to test the descriptive and predictive
capabilities of simple mathematical models (such as (1.54)) for the dynamics of
a real biological population and thereby to study the role of nonlinear theory
in population dynamics. To connect the LPA model (1.54) to time series data
in the manner described in section 1.4.1 a stochastic version of the model is
needed.
In [132] the stochastic model

was parameterized and validated using a historical data set. Here the Ei(t] are
normal random variables (uncorrelated through time) with mean 0 arid variances
and covariances to be estimated as model parameters from time series data.
Based on that preliminary study and its parameter estimates, an experiment
was devised in which the adult death rate /.IQ was manipulated in such a way that
the dynamics were predicted by the deterministic LPA model (1.54) to undergo
a specific sequence of bifurcations. The experiment, which included controls
and replicates, lasted 36 weeks. The data from half of the replicates were used
for a parameterization (using both maximum likelihood and conditioned least
squares methods) and model fit analysis, and the data from remaining replicates
were used for a model prediction analysis.
The results are reported in [72] and [133], where details of the laboratory
experiments and the parameterization and validation of the model can be found.
The maximum likelihood parameter estimates (with 95% confidence intervals)
for one of the two different genetic strains of Triboliwni castaneum (Hcrbst)
used in this experiment appear in Table 1.1. (An estimate for p,a was obtained
from a direct count of dead individuals in the control cultures and found to be
4.210 x Ifr 3 with a 95% confidence interval of (3.2 x 1(T3,5.2 x 1(T3).)
The matrix of variances and covariances of the EI was estimated for the
manipulated treatments to be
58 CHAPTER 1

TABLE 1.1
| Parameter ML estimate Confidence interval
b 7.876 (5.8,10.7)
M; 1.613 x KT1 1.0 x 10~1,2.2x KT1
cea 1.114 x 1(T2 1.0 x10- 2
,1.
2 x 10~
2

Cel 1.385 x 10~2 1.2 x1 0 -

2
,1.5x 10
2
-

Cpa 4.348 x 10~3 3.6 x 10-3,4.8x 10~3

With these parameter values the LPA model equations (1.54) predict a 2-
cycle bifurcation as /xa is increased through a critical value of approximately
0.1 as shown in the bifurcation diagram of Fig. 1.17. The data obtained from
experiments performed at adult death rates /ia = 0.04, 0.27, and 0.50 are shown
in Fig. 1.18 plotted in three-dimensional phase space together with the model
predicted attractors [133]. Clearly these data support the predicted 2-cycle
bifurcation.
Experiments of the type described above were repeated, for a longer length
of time, in which both adult mortality (held at na = 0.96) and adult recruit-
ment (i.e., Cpa) were manipulated. The model predicted bifurcations as cpa is
increased (with fj,a held at 0.96 and other parameter values taken from Table
1.1) are seen in Fig. 1.19 to be considerably more complicated than those in
Fig. 1.17. A stable equilibrium exists at cpa = 0, but a discrete Hopf (invariant
loop) bifurcation occurs at a very small positive value of cpa. For example, an
invariant loop occurs at cpa = 0.05. As cpa is further increased invariant loops
are interspersed with "windows" where periodic attractors of high period occur.
Ultimately, chaotic attractors appear. For example at cva = 0.35 the attractor
has Liapunov exponent 0.1146 and box dimension approximately 1.25. For large
values of cpa there is a distinctive 3-cycle. Eighty weeks of data (with transients
left out) from four of the eight treatments reported in [73] are plotted in Fig.
1.20 together with the corresponding model predicted attractor. In this case,
none of the data is used for parameter estimation (so that the results deal with
the prediction capabilities of the previously fitted and validated model)! The
model attractors are computed from the parameter estimates in Table 1.1 ob-
tained from the earlier experiment. Again one sees a convincingly close match
between the experimental data and the deterministic model predicted sequence
of bifurcations. These results are the first convincing evidence of the possibility
of chaotic dynamics in a biological population. See [62], [185], [255], [360].
1.4.3 Periodic habitats. Autonomous models assume a constant environ-
ment. If environmental fluctuations cause a model's parameters to vary in time,
then the model becomes nonautonomous. For example, if fertility and mortal-
ity rates are seasonal, then it might be reasonable to consider a periodically
forced model in which fertility and mortality coefficients are periodic functions
of time. Since regular environmental fluctuations (seasonal, daily, etc.) are
 DISCRETE MODELS 59

FIG. 1.17. A bifurcation diagram, plotting total population size p against the adult death
rate. p,a, for the flour beetle model (1.54) using parameter values from Table 1.1 shows a 2-
cycle bifurcation. Arrows indicate where laboratory experiments were performed: /ia = 0.04,
0.27, 0.50.

FIG. 1.18. Open circles show data (with transients removed) taken from laboratory ex-
periments using flour beetles with adult death rates p,a manipulated to the indicated values.
Solid circles show the model predicted attractor.
60 CHAPTER 1

FIG. 1.19. The bifurcation diagram of the flour beetle model (1.54) using parameter values
from Table 1.1 with na = 0.96. Total population size is plotted against cpa. Arrows indicate
where laboratory experiments were performed.

FIG. 1.20. Open circles show data (with transients removed) taken from experiments in
which Cpa was manipulated. Also plotted are the model (1.54) predicted attractors, which
are: an equilibrium for cpa = 0.00, an invariant loop for cpa — 0.05, a chaotic attractor for
Cpa = 0.35, and a 3-cycle for Cpa = 1.0.
 DISCRETE MODELS 61

certainly important for the dynamics of many species, one might expect to find
more periodically forced models in the literature. While there is a small body of
literature on periodically forced models in population dynamics and ecology, the
vast majority of models are autonomous. (This is no doubt due to the fact that
autonomous models are more tractable than periodically forced models.) An in-
teresting general question is, how many of the fundamental tenets of theoretical
ecology remain valid in a periodically (or otherwise) fluctuating environment?
See the references cited in [232] for literature on periodically forced models.
An opportunity to study a population in a periodically fluctuating environ-
ment is provided by the results of some laboratory experiments with flour beetles
performed by Jillson [252]. In these experiments the same experimental protocol
described above was used, except that the beetles were not always returned to
20g of flour. Instead the flour amount was changed at each census time (2 weeks)
in a periodic manner so that the average amount was 20g. One experiment, for
example, involved a flour amount schedule 32g-8g-32g-8g- • • with period 2, i.e..
a period of four weeks. Other experiments involved longer periods. All were
carried out with replicates and constant habitat controls.
Given the success of the autonomous model (1.54) in a habitat consisting of
a constant amount of flour, it seems reasonable to hope that an appropriate,
periodically forced, modification of this model might accurately describe the
data in Jillson's experiments, explain any interesting and unusual aspects in
this data, and even predict dynamical phenomena that might be successfully
observed in new flour beetle experiments. In fact, these expectations have been
remarkably well fulfilled.
The following periodically forced version model (1.54) was proposed and stud-
ied in [232], [74] as a model for flour beetle dynamics in a fluctuating amount
of flour with period 2:

This model is built upon the assumption that the cannibalism coefficients (i.e.,
the probability of a cannibalistic event) are inversely proportional to the amount
of flour. This assumption has been validated by laboratory experiments [74],
For example we write cei — Kei/V, where V is the volume occupied by 20g of
flour under laboratory conditions. If V is the average flour volume, then in
the autonomous model CKI = Kei/V. If, however, the flour volume fluctuates
with an average V, a period of 2 and a relative amplitude of a (0 < a < 1),
then V = V (1 + a( — 1)*) and the larval/egg cannibalism coefficient becomes
K,ei/V (1 + a( — \Y) = Cfij (1 + a( — 1)*). A similar derivation, when applied to
the other cannibalism coefficients, results in the periodic LPA model (1.55). The
autonomous model (1.54) corresponds to a = 0.
An interesting consequence of periodically fluctuating flour amounts was re-
62 CHAPTER 1

ported in [252]. For the experiment in which the period was 2, the total popula-
tion size was significantly larger than that found in the constant habitat control
with the same average amount of flour (20g). This surprising result is contrary
to an often held theoretical tenet asserting deleterious effects of habitat fluc-
tuations. This tenet is based upon studies of periodically forced, unstructured
models, such as the classical logistic equation [309], [331]. (See, however, [92],
[364].) Can the periodic LPA model explain this unexpected result?
The periodic LPA model (1.55) is analytically studied in [232] by means of
bifurcation and perturbation techniques. (Also see [107] and [227].) A global
continuum of positive 2-cycles bifurcates from the extinction equilibrium at the
critical value bcr — /j a / (1 — /ij) of the parameter b. This is the same bifurca-
tion point as that for the continuum of positive equilibria in the autonomous
LPA model (1.54). The reason for this is that the Jacobians of these two mod-
els evaluated at the extinction equilibrium are identical. The continuum of
positive 2-cycles bifurcates to the right and is stable. These bifurcation re-
sults can be proved using the same general techniques used for the autonomous
model. Liapunov/Schmidt expansions can be utilized to compare the (average)
total population sizes for the periodic and the autonomous LPA models near
the bifurcation point. Let (La,Pa,Aa) denote the averages of a 2-cycle so-
lution (La(t),Pa(t),Aa(t)) of the periodic LPA model (1.55) with amplitude
a < 1, i.e., La = £,l=0La(t)/2, etc. For a small parameter ea > 0 the Lia-
punov/Schmidt expansion near the bifurcation point yields [232]

and

We wish to compare the averages of La(t), Pa(t), and Aa(t) for a = 0 to those
for a > 0 for the same parameter values. To obtain the same value of b to first
order, we take ea = SQ (l - a2) and find that

Thus, for small values of b near bcr we have I/o > La, PQ > Pa, and AQ > Aa.
This negative effect of habitat fluctuation, in which the average total population
size pa = La + Pa +Aa decreases in a periodic habitat, is the opposite of
Jillson's findings. It is shown analytically in [232] that a positive effect pa > p0
can occur for the periodic LPA model (1.55) for sufficiently large values of
6 > bcr, at least for small relative amplitudes a. This is done by calculating
 DISCRETE MODELS 63

FIG. 1.21. The left graph shows the stable equilibria of the autonomous LPA model (solid
line) and the stable 2-cycles of the periodic LPA model (dotted lines) plotted against b. The
remaining parameters are given by (1.56), and the relative amplitude is a = 0.6 (as in Jillson's
experiment). The averages of the 2-cycles (dashed line) exceed the stable equilibria for b greater
than approximately 0.8 > 0.293 = bcr- In the right graph the stable 2-cycles of the periodic
LPA model with parameter estimates (1.56) are plotted against the relative amplitude a. For
a > 0 less than approximately 0.45 there exist two stable 2-cycles.

lower order terms in an a expansion of the 2-cycle solution near an equilibrium

for a = 0. This result shows that the periodic LPA model has at least the
theoretical capability of predicting positive effects of habitat fluctuations. See
Fig. 1.21. However, parameter estimates computed from the control treatments
in Jillson's experiment yield the values

For these parameter values the autonomous LPA model (1.54) has an unstable
equilibrium and a stable 2-cycle. Thus, the 2-cycles for small amplitudes a
obtained in [232] are unstable and cannot explain Jillson's result.
In [74] the periodic LPA model (1.55) is used to explain the positive effect
observed in Jillson's experiment. When, as for the parameter estimates (1.56),
the autonomous LPA model has a stable 2-cycle, perturbation methods show
that stable 2-cycles for the periodic LPA model exist for (at least small) relative
amplitudes a > 0. In fact, there are two stable 2-cycles for small a correspond-
ing to the two phase shifted 2-cycles of the autonomous LPA model. For the
periodic LPA model these two stable 2-cycles are not phase shifts of one an-
other and they have different amplitudes and, importantly, different averages.
64 CHAPTER 1

In Fig. 1.21 the stable 2-cycles are plotted as a function of a for the periodic
LPA model with the parameter estimates (1.56). It is seen from these graphs
that one 2-cycle has an average total population size greater than that of the
2-cycle in the constant habitat (a = 0) while the other has a smaller average
total population size. Thus, the periodic LPA model, parameterized to the Jill-
son experiment, predicts two stable 2-cycles. one implying a positive effect and
the other implying a negative effect. However, as can be seen in Fig. 1.21, as
a is increased the stable 2-cycle with the smaller average abruptly disappears,
leaving only the stable 2-cycle with the larger average. (It is conjectured that
a saddle-node bifurcation causes the disappearance of the smaller average 2-
cycle as it "collides" with the unstable 2-cycle that emanates from the unstable
equilibrium.) For a = 0.6, as in Jillson's experiment, the periodic LPA model
predicts only the stable 2-cycle with the larger average, and in this way the
model explains the positive effect observed by Jillson.
An interesting result of the above analysis of the periodic LPA model is the
prediction of two stable attractors for small relative amplitudes a > 0. This is
unexpected result has been recently confirmed by laboratory experiments [75].

1.5 Multispecies interactions

To this point we have considered structured models for the dynamics of only
a single population or species. Matrix models for the interaction of several
structured species can be readily constructed by assuming that some entries in
the fertility and/or transition matrices of each species depend upon (at least
some) class densities of other species. In multispecies models the structuring
variables do not have to be of the same kind nor do the numbers of categories
used have to be the same for all species. Coupled matrix equations of this
kind can be used to model interactions, such as predation and competition,
that are specific to structuring classes. For example, only certain classes of
individuals might be vulnerable to predation and among those class predation
rates might vary considerably (e.g., younger and older individuals are often much
more susceptible to predation). Similarly, only certain classes of individuals
(e.g., juveniles) might experience competitive pressure from other species for
resources. Individuals from two species may compete when small in size (young
in age), but not when they are larger (older) [435]. Indeed, structured models
can account for interactions that defy the classical typecasting for mnltispecies
interactions. In [435] many examples are given in which species exhibit mixed
competition/predation interactions in which individuals of a species, when small
(young), compete with those of another species on whom they eventually will
prey when larger (older). Such class specific interactions can profoundly affect
the population level dynamics of a community of species.
In this section coupled matrix equations will be considered from a bifurcation
theory point of view. The approach will be such that the methods and gen-
eral theorems used for the single species models are applicable to multispecies
models.
Consider a community of one or more "species" whose state variables are
 DISCRETE MODELS 65

contained in a column matrix y = y(t) e /?,™2. A ''species'' here can be a

biological species or a nonbiologieal quantity such as a food resource. Assume
that this community has an equilibrium ye > 0. To this "resident" community
we add a (possibly) structured species with state variables x = x(i) £ -R™ 1 ,
mi > li whose dynamics are described by a matrix model of the type considered
in this chapter with a projection matrix P. The species x and the species in the
community y are assumed to interact so that P = P(x, y). Thus, we are led to
consider a system of the form

Here the community dynamics in the absence of the species x are governed by
the equation y(t + 1) = g ( 0 , y ( t ) ) and g(0,yc) = ye- If the resident community
consists of one or more structured populations, then the dynamics g are given
by a projection matrix so that in place of (1.57) we have

This in turn could be written as a single matrix equation by combining the

vectors x and y and creating a block diagonal projection matrix with P and Q
along the diagonal.

1.5.1 Some equilibrium theory. The system (1.57) has the ''trivial"
equilibrium y - ye > 0, x = 0. We will refer to this as the "extinction"
equilibrium, meaning by this that x is absent in this equilibrium state. Our
approach to the existence of positive equilibria, in which x is therefore not
extinct, is to establish their bifurcation from the extinction equilibrium as a
model parameter, related to the "inherent" dynamics of species x, is varied.
This will be done in a manner analogous to that used for the single species case
in this chapter. Although a global result analogous to Theorem 1.2.4 by an
extension of the proof of Theorem B.I.I (see Appendix B.I) is possible, we will
present only simpler local bifurcation results.
P(().yK) is called the inherent projection matrix for x at ye. As in section 1.2
we assume that a bifurcation parameter A of interest appears linearly in this
matrix. Thus, we consider the system

where for some integer k £ 7[1, +oc)

66 CHAPTER 1

where fi is an open set containing R™, m = mi + 7712, and

The equilibrium equations associated with (1.58) are

By assumption y = ye is an equilibrium of

We say the equilibrium ye is nondegenerate if 1 is not an eigenvalue of the

Jacobian Jyg(Q, ye) of g with respect to y evaluated at x = 0, y = ye > 0.
Note that if all eigenvalues of Jyg satisfy |£| < 1, and hence ye is (locally
asymptotically) stable in the absence of x, then ye is nondegenerate.
If ye is nondegenerate, then by the implicit function theorem equation (1.59b)
can be solved for y = y ( x ) . Specifically, there exists such a solution y 6
Ck+1(N(Q),Rm*), y(0) = 0, where N(0) is an open neighborhood of 0 in
.R™1. For x e N(0) equation (1.59a) becomes x = P(X,x,y(x))x or x =
n

(A + XB) x + r (A, x), where r(A, x) = O(\x\ ) near x = 0. If / — A is invert-

ible, then this equation can be written in the form x — \Lx + h(X,x), where
L = (I - A)~1B and h(X,x) = (I - A)~lr(X,x) to which Theorem B.2.1 (see
Appendix B.2) can be applied. The result is that Theorems 1.2.5 and 1.2.6 are
valid concerning the bifurcation of solutions x from 0 at a critical value AO of
A + XB. Coupling these results with the solution y = y(x) we get the following
local bifurcation result for the equilibria of the multispecies system (1.58).
THEOREM 1.5.1. Suppose that A0 is a critical value of A + XB, I - A is
invertible, and WTBv / 0. Suppose that ye > 0 is a nondegenerate equilibrium
of (1.60). Then a branch of positive equilibria x and y of the multispecies matrix
model (1.58) bifurcates from the extinction equilibrium x = 0 and y = ye > 0.
This branch has the form

where for some £Q > 0

Moreover, ifk>2, then

 DISCRETE MODELS 67

The formula for ?/(0) can be obtained by substituting the expansions for A,
x, and y in Theorem 1.5.1 into equation x = P (A, x, y(x}) x, differentiating the
twice with respect to £ and setting e = 0 in the result. This yields the equation

for z"(0). Since the null space of / — (A + Ao-B) is spanned by v, it is necessary

for the solution of this equation that the right-hand side be orthogonal to w.
We turn now to the stability of the extinction equilibrium and of the positive
equilibria in Theorem 1.5.1. The Jacobian J ( X , x , y ] of (1.58) evaluated at the
extinction equilibrium x — 0, y = ye is the block diagonal matrix

The eigenvalues of J(0, ye) are those of the individual blocks A + \B and
J y g ( Q , y e ) . Thus, if Jyg(0,ye) has an eigenvalue with |C| > 1, then the ex-
tinction equilibrium is unstable for all A.
Let us consider the problem of a species x trying to "invade" a community
which is at stable equilibrium. By "invade" we mean that there is a positive
stable equilibrium. Since we are considering only local bifurcations of positive
equilibria from the extinction state (hence small amplitude x equilibrium states)
and only local stability, the ecological problem is that of population x attempt-
ing to invade a community at low population numbers. The community y is
supposed stable in the absence of x, so it is assumed that all eigenvalues of
Jyg(0,ye) satisfy |£| < 1. The stability of the extinction equilibrium is there-
fore determined by the magnitude of the eigenvalues of block A + \B. Theorem
1.2.3 applies and tells us that the extinction equilibrium loses stability as either
A increases (wTBv > 0) or decreases (wrBv < 0) through the critical value AQ.
We turn now to the local stability of the positive equilibria given in Theorem
1.5.1. Based on the results for single species matrix models we expect that
stability is related to the direction of bifurcation and hence to the sign of Aj.
An analog of Lemma 1.2.2 can be proved by a straightforward, but tedious,
calculation. Thus, if £ = £(e) is the dominant eigenvalue of the Jacobian of
J ( X , x , y ) evaluated at the positive equilibria in Theorem 1.5.1, then it turns
out that

for a positive constant K > 0.

THEOREM 1.5.2. Suppose, in addition to the assumptions of Theorem 1.5.1,
that all eigenvalues of Jyg(0.ye} satisfy |("[ < 1 (so ye > 0 is stable). The
extinction equilibrium x = 0, y = ye of the rnultispecies matrix model (1.58)
loses stability as A increases through AQ, ifwTBv > 0, or as A decreases through
A 0 ,i f w T B v < 0 .
Suppose further that AQ is a strictly dominant critical value, k > 2 and X\ ^ 0.
If wrBv > 0, then the bifurcation described in Theorem 1.5.1 is stable if it is to
68 CHAPTER 1

the right (Ai > 0) and unstable if it is to the left (Ai < 0). IfwTBv < 0, then
the bifurcation is stable if it is to the left and unstable if it is to the right.
One choice of the parameter A is the inherent net reproductive number n of
species x at the extinction equilibrium (0,ye). Write

where T(x,y) is the transition matrix and n®(x,y) is the fertility matrix for
species x. Here <E> is normalized so that n is the inherent net reproductive
number for the species x when y is at the equilibrium ye; i.e., 1 is the dominant
eigenvalue of T(0,j/ e ) + ® ( Q , y e ) . In this case B - $(0,j/ e ) > 0 (^ 0) and
wTBv > 0 so that the extinction equilibrium loses stability as n is increased
through 1 and the bifurcation is stable if and only if it is to the right, i.e.,
WT [dij] v < 0.
As in the case of a single species model, the stability properties in a neighbor-
hood of the bifurcation point may change outside that neighborhood through
various types of bifurcations depending on how the eigenvalues of the Jacobian
leave the unit circle.
Remark. Theorems 1.5.1 and 1.5.2 remain valid as stated if A appears in
g = g ( X , x . y ) and if there exists an equilibrium ye > 0 independent of A (at
least for A near AQ). In this case ye nondegenerate means 1 is not an eigenvalue
of the Jacobian Jyg(^n, 0. ye).

1.5.2 Applications
Interactions of two structured species. Consider two interacting species
described by the coupled matrix equations

where

are weighted total population sizes of x(t) and y ( t ) . The projection matrices

are assumed, in addition to satisfying the conditions in (1.58), to be irre-

ducible and primitive and to have net reproductive numbers Uj(pi,p2) at equi-
librium (pi,pa) (the positive, strictly dominant, simple eigenvalues values of
Fi(p 1; p2)(/—Tj(pi,p2))~ 1 ). As in section 1.2.5, at equilibrium we haven i(pi,p2)
= 1 for both i = l,2 (see (1.32)).
Let P = Pi(p\,p-2). g = P2(pi,p2)y in (1.58). Assume that species y has
a stable equilibrium in the absence of species x. That is to say, we assume
 DISCRETE MODELS 69

that y(t + 1) = P2(Q,P2(t))y(t) has a stable positive equilibrium ye > 0 (i.e.,

the eigenvalues C of the Jacobian of the right-hand side evaluated at yK satisfy
|C| < 1). Thus, we view y as a "resident" species in stable equilibrium and x
as a species trying to "invade" at low density. Denote by p2 > 0 the weighted
total population size of the equilibrium ye, and write

PI (pi ,P2) = TI (pi, p2) + n$i (pi, p2),

where n is the inherent net reproductive number of species x at the extinction

equilibrium pi = 0, p2 = p% > 0 and $1 is the normalized fertility matrix (so 1
is the dominant eigenvalue of 71(0,^2) + n^i(Q,p^)). Use A = n in Theorems
1.5.1 and 1.5.2.
The direction of bifurcation at n = 1 can be related to the properties of the
net reproductive numbers n,; as follows. Assume that both species have density
self regulation in the sense tha.t

Since ni(pi,p2) = ni/i(pi.p2), where i/i is the net reproductive number of

Ti(pi,p2) + $i(pi,p2), at equilibrium we have

For the equilibria in Theorem 1.5.1 the bifurcation will be to the right (and
stable) if V i ( p i , p 2 ) < v\(^iP2) = 1 f°r (p\,Pz), P\ > Oi close to (Q,p2). Using
the implicit function theorem we can solve the second equation in (1.61) for
P2 = P2(pi}~ P2(0) = p% and substitute the answer into the first equation in
(1.61) to obtain

The bifurcation will be to the right (and stable) if the factor i/i(pi,p2(pi)) is
decreasing at pi — 0. i.e..

where the partial derivatives are evaluated at ( p i , p 2 ) ~ (0,p|) and n = I .

On the other hand, detM < 0 implies that the bifurcation is to the left (and
unstable). See Fig. 1.22.
We can define the two most fundamental types of ecological interactions using
the net reproductive numbers of the species in the following way:
70 CHAPTER 1

FIG. 1.22. The bifurcation scenarios for a two species interaction model are schemati-
cally represented by bifurcation diagrams on the left and phase portraits on the right. In the
bifurcation diagrams the distance from the positive equilibria (pi, P2) to the extinction equi-
librium (Ojpj) is plotted against the inherent net reproductive number n of species pi at the
equilibrium (0,p|).

In a predator-prey interaction, the invading species p\ is the predator and

the resident species P2 is the prey. In this case detM > 0 and the predator
can successfully invade (i.e., coexist in stable equilibrium with) the prey at low
densities provided the predator's inherent net reproductive number n at the prey
equilibrium (0,p|) exceeds 1. If. on the other hand, n < 1, then low population
densities of the predator will go extinct. See Fig. 1.22.
In a competition interaction detM = dplnidp2nz — dp2nidpln<2 can be of ei-
ther sign, depending upon the relative magnitude of the product dp^n\dp2n2 (as
a measure of the strength of mtfraspecific competition in the system) and the
product dp2nidpln-2 (as a measure of the strength of the mierspecific competi-
tion). In either case, low population densities of the invading species p\ will go
extinct if its inherent net reproductive number n at the resident competitor's
equilibrium (0,^) is less than 1. If n > 1, then species p\ can successfully "in-
vade" the resident competitor species P2 at low population densities in the sense
that low initial population densities of pi will not go extinct. If intraspecific
 DISCRETE MODELS 71

competition outweighs interspecific competition (i.e., det M > 0). then the two
species will coexist in stable equilibrium if n > 1. If interspecific competition
outweighs intraspecific competition (i.e., detAf < 0), then there is no stable
positive equilibrium near the bifurcation point and the asymptotic outcome
when n > 1 cannot in general be determined from a local bifurcation analysis.
In this case it is often true in competition models that all orbits tend to an
axis equilibrium (except perhaps for those on a stable manifold of an unstable
equilibrium) so that one species always goes extinct, in keeping with the cel-
ebrated competitive exclusion principle in ecology. However, in this case it is
also possible for two competing species coexist in a nonequilibriurn manner. See
section 3.1.1 for an example involving two competing size-structured species.
The analysis above, based upon the net reproductive rates n\ and 71%, bears
a striking similarity to the standard analysis of the famous Lotka-Volterra (un-
structured) differential equation model for two species competition. The matrix
M is analogous to the so-called community matrix in that classical theory. Un-
like the classical theory, however, the criteria above can be related to class
specific transition and fertility rates through the net reproductive numbers n\
and n-i-

A host—parasitoid interaction. Parasitoids often attack a specific larval in-

star of their host. This fact can be of use in selecting parasitoids for a biological
control program in which the host is a pest species. Given that a selection of
parasitoids is available, which larval stage is optimally attacked if the goal is to
reduce or eliminate the pest species? We will consider one of a variety of discrete
age-structured host-parasitoid matrix models that were derived and studied by
Barclay [22] with this question in mind.
Consider a multispecies system of the form (1.58) in which x is a vector of
adult parasitoid life cycle stages and y is a vector of host life-cycle stages. The
first m-2 — 1 components 2 / 1 , . . . ,ym,2-i are iiistars of nonparasitized larva den-
sities and ym.2 is the adult (pest) host density. In this model g ( x , y ) = Q(x,y)y,
where the host projection matrix is a nonlinear Leslie matrix of the form

The transition probability a < 1 between age classes is taken to be independent

of age. and the inherent per capita fertility rate is /. Note that adult fertility
is density dependent (on adult population density only), as is expressed by the
term cxp(— gym^)- Parasitoids of all stages are assumed to attack a single host
instar, namely, the instar of age o, which is modeled by the term exp(—qpi) lo-
cated in the (a + l) s( row and ath column of the matrix Q(x, y). In the absence
72 CHAPTER 1

of the parasitoids, the host population has a unique positive equilibrium

provided fam^~l > 1. Let yea = a0-'™"* ±\n(f am*-1) denote the ath component
of ye. This equilibrium is stable if all eigenvalues £ of the Jacobian Jyg(Q,ye)
satisfy |£| < 1. These eigenvalues turn out to be C = (1 - Int/cr™ 2 " 1 )) 1 /™ 2 and
hence ye > 0 is stable if

which we assume to hold.

