Modeling Population Dynamics

Andr M. de Roos e

Modeling Population Dynamics

Andr M. de Roos e
Institute for Biodiversity and Ecosystem Dynamics Population Biology Section, University of Amsterdam Kruislaan 320, 1098 SM Amsterdam, The Netherlands E-mail: [email protected]

23 december 2011

Inhoudsopgave
I
1

Preface and Introduction

Introduction 1.1 Some modeling philosophy . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1
3 3

II
2

Unstructured Population Models in Continuous Time

Modelling population dynamics 2.1 Describing a population and its environment . . . . . . . . . . . . . . . . . . 2.1.1 The population or p-state . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.2 The individual or i-state . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.3 The environmental or E-condition . . . . . . . . . . . . . . . . . . . . . 2.2 2.3 2.4 2.5 Population balance equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . Characterizing the population . . . . . . . . . . . . . . . . . . . . . . . . . . . Population-level and per capita rates . . . . . . . . . . . . . . . . . . . . . . . Model building . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.1 Exponential population growth . . . . . . . . . . . . . . . . . . . . . . . 2.5.2 Logistic population growth . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.3 Two-sexes population growth . . . . . . . . . . . . . . . . . . . . . . . . 2.6 2.7 Parameters and state variables . . . . . . . . . . . . . . . . . . . . . . . . . . Deterministic and stochastic models . . . . . . . . . . . . . . . . . . . . . . .

5
7 7 7 8 9 10 11 12 15 15 16 17 18 19 21 21 23 24 28 32 35 36 37 38

Single ordinary dierential equations 3.1 3.2 3.3 3.4 3.5 3.6 3.7 Explicit solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Numerical integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Analyzing ow patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Steady states and their stability . . . . . . . . . . . . . . . . . . . . . . . . . . Units and non-dimensionalization . . . . . . . . . . . . . . . . . . . . . . . . . Existence and uniqueness of solutions . . . . . . . . . . . . . . . . . . . . . . Epilogue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Competing for resources 4.1 Intraspecic competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i

ii

INHOUDSOPGAVE 4.1.1 Growth of yeast in a closed container . . . . . . . . . . . . . . . . . . . . 4.1.2 Bacterial growth in a chemostat . . . . . . . . . . . . . . . . . . . . . . . 4.1.3 Asymptotic dynamics 4.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.4 Phase-plane methods and graphical analysis . . . . . . . . . . . . . . . . Interspecic competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Lotka-Volterra competition model . . . . . . . . . . . . . . . . . . . . . . 4.2.2 Competition for resources . . . . . . . . . . . . . . . . . . . . . . . . . . 38 40 43 44 52 52 57 77 77 78 81 84 85 90 94 96 99 100 104 108 117 119 123 127

Systems of ordinary dierential equations 5.1 5.2 5.3 5.4 Computation of steady states . . . . . . . . . . . . . . . . . . . . . . . . . . . Linearization of dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Characteristic equation, eigenvalues and eigenvectors . . . . . . . . . . . . . . Phase portraits of dynamics in planar systems . . . . . . . . . . . . . . . . . . 5.4.1 Two real eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.2 Two complex eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5 5.6 Stability of steady states in planar ODE systems . . . . . . . . . . . . . . . . Models with 3 or more variables . . . . . . . . . . . . . . . . . . . . . . . . . .

Predator-prey interactions 6.1 The Lotka-Volterra predator-prey model . . . . . . . . . . . . . . . . . . . . . 6.1.1 Incorporating prey logistic growth . . . . . . . . . . . . . . . . . . . . . 6.1.2 Incorporating a predator type II functional response . . . . . . . . . . . 6.2 Confronting models and experiments . . . . . . . . . . . . . . . . . . . . . . . 6.2.1 Predator-controlled, steady state abundance of prey . . . . . . . . . . . 6.2.2 Oscillatory dynamics at high prey carrying capacities . . . . . . . . . . . 6.2.3 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

III
7

Bifurcation theory
Continuous time models 7.1 7.2 7.3 7.4 7.5 General setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Transcritical bifurcation and branching point . . . . . . . . . . . . . . . . . . Saddle-node bifurcation and limit point . . . . . . . . . . . . . . . . . . . . . Hopf bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

129
131 132 140 141 142 144

IV
8

Exercises (in Dutch)

Oefenopgaven 8.1 ODEs (dierentiaalvergelijkingen) in 1 dimensie . . . . . . . . . . . . . . . . . 8.1.1 Populatiegrootte en populatiegroeisnelheid . . . . . . . . . . . . . . . . .

145
147 147 147

INHOUDSOPGAVE 8.1.2 Grasche analyse van een dierentiaalvergelijking . . . . . . . . . . . . . 8.1.3 Een model opstellen en grasch analyseren . . . . . . . . . . . . . . . . . 8.2 8.3 Evenwichten in 1D ODEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.1 Wiskundige analyse van een dierentiaalvergelijking . . . . . . . . . . . Het onderzoeken van ODEs in 2D . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.1 Grasche analyse van een stelsel dierentiaalvergelijkingen . . . . . . . . 8.3.2 Wiskundige analyse van een stelsel dierentiaalvergelijkingen . . . . . . 8.4 Voorbeeld tentamen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

iii 147 149 151 151 153 153 154 155

V
9

Instructions for using CONTENT

CONTENT: continuation software 9.1 9.2 9.3 9.4 How to use this manual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . How to start using CONTENT . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2.1 Setup of CONTENT program . . . . . . . . . . . . . . . . . . . . . . . On startup of Content 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Basic system manipulation in Content 1.5 . . . . . . . . . . . . . . . . . . . . 9.4.1 Entering the system management window . . . . . . . . . . . . . . . . . 9.4.2 Create a new system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4.3 Manipulating already dened systems . . . . . . . . . . . . . . . . . . . 9.5 Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.5.1 Orbits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.5.2 Equilibria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.5.3 Bifurcation points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.6 9.7 Storage of computed curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . Special cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.7.1 Limit cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.7.2 Continuing branching points . . . . . . . . . . . . . . . . . . . . . . . . . 9.8 Visualizing output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.8.1 Special case 1: plotting limit cycles . . . . . . . . . . . . . . . . . . . . . 9.8.2 Special case 2: plots of eigenvalues in the complex plane . . . . . . . . .

171
173 173 173 173 174 174 174 174 175 175 176 176 177 177 178 178 178 180 180 181

VI
10

Computer Labs
Computer Labs 10.1 Harvesting Cod . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2 The spruce-budworm(Choristoneura fumiferana) . . . . . . . . . . . . . . . . 10.3 Interspecic competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4 Vegetation catastrophes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.5 Lotka-Volterra Predation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.6 The Paradox of Enrichment . . . . . . . . . . . . . . . . . . . . . . . . . . . .

183
185 185 187 189 190 191 192

iv

INHOUDSOPGAVE 10.7 Chaotic dynamocs of Hare and Lynx populations . . . . . . . . . . . . . . . . 10.8 The Return of the Paradox of Enrichment . . . . . . . . . . . . . . . . . . . . 10.9 Cannibalism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.10A predator-prey model with density-dependent prey development . . . . . . . 194 195 196 197 199

Bibliography

Lijst van guren

2.1 2.2 2.3 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 Growth of the world population over the last century . . . . . . . . . . . . . . Population age distribution in the Netherlands in 1950, 1975 and 2000 . . . . Total number of births and death per 5 year age class in the Netherlands in 1999 Growth of the yeast Schizosaccharomyces kephir over a period of 160 h . . . . Schematic layout of a chemostat for continuous culture of micro-organisms . . Geometric representation of a system of ODEs in the phase plane . . . . . . . The phase plane of the bacterial growth model in a chemostat . . . . . . . . . Operating diagram of a chemostat . . . . . . . . . . . . . . . . . . . . . . . . Four isocline cases for the Lotka-Volterra competition model . . . . . . . . . . Four isocline cases with steady states and ow patterns for the Lotka-Volterra competition model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Michaelis-Menten relationship as a function of resource concentration . . . . . Zero net growth isoclines for a one consumer-two resource model . . . . . . . 8 9 10 38 41 46 50 52 54 55 59 62 63 66 70 71 71 73 73 74 74 75 75 76

4.10 Steady state location in a one consumer-two resource model . . . . . . . . . . 4.11 Steady state location in a two consumer-two resource model . . . . . . . . . . 4.12 Predicted and observed outcomes of competition for phosphate and silicate by Asterionella formosa and Cyclotella meneghiniana . . . . . . . . . . . . . . . 4.13 Single-species Monod growth experiments for four diatoms under conditions of limited silicate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.14 Single-species Monod growth experiments for four diatoms under conditions of limited phosphate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.15 Predicted and observed outcomes of competition between Asterionella formosa and Synedra liformis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.16 Predicted and observed outcomes of competition between Asterionella formosa and Tabellaria occulosa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.17 Predicted and observed outcomes of competition between Asterionella formosa and Fragilaria crotonensis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.18 Predicted and observed outcomes of competition between Synedra and Fragilaria 4.19 Predicted and observed outcomes of competition between Synedra and Tabellaria 4.20 Predicted and observed outcomes of competition between Fragilaria and Tabellaria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.21 The four cases of resource competition in a two consumer-two resource model v

vi 5.1 5.2 5.3 5.4 5.5 6.1 6.2 6.3 6.4 6.5 6.6 6.7 6.8 6.9

LIJST VAN FIGUREN Geometric representation of two real-valued eigenvectors in a planar system . Characteristic ow patterns in the neighborhood of the steady state with realvalued eigenvalues and eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . Geometric representation of the two basis vectors in a planar system with complex eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Characteristic ow patterns in the neighborhood of the steady state with complex eigenvalues and eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . Summary of stability properties for planar ODE systems . . . . . . . . . . . . Solution curves in the phase plane of the Lotka-Volterra predator-prey model Prey dynamics predicted by the Lotka-Volterra predator-prey model . . . . . Solution curves in the phase plane of the Lotka-Volterra predator-prey model with logistic prey growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Functional response of Daphnia pulex on three algal species . . . . . . . . . . Nullclines of the Rosenzweig-MacArthur, predator-prey model . . . . . . . . . Solution curves in the phase plane of the Rosenzweig-MacArthur predator-prey model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Solution curve in the phase plane of the Rosenzweig-MacArthur predator-prey model for high carrying capacity . . . . . . . . . . . . . . . . . . . . . . . . . Oscillatory dynamics of the Rosenzweig-MacArthur predator-prey model for high carrying capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Daphnia pulex carrying asexual eggs in the brood pouch . . . . . . . . . . . . 87 88 91 92 95 102 103 107 109 112 116 117 118 119 120 122 123 124 125 125 126 127 128 128

6.10 Schematic setup of the experiments by Arditi et al. (1991) . . . . . . . . . . . 6.11 Population dynamics of Daphnia, Ceriodaphnia and Scapholeberis in a chain of semi-chemostats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12 Population equilibria of Daphnia and Ceriodaphnia in a chain of semi-chemostats 6.13 Predicted and observed equilibrium values of algae in the presence of Daphnia 6.14 Population dynamics of the snowshoe hare and the lynx in northern Canada . 6.15 Population dynamics of two species of voles in northern Finland . . . . . . . 6.16 Cycle amplitudes (log) observed for Daphnia and algal populations from lakes and ponds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.17 Examples of the dynamics of Daphnia and algae in nutrient-rich and nutrientpoor, experimental tanks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.18 Large- and small-amplitude cycles of Daphnia and edible algae in the same global environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.19 Energy channelling towards sexual reproduction prevents the occurrence of largeamplitude predator-prey cycles in Daphnia . . . . . . . . . . . . . . . . . . . 7.1 7.2 7.3 7.4 Eigenvalue positions in the complex plane corresponding to a stability change in a 2 ODE model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bifurcation structure of the cannibalism model for < . . . . . . . . . . . . Bifurcation structure of the cannibalism model for > . . . . . . . . . . . .

134 138 139

Schematic, 3-dimensional representation of the Hopf-bifurcation in the RosenzweigMacArthur model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

Lijst van tabellen

4.1 4.2 4.3 5.1 5.2 7.1 State variables and parameters of a bacterial growth model in a chemostat . . State variables and parameters of Tilmans resource competition model . . . Silicate and phosphate parameters for four Lake Michigan diatoms . . . . . . Important characteristics of systems consisting of two linear(ized) dierential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Routh-Hurwitz stability criteria (following May 1974) . . . . . . . . . . . 42 58 72

86 97

Characteristics of all possible types of steady states in a 2-dimensional, continuoustime model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

vii

viii

LIJST VAN TABELLEN

Deel I

Preface and Introduction

Hoofdstuk 1

Introduction
This course is intended as an introduction to the formulation, analysis and application of mathematical models that describe the dynamics of biological populations. It starts at a very basic level, probably repeating some material that is also part of an introductory ecology course. Nonetheless, I think this rehearsal is necessary and useful, while it allows me to emphasize some theoretical aspects that are certainly not part of an introductory ecology course. The course material brought together here stems to a larger or lesser extent from earlier textbooks on theoretical ecology. Most notably, I have used the books by Edelstein-Keshet (1988), Yodzis (1989) and Murray (1989) as sources for the text presented. However, all of these books have slightly dierent approaches and put emphasis on slightly dierent aspects of theoretical ecology. My intention is to emphasize more the biological and the conceptual aspects of the theory in that I devote quite some attention to the formulation of models (the model building stage) and the interpretation of the mathematical analysis in terms of biological conclusions. This approach is probably most similar to the approach in Yodzis (1989), although I also add some new material and I hope to put even more emphasis on biological case studies.

1.1

Some modeling philosophy

In my opinion it is fair to say that every scientist in one way or another uses models, although they are by far not all mathematical. It is indeed hard to imagine to do science without a model: even the most empirical scientist will have some idea or hypothetical/stylized representation in his mind of the system that he is working on. This mental representation subsequently guides him in making new observations that lead to a better understanding of the system studied. In my opinion a model should hence not strife for a description, but rather a conceptualization of a system. Moreover, this conceptualization should not just be a static reection, but should incorporate the workings or mechanics of the system. I hence as much as possible advocate a mechanistic approach to model building. Given that a model is a conceptualization or abstraction, by denition it is also an incomplete and often even a false representation of the system. The analysis of models that turn out to be at odds with observations often elucidates more insight about the system than a model whose predictions are roughly in line with observations. In this respect, I tend to compare models with the H0 and H1 hypotheses in statistics: Only a rejection of the H0 -hypothesis in favor of the H1 -hypothesis makes a strong statement about a particular phenomenon. Loosely speaking one could therefore say: The best model is a wrong one! To students that are new to modelling, models always seem very simplistic and hence a poor reection of the real world. This often leads to the misconception that experimental results 3

HOOFDSTUK 1. INTRODUCTION

convey more information about a particular system than the analysis of a model. Moreover, many people are likely to think that making a model and analysing it is a quick and easy job, while experiments take much more time and eort. I would strongly argue against both these misconceptions. First of all, any attempt, be it experimental or theoretical, to comprehend how a particular system works is forced to use a simplied conceptualization of the system. In my opinion the main dierence is that, especially mathematical, models are extremely explicit about the assumptions they make. Experimental studies often make similar assumption, but very implicitly. Caswell (1988) has nicely discussed this issue using data that he obtained from a purely empirical poster session during the IVth International Congress of Ecology in Syracuse, N.Y., in 1986: The majority of experimental studies presented at the poster session only considered one or two, at most three factors in their experimental setup. These studies did not assume that other factors were unimportant, but they were simply not part of the study. While building a mathematical model, a theoretician would quickly be forced to make this particular situation explicit by making a statement like I assume that all other factors are constant. Even though in essence the experimental and theoretical study are not dierent in their basic assumptions, the explicitness with which a modeller states his assumptions often meets with a lot of opposition. Also the perception that modeling studies are quick and easy is rather misleading. It is indeed true that it only takes an afternoon or two to write down some equations that with a lot of handwaving might be argued to reect a biological system. Given the availability of modern software packages, the same afternoon would also provide sucient time to analyse the characteristics of this set of equations. Such an exercise would be comparable with a very simple pilot experiment, which could in some cases also be carried out in a single afternoon. It is very unlikely that either of the two, the quick and easy modeling exercise and the pilot experiment, would yield results that are worth publishing. From my own experience of using models I can only conclude that using mathematical models for developing theory which has a sound logical basis and moreover deepens our insight about a particular system in my opinion takes just as long as an experimental approach to gaining the same insight (if it were feasible). There is essentially less dierence between the two approaches as one might think of at rst. There are good reasons for mathematical models to be so widely used in ecology. Ecosystems tend to be very complex and governed by many intricate and usually non-linear mechanistic interactions. Thinking through these complex relationships might be attempted with just a verbal reasoning approach. However, as Yodzis (1989) phrases it, this would be too mind-boggling confusing. Mathematics is ideally suited to not only express these complex relationships in a succinct way, but mathematics also forces one to be exact in his or her statements of a system. As mentioned before, a mathematical model is very explicit about what exactly is assumed about a system and what not. Moreover, once a model has been formulated mathematics oers the appropriate tools to analyse its consequences, again exactly. Mathematics can hence be viewed as a language that is most appropriate for logical reasoning and logical analysis of problems.

Deel II

Unstructured Population Models in Continuous Time

Hoofdstuk 2

Modelling population dynamics

Population ecology is concerned with developing theory and insight about the persistence, structure and dynamics of biological communities. Often these communities are made up by a large number of species. Necessarily, population ecology is therefore dominated by a focus on interspecic interactions such as competition and predation. Studies that focus on developing mathematical models to describe population dynamics hence also often consider more than a single species. From this focus on multispecies communities one could easily get the impression that modeling the dynamics of a single species is not important or uninteresting. Or one could think that there is not much to discover in models of single populations. There are, however, many cases in which relatively simple processes of population growth are of fundamental importance, both from a scientic and an applied point of view. In this respect one can, for example, think about the following cases where the growth of a single population is the predominant process: Exponential population growth of the human population Invasions of exotic or genetically modied organisms in natural environments Epidemics of infectious diseases and strategies for their prevention. In the following sections I will subsequently discuss what it takes to model the dynamics of populations and what can be learned from the analysis of the models derived. The aim is to introduce some basic principles of population modeling and some techniques for model analysis.

2.1

Describing a population and its environment

In this section I will introduce some basic concepts of population dynamic models. The discussion in this section is rather formal as it does not directly relate to a biological system or idea.

2.1.1

The population or p-state

Formulating a model for the dynamics of any population is equivalent to specifying a recipe for population change. Basically, a population dynamic model answers the question how a population is going to change in the (near) future, given (1) its current status and (2) the environmental conditions that the population is exposed to. These changes in the population may be changes in the total number of individuals present, i.e. in the number of population members, but may also pertain to changes in the composition of the population. For example, the population model may be a recipe for how the number of small and large individuals, or juvenile 7

HOOFDSTUK 2. MODELLING POPULATION DYNAMICS

10

10

World population (in billions)

World population (in billions)

1500 2000 1 1900

1000

1950

2000

Year

Year

Figuur 2.1: Growth of the world population over the last century. Data: WHO. and adult individuals is going to change. Alternatively, the population may change in overall abundance, while the relative frequency of old and young individuals stays constant. As an example, consider the changes in the human population over the last 100 years (see Figure 2.1). Not only has there been a lot of interest in the changes in the total number of people in the world, as it grows virtually exponentially to what might be unsustainable abundances, but also has there been a lot of interest in the population age structure or age distribution, leading to expressions like baby boomers or aging population (see Figure 2.2). The two important characteristics of a population for its future developments are (1) the total number of individuals in the population and (2) the composition of the population in terms of old/young, small/large or juvenile/adult individuals. More formally, the rst of these two aspects is often referred to as the population size or population abundance, while the second is referred to as the population structure. Together the population abundance and structure dene the population state or pstate. The p-state is the characterization of the population in terms of how many individuals of which type (e.g. age, size, sex) are present in the population.

2.1.2

The individual or i-state

The individual organism itself is the fundamental entity in the dynamics of the population, since changes in the population can only come about because of events that happen with individual organisms. For example, changes in the population abundance are the direct consequence of birth, death, immigration and emigration of individual organisms, while changes in the population age- or size-distribution are the result of aging or growth in body size of individual organisms. If we want to keep track of the age- or size-distribution of a population it is necessary to distinguish the individual members of the population on the basis of their age or body size, respectively. On the other hand, in many cases the chance that a particular individual will give birth or will die strongly depends on its age or its body size. For example, in humans living

2.1. DESCRIBING A POPULATION AND ITS ENVIRONMENT

Number of individuals (in millions)

1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0

1950

1975

2000

5 year age class

Figuur 2.2: Population age distribution in the Netherlands in 1950, 1975 and 2000. Every bar represents the number of individuals in the Netherlands in a 5 year age class, starting with 0-4 year and 5-9 year old individuals and ending with 90-94 year old individuals and individuals of 95 year and older. Data: CBS, The Netherlands. in a developed country giving birth mainly occurs between the age of 18 and 45 years (for the mother), while death occurs mainly at older ages (see Figure 2.3). For a variety of reasons we therefore often want to distinguish individual organisms from each other on the basis of a number of physiological characteristics, such as age, body size or sex. The individual state or i-state is the collection of physiological traits, that are used to characterize individual organisms within a population and that inuence its life history in terms of its chance to reproduce, die, grow or migrate. The individual state may be any collection of variables that characterize individuals, but its choice is usually kept limited to one or two physiological variables (e.g. age and/or size).

2.1.3

The environmental or E-condition

The environmental conditions that a population is exposed to are important for its dynamics, while it usually sets the limits for its development. Environmental conditions can pertain to biotic and abiotic factors, for example, temperature, humidity, food abundance and the number of predators or competitors around. Since the individual organism is the fundamental entity in population dynamics, it is actually more appropriate to consider the environmental conditions that an individual member of the population is exposed to. For a specic individual organism, the environment is not only made up by the ambient temperature or food abundance that it is exposed to, but also by the number and type of fellow members within the population. This aspect of its environment is, for example, important if the individual reproduces via sexual reproduction. As an another example, if a species is potentially cannibalistic, the number of fellow population members may inuence the risk of dying for a particular young and small

10

HOOFDSTUK 2. MODELLING POPULATION DYNAMICS

1.0

Number of births/deaths (x 10 )

0.8

0.6

0.4

0.2

0.0

5 year age class

Figuur 2.3: Total number of births and death per 5 year age class in the Netherlands in 1999. Every bar represents the number of individuals born (red) or died (blue) in the Netherlands in a 5 year age class, starting with 0-4 year and 5-9 year old individuals and ending with 90-94 year old individuals and individuals of 95 year and older. Data: CBS, The Netherlands. individual. The environmental or E-condition of a particular individual hence comprises the collection of abiotic and biotic factors, external to the individual itself, that inuences its life history. The i-state of a particular individual could hence be viewed as determining, for example, its potential reproduction, while the E-condition will modulate this potential and thus set the realized reproduction. Similar arguments hold for the chances of a particular individual to die or grow.

2.2

Population balance equation

If we consider a population to be the collection of individuals of a particular species that lives within a well-dened area, any changes in the number of individuals within this population comes about by reproduction, death or migration of individual organisms. More formally, we can express a population dynamic model with the following (semi-)equation: Population change = Births Deaths + Immigration Emigration (2.1)

This equation reects the general structure of a population balance equation. The phrase balance equation refers to the fact that changes in population abundance are a balance between

2.3. CHARACTERIZING THE POPULATION

11

processes that decrease this abundance (e.g. death and emigration) and processes that increase the abundance (e.g. reproduction and immigration). Both immigration and emigration of individuals can be neglected if the area in which the population is considered to live is closed o to any movement of individual across its perimeter. Also when the size of the area is very large relative to its circumference, immigration and emigration might be of negligible importance. In those cases, population dynamics is just the balance between the number of birth and death cases of individual organisms. Population for which immigration and emigration can be neglected are usually referred to as closed populations or closed systems, as opposed to open populations or open systems, that are open to migration. Good examples of closed systems are lakes and many of the population living in them, while rivers or tidal zones on the shore are typical examples of open systems.

2.3

Characterizing the population

Before we can make the general population balance equation (2.1) more explicit we will have to decide on how to characterize the population that we want to model. This process is the rst step in building or formulating a model. Model building is a crucial aspect of theoretical population biology, as it forces a modeler to think carefully about the important aspects of the system that he/she wants to describe. Important issues to consider while choosing a particular representation of a population are, for example: Is it necessary to distinguish individual organisms from each other and if yes, what variables are used for this purpose? What individual traits predominantly inuence reproduction and death onf the individuals within the population? Are there biotic or abiotic factors in the environment that strongly inuence the life history of individuals? Answering these questions is not straightforward and is actually an art in itself (the art of modeling). Many people that are new to modeling are tempted to say that a model should include as many variables as possible to describe a particular system. However, ultimately such an approach would just lead to a mathematical copy of reality, which would probably be impossible to investigate and from which we could only learn just as much as we can learn from studying the real-world system we want to model. What then is the aim to develop a model? Therefore, building a population dynamic model forces a modeler to make judicious choices about which aspects of the study system to include into the model and which to neglect. Such choices should be decided upon on the basis a number of dierent considerations: What is the aim of the model to be developed? For example, if a model is developed to study the age-distribution of the human population it is necessary to distinguish individuals of the population on the basis of their age. However, for many other species (think about algae or mosquitoes) characterizing individual organisms with their age is in most cases not very relevant. Even more, it may be unnecessary to distinguish individual organisms from each other at all! What are the important processes that inuence the population dynamics? For example, does the model have to account for immigration or emigration or is it appropriate to consider a closed population?

12

HOOFDSTUK 2. MODELLING POPULATION DYNAMICS Which factors inuence the chance that individuals give birth or die? For example, are there any predators present that individuals might fall prey to? What is technically feasible? This is often a very important modeling consideration. Ideally, a model should be as simple as possible, while still capturing the essentials of the system that is to be modeled (the most parsimonious model). It is, however, obvious that there will always be a trade-o between model simplicity and the amount of detail that a model can incorporate. If the model incorporates too much detail, it becomes impossible to analyze while with too much simplicity the model looses its meaning.

Even though the individual organism is such a fundamental entity in both ecology and evolution, most population dynamic models do not at all distinguish between dierent individuals. This is very much the result of technical limitations: accounting for population structure makes the more explicit specication of the population balance equation (2.1) just so much more dicult and its analysis a daunting task. In general, ecological and evolutionary theory is therefore based on so-called unstructured population models, that is, models that ignore the presence of population structure. Developing theory that does account for dierences among individuals within the same population, i.e. on the basis of structured population models that do account for population structure, is very much the cutting edge of contemporary ecological research. Unstructured population models are hence based on an assumption that all individuals within a population are identical1 . The population state or p-state is in this case simply the same as the total number of individuals within a population. This total number at a specic point in time t will frequently be denoted by N (t). Often the explicit occurrence of the time t as an argument is left out and we simply write N to indicate the total population abundance.

Restrictions: To start with we will ignore any population structure and hence consider the population abundance N as the quantity that species our population state. Only in later discussions we will sometimes take into account population structure. Moreover, we start by considering closed populations and hence neglect immigration and emigration of individuals to and from our population under study.

2.4

Population-level and per capita rates

Consider now the population abundance at time t, indicated by N (t), and its abundance some time later, indicated by N (t+t). t refers here to the, possibly small, interval of time. Making the balance equation (2.1) more explicit, we can now write N (t + t) N (t) = Number of births during t Number of deaths during t (2.2)

Obviously, the change in the population state (which is the same as the population abundance due to our restrictions) is the dierence between N (t + t) and N (t). Because we have assumed
Strictly speaking, it is fair to say that unstructured models assume that all individuals within a population can be represented by some average type.
1

2.4. POPULATION-LEVEL AND PER CAPITA RATES

13

that there is no immigration and emigration, this change should equal the dierence between the number of individuals that have been born during the time interval t and the number that have died during that interval. Note that the quantities in the above equation all relate to numbers of individuals. In the previous section we imposed two restrictions on the class of models that we are going to discuss: (1) we decided to focus on unstructured population models, which ignore population structure, and (2) we decided to focus on closed populations. These restrictions can be viewed as dichotomies between dierent classes of models: the rst restriction sets the class of unstructured population models apart from the class of structured population models, while the second restriction sets the class of models for closed populations apart from those for open populations. Here we encounter another dichotomy, which sets apart the class of discrete-time models from continuous-time models. Many text books on population modeling start by considering population dynamics in discrete time. This is a very useful assumption, for example, when the aim is to model a population of annual plants. If reproduction only occurs once a year (or once a season) we could simply choose to specify the changes in the population state from year to year without specifying how the number of individuals in the population changes within a year. Take as an example a population of annual plants that reproduce by producing seeds, which overwinter and germinate in the next year. We could model this population by observing the number of plants present at the beginning of a growing season, say May 1st of every year. N (t) would in this case be the number of plants in one year and N (t + t) the number in the year after, while t equals exactly 1 year. Specifying the number of deaths during t (= 1 year) is straightforward, as all plants are annual and hence die before the next census time at t + t. The modeling of the dynamics in this case would boil down to specifying how the number of plants in a specic year (at time t + t) is related to the number of plants in the year before (at time t). This involves specifying the relationship between the number of seeds produced by a population of size N (t), the probability that a seed germinates into a seedling and the probability that this seedling grows into a new plant. The essence is that the model only describes what the state of the population is at May 1st of each year. Two dierent species might show the same population dynamics (when censused at May 1st of each year), while the one species owers only in May with all plants dying before the end of June and the other owers all summer and plants die only during winter. In other words, discrete time models only determine the state of the (modeled) population at specic points in time and do not tell what happens inbetween. None of the above arguments is really problematic and discrete-time models have been widely and successfully used to develop many pieces of important ecological theory. There are, however, some subtle but fundamental dierences between the mathematical theory on discrete-time models and continuous-time models. In my opinion, these dierences have often caused quite a bit of confusion with students that were new to modeling population dynamics. Hence, I decided not to treat both discrete-time and continuous-time models next to each other as separate model classes, but instead to discuss one of them at length and in detail. Hence, I will discuss basic theory about population models using continuous-time models, ignoring for the time being the considerable amount of theory on discrete-time models. A continuous-time version of the balance equation (2.2) can be derived by dividing both sides of the equation by t: Number of births Number of deaths N (t + t) N (t) during t during t = t t t By taking the limit t 0, the left-hand side of this equation becomes the derivative of N (t) with respect to time t, while the right-hand side becomes the dierence between the rate with

14

HOOFDSTUK 2. MODELLING POPULATION DYNAMICS

which individuals are born into the population and the rate with which individuals disappear from the population due to death. Hence, the continuous-time, population balance equation can be written as: d N (t) = B(N ) D(N ) (2.3) dt In this balance equation, I have written the birth and death rate B(N ) and D(N ), respectively, as explicit function of the total number of individuals N to indicate that the number of individuals present in the population usually to a very large extent determine the number of births and deaths that do occur during a certain time period. Equation (2.3) is an ordinary dierential equation or ODE for short. Here the model is specied by a single ODE, but they may also occur as systems of ODEs in more complex situations, for example, when we want to model more than a single population. The formulation and analysis of ODEs, both single equations and system of two or more ODEs, is a major focus of the following chapters. The function B(N ) and D(N ) are the population birth rate and population death rate, respectively. These are therefore population-level quantities, because they, for example, refer to the total number of ospring produced by the entire population. Because all individuals in the population are assumed identical any way, the balance equation (2.3) can be rewritten in terms of individual-level birth and death rates, the so-called per capita birth rate and per capita death rate, respectively. By dening the per capita birth rate b(N ) as: b(N ) = and the per capita death rate d(N ) as: d(N ) = the balance equation (2.3) can be rewritten as: d N (t) dt = b(N ) N d(N ) N (2.6) D(N ) N (2.5) B(N ) N (2.4)

The equation above is a general population balance equation for changes in the population abundance occurring in continuous time. It will be the basis of many of the population models in the forthcoming sections. The rates that occur in the population balance equation are not as easy to interpret as the number of individuals born or dying during a particular period of time which occur in the discrete-time balance equation (2.2). Both the population-level birth and death rate B(N ) and (D(N ), respectively, and the per capita birth and death rate b(N ) and d(N ), respectively, have a formal interpretation as a probability per unit time. For the per capita death rate d(N ) this means, for example, that within an timespan t that is innitesimally short, an individual has a probability to die equal to d(N )t. Also, on average the time elapsing between two death events equals 1/d(N ). Every individual of the population thus has an expected lifetime of 1/d(N ) units of time. Similar arguments hold for the other rate functions, such that the average time between two consecutive times at which a particular individual gives birth equals 1/b(N ). The population balance equation (2.6) does not specify a complete population dynamic model yet, as it only determines how the population abundance is going to change over time. We therefore still have to specify from what value it is going to change to start with. In other words, we have to specify an initial state of our population. Usually, this is done by specifying the number of individuals present at some particular point in time, which then simultaneously is chosen to equal the start of our time axis (t = 0): N (0) = N0 (2.7)

2.5. MODEL BUILDING

15

In this initial state equation, the quantity N0 is a known value from which the population dynamics is going to develop.

2.5

Model building

Building a model to describe the dynamics of a particular population or in more complicated cases a collection of populations can be separated into two distinct steps: First, we have to choose the mathematical representation of the population in our model. This issue was addressed in section 2.3, where we also discussed some issues to consider when choosing a particular representation. For example, it was discussed whether to include any more aspects of the population in our model than the total number of individuals (i.e. the abundance). Second, once we have chosen a particular representation for the population it is necessary to write down the population balance equation and to specify the exact form of the rates (either per capita or population-level rates) that occur in it. It should be noted that equation (2.6) is specic for a model in which the population is only characterized by its abundance N . If we choose a slightly more complicated representation of our population the corresponding balance equation will be analogous but slightly dierent. For example, if we would choose to keep track of both juvenile and adult individuals in the population a balance equation for both classes of individuals should be specied. This is still a relatively simple extension of the basic equation (2.6), but things can be much more complicated if we, for example, want to keep track of the entire age distribution of the population. The second step in formulating a model involves specifying the rate functions that occur in the population balance equation. Determining their actual form as dependent on the population abundance N is again where modeling becomes an art instead of a scientic procedure. As was the case for choosing the appropriate population representation there is no simple recipe how to do it and the process of model building can hence only be discussed by example. I will introduce here 3 examples which will be used in the next chapter to illustrate the steps in model analysis.

2.5.1

Exponential population growth

Malthus (1798) investigated the birth and death register of his parish and concluded that the population of his parish doubled every 30 years. He considered the following population balance equation: d N (t) = N N (2.8) dt in which and are now the specic form chosen for the per capita birth and death rate b(N ) and d(N ). Implicitly, Malthus (1798) hence made two very specic assumptions about the per capita birth and death rate b(N ) and d(N ) in that that they were constant and hence independent of N (It is perhaps fairer to say that he did not consider at all that b(N ) and d(N ) could depend on N , but the end eect is the same). Because in this case b(N ) = and d(N ) = do not depend on the population density N , this model is density independent. In a density independent model the per capita rates are not inuenced by any aspect of the population at all, neither direct nor indirect. An indirect population eect on the per capita rates might occur if the population feeds on a food resource, depleting it to low levels, and the food density itself subsequently aects the per capita birth and death rate. As an aside, note that when the per

16

HOOFDSTUK 2. MODELLING POPULATION DYNAMICS

capita birth or death rate is density independent, the population-level birth and death rates are linear in (i.e. proportional to) N . The population balance equation (2.8) used by Malthus (1798) can be written more succinctly as: d N (t) = rN (2.9) dt where r = (2.10)

Equation (2.9) is the famous exponential growth equation or Malthus growth law. The quantity r is known as the population growth rate or the Malthusian parameter . Equation (2.9) can easily be solved once we have also specied the initial state of the population (i.e. the population abundance N0 at time t = 0; see eq. (2.7)). As an explicit function of time t the solution is: N (t) = N0 ert (2.11) On the basis of Malthus observation that the population of his parish doubled every 30 years we can now estimate the parameter r to equal r = 0.0231 year1 (try to derive this estimate!). In other words, the population of Malthus parish was estimated to grow with 2.31% a year, which is not too far o the current growth rate of the human population (Fig. 2.1)!

2.5.2

Logistic population growth

The story goes that the prediction by Malthus (1798) led to some concern, because it implied that the human population would quickly exhaust its natural resources. Verhulst (1838) claimed that the model proposed by Malthus (1798) was too simplistic as it only included linear terms. Verhulst (1838) hence wrote as an alternative model: d N (t) = a N b N2 dt including an additional quadratic term with a negative coecient. There are a number of dierent ways in which this particular form of the population balance equation might come about. Let me suggest here one scenario that leads to the form proposed by Verhulst (1838) by considering a population in which the per capita birth rate decreases with population abundance. The simplest way to achieve such a density dependence is by assuming that the parameter that was already introduced in Malthus equation (2.8) is actually the per capita birth rate at very low (actually innitessimaly small) population abundances. Moreover, with increasing values of the population abundance N it is assumed to decrease linearly with N to reach a value of 0 at some arbitrary population density N = . If I in addition assume that the per capita death rate is again density independent, these assumptions lead to the following balance equation: N d N (t) = N 1 N (2.13) dt With a little bit of algebraic manipulation this equation can be rewritten in a far more familiar form, the logistic growth equation: d N (t) dt = rN 1 N K (2.14) (2.12)

2.5. MODEL BUILDING

17

in which the parameters r and K, representing the population growth rate and its carrying capacity, respectively, are related to the parameters , and from equation (2.13) by: r = K = (2.15) (2.16)

The logistic growth equation (2.14) must be familiar to any student that has followed an introductory ecology course. However, it can be conveniently used to discuss some techniques to analyze population dynamic models in terms of ODEs. The logistic growth equation (2.14) can not be taken seriously as a model that quantitatively describes the dynamics of any real-life population. There are examples of isolated laboratory populations (usually bacteria), whose growth in time, i.e. the function N (t), can be tted quite well by the solution of the logistic equation, but a number of alternative equations would t such data with similar goodness-of-t. In other words, the measurement noise in population data often is far too big to unambiguously decide that the growth of the population is obeying a particular model. The real utility of the logistic equation is, rather, as an embodiment of a qualitative behavior which is not so uncommon (except for humans): a population that starts out with a small number of individuals will ultimately grow to a maximum density, its carrying capacity, which is set by environmental conditions.

2.5.3

Two-sexes population growth

Here I also formulate a model for a population that reproduces by means of sexual reproduction. Unlike the exponential and logistic growth model, this one is not an established population dynamic model that has been widely used in the population biological literature. Rather, it is introduced here mainly to illustrate some techniques for model analysis in the next chapter. The assumptions on which this model is based and its mathematical form are also much more debatable than the the previous two models introduced. To model a population with sexual reproduction it is necessary to derive a mathematical representation for the process of two individuals encountering each other, mating and producing one or more ospring. Especially how to describe the rate at which encounters between sexual partners take place is not an easy or straightforward task. I follow an approach that has a long history in the modeling of chemical reactions: when two compounds A and B in a dilute gas or a solution react to yield a product C, the rate at which this chemical reaction takes place is assumed to be proportional to the densities of both reactants A and B. Hence, for the chemical reaction A + B C the rate with which the product C is produced is proportional to [A] [B] where [ ] refers to the concentration of a substance in the medium. The assumption that the rate of product formation is proportional to the product of the reactant concentrations has become known as the law of mass action or mass action law . Two requirements for it to hold is that the reactants move about randomly and are uniformly distributed through space. Analogously to this approach from chemical reaction kinetics, I assume that the rate at which sexual partners encounter each other is proportional to the product of their abundance. If I

18

HOOFDSTUK 2. MODELLING POPULATION DYNAMICS

furthermore assume that the sex ratio in the population is constant, the rate of encounter is proportional to the squared abundance N 2 . On encounter I assume that the partners produce ospring at a rate that decreases linearly with the population abundance N , as in the logistic growth equation. Hence, population reproduction can be described by a term N2 1 N

in which the term (1 N/) represents the density dependent reduction in ospring production. To describe the death of individuals I simply assume that the rate of mortality equals , as has been similarly assumed in the exponential and logistic growth equation. The rate at which individuals disappear from the population through death hence equals N . Putting these model pieces together, the dynamics of a population reproducing by means of sexual reproduction could therefore be described by the following ODE: d N (t) dt = N2 1 N N (2.17)

2.6

Parameters and state variables

Above I already used the phrases parameter and state variables, without really discussing their distinction. It is very important to distinguish what is a parameter and what is a state variable in a model. State variables are those quantities (1) that characterize the state of the system, in our case the biological population, and (2) that change over time. It is the change in these quantities that the model determines. Parameters, on the other hand, are also characteristic for a specic system (i.e. population), but parameters do not change over time. They hence occur in the population dynamic models above, but the models do not specify that they change over time. They can be viewed as a kind of innate properties of the system. In all the examples above, there was only a single state variable, i.e. the population abundance indicated with N . On the other hand, the quantities r (in both the exponential and logistic growth equation), K (in the logistic growth equation), (in the logistic and two-sexes population growth model) and and (in all three models) are all parameters. The parameters r and K can furthermore be identied as compound parameters, because they are in reality a combination of the low-level parameters , and (see equation (2.10), (2.15) and (2.16)).

Important: In the following chapters I will adopt the convention, which is common to most if not all population dynamic models, that model parameters are introduced in such a way that they can only meaningfully take on positive values. Since I have earlier on imposed the restriction that the populations considered are only represented by their abundance, the same positivity condition will hold for the state variables in the models as well.

2.7. DETERMINISTIC AND STOCHASTIC MODELS

19

2.7

Deterministic and stochastic models

In the following chapters I will mainly formulate and analyze deterministic models, as opposed to stochastic models. Given an initial state of the population, deterministic models specify a unique dynamic path of the system. Such a dynamic path is called a trajectory. Hence, a deterministic model attaches to every single initial condition, a single and unique trajectory. On the other hand, in stochastic models the initial state of the population determines an entire family of trajectories and every one of these may occur with a given probability. A simple example may illustrate the distinction between these two model frameworks. Imagine the fate of the last two individuals of an endangered species in a particular habitat. Each of these two individuals might have a chance to die, say, 0.1% per day. For simplicity I will assume that the two individuals are of the same sex and can hence not reproduce. How long these two individuals will survive, when the rst and last one will die, is the outcome of a chance process. In other words, given the starting condition (the last two individuals) there are many dierent possible outcomes and which one will actually happen is not a priori determined. The only thing possible to do is to calculate the probability that one or the other scenario (or trajectory) will occur. Now imagine another population of individuals with a probability of 0.1% per day to die. However, this population consists of a very large number of individuals, say 1010 . Given this large number of individuals we can expect that approximately 107 of these individuals will die every day. Because it is the outcome of a large number of independent chance processes (for every individual one) this number will also be relatively constant (This is due to the fact that the variance in the mean outcome of a large number of independent trials decreases with the number of trials). We can hence rather faithfully represent the death process in this population by a deterministic rate of 0.1% of the total population per day. Most, if not all, population dynamic processes are stochastic: individuals usually have a chance to die or give birth and hardly ever have a predestined time of reproduction or death. Nonetheless, the example above indicates that such stochastic dynamics can in principle be faithfully described by a deterministic model as long as the number of individuals is large. Hence, a deterministic model can be viewed as the limit of a particular stochastic model for when the number of individuals in the population becomes large, a fact which is usually referred to as the law of large numbers. I will focus on deterministic population models, because there exist tools and techniques to analyze there behavior. The analysis of stochastic population models is technically much more demanding and the only thing one can usually do is just simulate there dynamics. In section 3.2 I will argue that such simulations (often also referred to as Mont Carlo simulations) have sometimes very limited power to elucidate the dynamic possibilities that a model allows. Deterministic models allow a much more rigorous and detailed investigation of the model potential.

