Mathematical Ecology

Joachim Hermisson, Claus Rueffler & Meike Wittmann∗

March 3, 2019

Literature and Software

• Sarah P. Otto, Troy Day: A Biologist’s Guide to Mathematical Modeling in Ecology
and Evolution, Princeton University Press (∼ 72 Euro)
• Mark Kot: Elements of Mathematical Ecology, Cambridge University Press (∼ 62
Euro)
• Josef Hofbauer and Karl Sigmund: Evolutionary Games and Population Dynamics,
Cambridge University Press (∼ 49 Euro)
• Linda Allen: An Introduction to Stochastic Processes with Applications to Biology,
Prentice Hall (∼ 70 Euro)
• Peter Yodzis: Introduction to Theoretical Ecology (1989), Harper & Row.
This book is out of print. A pdf can be downloaded from
www.rug.nl/research/institute-evolutionary-life-sciences/tres/ downloads/bookyodzis.pdf
• Gerald Teschl: Ordinary Differential Equations and Dynamical Systems, American
Mathematical Society, pdf online at www.mat.univie.ac.at/ gerald/ftp/book-ode/
• Populus simulation and visualization software: http://cbs.umn.edu/populus/overview

Ecology
Oikos = house, dwelling place. Logos = word, study of. Ecology refers to the scientific
study of living organisms in their natural environment. It is a diverse scientific discipline
and covers various levels of biological organization.
• On the individual level, physiological ecology discusses the influence of food, light,
humidity, pesticide concentrations, etc, on the life histories of individuals.
∗
First version 2012 by JH and CR, revised and extended JH and MW 2015, edits JH 2018.

1
 • Population ecology studies the interactions of populations with their environment,
with consequences on population structure and demography. On the same level,
behavioral ecology discusses the consequences of different behavioral strategies.

• Finally, community ecology and ecosystems ecology treat the fate of complex ecosys-
tems with anything from two to tens of thousands of interacting species and groups
of species.

Ecology is closely related to evolution and the interactions of population dynamics and
evolution are the subject of evolutionary ecology. Many branches of ecological research
use mathematical models. For example, behavioral ecology makes use of game theoretical
methods to explore the impact of behavioral strategies. Evolutionary ecology draws heavily
on the mathematical models of evolutionary genetics. The focus of this lecture must be
much more narrow. It will mainly be on population ecology, where we study the dynamics
of population sizes, equilibria, growth and extinction, under various ecological boundary
conditions. We will make a few side-steps into evolutionary ecology, but we won’t treat
aspects of behavior and we won’t cover inheritance and the dynamics of genotypes. These
topics are devoted to the specialized lectures on game theory and on population genetics.

Ecological Modeling
Any biological model is a map of some part of Nature to a mathematical formalism. Models
are always abstractions, i.e. simplifying representations of reality. Modeling thus starts
with a series of model assumptions: some aspects of Nature are integrated into the model,
because we assume that they are essential for the problem at hand. Many other aspects
are ignored (or abstracted from), either because they are much less important or because
we want to take a reductionist perspective. In the latter case, we hope that we can
understand a complex system by studying of several (sets of) factors one by one. As an
example, if we want to model future population size in Austria, the current size and age
structure are certainly essential. Other factors like progress in medical treatment might
also have some impact on death rates, but can be ignored in a simple model. Still other
factors, such as immigration, are likely important, but a treatment without immigration
may already provide us with some valuable information and we may want to study the
impact of immigration in a separate step.
With an increasing number of factors included, a model gets more precise and spe-
cific. This is needed, in particular, for reliable quantitative predictions (weather forecast,
demographic models). However, added complexity always means reduced manageability
and often also reduced generality. From a model that is as complex as the system that it
represents we cannot obtain any new insights. Complex quantitative models that are used
for predictions can usually only be treated by computer simulations. In contrast, many
questions we might ask are of qualitative nature (e.g., whether population size approaches
an equilibrium or whether there will be cycles). In these cases, one often aims for a minimal
set of factors to explain a phenomenon.

2
 The art of modeling thus consists of selecting the essential factors to include in a model.
On the one hand, this requires experience and some knowledge of the biological system
of interest. On the other hand, this also requires an understanding of the mathematical
mechanism, in order to see which factors can have crucial consequences, even if they may
look like small effects initially. As such, ecological modeling relies on a broad mathematical
tool-box, including elements from the theory of stochastic processes, dynamical systems,
differential equations, and statistics.

1 Dynamics of single, unstructured populations

The dynamical process of population growth and decline is a function of factors that are
intrinsic to a population (e.g., its potential to reproduce, its life-cycle, or its density) and
the environmental conditions. The environment comprises all resources that are essential
for a population to thrive, like food and space, and factors that may reduce its size, such
as predation and disease. In nature, many of these factors are indeed reproducing popu-
lations themselves, which can act as predators, competitors, or as food resource. As such,
these populations should follow their own population dynamics. Since the dynamical pro-
cesses of (e.g.) predators and prey interact, we quickly obtain a complex multi-dimensional
problem. We deal with these complexities further on. As our first step, we make the sim-
plifying assumption that we can ignore all interactions with other dynamical aspects of
the environment and just model the dynamics of a single population. This can sometimes
be justified if the dynamics of all interacting populations happens on different time scales:
either much faster, such that we can always assume that the interacting population is at
a dynamical equilibrium, or much slower, such that the size of an interacting population
does not change much over time-spans of interest. We also assume that the population is
unstructured. This means, all individuals of the population are treated as equal. In par-
ticular, there are no age classes, no phenotypic differences (of relevance to the dynamics),
and we can ignore the distribution of the population across physical space.

1.1 Birth and death processes

We describe the development of a population through time as a dynamical process. For
a single, unstructured population, we have a single dynamic variable N (t), measuring the
population size or population density (individuals per square meter) at time t. The variable
N (t) may be affected by various demographic events, such as:
• birth
• death
• immigration and emigration
Demographic events in nature are stochastic. In the most explicit “individual based” de-
mographic models, the population dynamics is therefore described as a stochastic process.

3
 Define PN (t) as the probability to observe N individuals at time t. We assume that
each individual can give birth at a constant rate b and may die at rate d. Birth and death
occurs for all individuals independently of all other individuals and independently of age.
Formally, the process then follows a continuous-time Markov chain with states N ∈ N and
time-homogeneous transition probabilities. We have:

PN (t + ∆t) = (N − 1)b∆tPN −1 (t) + (N + 1)d∆tPN +1 (t) + (1 − N b∆t − N d∆t)PN (t) (1)

and thus in the limit ∆t → 0 (Kolmogorov forward equation or Master equation):

∂PN (t)
ṖN (t) = = d(N + 1)PN +1 (t) + b(N − 1)PN −1 (t) − N (b + d)PN (t) (2)
∂t

with some initial condition PN (0) = 1 for N = N0 and PN (0) = 0 else. (In particular, we
have PN (t) = 0 for all N < 0 and all t.) The Master equations are a system of infinitely
many ordinary differential equations. We consider the expected population size
∞
X
N̄ (t) = N PN (t). (3)
N =0

From the Master equation follows

∞
∂ N̄ X ∂PN (t)
= N
∂t N =0
∂t
∞
X
dN (N + 1)PN +1 (t) + bN (N − 1)PN −1 (t) − N 2 (b + d)PN (t)


=
N =0
X∞
d(N − 1)N PN (t) + b(N + 1)N PN (t) − N 2 (b + d)PN (t)


=
N =0
X∞


= N (b − d)PN (t) = (b − d)N̄ (t) (4)
N =0

Defining the net growth rate r = b − d, we obtain the solution

N̄ (t) = N0 · exp[rt]. (5)

We thus see that the expected value of the stochastic process follows simple exponential
growth. The long-term behavior follows a simple dichotomy: the expected population size
declines to zero as the population dies out for d > b, while it grows without bounds for
b > d. However, the behavior of the stochastic process is richer than predicted just by the
expected value. Similar to the derivation above, we can derive the variance. We start with

4
the dynamics of the second moment:
∞
∂N 2 X ∂PN (t)
= N2
∂t N =0
∂t
∞
X
dN 2 (N + 1)PN +1 (t) + bN 2 (N − 1)PN −1 (t) − N 3 (b + d)PN (t)


=
N =0
X∞
d(N − 1)2 N PN (t) + b(N + 1)2 N PN (t) − N 3 (b + d)PN (t)


=
N =0
X∞
(2N 2 (b − d) + N (b + d))PN (t) = 2(b − d)N 2 (t) + (b + d)N̄ (t).


= (6)
N =0

Defining N 2 (t) := Q(t)N0 exp(rt) and with the solution for N̄ (t) from above, we obtain:
Q̇(t) = (b − d)Q(t) + (b + d) (7)
which is solved by
b+d
Q(t) = C exp(rt) − .
b−d
With Var[N (t)] = N 2 (t) − (N̄ (t))2 and the boundary condition Var[N (0)] = 0 we now
obtain C = N0 + (b + d)/(b − d) and thus
b+d

Var[N (t)] = exp(rt) − 1 exp(rt)N0 . (8)
b−d
Just like the expected value, the variance increases exponentially with time if b > d. What
kind of biological consequences does this stochastic scattering around the expected value
have? Most significantly, since extinction (N = 0) is a stationary point of the process, we
may ask for the probability that this point is reached from a starting population N = N0 .
Define uN as the probability that a population of current size N will go extinct at some
time in the future. How can we derive uN ? We know the following:
1. Birth and death is independent for all individuals in our model. The population will
go extinct if and only if none of the individuals in the starting population will leave
offspring in the distant future. Let u = u1 be the probability that a single individual
does not leave any offspring. Because of independence, we have uN = uN .
2. Consider a population currently in state N . We may ask for the state of our popula-
tion right after the next demographic event, i.e., after a single birth or death. From
our current state N , we can only reach one of the neighboring states. For given birth
and death rates b and d, we will be in state N + 1 with probability b/(b + d) and in
state N − 1 with probability d/(b + d). We thus obtain the following recursion for
our extinction probabilities,
b d
uN = uN +1 + uN −1 for N ≥ 1, (9)
b+d b+d

5
 with boundary condition u0 = 1. Formally, this corresponds to the transition matrix
of a one-dimensional random walk with absorbing state at 0.

We thus obtain the condition

(b + d)u = bu2 + d (10)
The two solutions of this quadratic equation are

d
u(1) = ; u(2) = 1.
b
For d > b, u = 1 is the only valid solution and we see that extinction is certain. For
b > d, the population can escape extinction. However, there is still a probability given
by uN0 = uN0 = (d/b)N0 that the population dies out – even if the expected value grows
without bounds. Note that this probability becomes uN0 = 1 for r = 0, although the
expected population size stays constant in this case. However, for b > d and sufficiently
large N0 the extinction probability becomes very small. [A formal proof that u = d/b is
the correct solution for d < b requires some further arguments, see lecture on stochastic
processes or the book by Kot.]

Stochastic and deterministic models We can interpret a deterministic population

model as an approximation to a more detailed stochastic process. For a (linear) birth-
death model we have seen that the dynamics of the expected population size reproduces
deterministic exponential growth. However, there are important aspects of the stochas-
tic dynamics that are not covered by a deterministic model. These differences are most
important in small populations, where random fluctuations can easily drive a population
to extinction. Once the population size is sufficiently large, the deterministic system is a
valid approximation of the stochastic model. This can be justified from the coefficient of
variation, which derives as
p r
Var[N (t)] 1 b+d

CV[N (t)] = = 1 − exp(−rt) . (11)
N̄ (t) N0 b − d

We see that for r > 0 CV[N (t)] is limited for all t and gets small for large initial population
size N0 . It is straightforward to define more complex stochastic models for population
dynamics – even with multiple interacting populations. [Some theory that exists along
these lines is summarized in the book by Allen]. However, most of these models cannot be
solved and our potential to obtain explicit analytical results is very limited. For this reason,
the bulk of population dynamical research uses a deterministic approach. For a general
stochastic model of a single, unstructured population with birth, death, and immigration,
we obtain

N (t) → N (t) + 1 : b(t, N (t)) · N (t) + m(t, N (t)) (12)

N (t) → N (t) − 1 : d(t, N (t)) · N (t) (13)

6
where b(t, N (t)), d(t, N (t)) and m(t, N (t)) are birth-, death-, and immigration rates, re-
spectively, which can depend on the population density and on time. (Note that emigration
can be subsumed in the death rates). We then have the expected change
 
E[∆N |N (t)] = r(t, N (t)) · N (t) + m(t, N (t)) ∆t . (14)

This can be transformed into a deterministic process using the following correspondence:

Ṅ (t) = r(t, N (t)) · N (t) + m(t, N (t)) . (15)

We thus obtain an ODE, which in general can be non-linear and time-dependent. This
ODE approximates the dynamics of the stochastic system for large population sizes. (We
note that it is generally not true that it represents the - exact - dynamics of the expected
population size, like in the linear birth-death model). A general strategy in ecological
modeling is to use a deterministic model for derivations and to use stochastic computer
simulations as back-up, in order to test the robustness of the results.

1.2 Deterministic models in continuous time

The general deterministic model for the dynamics of a single, unstructured population
reads:
Ṅ (t) = f (N, t) = f (N (t), t) (16)
N (t) ∈ R is usually interpreted as population density rather than the total size. There are
(only) a few models with explicit solutions:
• The most basic model is exponential growth with

Ṅ (t) = rN (t)

and explicit solution N (t) = N0 exp(rt). The net growth rate r is also called the
Malthusian parameter (Thomas Malthus, 1798: Essay on the Principle of Popula-
tion). Exponential growth with r > 0 leads to population explosion and is unsus-
tainable in Nature. With r < 0, there is a trivial stable equilibrium at N = 0.
• The simplest model with a stable equilibrium population derives from a dynamics
with immigration and death (negative growth),

Ṅ (t) = c − dN (t)

with c, d > 0 and explicit solution

c  c
N (t) = + N0 − exp[−dt] . (17)
d d
The model describes the dynamics in a population sink. It also describes pure mi-
gration if immigration is constant and emigration occurs with a constant per-capita
rate.

7
 • The simplest model with population regulation of a population that can sustain itself
is the logistic growth model with
 N (t) 
Ṅ (t) = rN (t) 1 − .
K
It is characterized by linear density dependence with per-capita growth rate r(1 −
N/K). The parameter r is also called the intrinsic growth rate. K is the carrying ca-
pacity. The explicit solution for the logistic growth model follows (e.g. by separation
of variables) as
K
N (t) = . (18)
1 − (1 − K/N0 ) exp[−rt]
The logistic growth model can even be solved with Rtime-dependent per-capita growth
t
rate r(t), simply by replacing the exponent rt by 0 r(t)dt.

Very few other models of relevance in the ecological literature have explicit solutions.
Luckily, we are usually not so much interested in the explicit dynamics, but rather in
key qualitative properties. The most fundamental property of a dynamical model is its
equilibrium structure, which governs the long-term behavior and defines the landmarks of
the dynamics.

1.3 Equilibria and stability of single-species models in constant

environments
Consider a general (unstructured) population model in one dimension with autonomous
dynamics, i.e.,
Ṅ (t) = f (N ) (19)
The long-term behavior of a general population model in one dimension with an au-
tonomous dynamics can be read off from the phase-line diagram, where Ṅ = f (N ) is
plotted as a function of N . See Figure 1 for the case of logistic growth.

Definition In general, we define N = N ∗ is an equilibrium point (fixed point) of the

dynamics if
f (N ∗ ) = 0

• The equilibrium is Lyapunov stable if for any  > 0 there is a δ > 0, such that

|N (t0 ) − N ∗ | < δ ⇒ |N (t) − N ∗ | <  ∀t > t0 .

This means: if we start close, we stay close.

• A (Lyapunov) stable equilibrium is called asymptotically stable if an δ > 0 exists,

such that
|N (t0 ) − N ∗ | < δ ⇒ lim |N (t) − N ∗ | = 0
t→∞

8
 NHtL
g(x) 20 000

λK/4
15 000

10 000

x 5000
K/2 K

t
5 10 15 20 25 30

Figure 1: Phase-line diagram and dynamics for logistic growth (with K = 10000 and
r = 0.4). The growth dynamics has an inflection point at K/2: maximal absolute increase
in population density.

This means that sufficiently small deviations will only produce (small) excursions
that eventually return to the equilibrium. In the biological literature, this is often
also called “locally stable”.