The projection matrix of the parasitoid species has the form T(x, y) + F(x, y),
where the transition and fertility matrices are

The normalized fertility matrix and the inherent net reproductive number of
the parasitoid (at the host equilibrium ye) are

In this parasitoid model it is assumed that the hosts do not reach the adult
stage before mummification and emergence of the adult parasitoid. Prior to
emergence parasitoid mortality is assumed the same as that as the host, which
 DISCRETE MODELS 73

IK a for k time units, k < 6 (the developmental time of the parasitoids), during
which there is no change in host mortality due to parasitization, and u < rr after
k time units. After emergence, the adult parasitoid survival probability s < 1 is
independent, of stage and density.
In Theorems 1.5.1 and 1.5.2 with A = n, AO = 1 is a strictly dominant critical
value with

and therefore

Moreover, a calculation shows that

and hence A] > 0 by (1.62). It follows from Theorems 1.5.1 and 1.5.2 that there
is a stable bifurcation of positive equilibria from the extinction equilibrium x —
0. y = ye > 0 as n increases through 1. Such an equilibrium represents a stable
coexistence state in which the parasitoid survives and reduces the equilibrium
level of the host pest.
Returning to the question posed above, we wish to determine which host
class a should be attacked if the adult component ym2 = ym,2 (a) of the positive
equilibrium is to be minimized, holding all other parameters fixed (including n}.
In absence of an explicit formula for the positive equilibrium, we can use the
Liapunov/Schmidt parameterization in Theorem 1.5.1, mathematically treating
a as a continuous variable, to find

Thus, this derivative is positive for e > 0 small, which shows yni2 is minimized
at a — 0. That is to say. the "optimal'1 strategy is to parasitize the youngest
larval instar. These result agrees with the conclusions in [22].
The age-structured model above assumes that only one larval stage is para-
sitized and then asserts that to parasitize the youngest stage is optimal. How-
ever, should only one larval instar stage be parasitized for the minimization of
the adult host equilibrium level ;i/,,l2? That is to say, for a given parasitization
effort, could a further reduction in ym,2 be achieved by an optimal distribution of
parasitization stages? This question is studied in [80], where a host parasitoid
(Usher type) model is studied with a distribution of parasitization stages and it
is proved that it is optimal to parasitize only one larval stage, namely, the stage
that produces maximal emergence numbers of parasitoids.
74 CHAPTER 1

A discrete size-structured model for the chemostat. A discrete model

for size-structured microbial growth in a chemostat of the form (1.57) is derived
and studied in [178], [386]. In that model the microbial population is structured
by mi size classes and y is the limiting nutrient concentration (y is a scalar so
ma = 1). The components of x are the biomasses of these size classes at time t.
In (1.57)

where v = 2 1 /™ 1 , E e (0,1) is the "washout rate," ye > 0 is the nutrient

concentration supplied to the chemostat, and u(y) > 0 is the nutrient uptake
rate per unit biomass per unit time (which, because it is assumed that biomass
is measured in nutrient-equivalent units, is also the rate of increase in biomass
per unit time per unit biomass). Assume that

The prototypical nutrient uptake rate u is the Michaelis/Menten function

where umsjf is the maximum uptake rate and a > 0 is the "Michaelis-Menten"
or "half saturation" constant. The model parameters are required to satisfy
certain constraints in order to make the model meaningful (e.g., so that the
nutrient uptake in one time step does not exceed what is available). One such
constraint is, for example, umax < v — 1. Others are more complicated and will
not concern us here. See [178], [386] for more modeling details.
Take A = 1 — E and the model equations become
 DISCRETE MODELS 75

where

Note that the model has the extinction equilibrium x = 0, y — ye for all A €
(0,1). Referring to the remark at the end of section 1.5.1 we apply Theorems
1.5.1 and 1.5.2 with A = 0 and •

is a strictly dominant critical value of A + \B = \B with eigenvectors

and hence WTBv = mi (1 + u (yc)) > 0. Since Jyg(\o, 0, ye) = 1 — A0 ^ 0, Theo-

rem 1.5.1 implies the existence of a continuum of positive equilibria bifurcating
from the extinction equilibrium at A = A 0 , i.e., at E = E0 = l^,"'> *. & (0,1).
Theorem 1.5.2 implies that the extinction equilibrium loses stability as A in-
creases through AO (or as the washout rate E decreases through EQ). A calcu-
lation yields

and hence the bifurcation is stable. (Note that u'(yK) < 0 would imply an
unstable bifurcation.)
We have shown that for E > EQ (E < EQ) the extinction equilibrium is
locally asymptotically stable (unstable) and for E < EQ, but close to EQ, the
76 CHAPTER 1

FIG. 1.23. The bifurcation diagram for the discrete chemostat model (1.57)—(1.63) is
schematically represented by a plot of the distance between the positive equilibrium pairs and
the extinction equilibrium pair (0,Eo) against the parameter E.

positive equilibrium is locally asymptotically stable (see Fig. 1.23). In [386]

these results, under the appropriate constraints on the parameters, are shown
to be global in that the positive equilibrium stability is global for all E € (0, EQ)
and the extinction equilibrium stability is global for all E 6 (EQ, I).
The competition of several microbial species in a chemostat will be considered
in section 3.1.1.
 CHAPTER 2

Continuous Models

The matrix equations studied in Chapter 1 have several advantages when used
as models in population dynamics. Models are relatively easy to construct from
life-cycle information about individuals. Numerical simulations are particularly
easy to carry out. Also, matrix models lend themselves to a description of
general structuring variables and they are often (but certainly not always) more
analytically tractable than continuous models.
Matrix models are best applied to populations with distinct "stages" with
respect to the structuring variable (age, body size, etc.); they are at a disad-
vantage when such stages are not present. Matrix models cannot, of course,
describe the dynamics of a population that occur between the discrete time
steps of the model and do not lend themselves to modeling based on instanta-
neous rates of change. For a discussion of discrete and continuous structured
models, their advantages and disadvantages, see [408].
In this chapter we consider structured models continuous in both the structur-
ing variable and time. The general theory of continuous structured population
dynamics involves partial differential equations and therefore is mathematically
more difficult than that of matrix models. Even the most basic of questions, such
as the existence, uniqueness, and positivity of solutions of initial value problems,
involve formidable mathematical difficulties arid technicalities. Moreover, the
partial differential equation problems that arise are usually not of a classical type
in that they are "nonlocal" (i.e., they are integro-partial differential equations)
and may contain nonlinear, nonlocal boundary conditions. For these reasons, we
restrict our attention to the mathematically simplest and most well-developed
class of continuous models, namely, age-structured models. Moreover, we focus
on asymptotic dynamics (specifically, equilibria and stability) and do not dwell
on the mathematical theory of initial value problems (see [429], [244]).
We begin by deriving equations for continuous age-structured populations in
continuous time (often called the McKendrick equations [318]) from the discrete
Leslie matrix model by the limiting process of infinitesimally refining the age
classes and the census time interval. (For other derivations see [236], [323].)
We then give an equilibrium theory based on a bifurcation theoretic approach
that closely parallels that given in Chapter 1 for matrix models. (Indeed, the
same finite-dimensional Theorem B.I.I will be used.) Applications illustrate

77
78 CHAPTER 2

the theory and the two basic types of bifurcations that can occur.
In this chapter we work directly with the partial differential equation form
of the McKendrick model. Often the McKendrick model equations are used
instead as a starting point for the derivation of other, more tractable, types
of model equations, such as ordinary or delay differential equations. This is
done by taking advantage of special mathematical features that derive from
special modeling assumptions. Some examples of this approach will be studied
in Chapter 3.

2.1 Age-structured models

In the classical Leslie matrix model for age-structured populations the projection
time interval and the length of the age classes are equal and, for mathematical
convenience, taken to be the unit of time. If the projection time interval and
the length of the age classes are instead both equal to h > 0, then the Leslie
matrix model takes the form

where xa(t) denotes the number of individuals whose ages are between a — h
and a at time t. Here mh = OM is the maximal possible age of an individual
and
per capita number of (surviving) offspring produced during a
fa(k) =
time interval of length h by individuals of ages a — h to a,

the fraction of those individuals of ages

ta(h) =
a — h to a who survive h units of time.

Thus, 1 — ta(h) is the fraction of those individuals of ages a — h to a who do not

survive h units of time. We make the following assumptions concerning these
vital rates:
 CONTINUOUS MODELS 79 >

Finally, we assume that a density function p ( t , a ) exists (with respect to age a),
defined for a, t > 0, such that

That is to say, we assume that the population can be described by a function

p ( t , a) such that the number of individuals of ages between a\ and a^ is given
by Jafl2 p(t, a)da. Our goal is to derive an equation for p(t, a) by passing h -^ 0
in the Leslie matrix model (2.1).
Consider a fixed age a = ih > 0. From the Leslie model with projection
matrix (2.1) we have

From (2.2) these equations become

or

Subtracting fa_hp(t,a)da from both sides and dividing the result by h2, we
obtain

In order to pass to the limit h —» 0 we would need to make technical assumptions

about the density function p(t, a). We will not concern ourselves here with these
technicalities, but simply proceed heuristically. Assume that the directional
derivative

exists for t, a > 0. Then (formally) letting h —> 0 in (2.3) we obtain the equation

If p ( t , a ) has partial derivatives at (£,a), this equation can be written

This is the most commonly used form of the differential equation for p, and we
will use it here (although for rigorous treatments of the equation the directional
derivative, or some "weak" formulation of the differential equation, is used).
80 CHAPTER 2

From the nonnewborn classes o > 0 of the Leslie matrix model (2.1) we have
heuristically arrived at a first-order partial differential equation for the density
function p ( t , a). For the newborn class a = 0 we have from (2.1) for t > 0

and

Now

Thus, (2.4) yields in the limit as h —> 0 the equation

The equations for the density p ( t , a) of an age-structured population, together

with an initial condition, are

If the maximal age is finite, OM < +00, it is natural to require the additional
boundary condition

In many treatments of continuous age-structured models it is assumed that

there is no maximal age, i.e.. a^ = +00. In this case, this boundary condition
is replaced by some kind of control on p(t, a) for large a (e.g., p ( t , •) 6 L1).
The vital death and fertility rates 6 and 0 are allowed to depend explicitly on
time t and on p(t, a). In the latter case, the partial differential equation and/or
the integral boundary birth condition are nonlinear. In population dynamics a
reasonable assumption is that these vital rates depend on one or more function-
als (weighted total population sizes) of the form /QaM w(a)p(t, a)da. An example
is the total population size J^M p(t, a)da. In this case the McKendrick model
 CONTINUOUS MODELS 81

(2.5) involves a nonlinear hyperbolic partial integro-differential equation with a

nonlinear integral boundary condition.
Uribe [409] has proved that the Leslie matrix model, in the derivation above,
is a consistent and stable discretization of the McKendrick equations. He makes
rigorous the derivation of (2.5) above by showing that the Leslie model gives
rise to a solution that converges to a (weak) solution of the McKendrick model.
Rigorous treatments of the basic theory of the age-structured McKendrick
equations (2.5) can be found in [241]. [429] (for aM — +oc). and [244] (for
O-M < +00). Also see the seminal paper of Gurtin and MacCamy [198]. (See
[248] and [404] for treatments of the size-structured McKendrick model.) These
theories establish the well posedness of the initial/boundary value problem
above, i.e., give conditions under which solutions exist and are unique. As
is to be expected, these conditions involve appropriate Lipschitz conditions on
the vital rates as functions of the density p. (Note that these difficult tech-
nicalities of existence and uniqueness do not arise for discrete matrix mod-
els, since as recursive formulas they obviously define unique solutions.) More-
over, these references establish the basic linearization principle for the study of
stability of equilibria, either by nonlinear semigroup theory or by classical
analysis.

2.2 Autonomous age-structured models

If the vital rates 6 and ,3 do not depend explicitly on time t, then the McKendrick
model (2.5) is autonomous. We consider an autonomous McKendrick model in
which 8 and j3 depend on a finite number of weighted total population sizes and
a finite maximal age

Here the p, — Pi(t) are / functionals (weighted total population sizes) of the
form

with weights 0 < w,(a) g i^O. <IM\. We are interested in asymptotic dynam-
ics, and we will begin by studying the existence and stability of equilibrium
solutions.
Conditions sufficient to guarantee the existence of a unique local solution
p(-.,a) e C([Q,T}.Ll(Q,a,M)), for an initial density 0 < p0(a) e L ] (0, aM), arc
82 CHAPTER 2

given in [244, p. 51], namely, for (pi,... ,p;) e Q and a € [0, UM]

6 and /? are locally Lipschitz continuous with

respect to each Pi uniformly for a € [0,0^],

Here ft is an open neighborhood of Rl+. (A solution has a directional derivative

Dp in place of dtp + dap-) We assume in addition that for some integer fc €
/[O, +00) and each a 6 [0, O.M]

2.2.1 The extinction equilibrium. The McKendrick model (2.6) has the
extinction equilibrium p = 0. The linearization of (2.6) at this equilibrium is

where 6(a) = 6(a, 0 . . . . 0) and (3(a) = /3(a, 0 , . . . 0). Treatments of this linear
McKendrick model can be found in [236] for <IM = +00 and [244] for <IM < +00.
We will briefly outline the latter treatment.
The probability of living to age a is

Letting if>(t,a) = p(t,a)/Tl(d), we find from (2.9(a)) that ifr satisfies the partial
differential equation dttjj + datp = 0. The "general" solution of this equation is
ip = tp(t - a), where ip is an arbitrary function. Thus, p ( t , a ) = ip(t - a)II(a),
where (p is a function to be determined by the boundary and initial conditions
(2.9(b),(c)). The boundary condition (2.9(d)) is satisfied by the assumption on
6 in (2.7). The initial condition (2.9(c)) is satisfied by defining (p(t) for negative
t by (p(t) — p0(—t)/H(—t), where p0(a) is extended to be 0 for a > UM- Finally,
for t > 0, ip(t) is determined by the birth condition (2.9(b)). With /3(a) and
II(a) extended to be 0 for a > OM, the equation for (p(t) obtained from (2.9(b))
is
 CONTINUOUS MODELS 83

where for t > 0

Equation (2.10) is a Volterra integral equation for tp(t). The relationship be-
tween the solution of this integral equation and the solution of the linear McK-
endrick model (2.9) is made mathematically rigorous in [244, Chapters I and
II]. A solution of (2.10) yields a solution of (2.9)

and vice versa. (Here we ignore the mathematical technicalities in denning a

solution of (2.9).) Equation (2.10) can be studied using Laplace transforms.
The transform of the solution is given by ^p(z) = (p0(z)/(l — /?(z)). Consider the
"characteristic equation" 'K(Z) = 1, i.e.. the equation

for complex roots z. With z = x restricted to real numbers, from assumptions

(2.7) it follows that K(X) > 0 is strictly decreasing for x € (—oo, +00) with
lim:c__00 K(X) =• +00 and lim x _ +00 K(X) = 0. Thus, the characteristic equation
(2.11) has a unique real root x = r. If z = x + iy is any other root,

which implies that r > x = Re z. Thus, there exists a strictly dominant real
root of the characteristic equation (which is analogous to the strictly dominant
eigenvalue of a projection matrix and hence the inherent growth rate for discrete
matrix models). It is shown in [244] (also see [236] for the same result when
OM = +00) that the solution of (2.10) has the form

Thus, the solution of the linear McKendrick (2.9) is

84 CHAPTER 2

The asymptotic dynamics of p(t, a) are determine by the dominant root r.

Note that there is a constant £ > 0 such that p(t) < £e rt p(0), where

is the total population size at time t. If r < 0. then for each e > 0, p(0) < 6 = e/£
implies that p(t) < £ for all £ > 0 and \mit^+ocp(t) = 0. In this sense, the
extinction equilibrium is asymptotically stable if r < 0. On the other hand, if
r > 0, then p(t) is exponentially unbounded as t —> +00 and the extinction
equilibrium is unstable. Thus, the extinction equilibrium loses stability as r
increases through the critical value 0. Moreover, if r = 0, then there exist
infinitely many equilibrium solutions of (2.9) given by p ( t , a ) — kH(a), where k
is an arbitrary constant.
Since H(a) is the probability of living to age a, /3(a)II(a) is the number of
offspring per individual given that reaches age a. Thus,

is the expected number of offspring per individual per lifetime, which we define
to be the inherent net reproductive number. Since K(X) is strictly decreasing to
0 as x —> +00, we see from the characteristic equation (2.11) that r < 0 if (and
only if) n < 1, r > 0 if (and only if) n > 1 and r = 0 if (and only if) n = 1. For
the McKendrick age-structured model this is the equivalent of Theorem 1.1.3
for matrix equations.
From (2.12) for t > HM we have

and hence that the normalized age distribution of any nontrivial solution (c ^ 0)
satisfies

Thus, the normalized distribution obtained from all solutions tends to the same
asymptotic limit, called the stable age distribution. This property is analogous
to property (1.7) for general linear matrix models. Furthermore, an integration
of (2.9) from a = 0 to UM yields
 CONTINUOUS MODELS 85

Replacing the normalized distribution by its limit (2.13) we obtain the limiting
equation

Observing that

we see that the limiting equation reduces to

This scalar ordinary differential equation correctly predicts the asymptotic dy-
namics of the population level quantity p(t).
From the above results we see that the theory of linear age-structured McK-
endrick equations (2.9) is analogous to the theory of linear matrix equations
given in section 1.1.2. In the following section we pursue this analogy with re-
gard to equilibrium theory for nonlinear age-structured McKendrick equations
(2.6).

2.2.2 Positive equilibria. In analogy to the general bifurcation results for

matrix equations in section 1.2.3 we anticipate that positive equilibria of the
general McKendrick model (2.6) will bifurcate from the extinction equilibrium
as a parameter is varied in such a way as to cause the dominant root r to change
sign (or equivalently to cause the inherent net reproductive number n to pass
through 1). We will prove that this is in fact true by using Theorem B.I.I
in Appendix B.I; the same theorem we used for the matrix model analysis in
Chapter 1.
With

where b is normalized to satisfy

n is the inherent net reproductive number. Introducing this expression for /3

into (2.6) we obtain the equilibrium equations
86 CHAPTER 2

From (a) and (c) we have

Note that II(a) = Il(a,0,... ,0). Placing this expression into (b) and (d) we
obtain the / + 1 algebraic equations

for the / +1 unknowns c, pi,... , pi. From a solution of these algebraic equations
we define a function p(a) by (2.17), which by (2.7) is continuous for a e [0, OM]-
This function solves the original equilibrium equations at points where it is
differentiable (a further smoothness condition that depends on the smoothness
of 8 as a function of a). We define an equilibrium of (2.6) to be a function given
by (2.17) and (2.18).
The system of equations (2.18) is equivalent to the system obtained by sub-
stituting the right-hand side of the first equation for c in the second equation.
The resulting system can be written

where

and hi € C fc+1 (f2,R+) satisfy |/^(pi,... ,pi)\ = O(S'_1 \pi\) near the origin.
This system of algebraic equations for c and pi have the form of the equation
studied in Appendix B.I, namely,

where

The only characteristic value of the (/ + 1) x (/ + 1) matrix L is n = 1 and it

has the positive characteristic vector
 CONTINUOUS MODELS 87

By Theorem B.I.I (see Appendix B.I) there exists a continuum £+ of non-

negative equilibrium pairs (n.x) containing (1,0) for which £ + / {J?1 x <9.R+}
is nonempty (and consists of positive equilibrium pairs) and for which the fol-
lowing alternatives hold : (i) either £ + / {(1,0)} is unbounded in Rl x Rl+ and
contains only positive equilibrium pairs, or (ii) C+ contains a boundary pair
(n*.x*) 6 7?,1 x dRl+ for which x* ^ 0. The second alternative can be ruled
out. however, since the only solution of (2.18) lying on the boundary +dRl is
x* = 0. (To see this, note from (2.18) that c = 0 implies that all Pi — 0 and if
at least one pi = 0, then c = 0 and hence all other pi = 0.) Prom the second
equilibrium equation in (2.18), together with 0 < II(a,pi,.. .pi) < 1. we see
that if any one component in x is unbounded along £ + , then all components are
unbounded, including c. The continuum £+ therefore defines a continuum C+
of equilibrium pairs (n, p) € R x C°[0, a^], from (2.17), which is unbounded
(using, for example, the norm n\ + max[0,OM] |p(a)|).
These results are summarized in the following theorem.
THEOREM 2.2.1. Under assumptions (2.7) and (2.8), the McKendrick model
(2.6) has an unbounded continuum C+ of equilibrium pairs (n, p] such that
(1.0) 6 C+ and C l+ /{(1.0)} consists only of positive equilibrium pairs.
From the first equation in (2.18) we have, for any positive equilibrium,

where

The quantity nf ( p i , . . . ,pi) is the net reproductive number at equilibrium, and

(2.19) means that at equilibrium each individual exactly replaces itself. Math-
ematically, this equation can be used to determine properties of the bifurcating
continuum of equilibria in the way the analogous equation (1.35) is used for ma-
trix models in section 1.2.5. For example if b and 8 are such that v(p\,... ,p/) —>
0 as E< =1 \pi\ -» +oc, then by (2.19) the spectrum a(C+] = {n : ( n , p ) e C+}
associated with C+ must be infinite. Since 0 ^ a (C + ), the spectrum must be
an interval of the form a (C+) = (HQ, +00) or [no, +00) for 0 < no < 1.
Properties of equilibria near the bifurcation point can be determined from the
Liapunov/Schmidt parameterization of the branch (£+. See Theorem B.2.1 in
Appendix B.2. If k > 2 in (2.8), then for e > 0 small

Substituting these expansions in (2.19) and differentiating the result with re-
spect to s, we obtain, upon setting e = 0,
88 CHAPTER 2

where we have defined

The bifurcation is to the right (to the left) if n > 1 (if n < 1) for equilibrium
pairs from C+ near the (n,p) = (1,0). It follows that the bifurcation is to the
right if n\ > 0 and is to the left if HI < 0.
We might expect a relationship between the direction of bifurcation and the
stability of the equilibria from pairs near the equilibrium point, just as there
is for nonlinear matrix equations. We consider the stability of the positive
equilibria on the bifurcating continuum C+ by a formal linearization procedure.
For a rigorous proof of the validity of the linearization procedure see [244, p. 69).
(Also see [429] for the case OM = +00.)
If pe(a) is an equilibrium, then the linearization of (2.6) with /? = nb is ob-
tained by substituting p(t,a) = z(t,a) + pe(a) into the equations and dropping
all nonlinear terms in the resulting equation for z(t,a). This results in the equa-
tions

where pf e(a)da. =
We lookf^Mfor solutions Ui(a)p
of the form z(t,a) =
y(a) exp (—£t). A substitution leads to the equations

for y(a). For the equilibria from C+ we have from (2.20) and (2.17)
 CONTINUOUS MODELS 89

and consequently y and £ have expansions

We know the extinction equilibrium loses stability at £ = 0 (i.e., n = 1) at which

the dominant root of the characteristic equation (2.11) equals 0. Thus, C0 = 0
and we are interested in the sign of £ for e > 0, i.e., in the sign of the coefficient
Ci. A substitution of these expansions into (2.22) leads, from the lowest-order
terms in e, to yo(a) = IT(a) and. from the first-order terms in e, to the equations

From the first equation (which is a scalar, linear nonhomogeneous first-order

ordinary differential equation for y\) we obtain

which when substituted into the second equation yields

Solving this equation for (\ and using the formula (2.21), we obtain

This shows that the equilibria (2.23) are locally asymptotically stable if the
bifurcation is to the right (i.e., if n\ > 0).
Recall that the bifurcation of C+ is called stable (unstable) if the positive
equilibria from C+ near the bifurcation point (n,p) = (1,0) are locally asymp-
totically stable (unstable).
THEOREM 2.2.2. Assume (2.7) and (2.8).
(a) The extinction equilibrium p = 0 of the McKendrick model (2.6) is (locally
asymptotically) stable if n < 1 and is unstable if n > 1.
(b) Assume that k > 2 and HI ^ 0. The bifurcation of C+ is stable if it is to
the right (n\ > 0) and unstable if it is to the left. (HI < 0).
The expansions (2.23) can be used to study the nonlinear effects on the nor-
malized age distribution near the bifurcation point. To do this the expansion
for the equilibrium p e (a) must be carried out to order e'2. This yields
90 CHAPTER 2

From pe = f£M pe(a)da and d/de = (dn/de)(d/dn) we get

For example, if nonlinear effects serve to increase the death rate of all ages
(DS(a) > 0 for all a) and if the bifurcation is stable (n\ > 0), then for n K I

which shows that the equilibrium population gets younger as n > 1 increases.
For similar results with more general dependencies of S and b on p see [88],
[89],
Often the assumption of a finite maximal age CM < +00 is not made and
individuals of all ages are theoretically allowed. To consider McKendrick models
with aw = +00 one has to "control" p, 6, and 6 at a = +oc in some manner.
For example, in place of the last assumption in (2.7) (which is needed for the
boundary condition p(t,aM) = 0) it is often required that p G L 1 (0,+00); e.g.,
see [429]. Under these conditions Theorems 2.2.1 and 2.2.2 remain valid.

2.2.3 Hopf bifurcation. The stability properties of the positive equilibria

near the bifurcation point described in Theorem 2.2.2 may not persist globally
along the continuum C+. Should the positive equilibria change stability as n
is changed, due to a root (, of the characteristic equation of the linearization
at the equilibria crossing the imaginary axis, then (usually) another bifurcation
occurs. If the root moves through £ = 0, the bifurcation involves equilibria, the
possibilities being those discussed for discrete matrix models in section 1.2.3
(saddle-node, transcritical, and pitchfork). If, however, a complex conjugate
pair crosses the imaginary axis when n = ncr through points C = ±iO, & ^ 0,
and it does so transversely (i.e., dRe£(ncr)/dn ^ 0), then a classical Hopf bi-
furcation occurs in which time periodic solutions are created. These periodic
solutions arise as small amplitude oscillations around the destabilized equilib-
rium with period approximately equal to 6/lit. Their stability or instability can
be determined, in principle, by known formulas, the use of which is unfortu-
nately difficult in applications. Numerical investigations are commonly used to
investigate Hopf bifurcations. A Hopf bifurcation theorem specifically for the
McKendrick model (2.6) can be found in [87].
An illustration of a Hopf bifurcation to a limit cycle (followed by further
bifurcations to chaos) can be seen in Figs. 2.1 and 2.2, where solutions of the
 CONTINUOUS MODELS 91

FIG. 2.1. For the delay equations (2.24) the inherent net reproductive number is n = bS ^ .
The. population goes extinct for n = 0.5, equilibrates to a positive equilibrium for n = 7.5,
and (after a Hop/ bifurcation occurs) oscillates periodically for n = 9.5. Plots are shown for
S = 2, c = 0.5, and T = 10.

McKendrick model (2.6) are plotted for

Here Xr( a ) ^s t ne Dirac function at T > 0, and hence in this model only individ-
uals of age a = T are fertile. The nonlinear density effects in this model occur
in an individual's fertility rate, which is dependent on the total population size
at birth. The model equations

lead to the system of delay differential equations

for total population size p(i) and the total birth rate B(t) == p(t, 0).

2.3 Some applications

We look briefly at two applications that involve two fundamentally different
types of nonlinear interactions in population dynamics: "negative feedback" (or
deleterious) density effects arid advantageous density effects (or Allee effects).
92 CHAPTER 2

FIG. 2.2. With further increases in n, the periodic solution in Fig. 2.1 becomes more
complicated for n = 11.5 and an apparent chaotic oscillation occurs for n = 12.5.

Negative feedback effects of density lead to stable bifurcations while Allee ef-
fects usually lead to unstable bifurcations. In the first application we consider a
McKendrick model which allows for competition for a limited resource between
juveniles and adults. The competition is viewed as deleterious in the sense that
increased population levels result in decreased fertility and/or death rates. The
second application involves an Allee effect [5], [130] and an unstable bifurcation.
In this application unstable equilibria on the bifurcating continuum are stabi-
lized at a saddle-node bifurcation at a critical value n — ncr < 1. This is often
the case in population models with an Allee effect at low population densities
but negative feedback effects at high population densities.

Juvenile versus adult competition. Suppose that individuals in a popula-

tion are nonreproducing prior to reaching a maturation age (or period) a, < OM •
Then in the McKendrick model equations (2.5) b — 0 for all ages a € [0, a,-] an
b > 0 for a e (O^OM]- Individuals of ages 0 to a,j are called "juveniles" while
those of ages greater than a., are called "adults." We are interested in competi-
tive interactions between juveniles and adults and how this competition effects
the dynamics of the population. To do this the birth and death rates are as-
sumed age specific and dependent on weighted total numbers of juveniles and/or
adults

for nonnegative weights ujj and uA- For weights wj = UA = 1 PJ and PA reduce
to total juvenile and adult numbers, respectively. These more general weighted
 CONTINUOUS MODELS 93

class sizes allow differential effects of age classes on vital rates.

We consider the case when only the birth rate b is density dependent and
hence the death rate 6 = 6(a) is a function of age only; for other cases see [110].
Specifically, we take

in the McKendrick model (2.6) where 6 satisfies the normalization (2.15) and
n is the inherent net reproductive number. We are interested in iritraspecific
competition, so it is assumed that dpb < 0. The model parameter £ measures
the relative effect that the populations of adults and juveniles have on the adult
birth rates. An increase in £ means greater intraclass competition between
juveniles and adults.
Theorem 2.2.1 implies that there exists an unbounded continuum of positive
equilibria that bifurcates from the extinction equilibrium p = 0 at n = 1, From
formula (2.21)

where

and hence the bifurcation is to the right and stable. In fact, dpb < 0 and II(a) < 1
imply that v(p) < 1 in (2.19) and it follows that n > I for all positive equilibria.
An interesting question concerns the "stabilizing'' or "destabilizing" effects of
juvenile versus adult competition. In a seminal work, Ebenman concluded from
a simple (semelparous) discrete matrix model that such competition is "desta-
bilizing" [151], [152], [154] (but see [299]). Discussion of this question requires,
of course, a careful discussion of what is meant by "stabilizing" or "destabiliz-
ing" effects. For this purpose Ebenman uses the size of the parameter region
in which stable positive equilibria occur and treats a shrinkage of this region as
"destabilizing.1' Other measures of destabilization can be used. For example,
the effect on equilibrium of total population size or the "resilience" of the equi-
librium (the magnitude of the eigenvalue with smallest absolute value) [110]. A
decrease in equilibrium level or resilience can also be considered destabilizing.
In the juvenile/adult model above we find that
94 CHAPTER 2

for those n > 1 such that 1/n lie in the range of i/(p) = f®M b(a,p)H(a}da, i.e.,
for 1 < n < (limp_,+oc v(p)} • The sign of the derivative

can be positive or negative. For example, for all age classes a

That is to say, if the maturation period a., is long, then an increase in the
interclass competitive effects results in decreased equilibrium levels for all age
classes. On the other hand, if the maturation period Oj is short, then such an
increase has the opposite effect of increasing equilibrium levels. We conclude
for this case that interclass juvenile and adult competition is deleterious if the
juvenile period is long but is advantageous if it is short. On the other hand, it
is shown in [110] that equilibrium resilience for n > 1 near 1 is always decreased
by an increase in the interclass competition.
Furthermore, it is interesting to note that

and hence that the normalized age distribution is unaffected by a change in the
intraclass competition between juveniles and adults.
From the example above we see that the effects of juvenile/adult competition
can be varied. The consequences of such a competition can be even more com-
plicated if, unlike in the case considered above, the death rate 8 is also affected
by the competition. When 8 also depends on a weighted total population size,
say 8 = 8(a,p), dp(a,p) < 0, where p = (1 - e)pj(i] + epA(t), 0 < e < 1,
Theorems 2.2.1 and 2.2.2 still imply the bifurcation of an unbounded contin-
uum of positive equilibria which is to the right and stable. Equilibrium levels
and resiliency in this case are studied in [110] where either deleterious or ad-
vantageous effects of increased interclass juvenile/adult competition can occur
depending on the detailed age-specific nature of the competition. For example,
if the population has two distinct reproduction ages which are sufficiently far
apart, then juvenile/adult competition leads to increased equilibrium resilience
[110].

Allee effects. Consider a McKendrick model (2,6) in which there is an Allee

effect in adult fertility and death rates as functions of a single weighted total
population size
 CONTINUOUS MODELS 95

That is to say, b = 6(a,p) > 0, where

These assumptions describe an Allee effect at low population densities p < pcr
but a negative feedback effect at high densities p > pcr. Note that dpb(a,pcr) =
dpS(a,pcr) = 0 and d%6(a,pcr) > 0, dpb(a,pcr) < 0. For technical reasons we
assume that

From (2.21), ni < 0 and the bifurcation in Theorems 2.2.1 and 2.2.2 is to the
left and unstable.
By (2.19) we have

where, from the assumptions made,

Note that for a given pair (n,p) 6 R^_ satisfying nv(p) = 1 there exists a
unique positive equilibrium

Define

From nv(p) = 1 and (2.25) we deduce the following facts (see Fig. 2.3).
(1) The extinction equilibrium is (locally asymptotically) stable for n e (0,1)
and unstable for n € (1, +00).
(2) For n £ (0, n cr ) there exists no positive equilibrium.
(3) For n e (n c r ,l) there exist two positive equilibria p1(a,n) and p 2 (a,n).
Let pi(n) = /QaM cj(a)p,(a.n)da. Then 0 < p\(n) < p<2(n) and v'(p\(n)) > 0 and
96 CHAPTER 2

FIG. 2.3. Equilibria are given by the intersection of the graph of v(p) with a horizontal
line located at 1/n.

i/'(p2(n)) < 0. By Theorem 2.2.2 the "lower" equilibrium p,(a,n) is unstable

for n w 1. Also, lim n _ipi(n) = 0 and limra^ner pi (n) = \imn^ncr p2(n) =pcr-
(4) For n € (l,+oo) there exists only one positive equilibrium p2(a,n) and
v'(pi(n}) < 0.
In specific models, it is usually the case that the "lower" equilibrium p\(t,n)
is unstable for all n € (ncr, +00) and the "upper" equilibrium, which meets
the lower equilibrium at n = ncr, is (locally asymptotically) stable at least for
n K ncr. That is to say, usually a saddle-node bifurcation occurs at the "turn
around" point n = ncr.
As an example, consider the case 6 = 8(a) when the death rate is not density
dependent. The linearization of the McKendrick model at an equilibrium

yields the equations (2.22), namely,

for complex £ and nontrivial y(a). The solution from the first equation y(a) =
J/(0)e~^°fJ(fl), J/(0) ^ 0, when substituted into the second equation, yields the
 CONTINUOUS MODELS 97

characteristic equation c(£,p) = I for (", where

The equilibrium is (locally asymptotically) stable if there exist no roots £ satis-

fying Re £ > 0 and is unstable if there exists at least one root satisfying Re ( > 0.
Note that

As noted above. v'(p\] > 0 for the lower equilibrium p = p± with n 6 (ncr, I ) .
Hence e(O.pi) > 1 and it follows that there exists a positive real root £ > 0 of
c(C.pi) = 1. Thus, the lower equilibrium p1 is unstable for all n 6 (nCT. I ) .
To illustrate the saddle-node bifurcation at n = ncr we consider a specific
example. Take UM = +00, 6 = constant > 0, arid uj(a) = 1. Take b(p) =
<5(! + \P) [1 - PL . 0 < c < 1, where [x]+ = x if x > 0 and [x\+ = 0 if x < 0.
Then v(p) = (1 + \p) (I - p}+ and (2.19) becomes n(l + \p) [1 - p}+ = I. For
ncr < n < 1, ncr = 4c(c + 1)~ 2 < 1, the solution of this quadratic yields
two positive roots, p\ and p^- that satisfy p\ < pcr < p? < 1, where prr =
(1 — c) /2 < 1. At n = ncr both equilibria equal pcr. A calculation yields

and the roots C = %kpi (pcr ~ Pi) (c + Pi)'1 (1 - Pi)"1 of the characteristic equa-
tions c(C,,pi) — 1. Thus, C > 0 for pi and this lower equilibrium is unstable and
£ < 0 for p2 and the upper equilibrium is stable.