20

HOOFDSTUK 2. MODELLING POPULATION DYNAMICS

Hoofdstuk 3

Single ordinary dierential equations

In the previous chapter we have presented three dierent models for the growth of a single population, that can generally be written in the form: d N (t) = f (N ) dt The function f (N ) diered among the three dierent models introduced Exponential growth equation: Logistic growth equation: f (N ) = r N f (N ) = r N f (N ) = N 2 1 N K N N (3.2) (3.3) (3.1)

Two-sexes growth equation:

(3.4)

In this chapter we will discuss basic techniques that can be used to analyze the predictions of these models. Although the models itself are overly simply and in any case much simpler than the population dynamic models that are used in scientic studies, the basic techniques can conveniently be introduced using these simple growth equations. It should be realized that the process of model analysis as such is a purely mathematical procedure. Hence, it is possible to give a more or less complete recipe for analyzing the equations that determine a population dynamic model. However, the analysis of the equations can never be the nal result of a population dynamical study (unless you are a mathematician and just interested in the equations). Once the analysis of the equations is complete, the results of that analysis have to be translated into biological conclusions about the system, that the model was developed for. Even though many beginning students may think the mathematics itself a stumbling block, the step preceding and following the mathematical analysis, i.e building the model and interpreting the mathematical results into biological conclusions, turn out to be usually far more dicult and time consuming and in general requires far deeper thinking than the mathematical analysis does! As for model building, there is also no clear-cut recipe how to interpret the mathematical results in biological terms. Again, this can only be discussed by example.

3.1

Explicit solutions

Ideally, the analysis of a mathematical model should simply consist of specifying its explicit solution. This means that given the ODE for N (t) and the density of individuals at time t = 0 21

22

HOOFDSTUK 3. SINGLE ORDINARY DIFFERENTIAL EQUATIONS

(N (0)), we should write down the explicit expression for N (t). An explicit solution would allow a very complete analysis of the properties of the model, including numerical studies showing the development of the population over time. However, an explicit solution is seldom possible unless the right-hand side of the ODE (i.e. the function f (N ) in the ODE (3.1)) is linear in N . In the previous chapter we have discussed that a linear form of f (N ) implies that the per capita birth and death rates are density independent, i.e. independent of N itself. For the exponential growth equation (3.2) it is hence possible to write down the explicit solution: N (t) = N0 ert (3.5)

where N0 is the population abundance at t = 0 and the parameter r is the population growth rate. This explicit solution can be obtained by re-writing the ODE to separate the two variables occurring in the ODE, i.e. time t and population abundance N , respectively. I leave a complete treatment of this integration to the reader as an exercise. It must be said that also the logistic growth equation (3.3) allows for an explicit solution in terms of the population abundance N as a function of time t. This solution can be obtained by rst re-writing the ODE as: K dN = r dt . N (K N ) Subsequently, the left-hand side of this equation can be separated in a term with denominator N and a term with denominator K N : 1 1 + N K N
t

dN = r dt .

The left- and right-hand side of this last equation can be integrated to yield:
t

ln(N )
0

ln(K N )

= rt
0

From this the following explicit solution to the logistic growth equation (3.3) can be obtained: N (t) = N0 K N0 + (K N0 ) ert (3.6)

In the explicit solution the initial condition has already been used to substitute N (0) by the value N0 . The logistic growth equation: d N (t) = rN dt 1 N K (3.7)

is an example of a non-linear dierential equation, while the exponential growth equation: d N (t) = rN dt (3.8)

is a linear dierential equation with constant coecients (The phrase constant coecient refers to the fact that the only parameter in the equation r does not depend on time t). The distinction between linear and non-linear ODEs is a very important one, as we have seen that linear ODEs with constant coecients can be solved explicitly. This even holds for systems of coupled, linear ODEs, which may occur in models with multiple species that interact with each other (see later chapters). Non-linear ODEs, on the other hand, can only be solved in exceptional cases, such as in case of the logistic growth equation. Coupled systems of non-linear ODEs are virtually never possible to solve explicitly.

3.2. NUMERICAL INTEGRATION

23

What then is exactly a linear dierential equation? The denition of a linear ODEs requires the right-hand side f (N ) to have the following properties: f (N1 + N2 ) = f (N1 ) + f (N2 ) f (aN1 ) = a f (N1 ) (3.9) (3.10)

In words, if the right-hand side function is applied to the sum of two quantities, the results should be the same as the sum of the function applied to the two quantities separately. Moreover, applying the right-hand side function to a value of N that is twice as large, should be identical to twice the result of applying the function to N itself. It should be noted that these two requirements decide upon the linearity of both single and systems of ODEs. In the latter case, the quantity N represents a vector and f (N ) is a vector-valued function.

3.2

Numerical integration

In the previous section it was explained that only a limited number of, mostly linear, ODEs allow an explicit solution as a function of time t. Moreover, even if an explicit solution can be given, how do we proceed to gain insight into the properties of the dierential equations? First of all, what are the properties of an ODE? Of course, modern computer facilities make it very easy to just draw an explicit solution N (t) as a function of time t in a simple graph. A requirement to do this is that all parameters in the model, i.e. the growth rate r and the carrying capacity K in the logistic growth equation, and the initial state N (0) are given specic numerical values. Even more complex models, consisting of multiple species that interact or a single species with an age structure taken into account, can be studied in this way: once all parameters and the initial condition of a model are given explicit values, the solution as a function of time t can usually be obtained in a rather straightforward manner. If the explicit solution of the ODE is not available, as is the case for most systems of ODEs, numerical integration methods are readily available to generate the numerical solution of a system of ODEs (see, for example, Press et al. 1988, for a selection of numerical methods for ODEs). This procedure of obtaining a numerical solution for a system of ODEs for which both the parameters and the initial state are given explicit numerical values, is referred to as numerical integration or sometimes numerical simulation (the rst term is actually more appropriate, while numerical simulation often implies that the model is stochastic). Methods for numerical integration come in a large variety, which I will not further discuss here. Numerical integration of a model is a very good way to quickly gain some supercial insight into the behavior of the model. By just repetitively choosing parameter values and initial conditions, it can show whether the populations will persist for the chosen values or not, whether they will approach constant values or whether they uctuate indenitely over time. However, a numerical integration only gives insight about one particular set of parameters and one particular set of initial conditions. In later chapters we will encounter models that show drastically dierent behavior for two slightly dierent values of a single parameter. Lets for the moment assume that you are investigating a model that can indeed show these drastically dierent types of behavior for parameter values that are only slightly dierent. What should you do? One of the rst reactions that students new to modeling show is to ask what the appropriate parameter value is that holds for the natural situation. This is like asking: Does the population growth rate in the logistic equation a value of r = 0.3 or r = 0.29? There are at least two good reasons why it is not very useful to ask what the real value of r for the ecological system you are studying:

24

HOOFDSTUK 3. SINGLE ORDINARY DIFFERENTIAL EQUATIONS First of all, parameter values are always hard to extract from experimental data. Given the noise that is usually present in measurements on a system, it is, for example, undoable to decide whether r = 0.3 is more appropriate than r = 0.29 in an exponential growth equation. Second of all, I have argued in chapter 1 that a model is a conceptualization of a system. As such it is an abstraction and hence parameters are not necessarily objects that can be unambiguously identied in the system, as they might only exist in our conceptualization of it. This is a rather dicult issue to explain when encountered for the rst time, but in my opinion it is fair to say that the existence of such a thing as a carrying capacity K of a natural population is debatable.

Basically, these two reasons state that there is a lot of uncertainty, not just in the value of model parameters, but even in the conceptualization of the system: the same ecological system might be equally well represented by a conceptualization that incorporates slightly dierent mathematical functions. Because of these uncertainties it is often much more useful to try and gain an understanding of the qualitative behavior of the model. This means that one tries to answer the questions: What dierent types or classes of dynamics can the model exhibit? In other words, does the model predict uctuating population abundances for certain parameter values? What are approximately the ranges of parameter values and initial conditions for which the dierent types of dynamics occur in the model? At which parameter values do transitions between the dierent types of dynamics occur? I will loosely refer to these investigations into the qualitative behavior of the population model as a bifurcation analysis of the model, even though the exact meaning of bifurcation analysis is arguably slightly dierent. Also, why it is called bifurcation analysis will only become clear in later chapters.

3.3

Analyzing ow patterns

A rather straightforward and intuitive way of analyzing the qualitative properties of a population dynamic model is by means of graphical methods. Graphical analysis is especially useful for simple models formulated in a single or at most two dierential equations. The aim is to evaluate the right-hand side f (N ) of the ODE (3.1) and draw in gures for which values of N the population abundance will increase and for which it will decrease. Because it is aimed at resolving in which direction N will change given its current value, the approach is also referred to as the analysis of ow patterns. Imagine now the exponential growth equation (3.2). Depending on the values of the birth and death rates and , respectively, the right-hand side f (N ) is either a linearly increasing, strictly positive or a linearly decreasing, strictly negative function of N (or zero for all values of N when r = 0, but this case is too exceptional to discuss). For positive values of f (N ) the population abundance will increase, while for negative values the abundance will decrease, as is shown in the following sketch:

3.3. ANALYZING FLOW PATTERNS

f (N ) f (N )

25

Now consider the logistic equation (3.3). A graph of the the right-hand side f (N ) as a function of N is given below:
f (N )

Obviously, the right-hand side f (N ) equals 0 for both N = 0 and N = K. Hence, for these two values of the population abundance the model predicts no change. If the initial population abundance would be either N (0) = 0 or N (0) = K the abundance would remain 0 and K for all times. Such a value for the population abundance is referred to as a steady state or equilibrium. More generally, a steady state is such a state that, if started from it initially, the population would keep for all times. However, even if the population would remain in a steady state once started in that state, it does not imply that the population would also approach that state if the initial population state is only close to the steady state. Or if the population would be displaced from the steady state only a little bit, it will not necessarily return to a steady state. The graph of the logistic growth equation shown above, indicates that if the population abundance would be displaced away from N = 0 to a small positive value, the population abundance would actually increase even further, because the value of f (N ) is positive for 0 < N < K. On the other hand, if the population abundance would be displaced from the steady state N = K to a slightly smaller

26

HOOFDSTUK 3. SINGLE ORDINARY DIFFERENTIAL EQUATIONS

population abundance, the population growth rate would change from f (N ) = 0 to a positive value. If the population abundance would be displaced from the steady state N = K to a slightly larger population abundance, the population growth rate would change from f (N ) = 0 to a negative value. Hence, if the population abundance is perturbed away from the steady state N = K in whatever direction (positive or negative) it would approach the steady state anew. The steady state N = K is hence referred to as a stable equilibrium. The steady state N = 0, on the other hand, is unstable, as the population abundance would grow away from it when perturbed. Clearly, the analysis of the ow of the population abundance (i.e. the arrows in the graphs above) in the neighborhood of a steady state has shown us the general behavior of the logistic growth equation: when starting from a non-zero positive population abundance, the model predicts that the population will grow asymptotically to the stable equilibrium N = K (Note that mathematically the population abundance will never become equal to K, but will only come innitesimally close to it). Even if the population abundance is perturbed away from this equilibrium, it will ultimately return to it. Finally, lets consider the two-sexes population growth model. First we have to try to sketch a graph of the function f (N ) (see equation (3.4)) as a function of its argument N . How to do this is the subject of basic function analysis and will not be discussed here in detail. The points to note about the function f (N ) are: the growth rate f (N ) equals 0 for N = 0 (as should be the case for any dynamic model of a closed population), for non-zero, but very small, positive values of N f (N ) decreases to negative values, for very large values of N the value of f (N ) becomes very large, but negative, and as long as > 4

the function f (N ) has two additional roots for which f (N ) = 0. These roots will be indicated with the symbols N and N , respectively. Altogether, this leads to the following qualitative graph of f (N ) as a function of N :
f (N )

From this graph we can infer that the two-sexes population model has 3 steady states: N = 0, N = N and N = N , respectively. As before, if the population would initially have an abundance

3.3. ANALYZING FLOW PATTERNS

27

equal to one of these three values, it would keep the particular abundance for all times. However, the graph above also shows that not all of these 3 steady states are stable. In contrast to the logistic growth equation, if the population abundance is perturbed away from N = 0 to a small positive abundance, its growth rate will be negative. Hence, it will decrease and approach N = 0 again. The steady state N = 0 is therefore stable. Similarly, if the population abundance is perturbed away from N = N to a slightly larger value, it will also have a negative growth rate, hence decrease and approach N = N anew. If perturbed away from N = N to a slightly smaller value, the population growth rate will be positive and N = N is approached as well. The steady state N = N is therefore also stable. The steady state N = N is an entirely dierent matter: if perturbed away from this abundance to a slightly smaller value of N , the population growth rate will be negative and the abundance will decrease and eventually approach the steady state N = 0. On the other hand, if perturbed away from this abundance to a slightly larger value, the population growth rate will be positive and the abundance will increase to eventually approach the steady state N = N . The steady state N = N hence is an unstable steady state, because whenever the population abundance is perturbed away from this steady state, it will only move farther and farther away. In addition, it has the character of a breakpoint: it separates values of N that would eventually lead to approaching the stable steady state N = 0 from those values of N that would eventually lead to approaching the stable steady state N = N . An unstable steady state like the one N = N encountered here is called a saddle point. From the graphs of the logistic and the two-sexes growth equations shown above, we can also deduce that at stable steady states the curve of f (N ) as a function of N has a negative slope, while at unstable steady states (i.e. saddle points) the slope of f (N ) is positive. Two more points can be deduced from the graphs presented here: Every stable steady state has a basin of attraction. This is the technical term for those initial states of a population dynamic model that would eventually lead to the particular steady state considered. Hence, the basin of attraction of the steady state N = 0 are all those population abundances for which 0 < N < N. while the basin of attraction of the steady state N = N are all those population abundances for which N > N. The boundary between the basins of attraction of two dierent, stable steady states is referred to as a separatrix . When, as in the cases considered here, the model is only onedimensional (meaning it consists of only a single ODE) a separatrix is only a single point, in the two-sexes population model considered here, the saddle point N = N . If the model is, however, of higher dimensions (i.e.described by more ODEs) separatrices can be lines, (curved) planes or even more complicated geometrical objects. The stability of a stable steady state may only be a relative concept: if the population abundance would initially be in the stable steady state N = N , but would be drastically perturbed to a value below N = N , the population abundance would approach the other stable steady state N = 0. Hence, that a steady state is stable against perturbations only holds for small perturbations that do not bring the abundance outside the basin of attraction of the steady state. Mathematically, we therefore consider only the local stability of a steady state. It would hence be more appropriate to refer to a steady state as locally stable. Nonetheless, the adjective locally is often dropped in the ecological

28

HOOFDSTUK 3. SINGLE ORDINARY DIFFERENTIAL EQUATIONS literature. This has led to quite some confusion and dierent interpretations of the phrase stability. When discussing steady states, I will always use the adjective stable in the sense of locally stable, unless explicitly indicated otherwise. For example, for the logistic growth equation the population would always approach the steady state N = K, whatever the initial population abundance is. This steady state is not only locally stable but also globally stable.

A remarkable feature of the two-sexes population growth model is that it can not grow away from N = 0, which means that for small, but positive abundances N the population will actually decline, even though for larger abundances (N > N ) the population will eventually grow to a positive steady state. This feature is referred to as an Allee eect (Allee 1931). In this particular model, the biological explanation for its occurrence is that at low population abundances, individuals do not stand a chance to nd a partner. Hence, reproduction is very much reduced. More generally, the phrase Allee eect is used to indicate a situation in which individuals at very low density are actually performing worse than at slightly higher densities. Summarizing, in this section we have learned the following qualitative theory about population models of the type (3.1): 1. At those values of N for which f (N ) is positive, the dynamics of the ODE (3.1) will cause the population abundance N to increase. 2. At those values of N for which f (N ) is negative, the dynamics of the ODE (3.1) will cause the population abundance N to decrease. 3. If for some value of Ne the value of f (Ne ) equals 0, then this population abundance Ne is called a steady state of the ODE (3.1). If initially the population abundance equals this value Ne , it will remain at this value (because dN/dt = f (Ne ) = 0, meaning there is no change), until the population abundance is displaced away from Ne . 4. If the slope of the curve f (N ) at the value of Ne is negative: df dN < 0
Ne

the equilibrium Ne is stable. After a small displacement away from Ne the population abundance will return to it. The value Ne is hence called a (locally) stable equilibrium of the ODE (3.1). 5. If the slope of the curve f (N ) at the value of Ne is positive: df dN > 0
Ne

the equilibrium Ne is unstable. After a small displacement away from Ne the population abundance will grow or decline further away from it. The value Ne is hence called an unstable equilibrium of the ODE (3.1).

3.4

Computation of steady states and analysis of their stability

In this section we discuss what is usually the core part of the analysis of a population dynamic model. At the same time it is also mathematically the most demanding part of the analysis.

3.4. STEADY STATES AND THEIR STABILITY

29

The key problem is that the graphical analysis discussed in section 3.3 is very useful for models consisting of one or two ODES, but is hardly usable for models of higher dimensions. In general, in more complicated situations the graphical method of analysis is very dicult, if not impossible, to apply. The techniques discussed in this section constitute a more rigorous and robust method of analysis with the same aim as the graphical method of analysis: determining the qualitative behavior of the population dynamic model. Basically, the graphical analysis discussed in section 3.3 resulted in: the identication of steady states, i.e. states that a population would keep for all times, if initially started in it, and the determination of the local stability of these steady states, i.e. analyzing whether the population abundance would return to the steady state, if displaced away from it by a (innitesimally) small, but otherwise arbitrary amount. These same results can be obtained by means of a mathematical analysis, that furthermore generalizes to more complicated situations (unlike the graphical analysis). The identication of steady states is relatively straightforward. It boils down to determining all values of N , for which: f (N ) = 0 (3.11) Implicitly, we already used this property of steady states to draw the graphs in section 3.3 for the logistic and the two-sexes population growth equations. Step 1 in a mathematical analysis of a population dynamic model is to determine its steady states by guring out for which values of the population abundance(s) N the right-hand side of the ODE, f (N ), vanishes. Doing some basic analysis on the function f (N ) in equation (3.4) for the two-sexes population model, leads to the following 3 steady state values for the population abundance N : N= 0 N= N = 1 2 1 2 1 14 (3.12a) (3.12b)

N= N =

1 +

14

(3.12c)

Modern computer software for symbolic calculations, like Mathematica or Maple, allows for a rather rapid and easy derivation of such steady state values from the model equations. More challenging is the question, how to determine the stability of a steady state in a mathematically rigorous way that can also be applied in more complicated cases, for example, when the dynamics is described by 3 or even more ODEs. To analyze the stability of the steady states presented in equations (3.12), we make use of two crucial mathematical properties: 1. Linear ODEs and even systems of linear ODEs can be solved explicitly, as was discussed in section 3.1. 2. Every mathematical function f (N ) can be approximated in a small neighborhood around a particular value N = N by its Taylor expansion: f (N ) = f (N ) + f (N )N + 1 2 f (N )N + O 3 N 2 (3.13)

30

HOOFDSTUK 3. SINGLE ORDINARY DIFFERENTIAL EQUATIONS in which I have used N as a shorthand notation for: N := N N An easy approximation of the function f (N ) is obtained by dropping all terms that incorporate the quantity N with a power of 2 or higher: f (N ) f (N ) + f (N )N or written without the shorthand notation N : f (N ) f (N ) + f (N ) (N N ) (3.15) (3.14)

The right-hand side of this equation is referred to as the rst-order Taylor approximation of the function f (N ) in the neighborhood of N = N . Note that the rst-order Taylor approximation to the function f (N ) in the neighborhood of N = N is a linear function of the population abundance N , since both f (N ) and f (N ) have constant values. In the section 3.3 the stability of a steady state was essentially determined by investigating whether the population state would return to the steady state after being displaced away from this steady state by an innitesimally small amount. The displacement had to be very small to avoid ending up outside the basin of attraction of the steady state. Lets dene the small displacement away from the steady state by N (t): N (t) := N (t) Neq (3.16)

Here I have used the notation Neq to indicate some arbitrary steady state of the population dynamic model, the stability of which I want to investigate. The ODE (3.1) that represents our population dynamic model not only species the dynamics of the population abundance N (t), but because N (t) is dened in terms of N (t) and a constant value Neq , it also species the dynamics of this small displacement N (t). Mathematically, this is expressed by the following: d (N (t) Neq ) dN (t) dNeq dN (t) dN (t) = = = dt dt dt dt dt (3.17)

Hence, the right-hand side of the ODE describing the dynamics of N (t) is the same as the right-hand side of the ODE for N (t): dN (t) = f (N ) dt Using the relation between N (t) and N (t) (3.16), this ODE can be rewritten as: dN (t) = f (Neq + N ) dt (3.19) (3.18)

This last ODE still exactly describes the dynamics of the small displacement N (t) away from the equilibrium Neq . Moreover, as the function f (N ) will generally be a non-linear function that is impossible to solve explicitly, also this exact ODE for the dynamics of N (t) can not be solved explicitly. By rewriting the original ODE of the population model in terms of the dynamics of the small displacement N (t) I have ended up with an ODE that is just as complicated and essentially I have not gained anything. However, using the fact that N (t) is small I can gain a lot of analytic power, because it allows me to approximate the ODE (3.19) with an ODE in which I substitute the right-hand side f (Neq + N ) by its rst-order Taylor expansion around Neq : dN (t) f (Neq ) + f (Neq )N (3.20) dt

3.4. STEADY STATES AND THEIR STABILITY Because Neq is an equilibrium, f (Neq ) equals 0. Hence, the ODE above simplies to: dN (t) = f (Neq )N dt

31

(3.21)

This nal ODE is a linear one, which we have learned to solve explicitly in section 3.1. Hence, the dynamics of N (t) is given by: N (t) = N (0) exp f (Neq ) t (3.22)

in which N (0) is the initial displacement away from the steady state Neq (compare the derivation of equation (3.5)). From this solution we can infer that the steady state is stable if the derivative of the function f (N ) at the steady state value N = Neq is negative, while the steady state is unstable if this derivative is positive (When the derivative exactly equals 0 the steady state is at the edge of stability and instability and hence its stability is undetermined. This exceptional situation will not be discussed further here). It should be noted that the relationship between the sign of the derivative of the function f (N ) at the steady state value N = Neq and the stability of the steady state, was already discovered in section 3.3. There it was found that a steady state was stable if the slope of the curve f (N ) as a function of N was negative, while the steady state was unstable if this slope was positive. Essentially, by deriving the linear ODE (3.21) for the small displacement N (t) we have approximated the curve of f (N ) as a function of N by a straight line through the point N = Neq , i.e. the tangent line in this point. The derivative of the function f (N ) at N = Neq is exactly the slope of this tangent line and hence the slope of the curve f (N ) as a function of N . The approximation of the term f (Neq + N ) by its rst-order Taylor expansion equal to f (Neq )N is therefore based on the assumption that in a very small neighborhood of the steady state N = Neq we can approximate the curve f (N ) by its tangent line in N = Neq , ignoring any higher order curvature. The above process of deriving a linear ODE for the dynamics of a small displacement N (t) in the neighborhood of a steady state N = Neq is referred to as local linearization of the dynamics, as the full model dynamics is locally represented by a linear type of dynamics. In mathematical literature the analysis is also referred to as linear stability analysis to indicate that the stability of a steady state is determined by a linear analysis. The proof that the linear analysis explained here indeed determines the stability of a steady state and the conditions that have to hold for it to apply or to fail (which indeed occurs!) will not be discussed here, as they are mathematically too complex (for details see, for example, Kuznetsov 1995). In later chapters, when analyzing models that are formulated in terms of more than a single ODE, a similar procedure will be followed to investigate the stability of steady states: the system of ODEs describing the model dynamics will be linearized in the neighborhood of a particular steady state to determine the fate of a small but arbitrary displacement (perturbation) away from the steady state. However, in contrast to the procedure described above leading to equation (3.22), the full solution of the linearized system of ODEs will not be derived or written down. Instead, based on the insight that linear ODEs yield solutions of exponential form, a trial solution of the form Ce t (3.23) will be substituted into the linearized system of ODEs (C is here some arbitrary constant). This will allow us to derive an expression of the exponential growth rate(s) that are characteristic for the linearized dynamics in the neighborhood of the steady state. Again, it will turn out that these growth rates should all be negative for a steady state to be stable. I will illustrate this procedure for the linearized ODE (3.21) even though its explicit solution has already been given. Substituting the trial solution (3.23) for N (t) into the linearized ODE

32 yields:

HOOFDSTUK 3. SINGLE ORDINARY DIFFERENTIAL EQUATIONS

dCe t = f (Neq )Ce t . dt The time derivative in the left-hand side of this ODE can be simplied to yield: Ce t = f (Neq )Ce t .

(3.24)

(3.25)

After dividing both sides of the resulting equation by C exp(t), the following equation is obtained: = f (Neq ) (3.26) This equation is referred to as the characteristic equation, as it species the characteristic growth rate of the linearized dynamics. As was already concluded, if is positive the steady state is unstable, while it is stable if is negative. The quantity , which you can interpret as a characteristic growth rate, is called the eigenvalue of the linearized dynamics. The entire discussion of characteristic equation and eigenvalue in the context of the single ODE models that are the topic of this chapter is a little bit overdone. However, it illustrates nicely the approach taken with more complicated models, formulated in terms of systems of ODEs. In those cases, the characteristic equation is often a more complicated, matrix equation which will only implicitly determine the eigenvalues. Surely, it will in general not be possible to specify the eigenvalues as explicitly as it is done in equation (3.26). Nonetheless, the idea is the same: a characteristic equation is derived for the linearized dynamics in the neighborhood of a steady state. This characteristic equation determines, usually in a rather dicult way, the eigenvalues (or characteristic growth rates) , which all have to be negative for the particular steady state to be stable.

Important:
It should be noted that the eigenvalues pertain to a particular steady state of a particular model. Hence, if a model has multiple steady states, as we encountered before, to each of these steady states belongs a unique set of eigenvalues.

3.5

Units of measurement and non-dimensionalization

The analysis of the physical dimensions in which the model equations are expressed is a powerful tool for checking their validity. There are some simple rules that the equations should conform to: If an ODE is written down both the right-hand side and the left-hand side of the ODE should carry the same physical dimension. Since the left-hand side (dN/dt) usually carries the dimension of number per unit of time or density per unit of time, the right-hand side of the ODE (i.e.the function f (N )) should carry this dimension as well. In general the right-hand side f (N ) is a collection of various terms that are added or subtracted. When adding or subtracting terms, these should have the same dimension. This rule allows us to deduce, for example, the dimension of the parameter K in the

3.5. UNITS AND NON-DIMENSIONALIZATION

33

logistic growth equation (3.3). Since the ration N/K in this equation is subtracted from the dimensionless constant 1, both N and K have to have the same dimension. The term (1 N/K) as a whole is therefore dimensionless. Because the entire right-hand side should have the dimension of number per unit of time, the dimension of the parameter r is hence per unit time. Notice that such an analysis shows that the birth parameter in the logistic equation (2.13) has a dierent dimension than the same parameter in the two-sexes population growth equation (2.17) (I leave it to the reader to gure out the correct dimensions of these two parameters). If two quantities are multiplied, the result carries the product of their dimensions. Dividing two quantities is more subtle: if two quantities with dierent dimensions are divided the result carries the ratio of their dimensions. For example, the number of individuals in a population divided by the surface of the habitat they live in yields the population density in terms of the number of individuals per unit area. On the other hand, if two quantities with identical dimensions are divided the result is dimensionless. Whether dimensions are dierent or identical is sometimes a bit subtle, as the following example illustrates: a frequently encountered parameter in population models is the assimilation eciency. This is the eciency with which ingested food is transformed into assimilated energy. If both food and assimilated energy are measured in the same units, for example both are measured in terms of their caloric value, i.e. in Joules, the ratio of assimilated energy over ingested food is truly dimensionless. On the other hand, the per capita rate with which a predator consumes prey individuals has the dimension of number of individuals per individual per unit of time, ind/(indtime), and not simply per unit of time. The reason is that the one type of individuals are prey (in the numerator) while the other type of individuals (in the denominator) are predators. Hence, these dimensions do not cancel. As a rule if two quantities could be added to each other on the basis of their dimensions, their dimensions would cancel out when their ratio is computed. Using the above rules it is good practice to check whether a set of model equations bear the correct dimensions and whether the parameters in the model indeed are expressed in terms of the correct dimensions. A lack of consistency in dimensions often points at some inconsistency in the model formulation. The choice of physical dimensions for model variables are to some extent arbitrary. For example, to express a density of 105 bacteria in a volume of 1 liter growth medium, any of the following dimensions can be used: N = 105 cells/liter = 0.1 million cells/liter = 0.1 cells/mm3 = 100 cells/milliliter Similarly, the variable time t in the model could be measured in minutes, days, weeks or years and this choice is absolutely arbitrary. This freedom in choosing model dimensions allows us to write all measured quantities in the model as a product of a scalar multiple and a unit carrying dimensions. For example, the density N can be written as N = N N (3.27)

in which N determines the scale of measurement, for example, cells/liter, cells/mm3 or cells/milliliter and carries the physical dimensions. The quantity N takes on dierent values, i.e. 105 , 0.1 or

34

HOOFDSTUK 3. SINGLE ORDINARY DIFFERENTIAL EQUATIONS

100, depending on the scale set by N . Similarly, the time t can be expressed as the product of and a quantity setting the scale of measurement and carrying the physical a scalar multiple t dimension t: t = t t (3.28) Over time the model species that the values of both N and N will change, but the quantity N remains constant. Similarly, the value of the time variable t and t will change, but not the scaling variable t. The choice of units in which to express the state variables in the model will determine the particular value of parameters: if a population growth rate r is to be specied, it should be in terms of the unit of time adopted in the model, i.e. either, say, r = 0.1 per day or equivalently r = 0.7 per week. On the other hand, it is obvious that the qualitative dynamics of the population model should be independent of this choice of units. The invariance in model dynamics under a change of measurement units can be exploited to reduce the number of parameters in the model and hence to reduce the number of quantities that determine the ultimate population dynamics. Consider for example the logistic growth equation: dN = rN dt 1 N K (3.29)

and substitute the expressions (3.27) and (3.28) for N and t, respectively: d (N N ) = r (N N ) t) d (t 1 (N N) K (3.30)

Because both N and t are time-independent constants, this ODE can be rewritten as: N d N = r (N N ) t d t 1 (N N ) K

Multiplying both sides of this latter ODE by t/N yields the ODE: d N = rt N d t 1 (N N ) K (3.31)

Notice that we are still free in our choice of measurement units N and t, as we did not yet make any assumption about them and their choice is arbitrary. There are some smart (or rather judicious) choices we could make for these two scales of measurement. For example, choosing: t = 1/r N = K leads to an ODE in terms of dimensionless quantities N and t : d N = N (1 N ) d t which does not contain any parameters anymore. By dropping the starred superscripts on the quantities N and t the ODE can be written in the following form: dN = N (1 N ) dt (3.34) (3.32) (3.33)

3.6. EXISTENCE AND UNIQUENESS OF SOLUTIONS

35

The key point of the procedure just described is that by making judicious choices for our units in which the state variables in the model are expressed, we have been able to reduce the number of parameters by 2: For whatever values of r and K, the dynamics of the model (3.29) is identical to the dynamics of the ODE (3.34) except for some scaling of both the axis of time t and density N. By appropriately choosing our scales of measurement in the model, it has been possible to arrive at a dimensionless form of the logistic growth model from which we eliminated two parameters and which no longer depends on parameters at all. The resulting equation is much simpler and does not contain any degrees of freedom, which severely reduces the amount of work needed to analyze the model properties. As a rule, by rewriting a population dynamic model in a dimensionless form we can reduce the parameters set in the model by a number that equals the number of state variables in the model (i.e. all population densities plus the time variable). For example, a dimensionless form of the two-sexes population growth model (2.17) will ultimately contain only a single parameter. As shown in this section, rewriting population dynamic models in a dimensionless form is a powerful approach to reduce some of the complexity of the model and to do away with some of its degrees of freedom.

3.6

Existence and uniqueness of solutions

The existence and uniqueness of solutions for particular ODEs is usually a topic that mathematicians are concerned about and hence investigate. Ecological modelers are much less or perhaps even not at all concerned about this. Nonetheless, it is appropriate to at least warn for the possibility that after formulating a population dynamic model we could potentially end up with a (system of) ODE(s) for which no unique solution exists or for which no solution exists at all! The existence and uniqueness of solutions in general not only depends on the ODE itself, but on the combination of the ODE and the initial condition. I will not discuss this issue in detail, but I will give just two examples that may give some idea of the problems that you can run into: A solution of an ODE is not necessarily always a nice function of the time t. For example, had we formulated the two-sexes population growth equation (3.4) with a density independent reproduction rate and had we ignored the mortality rate, we would have ended up with the ODE: d N (t) = N2 (3.35) dt The solution of this ODE for an initial population abundance N (0) = N0 equals: N (t) = N0 . 1 N0 t (3.36)

This solution only exists for t < 1/N0 . At t = 1/N0 the ODE (3.35) does not hold any longer, while the solution (3.36) in the meanwhile became innitely large. The reason is that the right-hand side of the ODE (3.35) increases too quickly with increasing values of N . Explosions of this kind (and also implosions that are qualitatively similar) are in any case impossible if the right-hand side of the ODE fullls the condition: m kN < f (N ) < m + kN for some arbitrary, but positive constants k and m and N > 0 (a similar condition can be formulated for N < 0, but is considered uninteresting in the current context).

36

HOOFDSTUK 3. SINGLE ORDINARY DIFFERENTIAL EQUATIONS An example of an ODE without a unique solution is: d N (t) dt = N, N (0) = 0 (3.37)

with c an arbitrary, positive constant species an entire family of solutions to the ODE: for whatever value of c, equation (3.38) is a solution for the ODE. Hence, these solutions are never unique. The problem is that the right-hand side of the ODE (3.37) increases too quickly for N = 0, since the tangent to N in this point is a vertical line. This type of non-uniqueness is guaranteed not to occur if the derivative of the right-hand side function f (N ) exists and is continuous for all permissible values of its argument N .

Even though it is hard to imagine that this ODE represents a population dynamic model, it suits our purpose to illustrate the non-uniqueness of solutions. The equation 0 for t c 2 N (t) = (3.38) tc for t c 2

3.7

Epilogue

In principle this chapter has discussed most of the theory about how to analyze population dynamic models that are formulated in terms of (systems of) ODEs. Also, most of the material discussed allows generalization to models that are formulated in terms of other frameworks, such as dierence equations or integral equations. When models become of higher dimension, new concepts will be needed to deal with technical complications, making the analysis also technically more dicult. For example, when analyzing systems of ODEs matrix calculus is needed to derive the characteristic equation and the eigenvalues. Nonetheless, the basic ideas are the same as discussed here for one-dimensional models (i.e. in terms of a single ODE): Models can be reformulated as non-dimensional analogues. These analogues are phrased in terms of scaled state variables and scaled time with the minimum number of (scaled) parameters, but describe exactly the same dynamics. Steady states are determined by nding those states for which the right-hand side of the ODEs vanish. The stability of the steady states can be determined by studying the linearized version of the model in the neighborhood of the steady states. For this linearized system a characteristic equation can be derived, which the eigenvalues (or characteristic growth rates) of the model in the neighborhood of the steady state. One aspect of model analysis has not been discussed in this chapter: studying the behavior of the model for dierent parameter values. This addresses the question how the model dynamics is going to change when particular parameters are increased or decreased, what type of dierent dynamics can occur over the full range of relevant parameter values, how one type of dynamics transforms into the other, etc. These questions are within the domain of bifurcation theory which will be discussed in part III.

Hoofdstuk 4

Competing for resources

Even though the concept competition is central to population dynamics and ecology, its precise meaning is less clear cut. In a general textbook on ecology (Begon et al. 1996), competition is dened as an interaction in which one organism consumes a resource that would have been available to, and might have been consumed by, another. One organism deprives another, and as a consequence, the other organism grows more slowly, leaves fewer progeny or is at greater risk of death. The above denition suggest that competition is about food, as the denition refers to organisms consuming resources. However, competition may be about many dierent things that have nothing to do with feeding, such as nesting sites, territoria and mating partners. Basically, competition is always an act that one individual takes, eats or uses something, which is subsequently unavailable to other individuals of the same or other species. That something is referred to as a resource. Competition may be between individuals of the same species, in which case we call it intraspecic competition, or of dierent species, in which case it is called interspecic competition. Moreover, competition may be direct or indirect. An example of direct or interference competition is when two individuals ght for the same piece of food that one of them has just caught. Also competition for sexual partners is usually a form of interference competition. On the other hand, indirect competition, which is also referred to as exploitation competition does not involve direct contact among the competitors. The most obvious example is when individuals feed on a shared food source: the food eaten by one individual is unavailable to the others and hence there is a competitive interaction. In this chapter we will mainly deal with exploitation competition. After this introduction of the most common terminology regarding competition we return to the question what exactly is a resource? Answering this question is actually very hard. Instead of elaborately discussing the meaning of the term resource I will just give some examples of what can rightfully be called resources. Food is denitely a resource for all heterotrophic organisms. Appropriate types of shelter or space have also been termed resources if they have an eect on an individuals probabilities of birth and death. Common resources for plants are light, mineral nutrients, water, pollinators and seed dispersal agents. Within a species, members of one of the sexes have been called a resource for the opposite sex. Basically, two requirements can be distinguished for something to be considered a resource: 1. it should aect the survival or reproduction of an individual, and 2. it should be consumed or used and hence depleted, making it unavailable to other individuals. 37

38

HOOFDSTUK 4. COMPETING FOR RESOURCES

Figuur 4.1: Growth of the yeast Schizosaccharomyces kephir over a period of 160 h. The circles are experimental observations. The solid curve is the tted curve N (t) = 5.8/(1 + exp(2.47 0.0607t)). From Gause (1934). This chapter is devoted to formulating various models that incorporate intra- and/or interspecic competition. The overall aim of the chapter is on the one hand to introduce a number of concepts and ideas that have been developed as part of a competition theory. On the other hand, while discussing the ins and outs of competition within and among species, a number of new techniques or extensions of already discussed techniques will be introduced. The rst part of the chapter will deal with intraspecic competition, while the second part will focus on interspecic competition.

4.1
4.1.1

Intraspecic competition in a single population

Growth of yeast in a closed container

Some classic examples about population growth and intraspecic competition are the experiments carried out by Gause (1934) with the yeast Schizosaccharomyces kephir (see Fig. 4.1). These experiments have been carried out in closed containers such as test tubes and Erlenmeyer asks. The experiments were started with a virgin container with growth medium and a small inoculation of yeast. Figure 4.1 shows that at rst the yeast population starts to grow exponentially, but that growth decelerates and population abundance reaches a constant level. The tted curve in Fig. 4.1 is actually the logistic growth equation (2.14). Both the logistic an two-sexes population growth model (Eq. (2.14) and (2.17), respectively) are examples of models that implicitly incorporate intraspecic competition, because in both cases the per capita reproduction rate decreases with population abundance. These model hence do specify that the presence of other individuals around decreases the reproductive output of single individuals, i.e. individuals in the population do compete, but the models leave unspecied what the precise mechanism of this decrease is. It could hence be because food availability declines or because individuals at higher densities spend more time ghting each other. For the experimental design used by Gause (1934) it is possible to formulate an explicit and

4.1. INTRASPECIFIC COMPETITION

39

mechanistic model for the growth of the yeast in the closed container. In the end this model will be the same as the logistic growth model (2.14), but it will be derived on the basis of an explicit conceptualization of the competition process going on among the yeast cells. To this end assume that the growth of the yeast is limited by a single nutrient (resource), whose concentration at time t is indicated by R(t), every yeast cell contains units of the limiting nutrient, the initial concentration of the limiting nutrient equals R(0), while the initial abundance of yeast cells equals N (0), a single yeast cell takes up the limiting nutrient at a rate R(t) that is proportional to the current nutrient concentration, and a single yeast cell divides into 2 identical daughter cells after having taken up units of the limiting nutrient. At the start of the experiment, the total amount of limiting nutrient in the experimental vessel equals: R0 := R(0) + N (0) (4.1) which is the sum of the nutrient amount in solution and the nutrient amount contained in the initial yeast population. Since nutrient only disappears from the solution through uptake by the yeast, at any time t the total amount of nutrient in the vessel (both in solution and incorporated into the yeast population) should equal R0 : R(t) + N (t) = R0 (4.2)

The uptake rate of the limiting nutrient by a single yeast cell equals R(t). As a consequence the time it takes to take up sucient nutrients for a new daughter cell equals , R(t) such that the rate at which a single yeast cell produces new daughter cells equals R(t) . This leads to the following ODE, describing the dynamics of yeast cells: R(t) dN = N (t) dt If now the relation (4.2) is substituted in this ODE we obtain: dN = dt which can be rewritten as: R0 dN = dt R0 N (t) N (t) R0 / N (t) (4.4) (4.3)

N (t)

(4.5)

40

HOOFDSTUK 4. COMPETING FOR RESOURCES

This latter ODE is identical to the logistic growth equation (2.14) with the parameters r and K dened as: R0 R0 K= r= (4.6a) (4.6b)

The dynamics of both the limiting nutrient concentration R(t) and the yeast cell abundance N (t) could have been expressed as a system of two, coupled ODEs: dR = R(t)N (t) dt dN R(t) = N (t) dt (4.7a) (4.7b)

The derivation of the logistic growth equation (4.5) for the yeast cell abundance N (t) actually represents a way of solving the coupled system of ODEs (4.7). In a more formal way of solving the system of ODEs, the rst equation (4.7a) is rewritten by using the second ODE (4.7b): dN dR = dt dt (4.8)

By integrating both the left-hand and right-hand side of this latter ODE we actually end up with:
t 0

dR dt = dt

t 0

dN dt dt

R(t) R(0) = (N (t) N (0)) R(t) = N (t) + R(0) + N (0) which is equivalent to equation (4.2) relating the current concentration of nutrient in solution and the current yeast cell abundance to the total amount of nutrient initially present in the experimental vessel. The above derivation shows that the logistic growth model can actually have a solid mechanistic basis in terms of competition for a limiting nutrient supply.