Elementary facts

• If f (N ) is continuously differentiable at N = N ∗ then an equilibrium point N ∗ is

asymptotically stable if f 0 (N ∗ ) < 0. It is unstable if f 0 (N ∗ ) > 0. For f 0 (N ∗ ) = 0,
stability depends on higher derivatives (e.g., an internal equilibrium is unstable in
this case if f 00 (N ∗ ) 6= 0, etc). In general: An equilibrium N ∗ is asymptotically stable
if and only if f (N ∗ + ) < 0 and f (N ∗ − ) > 0 for some δ > 0 and all 0 <  < δ. –
Proof: obvious from the phase-line diagram, or see Kot.

• Let f (N ) be continuously differentiable. We call an unstable equilibrium a breakpoint

if and only if f (N ∗ +) > 0 and f (N ∗ −) < 0 for some δ > 0 and all 0 <  < δ. If f (N )
has a finite number of equilibria, asymptotically stable equilibria and breakpoints
are always interlaced (i.e., there is exactly one stable fixed point between any two
breakpoints and vice-versa).

• Assume that all fixed points are either asymptotically stable equilibria or break-
points (this is the generic case). Then the domain (or basin) of attraction of any
stable equilibrium extends to the neighboring breakpoints, or to the boundaries of
the parameter space.

• Asymptotically stable equilibria and breakpoints are structurally stable in the sense
that small perturbations to f (N ) may lead to small changes in the positions of the
equilibria, but they persist otherwise. All other equilibria are structurally unstable.

9
1.4 Example: Malaria infection model
In 1897, Sir Ronald Ross (1857 – 1932, Nobel prize 1902) discovered that malaria is trans-
mitted by Anopheles mosquitoes. He suggested to fight the mosquitoes to get rid of the
disease. However, it is clear that a total extinction of all mosquitoes is not realistic. Peo-
ple argued that Plasmodium, the malaria parasite, would survive in some mosquitoes and
return together with the mosquitoes after a (costly) program to fight the mosquitoes has
ended. Sir Ronald designed a mathematical model to convince his contemporaries that
this is not necessarily true.
In a simplified version (e.g., ignoring immunity), the model works like this:

• The model assumes a human population with two parts: infected and uninfected
people. Let p be the proportion of infected people.

• Uninfected people can get malaria from infected people via Anopheles mosquitoes.
Let the density of (human active) mosquitoes be α. Then the infection rate (per
uninfected) is αp.

• Infected people can also recover at rate µ

We can summarize the model in the following figure:

α p(1−p)

p 1−p
infected not infected
µp

• Construct an ODE model for the dynamical variable p(t) and discuss the equilibria.

• How does the phase diagram look like?

• How does the dynamics depend on the mosquito density α? Draw a figure for the
equilibria p∗ as a function of α.

• What are the prospects for getting rid of the disease without getting rid of Anopheles?

We have  p 
ṗ = αp(1 − p) − µp = (α − µ)p 1 − . (20)
(α − µ)/α
For α > µ, this is the dynamics of logistic growth with r = α − µ and K = (α − µ)/α.
We then get a stable equilibrium of infected people at the carrying capacity K. This is
the so-called endemic equilibrium. In contrast, for α < µ, we have a stable equilibrium at
p = 0. The disease is extinct although the mosquitoes are still there at some density. Sir
Ronald Ross always considered his mathematical models his most important contribution

10
 g(p) g(p)

p p

Figure 2: Phase-line diagram for the infection model (20). Left: α > µ, right: α < µ.

to science! Only very few of his contemporaries understood and appreciated these models.
Only two decades after his death, people started to construct mathematical models for
infection and disease in his tradition. For malaria, the qualitative finding that the disease
cannot persist with a low mosquito density is confirmed (e.g., parts of USA and Italy). Of
course, modern models used in the fight against malaria in Africa and Asia are many levels
more complex.

1.5 Bifurcations
For qualitative analysis, it is often of central interest how the equilibrium structure of
a model changes if we change some of the model parameters. As a consequence, new
equilibria may be created or old ones destroyed, or equilibria may collide and exchange
their stability. The resulting qualitative changes in the behavior of the dynamical systems
are called bifurcations. We can visualize this “dynamics of the equilibria” in a so-called
bifurcation diagram. In the malaria example above, we have α as a model parameter and
the following equilibria: p∗1 (α) = 0 and p∗2 (α) = K = 1 − µ/α. This corresponds to a so-
called transcritical bifurcation. We consider a “sufficiently smooth” (at least continuously
differentiable) function f (N ). The basic types of bifurcations in one dimension are the
following:

1. Transcritical bifurcation: two equilibria cross and exchange stability properties. The
normal form of a transcritical bifurcation at x = 0 (think of x = N − N ∗ ) and
perturbation parameter α reads

ẋ = αx − x2 . (21)

2. Saddle-node bifurcation: two equilibria (with opposite stability) collide and vanish
(or vice-versa), e.g.,
ẋ = α − x2 . (22)

11
 3. Pitchfork bifurcation: three equilibria collide. There are two possibilities. For the
supercritical case, the three equilibria are stable – unstable – stable and turn into a
stable equilibrium, e.g.,
ẋ = αx − x3 . (23)
For the subcritical case, the three equilibria are unstable – stable – unstable and turn
into a single unstable equilibrium,

ẋ = αx + x3 . (24)

4. Higher-order bifurcations with four or more equilibria colliding are in principle pos-
sible, but non-generic.

5. In addition to bifurcations, equilibria can also enter or leave the biological state space.
For a population density N the critical boundary is usually at N = 0. Equilibria
can also emerge (or vanish) from infinity. However, most biologically relevant models
have limN →∞ f (N ) << 0, in which case this is not possible.

1.6 Exercise 1: Harvest models

Consider a population of fish with dynamics according to logistic growth. We want to use
this population as a resource and we are looking for a harvesting strategy that guarantees
large and stable yield. We consider the following two strategies:

1. With constant-rate harvesting (e.g., because of fishing quota), we have

 N
Ṅ = rN 1 − −H. (25)
K

2. With relative-rate harvesting, we will catch fish proportional to the stock size,
 N
Ṅ = rN 1 − − EN, (26)
K
where E measures the fishing effort (this could be a quota on fishing boats).

We define the maximum sustainable yield (MSY) as the largest yield that can be taken
from the species’ stock over an infinite period.

• Make a bifurcation analysis for both harvesting strategies with H and E as param-
eters (diagrams and formulas for bifurcation points). What kinds of bifurcations
occur?

• What is the MSY and for which parameter values of H and E do we get this yield?
Which implications for harvesting strategies do you see?

12
From the solution we see that constant rate harvesting at MSY is dangerous, since any
higher rate can lead to a collapsing fish population. With relative rate harvesting, the
risk of a sudden crash is avoided. Here, we get the maximal cumulative yield for an
intermediate cumulative effort E. It is interesting to consider a situation where multiple
fisheries (or countries) compete and try to maximize their share of the harvest. In many
cases, we will find that the individual yield increases monotonically with the individual effort
– although the cumulative yield decreases. This is an example of the so-called “tragedy of
the commons” (see lecture on game theory).

Allee effect In the models we have considered so far, the per-capita growth rate g(N ) =
Ṅ /N was either independent of N (as in exponential growth) or decreased with increasing
N (as in logistic growth). In some species, however, the per-capita growth rate is reduced at
small population densities, for example because individuals have difficulty finding mates,
because group defense (like swarming) or hunting do not work properly, or because of
increased predation pressure. Formally, a population experiences a demographic Allee
effect if limN →0 dg/dN > 0. If the reduction in per-capita growth rate at small densities
is so strong that populations below a certain critical density have a negative per-capita
growth rate, we say that there is a strong demographic Allee effect. Otherwise the Allee
effect is called weak and small populations can still grow. A typical model with strong
Allee effect reads N  N
Ṅ = rN −1 1− , (27)
K0 K
where K0 < K defines the critical density for which growth becomes positive.
• We can consider relative-rate harvesting on such a population, i.e.,
N  N
Ṅ = rN −1 1− − EN. (28)
K0 K
We get equilibria at N = 0 and at the solutions of E = r((N/K0 ) − 1)(1 − (N/K)).
A bifurcation analysis reveals that an allee effect can lead to qualitative changes in
the dynamics (see exercises).

1.7 Nondimensionalization
Understanding the qualitative behavior of a model with several parameters can be difficult.
If we vary one parameter at a time while fixing the other parameters at certain values,
we cannot be sure that we get the full picture of the model’s behavior. Fortunately, often
the only quantity that matters is the relative magnitude of parameters. This is exploited
by a model simplification technique called nondimensionalization. The idea is to rescale
variables and time such that the rescaled model has fewer parameters but maintains all
other model features. As an example, we will nondimensionalize the Allee effect model (27)
above. First, we rescale population size by expressing it relative to the carrying capacity
K (scaling by K0 would work as well). Now our new variable is x = N/K. Second, we

13
define a new time scale t̃ = r · t, i.e. rate r on the original time scale corresponds to rate 1
on the new time scale.
With this, we get the new differential equation
 
dx dx dt 1 dN Kx
= · = = x(1 − x) −1 . (29)
dt̃ dt dt̃ Kr dt K0
We set k0 = K0 /K and drop the tilde (keeping in mind the difference in time scale):
 
x
ẋ = x(1 − x) −1 . (30)
k0
Our new model has just one parameter, which makes the computation of equilibria, bi-
furcation analyses etc. more transparent. After we have completed our analyses, we can
translate the results back to the original scale. This is especially important if we want to
relate them to biological data.

1.8 Bifurcations and structural stability

We have seen that for relative-rate harvesting the inclusion of the Allee effect transforms
a transcritical bifurcation into a saddle-node bifurcation. This is indeed typical: Consider
a small constant perturbation to a transcritical bifurcation, i.e.,
ẋ = αx − x2 +  . (31)
√
There are equilibria for 1/2(α ± α2 + 4). We see that there is no bifurcation for  > 0,
because the equilibria never cross. And there √ are two saddle-node bifurcations for  < 0,
where there is no equilibrium at all for |α| < 2 −. In an extended parameter space with
two perturbation parameters α and , the transcritical bifurcation will only be seen on a
low-dimensional manifold (for  = 0). [In dynamical systems theory, this kind of analysis is
called an “unfolding” of a bifurcation.] We say, the transcritical bifurcation is structurally
unstable with respect to constant perturbations. There is a similar effect for the pitchfork
bifurcation, but not for the saddle-node bifurcation. In this sense, saddle-node bifurcations
represent the most “generic” bifurcation type in one dimension.

1.9 Functional response

Define the functional response F (N ): the per capita resource consumption as a function
of resource density N per unit time. We set
Ts · eN
F (N ) = (32)
Ts + Th
where Ts is the time spent searching for food, e the search efficiency, and Th the handling
time, i.e., the time needed to process all the attacked prey items. With Th = eN Ts · th ,
where th is the time needed to process a single prey item, we get
eN
F (N ) = . (33)
1 + eN th

14
After Holling, we can distinguish three types of feeders

1. th = 0: Type 1 functional response, e.g. for filter feeders.

2. th > 0: Type 2 functional response, with asymptote at 1/th .

3. th > 0 and e = e(N ) = e0 N : Type 3 functional response. Here, the search efficiency
increases with the prey density, for example because a search image is created as the
predator get “practice”. This type of functional response is often assumed for higher
organisms like vertebrates.

Based on these functional responses, we can now consider the resource dynamics as a
function of the consumer density. For the type 1 functional response, this coincides with
the relative-rate harvesting model. For type 2 or type 3, we obtain a different behavior.

Resource dynamics for type 2 functional response Consider a model for a resource
population with logistic growth, a constant consumer density C, and type 2 functional
response:
 N CeN
Ṅ = rN 1 − − . (34)
K 1 + eth N
• Explore the model by plotting Eq. (34) for different parameter combinations.

• Set r = 0.1, K = 1000, th = 0.25, e = 0.05. Make a graphical bifurcation analysis for
C. If you want, you can use mathematica, R, maple or some other software package.

• For which consumer densities C does the resource experience an Allee effect? Is the
Allee effect weak or strong?

Resource dynamics for type 3 functional response For a type 3 response, the model
looks as follows:  N CgN 2
Ṅ = rN 1 − − 2 (35)
K N0 + N 2
where C is the number of consumers, g the asymptote of the response curve, and N0 the
half saturation point. We may, for example, think of C as the number of sheep on a meadow
and N as the grass biomass. We can analyze this case qualitatively purely graphically.

Take-home message from bifurcation theory Small changes can have a large effect.
This is typical for ecosystems, where these abrupt shifts are frequently seen. The fact that
continuous change may lead to discontinuous consequences is counterintuitive for most
people. This can have dangerous consequences. Providing insight into the bifurcation
structure of natural systems is one of the most important tasks of theoretical ecology.
Examples:

• Fisheries: extinction risk because of overfishing.

15
 • Grasslands turning into deserts because of overgrazing. Very difficult to revert after
erosion of top soil.

Of course, it will often be difficult to obtain quantitative estimates of the positions of bi-
furcation points. Note also that natural systems exhibit environmental and demographic
stochasticity. For this reason, it will be wise to stay at a safety distance from any catas-
trophic bifurcation point.

1.10 Stability revisited

We have seen that for asymptotically stable equilibria the population will return to the
equilibrium after a single and “sufficiently small” disturbance. However, disturbances in
biology are frequent and not necessarily small. How can we deal with this problem? We
will treat the case that perturbations are still small, but can be frequent. The relevant
question for a second (and each further) perturbation is whether the system has already
returned to the equilibrium after the previous perturbation. With a small perturbation
N (0) = N ∗ + δ, we can linearize the dynamics around N ∗ ,

Ṅ (t) ≈ Λ(N ∗ − N (t)) (36)

with
∂f (N )
Λ=− . (37)
∂N N =N ∗
For an asymptotically stable equilibrium, we have Λ > 0. Now, Eq. (36) is precisely the
dynamics of an immigration-death model (see above) and can be solved exactly. We get

N (t) = N ∗ + δ exp[−Λt] . (38)

We thus obtain an exponential return to the equilibrium with rate given by Λ. We can
define a characteristic return time to the equilibrium as
1
TR = . (39)
Λ
For t  TR we will be close to N ∗ , for t  TR , we are still about a distance δ away. We
can now compare TR with the average time TD between disturbances. For TR < TD , the
system is stable and will remain in the neighborhood of N ∗ . In contrast, for TD < TR , we
get larger excursions away from N ∗ . In this case, the exact behavior cannot be predicted
anymore from a local analysis alone.

1.11 Systems with time delay

One-dimensional autonomous ODE models have a simple equilibrium structure: only stable
or unstable equilibrium points can occur. Stable equilibria have simple intervals as domains
of attraction. In particular, there are no oscillations or chaotic behavior. All this quickly

16
changes if the dynamics of a focal population can depend on other factors than just its
current size. As already mentioned, Ṅ could depend explicitly on time (non-autonomous
systems), or on the dynamics of a different population (multidimensional ODEs). As
yet another possibility, we consider the following dynamics of a population with damped
oscillations:
N (t) = N ∗ + (N0 − N ∗ ) cos[γt] · exp[−dt] . (40)
We then have

Ṅ (t) = (N0 − N ∗ ) exp[−dt] (−γ sin(γt) − d cos(γt)) (41)

= (N0 − N ∗ ) exp[−dt] (γ cos[γ(t − 3π/2γ)] − d cos(γt)) . (42)

With τ = 3π/2γ and using Eq. (40) to solve for and substitute cos[γ(t − 3π/2γ)] and
cos(γt), we obtain

Ṅ (t) = γ exp[−dτ ]N (t − τ ) − dN (t) + (d − γ exp[−dτ ])N ∗ . (43)

Although Ṅ (t) only depends on the size of the focal population, it is not only its current
size, but also the size at some earlier time t − τ . Defining b := γ exp[−dτ ], we recognize
this as a time delay in the effect of birth events. Biologically, this can easily be understood
if N measures adult individuals, and if it takes some time for juveniles to grow up to
reproductive age. There is a lot of theory on ODE’s with time delay (also called difference-
differential equations), but analytical results are sparse. Formally, these equations are
equivalent to infinite-order ODE’s, which depend on initial conditions on a whole time
interval. The math gets quite complex. Nevertheless, delay effects can obviously be relevant
in biological systems. If this is the case, they need to be included into a model. As it turns
out, however, this is most easily done in a discrete time model using difference equations.