2.4 Multispecies interactions

The interaction of several age-structured species can be modeled by McKendrick
equations of the form (2.6) by allowing that a species' vital birth and death
rates, J3 and 6, depend on the total population sizes of other species. Such
models permit the description of age specific interactions among species, e.g..
age specific differences in vulnerability to predation, in competitive efficiency for
resources, etc. Such an assumption leads to a coupled system of McKendrick
models. For example, a system of this kind is
98 CHAPTER 2

where pi — pi(t, a) is the age density function for the ith species, i e /[I, k], and

are weighted total population sizes.

One approach to the equilibrium existence and stability question is that taken
in section 1.5 for multispecies interactions modeled by discrete matrix equations.
Namely, the subcommunity of species i e I [2, k] is assumed to have a stable pos-
itive equilibrium p°(a) in the absence of the first species pl, which is viewed as
an "invader" species. The loss of stability of the subcommunity equilibrium
pj (a) and the possible bifurcation of a positive equilibrium for the full commu-
nity of species including pl can be studied as a function of a model parameter.
say the inherent net reproductive number n of pl when the subcommunity is at
its equilibrium.
Toward this end, we write

0! — n&i(a,pn,pi2,. • . ,Pifc)>
where \>\ is normalized by

Bifurcation results follow immediately from theorems in [91], where even more
general systems are considered. Or, alternatively, one can take advantage of
the dependence of <5, and (3i on the functional Pij(t) in (2.26), which allows
the equilibrium equations to be reduced to algebraic equations in the same way
they were for the single species case in section 2.2, and independently derive
bifurcation results using the theorems in Appendix B in a manner analogous
to that carried out for multispecies matrix models in section 1.5. Either way,
the end result is a continuum of positive (coexistence) equilibria, in which pv
is present, that bifurcates at the critical value n = 1 from the subcommunity
"trivial" equilibrium (0, p<], • l •is•absent. >Pfc)i Thismcontinuum
whichcan p
be parameterized locally (by Liapunov/Schmidt expansions) and the direction
of bifurcation determined, i.e., whether the positive coexistence equilibria near
the bifurcation equilibrium exist for n > 1 or n < 1.
The equilibrium stability question is rigorously a difficult one for partial dif-
ferential equations like (2.26); see [429] for a nonlinear semigroup treatment of
stability for McKendrick equations. Formally, local (linearized) stability can be
studied by investigating (time) exponential solutions of the equations linearized
at the equilibrium of interest. This is done in [91] with the result that linearized
stability of the positive equilibria can be related to the direction of bifurcation,
 CONTINUOUS MODELS 99

at least in a neighborhood of the bifurcation point. Thus, it turns out (ignoring

some technical details) that the subcommunity equilibrium loses stability as n
is increased through n = 1 and the bifurcating coexistence (positive) equilibria
are stable if the bifurcation is to the right and unstable if the bifurcation is to
the left.
Consider the case of two interacting species (k = 2). Assume that the vital
birth and death rates depend on only one weighted total population size of each
species. Thus,

where

and

From the equilibrium equations

follow n;(pi,p 2) = 1, where nt (pi,p2) = f^M' h (a,pi,p2) Hj (a,pi,p2) da or

where

These algebraic equations are identical to those that arose in the two species
matrix model studied in section 1.5.2. namely, (1.61). Thus, if species p2 has
negative feedback density effects (at least near equilibrium) so that

then
det M > 0 implies a stable bifurcation at n = 1,
detM < 0 implies an unstable bifurcation at n = 1,
100 CHAPTER 2

where

and the partial derivatives are evaluated at ( p i , p z ) = (0,p|) ail<^ n = I . (If

dp2n2(Q,p?,) > 0, these inequalities are reversed.)
Thus, if det M > 0, then the species pl can successfully invade the second
species (at low density) if n > 1 but not if n < 1.
The remarks in section 1.5.2 concerning predator-prey interactions and com-
petition interactions apply to this continuous age-structured model. For exam-
ples of continuous age-structured predator-prey (and host-parasite) see [108],
[216], [217], [218], [231], [286]. An age-structured competition model is studied
in [109].

2.5 Other structured models

The McKendrick age-structured model (2.5) is derived heuristically from the
Leslie matrix model in section 2.1 by a limiting process. The derivation is made
rigorous in [409]. The Usher matrix model is also subjected to a similar limiting
process in [409] in order to derive a continuous size-structured model. Under
the assumption that all newborns have the same size Sb, the result is the set of
model equations

Here s is a measure of an individual's "body size" (length, surface area, volume,

biomass, etc.) and p = p(t, s) is a size-specific density function. In these
equations 6 = 6(t, s) and j3 = /3(t. s} are size-specific death and birth rates. The
coefficient 7 = f ( t , s) is the "growth" rate of an individual of size s at time t.
An initial condition p(Q, s) = PQ(S) is added to these equations. If the maximal
size SM < +00, then it is natural to require the additional boundary condition
p(t,s) = 0. t > 0, s > SM- If any of the vital rates 6, /3, or 7 depend on the
density p, then the model becomes nonlinear. This generalized McKendrick
model is equivalent to the age-structured McKendrick model (2.5) if 7 si. It
can also be derived from first principles as a continuity equation (balance law)
[323]. If all newborns are not of the same size, then the boundary condition
must be modified; see [323]. Models in which there is more than one structuring
variable have a similar form. For example, if both age and size are used as
structuring variables, then the partial differential equation for p = p(t,s,a)
becomes

For a treatment of age-size-structure equations see [404].

With regard to equilibria, the bifurcation theoretic approach used above for
both discrete and continuous age-structured models has been applied in [409] to
 CONTINUOUS MODELS 101

the autonomous and nonlinear size-structured model (2.30) with SM = +00, in

which 8 — 6(s. p), {3 = /3(s, p) and 7 = 7(,s, p) depend on population density. A
global continuum of positive equilibria is shown to bifurcate from the extinction
equilibrium as a function of the inherent net reproductive number

If, as a special case, the density effects are through a dependence on one or more
weighted total population sizes b — <5(s,pi,... , p k ) , 0 — j 3 ( s , p i , . . . , p k ) , and
7 = 7 ( s , p i , . . . ,pfc), then the finite dimensional bifurcation theorem Theorem
B.I.I (see Appendix B.I) can once again be used on the equilibrium equations in
a manner almost identical to that used on the McKendrick model (2.6) in section
2.2.2. For an application that utilizes these kinds of continuous size-structured
models see section 3.2.
Finally, we point out that the bifurcation approach to equilibria for au-
tonomous age-structured models taken above has been applied to periodic so-
lutions of periodic age-structured models [90]. Periodic structured models are
ones in which the birth and/or death rates are periodic functions of time, as
might be appropriate for a population in a periodically fluctuating (e.g., sea-
sonal) environment.
This page intentionally left blank
 CHAPTER 3
Population Level Dynamics

One of the basic goals of structured population dynamics is to form a bridge

from the level of individual organisms to the total population level. The mod-
eling methodologies in Chapters 1 and 2 set up dynamical equations for a class
distribution or a density function of individuals. Often these equations are not
themselves utilized and studied, but instead are used as a starting point for
the derivation of dynamical equations for a population level quantity (such as
total population numbers, biomass, etc.). In this way the gap between the in-
dividual level and the population level is mathematically bridged. Moreover,
the equations at the population level are sometimes more analytically tractable
than those at the individual level, being of a "simpler" type (e.g., an ordinary
rather than a partial differential equation) or of lower dimension. In this chap-
ter we look at several classes of structured models for which it is possible to
derive equations for total population level quantities. These types of models are
specialized, of course, but several applications in which these types arise will be
given.

3.1 Ergodicity and nonlinear models

For general linear matrix equations we found in section 1.1.2 that normalized
distributions always equilibrate and that the dynamics of total population size
could be determined by a simple linear scalar difference equation. We saw in
section 2.2.1 that an analogous result holds for linear continuous age-structured
models where, in this case, the dynamics of total population size can be deter-
mined from a scalar linear ordinary differential equation. In this section we will
see how, under certain restrictive conditions, a similar jump from individual
level dynamics to population level dynamics can be made for certain types of
nonlinear models.

3.1.1 Discrete matrix models. In one of his seminal papers, P. H. Leslie

considered a nonlinear modification of the linear age-structured matrix model
that now bears his name [282]. This modification describes a specialized situ-
ation in which density effects on mortality are independent of age. Thus, each
103
104 CHAPTER 3

transition probability ij.j_i in the Leslie projection matrix

is reduced by the same fraction l/q(p), a fraction that depends on total popu-
lation size p = x\. Moreover, since the entries fa are the per capita number of
surviving offspring (per unit time), each of these entries in the matrix must also
be multiplied by the factor l/q(p). This results in a nonlinear Leslie projection
matrix

where q(p) > I for p > 0. If <?(0) = 1, then the entries i^-i are inherent
transition rates, i.e., the probabilities of surviving from one age class to the
next when the effects of density are ignored (e.g., when the population density
is very low). Similarly, the f\j are the inherent fertilities, i.e., fertility rates at
very low population densities.
Leslie studied the case q(p) — I + cp, c > 0. He noted that the resulting
nonlinear matrix model has a stable age class distribution, just as in the Funda-
mental Theorem of Demography (see Theorem 1.1.2) for linear matrix models,
but unlike linear models the total population size p(t) always asymptotically
equilibrates. We can see why this is true by considering the matrix equation

where P is a (constant) irreducible and primitive matrix and

From

we see that the normalized class distribution

 POPULATION LEVEL DYNAMICS 105

satisfies the recursive formula

Similarly, it is shown that the normalized class distribution

of the solution of the linear matrix equation

also satisfies (3.5). Since recursion formula (3.5) has a unique solution it follows
that (f>(t) — tp(t) and (see section 1.1.2)

where 7; is a positive eigenvalue associated with the dominant eigenvalue r > 0.

Moreover.

This is a nonautonomous scalar difference equation for the dynamics of total

population size p(t) which is coupled to the normalized class distribution <p(t)
(i.e., to the level of individuals). However, since <p(t) stabilizes, it can be replaced
by its limit (3.6) in equation (3.7). This yields the limiting equation

for total population size [278]. This limiting equation is a scalar autonomous
difference equation that is uncoupled from the dynamics of the normalized class
distribution.
Although it is natural to conjecture that the asymptotic dynamics of the
original equation (3.7) for total population size p(t) and those of its limiting
equation (3.8) for q(t) are closely related, this mathematical question does not
seem to have been thoroughly studied. When the limiting equation (3.8) has
"tame'' asymptotic dynamics, a rigorous connection to the dynamics of equation
(3.7) is made in the following theorem [95] (also see 7]).
THEOREM 3.1.1. Assume that P > 0 is irreducible and primitive and h sat-
isfies (3.4). Suppose that the limiting equation (3.8) has at most a finite numbe
of cycles in any compact interval, all of which are hyperbolic, and every forward
bounded solution tends to a cycle as t —> +00. Then a forward bounded solution
p(t) of (3.7) tends to a cycle of its limiting equation (3.8) as t —-» +00. Moreover,
if the extinction equilibrium q = 0 of the limiting equation is unstable, then the
solution p(t) of (3.7) does not tend to 0 as t —» +oc.
106 CHAPTER 3

The cycle to which the bounded solution of (3.7) tends asymptotically in

Theorem 3.1.1 is not necessarily a stable cycle of the limiting equation. Nor is
it necessarily true that a solution p(i) of equation (3.7) that starts sufficiently
near a stable cycle of the limiting equation (3.8) will tend asymptotically to that
cycle. This is because the dependency of the nonautonomous equation (3.7) on
t can cause significant initial differences between its solutions and those of its
limiting equation (3.8). However, if the total population size p(t) starts near a
stable cycle of the limiting equation and the normalized class distribution </?(£)
starts near the stable class distribution v/ \v\, then presumably such differences
would not arise. This is the content of the following theorem [95].
THEOREM 3.1.2. Suppose that q(t) is an asymptotically stable cycle of
the limiting equation (3.8). Under the assumptions of Theorem 3.1.1 there
exists a S > 0 such that |p(0) — g(0)| < 6 and \<f>(t) — v\ < 6 imply that
limt-H-oo |p(t) - </(*)|=0.
Thus, in general, to guarantee that the population asymptotically approaches
a stable cycle of the limiting equation it is not sufficient that the total population
size start near the cycle, but it is required in addition that the initial normalized
class distribution start near the stable class distribution. Or, put another way, a
small perturbation of total population from a stable cycle might be destabilizing
if the disturbance significantly perturbs the class distribution from the stable
class distribution.
In Theorems 3.1.1 and 3.1.2 an equilibrium is a cycle of period 1. For more
on limiting equations see [77], [278].
If h = h(p) is a function of a weighted total population size

then the limiting equation is

In particular, if h — h(p) is a function of total population size (wj = 1), then

the limiting equation is

The age-structured matrix equation with h(p) = 1 + cp considered by Leslie

yields the Beverton-Holt limiting equation

where r is the strictly dominant eigenvalue of the Leslie matrix (3.1). This
equation has only equilibrium dynamics, namely, lim.t-,+00 Q(t) = 0 if r < 1 and
lim^+oo q(t) — (r — 1) /c if r > 1. Theorem 3.1.1 implies that p(t) satisfies
these same alternatives and, in either case, the normalized class distribution
 POPULATION LEVEL DYNAMICS 107

normalized distributions
FIG. 3.1. The normalized distributions ip(t) from a 3x3 Leslie model (3.2) with a nonlinear
factor l/g(p) = exp(—p) stabilize to v/ v\ while the total population size p(t) exhibits several
different asymptotic dynamics. In (a) /n = 0, /i2 = 2.25, /is = 1, p2i = 0.25, ps2 = 0.75
with dominant eigenvalue r = 0.88 and v/\v\ = (0.65,0.19,0.16) and the population goes to
extinction. In (b) /] j =0. /i2 = 10, /is = 10, p2i = 0.9, pa2 = 0.75 with r = 3.3 and v/ \v =
(0.75,0.20,0.05) and the total population size tends to a positive equilibrium. In (c) /n = 0,
/12 = 75, /i3 = 50, pal = 0.9, p32 = 0.75 with r - 8.5 and v/ \v\ - (0.896,0.095,0.009) and
the total population tends to a cycle of period 1. In (c) /n = 2, /i2 = 750, /is = 1, p2i = .9,
P32 = 0.75 with r - 27.0 and v/ v = (0.9669,0.0322,0.0009) and the total population size
appears chaotic.

stabilizes to the unit eigenvector v/ \v\ > 0 of the Leslie matrix (3.1) associated
with the dominant eigenvalue.
If in Leslie's nonlinear model h(p) — exp( — cp) is used instead, the limiting
equation is the Ricker equation

for which stable attractors other than equilibria are possible. For r < 1.
limt_+oo <l(t) — 0, and for 1 < r < e2 there is a unique positive equilibrium
c^ 1 lnr which is (globally) asymptotically stable. Theorem 3.1.1 implies that
the same alternatives hold for the total population size p(t). For r > e2 there
occurs the familiar period doubling cascade of stable cycles up to the critical
value rcr « e2-6924 ^ 14 767 (see [313])_ Theorems 3.1.1 and 3.1.2 are applicable
for r < rcr. For r > rcr chaotic dynamics occur for the limiting equation and
although in this case we have no theorems relating such dynamics to the dy-
namics of p ( t ) . the numerical simulations in Fig. 3.1 suggest that p(t) can also
be "chaotic."
108 CHAPTER 3

A stable class distribution result can be obtained for more general equations
than (3.3) from the following theorem (a proof of which appears in Appendix
C) [98].
THEOREM 3.1.3. Consider the equation

where
(a) 0 < a(t) < aa, 0 < bo < b(t) for constants O.Q and bo and t e /[O, +00);
(b) / is the m x TO identity and L is an m x m matrix that has a strictly
dominant, simple eigenvalue r > 0 with a positive eigenvector v > 0.
Suppose that x(t) is a solution satisfying 0 / x(t) > 0 for all t £ /[O, +00)
and p(t) = wTx(t) = Y^i^iX^), 0 ^ w »> Y^iLi^i ^ 0, is a weighted total
population size. Then

This theorem can be applied when a and b are functions of x and equation
(3.9) is nonlinear. Using (3.10) we can derive a limiting equation for the weighted
total population size p(t) as follows. Substituting x(t) = <p(t)p(t) into

and replacing (p(t) by its limit V/UTV in the result, we obtain (relabeling p as q)

and hence the scalar limiting equation

If a = a(p) and b — b(p) are functions of p, then the limiting equation simplifies
to the scalar equation

In this theorem the dependence of the coefficients a and b on t can be implicit

through a dependence on x. Thus, equation (3.3) is included in (3.9) by settin
a(t) = 0, b(t) = h ( x ( t ) ) , and L = P.
Theorem 3.1.3 was motivated by size-structured models of the following type.
In a matrix equation with projection matrix P — T + F write the transition
rates i^ in T = [tij] as

where 7^- is the fraction that leaves the jth class per unit time, TJJ is the fraction
of those who leave that move to class i, and ~KJ is the survival rates per unit
 POPULATION LEVEL DYNAMICS 109

time. Make the following two simplifying assumptions: 7^ and the class specific
fertility rates fa in the fertility matrix F are proportional to a common resource
consumption rate u > 0 and the survival rates TTJ are class independent. Thus,

where it is required that TJU < 1. In this way assumptions about how indi-
viduals of each structuring class consume a designated resource and how that
resource consumption affects both fertility and the movement between classes
(e.g., growth to other size classes or changes to other life-cycle stages) can be
included in the model by construction of a submodel for u and its dependence
on resource density, competition factors, etc.
Selecting any index k we can write the projection matrix in the form a(t)I +
b(i}L, where

and L is the matrix

If k chosen as the index for which T^ is maximal (T> > T; for all i ) , then L > 0. If
the final simplifying assumption is made that the only model parameter affected
by population density is the resource uptake function u = u(x), where u(0) = 1,
then the model equations take the form (3.9) with a(t) == TT (1 — T k-u(x(t))),
b(t) = 7rti(z(t)), i.e.,

If L satisfies the requirement in Theorem 3.1.3 (which it will, for example, if it

is rioimegative, irreducible, and primitive) and if, for a given solution x ( t ) , the
coefficients a and b satisfy the requirements of this theorem, then x(t) will have
a stable normalized class distribution (3.10). The requirement on a is satisfied
with ay — 1- The condition on b requires, for each solution x(t) > 0. that the
uptake function u(x(t)) be bounded away from 0 for all t. This condition will
be met, for example, if u(x) > 0 is continuous for all x > 0 and the solution is
bounded.
The limiting equation is

where v(q) = u(vq/u!Tv), u(0) — 1, and 0 == r — T£. While this is the limiting
equation for the population level quantity p(t), individual (class) level param-
eters contribute to the dynamics through the eigenvalue r and eigenvector v
110 CHAPTER 3

of the matrix L. Note that if u = u(p) is a function of the weighted total

population size p, then v(q) = u(q) in the limiting equation.
As an example consider the self regulatory case when the resource uptake
function u(x) > 0, w(0) = 1, is a decreasing function of each class density
Xi. Then v(q) is a decreasing function of the real variable q > 0 and con-
sequently v(q) < 1 for q > 0. From the inequality q(t + 1) < TT(! + 6) q(i)
we find lim(^+00 q(t) = 0 for all </(0) > 0, and the population goes extinct if
9 < 6cr = (1 — 7r)/7r (Theorem 3.1.1). If Q > 9cr, then there exists a unique
positive equilibrium of the limiting equation q = v~l (6CT /9). This equilibrium
is asymptotically stable near the bifurcation point 9 « 6cr, but might lose sta-
bility for larger 6 (for which the familiar period doubling route to chaos might
occur). For example, for the case u(p) = 1 + cp studied by Leslie the positive
equilibrium is globally asymptotically stable for all 0 > BCT, while for the Ricker
nonlinearity u(p] = exp(—cp) a period doubling route to chaos occurs in the
limiting equation as 6 is increased.
For other generalizations of Theorem 3.1.3 and applications see [78], [79],
[289]. For more on ergodic theorems see [7], [64], [65], [66], [67], [177], [186].

A size-structured model. A size-structured model of the form (3.12) is stud-

ied in [98], [101]. In that model individuals are classified according to body
length and the projection matrix is a nonlinear Usher matrix of the form

All newborns lie in the smallest (first) size class and an individual either remains
in its size class or grows to the next in one unit of time. No individual grows
larger than the final size class m.
The parameters a; and fti are called the reproductive efficiency and growth
efficiency coefficients, respectively. They are related, in this model, to various
size specific parameters, namely,

where TJ is the inherent (low density) resource uptake rate (per unit time per
unit body surface area) of an individual from the jth size class; e~d denotes the
fractional decrease in resource uptake of any individual (per unit time per unit
body surface area) due to a unit surface area of a competitor; Sj is the body
length of an individual from the jth size class (a representative size from the
size interval, say the middle point); 8j is the size class interval length; Oj is a
proportionality factor that relates body surface area to the square of body length
 POPULATION LEVEL DYNAMICS 111

sj (and hence depends on the geometry of an individual): //^ is a body density

per unit volume (so that /z^ is body weight); KJ is the fraction of consumed
resource that individuals of size Sj allocate to growth, and 1 — KJ is the fraction
of consumed resource that individuals of size Sj allocate to reproduction; rjj is
a conversion factor of resource units to body weight; u>j is a conversion factor of
resource units to offspring body weight; and finally w\ — /t^s? is the weight at
birth. In this model p is the weighted total population size p — X^IHi ffis"ixi(t}';
i.e., p is the total population surface area.
These details are given in order to show the considerable amount of class spe-
cific (individual level) information that is included within the strictly dominant
eigenvalue r of the matrix

and hence within the single parameter 9 — r — /3k that determines the dynamical
properties of the limiting equation

for total population surface area. In this way the effect that a particular size
specific parameter has on the population level dynamics can be determined by
studying the effect that the parameter has on f).
As a function of 6, solutions of the limiting equation tend to 0 if 0 < Ocr. For
9 > Ocr solutions persist, but there is a period doubling cascade to chaos as 6 is
increased. Note that 9 is the eigenvalue of

with largest real part.

Perhaps the simplest example is the case of m = 2 size classes in which the
smaller consists of juveniles (KI = 1 and a\ — 0) and the larger consists of
adults («2 < 1 and a-i > 0). Then the projection matrix is
112 CHAPTER 3

(in which unnecessary subscripts have been dropped: 0i — 0, a? — a and

«2 = K) and

from which we calculate

As seen above the population goes extinct if 0 < Ocr = (1 — 7r)/7r and persists
if 9 > Ocr, so it is advantageous for the population to increase 0. Since 0
is an increasing function of each coefficient a and (3, it is to the population's
advantage to increase either coefficient. How changes in the size class specific
parameters listed above cause increases or decreases in these reproductive or
growth coefficients can be determined from the formulas (3.14), i.e.,

(where once again unnecessary subscripts have been dropped).

Consider, for example, 9 as a function of the growth allocation fraction n of
adults. Writing

where

are dimensionless parameters, it is straightforward to show that O(K) has a

unique maximum value of 0max = cic 2 ( v /ci'+l)~ 1 /2 which occurs for the unique
fraction K = K max (c 2 ) = \(c<2 — \fcz) (C2 ~ 1)~ • It is interesting to note that
ftmax lies between 0 and 0.5 and consequently the maximum value of 6 is attained
in this model by using (and only by using) an adult resource allocation that
assigns more than half to reproduction. The fraction K m ax(c2) is an increasing
function of c2 > 0 satisfying K max (0) = 0, lim C2 _ +00 « max (c 2 ) = 0.5. Thus,
for c2 large 6 is maximized by nearly an equal resource allocation to growth
and reproduction, while for c2 small 9 is maximized by allocating very little
consumed resource to growth and nearly all to reproduction. See Fig. 3.2.

An interference competition model and nonequilibrium coexistence.

In a coupled system of matrix equations for the dynamics of interacting struc-
tured species, scalar limiting equations can be derived for the total population
sizes of those species whose matrix equation is of the type in Theorem 3.1.3.
See [79] for examples. We consider a competition example in which all species
 POPULATION LEVEL DYNAMICS 113

FIG. 3.2. For small c? the maximum of 0 is attained when only a small fraction K,max of
resource is allocated to adult growth. For large 02 the maximum of 6 is attained when the
resource allocations to adult growth and reproduction are nearly equal.

are described by matrix equations of this type and hence the system of matrix
equations for species class distributions can be replaced by a limiting system of
scalar equations.
Consider a community of n species, each of whose dynamics are modeled by
a matrix equation of the type (3.13). Assume the species are in competition
for a resource and that this competition is expressed by the dependence of a
common uptake function u on the weighted population sizes Pj(t) of all species.
The weighted population sizes pj do not necessarily have the same weights.
Thus, we write u = u(p\,... .pn). u(Q,... , 0) = 1, and under the assumptions
of Theorem 3.1.3 we obtain the system of limiting equations

where TT^ is now the class independent survival rate for the ith species and (9,
is computed from the L matrix for species i. How the dynamical properties
of this population level competition model depend on class specific (individual
level) parameters can be studied through the formulas (3.14) that relate how #,•
depends on the entries in the matrix (3.15) for the ith species.
Since we are interested in competition, we assume that u(p\^. . . , / ; „ ) > 0 is
a decreasing function of each of its arguments

It follows that
114 CHAPTER 3

Define

The first observation to make is that, arguing exactly as above for a single
species, any species for which Bi < 9^r will asymptotically go extinct. Therefore,
we might as well assume that

In this case, the limiting equation for each species has a positive equilibrium
Qi = ql > 0 given by the unique solution of the equation

The inequality (3.21) and the monotonicity assumption (3.20) imply that this
equation has a unique positive equilibrium.
The second observation is that if a nonextinction equilibrium ( q ± , . . . , qn) > 0
exists, then the ratios 9^ jOi would have to be equal for all subscripts corre-
sponding to nonzero components g, > 0. Ignoring this unlikely ("nongeneric")
case it follows that the only positive nonextinction equilibria are those with a
single positive component & > 0, which is in fact an equilibrium for the ith
species in the absence of the other species, i.e., Qi = qf. It is impossible, in this
model, for two or more species to coexist in a state of equilibrium, a fact in
keeping with the famous competitive exclusion principle in theoretical ecology.
The local stability of an equilibrium ( 0 , . . . , <j|,... , 0) is determined by the n
eigenvalues

and

associated with the linearization of (3.19) at this equilibrium. Suppose that

the jth species has a stable equilibrium in the absence of the other species;
specifically assume that the absolute value of the eigenvalue (3.23) is less than
1. Stability is then determined by the remaining eigenvalues (3.22). Since

implies that

these eigenvalues can be written as

 POPULATION LEVEL DYNAMICS 115

These eigenvalues are less than 1 if and only if

Thus, one and only one of the equilibria is (locally asymptotically) stable and
it is the one corresponding to the largest ratio 6j/9c^, say for j = k.
These results only suggest that all species except the kth species go extinct,
since the stability analysis is only local. Under certain circumstances, how-
ever, the stability of the kth species equilibrium can be shown to be globally
attracting. Namely, suppose

Then Qkl^k is the largest ratio and 6 = max^fc {i^i/^k} < 1- Let q i ( t ) , i G
/[l,n], be a (forward) bounded solution of (3.19) with <?,•(()) > 0. For i / k we
have from (3.19) that

Consequently q i ( t ) / q k ( t ) —» 0 and q%(t] —» 0 as t —> +00.