4.1.2

Bacterial growth in a chemostat

Chemostats are experimental systems that are used to sustain a continuous culture of bacteria. In contrast to the closed container experiments discussed in the previous section, a chemostat is an open system with a continuous supply of nutrients. Figure 4.2 shows a schematic layout of a chemostat. A stock nutrient solution, in which the concentration of limiting nutrient equals R0 , is pumped into the culture vessel at rate F . An equal eux keeps the total volume V in the culture vessel constant over time. Note that through the outow both nutrient and bacteria leave the culture vessel. Apart from the importance of the chemostat for sustaining a continuous culture of microorganisms (primarily bacteria), it also constitutes a highly idealized system for modeling and studying competition within and among species, which has given rise to even textbooks on the theory of the chemostat (Smith & Waltman 1994). To model a single bacterial population in continuous culture we make the following assumptions:

4.1. INTRASPECIFIC COMPETITION

Stock Nutrient Reservoir F R
0

41

Inow Volume V

Culture Chamber

N R

Euent

Figuur 4.2: Schematic layout of a chemostat for continuous culture of micro-organisms. A stock nutrient solution, in which the concentration of limiting nutrient equals R0 , is pumped into the culture vessel at rate F . An equal eux keeps the total volume V in the culture vessel constant over time. It is assumed that the culture vessel is kept well stirred, such that there are no spatial inhomogeneities in the concentration of either nutrient or bacteria. This assumption allows us to describe the growth process entirely in terms of ordinary dierential equations. Although the nutrient medium may contain a number of components, we will focus attention on a single growth-limiting nutrient whose concentration will determine entirely the growth of the bacteria. The current concentration of this limiting nutrient in the culture vessel at time t will be denoted by R(t). Bacteria take up the limiting nutrient at a rate R(t) that is proportional to the current nutrient concentration. Every bacterial cell is assumed to contain units of the limiting nutrient. A bacteria divides into 2 identical daughter cells after having taken up units of the limiting nutrient. These assumptions are very much analogous to the assumptions made while formulating the yeast growth model in a closed container (see section 4.1.1). The essential dierence between the model for the chemostat and the yeast growth model, is that now fresh nutrient medium is pumped into the culture vessel at rate F , while both nutrient and bacterial cells leave the culture chamber at the same rate F . Table 4.1 lists the state variables and parameters of the model along with their interpretation and dimension. As a rst start, we can use the model equations (4.7) as a basis to describe the dynamics of both the nutrient concentration and the bacterial population abundance in the chemostat. The only terms that have to be added are the ones accounting for inow of fresh nutrient from the reservoir and outow of nutrient and bacteria from the culture chamber. A rst attempt for the model equations hence is: dR = F R0 F R RN dt R dN = N FN dt (4.9a) (4.9b)

In these equations I have dropped the explicit dependence of R and N on time t in the righthand side of the ODE. This is an established practice as the ODE itself indicates that these quantities are functions of time. The rst ODE of this system contains the terms F R0 and F R, respectively, modeling the inow and outow of nutrients. Similarly, the second ODE of

42

HOOFDSTUK 4. COMPETING FOR RESOURCES Tabel 4.1: State variables and parameters of a bacterial growth model in a chemostat Symbol Description Dimension

State variables R(t) Nutrient concentration in culture chamber N (t) Bacterial population abundance Parameters R0 Nutrient concentration in reservoir F Input and output ow rate V Volume of the culture chamber Proportionality constant for bacterial nutrient uptake Nutrient content of single bacteria

mass/volume number/volume

mass/volume volume/time volume volume/time mass

the system incorporates the term F N , which models the outow of bacteria from the culture chamber. Even though these ODEs contain all the necessary terms, they are not correct. This becomes clear when we carefully check the dimensions of all terms in the above system of ODEs (cf. section 3.5). First of all we should gure out what the dimension of the parameter is, while that is not immediately clear from the assumptions. Because the left-hand side of equation (4.9b) bears the dimension number/(volumetime), the right-hand side should bear this dimensions as well. Given the dimensions of N , R and , occurring in the rst term of the right-hand side (4.9b), we can infer that needs to bear the dimensions volume/time, as is already listed in table 4.1. However, the second term in the right-hand side of equation (4.9b) has dimension number/time which lacks a volumetric dimension in the denominator. Similarly, the left-hand side of equation (4.9a) bears dimension mass/(volumetime), while the rst and second term of its right-hand side have a dimension mass/time. Again a volumetric dimension is missing in the denominator. (Supercially, the last term in the right-hand side of equation (4.9a) also has the wrong dimension, since straightforward application of the dimensions listed in table 4.1 would suggest that this term carries the dimension (numbermass)/(volumetime). It should be realized, however, that a number is essentially dimensionless such that the dimension is indeed correct: mass/(volumetime)). The mistake made in the derivation of the system of ODEs (4.9) is that the rate F with which fresh nutrient supply of concentration R0 enters the culture chamber distributes itself over the entire volume V of the culture vessel. The total input of fresh nutrients, equal to F R0 , hence distributes itself over a volume V and cause a change in the nutrient concentration equal to F R0 /V . Similar arguments hold for the outow rates of nutrients and bacteria from the culture vessel, F R and F N respectively, which also have to be divided by the vessel volume V to correctly express the eect of outow on the nutrient concentration R(t) and cell abundance N (t). Therefore, the correct set of equations describing the dynamics of nutrients and bacteria in the chemostat is: F R0 FR dR = RN dt V V R FN dN = N dt V (4.10a) (4.10b)

4.1. INTRASPECIFIC COMPETITION

43

This set of ODEs is the rst system of coupled ODEs that we encounter. Such coupled sets of equations occur in all models that describe the dynamics of more than a single component. These components can be multiple species or nutrients and populations as in the chemostat example discussed here. In the following sections we will discuss a number of dierent methods to analyze the behavior prescribed by the model.

4.1.3

Asymptotic dynamics

One approach to analyzing a system of ODEs like (4.10) is to attempt an analogous simplication as was discussed in section 4.1.1 (see equation (4.8)) for the growth of a yeast population in a closed container. In other words, can we solve the ODE (4.10a) for the nutrient concentration R(t) in terms of the cell abundance N (t)? The background of equation (4.8) was that the total nutrient concentration, that is the sum of nutrients in solution plus nutrients contained in yeast, remained constant over time, as the dynamics were taking place in a closed container. By dening the total nutrient concentration in the chemostat (both free in solution and bound in bacterial cells) as: TR (t) := R(t) + N (t) , (4.11) an ODE can be derived for TR (t) as follows: d(R + N ) dTR = dt dt = = dN dR + dt dt F R0 FR RN + V V FR FN F R0 V V V F R0 F TR V V R FN N V

= =

Hence, eventually the total nutrient concentration in the culture chamber will follow the ODE: F dTR = R0 TR dt V From this ODE we can easily see that
t

(4.12)

lim TR (t) = R0

(4.13)

The value TR = R0 is actually the steady state value for the total nutrient concentration TR , i.e. the concentration for which dTR /dt = 0. We could now simply assume that dTR 0 dt holds for all times t. Especially if the the ow rate F is large and the volume is not too large, the quotient F/V can be large in comparison with the growth rate of the bacteria, which equals R/: R F >> V

44

HOOFDSTUK 4. COMPETING FOR RESOURCES

In this case, the total nutrient concentration TR (t) will have approximated its steady state value R0 long before the cell abundance N (t) has even changed a little from its initial value. In this case we can simply pretend that TR = R(t) + N (t) R0 This latter equality can subsequently be used to eliminate the nutrient concentration R(t) from the ODE (4.10b) for the dynamics of the bacterial cell abundance. This leads to: (R0 N ) FN dN = N dt V = R0 F N V N

After some algebraic manipulation we see that the resulting ODE can be written as a logistic growth equation (2.14) with parameters: F R0 V R0 F K= V r=

(4.14a) (4.14b)

The logistic growth equation can hence show up even though the underlying mechanisms are entirely dierent. The expressions for the parameters (4.14) also make clear that the parameters r and K in this case have no interpretation that is intrinsic to the bacteria, since also the ow rate F , the nutrient supply concentration R0 and the vessel volume V show up in both r and K.

4.1.4

Phase-plane methods and graphical analysis

The aim of this chapter is to study the behavior of systems of coupled ODEs mostly through graphical and geometric approaches. Hence, instead of more rigorous algebraic manipulations we will exploit geometric insights and some intuition to chart out the qualitative behavior of a model in a pictorial form. Such pictures are generally far more informative than mathematical expressions and lead to a better understanding of the model dynamics and how it is inuenced by its parameters. For systems of two ODEs it is possible to carry out a graphical analysis much along the lines that was discussed in section 3.3 (Some people even claim to analyze systems of three ODEs in this manner, but such experience is simply beyond me). However, since there are now 2 state variables it is not much use any longer to plot the right-hand side of the ODE as a function of a single state variable (cf. section 3.3). Instead, our basic analysis will use graphs spanned by an x- and a y-axis on which the two state variables R(t) and N (t) are indicated. The choice which state variable to represent with which axis is arbitrary, but I will consistently use R(t) on the x-axis and N (t) on the y-axis of the graphs:

4.1. INTRASPECIFIC COMPETITION

N

45

Such a graph spanned by axes that represent state variables is called a phase plane graph. Its name is derived from the fact that all the combinations (R, N ) depicted in the graph represent all possible states that the system we are modeling can adopt. If the model we are studying becomes higher dimensional, for example, in terms of three or more variables, we are usually referring to a phase space or state space instead of a phase plane, because the set of all possible states that the system can adopt also becomes higher dimensional. In the following I will refer to the phase plane as the (R, N )-plane. First, let us rewrite the chemostat model equations (4.10) as: dR = f (R, N ) dt dN = g(R, N ) dt in which the right-hand side functions f (R, N ) and g(R, N ) are dened as: f (R, N ) = g(R, N ) = F R0 FR RN V V FN R N V (4.16a) (4.16b) (4.15a) (4.15b)

Technically, we have to assume that f (R, N ) and g(R, N ) are continuous functions and that their derivatives with respect to both R and N exist to ensure that the set of ODEs actually species a unique solution for an particular initial conditions R(0) and N (0) (see also section 3.6). Two, time-dependent functions, R(t) and N (t), would be a solution to the system of ODEs (4.15) if they satisfy the set of ODEs (4.15) together with the initial conditions R(0) and N (0) imposed (if there are any). In the phase plane any combination of R(t) and N (t) represents a point, which we can equivalently represent by a vector x(t) dened as: x(t) := R(t) N (t) (4.17)

The system of ODEs (4.15) is a recipe that species how the values R(t) and N (t) change over time. At the same time, it therefore species how the point (R(t), N (t)) or vector x(t) moves over time through the (R, N )-plane. By dening a vector-valued function H(x) as follows: H(x) := f g = dR/dt dN/dt (4.18)

46

HOOFDSTUK 4. COMPETING FOR RESOURCES

N (R(0), N (0))

(R(t), N (t))

dx/dt = (dR/dt, dN dt) x(t)

R
Figuur 4.3: Geometric representation of a system of ODEs in the phase plane. The curve starting in the point R(0), N (0)) represents some imaginary solution curve to a system of ODEs, which also passes through the point (R(t), N (t)) at time t. The vector x(t) is the vector representation of this latter point. Also indicated is the velocity vector dx/dt determined by the system of ODEs. the systems of ODEs (4.15) could therefore also be written in vector form as: dx = H(x) dt (4.19)

(Notice that I have left out the explicit dependence of the functions f and g in equation (4.18) on the state variable R and N ). Loosely speaking, the left-hand side of ODE (4.15a) and (4.15b), dR/dt and dN/dt, respectively, can be interpreted as the change in the nutrient concentration R(t) and bacterial cell abundance N (t) during an innitesimally small time interval dt. In the phase plane these two quantities determine the direction of a curve through the point (R(t), N (t)), which is the solution curve to the system of ODEs (see also Figure 4.3). This direction specied by dR/dt and dN/dt can again be represented by a vector, which in equation (4.18) is dened as H(x). This vector is tangent to the solution curve in the point x(t) = (R(t), N (t)). These geometric representations of the system of ODEs (4.15) and (4.19) are illustrated in the phase plane graph in Fig. 4.3. Verbally they can be summarized as follows: 1. The pair of values (R(t), N (t)) that forms a solution to the system of ODEs (4.15) represents a curve in the (R, N )-phase-plane. Each position on this curve is uniquely determined by the time value t (i.e. the curve is parameterized by t). 2. The point (R(t), N (t)) can also be represented by a vector x(t), starting in the origin and pointing to a position along the solution curve (R(t), N (t)), which is the state of the system at time t. 3. The vector dx/dt, which is just a short-hand notation for (dR/dt, dN dt) is the tangent vector to the solution curve in the point x(t) and thus represents the curve direction.

4.1. INTRASPECIFIC COMPETITION

47

Its magnitude, which is indicated by |dx/dt|, represents the speed of motion of the point (R(t), N (t)) along the solution curve. The vector dx/dt is hence also referred to as the velocity vector or direction vector , while the vector x(t) is referred to as the position vector . 4. Using x(t) the system of ODEs (4.15) for the change in nutrient concentration R(t) and cell abundance N (t) can als be written in the shorter, vector form dx = H(x) dt Here the vector function H(x), dened in equation (4.18), assigns a velocity vector to every location x in the phase plane. Having determined that the right-hand side of a system of ODEs like the one in equation (4.15) determines a tangent velocity vector to the unique solution curve through any point x = (R, N ), one way to proceed with a graphical analysis is by drawing a large number of these velocity vectors in the (R, N )-phase-plane. Together such a collection of velocity vectors is referred to as a direction eld or a phase plane portrait. Drawing phase planes with direction elds gives, after some training, good insight into the dynamics of systems of 2 ODEs (see, for example, EdelsteinKeshet 1988, section 5.4 on page 175-178). Another way to gain insight into the dynamics is by delineating in the phase plane graph the regions with (R, N ) combinations for which: both the nutrient concentration R and the bacterial cell abundance N increase, both the nutrient concentration R and the bacterial cell abundance N decrease, the nutrient concentration R increases, but the bacterial cell abundance N decreases, and the nutrient concentration R decreases, but the bacterial cell abundance N increases. Delineating these 4 dierent regions will be carried out in two steps: rst we will nd those combinations of (R, N ) for which the nutrient concentration R increases or decreases. In the second step, the same procedure will be repeated for the bacterial cell abundance N . For which values of R and N the nutrient concentration R increases or decreases can be derived by considering for (R, N ) combinations it does not change. In other words, we rst derive for which values of R and N dR = 0 dt Substituting the right-hand side of the ODE (4.10a) in the equality above leads to the equation: F R0 FR RN = 0 V V With some simple algebraic manipulation this equality can be rewritten as: N = F V R0 1 R (4.20)

This last equation determines a curve in the (R, N )-phase plane, indicating all those values of R and N for which the nutrient concentration R does NOT change. Such a curve, for which dR/dt = 0 holds, is called the isocline for the nutrient concentration R. Isoclines or nullclines are curves in the phase plane for which one of the state variables (R and N in the case considered here) does not change. Equation (4.20) determines a curve that intersects the x-axis of the phase plane (i.e. the axis N = 0) for the value R = R0 . On the other hand, if R decreases to 0, R 0 the bacterial cell

48

HOOFDSTUK 4. COMPETING FOR RESOURCES

abundance N increases unboundedly, N . If we substitute the values R = 0 and N = 0 in the right-hand side of the ODE (4.10a), the rate of change of the nutrient concentration R is found to equal: F R0 dR = dt V which is clearly positive. Therefore, for values of R and N close to 0 the rate of change dR/dt is clearly positive. On the other hand, if R and N are both very large, it is easily inferred from the right-hand side of the ODE (4.10a) that the rate of change dR/dt is negative. These characteristics of the ODE (4.10a) can now be summarized in the following phase plane graph:
N

R0

The thick solid curve drawn in the graph above is the isocline for the nutrient concentration R. Hence, for those combination of R and N the rate of change dR/dt equals 0 (i.e. not change in nutrient concentration). The arrows below and above the curve indicate that in these regions with (R, N )-combinations the rate of change dR/dt is positive (i.e. R will increase with time) or negative (i.e. R will decrease with time), respectively. Next we repeat the same type of analysis for the ODE (4.10b) determining the rate of change dN/dt of the bacterial cell abundance N . Equating dN/dt to 0 leads to the relation: FN R N = 0 V which shows that it is possible to infer two relations for which dN/dt = 0: N = 0 or R = F V (4.22) (4.21)

The curves of these relations in the (R, N ) phase plane are again referred to as the isoclines, now for the bacterial cell abundance N . Obviously, one isocline for the bacterial cell abundance N is given by N = 0: if there are no bacteria at all present there will be neither growth or decline and hence dN/dt obviously equals 0 (this zero isocline is encountered in almost every population dynamical model, in which population growth or decline only depends the population abundance and there hence is no immigration). In addition, equation (4.22) determines a second isocline for the bacterial cell abundance N which constitutes a vertical line in the (R, N ) phase

4.1. INTRASPECIFIC COMPETITION

49

plane. For high values of nutrient concentration R and non-zero bacterial densities N > 0, the right-hand side of the ODE (4.10b) clearly shows that dN/dt is positive and hence the bacterial cell abundance will increase. For values of nutrient concentration R close to zero and non-zero bacterial densities N > 0, dN/dt is negative and N will decrease. These properties of the ODE (4.10b) are summarized in the following graph:
N

F/(V )

The isoclines in this graph are indicated with thick dashed line, one isocline being N = 0, the other R = F/(V ). To the left of the latter isocline the bacterial abundance will decline, to the right of it the bacterial cell abundance N will increase. The next step is that we combine the two graphs shown above with the isoclines and ow patterns of the nutrient concentration R and the bacterial cell abundance N into a single phase plane graph. Figure 4.4 shows the resulting graph. The two intersections of the R-isocline with the N -isoclines are the steady states of the model (4.10), as for these two combinations of R and N neither the nutrient concentration, nor the bacterial cell abundance will change (i.e. both dR/dt and dN/dt equal 0). The two steady states are indicated by solid circles in the graph above. The phase plane with positive value of R and N (usually referred to as the positive quadrant or more generally the positive cone) is hence divided into 4 dierent regions by the R- and N -isoclines. In each of these regions a single pair of a vertical and a horizontal arrow in the graph above indicate whether the nutrient concentration R and the bacterial cell abundance N increase or decrease, respectively. Mathematically, the two steady states are given by: Extinct steady state: Internal steady state: {R, N } = {R0 , 0} {R, N } = {F/(V ), R0 / F/(V )} (4.23) (4.24)

The bacterial cell abundance N for the internal steady state is obtained by substituting R = F/(V ) into equation (4.20) and simplifying the resulting expression. From the isoclines and ow patterns it is sometimes possible to work out whether the steady states are stable and what the dynamics will eventually be in the neighborhood of the steady states. This can be done for the extinct steady state {R, N } = {R0 , 0}. The phase plane graph (Fig. 4.4) shows that if the bacterial cell abundance N would be slightly perturbed (displaced) to a very small positive value, the abundance would actually increase (the ow arrows are pointing upwards and to the left). Hence, the cell abundance N would increase even further, while the nutrient concentration R will decrease. On the other hand, if the nutrient concentration would be slightly perturbed to a lower or higher value than R0 , the nutrient concentration would return

50
N

HOOFDSTUK 4. COMPETING FOR RESOURCES

R0

F/(V )

Figuur 4.4: The phase plane of the bacterial growth model in a chemostat. Thick dashed lines indicate the zero isoclines for the change in bacterial cell abundance, the thick solid line indicates the zero isocline for the change in nutrient concentration. See text for further explanation. to R0 , as long as the bacterial cell abundance is kept at 0. This assessment can be derived from the ODEs (4.10) and by studying the phase plane graph in Fig. 4.4. Therefore, the extinct steady state is in general unstable, but it is stable against perturbations along the x-axis of the nutrient concentrations R. A steady state with the property that it can withstand perturbations in particular directions, but not in all and that is hence unstable, is called a saddle point. A similar analysis for the internal steady state is, however, in general not possible. The following graph zooms in at the phase plane in the neighborhood of this steady state:

In this graph I have also drawn direction vectors for 4 points on the two isoclines. The vectors for the points at the R-isocline necessarily point in a vertical direction, since for all the points on the R-isocline dR/dt = 0. Similarly, the vectors for the points at the N -isocline necessarily point in a horizontal direction, since for all the points on this isocline dN/dt = 0. The graph above seems to indicate that the solution curves tend to spiral around the steady state point, but nothing in the graph indicates whether they will spiral inward towards the steady state point (in which case the latter is stable) or outward away from the steady state point (in which

4.1. INTRASPECIFIC COMPETITION

51

case it is unstable). Graphical methods do not suce to resolve this issue and we will have to build up some more analytical understanding of dynamics in the neighborhood of a steady state before we can determine whether the state of the system will eventually approach the steady state or not (see also chapter 5). Actually the bacterial population in a chemostat is an exception to the above rule that we cannot work out the stability of a steady state from a phase plane analysis. In section 4.1.3 we already learned that the total nutrient concentration in the chemostat, both free in solution and bound in bacteria, obeys the very simple ODE (4.12). This total nutrient concentration TR equals R + N and the ODE (4.12) species that it will never cross the value TR = R0 : if the total nutrient concentration TR at time 0 is larger than R0 it will always remain larger and if it is smaller it will always remain smaller than R0 . In the phase plane graph this means than no solution curve will cross the line which is determined by the relation: R + N = R0 (4.25)

Using this fact we can actually work out the dynamics in the neighborhood of the internal steady state in this particular case, as is shown in the following diagram:

R + N = R0

Here I have added the line determined by (4.25) to the phase plane and have sketched two representative solution curves that approach the internal steady state (note that these curves only qualitatively represent the actual trajectories of the model). One of these solution curves start with a total nutrient concentration TR larger than R0 , the other with TR smaller than R0 . In the entire discussion of the chemostat model up to this point, I have pretended that the vertical N -isocline was located at a nutrient concentration R that is smaller than R0 . The N isocline was located at R = F/(V ). As long as R0 > F/(V ) there exists an internal steady state, because the N -isocline and the R-isocline intersect in the positive quadrant (see Fig. 4.4). However, if R0 < F/(V ) an internal steady state is impossible to maintain. Under these conditions the input nutrient concentration is too low to sustain a bacterial population in the chemostat. The relation F = V R0 (4.26)

therefore determines a critical boundary, separating those parameter values R0 , F , and V for which a bacterial population can be maintained in the chemostat from those parameter values for which the bacteria will go extinct. Because the critical boundary is also determined by the

52
Flow rate F

HOOFDSTUK 4. COMPETING FOR RESOURCES

Bacterial wash-out

m p ti

al

ow

e rat

Nutrient input concentration R0

Figuur 4.5: Operating diagram of a chemostat. The graph gives a schematic representation of those parameter combinations F and R0 for which wash-out of bacteria will occur and which ow rate F will give a maximal yield of bacteria (dashed curve). See text for further explanation. ow rate F of the chemostat, the extinction of the bacteria is also referred to as wash-out, since too high a ow rate will cause the bacteria to go extinct. If an internal steady state is possible (i.e. F < V R0 /), the steady state outow of bacteria equals: F2 R0 F F Nss = V where Nss is used to indicate the steady state bacterial abundance (see also equation (4.24)). This outow of bacteria is referred to as the yield of the chemostat, while the idea of the chemostat is to harvest the bacteria growing in the culture vessel. The yield of bacteria is maximal for a ow rate F equal to: V R0 (4.27) F = 2 which is exactly half the ow rate above which wash-out of the bacteria will occur. Schematically these relations are summarized in Figure 4.5 which represents a parameter plane spanned by a x-axis representing the nutrient concentration in the inow R0 and a y-axis representing the ow rate F . The gure gives an overview of all parameter combinations for which bacterial wash-out will occur and graphs the relation (4.27) for the optimal ow rate of the chemostat.

4.2
4.2.1

Interspecic competition
Lotka-Volterra competition model

The most well-known model for interspecic competition has been proposed by Lotka and Volterra and has been studied extensively by Gause (1934). In the Lotka-Volterra model the competition between two species is represented without any reference to resources. For a particular species the presence of a competitor is simply assumed to reduce its growth.

4.2. INTERSPECIFIC COMPETITION

53

Let us denote the two species by their population abundance N1 and N2 , respectively. In the absence of the competitor, both species 1 and 2 are assumed to grow following the logistic growth model (2.14). When both species are present their dynamics are assumed to be described by the following system of ODEs: dN1 = r1 N1 dt dN2 = r2 N2 dt 1 1 N1 + 12 N2 K1 N2 + 21 N1 K2 (4.28a)

(4.28b)

In these equations r1 and r2 are the logistic, population growth rate of species 1 and 2, respectively. K1 and K2 are the carrying capacity of both species. The term 12 N2 in equation (4.28a) can be thought of as the decrease in growth rate of species 1 due to the presence of species 2. The parameter 12 represents the per capita decline (per individual of species 2 ). Similarly, the term 21 N1 represents the decrease in growth rate of species 2 due to the presence of species 1 with the parameter 21 is the per capita decline (per individual of species 1 ). Under specic assumptions the Lotka-Volterra model can be shown to mechanistically represent competition for resources by two competing species, but it has mostly been studied without any reference to mechanisms. Its analysis can be carried out graphically and the results constitute an important part of competition theory. To analyze the properties of the Lotka-Volterra model we will carry out a phase-plane analysis of the model equations (4.28) by the following steps: determine the isoclines for both N1 and N2 and represent them in a (N1 , N2 )-phase-plane, determine the steady states as the intersection points of the isoclines,and determine the stability properties of these steady states as far this is possible using graphical methods. From equation (4.28a) we observe that dN1 /dt equals 0 yields the following expressions for the N1 -isoclines:

N1 = 0 and K1 N1 12 N2 = 0 Similarly, equation (4.28b) yields the following expressions for the N2 -isoclines:

(4.29a)

(4.29b)

N2 = 0 and K2 N2 21 N1 = 0

(4.30a)

(4.30b)

The rst isocline for both species 1 (4.29a) and species 2 (4.30a) simply are the N1 - and N2 axis, respectively. The second isocline for species 1 (4.29b) intersects the axis where N2 = 0

54
N2

HOOFDSTUK 4. COMPETING FOR RESOURCES

N2
I II

K2

K1 /12

K1/12

K2

K1 N2

K2/21

N1 N2
III

K2/21

K1

N1

IV

K2

K1 /12

K1/12

K2

K2 /21

K1

N1

K1

K2/21

N1

Figuur 4.6: Four isocline cases for the Lotka-Volterra competition model. Thick dashed lines indicate N1 -isoclines; thick solid lines represent N2 -isoclines. Cases are numbered according to the conditions (4.31) in the text. at the carrying capacity of species 1 N1 = K1 , while it intersects the axis where N1 = 0 at N2 = K1 /12 . Similarly, the second isocline for species 2 (4.30b) intersects the axis where N1 = 0 at the carrying capacity of species 2 N2 = K2 , while it intersects the axis where N2 = 0 at N1 = K2 /21 . One pair of these four intersection points lies on the N1 -axis, while the other pair lies on the N2 axis. This gives rise to four distinct possible constellations, depending on which of the two members of each pair is larger. The four cases are characterized by: K2 > K1 21 K2 K1 > 21 K2 K1 > 21 K2 > K1 21 K1 12

case I: case II: case III: case IV:

and and and and

K2 >

(4.31a) (4.31b) (4.31c) (4.31d)

K1 > K2 12 K1 K2 > 12 K1 > K2 12

Figure 4.6 illustrates these four possible cases of the isoclines in the Lotka-Volterra model. All isoclines are straight lines as could already be inferred from their mathematical expressions (4.29)

4.2. INTERSPECIFIC COMPETITION

N2
I

55
N2
II

K2

K1 /12

K1/12

K2

K1 N2

K2/21

N1 N2
III

K2/21

K1

N1

IV

K2

K1 /12

K1/12

K2

K2 /21

K1

N1

K1

K2/21

N1

Figuur 4.7: Four isocline cases with steady states and ow patterns for the Lotka-Volterra competition model. Thick dashed lines indicate N1 -isoclines; thick solid lines represent N2 isoclines. Cases are numbered according to the conditions (4.31) in the text. and (4.30). From Figure 4.6 we can deduce the possible steady states of the Lotka-Volterra model as the intersection points of a N1 -isocline and a N2 -isocline. All four cases allow the following three steady states:

Extinct state: N1 -only state: N2 -only state:

(N1 , N2 ) = (0, 0) (N1 , N2 ) = (K1 , 0) (N1 , N2 ) = (0, K2 )

(4.32a) (4.32b) (4.32c)

In the extinct state both species are absent. In the N1 -only state, species 1 reaches its carrying capacity, while species 2 goes extinct, while in the N2 -only state species 1 goes extinct and species 2 tends to its carrying capacity. The cases III and IV allow in addition to these three steady states in which at least one of the species is always absent, an internal steady state in which both species have non-zero abundance: K1 12 K2 K2 21 K1 , 1 12 21 1 12 21

Coexistence state:

(N1 , N2 ) =

(4.33)

56

HOOFDSTUK 4. COMPETING FOR RESOURCES

Figure 4.7 is almost identical to Figure 4.6, but indicates in addition the steady states for the Lotka-Volterra competition model and the direction elds in the phase plane. To draw these direction elds, observe that: in the absence of species 2, species 1 grows logistically to its carrying capacity K1 , in the absence of species 1, species 2 grows logistically to its carrying capacity K2 , across the N1 -isoclines the ow in the direction of the N1 -axis is always 0, i.e. the ow is always parallel to the N2 -axis (vertical direction in Figure 4.7), and across the N2 -isoclines the ow in the direction of the N2 -axis is always 0, i.e. the ow is always parallel to the N1 -axis (horizontal direction in Figure 4.7). In Figure 4.7 these ows across the N1 - and N2 -isoclines have been drawn for all four cases of the isocline conguration. From these ows across the isoclines the direction elds within each separate region in the phase plane can be worked out. For example, in case I it is obvious that all solution curves will eventually end up inbetween the two slanted lines that form the isocline for species 1 and 2, respectively. Within the area bounded by these two lines the ow is towards the top-left corner (see Fig. 4.7), which implies that all solutions will tend to the N2 -only state (N1 , N2 ) = (0, K2 ). Similarly, in case II all solution curves will also end up inbetween the two slanted lines, but the ow within the enclosed region is towards the bottom-right corner (see Fig. 4.7). In case II therefore all solutions will tend to the N1 -only state (N1 , N2 ) = (K1 , 0). The cases III and IV are more dicult to analyze, but from the ow patterns in Figure 4.7 the following observations can be made: in case III, both the N1 -only state (N1 , N2 ) = (K1 , 0) and N2 -only state (N1 , N2 ) = (0, K2 ) are stable. This can be seen from the fact in the neighborhood of these two steady states the ow is towards the steady state from all directions. Hence, after a small perturbation away from any of these two steady states the steady state will be approached anew. in case III, the internal steady state is a saddle point. This can be deduced from the fact that the direction eld points away from the steady state within the triangular regions bounded by the two slanted lines that represent the N1 - and N2 -isoclines. A small perturbation that would displace the state of the systems into one of these two triangular regions would hence increase with time and the system would move away from the internal steady state. Notice that such a perturbation into the lower triangular region (towards the N1 -only state (N1 , N2 ) = (K1 , 0)) will cause species 2 to go extinct, while a perturbation into the higher triangular region (towards the N2 -only state (N1 , N2 ) = (0, K2 )) will cause species 1 to go extinct. Also notice that after a perturbation which displaces the system away from the internal steady state, but not into one of the two triangular regions bounded by the N1 - and N2 -isoclines, the state of the system will approach the internal steady state anew, as the ow from those directions is towards the steady state. The internal steady state is hence a saddle point. in case IV, the ow around the internal steady state is always in the direction of the steady state (see Fig. 4.7). The internal steady state is hence stable. On the other hand, a small displacement of the system away from either the N1 -only state (N1 , N2 ) = (K1 , 0) or the N2 -only state (N1 , N2 ) = (0, K2 ) steady state will always grow, because the direction eld around these steady states points away from them. Both the N1 -only and the N2 -only steady state are hence unstable. Together, these deductions lead to the following outcomes of the competition between species 1 and 2:

4.2. INTERSPECIFIC COMPETITION case I: case II: Species 2 outcompetes species 1. Species 1 outcompetes species 2.

57

case III: Species 1 can outcompete species 2, but species 2 can also outcompete species 1. The outcome depends on the initial condition. case IV: Species 1 and 2 coexist.

Hence, only in case IV both species can coexist, but they will do so at abundances below their carrying capacity (see equation (4.33)). In all other cases, one species will outcompete the other. From the conditions (4.31) delineating the four dierent cases it can be seen that for case IV to occur the interspecic competition parameters 12 and 21 can not be large. If the two carrying capacities are equal K1 = K2 the condition (4.31d) simplies to 21 < 1 and 12 < 1 (4.34)

which states that the interspecic competition is less intense than the intraspecic competition (note the coecients of 1 in equation (4.28a) and (4.28b)). If in any way, interspecic competition is stronger than intraspecic competition, that is one of the species is more aggressively competing with the other than with its conspecics, one of the two species will be excluded. Since the model predicts that competitive exclusion (i.e. one species outcompeting the other and driving it to extinction) will be observed in many situations, an important topic in population biology is to search for explanations why so many, apparently competing species can coexist in the same habitat. In other words, an important eld of investigations in population biology addresses the question which mechanism(s) prevents competitive exclusion of competing species. Of course, the Lotka-Volterra model is very simplistic and its neglects many complicating factors, but its strong predictions, which hold true in some experimental situations (see the next section), have reshaped some preconceptions about coexistence and species interactions. The model has hence played an important role in rephrasing the type of research questions being investigated in communities of interacting species that apparently seem to persist without excluding each other.

4.2.2

Competition for resources

An extensive and important body of population biological theory concerns the competition for shared resources between dierent consumer species. This body of theory has been mainly developed by Tilman (1980, see also Tilman 1981, 1982) on the basis of competition experiments with dierent algal species in chemostats. The theory has been primarily developed using a graphical approach (Tilman 1980). Hence, a discussion of this body of theory is appropriate in the current chapter, which to a large extent depends on graphical methods as well. The focus of the following presentation of Tilmans competition theory will be on the ecological aspects of the theory, as opposed to its theoretical and mathematical issues. These theoretical and mathematical ins and outs will hence not be discussed in too much a depth. One consumer and one resource The basis of the theory is the interaction between consumer species and abiotic resources. Algae that use nutrients like phosphate and silicate for their growth are species to which the theory applies best and which hence have been used extensively to test and show the validity of the theory. The basic set of equations that is used to model this consumer-resource interaction is

58

HOOFDSTUK 4. COMPETING FOR RESOURCES Tabel 4.2: State variables and parameters of Tilmans resource competition model Symbol Description

State variables R(t) Available concentration of the resource N (t) Abundance of consumers Parameters S Maximum concentration of resource supplied to the habitat For a chemostat: the inuent concentration of resource F Flow or supply rate of the resource D K Q Maximum population growth rate of consumers Per capita death or mortality rate of consumers Half-saturation constant for consumers feeding on resource Amount of resource required to produce one consumer individual

similar to the system of equations (4.10), describing the growth of bacteria in a chemostat. The coupled system of ODEs central to the theory is the following: dN = dt R D K +R

(4.35a)

R dR = F (S R) Q N dt K +R

(4.35b)

In the above equations the state variable N represents the abundance of consumers. It is assumed that the growth of the consumers is limited by a single resource, of which the actual concentration is denoted by R. The parameter S represents the maximum nutrient concentration that is possible in the habitat, which in a chemostat is identical to the inow nutrient concentration of resource. The parameter F is the ow or supply rate of resource, which is expressed per unit volume habitat volume (see the discussion in section 4.1.2). The parameter Q indicates the amount of resource that is needed to produce a single consumer individual. Its inverse 1/Q is often referred to as the yield, since it represents the number of consumer individuals that a single unit of resource can yield. In the ODE (4.35a), the parameter represents the maximum population growth rate and D the per capita mortality or death rate of consumers. The realized population growth rate is assumed to depend on the resource concentration following a MichaelisMenten equation, R/(K + R). For very large values of the resource concentration this equation approximates the value 1. For a resource concentration equal to K it has the value of 1/2, which is the reason that K is referred to as the half-saturation constant. The Michaelis-Menten equation is a well-known relationship from enzyme kinetics. It represents a saturating production rate as a function of substrate concentration. Figure 4.8 illustrate the form of this Michaelis-Menten relation as a function of the resource concentration R. In chapter 6 we will see that the same function is widely used to model the predator functional response, which is the relationship between the amount of prey eaten by a single predator and the current prey abundance. In that chapter we will also discuss a mechanistic basis for this functional relationship. Table 4.2 summarizes and lists all parameters used in Tilmans competition model (4.35).

4.2. INTERSPECIFIC COMPETITION

R K +R

59

1/2

Figuur 4.8: Michaelis-Menten relationship as a function of resource concentration. The parameter K indicates the resource concentration at which the function reaches half its maximal value (which is 1). The properties of the single species version of Tilmans model (4.35) are relatively simple: For all initial conditions the state of the system will approach a stable equilibrium (in this chapter we will not proof the stability of this equilibrium). The resource concentration in this stable equilibrium can be inferred from ODE (4.35a) by equating its right-hand side to 0, yielding: N = 0 (4.36)

R =

DK D

(4.37)

The latter equilibrium value for the resource concentration R is the solution of the equation: If the value of R is too large, i.e. if R > S the consumer population can not persist and the extinct steady state N = 0 is stable. If, on the other hand, the maximum resource concentration in the habitat S is suciently large, i.e. if the habitat is suciently productive, the consumers can establish a stable population and control the resource concentration in the habitat at the concentration equal to R . The consumer abundance in this internal equilibrium can be derived from ODE (4.35b) by substitution of the value R = R . This leads to N = F (S DK/( D)) F (S R ) = DQ DQ (4.38) R = D K +R

This equilibrium consumer abundance, however, does not play a prominent role in Tilmans theory on resource competition and is hence hardly ever discussed. The crucial quantity is the equilibrium resource concentration R (eq. (4.37)). Hence, the theory is also referred to as the R -theory.

60 Two consumers and one resource

HOOFDSTUK 4. COMPETING FOR RESOURCES

Adding a second species of consumer to the consumer-resource model (4.35) relatively straightforwardly leads to the following system of ODEs: dN1 = dt dN2 = dt 1 R D1 K1 + R R D2 K2 + R N1 (4.39a)

N2

(4.39b)

dR R R = F (S R) Q1 1 N1 Q 2 2 N2 dt K1 + R K2 + R

(4.39c)

This model simply includes an additional consumer ODE for the second consumer species. The ODE itself is of identical form for both consumers, only the parameters are assumed to dier between the two consumer species. The abundance of the two consumers is now referred to as N1 and N2 , while all parameters bear an index 1 or 2 to indicate which consumer species they apply to. In the absence of the competing consumer species, each of the two species would establish a steady state population, as long as we assume that the habitat is suciently productive. Each of the two consumer species would impose an equilibrium resource concentration equal to its R value. These R values will be referred to as
R1

and
R2 ,

for species 1 and 2, respectively. Only in very exceptional cases, the two R values will be equal to each other. In general, either R1 or R2 will be the smaller of the two:
R1 < R2

or

R2 < R1

(4.40)

In the rst case, species 1 can sustain a stable population at a resource level that is too low for species 2 to persist. Hence, species 2 will go extinct. In the second case, the opposite occurs that species 2 drives species 1 to extinction due to its ability to persist at lower resource concentrations. In other words, it is the species with the lowest value of R that outcompetes the other consumers and drives to extinction. The latter result can be generalized to multiple species. If there are p consumer species competing for the same resource, they can be ordered and indexed on the basis of their R -values:
R1 < R2 < R3 < R4 . . . < Rp

Consumer species 1 will be able to establish and sustain a stable population at a resource level below the resource concentration that the other species need for persistence. Hence, species 1 will competitively displace all other consumer species. This result exemplies the competitive exclusion principle, which states that:

Competitive exclusion principle: p consumer species cannot coexist in a stable equilibrium state on fewer than p resources

4.2. INTERSPECIFIC COMPETITION

61

One consumer and two resources An extension of Tilmans model for resource competition to a single consumer feeding on two dierent resources is represented by the following system of ODEs: dN1 = dt 1 min R1 R2 , K11 + R1 K12 + R2 D1 N1 (4.41a)

dR1 = F (S1 R1 ) Q11 1 min dt dR2 = F (S2 R2 ) Q12 1 min dt

R1 R2 , K11 + R1 K12 + R2 R1 R2 , K11 + R1 K12 + R2

N1

(4.41b)

N1

(4.41c)

In these equations the abundance of consumers is represented by N1 (the subscript 1 is retained in these equations because of the models formulated in the previous and the following section, which both deal with two consumer species). R1 and R2 represent the current concentration of resource 1 and 2, respectively. The parameter F again represents the ow or supply rate of resources per unit volume habitat volume. It is assumed that F is the same for both resources, as would be the case if the model was to mimic the dynamics of competitors in a chemostat. The parameters S1 and S2 represent the maximum nutrient concentration of resource 1 and 2, respectively, that is possible in the habitat. In a chemostat it is identical to the inow nutrient concentration of resources. The parameters Q11 and Q12 indicate the amount of resource 1 and 2, respectively, that is needed to produce a single consumer individual, their inverse again represent the yield (see section 4.2.2). Note that in case a parameter carries two subscripts, the rst one refers to the consumer species, while the second one refers to the resource. Hence, Q12 represents the amount of resource 2 that a consumer individual of species 1 needs to produce an ospring. In the ODE (4.41a), the parameter 1 again represents the maximum population growth rate and D1 the per capita mortality or death rate of consumers. It is assumed that both resources are essential, which means that they are both needed for population growth and one resource can not be used to substitute the other. Essential resources for algae are, for example, light and carbon dioxide: both substances are required for photosynthesis and for algal growth. Neither of them can make up for the absence of the other. The opposite of essential resources are substitutable resources. Of two substitutable resources an organism will only need one to grow and reproduce. For many heterotrophic organisms fat and sugars are to a large extent substitutable resource as they both provide energy that is used for growing and reproducing. If two resources are truly essential and non-substitutable, the one of them that is shortest in supply, will be limiting population growth. In the above ODEs this expressed by the term R2 R1 , 1 min K11 + R1 K12 + R2 which represents the realized population growth rate. The fractions R1 /(K11 +R1 ) and R2 /(K12 + R2 ) are the Michaelis-Menten relationships relating the consumer population growth rate to the current concentration of resource 1 and 2, respectively. The population growth rate that is realized is the minimum of the two possible values (Liebigs law of the minimum). The parameters K11 and K12 represent the half-saturation concentration of resource 1 and 2, respectively for consumer species 1, i.e. those concentrations at which the fractions reach the value 1/2.

62

R2

R12

1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000 1111111111111111 0000000000000000

HOOFDSTUK 4. COMPETING FOR RESOURCES

R11

R1

Figuur 4.9: Zero net growth isoclines for a one consumer-two resource model. Resource combinations in the hatched area allow consumer growth, combinations in the unhatched cause consumer decline. The thick solid lines between these regions are the zero-growth isoclines.

The analysis of the model (4.41) is an important rst step for the case considered in the next section, where we will analyze the outcome of competition of two consumers for two essential resources. Hence, we also retained the subscript 1 to denote the consumer species. To infer the dynamics predicted by the system of ODEs (4.41), notice that for both resources it is possible to derive an R -value for consumer species 1:
R11 =

D1 K11 1 D1 D1 K12 1 D1

(4.42a) (4.42b)

R12 =

These R -value only dier in the parameters K11 and K12 . In other words, the dierences in R -value are solely determined by the anity of the consumer species for both resources. In steady state the per capita birth rate of the consumers should equal their death rate, which implies that the righthand side of the ODE (4.41a) should equal 0. Ignoring the trivial steady state in which no consumers are present and equating the righthand side of (4.41a) with 0 leads to the following steady state condition: 1 min R1 R2 , K11 + R1 K12 + R2 = D1 (4.43)

This steady state equation does not specify a single combination of R1 and R2 , but a large number of such combinations. More specically, this steady state condition determines two line segments that are perpendicular to each other and are located at R1 = R11 and R2 = R12 , respectively, as shown in Figure 4.9. The two thick solid lines in this gure are all combinations of R1 and R2 that fulll the steady state condition (4.43). The two perpendicular line segments are together referred to as the zero net growth isocline, which is hereafter abbreviated to ZNGI.