Further example: Lagged logistic growth We can also introduce a time-lag into the
logistic growth model, e.g., if the negative feedback of the population density on population
growth is not immediate, but only sets in once individuals have grown up. For a lag time
τ we get
 N (t − τ ) 
Ṅ (t) = rN (t) 1 − . (44)
K
There is no analytical solution of this ODE, but we can study its behavior numerically.
• Use the ”Populus” software to plot numerical solutions for the lagged logistic model
(single species/density-dependent growth). Which types of behaviors do you get
when you vary the model parameters?

2 One-dimensional models in discrete time

So far, we have assumed that demographic events happen continuously over time and
can thus be written as rates. For many biological species with overlapping generations

17
and continuous reproduction this is a valid approximation. However, many other popula-
tions, like annual plants, have discrete generations and many species have clearly defined
breeding seasons. These are better modeled in a discrete time framework. More funda-
mentally, we have already discussed that the effect of biological events (like birth) on the
population dynamics is often not immediate as it is assumed in an ODE system, but only
becomes apparent after some time delay. While in continuous time this leads to complex
difference-differential equations, effects of time delay (by one time unit or generation) are
automatically included in discrete time models. In this section, we will consider a popula-
tion with dynamics given by the following first-order, autonomous (= constant coefficient)
difference equation,
Nt+1 = F (Nt ) , t ∈ N . (45)
An initial value N0 is specified and the reproduction function F can take any shape. The
simplest example is the linear difference equation

Nt+1 = F Nt (46)

of geometric growth (for F > 1) or decline (for 0 < F < 1). Geometric growth has the
explicit solution
Nt = F t N0 . (47)
Very few other discrete time systems have explicit solutions. Next, we need to include
density-dependent population regulation. Rather than “dropping down” a model, we want
to construct a model from biological principles.

2.1 Non-linear discrete population growth

Consider a fish population of adult stock size Nt . From one breading season to the next (t
to t+1), the survival probability for any adult is given by s. (For s = 0, we have the special
case of no generation overlap). During breeding season, the adults produce a number of
Lt = bNt larvae. We assume that larvae grow up to reproductive age during one season.
Surviving larvae are recruited to the adult population. We thus need a model for larval
survival.

1. Assume that the larvae L develop according to (continuous time) logistic decline, i.e.,

∂Lt (τ )
= −m1 Lt (τ ) − m2 L2t (τ ) , (48)
∂τ
where m1 measures mortality due to extrinsic (e.g. abiotic) factors and m2 the density
dependent component (e.g., due to competition for resources). With Lt (0) = bNt and
a time ∆ between seasons we get

bNt m1
Lt (∆) = . (49)
(bNt m2 + m1 ) exp[m1 ∆] − bNt m2

18
 We thus get
c1 Nt
Nt+1 = sNt + Lt (∆) = sNt + (50)
1 + c2 Nt
with positive constants
m2
c1 = b exp[−m1 ∆] , c2 = b (1 − exp[−m1 ∆])
m1
This is the Beverton-Holt model for stock recruitment. For s = 0, the model is also
called the Verhulst model.

2. An alternative approach assumes that larvae (or juveniles) do not compete primarily
among themselves, but with their adult conspecifics. This will be true, in particular,
if adult fish eat eggs and larvae of their own species (which is indeed true for many
species, where predation is purely size-dependent). We then get

∂Lt (τ )
= −m1 Lt (τ ) − m2 Nt Lt (τ ) , (51)
∂τ
which simply results in

Lt (∆) = bNt exp[−(m1 + m2 Nt )∆] . (52)

We then get a so-called Ricker model for stock recruitment

Nt+1 = sNt + Nt exp[r(1 − Nt /K)] (53)

with constants
log[b] − m1 ∆
r = log[b] − m1 ∆ , K= .
m2 ∆
3. Finally, we can assume that density-dependent competition is entirely among adults
(e.g. for breeding places). Then we get a discretized version of logistic growth
 Nt 
Nt+1 = rNt 1 − , (54)
K

where r = s + b accounts for the surviving adult and juvenile individuals from the previous
generation.
Recruitment in the Beverton-Holt model is called normal compensation. It increases
monotonically with stock size and reaches an asymptote for Nt → ∞ at Lmax (∆) = c1 /c2 .
In contrast, recruitment for the Ricker model is over-compensating. Here, Lmax (∆) runs
through an intermediate maximum and approaches zero for Nt → ∞. As a consequence,
the reproduction function of the Ricker model is usually non-monotonic. The same holds
true for the discrete logistic growth model. All three models can be derived in many
different ways and for various underlying mechanisms. They are archetypical models of
theoretical ecology.

19
2.2 Discrete versions of continuous models
One special application of discrete models consists simply of an observation in discrete
time intervals of a population with continuous birth and death. Thus, we simply set
Nt = N (t · ∆). An explicit reproduction function can be derived if and only if the ODE
can be solved. For logistic growth, in particular, we obtain
K λNt
Nt+1 = = (55)
1 − (1 − K/Nt ) exp[−r] 1 + (λ − 1)Nt /K
with λ = exp[r]. We recognize this as a special case of the Beverton-Holt model with
s = 0, c1 = λ, and c2 = (λ − 1)/K. Note that we do not get the discrete logistic growth as
the corresponding model. This correspondence also implies that the Beverton-Holt model
(with s = 0) is explicitly solvable. In more general, we can define a corresponding discrete
model for every autonomous ODE model. The converse is not true, however. In this sense,
there are “more” discrete models than continuous models. As we will see, they also show
a larger variety of phenomena.

2.3 Equilibria and Stability of Discrete Dynamics

Consider a general discrete dynamical system (1-dim., etc) with continuously differentiable
reproduction function,
∂F (N )
Nt+1 = F (Nt ) ; F 0 (N ) := continuous (56)
∂N
We can then characterize equilibria as follows
1. Equilibria (or fixed points) of the dynamics are all intersections of the reproduction
function with Id(N ) (the identity line with slope 1),
F (N ∗ ) = N ∗ (57)

2. Let
∂F (N )
Λ := .
∂N N ∗
Then N ∗ is locally (asymptotically) stable for |Λ| < 1 and locally unstable for |Λ| > 1.

Proof: Consider some starting value N0 close to an equilibrium point N ∗ . Due to the
mean value theorem, there is a c ∈ (N0 , N ∗ ) such that
∂F (N ) F (N0 ) − F (N ∗ ) F (N0 ) − N ∗
= = .
∂N N =c N0 − N ∗ N0 − N ∗
Since the derivative of F (N ) is continuous, we have
∂F (N )
<1
∂N c

20
if |Λ| < 1, and thus |F (N0 ) − N ∗ | < |N0 − N ∗ |. We thus get a sequence Ni that converges
to the equilibrium point N ∗ . Similarly, we are (locally) driven away from the optimum for
|Λ| > 1.

Figure 3: Cobwebbing method to determine the stability of equilibria.

Cobwebbing We can represent the discrete dynamics easily using the cobwebbing method:
• Start at N0 , go (vertically) to F (N0 ).
• Project (horizontally) to the identity line Id(N ) to get N1 = F (N0 ).
• Go vertically to F (N1 ) and iterate.
Depending on the slope Λ near an equilibrium point, we get four types of characteristic
behavior.
1. 0 < Λ < 1: monotone convergence.
2. −1 < Λ < 0: oscillating convergence (damped oscillations)
3. Λ > 1: monotone divergence
4. Λ < −1: oscillating divergence (driven oscillations)

21
2.4 Analysis of the Beverton-Holt and the Ricker model
Beverton Holt model Consider the dynamics of the Beverton-Holt model defined by
c1 Nt
Nt+1 = F (Nt ) = sNt + . (58)
1 + c2 Nt
• We have
N (1 − s)(1 + c2 N ) = c1 N
and thus N1∗ = 0 and
c1 − 1 + s
N2∗ =
c2 (1 − s)
• With
∂F (N ) c1
F 0 (N ) = =s+ 2
≤ s + c1 = F 0 (0)
∂N (1 + c2 N )
c1 (1−s)(1−s−c1 )
and F 0 (N2∗ ) = s + (1+(c1 −1+s)/(1−s))2 = 1 + c1
< 1 for s + c1 > 1 and thus
∗ ∗
N2 > 0. We get monotonic approach of N2 for s + c1 > 1 and monotonic approach
of N1∗ = 0 otherwise.

Lemma In general, if F (N ) is continuous and monotonic on a closed interval I ⊆ R+

(e.g., I = [0, ∞)) with F (I) ⊆ I, then every sequence Ni = F i (N0 ) with N0 ∈ I converges
monotonically to a fixed point of F or to ∞.

Proof For F (N0 ) > N0 , we have F (F (N0 )) ≥ F (N0 ) because of monotony, and thus
N0 < F (N0 ) ≤ F (F (N0 )) ≤ .... The sequence will either converge to some N ∗ or diverge
to infinity. In the former case, we have F (N ∗ ) = N ∗ since F is continuous. The case
F (N0 ) < N0 works analogously, F (N0 ) = N0 is trivial.

Ricker model Consider now the Ricker model (we use s = 0, without restriction under
rescaling),
Nt+1 = F (Nt ) = Nt exp[r(1 − Nt /K)] . (59)
• We obtain equilibria at N1∗ = 0 and N2∗ = K.
• We have  Nr
F 0 (N ) = 1 − exp[r(1 − N/K)] .
K
and thus

Λ1 = F 0 (0) = exp[r] > 1, (60)

Λ2 = F 0 (K) = 1 − r. (61)

• We have a unique maximum of F (N ) at Nmax = K/r (and, of course a minimum

Nmin = 0).

22
We thus can distinguish three dynamical regimes (see, for example, Populus, density-
dependent population growth, “discrete logistic” = Ricker)

1. For r < 1, the interval I = [0, Nmax ] contains both equilibrium points. Since we
have F (N ) monotonically increasing in I, with F (I) ⊆ I, we obtain monotonic
convergence to the stable equilibrium N2∗ = K for every start value N0 ∈ I. For start
values N0 ∈
/ I, we have F (N0 ) ∈ I and the same applies after the first step.

2. For 1 < r < 2, we have no longer monotonic convergence, but still |Λ2 | < 1. Hence,
N2∗ = K is still asymptotically stable and approached by damped oscillations.

3. For r > 2, we have |Λ2 | > 1 and therefore do no longer have a stable equilibrium. In
particular, we get oscillating divergence for the equilibrium at N2∗ = K. Since the
Ricker model has a finite maximum population size, we can ask what happens.

2.5 Limit cycles

The basic idea to gain further insights of the long-term dynamics of the Ricker model and
similar systems is to consider the iterated map

F (2) (N ) = F (F (N )) resp. F (k) (N ) = F (. . . F (F (N ))) k − fold (62)

across k seasons. Obviously, this is again a discrete dynamical system and can be studied
in the same way as the original one. Concerning its long-term behavior, we observe the
following elementary facts:

1. Equilibria of F are also equilibria of F (2) and of every higher iteration F (k) .

2. For the derivative of F (2) at the fixed points, we have, using the chain rule,

∂F (2) (N ) ∂F (Z) ∂F (N )
= · . (63)
∂N ∂Z Z=F (N ) ∂N

and accordingly iterated for F (k) . For equilibria N ∗ of F (N ), in particular, this

means
k
∂F (k) (N )

∂F (N )
= . (64)
∂N N =N ∗ ∂N N =N ∗
Thus, stable equilibria of F remain stable for all higher iterations, and unstable
equilibria remain unstable.

3. Additional (stable or unstable) equilibria of iterated maps F (k) can occur if the
original dynamics F leads to cycling.

23
Definition and basic properties: Limit Cycles

1. A point N0 is called a point of period k or a k-cycle point if it is a fixed point

(equilibrium) of the k-fold iterated map F (k) (N ), but not a fixed point of any map
0
F (k ) with 1 ≤ k 0 < k. Its orbit

{N0 , F (N0 ), . . . , F (k−1) (N0 )} =: {N0 , N1 , . . . , Nk−1 }

is called the corresponding k-cycle. Note that all points Ni in the cycle are fixed
points of F (k) .

2. A limit cycle is asymptotically stable, if and only if the corresponding cycle points
are asymptotically stable equilibria of the k-fold mapping. In particular, we obtain
via the chain rule
k−1
(k) ∂F (k) Y ∂F (N )
Λ = = (65)
∂N i=0
∂N N =Ni

for the derivative of F (k) at all points of the cycle. The cycle and the corresponding
fixed points of F (k) are asymptotically stable if |Λ(k) | < 1 and unstable if |Λ(k) | > 1.

3. We define the characteristic exponent (or Floquet exponent)

k−1
1 1X ∂F (N )
λ= log Λ(k) = log . (66)
k k i=0 ∂N N =Ni

Obviously, the cycle is stable for λ < 0 and unstable for λ > 0.

2.6 Example: Discrete logistic growth

As an example, we consider logistic growth
Nt 
Nt+1 = F (Nt ) = rNt 1 − , (67)
K
which behaves similar as the Ricker equation, but is easier to analyze. We have a maximum
of F (N ) at N = K/2, where F (K/2) = rK/4. There are two equilibria at N1∗ = 0 and at
N2∗ = K(r − 1)/r. With F 0 (N ) = r(1 − 2N/K), we obtain F 0 (0) = r and F 0 (N2∗ ) = 2 − r.

1. For 0 < r < 1, we have a single stable equilibrium at N1∗ = 0 and the population dies
out.

2. For 1 < r < 3, a second equilibrium N2∗ appears and is stable, while N1∗ is unstable.
Approach to the stable equilibrium is monotonic for r < 2 and oscillating for r > 2.

3. For r > 3, both equilibria N1∗ and N2∗ are unstable and we can expect limit cycles or
other types of limit behavior.

24
Finally, for a reasonable biological model, r should not be larger than 4 since otherwise
F (N ) can get larger than K and then negative in the next iteration. To study the dynamics
via iterated maps, assume K = 1 for simplicity. We then obtain:

F (N ) = rN (1 − N ) (68)
(2) 2 2
F (N ) = r(rN (1 − N ))(1 − rN (1 − N )) = r N (1 − N )(1 − rN + rN ) (69)
(2)
∂F (N )

= r 1 − 2rN (1 − N ) r(1 − 2N ) (70)
∂N

(Note that, in general, we have the symmetry F (k) (1 − N ) = F (k) (N ).) We find equilibria
F (2) (N ) = N for
p
r − 1 1 + r ± (r − 1)2 − 4
N1∗ = 0 , N2∗ = , N3,4∗
= (71)
r 2r
∗
Obviously, N3,4 exist for r ≥ 3. We also find that F (N3∗ ) = N4∗ and vice-versa F (N4∗ ) = N3∗ .
The corresponding derivatives are
(2) (2) (2)
Λ1 = r 2 Λ2 = (2 − r)2 , Λ3,4 = 4 + 2r − r2 (72)

• At r = 3, we have N2∗ = N3∗ = N4∗ . We see that, by increasing r beyond this

threshold, the previously stable equilibrium N2∗ turns unstable. At the same time
(2)
two new equilibria of F (2) (N ), N3∗ and N4∗ , appear and are stable: |Λ3,4 | < 1 for
√
3 < r < 1 + 6 ≈ 3.45. This is the typical signature of a pitchfork bifurcation. For
the original map F (N ), the number of equilibria does not change. The new equilibria
for F (2) correspond to a stable limit cycle with period 2.
√ (2)
• For r > 1 + 6, we have Λ3,4 < −1 and the equilibria N3∗ and N4∗ of F (2) turn
unstable. For the iterated map F (4) this means that the slope at both equilibria
(2)
increases to (Λ3,4 )2 > 1, thereby generating, once again, two new equilibria in a
pitchfork bifurcation. We thus obtain a 2-cycle for F (2) , corresponding to a 4-cycle
for F (N ), in each case. In general, every time when a stable equilibrium of F (k)
becomes unstable, we get two new stable equilibria of F (2k) , and thus a limit cycle
with the period 2k. This is the so-called period doubling cascade of the logistic map
and many similar discrete maps (such as the Ricker model).

2.7 Bifurcations for discrete time dynamics

We have encountered several cases of bifurcations in the examples, where fixed points
originate or disappear or change their stability. They can be summarized as follows. First,
the same three types of bifurcations that exist for the continuous time dynamics, Ṅ = f (N ),
also exist in discrete time with iteration function Nt+1 = F (Nt ). Instead of the zeros of
f (N ), we now need to consider the zeros of F (N ) − N . We obtain a

25
 • transcritical bifurcation if two zeros of F (N ) − N cross and the corresponding fixed
points exchange stability (e.g. Beverton Holt model for s + c1 = 1);

• saddle node bifurcation if two fixed points (one stable, one unstable) are either newly
generated or annihilated. This happens if Id(N ) is tangent to F (N ) and F (N ) then
either crosses or moves away from this line;

• pitchfork bifurcation if a single fixed point splits into three fixed points or, vice-versa,
if three fixed points merge and a single one remains (e.g. bifurcation at r = 3 for the
iterated logistic map F (2)).