Through the relationship of the ratio ®ijff? to the size class specific param-
eters for the iih species in the model ( a j , Sj, 77^, etc.), obtained through the
formulas (3.14) for the entries in the matrix (3.15), one can study how any one
of these size class specific parameters affects the competitive outcome of the
system.
As an example, consider the competition of two size-structured species, each
structured by m = 2 size classes, modeled by a coupled system of two matrix
equations with projection matrices of the form (3.16). Thus, the smallest size
classes consist of (nonreproducing) juveniles. If the species are in competition
for the resource, then u — exp (—d\pi — dipz), where pl is the total population
surface area of the ith species and the limiting equations of each species form
the coupled system

The surviving species is the one with the largest ratio Qi/9c,r > 1, where each
&i is given by the formula (3.17) using the reproductive and growth coefficients
for that species, i.e.,

Consider two species that are identical except for their adult growth allocation
fractions, which are different. By this we mean the two species have the same
model parameter values (a, 6, 77, etc.) except for K. Then both Oi = #(KJ) are
116 CHAPTER 3

given by the same function of K (namely, (3.18)). It follows that if one species
utilizes the allocation fraction K max , then that species "wins1' the competition.
The competition model (3.19) is consistent with the fundamental competitive
exclusion principle from theoretical ecology insofar as equilibrium dynamics are
concerned. However, from the broader perspective of nonequilibrium dynamics
this famous principle does not always hold.
Prom a bifurcation theory point of view the competitive exclusion principle
can be explained, for the type of competition models being considered here, as
follows. Consider the two equations

for a positive equilibrium of the limiting equations (3.19) for m = 2 competing

species. Using 9\ as a bifurcation parameter we see that these equations have
a solution if and only if Q\ = B^Q^/Q'z , in which case positive solutions are
obtained from the equation u(qi, 92) = fff JQ\. By the implicit function theorem
this equation has a solution q\ = 91(92), 9i(<?I) = 0- That is to say, there
is a branch of positive equilibria that bifurcates from the equilibrium point
(91,92) = (0,9|) as a function of 9\ at the critical value #1 = Oc\ 02/^2"• But
the bifurcation is "vertical" in the sense that the positive equilibria exist only
for a single value of Q\ and it is this "point" spectrum that gives rise to the
competitive exclusion principle in this model. (Many competition models in
the literature can be viewed in a similar manner.) Thus, in trying to obtain a
bifurcation from a stable equilibrium (0,9^) to stable positive equilibria in which
both species are present, we find that the spectrum for QI is a single point and
hence these positive coexistence equilibria are "nongeneric."
The same bifurcation point of view can be taken when the isolated species
92 has a stable attractor different from an equilibrium, such as, for example,
a stable cycle. If we attempt to find positive coexistence cycles that bifurcate
from this stable cycle by using &i as a bifurcation parameter, will the spectrum
still be a single point? Or can the bifurcation in this case not be vertical and
the spectrum an interval, so that coexistence is possible over a set of 61 values
that does not have measure 0? The existence of a bifurcating branch of cycles
can be proved, and its direction of bifurcation determined, in a manner identical
to that used for equilibria by considering composites of the modeling equations
(a fixed point of the first composite, that is riot an equilibrium, is a 2-cycle; a
fixed point of the second composite that is not an equilibrium is a 3-cycle, etc.);
see [76]. We will be content here with numerical examples that illustrate such
bifurcations.
In Fig. 3.3 graphs of simulations of the limiting equations (3.24) are shown
for selected values of the parameters in which species 92 has a stable 2-cycle
in the absence of species q\. For small values of 9\ species 91 goes extinct and
species 92 survives while the opposite is true for large values of 0\. However,
for intermediate values of 9\ there is a positive stable 2-cycle of the limiting
equations which represents an oscillatory coexistence state. The coexistence 2-
cycles bifurcate from the 2-cycle on the 92 axis (which loses its planar stability
at the bifurcation point) into the positive quadrant and continue to exist as
 POPULATION LEVEL DYNAMICS 117

FlG. 3.3. In thep\,p2 plane attractors are shown for several values oj Q\. (Dotted lines
connect 2-cycle attractors.) The other parameter values in the model equations (3.24) are
TTi = 0.75, 7T2 = 0.25, 62 = 100, c/i =1^2 = 1. For 0\ — 1 species p\ goes extinct while
species P2 tends to a stable 2-cycle. For 0\ = 12 species P2 goes extinct while species pi
tends to a stable equilibrium. For intermediate values ofOi the two species coexist in a stable
2-cycle, in contradiction to the fundamental tenet of competitive exclusion (which is based
upon equilibrium dynamics).

6*1 is increased until they are destroyed by a saddle-node bifurcation with the
unstable positive equilibrium.
Similar bifurcations into the positive quadrant of higher period cycles can be
seen in this model if q-2 has stable cycles of higher periods in the absence of q\.
Even strange and chaotic attractors can bifurcate into the positive quadrant
if q-2 has a strange attractor in the absence of q\\ see Fig. 3.4.
Thus, we see tha.t the classical competitive exclusion principle is very much
an equilibrium theory. For situations in which stable nonequilibrium dynamics
occur for one of the species it is theoretically possible for two (or more) species
to coexist when competing for a single resource, which is in violation of this
principle.

An exploitative competition model. Consider the discrete cheraostat

model in section 1.5.2 in which a micro-organism x is size-structured. This mo-
del can be put in the form (3.9) suitable for Theorem 3.1.3 in the following
way. Given that u max < v — I in the Michaelis/Menten uptake function u in
(1.64) choose any small number e > 0 so that (1 + s) umax < v — 1 and write
the projection matrix (1.63) in the form (3.9) with
118 CHAPTER 3

FIG. 3.4. In the pi,p2 plane attractors are shown for several values of Q\. (Dotted lines
connect the two 2-cycle attractors). The other parameter values in the model equations (3.24)
are the same as in Fig. 3.3 except that 62 = 500. For Oi = 20 species pi goes extinct while
species P2 tends to a chaotic attractor. For 6\ = 60 species p2 goes extinct while species
pi tends to a stable equilibrium. For intermediate values of 6\ the two species coexist in a
chaotic attractor (at 61 = 27) or a stable 2-cycle (at 6\ = 30 and 40), in contradiction to the
fundamental tenet of competitive exclusion.

and

Here we have relabeled the substrate concentration y in section 1.5.2 as S. The

strictly dominant eigenvalue and eigenvectors of L are r = v + e and
 POPULATION LEVEL DYNAMICS 119

Thus, the population has a stable normalized class distribution and its total
biomass p(t) = Y^i X»M nas ^ie limiting equation (after some simplification)

This equation, together the limiting equation for the substrate is given by g(x, S)
in (1.63) and the Michaelis/Menten expression (1.64) for u(S), yields a planar
system of limiting equations for the population level dynamics of the discrete
chemostat

This system has the extinction equilibrium (q, S) = (0, Se) for all parameter
values. It has a positive equilibrium

if and only if

The eigenvalues of the Jacobiau evaluated at the extinction equilibrium (q. S) =

(0, Se) are the positive numbers

both of which are less than 1 if and only if E > EQ.

An equivalent system is obtained if the second equation is replaced by the
sum of the equations, namely,

The second equation implies that q(t) + S(t) —» Se, and thus, we replace S(t)
in the first equation by Se — q(i) to obtain another limiting equation

(This can be viewed as the dynamics on the line segment q + S = Se lying in the
positive quadrant to which all positive solutions tend.) We have reduced the
discrete size-structured chemostat model, regardless of its size (i.e., the number
120 CHAPTER 3

of size classes used), to the study of a single scalar map! The limiting equation
(3.27) and its mathematical relationship to the dynamics of the original size-
structured model are rigorous and thoroughly studied in [386], where it is shown,
under appropriate restrictions on the model parameters, that the extinction
equilibrium q = 0 is globally stable if E > EQ and a unique positive equilibrium
is globally stable if E < EQ.
If n species are each modeled by a discrete size-structured model of the type
above, then each species has a stable normalized size class distribution and each
matrix equation has a limiting equation of the form (3.25). The result is a
system of n + 1 limiting equations for the substrate and the total biomasses of
the species

Although this system is still rather formidable, it is nonetheless of considerably

lower dimension than the original size-structured model. The case of n — 2
species is studied in detail in [386]. It is shown there that the familiar chemostat
competition result holds, namely, that only one species survives asymptotically
and that is the species with the smallest "break-even concentration of nutrient."
This break-even concentration is denned to be the level Aj of substrate at which
the coefficient in the equation for g$ is equal to 1, i.e.,

3.1.2 Continuous age-structured models. The stable normalized distri-

bution result for the nonlinear discrete model (3.3) has an analog for nonlinear
continuous age-structured models [378] (also see [41], [244]). The modeling as-
sumption in the McKendrick age-structured model (2.5) analogous to the scalar
multiplicative assumption in the discrete model (3.3) is that the death rate has
an additive decomposition into an age specific term plus a density dependent
term, namely,

where pi is a weighted total population size

as in section 2.2. For this reason the resulting McKendrick age-structured model
 POPULATION LEVEL DYNAMICS 121

is often called a separable model. Biologically the model implies that the effects
of density on the death rate act uniformly over all age classes. Note that the
fertility rate (3 (a), while age dependent, is density independent in this model.
The key to obtaining a stable age distribution result and a limiting equation for
total population size

is found in an analog to equation (3.5). Straightforward calculations show that

the normalized distribution

of the solutions of both the separable model (3.28) and the related linear model
(2.9) satisfy the same equations

It is here that the additive separability of the density and age-specific terms in
the death rate is crucial. It follows that the normalized age distribution of the
separable equation has the same stable age distribution as that of the linear
equation in section 2.2.1. namely, (2.13) holds.
The nonlinear analog of the linear limiting equation (2.14) (and the continuous
analog of the discrete limiting equation (3.11)) can be found by substituting
p ( t , a ) = p(t)(p(t.a) into equations (3.28) and making use of equations (3.29)
and

The result is the ordinary differential equation

for total population size p — p(t). This equation is coupled with the normalized
distribution (p. The limiting equation for p is obtained by replacing <p by its
asymptotic limit

where
122 CHAPTER 3

and r > 0 is the dominant positive real root of the characteristic equation

Using

we obtain the limiting equation (replacing p(t) by q(t))

where

Solutions of this scalar autonomous ordinary differential equation have only

monotonic dynamics and the only possible asymptotic states for bounded solu-
tions are equilibria.
A special case occurs when the death rate ^ = (j,(p) depends on total pop-
ulation size. In this case, the limiting equation simplifies to the "logisticlike"
equation

Multispecies separable equations result if the death rate of each species is

additively separable and the density term depends on the population densities
of other species. For example, if for the ith species fj,t = /^(PI, •.. ,pn), where
Pi is the total population size of the ith species, then the limiting equations for
each species form a system of autonomous differential equations

of the classical (Kolmogorov) type common in theoretical ecology. A linear

dependence of /^ on population sizes result in the famous Volterra system of
differential equations.
Rigorous treatments of separable McKendrick age-structured equations can be
found in [244] for a^ < +00 and [41], [378] for UM — oo. Theorems concerning
the existence and uniqueness of solutions, limiting equations, etc. can be found
in these references.
 POPULATION LEVEL DYNAMICS 123

3.2 The linear chain trick

If the birth and death rates >3 and S in the McKendrick age-structured model

are independent of age, then an integration of the partial differential equation

from a = 0 to +oc (assuming lim a _^ +00 p(t, a) = 0) yields the ordinary differen-
tial equation

for total population size p(t).

If f3 and 6 depend on age a, but do so in a certain mathematical way, inte-
grations of the partial differential equation can still yield ordinary differential
equations for p and/or other weighted total population sizes. The procedure,
called the linear chain trick, will only be illustrated here by means of selected

the right circumstances, it can also be used to derive dynamical equations for
population level quantities from models using structuring variables other than
age a as well (e.g., see the size-structured competition application below). The
linear chain trick relies on a special mathematical type of dependence of the
vital birth, death, and/or growth rates on the structuring variable, namely, a
polynomial multiplied by an exponential.
The linear chain trick has been extensively applied in the analysis of delay
differential equations [86], [303]. Also see [42J, [145], [199], [323].
Consider the McKendrick model equations

in which 6 is assumed independent of age a and

For this submodel fertility is an exponentially decreasing function of age. Define

the weighted population sizes

Then (assuming lim a ^ +00 /9(t,a) = 0)

and p(t,0) = bp2(t) yield the equation p[(t) = bp^t) -6pi(i) for p i ( t ) . An equa-
tion for p2 can be derived by first multiplying the partial differential equation
124 CHAPTER 3

for p by e ca before integrating. Thus,

The result is the two-dimensional system

for the population level quantities pi and p%. If f> and j3 are constants, this is a
linear system for which the origin is (globally asymptotically) stable if and only
if the inherent net reproductive number n = -^ < 1. If n > 1, both p\(t) and
Pz(t) grow exponentially.
If S and f3 are dependent on total population size pi (and/or p%), i-e., if

then we have the nonlinear plane autonomous system

From a knowledge of p\ and P2 the age distribution p is known from the formu-
las6

A biologically more realistic fertility function would vanish for newborns and
increase to a maximum at some age am > 0 before decreasing to 0. For example,
we could take

6
A "solution" p of the McKendrick model equations gives rise to a solution pair pi,p2 of
the planar system of ordinary differential equations. Since the converse is not true (the initial
conditions for pi and p% are not arbitrary), the two mathematical problems are not equivalent
in the sense that a solution of one yields a solution of the other. The relationship between
the asymptotic dynamics of the two systems of equations is studied in [42].
 POPULATION LEVEL DYNAMICS 125

Define

Differential equations for these weighted total population sizes can be derived
from integrations of the McKendrick partial differential equation (first multi-
plied by e~ca for p2 and by ae~ca for p3). The result is the system

of ordinary differential equations for the p,-.

More generally this procedure can be applied when age-specific, fertility is
modeled by expressions of the form

The integer k inversely measures the "width" of the "reproductive window'';

that is to say, for large k fertility is concentrated near the age am. Define the
weighted population sizes

Multiplying the McKendrick partial differential equation by ale. ca and inte-

grating the result from a = 0 to +oc, we obtain a system of ordinary differential
equations for the weighted population sizes Pi(t)

The derivation of this system of ordinary differential equations remains valid

when 6 = 6(pi) and b = b(pi), in which case the system becomes a nonlinear
system of ordinary differential equations for the fc + 2 weighted population sizes
PJ. The age distribution is given bv
126 CHAPTER 3

We consider two applications in which the linear chain trick is used. The
first involves an age-structured single species model. The second involves a
size-structured multispecies competition model. For an application of the linear
chain trick to age-structured predator-prey models see [201].

Age-specific fertility windows. In order to investigate some effects of age

dependent fertility we consider the McKendrick model with submodels

6 = constant > 0, (3 = b(pi)ake-ca, c=-£-, k£ 7[0, +00),

6(0) > 0, b'(p) < 1, limp_+00 b(p) = 0.

Here pi(t) = JQ °° p(t,a)da is total population size. The focus is on age de-
pendent fertility, so age dependence in mortality is ignored. The birth rate is
dependent both on age and on total population size p\. At any population size
Pi maximum fertility occurs at age am and as k gets larger fertility becomes
more concentrated near this age.
The Jacobian of the differential system (3.32) for the weighted population
sizes pi (3.31) is, at the extinction equilibrium Pi = 0,

The eigenvalues of this Jacobian are

Thus, the extinction equilibrium is (locally asymptotically) stable if the inherent

net reproductive number

satisfies n < 1 and unstable if n > 1.

There is a positive equilibrium if (and only if)
 POPULATION LEVEL DYNAMICS 127

namely,

We know from the general bifurcation results in section 2.2 that this positive
equilibrium is (locally asymptotically) stable for n close to 1. In fact this equi-
librium is stable for all n > 1. To see this, we calculate the Jacobian of the
differential system (3.32) at the positive equilibrium and obtain

The characteristic equation

can be rewritten as

For any complex number with Re £ > 0 the magnitude of the right-hand side
is less than 1 while that of the left-hand side is strictly greater than 1. Thus,
there is no root (eigenvalue) with Re£ > 0 and the equilibrium is (locally
asymptotically) stable.
An interesting conclusion suggested by this example is that age-dependent
fertility is not destabilizing no matter how narrowly defined the reproductive
window is or how late the "maturation period delay" am is. See [85] for a more
general consideration of this assertion.

Size-structured competition. In [94] and [96] a competition model for size-

structured species is considered (also see [387]). One motivation for considering
size-structured species comes from the question of whether body size confers
any competitive advantage to a species. See [129], [148], and [212] (and the
references therein) concerning this issue.
The model in [94] is for exploitative competition among species; that is to
say, individuals do not directly interfere with each other in their competition
128 CHAPTER 3

for the limiting resource, but only compete indirectly through their mutual
consumption of the resource. Mathematically, this means the equations for the
species dynamics are not coupled to one another. For example, competition in
a chemostat is usually considered exploitative [387].
The abundance (concentration) of the limiting resource is R = R(t) and
its dynamics are assumed governed by an ordinary differential equation R' =
f ( R ) in the absence of the species. Prototypical resource dynamics are the
"chemostat" model f ( R ) = r (R$ — R) arid the "renewable resource (logistic)"
model f ( R ) = r (l - f ) R.
Each size-structured species is modeled by equations of the form (2.30), i.e.,

where s is some measure of body size (to be specified below). All newborns are
assumed to have the same size S(, at birth. The submodels for the vital rates
b and 7 are based on the assumption that growth, reproductive, and metabolic
rates scale exponentially to body length / [435]. We will assume, for simplicity,
that the death rate 6 is not size specific, i.e., 6 = constant > 0.
Birth and growth rates depend on the resource uptake rate. It is assumed that
the resource uptake rate u(.R)/ M scales to body length I and that some portion of
the consumed resource is utilized for metabolism, leaving a net amount n(R)l^
available for growth and reproduction [212]. If w denotes an individual's weight
(assumed proportional to volume / 3 ), then weight change is proportional to
n(R)l^, i.e.,

where K € (0,1) is the fraction of consumed resource that is allocated to growth

(and 1 — K is the fraction allocated to reproduction) and ij is a conversion factor
relating weight to resource units. Then since 7 = ^f, we have

and

where £ is a resource-to-offspring conversion factor and Wb is the weight of

newborns. (Notice the simplifying assumption has been made that individuals
of all sizes s > s& are reproductive.) The coefficients
 POPULATION LEVEL DYNAMICS 129

are the reproductive and growth efficiency coefficients, respectively.

We are interested in the competition of m species for the resource R. There-
fore the quantities in the model equations above are subscripted by i £ I[l,m].
Let Si denote the size st, at birth for the iih species. The model equations for
m size-structured species are

In [96] the cases p, — 3 and JJL — 2 are considered where resource consumption
scales to body volume and to body surface area, respectively. In the first case
the structuring variable is taken to be body volume so that s — I 3 . In the second
case s = I. Both cases are amenable to the linear chain trick analysis. We will
consider here only what turns out to be the mathematically simpler case \i = 3.
See [94] or [387] for the details of the case // = 2.
For n = 3 the model equations become

Differential equations for the total population sizes and the total population
volumes

can be derived as in the age-structured examples above, namely, by integrating

the partial differential equation from s — s, to -foe (multiplying the equation
by s in the case of t>j). The result is the system of ordinary differential equations

Since the first of these equations is uncoupled from the remaining equations we
can determine the asymptotic dynamics from the system
130 CHAPTER 3

Total population numbers are then determined from the variation of constants
formula

A straightforward calculation of the derivative of the average volume

shows that

Thus, on the time scale T» = J0 [ni(R(v))} 1 dv, the average volume of an indi-
vidual Vi satisfies the famous logistic equation

and hence equilibrates to

This limiting average volume v°° for individuals of the species is independent
of the asymptotic dynamics of the species as a component of the system (3.33)
and, in particular, is independent of whether the species goes extinct or survives,
equilibrates or oscillates! We define v°° to be the "size" of the species. This
"species size" is related to the original model parameters by the formula

The outcome of the competition described by the system (3.33) depends on

the designated submodels for the resource dynamics f ( R ) and for the uptake
rates Ui(R) and rii(R). Most submodels will satisfy the minimal conditions

First of all, we note that the positive cone R™+1 is forward invariant under
the flow denned by (3.33). That t>;(0) > 0 implies that Vi(t] > 0 for all t follows
from
 POPULATION LEVEL DYNAMICS 131

Suppose that R(0) > 0. If /(O) = 0, then

for all t. On the other hand, if /(O) > 0 and if t\ > 0 were the first time at
which R(t) vanishes, then 0 > R'(ti) = /(O), a contradiction that implies that
no snch first time t\ exists.
Second, if we assume that
(3.35) solutions of R' = f ( R ) , R(0) > 0, are bounded for t > 0,
then we can argue that positive solutions of (3.33) remain bounded for t > 0.
That R(t) is bounded for t > 0 follows from R' < f ( R ) and standard comparison
-1
theorems. To show that i>,;(£) is bounded, define v = R + X)I=i (ai + ft) vi
and obtain, setting 6 = min£j > 0.

or finally the inequality

In as much as R, and hence /(-R), have been shown to be bounded for t > 0
it follows that v(t) is bounded for t > 0. Since v^ is known to be positive, it
follows from the definition of v that each v^ is bounded for t > 0.
In summary, under the assumptions (3.34) and (3.35) all solutions of (3.33)
with R(Q) > 0 and £,(0) > 0 are positive and bounded for t > 0.
Consider the case when the resource dynamics are given by

With this submodel for /, (3.33) is a model for micro-organisms growing in

a chemostat [387]. In this case, the resource equilibrates exponentially to RQ
in the absence of the micro-organism populations from the chemostat. Let us
assume a rnonotoiiic uptake rate, u'^R) > 0. The so-called Michaelis/Menten
(or Holling II or Monod) uptake rate u — "O^R is the most famous example.
Furthermore, we assume that the net uptake rate rij is a fixed fraction ipt of the
total uptake rate, so that n,.(_R) = ^^(R). Finally, we assume that 6t — r for
all i. This latter assumption, almost always made in chemostat models, means
that the removal of micro-organisms per unit time due to the washout rate r
in the chemostat far exceeds the mortality rate of the micro-organisms per unit
time (which is then ignored in the model). The system (3.33) now becomes
132 CHAPTER 3

This competition system has been well studied [387], [43]. It is known to have
global equilibrium dynamics only and that at most one species can survive (in
keeping with the competitive exclusion principle). The asymptotic dynamics are
determined by the so-called break-even concentrations \i denned as the (unique)
solution of the equation

All species will die out except the one with the smallest break-even value Aj (pro-
vided it exceeds a minimum threshold value, namely, the input concentration
RO of the limiting resource). Thus, if A^ = min Aj, then

and

In the second case

defines the equilibrium state of the surviving species.

How does species size vf° relate to competitive efficiency? Does the winning
species, i.e., the species with the smallest break-even concentration A i? have the
largest species size vf^l The answer is that in this model there is no simple
relationship between species size and competitive success, at least not without
further restrictive assumptions about the species involved. For example, con-
sider m similar species, similar in the sense that they all have the same resource
uptake rate Ui(R) = u(R) and the same metabolic demands tl>i = 1/1. In this
case the species with largest value of oti + fti will have the smallest break-even
concentration

However, the species with the largest on + f3i is not necessarily the species with
the largest species size
 POPULATION LEVEL DYNAMICS 133

For more discussion of this issue see [94]. Similar conclusions are obtained in
the case \i = 2 for which the model equations are

3.3 Hierarchical models

For those models considered so far in this chapter the effects of population
density have been modeled by assuming that vital rates are functions of one or
more weighted total population sizes of the forms

for discrete and continuous models, respectively. A more general situation is

one in which the effects of population density that are felt by an individual
are dependent on that individual's class (age. size, etc.), so that the weights
Lj,; (or uj(ct)) depend on the structuring variable. Weights w, v (or o;(f*,a)) then
measure the effect that individuals of class i (or a) have on the vital rates of
individuals of class j (or a). One situation in which this is the case occurs when
there is a hierarchy established by the structuring variable that determines an
individual's access to resources.
For example, contest competition can be defined as the situation when no
individual in a class of lower rank (say. of lower age or smaller size) can affect
the amount of resource available to an individual of greater rank (of greater age
or size). Scram,ble competition, on the other hand, can be defined as the opposite
extreme when every individual can affect the amount of resource available to
any other individual in the population [297]. In these circumstances the vital
rates in a structured model would depend on functionals of the form

in discrete models and

in continuous models. We call such models hierarchical models.

For both discrete and continuous hierarchical models, it turns out to be pos-
sible to derive dynamical equations for total population size.
134 CHAPTER 3

3.3.1 Continuous age-structured models. In a McKendrick age-struc-

tured model with a^ = +00 the integrals

give the number of individuals of age less than a and the number of individuals
of age greater than a, respectively. In a hierarchical McKendrick model the
fertility and death rates are functions of Y and O, so that

Note that the total population size is given by

These models are not of the form (2.6) studied in section 2.2. However, models
of the form (3.37) are amenable to considerable analysis by means of a single
ordinary differential equation for p ( t ) , Heuristically, this equation can be derived
as follows. Define

Noting that daY(t, a) = p(t, a) we obtain

These equalities, together with an integration of the differential equation in

(3.37) from a = 0 to +00, lead to the scalar ordinary differential equation

for p — p ( t ) . The relationship between this equation and the model equations
(3.37) is rigorously studied in [104], where it is shown, under certain conditions,
that a solution p ( t , a) of (an appropriately formulated version of) the initial
value problem for (3.37) defines a solution p(t) — JQ p(t, a)da of (3.38) and
vice versa. These conditions are
 POPULATION LEVEL DYNAMICS 135

It is also shown in [104] that the initial value problem for (3.37) is well posed
under these conditions.
When 6 = t>(Y,O) and 13 = d(O,Y) are independent of time t, then (3.38)
is an autonomous, scalar ordinary differential equation and therefore has only
monotonic dynamics. Thus, the asymptotic dynamics at the population level
are easily analyzed by means of the real valued function B(p) - D(p) of the real
variable p.
If p(t) approaches an equilibrium p^ as t —» +00, the dynamics of Y(t, a) and
p ( t , a ) can be deduced as follows. From the definition of Y and (3.37) follows

or in coordinates (T, a) = (t - a, a)

This is a scalar ordinary differential equation satisfied by Y as a function of

Q with T as a parameter. Classical theorems for the continuous dependence
of solutions on parameters imply that lim^+30 Y(t,a) = yoo(a) uniformly for
a = a restricted to a compact interval where yx(a) is the solution of the scalar
ordinary differential equation

or equivalently

Note that 6 > 0 implies that lim a _ +00 yx(a) = p^,. With regard to p ( t , a), for
t>a

where

From lim t _ + 0 0 y(t) = px and \hnt-++00Y(t,a) = y<x>(«) it follows that p ( t , a )

approaches a limit as t —> +oc, uniformly for a restricted to any compact
interval, and this limit is given by the formula
136 CHAPTER 3

THEOREM 3.3.1. Consider the autonomous age-structured McKendrick model

(3.37) where 6 = S(Y, O) and 0 = 0(Y, O) satisfy (3.39). Then the total popula-
tion size p(t) satisfies the scalar ordinary differential equation

and consequently is monotonic in time t. If pit) asymptotically approaches an

equilibrium p^, then Iim4^+00 p ( t , a) = [^(a) uniformly for a restricted to any
compact interval where pcc(a) is defined by (3.40).

Contest versus scramble competition. One problem of interest concerns

the relative advantages or disadvantages of different types of intraspecific com-
petition among the members of a population. Lomnicki [297] argues that contest
competition is generally more advantageous and is therefore expected to be more
common in nature. If the competition hierarchy is based upon chronological age
(or at least correlates closely with chronological age), then one approach to the
study of contest competition is to utilize the hierarchical McKendrick model
(3.37) and the results in Theorem 3.3.1. Two issues are involved in address-
ing Lomriicki's tenet about the advantage of contest over scramble competition:
how are these two fundamental types of competition modeled using (3.37) and
how is an appropriate comparison made between the two models?
Let R denote the amount of a limiting resource available for consumption
by an age-structured population. Let c be the fraction of this amount that is
available to an individual. In the presence of competition this fraction c will
be dependent on population density in some way. A reasonable assumption for
age-structured contest competition, in which older individuals are dominant, is
that for an individual of age a the fraction c is a decreasing function of the
number of older individuals, i.e., c = cc(O), where O = O(t,a). This assumes
that younger individuals have no effect on an individual's resource availability.
(If the hierarchy is such that younger individuals are dominant, then c is a
decreasing function of Y = Y(t,a).) For scramble competition the fraction
c could be taken as a decreasing function c = cs(p) of total population size
p = p(t). Thus, under contest competition the amount of resource available to
an individual of age a is Rcc(O), while under scramble competition the amount
is Rcs(p). If u is the resource uptake rate, then for scramble competition we have
a resource uptake rate us — u(Rcs(p)) and for contest competition we have a
resource uptake rate uc = u(Rcc(O)).
The next modeling assumptions provide submodels for the vital rates 6 and 0
in the McKendrick model (3.37). We concentrate in this example on the effects
that resource competition have on fertility. Therefore, we make the simplifying
assumption that the death rate is a constant 6 = SQ > 0. With regard to
fertility, we assume that the birth rate 13 is proportional to the resource uptake
 POPULATION LEVEL DYNAMICS 137

rate. Thus,
scramble competition: (3 = /30u(Rcs(p}), i30 > 0. b' = bo > 0,

contest competition: ,3 — /30u(Rcc(O)), 80 > 0, 6' — 60 > 0.

where

and both competition fractions cs and cc satisfy

In order that the total amount of resource available to the whole population
remains less that R, i.e., that JQ Rcpda < R, the competition fractions shoul
satisfy

With these designations of the submodels in (3.37) we have the models for
intraspecific scramble and contest competition considered in [230].
In order to make an appropriate comparison between the two types of compe-
tition, we impose the criterion that for both types the same amount of resource
is utilized (for a given density function p ( t , a ) ) . Thus, we require

If R is independent of age a, this requires the relationship

between the competition fractions cs and c.c.

Given that both types of competition are utilizing the same total amount of
resource, which mode is more "advantageous" ? What is meant by ''advanta-
geous"? We consider two criteria for comparing the two types of competition:
equilibrium level and equilibrium resilience. Assuming the equilibrium to be
stable, resilience is defined to be the smallest magnitude of the real parts of all
eigenvalues of the Jacobian; the resilience provides a lower bound for the rate
of approach to the equilibrium. Larger resilience values mean a "more stable"
equilibrium in that the return to equilibrium from small perturbations is faster.
By Theorem 3.3.1 the dynamics of total population size, for scramble and for
contest competition, are governed by the ordinary differential equations
scramble competition:

contest competition:
138 CHAPTER 3

respectively, where

and where c satisfies (3.42). The quantity

is the inherent net reproductive number. From the assumptions placed on u and
c it follows that both fs and /c equal 1 at p = 0 and decrease monotonically to
0 as p —> +00. It is straightforward to deduce from these facts that each of the
scalar, autonomous ordinary differential equations in (3.43) has a unique positive
equilibrium for n > 1 and that this equilibrium is globally asymptotically stable
(for initial conditions p(Q) > 0). For n < I the extinction equilibrium p =
0 is globally asymptotically stable (for initial conditions p(0) > 0) for both
equations.
Consider the case n > 1, and let ps > 0 and pc > 0 denote the globally
stable positive equilibria of the scramble and contest competition models (3.43),
respectively. If u"(z) < 0 for 0 < z < R. it follows from Jensen's Inequality
that fc(p) < fs(p) for all p > 0 (see [230]). Since fc(pc) = n = fa(pa) and sinc
fc and fs are decreasing functions, it follows that pc < ps. A similar argumen
shows that if u"(z) > 0 for 0 < z < R, then pc > ps. This result shows that the
concavity properties of the nonlinearity in the resource uptake rate determines
which of the two types of competition has the higher equilibrium level.
For example, the Holling II (Michaelis/Menten or Monod) uptake rate

satisfies u"(z) < 0 for all z > 0. Thus, for this model scramble competition is
more advantageous than contest competition in the sense that a higher total
population equilibrium size is attained under scramble competition. In the case
of a Holling III uptake rate

there is a change in concavity in u as a function of z. For low resource levels R,

u"(z] > 0 for 0 < z < R and contest competition leads to a higher equilibrium
level. For larger resource levels it is possible that this can be reversed (see [230]
for an example).
A comparison can also be made of the equilibrium age densities under scram-
ble and contest competition. These densities are 5opse~6°a and 6opce~s°a, re-
spectively, which show that whatever relationship is borne between the popula-
tion level equilibria is also borne by every age class.
 POPULATION LEVEL DYNAMICS 139

The conclusions above concerning the relative advantage and disadvantages

of scramble and contest competition are based upon a comparison of relative
equilibrium levels. If other criteria are used, the conclusions may not be the
same. For example, as mentioned above, resilience of equilibrium can also be
used as a criterion, greater resilience being interpreted as advantageous. In this
case, it turns out (at least for n > 1 near 1) that the conclusions above are
exactly reversed [230]!
These results show that when making such assertions about relative advan-
tages or disadvantages, properties of the nonlinearity and the methods of com-
parison are crucial.
Intraspecific predation (cannibalism) can also be studied using hierarchical
models under the assumption that victims are younger (smaller) than cannibals.
See [30], [100], [104], [114].