4.2. INTERSPECIFIC COMPETITION

63

R2
(S1 , S2 )

(R1 , R2 )

S1 R1 S2 R2 R12 Q11 Q12

R11

R1

Figuur 4.10: Steady state location in a one consumer-two resource model. The thick solid line indicates the zero net growth isocline (ZNGI) at which consumer growth equals 0. The point indicated by (S1 , S2 ) represents the supply point, i.e. the maximum nutrient concentrations of resource 1 and 2, respectively, that are possible in the habitat. In a chemostat these would equal the concentrations of the two resources in the inowing nutrient solution. The vector Q11 Q12 represents the consumption vector of the consumer species. This vector indicates the direction in which the resource state (R1 , R2 ) declines due to consumer feeding. Both resources increase due to resource inow or regeneration which is always in the direction of the supply point (S1 , S2 ) when considered from the current combination of resource concentrations (R1 , R2 ). Hence, the direction in which the resource state (R1 , R2 ) changes due to inow or regeneration is given by the supply vector S1 R1 . The steady state is the unique point on the ZNGI where the S2 R2 consumption and supply vector are pointing in exactly opposite direction.

For combinations of R1 and R2 in the hatched area above and to the right the ZNGI, the consumer population will grow, while for combinations of R1 and R2 below and to the left of the ZNGI (the unhatched area) the consumer abundance will decline. The steady state of the model is located somewhere at the border of the hatched and unhatched area where the consumer growth rate is 0. To determine the steady state of the model, we could of course equate the righthand sides of the ODEs (4.41b) and (4.41c) to 0 and combine this with the steady state condition (4.43). After some considerable algebraic labor we will then nd the steady state expressions for the consumer abundance and the two resource concentrations. However, there also exists a graphical approach to determine the location of the steady state, which will be discussed here as it ts in with the graphical and intuitive methods of analysis presented in this chapter. In this graphical analysis the consumer abundance does not play an important role and is not discussed any further. The main aim of the graphical analysis is to determine the location of the resource concentrations R1 and R2 in steady state. In other words, where on the ZNGI shown in gure 4.9 the steady state is actually located. Figure 4.10 shows the same ZNGI as presented in gure 4.9. In addition, Figure 4.10 indicates the supply point, which is the set of concentrations of resource 1 and 2 that can maximally occur

64

HOOFDSTUK 4. COMPETING FOR RESOURCES

in the habitat. This combination of resource concentrations equals (S1 , S2 ). It can only occur in model (4.41) in the absence of consumer, i.e. when N1 = 0. In case the model represents the growth of a consumer species in a chemostat, the supply point represents the concentrations of resource 1 and 2 in the nutrient stock solution with which the culture chamber is fed. The graphical approach to locating the steady state of model (4.41) now proceeds by considering the direction in which the current resource state (R1 , R2 ) would move through the (R1 , R2 )-phase space due to consumer feeding and due to resource inow or regeneration. Taking the ratio of the last term in equation (4.41b) and the last term in equation (4.41c) shows that the ratio between the feeding pressure by consumers on resource 1 and 2, respectively, always equals Q11 /Q12 . In other words, consumer feeding on both resources will force the current resource state (R1 , R2 ) into the direction of the vector Q11 (4.44) Q1 := Q12

This vector is referred to as the consumption vector of the consumer species. Higher or lower concentrations of resource 1 or 2 and higher or lower densities of consumers would only imply a scaling of the absolute resource consumption rate by consumers, but would not change the relative rate with which resource 1 is consumed in comparison with resource 2. Hence, consumer feeding will always force the current resource state (R1 , R2 ) into the direction of the consumption vector. The direction in which the current resource state (R1 , R2 ) will change due to inow or regeneration can be determined by taking the ratio between the rst term in equation (4.41b) and the rst term in equation (4.41c). This shows that inow or regeneration will always force the current resource state (R1 , R2 ) into the direction of the supply point (S1 , S2 ). This direction is represented by the supply vector: S 1 R1 (4.45) S 2 R2 (Note that the assumption that the ow rate F is equal for both resources is crucial to obtain this result). In steady state resource consumption should nullify resource regeneration. Hence, the steady state of model (4.41) can only occur at the unique resource state (R1 , R2 ) on the ZNGI, where the supply vector is pointing in exactly the opposite direction as the consumption vector (see Figure 4.10). For all other resource combinations (R1 , R2 ) on the ZNGI the two vectors would have slightly dierent directions and would hence never cancel. The consumer abundance in steady state will be such that the absolute consumption rate equals the absolute supply rate. Hence, if the concentration of resource 1 in steady state is denoted by R1 the steady 1 is given by: state consumer abundance N S 1 R1 N1 = F Q11 D1 (4.46)

The above analysis shows that an intuitive and graphical consideration of the equations can yield substantial insight into model dynamics, to the extent that it has been possible to graphically locate the steady state in a three dimensional ODE model. Of course, the above analysis does not yield any insight into the stability of the steady state determined. For that aspect of model analysis we will have to use the analytical approach that is the subject of chapter 5. Such a stability analysis would show that the steady state determined in this section is stable whenever it exists, i.e. whenever the supply point (S1 , S2 ) is located within the growth region (the hatched area in Figure 4.9). On the other hand, the one consumer-two resource model is just the preliminary step towards the analysis of a two consumer-two resource model, which has been the central piece of research on which Tilmans competition theory is founded. This two consumer-two resource model is discussed next.

4.2. INTERSPECIFIC COMPETITION Two consumers and two resources

65

The model of the previous section can be straightforwardly extended to include a second consumer species by adding an ODE analogous to equation (4.41a) for the abundance of species 2, N2 . In addition, the equations describing the dynamics of each resource concentration should be extended with a term that represents the feeding of the second consumer species on both resources. The modeling of these consumption terms will be analogous to the consumption of both resources by species 1. The resulting set of ODEs hence becomes: dN1 = dt dN2 = dt R1 R2 , K11 + R1 K12 + R2 R1 R2 , K21 + R1 K22 + R2

1 min

D1 D2

N1

(4.47a)

2 min

N2

(4.47b)

dR1 = F (S1 R1 ) Q11 1 min dt Q21 2 min dR2 = F (S2 R2 ) Q12 1 min dt Q22 2 min

R1 R2 , K11 + R1 K12 + R2 R2 R1 , K21 + R1 K22 + R2 R1 R2 , K11 + R1 K12 + R2 R1 R2 , K21 + R1 K22 + R2

N1

N2

(4.47c)

N1

N2

(4.47d)

In these equations the abundance of consumers of species 1 is again represented by N1 , while the abundance of species 2 consumers is represented by N2 . As before, R1 and R2 represent the current concentration of resource 1 and 2, respectively. The parameters Q21 and Q22 indicate the amount of resource 1 and 2, respectively, that is needed to produce a single consumer individual of species 2, while the parameter 2 represents the maximum per capita growth rate of species 2. The parameters K21 and K22 represent the half-saturation concentration of resource 1 and 2, respectively for consumer species 2, i.e. those concentrations at which the Michaelis-Menten equations R1 /(K21 + R1 ) and R2 /(K22 + R2 ), respectively, equal 1/2. All other parameters have identical interpretations as in model (4.41) for a single consumer feeding on two resources. The ODE (4.47b) describing the dynamics of species 2 consumers is fully analogous to the ODE (4.47a) for species 1. Moreover, the consumption terms representing the feeding of species 2 on both resource, given by Q21 2 min and Q22 2 min R2 R1 , K21 + R1 K22 + R2 R2 R1 , K21 + R1 K22 + R2 N2

N2 ,

respectively, are completely analogous to the corresponding terms modeling the feeding of species 1 on both resources. Determining the steady state in the 4-dimensional model (4.47) and analyzing its stability is a daunting, analytical task, which goes beyond the purpose of the current discussion. Full details of these computations can be found in Tilman (1980). However, as in the previous section a

66
R2

HOOFDSTUK 4. COMPETING FOR RESOURCES

R2

II

III IV

II

III

IV

V
R12 R12

Q2
R22 R11

Q1

VI
R22

Q1 Q2

VI

R21

R1

R11

R21

R1

Figuur 4.11: Steady state location in a two consumer-two resource model. Left: Stable equilibrium case. The thick solid line indicates the zero net growth isocline (ZNGI) for species 1. The thick, dashed line is the corresponding ZNGI for species 2. The internal steady state is located at the intersection point of these two ZNGIs. R11 and R12 are the zero growth resource concen tration of species 1 for resource 1 and 2, respectively, while R21 and R22 are the corresponding values for species 2. In the steady state point the two consumption vectors, Q1 for species 1 and Q2 for species 2, are drawn, as well as two separation lines through the steady state point with the same direction as these consumption vectors. For supply points (S1 , S2 ) in region I both species will go extinct. For (S1 , S2 ) in region II species 1 will outcompete species 2, while for (S1 , S2 ) in region VI the opposite occurs. In region III, IV and V the resource supply (S1 , S2 ) is suciently high for both consumer species to persist in the absence of the other. However, in region III species 1 will outcompete species 2, while in region V the opposite occurs. On for supply points (S1 , S2 ) in region IV stable coexistence of both consumer species is possible. Right: Unstable equilibrium case. Completely analogous to the stable equilibrium case, except for supply points (S1 , S2 ) in region IV . Here, species 1 outcompetes species 2 or vice versa, dependent on initial conditions. Notice the ordering of the consumption vectors Q1 and Q2 in this case.

graphical analysis of the current model in the (R1 , R2 )-phase plane again allows us to locate the steady state and to gain substantial insight into the dynamics of the model. When considering the zero net growth isoclines (ZNGIs) for both species in the (R1 , R2 )-phase plane, it is clear that stable coexistence of both species is only possible if these ZNGIs intersect (see Figure 4.11). If the ZNGI for species 2 would be located completely to the right and above the ZNGI for species 1, the latter would outcompete the former for all possible values of the supply point (S1 , S2 ). On the other hand, if the ZNGI for species 1 would be located completely to right and above the ZNGI for species 2, the opposite would occur. This immediately determines the only possible steady state in the two-consumer-two resources model in which both species may coexist (see Fig. 4.11). In addition, this coexistence equilibrium is only feasible if the supply point (S1 , S2 ) is to the right and above the ZNGI of both species, i.e. if (S1 , S2 ) in region III, IV or V in Fig. 4.11. For supply points (S1 , S2 ) in region I, the maximum concentrations of both resources are insucient for both consumers to persist. For supply points (S1 , S2 ) in region II the maximum concentrations of both resources are sucient for consumer species 1, but insucient for species 2 to persist. For supply points (S1 , S2 ) in region VI, the opposite is

4.2. INTERSPECIFIC COMPETITION true (species 2 can persist, species 1 not).

67

If the two consumer species would coexist in a stable equilibrium, their combined consumption simply equals the sum of their individual consumptions. Hence, in a coexistence equilibrium the consumption vector for both consumer species together is a linear combination of their individual consumption vectors, where each individual consumption vector may be multiplied by a strictly positive constant. This implies that the combined consumption vector has a direction that is bounded by both individual consumption vectors Q1 and Q2 (see Fig. 4.11). Since consumption must cancel resource regeneration in steady state, it follows that only resource supply points (S1 , S2 ) in region IV may lead to coexistence of both consumers. Only for points in this region S R the supply vector S1 R1 in the steady state (R1 , R2 ) can be compensated by a combined 2 2 consumption vector that is within the region bounded by the individual consumption vectors Q1 and Q2 and that hence constitutes a linear combination of these two vectors. For supply S R points (S1 , S2 ) in region III or V the supply vector S1 R1 can never be compensated by a 2 2 consumption vector that is a linear, strictly positive combination (i.e. with positive coecients) of Q1 and Q2 (see Fig. 4.11). For supply points (S1 , S2 ) in region III, species 1 will outcompete species 2. Note that for supply points in this region a steady state with only consumer species 2 present would always be located at the vertical part of the ZNGI for this species (remember how the one-consumer-two-resources steady state was located in section 4.2.2). This vertical part is, however, located in the growth region of species 1. Species 1 would hence be able to invade such an equilibrium and subsequently outcompete species 2. An analogous argumentation holds for a species 1-only steady state and supply points (S1 , S2 ) in region V. A species 1-only steady state would be located at the horizontal part of its ZNGI which is within the growth region of species 2. For (S1 , S2 ) in this region species 2 therefore always outcompetes species 1. Even if an equilibrium with two coexisting consumer species is feasible, it is not necessarily a stable equilibrium. Determining the stability of the two-consumer equilibrium shown in Figure 4.11 can only be achieved by means of a far from trivial analysis. Tilman (1980) has carried out this local stability analysis for the coexistence equilibrium and has derived a set of necessary conditions for it to be stable. For a two-consumer-two resource equilibrium to be locally stable, it is necessary that the following conditions hold: For two consumers to stably coexist on two resources, each consumer species must, relative to the other, consume proportionally more of the one resource which more limits its own growth rate. The amounts of each resource consumed by individuals of each species may change only slightly in response to small changes in the availability of each resource. The rst of these conditions is graphically represented in the two panels in Figure 4.11. The left panel of this gure represents a stable two-species equilibrium, while the right panel represents an unstable equilibrium. For supply points (S1 , S2 ) in region IV (right panel, Fig. 4.11) the coexistence equilibrium is unstable and species 1 will either outcompete species 2 or species 2 will outcompete species 1. The precise outcome is dependent on the initial abundances of both consumers and both resources. Hence, both single-species steady states are stable, while the coexistence equilibrium is a saddle point. From Figure 4.11 it is possible to understand that the arrangement of the two consumption vectors Q1 and Q2 which is shown in the left panel of the gure implies that coexistence of both species will occur. For supply points in region IV, a species 2-only equilibrium is unstable, because it would be located at the (vertical) part of the ZNGI which is inside the growth region of species 1 (remember how the one-consumer-tworesources steady state was located in section 4.2.2). Similarly, a species 1-only equilibrium is unstable, because it would be located at the (horizontal) part of the ZNGI which is inside the

68

HOOFDSTUK 4. COMPETING FOR RESOURCES

growth region of species 2. Because both single-species steady states are unstable, it follows that coexistence is at least ensured. Moreover, consider the system to be in steady state. If the abundance of species 2, N2 , would be slightly increased, it would lead to a stronger decline in resource 1 if the consumption vectors Q1 and Q2 are ordered as in the left panel of Figure 4.11. Resource 1 is in this case the resource for which species 2 is the worst competitor, meaning that the increase of species 2 lead to conditions that favor species 1. Hence, the latter species can regain the balance which was perturbed with the increase in species 2. On the other hand, with the arrangement of consumption vectors Q1 and Q2 as in the right panel of Figure 4.11 a similar increase in species 2 would lead to conditions that favor itself, causing an even further deviation from the steady state. Even though this argumentation is by no means a full proof of the stability of the steady state, the rst condition for stability listed above is intuitively understandable. In contrast, the second condition for stability of the coexistence equilibrium follows from the mathematical analysis, but can not be motivated otherwise. It is hence less intuitive. The theory discussed above has been largely developed by Tilman (1980, see also Tilman 1981, 1982). Most importantly, he has tested the predictions of the theory with an extensive set of experiments with dierent diatom species competing for phosphate (PO4 ) and silicate (SiO2 ). For diatoms both nutrients are essential as they need silicate to build their hard outer wall and phosphate for photosynthesis. Figure 4.12 summarizes the results for competition experiments between Asterionella formosa and Cyclotella meneghiniana. For these two species coexistence is possible, but only for particular values of the resource supply (S1 , S2 ). The observed outcomes of the experiments match the theoretical predictions remarkably well (see Figure 4.12 and its legend for further details). In addition, Figure 4.13 to 4.20 show further experimental results for the competition between four dierent diatom species (Fragilaria crotonensis, Asterionella formosa, Synedra liformis and Tabellaria occulosa), competing for phosphate and silicate. Figure 4.13 and 4.14 show the results of single species growth experiments that are used to estimate the model parameters for all 4 species. The estimates for these parameters are presented in Table 4.3. The Figures 4.15 to 4.20 show the results of all 6 pairwise competition experiments that are possible between these 4 diatom species. On the basis of the estimated parameters, it was predicted that Tabellaria occulosa would never be able to coexist with any of the other species. Moreover, this species can never outcompete any of the other species. The results in Figure 4.16, 4.19 and 4.20 rigorously conrm these predictions. In many of the presented cases even the observed dynamics of the two competing species is accurately predicted by the model (see the gures). The two species Asterionella formosa and Fragilaria crotonensis turned out to have virtually identical R -values for silicate (see Table 4.3) and in addition have R -values for phosphate that dier only insignicantly from each other. Hence, the prediction is that either both species should be able to coexist for all possible values of the resource supply or Asterionella should win the competition on the basis of its slightly lower R -value for phosphate. Figure 4.17 that both species coexist for all supply conditions that have been examined. Since the characteristics of Asterionella formosa and Fragilaria crotonensis are so similar, the predictions regarding the competition of each of these two species with Synedra liformis are virtually identical(see Figure 4.15 and 4.18): Synedra has the lowest R -value for phosphate of all 4 diatom species and hence is predicted to outcompete all other species if the supply rate of phosphate is low. On the other hand, in comparison with Asterionella and Fragilaria Synedra has a high R -value for silicate, which leads to the prediction that Asterionella and Fragilaria will outcompete Synedra under low silicate supplies. In addition, there is a large region of resource supply values for silicate and phosphate that are predicted to lead to coexistence

4.2. INTERSPECIFIC COMPETITION

69

between Asterionella and Synedra and between Fragilaria and Synedra. Figure 4.15 and 4.18 conrm the theoretical predictions remarkably well and once more show that even the observed dynamics in the experiments is reasonable well matched by the predictions of the model. The combination of the convincing experimental test results and the logic behind the theory on resource competition provide a strong basis to the insights that have been revealed by studying the models in the last sections. In addition to the experimental results discussed above Tilman (1980) has extended the theory even further to include also substitutable and partly substitutable resources. For partly substitutable resources the ZNGIs will not consist of two perpendicular line segments any longer but will have a more continuous shape. As shown in Figure 4.21, this generically leads to 4 possible arrangements of the ZNGIs that strongly resemble the 4 cases of competitive outcome predicted by the Lotka-Volterra competition equations (compare Figure 4.21 with Figure 4.7). More generally, even with the perpendicular shape of the ZNGIs for essential resources, exactly 4 dierent arrangements of the ZNGIs are feasible and these are completely analogous to the 4 dierent cases of competitive outcome predicted by the Lotka-Volterra model for interspecic competition, which does not explicitly account for resources.

70

HOOFDSTUK 4. COMPETING FOR RESOURCES

Figuur 4.12: Predicted and observed outcomes of competition for phosphate and silicate by Asterionella formosa and Cyclotella meneghiniana. The mortality (ow) equals 0.25 day1 . The two-species equilibrium point occurs at 1.9 M SiO2 and 0.2 M PO4 . The consumption vectors (from Tilman 1977) have a slope (Si/P) of 87 for Asterionella and 6.2 for Cyclotella. For resource supply points in the region labeled 3, Asterionella should be dominant. For resource supply points in region 4, both species should coexist. Cyclotella should be dominant in region 5. Experiments (Tilman 1977) for which Asterionella was dominant are shown with an asterisk; those for which Cyclotella was dominant are shown with a diamond, and those for which both species coexisted are shown with a dot. A supply point o the graph (9.8 M PO4 , 15 M SiO2 ) was dominated by Cyclotella, as predicted. For this analysis, the observed maximal growth rates, reported in Tilman & Kilham (1976), were used even though the maximal rates under PO4 and SiO2 limitation for each species diered. The boundaries shown dier slightly from those of Tilman (1977) because of this dierence in maximal rates. (Figure 10 with legend from Tilman (1980)) .

4.2. INTERSPECIFIC COMPETITION

71

Figuur 4.13: Single-species Monod growth experiments for four diatoms under conditions of limited silicate. Curves were tted using a nonlinear regression. (Figure 2 with legend from Tilman (1981)) .

Figuur 4.14: Single-species Monod growth experiments for four diatoms under conditions of limited phosphate. Curves were tted using a nonlinear regression. (Figure 3 with legend from Tilman (1981)) .

72

HOOFDSTUK 4. COMPETING FOR RESOURCES

Tabel 4.3: Silicate and phosphate parameters for four Lake Michigan diatoms. K is the half saturation constant; r the maximal growth rate; Q the quotient; and R the calculated amount of nutrients required to grow at D = 0.25 d1 . The 95% condence intervals are in parentheses. (Table 1 with legend from Tilman (1981))
Species r (d1 ) K (mol/L) Q (mol/cell) R (mol/L)

Silicate limited experiments Fragilaria Asterionella Synedra Tabellaria 0.62 (.54-.70) 0.78 (.72-.84) 1.11 (.87-1.36) 0.74 (.44-1.04) 1.5 (.72.5) 2.2 (1.6-2.9) 19.7 (12.7-30.3) 19.0 (9.0-41.7) 9.7 107 1.5 106 5.8 106 6.3 106 1.0 (.7-1.5) 1.0 (.8-1.3) 5.7 (4.0-8.3) 9.7 (5.3-23.0)

Phosphate limited experiments Fragilaria Asterionella Synedra Tabellaria 0.80 (.72-.88) 0.59 (.53-.64) 0.65 (.61-.69) 0.36 (.30-.43) 0.011 (0-.024) 0.006 (.002-.011) 0.003 (0-.015) 0.008 (0-.04) 4.7 108 2.6 108 1.1 107 1.9 107 0.005 (.002-.008) 0.004 (.003-.007) 0.002 (.001-.006) 0.02 (.006-.07)

4.2. INTERSPECIFIC COMPETITION

73

Figuur 4.15: Predicted and observed outcomes of competition between Asterionella formosa and Synedra liformis. (A): predicted outcomes of resource competition between Asterionella (Af) and Synedra (Sf). (B) Observed results of competition experiments for Case 1 (low silicate) are shown with stars for Af and dots for Sf. The continuous thick line, labeled Af, and the continuous thin line, labeled Sf, show the predicted population dynamics. (C) As (B), but for Case 2 (intermediate concentrations of silicate and phosphate). (D) As (B), but for Case 3 (low phosphate concentrations). (Figure 4 with legend from Tilman (1981)) .

Figuur 4.16: Predicted and observed outcomes of competition between Asterionella formosa and Tabellaria occulosa. Af, stars, thick lines: Asterionella; Tf, squares, thin lines: Tabellaria, using the notation of Fig. 4.15. (Figure 5 with legend from Tilman (1981)) .

74

HOOFDSTUK 4. COMPETING FOR RESOURCES

Figuur 4.17: Predicted and observed outcomes of competition between Asterionella formosa and Fragilaria crotonensis. Af, stars, thick lines: Asterionella; Fc, triangles, thin lines: Fragilaria, using the notation of Fig. 4.15. (Figure 6 with legend from Tilman (1981)) .

Figuur 4.18: Predicted and observed outcomes of competition between Synedra and Fragilaria. Sf, dots, thick lines: Synedra; Fc, triangles, thin lines: Fragilaria, using the notation of Fig. 4.15. (Figure 7 with legend from Tilman (1981)) .

4.2. INTERSPECIFIC COMPETITION

75

Figuur 4.19: Predicted and observed outcomes of competition between Synedra and Tabellaria. Sf, dots, thick lines: Synedra; Tf, squares, thin lines: Tabellaria, using the notation of Fig. 4.15. (Figure 8 with legend from Tilman (1981)) .

Figuur 4.20: Predicted and observed outcomes of competition between Fragilaria and Tabellaria. Fc, triangles, thick lines: Fragilaria; Tf, squares, thin lines: Tabellaria, using the notation of Fig. 4.15. (Figure 9 with legend from Tilman (1981)) .

76

HOOFDSTUK 4. COMPETING FOR RESOURCES

Figuur 4.21: The four cases of resource competition in a two consumer-two resource model. These four cases are directly analogous to the four cases of the Lotka-Volterra competition equations. The outcomes of competition are labeled consistently in all four parts of this gure: 1 = neither species able to survive for resource supply points in this region; 2 = only species A able to survive; 3 = species A competitively displaces species B; 4 = stable coexistence of both species; 5 = species B competitively displaces species A; 6 = only species B able to survive. Case 1: Because the ZNGI (zero net growth isocline) of species A is inside that of species B, species A will always competitively displace species B. Case 2: Because the ZNGI of species B is always inside that of species A, species B will always win in competition with species A. Case 3: This equilibrium point is locally stable. Any resource supply point in region 4 will lead to both species stably coexisting at the resource equilibrium point. Case 4: This two-species resource equilibrium point is locally unstable. Any resource supply point in the region labeled 3 or 5 will eventually result in the competitive exclusion of either species A or B. The outcome of competition in this region depends on the starting conditions. (Figure 6 with legend from Tilman (1980)) .

Hoofdstuk 5

Systems of ordinary dierential equations

In section 4.1.4 (see the discussion following equation (4.24)) it was already stated that graphical methods of analysis in general are not sucient to determine the stability of steady states in models that are phrased in terms of 2 or more systems of ODEs. Chapter 4 discussed several models of 2 ODEs that are exceptions, such as the bacterial chemostat model and the LotkaVolterra competition model, but in general graphical methods do not suce. In this chapter I will rst focus on the linearization of the dynamics in the neighborhood of steady states, much along the lines as followed in section 3.4. Subsequently, I will discuss how the linearized approximation to the model dynamics in the neighborhood of a steady state can be solved explicitly and what this implies for the stability of the steady state. These results allow a complete classication of the types of dynamics that may be observed in models that are formulated in terms of a system of 2 ODEs. Such systems are also referred to as planar ODE systems, as their state space constitutes a plane spanned by two axes. The approach in this chapter will be more formal and rigorous than in the previous chapter. Indeed, it will be a purely analytical approach. Finally, in section 5.6 I will discuss how the theory generalizes to models in terms of 3 or more ODEs. Because this entire chapter deals with theory that is based on the linearized approximation to the dynamics in the neighborhood of steady states, the theory derived from it is referred to as linear stability theory. Before proceeding, it should be stressed once more that the analysis in this chapter are mostly valid only in the very close vicinity of steady states and hence only provide information about the local stability of the steady state. This means that the analysis can only reveal what will happen after an innitesimally small displacement of the system away from its steady state. Only in specic cases this local analysis will also give insight in the global behavior of the model, i.e. the model dynamics far away from the steady state.

5.1

Computation of steady states

In section 4.1.4 it was explained that a system of 2 ODEs could also be represented in vector notation (see equation (4.19)). I will use here two arbitrary symbols, y(t) and z(t), respectively, to indicate the state variables in the model. Moreover, I will assume that the dynamics are described by the following system of ODEs: 77

78

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS

dy = F (y, z) dt dz = G(y, z) dt

(5.1a) (5.1b)

the precise form of the functions F (y, z) and G(y, z) I will not specify any further. I will only assume that these functions are continuous and that their derivatives with respect to both y and z exist to ensure that the set of ODEs actually species a unique solution for an particular initial conditions y(0) and z(0) (see also section 3.6). The system of ODEs (5.1) can be written in vector form by dening a vector x: x := and a vector valued function H(x): H(x) := as: F (y, z) G(y, z) (5.3) y z (5.2)

dx = H(x) (5.4) dt This latter equation is simply a shorter and more abstract way of writing the system of ODEs which describes the model dynamics. Rewriting the ODEs (5.1) into the form (5.4) is hence only a matter of denitions and notation. The advantage of the rewriting is that it allows an immediate extension to systems of more than 2 ODEs, as equation (5.4) could as well describe the dynamics of a model with 3 or more variables (see section 5.6). Steady states of the model, specied by equation (5.4), are always characterized by the fact that the value of the state variables do not change over time. Hence, all steady states are determined by the equality: H(x) = 0 (5.5) Notice that in our case of a model in terms of 2 ODEs this equation represents actually two equations (H was a vector function!). Because population dynamic models usually incorporate non-linear relations between state variables, this system of equations is in general non-linear as well and hence hard to solve explicitly. Numerically, it is often possible, though, to calculate the value of a steady state for a particular set of parameters. Lets assume that a particular steady state of the model (5.4) is indicated by x, such that H() = 0 x (5.6)

5.2

Linearization of dynamics

How do we describe the dynamics in the neighborhood of the steady state? In section 3.4 this was achieved by studying the fate of a small perturbation or displacement away from the steady state with the help of the linearized dynamics around the steady state. The linearized dynamics was obtained by replacing the right-hand side function with its Taylor expansion around the steady state. For systems of 2 ODEs we follow exactly the same procedure. Lets dene a small perturbation away from the steady state: x (t) := x(t) x (5.7)

5.2. LINEARIZATION OF DYNAMICS Notice that x (t) again is a 2-dimensional vector which could also be dened as: x (t) := y (t) z (t) = y(t) y z(t) z = y(t) z(t) y z

79

In this equation I have used y and z to denote the steady state values of the state variables in the model. Following the same procedure as with single ODEs, the dynamics of the displacement vector x (t) can be described by: dx (t) d (x(t) x) dx(t) d x dx(t) = = = dt dt dt dt dt (cf. equation (3.17)). Using the model equation (5.4) this equality can be rewritten as: dx (t) = H(x) = H( + x ) x dt (5.9) (5.8)

This latter ODE describes the dynamics in the neighborhood of the steady state exactly. Therefore, the ODE is just as (un)solvable as the ODE (5.4) we started out with. Analogous to the procedure for single ODEs, in which the dynamics in the neighborhood of a steady state was approximated by the Taylor expansion of the function f (N ) around N (see equation(3.13)), we replace the function H(x) by its Taylor expansion around x, as follows: H(x) = H() + H () x + H () 2 + O 3 x x x x x (5.10)

Approximation to rst order is achieved by dropping all terms in x with a power of 2 or higher (cf. equation (3.15)): H(x) H() + H () x x x (5.11)

Substitution of this equality into the ODE (5.9) and noting that the equilibrium condition (5.6) implies that the rst term on the right-hand side vanishes, leads to the following ODE describing the approximate dynamics of x : dx (t) = H () x (t) x dt (5.12)

In this derivation we have followed the exact same steps as were taken for the analysis of single ODEs. Although this is formally correct, H(x) is a vector-valued function and its derivative H () is not a mere number like the derivative f (N ) in equation (3.15). For a function of two x variables, such as F (y, z), one of the elements of H(x), the Taylor expansion in the neighborhood of a particular steady state (, z ) is dened as (cf. equation (3.13)): y F y F z

F (y, z) = F (, z ) + y

y= y z= z

y +

y= y z= z

z +

1 2F 2 y 2

y= y z= z

2 + y

2F yz

y= y z= z

y z +

1 2F 2 z 2

y= y z= z

2 + O(3 ) z

The last term O(3 ) represents all higher order terms in the expansion, i.e. all those which at least include a term 3 , 2 z , y 2 or 3 . The above expression is the generalization of z z y y

80

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS

the Taylor expansion given in equation (3.13) to functions with two arguments. Dropping all terms of order 2 and higher yields the following expression for the rst order Taylor expansion of F (y, z) in the neighborhood of a steady state (, z ): y F (y, z) F (, z ) + y F y y + F z z (5.13)

y= y z= z

y= y z= z

Similarly, the function G(y, z) can in the neighborhood of the steady state (, z ), be approximay ted by: G(y, z) G(, z ) + y G y y + G z z (5.14)

y= y z= z

y= y z= z

In vector notation the last two equations can be written as: F y + G y F z G z

F (y, z) G(y, z)

F (, z ) y G(, z ) y

y= y z= z

y= y z= z

y z

(5.15)

y= y z= z

y= y z= z

Comparing the last equation with equation (5.11) it is easy too see that the derivative H () x has to be interpreted as the matrix of partial derivatives, evaluated at the steady state: F y H () = x G y F z G z

y= y z= z

y= y z= z

(5.16)

y= y z= z

y= y z= z

We will denote this matrix of partial derivatives with the symbol A and its elements with a11 , a12 , a21 and a22 , respectively: a11 a12 a21 a22

The elements a11 , a12 , a21 and a22 are hence dened as: F y F z G y G z

A =

(5.17)

a11 = a12 = a21 = a22 =

y= y z= z y= y z= z y= y z= z y= y z= z

(5.18a) (5.18b) (5.18c) (5.18d)

5.3. CHARACTERISTIC EQUATION, EIGENVALUES AND EIGENVECTORS

81

The matrix of partial derivatives A is referred to as the Jacobian matrix (or in short Jacobian) of the function H(x), evaluated at the steady state x. The Jacobian plays an important role in the stability analysis of the steady state. In general, the dynamics of a tiny deviation x from a steady state x is thus described by the ODE: dx = A x dt (5.19)

Since the Jacobian matrix A is evaluated at the steady state x and thus constant, this ODE is linear in the perturbation x . It approximates the dynamics of the model in the neighborhood of the steady state by its linearized dynamics.

5.3

Characteristic equation, eigenvalues and eigenvectors

Having obtained a system of 2 linear ODEs describing the dynamics of small displacements y (t) and z (t), respectively, from the steady state (, z ), we analyse in this section whether y these displacements will ultimately grow or shrink in magnitude. If they shrink, the state of the system will ultimately return to the steady state value after a small perturbation. The steady state is in this case locally stable. On the other hand, if small perturbations to the state of the system in the close neighborhood of the steady state grow over time, the steady state is unstable. In section 3.4 we have already discovered that linear ODEs yield solutions of exponential form. For this reason, we analyze what results from substitution of an exponential trial solution v et (5.20)

into the linearized ODE (5.19) (cf. the derivation in equations (3.23)-(3.26)). In the above equation v refers to the initial perturbation (y (0), z (0) to the state of the system at t = 0. Since this initial perturbation does not depend on the time variable, the substitution of the trial solution into the ODE (5.19) yields: dv et = A v et dt v et = A v et Dividing by the common factor exp(t) yields the equation Av = v (5.23)

(5.21)

(5.22)

It should be noted that the quantity v cannot be canceled, like the common factor exp(t), since it represents a vector and accordingly Av represents a matrix multiplication. However, equation (5.23) can be rewritten as (A I) v = 0 in which the I represents the identity or unit matrix (5.24)

82

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS

Equation (5.24) is trivially satised when v = 0 (here, 0 refers to the null-vector (0, 0)). However, we are concerned with studying the fate of a perturbation away from the steady state of the system and hence we exclude this trivial case of a zero displacement. For non-trivial perturbations v, readers that are familiar with linear algebra will recognize that equation (5.24) can only be satised if det (A I) = 0 (5.25)

I =

0 1 . .. . . . . . . 0

.. . .. . .. .

.. .. .

0 . . . . . .

. 0

0 1

By substituting the expressions for the matrices A and I into the above equation, this equation can be rewritten as: a11 a21 a12 a22

and after computation of the determinant as

det

= 0

(5.26)

(a11 ) (a22 ) a12 a21 = 0

(5.27)

The latter equation can be rearranged to a quadratic polynomial in terms of the unknown quantity : 2 (a11 + a22 ) + (a11 a22 a12 a21 ) = 0 (5.28)

Readers that are not familiar with linear algebra may check that equation (5.24) corresponds to the following system of equations: (a11 ) y (0) + a12 z (0) = 0 (5.29a) (5.29b)

a21 y (0) + (a22 ) z (0) = 0

From these two equations the quantity z (0) can be eliminated by multiplying the rst and second equation with (a22 ) and a12 , respectively, and adding the results of these two operations. This leads to the following equation for y (0): ((a11 ) (a22 ) a12 a21 ) y (0) = 0 (5.30)

Since we have excluded the case of a zero initial perturbation, it is easily seen that this last equation is identical to equation (5.27), which was derived above. Equation (5.25) represents the characteristic equation of the model (cf. equation (3.26)). The characteristic equation determines the eigenvalues . Given that the characteristic equation (5.25) can be rewritten as a quadratic polynomial (cf. equation (5.28)), it is clear that there are two eigenvalues . These two eigenvalues we will denote by 1 and 2 , respectively, or by

5.3. CHARACTERISTIC EQUATION, EIGENVALUES AND EIGENVECTORS

83

i if the discussion applies to both of them. Generically, 1 and 2 have dierent values. If for particular choices of the model parameters they are identical, we will still consider them as two distinct entities, which only happen to coincide. As was discussed at the end of section 3.4 eigenvalues can be interpreted as characteristic growth rates. Their interpretation in the context of the 2-dimensional model is more dicult than in the case of a single ODE. In the latter, the perturbation away from the steady state is just a scalar quantity, for which the eigenvalue determined the rate of growth. In case of a 2-dimensional model, the eigenvalues 1 and 2 are not just the growth rate in the scalar quantities y (t) and z (t), respectively, but are both associated with a particular vector v, which is called the corresponding eigenvector . The eigenvector v determines a direction in the phase space spanned by the two variables y and z. The eigenvalues 1 and 2 represent the characteristic growth rate in the direction of their corresponding eigenvector, which we shall denote by v11 v21 , v1 = and v2 = v12 v22

respectively (see also section 5.4). The eigenvector vi associated with a particular eigenvalue i can be computed from the vector equation (5.24) (or equivalently the corresponding system of equations (5.29)). Inspection of this equation shows that we can multiply both sides of the equation with a constant without invalidating it. In other words, the equation does determine the direction of the eigenvector, but it does not determine its length or absolute magnitude. We are hence free to choose one of the elements of the vector vi equal to 1. For example, if the value of a12 = 0 and the value of vi1 is set to 1, we can solve equation (5.29a) for the other element vi2 of the eigenvector vi . Therefore, in the case of a 2-dimensional model the eigenvector vi corresponding to a particular eigenvalue i equals 1

if a12 = 0. In case a12 = 0 and a21 = 0 the eigenvector vi is given by i a22 a21 vi = 1

vi = i a11 a12

(5.31)

(5.32)

which can be derived from equation (5.29b) by substitution of vi2 = 1. Once the eigenvalues 1 and 2 and the corresponding eigenvectors v1 and v2 are found, we can write down a general solution for the linearized dynamics of the perturbation (y (t), z (t)) in the neighborhood of the steady state (, z ). Provided that 1 = 2 , any initial perturbation y (y (0), z (0)) can be expressed as a linear combination of the two eigenvectors v1 and v2 , say by y (0) = C1 v1 + C2 v2 (5.33) x (0) = z (0) In this equation C1 and C2 are constants that are unique for a specic initial perturbation (y (0), z (0)). From this decomposition it follows that the perturbation (y (t), z (t)) at time t is given by y (t) = C1 v1 e1 t + C2 v2 e2 t (5.34) x (t) = z (t)

84

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS

which is a shorthand version of the following: y (t) = C1 v11 e1 t + C2 v21 e2 t z (t) = C1 v12 e1 t + C2 v22 e2 t (5.35a) (5.35b)

Equation (5.34) and (5.35) follow from the so-called superposition principle, which holds for linear ODEs. This principle states that if two arbitrary functions k(t) and l(t) are both solutions of a particular, linear ODE, any linear combination p k(t) + q l(t) with p and q arbitrary constants constitutes a solution of the ODE as well. Assume that the ODE can be written as dN = N dt Since both k(t) and l(t) represent solutions, it follows that dk(t) = k(t) dt and dl(t) = l(t). dt

For an arbitrary linear combination h(t) = p k(t) + q l(t), we can derive d (p k(t) + q l(t)) dk(t) dl(t) dh(t) = = p +q dt dt dt dt Given that both k(t) and l(t) represent solutions, the right-hand side of this equation can be written as p k(t) + q l(t) = (p k(t) + q l(t)) = h(t) which shows that also the linear combination h(t) = p k(t) + q l(t) is a solution of the ODE. What follows from the superposition principle is that if a particular initial condition x (0) can be written as a linear combination C1 v1 + C2 v2 of two quantities v1 and v2 and we know the solutions of the ODEs that have v1 and v2 as initial condition, the solution x (t) can be expressed as the linear of combination of these particular solutions for v1 and v2 . Equation (5.34) expresses this result mathematically. The eigenvectors 1 and 2 and the corresponding eigenvectors v1 and v2 therefore completely determine the fate of any arbitrary perturbation.

5.4

Phase portraits of dynamics in planar systems

In the previous section we have derived the characteristic equation, which determines the eigenvalues pertaining to a particular steady state in a 2-dimensional model. It was argued that these eigenvalues and the corresponding eigenvectors completely determine the fate of any arbitrary perturbation in the neighborhood of a particular steady state x. In this section we will analyse the relation between the eigenvalues and the qualitative properties of the dynamics close to a steady state and how its stability properties depend on these eigenvalues. Based on equation (5.28) the characteristic equation can be expressed as 2 + = 0 in which the quantities and are dened as (5.36)

= a11 + a22 = a11 a22 a12 a21

(5.37) (5.38)

5.4. PHASE PORTRAITS OF DYNAMICS IN PLANAR SYSTEMS

85

if a12 = 0, and by

equals the trace of the Jacobian matrix A, while is identical to the determinant of A. Given a particular eigenvalue i its corresponding eigenvector is determined by a12 vi = (5.39) i a11 vi = i a22 a21 (5.40)

in case a12 = 0 and a21 = 0.

The roots i of the characteristic equation (5.36) are given by

1 = 2 =

+ 2 2

(5.41)

(5.42)

where the quantity equals the discrimant of the Jacobian matrix A: = 2 4 (5.43)

Clearly, if > 0 equation (5.41) and (5.42) specify two distinct, real-valued eigenvalues. In contrast, if < 0 both eigenvalues 1 and 2 are complex. In the next sections we will study these two situations separately. The case, in which = 0 and hence 1 equals 2 , will not be discussed any further here. With two identical eigenvalues 1 and 2 also the eigenvectors are identical. This non-generic case requires special treatment, which can be found in any text book on ODEs. Table 5.1 summarizes the most important properties of systems with two linear(ized) dierential equations.