All these bifurcation readily occur also for monotonically increasing iteration functions
F (N ). With decreasing F (N ), another type of bifurcation is possible when a fixed point
becomes unstable as F 0 (N ) drops below −1. As described for the logistic map, this typically
leads to a pitchfork bifurcation for F (2) . For F (N ), it leads to a (stable) limit cycle instead
of a single fixed point. This type of bifurcation that cannot occur in the continuous time
dynamics. More generally, we call it a

• period doubling bifurcation if a stable fixed point or an existing stable limit cycle
becomes unstable and a new stable limit cycle (with a period of double length)
emerges.

2.8 Chaos
√
For the logistic map, we get a series of critical values r2k (with r2 = 3 and r4 = 1 + 6 ≈
3.45), above which stable cycles of period 2k exist. As it turns out, this sequence of period
doubling bifurcation points quickly converges to a finite value rc = r∞ ≈ 3.57. We can ask
what happens beyond this point. As an example, we can consider the logistic map with
r = 4 and dissect the interval [0, 1] as follows

1. Two intervals I0 = [0, 1/2] and I1 = [1/2, 1]. Both intervals are mapped by F (N ) to
the full domain [0, 1].

2. Dissect I0 such that I00 = [0, q] is mapped by F to I0 and I01 = [q, 1/2] is mapped
to I1 . Similarly, dissect I1 into I10 and I11 .

3. Iterate this: I000 ⊂ I00 is mapped by F to I00 and by F (2) to I0 ; I001 ⊂ I00 is mapped
to I01 and then to I1 , etc. In general, Ii0 i1 i2 ...ik ⊂ Ii0 with ij ∈ {0, 1} is mapped by F
to Ii1 i2 ...ik ⊂ Ii1 and by F (k) to Iik .

4. In general, if (e.g.) N ∈ I0100 , then N ∈ I0 , F (N ) ∈ I1 , F (2) (N ) ∈ I0 , F (3) (N ) ∈ I0 .

We can identify each point with an infinite interval nesting (dyadic transformation).

5. Points with periodic nesting correspond to periodic orbits. There are points with ev-
ery period, starting with I0000... and I1111... , which encode the (unstable) fixed points.

26
 6. All points on interval boundaries are attracted by 0 (dense set: population on the
verge of extinction; this is special for r = 4).
7. Almost all points (with no period in the nesting) must have orbits of infinite length.
We also see that orbits of very close points can differ widely after even a few gener-
ations. This is the signature of chaos.
8. For the special case r = 4, there is an explicit solution for the dynamics
1 1/2
Nk = sin2 [2k−1 πθ] where θ = sin−1 [N0 ] . (73)
π
We see that we get aperiodic orbits for all irrational θ.
These observed phenomena are not specific to r = 4, but occur for most values r > rc . We
can summarize them as follows:
• There are infinitely many periodic points, with all kinds of periods, even and odd.
• For most r values, we obtain chaos, indicated by orbits of infinite length that never
come back to a starting point. Another characteristic of chaotic solutions is that
very small initial differences are magnified due to the mapping. This means that the
long-term predictability of the system gets lost (butterfly effect).
• As analog to the Floquet exponent for finite cycles, one defines the Lyapunov exponent
k−1
1X ∂F (N )
λ = lim log (74)
k→∞ k ∂N N =Ni
i=0

for any point N0 with infinite orbit. A positive Lyapunov exponent indicates that
small initial differences are magnified and the resulting orbit is chaotic.
• The chaotic regimes are interlaced by intervals with stable limit cycles with low
period (e.g. 3-cycles).

Periodicity and Chaos in Biology Our examples show that even very simple models
(in one dimension!) can give rise to very complex phenomena. Periodic dynamics are a
frequent observation in biology, also in population dynamics. There can be multiple causes
for such cycles. With discrete dynamics, cycling is created by an overshooting of the equi-
librium, which is typical of many systems where regulation is not immediately effective,
but acts with a certain delay. Essentially the same behavior is seen in continuous delay
ODE’s. Chaotic behavior in biology has been speculated a lot. From a theoretical point
of view, we should expect chaos in particular for systems with higher dimension (multi-
ple interacting species). Even with continuous dynamics, chaos exists in dimension ≥ 3.
However, convincing empirical evidence is hard to obtain. In particular, it is often difficult
(or impossible) to distinguish long cycles and/or stochastic noise from “real” deterministic
chaos. Finally, true chaos is never strictly possible in populations with a discrete number
of individuals. Nevertheless, chaos is a field where biology has inspired mathematics and
has helped to found chaos theory as a research field (Robert May 1974, 1976).

27
3 Structured populations
In all dynamical models considered so far, we have assumed that all individuals behave
equally with respect to the population dynamics. All have the same birth and death rates,
and all contribute to competition in the same way. Although for the discrete dynamics,
we did consider individuals at different life stages (juveniles and adults), this only had
an indirect effect on the recurrence equation F (N ). Indeed, there was at least one time
point during the life cycle when all individuals were equivalent (adults just before giving
birth). We could thus use this life stage for the dynamics, while other stages (juveniles)
were “integrated out”. However, this is no longer possible if individuals at different life
stages coexist at all times.

Age structure and Leslie matrices The typical example is a population with an
age-structured demography. Let xi the number (or density) of individuals in age-class i,
i ∈ {1, . . . , n}. Let pi be the survival probability from age class i to i + 1 and let fi be the
fertility (expected number of offspring) of each member of age-class i. Then

xi+1 (t + 1) = pi xi (t) i = 1, . . . , n − 1 (75)

X
x1 (t + 1) = fi xi (t) (76)
i

which can be summarized as

x(t + 1) = Lx(t) (77)
with the Leslie matrix (Leslie 1945)
 
f1 f2 · · · fn−1 fn
p1 0 · · · 0 0
 
L =  0 p2 · · · 0 0 (78)


 .. . . .. 
. . .
0 0 · · · pn−1 0

Other examples of structure

1. Spatial structure: A population lives in a structured habitat. Birth and death rates
depend on the local characteristics and there is migration among the demes.

2. Different genotypes: Birth and death rates depend on the heritable genotype. Tran-
sition between genotype classes can occur due to mutation during reproduction.

3. Size structure: instead of age as such, for many species (like trees) the size is the
more decisive factor.

28
 • In general, the art of modeling is to decide which structure best describes a popula-
tion. It is not necessary that individuals in one class have all the same demographic
parameters: we can consider average quantities. However: these average quantities
should only depend on the state the individuals are in and not on the history (i.e.,
where they came from): We thus want to make sure that the population dynamics
can be described by a generalized Markov model (but note that the transition matrix
is not normalized).
• We can illustrate all these models as graphs. Arrows between states represent non-
zero transition probabilities. The transition matrix is also called the population
projection matrix.

General formalism: linear model Let xi (t) the number of individuals in class i at
time t, i ∈ {1, . . . , n}. Let mij := mi←j be the transition rate from state j to state i and
M the corresponding population projection matrix. We assume that the transition rate (=
demographic parameters) are constant. We thus consider a (time)-homogeneous process.
In particular, we also assume ignore any density- or frequency-dependence. We then have
a linear model with explicit solution
x(t) = Mt x(0) . (79)
We ask the following questions:
• What is the long-term growth rate of the population? Do we get growth or extinction?
• What kind of distribution over the states xi does result in the long run? −→ Analog
to stationary distribution of a Markov process.

Definitions: Irreducible, primitive, spectral radius

• A non-negative matrix M is called irreducible if for each pair [i, j] there is some
k(i, j) ∈ N, such that [Mk(i,j) ]i,j is positive.
Biologically, a population projection matrix is irreducible if and only if a path exists from
every state to any other state in the corresponding (life-cycle) graph. Most population
models are irreducible. Exceptions are life-cycles with post-reproductive states and spa-
tially structured models with sources and sinks.
• A non-negative matrix M is called primitive if there is some k ∈ N, such that Mk is
positive.
Obviously, primitive implies irreducible. Irreducible, but non-primitive projection matrices
are characterized by loops in their corresponding graphs. An irreducible matrix M is
primitive if and only if the greatest common divisor of the loop lengths is 1. Sufficient
conditions for an irreducible matrix to be primitive are (1) any self-loop (2) positive fertility
of adjacent age-classes in a Leslie matrix.
• The spectral radius of a matrix M is r(M) := maxi |λi |, where λ1 , . . . , λn are the
eigenvalues of M.

29
Theorem: Perron-Frobenius (Perron 1907, Frobenius 1912) Let M := (mij )1≤i,j≤n a
matrix with real, non-negative entries mij ≥ 0. Then
1. M has a real leading eigenvalue λ1 = r(M) (called dominant EV). Right and left
eigenvectors v and u corresponding to λ1 with non-negative entries exist.
2. If further M is irreducible, it has a dominant EV λ1 > 0, which is simple (has
geometric and algebraic dimension one). All entries of the corresponding eigenvectors
v and u are strictly positive.
3. Assume further that M is primitive. Then λ1 is strictly dominant, i.e., larger than
all other eigenvalues, λ1 > |λi |, i 6= 1.

Proof
(a) Convergent matrices
For any (square) matrix M,
r(M) < 1 ⇔ Mk → 0 for k → ∞ (80)
Proof: Obvious for any matrix D in Jordan from and Mk = SDk S−1 (a matrix
M with this property is called convergent).
(b) Lower bound for the spectral radius
Let M ≥ 0 and 0 6= v ≥ 0 and λ ≥ 0. Then
n
X
∀i : [Mv]i = mij vj > λvi ⇒ r(M) > λ . (81)
j=1

Proof: ∃  > 0 : Mv > (λ + )v (for every component). Thus

M  M 2  M k
v< v< v< v ∀k > 2
λ+ λ+ λ+
because Mv with mij ≥ 0 is monotonic in every component vi . For k → ∞ we see
that M/(λ + ) has a lower bound and is thus not convergent. With (a) we conclude
that r(M) ≥ λ +  > λ.
(c) Positivity of the leading eigenvalue and -vector
Let M > 0 and λ1 eigenvalue with |λ1 | = r(M). Then λ1 = r(M) and any corre-
sponding right eigenvector is strictly positive, v > 0 (up to a common scalar factor
for all entries of v). The same holds for any dominant left eigenvector.
Proof: Let λ1 v = Mv and assume that the entries of v are not collinear in the
complex plane (where collinear means positive real or zero up to a common scalar
factor). Then
n
X n
X
|λ1 ||vi | = mij vj < mij |vj | ∀i = 1, . . . , n .
j=1 j=1

30
 Thus |λ1 ||v| < M|v| (in each row) and thus |λ1 | < r(M) due to (b), contradicting
the assumption. Thus all entries of v are collinear and can be chosen as ≥ 0. Since
n
X
mij vj = λ1 vi
j=1

also λ1 must be real and positive. Since mij > 0, we also have vi > 0 ∀i. Proof for
the left eigenvector follows by switching to MT .
(d) Geometric and algebraic dimension of the dominant eigenvalue
Let M > 0. Then the geometric and algebraic dimension of the dominant eigenvalue
λ1 = r(M) is one.
Proof: Geometric: Assume that there are two eigenvectors v and w corresponding
to λ1 . Then a linear combination v + κw exists with at least one component zero,
contradicting (c). The proof for the left eigenvector follows again by switching to MT .
Algebraic: Assume that w exists, such that (M − λ1 I)2 w = 0, but (M − λ1 I)w 6=
0 (generalized eigenvector). Then x = (M − λ1 I)w is an eigenvector of M with
eigenvalue λ1 = r(M). Due to (c) we can choose x > 0. Let u > 0 be the dominant
left eigenvector. Then we obtain a contradiction with
0 = uT · (M − λ1 I)w = hu; xi > 0 .
Remark: c) and d) imply strict dominance of λ1 . This completes the proof of the
Perron-Frobenius Theorem for strictly positive matrices.
(e) Primitive matrices
Let M be primitive and Mk > 0. Let λ1 be some eigenvalue of M with |λ1 | = r(M)
and corresponding eigenvector v. Then v is also eigenvector of Mk with eigenvalue
λk1 = r(Mk ). From the theorem for positive matrices we know v > 0, thus also λ1
is strictly dominant, real and positive and has dimension one. This proves part 3 of
the Theorem.
(f) Irreducible matrices
Let M be irreducible. Then M + I is primitive. Let λ = λ +  be the strictly
dominant eigenvalue of M + I with corresponding eigenvector v. Note that v does
not depend on  and is, in particular, also eigenvector of M with eigenvalue λ. Since
v > 0, we must have λ > 0. Since I just shifts the spectrum, λ has dimension one
and we have λ = r(M) due to continuity, but other eigenvalues |λ̃| = r(M) may
exist.
(g) Non-negative matrices
Let M ≥ 0 and  > 0. The theorem for positive matrices holds for any matrix
M = (mij + )1≤i,j≤n . Let λ = r(M ) be the strictly dominant eigenvalue and
v > 0 the corresponding eigenvector. For  → 0 we get v → v ≥ 0, where v
is eigenvector of M with eigenvalue λ = r(M). Note that λ need not be strictly
dominant and its dimension need not be one.

31
3.1 Long term behavior
Consider the linear population dynamics
x(t) = Mt x(0) . (82)
Assume first that M is primitive. (This is the case for most models.) Then there is a
strictly dominant eigenvalue λ1 = r(M) with positive left and right eigenvectors u1 and
v1 . We can normalize these according to kv1 k = 1 and hu1 ; v1 i = 1. We then have
 M t
−→ v1 ⊗ uT1 for t → ∞ . (83)
λ1
If M can be diagonalized with eigenvalues {λi }i , and corresponding sets of left and right
eigenvectors, {ui }i and {vi }i , respectively, with hui ; vj i = δij . Then
n
X
t
M = λti vi ⊗ uTi . (84)
i=1

Proof. Let U be the orthogonal complement of the vector uT1 . Note that M maps U
onto itself. Let M0 be the restriction of M to U. Then the spectral radius of M0 /λ1 is
smaller than one, and the matrix hence convergent. Since v1 is not orthogonal to u1 we
can chose v1 and any basis of U as basis of the full vector space. The proof follows from
the action of (M/λ1 )t on this basis in the limit t → ∞. The proof of (84) is analog, using
the pairwise orthogonality of sets of left and right eigenvectors. It also holds if M is not
primitive.

Consequences.
• For any population vector x(0) and any irreducible matrix M observe that hu1 ; x(0)i >
0. For any primitive matrix M we then obtain the long-term growth rate
kx(t + 1)k1 k(M/λ1 )(t+1) x(0)k1
r∞ = lim = λ1 lim = λ1 . (85)
t→∞ kx(t)k1 t→∞ k(M/λ1 )t x(0)k1

• The limit distribution reads

x(t)
lim = v1 . (86)
t→∞ kx(t)k1

For age structured models, this is also called a stable age structure.
• We can estimate the rate at which the limit distribution is approached by considering
the second largest eigenvalue |λ2 |, as follows
t

λ2
rt = r∞ + C x(0) (87)
λ1
with a constant C that depends on the initial conditions. An analog results holds for
the limit distribution.

32
 • For any irreducible, but periodic matrix, Eq. (85) still holds, but there is no unique
limit distribution (86). We can still characterize the long-term behavior by consid-
ering all eigenvectors to eigenvalues |λi | = r(M) and their linear combinations.

• For reducible matrices, we first try to identify absorbing states or subspaces and then
consider the dynamics on irreducible subsystems defined by this absorbing states and
their complements.