3.3.2 Discrete matrix models. An analog of Theorem 3.3.1 for hierarchical

age-structured populations can be obtained for a general class of structured
matrix models. Consider a projection matrix P = T + F with

where GJ is the probability that an individual of class j survives one unit of time,
Tij is the fraction of those survivors that moves to class i (hence, X^i Tij = 1
for all j), (3j is the number of surviving offspring from an individual in class j,
and yi • is the fraction of the offspring that lies in class i (hence Y^iLi ^Pij ~ 1 f°r
all j } . Submodels for survival <TJ and fertility (3i are derived under the assumption
that these quantities depend on the density yi of individuals of rank less than
?., i.e.,

in a population of size p — ym+i- This is done in the following way. Let (3 £

C° (.R2, -R+) be a function such that the per capita birth rate of an individual
is P ( z , p ) when the total population size is p and the density of individuals of
lower rank is z. Define (3j to be the average

For example, if rank is determined by age, the birth rate of individuals in the
jth age c}ass ranges from /3(yj,p] for the youngest in the class to (3(yj^i,p] =
(3(yj + Xj,p) for the oldest in the class; /?• is the average for the jth class.
Similarly, define
140 CHAPTER 3

where a € C°(R2, [0,1]) is a function such that v(y,p) is the probability an

individual will survive one unit of time when the density of younger individuals
is y and the total population size is p.
Cumulative sums of the equations

yield

Since tne ^ast equation in this system

of difference equations uncouples as a scalar equation for total population size.
Thus, letting

we obtain the system of difference equations

for the cumulative distributions j/i(t), i e I[2,m], and the uncoupled scalar
equation

for total population size p(t) = ym+i(t), where s, b e C° (Rl,R+) are functions
defined by

This is a generalization of the result in [444] (where it is assumed that all

newborns lie in the first class).
The dynamics of total population p(t) can be studied by means of the many
analytical and graphical techniques available for one-dimensional maps. If the
 POPULATION LEVEL DYNAMICS 141

dynamics of p(t) are known, the dynamics of y(t) (and consequently of the
original class distribution x ( t ) ) can be studied by means of the resulting nonau-
tonomous system (3.46). For example, if p(t) tends to an equilibrium p, then
the limiting equation of (3.46) is the autonomous system

For the special case of a Leslie matrix it is shown in [444] that if p(t) equili-
brates, then y i ( t ) and Xi(t) also equilibrate. The following theorem is a discrete
analog of Theorem 3.3.1 [444].
THEOREM 3.3.2. Suppose that the projection matrixT+F described by (3.45)
is a Leslie matrix, i.e., TJJ = 0 for all j ^ i + 1, T^J+I = I and (p^ = 0 for
all i ^ 1 and j, <^1 • = 1. Assume that the functions (3, a and their partial
derivatives dp(3, dpa are bounded on R2^. Suppose that a solution p(t] > 0,
t G 7[0,+00), of (3.47) satisfies \imt^+00p(t) = p > 0, and let Xi(t) be the
solution of the matrix model whose initial conditions satisfy Y^iLi xi(ty = P(0)i
Xi(0] > 0. Then lim t _ +oc yi(t) = ^ and l\mt^+oo xl(t) = xt exist and are given
by the formulas

Contest versus scramble competition revisited. Consider the comparison

of contest and scramble competition in a hierarchical population. Assume that
survivorship is a constant a(z.p) — CTQ > 0. For scramble competition let
0 = /30u(Rcs(p)), where Rcs(p) is the per capita resource availability for any
individual and u(Rcs(p)} is the amount of available resource consumed by an
individual. For contest competition let 13 = j3Qu (Rcc(p — z } ) , where Rcc(p — z]
is the resource available to an individual in a population of size p with a density
z of lower ranking individuals and u(R.cc(p — z)} is the amount of resource
consumed by an individual. In both cases (30 > 0 is the number of offspring
produced per unit resource. The resource uptake rate u and the functions cs
and cc are assumed to satisfy (3.41) and (3.42). The average per capita resource
availability for the ith class under scramble competition is u (R,cs(p)} and under
contest competition is

We require that the same amount of resource be divided up under both types
of competition, so that
142 CHAPTER 3

and hence

From (3.47) we have the scalar difference equations

scramble:

contest:

for total population sizes in the two cases. Here the subscript on cc has been
dropped. The equilibrium equations are

scramb
contest:

where n = u(R)/30 (1 — a)~ and fs and fc are defined by (3.44). These equi-
librium equations have the same form as those of the continuous age-structured
model considered in section 3.3.1. As a result we can draw the same conclusions
concerning the relationship between the equilibrium levels for scramble and con-
test competition and how this relationship depends on the concavity of u. This
discrete matrix hierarchical model is not necessary age structured, however.

3.4 Total population size in age-structured models

In section 3.1.2 an uncoupled limiting equation for total population size is de-
rived for separable McKendrick age-structured equations (3.28). In those types
of equations the fertility rate /? is independent of population density. In this
section we consider some models in which the focus is instead on density depen-
dence in fertility and for which a limiting equation for total population size can
be derived [85].
Consider the McKendrick model (2.5) with OM = +00 and

where the normalization (2.15) holds, i.e.,

so that n is the inherent net reproductive number. If we assume that newborns

(age o = 0) are not reproductive and fertility is bounded, then we have
 POPULATION LEVEL DYNAMICS 143

From the McKeridrick partial differential equation (2.5(a)) and its initial con-
dition p0(a) € Ll (R\, R\) we obtain

where the total birth rate p ( t , 0) is determined by the birth equation in the
McKendrick model (2.5(b)), namely,

where

An integration of the partial differential equation (2.5(a)) with respect to a from

a — 0 to +00 yields, after an integration by parts, the equation

The equations (3.48)- (3.49) constitute a coupled system of equations for the
total birth rate /o(t,0) and total population size p(t). This system is equivalent
to the system consisting of (3.48) and the equation

This equation can be rewritten

and simplified, by an integration by parts, to the uncoupled equation

where £(x,y,t) ==• t/>(x,i) — nb(x,t')ye 6t. Since £(x,y,t) tends to 0 (uniformly
for x and y on compact intervals) as t —> +00. we obtain the limiting equation

for total population size [278], [325] to which we now turn our attention.
The integro-differential equation (3.50) has, of course, the extinction equilib-
rium q = 0. Positive equilibria are solutions q €E R\ of the equation
144 CHAPTER 3

or equivalently of the equation

Stability can be studied formally by linearizing (3.50) at an equilibrium q > 0

to obtain the linear integro-differential equation

where

and investigating solutions y = z exponential in time. This results in the

characteristic equation

for complex z. If the characteristic function f ( z ) has no complex roots satisfying

Re z > 0, then the equilibrium q is stable; if there exists a root with Re z > 0,
then the equilibrium is unstable [325].
For the extinction equilibrium q = 0 the roots of the characteristic function

are z = — 6 < 0 and the solutions of the equation

If n < 1 and Rez > 0, the left-hand side is less than 1 in magnitude. Hence
the equation has no roots satisfying Rez > 0. If n > 1, then the left-hand side,
as a function of x — Rez, is greater than 1 at x — 0, strictly decreases to 0
as x —» +00, and consequently has a unique positive real root. As a result we
obtain the familiar result that the extinction equilibrium loses stability as n is
increased through 1.
Let q > 0 be a positive equilibrium. By Theorem C.O.I (see Appendix C) q
is unstable if

Suppose, on the other hand, that /(O) > 0. A calculation shows

 POPULATION LEVEL DYNAMICS 145

FlG. 3.5. A plot of the graph defined by ni>(q) — 1 gives the equilibrium bifurcation diagram
in the n, q plane together with stability properties.

Theorem C.O.I (see Appendix C) implies that the number of roots of f(z) lying
in the right half complex plane is equal to 2m, where m is the number of times
the image of the upper imaginary axis winds around the origin. Consider f(ir]
for r > 0. Note that the second factor in the product on the right-hand side of
f ( i r ) (which by (3.51) lies in the unit circle centered at z = 1) has an argument
lying between —?r/2 and Tr/2 and that the first factor has an argument lying
strictly between —Tr/2 and Tr/2. Thus, the product has an argument lying
strictly between —TT and TT, which means that the image f ( i r ) of the imaginary
half axis ir, r > 0, does not cross the negative real axis. Thus, m — 0 and there
are no roots in the right half complex plane. It follows that the equilibrium
q > 0 is (locally asymptotically) stable.
In summary q = 0 loses stability as n increases through 1 and a positive
equilibrium q > 0 is stable if dpv(q) < 0 and unstable if dpv(q] > 0.
This result has a nice interpretation in a bifurcation diagram that plots the
equilibrium q against the inherent net reproductive number n. Since equation
(3.51) defines the graph of the positive equilibrium, an implicit differentiation
shows that those equilibria lying on an increasing branch of the bifurcation curve
are stable, while those lying on a decreasing branch of this curve are unstable.
See Fig. 3.5. Note in particular that the common, density regulation assumption
dpb(p. a) < 0 implies that all positive equilibrium are stable. Thus, in such a
model, age-dependent fertility (and, in particular, a maturation period delay)
cannot destabilize an equilibrium.
This page intentionally left blank
 APPENDIX A
Stability Theory for Maps

A.I Linear maps

Consider the linear nonhomogeneous equation

where P is an 77; x m matrix of real numbers. For a given sequence g :

/[O, +oc) —» Rm a forward solution is defined to be a sequence x : /[O. +DC) —»
Rm that satisfies (A.I) for all t e /[(), +00). For a given sequence g : /(-oc, -1]
—>• /?."' a backward solution is a sequence x : J(—oc,0] —» R'" that satisfies (A.I)
for alH e /(-OG,-1].
LEMMA A. 1.1. (Variation of constants formula). For any g : 7[0,+OG)
—> 7?T" £/ie initial value problem

has a unique forward solution, and this solution is given by the, formula

Proof. For t = 0 we have x ( l ) = Px0+g(0) = Px(0) I-e/(0). For t e J[l. +oc)

Thus, (A.2) satisfies ( A . I ) for all t e 7[0.+oc).

We are interested in forward solutions of (A.I) that are bounded and also
those that tend to 0 as t —> +00. For a sequence x : /[O. +00) —> Rm define
147
148 APPENDIX A

are Banach spaces under the norm ||-||+ . We are also interested in backward
solutions of (A.I) that are bounded or tend to 0 as t —> — oo. For a sequence
x : 7(-oo,0] -> Rm define ||x||_ = sup^^^ x ( t ) \ . The sets

are Banach spaces under the norm ||-||_ . Our first goal is to develop modified
variation of constants formulas for solutions of (A.I) that lie in these spaces.
Assume that P is hyperbolic, i.e.. P has no eigenvalues satisfying |A| = 1.
Assume that P has Jordan form, i.e.,

where

each eigenvalue £ of Ju satisfies |£| > 1, and each eigenvalue A of Js satisfies

|A| < 1. (This can be done without loss of generality since a linear change of
coordinates will put P into Jordan form.) We write J = Ps + Pu, where

For any integer t € /[I, +00)

and Pt = P* + PU- For any constant r]s satisfying

there exists a constant cs > 0 such that

For notational convenience define

 STABILITY THEORY FOR MAPS 149

and

where Im, and Im.n are the m., x ms and mu x mu identity matrices, respectively.
Note that P® = /, and P® = Iu. For any constant i]v satisfying

there exists a, constant cu > 0 such that

For t,, t'2 G 7(-oc,+oc)

and for t £ /(-oc. +00)

The following lemma contains a modified variation of constants formula for

forward bounded solutions of (A.I).
LEMMA A.1.2. Assume that P is hyperbolic and has Jordan form (A.3). For
a e Rm and g £ BS+ define.

where Ps anil Pu are given by (A.4). Then

(a) ,r is a forward solution. o/(A.l);
(b) x 6 BS^:
(c) g e BS^ implies that x e BS^.
Conversely.
(d) «/:r e /^S"+ (or BS^) is a forward solution of ( A . l ) , i/^en i/;ere is an
a £ /?'" such that ,r is given by (A.9).
Proof, (a) First of all. we note that the infinite series in the definition (A.9) of
;r(/j is convergent. Let ;/ = niax{?/ s . ?/ u } arid c. — max{cs.cu}. Then 0 < r\ < 1
and
150 APPENDIX A

For t = 0 we have

For t £ /[I, +00) we have

(b) For t e /[I, +00)

Thus, x e BS+.
(c) Let e > 0 be arbitrary and choose T = T(e] > 1 so that t > T(e) implies
that
 STABILITY THEORY FOR MAPS 151

Then for t > T(e] + I

and therefore 0 < lirnsup t ^ +00 \x(t)\ < £. Since e is arbitrary it follows that
lim t _^ +00 \x(t)\ — 0 and consequently x G BS^.
(d) From the variation of constants formula (A.2) we have, for t G /[I, -f oo).

The last three terms, on the right-hand side, are bounded for t G /[I, +oc), and
therefore the first term must be bounded for t G 7[l.+oo). This implies that
the initial condition XQ must satisfy

As a result
152 APPENDIX A

Let a be any vector in Rm such that

Then P^XQ = P*a and x(t) is given by the formula (A.9).

The following lemma contains a modified variation of constants formula, for
backward bounded solutions.
LEMMA A.1.3. Assume that P is hyperbolic and has Jordan form (A.3). For
a£Rm and g e BS~ define

where Ps and Pu are given by (A.4). Then

(a) x is a backward .solution of (A.I);
(b) xeBS";
(c) g e BSy implies that x G BS$ .
Proof, (a) First of all, we note that the infinite series in the definition of x is
convergent. Let r\ = max{r?s,77u} and c = max{cs, cu}. Then rj < 1 and

F o r t e /(-oo.-1]

Therefore, x(t) is a backward solution of (A.I),

(b) For ie/(-oo,0]

Thus, x e BS-.
 STABILITY THEORY FOR MAPS 153

(c) Let £ > 0 be arbitrary and choose T = T(e) < 0 so that t < T(e) implies
that

For t < T(e)

and therefore 0 < limsup t __ oc x(t)\ < s. Since e > 0 is arbitrary, it follows
that lim t ^_,c x(t)\ — 0 and consequently x e BS^.
A linear change of coordinates in Lemmas A.1.2 and A.1.3 produce formulas
for matrices P not in Jordan form.

A.2 Linearization of maps

Consider the nonlinear difference equation

For r > 0 and xe £ Rm denote

Assume that

for some r-o > 0. Write equation (A. 10) as

154 APPENDIX A

where x ( t ) has been replaced by x(t) — xe, A = J x f ( x e ) is the Jacobian of /

evaluated at xe, and h : B(TO, 0) —> Rm is twice continuously differentiable with
/i(0) = 0 and Jxh(0) = 0.
The equilibrium xe e .R™ is called stable if for each £ > 0 there exists a
6 — 6(e) > 0 such that \x(0) — xe < 8 implies that \ x ( t ) — xe < £ for all
t € /[O,+00). If x is not stable, it is called unstable. The equilibrium xe
is an attractor if there exists a 6 > 0 such that \x(0) — xe\ < 6 implies that
linit^+oo x ( t ) — xe\ — 0. If xe is a stable attractor it is called (locally) asymp-
totically stable.
A — J x f ( x e ) is hyperbolic if no eigenvalue satisfies |A| = 1. Denote the spec-
tral radius of A by s(A) = {max |A| : A is an eigenvalue of A} . Theorem A.2.1
below appears in [156, Theorem 4.21] although with the unnecessary assumption
that A 7^ 0 (i.e., that / is locally invertible). This theorem can also be found
in [278, Theorem 9.14], where a proof is given using Liapunov functions. We
will give a proof based on the variation of constants formula and the following
discrete version of Gronwall's inequality.
LEMMA A.2.1. // k(i) > 0 and y ( i ) > 0, i = 0,1, 2 . . . . , are sequences of real
numbers such that y(Q) < m and

for some constant m > 0, then

Proof. Let the sequence z ( t ) > 0 be denned by z ( 0 ) = m and

By induction

For t 6 /[O, +oc)

and by induction

From (A. 14) we obtain (A. 13).

 STABILITY THEORY FOR MAPS 155

THEOREM A.2.1. (Fundamental Theorem of Stability). Assume. (A.11). //

s ( J x f ( x e ) ) < l j then the equilibrium x — xe of (A.10) is asymptotically stable.
Proof. We show that x = 0 is an asymptotically stable equilibrium of equation
(A.12). Let A = J x f ( x e ) . It is left as an exercise to show that a sequence solves
(A. 12) if and only if it solves

For any // satisfying ,s (Jxf(xK)} < r] < 1 there is a constant c > 0 such that
||A*|| <crf. t e 7[0.+oc). Thus,

Let c' = max {1, e}. For £ = -^ ( \ - rj) choose b > 0 so that for all x € B(6, 0)
the inequality

holds. This is possible because h is twice continuously differentiable and h(0) =

0 and ,/x/i(0) = 0. For as long as x(t) e B(6,Q). t e 7(1, +00), we have

or, letting y(i) =7/-*|ar(t)|,

Using rn = e! XQ , k(i) = c'rj le in Lemma A.2.1 we obtain

and hence

for as long as x(t) e B(6,0), t € /[O,+00). Note that ( r / + l ) / 2 < 1. By

choosing XQ < mm {6,8/<•'}. we have (by induction) that x(t) & B(S, 0) and
hence (A.15) holds for all t 6 /[O.+oc). It follows from inequality (A.15) that
the equilibrium x = 0 of (A. 12) is both stable and an attractor.
What happens if s ( J x f ( x e } ) > 1 ? A l o c a l s t a b l e m a n i f o l d t h e o r e m w h e n /
is invertible can be found in [9], [193]. [336]. We prove a local stable manifold
theorem that does not require / to be invertible.
156 APPENDIX A

Let x(t, XQ) denote the solution of (A.10) starting initially at XQ, i.e., x(0, XQ) =
XQ. If f ( x ) is q times continuously difFerentiable, then an induction argument
shows that, for each t £ 7[0, +00), the solution x = x(t, XQ) of the initial value
problem x(t + 1) — f ( x ( t ) ) , x(0) = XQ, is q times continuously differentiable in
•TO-
Let Es denote the span of the generalized eigenvectors associated with eigen-
values of A satisfying |A| < 1. Es is the stable manifold of the linearization, i.e.,
of the linear homogeneous system

Let Eu denote the span of the generalized eigenvectors associated with eigen-
values of A satisfying |A| > 1. Eu is the unstable manifold of (A.16). Define

THEOREM A.2.2. (Stable Manifold Theorem). Assume (A.11) andx fJ( x e )

is hyperbolic. In a sufficiently small neighborhood of the equilibrium xe, there is
(a) a manifold Ws of dimension dimEs passing through xe tangentially to
Es such that XQ £ Ws implies that lim^ +00 x(t,a:o) — xe, and
(b) a manifold Wu of dimension dim Eu passing through xe tangentially to
Eu such that XQ 6 Wu implies the existence of at least one backward solution
x(t,xo) such that \\mt-t-oo X(I,XQ) = xe.
Ws is called the (local) stable manifold of x = xe and Wu the (local) unstable
manifold of x = xe.
Proof. Without loss of generality assume that a linear change of variables
has been performed so that A has Jordan form (A.3) with Ps and Pu defined
by (A.4). Choose any a 6 Rm. By Lemma A.1.2 a solution x 6 BS+ of the
equations

is a forward bounded solution of (A.12). For fixed a let N(x,a) denote the
operator from BS+ to BS+ defined by the right-hand side of (A.17). First we
show that AT is a contraction map from a ball £Q"(£) into itself, at least for
sufficiently small |a .
From (A. 11) it follows that corresponding to any e > 0 there exists a S =
<5(e) > 0 such that for all x, y <E £^(<5)
 STABILITY THEORY FOR MAPS 157

Choose £ = (1- T ] ) ( l + r i ) ' l / 2 c ' , where c' = m a x j l . c } . For x £ S^(6(£),0) we

have

Therefore, for all / 6 /[O, +00}

If a is chosen so that

then ||A r (.r,o)|| + < 6. Since h(x(t)) £ BS^. it follows from Lemma A.1.2(c)
that N ( x ( t ) . a ) e BSt. Thus, for a e Rm satisfying (A.18)

Next we show that N is a contraction. For x.y & ^(6(e)) the inequalities

show that

This shows that for all a satisfying (A. 18) the operator N is a contraction from
Eo"((5(e)) into itself. We conclude that for all such a, the equation (A.17) has
a unique fixed point x(t.a) G £j(<5(c)). This fixed point is a forward bounded
solution of (A. 12) that tends to 0 as t —> +oc.
Finally, we need to describe the set of initial conditions XQ corresponding to
the set of fixed points x(t) = x(t.xo) found above. From (A.17) XQ satisfies the
equation
158 APPENDIX A

What kind of solution set do these equations define near XQ = 0? Let

where

If we denote x(t,xo) = x(t,xs,xu), then (A.19) can be written

from which it is seen that xs = as is arbitrary and xu = xu(as) satisfies the

equation

By the implicit function theorem this equation has a twice continuously differ-
entiable solution xu(as) satisfying xu(0) = 0 for as K 0. Thus, the manifold of
initial conditions XQ is described parametrically by the equations

From equation (A.20) and the fact that hu is of second order near x = 0 it
follows that V 0 ,z u (0) = 0. This proves part (a).
(b) Choose any a £ Rm. By Lemma A. 1.3 a solution x £ BS~ of

is a backward bounded solution of (A.12). For fixed a, let N(x,a) denote the
operator from BS~ to BS~ defined by the right-hand side of this equation.
Arguments similar to those in (a) above show that N is a contraction map from
a ball £J7(<5) m*° itself, at least for sufficiently small [a . The unique fixed point
x = x(t,xo) is a backward solution that tends to 0 as t —»• —oo. The initial
conditions of this set of solutions are parameterized by

where xs(au) solves the equation

 STABILITY THEORY FOR MAPS 159

What can be said about solutions whose initial points do not lie on either
of the two manifolds described in Theorem A.2.2? For flows this question is
answered by the classical Hartman/Grobman Theorem. This theorem has an
analog for maps that are invertible [193]. For noninvertible maps this classical
theorem does not hold, as the following example shows.
Consider the m = I dimensional map x(t + 1) = x2(i). The Jacobian at the
equilibrium x = 0 is A — 0 and the linearization is y(t + 1) = 0. Does there
exist a homeomorphism H : R —> R such that H(x(t + I , Z D ) ) = Al+1H(xo).
i.e.. such that H ( x 2 ( t ) ) = 0 for all t € /[O, +oc)? Such a map would have to
send all positive x to 0 and therefore could not be one-to-one.
We will replace the Hartman/Grobman Theorem with theorems based on the
following lemma (which is a special case of Lemma 1 in [84]).
LEMMA A.2.2. Suppose, that X and Y are Banach spaces and L : X —> Y
is a dosed linear operator. Suppose that there exists a subspace S of X such
that the restriction of L to S, denoted by LS, is one-one and onto Y. Suppose
that h : X -> Y , h(0) - 0, satisfies \h(x) - h(y)\x < e x - y\y for all x,
y e £(<5) = (;r e X : \\x\\ < 6} and some constants e, 6 > 0. If £\Lsl\\ < 1,
then there exists a, constant c > 0 such that there is a one-one bicontinuous map
between all solutions of Lx = 0 in H(c<5) and all solutions of Lx = h(x) in £(<*>).
Let X = Y = BS+. Lemma A.2.2 can be applied to the bounded linear
operator L : X —> Y defined by

in the following way. Let Xi C R"' be the subspace of initial conditions x(0) for
which the linear homogeneous system x(t + 1) = Ax(t) has a forward bounded
solution. Let X-2 be any supplementary subspace: R1" = X\ © X?. Define
S = {x G BS+ : x(0) (E X-2\ . To apply Lemma A.2.2, for an appropriate h, we
need only show that the restriction of L to S is one-one and onto Y. For this
purpose we assume that

for all g 6 DS+ there exists at least

(A.21) one forward solution x e BS+ of
x(t + l ) = Ax(t)+g(t).

Under this assumption write ,r(0) G Xi (B X? as x(0] — x± + x%, X{ G A'; and

define x\(€) to be the solution of (A. 16) with ./,'i(0) = x-\. By definition of X\,
it follows that x i ( t ) e BS+. Then z ( t ) - x ( t ) - x i ( t ) lies in BS+ and satisfies
z(t + 1) = Az(t) + g(t], z(0) = x-2 <E X<2, i.e., Lz — g and z e S. This shows
that LS is onto Y = BS+. In order to show that LS is also one-one, suppose for
g 6 BS+ there exist x, y 6 S such that Lx = Ly = g. Then L(x — y) — 0 and the
difference x — y £ S is a forward bounded solution of (A.16) with initial condition
x(0) - y(0) 6 X2. By the definition of A'2 this implies that :c(0) - j/(0) = 0 and
hence x ( t ) = y ( t ) , t e /[O, +00).
We are interested in the case when h is higher order at x = 0. In this case
the condition on h in Lemma A.2.2 is satisfied for s, 6 > 0 sufficiently small. By
160 APPENDIX A

Lemma A.1.2 condition (A.21) holds if A is hyperbolic. In other words, Lemma

A.2.2 applies to (A.lO)-(A.ll) when A = J x f ( x e ) is hyperbolic.
Similar (virtually identical) arguments apply with BS+ replaced by BS^.
THEOREM A.2.3. Assume that (A.11) holds and A = J x f ( x e ) is hyperbolic.
There exist constants c, 6 > 0 such that the following hold.
(a) There is a one-one bicontinuous map between forward solutions of (A. 10)
lying in £ + (<5) and forward solutions of its linearization (A. 16) lying in £ + (c<5).
(b) There is a one-one bicontinuous map between forward solutions o/(A.10)
lying in £0"(6) and forward solutions of its linearization (A. 16) lying in EQ"(C£).
Similar arguments can be made for the bounded linear operator L : X —> Y
defined by

with X = Y = BS~ or BS^. In this case XT. C Rm is the subspace of

initial conditions x(Q) for which the linear homogeneous system (A. 16) has
a backward bounded solution. X% is any supplementary subspace, and S =
(x € BS- : z(0) & X2} .
THEOREM A.2.4. Assume that (A.11) holds and A - J x f ( x e ) is hyperbolic.
There exist constants c, b > 0 such that the following hold.
(a) There is a one-one bicontinuous map between backward bounded solutions
o/(A.10) lying in S~(<5) and backward solutions of its linearization (A.16) lying
m£-(c<5).
(b) There is a one-one bicontinuous 'map between backward solutions of (A. 10)
lying in £^(<5) and backward solutions of its linearization (A.16) lying in £^"(c<5).
 APPENDIX B
Bifurcation Theorems

B.I A global bifurcation theorem

The set of characteristic values of an m x m matrix L is the set of reciprocals
of the nonzero eigenvalues of L. If AO is a characteristic value of L, then there
exist nonzero left and right characteristic vectors q and v such that v = X0Lv
and qr = \oqTL. Necessarily a characteristic value is nonzero. An algebraically
simple characteristic value is a simple root of the characteristic polynomial
det (/ — XL). A characteristic value AO is geometrically simple if the left and
right null spaces of I — X^L have dimension 1 and are therefore spanned by
single characteristic vectors q and v, respectively. Denote

An algebraically simple characteristic value is geometrically simple, but the

converse is not necessarily true.
Consider the algebraic equation

where

uniformly on compact A intervals.

Here ft is an open neighborhood of B™ (the closed positive cone). A pair
(A,.r) 6 R1 x ft that satisfies (B.2) is called a solution pair. Any trivial pair
(A,0) is a solution pair for all A € -R1. A solution pair (A,:r) is called nontrivial
if x ^ 0. The closure of the set of nontrivial solution pairs is denoted by S. A
solution pair is nonnegative (positive) if x > 0 (x > 0). A continuum of pairs
(A, x) € Rl x Rm is a closed and connected set in R1 x Rm. Let OR™ denote the
boundary of fi™. i.e., the set of nonnegative vectors x > 0 that are not positive.
A pair (A, x) G R1 x dR™ is called a boundary pair.
We are interested in the existence of nonnegative nontrivial (and also of pos-
itive) solution pairs of (B.2). Specifically, we will look for a continuum in 5
161
162 APPENDIX B

that contains a trivial solution pair (Ao,0) for a critical value AQ (i.e., that
"bifurcates" from (A 0 ,0)).
First of all, we observe that the only candidates for such a critical bifurcation
value of A are the characteristic values of the matrix L. To see this suppose there
exists a sequence (Aj,Xj) of nontrivial pairs such that limj_, +00 (Ai,Xi) = (Ao,0).
Define the unit vectors y; = x*/ |xj|. The unit sphere in Rm is compact so we
can assume (by extracting a subsequence if necessary) that lim^+00 j/j = y^,
Ij/oo = 1. Since (B.3) implies that limt_,+ 0 0 /i(Aj,Xi)/|xj| = 0, we find from
(B.2), after dividing by |xj and passing to the limit i —*• +00, that y^ = XoLy^.
Thus, AQ is a characteristic value of L.
Is it a sufficient condition for a bifurcation to occur that AO is a characteristi
value of L? The answer is in general "no," as the example

shows. In this example AO = 1 is the only characteristic value of L (which is

the 2 x 2 identity matrix). For any A, x\ — 0 implies that x% = 0 and vice
versa. Thus, for a nontrivial solution pair (A, x) we must have x\ ^ 0 and
x'2 7^ 0 and therefore 1 — A — xj^xf 1 = —xfxj 1 , which leads to the contradiction
x\ + x\ = 0. This example has no nontrivial solution pairs. Note that AQ = 1 is
not a geometrically simple characteristic value.
We will use theorems from [354] to obtain the existence of nontrivial solutions
that bifurcate from (Ao, 0) when A 0 is a geometrically simple characteristic value
of L. (In fact, theorems from [354] more generally allow for AO to be of add
geometric multiplicity.) Define the continuous function a : Rm —> #"1 by

Since a(x) = x for x > 0, the set of nonnegative (positive) solution pairs of the
equation (B.2) is identical to that of the equation

The advantage of working with this equation is that it is globally defined in x,

i.e., Jl = Rm. The function g satisfies (B.3). If we assume that AO has a positive
characteristic vector, then Theorem 1.25 of [354] implies that S, the closure
of the set of nontrivial solution pairs of this equation, contains a continuum
C that "bifurcates" from (Ao,0) (i.e., (Ao,0) € C) such that in some open
neighborhood of (Ao, 0) the solution pairs from C'/ {(Ao, 0)} are positive. These
positive solution pairs are then positive solution pairs of the original equation
(B.2). Thus, we have a local bifurcation result for positive solution pairs. We
turn now to the problem of determining the global extent of these positive
solution pairs.
 BIFURCATION THEOREMS 163

Theorem 1.40 in [354] implies that the (locally positive) continuum C satisfies
one of two alternatives: C is either (1) unbounded in R^ x Rm or (2) it contains
a trivial solution pair (Ai,0) for which AI ^ AQ is another characteristic value
of L. (This theorem also implies the existence of another continuum satisfying
these same alternatives which is locally negative. The theorem guarantees that
these two continua are globally distinct, however, i.e., they do not meet except
at the point (Ao,0).)
We are interested in only nonnegative (and positive) solution pairs from the
continuum C. Let C have the topology inherited from R1 x Rm. The subset
S+ = {(A,x) 6 C : x > 0} of positive solution pairs from C is open. Let SQ~
be the maximal open subset of S+ whose closure C+ is connected and contains
(Ao,0). The set SQ is nonempty since the continuum C is positive in a neigh-
borhood of (A 0 ,0). By considering the two alternatives (1) and (2) above we
conclude the following about the continuum C+: either <7 + /{(Ao,0)} is un-
bounded arid contains only positive solution pairs or C+ contains a boundary
point (A*,x*) ^ (Ao,0). In the latter case we can say more. If x* — 0, then,
as we showed above. AI is a characteristic value of L and it has a nonnegative
characteristic vector (since y% > 0 implies that y^, > 0).
We summarize these results in the following theorem.
THEOREM B.I.I. Suppose that (B.3) holds and L has a geometrically simple
characteristic value \Q which has a positive right characteristic vector v > 0.
Then there exists a continuum C+ of nonnegative solution pairs of (B.2) con-
taining (AQ.O) for which C+/ {.R1 x dR™} is nonempty and consists of positive,
solution pairs. The following alternatives hold: either
(i) CQ~ = C + /{(Ao,0)} is unbounded in Rl x R'™ and contains only positive
solution pairs, or
(ii) C+ contains a boundary pair (A*,x*) / (Ao,0).
In case (ii) if x* = 0, then A* =£ AQ is another characteristic value of L which
has a nonnegative characteristic vector.
This theorem could be succinctly described by stating that, under the assumed
conditions, there exists a "positive branch" of solutions of (B.2) that bifurcates
from (Ao,0) and "connects" to the boundary dR™ of the positive cone R™ (oo
being considered as on the boundary).