5.4.1

Two real eigenvalues

Equation (5.41) and (5.42) show that both eigenvalues are real-valued, provided the discriminant of the matrix A is positive: Disc A = = 2 4 > 0 which always holds if the determinant of the matrix A is negative, < 0. If the eigenvalues are real, also both elements of the two corresponding eigenvectors are real-valued (see equation (5.39) and (5.40)). Figure 5.1 schematically depicts the arrangement of the eigenvectors in the phaseplane in the neighborhood of the steady state x. In the gure it is assumed that the steady state is located at the origin (0, 0), which can always be achieved through a translation of the axes. The phase-plane shown in gure 5.1 is therefore spanned by the perturbations in the variables y and z, y and z , respectively. Figure 5.1 shows that each of the eigenvectors determines a line in the phase-plane. Recall from equation (5.34) that the solution of the linearized system of equations in the neighborhood of the steady state can be written as x (t) = C1 v1 e1 t + C2 v2 e2 t

86

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS

Tabel 5.1: Important characteristics of systems consisting of two linear(ized) dierential equations Vector notation Linear(ized) equations dx = Ax, dt A= a11 a12 a21 a22 Algebraic notation dx1 = a11 x1 + a12 x2 dt dx2 = a21 x1 + a22 x2 dt = a11 + a22 = a11 a22 a12 a21 = 2 4 2 + = 0 1,2 = 2

Signicant quantities

Tr A Det A Disc A

Characteristic equation Eigenvalues Eigenvectors

det(A I) = 0 1,2 = Tr A Disc A 2

v1 and v2 , such that Avi = i vi

if a12 = 0: a12 , 1 a11 1 a22 , a21 a12 2 a11 2 a22 a21

if a12 = 0 and a21 = 0:

Solutions

x(t) = C1 v1 e1 t + C2 v2 e2 t

if a12 = 0: x1 (t) = C1 a12 e1 t + C2 a12 e2 t x2 (t) = C1 (1a11 )e1 t + C2 (2a11 )e2 t if a12 = 0 and a21 = 0: x1 (t) = C1 (1a22 )e1 t + C2 (2a22 )e2 t x2 (t) = C1 a21 e1 t + C2 a21 e2 t

in which C1 and C2 are constants determined by the initial perturbation. If the initial perturbation x (0) = (y (0), z (0)) is on the line spanned by v1 , the constant C2 equals 0. The dynamics of the perturbation over time are in this case described by x (t) = C1 v1 e1 t For any value of t x (t) is then a scalar multiple of v1 , which means that x (t) remains on the line spanned by this eigenvector. The scalar multiple is determined by the product of C1 and the last, exponential term in the equation above. If 1 is negative this scalar quantity will decay for large values of t, if 1 is positive it will grow. In the limit for t , the state of the system will hence approach the steady state x following the line spanned by v1 , if and only if

5.4. PHASE PORTRAITS OF DYNAMICS IN PLANAR SYSTEMS

87

z v2

v1 x y

Figuur 5.1: Geometric representation of two real-valued eigenvectors in a planar system. The steady state x is located at the origin (0, 0), possibly after translation of the axes. The phase plane is therefore spanned by the perturbations in the variables y and z, y and z , respectively. v1 and v2 represent the two eigenvectors corresponding to the two eigenvalues 1 and 2 , respectively. 1 < 0. Analogously, if the initial perturbation x (0) = (y (0), z (0)) is on the line spanned by v2 , the constant C1 equals 0 and the dynamics are described by x (t) = C2 v2 e2 t Also in this case, in the limit for t , the state of the system will approach the steady state x following the line spanned by v2 , if and only if 2 < 0. It follows that the dynamics of any perturbation that is initially on one of the lines through the steady state spanned by either v1 or v2 will stay on that line for all t and will approach or move away from the steady state, depending on whether 1 or 2 are negative or positive, respectively. Note that because terms like exp(t) only approach 0 and never become identical to it, the steady state can be approached for t , but will never be reached exactly.

Perturbations that are not initially part of one of the lines spanned by the eigenvectors tend to be curved, since their dynamics are described by a weighted sum of terms involving both eigenvectors and eigenvalues (see equation (5.34)). These dynamics will be dominated by the exponential term with the largest eigenvalue, because this term will grow most rapidly or decay most slowly. Hence, the trajectories in the phase-plane determined by such initial perturbations will curve toward the line spanned by the eigenvector, which is associated with the largest eigenvalue. Depending on the signs of 1 and 2 3 dierent cases can now be distinguished, which are illustrated with their ow patterns in gure 5.2. Case 1: Two negative eigenvalues: 1 < 0, 2 < 0

For both eigenvalues to be negative, we can infer from equation (5.41) and (5.42) that at least has to be negative as well, < 0 (Note that in this section we restrict ourselves to

88

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS

1 < 0, 2 < 0 Stable node

1 < 0, 2 > 0 Saddle node

1 > 0, 2 > 0 Unstable node

Figuur 5.2: Characteristic ow patterns in the neighborhood of the steady state with realvalued eigenvalues and eigenvectors. the case that > 0 and that hence represents a positive, real value). In addition, should be smaller in absolute value than , otherwise the numerator in equation (5.41), and thus 1 , will be positive. The inequality = 2 4 < || can clearly only be satised if > 0. Hence, the necessary and sucient conditions for 1 < 0 and 2 < 0 are: < 0, > 0 and = 2 4 > 0 (5.44) The resulting ow pattern in this case is illustrated in the left panel of gure 5.2. Along the lines that are spanned by both eigenvectors v1 and v2 the ow is toward the steady state x, since both exp(1 t) and exp(2 t) with negative 1 and 2 are decreasing functions of times. As a consequence, the ow from anywhere in the neighborhood of the steady state x is directed toward the steady state. Any arbitrary displacement of the state away from the steady state will ultimately decay over time and the steady state is locally stable. In this particular case, the steady state is referred to as a stable node or a sink , a term which more or less captures the characteristic ow pattern in its neighborhood. Case 2: One positive, one negative eigenvalue: 1 > 0, 2 < 0 or 1 < 0, 2 > 0

For one of the eigenvalues to be positive, the equations (5.41) and (5.42) show that it does not matter what the value of is, as long as the term is in absolute value larger than . If > || and < 0, equation (5.41) for 1 yields a positive value, while for > || and > 0 equation (5.42) for 2 results in a negative value. Therefore, the only condition apart from the condition > 0 to hold for this case is that > ||. From the analysis of the previous case we can immediately conclude that this inequality holds if and only if < 0. The necessary and sucient conditions to have one positive and one negative eigenvalue, either 1 < 0 and 2 > 0 or 1 > 0 and 2 < 0, hence are: < 0 and = 2 4 > 0 (5.45)

The ow pattern for the case 1 > 0 and 2 < 0 is shown in the middle panel of gure 5.2. Along the line spanned by the eigenvector v2 the ow is directed toward the steady state x because the factor exp(2 t) with 2 < 0 is decaying over time. In contrast, the ow along the line spanned by the eigenvector v1 is directed away from the steady state since the term exp(1 t) with 1 > 0 only increases with time. These dierent responses along the

5.4. PHASE PORTRAITS OF DYNAMICS IN PLANAR SYSTEMS

89

two eigenvectors result in a characteristic ow pattern in the neighborhood of the steady state: Trajectories are curving toward the steady state when starting close to the line spanned by v2 , but bend away from it when coming close to the steady state, leaving its vicinity following the line spanned by the eigenvector v1 that is associated with the positive eigenvalue 1 . As a result, the steady state is initially approached, but subsequently left behind. This bi-phasic response is only apparent for initial states close to the line spanned by the eigenvector that is associated with the negative eigenvalue. Initial states further away from it show this bi-phasic response less and less, such that initial states close to the line spanned by the the eigenvector that is associated with the positive eigenvalue, only show the phase in which the state moves away from the steady state. Because of this bi-phasic ow pattern in its neighborhood, a steady state x characterized by one positive and one negative eigenvalue is called a saddle node or saddle point. Since eventually perturbations to the state of the system will cause the state to leave its neighborhood, the steady state is unstable. Case 3: Two positive eigenvalues: 1 > 0, 2 > 0 This case is by and large the mirror image of case 1. For both eigenvalues to be positive, equation (5.41) (5.42) show that has to be positive, > 0 (Note again that > 0 and and that hence > 0). Like in case 1, should be smaller in absolute value than , otherwise the numerator in equation (5.42), and thus 2 , will be negative. For the case 1 it was already deduced that < || can only be satised if > 0. Hence, the necessary and sucient conditions for 1 > 0 and 2 > 0 are: > 0, > 0 and = 2 4 > 0 (5.46)

The right panel in gure 5.2 shows the characteristic ow pattern for this case in the neighborhood of the steady state x. It is analogous to the ow pattern around the sink discussed as case 1 (see left panel of gure 5.2) with only the direction of movement on all trajectories reversed. Along both lines spanned by v1 and v2 the ow is away from the steady state x, since both exp(1 t) and exp(2 t) with positive 1 and 2 are increasing functions of times. As a consequence, the ow from anywhere in the neighborhood of the steady state x is directed away from the steady state. Any arbitrary displacement of the state away from the steady state will ultimately grow over time and the steady state is unstable. Because all trajectories seem to originate from it, the steady state in this case is referred to as an unstable node or a source.

Summary: In the neighborhood of a steady state, which is characterized by two real eigenvalues 1 and 2 , these eigenvalues determine the rate with which the steady state is approached (in case of a negative eigenvalue) or left behind (in case of a positive eigenvalue) along the two characteristic directions, spanned by the eigenvectors v1 and v2 that are associated with 1 and 2 , respectively.

90

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS

5.4.2

Two complex eigenvalues

If the discriminant of the matrix A is negative, the factor in equation (5.41) and (5.42) and therefore both eigenvalues 1 and 2 have a complex value. Moreover, the two eigenvalues 1 and 2 are each other complex conjugate, because the same factor occurs in the expressions for the eigenvalues, but with a dierent sign. We can therefore express 1 and 2 as 1 = p + i q and 2 = p i q (5.49) in which i = 1 and both factors p and q are real-valued, scalar quantities, dened as p = q = 2 || = 2 (5.50) 2 (5.51) (5.48) Disc A = = 2 4 < 0 (5.47)

Even for more general models in terms of more than 2 ODEs it holds that if a particular eigenvalue is complex, its complex conjugate is an eigenvalue as well. Thus, complex eigenvalues always occur in conjugate pairs. For the discriminant of the matrix A to be negative, equation (5.47) shows that has to be larger than 2 /4, which implies that at least has to be positive. Substitution of equation (5.48) and (5.49) into equation (5.39) shows that the eigenvectors corresponding to 1 and 2 can be written as a12 a12 0 = + i v1 = (5.52) p + iq a11 p a11 q 0 = i v2 = p iq a11 p a11 q a12 a12 (5.53)

and

and

Note that the requirement for equation (5.39) to hold, i.e. a12 = 0, is always satised. Otherwise, for a12 = 0 the determinant of the Jacobian matrix A equals a11 a22 , which implies that 1 and 2 would both be real-valued and equal to a11 and a22 , respectively (see equation (5.36)-(5.38)). Dening two real, constant vectors r and c as a12 r = (5.54) p a11 0 c = , q (5.55)

respectively, the eigenvectors v1 and v2 can now be expressed as v1 = r + i c v2 = r i c (5.56) (5.57)

5.4. PHASE PORTRAITS OF DYNAMICS IN PLANAR SYSTEMS

91

z r

c x y

Figuur 5.3: Geometric representation of the two basis vectors in a planar system with complex eigenvectors. The steady state x is located at the origin (0, 0), possibly after translation of the axes. The phase-plane is therefore spanned by the perturbations in the variables y and z, y and z , respectively. r and c represent the real and complex part of the two eigenvectors v1 and v2 (see equation (5.56) and (5.57)) corresponding to the two eigenvalues 1 and 2 , respectively. The closed loop represents the trajectory determined by the two functions u(t) and w(t) (see equation (5.58) and(5.59)). The complex nature of the eigenvalues 1,2 = p iq and the eigenvectors v1,2 = r ic raise the question how equation (5.34) can still represent the general solution from any arbitrary point in the neighborhood of the steady state. Complex eigenvectors have no geometrical meaning, like in the case for real eigenvectors (see gure 5.1) and we can hence not draw initial points as linear combinations of them. How do we then represent such a particular perturbation in terms of the eigenvectors and what is there dynamics over time? To answer these questions we substitute the expressions for the eigenvalues and eigenvectors, 1,2 = p iq and v1,2 = r ic, respectively, into equation (5.34) and substitute for the exponential functions the identity e(p iq)t = ept (cos qt i sin qt) With these substitutions equation (5.34) can be rewritten as follows x (t) = C1 v1 e1 t + C2 v2 e2 t = C1 (r + i c) e(p + iq)t + C2 (r i c) e(p iq)t = ept {C1 (r + i c) (cos qt + i sin qt) + C2 (r i c) (cos qt i sin qt)} = ept {(C1 + C2 ) (r cos qt c sin qt) + i (C1 C2 ) (r sin qt + c cos qt)} If we dene two real, vector-valued functions u(t) and w(t) as u(t) = r cos qt c sin qt w(t) = r sin qt + c cos qt (5.58) (5.59)

92

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS

Re 1 = Re 2 < 0 Stable spiral

Re 1 = Re 2 = 0 Neutral center

Re 1 = Re 2 > 0 Unstable spiral

Figuur 5.4: Characteristic ow patterns in the neighborhood of the steady state with complex eigenvalues and eigenvectors. x (t) can hence be written as x (t) = ept {(C1 + C2 ) u(t) + i (C1 C2 ) w(t)} (5.60)

This equation leads to the conclusion that as long as we choose the constants C1 and C2 as each others complex conjugate, all complex terms in the expression for x (t) cancel. More specically, if we assume that C1 and C2 are dened as C1 = a ib , 2 and C2 = a + ib , 2

respectively, equation (5.60) simplies to x (t) = ept {a u(t) + b w(t)} (5.61)

in which both constants a and b as well as both vector-valued functions u(t) and w(t) are real. From equation (5.58) and (5.59) we can infer that for t = 0 u(t) evaluates to the vector r, while w(t) evaluates to c. These two vectors r and c constitute a set of independent basis vectors. By choosing the (real) constants a and b appropriately, any initial state in the neighborhood of the steady state x can be represented as a linear combination of them (see gure 5.3). Therefore, we can conclude that equation (5.61) represents the general form of the solution to the linearized dynamics (5.19) from any arbitrary initial point in the neighborhood of the steady state x. To investigate the characteristics of the general solution we note that the term exp(pt) represents a growing or decaying function of time, depending on the value of the (real) constant p (remember that p equals the real part of the eigenvalues 1 and 2 ). Equation (5.58) shows that the function u(t) evaluates to the vector r at t = 0, to the vector c for qt = /2, to the vector r for qt = , to the vector c for qt = 3/2 and once again to the vector r at qt = 2. In other words, the function u(t) describes a cyclic pattern in the phase-plane, cycling between the points specied by r, c, r and c with a periodicity equal to 2/q (see gure 5.3). Analogously, we can infer from equation (5.59) that the function w(t) evaluates to c at t = 0, to the vector r for qt = /2, to the vector c for qt = , to the vector r for qt = 3/2 and once again to the vector c at qt = 2. The function w(t) hence describes the same cyclic pattern in the phase-plane, cycling between the points specied by c, r, c and r with a periodicity equal to 2/q, but lagging behind a quarter period (i.e. with a phase shift equal to /2) in comparison to the function

5.4. PHASE PORTRAITS OF DYNAMICS IN PLANAR SYSTEMS

93

u(t). In terms of the phase-plane shown in gure 5.3, the two functions u(t) and w(t) describe the same closed trajectory over time, which is followed with a periodicity equal to 2/q. The dierence between the two functions is the lag between them of a quarter period. In addition to the 3 cases with real eigenvalues discussed in section 5.4.1, the conclusions of the analysis above allow us to distinguish 3 additional cases for the dynamics in the neighborhood of a steady state that is characterized by a complex set of eigenvalues and eigenvectors. The ow patterns in the neighborhood of the steady state are illustrated in gure 5.4. Case 4: Negative real parts: Re 1 = Re 2 < 0 If the real part of 1 and 2 , i.e. the constants p, is negative, the solution for x (t) is the product of the oscillatory function of time, given by a u(t) + b w(t) and the decaying exponential function of time ep t Therefore, the trajectories in the neighborhood of the steady state x circle around the steady state with a period that is determined by the factor 2/q, while the amplitude, i.e. roughly speaking the distance to the steady state, decreases at a rate p. These trajectories hence represent spiraling dynamics, approaching the steady state for t (see the left panel in gure 5.4). The steady state is clearly stable, as small perturbations will ultimately decay with time. This type of steady state is appropriately referred to as a stable spiral point. From equation (5.50) it is clear that this case occurs, if in addition to the condition (5.47), is negative. The necessary and sucient conditions hence are < 0 and = 2 4 < 0 (5.62)

Case 5: Positive real parts: Re 1 = Re 2 > 0 This case is the mirror image of the previous case. Therefore, the solution for x (t) is the product of the oscillatory function of time, given by a u(t) + b w(t) and ep t which represents an increasing function of time. The trajectories in the neighborhood of the steady state x circle around the steady state with a period that is determined by the factor 2/q, while the amplitude (distance to the steady state) increases at a rate p. Altogether, the trajectories represent spiraling dynamics, moving away from the steady state for t (see the right panel in gure 5.4). The steady state is unstable, as small perturbations will increase with time. This type of steady state is referred to as an unstable spiral point. The necessary and sucient conditions, for which this case occurs, are > 0 and = 2 4 < 0 (5.63)

94

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS Case 6: Zero real parts: Re 1 = Re 2 = 0 If the real part p of 1 and 2 equals 0, the amplitude of the oscillatory dynamics in the neighborhood of the steady state neither grows nor declines. The solution for x (t) hence equals a u(t) + b w(t) This function describes purely cyclic dynamics with a period equal to 2/q. The trajectories form closed loops around the steady state (see the middle panel in gure 5.4). The initial perturbation determines the constants a and b and hence which closed loop is followed. Since initial perturbations neither grow nor decline in magnitude, the steady state is neutrally stable. This type of steady state is referred to as a neutral center . The necessary and sucient conditions, for which this case occurs, are = 0 and = 2 4 < 0 (5.64)

This case is rather special and generically only occurs for limited sets of parameters. In essence, it is as special as the case, in which two real eigenvalues are equal, 1 = 2 . This latter case was not discussed in section 5.4.1, precisely because it only occurs for restricted parameter combinations and hence is not generic. The reason that we nonetheless distinguish and discuss the case of complex eigenvalues with zero real part follows from the fact that the classical Lotka-Volterra model for the interactions between predators and their prey yields precisely this class of dynamics (see section 6.1).

Summary: For a steady state which is characterized by two complex eigenvalues 1 and 2 , the imaginary part q determines the period 2/q of the oscillatory dynamics that occur in the neighborhood of the steady state. The real part p of the eigenvalues determines the rate of growth or decline of the amplitude of the oscillatory dynamics, which roughly speaking corresponds to the distance to the steady state.

5.5

Stability of steady states in planar ODE systems

In section 5.4.1 it was shown that in case the eigenvalues are real-valued, a steady state of a planar ODE system is stable if and only if both eigenvalues 1 and 2 are negative. In section 5.4.2 it was shown that in case the eigenvalues are complex, the steady state is a stable spiral if and only if the real parts of the two eigenvalues 1 and 2 , which constitute a complex, conjugate pair, is negative. To summarize we can conclude

5.5. STABILITY OF STEADY STATES IN PLANAR ODE SYSTEMS

95

= 2 4 = 0

(a)

(b)

(d)

(c)

(e)

(f )

(f )

Figuur 5.5: Summary of stability properties for planar ODE systems. The stability properties of the steady state are indicated with the characteristic ow patterns in the neighborhood of the steady state, as a function of the trace , the determinant and the discriminant of the Jacobian matrix A. Types of steady states: (a): Stable spiral; (b): Unstable spiral; (c): Neutral center; (d): Stable node; (e): Unstable node; (f ): Saddle node. Note that a neutral center only occurs for = 0.

Important: In a continuous-time model in terms of ordinary dierential equations, a steady state will be stable provided that the eigenvalues of the characteristic equation, which is associated with the linearized dynamics in the close neighborhood of the steady state, are all negative (if real) or have negative real part (if complex). That is, Re i < 0 for all i.

96

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS

On purpose the conclusion above is phrased without a reference to a 2-dimensional, planar system of ODEs. The conclusion generalizes without any change to models of higher, even innite, dimension. For a model in terms of 2 ODEs we derived in section 5.4.1 and 5.4.2 a complete set of necessary and sucient conditions determining whether the steady state was a stable or unstable node, a saddle node, a neutral center, a stable or unstable spiral point. As shown in gure 5.2 and 5.4 the type of steady state determined completely the ow of the trajectories in the close neighborhood of the steady state. The conditions were phrased in terms of 3 quantities: , the trace of the Jacobian matrix A, the determinant of A and = 2 4, its discriminant. Figure 5.5 summarizes these conditions by plotting in the plane spanned by the quantities and examples of the characteristic ow patterns in the neighborhood of a steady state. All types of steady states occur for measurable sets of (, )-combinations, except for neutral centers, which only occur on the -axis in gure 5.5, i.e. for = 0. The quantities and determine the values of the eigenvalues and the corresponding eigenvectors (see equation (5.41) and (5.42)). As we concluded before, these eigenvalues and eigenvectors determine completely the dynamics of the model in the neighborhood of the steady state.

5.6

Models with 3 or more variables

In general, any model specifying the dynamics of 3 or more variables can be abstractly represented in vector notation by the same equations as presented in section 5.1. Assuming that we are studying a model in terms of n variables, denoted by x1 , . . . xn , the dynamics can hence be described by the following, n-dimensional system of ODEs dx = H(x) dt (compare equation (5.4)) in which x represents an n-dimensional column vector and H(x) an n-dimensional vector-valued function: x1 . x = . , . xn H1 (x) . . H(x) = . Hn (x)

The steady state equation of such a model can again be represented by H() = 0 x (compare equation (5.5)), which is now a short-hand notation for a system of n algebraic equations. The linearization of the dynamics in the neighborhood of a steady state, which is described in section 5.2 for a 2-dimensional model, also generalizes to models of any dimension. Therefore, in the neighborhood of a particular steady state x, the dynamics can be approximated by the linearized dynamics, described by the ODE: dx = A x dt (5.65)

Here, x represents the (n-dimensional) perturbation to the steady state x. The Jacobian matrix A equals the matrix of partial derivatives of the function H(x), evaluated at the steady state x. A is therefore dened as

5.6. MODELS WITH 3 OR MORE VARIABLES Tabel 5.2: The Routh-Hurwitz stability criteria (following May 1974) n 1 2 3 4 5 Stability criteria c1 > 0 c1 > 0, c2 > 0 c1 > 0, c3 > 0, c1 c2 > c3 c1 > 0, c3 > 0, c4 > 0 c1 c2 c3 > c2 + c2 c4 3 1 ci > 0 for all i, c1 c2 c3 > c2 + c2 c4 , 3 1 (c1 c4 c5 )(c1 c2 c3 c2 c2 c4 ) > c5 (c1 c2 c3 )2 + c2 c5 3 1 1

97

a11 . .. A = . . . an1

a1n . = . . ann

Each element aij of the matrix A equals the partial derivative of the function Hi (x) with respect to the variable xj , evaluated at the steady state x. Substitution of exponential trial solutions of the form v et into the ODE (5.65) yields, in vector notation, the same characteristic equation as derived for a 2-dimensional model: det (A I) = 0 (5.67)

Hn x1

H1 x1 . . .

x= x

.. .

H1 xn . . . Hn xn

x= x

(5.66)

x= x

x= x

As before, this characteristic equation determines the eigenvalues of the model pertaining to a particular steady state. The matrices A and I are now n n-matrices, as opposed to the 2 2-matrices encountered in section 5.3. Therefore, computation of the determinant in the characteristic equation will eventually lead to a polynomial expression in of degree n, which can be written as n + c1 n1 + c2 n2 + + cn2 2 + cn1 + cn = 0 (5.68)

The n roots of this polynomial equation correspond to the eigenvalues. Note that the number of eigenvalues always equals the dimension of the model. For a 2-dimensional model we have found in section 5.4 that c1 = = (a11 + a22 )

c2 =

= a11 a22 a12 a21

Given a particular eigenvalue i its corresponding eigenvector vi can be computed from A vi = i v

98

HOOFDSTUK 5. SYSTEMS OF ORDINARY DIFFERENTIAL EQUATIONS

which is analogous to equation (5.23), as derived in section 5.3 for a 2-dimensional model. Apart from writing down the equations that determine the eigenvalues and eigenvectors of a model with more than 2 variables, further analysis of the stability properties of steady states is very dicult and often not possible analytically. However, as in the 2-dimensional model studied in the previous sections, the general solution of the ODE (5.65), describing the dynamics of the perturbation x over time, can be written as a sum of eigenvectors and exponentials of the eigenvalues (compare equation (5.34)):
n

x (t) =
i=1

Ci vi ei t

(5.69)

in which the constants Ci depend on the value of the initial perturbation x (0) (see equation (5.33)). Strictly speaking, the equation above only holds for the case that all eigenvalues, and hence all eigenvectors, are distinct, similar to the results derived for 2-dimensional models. The case where eigenvalues are not distinct is seldom of substantial practical interest in modeling since an innitesimal change in one of the matrix elements (for example, due to a very small change in one of the model parameters) will normally remove the problem. Generically, we may thus conclude from equation (5.69) that perturbations to a steady state x will decay and the steady state will thus be stable, provided that all eigenvalues of the matrix A have negative real parts. If any of the eigenvalues has a positive real part, the steady state is unstable and perturbations will grow with time. The roots of the polynomial (5.68) cannot be evaluated conveniently, unless n = 1 or n = 2. However, there are some very useful mathematical theorems which provide necessary and sucient conditions for all the eigenvalues of a matrix to have negative real parts, without involving explicit calculation of the eigenvalues (May 1974). These stability conditions are known as the Routh-Hurwitz criteria and are expressed in terms of the coecients ci . Table 5.2 lists these criteria for n = 1, . . . , 5. These conditions themselves are not very enlightening, but for small values of n the inequalities constitute one of the most powerful and widely used theoretical tools to analytically evaluate the stability of steady states.

Hoofdstuk 6

Predator-prey interactions
In 1926 the Italian mathematician Vito Volterra wrote a paper, entitled Fluctuations in the abundance of a species considered mathematically, which was inspired by a question from his son-in-law, Dr. U. dAncona. The question posed by Dr. dAncona was why the predatory sh species the Adriatic Sea had increased in abundance after sheries had ceased during World War I, while the opposite was true for the prey species these predators fed upon. Volterra (1926) developed a model to describe the interaction between a predator and a prey, based on the following assumptions: prey would multiply indenitely (i.e. grow exponentially) in the absence of predators, predators would go extinct if prey were absent, due to a lack of food, the proportional rate of increase of the prey decreases as the number of predators increases, and the growth rate of predators increases when the number of prey increases. Without stating any equations, his model analysis led Volterra (1926) to the following 3 conclusions: 1. The two species uctuate periodically in abundance, the period only depending on the coecients of increase and of destruction of the two species, and on the initial numbers of the individuals of the two species. 2. The average numbers of the two species tend to constant values, whatever the initial may have been, so long as the coecients of increase or of destruction of the two species and also the coecients of protection and attack remain constant. 3. If we try to destroy individuals of both species uniformly and proportionally to their number, the average number of individuals of the eaten species grows and the average number of the eating species diminishes. But increased protection of the eaten species increases the average numbers of both. Volterra (1926) closed his short paper with the modest statement: Seeing that a great number of biological phenomena are characteristic of associations of species, it is to be hoped that this theory may receive further verication and may be of some use to biologists. 99

100

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Despite the brevity of the paper and the modest closing sentence, Volterra established with his study a corner-stone for the theory about predator-prey interactions. The Lotka-Volterra predator-prey model, as it has been referred to since Volterras contribution, forms a basis on which most if not all models of such interactions have been founded. In addition, it has become clear that predator-prey interactions are one of the most important causes of oscillations in species abundance. Hence, not only the model that Volterra (1926) studied, but also the uctuations he reported are to the present day important focal points for studying the dynamics of interacting species. In this chapter the basic Lotka-Volterra predator-prey model will be discussed rst, together with some of the models derived from it. Subsequently, I will discuss the relevance of the model conclusions for natural systems and more specically the occurrence in natural systems of population cycles that might be the result of the interactions between predators and their prey.

6.1

The Lotka-Volterra predator-prey model

Taking the simplest set of ODEs consistent with his assumptions, Volterra (1926) studied the dynamics of the following model: dF = rF aF C dt dC = aF C C dt

(6.1a) (6.1b)

In these equations F and C represent the abundance of prey (food) and predators (consumers), respectively. The parameter r represents the exponential growth rate of prey in the absence of the predator, while represents the death or mortality rate of the predators in the absence of prey. Encounters between prey and predators are assumed to follow the mass action law (see section 2.5.3) and are hence proportional to the product of the abundances F and C. The parameter a represents the attack rate of predators, which equals the area or volume that a predator searches through during a single unit of time. The parameter represents the conversion eciency, i.e. the eciency with which predators convert consumed prey into ospring. The notation used in the above equations is not identical to the notation used by Volterra (1926), but chosen for agreement with variants of the model discussed below. Notice that only positive parameter values are meaningful. The nullclines of the Lotka-Volterra predator-prey model (6.1) where dF/dt = 0 are given by: F =0 and C= r a a (6.2)

while the nullclines where dC/dt = 0 are given by: C=0 and F = (6.3)

These nullclines imply that the model possesses two steady states, the trivial steady state: (F 1 , C 1 ) = (0, 0) and the internal steady state (F 2 , C 2 ) = r , a a (6.4)

(6.5)

6.1. THE LOTKA-VOLTERRA PREDATOR-PREY MODEL

101

The Lotka-Volterra predator-prey model makes two counterintuitive predictions regarding the internal steady state: The steady state value of the prey is independent of its own growth or mortality rate, but is completely determined by the characteristics of the predator, while increased prey growth rates or increased protection of prey against predation (i.e. lower values of a) only result in higher predator densities at the steady state. The strong, regulating coupling between prey and predator induces that at steady state the available prey has to be just sucient to yield a predator reproduction rate that balances its mortality. Hence, if predators experience a higher mortality a higher steady-state prey abundance is needed to balance it. On the other hand, in steady state there should be suciently many predators to keep the total prey production under control. If prey are less sensitive to predator attacks, this implies that there simply have to be more predators around to exert this control. Next, consider the stability of these two steady states. The Jacobian matrix with partial derivatives is given by: r aC aF J = (6.6) aC aF Evaluated at the trivial steady state, this yields: r 0 J1 = 0 and the following characteristic equation: (6.7)

(r )( ) = 0 The eigenvalues pertaining to the trivial steady state hence equal: 1 = r 2 = The corresponding eigenvectors are v1 = v2 = 1 0 0 1

(6.8)

(6.9a) (6.9b)

(6.10)

for 1 = r and

(6.11)

2 = . These values for the eigenvalues and eigenvectors imply that the trivial steady state will be stable against perturbations in the predator density, but unstable against perturbations in the prey density. It is therefore a saddle point for all values of the parameters. Substitution of the equilibrium densities (6.5) in the expression for the Jacobian matrix (6.6) yields for the internal steady state: 0 (6.12) J2 = r 0

102

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Predator density C
1 dC/dt = 0

r/a

dF/dt = 0

/(a)

Prey density F
Figuur 6.1: Solution curves in the phase plane of the Lotka-Volterra predator-prey model. Two isoclines where dF/dt = 0 and dC/dt = 0 are drawn (the isoclines coinciding with the x- and y-axis have been omitted). The isoclines intersect in the neutrally stable state. All solutions are closed elliptic curves around this state, each curve completely determined by the initial state (F0 , C0 ). Parameter values: r = 0.5, a = 1.0, = 0.5 and = 0.1. and the following characteristic equation: 2 + r = 0 (6.13)

The following pair of complex eigenvalues therefore governs the dynamics in the neighborhood of the the internal steady: (6.14) 1,2 = r i Because the two eigenvalues are pure imaginary for all values of the parameters the internal steady state is always neutrally stable. Due to the product terms aF C and aF C in the ODE for the prey and predator abundance, respectively, the model is non-linear. For a linear systems a steady state with purely imaginary eigenvalues is always a neutral center (see chapter 5). The non-linearities in the Lotka-Volterra model could theoretically imply that the steady state is a spiral point, from which trajectories would spiral away. It is possible, however, to show that the internal steady state is in fact a neutral center through the analysis of a dierential equation dF/dC that can be derived by dividing the left-hand sides and right-hand sides of the system of ODEs (6.1) rF aF C dF = (6.15) dC aF C C

6.1. THE LOTKA-VOLTERRA PREDATOR-PREY MODEL

103

Prey density F
1.0 0.8 0.6 0.4 0.2 0.0 0 10 20 30 40 50

Time
Figuur 6.2: Prey dynamics predicted by the Lotka-Volterra predator-prey model. The initial predator density has been taken equal to its internal steady state value for the used parameter values, C = 0.5. Dierent initial prey densities lead to solutions that dier in both period and amplitude. The vertical line indicates the oscillation period predicted by the pair of complex eigenvalues, = 2/ r Parameter values: r = 0.5, a = 1.0, = 0.5 and = 0.1. The variables in this latter ODE can be separated to yield: r aC aF dF = dC F C Integrating both sides from the abundances at time 0 to the abundances at time t yields:
F (t) F0

dF + F

F (t)

C(t)

a dF =
F0 C0

r dC C

C(t)

a dC
C0

Where F0 and C0 are the initial abundances of prey and predators, respectively, and F (t) and C(t) are the same abundances at time t. From this last equation it can be inferred that any solution curve (F (t), C(t)) obeys the equation: ln F (t) + aF (t) r ln C(t) + aC(t) = ln F0 + aF0 r ln C0 + aC0 (6.16)

In the (F, C)-phase plane this equation describes an elliptic curve through the initial state (F0 , C0 ). Hence, the internal steady state is not a spiral point, but a neutral center. All solution curves are closed orbits around this center. Figure 6.1 shows some possible solution curves. The initial state (F0 , C0 ) uniquely determines which elliptic curve is followed. The analysis thus shows that not only the steady state is neutrally stable, but that also all solution orbits are neutrally stable: any small perturbation to the system will force it out of the orbit around the steady state it is currently following into a new orbit. The imaginary eigenvalues (6.14) predict that the period of oscillations are proportional to ( r)1 . However, the real period of the oscillations also depends on the distance of the orbit to the steady state, as is shown in Figure 6.2. Only innitesimally close to the steady state is the period of oscillation exactly equal to ( r)1 , it increases with the distance from the orbit to the steady state.

104

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Nonetheless, a higher prey growth rate r or a higher predator mortality rate lead to more rapid oscillations in prey and predator abundance. The basic Lotka-Volterra predator-prey model thus makes two important predictions: Steady-state prey abundances are completely controlled by life-history characteristics of the predator, and Oscillations in both prey and predator abundance are the dominant type of dynamics expected. In the following two section we will see to what extent these conclusions are inuenced by the neutral stability of the internal steady state.

6.1.1

Incorporating prey logistic growth

A slightly more complicated and realistic version of the Lotka-Volterra predator-prey model (6.1) assumes that prey do not grow indenitely in the absence of predators, but will ultimately reach a maximum prey abundance. Hence, prey do nog grow exponentially, but follow, for example, a logistic growth equation, as in the following set of ODEs: dF = rF dt 1 F K aF C (6.17a)

dC = aF C C dt

(6.17b)

The only dierence of this set of equations with the standard model (6.1) is the term (1 F/K), characterizing the logistic growth process of prey up to a carrying capacity K in ODE (6.17a). The nullclines where dF/dt = 0 for this modied Lotka-Volterra predator-prey model are given by: F r 1 (6.18) F =0 and C= a K while the nullclines where dC/dt = 0 are the same as before (cf. equation (6.3)): C=0 and F = a (6.19)

These nullclines imply that the modied model possesses three steady states, the trivial steady state: (6.20) (F 1 , C 1 ) = (0, 0) a prey-only steady state: (F 2 , C 2 ) = (K, 0) and the internal steady state (F 3 , C 3 ) = r 1 , a a aK . (6.22) (6.21)

The prey abundance in the internal steady state is again completely determined by the parameters, i.e. the life-history characteristics, which pertain to the predator. The predator abundance in the internal steady is now also determined by the steady-state prey abundance and therefore indirectly dependent on the predator parameters as well. Nonetheless, the steady-state predator abundance again is just sucient to keep the total prey growth in steady state under control.

6.1. THE LOTKA-VOLTERRA PREDATOR-PREY MODEL The Jacobian matrix with partial derivatives of the modied model (6.17) is given by: r2 rF aC aF K aC aF

105

J =

(6.23)

The dynamic properties of the trivial steady state are therefore not aected by adopting a logistic growth process for the prey: the trivial steady state is again a saddle point for all values of the parameters, which is stable against perturbations in the predator density, but unstable against perturbations in the prey density (see the discussion following equation (6.7)). For the prey-only steady state the Jacobian evaluates to: r 0 aK aK

Evaluated at the trivial steady state, this yields the same Jacobian as for the basic Lotka-Volterra model (6.1) (cf. equation (6.7)): r 0 (6.24) J1 = 0

J2 =

(6.25)

and the corresponding characteristic equation: (r )(aK ) = 0 (6.26)

The dynamics in the neighborhood of the prey-only steady state is hence determined by two real-valued eigenvalues: 1 = r 2 = aK As long as, K < Kc := a (6.27a) (6.27b)

(6.28)

both eigenvalues are negative, which implies that the prey-only steady state is a stable node. If K is larger than the critical value of the prey carrying capacity Kc , one of the eigenvalues (2 ) becomes real, but positive, turning the prey-only steady state into a saddle point. From equation (6.22) it can be seen that the critical carrying capacity Kc equals the prey abundance in the internal steady state. For K < Kc the internal steady state is biologically uninteresting, since the steady state predator abundance adopts a negative value for a prey carrying capacity that low (see eq. (6.22)). Only if K > Kc the predator abundance in the internal steady state is positive and is the steady state biologically relevant. In other words, if the maximum prey abundance is lower than the abundance needed by the predator to survive (i.e. the steady-state abundance of prey), the internal steady state assumes negative and thus biologically uninteresting values. For these carrying capacity values the preyonly steady state is the only stable steady state, which is ultimately approached whatever the initial condition of the model. As soon as the internal steady state adopts positive and thus biologically relevant values, the prey-only steady state becomes a saddle point, which allows the predator population to establish itself.

106

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS internal steady state evaluates to: r aK = 0 r 1 aK + r 1 aK = (6.30) r + r 1 + aK aK

2

yielding the following characteristic equation: r aK

For K > Kc the Jacobian pertaining to the F3 aF 3 r K J3 = 0 aC 3

(6.29)

= 0

From the characteristic equation the eigenvalues pertaining to the internal steady state can be derived: 1 = 2 = 1 r 1 2 aK 2 1 1 r + 2 aK 2 r aK r aK
2

4r 1 4r 1

aK aK

(6.31a)

(6.31b)

Let refer to the expression below the square root sign: := is positive as long as r aK which can be simplied to: > 2 aK r or equivalently K < Ks := a 1 1 + 2 2 1+ r (6.33) 1+ r 1 (6.32)
2

r aK

4r 1

aK

+4

aK

4 > 0

Ks is the critical value of the prey carrying capacity, below which the quantity is positive. A positive value of implies that the eigenvalues 1 and 2 are both real-valued. Moreover, because K > Kc = /(a) the term 1 aK is always positive as well, which implies that both eigenvalues 1 and 2 are real-valued, but negative. For K < Ks the internal steady state is thus a stable node. On the other hand, if K > Ks the quantity is negative, which implies that the two eigenvalues 1 and 2 constitute a complex pair. Their real part, however, is negative (see eq. (6.31)) which leads to the conclusion that for K > Ks the internal steady state is a stable spiral. The rst term in parentheses in the expression for Ks (eq. (6.33)) equals the prey abundance in the internal steady state, F 3 , which is also equal to the critical carrying capacity Kc . The

6.1. THE LOTKA-VOLTERRA PREDATOR-PREY MODEL

Predator density C
1 dC/dt = 0

107

Predator density C
1 dC/dt = 0

r/a

r/a

dF/dt = 0

dF/dt = 0

/(a)

/(a)

Prey density F Predator density C

1 dC/dt = 0

Prey density F Predator density C

1 dC/dt = 0

r/a

r/a

dF/dt = 0

dF/dt = 0

/(a)

/(a)

Prey density F

Prey density F

Figuur 6.3: Solution curves in the phase plane of the Lotka-Volterra predator-prey model with logistic prey growth. Two isoclines where dF/dt = 0 and dC/dt = 0 are drawn (the isoclines coinciding with the x- and y-axis have been omitted). The isoclines intersect in the internal steady state. Top-left panel: K = 0.15; the internal steady state is biologically irrelevant, the prey-only equilibrium is a stable node. Top-right panel: K = 0.3; the internal steady state is a stable node. Bottom-left panel: K = 1.0; the internal steady state is a stable spiral. Bottomright panel: K = 10.0; the internal steady state is a stable spiral. Compare this last gure with Figure 6.1. Other parameter values: r = 0.5, a = 1.0, = 0.5 and = 0.1. second term in parentheses at least larger than 1 and increases with an increasing ratio r/. This leads to the following conclusions regarding the dynamics of the modied Lotka-Volterra model (6.17): For K < Kc the maximum prey abundance is insucient to allow predators to persist. The internal steady state adopts negative and hence biologically irrelevant values. The prey-only equilibrium is a stable node. For Kc < K < Ks the internal steady state is biologically feasible and is a stable node. The prey-only equilibrium is a saddle point. For K > Ks the internal steady state is still the only biologically feasible and stable steady state, but is has become a stable spiral. Hence, the approach to the steady state is always

108 oscillatory.

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Figure 6.3 shows solution curves in the (F, C)-phase-plane for a number of dierent values of K, illustrating the conclusions above. If K is large the approach to the internal steady state is oscillatory. Moreover, the larger the value of K, the smaller the real part of the (complex) eigenvalues. For larger K-values the oscillations around the internal steady state therefore damp out more slowly. With respect to the two important predictions that were derived from the basic Lotka-Volterra predator-prey model, the model with logistic prey growth does not change the prediction that steady-state prey abundances are completely controlled by life-history characteristics of the predator. However, incorporating a logistic prey growth stabilizes the internal steady state, such that oscillation are not to be expected any longer. Even for very larger values of the prey carrying capacity the internal steady state is ultimately approached, albeit slowly. The slightest amount of density dependence in the prey growth hence stabilizes the oscillations of the basic Lotka-Volterra model. Because a very small perturbation to the model structure (i.e. adding even a tiny amount of density dependence in prey growth) changes the neutral stability of the steady state and the neutrally stable oscillations, the basic Lotka-Volterra model is not considered structurally stable. Structural stability refers to the property of a model that a small change to its structure does not cause signicant changes in its predictions.