3.2 Eigenvalues of Leslie matrices

In general, it is difficult (or impossible) to derive the eigenvalues and eigenvectors of high-
dimensional matrices. For the special form of the Leslie matrix, however, this is simplified.
Here, we have from Lx = λx,
p1
p1 x1 = λx2 ⇒ x2 = x1 (88)
λ
p2 p1 p2
p2 x2 = λx3 ⇒ x3 = x2 = 2 x 1 (89)
λ λ
pn−1 p1 · · · pn−1
pn−1 xn−1 = λxn ⇒ xn = xn−1 = x1 (90)
λ λn−1
We thus have

λx1 = f1 x1 + . . . + fn xn (91)
 p1 p1 p2 pn−1 · · · p1 
= x1 f 1 + f 2 + f 3 2 + . . . + f n . (92)
λ λ λn−1
If we define the probability for a newborn to survive until age i as l1 = 1 and li =
p1 · · · pi−1 , i > 1, we obtain the so-called Euler-Lotka equation
n
X
1= fi li λ−i =: φ(λ) . (93)
i=1

The Euler-Lotka equation relates the asymptotic growth rate (also called the fitness of an
individual) to the life-history parameters fi and pi (or li ). It is easy to see that the expres-
sion on the right-hand-side is strictly monotonically decreasing. We thus obtain exactly
one solution for a real eigenvalue λ = r(L). One can further show that Leslie matrices are
always diagonalizable; all other eigenvalues are complex. There is still no general explicit
solution for λ1 , but we can at least determine whether λ1 > 1 (the population is grow-
ing) orPλ1 < 1 (shrinking) population. For this, define the life-time reproductive success
R0 = i fi li . We see that R0 is the value of the Euler-Lotka equation at λ = 1. From
the shape of the graph, we immediately see that λ1 > 1 ⇔ R0 > 1 (which makes intuitive
sense). Once we know λ, it is easy to obtain the corresponding eigenvector
l l2 ln T
1
v1 = , , . . . , (94)
λ1 λ21 λn1

33
3.3 Application: Life history evolution
Typical problems of life history theory are as follows:
• Why do we age? Why are there widely different life spans among species?
• At a given age class: is it better to invest into survival or into reproduction?
• Is it better to have many offspring once or fewer offspring more often during life?
All these problems ask why certain life-history parameters in a Leslie matrix take their
given values. We can try and approach this problem from an evolutionary point of view:
Consider a resident population that is characterized by a certain set of life-history pa-
rameters. Assume now that a mutant appears in this population with a different set of
parameters. We can always ask: Will the mutant replace the resident? To answer this
question, we first need to know how individuals of different types compete and how popu-
lation regulation is included into our framework of structured models. Since this leads to
non-linear models, this can be very complicated, indeed! So simplify things, we will assume
here that population regulation occurs in an independent step and affects all individuals
in the same way, i.e.,

x(t) → x(t + 1/2) = Lx → x(t + 1) = f (N )x(t + 1/2) (95)

P
where N = i xi is the total population size and f (N ) takes care of density regulation.
For example we could have
1
f (N ) = . (96)
1 + cN
This scheme has two important properties
1. First, since density regulation acts only as a factor on the population vector, it does
not alter the distribution. In the long term, we will thus reach the same stable age
distribution as with the linear model.
2. Assume now that this stable age distribution has been reached. We then have

N (t + 1) = λN (t) · f (N (t)) (97)

where λ is the long-term growth rate of the linear model.

We thus can separate our problem and treat both parts (age distribution and density
regulation) independently. Indeed, this is also the reason why models without age structure
make some sense at all. We simply assume (implicitly) that a stable age distribution has
been reached. Now, what will happen with a rare mutant in such a case? (We assume no
recurrent mutation)
• A mutant will start its own clone and both clones will reach a stable age distribution.
Since they are subject to joint density regulation, the clone with the higher limiting
growth rate (higher fitness) will win and displace the other type.

34
Example: Iteroparous versus semelparous reproduction Why do many plants and
some animals reproduce only once in their lifetime and then die (semelparous reproduction),
while others reproduce multiple times (iteroparous reproduction)? – In principle, being
able to reproduce multiple times is of course be an advantage and should be selected.
However, there may be a cost: concentration on a single reproduction event could increase
the expected litter size. In a much simplified version, compare two strategies:

1. On the one hand, consider an immortal iteroparous clone with fertility fit every
season.

2. On the other hand, consider a semelparous clone, where individuals reproduce after
the first season with fertility fsem and then die.

How large must fsem be that the semelparous clone matches the growth rate of the iteroparous
clone? (Assume that all our results can be extended to infinite dimensional matrices in
this case.)

• Iteroparous: Euler-Lotka equation

∞
X λ−1
1 = fit λ−i = fit
i=1
1 − λ−1

and hence λ = fit + 1

• Semelparous:
1 = fsem λ−1
and thus λ = fsem .

We thus see that only a single individual more in the first year is worth as much as eternal
life and offspring every year. The question arises why there are so many iteroparous species!
More generally, we can derive the predictions for an adult survival probability pa and a
juvenile survival probability pj . Here, we assume that pj is the probability that juveniles
survive even until the first census. then pj acts like a factor to the fertility. We obtain:

• Iteroparous:
∞
X
1 = fit pj (p(i−1)
a λ−i )
i=1

and thus λ = pa + pj fit .

• Semelparous:
1 = fsem pj λ−1
thus λ = fsem pj .

35
We get equal long-term growth rates for
pa
fsem = fit + .
pj

This leads to the prediction that species with a higher ratio of adult / juvenile survival
probability should more likely be iteroparous (or have more breeding seasons, on average).
This is, indeed, confirmed by data.

3.4 Reproductive values and sensitivity analysis

How important are the different (age-)classes for the population growth? Or: How much
does an individual in the initial population contribute to future population growth? We
have X
x(t) ≈ λt v ui xi (0) (98)
i

where v is the leading right eigenvector and ui are the entries of the leading left eigenvector.
We thus see that the ui are weighting factors that determine the importance of age-class i
for future population growth. They are called reproductive values. For the Leslie matrix,
we can derive the reproductive values as follows. Let u1 be the reproductive value of
newborns. In the distant future, the clones started from any individual at time zero will
grow with factor λ per time step. But initially, the relative growth rates are different. We
can relate the size of a clone started from an individual in age class i to the size of a clone
started from a newborn as follows:
 
fi li+1 fi+1 li+2 fi+2 ln fn
ui = + + + ... + u1 (99)
λ li λ2 li λ3 li λn−i+1
n
λi−1 X
= lj fj λ−j u1 . (100)
li j=i

The factors λj in the denominator accounts for the delay of j time steps for the newborn
offspring individuals produced by our focal individual in age class i. Note that the equation
for u1 , in particular, reproduces the Euler-Lotka equation. With the choice u0 = λ, we
have
n
λi X
ui = lj fj λ−j (101)
li j=1

Typically, the reproductive value increases until an individual reaches maturity and then
decreases. For post-fertility age-classes it is zero. Reproductive values also tell us how
severe perturbations in the number of individuals in a given age class are for future pop-
ulation dynamics (relevant for population management). Another use of the reproductive
value is in sensitivity analysis.

36
Sensitivity analysis In general, we can ask, which demographic parameter “matters
most” for the long-term growth rate. In an evolutionary perspective, this corresponds to
the parameters that experience the strongest selection pressure. For a general projection
matrix M = (mij )1≤i,j≤n , the answer is given by the so-called sensitivities
∂λ
sij := . (102)
∂mij
Frequently, sensitivities are measured on a scale of relative change rather then absolute
change. We then obtain the so-called elasticities
∂ log λ mij
eij := = sij . (103)
∂ log mij λ
With
uT · Mv
λ= (104)
uT · v
we simply get
ui vj
sij = . (105)
uT · v
In general, if several entries of the projection matrix depend on some model parameter x,
mi,j = mi,j (x), we obtain the sensitivity with respect to any such x as a linear combination
of the sij ,
∂λ X ∂mi,j
sx = = sij · . (106)
∂x i,j
∂x
For Leslie matrices M = L, we have
n X
X n n
X
−j
T
u ·v = lj fj λ = k fk lk λ−k (107)
k=1 j=k k=1

and hence
−k
(lj /li )λi−j
P
k=i lk fk λ
sij = Pn . (108)
k=1 k lk fk λ−k
In particular, this gives
∂λ lj λ1−j
= s1j = n
P −k
(109)
∂fj k=1 k lk fk λ
and
(lj /lj+1 )λ nk=j+1 lk fk λ−k
P
∂λ
= sj+1,j = Pn −k
. (110)
∂pj k=1 k lk fk λ
We can make the following observations
• Early-in-life fecundities are under stronger selection than late-in-life fecundities (for
λ ≥ 1).

37
 • Also selection on viabilities monotonically decreasing and only > 0 as long as the
fertility later in life is still positive. As a consequence, we have weak sekection
against deleterious mutations with an effect late in life: evolutionary explanation for
senescence and death.

• Assume that there is a trade-off between viability and fecundity at age i. We have
 ∂λ   ∂λ 
/ = uj /λ (111)
∂pj ∂fj

Thus, age classes with high reproductive value should invest more into survival than
age classes with low reproductive value. In particular, survival at maturity is more
strongly selected than survival of newborns.

3.5 Other formalisms to model population structure

We have discussed in some detail the projection matrix approach to population struc-
ture. This formalism works in discrete time and it assumes that the ppopulation can be
structured into discrete classes. Both of these assumptions are convenient, but analogous
approaches, which do not make these assumptions also exist. This leads either to so-called
Lotka integral equation or to the McKendrick-von Foerster partial differential equation.
Similar to the matrix approach, both formalisms allow for (partial) solutions as long as
the dynamics is linear. In general, also similar conclusions can be drawn.

38
4 Dynamics of interacting populations
So far, we have considered the dynamics of a single species. However, the major part of
ecology is really about interactions of species that live in the same ecosystem. And even if
we are only interested in a single species, we need to take into account that its dynamics in
nature will crucially be influenced by the presence (or absence) of other species. Whenever
these other species dynamics on a similar time scale, we need to include them into the
model. For a pair of species, there are three basic types of interactions:

1. Competition (e.g. for food or for nesting sites). Presence of individuals from the
other species reduces population growth.

2. Mutualism. Both species profit from the presence of the respective other species.
Typical examples are plant-pollinator interactions, or seed-dispersal systems, where
plants produce fruit to reward animal-aided dispersal. Another class are the interac-
tions of endosymbionts (bacteria, fungi) and their hosts (most animals).

3. Host-parasite or predator-prey. Asymmetric consequences of the interactions.

Presence of the host/prey is good for the predator – but not vice-versa.

In these scenarios, the density regulation for a given species will typically not only depend
on the total density of individuals across all species, but on the individual densities of
all species: for a predator it will matter a lot whether other individuals around are prey
or rather competing predators. We therefore cannot use the simplistic modeling scheme
of the previous section, where we have assumed that density regulation depends only on
the total number of all individuals (in various age classes) and affects all these individuals
in the same way. We will therefore introduce a new class of models for the population
dynamics of interacting species. We will use the formalism in continuous time, which leads
to coupled ordinary differential equations. Let xi (t) be the density of species i at time t
and 1 ≤ i ≤ n. Then
ẋi (t) := fi (x1 (t), x2 (t), . . . , xn (t)) (112)
with fi continuously differentiable on some subset of Rn . Collectively, we can write

ẋ(t) = f (x(t)) (113)

defining a time-homogeneous, coupled, first-order ODE system. There is a unique solution

for each initial condition x, which can be represented as a line x(t) in the phase space,
where f (x) are tangent vectors to x(t) defining the velocity of the dynamics. In this
notation, x(t), y(t) etc. are solutions of a system of differential equations ẋ = f (x) with
the same f , but with different initial conditions x and y. In general, solutions need not
exist for all times t (but only for an open time interval). However, if the dynamics keep
x(t) in a compact set, the solution exists for all times −∞ < t < ∞. Since population
densities are non-negative and also won’t increase without bounds, this is always the case
for a meaningful biological model.

39
 10000

predator population
8000

6000

4000

2000

0
0 2000 4000 6000 8000 10000 12000
prey population

Figure 4: Phase space with vector field for the Lotka-Volterra predator-prey model (rx = 1,
cx = −10−4 , cxy = −10−4 , ry = −0.05, cy = −10−5 , cyx = 10−5 ) with solutions for various
starting conditions.

4.1 Lotka-Volterra model

As our main example, we will consider interactions between two species, x(t) and y(t),
following the Lotka-Volterra dynamics,

ẋ = fx (x, y) = rx x + cx x2 + cxy x y , (114a)

ẏ = fy (x, y) = ry y + cy y 2 + cyx y x . (114b)

The model has the following properties:

• rx and ry describe the intrinsic growth rates of both species while rare and in the
absence of the other species. These growth rates can be positive or negative, for
example for a predator in the absence of prey or for two mutualists that crucially
depend on each other.

• The constants cx and cy describe the consequences of within-population interactions.

These are usually (but not always) dominated by competition, such that cx and cy
are negative.

• If the interaction parameter cxy (or cyx ) is zero, the x (or y) population follows a
single-species dynamics. With cx = −rx /Kx (resp. cy = −ry /Ky ), we obtain 1-dim.
logistic growth.

40
 • In general, the sign of the interaction parameters determines the type of the biological
interaction:
1. For cxy , cyx < 0 we obtain a competition model;
2. cxy < 0 and cyx > 0 corresponds to an x-host-y-parasite (or x-prey-y-predator)
model;
3. finally, cxy , cyx > 0 indicates mutualism.
• The classical Lotka-Volterra model has rx > 0, ry < 0, cx = cy = 0, cxy < 0, and
cyx > 0. It was designed in the 1920’s by Vito Volterra to describe the increased per-
centage of sharks and rays among all caught fish in the Adriatic Sea after fishing had
been interrupted for multiple years due to WW1. The equations were independently
derived and analyzed by Alfred Lotka (from Lemberg, Austria-Hungary).

4.2 Equilibria
All points x∗ with ẋ∗ = 0 are called equilibrium points (or fixed points, rest points,
stationary states) of the dynamics. For example, the Lotka-Volterra model has a trivial
equilibrium at x = 0. In general, we consider the so-called isoclines. The xi -isocline is the
set of all points where ẋi = 0, i.e. where the vector field is vertical to the xi axis. For the
Lotka-Volterra model, we obtain:
• x-isocline for fx (x, y) = 0:
rx + cx x
x = 0 and y = − (115)
cxy

• y-isocline for fy (x, y) = 0:

ry + cyx x
y = 0 and y = − (116)
cy

The equilibria are given by the intersections of the x- and y-isoclines (in general: of all
xi -isoclines). We obtain:
(x∗1 , y1∗ ) = (0, 0) ; (x∗2 , y2∗ ) = (0, −ry /cy ) ; (x∗3 , y3∗ ) = (−rx /cx , 0) (117)
c r −c r c r −c r 
xy y y x yx x x y
(x∗4 , y4∗ ) = , . (118)
cx cy − cxy cyx cx cy − cxy cyx
Only equilibrium points with x∗i , yi∗ ≥ 0 are attainable and make biological sense. With
this information about equilibria and isoclines, we already obtain a qualitative graphical
representation of the dynamics in the phase plane. For concreteness, we consider the
predator-prey case with rx > 0, ry < 0, cx , cy < 0, cxy < 0, and cyx > 0. We see that the
(second part of the) x-isocline is monotonically decreasing, starting from (0, |rx /cxy |) on
the y-axis. The second part of the y-isocline is monotonically increasing, starting from the
unbiological equilibrium (0, −|ry /cy |) on the negative y-axis. It intersects with the x-axis
at x̂ = −ry /cyx > 0. There are two fundamentally different dynamical regimes.

41
 • For x̂ > x∗3 = −rx /cx > 0, the x and y isoclines intersect in the negative-y half plane
and the forth equilibrium point (x∗4 , y4∗ ) is unbiological (and cannot be reached from
positive starting values). Adding schematic velocity arrows to the various regions
of the phase plane, we quickly see that (x∗3 , y3∗ ) = (−rx /cx , 0) is the only stable
equilibrium of the dynamics: predators cannot be maintained and die out.

• For x̂ < x∗3 , the x and y isoclines intersect for positive y and the forth equilibrium
point (x∗4 , y4∗ ) is biologically meaningful. Adding velocity arrows to the phase plane,
we see that (x(t), y(t)) trajectories will oscillate around this point, but it remains
unclear whether the equilibrium is stable or unstable. A more detailed analysis is
needed.

4.3 Stability
Extending the one-dimensional case, we define:

• A fixed point x∗ of an ODE ẋ = f (x) is (Lyapunov) stable iff, for any  > 0 there is
a δ > 0, such that
x(0) ∈ B(x∗ , δ) ⇒ x(t) ∈ B(x∗ , )
for all t ≥ 0.

• A fixed point is asymptotically stable iff it is stable and

lim |x(t) − x∗ | = 0
t→∞

for all x(0) ∈ B(x∗ , ) and some  > 0.