B.2 Local parameterization

In this section we consider a parametric representation of the bifurcating branch
of nontrivial solutions guaranteed by Theorem B.I.I. This representation is
sometimes called the Liapunov/Schmidt expansion or parameterization of the
branch.
Consider first the linear nonhomogeneous algebraic equation x = Lx + / for
an unknown x £ Rm, with / £ Rm and the mxm matrix L given. There exists
a unique solution x = (I — L) / i f and only if I — L is nonsingular. Suppose
however, that / — L is singular and that its null spaces (as in (B.I) with AQ = 1)
have dimension 1. If x = Lx + f has a solution, then <fx = qrLx + if j implies
that (f (I — L) x — <f f and, since q £ Q. that qT f = 0. Conversely, if qr f = 0,
164 APPENDIX B

then the general solution has the form x = ev + zt where e € Rl is an arbitrary

scalar and z is any particular solution. Out of this general solution we can
extract a unique solution that satisfies VTX = 0 (by choosing £ = -vrz/vrv).
Let V-1 and Qx denote the spaces orthogonal to V and Q, respectively, i.e.,

Then Rm = Q@QL = V®V^. For each / e QL the equation x = Lx + f has a

unique solution x 6 V1-. This correspondence defines a linear transformation of
Q-1 into I/-1. Let G be the matrix associated with this transformation so that
Gf is the unique solution lying in V^. The general solution of x = Lx + f has
the form x — ev + Gf: where e & Rl is arbitrary.
Consider the nonlinear algebraic equation

when / — L is singular and its null space is one dimensional. Suppose that
x e Rm is a solution. Since Rm = V ® V^ we can write x — ev + z, e e R1,
z € V-1. It must be the case that f ( x ) e Q-1, i.e.. q r f ( x ) = 0. and therefore
that z = Gf(x). Thus, e and z € V^ satisfy the pair of equations

Conversely, if e e .R1 and z € V^- satisfy these two equations, then x = ev + z

satisfies the equation (B.4). In this sense equation (B.4) is equivalent to the
pair of equations (B.5).
Consider the equation with parameter

The problem is to find nonzero solution pairs (A,x) = (X,ev + z) g Rl x V ©V^

for A near a characteristic value AO of L. i.e., for a value such that 7 — \oL is
singular. Suppose that AQ is a geometrically simple characteristic value of L
with left and right characteristic vectors q and v. Define 77 = A — AO, and writ

With f ( x ) taken to be rjLx-\-h(\o+Tj, x) the equivalent pair (B.5) can be written

We view this as a system to be solved for rj & Rl and z € V^ as functions of

£ » 0. Using the implicit function theorem on the second equation (B.7(b)), we
can locally solve uniquely for
 BIFURCATION THEOREMS 165

where z is just as smooth in its arguments as h is in its arguments. Assume

that

times continuously
differentiable near
and near x — 0 uniformly on
compact A intervals.

Then z ( s , 77) is fc + 1 times continuously difi'erentiable near e = r/ — 0. We can use

the implicit function theorem because the right-hand side of the second equation
(B.7(b)) vanishes when z = 0, r/ = £ — 0 and has an iiivertible Jacobian with
respect to z at this point (equal, namely, to the identity). Notice that if e — 0
in equation (B.7), then z = 0 is a solution. By the uniqueness of the solution
obtained from the implicit function theorem it follows that 2(0,77) = 0 for all
small 77 and we can "factor an e out of z'\ i.e., we can write z(e.r/) = ££(£,7
where ((£,77) is k times continuously differentiable near e — r/ = 0. Also, from
equation (B.7(b)) it follows that ((0,0) = 0. With equation (B.7(b)) solved
the equivalent system (B.7) reduces, in a neighborhood of (Ao,0), to the single
scalar equation obtained from (B.7(a)), i.e.,

where h(e, rj) = e~lh (Ao + 77, £t' + e((e, r/)) is k times continuously differentiabl
near e = 77 = 0 and /i(0,77) = 0. By the implicit function theorem this equatio
can be locally solved for a (unique) Ck solution 77 — rj(e), rj(0) = 0 provide
qTv ^ 0 (since the right-hand side of the equation vanishes at £ ~ v — 0
and has a derivative with respect to 77 equal to Ay 1qTv at this point). Using
z(e] = £((£, »7(e)) we obtain a Ck solution pair A = AQ +7/(c), x — ev + z(e) of
(B.6) for e near 0.
THEOREM B.2.1. Suppose that (B.8) holds and L has a geometrically simple
characteristic value AO which has left and right characteristic vector K q and v.
Assume that v > 0 and qTv ^ 0. Then in a neighborhood of (\,x) = (Ao,0) the
branch of nontrivial solutions of equation (B.6) guaranteed by Theorem B.I.I
can he written, for e G ( — £ o , e < j ) , £Q > 0, in the form

where 77 e C f c ( ( - £ 0 , £ o ) , R 1 ) , z e Ck((-£0,e0).V-L) and 77(0) = 0, z(0) =

z'(0) = 0.
Approximations of the bifurcating branch can be obtained by a calculation of
r/(0) and 2"(0). These quantities can be found by differentiating the equation
x — XLx + h(X, x) twice with respect to e and evaluating the result at e = 0.
This yields
166 APPENDIX B

Here h(v, v) is the vector whose ith component is

where Hi = [dkdjhi(\o,0)] is t h e m x m Hessian of hi with respect to x evaluated

at (A, x) = (A 0 ,0). Since AQ is a characteristic value of L, the necessary condition
for the solvability of this equation for z'(0) is qr (2r]'(Q)Lv + h(v, v)) = 0, which
yields the formula

The solution for 2"(0) is then

Thus, if fc > 2, we have the parameterization

 APPENDIX C

Miscellaneous Proofs

Proof of Theorem 1.1.2(b). Part (b) of Theorem 1.1.2 follows immediately from
the following lemma.
LEMMA C.O.I. Suppose that q(t) satisfies the scalar equation q(t + 1) =
c(t)q(t), t 6 /[O. +oc), where lim t _^ +00 c(t) = 0^.
(a) //|coo < 1. then limt^+oc \q(t)\ = 0 (exponentially).
(b) Suppose that c^] > 1. Ifc(t) ^ 0 for all t € 7[0,+oc), then\\mt^+OQ \q(t)\
= +00 (exponentially).
Proof, (a) Pick e > 0 such that \c00\+e<\. There exists a T = T(e) > 0 such
that t e /[T, +00) implies that \c(t)\ < c^ +e. For t € /[T, +00), g(f + 1) =

0.
(b) Pick e > 0 such that c^l - e > 1. There exists a T = T(e) > 0 such
that t e /[T, +00) implies that c(t)| > Ic^ - e. For i 6 I[T, +cx). g(i + 1) -
<l(T)^Tc(i) and |9(t + l)| > |g(r)|(| Coc |- £ r T+1 . Hence lim(^+oc |g(t)| =
+00.
Proof of Theorem 1.2.l(b). We show that the origin 0 is (a) an isolated
invariant set in R™ and (b) equal to its own stable set. It follows from Theorem
4.1 in [235] that the origin is uniformly persistent with respect to the origin.
By definition the set {0} is isolated if there exists a closed neighborhood U
of {0} in R^ such that {0} is the largest invariant set in U. Let M' denote
the largest invariant set in U and choose x(Q) € M'. Then x(0) > 0 and the
forward orbit of .r(0) remains in M' and hence is bounded. Thus x(0) lies on the
stable manifold of 0. However, according to the remarks prior to the statement
of Theorem 1.2.1, the local stable manifold intersects the closed positive cone
at the origin only when r > 1. Thus, x(0) = 0 and M' = {0}.
(b) By the Theorems A.2.2, A.2.3, and A.2.4 there is no point z(0) e U/ {0}
such that limt_ +00 x(t)\ = 0. Thus, the stable set of {0} is itself.
Proof of Lemma 1.2.2. For small e > 0 the Jacobian Jx(\,x) = A + \B +
Jxr(\,x) evaluated at the equilibrium (1.31) can be written Jx(X,x) = JQ +
Ji£ + O(e2). From

167
168 APPENDIX C

we obtain

The dominant eigenvalue and eigenvector of the Jacobian Jx(\,x) also have e
expansions

Placing these expressions into Jx(\,x)e = C,e we obtain, to first order in e, the
equation

Necessary for the solution of this linear algebraic system for e\ is the orthogo-
nality condition u'r(('(0)t; - Jiv) = 0. which yields

The numerator is

Using the formula in (1.31) for AX, we obtain

and hence

Proof of Theorem 3.1.3. Define p(t) = a(t)/b(t). Then from equation (3.9)

follows. It is not difficult to show that the solution of this equation is (p(t) —
il}(t)/wTil>(t), where i/> is the solution of the equation

For simplicity we give a proof when L has a basis of eigenvectors v\ = v,

v % , . . . ,vm. For the more general case see [76], [289]. Write </?(0) = ^2'iLi civi-
The matrix p(i)I + L has eigenvalues Aj + p(t) and eigenvectors DJ. From the
solution
 MISCELLANEOUS PROOFS 169

we have

Divide numerator and denominator by H/Uo (^ + P(^))' an(i define

Then

By the following lemma, lim t _ +00 Ui(t) = 0 for i e I[2,m] and limit (3.10)
follows.
LEMMA C.0.2. / / A T > |A t | for i e /[2,m] and 0 < p(^) < p0 < +oc /or
t e /[O,+00). tfienlimt^+3C !!,(£) = 0 /ori e /[2,m].
Proof. The ratio |Aj +x| / (Ai + x) is continuous for x > 0. Moreover, the
assumption on AI implies that this ratio is less than 1 and therefore it is bounded
away from 1 on bounded x intervals. Thus, 0 < |A; + x / (Ai + x) < m; for some
ml < 1 for all x € [0,/?0]. It follows that H(i)| < m*.

Roots of characteristic polynomials. The following theorem shows that

under certain conditions the number of roots of f ( z ) = z + c + K ( z ) lying in
the right half complex plane is related to the number of times the image /(ir),
r > 0. of the upper imaginary axis ''winds around the origin."
THEOREM C.O.I, (see [85]). Assume that ki = J0+oc \k(a)\ da < +00, and
consider the analytic function f (z) = z+c+K(z), where K(z) = f0 k(a)e~azd
is the Laplace transform ofk(a).
(a) // /(O) < 0. then f ( z ) has a positive real root.
(b) Assume that /(O) > 0 and /() °°a\k(a)\da < +00. Assume that f ( z )
has no purely imaginary roots. Then arg/(+ioo) — \— 2mir for some integer
m € I (—00, +oc) and the number of roots of f ( z ) lying in the right half complex
plane is 2m.
Proof, (a) For real z — x, lirn x ^ +00 f ( x ) — +oc and the result follows by the
intermediate value theorem.
(b) By the argument principle the number of roots lying inside the semidisk
\z < r, Rez > 0, is

where d(r) is the boundary of the semidisk. Consider first the integral over the
semicircular part \z — r. Rex > 0. of the boundary d ( r ) . which we denote by
170 APPENDIX C

9i(r). Consider the difference

For r > \c\ + ki

since \K(z}\ < fci for Re 2 > 0. Thus,

Now K'(z) = J+°°ae-zada and limr_+00 \K'(rew)\ = 0 for -f < 6* < f.

By the dominated convergence theorem linir-^+oo /0_rTw 2 |^'('*eie)|d0 = 0 and
hence

It follows that

Thus, the number of roots v of f ( z ) in the right half plane Re z > 0 is

Since f (z) — f ( z ) it follows that arg/(—ir) = — arg/(ir) and v = | -

^arg/(+icxD), where arg/(+zoo) = lim^^+oo arg/(zr).
Since limr_^+oc Im /(ir) = limr_>+00 (ir + c + K(ir)} = +zoo, it follows that
arg/(+ioo) = ^ — 2m7r for some integer m € /(—oo,+CXD) and the number of
roots in the right half plane equals 2m.
 Bibliography

[1] J. D. Aber, Why don't we believe the models?, Bull. Ecol. Soc. Amer. 78 (1997), 232-
233
[2] D. M. Adams and A. R. Ek, Optimizing the management of uneven-aged forest stands,
Can. J. For. Res. 4 (1974), 274-287
[3] F. R. Adler, Coexistence of two types on a single resource in discrete time, J. Math.
Biol 28 (1990), 695-713
[4] W. G. Aiello, H. I. Freedman, and J. Wu, Analysis of a model representing stage-
structured population growth with state-dependent time delay, SI AM J. Appl. Math.
52 (1992), 855-869
[5] W. C. Allee, Animal Aggregations, University of Chicago Press, Chicago, 1931
[6] W. C. Allee. Cooperation Among Animals, Henry Schuman, New York, 1951
[7] L. J. S. Allen, A density-dependent Leslie matrix model, Math. Biosci. 95 (1988), 179-
187
[8] L. ,1. S. Allen, M. P. Moulton, and F. L. Rose, Persistence in an age-structured popula-
tion for a patch-type environment, Nat. Res. Modeling 4 (1990), 197-214
[9] K. T. Alligood, T. D. Sauer, and J. A. Yorke, Chaos: An Introduction to Dynamical
Systems, Springer-Verlag, Berlin, 1997
[10] O. Arino, D. K. Axelrod. and M. Kimmel (eds.), Mathematical Population Dynamics,
Lecture Notes in Pure and Appl. Math. 131, Dekker. New York, 1991
[11] M. Artzrouni, On the dynamics of a population subject to slowly changing vital rates,
Math. Biosci. 80 (1986), 265-290
[12] P. Auger, Stability of interacting populations with class-age distributions, J. Theor.
Biol. 112 (1985), 595-605
[13] H. T. Banks, Some remarks on estimation techniques for size-structured population
models, Frontiers in Mathematical Biology (S. A. Levin, ed.), Lecture Notes in Biomath.
100, Springer-Verlag, Berlin, 1995, 609-623
[14] H. T. Banks, L. W. Botsford, F. Kappel, and C. Wang. Modeling and estimation in
size structured population models, Mathematical Ecology, World Scientific Publishing,
Teaneck, N.I. 1988, 521-541
[15] H. T. Banks, L. W. Botsford, F. Kappel, and C. Wang, Estimation of growth and
survival in size-structured cohort data: An application to larval striped bass (Morone
Saxatitis), J. Math. Biol. 30 (1991), 125-150
[16] H. T. Banks and B. G. Fitzpatrick, Statistical methods for model comparison in param-
eter estimation problems for distributed systems. J. Math. Biol. 28 (1990). 501 527
[17] H. T. Banks and B. G. Fitzpatrick, Estimation of growth rate distributions in size
structured population models, Quart. Appl. Math. 49 (1991), 215—235

171
172 BIBLIOGRAPHY

[18] H. T. Banks, B. Fitzpatrick, and Y. Zhang, Estimation of distributed individual rates

from aggregate population data, Differential Equations and Applications to Biology and
to Industry (M. Martelli, K. Cooke, E. Cumberbatch, B. Tang, and H. Thieme, eds.),
World Scientific Publishing, Singapore, 1996. 13-22
[19] H. T. Banks, F. Kappel, and C. Wang, Weak solutions and differentiability for size
structured population models, Internal. Ser. Nurner. Math. 100 (1991), 35-50
[20] H. T. Banks and K. Kunisch, Estimation Techniques for Distributed Parameter Systems,
Birkhauser, Boston, 1989
[21] T. H. Barr, Approximation for age-structured population models using projection meth-
ods, Comput. Math. Appl. 21 (1991), 17-40
[22] H. Barclay, Models of host-parasitoid interactions to determine the optimal instar of
parasitization for pest control, Natural Resource Modeling 1 (1986), 81—103
[23] J. R. Beddington and C. A. Free, Age structure effects in predator-prey interactions,
Theoret. Population Biol. 9 (1976), 15-24
[24] S. Bhattacharyya and C. G. Chakrabarti, On the dynamics of age-structured population,
Bull. Calcutta Math. Soc. 83 (1991), 297-304
[25] J. Beddington, Age distribution and the stability of simple discrete time population
models, J. Theor. Biol. 47 (1974), 65-74
[26] J. Beddington and C. A. Free, Age structure effects in predator-prey interactions, The-
oret. Population Biol. 9 (1976), 15-24
[27] T. S. Bellows Jr. and M. P. Hassell, The dynamics of age-structured host-parasitoid
interactions, J. Anim. Ecol. 57 (1988), 259-268
[28] J. Bengtsson, Smaller zooplankton species are not superior in exploitative competition:
A comment on Persson, Amer. Nat. 129 (1987), 928-931
[29] M. O. Bergh and W. M. Getz, Stability of discrete age-structured and aggregated delay-
difference population models, J. Math. Biol. 26 (1988), 551-581
[30] K. Blayneh, A Hierarchically Size-Structured Population Model, Ph.D. dissertation,
University of Arizona, Tucson, AZ, 1995
[31] S. P. Blythe, R. M. Nisbet, and W. S. C. Gurney. Formulating population models with
differential aging, Population Biology (H. I. Freeman and C. Strobeck, eds.), Lecture
Notes in Biomath. 52 (1983), Springer-Verlag, Berlin, 133-140
[32] S. P. Blythe, R. M. Nisbet, and W. S. C. Gurney. The dynamics of population models
with distributed maturation periods, Theoret. Population Biol. 25 (1984), 289-311
[33] L. W. Botsford, Optimal fishery policy for size-specific, density-dependent population
models, J. Math. Biol. 12 (1981), 265-293
[34] L. W. Botsford. Individual state structure in population models, Individual-Based Mod-
els and Approaches in Ecology (D. L. DeAngelis and L. J. Gross, eds.), Chapman and
Hall, New York, 1992, 213-236
[35] E. N. Boulanger, Small perturbations in nonlinear age-structured population equations,
J. Math. Biol. 32 (1994), 521-533
[36] F. Brauer, Nonlinear age-dependent population growth under harvesting, Int. J. Com-
put. Math. Appl. 9 (1983), 345-352
[37] F. Brauer and M. Zhien, Stability of stage-structured population models, J. Math. Anal.
Appl. 126 (1987), 301-315
[38] M. Brokate, Pontryagin's principle for control problems in age-dependent population
dynamics, J. Math. Biol. 23 (1985), 75-101
[39] J. L. Brooks and S. I. Dodson, Predation, body size, and composition of plankton,
Science 150 (1965), 28-35
[40] J. Buongiorno and B. R. Michie, A matrix model of uneven-aged forest management,
Forest Sci. 26 (1980), 609-625
 BIBLIOGRAPHY 173

[41] S. N. Busenberg and M. lannelli, Separable models in age-dependent population dy-

namics, J. Math. Bioi 22 (1985), 145-173
[42] S. N. Busenberg and C. C. Travis, On the use of reducible-functional differential equa-
tions in biological models, J. Math. Anal. Appl. 89 (1982), 46-66
[43] G. J. Butler and G. S. K. Wolkowicz, A mathematical model of the chemostat with a
general class of functions describing nutrient uptake. SIAM J. Appl. Math. 45 (1985),
138-151
[44] W. A. Calder III, Size, Function, and Life History, Harvard University Press. Cam-
bridge, MA, 1984
[45] A. Calsina and J. Saldana, A model of physiologically structured population dynamics
with a nonlinear individual growth rate, J. Math. Biol. 33 (1995), 335-364
[46] A. Calsina and J. Saldana, Asymptotic behaviour of a model of hierarchically structured
population dynamics, J. Math. Biol. 35 (1997), 869-907
[47] Y. L. Cao, J. Fan, and T. C. Gard, The effects of state-dependent time delay on a
stage-structured population growth model, Nonlinear Anal. 19 (1992), 95-105
[48] Y. L. Cao and H. I. Freedman, Oscillation and global attractivity in stage-structured
population models, Canad. Math. Bull. 36 (1993). 129 138
[49] H. Caswell, A general formula for the sensitivity of population growth rate to changes
in life history parameters, Theoret. Population Biol. 14 (1978), 215-230
[50] H. Caswell, Stable population structure and reproductive value for populations with
complex life cycles, Ecology 63 (1982), 1223 1231
[51] H. Caswell, Optimal life histories and the maximization of reproductive value: A general
theorem for complex life cycles. Ecology 63 (1982), 1218-1222
[52] H. Caswell, Life history theory and the equilibrium status of populations, Ame.r. Nat.
120 (1982), 317-339
[53] H. Caswell, Life cycle models for plants, Lectures on Mathematics in the Life Sciences
18 (1986), 171-233
[54] H. Caswell, Approaching size and age in matrix population models. Size-Structured
Populations (B. Ebenman and L. Persson, eds.), Springer-Verlag, Berlin, 1988, 85-105
[55] II. Caswell. Matrix Population Models, Sinauer Associates, Inc. Publishers, Sunderland.
MA, 1989
[56] H. Caswell and A. M. John, From the individual to the population in demographic
models, Individual-Based Models and, Approaches m Ecology (D. L. DeAngelis and L.
J. Gross, eds.), Chapman and Hall, New York, 1992, 36-61
[57] W.-L. Chan and B.-Z. Guo, Global behaviour of age-dependent logistic population mod-
els, J. Math. Biol. 28 (1990), 225-235
[58] B. Charlesworth, Natural selection in age-structured populations, Lectures on Mathe-
matics in the Life Sciences 8 (1976), 69-87
[59] B. Charlesworth. Evolution in Age Structured Populations, Cambridge University Press,
Cambridge, MA, 1980
[60] M. Chipot. On the equations of age-dependent population dynamics. Arch. Rational
Mech. Anal. 82 (1983), 13-26
[61] M. Chipot and L. Edelstein, A mathematical theory of size distributions in tissue culture,
J. Math. Biol. 16 (1983), 115-130
[62] B. Cipra, Chaotic bugs make the leap from theory to experiment, SIAM News, Vol. 30,
No. 6 (July/August, 1997)
[63] A. J. Coale, The Growth and Structure of Human Populations: A Mathematical Inves-
tigation. Princeton University Press, Princeton, NJ, 1972
[64] J. E. Cohen, Ergodicity of age structure in populations with Markovian vital rates T:
Countable states, J. Amer. Statist. Assoc. 71 (1976), 335 339
174 BIBLIOGRAPHY

[65] J. E. Cohen, Ergodicity of age structure in populations with Markovian vital rates II:
General states, Adv. in Appl. Probab. 9 (1977), 18-37
[66] J. E. Cohen, Ergodicity of age structure in populations with Markovian vital rates III:
Finite-state moments and growth rate; An illustration, Adv. in Appl. Probab. 9 (1977),
462-475
[67] J. E. Cohen, Ergodic theorems in demography, Bull. Amer. Math. Soc. 1 (1979), 275-
295
[68] B. D. Coleman, On the growth of populations with narrow spread in reproductive age.
I. General theory and examples, J. Math. Biol. 6 (1978), 1-19
[69] C. S. Coleman and J. C. Prauenthal, Satiable egg eating predators, Math. Biosci. 63
(1983), 99-119
[70] D. Cooke and J. A. Leon, Stability of population growth determined by 2 x 2 Leslie
matrix with density-dependent elements, Biometrics 32 (1976), 435-442
[71] R. F. Costantino and R. A. Desharnais, Population Dynamics and the 'Tribolium'
Model: Genetics and Demography, Monographs on Theoretical and Applied Genetics,
Vol. 13, Springer-Verlag, Berlin, 1991
[72] R. F. Costantino, J. M. Gushing, B. Dennis, and R. A. Desharnais, Experimentally
induced transitions in the dynamics behaviour of insect populations, Nature 375 (May
18, 1995), 227-230
[73] R. F. Costantino, R. A. Desharnais, J. M. Gushing, and B. Dennis, Chaotic dynamics
in an insect population, Science 275 (January 17, 1997), 389-391
[74] R. F. Costantino, J. M. Cushing, B. Dennis, R. A. Desharnais, and S. M. Henson,
Resonant population cycles in alternating habitats, Bull. Math. Biol. 60 (1998), 247-
275
[75] S. M. Henson, R. F. Costantino, J. M. Cushing, B. Dennis, and R. A. Desharnais,
Multiple Attractors, Saddles and Population Dynamics in Periodic Habitats, preprint
[76] K. M. Crowe, A Discrete Size-Structured Competition Model, Ph.D. dissertation, In-
terdisciplinary Program in Applied Mathematics, University of Arizona, Tucson, AZ,
1991
[77] K. M. Crowe, Persistence in Asymptotically Autonomous Systems of Difference Equa-
tions, Fields Institute Technical Report, Toronto, ON, Canada, 1993
[78] K. M. Crowe, A nonlinear ergodic theorem for discrete systems, J. Math. Biol. 32 (1994),
179-191
[79] K. M. Crowe, Nonlinear ergodic theorems and symmetric versus asymmetric compe-
tition, Structured-Population Models in Marine, Terrestrial, and Freshwater Systems
(Tuljapurkar and Caswell, eds.), Chapman and Hall, New York, 1997, Chapter 17
[80] K. M. Crowe and J. M. Cushing, Optimal instar parasitization in a stage-structured
host-parasitoid model, Natural Resource Modeling 8 (1994), 119-138
[81] P. Cull and A. Vogt, Mathematical analysis of the asymptotic behavior of the Leslie
population matrix model, Bull. Math. Biol. 35 (1973), 645-661
[82] P. Cull and A. Vogt, The periodic limit for the Leslie model. Math. Biosci. 21 (1974),
39-54
[83] P. Cull and A. Vogt, The period of total population, Bull. Math. Biol. 38 (1976), 317-
319
[84] J. M. Gushing, An operator equation and bounded solutions of integro-differential sys-
tems, SIAM J. Math. Anal. 6 (1975), 433-445
[85] J. M. Cushing, Six Lectures on Integrodifferential Equations in Population Dynamics,
Proc. CIME Session on Mathematics in Biology II, Ciclo 1979, Florence, Italy
[86] J. M. Cushing, Integrodifferential Equations and Delay Models in Population Dynamics,
Lecture Notes in Biomathematics 20, Springer-Verlag, Heidelberg, 1979
 BIBLIOGRAPHY 175

[87] J. M. Gushing, Bifurcation of time periodic solutions of the McKendrick equations

with applications to population dynamics, Camp, and Math. Appl. 9 (1983), 459-478
(Reprinted in Advances in Hyperbolic Differential Equations, Vol. 1, M. Witten, ed.,
Pergamon Press, Elmsford, NY, 1983.)
[88] J. M. Gushing, Equilibria in structured populations, J. Math. Biol. 23 (1985), 15-39
[89] J. M. Gushing, Global branches of equilibrium solutions of the McKendrick equations
for age-structured population growth, Comp. and Math. Appl. 11 (1985), 175-188
(Reprinted in Advances in Hyperbolic Differential Equations, Vol. 2, M. Witten, ed.,
Pergamon Press, Elmsford, NY, 1985.)
[90] J. M. Gushing, Periodic McKendrick equations for age-structured population growth,
Comp. and Math. Appl 12A (1986), 513-526
[91] J. M. Gushing. Equilibria in systems of interacting age-structured populations, J. Math.
Biol. 24 (1987), 627-649
[92] J. M. Gushing, Oscillatory population growth in periodic environments, Tkeoret. Pop-
ulation Biol. 30 (1987), 289-308
[93] J. M. Gushing, The Alice effect in age-structured population dynamics, Mathematical
Ecology (S. A. Levin, T. Hallam, eds.), World Scientific Publishing, Teaneck, NJ, 1988,
479-505
[94] J. M. Gushing, A competition model for size-structured species, SIAM J. Appl. Math.
49 (1989), 838-858
[95] J. M. Gushing, A strong ergodic theorem for some nonlinear matrix models for the
dynamics of structured populations, Natural Resource Modeling 3 (1989), 331-375
[96] J. M. Gushing, Some competition models for size-structured populations, Roc. Mount.
J. of Math. 20 (1990), 879-897
[97] J. M. Gushing, Some delay models for juvenile vs. adult competition. Differential Models
in Biology, Epidemiology, and Ecology (S. Busenberg and M. Matelli, eds.), Springer-
Verlag, Berlin, 1991, 177-188
[98] J. M. Gushing, Competing size-structured species, Mathematical Population Dynamics
(O. Arino, D.E. Axelrod, and M. Kimmel, eds.), Marcel Dekker, Inc., NY, 1991, Chapter
3
[99] J. M. Gushing, A simple model of cannibalism, Math. Biosci. 107 (1991), 47-72
[100] J. M. Gushing, A size-structured model for cannibalism, Theoret. Population Biol. 42
(1992), 347-361
[101] J. M. Gushing, A discrete model of competing size-structured species, Theoret. Popu-
lation Biol. 41 (1992), 372-387
[102] J. M. Gushing, Nonlinear Matrix Models for Structured Populations. Mathematical
Population Dynamics: Analysis of Heterogeneity, Volume. Ill: Mathematical Methods
and Modelling of Data (O. Arino, D. Axelrod, M. Kimmel, and M. Langlais, eds.),
Wuerz Publishing Ltd, Winnipeg, MB, Canada, 1994, Chapter 26
[103] J. M. Gushing. Oscillations in age-structured population models with an Allee effect, J.
Comp. Appl. Math. 52 (1994), 71-80
[104] J. M. Gushing, The dynamics of hierarchical age-structured populations. J. Math. Biol.
32 (1994), 705-729
[105] J. M. Gushing, Systems of difference equations and structured population dynamics,
Proc. of the First Int. Conf on Difference Equations (S. N. Elaydi, J. R. Graef, G.
Ladas, and A. C. Peterson, eds.), Gordon and Breach Publishers, Amsterdam, 1995,
123-132
[106] J. M. Gushing, Structured population dynamics, Frontiers in Mathematical Biology (S.
A. Levin, ed.), Lecture Notes in Biomath. 100, Springer-Verlag, Berlin, 1995, 280-295
[107] J. M. Gushing, Periodically forced nonlinear systems of difference equations, J. Differ-
ence Equations and Applications, 3 (1998), 547-561
176 BIBLIOGRAPHY