6.1.2

Incorporating a predator type II functional response

Both the basic Lotka-Volterra predator-prey model and its variant with logistic prey growth assume that the total predation of prey equals aF C This assumption implies that the feeding rate of a single predator equals aF , the product of the per-capita attack rate and the current prey abundance. Hence, according to this formulation predators never get satiated as they will eat more and more the more prey there are around: for innitely large prey abundances the predator feeding rate will also become innite. The amount of prey eaten by a single predator per unit of time is referred to as the predators functional response. The basic Lotka-Volterra assumes that the predator functional is a linear function of the prey abundance. This form is also known as a type I functional response: 1 (F ) = aF (6.34)

For larger values of the prey abundance this functional response is not very realistic, as predators get quickly satiated when prey availability is very high. Figure 6.4 shows examples for the waterea Daphnia pulex feeding on 3 dierent algal species. Even though there is substantial scatter between the measurements at each food concentration, these relations between feeding rates and food concentration show a uniform shape: at low food concentration the feeding rate increases rapidly, but levels o at higher food concentrations. Whether at high food densities the feeding rate remains at a constant, high level, continues to increase slowly or even decreases a bit is debatable, but the feeding rate certainly does not increase indenitely. A formulation of the predators functional response which accounts for the satiation of predators at high food densities is due to Holling (1959). This particular form of the functional response is referred to as Hollings type II functional response or simply a type II functional response. The most mechanistic mathematical formulation of the type II functional response is: 2 (F ) = aF 1 + ahF (6.35)

6.1. THE LOTKA-VOLTERRA PREDATOR-PREY MODEL

109

Figuur 6.4: Functional response of Daphnia pulex on three algal species. In spite of the dierent algal species and the scattering of the points the shape of the functional response as a function of food concentration is in principle uniform. (Vertical bars: 95% condence limits of estimates). Source: Figure 2 in Lampert (1977) in which the parameter a again denotes the rate at which a single predator searches for (i.e. attacks) prey whenever it is not currently consuming a prey item and the parameter h is the average time span a predator uses to consume a prey it has caught. The type II functional response as presented by equation (6.35) can be derived as follows: Consider the number of prey items X that a single predator can catch and consume within a total period of time T . Whenever the predator is searching for prey it is assumed to encounter new prey items at a rate aF just like the assumption in the basic Lotka-Volterra model. However, if the predator needs on average h time units to consume a prey item caught, it will only have T hX time available to eectively search for new prey items. Therefore, the total number of prey items X caught by a single predator during a total time period T obeys the following equality: X = aF (T hX) (6.36) The total number of prey items caught X can be solved from this equality as: X = aF T 1 + ahF

The type II functional response (eq. (6.35)) represents the rate at which the predator consumes prey items (i.e. the quotient X/T ), given that the predator needs on average h time units to handle each single item. The type II functional response is hence based on the assumption that at very high prey densities predators become limited by their handling time h. Indeed, for F the limit of the functional response equals the inverse of the handling time, 2 (F ) 1/h. Although the derivation presented here is based on an argument that the predator needs some

110

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

non-negligible amount of time to physically consume a caught prey item, the same mathematical relationship for the functional response can be derived by considering the digestion of consumed prey items the limiting step in the feeding process. For examples of such a derivation we refer to Metz & Diekmann (1986) and Claessen et al. (2000). The Holling type II functional response is mathematically identical to the Michaelis-Menten equation that was used to model the nutrient uptake by bacteria, which was used in Tilmans competition model (4.35). The form of the Michaelis-Menten equation was already presented in Figure 4.8. Indeed, Hollings type II functional response is also often presented in the following form: F (6.37) 2 (F ) = Im Fh + F in which the parameter Im represents the maximum predator ingestion rate and the parameter Fh represents the prey density at which the predator realizes 50% of its maximum food intake. This representation of the type II functional response is completely analogous with the MichaelisMenten function R K +R which was used to model nutrient uptake in Tilmans competition model (4.35). The two representations of the type II functional response are identical given the parameter equivalences: Im = and Fh = 1 h 1 . ah

The derivation of the type II functional response given above uses the time budget of the predator to determine its net rate of prey consumption. Such a derivation is not easily extendable to more complicated cases, where the predator is also involved in some other activities. For example, if one wants to derive the formulation of a predators type II functional response in case the predators next to catching and consuming prey also interfere with each other during the searching process, a time budget analysis is not feasible any longer. In these cases the functional response can be derived by formulating and analyzing a behavioral model that describes the changes in the predators state (i.e. whether it is feeding or consuming) on a short time scale. Here I illustrate this alternative procedure for the type II functional response derived above. Consider that a predator can be in two behavioral state: it is either actively searching for prey (state S) or it is consuming a prey item it has just caught (state H). Let the probability that the predator is in state S be given by Ps and the corresponding probability that the predator is in state H by Ph . On a short time scale we can assume that the prey abundance F does not change signicantly. If predators encounter prey at a rate aF whenever they are busy searching for it, the probability that a predator individual is actively searching decreases at a rate aF Ps . Similarly, when predators are busy consuming a caught prey item and they need on average h time units to consume a prey item, the probability that a predator individual is actively searching increases at a rate Ph /h (note that h is a time and hence 1/h represents a rate). Altogether the changes on a short time scale in the probability that a predator is actively searching, Ps , can be described by the following ODE: Ph dPs = aF Ps + dt h Because we assume predators are either searching or handling Ps + Ph = 1

6.1. THE LOTKA-VOLTERRA PREDATOR-PREY MODEL the ODE can be rewritten as

111

1 Ps dPs = aF Ps + dt h This behavioral model for the probability Ps that a predator is in the searching state S leads to a steady state value for Ps given by: Ps = 1 1 + ahF

Next we assume that the changes in the predators state (searching or consuming) are so rapid that on the population dynamical time scale which is relevant for births and deaths of predators, Ps is approximately equal to its steady state value P s . This assumption is called the pseudosteady-state assumption. As a consequence, in the population dynamic model we must take into account that the probability that a single predator is actively searching for prey equals P s and hence that its net rate of food intake equals aF P s = aF 1 + ahF

Again, we end up with the same formulation of the type II functional response as was derived before. The alternative derivation using a model for the changes in behavioral state of the predator seems more complicated than the time budget analysis used at rst. Its advantage is that it can also be applied if the predator can be in other behavioral states than just searching or consuming. Ruxton et al. (1992) have successfully used the behavioral model approach, for example, to derive various forms of the functional response in case predators waste time ghting with each other over food items. Such derivations are not possible using a time budget analysis. The predator-prey model that accounts for both logistic prey growth and a predator type II (or satiating) functional response can now be represented by the following system of ODEs: dF = rF dt F K aF C 1 + ahF

(6.38a)

dC aF = C C dt 1 + ahF

(6.38b)

This model is also referred to as the Rosenzweig-MacArthur model after two authors that have investigated the dynamics of this model in great detail (see, for example, Rosenzweig (1971)). The nullclines where dF/dt = 0 for the Rosenzweig-MacArthur, predator-prey model are given by: F r 1 (1 + ahF ) (6.39) F =0 and C= a K while the nullclines where dC/dt = 0 are given by: C=0 and F = a( h) (6.40)

Disregarding the two trivial nullclines F = 0 and C = 0, respectively, these expressions for the nullclines of the model indicate that the nullcline where dC/dt = 0 constitutes a vertical line in the (F, C)-phase-plane, as was the case for the two predator-prey models discussed before. On the other hand, the non-trivial nullcline where dF/dt = 0, described by the relationship: C := H(F ) = r a 1 F K (1 + ahF ) (6.41)

112

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Predator density C
1 dC/dt = 0

dF/dt = 0

r/a

a(h)

Prey density F
Figuur 6.5: Nullclines of the Rosenzweig-MacArthur, predator-prey model. Two isoclines where dF/dt = 0 (solid curves) and dC/dt = 0 (dashed curve) are drawn (the isoclines coinciding with the x- and y-axis have been omitted) for 4 dierent values of the prey carrying capacity K. The isoclines where dF/dt = 0 have increasing values of K from the left- to the right-most curve. Closed circles: Stable nodes; Open circles: Saddle points; Closed triangles: Stable spirals; Open triangles: Unstable spirals. is a quadratic function H(F ) of the prey abundance F . The function H(F ) adopts the value r/a for F = 0. Therefore, at negligible prey densities, the isocline has the same value of C as for the basic Lotka-Volterra model and the model with logistic prey growth (see eq. (6.2) and eq. (6.19)). The top of the parabola H(F ) occurs at F = Fm := As long as 1 2 K 1 ah (6.42)

1 ah the function H(F ) is therefore strictly decreasing for positive prey abundances F . It is obvious that H(F ) becomes zero when the prey abundance equals its carrying capacity, i.e. when F = K. Figure 6.5 illustrates the shape of these isoclines in the (F, C)-phase-plane for 4 dierent values of the prey carrying capacity K. Notice that the intersections of the nullclines represent the steady states of the model, whose stability will be considered next. K < The nullclines imply that the Rosenzweig-MacArthur model possesses three steady states as was the case for the Lotka-Volterra model with logistic growth. First of all, the trivial steady state: (F 1 , C 1 ) = (0, 0) (6.43)

6.1. THE LOTKA-VOLTERRA PREDATOR-PREY MODEL a prey-only steady state: (F 2 , C 2 ) = (K, 0) and the internal steady state (F 3 , C 3 ) = r (aK( h) ) , a( h) a2 ( h)2 K .

113

(6.44)

(6.45)

The prey abundance in the internal steady state is again completely determined by the parameters, i.e. the life-history characteristics, which pertain to the predator, while the predator abundance is determined by parameters, i.e. life-history characteristics, of both predator and prey. The Jacobian matrix with partial derivatives of the Rosenzweig-MacArthur model is given by: a aF rF r 2 K (1 + ahF )2 C 1 + ahF J = a aF 2 C 1 + ahF (1 + ahF )

(6.46)

Evaluated at the trivial steady state, this yields the same Jacobian as for the basic Lotka-Volterra model (6.1) (cf. equation (6.7)): r 0 (6.47) J1 = 0

The dynamic properties of the trivial steady state are therefore not aected by adopting a type II predator functional response: the trivial steady state is again a saddle point for all values of the parameters, which is stable against perturbations in the predator density, but unstable against perturbations in the prey density (see the discussion following equation (6.7)). For the prey-only steady state this Jacobian evaluates to: aK r 1 + ahK J2 = aK 0 1 + ahK and the corresponding characteristic equation: (r ) aK 1 + ahK = 0

(6.48)

(6.49)

The dynamics in the neighborhood of the prey-only steady state is hence determined by two real-valued eigenvalues, as was the case for the Lotka-Volterra model with logistic prey growth (cf. eq. (6.27)): 1 = r 2 = aK 1 + ahK (6.50a) (6.50b)

From these expressions it is easy to see that there is again a critical value of prey carrying capacity Kc , which equals the prey abundance in the internal steady state (cf. eq. (6.28)): Kc = F 3 = a( h) (6.51)

114

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

For K < Kc both eigenvalues are negative, which implies that the prey-only steady state is a stable node. If K is larger than the critical value of the prey carrying capacity Kc , one of the eigenvalues (2 ) becomes real, but positive, turning the prey-only steady state into a saddle point. As before, for K < Kc the internal steady state is biologically uninteresting, since the steady state predator abundance adopts a negative value for a prey carrying capacity that low (see eq. (6.45)). Only if K > Kc the predator abundance in the internal steady state is positive and is the steady state biologically relevant. In other words, if the maximum prey abundance is lower than the abundance needed by the predator to survive (i.e. the steady-state abundance of prey), the internal steady state assumes negative and thus biologically uninteresting values. For these carrying capacity values the prey-only steady state is the only stable steady state, which is ultimately approached whatever the initial condition of the model. As soon as the internal steady state adopts positive and thus biologically relevant values, the prey-only steady state becomes a saddle point, which allows the predator population to establish itself. These conclusions are completely analogous to the conclusions held for the Lotka-Volterra model with logistic prey growth (see the discussion following eq. (6.28) on page 105). By substituting F = F 3 and C = C 3 in the Jacobian matrix (6.46), the Jacobian of the internal steady state evaluates for K > Kc to: h + 2 r aK aK( h) (6.52) J3 = 0 r h aK This leads to the following characteristic equation: r h + 2 aK aK( h) + r h aK = (6.53) r + h aK = 0

2 r

h + 2 aK aK( h) 1 r 2 1 r 2

The two eigenvalues pertaining to the internal steady state can hence be represented as: 1 = h + 2 aK aK( h) h + 2 aK aK( h) + 1 2 1 2 (6.54a)

2 =

(6.54b)

where the quantity is dened as: := r 2 h + 2 aK aK( h)

2

aK

(6.55)

Since we only consider cases where K > Kc (see eq. (6.51)), it follows that h is positive and consequently that < r
2

aK

h + 2 aK aK( h)

(6.56)

6.1. THE LOTKA-VOLTERRA PREDATOR-PREY MODEL

115

From the expressions (6.54) for the eigenvalues 1 and 2 it can now be inferred that the stability of the internal steady state is entirely determined by the sign of the expression r h + 2 aK aK( h) (6.57)

If this expression is negative and the quantity is positive, both eigenvalues 1 and 2 are real-valued but negative, since < r h + 2 aK aK( h)

(refer to inequality (6.56)). In this case the internal steady state is a stable node. If, on the other hand, the quantity in expression (6.57) is negative and the quantity is negative as well, both eigenvalues 1 and 2 form a complex, conjugate pair with negative real part. The internal steady state is then a stable spiral. However, if the quantity in equation (6.57) changes from negative to positive, it is certain that is negative (at least initially; see the denition of in equation (6.55)). In this case, both eigenvalues 1 and 2 form again a complex, conjugate pair with a real part that now turns from negative to positive. The internal steady state therefore changes from a stable spiral point into an unstable spiral point. This change in stability occurs when the carrying capacity K becomes larger than the critical value Ks , dened as: Ks := 1 1 + h = + 2 = + 2F3 ah ( h) ah a ( h) ah (6.58)

In other words, the change in stability occurs when F 3 becomes smaller than 1 2 K 1 ah

which is exactly equal to the prey abundance F at which the prey isocline H(F ) reaches its maximum (refer to eq. (6.42)). In the context of the location and the shape of the isoclines of the model, this implies that the internal steady state changes from a stable to an unstable spiral point, when the (vertical) predator isocline moves from a location to the right of the top in the (quadratic) prey isocline to a location to the left of this maximum (see Figure 6.5). Figure 6.6 illustrates the series of changes in dynamics that occur for increasing values of the carrying capacity K: For K < Kc (Fig. 6.6; top-left panel) the internal steady state is biologically irrelevant, as the predator abundance in steady state would be negative (see eq. (6.45)). The prey carrying capacity is below the abundance needed by the predator to sustain itself. For Kc < K < Ks (Fig. 6.6; top-right panel) and K only slightly larger than Kc the internal steady state is a stable node. From every initial condition it is approached smoothly, i.e. in a non-oscillatory fashion. For larger K, but still K < Ks (Fig. 6.6; bottom-left panel) the internal steady state turns into a stable spiral and is approached from any initial condition in an oscillatory manner. For K > Ks (Fig. 6.6; bottom-right panel) the internal steady state has become an unstable spiral. Starting from an initial condition close to the steady state, the trajectory spirals away from the steady state and approaches a stable limit cycle surrounding the steady

116
Predator density C
0.3 dC/dt = 0

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Predator density C
0.3 dC/dt = 0

0.2

0.2

dF/dt = 0 0.1 dF/dt = 0 0.1

0.1

0.2

0.3

0.1

0.2

0.3

Prey density F Predator density C

0.3 dC/dt = 0

Prey density F Predator density C

0.3 dC/dt = 0

0.2

0.2

dF/dt = 0 0.1 0.1

dF/dt = 0

0.1

0.2

0.3

0.1

0.2

0.3

Prey density F

Prey density F

Figuur 6.6: Solution curves in the phase plane of the Rosenzweig-MacArthur predator-prey model. Two isoclines where dF/dt = 0 (dashed line) and dC/dt = 0 (dotted line)are drawn (the isoclines coinciding with the x- and y-axis have been omitted). The isoclines intersect in the internal steady state. Top-left panel: K = 0.08; the internal steady state is biologically irrelevant, the prey-only equilibrium is a stable node. Top-right panel: K = 0.15; the internal steady state is a stable node. Bottom-left panel: K = 0.25; the internal steady state is a stable spiral. Bottom-right panel: K = 0.3; the internal steady state is an unstable spiral. In this latter gure two trajectories are drawn: one starting close to the unstable steady state, the other starting far away from it. Both approach the unique limit cycle (thick solid line). Other parameter values: r = 0.5, a = 5.0, h = 3.0, = 0.5 and = 0.1.

state. This stable limit cycle arises when the steady state changes from a stable into an unstable spiral. From all initial conditions the trajectories eventually approach this stable limit cycle. Hence, when an initial condition outside the limit cycle is chosen, the trajectory spirals inwardly towards the limit cycle (as shown in Fig. 6.6). The limit cycle itself is invariant, which means that if an initial state is chosen exactly on the limit cycle, the state of the system would return to exactly this initial state again after going through exactly one period of oscillation. With K becoming larger and larger, the amplitude of the limit cycle grows very rapidly, such that eventually it passes very close to both the xand the y-axis of the phase plane (not shown).

6.2. CONFRONTING MODELS AND EXPERIMENTS

Predator density C
0.6 dC/dt = 0

117

0.4

dF/dt = 0 0.2

0.2

0.4

0.6

Prey density F

Figuur 6.7: Solution curve in the phase plane of the Rosenzweig-MacArthur predator-prey model for high carrying capacity. As Figure 6.6 but for K = 0.6. Other parameter values: r = 0.5, a = 5.0, h = 3.0, = 0.5 and = 0.1. Note that the scales have been doubles as compared to Fig. 6.6. It is possible to determine exactly for which value of K with Kc < K < Ks the internal steady state changes from a stable node into a stable spiral from the equality = 0 (see eq. (6.55)). However, the resulting expression for K is rather complicated and does not give any particular insight. Its derivation has therefore been omitted here. The analysis would show that eventually for very large values of K > Ks the two eigenvalues 1 and 2 would both become real-valued once again. From equation (6.54) it is clear that 1 and 2 are then both positive. For very large values of K > Ks the internal steady state thus becomes an unstable node or source. This change in character of the steady state is not very relevant, though, since it was already an unstable spiral point, surrounded by a stable limit cycle. The change only implies that with a initial state close to the steady state the trajectory will not spiral away from the steady state, but move away smoothly from the steady state to nally approach the stable limit cycle.

6.2

Confronting models and experiments

All three predator-prey models discussed above predict that in a steady state where both prey and predators coexist, the prey abundance is completely determined by the parameters, i.e. life-history characteristics, of the predator. In other words, even if the environment is richer in nutrients, leading to a higher growth potential (i.e. productivity) of the prey, the prey does not benet from this enrichment. On the contrary, it is the predator that benets as its abundance should increase suciently to counter the increased productivity of the prey. In conclusion, the predator-prey models of the Lotka-Volterra type predict that predators completely control the steady state abundance of their prey at a constant level and that increases in prey productivity

118
0.3

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Predator/prey abundance

0.2

0.1

50

100

150

200

Time

Figuur 6.8: Oscillatory dynamics of the Rosenzweig-MacArthur predator-prey model for high carrying capacity. Fluctuations in prey (black solid line) and predator abundance (red dashed line) as predicted by the Rosenzweig-MacArthur model. Notice the characteristic phase shift between the oscillations in the two populations. Parameter values: r = 0.5, K = 0.3, a = 5.0, h = 3.0, = 0.5 and = 0.1.

lead to a higher predator abundance. This prediction is the rst one that will be confronted with empirical and experimental observations below. A second prediction of the predator-prey models discussed pertains to the occurrence of limit cycles in predator and prey abundance. The basic Lotka-Volterra model predicts that abundances should always cycle, while the Rosenzweig-MacArthur model predicts that stable cycles in abundance should occur at higher levels of prey productivity (i.e. higher carrying capacities of the prey). A higher value of the prey carrying capacity can be thought of as representing a natural environment that is richer in basic nutrients. The phenomenon that with increasing values of the prey carrying capacity a predator-prey system would change from approaching a stable steady state to displaying limit cycles in abundance with rapidly increasing amplitude the larger the value of the carrying capacity K, has been called the Paradox-of-Enrichment by Rosenzweig (1971). Rosenzweig (1971) argued that intuitively a natural environment that is richer in nutrients is expected to be less prone to extinction of one of its species. However, with very large values of the prey carrying capacity K the Rosenzweig-MacArthur model exhibits large-amplitude limit cycles in which the predator and especially the prey abundance passes through values close to 0 (see Fig. 6.7). The slightest perturbation in these phases of very low densities may cause the prey to go extinct, subsequently leading to the extinction of the predator as well. In this way the Rosenzweig-MacArthur model predicts that environments richer in nutrients may actually be more prone to species extinction, in contrast to the intuitive expectation. Figure 6.8 shows that the resulting oscillations in prey and predator density also have a characteristic pattern: the maxima and minima in predator abundance are always delayed as compared to the maxima and minima in prey abundance. This phenomenon is referred to as a phase shift between the oscillations in abundance of the two populations. Characteristically, the phase shift between prey and predator abundance is roughly 1/4 of the cycle period in the RosenzweigMacArthur model (see Fig. 6.8)

6.2. CONFRONTING MODELS AND EXPERIMENTS

119

Figuur 6.9: Daphnia pulex carrying asexual eggs in the brood pouch. The ubiquity of cyclic dynamics in predator-prey models has drawn a lot of attention and has led to a large body of scientic research focusing on the question whether these cycles also occur in natural systems and if not, which mechanisms are responsible for this lack of occurrence. Here, I will discuss some empirical evidence, showing that the model prediction is in principle correct, but that the cyclic dynamics are often countered by many dierent mechanisms. The empirical evidence that will be used in the following discussion mainly relates to observations on species of the genus Daphnia (Cladocera; see Fig. 6.9). Daphnids or watereas, as they are more commonly known, have been used extensively in population dynamical studies for a variety of reasons. First of all, watereas often occupy an important position in freshwater foodwebs, as they are the main consumers of the algal primary producers. In addition, daphnids can be easily kept in the laboratory and through their short generation time lend themselves quite well for population dynamical experiments. Last but not least, daphnids have been the subject of many modeling studies, both to formulate models of individual physiology and life-history and for the derivation of population dynamic models. Therefore, daphnids are probably the best studied organism when it comes to population dynamics.

6.2.1

Predator-controlled, steady state abundance of prey

An elegant test of the prediction that steady state prey abundances are completely controlled by the predator life-history has been carried out by Arditi et al. (1991). Arditi et al. (1991) cultured several species of Daphnia and related organisms in semi-chemostat type culture vessels that were arranged in a serial order (see Figure 6.10). The rst culture vessel was fed with a stock solution with a high algal concentration. All following vessels were fed with the outow of the preceding culture vessel. At the start of the experiments, initial populations were established whereby culture vessels towards the end of the chain started out with lower abundances. The development of the populations in the various vessels in the chain were subsequently monitored. Several species of cladocerans, among which were two waterea species Daphnia magna and Ceriodaphnia reticulata, were tested in this experimental setup. Without explicitly writing down a model for the dynamics of the populations in the chain of culture vessels, the results obtained in the previous section are sucient to formulate a prediction about the outcome of the experiments. In section 6.1.1 and 6.1.2 it was shown that a steady state in which both prey and predator coexist is only possible if the maximum prey abundance in the absence of any predators (i.e. the prey carrying capacity) is larger than the prey abundance in the internal steady state. For the Lotka-Volterra model with logistic prey growth and for

120
I0

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

I1

I2

I3

I4

Figuur 6.10: Schematic setup of the experiments by Arditi et al. (1991). Populations of Daphnia and some related species are cultured in semi-chemostat type culture vessels that are arranged in a serial order. The outow of the culture vessels is only open to algae. Only the rst vessel in the chain is fed by a stock solution with high algal density. All following vessels are fed by the outow of the previous vessel. the Rosenzweig-MacArthur model it was derived that the internal steady state only adopted biologically meaningful (i.e. positive) values, i K > Kc := F 3 (see eq. (6.28) and (6.51)). In the experiments of Arditi et al. (1991) the prey growth is not logistic but follows a semi-chemostat growth equation (see eq. (4.9)). Nonetheless, the same model prediction holds: only if the maximum prey abundance is larger than the prey abundance in the internal steady state, a population of predators can persist. The maximum prey abundance in a semi-chemostat culture vessel is obviously equal to the abundance in the inow. If the prey abundance in steady state F 3 is indeed completely controlled by the life-history parameters of the predators only the rst culture vessel in the series will be subject to a prey inow that is larger than the steady state prey abundance. The inow of the second vessel in the chain is coupled to the outow of the rst, in which the prey abundance equals the predator-controlled steady state value F 3 . Therefore, if the model prediction holds that the prey abundance in steady state is controlled by the predator at its subsistence level, only the predator population in the rst culture vessel in the chain can survive, as the other vessels are subject to a prey inow that equals the predator subsistence level F 3 and is hence not sucient for the establishment of a predator population. A more formal analysis involves writing down a coupled system of ODEs for the algal and daphnid populations in each culture vessel. Since daphnids tend to feed on algae following a type II functional response (see Fig. 6.4) this set of equations is very similar to Tilmans

6.2. CONFRONTING MODELS AND EXPERIMENTS

121

competition model (4.35). More specically, the dynamics of algae and daphnids in compartment n can be described by: aFn dFn = D (In Fn ) Cn dt 1 + ahFn dCn aFn = Cn Cn dt 1 + ahFn

(6.59a) (6.59b)

in which D is the ow-through rate, In refers to the algal abundance in the inow of compartment n, Fn and Cn represent he algal and daphnid abundance in this compartment and the parameters a, h, and have the same interpretation as in the Rosenzweig-MacArthur model. Obviously, the internal steady state predicted by the model obeys: a( h)

F n = F := Cn =

(6.60a) (6.60b)

D In F

Since for n > 1 the inow In is the outow of the previous compartment, which equals F , it is clear that in all culture vessels except the rst, persistence of the daphnid populations is not possible. Figure 6.11 and 6.12 show the experimental results that Arditi et al. (1991) obtained: Clearly all populations of Daphnia magna in culture vessel 2 and higher go extinct. The same result holds for Ceriodaphnia reticulata, although here the extinction of these populations takes much longer. Daphnia magna is a much larger and more voracious lter-feeder on algae than Ceriodaphnia reticulata, which may explain the dierence in transient time. Nonetheless, the experimental results t the predictions of the predator-prey models quite well. Arditi et al. (1991) also obtained results for Scapholeberis kingi, a cladoceran species closely related to Daphnia, that did not t the model predictions at all: predator populations persisted in all compartments in the chain. This species diered, however, in one important aspect from Daphnia in that the organisms tend to be surface-dwelling. Hence, the predator-prey models discussed in this chapter might not apply to these species, as they are based on an assumption that both the predator and the prey live in a homogeneous, well-mixed environment. Murdoch et al. (1998) provide another test of the prediction that the prey abundance at steady state is completely controlled by the life-history parameters of the predator and not aected by the prey growth capacity. Murdoch et al. (1998) analyzed a number of datasets documenting the dynamics of Daphnia species in experimental tanks, ponds and lakes. Only those data sets were included for which it was ascertained that sh predation was either absent or could be corrected for. Moreover, only data sets for which there was complete information on the assemblage of algal species found in the studied system were included. Daphnia does not eat and is actually hampered in its feeding behavior by the presence of larger, inedible species of algae. The precise information on the algal assemblage made it possible that an accurate estimate was derived for the abundance of Daphnias prey, i.e. the edible algal species. Figure 6.13 shows that over a large range of nutrient levels the mean abundance of edible algae in the presence of Daphnia remains relatively constant. The nutrient level of the system was estimated from the maximum algal abundance observed in the lake or pond during the spring algal bloom, a short period of time in the beginning of the year when the algae are not controlled

122

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Figuur 6.11: Population dynamics of Daphnia, Ceriodaphnia and Scapholeberis in a chain of semi-chemostats. Population dynamics (log scale) of the three species in each chamber. Dotted lines denote extinctions. From the chosen initial conditions, equilibria were reached rapidly and without oscillations. In (a) and (b), extinct chambers were re-inoculated every week but became extinct again (data not shown). Source: Figure 2 in Arditi et al. (1991).

by zooplankton yet. This estimate can hence be interpreted as an estimate of the carrying capacity of the algae. A Rosenzweig-MacArthur model for the interaction between Daphnia and algae had been derived previously (Nisbet et al. 1991). The parameters of this model were estimated on the basis of independent, mostly laboratory experiments. Hence, the model provided a completely independent estimate of the expected algal abundance in equilibrium, amounting to 0.05 mg C/L (prey abundance was consistently expressed as biomass in terms of amount of carbon). This estimate ts the observed algal densities reasonably well. A careful analysis of the data on Daphnia and its algal prey thus shows that the observations agree with the hypothesis that prey abundances in steady state are completely controlled and depend on the life-history parameters of the predator. Notice that this assessment only holds true after some complicating mechanisms, such as dierent levels of sh predation and the presence of inedible algae, have been corrected for. In other words, even though the model conclusion of a predator-controlled prey abundance in steady state in principle holds true, it may often be

6.2. CONFRONTING MODELS AND EXPERIMENTS

123

Figuur 6.12: Population equilibria of Daphnia and Ceriodaphnia in a chain of semi-chemostats. Populations declined to extinction in all compartments beyond the rst one, in agreement with model predictions of the Lotka-Volterra type. Each point is the average of four values (last two weeks of two replicates). Source: Figure 3 in Arditi et al. (1991). obscured by mechanisms that are not accounted for in the basic model.

6.2.2

Oscillatory dynamics at high prey carrying capacities

The Rosenzweig-MacArthur model predicts that with increasing prey carrying capacity oscillatory dynamics are more likely to occur. The question whether cycles in population abundance also occur in natural populations has been a topic of many ecological studies. The dynamics of the snowshoe hare and the lynx population in northern Canada (see Figure 6.14) has been often cited as a classical example of cycles in population abundance in a predator-prey system. The data on these two species constitute one of the longest time series in ecology. The basis of this dataset are the fur trading records of the Hudson Bay Company in northern Canada. However, despite the uctuations in the number of hare and lynx furs that the Hudson Bay Company bought up from trappers each year, the cycles do not have the characteristics of a predator-prey cycle. Figure 6.14 shows clearly that a phase shift between the predator and prey abundances, as is typical for the Rosenzweig-MacArthur model (see Fig. 6.8) is hard to detect in the data on hare and lynx. Indeed, Gilpin (1973) has shown that when the dynamics are represented in the phaseplane, the trajectory does not spiral in the direction that is typical for a predator-prey model (see, for example, Fig. 6.7). Another system that has often been the focus of studies on population cycles, are the microtine rodent (vole and lemming) populations in northern Europe (see Fig. 6.15). Microtine rodent populations in Scandinavia have exhibited frequent population cycles with huge dierences in abundance from year to year, which have not only captured the imagination of ecologists, but

124
0.3

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Equilibrial edible biomass (mg C/L)

0.2

0.1

1.0

2.0

3.0

Edible algal carrying capacity, K (mg C/L)

Figuur 6.13: Predicted and observed equilibrium values of algae in the presence of Daphnia. Only edible algae have been included in these estimates. Over a ten-fold range of nutrient levels, as expressed by the edible algal carrying capacity K, the observed algal abundance in steady state is relatively constant. The solid line indicates the prediction of an independently parameterized, Rosenzweig-MacArthur model. Source: Figure 7 in Murdoch et al. (1998).

also of the general public. Turchin et al. (2000) have argued convincingly that the cycles in abundance in voles are predator-prey cycles, in which the vole are the prey of their specialized predator, the weasel. However, data on uctuations in abundance of the weasel are generally not available, hence it is not possible to compare the cycle characteristics with those of the limit cycles predicted by the Rosenzweig-MacArthur model. The uctuations in abundance of Daphnia and algae is a third system that has been used to study the occurrence of cyclic dynamics predicted by the Rosenzweig-MacArthur model. McCauley & Murdoch (1990) noted that over a wide range of nutrient levels, the amplitude in the uctuations in Daphnia and algae does not increase with nutrient level (see Fig. 6.16). The Rosenzweig-MacArthur model predicts that with increasing prey carrying capacity K both the amplitude in the predator-prey cycles and the period of the cycles increase. Neither of these two increases were distinguishable in the dataset collated by McCauley & Murdoch (1990). In addition, McCauley & Murdoch (1990) set up experiments in stock tanks in which they studied the dynamics of Daphnia and algae under nutrient-rich and nutrient-poor conditions. The nutrient-rich conditions had an estimated algal carrying capacity that was 10 times higher as in the nutrient-poor conditions. Under nutrient-rich conditions, the average biomass of Daphnia was signicantly higher than under nutrient-poor conditions. However, when corrected for this increase in average biomass, no increase in cycle amplitude could be observed between the two sets of systems (see Fig. 6.17). Hence, even though cycles are observed in Daphnia and algae there seems no evidence that the cycles are of the predator-prey type predicted by the Rosenzweig-MacArthur model, since neither cycle amplitude nor cycle period tend to increase with increasing algal carrying capacity. McCauley & Murdoch (1990) argued that the cycles that are observed in Daphnia are caused by the delay between the birth of an individual waterea

6.2. CONFRONTING MODELS AND EXPERIMENTS

3.0

125

Population abundance (x 105)

2.0

1.0

0 1800 1850 1900

Time in years

Figuur 6.14: Population dynamics of the snowshoe hare and the lynx in northern Canada. Squares with solid line: snowshoe hare abundance; Triangles with dotted line: lynx abundance. Data are partially derived from the fur trading records kept by the Hudson Bay Company. Source: NERC Centre for Population Biology, Imperial College (1999), The Global Population Dynamics Database, http://www.sw.ic.ac.uk/cpb/cpb/gpdd.html.
40

30

Population abundance

20

10

0 1950 1960 1970 1980 1990

Time in years

Figuur 6.15: Population dynamics of two species of voles in northern Finland. Squares with solid line: Grey-sided vole (Clethrionomys rufocanus); Triangles with dotted line: Common vole (Microtus). Both species constitute important prey species for weasels. Source: NERC Centre for Population Biology, Imperial College (1999), The Global Population Dynamics Database, http://www.sw.ic.ac.uk/cpb/cpb/gpdd.html.

126

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Figuur 6.16: Cycle amplitudes (log) observed for Daphnia and algal populations from lakes and ponds. Data collated by McCauley & Murdoch (1987). Comparable studies are plotted (that is, Daphnia and algae estimated as density and biomass, respectively). Cycle periods are not correlated with changes in algal biomass among systems. The average total algal biomass is used as an index of the level of enrichment. Source: Figure 1 in McCauley & Murdoch (1990).

and the onset of its reproduction, since the cycle period was consistently close to the duration of the juvenile period. The experiments by McCauley & Murdoch (1990) suered from a complicated factor, though, in that in the experimental tanks also signicant densities of inedible algae were present. As mentioned before, these algae are not only inedible for Daphnia, but they also interfere with the foraging of Daphnia on edible algae. Because this interference may lead to a reduced attack rate of Daphnia on edible algae and hence to stable dynamics, McCauley & Murdoch (1990) suggested that the presence of inedible algae was a potential mechanism explaining the absence of predator-prey cycles in Daphnia. This explanation has been tested more recently by McCauley et al. (1999) using similar studies with laboratory populations of Daphnia in experimental tanks. McCauley et al. (1999) avoided the interference of inedible algae by preventing any algae to grow on the walls of the tanks. These walls are thought to be the main spot in the tanks were large-bodied, inedible algae can establish themselves. Indeed, the experimental control of inedible algae led to the occurrence of large-amplitude, predator-prey cycles that resemble the dynamics exhibited by the Rosenzweig-MacArthur model (see panel A and B in Fig. 6.18). However, among the replicate experiments only about half of the time series of Daphnia and inedible algae displayed such predator-prey cycles. The other half displayed cycles with approximately the duration of the juvenile period of Daphnia that are attributed to the delay between birth and the onset of reproduction (McCauley & Murdoch 1990). Hence, the same experimental and initial conditions led to two dierent patterns of uctuations. The occurrence of the smallamplitude cycles in Daphnia and inedible algae were observed to coincide with the occurrence of larger densities of female Daphnia carrying resting eggs. Normally, Daphnia is parthenogenetic with females producing large clutches of asexual eggs (see Fig. 6.9). Under adverse temperature and food conditions, though, females switch to producing sexual eggs that form a resting stage

6.2. CONFRONTING MODELS AND EXPERIMENTS

127

Figuur 6.17: Examples of the dynamics of Daphnia and algae in nutrient-rich and nutrientpoor, experimental tanks. Daphnia is measured as biomass (mg dry weight per liter), while algae are measured as chlorophyll a density (g per liter). Only edible algae are included. Nutrientrich tanks had approximately 10 times higher nutrient levels than nutrient-poor tanks, leading to an estimated increase in algal carrying capacity of one order of magnitude. Abundances are represented as residual deviations from a long-term, simple seasonal trend. Source: Figure 2 in McCauley & Murdoch (1990). and can overwinter on the bottom of a lake or pond. It is these resting eggs that give rise to a new population of Daphnia each year. The production of such resting eggs implies a channelling of energy away from immediate reproduction of Daphnia, which might be another mechanism preventing large-amplitude predator-prey cycles. Indeed, when experimental tanks were manipulated such that females with sexual resting eggs were continuously replaced with similarly sized females carrying parthenogenetic eggs, all replicates exhibit large amplitude predator-prey cycles (see Fig. 6.19). Hence, when both inedible algae and the occurrence of sexually reproducing females in Daphnia were controlled for, the populations of Daphnia and edible algae indeed exhibited the type of dynamics predicted by the Rosenzweig-MacArthur, predator-prey model.

6.2.3

Concluding remarks

The above discussion and presentation of experimental evidence suggests that the two most fundamental predictions of the predator-prey models discussed in this chapter, i.e. the predatorcontrolled prey abundance in steady state and the occurrence of large-amplitude predator-prey cycles at high algal carrying capacity in homogeneous environments, can indeed be conrmed to hold in natural systems under specic conditions. However, as the discussion has also made clear there are many dierent, sometimes even very subtle, mechanisms that may obscure the observation of these cycles or may prevent them to occur. It is exactly the discrepancy between the basic predictions of the model and the experimental observations that leads to further study of the dynamics of a population, ultimately revealing new insight about factors that shape the pattern of dynamics observed. This has been most clearly demonstrated by the phenomenon that the production of sexual eggs in Daphnia can prevent the occurrence of large-amplitude predator-prey cycles, as discussed in the last section.

128

HOOFDSTUK 6. PREDATOR-PREY INTERACTIONS

Figuur 6.18: Large- and small-amplitude cycles of Daphnia and edible algae in the same global environment. All four data sets stem from comparable, nutrient-rich systems. Red triangles: Daphnia; Green squares: algae. The solid lines are spline ts to the time series. A,B : Examples of large-amplitude predator-prey cycles. C,D : Examples of small-amplitude, stage-structured cycles. The initial biomass of all replicates is similar. Source: Figure 1 in McCauley et al. (1999).

Figuur 6.19: Energy channelling towards sexual reproduction prevents the occurrence of largeamplitude predator-prey cycles in Daphnia. A: Dynamics of replicate Daphnia populations illustrating that the replicates are asynchronous in time and can either be of large or small amplitude. B : Dynamics of the Daphnia populations in which energy channelling towards asexual reproduction is maintained by swapping adult females producing ephippial resting eggs with similarly sized females carrying asexual eggs. Source: Figure 5 in McCauley et al. (1999).

Deel III

Bifurcation theory

129

Hoofdstuk 7

One parameter bifurcations in continuous time models

Imagine yourself sitting on a swing, happily swinging back and forth. The strain exerted on the ropes with which the swing is fastened to the overhead beam can in principal be computed using well known relations from Newtonian mechanics. This strain is the resultant of the gravitational force and the force induced by the swinging motion. It will increase with increasing swinging speed. Even though the mechanical laws can be used to compute the strain on the ropes with good accuracy, you will probably be much more interested in the question whether the rope will break or not and what the critical swinging speed is at which the rope will break. In other words, your main interest will be whether an abrupt change in the resulting motion can occur and, if yes, at which values of the conditions this change will occur. In modeling biological populations a similar argument applies: Often the main focus of a population dynamical study is whether abrupt changes in the dynamics of the populations occur and, if yes, at which values of parameters such changes do occur. An abrupt change in model dynamics with a (slight) change in parameters is called a bifurcation. The change of a steady state from a stable node into a saddle is an example of such a bifurcation. Bifurcation theory is the part of theory about population dynamic models, or more generally dynamical systems, that deals with classifying, ordering and studying the regularity in these changes. Bifurcations that occur in dynamical systems are generally robust against (small) changes in model structure. The dynamics of a particular model will of course depend on its precise formulation and on the chosen set of parameters. Moreover, changes in model formulation are likely to change the dynamics observed for a given set of parameters. But if a particular model exhibits a change in dynamics from a stable equilibrium to limit cycle behavior with a change in a particular parameter, this qualitative change in dynamics will in general also be observed in models that are analogous, but slightly dierent in formulation. The scientic questions that are sought to be answered with a population dynamical model often translate into questions regarding bifurcations. For example, when investigating whether the fragmentation of natural habitat will lead to species extinction, one is actually asking whether an internal steady state with positive abundances will be predicted by the model as opposed to a steady state with zero population abundance. Similarly, when considering the question whether an epidemic of a particular disease is likely to occur, one is also asking about the likelihood of two dierent types of steady states: an internal steady state and an extinct steady state. The transition at a particular parameter value between a parameter domain where an extinct steady state is stable to a domain where an internal steady state is stable is an example of a transcritical bifurcation or branching point (see section 7.2). In many cases, population dynamical studies

131

132

HOOFDSTUK 7. CONTINUOUS TIME MODELS

thus aim at identifying qualitative changes in the long-term dynamics predicted by the model, i.e. changes in the likelihood and stability properties of the model steady states. From a biological point of view the robustness of bifurcations, as discussed above, implies that the occurrence of a particular change in the observed population dynamics will not sensitively depend on the precise biological mechanisms assumed in the model. Biological conclusions based on bifurcations are hence more robust than conclusions that are based on the dynamics of the model for a particular set of parameters. In this chapter I will discuss the types of changes in dynamics that can occur in continuous-time models of biological populations when a single model parameter is varied. Only three types of such bifurcations can occur and these three types posses particular characteristics that hold independent of the model they occur in. For presentational purposes I will, however, focus on models that are formulated in terms of 2 ODEs. This chapter will essentially not introduce any new techniques or methods, as the discussion below depends entirely on the computation of steady states, their eigenvalues and stability properties, such as introduced in chapter 5. A new perspective will be added, because we will focus on the question how these model properties change when a parameter in the model is changed. For the discussion below it is important to note the following:

Important: If one of the parameters in the model changes, the value of the steady state(s), the corresponding eigenvalues and hence the stability characteristics may change as well. However, these changes are in general continuous, which means that plotted as a function of the parameter value the value of the steady state(s) form a smooth and continuous curve. The same holds for the value of the corresponding eigenvalues (both the imaginary and real parts!).

7.1

General setting

In chapter 5 it was discussed that a steady state X of a population dynamic model, formulated as, dx = H(x) (7.1) dt (see eq. (5.4)) is stable if all the eigenvalues pertaining to the steady state have a negative real part. It was also explained that the number of eigenvalues pertaining to a steady state equaled the dimension of the model, for example, in a model in terms of 2 ODEs every steady state is characterized two, unique eigenvalues 1 and 2 . In such a 2-dimensional model there are 5 possible congurations of the eigenvalues, dependent on whether the steady state is a stable or unstable node, a stable or unstable focus or a saddle point. These congurations are presented in Table 7.1 together with the phaseportrait of the trajectories in the neighborhood of the steady state. A change in stability properties of the steady state occurs when a change in a particular model parameter causes the real part of an eigenvalue to change from negative to positive. Represented in the complex plane, this implies that a change in stability of a steady state, and hence a

7.1. GENERAL SETTING

133

Tabel 7.1: Characteristics of all possible types of steady states in a 2-dimensional, continuoustime model. For each type of steady state the number of eigenvalues with a positive (N+ ) and negative (N ) real part are indicated, their position in the complex plane, the phaseportrait in the neighborhood of the steady state and its stability.