• The set of points {x} with x(t) → x∗ is called the domain (or basin) of attraction
of x∗ . The fixed point is said to be globally attracting iff its domain of attraction is
the entire state space.

• As generalization of a fixed point, we define the ω-limit of any point x as the set of
all accumulation points of x(t) for t → ∞ with x(0) = x. Obviously, the ω-limit
of any point in the domain of attraction of a fixed point x∗ is just x∗ . In general,
ω-limits can also be empty, but this is not possible if the dynamics is confined to a
compact set.

• Analogous to ω-limits, we define α-limits as accumulation points for t → −∞.

As in the one-dimensional case, we can relate stability to local properties of the dynamics
in the neighborhood of a fixed point. To this end, we consider the linearized dynamics for
some point x ∈ B(x∗ , ) according to a Taylor expansion
 
∗ ∂fi ∂fi
· x − x∗ + O 2

 
fi (x) = fi (x ) + ,..., (119)
∂x1 x=x∗ ∂xn x=x∗

42
and
f (x) = f (x∗ ) + J · x − x∗ + O 2

 
(120)
with the Jacobian matrix  ∂f1 ∂f1

∂x1
··· ∂xn
J =  ... ..  (121)

. 
∂fn ∂fn
∂x1
··· ∂xn
x=x∗

For a fixed point, we have f (x∗ ) = 0 and defining y := x − x∗ we obtain the local linear
map
ẏ = Jy . (122)

Linear ODE’s Linear differential equations are not capable of describing meaningful
interactions among species. However, as we have seen, any ODE system can be approxi-
mated by a linear system locally around a fixed point. It is therefore worthwhile to consider
the linear dynamics. Let
ẋ = Jx . (123)
with real-valued n × n matrix J. Then the solution can be written as
 
x(t) = exp Jt x(0) (124)

with the matrix exponential

2
  t 2t
exp Jt = I + J + J + ··· . (125)
1! 2!
Eigenvalues of J can either be real or occur in complex conjugate pairs. Note that if λ
is an eigenvalue of J with corresponding eigenvector v, then v is also an eigenvector of
exp[Jt] with eigenvalue exp[λt]. Representing x(0) as linear combination of (generalized)
eigenvectors of J, we see that the components xi (t) of the solution x(t) can be written as
linear combinations of the following functions:
(i) exp[λt] whenever λ is a real eigenvalue of J.
(ii) exp[at] cos[bt] and exp[at] sin[bt] whenever a ± ib are eigenvalues of J.
(iii) tj exp[λt], or tj exp[at] cos[bt] and tj exp[at] sin[bt], with 0 ≤ j < m, whenever the
eigenvalue λ or the eigenvalue a ± ib occurs with (algebraic) multiplicity m.
Obviously, x∗ = 0 is a fixed point of the linear ODE (the only one if det A 6= 0). We can
note that positive real parts in the eigenvalues (λ > 0 or a > 0) of J lead to components
in x(t) that diverge away from 0, while negative real parts lead to components that vanish
in the limit t → ∞. Any imaginary part b 6= 0 leads to oscillations, which are damped if
and only if a < 0. We can define the following types of fixed points:
• A fixed point is called hyperbolic if none of the eigenvalues has zero real part. A
hyperbolic fixed point is...

43
 1. A sink if the real parts of all eigenvalues are negative. In this case,

lim x(t) = 0 (126)

t→∞

for any starting vector x(0). Hence, 0 is globally stable (the ω-limit of every
point). A sink can be either a stable node (all eigenvalues real) or a stable
spiral/sink focus (some eigenvalues complex).
2. A source if the real parts of all eigenvalues are positive. In this case any orbit
x(t) grows exponentially and 0 is the α-limit of every point. A source can be
either an unstable node (all eigenvalues real) or an unstable spiral/source-focus
(some eigenvalues complex).
3. A saddle point if some eigenvalues have positive real parts and some have neg-
ative real parts, but none are on the imaginary axis. The points with ω-limit 0
and the points with α-limit 0 each form a linear submanifold of Rn , called the
stable and unstable manifold, respectively, which together span Rn . Sometimes,
saddle points are also further characterized, for example by attaching the word
“focus” when there are complex eigenvalues. However, this is only possible in
more than two dimensions.

• Eigenvalues on the imaginary axis correspond to degenerate solutions. In particular,

an eigenvalue λ = 0 leads to a manifold of fixed points and a purely imaginary
eigenvalue to a manifold of periodic orbits.

• Hyperbolic eigenvalues are structurally stable in the sense that small changes of the
parameters lead to (only) small changes of the equilibrium structure. In contrast,
non-hyperbolic equilibria are structurally unstable with respect to any perturbation
that introduces a small component on the real axis.

Two-dimensional systems. A classification of fixed points depends on the roots of the

characteristic polynomial. For a real-valued two-dimensional matrix
 
a b
A= (127)
c d

this can be written in terms of the trace and the determinant as

λ2 − λtr[A] + det[A] (128)

and hence p
tr[A]2 − 4 det[A]
tr[A] ±
λ1,2 = . (129)
2
According to the so-called Routh-Hurwitz criterion, we have

Re(λ1,2 ) < 0 ⇔ tr[A] < 0 and det[A] > 0 (130)

44
 A

Figure 5: Classification of the asymptotic equilibrium behavior of 2-dim linear ODE

systems. For each part of the trace-determinant space, the figure displays the lo-
cation of eigenvalues in the complex plane (real part on the x-axis and imag-
inary part on the y-axis) and a representative phase portrait. Modified from
http://www.augustana.ualberta.ca/ hackw/mat332/exhibit/fxdpt2d.html.

(proof immediate). Further, both roots will be real if and only if tr[A]2 ≥ 4 det[A]. We
obtain a full classification of possible equilibria of the linear system (Fig. 5).
So far, we have derived a classification framework for equilibria in linear systems of
ODEs. The Hartman-Grobman theorem guarantees that this classification can also be
used to characterize the fixed points of nonlinear ODE system. In particular,

• A fixed point is a sink if all eigenvalues of the corresponding Jacobian have negative
real parts. In this case, the fixed point is asymptotically stable.

• Any eigenvalue of the Jacobian with strictly positive real part leads to an unstable
fixed point.

• For fixed points, where some (or all) eigenvalues of the Jacobian have real part zero,
and all others negative real parts, the Hartman-Grobman theorem does not apply
and stability depends on higher order terms.

With this, we can now apply this classification framework to equilibria in the Lotka-Volterra
model.

45
 rx
−
cxy

rx ry ry rx
− − − −
cx cyx cyx cx

Figure 6: The three isocline configurations for the Lotka-Volterra predator-prey model.
(a) Isoclines do not intersect. The point (−ry /cyx ) equilibrium is globally stable. (b)
Isoclines intersect but prey grows exponentially in the absence of the predator. The interior
equilibrium is neither a sink nor a source but a center. The whole interior of the phase
plane is filled with periodic orbits (hatched lines). (c) Isoclines intersect. The interior
equilibrium (x∗ , y ∗ ) is globally stable.

4.4 Stability for the Lotka-Volterra model

We consider the general Lotka-Volterra model. We have the Jacobian
!
∂fx ∂fx
∂x
= rx + 2cx x + cxy y ∂y
= cxy x
J= ∂fy ∂fy (131)
∂x
= cyx y ∂y
= ry + 2cy y + cyx x

which can be analyzed for arbitrary parameters. We will discuss the three models for the
basic biological interactions in turn.

Predator-prey We start with the predator-prey model, defined by cxy < 0 and cyx > 0.
We assume logistic growth of the prey with rx > 0 and cx = −rx /Kx < 0. The predator
cannot grow on its own. Thus ry < 0. For computational simplicity, we chose cy = 0. For
the trivial fixed point (0, 0), we obtain
 
∗ rx 0
J(x1 ) = . (132)
0 ry
It is thus a saddle point and unstable. The second fixed point is unbiological. For the third
one,  
∗ −rx −rx cxy /cx
J(x3 ) = , (133)
0 ry − rx cyx /cx
with eigenvalues −rx < 0 and ry − rx cyx /cx . This fixed point is stable if and only if
rx ry
Kx = < . (134)
cx cyx

46
 x∗ y∗

ry K ry K
− −
cy cy
Figure 7: Bifurcation diagram for the prey x∗ and predator y ∗ equilibrium density with
bifurcation parameter K = −rx /cx . Solid lines correspond to asymptotically stable equi-
libria, hatched lines to unstable equilibria. At Kx = −ry /cy a transcritical bifurcation
occurs.

This means that the prey carrying capacity is too small to satisfy the needs of the predator,
which will then die out. For the fourth equilibrium

∗
 ry cyx rx − cx ry 
x4 = − ,− (135)
cyx cxy cyx

we have  
−cx ry /cyx −cxy ry /cyx
J(x∗4 ) = . (136)
−(cyx rx − cx ry )/cxy 0
We have tr[J] = −cx ry /cyx < 0 and

ry (cyx rx − cx ry ) rx ry
det[J] = − >0 ⇔ Kx = > . (137)
cyx cx cyx

Using the Routh-Hurwitz criterion, we see it is stable if and only if the fixed point x∗3 is
unstable. This is the case if and only if there is a positive predator frequency y4∗ > 0. The
fixed point can be either a node (for 4 det < tr2 ) or a spiral. We get spirals for

|cx | q 
<2 1 + rx /|ry | − 1 , (138)
cyx

this is, if intraspecific competition for the prey is weak (little self-regulation) relative to
the predator’s benefit from the prey. In the limit cx → 0, we get center-type behavior (Fig.
6b). This is the original model by Volterra. Since it is “on the boundary” between stable
and unstable spirals, it is structurally unstable. Note that the prey equilibrium density
depends only on growth parameters of the predator. [This holds strictly only for cy = 0.] In
particular, only the predator profits from increased growth (large rx ) and carrying capacity
(low cx ) of the prey. This served as the explanation of the Adriatic fish-and-shark data
after WW1.

47
Figure 8: Four possible isocline configurations in the Lotka-Volterra two-species competi-
tion model. Open and filled circles indicate unstable and asymptotically stable equilibria,
respectively. (a) Species x dominates species y. (b) Species y dominates species x. (c)
Each species dominates when frequent. The interior equilibrium is a saddle point. (d) Each
species increases in frequency when rare. The interior equilibrium point is asymptotically
stabe.

Competition With the Lotka-Volterra competition model, we have the parameters rx , ry >
0, cx , cy < 0, and cxy , cyx < 0. It is easy the determine the whole dynamics purely graphi-
cally.

• First, we have simple logistic growth for both species in the absence of the other one.
This determines the dynamics on the axes with the carrying capacities Kx = |rx /cx |
and Ky = |ry /cy | as stable equilibria.

• Second, presence of some individuals of the other species contributes to a reduction

in the population size that can be maintained. We can conclude that both the x-
and y-isoclines have a negative slope. For the Lotka-Volterra model, both isoclines
are linear.

• The x-isocline intersects the x axis at Kx . The intersection point |rx /cxy | with the y
axis can either be larger or smaller than Ky . Similarly, the y-isocline extends from Ky
to some intersection point |ry /cyx | with the x axis that may be larger or smaller than

48
 Kx . The direction of the flow in the various domains follows easily from extrapolation
of the flow on the axes.

We obtain four different possibilities (see Fig. 8), corresponding to three different patterns:

1. Dominance If both isoclines do not cross, one species (with the higher isocline)
always dominates the other. (For example: Kx = |rx /cx | > |ry /cyx | and |rx /cxy | >
|ry /cy | = Ky for a dominating species x.)

2. Mutual exclusion If both isoclines intersect below the carrying capacity of the
other species, both species dominate when frequent. Biologically, this is the case
if competition against the other species is stronger than competition against con-
specifics (weighted by intrinsic growth): |cyx /ry | > |cx /rx | and |cxy /rx | > |cy /ry |. As
a result, the intersection of both isoclines corresponds to an unstable saddle point.

3. Coexistence If both isoclines intersect above the carrying capacity of the other
species, both species dominate when rare. We then obtain a stable equilibrium where
both species coexist. Biologically, this means that competition among conspecifics is
stronger than competition with the other species, |cyx /ry | < |cx /rx | and |cxy /rx | <
|cy /ry |. This is possible if both species also have their own “private” resource and
only compete for some (shared) resources. We then say that each species has its
own ecological niche. To this end, both species must not be “too similar” in an
ecological sense. This is known as the principle of limiting similarity as condition for
coexistence.

Mutualism See Exercise 5.

4.5 Global stability

So far, we have analyzed only the local stability of equilibria. For a full understanding of
the long-term dynamics of an ODE system, we need to know the ω-limit for each point in
the interior of the state space. In particular, a locally stable equilibrium point is globally
stable iff it is the ω-limit of all points in (the interior of) the state space. From our
graphical analysis of the LV competition scenario, it seems to be clear that asymptotic
stability implies global stability whenever only a single locally stable equilibrium exists.
For mutual exclusion, where two stable equilibria exist, the attraction basins of these
equilibria divide the state space (a more formal proof in provided in the next chaper).
In general, however, the global dynamics cannot easily be inferred from local stability
properties alone, even if there is only a single stable equilibrium. For example, given
our current knowledge, we cannot exclude periodic orbits (running around the stable x∗4
equilibrium) for the coexistence case of the LV predator-prey model. We therefore need
methods that go beyond a local analysis. In the following, we use the short-hand notation
that x(t), y(t) etc. are solutions of a system of differential equations ẋ = f (x) with the
same f , but with different initial conditions x ≡ x(0) and y ≡ y(0).

49
Elements of ω-limits In addition to fixed points , x∗ with ẋ∗ = 0, ω-limits can contain
various further points.
• Periodic points x such that x(kT ) = x for some T > 0 and k ∈ N, but x(t) 6= x
for 0 < t < T . Then all other points on the orbit x(t) are also periodic points and
together they form a periodic orbit γ. A periodic orbit γ is called
– limit cycle if ∃ x ∈
/ γ such that ω(x) = γ and
– attractor if ω(x) = γ ∀x in some neighborhood of γ.
• It is also possible for ω-limits to contain points that are neither fixed points nor
periodic points (e.g., with chaotic behavior).
For the Lotka-Volterra model and most other biological models of interest, we do not
know the full solution x(t) of the nonlinear system of ODEs. Therefore, we cannot simply
obtain ω-limits as accumulation points of the solution. Fortunately, there are two results
that allow us to determine (or at least to narrow down) ω-limits even without knowing the
full solution. Importantly, these results also apply in cases where the Jacobian matrix has
eigenvalues with zero real part such that the Hartman-Grobman theorem does not apply.

Theorem: Every ω-limit is invariant: y ∈ ω(x) ⇒ y(t) ∈ ω(x).

Proof Let y be in ω(x). Then x(tk ) → y for some sequence tk . Since solutions are
continuous functions of their initial condition, we have x(tk + t) → y(t) for arbitrary t’s
and for k → ∞. Therefore, y(t) belongs to ω(x).

The second result is based on the concept of Lyapunov functions. The idea is as follows.
We construct a function V of x that has a maximum at x∗4 . Thus, the “peak of a mountain”
in the plane spanned by x is located at x∗4 . The aim is to construct this function V such
that a solution of the dynamical system under consideration can only travel in an uphill
direction on this landscape. Alternatively, we can construct V in such a way that we travel
always downhill and the focal equilibrium is a minimum.

Theorem: Lyapunov Let ẋ = f (x) be a time-independent ODE defined on some

subset G of Rn . Let V : G → R be continuously differentiable. If the derivative V̇ of the
map t → V (x(t)) satisfies the inequality V̇ ≥ 0 (or V̇ ≤ 0) for all solutions x(t), then
ω(x) ∩ G is contained in the set {x ∈ G : V̇ (x) = 0} (and so is α(x) ∩ G). In particular,
if this set only contains a single point, this point is a globally stable equilibrium.

Proof Let y ∈ ω(x) ∩ G. Then there exists a sequence tk → +∞ with x(tk ) → y.

Since V̇ ≥ 0 along the orbit of x, we have V̇ (y) ≥ 0 by continuity. Also by continuity,
V (x(tk )) converges to V (y), and hence

V (x(t)) ≤ V (y) (139)

50
for every t ∈ R. To prove the statement by contradiction, assume V̇ (y) > 0. Since the
value of V can never decrease along the orbit, this implies
V (y(t)) > V (y) (140)
for t > 0. From x(tk ) → y it follows by continuity that x(tk + t) → y(t) and hence
V (x(tk + t)) → V (y(t)). (141)
With Eq. (140) it follows that
V (x(tk + t)) > V (y) (142)
for k sufficiently large. This is in contradiction to Eq. (139).