[108] J. M. Gushing and M. Saleem, A predator-prey model with age structure, J. Math. Biol.
14 (1982), 231-250
[109] J. M. dishing and M. Saleem, A competition model with age structure, Lecture Notes
in Biomath. 54 (1982), 178-192
[110] J. M. Gushing and J. Li, On Ebenman's model for the dynamics of a population with
competing juveniles and adults, Bull. Math. Biol. 51 (1989), 687-713
[111] J. M. Gushing and J. Li, Juvenile versus adult competition, J. Math. Biol. 29 (1991),
457-473
[112] J. M. Gushing and J. Li, Intra-specific competition and density dependent juvenile
growth, Bull Math. Biol 54 (1992), 503-529
[113] J. M. Gushing and J. Li, The dynamics of a size-structured intraspecific competition
model with density dependent juvenile growth rates, Individual Based Modeling: Con-
cepts and Models (D. DeAngelis and L. J. Gross, eds.), Routledge, Chapman, and Hall,
New York, 1992, 112-125
[114] J. M. Gushing and J. Li, Oscillations caused by cannibalism in a size-structured popu-
lation model, Canad. Appl. Math. Quart. 3 (1995), 155-172
[115] J. M. Gushing and Z. Yicang, The net reproductive value and stability in matrix pop-
ulation models, Nat. Res. Mod. 8 (1994), 297-333
[116] J. M. Gushing, B. Dennis, R. A. Desharnais, and F. R. Costantino, An interdisciplinary
approach to understanding nonlinear ecological dynamics, Ecol. Modeling 91 (1996),
111-119
[117] J. M. Gushing, B. Dennis, R. A. Desharnais, and R. F. Costantino, Moving toward an
unstable equilibrium: Saddle nodes in population systems, J. Anim. Ecol. 67 (1998),
298-306
[118] D. L. DeAngelis, D. K. Cox, and C. C. Coutant, Cannibalism and size dispersal in
young-of-the-year large mouth bass: Experiment and model, Ecol. Modelling 8 (1979),
133-148
[119] D. L. DeAngelis and L. J. Gross (eds.). Individual-Based Models and Approaches in
Ecology, Chapman and Hall, New York, 1992
[120] D. L. DeAngelis and J. S. Mattice, Implications of a partial differential equation cohort
model, Math. Biosci. 47 (1979), 271-285
[121] D. L. DeAngelis, K. A. Rose, and M. A. Huston, Individual-oriented approaches to
modeling ecological populations and communities, Frontiers in Mathematical Biology
(S. A. Levin, ed.), Lecture Notes in Biomath. 100, Springer-Verlag, Berlin, 1995, 390-
410
[122] D. L. DeAngelis, L. J. Svoboda, S. W. Christensen, and D. S. Vaughan, Stability and
return times of Leslie matrices with density-dependent survival: Applications to fish
populations, Ecol. Modelling 8 (1980), 149-163
[123] G. di Blasio, Nonlinear age-dependent population growth with history-dependent birth
rate, Math. Biosci. 46 (1979), 279-291
[124] G. di Blasio, M. lannelli, and E. Sinestrari, Approach to equilibrium in age structured
populations with an increasing recruitment process, J. Math. Biol. 13 (1982), 371-382
[125] K. Deiinling, Equilibria of an age-dependent population model, Nonlinear Differential
Equations: Invariance, Stability and Bifurcation (P. de Mottoni and L. Salvador!, eds.),
Academic Press, New York, 1981
[126] L. Demetrius, Primitivity conditions for growth matrices, Math. Biosci. 12 (1971), 53-58
[127] L. Demetrius. Natural selection and age-structured populations, Genetics 79 (1974),
535-544
[128] L. Demetrius, Demographic parameters and natural selection, Proc. Nat. Acad. Sci. 71
(1974), 4645-4647
 BIBLIOGRAPHY 177

[129] W. R. DeMott and W. C. Kerfoot, Competition among cladocerans: Nature of the

interaction between Bosmma and Daphma, Ecology 63 (1982), 1949-1966
[130] B. Dennis, Allee effects: Population growth critical density, and the chance of extinction,
Nat. Res. Modeling 3 (1989), 481-538
[131] B. Dennis, P. O. Munholland, and J. M. Scott, Estimation of growth and extinction
parameters for endangered species. Ecol. Monogr. 61 (1991), 115—143
[132] B. Dennis, R. A. Desharnais, J. M. Gushing, and R. F. Costantino, Nonlinear demo-
graphic dynamics: Mathematical models, statistical methods, and biological experi-
ments, Ecol Monogr. 65 (1995), 261-281
[133] B. Dennis, R.. Desharnais, J. M. Gushing, and R. F. Costantino. Transitions in pop-
ulation dynamics: Equilibria to periodic cycles to aperiodic cycles, J. Anim. Ecol. 66
(1997), 704-729
[134] A. M. de Roos. Numerical methods for structured population models: The escalator
boxcar train, Numer. Methods Partial Differential Equations 4 (1988), 173-195
[135] A. M. de R.OOS, J. A. J. Metz, E. Evers, and A. Peipoldt, A size dependent predator-prey
interaction: Who pursues whom?, J. Math. Biol. 28 (1990), 609-643
[136] A. M. de Roos, O. Diekmann, and J. A. J. Metz, Studying the dynamics of structured
population models: A versatile technique and its application to Daphnia. Amer. Nat.
139 (1992), 123-147
[137] R. A. Desharnais arid J. E. Cohen, Life not lived due to disequilibrium in heterogeneous
age-structured populations, Theoret. Population Bwl. 29 (1986), 385-406
[138] R.. A. Desharnais and L. Liu, Stable demographic limit cycles in laboratory populations
of Tribohum castaneum, J. Anim. Ecol. 56 (1987). 885-906
[139] G. di Blasio, M. lannelli, arid E. Sinestrari, Approach to equilibrium in age structured
populations with an increasing recruitment process, J. Math. Biol. 13 (1982), 371-382
[140] G. Di Cola and F. Nicoli, Parameter estimation in age-structured population dynamics,
Riv. Mat. Univ. Parma (4) 9 (1983), 213-222
[141] O. Diekmann, The stable size distribution: An example in structured population dy-
namics, Lecture Notes in Biomath. 54 (1982), 90 96
[142] O. Diekmann, The dynamics of structured populations: Some examples, Mathematics in
Biology and Medicine (V. Capasso, E. Grosso, and S. L. Paveri-Fontana, eds.), Lecture
Notes in Biomath. 57, Springer-Verlag, Berlin, 1985, 7 18
[143] O. Diekmann, M. Gyllenberg, J. A. J. Metz, and H. Thieme, The "cumulative" formu-
lation of (physiologically) structured population models, Evolution Equations, Control
Theory and Biomathematics (Ph. Clement and G. Lumer, eds.), Lecture Notes in Pure
and Appl. Math. 155, Marcel Dekker, New York, 1994, 145-154
[144] O. Diekmann, H. A. Lauwerier, T. Aldenberg, and J. A. J. Metz, Growth, fission and
the stable size distribution, J. Math. Biol 18 (1983), 135-148
[145] O. Diekmann and H. Metz, Exploring linear chain trickery for physiologically structured
populations, CWI Quarterly 1 (1989), 3-14
[146] O. Diekmann and J. A. J. Metz, On the reciprocal relationship between life histories
and population dynamics, Frontiers in Mathematical Biology (S. A. Levin, ed.), Lectvire
Notes in Biomath. 100, Springer-Verlag, Berlin, 1995, 263-279
[147] O. Diekmann, R. M. Nisbet, W. S. C. Gurney, and F. van den Bosch, Simple mathemat-
ical models for cannibalism: A critique and a new approach, Math. Biosci. 78 (1986),
21-46
[148] S. I. Dodson, Zooplankton competition and predation: An experimental test of the
size-efficiency hypothesis, Ecology 55 (1974), 605-613
[149] J. Douglas Jr. and F. A. Milner, Numerical methods for a model of population dynamics,
Calcolo 24 (1987). 247-254
178 BIBLIOGRAPHY

[150] W. Ebeling, A. Bngel, and V. G. Mazenko, Modeling of selection processes with age-
dependent birth and death rates, BioSystems 19 (1986), 213-221
[151] B. Ebenman, Niche difference between age classes and intraspecific competition in age-
structured populations, ,/. Theor. Biol, 124 (1987), 25-33
[152] B. Ebenman, Competition between age classes and population dynamics, J. Theor.
Biol. 131 (1988), 389-400
[153] B. Ebenman, Dynamics of age- and size-structured populations: Intraspecific compe-
tition, Size-Structured Populations: Ecology and Evolution (B. Ebenman, L. Persson,
eds.), Springer-Verlag, Berlin, 1988, 127-139
[154] B. Ebenman and L. Persson, Size-Structured Populations: Ecology and Evolution,
Springer-Verlag, Berlin, 1989
[155] L. Edelstein-Keshet, Mathematical theory for plant-herbivore systems, J. Math. Biol.
24 (1986), 25-58
[156] S. N. Elaydi, An Introduction to Difference Equations, Springer-Verlag, New York, 1996
[157] R. H. Elderkin, Structured populations with nonlinear, nonlocally dependent dynamics:
Basic theory, J. Math. Anal. Appl. 192 (1995), 392-412
[158] M. A. Elgar and B. J. Crespi, Cannibalism: Ecology and Evolution Among Diverse
Taxa, Oxford University Press, Oxford, 1992
[159] Y. S. Fan and L. S. Chen, Periodic solutions and period-similarity of a growth model
with age-structured population, Systems Sci. Math. Sci. 4 (1991), 148—157
[160] M. E. Fisher and B. S. Goh, Stability results for delayed-recruitment models in popu-
lation dynamics, J. Math. Biol. 19 (1984), 147-156
[161] B. G. Fitzpatrick, Statistical tests of fit in estimation problems for structured population
modeling, Quart. Appl. Math. 53 (1995), 105-128
[162] L. R. Fox, Cannibalism in natural populations, Ann. Rev. Ecol. Syst. 6 (1975), 87-106
[163] J. E. Franke and A.-A. Yakubu, Global attractors in competitive systems, Nonlinear
Anal. 16 (1991), 111-129
[164] J. E. Franke and A.-A. Yakubu, Mutual exclusion versus coexistence for discrete com-
petitive systems, J. Math. Biol. 30 (1991), 161-168
[165] J. E. Franke and A.-A. Yakubu, Pioneer exclusion in a one-hump discrete pioneer-climax
competitive system, J. Math. Biol. 32 (1994), 771-787
[166] J. E. Franke and A.-A. Yakubu, Extinction in systems of bobwhite quail populations,
Canad. Appl. Math. Quart. 3 (1995), 173-201
[167] J. E. Franke and A.-A. Yakubu, Dominance conditions in age-structured discrete com-
petitive systems, Proc. First Int. Conf. on Diff. Equations and Appl. (S. N. Elaydi, J.
R. Graef, G. Ladas, and A. C. Peterson, eds.), Gordon and Breach Publishers, New
York, 1995, 197-211
[168] J. E. Franke and A.-A. Yakubu, Dominance conditions in age-structured discrete non-
cooperative systems, J. Math. Biol. 34 (1996), 442-454
[169] J. E. Franke and A.-A. Yakubu, Extinction and persistence of species in discrete com-
petitive systems with a safe refuge, J. Math. Anal. Appl. 203 (1996), 746-761
[170] J. E. Franke and A.-A. Yakubu, Extinction of species in age-structured discrete nonco-
operative system, J. Math. Biol. 34 (1996), 442-454
[171] J. E. Franke and A.-A. Yakubu, Global asymptotic behavior and dispersion in age-
structured, discrete competitive systems, to appear in Canad. Appl. Math. Quart. 5
[172] J. E. Franke and A.-A. Yakubu, Principles of competitive exclusion for discrete pop-
ulations with reproducing juveniles and adults, Proc. World Congress of Nonlinear
Analysts, 30 (1997), 1197-1206
[173] J. E. Franke and A.-A. Yakubu, Exclusionary population dynamics in size-structured,
discrete competitive systems, J. Differ. Equations Appl., in press
 BIBLIOGRAPHY 179

[174] J. C. Frauenthal, Some simple models of cannibalism, Math. Biosci. 63 (1983), 87-98
[175] H. I. Freedman and J. W.-H. So, Persistence in discrete semidynamical systems, SI AM
J. Math. Anal. 20 (1989), 930-938
[176] G. Frobenius, Uber matrizen aus nicht negativen elernenten, S.-B. Deutsch. Akad. Wiss.
Berlin. Math-Nat. Kl (1912), 456-477
[177] T. Fujimoto and U. Krause. Asymptotic properties for inhomogeneous iterations of
nonlinear operators, SIAM J. Math. Anal. 19 (1988), 841-853
[178] T. B. K. Gage, F. M. Williams, and J. B. Horton, Division synchrony and the dynamics
of microbial populations: A size-specific model. Theoret. Population Biol. 26 (1986),
296-314
[179] F. R. Gantmacher, The Theory of Matrices, Vol. 2, Chelsea, New York, 1960
[180] J. Gerritsen, Size efficiency reconsidered: A general foraging model for free-swimming
aquatic animals, Amer. Nat. 123 (1984), 450-467
[181] W. M. Getz, Optimal harvesting of structured populations, Math. Biosci. 44 (1979),
269-291
[182] W. M. Getz, Harvesting discrete nonlinear age and stage structured populations, J.
Optim. Theory Appl. 57 (1988), 69-83
[183] W. M. Getz, A metaphysiological approach to modeling ecological populations and
communities, Frontiers in Mathematical Biology (S. A. Levin, ed.), Lecture Notes in
Biomath. 100, Springer-Verlag, Berlin, 1995, 411-442
[184] W. M. Getz and R. G. Haight, Population Harvesting: Demographic Models of Fish,
Forest, and Animal Resources, Monographs in Population Biology 27, Princeton Uni-
versity Press, Princeton, N.I, 1989
[185] C. Godfray and M. Hassell, Chaotic beetles, Science (January 17, 1997), 323
[186] M. Golubitsky. E. B. Keeler. and M. Rothschild, Convergence of the age-structure:
Applications of the projective metric, Theoret. Population Biol. 7 (1975), 84-93
[187] L. A. Goodman, On the sensitivity of the intrinsic growth rate to changes in the age-
specific birth and death rates, Theoret. Population Biol. 2 (1971). 339-354
[188] K. Gopalsamy, Age specific coexistence in two species competition. Math. Biosci. 61
(1982), 101-122
[189] G. Greiner, A typical Perron-Frobenius theorem with applications to an age-dependent
population equation, Infinite-dimensional Systems (F. Kappel and W. Schappacher,
eds.), Lecture Notes in Math. 1076, Springer-Verlag, Berlin, 1984, 86-100
[190] G. Greiner, J. A. P. Heesterbeek, and J. A. J. VIetz, A singular perturbation theorem for
evolution equations and time-scale arguments for structured population models, Canad.
Appl. Math. Quart. 1 (1994), 435-459
[191] D. H. Griffel, Age dependent population growth, IMA J. Appl. Math. 17 (1976), 141-152
[192] J. Guckenheimer, G. Oster, and A. Ipaktchi, The dynamics of density dependent pop-
ulation models, J. Math. Biol. 4 (1977), 101-147
[193] J. Guckenheimer and P. Homes, Nonlinear Oscillations, Dynamical Systems, and Bi-
furcations of Vector Fields, Springer-Verlag, Berlin, 1983
[194] W. S. C. Gurney and R. M. Nisbet, Age and density-dependent population dynamics
in static and variable environments, Theoret. Population Biol. 17 (1980), 321-344
[195] W. S. C. Gurney and R. M. Nisbet, The systematic formulation of delay-differential
models of age or size structured populations, Population Biology (Freedrnan and
Strobeck, eds.), Lecture Notes in Biomath. 52, Springer-Verlag, Berlin, 1983
[196] W. S. C. Gurney and R. M. Nisbet, Fluctuation periodicity, generation separation and
the expression of larval competition, Theoret. Population Biol. 28 (1985), 150-180
[197] W. S. C. Gurney, R. M. Nisbet, and J. H. Lawton, The systematic formulation of
tractable single species population models incorporating age structure, J. Anirn. Ecol.
52 (1983), 479-495
180 BIBLIOGRAPHY

[198] M. E. Gurtin and R. C. MacCamy, Nonlinear age-dependent population dynamics, Arch.

Rational Mech. Anal. 54 (1974), 281-300
[199] M. E. Gurtin and R. C. MacCamy, Some simple models for nonlinear age-dependent
population dynamics, Math. Biosci. 43 (1979), 199-211
[200] M. E. Gurtin and R. C. MacCamy, Population dynamics with age dependence, Nonlinear
Analysis and Mechanics, Herrot-Watt Symposium III, Pitman, Boston, 1979
[201] M. E. Gurtin and D. S. Levine, On predator-prey interactions with predation dependent
on age of prey, Math. Biosci. 47 (1979), 207-219
[202] M. E. Gurtin and L. F. Murphy, On the optimal harvesting of persistent age-structured
populations, J. Math. Biol 13 (1981), 131-148
[203] M. E. Gurtin and L. F. Murphy, On the optimal harvesting of age-structured popula-
tions: Some simple models, Math. Biosci. 55 (1981), 115-136
[204] M. Gyllenberg, Nonlinear age-dependent population dynamics in continuously propa-
gated bacterial cultures, Math. Biosci. 62 (1982), 45-74
[205] M. Gyllenberg, Stability of a nonlinear age-dependent population model containing a
control variable, SIAM J. Appl. Math. 43 (1983), 1418-1438
[206] M. Gyllenberg, G. Soderbacka, and S. Ericsson, Does migration stabilize local popula-
tion dynamics? Analysis of a discrete metapopulation model, Math. Biosci. 118 (1993),
25-49
[207] M. Gyllenberg and G. F. Webb, A nonlinear structured population model of tumor
growth with quiescence, J. Math. Biol. 28 (1990), 671-694
[208] K. P. Hadeler, Pair formation in age-structured populations, Acta Appl. Math. 14 (1989),
91-102
[209] J. W. Haefner, Modeling Biological Systems: Principles and Applications, Chapman
and Hall, New York, 1996
[210] A. Haimovici, On the growth of a population dependent on ages and involving resources
and pollution, Math. Biosci. 43 (1979), 213-237
[211] J. K. Hale, Asymptotic Behavior of Dissipative Systems. Mathematical Surveys and
Monographs 25, AMS, Providence, RI, 1985
[212] D. J. Hall, S. T. Threlkeld, C. W. Burns, and P. H. Crowley, The size-efficiency hypoth-
esis and the size structure of zooplankton communities, Ann. Rev. Ecol. Syst. 7 (1976),
177-208
[213] T. G. Hallam and S. M. Henson, Extinction in a structured predator-prey model with
size-dependent predation, Differential Equations and Applications to Biology and to
Industry (M. Martelli, K. Cooke, E. Cumberbatch, B. Tang, arid H. Thieme, eds.),
World Scientific, Singapore, 1996, 157-172
[214] T. G. Hallam, R. R. Lassiter, J. Li, and W. McKinney, Physiologically structured popu-
lation models in risk assessment, Biomathematics and Related Computational Problems,
Kluwer Academic Publishers, Dordrecht, 1988, 197-211
[215] K. B. Hannsgen and J. J. Tyson, Stability of the steady-state size distribution in a
model of cell growth and division, J. Math. Biol. 22 (1985), 293-301
[216] A. Hastings, Simple models for age dependent predation, Lecture Notes m Biomath. 54,
Springer-Verlag, Berlin, 1984, 114-119
[217] A. Hastings, Age-dependent predation is not a simple process I. Continuous time models,
Theoret. Population Biol. 23 (1983), 347-362
[218] A. Hastings, Age-dependent predation is not a simple process II. Wolves, ungulates,
and a discrete time model for predation on juveniles with a stabilizing tail, Theoret.
Population Biol. 26 (1984), 271-282
[219] A. Hastings, Simple models for age-dependent predation, Mathematical Ecology (S. A.
Levin and T. G. Hallam, eds.), Lecture Notes in Biomath. 54, Springer-Verlag, Berlin,
1984, 114-119
 BIBLIOGRAPHY 181

[220] A. Hastings, Cycles in cannibalistic egg-larval interactions, J. Math. Biol. 24 (1987),

651-666
[221] A. Hastings, Age-dependent dispersal is not a simple process: Density dependence,
stability and chaos, Theoret. Population Biol. 41 (1992), 388-400
[222] A. Hastings and R. F. Costantino, Cannibalistic egg-larva interactions in Tribolium: An
explanation for the oscillations in population numbers, Ame-r. Nat. 130 (1987), 36—52
[223] A. Hastings and R. F. Costantino, Oscillations in population numbers: Age-dependent
cannibalism, J. Anim. Ec.ol. 60 (1991), 471-482
[224] A. Hastings and D. Wollkind, Age structure in predator-prey systems. A general model
and a specific example, Theoret. Population Biol. 21 (1982), 44-56
[225] H. J. A. M. Heijmans, On the stable size distribution of populations reproducing by
fission into two unequal parts. Math. Biosci. 72 (1984), 19 50
[226] H. J. A. M. Heijmans, Structured populations, linear semigroups and positivity, Math.
Z. 191 (1986), 599-617
[227] S. M. Henson, Existence and stability of nontrivial periodic solutions of periodically
forced discrete dynamical systems, J. Differ. Kquation.i Appl. 2 (1996), 315-331
[228] S. M. Henson, Cannibalism can be beneficial even when its mean yield is less than one,
Theoret. Population Biol. 51 (1977), 109-117
[229] S. M. Henson, A Continuous, Age-Structured Insect Population Model, preprint
[230] S. M. Henson and J. M. Gushing, Hierarchical models of intra-specific competition:
Scramble versus contest, J. Math. Dial. 34 (1996). 755-772
[231] S. M. Henson and J. M. Gushing, Optimal parasitization in a size-structured host-
parasitoid model, Nonlinear World 3 (1996), 1-18
[232] S. M. Henson and J. M. Gushing. The effect of periodic habitat fluctuations on a non-
linear insect population model, ,/. Math. Biol. 36 (1997), 201-226
[233] S. M. Henson and T. G. Hallam, Survival of the fittest: Asymptotic competitive ex-
clusion in structured population and community models. Nonlinear World 1 (1994),
385-402
[234] J. Hofbauer, V. Hutson, and W. Jansen, Coexistence for systems governed by difference
equations of Lotka-Volterra type, J. Math. Biol. 25 (1987), 553-570
[235] J. Hofbauer and J. W.-H. So, Uniform persistence and repellors for maps, Proc. AMS
107 (1989), 1137-1142
[236] F. Hopperisteadt. Mathematical Theories of Populations: Demographics, Genetics and
Epidemic.'!. Regional Conference Series in Applied Mathematics 20, SIAM. Philadelphia.
PA, 1975
[237] F. Hoppensteadt, Mathematical Methods of Population Biology, Cambridge University-
Press, Cambridge, 1982
[238] J. W. Horweed and J. G. Shepherd, The sensitivity of age-structured populations to
environmental variability, Math. BioKci. 57 (1981), 59-82
[239] F. Houllier and J.-D. Lebreton, A renewal-equation approach to the dynamics of stage-
grouped populations, Math. Biosci. 79 (1986), 185-197
[240] Y. H. Hsieh, On the evolution of altruism in an age-structured population, Math. Com,-
put. Modelling 11 (1988), 472-475
[241] X. C. Huang, An age-dependent population model and its operator, Phys. D. 41 (1990),
356-370
[242] T. P. Hughes and J. H. Connell, Population dynamics based on size or age? A reef-coral
analysis, Amer. Nat. 129 (1987), 818 829
[243] M. laimelli, Mathematical problems in the description of age structured populations.
Mathematics in Biology and Medicine (V. Capasso, E. Grosso and S. L. Paveri-Fontana.
eds.), Lecture Notes in Biomath. 57, Springer-Verlag. Berlin, 19 32
182 BIBLIOGRAPHY

[244] M. lannelli, Mathematical Theory of Age-Structured Population Dynamics, Applied

Mathematics Monographs 7, Giardini Editori e Stampatori, Pisa, 1994
[245] J. Irapagliazzo, Deterministic Aspects of Mathematical Demography, Biomathematics
13, Springer-Verlag, Berlin, 1980
[246] H. Inaba, A semigroup approach to the strong ergodic theorem of the multi-state stable
population process, Math. Pop. Studies 1 (1988), 49-77
[247] H. Inaba, Weak ergodicity of population evolution processes, Math. Biosci. 96 (1989),
195-219
[248] U. Inprasit, Equilibria in Size-Structured Population Models, Ph.D. dissertation, Uni-
versity of Arizona, Tucson, AZ, 1996
[249] K. Ito, F. Kappel, and G. Peichl, A fully discretized approximation scheme for size-
structured population models, SIAM J. Numer. Anal. 28 (1991), 923-954
[250] V. A. A. Jansen, R. M. Nisbet, and W. S. C. Gurney, Generation cycles in stage struc-
tured populations, Bull. Math. Biol. 52 (1990), 375-396
[251] A. L. Jensen, Leslie matrix models for fisheries studies, Biometrics 30 (1974), 547-551
[252] D. Jillson, Insect populations respond to fluctuating environments, Nature 288 (1980),
699-700
[253] J. N. Kapur, Age-structured population models with density dependence, Bull. Can.
Math. Soc. 74 (1982), 207-215
[254] J. N. Kapur, Age-structured population models with density dependence, Bull. Calcutta
Math. Soc. 74 (1982), 207-215
[255] P. Kareiva, Predicting and producing chaos, Nature 375 (May, 1995), 189-190
[256] W. O. Kermack and A. G. McKendrick, A contribution to the mathematical theory of
epidemics, Proc. Roy. Soc. A 115 (1927), 700-721
[257] W. O. Kermack and A. G. McKendrick, Contributions to the mathematical theory of
epidemics II. The problem of endemicity, Proc. Roy. Soc. A 138 (1932), 55-83
[258] W. O. Kermack and A. G. McKendrick, Contributions to the mathematical theory of
epidemics III. Further studies on the problem of endemicity, Proc. Roy. Soc. A 141
(1933), 92-122
[259] N. Keyfitz, Introduction to the Mathematics of Populations, Addison-Wesley, Reading,
MA, 1968
[260] N. Keyfitz and J. A. Beekman, Demography through Problems, Springer-Verlag, New
York, 1984
[261] N. Keyfitz, Applied Mathematical Demography, Springer-Verlag, Berlin, 1985
[262] J. Knight, Boom time for beetles, New Scientist 156, No. 2110 (November, 1997), 19
[263] S. A. L. M. Kooijman, Dynamic Energy Budgets in Biological Systems, Cambridge
University Press, Cambridge, 1993
[264] T. Kostova, Oscillations in age dependent population dynamics, Differential Equations
and Applications to Biology and to Industry (M. Martelli, K. Cooke, E. Cumberbatch,
B. Tang, and H. Thieme, eds.), World Scientific, Singapore, 1996, 253-260
[265] T. V. Kostova and N. H. Chipev, A model of the dynamics of intramolluscan trematode
populations: Some problems concerning oscillatory behavior, Comp. Math, and Appl.
21 (1991), 1-15
[266] T. Kostova and F. Milner, Nonlinear age-dependent population dynamics with constant
size, SIAM J. Math. Anal. 22 (1991), 129-137
[267] M. Kot, Discrete-time travelling waves: Ecological examples. J. Math. Biol. 30 (1992),
413-436
[268] M. Kot and W. M. Schaffer, Discrete-time growth-dispersal models, Math. Biosci. 80
(1986), 109-136
[269] S. V. Krishna and S. Bhattacharyya, Study of structured populations with quiescent
state, Differential Equations Dynam. Systems 1 (1993), 325-340
 BIBLIOGRAPHY 183

[270] Y. Kuan, Delay Differential Equations: With Applications in Population Dynamics,

Academic Press, Boston, 1993
[271] Y. Kuang and J. M. Gushing. Global stability in a nonlinear difference-delay equation
model of flour beetle population growth. J. Differ. Equations Appl. 2 (1995), 31-37
[272] R. H. Lamberson, K. McKelvey, B. R. Noon, and C. Voss, A dynamic analysis of
Northern Spotted Owl viability in a fragmented forest landscape, Conservation Biology
6 (1992), 505-512
[273] L. Lamberti and P. Vernole, Existence and asymptotic behaviour of solutions of an age
structured population model, Boll. Un. Mat. Ital C (5) 18 (1981), 119-139
[274] H. D. Landahl and B. D. Hansen, A three stage population model with cannibalism,
Bull. Math. Biol. 37 (1975). 11-17
[275] R. Lande, Demographic models of the northern spotted owl (Strix occidentalis caurina),
Oecologia 75 (1988), 601-607
[276] R. Lande and S. H. Orzack, Extinction dynamics of age-structured populations in a
fluctuating environment, Proc. Nat. Acad. Sci. 85 (1988), 7418-7421
[277] H. L. Lanhaar, General population theory in the age-time continuum. J. Franklin Inst.
293 (1972), 199-214
[278] J. P. LaSalle, The Stability of Dynamical Systems, Regional Conference Series in Applied
Mathematics 25, SIAM, Philadelphia. PA, 1976
[279] R. Law, A model for the dynamics of a plant population containing individuals classified
by age and size, Ecology 64 (1983), 224-230
[280] L. P. Lefkovitch, The study of population growth in organisms grouped by stages,
Biometrics 25 (1965), 309-315
[281] P. H. Leslie, On the use of matrices in certain population mathematics, Biometrika 33
(1945), 183-212
[282] P. H. Leslie, Some further notes on the use of matrices in population mathematics.
Biometrika 35 (1948), 213-245
[283] S. A. Levin and C. P. Goodyear, Analysis of an age-structured fishery model, J. Math.
Biol. 9 (1980), 245-274
[284] S. A. Levin, Age-structure and stability in multiple-age spawning populations. Renew-
able Resource Management (T. L. Vincent and J. M. Skowronski, eds.), Springer-Verlag,
Berlin, 1981, 21-45
[285] D. S. Levine, On the stability of a predator-prey system with egg-eating predators,
Math. Biosci. 56 (1981), 27-46
[286] D. S. Levine, Bifurcating periodic solutions for a class of age-structured predator-prey
systems, Bull Math. Biol. 45 (1983), 901-915
[287] D. S. Levine. Models of age-dependent predation and cannibalism via the McKendrick
equation, Comp. Math. Appl. 9 (1983), 403-414
[288] E. G. Lewis, On the generation and growth of a population, Sankhya 6 (1942), 93-96
[289] S. J. Lewis, A note on the strong ergodic theorem of some discrete models, J. Differ.
Equations Appl. 3 (1997). 55-65
[290] J. Li, Persistence and extinction in continuous age-structured population models, Comp.
Math. Appl 15 (1988), 511-523
[291] J. Li, Persistence in discrete age-structured population models, Bull. Math. Biol 50
(1988), 351-366
[292] J. Li and T. G. Hallam, Survival in continuous structured population models, J. Math.
Biol. 26 (1988), 421-433
[293] X. R. Li and Z. E. Ma, An improved model of age-structured population dynamics,
Math. Comput. Modelling 20 (1994), 143-150
[294] L. Liu and J. E. Cohen, Equilibrium and local stability in a logistic matrix model for
age-structured populations, J. Math. Biol. 25 (1987), 73—88
184 BIBLIOGRAPHY