Name

(N , N+ )

Eigenvalues

Phaseportrait

Stability

Node

(2, 0)

Stable

Focus

(2, 0)

Stable

Saddle

(1, 1)

Unstable

Node

(0, 2)

Unstable

Focus

(0, 2)

Unstable

134

HOOFDSTUK 7. CONTINUOUS TIME MODELS

Figuur 7.1: Eigenvalue positions in the complex plane corresponding to a stability change in a 2 ODE model. Left: Re() = Im() = 0. A real eigenvalue passes the (real) value 0. This change in stability occurs at a saddle-node bifurcation and at a transcritical bifurcation (or branching point). Right: Re(1 ) = Re(2 ) = 0 and Im(1 ) = Im(2 ) > 0. A complex, conjugate pair of eigenvalues crosses the imaginary axis from left to right. This occurs at a Hopf bifurcation. A Hopf bifurcation can only occur in models that are of dimension 2 or higher. change in long-term dynamics of the model, occurs when an eigenvalues crosses the imaginary axis from the left-half plane to the right-half plane. Because the value of the steady state and its eigenvalues change continuously with a change in model parameters and complex eigenvalues always occur as a pair of complex conjugates, there are only two generic ways in which such a shift of eigenvalues from the left- to the right-half complex plane can come about (see Fig. 7.1): a real-valued eigenvalue changes from negative to positive, crossing the (real) value 0, a pair of complex, conjugate eigenvalues 1 and 2 crosses the imaginary axis at a non-zero imaginary value i and i, respectively. Referring to Table 7.1 this implies that a stable node can only destabilize by changing into a saddle point, and a stable focus can only destabilize by changing into an unstable focus. The rst of these situations is characteristic for two dierent types of bifurcations: the transcritical bifurcation (or branching point; see section 7.2) and the saddle-node bifurcation (or limit point; see section 7.3). The last of these situations is characteristic for a Hopf bifurcation (see section 7.4). The Hopf bifurcation can only occur in models of at least 2 ODEs, as it involves always two eigenvalues. The type of stability change is thus intimately tied to the changes in the eigenvalues that characterize the steady state. Sections 7.2-7.4 discuss these three types of bifurcations using two example models. The transcritical bifurcation and the saddle-node bifurcation will be illustrated using a model for the dynamics of a cannibalistic population, in which adult individuals forage on juvenile conspecics, next to feeding on an alternative resource. Here, I will rst introduce and discuss this particular model. Let J denote the abundance of juvenile individuals and A the abundance of adults. The dynamics of the population can then be described by the following system of 2 ODEs:

7.1. GENERAL SETTING

135

dJ = A J J A dt dA = J A dt + J

(7.2a) (7.2b)

In these equations denotes the per capita reproduction rate of the adult individuals, which implies that A models the total number of ospring produced by all the adults per unit of time. represents the developmental rate of the juvenile individuals, which implies that J models the number of juvenile individuals that mature per unit of time. JA represents the cannibalism of adult individuals on the juveniles. Adult individuals are assumed to forage also for an external, non-cannibalistic food source, the density of which is represented by . The death rate of adult individuals is assumed to be inversely proportional to their total food intake rate, which equals + J. The proportionality constant is denoted by . Hence, A/( + J) is the total number of adult individuals that are dying per unit of time. Notice that juvenile individuals are assumed to die from no other causes, but cannibalism. From equation (7.2b) we can infer that in steady state the abundance of adult individuals A is always positively correlated with the abundance of juveniles J: A = J + J (7.3)

This implies that given a steady state value J, the steady state abundance A is uniquely determined. Substituting the relation (7.3) into the right-hand side of the ODE (7.2a) and equating it to 0, yields the following equation for the steady state abundance J: J + ( ) J 2 J The model hence possesses a trivial steady state: (J, A) = (0, 0) and non-trivial steady states that have to fulll the conditions:
2 2

= 0

(7.4)

(7.5)

2 J + ( ) J + ( ) = 0 A = J + J

(7.6a) (7.6b)

In the following the non-trivial steady state(s) and their stability will be investigated as a function of the parameter , which scales the food-dependent mortality of the adult individuals. Hence, if increases adult mortality is higher and increases more rapidly with decreasing food conditions. To construct graphs of the steady state abundance J as a function of the parameter it will turn out to be most convenient to rst construct the graph of as a function of J. The latter relation between , as the dependent variable, and J , as the independent variable, is given by the following quadratic equation:

136

HOOFDSTUK 7. CONTINUOUS TIME MODELS

= + ( ) J 2 J

(7.7)

This parabola is easy to construct and can subsequently be ipped around the line = J to yield the required relation between J, as the dependent variable, and , as the independent variable. The Jacobian matrix of the cannibalism model is given by the equation: A + +J
2

J A +J

J =

(7.8)

Substitution of the trivial steady state values, (J, A) = (0, 0), into this Jacobian matrix yields the following equation for the eigenvalues: 2 + ( + ) + ( ) = 0 (7.9)

(Check this by carrying out the substitution and determining the determinant of the resulting matrix!) From equation (7.9) it can be deduced that the trivial steady state is characterized by two real, but negative eigenvalues if: < 1 / (7.10)

The left-hand side of this inequality can be interpreted as the expected number of ospring produced by a single individual during its entire life under conditions of very low population abundance. When abundance is low, density-dependent eects are absent and neither is juvenile mortality increased, nor is adult mortality decreased through cannibalism. The left-hand side of the inequality hence equals the expected lifetime reproduction when individuals are living more or less on their own. This quantity is usually denoted by the symbol R0 . For the cannibalism model R0 is given by: R0 = / (7.11)

The expression for R0 can be derived as follows. Under conditions of low population abundance, death of juveniles is negligible since there are no adults to cannabilise a juvenile individual. For an adult individual, the probability to survive for a period of time t equals: S(t) = e(/)t (7.12)

S(t) represents here the probability that an individual is still alive after t units of time and is obtained by solving ODE (7.2b) when J is taken equal to 0 in that equation (Remember the assumption of low population abundance!). Per unit of time an adult individual produces on average newborn ospring. Hence, the expected number of ospring produced during its entire life equals the integral of the S(t) for t = 0 to t = (i.e. the integral over the entire adult lifespan):

R0 =
0

e(/)t dt =

(7.13)

7.1. GENERAL SETTING

137

Condition (7.10) now states that the trivial steady state is stable if the expected lifetime reproduction on the basis of alternative, non-cannibalistic food, /, is smaller than 1, i.e. if an individual cannot on average replace itself on the basis of non-cannibalistic food. If / is larger than 1 the trivial steady state is unstable and the population is capable of growing away from low population abundances. Given the biological interpretation of the quantity R0 = / this is easily understood. Note that with increasing the adult mortality increases and hence their expected lifespan decreases. This implies that only for lower values of a population will be able to persist at non-zero population abundances. The characteristic equation (7.9) for the trivial steady state thus indicates that at = the stability properties of the trivial steady state change. Substitution of this value = into the condition for the juvenile abundance in the non-trivial steady state (eq. (7.7)) shows that condition (7.7) yields a solution J = 0, as well as a second solution: J = (7.14)

For = both the trivial steady state condition (7.5) and the internal steady state condition (7.7) hence yield the same solution J = 0. If the value of the trivial steady state and the internal steady state is plotted as a function of the parameter this implies that the curve representing the internal steady state intersects the curve for the trivial steady state at the parameter value = . In addition, equation (7.14) indicates that a second, positive, non-trivial steady state exists for = if > . (7.15)

Figure 7.2 and 7.3 illustrate the relationship between the steady state abundance J and the adult mortality in case < and > , respectively. Each of these two gures represents two curves: one curve indicating the trivial steady state and one curve representing the internal steady state, as a function of the parameter . For each of these curves the gures also indicate for which parameter values the particular steady state is stable and for which values it is unstable. The two gures are representative for the two distinct situations that can occur dependent on and : If < the curve representing the internal steady state bends to the left at the point (, J ) = (, 0) where it intersects the curve that represents the trivial steady state. For these parameter values condition (7.7) determining the internal steady state has a second, non-trivial solution at a negative juvenile abundance J = ( )/ for = . In accordance with this assessment we can infer that the quadratic relation (7.7), which species as a function of J , is a parabola with a maximum at a negative value of J. Hence, for increasing J this condition determines that decreases monotonously. Flipping the parabola around the line = J yields the relationship illustrated in Figure 7.2. If > the curve representing the internal steady state bends to the right at the point (, J ) = (, 0) where it intersects the curve that represents the trivial steady state. In other words, the curve initially extends towards higher values of the mortality parameter and curves back

138

HOOFDSTUK 7. CONTINUOUS TIME MODELS

J
C

=
C

Figuur 7.2: Bifurcation structure of the cannibalism model for < . Both the positive, internal steady state and the trivial steady state are indicated with thick solid lines, when they are stable and thick dashed lines when they are unstable. The insets show the position of the eigenvalues in the complex plane for the dierent parts of the steady state curves. to lower values of at a positive value of J (see Figure 7.3). In this case there is a second, positive internal steady state abundance J = ( )/ for = . When > the quadratic relation (7.7), which species as a function of J , is a parabola with a maximum at a positive value of J , which yields the relationship as illustrated in Figure 7.3 when around the line = J. For positive value of both J and A it is immediately clear that the trace of the Jacobian matrix (7.8) is negative. Hence, the internal steady state can only destabilize when a real-valued eigenvalue becomes positive (see the left panel in Figure 7.1). The internal steady state is hence either a stable node or focus or it is a saddle point. The determinant of the Jacobian (7.8) evaluates to 3 2 J det J =
2

+ 2 ( ) J + ( ) + J

(7.16)

which expression has been obtained after substituting expression (7.3) for A. The expression for the determinant det J can be simplied further by using the steady state relation (7.7) between and J to:

7.1. GENERAL SETTING

139

Figuur 7.3: Bifurcation structure of the cannibalism model for > . Both the positive, internal steady state and the trivial steady state are indicated with thick solid lines, when they are stable and thick dashed lines when they are unstable. The insets show the position of the eigenvalues in the complex plane for the dierent parts of the steady state curves.

det J =

J 2 J + ( ) + J

(7.17)

For small, positive values of J the determinant is positive as long as < and negative if the opposite inequality holds. Hence, for < the internal steady state is always stable, as the determinant of the Jacobian matrix is positive while its trace is negative (see Table 5.1 and Figure 5.5 in chapter 5). If, on the other hand, > the internal steady state is a saddle point, since the determinant det J is negative for small, but positive values of J . With increasing values of J the value of the determinant increases as well to become equal to 0 when J = 2 (7.18)

(see equation (7.17)). This value of J is exactly the value where the curve representing the internal steady state in Figure 7.3 reaches its maximum value of and bends back towards lower -values. The equilibrium relation between and J indicates that this extremum is given by: (, J ) = ( + )2 , 4 2 (7.19)

140

HOOFDSTUK 7. CONTINUOUS TIME MODELS

At this maximum value of (see Fig. 7.3) the internal steady state hence changes from a saddle point into a stable node. With further increases in J the determinant will only increase further, while the trace of the Jacobian matrix will remain negative. Hence, for J > the internal steady state will always be stable. Figure 7.2 and 7.3 summarize the results derived above for the cannibalism model. The gures clearly indicate that there are two characteristic changes in stability of the internal steady state: one at the intersection point between the curve representing the internal and the trivial steady state and one where the curve of the internal steady state exhibits an extremum (see Fig. 7.3). These characteristic changes in stability, corresponding to a transcritical bifurcation and a saddle-node bifurcation, respectively, will be discussed next. 2

7.2

Transcritical bifurcation and branching point

From the analysis of the cannibalism model discussed above, we can infer that at the value = the curves representing the trivial steady state and the internal steady state, respectively, intersect. For decreasing values of the stability of the trivial steady state changes at this intersection from being a stable node into a saddle point, while the internal steady state changes from a saddle point into a stable node. In other words, the trivial and internal steady state exchange their stability characteristics at the intersection point. In addition, the internal steady state enters the positive cone, i.e. the part of the (J, A)-phaseplane with positive abundances for both juveniles and adults. Each of the steady states is characterized by a pair of eigenvalues that are both real. At the intersection point (, J ) = (, 0) from these two pairs of eigenvalues (i.e. one pair pertaining to the trivial steady state and the other pertaining to the internal steady state) always one member of the pair is negative, while the second member has a zero value. In other words, at the intersection point two eigenvalues equal to 0 coincide, but they pertain to two dierent steady states. When decreasing the value of from above to below its critical value the zero eigenvalue pertaining to the trivial steady state moves to the right (see Fig. 7.1) into the half of the complex plane where real parts are positive. Simultaneously, the zero eigenvalue pertaining to the internal steady state moves to the left, towards the half of the complex plane where real parts are negative. The changes in steady state characteristics occurring at the value = are representative for a transcritical bifurcation. The intersection point itself, (, J ) = (, 0), is referred to as a branching point. Transcritical bifurcations are very common in population dynamic models in general. The characteristic properties of a transcritical bifurcation are: it is an intersection point, referred to as branching point, of two curves representing dierent steady states of the model, at the intersection point both steady states have a zero eigenvalue, when moving through the intersection point one of the two zero eigenvalues moves into the left half of the complex plane, the other one into the right half of the complex plane. Since they belong to two dierent steady states, the steady states exchange their stability characteristics. These properties are independent of the detailed specication of the model, but depend solely on the changes in the value of the eigenvalues around the branching point. Hence, when an

7.3. SADDLE-NODE BIFURCATION AND LIMIT POINT

141

intersection point of two steady state curves is found in a particular model, we can immediately infer that an exchange in stability characteristics has to occur there.

7.3

Saddle-node bifurcation and limit point

For > the curve representing the internal steady state has an extremum at the point (, J ) = ( + )2 , 4 2

(see Figure 7.3). For -values slightly larger than ( + )2 /4 there are no internal steady states at all, while for slightly smaller -values there are two steady states, a saddle point and a stable node. At this point the internal steady state exhibits a change in stability from a stable node into a saddle point, when we follow its curve through the (, J)-plane (see Fig. 7.3). The extremum point of the curve is called a limit point. The changes in steady state characteristics that occur around the limit point are referred to as a saddle-node bifurcation, because a saddle point and a stable node merge and disappear when increases past the limit point at = ( + )2 /4. The changes in stability again relate to the fact that a real-valued eigenvalue changes sign when the steady state curve is followed around the limit point. In general, we can state that a limit point is a point in the steady state curve, at which an extreme parameter value is reached, and at the limit point a saddle point and a stable node merge and disappear, or in other words, when following the steady state curve around the limit point, the steady state changes its stability, because a real-valued eigenvalue changes sign. Limit points are frequently occurring constructs in dynamic systems of any kind. The characteristics of the dynamics in the neighborhood of the limit point are independent of the model specics and are completely determined by the changes in the eigenvalues characterizing the steady states. Therefore, when during the analysis of a model, a steady state curve exhibits a limit point, it immediately gives information about the stability properties of the steady state in the neighborhood of this point. Alternatively, if during numerical simulations of a dynamic model a steady state is reached for a particular range of parameter values, but a small increase or decrease in the parameter outside the range will cause the steady state to suddenly disappear, one has to strongly suspect the presence of a limit point and hence the presence of a saddle point close to the steady state that was initially approached. The presence of a limit point is hence a likely explanation for rather catastrophic changes in model dynamics with small changes in parameters. From a biological point of view Figure 7.3 illustrates the most interesting eect of cannibalism on the population dynamics. For a range of -values larger than = the cannibalistic population can sustain itself in an internal steady state, while adult individuals are incapable to produce more than a single ospring during their life on the basis of the alternative, non-cannibalistic food source. Through cannibalism the adult individuals can make up for this deciency such that on average they do replace themselves during their lifetime. This eect has been dubbed the life-boat eect of cannibalism (Van den Bosch et al. 1988).

142

HOOFDSTUK 7. CONTINUOUS TIME MODELS

K Kc Ks
Figuur 7.4: Schematic, 3-dimensional representation of the Hopf-bifurcation in the RosenzweigMacArthur model. The axes represent the carrying capacity K, the prey abundance F and the predator abundance C. The thick, solid line in the bottom plane indicates the prey-only steady state (F , C) = (K, 0) for those values of K where this steady state is stable. The positive, internal steady state is indicated with a thick solid line when stable and with a thick dashed line when it is unstable. With the destabilization of the internal steady state a limit cycle occurs. In the (F, C)-phaseplane this limit cycle constitutes a closed curve. Every initial condition will ultimately approach this limit cycle, when it exists. With increasing value of K the amplitude of the limit cycle increases. The set of limit cycles for all K values forms a parabolic structure in the 3-dimensional space as sketched. Kc and Ks indicate the critical values of K where a transcritical and a Hopf-bifurcation occur, respectively.

7.4

Hopf bifurcation

The last bifurcation that can occur in a continuous-time model when varying a single parameter, is the change from a stable to an unstable focus (see Table 7.1). This change in stability occurs when a pair of complex, conjugated eigenvalues crosses the imaginary axis from the left-half plane into the right-half plane (see Figure 7.1). This bifurcation has been encountered in the Rosenzweig-MacArthur model, discussed in section 6.1.2. The model analysis that has been presented in that particular section can be summarized as: The model possesses a trivial steady state (F 1 , C 1 ) = (0, 0) which is always a saddle point (see equation (6.47) and the discussion following it).

7.4. HOPF BIFURCATION The model possesses a prey-only steady state (F 2 , C 2 ) = (K, 0) which is stable as long as K < Kc = a( h)

143

(see equation (6.48) and the discussion following it). For larger values of K the prey-only steady state is a saddle point. The model possesses a internal steady state (i.e. with positive abundances of both prey and predator) r (aK( h) ) . , (F 3 , C 3 ) = a( h) a2 ( h)2 K when K > Kc (see equation (6.51) and the discussion following it). This steady state becomes unstable and limit cycles occur when the carrying capacity K exceeds the value Ks = + h ah ( h)

(see equation (6.58) and the discussion preceding and following it). From the above conclusions it is easy to infer that a transcritical bifurcation occurs at K = Kc , where the curves representing the prey-only steady state and the internal steady state, respectively, intersect and moreover the internal steady state turns positive. At K = Kc hence a branching point occurs (see section 7.2), where the prey-only steady state changes from a stable node into a saddle point and the internal steady state exhibits the opposite change (see Figure 7.4). At K = Ks the internal steady state changes stability through a Hopf-bifurcation. This type of bifurcation is characterized by the fact that: a pair of complex, conjugated eigenvalues cross the imaginary axis from the left half into the right half of the complex plane (see Fig. 7.1), such that their real part becomes positive, and at the bifurcation a limit cycle originates. A limit cycle is a closed, invariant loop in the state space. This means that if an initial point is chosen right on the limit cycle, the state of the system will move along the limit cycle for all time. At least for parameter values close to the bifurcation point the limit cycle will be an attractor of the model, meaning that starting from initial conditions close to the limit cycle, the state of the system will in the long run approach the limit cycle and follow it at smaller and smaller distance. In terms of dynamics over time, the limit cycle implies that population abundances will uctuate (see the predator-prey cycles in Fig. 6.8). At the bifurcation point the amplitude of the limit cycle will initially be innitely small, but will increase when moving away from the bifurcation point into the unstable parameter region. Figure 7.4 illustrates the Hopf-bifurcation occurring in the Rosenzweig-MacArthur model (6.38) in a 3-dimensional space spanned by the prey carrying capacity parameter K and the prey and predator abundances, F and C, respectively. For a number of parameter values K the limit cycle is also indicated in this gure. With increasing values of K the amplitude of the limit cycle increases. The (innite) collection of limit cycles for all parameter values K constitutes a parabolic object in the 3-dimensional space, as shown in Figure 7.4.

144

HOOFDSTUK 7. CONTINUOUS TIME MODELS

7.5

Concluding remarks

Even though they are discussed here in the context of 2-dimensional models, these bifurcations are the only types of changes in equilibrium characteristics which can occur in a model when changing a single parameter, irrespective of its dimension. Hence, bifurcation theory learns us that there is order and regularity in the changes that can occur in equilibrium characteristics when changing parameters. It is this order and regularity that helps in understanding the properties of more complex models.

Deel IV

Exercises (in Dutch)

145

Hoofdstuk 8

Oefenopgaven
8.1 ODEs (dierentiaalvergelijkingen) in 1 dimensie

1. Trefwoorden: N (t) en f (N ) populatie en populatiegroeisnelheid evenwichten stabiliteit oplossingen, banen, orbits, trajectorin asymptotisch dynamisch gedrag intile condities e e dynamisch systeem 2. Vragen: Wat is een dierentiaalvergelijking? Hoe analyseer ik een dierentiaalvergelijking?

8.1.1

Populatiegrootte en populatiegroeisnelheid

In deze opdracht bekijken we de populatiegrootte N (t) en de populatiegroeisnelheid dN/dt. Stel dat de populatiegrootte N (t) beschreven wordt door N (t) = 0.16 e0.5t 1. Schets N (t). 2. Wat is de populatiegroeisnelheid op tijdstip t? 3. Druk de populatiegroeisnelheid uit in N (t), en schrijf het antwoord op als een dierentiaalvergelijking. 4. Wat is de per capita groeisnelheid in dit model?

8.1.2

Grasche analyse van een dierentiaalvergelijking

Beschouw de volgende dierentiaalvergelijking: dN = f (N ) dt Voer onderstaande, algemene opdracht uit voor elk van de specieke uitdrukking van f (N ), die beschreven zijn in de erna volgende opdrachten 8.1.2.1 tot en met 8.1.2.6: 1. Schets f (N ) voor deze dierentiaalvergelijking. 147

148

HOOFDSTUK 8. OEFENOPGAVEN

2. Teken in hetzelfde plaatje op de N -as pijltjes die aangeven in welke richting N verandert. 3. Hoeveel evenwichten zijn er? 4. Teken in het N/t vlak pijltjes (vectoren) die de snelheid van verandering in N aangeven, voor verschillende waarden van zowel N als t. 5. Schets oplossingen (banen, orbits, trajectorin) van N (t) in het N/t vlak. e 6. Is het asymptotisch gedrag van N (t) afhankelijk van de initile condities? (asymptotisch e gedrag is het gedrag van het systeem als t naar oneindig gaat) 7. Als er evenwichten zijn; welke is/zijn stabiel? 8. Teken de per capita groeisnelheid als functie van de populatiedichtheid. 8.1.2.1 Immigratie en emigratie f (N ) = I E met I en E constant. Voer de analyse uit voor: a. neem aan dat I > E b. neem aan dat I < E 8.1.2.2 Semi-chemostaat f (N ) = r (K N ) met r > 0, K > 0. 8.1.2.3 Logistische groei f (N ) = rN met r > 0, K > 0. 8.1.2.4 Het Allee-eect f (N ) = bN 2 1 N mN 1 N K

met b > 0, > 0, m > 0, en neem aan dat b > 4m. 8.1.2.5 De drempelvlieg, Bizarrus periodicus alleei f (N ) = sin(N ) N

8.1. ODES (DIFFERENTIAALVERGELIJKINGEN) IN 1 DIMENSIE 8.1.2.6 Structureel instabiel (een pathologisch geval) f (N ) = bN 2 1 N mN

149

maar dan met b = 0.1, = 10, m = 0.25. Voer voor dit laatste model ook nog de volgende opdrachten uit: 9. Vergelijk de volgende twee banen: de eerste begint met N (0) = 1 + , de tweede met 2 1 N (0) = 2 , waarbij een heel klein getal is. Waar eindigen de twee banen voor t ? 10. Wat gebeurt er met het aantal evenwichten bij een kleine verandering in n van de paraee meters?

8.1.3
8.1.3.1

Een model opstellen en grasch analyseren

Zilvermeeuwen

Stel je een populatie zilvermeeuwen (Larus argentatus) voor op Schiermonnikoog. Ze broeden in een kolonie op de grond in een beperkte ruimte, hetgeen leidt tot felle gevechten tussen broedparen om het beperkte aantal nestplaatsen. Hun voedsel daarentegen vinden ze over het hele eiland. Gedurende enkele tientallen jaren hebben studenten van de Rijksuniversiteit Groningen in de Zologiecursus meeuwengedrag in de kolonie bestudeerd, en de geboortesnelheid o (per capita reproductiesnelheid) en sterftesnelheid (per capita mortaliteitssnelheid) gemeten door de demograe van de kolonie bij te houden. Deze dataset laat zien dat doodsoorzaak nummer 1 verwonding is, als gevolg van gevechten om nestplaatsen. De sterftesnelheid is daarom dichtheidsafhankelijk, lineair toenemend met de populatiegrootte met een helling (die je je voor kan stellen als een constante kans om te sterven per gevecht), en een achtergrondsterftesnelheid 0 . De geboortesnelheid daarentegen is niet dichtheidsafhankelijk (waarschijnlijk omdat het voedselaanbod op het eiland geen limiterende factor is voor de populatie zilvermeeuwen): het is een constante snelheid . (Neem hierbij aan dat > 0 ). 1. Stel de balansvergelijking op. 2. Analyseer de dierentiaalvergelijking (d.w.z., voer de opdrachten uit zoals aangegeven in 8.1.2). 8.1.3.2 Intraspecieke concurrentie en interspecieke predatie

De gewone pad (Bufo bufo) is een belangrijke voedselbron voor de ringslang (Natrix natrix). De padden concurreren met elkaar om ruimte en voedsel. Bij deze padden hangt het nadelige eect van concurrentie af van hoe vaak ze elkaar tegenkomen (interference competition). 1. Stel de balansvergelijking op voor de dynamica van de populatiedichtheid van de padden P (per ha). Neem aan dat de slangen een constante populatiedichtheid van S (per ha) hebben. Gebruik de volgende aannamen: De per capita reproductiesnelheid van padden is b (per dag), en de per capita sterftesnelheid m. Hoe vaak padden elkaar tegenkomen hangt af van de populatiedichtheid van de padden, P , en hun beweeglijkheid w (ha per dag). Neem aan dat elke ontmoeting een nadelig eect ter grootte van c heeft op de per capita reproductiesnelheid. De slangen hebben een zoeksnelheid (attack rate) (ha per dag). Uiteraard hangt de mortaliteit door predatie af van hoe vaak padden slangen tegenkomen.

150 2. Teken f (P ) en het faseportret.

HOOFDSTUK 8. OEFENOPGAVEN

3. Herschrijf je model in de vorm van het logistische groeimodel. dP = rP dt 1 P K

Dus: druk de intrinsic rate of increase (r) en de carrying capacity (K) uit in de parameters van jouw model.

8.2. EVENWICHTEN IN 1D ODES

151

8.2

Evenwichten in 1D ODEs

1. Trefwoorden: evenwichtscurve parameterafhankelijkheid linearisatie eigenwaarden evenwichten stabiliteit structureel stabiel/instabiel 2. Vragen: Hoe analyseer ik een dierentiaalvergelijking? Hoe bepaal je de stabiliteit van evenwichten analytisch? 3. Denities: Faseportret: een tekening van representatieve banen samen met richtingspijltjes en evenwichten.

8.2.1

Wiskundige analyse van een dierentiaalvergelijking

Beschouw de volgende dierentiaalvergelijking: dN = f (N ) dt Voer onderstaande, algemene opdracht uit voor elk van de specieke functies f (N ), die beschreven zijn in de voorgaande opdrachten 8.1.2.2 tot en met 8.1.2.6: 1. Druk de evenwichtspunten uit in termen van de parameters. 2. Bepaal de stabiliteit van de evenwichten. Dat wil zeggen: bereken voor elk evenwicht de bijbehorende eigenwaarde(n). 8.2.1.1 Waterplanten en golferosie

1.4 1.2 1 0.8 0.6 0.4 0.2 0 0.2 2 4 6 V 8 10 12

The relative growth rate of submerged vegetation declines linearly with vegetation biomass (V ) due to competition in a logistically growing vegetation. Relative erosion mortality of plants caused by uprooting by waves or animals from unstable sediment approaches zero with increasing vegetation biomass due to consolidation of the sediment, dampening of wave action and exclusion of benthivorous sh. Uit: Ecology of Shallow Lakes van Marten Scheer. 1. Zijn in de guur populatie- of per capita groeisnelheden weergegeven? (Hint: lees de tekst zorgvuldig).

152 2. Schets f (V ) en het faseportret.

HOOFDSTUK 8. OEFENOPGAVEN

3. Om het zwemmen aangenamer te maken, legt men op een goede dag een windsingel aan rond een Fries meertje. Tot ieders opperste verbazing, en het ongenoegen van de dijkgraaf, groeide het meertje binnen korte tijd geheel vol met waterplanten. De waterplanten bleken hardnekkig; verwijdering leidde slechts tot her-invasie. Verklaar dit fenomeen aan de hand van bovenstaand model. 8.2.1.2 Pitchfork bifurcation dx = x x3 dt en is een constante parameter. 1. Hoeveel evenwichten (inclusief negatieve x) zijn er voor: a. > 0 b. < 0 c. = 0 2. Teken f (x) en het faseportret voor = 1 over het domein x (2, 2). Geef ook de stabiliteit van de evenwichten aan. Schets banen in het x/t vlak. 3. Doe hetzelfde voor = 1. 4. Parameterafhankelijkheid, evenwichtscurven: Teken een plaatje dat aangeeft hoe de evenwichten afhangen van . Dat wil zeggen: teken in het x/ vlak voor ieder evenwicht een curve die de lokatie van het evenwicht weergeeft. 5. Bereken de eigenwaarden van de evenwichten als functie van . Bij welke verandert het nulevenwicht van stabiliteit? En wat is de stabiliteit van eventuele andere evenwichten? Geef in de guur die je bij d gemaakt hebt de stabiliteit van de evenwichten aan.

8.3. HET ONDERZOEKEN VAN ODES IN 2D

153

8.3

Het onderzoeken van ODEs in 2D

1. Trefwoorden: stelsels van dierentiaalvergelijkingen fasevlak nul(iso)clines evenwichten faseportret linearisatie jacobiaan eigenwaarden stabiliteit 2. Vragen: Hoe analyseer ik een stelsel van dierentiaalvergelijking? Wat is de fasevlakmethode?

8.3.1

Grasche analyse van een stelsel dierentiaalvergelijkingen

Beschouw het volgende stelsel van dierentiaalvergelijkingen: dx = f (x, y) dt dy = g(x, y) dt

Voer onderstaande, algemene opdracht uit voor elk van de specieke functies f (x, y) en g(x, y), die beschreven zijn in de erna volgende opdrachten 8.3.1.1 tot en met 8.3.1.3: 1. Schets de nulisocline(s) voor f (x, y) en g(x, y) in het x/y vlak (het fasevlak). Let erop dat je aangeeft welke nulisoclines van f (x, y) zijn en welke van g(x, y) (dit kan door de lijnen verschillende kleuren te geven, of getrokken en onderbroken lijnen te gebruiken, etc.). 2. Hoeveel evenwichten zijn er? 3. Teken in het fasevlak vectoren die de verandering in zowel x als y aangeven. De nulisoclines verdelen het x/y vlak in verschillende deelvlakken; teken minstens 1 vector per deelvlak. Het tekenen van vectoren op de isoclines helpt ook. 4. Schets oplossingen (banen, orbits, trajectorin) van [x(t), y(t)] in het fasevlak. e 5. Is het asymptotisch gedrag van banen afhankelijk van de beginwaarden van x en y (de initile condities)? e 6. Bepaal, indien mogelijk, met het faseportret de stabiliteit van de evenwichten. Hoeveel zadelpunten zijn er? 8.3.1.1 Lotka-Volterra competitie f (x, y) = x(1 x y) g(x, y) = y(1 y x)

met x, y, en positief. De volgende 4 gevallen worden onderscheiden: a. < 1, < 1 b. < 1, > 1 c. > 1, < 1

154 d. > 1, > 1

HOOFDSTUK 8. OEFENOPGAVEN

Vergelijk de bovenstaande vier gevallen aan de hand van de volgende vragen: 1. Welke biologische conclusies kun je uit deze modelresultaten trekken ten aanzien van competitie tussen twee soorten? 2. Wat voor ecologische interactie beschrijft het model als en negatief zijn? 3. Neem aan dat < 0, en < 0. Onder welke voorwaarde (uitgedrukt in en ) bestaat er een positief, intern evenwicht? 8.3.1.2 Wortels en kwartels f (x, y) = x2 y met x en y rele getallen. e 8.3.1.3

g(x, y) = x + y 2

Lotka-Volterra prooi-roofdier f (x, y) = x xy g(x, y) = xy y

met x en y positieve variabelen, en , , en positieve parameters.

8.3.2

Wiskundige analyse van een stelsel dierentiaalvergelijkingen

Beschouw het volgende stelsel van dierentiaalvergelijkingen: dx = f (x, y) dt dy = g(x, y) dt

Voer onderstaande, algemene opdracht uit voor elk van de specieke functies f (x, y) en g(x, y), die eerder beschreven zijn in de opdrachten 8.3.1.1 tot en met 8.3.1.3. Bij de stabiliteitsberekening van het interne evenwicht (x = 0, y = 0) mag je uitgaan van de volgende twee gevallen: a. = 2/3, = 3/4, b. = 3/2, = 2. 1. Hoeveel evenwichten zijn er? Druk deze uit in termen van de parameters. 2. Bepaal de stabiliteit van de evenwichten. Dat wil zeggen: bereken voor elk evenwicht de bijbehorende eigenwaarde(n).

8.4. VOORBEELD TENTAMEN

155

8.4

Voorbeeld tentamen

Opgave 1
We bestuderen een wiskundig model dat bestaat uit twee gewone dierentiaalvergelijkingen (ODEs). Het volgende plaatje geeft de evenwichten weer als functie van n van de modelparaee meters:

111 0005 111 000 111 000 111 000 111 000 111 000 C 111 000 111 000

11 00 11 00 11 00

111 000 111 000 111 000 111 000

4 11 00 11 00 11 00 11 00 11 00 3
11 00B 11 00 11 00

111 000 111 000 111 000 111 000

111 000 111 000 111 000 111 000 111 000

111 000 111 000 111 000 111 000

111 000 1 111 000 111 000 111 000

Op de x-as van dit plaatje staat de waarde van de modelparameter, op de y-as de waarde van n van de twee variabelen in het model. De dikke lijnen representeren evenwichtscurves, ee d.w.z. ieder punt op een dikke lijn correspondeert met een evenwicht bij de gegeven waarde van de modelparameter. A, B, C, en D duiden alle bifurcatiepunten aan (cirkels). De kruisen, genummerd van 1 tot en met 6, duiden willekeurige posities (geen bifurcatiepunten) op de evenwichtscurves aan. Neem aan dat het evenwicht dat aangeduid wordt met het kruis met nummer 1 een zadelpunt is. 1. Benoem de bifurcatiepunten A, B, C, en D (d.w.z. geef het type bifurcatiepunt aan). 2. Beschrijf wat er gebeurt met de eigenwaarden van het evenwicht (of de evenwichten) bij punt A en bij punt D als het bifurcatiepunt gepasseerd wordt van links (lage parameterwaarde) naar rechts (hoge parameterwaarde). 3. Beschrijf wat er gebeurt met de eigenwaarden van het evenwicht (of de evenwichten) bij punt B en bij punt C als het bifurcatiepunt gepasseerd wordt van beneden (lage evenwichtswaarde) naar boven (hoge evenwichtswaarde). 4. Teken de fasevlakplaatjes van de evenwichten die met de kruisen 1 tot en met 6 zijn gemarkeerd. Hiermee wordt bedoeld dat voor ieder van de met de kruisen gemarkeerde evenwichten het patronen van de oplossingsbanen (orbits of trajectories) in de buurt van het betreende evenwicht wordt getekend.

156

HOOFDSTUK 8. OEFENOPGAVEN

Opgave 2
In een model voor een enkele populatie, die zich voortplant via seksuele reproduktie, wordt de per capita reproduktiesnelheid gegeven door 1 + N 1 + + N In deze formule geeft N de dichtheid van de populatie aan en de sterkte van het Allee eect dat ontstaat door seksuele reproduktie. Voor = 0 is er geen Allee eect en is de reproduktie van de populatie niet anders dan een populatie die zich aseksueel voortplant. Voor > 0 is bij lage dichtheden N de reproduktie geremd doordat individuen moeite hebben een partner voor de voortplanting te vinden. 4.5 Voor de per capita sterftekans per tijdseenheid van de individuen nemen we aan dat die gelijk is aan (1 + N ) 1. Stel de dierentiaalvergelijking (ODE) van het model op. 2. Teken voor = 0 in n graek de per capita reproduktiesnelheid en de per capita sterfee tesnelheid als functie van de dichtheid N . 3. Geef in deze graek alle evenwichten aan en geef aan wat hun stabiliteitseigenschappen zijn. 4. Neem aan dat . De per capita geboortesnelheid wordt in dit geval gegeven door de formule N 4.5 1+N Teken voor dit geval in n graek de per capita reproduktiesnelheid en de per capita ee sterftesnelheid als functie van de dichtheid N . 5. Geef in deze graek alle evenwichten aan en geef aan wat hun stabiliteitseigenschappen zijn. 6. Vergelijk de onder c. en e. gevonden graeken. Bepaal voor welke waarde van (groter dan 0, maar kleiner dan ) er een drastische verandering optreedt in de evenwichten. 7. Schets het bifurcatieplaatje als functie van .

Opgave 3
We gaan een model formuleren voor de dynamiek van een prooi en een predator populatie. Met F wordt de dichtheid van prooien aangegeven, met C de dichtheid van predatoren. Neem aan dat de prooipopulatie in afwezigheid van predatoren logistisch groeien met groeisnelheid r en carrying capacity K. Neem verder aan dat de sterftekans per tijdseenheid van een predator individu constant is en aangegeven wordt met de parameter d. Predatoren zetten gevangen prooien om in nieuwe predatoren met een conversie-ecintie, die aangegeven wordt met de e parameter . De functionele respons van de predatoren beschrijven met de functionele respons formule van Ivlev. Deze onderzoeker ging er vanuit dat predatoren een maximale snelheid van eten hebben, die aangegeven wordt met de parameter M . Verder nam Ivlev aan dat met toenemende dichtheid van prooien de werkelijk hoeveelheid prooi, die een enkele predator eet, toeneemt van 0 tot dit maximale niveau M volgens een exponentiel verband met het product van een parameter a en e de prooidichtheid F .

8.4. VOORBEELD TENTAMEN

157

1. Schrijf de volledige modelspecicatie op, d.w.z. formuleer de dierentiaalvergelijking (ODE) voor zowel de prooi- als de predatordichtheid.

Door het schalen van de variabelen kan dit model vereenvoudigd worden tot het volgende dimensieloos model:

df dt dc dt

= f

1 ef

(8.1)

1 ef

c c

(8.2)

Dit dimensieloze model zullen we hieronder verder onderzoeken.

2. Geef een vergelijking voor alle nullclines voor de prooi.

3. Geef een vergelijking voor alle nullclines voor de predator.

4. Geef uitdrukkingen voor alle evenwichten die in het model voorkomen.

5. Leid de Jacobiaan van het stelsel dierentiaalvergelijkingen af.

6. Bereken de eigenwaarden voor het triviale evenwicht en het evenwicht waarin alleen prooien aanwezig zijn.

7. Bereken voor deze twee evenwichten ook de corresponderende eigenvectoren.

8. Onder welke voorwaarden bestaat er een evenwicht met zowel prooien als predatoren?

9. Geef aan of en, zo ja, onder welke condities het niet-triviale (interne) evenwicht stabiel is.

10. Schets het bifurcatieplaatje waarin de evenwichtswaarde van de predatoren c staat uitgezet als functie van de parameter .

158

HOOFDSTUK 8. OEFENOPGAVEN

Antwoorden

ANTWOORDEN

159

!
"# $ $%& !

160

HOOFDSTUK 8. OEFENOPGAVEN

'

(

)*+" ! ,

+ - -)#

)*./ * 0)

)0 /



/

/

/

+ - -)#

)*./ .


/ .

ANTWOORDEN

161

)*+" ! ,

)*+" ! ,
/

)

. / 1 / / /

)

/

0

/

)*+" ! ,
/

/

2

/1

)

162

HOOFDSTUK 8. OEFENOPGAVEN

!" 3 3 / / '2 / .(

+

-

/
'2 / .(

+

-

ANTWOORDEN

163

"#$
4

% & ' '

5
*



)


)
$ )




6$ ! 0 '. 1 -)( & *0
) *
$ )
'
)
$
( &


7

(#$' )!#' 7

( . .

/

.

/

(

0

164

HOOFDSTUK 8. OEFENOPGAVEN

ANTWOORDEN

165

./

/

166

HOOFDSTUK 8. OEFENOPGAVEN

* '
+'

,'-.' #'&

+

.

- 81 / -) / * 0)

)0 .

9*) - 1

/ /

/ /

/

/

/

/

)

* . / / / / /

1

/ . . /

. 1 /

1

* )

ANTWOORDEN

167

%' "

/

/

+

-

: *$!$ )$-$ 9*) - /
/ $ $ /
/ /

$ $ / $/ $

/

)

/

.

/

*

,'-.' &''-''

&
0 ' )
(
* - 8 -) * 0)

)0

168

HOOFDSTUK 8. OEFENOPGAVEN

.'')

& 7 ;

<'>(;

)
)
< > ?0
)
'@( & #)
)
>7
AA
C


'
(

'

( 6 *

AA
C

$

'
* (

'
( >& )
0 !

$

*


0 ! * & C )

>>
AA
C

$

'
* (

'

( >?
AA
C

$

'
* (

'
( > 1

*
)
.6 E F/.



*
)
6 E F

*
)
6 E '>





*
)


/6 E * * (

& 1

$
)
G3
1 $


G3
0
H)
-$ 3
3.

ANTWOORDEN

169

$
)

3
3 3.

3/

&
! "! !
!

%#

$

$
" & &

170

HOOFDSTUK 8. OEFENOPGAVEN

'

( ( / ) ( .

'

" $ ' ) $ ' ')

!

F



0 * " &
&

" /"

!

" */ ' * " */
#

6
" I !

#

0
I

"

!

!

6

0

) )

* "

* / "

)$ 0

!
" /" #

/ #
)

*0
: ))E 1 -) $

# !$

Part V

Instructions for using CONTENT

171

Chapter 9

CONTENT: continuation software

9.1 How to use this manual

The rst part of this document gives a step-by-step description of several basic system manipulation features of Content 1.5. Until you get the hang of it, it is a good idea to follow the steps in the manual. The second part is more of a general howto-document. It can be read as a guideline, and is by no means comprehensive. It provides a springboard to get you familiarized with the program, from which you should be able to start exploring its darker corners. Finally, CONTENT has an built-in Help function, which you always consult. However, the text is rather technical and less helpful for the beginning user. Use therefore this manual, and ask for help if you still have trouble.

9.2
9.2.1

How to start using CONTENT

Setup of CONTENT program

1. Go to Blackboard/Course Documents/Files for Computer Practical. 2. Download ContentProgram (save the .zip le on your Desktop). 3. Extract the .zip le. 4. In the Content folder, nd the program called SETUP; double-click the icon. 5. Hit Run in the window that pops up 6. In the content 1.5 setup window, mark Full under the header Installation. 7. Click Setup. 8. After this process is nished, you will nd a program called CONTENT in the same folder (le type is Application). 9. Create a shortcut to the program on your Desktop (you can do this by right-clicking the CONTENT icon, and selecting the option Create Shortcut, then drag the shortcut to your desktop). 10. Start the program by double-clicking the shortcut. 173

174

CHAPTER 9. CONTENT: CONTINUATION SOFTWARE

9.3

On startup of Content 1.5

1. Select the main window (titled Content 1.5). 2. Choose Select from the list of menus. 3. Choose Class and check if the selected class of models corresponds to the system type you intend to investigate. If it doesnt, change it. If you start up Content for the rst time, the class is set to Iterated maps. Make sure that you set this to ODEs (i.e. ordinary dierential equations) to follow the steps in the next subsections. 4. Now choose system from the same menu. A new window appears titled System. 5. Select the system you want to load by clicking on it and then click Ok. Alternatively, you can click Cancel to return to the main interface window.

9.4
9.4.1

Basic system manipulation in Content 1.5

Entering the system management window

1. Select the main window (titled Content 1.5). 2. Choose Select from the list of menus. 3. Choose Class and check if the selected class of models corresponds to the system type you intend to work with (i.e. ODEs). If not, change it so that it does. 4. Now choose system from the same menu. A new window appears titled System. From here you can proceed to manipulate existing systems or create new ones.