Lyapunov function for the Lotka-Volterra predator-prey model Lyapunov’s the-

orem does not tell us how to find a Lyapunov function V and unfortunately no general
recipe exists. In mechanical or electrical systems, energy is often a Lyapunov function.
We will now attempt to construct a Lyapunov function to determine whether or not the
interior equilibrium of the LV model given by Eq. (114) is globally stable. For this, we
look for a function V that has a maximum at
 
∗ ∗ ∗ cxy ry − cy rx cyx rx − cx ry
x4 = (x4 , y4 ) = , .
cx cy − cxy cyx cx cy − cxy cyx
We do some reverse engineering. A simple function would have the form
V (x, y) = aH(x) + bG(y) (143)
for some a, b > 0. To get a maximum at x∗4 , we need
dV dH dV dG
0= =a , 0= =b .
dx x=x∗ dx x=x∗ dy x=x∗ dy y=y ∗

Simple derivatives fulfilling these requirements are

dH x∗ dG y∗
= −1 and = −1
dx x dy y
resulting in
V (x, y) = a(x∗ ln x − x) + b(y ∗ ln y − y), (144)
which is defined on R2+ . Checking the second derivates shows that this choice produces in-
deed a maximum. Does this V do the job? The derivative of the function t → V (x(t), y(t))
is
∂V ∂V
V̇ (x, y) = ẋ + ẏ
∂x ∂y
 ∗   ∗ 
x y
=a − 1 x(rx + cx x + cxy y) + b − 1 y(ry + cy y + cyx x) (145)
x y

51
Recall that the nontrivial isoclines of Eq. (114) are given by

rx = −cx x − cxy y (146a)

ry = −cy y − cyx x (146b)

and that x∗4 is the solution of Eq. (146). Thus, we can replace rx and ry with −cx x∗ − cxy y ∗
and −cy y ∗ − cyx x∗ , respectively. This yields

V̇ (x, y) =a(x∗ − x)(−cx x∗ − cxy y ∗ + cx x + cxy y) + b(y ∗ − y)(−cy y ∗ − cyx x∗ + cy y + cyx x)

= − acx (x − x∗ )2 − bcy (y − y ∗ )2 + (acxy + bcyx )(x − x∗ )(y − y ∗ ). (147)

The sign of (147) is not obvious. However, if we choose a = cyx and b = −cxy , then the
last term in (147) becomes zero and we have

V̇ (x, y) = −cx cyx (x − x∗ )2 + cy cxy (y − y ∗ )2 ≥ 0. (148)

We may therefore apply Lyapunov’s theorem. The ω-limit of every orbit in R2+ is contained
in the set K = {(x, y) : V̇ (x, y) = 0}. We can distinguish three cases:

1. For cx , cy < 0, K consists only of the point x∗4 . All trajectories in R2+ converge to x∗4 .
This point is globally stable.

2. For cy = 0, but cx < 0 (no self limitation in the predator, the case studied above),
K = {(x, y) ∈ R2+ : x = x∗4 }. However, the ω-limit must be an invariant subset of K,
and therefore reduces to x∗4 . This point is again globally stable. The same holds for
the case cy < 0 and cx = 0.

3. If both cy = 0 and cx = 0 (Lotka’s original model in which both species are not
self-regulated), we have K = R2+ and thus

V (x(t), y(t)) = const.

along any orbit x(t). In this case, V is called a constant of motion. V has a unique
maximum at x∗4 and approaches −∞ along every half-line starting at x∗4 . The con-
stant level sets {(x, y) ∈ R2>0 : V (x, y) = const.} are closed curves around x∗4 . Hence,
the orbits are periodic and x∗4 is a center (Fig. 6 b). It is neither locally nor globally
stable.

In particular, we see that stable limit cycles are impossible in the LV predator-prey
system. We have previosly seen that this also holds for the competition model. In general,
one can show also for versions of the LV model that do not have a biological interpretation:

Theorem The two-dimensional Lotka-Volterra Eq. (114) has either no periodic orbits
or a continuum of periodic orbits. [For a proof see Hofbauer & Sigmund, 1998, p. 33].

52
5 The Dynamics of Two Interacting Populations: Be-
yond the Lotka-Volterra model
Many insights into the dynamics of interacting species rely on the Lotka-Volterra frame-
work. This raises the question how reliable and general these results can be. Indeed, the
model makes several rather restrictive assumptions, such as a linear dependence of per
capita birth rates on the densities (e.g. no Allee effect) and no density dependence of the
pair interaction parameters (cx , cy , cxy , cyx do not depend on x or y). This can have unbio-
logical consequences, e.g. for mutualism, where populations can grow without bounds for
some parameter choices. In more general, we should ask for any model whether changes
in the modeling assumptions can lead to qualitatively new effects.

5.1 A general two-species competition model

A general model of interacting populations (without immigration) is of the form

ẋ = x F (x, y) (149a)
ẏ = y G(x, y), (149b)

with continuously differentiable functions G(x, y) and F (x, y). For most natural systems
it will be impossible to determine the exact functional form of F and G. It is therefore
desirable to obtain results without specifying F and G but only constraining them by
certain boundary conditions. Here, we follow this approach for the case of competition
between two populations. We make the following assumptions:
1. Competing populations affect each others growth rates negatively. Therefore,
∂F (x, y) ∂G(x, y)
<0 and < 0. (150)
∂y ∂x

2. Competition occurs due to shared resources (e.g., food, nesting sites). Here, we
assume that the two competing types both require the same essential resource. Thus,
if one competitor is very common both species have a negative growth rate: there
exists a K > 0 so that F (x, y) < 0 and G(x, y) < 0 if x > K or y > K.
In the following we derive the qualitative dynamics of a competition model based on
assumptions (1) and (2). The following statements can be made about the isoclines.
• The non-trivial x-isocline is given by

F̄ = {(x, y) ∈ R2>0 : F (x, y) = 0}. (151)

Above F̄ we have ẋ < 0 and below F̄ we have ẋ > 0. Indeed, for x fixed, F (x, y)
decreases as a function of y (cf. Eq. 150). Therefore, a vertical line through a point
(x, y) intersects with F̄ at most once.

53
 α

F̄
β
Ḡ

Figure 9: Four examples for possible isocline configurations in the general two-species
competition model. Open and filled circles indicate unstable and asymptotically stable
equilibria, respectively. (a&b) Isoclines do not intersect. (b&c) Both species show an
Allee effect. (d) Example with a stable interior equilibrium point P . The gray area B
is positively invariant: all orbits that cross its ordinary boundary points go into B. The
isocline for the x-species is labelled F̄ . It passes through the equilibria that lie on the
x-axis. The isocline for the y-species is labelled Ḡ. It passes through the equilibria that
lie on the y-axis.

• The non-trivial y-isocline is given by

Ḡ = {(x, y) ∈ R2>0 : G(x, y) = 0}. (152)

To the right of Ḡ we have ẏ < 0 and to the left of Ḡ we have ẏ > 0. For y fixed,
G(x, y) decreases as a function of x (cf. Eq. 150). Therefore, a horizontal line through
a point (x, y) intersects with Ḡ at most once.

Both F̄ and Ḡ are continuous. Due to our second assumption above, they both lie within
the square
{(x, y) : 0 ≤ x ≤ K, 0 ≤ y ≤ K}.
Every orbit in R2>0 flows into this square.

• The situation is simple if F̄ and Ḡ do not intersect. Then either one or both species
go extinct (Fig. 9a,b).

54
 • Next, consider the case that F̄ and Ḡ intersect finitely many times (this is the generic
case). The points of intersection are equilibria of Eq. (149). The isoclines F̄ and Ḡ
divide R2>0 into regions B in which the signs of ẋ and ẏ do not change. The boundary
of these regions consists of equilibrium points, of segments of the coordinate axes,
and of points that belong to F̄ or Ḡ, but not to both. Let us call the latter ordinary
boundary points.

Lemma For each region B, either all orbits passing through its ordinary boundary
points enter B, or all orbits passing through its ordinary boundary points leave B.

Proof Four cases have to be distinguished: (i) ẋ > 0, ẏ > 0; (ii) ẋ > 0, ẏ < 0; (iii)
ẋ < 0, ẏ > 0; (iv) ẋ < 0, ẏ < 0. Consider (ii), a region B with ẋ > 0 and ẏ < 0 (gray
region in Fig. 9d) . This means that B lies below F̄ (this is where we have ẋ > 0) and to
the right of Ḡ (this is where we have ẏ < 0). Next we consider an ordinary boundary point
P . If P ∈ F̄ , then P is to the right of Ḡ: the orbit through P goes vertically downwards
into B. If P ∈ Ḡ, then P is below F̄ : the orbit through P goes horizontally to the right
into B. In this case, all orbits passing ordinary boundary points P of B go into B. The
argument is analogous for the other three cases. We find that regions of type (ii) and (iii)
can never be left again: they are positive invariant. Conversely, regions of type (i) and (iv)
can never be entered from outside (negative invariant).

For the asymptotic behavior, we can conclude the following:

• Because the signs of ẋ and ẏ are constant in the interior of a region B, the ω-limit of
any orbit x(t) can never be contained in the interior of a single region. In the positive
invariant regions all orbits converge toward an equilibrium. In regions of type (i) and
(iv), orbits either also converge to an equilibrium or they enter a positive invariant
region through an ordinary boundary point. In conclusion, in the general competition
model the dynamics always converges towards an equilibrium.

• Which equilibria are asymptotically stable? It is easy to see that equilibria at which
both F̄ and Ḡ have a negative slope with F̄ steeper than Ḡ are asymptotically stable
(Fig. 9d). An investigation of all other cases in which F̄ and Ḡ can intersect each
other shows that the only other asymptotically stable fixed point is (Kx , 0) when Kx
is to the right of Ḡ and (0, Ky ) when Ky is above F̄ .

• How do ω-limits depend on initial conditions? In Fig. 9(d) point Q is a saddle point.
Then exactly two trajectories, α and β, exist that approach Q. These are known as
stable manifolds of Q or separatrices of Q. The separatrix α divides the unbounded
region into a part with points x with ω(x) = (0, Ky ) and a part with points x with
ω(x) = P . Thus, two points x and x0 that are very close to each other but separated
by α have different ω-lime sets.

55
Summarizing, we see that the qualitative dynamics of the Lotka-Volterra competition
model matches the general competition model in that there are no limit cycles or chaotic
orbits. However, in the general case multiple stable equilibria in the interior and on the
boundary can exist. Biologically, whether species can coexist or one species goes extinct
can depend on the starting conditions.

5.2 Equilibria in 2-dimensional ODEs

In the last section, we have determined which kind of equilibrium behavior can result from
a general competition model. Before we try to obtain similar results for other biological
scenarios, we should first get a better idea about which types of equilibria are possible at
all for a 2-dimensional dynamical system. We thus assume a general ODE

ẋ = f (x)

on an open subset M ⊂ R2 and f ∈ C 1 (M, R2 ). Which kind of ω-limits can we get for such
a system? In one dimension, all ω-limits are fixed points. In two dimensions, the following
theorem shows that the situation is still relatively simple.

Poincaré-Bendixson Theorem Let ẋ = f (x) be a time-independent 2-dimensional

ODE as defined above with at most finitely many fixed points. Let x ∈ M and ω(x) be a
non-empty, connected, and compact ω-limit set. Then one of the following applies:

1. ω(x) is a fixed point.

2. ω(x) is a periodic (or closed) orbit.

3. ω(x) consist of finitely many fixed points and orbits connecting these fixed points
(with these fixed points as ω- and α-limits).

We note the following

• One can easily show that for any orbit that is bounded forward in time the corre-
sponding ω-limit is non-empty, connected, and compact. Since this should be the
case for any realistic population dynamical model, the Poincaré-Bendixson Theorem
will apply. For orbits that are bounded backward in time, the analogous result holds
for the α-limit.

• A direct consequence of the theorem is that chaotic attractors cannot occur in two
dimensions.

• The theorem can be used to demonstrate the existence of periodic orbits: if we can
identify a compact trapping region, which can only be entered, but never left by any
orbit, and if this region does not contain a fixed point, then it must contain at least
one periodic orbit.

56
 • Since a periodic orbit and its interior is a compact region that cannot be entered
or left (forward and backward invariant), a fixed point must exist in the interior of
each closed orbit. (Intuitively: An orbit starting from an internal point must have
an ω-limit or α-limit in the interior of the region. This limit is either a fixed point
or a closed orbit. If it is a closed orbit, iterate the argument.) This implies that
periodic orbits can only occur around fixed points. A compact trapping region to
demonstrate the existence of a periodic orbit must therefore always contain a hole
(with the fixed point in it).
• One can further show that in 2d dynamical systems all isolated closed orbits are in-
deed limit cycles. Vice versa, closed orbits that are not isolated (center-type behavior
like in the Lotka-Volterra predator-prey model) are not limit cycles.
• For two dynamical variables x1 and x2 with derivatives ẋ1 and ẋ2 given as polynomials
of x1 and x2 with finite degree n, the question how many (isolated) limit cycles can
exist is known as Hilbert’s 16th problem and no answer is known for n > 1 (where
no limit cycle can exist). For n = 2, systems with up to four limit cycles are known.
For general n, it is at least known that the maximal number is always finite.
• A formal proof of the Pointcaré-Bendixson Theorem can be found, for example, in
the book by Gerald Teschl Ordinary Differential Equations and Dynamical Systems,
available form his webpage. The proof relies on the Jordan Curve Theorem, which
states that any closed curve divides the plane into two disjoint sets: the interior and
the exterior. Any two points in the interior can be connected by a path which never
intersects the curve, but every path between interior and exterior has to cross the
curve. This seems a trivial observation which, however, is very difficult to prove.
• The theorem uses that solutions of ODEs in the plain are subject to strong geometric
constraints. Since the orbits can never cross (uniqueness of the solution) they can
easily be trapped in the interior of some region where they can no longer escape.
Since this constraint does not exist in higher dimensions, the equilibrium structure
of dynamical systems in 3 or more dimensions can be much more complex (including
chaotic behavior).
The first step in the analysis of a dynamical system is usually to characterize the fixed
points and their stability. In simple cases, we can still do this analytically. For a full
characterization of the stationary behavior, we need to account for limit cycles, which is
generally a more difficult task. To exclude cycling, we can try to find a Lyapunov function,
as explained above. An alternative criterion is provided by the following theorem.

Bendixson-Dulac Theorem Let ẋ = f (x) be an ODE defined on a simply connected

set (i.e., a set without holes) G ⊂ R2 . Let B be a positive scalar function defined on G. If
the divergence of the vector field Bf ,
∂Bf1 ∂Bf2
div[Bf ](x) = (x) + (x) (153)
∂x1 ∂x2

57
has the same sign (strictly positive or strictly negative) on G almost everywhere, then
ẋ = f (x) admits not periodic orbits in G.
• A function B with the above property is known as Dulac function. For B = 1,
expression (153) equals the divergence of f and is given by the trace of the Jacobian
matrix. Indeed, divf (x) 6= 0 means that the flow defined by the ODE is area
expanding or contracting, i.e., averaging over the boundary, leaves or enters every
compact and simply connected set. This is not possible with a closed orbit, which
defines a region that can neither be entered nor left by the flow.
• The proof of the theorem follows from Green’s theorem, see Hofbauer & Sigmund for
details. Note that a positive Dulac function B(x) can always be chosen since the
transformed ODE
ẋ = B(x)f (x)
has the same orbits as the original ODE, with a change in the velocity defined by B.
In particular, the stationary behavior of both ODE systems is the same.

5.3 General predator-prey models and Hopf-bifurcations

In this section, we revisit predator-prey models. In the Lotka-Volterra predator-prey model
isolated (and stable) limit cycles do not exist. However, predator-prey cycles seem to
be common in nature. Famous examples: lynx and snowshoe hare in Canada, many
populations of voles and their predators. Is the Lotka-Volterra predator-prey model missing
something?