[295] M. Lloyd, Self regulation of adult numbers by cannibalism in two laboratory strains of
flour beetles (Tribolium castaneum), Ecology 49 (1967), 245-259
[296] D. O. Logofet, Matrices and Graphs, Stability Problems in Mathematical Ecology, CRC
Press, London, 1992, 552-564
[297] A. Lomnicki, Population Ecology of Individuals, Monographs in Population Biology 25,
Princeton University Press, Princeton, NJ, 1988
[298] A. Lomnicki and S. Sedziwy, Do individual differences in resource intakes without mo-
nopolization casue population stability and persistence?, J. Theor. Biol. 136 (1989),
317-326
[299] M. Loreau, Competition between age classes, and the stability of stage-structured pop-
ulations: A re-examination of Ebenman's model, J. Theor. Biol. 144 (1990), 567-571
[300] M. Loreau and W. Ebenhoh, Competitive exclusion and coexistence of species with
complex life cycles, Theoret. Population Biol. 46 (1994), 58-77
[301] A. J. Lotka, The stability of the normal age distribution, Proc. Nat. Acad. Sci. 8 (1922),
339-345
[302] M. Lynch, Complex interactions between natural coexpoliters-Dap/mia and Ceriodaph-
nia, Ecology 59 (1978), 552-564
[303] N. MacDonald, Time Lags in Biological Models, Lecture Notes in Biomathematics 27.
Springer-Verlag, Berlin, 1978
[304] P. Marcati, Asymptotic behavior in age-dependent population dynamics with hereditary
renewal law, SIAM J. Math. Anal. 12 (1981), 904-916
[305] P. Marcati, On the global stability of the logistic age-dependent population growth, J.
Math. Biol 15 (1982), 215-226
[306] J. E. Marsden and M. McCracken, The Hopf Bifurcation and Its Applications, AMS 19,
Springer-Verlag, Berlin, 1976
[307] C. F. Martin, L. J. S. Allen, and M. Stamp, An analysis of the transmission of chlamydia
in a closed population, J. Differ. Equations Appl. 2 (1996), 1-20
[308] S. Matucci, Existence, uniqueness and asymptotic behaviour for a multi-stage evolution
problem of an age-structured population. Math. Models Methods Appl. Sci. 5 (1995),
1013-1041
[309] R. M. May, Stability and Complexity in Model Ecosystems, Princeton University Press,
Princeton, NJ, 1973
[310] R. M. May, Biological populations with nonoverlapping generations: Stable points,
stable cycles, and chaos, Science 186 (1974), 645-647
[311] R. M. May, Simple mathematical models with very complicated dynamics, Nature 261
(1976), 459-467
[312] R. M. May, G. R. Conway, M. P. Hassell, and T. R. E. Southwood, Time delays, density-
dependence and single-species oscillations, J. Amm. Ecol. 43 (1974), 747-770
[313] R. M. May and G. F. Oster. Bifurcations and dynamic complexity in simple ecological
models, Amer. Nat. 110 (1976), 573-599
[314] J. Maynard Smith, Models in Ecology, Cambridge University Press, Cambridge, 1975
[315] J. Maynard Smith, Mathematical Ideas in Biology, Cambridge University Press. Cam-
bridge, 1977
[316] N. G. Medhin and M. Sambandham, Analytical and numerical investigation of age
structured population, Dynam. Systems Appl. 2 (1993), 435-450
[317] A. G. McKendrick and M. K. Pai, The rate of multiplication of micro-organisms: A
mathematical study, Proc. Roy. Soc. Edinburgh 31 (1910), 649-655
[318] A. G. McKendrick, Applications of mathematics to medical problems, Proc. Edin. Math.
Soc. 44 (1926), 98-130
[319] J. N. McNair, Stability effects of a juvenile period in age-structured populations, J.
Thoer. Biol. 137 (1989), 397-422
 BIBLIOGRAPHY 185

[320] J. N. McNair and C. E. Gouldcn, The dynamics of age-structured populations with a

gestation period: Density-independent growth and egg ratio methods for estimating the
birth rate, Theoret. Population Biol 39 (1991), 1-29
[321] J. A. J. Metz, A. M. de Roos, and F. van den Bosch, Population models incorporating
physiological structures: A quick survey of the basic concepts and an application to size-
structured population dynamics in waterfleas, Size-Structured Populations: Ecology and
Evolution (B. Ebenman and L. Persson, eds.), Springer-Verlag, Berlin, 1989
[322] J. A. J. Metz and A. M. de Roos, The role of physiologically structured population
models within a general individual-based modeling perspective, Individual-Based Models
and Approaches in Ecology (D. L. DeAngelis and L. J. Gross, eds.), Chapman and Hall,
New York, 1992, 88-111
[323] J. A. J. Metz and O. Diekmann, The Dynamics of Physiologically Structured Popula-
tions. Lecture Notes in Biomath. 68, Springer-Verlag, Berlin, 1986
[324] J. A. J. Metz and O. Diekmann, Exact Finite-Dimensional Representations of Models
for Physiologically Structured Populations I. The Abstract Foundations of Linear Chain
Trickery, Lecture Notes in Pure and Appl. Math. 133, Dekker, New York, 1991
[325] R. K. Miller, Asymptotic stability and perturbations for linear Volterra integrodifferen-
tial systems, Delay and Functional Differential Equations and Their Applications (K.
Schmit, ed.), Academic Press, New York, 1972
[326] F. A. Milner. Age structured populations with history dependent mortality and natality.
Calcolo 30 (1993), 29-39
[327] W. W. Murdoch, R. M. Nisbet, S. P. Blythe, W. S. C. Gurney, and J. D. Reeve, An
invulnerable age class and stability in delay-differential parasitoid-host models, Amer.
Nat. 129 (1987). 264-282
[328] L. F. Murphy, A nonlinear growth mechanism in size structured population dynamics,
J. Theor. Biol. 104 (1983), 493-506
[329] L. F. Murphy, Density dependent cellular growth in an age-structured colony, Comput.
Math. Appl. 9 (1983), 383-392
[330] L. F. Murphy and S. J. Smith, Optimal harvesting of an age-structured population, J.
Math. Biol 29 (1990), 77-90
[3311 R. M. Nisbet and W. S. C. Gurney, Modelling Fluctuating Populations, Wiley and Sons,
New York, 1982
[332 R. M. Nisbet and W. S. C. Gurney, The systematic formulation of population models for
insects with dynamically varying instar duration, Theoret. Population Biol. 23 (1983),
114-135
[333] R. M. Nisbet, W. S. C. Gurriey, S. P. Blythe, and J. A. ,7. Metz, Stage structure models
of populations with distinct growth and development processes, IMA J. Math. Appl.
Biol. Med. 2 (1985), 57-68
[334] R. M. Nisbet, W. S. C. Gurney, E. McCauley, and W. W. Murdoch, Structured pop-
ulation models: A tool for linking effects at individual and population levels, Biol. J.
Lmnean Soc. 37 (1989), 79 99
[335] R. M. Nisbet and L. C. Onyiah, Population dynamic consequences of competition within
and between age classes, J. Math. Biol. 32 (1994), 329-344
[336] Z. Nitecki, Differentiate Dynamics, MIT Press, Cambridge, MA, 1971
[337] P. M. North, A computer modelling study of the population dynamics of the screech
owl (Otus asio), Ecol. Model. 30 (1985), 105-114
[338] D. O'Regan and S. J. Smith. The maximum principle for continuous age-structured
populations with unconstrained harvest rates, Dynam. Systems Appl. 3 (1994), 189
206
[339] G. Oster, Modern modeling of continuum phenomena, Lectures in Population Dynamics
(R. C. diPrima, ed.), AMS, Providence, RI, 1977, 149-190
186 BIBLIOGRAPHY

[340] G. Oster and Y. Takahashi, Models for age-specific interactions in a periodic environ-
ment, Ecol. Monogr. 44 (1974), 483-501
[341] S. W. Pacala and J. Weiner, Effects of competitive asymmetry on a local density model
of plant interference, J. Theor. Biol 149 (1991), 165-179
[342] C. J. Pennycuick, R. M. Compton, and L. Beckingham, A computer model for simulating
the growth of a population, or of two interacting populations, J. Theor. Biol. 18 (1968),
316-329
[343] O. Perron, Uber matrizen, Math. Ann. 64 (1907), 248-263
[344] M. Pilant and W. Rundell, Age structured population dynamics. Inverse methods in
action, Inverse Probl. Theor. Imaging, Springer-Verlag, Berlin, 1990, 122-129
[345] R. E. Plant and L. T. Wilson, Models for age structured populations with distributed
maturation rates, J. Math. Biol. 23 (1986), 247-262
[346] G. A. Polis, The evolution and dynamics of intraspecific predation, Ann. Rev. Ecol.
Syst. 12 (1981), 225-251
[347] G. A. Polis, Age structure component of niche width and intra-specific resource parti-
tioning: Can age groups function as ecological species?, Amer. Nat. 123 (1984), 541-564
[348] J. H. Pollard, Mathematical Models for the Growth of Human Populations, Cambridge
University Press, London. 1975
[349] T. Prout and F. McChensney, Competition among immatures affects their adult fertility:
Population dynamics, Amer. Nat. 126 (1985), 521-558
[350] J. Priiss, Equilibrium solutions of age-specific population dynamics of several species,
J. Math. Biol. 11 (1981), 65-84
[351] J. Priiss, On the qualitative behaviour of populations with age-specific interactions,
Comput. Math. Appl. 9 (1983), 327-339
[352] J. Pruss, Stability analysis for equilibria in age-specific population dynamics, Nonlinear
Anal. 7 (1983), 1291-1313
[353] J. Priiss and W. Schappacher, Semigroup methods for age-structured population dy-
namics, Jahrbuch Uberblicke Mathematik, Vieweg, Braunschweig, 1994, 74-90
[354] P. H. Rabinowitz, Some global results for nonlinear eigenvalue problems, J. Fund. Anal.
7 (1971), 487-513
[355] S. Rai and D. Brewer, Analysis of a nonlinear FDE for an age-structured population,
J. Natur. Geom. 11 (1997) 57-74
[356] J. Reed and N. C. Stenseth, Evolution of cannibalism in an age-structured population,
Bull. Math. Biol. 46 (1984), 371-377
[357] W. J. Reed, Recruitment variability and age structure in harvested animal populations,
Math. Biosci. 65 (1983), 239-268
[358] E. Renshaw, Modelling Biological Populations in Space and Time, Cambridge Univer-
sity Press, Cambridge, 1991
[359] D. J. Rodriquez, Models of growth with density dependence in more than one stage,
Theoret. Population Biol. 34 (1988), 93-119
[360] P. Rohani and D. J. D. Earn, Chaos in a cup of flour, Trends m Ecol. and Evol. 12
(1997), 171
[361] C. Rorres. Stability of an age specific population with density dependent fertility, The-
oret. Population Biol. 10 (1976), 26-46
[362] C. Rorres, A nonlinear model of population growth in which fertility is dependent on
birth rate, SIAM J. Appl. Math. 37 (1979), 423-432
[363] C. Rorres, Local stability of a population with density-dependent fertility, Theoret.
Population Biol. 16 (1979), 283-300
[364] S. Rosenblat, Population models in a periodically fluctuating envrinoment, J. Math.
Biol 9 (1980), 23-36
 BIBLIOGRAPHY 187

[365] M. Rotenberg, Equilibrium and stability in populations whose interactions are age-
specific, J. Theor. Biol. 54 (1975), 207-224
[366] S. I. Rubinov, Age-structured equations in the theory of cell populations, Studies in
Mathematical Biology II (S. A. Levin, ed.), MAA, Washington, 1978
[367] W. Rundell, Determining the death rate for an age-structured population from census
data, SIAM J. Appl. Math. 53 (1993), 1731-1746
[368] K. Saints, Discrete, and Continuous Models of Age-Structured Population Dynamics,
Senior thesis, Harvey Mudd College, Claremont, CA. 1987
[369] M. Saleem, Predator-prey relationships: Egg-eating predators, Math. Biosci. 65 (1983),
187-197
[370] M. Saleem, Egg-eating age-structured predators in interaction with age-structured prey,
Math. Biosci. 70 (1984), 91-104
[371] M. Saleem, Predator-prey relationships: Indiscriminate predation, ,7. Math. Biol. 21
(1984), 25-34
[372] M. Saleem, On interacting age-structured populations, Math. Student 55 (1989), 232
240
[373] W. M. Schaffer, Selection for optima] life histories: The effects of age structure. Ecology
55 (1974), 291-303
[374] F. M. Scudo and J. R. Ziegler, The Golden Age of Theoretical Ecology: 1923-1940,
Lecture Notes in Biomath. 22, Springer-Verlag, Berlin, 1978
[375] F. R. Sharpe and A. J. Lotka, A problem in age-distribution, Phil. Magazine 21 (1911),
435-438
[376] H. R. Siegismund, V. Loeschcke, and J. Jacobs, Intraspecific competition and compo-
nents of niche width in age structured populations, Theoret. Population B'lol. 37 (1990),
291-319
[377] J. A. L. Silva arid T. G- Hallam, Compensation and stability in nonlinear matrix models,
Math. Biosci. 110 (1992), 67-101
[378] S. D. Simmes, Age Dependent Population Dynamics with Nonlinear Interactions, Ph.D.
dissertation, Carnegie-Mellon University, Pittsburgh, PA, 1978
[379] E. Sinestrari. Nonlinear age-dependent population growth, J. Math. Biol. 9 (1980),
331-345
[380] J. W. Sinko and W. Streifer. A new model for age-size structure of a population, Ecology
48 (1967), 910-918
[381] J. W. Sinko and W. Streifer, Applying models incorporating age-size structure of a.
population to Daphnia, Ecology 50 (1969). 608-615
[382] D. Smith and N. Keyfitz, Mathematical Demography, Springer-Verlag. New York, 1977
[383] H. L. Smith, Equivalent dynamics for a structured population model and a related
functional-differential equation. Rocky Mountain J. Math. 25 (1995), 491-499
[384] H. L. Smith, Reduction of structured population models to threshold-type delay equa-
tions and functional-differential equations: A case study, Math. Biosci. 113(1993), 1-23
[385] H. L. Smith, A structured population model and a related functional-differential equa-
tion: Global attractors and uniform persistence, J. Dynam. Differential Equations 6
(1994), 71-99
[386] H. L. Smith, A discrete size-structured model of microbial growth and competition in
the chemostat, J. Math. Biol. 34 (1996), 734-754
[387] H. L. Smith and P. Waltman, The Theory of the Chemostat: Dynamics of Microbial
Competition, Cambridge Studies in Mathematical Biology, Cambridge University Press,
Cambridge, 1995
[388] J. L. Smith and D. J. Wollkind, Age structure in predator-prey systems: Intraspecific
carnivore interaction, passive diffusion and the paradox of enrichment, ./. Math. Biol.
17 (1983), 275-288
188 BIBLIOGRAPHY

[389] R. H. Smith and R. Mead, Age structure and stability in models of prey-predator
systems, Theoret. Population Biol. 6 (1974), 308-322
[390] S. J. Smith, Maximum sustainable yield of a density dependent sex-age-structured pop-
ulation model, J. Math. Biol. 30 (1992), 801-814
[391] P. E. Smouse and K. M. Weiss, Discrete demographic models with density-dependent
vital rates, Oecologia 21 (1975), 205-218
[392] W. Streifer, Realistic models in population ecology, Advances in Ecological Research,
Vol. 8 (A. Macfadyen, ed.), Academic Press, New York, 1974
[393] D. Sulsky, R. R. Vance, and W. I. Newman, Time delays in age-structured populations,
J. Theor. Biol. 141 (1989), 403-422
[394] D. Sulsky, Numerical solution of structured population models. I. Age structure, J.
Math. Biol. 31 (1993), 817-839
[395] D. Sulsky, Numerical solution of structured population models. II. Mass structure, J.
Math. Biol. 32 (1994), 491-514
[396] J. H. Swart, The stability of a nonlinear, age dependent population model, as applied
to a biochemical reaction tank, Math. Biosci. 80 (1986), 47-56
[397] K. E. Swick, Stability and bifurcation in age-dependent population dynamics, Theoret.
Population Biol. 20 (1981), 80-100
[398] K. E. Swick, A nonlinear model for human population dynamics, SIAM J. Appl. Math.
40 (1981), 266-278
[399] H. R. Thieme, Well-posedness of physiologically structured population models for Daph-
nia magna. How biological concepts can benefit by abstract mathematical analysis, J.
Math. Biol. 26 (1988), 299-317
[400] H. R. Thieme, Analysis of age-structured population models with an additional struc-
ture, Mathematical Population Dynamics (O. Arino, D. E. Axelrod, and M. Kimmel,
eds.), Lecture Notes in Pure and Appl. Math. 131, Dekker, New York, 1991, 115-126
[401] R. W. Thompson, D. di Blasio, and C. Mendes, Predator-prey interactions: Egg-eating
predators, Math. Biosci. 60 (1982), 109-120
[402] W. O. Tschumy, Competition between juveniles and adults in age-structured popula-
tions, Theoret. Population Biol. 21 (1982), 255-268
[403] H. Tong, Nonlinear Time Series: A Dynamical System Approach, Oxford University
Press, Oxford, 1990
[404] S. L. Tucker and S. O. Zimmerman, A nonlinear model of population dynamics con-
taining an arbitrary number of continuous structure variables, SIAM J. Appl. Math. 48
(1988), 549-591
[405] S. D. Tuljapurkar, Population dynamics in variable environments IV: Weak ergodicity
in the Lotka equation, J. Math. Biol. 14 (1982), 221-230
[406] S. Tuljapurkar, Entropy and convergence in dynamics and demography, J. Math, Biol.
31 (1993), 253-271
[407] S. Tuljapurkar, Stochastic demography and life histories. Frontiers in Mathematical
Biology (Simon A. Levin, ed.), Lecture Notes in Biomath. 100, Springer-Verlag, Berlin,
1994, 254-262
[408] S. Tuljapurkar and H. Caswell, Structured-Population Models in Marine, Terrestrial,
and Freshwater Systems, Chapman and Hall. New York, 1997
[409] G. Uribe, On the Relationship Between Continuous and Discrete Models for Size-
Structured Population Dynamics, Ph.D. dissertation, Interdisciplinary Program in Ap-
plied Mathematics, University of Arizona, Tucson, AZ, 1993
[410] M. B. Usher, A matrix approach to the management of renewable resources, with special
reference to selection forests, J. Appl. Ecol. 3 (1966), 355-367
[411] M. B. Usher, A matrix approach to the management of renewable resources, with special
reference to selection forests-two extensions, J. Appl. Ecol. 6 (1969), 347-348
 BIBLIOGRAPHY 189

[412] M. B. Usher, A matrix model for forest management. Biometrics 25 (1969), 309-315
[413] M. B. Usher, Developments in the Leslie matrix model, Mathematical Models in Ecology
(J. M. R. Jeffers, ed.), Blackwell Scientific Publishers, London, 1972, 29-60
[414] M. B. Usher, Extensions to modes, used in renewable resource management, which
incorporate an arbitrary structure. J. Environ. Man. 4 (1976), 123-140
[415] M. B. Usher and M. H. Williamson. A deterministic matrix model for handling the
birth, death and migration processes of spatially distributed populations, Biometrics
26 (1970), 1-12
[416] F. van den Bosch and O. Diekmann, Interactions between egg-eating predator and prey:
The effect of the functional response and of age structure, IMA J. Math. Appl. Med.
Biol 3 (1986), 53-69
[417] F. van den Bosch, A. M. de Roos, and W. Babriel, Cannibalism as a life boat mechanism,
J. Math. Biol. 26 (1988), 619 633
[418] J. van Sickle, Analysis of a distributed-parameter population model based on physio-
logical age, J. Theor. Biol 64 (1977), 571 586
[419] R. R. Vance, W. I. Newman, and D. Sulsky. The demographic meanings of the classical
population growth models of ecology, Theoret. Population Biol. 33 (1988), 199-225
[420] J. Vandermeer, Choosing category size in a stage projection matrix, Oecologia 32 (1978),
79-84
[421: E. Venturino, Age-structured predator-prey models. Math. Mod. 5 (1984), 117-128
[422] M. O. Vlad, The occurrence of time-persistent distribution functions in structured pop-
ulation dynamics, J. Theor. Rial 126 (1987), 239-242
[423] M. O. Vlad, Separable models for age-structured population genetics, J. Math. Biol. 26
(1988), 73-92
[424] H. Von Foerster, Some remarks on changing populations, The Kinetics of Cellular
Proliferation (F. Stohlman, ed.), Grime and Stratton, New York, 1959. 382 407
[425] P. Waltmaii, Competition Models in Population Biology, Regional Conference Series in
Applied Mathematics 45, SIAM, Philadelphia, PA, 1983
[426] F. S. .1. Wang. Stability of an age-dependent population, SIAM J. Math. Anal. 11
(1980), 683-689
[427i G. F. Webb, Nonlinear age-dependent population dynamics in L 1 , J. Integral Equations
5 (1983), 309-328
[428] G. F. Webb, The semigroup associated with nonlinear age dependent population dy-
namics, Int. J. Camp. Math, and Appl 9 (1983), 487-497
[429] G. F. Webb, Theory of Nonlinear Age-Dependent Population Dynamics, Marcel Dekker,
New York, 1985
[430] G. F. Webb, Dynamics of populations structured by internal variables. Math. Z. 189
(1985), 319 335
[431] G. F. \Vebb, Logistic models of structured population growth. Hyperbolic Partial Dif-
ferential Equations, III. Comput. Math. Appl Part A 12 (1986), 527-539
[432] G. F. Webb, An operator-theoretic formulation of asynchronous exponential growth,
Trans. Amer. Math. Soc. 393 (1987), 751-763
[433] G. H. Weiss, Equations for the age structure of growing populations, Bull. Math. Bto-
phys. 30 (1968), 427-435
[434j P. A. Werner and H. Caswell, Population growth rates and age versus stage-distribution
models for teasel (Dipsa,cus sylvestria Huds.j, Ecology 58 (1977), 1103-1111
[435] E. E. Werner and J. F. Gilliam, The ontogenetic niche and species interactions in size-
structurcd populations. Ann. Rev. Kc.ol. Syst. 15 (1984), 393-425
[436] S. Wiggins. Introduction to Applied Nonlinear Dynamical Systems and Chaos, Texts in
Applied Mathematics 2, Springer-Verlag, Berlin. 1990
190 BIBLIOGRAPHY

[437] A. Wikan and E. Mjolhus, Periodicity of 4 in age-structured population models with

density-dependence, J. Theor. Biol 173 (1995), 109-119
[438] S. A. Wissinger, Niche overlap and the potential for competition and intraguild preda-
tion between size-structured populations, Ecology 73 (1992), 1431-1444
[439] D. Wollkind, A. Hastings, and J. Logan, Age structured in predator-prey systems II.
Functional response and stability and the paradox of enrichment, Theoret. Population
Biol. 21 (1982), 57-68
[440] I. O. Woodward, Modelling population growth in state-grouped organisms: A simple
extension of the Leslie model, Austr. J. of Ecol. 7 (1982), 389-394
[441] M. E. J. Woolhouse and R. Harmsen, A transition matrix model of the population
dynamics of a two-prey-two-predator acarid complex, Ecol. Model. 39 (1987), 307-323
[442] Z. Yicang and J. M. Gushing, Stability conditions for equilibria of nonlinear matrix
population models, J. Differ. Equations Appl, in press
[443] J. Yuan, Periodic solutions of a nonlinear autonomous age-dependent model of single
species population dynamics, Nonlinear Anal. 12 (1988), 1079-1086.
[444] B. Xu, A Discrete Nonlinear Model of Stage-Structured Populations, Ph.D. dissertation,
Interdisciplinary Program in Applied Mathematics, University of Arizona, Tucson, AZ,
1995
[445] D. W. Zachmann and J. A. Logan, Mathematical stability analysis for an age-structured
population with finite life span, J. Theor. Biol. Ill (1984), 131-148
[446] D. W. Zachmann, Unique solvability of an age-structured population model with can-
nibalism, Rocky Mountain J. Math. 17 (1987), 749-766
[447] A. A. S. Zaghrout and S. H. Attalah, Analysis of a model of stage-structured population
dynamics growth with time state-dependent time delay, Appl. Math. Comput. 77 (1996),
185-194
[448] A. D. Ziebur, Age-dependent models of population growth, Theoret. Population Biol.
26 (1984), 315-319
[449] M. F. Alameddine, Size-Structured Competition Models with Periodic Coefficients,
Ph.D. dissertation, Interdisciplinary Program in Applied Mathematics, University of
Arizona, Tucson, AZ, 1993
[450] Z. M. Alawneh, A Numerical Method for Solving Certain Nonlinear Integral Equations
Arising in Age-Structured Population Dynamics, Ph.D. dissertation, University of Ari-
zona, Tucson, AZ, 1990
 Index

Allee effect, 14, 26, 32, 34, 91, 94 Cannibalism, 37. 139
Allocation fractions, 111, 112, 128 Chaos, 57, 58, 90, 107, 110, 117
Asymptotic stability, 16, 154 Characteristic value. 3, 161
Attractor, 16, 154 Chemostat, 74, 117, 128, 131
Class distribution vector, 2
Beverton-Holt nonlinearity, 13, 106 Coexistence, 98
Bifurcation, 7, 24, 25, 46 competitive
direction of, 23, 24, 67, 99 equilibrium, 71, 116
discrete Hopf (Naimark/Sacker), periodic arid chaotic, 116
28, 34, 39, 58 host/parasite, 73
global, 163 predator/prey. 70
Hopf, 90 Competition
of 2-cycles, 27, 34, 39, 46, 57, exploitative, 120, 127
116 interference, 113
of chaotic attractors, 117 interspecific, 15, 64, 70, 100, 113,
of positive equilibria, 22, 23, 30, 120, 127
66, 67, 87, 89, 98 intraspccific, 34, 43, 70, 92, 110,
of strange attractors, 117 133. 136, 141
pitchfork. 25 juvenile versus adult. 43, 92
point. 7, 23, 88. 89 scramble versus contest, 133, 136,
saddle-node, 25, 43, 96 141
stable, 24, 27, 30, 34, 67, 89, 92 size-structured, 34, 110, 127
to the left or right, 23, 24, 30, Competitive exclusion principle, 71.
67, 88, 89, 99 114, 132
traiiscritical, 24, 25, 67, 89 violation of, 117
unstable, 24, 27, 30, 34, 42, 67. Continuum. 161
92 Critical value, 20, 22
vertical. 7, 116 strictly dominant, 20, 24, 45
Body size, 1, 44, 100, 127 Density effects, 2, 12
length, 110, 128 advantageous, 14
surface area, 110, 129 deleterious, 13, 31
volume, 111, 129 Dissipative, 17
weight, 111
Break-even concentration. 120. 132 Eigenvalue and eigenvectors, 3
191
192 INDEX

Equilibrium, 16, 84, 86 equilibrium, 16, 156, 160

asymptotically stable, 16, 154, matrix, 16, 148
155
attractor, 16, 154 Inherent growth rate, 7, 83
bifurcation of, 25 Inherent net reproductive number,
destabilization, 25, 90 7, 10, 19, 28, 68, 84, 98,
exchange of stability, 26, 89, 99 101, 126, 138, 142
extinction, 7, 16. 20, 65 for Leslie matrix, 7
hyperbolic, 16 for Usher matrix, 11
nondegenerate, 66 Inherent parameter, 19, 34, 42, 104
stable, 16, 154 Invariant loop, 27, 39, 58
trivial, 65, 98 Irreducible matrix, 3
unstable, 16, 154
Equilibrium pair, 21, 87 Jacobian, 16, 24, 25, 67, 154
boundary, 21, 22
continuum of, 21, 22, 87 Least squares parameterization, 54,
extinction, 21 57
global bifurcation of, 22, 29, 87 Leslie matrix, 4, 78, 81, 104, 106,
local bifurcation of, 23, 24, 66, 141
89 Liapunov/Schmidt expansion, 23, 66,
nonextinction, 21 87. 98, 163, 165
nonnegative, 21 Limiting equation, 7, 85, 105, 108,
positive, 21, 22, 29-31, 66, 87 113, 120, 122
Ergodic, 6, 103, 108, 110 Linear chain trick, 123
Logistic, 122, 128, 130
Fertility matrix, 2 LPA model, 37, 56
normalized, 19
Fertility window, 126 Matrix equation, 2
Flour beetle, 34, 37, 51, 54 Maturation
Fundamental Theorem of Demogra- period, 92, 127, 145
phy, 7, 104 size, 34, 44
Fundamental Theorem of Stability, Maximum likelihood parameter es-
154 timates, 49
McKendrick equation, 77, 81
Global extinction, 18. 19, 36 Allee effect, 94
Gronwall's inequality, 154 derivation of, 77, 78
Growth efficiency coefficient, 110, 115, extinction equilibrium, 82
129 hierarchical, 134
Hopf bifurcation, 90
Hartman/Grobman Theorem, 159 inherent growth rate, 83
Hierarchical models, 133 inherent net reproductive num-
continuous, 134 ber, 84
discrete, 139 limiting equation, 85, 142
Rolling uptake rates, 131, 138 linear, 82
Host/parasite, 71, 100 linear chain trick. 123
Hyperbolic multispecies, 97
 INDEX 193

net reproductive number, 87 Reproductive efficiency coefficient, 110,

nonlinear, 85 115, 129
normalized age distribution, 84 Residuals, 48, 51
positive equilibria, 87, 89 Ricker-type nonlinearity, 13. 36. 50.
separable, 121 51, 56, 107, 110
stable age distribution, 84
Michaelis/Menten, 74, 117, 131, 138 Solution, 2, 81, 83
Model validation, 53, 56 backward, 147, 156
bounded, 147
Net reproductive number, 28, 69, 87 existence, 16, 77. 81, 122, 156
for Usher matrix, 34 forward, 2, 16, 147, 156
Nonnegative matrix, 3 pair, 21, 161
Nonnegative vector, 3 periodic, 27, 90
Normalized distribution, 6, 84, 103. positivity, 16, 77
105, 108, 121 that tends to zero, 147
Normalized fertility matrix, 19, 29, uniqueness. 16, 77, 81, 122
68 Spectrum, 7, 19, 29, 30, 87, 116
Spotted owl. 15. 23, 26, 34. 41
Periodic habitat, 61 Stability, 16, 154
Periodic LPA model, 62 Fundamental Theorem of. 154
Perron/Frobenius Theorem. 4 Stable distribution, 6. 84, 121
Persistence, 18, 36, 38, 167 Stable manifold, 17, 156
Positive vector. 3 Stable Manifold Theorem. 156
Predator/prey, 15, 64, 70, 100, 126 Stochastic models, 48. 56, 57
Prediction fit, 53, 56, 58 Strictly dominant eigenvalue. 3, 83
Primitive matrix. 4
Projection matrix, 2 Total population size, 13, 31. 32, 80,
inherent, 16, 65 98, 105, 108, 121, 123, 136,
multispecies, 15, 65 139, 142
nonlinear, 12 Total population surface area, 111
Total population volume, 129
Range, 29, 40 Transition matrix, 2. 68
Rates
Unstable manifold, 156
death, 2, 37, 120, 123, 134, 142
Usher matrix, 5, 10. 32. 34, 100, 110
fertility, 2, 12, 15, 30, 32, 44,
80, 92, 94, 97, 100, 104, Variation of constants formula, 147
109, 123, 134, 139, 142 modified, 149. 152
growth, 100, 128 Volterra, 71, 83, 122
metabolism, 2, 128
resource consumption, 2, 109.
110, 128, 136
survival, 109, 140
transition, 12, 15, 30, 32, 80,
92, 94, 97, 100, 104, 108
washout, 131
Reducible matrix. 3