9.4.2

Create a new system

We will use the model of section 10.1, which is used to explore the consequences of dierent harvesting strategies for a population of Cod, as an example: 1. Go to the system management window. 2. Select the menu Actions and click on New 3. Enter a name for your model in the dialog box that appears and click Ok. You can choose any name you want. Given the example model we focus on, Cod would be an obvious choice for a name. 4. Enter the specication of your model, that is, dene its variables (Coordinates), parameters and the time variable and specify the equations of the right-hand side of your dynamic equation. 5. Entering a model in Content 1.5 is quite straightforward. For example, consider the model presented in section 10.1: dC C = rC(1 ) (9.1) dt K To implement this model in Content we would specify C in the box Coordinates of the model implementation window, r,K in the box Parameters and t in the box Time. In the main (largest) box of model implementation window the dierential equation has to be implemented as:

9.5. COMPUTATION C= r*C*(1-C/K);

175

Models are entered in Content 1.5 using the C programming language. Therefore, every line in the model implementation box has to end with a semi-colon (;). Furthermore, it is possible to use all functions and operators available in the C programming language. For example, if the dierential equation contains an exponential function of the model variable C, like eaC , this can be implemented in Content as exp(-a*C). 6. When youre done click Ok, if you change your mind select Cancel. When you are done with implementing your model you can continue with its analysis by following the step in section 9.5 below. The next subsection explain how you can edit or change, rename or remove a model that you have already implemented. In general you have to consult these sections only after you have already carried out some analysis with the program.

9.4.3

Manipulating already dened systems

To edit an existing system, follow the steps below: 1. Go to the system management window. 2. In the list of systems that appears, select the one you intend to edit. 3. Select the menu Actions and click on Edit 4. You are now presented with the system and you can start editing. Remember though that all curves and diagrams stored with this model will be deleted if anything is changed in the variables, parameters or time list. To delete or rename an existing system, carry out the following steps: 1. Go to the system management window. 2. In the list of systems that appears, select the one you intend to rename or delete. 3. Select the menu Actions and click on Delete to delete the selected system, or on rename to rename it. MAKE SURE you do not accidentally hit the Delete All option.

9.5

Computation

To compute curves using Content 1.5 several settings need to be dened. First of all, you have to tell the program the type of curve you intend to compute, and the type of point from which you will start the computation. This is set by choosing TypeInitial point and TypeCurve from the main Content 1.5 window. Notice that the currently selected point type and Curve type are also listed in the main Content 1.5 window. Second, you have to provide the initial values of all the necessary variables, parameters and computational settings. The computational settings have default values that suce for fairly standard computations, but may need tweaking for more complex systems.

176

CHAPTER 9. CONTENT: CONTINUATION SOFTWARE

9.5.1

Orbits

The most basic computation type that Content 1.5 performs is the Orbit. This is a trajectory of the system in time from set initial conditions. To perform this type of computation, go to TypeInitial point and select point, and then choose Orbit from the menu TypeCurve. Two new windows show up on your screen, called Starter and Integrator. As the names suggest, one contains all system-specic information (like parameters etc) needed to start the requested computation, while the other contains the specications of the computation itself, for example the length of time that you intend your orbit to continue. The integrator contains one menu item, called Method, which allows you to choose the integration method to apply. If all the settings for your calculation are entered, you can start the computation by selecting Computeforward or Computebackward from the main window or any output window. The computation will run until it reaches the amount of time you specied in the Integrator window, until you interrupt it, or until one of the error thresholds of the integration routine is crossed. You can interrupt a computation by hitting the space bar, or end it by hitting the escape button. Hitting the space bar while in pause mode will resume the paused calculation. Computeforward calculates in the same direction as the previous calculation, while Computebackward computes in the opposite direction. If there is no previous calculation, Computeforward goes towards innity while Computebackward goes towards minus innity. Computeextend computes the same curve as the last computed one, starting at the last point of the previous computation. Thus, if you initially calculated t = 0 to t = 100, Computeextend will compute t = 101 to 200. Note that this does not create a new curve, it appends the results to those of the rst curve.

9.5.2

Equilibria

If the system converges to an equilibrium point during the computation of an orbit, Content 1.5 allows you to compute the properties of that equilibrium when one or more parameters are changed. In order to do this, you need to tell Content 1.5 to use the end point of the computed orbit as the starting point of a new calculation. First of all, rename the orbit you have just computed. After you have done that, go to Select Initial point in the main window. A window appears showing all the curves in the current diagram, and a list of special points for each curve. For an orbit, the last point is the only point listed. Select it by clicking on it and then click on OK. To compute an equilibrium curve, select Point as the initial point type and Equilibrium as the curve type. Note that the two calculation windows that appear are dierent from those when calculating an orbit. The Starter window now gives you the opportunity to select the parameter that you want to vary. this is done by clicking the mouse on a parameter. If it is selected, it turns black. notice that you can select more that one parameter. For equilibrium curves however, only one parameter should be selected. Clicking on a selected parameter will deselect it. Furthermore, there are 2 new sections in the starter window that are of importance. In the section Monitor singularities you can turn on or o the option to check whether a point on the equilibrium curve is a bifurcation point. As the names suggest, branching detects branching points, limit point detects limit points (of fold bifurcations) and Hopf detects Hopf bifurcations. In the section Eigenvalues you can toggle the option to compute the eigenvalues of the system. This is a very handy option, as it allows you to plot equilibria in the complex plane. The sections that are not discussed do not matter for now, their default values suce for all computations you will encounter. Instead of an Integrator window, a new window named Continuer has appeared. Here the parameters are set for the algorithm that is used to continue the selected equilibrium in parameter space. Note that the step-size parameters are the same as when calculating an orbit.

9.6. STORAGE OF COMPUTED CURVES

177

The setting titled MaxNumPoints denes after how many integration steps the calculation stops. There is no 1:1 correspondence between the value of the varied parameter and the number of points, since the continuation algorithm uses a variable time step. Generally, it is safe to set the value for MaxNumPoints to 10000. Once all settings are entered correctly, you can compute the equilibrium curve by choosing Computeforward, Computebackward or Compute extend. When all the options under Monitor singularities are set to yes, the computation will automatically pause when one of these points is encountered.

9.5.3

Bifurcation points

These bifurcation points themselves can then be used as initial points and can be continued in parameter space. To do this, stop the current equilibrium calculation, select the bifurcation point as the initial point, and set TypeCurve and TypeInitial point so that they match the bifurcation you intend to continue. Lets assume you have encountered a Hopf bifurcation. After you have selected it as the initial point, and you have set TypeInitial point and TypeCurve to Hopf, a new Starter and Continuer window pop up. Both are essentially the same as when you computed an equilibrium curve. Note that in the Starter window, there are now 2 new entries in the Monitor singularities section, Generalized Hopf and BogdanovTakens. These are higher order bifurcations that will not be discussed here. To continue the Hopf bifurcation, you will need to select 2 parameters in the starter window, since a Hopf bifurcation is of order 2. Again, to compute the curve select Computeforward, Computebackward and Computeextend.

9.6

Storage of computed curves

Content does not contain the familiar options load and save to store and retrieve results that you have computed earlier. Instead, Content stores computed curves by giving them temporary names. Curves with such temporary names are referred to as untitled curves. To view and access the curves that you have computed open the Curve window, which you reach by choosing SelectCurve from the main Content 1.5 window. This window shows you a list of all the computations within the current diagram, with the most recent one on top. Names of untitled curves are as follows: ***Point type >Curve type(direction of calculation) There is however a maximum to the number of untitled curves that Content stores. If you compute more curves than the maximum number of untitled curves, the oldest curve will be deleted each time you calculate a new one. Note that calculating the same curve twice counts as calculating two curves. Also note that if you compute a new curve, it automatically becomes the currently selected curve. Thus it is not possible to overwrite a curve by computing a new one. You can change the maximum number of untitled curves by going to the menu Options on the main Content window and selecting Archive lter. In the option window that opens up the value of Maximum number of untitled curves determines the number of untitled curves that the program will store. Its default value is 5. Do not set this value too large, it rapidly consumes the computers memory and may slow down further computations. Path lter sets the number of data points that are remembered per curve. It is best to set the number of points to 32000 and step to 1. In the list of names you can edit which calculated numbers are remembered. In general you probably want to remember the time, all coordinates and all parameters, and the

178 eigenvalues.

CHAPTER 9. CONTENT: CONTINUATION SOFTWARE

You can permanently store an untitled curve by changing its name. Renaming and deleting curves is done from the Curve window, which you reach by choosing SelectCurve from the main Content 1.5 window. Once you have changed the name of a curve to something other that its temporary name (see above) it will only be deleted if you explicitly instruct the program to do so. A good way of organizing stored computations is by using diagrams. these are analogous to the more familiar folders. They can be dened anew, renamed and deleted, and may contain any number of computed curves. Keep in mind however that it is not possible to move computed curves from one diagram to another. To view the contents of the currently selected curve, choose WindowRedraw curve from the appropriate output window, or from the main Content 1.5 window to redraw it in all open output windows. To view the results of all curves within the current diagram, choose WindowRedraw diagram. See the section on viewing output for further details.

9.7
9.7.1

Special cases
Limit cycles

In addition to numerical investigations of xed points, Content can also be used to calculate properties of limit cycles. If you nd a Hopf bifurcation point, it is easy to pick up the limit cycle that departs from it, and continue it in parameter space. Select the Hopf bifurcation point as the initial point (select it by choosing from the list under Selectinitial point, and set TypeInitial point to Hopf), then set TypeCurve to limit cycle. Now, Computeforward, Computebackward and Computeextend can be used to continue the limit cycle and its properties in parameter space. Note that in this case, you must select only one parameter in the starter window. You can also start limit cycle continuation directly form an orbit computation, however, he procedure to do this more complex and usually it is easier to rst nd the Hopf bifurcation where the limit cycle originated, and continue it from there.

9.7.2

Continuing branching points

Unfortunately, it is not straightforward to compute curves in a two-parameter plane that represent the location of a branching point (BP) with the current version of Content. A work-around is needed to make Content compute these curves, in which we will exploit the feature of Content that it can deal with a user-dened function. To illustrate the procedure of computing branching point curves in a two parameter plane, consider the following classical predator-prey model: dF = rF dt F K aF C 1 + ahF

(9.2a)

aF dC = C C dt 1 + ahF

(9.2b)

This so-called Rosenzweig-MacArthur model is discussed in detail on page 118-124 of the syllabus. The per-capita growth rate of the predator population equals:

9.7. SPECIAL CASES

179

g(F ) =

aF 1 + ahF

(9.3)

and a branching point occurs when the equilibrium, in which prey and predator coexist, coincides with the equilibrium with only prey present. In other words, the condition determining the branching point is that g(F ) = 0 (9.4) and C = 0 (9.5) Follow these steps to continue this branching point as a function of the carrying capacity K and the predator mortality : 1. Dene a user function by selecting SelectUser functions.... This opens a new window that allows you to create, modify and delete user-dened functions. Create a new function by entering a label and a name, for example, pgC (use the same for both label and name) and pressing the Add button. In the edit portion of the window you will now see: *Value=?; return 0;

2. Program here the expression for the per capita growth rate of the predator. You can program in much the same way as when dening your system of equations. In the case of the Rosenzweig-MacArthur model the following will do: *Value=eps*a*F/(1+a*h*F)-mu; return 0; This depends of course on which names you have given to your parameters, when dening the system of equations in Content, but that is self-evident. 3. Press Ok. If you did not make any mistakes the program will now compile the userdened function. 4. When successful, the Starter window of Content will now contain an additional section User dened function, where the function pgC should now be visible. Next to the name there is an option that most likely reads ignored when you just dened the function, but which you can change to monitor, append or detect by clicking on the option. 5. Also, if you open a numeric window you can now make the valuable of the user-dened function visible by selecting WindowLayout and by selecting the function pgC under the category user-dened functions. Notice that the option in the Starter window should equal monitor, append or detect to really calculate the value of the userdened function and hence to display a value at all. 6. To continue the curve of branching points you start by locating a prey-only equilibrium, in which C = 0. This equilibrium you subsequently continue as a function of K, but before doing so choose the option of the user-dened function pgC equal to detect. 7. When computing the prey-only equilibrium as a function of K Content will now stop when it detects that value of K, for which pgC equals 0. As explained above, this occurs at the branching point.

180

CHAPTER 9. CONTENT: CONTINUATION SOFTWARE

8. Now you have the choice of continuing the computation, but you can also switch. If you choose switch notice that the option next to the function pgC changes to append. Content now appends the condition g(F ) = 0 (i.e. where pgC=0) to the system of equations and nds all the points as a function of two parameters with that property. 9. You have to activate two parameters, so highlight also the mortality parameter mu. Moreover, to prevent the occurrence of superuous messages turn of the detection of all types of bifurcations (Branching points, Limit points and Hopf bifurcation; In the Starter window you can set their detection to no). 10. Open up a graph window with K and on the axes and choose Computeforward and Computebackward. The resulting curve is the curve of parameter combinations, for which both C = 0 as well as its per capita growth rate equals 0 in equilibrium. These represent all branching points.

9.8

Visualizing output

The two main ways of visualization in Content 1.5 are the Numeric and Graph window. Both can be dened to show any computational results you are interested in. To open a new graph or numerical window, select Window2d graph, Window3d graph or Windownumeric. Numeric windows are quite straightforward. They show the current value of a number of properties of your system as the computation runs along. By selecting WindowLayout from a numeric window, you can change its name and the quantities it shows. 3d graphs are rarely used in Content, as they tend to be messy and hard to read, so I will not discuss them here. If you want to visualize complex output results in Content, conguring an appropriate graph window can take some time. In principle, graphs dened in Content are permanent, and are stored with the system you are studying. However, they are not written to disk until you exit the program. This implies that if Content crashes, any newly dened output windows are lost. To avoid any frustration, it is best to exit the program and restart it every time you dene a new output. Another important thing to consider is that if you click on the close button in the upper right corner of an output window, it is lost. To avoid getting a very crowded screen, use Windowhide in the appropriate output window to get it o your screen. You can then select Windowunhide from the main window to unhide it. When you use many graph windows simultaneously, it can also be convenient to change window names. This is done under Window Layout. Here you can also change the background color and several other properties of the graph window. To dene a graph that plots time versus a variable X, open a new 2d graph, then choose AttributesCurrent. You will see a window in which you can assign functions to the axes of the graph. To set time on the x-axis, enter it in the box next to Abscissa (the x-axis). Alternatively, click once in the Abscissa box, and then double click on time, which will show the name you assigned to the time in your model formulation (usually t). double click it to enter it as the x-axis variable. Notice that by clicking the New function button, you can add a new plot to the same graph window. It it thus possible to plot for examples all variables as a function of time in one graph window. In the lower section of the Attributes window you can change color and line properties for the current function. After you change anything here, use the Change color button to activate the changes.

9.8.1

Special case 1: plotting limit cycles

When Curvetype is set to limit cycle, a new item, Standard functions is added to the list of names in the upper right corner of the graph Attributes window. This contains

9.8. VISUALIZING OUTPUT

181

several characteristics of the limit cycle you may be interested in, such as the period, minimum and maximum. For plotting a limit cycle as a function of 1 variable and 1 parameter, it is usually more convenient to plot only its minimum and maximum value. You can do this by dening 2 new functions, one plotting the cycle minimum and the other the maximum versus the parameter.

9.8.2

Special case 2: plots of eigenvalues in the complex plane

In many situations these plots can give you much insight into exactly how equilibrium stability changes with parameters. To dene such a plot, rst go to Windowlayout and switch on the option draw unit cycle. This draws a circle with unit radius in the background of the plot window. In the graph window, set the scales for the x- and y-axis from -1.5 to 1.5. Then go to AttributesCurrent. In the list of names, double-click Eigenvalues. Supposing you are investigating a 2-dimensional system, this gives you a list consisting of Re 1, Im 1,Re 2 and Im 2, which are the real and imaginary parts of the rst and second eigenvalues of the system. Now dene 2 functions, each plotting the real parts of the 2 eigenvalues along the abscissa (xaxis) versus the corresponding imaginary parts on the ordinate (y-axis). Now you have a plot showing how the eigenvalues of your system dance around the origin in the complex plane, as you change a parameter (by calculating an equilibrium curve).

182

CHAPTER 9. CONTENT: CONTINUATION SOFTWARE

Part VI

Computer Labs

183

Chapter 10

Computer Labs
10.1 Harvesting Cod

In 1497 the ocean around Newfoundland (Canada) was once so full of cod that explorer John Cabot marveled that they virtually blocked his ship. In 1992 the Canadian cod industry collapsed. Years of over-shing had reduced the stock to dangerous levels, and the Canadian government was nally forced to put a stop to the drastic over-shing. 40,000 people lost their jobs. In 2001 Cod stocks in the North Sea are down to one-tenth of the level 30 years ago. Scientists fear that North Sea cod may go the way of those in the Grand Banks of Canada, where over-shed stocks disappeared in 1992 and have not revived. Consider the following dierential equation: dC = rC dt 1 C K

Here C stands for the Cod population (in 100.000 tons of cod), r is the intrinsic growth rate (per day), and K is the carrying capacity (in 100.000 tons of cod). This is the familiar logistic growth equation that describes the growth of most self-limiting populations. We are interested in describing the qualitative behavior of the cod population of the North Sea and wish to make a few statements about shing and more importantly about the risks of over-shing. For the sake of simplicity, we will start with this simplest of models. 1. Use Excel (the worksheet RHS-1-ODE.xls) to investigate the properties of the righthand side of the ODE for dierent values of r and K. Which equilibria can you nd? How do they change as a function of K and r? 2. Now implement the model in CONTENT. Compute some orbits for dierent parameter combinations of K and r. Use this as an exercise to familiarize yourself with CONTENT. 3. Subsequently, try to construct a graph of the equilibrium as a function of K.

Fishing - Constant eort

Lets play around with the logistic growth equation for a bit to study the eects of shing mortality. Assume that the shing mortality is such that sherman make a constant eort per day and so harvest a fraction (f ) of the actual cod population in the water. 4. Think about what this means. How would you change the logistic growth model so as to include this? 185

186

CHAPTER 10. COMPUTER LABS

5. Implement this model also in CONTENT & Excel. Parameters are: K = 10, r = 0.1, f = .01. 6. Start o by making several numerical integrations (orbits) of this model. Make sure to start at several dierent initial conditions. How many equilibria does the model have? How stable are they? What is the eect of increasing f ? 7. Now move on to continuing equilibria in CONTENT. Select a nal point of a computed orbit to start continuing the equilibrium curve. Do not forget to change the point and curve type. 8. Follow the equilibria of this model as a function of f . Can you link what you see here to what you have observed previously?

Fishing - Constant quota

The situation in the North Sea is not as quite as rose-colored as shown above, at least not from a shs point of view. The fact of the matter is that governments set xed, yearly, quotas (Q) for the capture of sh and so shing pressure is more or less a constant. 9. Think about what this means. How would you change the existing model for sh growth in the North Sea so as to include this constant shing pressure? 10. Implement the model in CONTENT. Parameters are: K = 10, r = 0.1, Q = 0.05. Do you still have a logistic or modied logistic growth model? 11. Make several orbits with dierent initial conditions. How has the behavior of the system changed from that of the logistic model? Next, set Q to 0.1, 0.2, 0.25, and 0.26. What happens to the sh population for each of these quotas? 12. Select an orbit for parameters that support a viable sh population. Select as initial point the end point of the orbit to start continuing an equilibrium curve. Do not forget to change the point and curve type. 13. Open a new 2D-plot window and select Q and C for the axis variables. Plot the equilibrium. What is stable? Unstable? Relate what you see now to what you have observed previously. 14. What is the eect of a) faster sh stock re-growth and b) an increased sh carrying capacity on the equilibrium curve? (tip: see the section on Bifurcation points in the CONTENT manual). 15. Is a constant quota a good thing for sh? For the shing industry? Compare with the constant eort model.

10.2. THE SPRUCE-BUDWORM(CHORISTONEURA FUMIFERANA)

187

10.2

The spruce-budworm(Choristoneura fumiferana)

The spruce budworm (Choristoneura fumiferana) is a moth that inhabits northern American pine forests. Normally, the density of these moths is not very high, and their impact upon the forest is small. However, approximately once every 40 years, their density increases explosively, and entire forests are picked clean of needles in a very short time span. Obviously, this has severe negative eects on local timber production. The major predators of these budworms are birds, and at high densities budworms make up a major part of the birds diet. The budworms themselves eat the newly emerging needles of pine and spruce trees. The following model has been proposed to explain the outbreaks of budworm population dynamics. N EP N 2 dN = rN (1 ) 2 dt K N0 + N 2 with K = cA and N0 = f A. Both K and N0 depend on A, which is the leaf (or needle) density of the forest. Of course, in reality A is not a constant. But the dynamics of the forest are on such a slow timescale compared to the dynamics of the budworms, that we treat it as such. P , the density of the predatory birds, is also assumed constant. this is because if budworm density is low, the birds either switch to other prey, or migrate out of the area. Either way, there is no numerical response of birds to changes in budworm density. A valid parameter setting for this model is: A = 0.5, c = 20, E = 0.314, f = 0.474, P = 0.7 and r = 0.1. This parameter setting reects the conditions in North American pine forests. 1. Implement the model in CONTENT and open an 2D-plot window with time on the x-axis and the density N on the y-axis. 2. Compute several orbits for the default parameter set, while varying the initial value of N . Make sure to study not only high population densities, but also to consider the range N = 0.0 . . . 0.1. 3. Repeat this procedure for various values of A between 0.1 and 5.0. What is the most remarkable feature that you encounter? 4. Compute now also several orbits for the default parameter set and various initial values of N , but going backward in time (Use Computebackward). What do you reach by carrying out this backward computation? 5. Now turn to Excel (use the worksheet RHS-1-ODE.xls) and investigate the right-hand side of the ODE as a function of N . Do this for a number of dierent values of A between 0.1 and 5.0. Explain the results you obtained by integration with CONTENT using the changes you observe in the graphs of the right-hand side of the ODE. 6. Turn back to CONTENT. Select one of the nal points of an orbit to start a continuation of the equilibrium as a function of A. Do not forget to change the point and curve type. Relate your ndings of the integrations to the picture you have constructed. 7. Explain in words the mechanism behind the regular outbreaks of the spruce budworm in North American forests. 8. One way to prevent outbreaks is to spray the forest with insecticides. Assume that this would make a smaller fraction of the foliage available for consumption by the budworms,

188

CHAPTER 10. COMPUTER LABS i.e. change the value of c. We can obtain insight about the dynamics of the model for various combinations of A and c by continuing the limit points (LP) as a function of these two parameters. Open a 2D-graph with the two parameters on the axis. As initial point now choose one of the limit points that have been computed and continue it.

9. Interpret the graph you have just constructed. To that end, sketch the bifurcation graphs of the density N as a function of A for dierent values of c. What does the point denoted Cusp represent? What type of structural change occurs at this particular point? 10. Repeat the limit point continuation for the parameters combination A and P . P could be increased by providing more nesting sites for birds. Is this an eective strategy to prevent budworm outbreaks?

10.3. INTERSPECIFIC COMPETITION

189

10.3

Interspecic competition

Competition is one of the most studied subjects in ecology. It is generally thought to be one of the key factors that shape community structure and dynamics. The foundations of competition between species are illustrated by the following model: dN1 dt dN2 dt N1 + 12 N2 K1 N2 + 21 N1 1 K2 1

= r1 N1 = r2 N2

1. Here N1 and N2 are competing species. r1 and r2 are the growth rates, and K1 and K2 are the carrying capacities of species N1 and N2 respectively. What is the function of the parameters 12 and 21 ? 2. Examine in CONTENT the behavior of the model (compute orbits for dierent initial conditions). Set the parameters r1 = r2 = 0.1, K1 = K2 = 100 and 12 = 21 = 1.0. What is going on with these species? Under what condition does one or the other win the competition, and why? 3. Compute an orbit starting from N1 (0) = 1 and N2 (0) = 3. Then change the value of K1 to 101 and compute the orbit anew. What is the dierence between these two orbits? 4. Repeat the previous exercise, but now computing the orbits from t = 0 to 7000. The orbit you just computed for K1 = 101 can be divided into two distinct types of dynamics. Which types? Which processes are important in each? 5. How does the equilibrium you just found depend on K1 ? 6. Use the Excel worksheet Analysis-2-ODEs.xls to examine the zero-isoclines and the vector eld of the system. Which qualitatively dierent isocline congurations are possible when K1 changes, and which types of dynamics do you expect for each conguration? 7. What is the isocline conguration for the symmetrical case (r1 = r2 = 0.1, K1 = K2 = 100 and 12 = 21 = 1.0) that you studied above? Does this corroborate your previous results? 8. Calculate the isoclines for r1 = 0.1, r2 = 0.08, K1 = K2 = 100 and 12 = 21 = 1.2. What are the dynamics you expect from the phaseplane analysis? Conrm your expectations in CONTENT. 9. Use CONTENT to construct a graph of all equilibria as a function of the parameter K1 . As starting points you can exploit the fact that both N1 = 0, N2 = K2 and N1 = K1 , N2 = 0 are obvious equilibria that might be stable for some ranges of parameters. Use the branching points that you encounter to locate all equilibrium curves. Explain for which ranges of K1 the equilibria you nd represent a stable or an unstable equilibrium. 10. What are the conditions (in words, not math) for coexistence between the species in this model?

190

CHAPTER 10. COMPUTER LABS

10.4

Vegetation catastrophes

In semi-arid regions of the world like the Sahel vegetated areas are slowly giving way to desert regions. One of the hypotheses concerning the disappearance of vegetation cover in these areas is that this is under the inuence of grazing pressure from cattle. A mystery has been why vegetated areas do not return when grazing stops. Semi-arid regions are characterized by sparse rainfall. Plant growth is water limited and so their growth rate is primarily a function of water availability in the soil. Water inltration into the soil in semi-arid areas has a positive relationship with plant density: the higher vegetation biomass is, the more water inltrates into the soil. Rietkerk et al (1997) described vegetation dynamics as a function of water inltration into the soil, plant growth and herbivory. dW dt dP dt

= Win (P ) cmax g(W ) P rw W = gmax g(W ) P d P b P

Win (P ) is a function describing inltration of water into the soil as a function of rainfall (R) a half saturation constant k2 , and Wo , the minimum water inltration in the absence of plants. g(W ) describes the growth of plants as a saturating function of water availability in the soil, with a half saturation constant k1 . Win (P ) = R g(W ) = P + k2 Wo P + k2

W W + k1

Default parameters are: gmax = 0.5, k1 = 3, d = 0.1, cmax = 0.05, R = 2, rw = 0.1, k2 = 5. The parameter b indicates the grazing pressure of cattle on the vegetation (i.e. the herbivory) and is human-controlled. The main question of this computer lab is how the interplay of grazing and water availability aects vegetation biomass. Steps you may take to investigate this are: 1. Start with a high minimum water inltration into the soil: Wo = 0.9 and no grazing. Describe the relationship between plant biomass and soil water availability with increasing herbivory. What do the isoclines look like: how do they change with increasing herbivory? 2. How does lowering the minimum water inltration into the soil (Wo = 0.2) aect the patterns you see? What happens as grazing pressure increases? 3. Illustrate the dierent qualitative behaviors of the system with pictures of isoclines. Summarize these dierent states as a function of Wo and b.

10.5. LOTKA-VOLTERRA PREDATION

191

10.5

Lotka-Volterra Predation

The classical equations concerning predator-prey interactions (also: consumer-resource interactions) are: dR dt dC dt

= rR aRC = caRC C

And are attributed to Lotka (1932) and Volterra (1926). Here C stands for consumers, and R for resources. r is the rate of resource replenishment. C attacks R with a type I functional response (non-saturating: i.e. constant): a. The conversion rate of sequestered R biomass to new C individuals is c, and is the consumer death rate. 1. What are the possible equilibria? Examine isoclines. 2. Implement the model in CONTENT with parameters r = 0.5, a = 0.2, c = 0.5 and = 0.1. Examine the stability of the equilibrium to disturbances: is this equilibrium stable? Unstable? Hint: (look at eigenvalues). 3. Examine prey growth in the absence of predators. What do you notice? Is this realistic? Why? We now assume that resources grow according to the logistic growth model, equilibrating at a carrying capacity K in the absence of consumption by consumers. 4. What are the equilibria now? What can you say about the feasibility of equilibria? Draw isoclines. How has the stability of the equilibria changed? 5. Implement the model (in CONTENT) with parameters as above and K = 10. How does the model behave dierently from the previous one? 6. Examine equilibria as a function of K. Pay special attention to the point BP in CONTENT: it marks the intersection of two equilibria, and a subsequent change in stability. 7. Examine the approach to the internal (two-species) equilibrium. What do you notice at small K (=1.1) and at very large K? 8. What do you notice about the isoclines at extremely large values of K? Why does this happen? Link the behavior of the two models.

192

CHAPTER 10. COMPUTER LABS

10.6

The Paradox of Enrichment

One of mans greatest eects on ecosystems is caused by non-point pollution of surface waters. Runo from farms and cities enrich lakes (and other bodies of water) with phosphorous and nitrogen. This has far reaching eects for species composition and the persistence of ecosystems. We will examine the eects of an increase of nutrient input into a lake populated by phytoplankton prey (P ) and Daphnia predators (D) feeding on phytoplankton. dP dt dD dt P K

= rP

f (P ) D

= f (P ) D D

Phytoplankton grows logistically with growth rate r and carrying capacity K. We assume that phytoplankton are nutrient limited, i.e. that a linear increase in phosphorous and nitrogen levels translates into a linear increase in the equilibrium levels that phytoplankton can attain if growing on their own (K). f (P )is a function determining how many phytoplankton are eaten by Daphnia given an encounter, is the conversion eciency of phytoplankton to Daphnia biomass, and is the Daphnia death rate. 1. Assume that Daphnia attack phytoplankton according to a satiating, type II functional response. At low prey densities Daphnia are limited by the amount of prey they can nd while at high prey densities, Daphnia are limited by the time it takes to handle and digest a prey individual. A function describing this is: f (P ) = aP 1 + a Th P

with a, the attack rate and Th the time it takes a predator to attack and digest a prey individual. 2. Implement the model rst in the Excel worksheet Fasevlak-2-ODEs.xls, with parameters r = 0.2, = 1, a = 0.02, = 0.25, and Th = 1. Examine the eect of increasing K. What equilibria are possible? Are they stable/unstable? When is coexistence possible? When not? Now examine your ndings in CONTENT. Make several orbits with dierent initial conditions. What is the eect of increasing carrying capacity? 3. Examine your equilibria as a function of K in CONTENT. 4. Examine eigenvalues of the model as a function of K. What can you say about the relationship between their values and the observed dynamics of your system with increasing K? Make a diagram of the types of observed dynamics in relationship to K. 5. What is a Hopf bifurcation? 6. Point out when a Hopf bifurcation occurs in terms of the conguration of your predator/prey isoclines. 7. What is the paradox of enrichment? 8. Why does the paradox of enrichment occur? (In words or in math).

10.6. THE PARADOX OF ENRICHMENT

193

9. Construct the Hopf bifurcation boundary as a function of K and . This boundary separates the parameter combinations for which limit cycles occur from those for which the equilibrium is stable. 10. Sketch with pen and paper a bifurcation diagram as a function of , based on your (, K)graph and the bifurcation diagram as a function of K. 11. A challenge: Use CONTENT to locate the existence boundary of Daphnia as a function of K and . 12. Another challenge: Use a 3D-plot to visualize the limit cycle as a function of K.

194

CHAPTER 10. COMPUTER LABS

10.7

Chaotic dynamocs of Hare and Lynx populations

The following set of equation describes the interaction between the hare (H) that graze the vegetation (V ) and are preyed upon by lynx (L). dV dt dH dt dL dt = aV 1 V Vmax a1 V H 1 + k1 V

= a1 V

H bH a2 HL 1 + k1 V

= a2 HL c (L Lmin )

You dont need to program this model yourself in Content, you can download a le with a readymade program from the Blackboard site. Download the le and store it somewhere on disk. The le can be imported into Content as follows: Choose SelectSystem and in the new window: ActionsImport. Now browse to the location where you stored the le Lynx.es and select. Press OK to import the system into Content. Subsequently edit the model such that Content uses numerical derivatives to carry out computations. 1. Click WindowShow hidden... and select L-t plot to make time orbits (1000 timesteps) of this system. Slowly increase c from 7 to 12. At what values of c do you start to see dierences in dynamics? 2. Now vary initial values of V by a very small amount ( 0.00001). Use these two slightly dierent initial values of V at every value of c. What dierence do you see in nal values of the variables when c is raised by 0.1 between 10 and 11? (Look at the numerical window at the end of each run and use dierent colors for dierent orbits). 3. Open the 3D phase space window and look at the changing variables V , H and L over long time-periods for high c values. 4. Choose SelectCurve and select the curve Start next maximum map. Vary c between 7 and 12 (The curve only appears after an interval of 1000 time-steps). What is graphed? What does it tell you about the predictability of the system? Sketch the graph on paper and discuss your ndings with one of the lab assistants. 5. Choose SelectCurve and select the curve Bifurcation plot. Also open the window called Lextreme-c plot. What is graphed here? What happens if c is raised? Re-run orbits with specic c values, if necessary. 6. A proper equilibrium bifurcation curve can only be computed partly. Select the curve Start equilibrium and open the window Bifurcation L-c plot. Compute forward and switch at every bifurcation point. The Hopf bifurcation point is the start of a limit cycle and thus from that point we need to change the curve type to Limit Cycle. After that we encounter PD s (Period Doubling). When switching there, we need to raise the value of the numerical setting called ntst (to be found in the Starter window) by 4 at every PD.

10.8. THE RETURN OF THE PARADOX OF ENRICHMENT

195

10.8

The Return of the Paradox of Enrichment

Recall the paradox of enrichment, where the interactions between Daphnia and phytoplankton lead to destabilization with increasing carrying capacity. The objective of this lab will be to examine the robustness of those conclusions to predation pressure by sh (F ) on Daphnia. Fish imposed mortality saturates according to a type III functional response. We assume that sh stocks are constant at F . P dP = rP 1 f (P ) D dt K dD dt in which f (P ) = and = f (P ) D d(D) aP 1 + a Th P

D2 1 + Ad D 2 Use parameters r = 2, = 0.5, Ad = 44.444, F = 6.667, a = 9.1463, Th = 0.667, = 0.1. d(D) = D + F NB: There is no particular order in the suggestions a-c given under the study questions. 1. Occurrence and stability of equilibria: Start with the qualitative behavior of the model for dierent levels of productivity (i.e. dierent values of K, between 0 and 1.0) by studying: a). b). c). d). e). Isocline congurations in the phase-plane (use the Excel spreadsheet). Time series starting from dierent initial conditions (use Excel and/or CONTENT). Bifurcation of equilibria as a function of K. Limit cycles (or the minima and maxima of these) as a function of K. The global behavior of the model.

Make sure you have a picture (mentally or on paper) of the isocline congurations in every dierent region of K 2. How does the bifurcation graph over K compare to the system without sh predation on Daphnia? (See the computer practical 10.6 The Paradox of Enrichment). Make an explicit comparison of the bifurcation graphs in the two systems and describe the qualitatively dierent regions. 3. Eects of predator-induced mortality: Continue with the qualitative behavior of the model for dierent values of F , by studying: a). Isocline congurations in the phase-plane (use the Excel spreadsheet). b). Time series starting from dierent initial conditions (use Excel and/or CONTENT). c). Two-parameter bifurcation of all (3 types of) special points found in the bifurcation over K. Deduce from the two-parameter plot a bifurcation graph of the equilibria in the system as a function of F, for dierent values of K; verify your interpretation with the help of CONTENT. 4. Biological interpretation: What is the eect of a top-predator on the paradox of enrichment?

196

CHAPTER 10. COMPUTER LABS

10.9

Cannibalism

The following model describes the dynamics of a cannibalistic population, composed of juveniles X and adults Y . The adult individuals eat juveniles in addition to alternative food. After scaling the density and time variables the model can be expressed by the following equations: dX dt dY dt = Y X X Y = X Y + X

in which is the (scaled) fecundity of adult individuals; the (scaled) availability of alternative, noncannibalistic food for the adults and represents the constant mortality rate of the adults. Analyze the model described above, such that you understand as much as possible which type of dynamics the model can exhibit for dierent parameter values. Use both Excel (worksheet LV1 Fasevlak-2-ODEs.xls) and CONTENT to develop this understanding. Here are some useful guidelines: 1. Use as default parameters: = 0.4, = 0.05 and = 0.05. 2. Use Excel to construct isocline graphs. 3. To get an initial understanding compute orbits for dierent starting values of X and Y . Use = 0.05 as parameter value. 4. Construct bifurcation graphs as a function of for dierent values of . Sketch these graphs, while indicating which equilibria are stable and which are not. 5. If you encounter any bifurcation points (limit points) construct a graph that shows the location of this bifurcation point as a function of two parameters, for example, as function of and .

10.10. A PREDATOR-PREY MODEL WITH DENSITY-DEPENDENT PREY DEVELOPMENT197

10.10

A predator-prey model with density-dependent prey development

Consider a prey population that is subdivided into juveniles and adults. We will denote the densities of juvenile and adult prey by J and A, respectively. We will assume that adult prey produce ospring at a per capita rate , while juvenile and adult prey die at a per capita rate of j and a , respectively. Both reproduction and mortality are hence density-independent processes. The regulation of the prey population in the absence of any predator is assumed to operate through density-dependent development. The per capita rate of maturation from juvenile to adult is assumed to follow the function: 1 + dJ 2 1. Formulate the system of two dierential equations describing the dynamics of juvenile and adult prey density J and A, respectively. Take as default values for the parameters: varying between 0.2 and 1.5, = 1, d = 1, j varying between 0 and 0.15 and a = 0.2. 2. Investigate the properties of the model as completely as possible, focusing on the inuence of the parameters and j . We will assume that predators, indicated with the variable P , of this prey population only attack adult prey following a linear functional response with an attack rate equal to a = 1 (default value). Predators convert the prey biomass they consume into ospring with conversion eciency equal to 1.0. Predators experience a per capita death rate of per unit of time. 3. Formulate the system of three dierential equations describing the dynamics of juvenile and adult prey density J and A their predators P , respectively. 4. Investigate the properties of the model taking as a bifurcation parameter for dierent values of and j . 5. In total there are 4 dierent parameter regions to be found, in which the number and type of equilibria diers from the other regions. Try to localize these 4 dierent areas by computing the boundaries between these regions as a function of and j . Extra: Now assume that in addition to a predator on adults, there is also a predator population, indicated with the variable Q, which exclusively forages on juvenile prey. Again assume a linear functional response with an attack rate symbolized by b = 1. Furthermore, assume that the conversion eciency of this predator on juveniles equals m. 6. Formulate the system of 4 dierential equations and investigate its properties.

Bibliography
Allee, W.C., 1931. Animal aggregations: A study in general sociology. University of Chicago Press, Chicago. Arditi, R., Perrin, N., & Saiah, H., 1991. Functional-responses and heterogeneities - an experimental test with cladocerans. Oikos 60: 6975. Begon, M., Harper, J. L., & Townsend, C. R., 1996. Ecology: Individuals, populations and communities. 3rd edition. Blackwell Scientic Publications, Oxford. Caswell, H., 1988. Theory and models in ecology: a dierent perspective. Ecol. Model. 43: 3344. Claessen, D., De Roos, A. M., & Persson, L., 2000. Dwarfs and giants: Cannibalism and competition in size- structured populations. Am. Nat. 155: 219237. Edelstein-Keshet, L., 1988. Mathematical Models in Biology. McGraw-Hill, New York. Gause, G.F., 1934. The Struggle for Existence. Williams & Wilkins, Baltimore (reprinted 1964 by Hafner, New York). Gilpin, M.E., 1973. Do hares eat lynx? Am. Nat. 107: 727730. Holling, C.S., 1959. Some characteristics of simple types of predation and parasitism. Can. Entomol. 91: 385398. Kuznetsov, Yu. A., 1995. Elements of applied bifurcation theory, volume 112 of Applied Mathematical Sciences. SpringerVerlag, Heidelberg. Lampert, W., 1977. Studies on the carbon balance of Daphnia pulex as related to environmental conditions. II. Arch. Hydrobiol. (Suppl.) 48: 310335. Malthus, T.R., 1798. An essay on the principle of population. Penguin Books, Harmondswordth, England, reprinted in 1970 edition. May, R.M., 1974. Stability and complexity in model ecosystems. Princeton University Press, Princeton, 2nd edition. McCauley, E., & Murdoch, W. W., 1987. Cyclic and stable populations: Plankton as paradigm. Am. Nat. 129: 97121. McCauley, E., & Murdoch, W. W., 1990. Predatorprey dynamics in rich and poor environments. Nature 343: 455457. McCauley, E., Nisbet, R. M., Murdoch, W. W., de Roos, A. M., & Gurney, W. S. C., 1999. Largeamplitude cycles of daphnia and its algal prey in enriched environments. Nature 402: 653656. Metz, J. A. J., & Diekmann, O., 1986. The dynamics of physiologically structured populations, volume 68 of Lect. Notes in Biomath. SpringerVerlag, Heidelberg. Murdoch, W. W., Nisbet, R. M., McCauley, E., De Roos, A. M., & Gurney, W. S. C., 1998. Plankton abundance and dynamics across nutrient levels: Tests of hypotheses. Ecology 79: 13391356. Murray, J. D., 1989. Mathematical Biology. SpringerVerlag, Heidelberg. Nisbet, R. M., McCauley, E., De Roos, A. M., Murdoch, W. W., & Gurney, W. S. C., 1991. Population dynamics and element recycling in an aquatic plantherbivore system. Theor. Pop. Biol. 40: 125147. Press, W. H., Flannery, B. P., Teukolsky, S. A., & Vetterling, W. T., 1988. Numerical recipes in C: The art of scientic computing. Cambridge University Press, Cambridge, UK. Rosenzweig, M. L., 1971. Paradox of enrichment: destabilization of exploitation ecosystems in ecological time. Science (Wash., D.C.) 171: 385387. Ruxton, G. D., Gurney, W. S. C., & De Roos, A. M., 1992. Interference and generation cycles. Theor.

199

200

BIBLIOGRAPHY

Popul. Biol. 42: 235253. Smith, H.L., & Waltman, P., 1994. The theory of the chemostat. Cambridge University Press, Cambridge (UK). Tilman, D., 1977. Resource competition between plankton algae: An experimental and theoretical approach. Ecology 58: 338348. Tilman, D., 1980. Resources: A graphical-mechanistic approach to competition and predation. Am. Nat. 116: 362393. Tilman, D., 1981. Tests of resource competition theory using four species of Lake Michigan algae. Ecology 62: 802815. Tilman, D., 1982. Resource competition and community structure. Princeton University Press, Princeton, N.J. Tilman, D., & Kilham, S.S., 1976. Phosphate and silicate growth and uptake kinetics of the diatoms Asterionella formosa and Cyclotella meneghiniana in batch and semicontinuous culture. J. Phycol. 12: 375383. Turchin, P., Oksanen, L., Ekerholm, P., Oksanen, T., & Henttonen, H., 2000. Are lemmings prey or predators? Nature 405: 562565. Van den Bosch, F., De Roos, A. M., & Gabriel, W., 1988. Cannibalism as a life boat mechanism. J. Math. Biol. 26: 619633. Verhulst, P.F., 1838. Notice sur la loi que la population suit dans son accroissement. Corr. Math. et Phys. 10: 113121. Volterra, V., 1926. Fluctuations in the abundance of a species considered mathematically. Nature 118: 558560. Yodzis, P., 1989. Introduction to Theoretical Ecology. Harper & Row, New York.