A (somewhat) general predator prey model: the Gause model

Georgy Gause was a Russian biologist who experimentally tested the Lotka-Volterra pre-
dator-prey model with protists and yeast. After he was not able to reproduce the predic-
tions of Lotka and Volterra, he proposed modifications to original Lotka-Volterra model.
(This is an interaction between theory and empiricism at its best!) As in the case of com-
petition, it is desirable to study a predator-prey model in which not all functional forms
are specified. For the Gause model, we study

ẋ = xg(x) − yf (x) (154a)

ẏ = y c(x) − d , (154b)

where
1. g(x) is the growth rate of the prey population in the absence of predation. Self-
regulation in the prey implies that there exists a Kx > 0 so that g(x) > 0 for
x < Kx , g(Kx ) = 0, and g(x) < 0 for x > Kx .
2. f (x) is the functional response of the predator, describing the effect of predation on
the prey. We require that f (0) = 0 and f (x) > 0 for x > 0.

58
 ∗
Kxx x∗ Kx x∗ Kx

Figure 10: The three generic isocline configurations in the Gause predator-prey model.
Open and filled circles indicate unstable and asymptotically stable equilibria, respectively.
(a) Isoclines do not intersect. (b) Prey isocline has a negative slope at the interior equi-
librium. The equilibrium is then asymptotically stable. (c) Prey isocline has a positive
slope at the interior equilibrium. The equilibrium is then unstable. For the special case
of the Rosenzweig-MacArthur model we prove that it is then surrounded by a stable limit
cycle (dotted line). The hatched line in (b) and (c) represents the unstable manifold of the
saddle point (Kx , 0).

3. c(x) describes consumption of prey and conversion into predator individuals. c(x)
may be independent of f (x). We require that c(0) = 0 and dc(x)/dx > 0 for x > 0.

4. d is the constant consumer mortality rate. There is thus no self-regulation in the

predator. Indeed, the basic assumption of the Gause model is that the predator
density remains low, such that its effect on the dynamics is only linear.

We can determine the isoclines for this model.

• Assume x∗ exists with c(x∗ ) = d. This simply means that predators can eat enough
to exist. The predator-isocline equals to the vertical line x = x∗ and the line y = 0
(absence of predators).

• The non-trivial part of the prey isocline {(x, y) ∈ R2>0 : ẋ = 0} is determined by the
equation
xg(x)
yiso = .
f (x)
We have yiso ≥ 0 (in the biological parameter range) for x ∈]0, Kx ]. This function
intersects the predator-isocline at (Kx , 0) (the prey equilibrium in the absence of
predators). Furthermore, it may have at most one intersection with the line x = x∗ .
Two scenarios are relevant:

1. x∗ > Kx : Isoclines do not intersect (Fig. 10). Then the predator goes extinct.

59
 2. x∗ < Kx : Isoclines have a unique point of intersection (x∗ , y ∗ ) ∈ R2≥0 (Fig.
10b&c). We focus on this case in the following.

The Jacobian matrix at (x∗ , y ∗ ) equals

H(x∗ ) −f (x∗ )
!
J= dc(x∗ )
y∗ 0
dx
where
dg(x) xg(x) df (x)
H(x) = g(x) + x − . (155)
dx f (x) dx
We apply the Routh-Hurwitz criterion (Eq. 130). We have det J > 0 and trace = H(x∗ ).
The point (x∗ , y ∗ ) is asymptotically stable if and only if H(x∗ ) < 0. Note that
d xg(x) H(x)
= . (156)
dx f (x) f (x)
Thus, H(x∗ ) < 0 if and only if
d xg(x)
< 0. (157)
dx f (x) x=x∗

• In conclusion, the interior equilibrium is a sink when the prey-isocline has a negative
slope at the equilibrium and a source when the prey-isocline has a positive slope (Fig.
10b&c).

• We see from the graph (and can easily derive) that the two other equilibria, (0, 0)
and (Kx , 0), are both are saddle points. Their stable manifolds are given by the x
and y axes.

• What can we conclude for the global behavior? Consider, in particular, a point on
the unstable manifold that departs from (Kx , 0) (i.e., a point for which (Kx , 0) is the
α-limit). The orbit starting from such a point is bounded. Indeed, for a point on
the manifold near (Kx , 0), the prey density x is initially declining and must reach
x∗ at some finite value ỹ ≥ y ∗ . At this point, the predator density cannot increase
any further and the orbit enters a region that is bounded by the axes, the unstable
manifold from (Kx , 0) to (x∗ , ỹ) and the line y = ỹ, that cannot be left anymore.
From the Pointcaré-Bendixson theorem, we conclude that there are two possibilities:

1. If ω(x) contains an equilibrium point it must be (x∗ , y ∗ ) (the other two equi-
librium points are out of question). In fact, then (x∗ , y ∗ ) is even globally stable
because all orbits have to circle around (x∗ , y ∗ ), but cannot intersect the unsta-
ble manifold of (Kx , 0) (Fig. 10b).
2. If ω(x) contains no equilibrium point, then ω(x) is a periodic orbit γ. This
orbit surrounds (x∗ , y ∗ ) and is a limit cycle: since other orbit cannot cross the
unstable manifold, every point in the exterior converges to γ (Fig. 10c).

60
 • One might think that the first scenario applies when (x∗ , y ∗ ) is a sink and that the
second scenario applies when (x∗ , y ∗ ) is a source. However, this need not to be the
case. For example, the limit cycle can be an attractor from the outside but not from
the inside (half-stable) or more than one limit cycle can exist. For the general model
the analysis cannot be completed. Below, we therefore specify the functions g(x),
f (x), and c(x).

The Rosenzweig-MacArthur model

The easiest choice for competition among prey is the logistic model with linear density
dependence in g(x). A natural choice for the interaction between predator and prey is
type two functional response, which, in contrast to the Lotka-Volterra model (with type
one functional response), assumes a saturation effect for large prey densities x. We then
arrive at the following model first studied by Rosenzweig and MacArthur.
 
x xycxy
ẋ = f1 (x, y) = xrx 1 − − (158a)
Kx a+x
xycyx
ẏ = f2 (x, y) = − dy, (158b)
a+x
where all parameters are positive.
• From the general model we know that the non-trivial prey isocline is determined by
the equation yiso = rx (1 − x/Kx )(a + x)/cxy .

• The non-trivial predator isocline is given by the vertical line x∗ = ad/(cyx − d).
If Kx ≤ ad/(cyx − d), the predator goes extinct and (Kx , 0) is globally stable. From now on
we focus on the case where an internal equilibrium (x∗ , y ∗ ) exists. From Eq. (157) already
know that it is a sink if Kx < a + 2x∗ and a source if Kx > a + 2x∗ . A more detailed
answer concerning global stability is provided by the following proposition, which makes
use of the Bendixson-Dulac theorem.

Proposition The internal fixed point (x∗ , y ∗ ) of Eq. (158) is globally stable if and only
if K ≤ a + 2x∗ .

Proof We consider the Dulac function

a + x γ−1
B(x, y) = y
x
with γ chosen appropriately. We derive
y γ−1
   
∂Bf1 ∂Bf2 a + 2x ∗
div[Bf ] = + = rx x 1 − − γ(cyx − d)(x − x) . (159)
∂x ∂y x Kx
We argue as follows:

61
 • The term rx x(1 − (a + 2x)/Kx defines a parabola in x that opens downwards and
intersects the x-axis at x = 0 and x = (Kx − a)/2. The term γ(cyx − d)(x∗ − x)
describes a line with slope −γ(cyx − d) that intersects the x-axis at x = x∗ .

• If Kx < a + 2x∗ then x∗ > (Kx − a)/2 and the line intersects the x-axis outside the
parabola. It is therefore possible to choose an γ so that the parabola lies below the
line (Fig. 11).

• Hence, the parabola and the line never cross and div[Bf ] always has the same sign
(< 0). Due to the Bendixson-Dulac theorem, this excludes cycles and with Pointcaré-
Bendixson we conclude that the internal fixed point must be globally attracting in
the interior of the parameter range.

• For Kx = a + 2x∗ , we still find an γ for which the line is tangent to the parabola and
thus div[Bf ] < 0 almost everywhere, but for the point x = x∗ where it equals zero.
Hence, periodic orbits can still be excluded and the interior fixed point is the global
attractor.

• If, Kx > a + 2x∗ then we have seen that the internal fixed point is unstable. With
the Poincaré-Bendixson theorem and our results for the general Gause model it is
clear that in this case a limit cycle around (x∗ , y ∗ ) must exist.

K−a K−a x∗ x
2 2

Figure 11: Sketch illustrating the proof that an γ can be chosen so that Expression (159)
is negative. The parabola is given by rx x(1 − (a + 2x)/Kx ) and the straight line by
γ(cyx − d)(x∗ − x) for an appropriately chosen value of γ. The right parabola applies for
K > a and the left one for K < a.

Let us summarize the stationary behavior. The Jacobian matrix J has a pair of complex
conjugate eigenvalues λ1,2 = α ± βi with α, β ∈ R.
• For Kx < a + 2x∗ we have α < 0. (x∗ , y ∗ ) is a sink and is globally stable (Fig. 10b).

• For Kx = a + 2x∗ we have α = 0. The linearized dynamics has a center at (x∗ , y ∗ ).

For the nonlinear dynamics, the point (x∗ , y ∗ ) is asymptotically stable.

62
 • For Kx > a + 2x∗ we have α > 0. (x∗ , y ∗ ) is a source that is surrounded by a limit
cycle (Fig. 10c). [In fact, one can prove that in this model only a single limit cycle
can exist but such a proof is beyond the scope of this lecture.]

x∗ y∗

ad
a + 2 cyxad−d Kx ad
a + 2 cyxad−d Kx
cyx −d cyx −d

Figure 12: Bifurcation diagram for the prey (x∗ ) and predator (y ∗ ) equilibrium density
with bifurcation parameter Kx . Solid lines correspond to asymptotically stable equilibria,
hatched lines to unstable equilibria. At Kx = ad/(cyx −d) a transcritical bifurcation occurs
and at Kx = a − 2ad/(cyx − d) as supercritical Hopf bifurcation occurs.

Some remarks on the equilibrium behavior of the Rosenzweig-MacArthur model:

1. It is clear that at Kx = a + 2x∗ a bifurcation point of the dynamics has to exist. This
type of bifurcation is known as Hopf bifurcation and is shown in Fig. 12.
2. Increasing the prey carrying capacity beyond the point ad/(cyx − d) (the minimal
density for predators to exist) benefits only the predator, but not the prey. This
“extreme” result is due to the vertical predator isocline at x = x∗ . However, very
steep isoclines readily occur in natural systems if the predator is primarily limited
by the availability of prey rather than by direct negative density dependence.
3. Increasing the prey carrying capacity even further (beyond the critical point of the
Hopf bifurcation) increases the chance for the prey to go extinct due to stochasticity
(paradox of enrichment).
4. Form an empirical point of view, it is still still not clear whether these models capture
the essence of predator-prey cycles in nature. In particular, it is an active research
area to determine whether predator-prey cycles are endogenously driven, i.e., by the
interactions of predators and their prey, or by exogenous factors.

Hopf bifurcations
The bifurcation that we have seen in the Rosenzweig-MacArthur model is an example of
a general type of bifurcations, where stable or unstable fixed points can split a fixed point
plus a (stable or unstable) limit cycle. For ODE systems in arbitrary (finite) dimensions,
they are characterized by the following theorem.

63
Hopf ’s theorem Let G be an open subset of Rn and

ẋ = f µ (x)

a family of n-dimensional ODEs on G depending on some parameter µ ∈ (−, ) and

f µ analytic (i.e. given by a power series). Let Pµ be an equilibrium point at which all
eigenvalues of the Jacobian matrix J have negative real parts but for one pair of complex
conjugate eigenvalues

λ1 = α(µ) + β(µ)i
λ2 = α(µ) − β(µ)i

with α(µ), β(µ) ∈ R and sign[α] = sign[µ] and β(0) 6= 0. Thus, Pµ is a sink for µ < 0
(and hence asymptotically stable) and a source for µ > 0 (and hence unstable). We also
require that ∂α/∂µ > 0, i.e. the eigenvalues cross the imaginary axis with non-zero speed.
We then have:

supercritical Hopf bifurcation If P0 is asymptotically stable, then, for sufficiently small

positive values of µ, the unstable equilibrium point Pµ is surrounded by a periodic
attractor (Fig. 13a).

subcritical Hopf bifurcation If P0 is unstable, then, for sufficiently small negative val-
ues of µ, the asymptotically stable equilibrium point Pµ is surrounded by a periodic
repellor (Fig. 13b).

0 µ 0 µ

Figure 13: (a) supercritical Hopf bifurcation, (b) subcritical Hopf bifurcation. Solid lines
correspond to asymptotically stable equilibria, hatched lines to unstable equilibria.

• At a supercritical Hopf bifurcation, a stable fixed point becomes locally unstable but
the flow sufficiently far away does not feel this local change and is still contracting.
As a result, a periodic orbit becomes the new attractor. (Analogous for subcritical.)

64
 • The direction of the “far flow” is indicated by the stability or instability of the
equilibrium point P0 for µ = 0. With vanishing real parts of λ1,2 , this equilibrium is
non-hyperbolic (neither a sink, a source, nor a saddle). Its stability therefore cannot
be deduced from linearization but depends on higher-order terms. These terms are
weaker than the linear ones in the direct vicinity of the fixed point, but stronger
at some distance. A small perturbation by moving µ away from zero will introduce
small, but non-zero linear terms, which dominate in a neighborhood, but not at some
greater distance.

• In practice, the dependence on higher order terms makes it difficult to determine

the stability of P0 . Often only a Lyapunov function or a Dulac function can provide
definite answers.

• The diameter of the limit cycle generated at the bifurcation points increases continu-
ously from zero and grows initially proportional to the square root of the bifurcation
√
parameter, µ.

6 Dynamics of more than two populations

In dynamical systems of dimension three or higher the Jordan curve theorem does not apply
and ω-limit sets of ODEs need not to be equilibrium points or periodic orbits anymore.
Instead, also strange attractors can exist and the dynamics can show chaos.

6.1 The General Lotka-Volterra Equations

The general Lotka-Volterra equation for n populations reads
n
!
X
ẋi = xi ri + aij xj , (160)
j=1

where xi and ri denote the density and the intrinsic growth rate of the ith species, respec-
tively. The aij describe the effect of the jth species on the growth rate of the ith species.
In matrix notation this becomes

ẋ = r ◦ x + x · Ax, (161)

where ◦ denotes the element-wise multiplication of two vectors (Hadamard product). The
matrix A is known as interaction matrix. While the two-dimensional Lotka-Volterra equa-
tions are completely classified, this has not been achieved for the general system due to
the potential for more complicated attractors. Indeed, chaotic motion is possible already
in the system with three species. Nevertheless, we can state some general results and also
obtain explicit solutions for some special cases of biological interest.

65
Equilibria
Note first that all faces of the state space xi = 0, which correspond to states where species
i is absent, are invariant under the flow (no individual can immigrate). Consequently, also
the flow in the interior of the state space (i.e., for xi > 0, ∀i) is invariant. The density xi (t)
may approach zero, however, which means extinction. Internal fixed points are solutions
of the linear equation system
n
X
y i = ri + aij xj = 0 i = 1, . . . , n (162)
j=1

with all components xj = x∗j > 0. Since the restriction of the dynamics to a boundary face
xi = 0 again leads to a Lotka-Volterra system (with one fewer dimension), fixed points on
the boundaries can be found in an analogous way.

• In general (i.e., if the linear system is not degenerate), there will be at most a single
internal fixed point for the Lotka-Volterra system. In the degenerate case, there is a
continuum of fixed points.

• A useful theorem states that if there is no internal fixed point, then there is no α-
or ω-limit in the interior of the state space at all. I.e., all orbits either flow to the
boundary (extinction of one species) or to infinity. For a proof, note that the convex
set K defined by the linear system (162) with xj > 0 is disjoint from 0 if there is
no internal fixed point. We then can can define a hyperplane H through 0 that is
disjoint from K and a vector c = (c1 , . . . , cn ) 6= 0 that is orthogonal to H such that
c · y > 0 for all y ∈ K. Then
X
V (x) = ci log xi
i
P P
is a Lyapunov function (indeed, V̇ = i ci ẋi /xi = i ci yi > 0). Since V (x) has no
internal maximum this proves the theorem.

• If there is a unique internal fixed point x∗ , one can further prove that all orbits
that neither converge to the boundary nor to infinity must have x∗ as their long
time average (see Hofbauer & Sigmund, chapter 5 for a proof). Hence, if closed or
chaotic attractors exist for an n-dimensional Lotka-Volterra model, they must always
surround a fixed point.

6.2 Lotka-Volterra Models for Food Chains

A simplified picture of an ecosystem is that of a food chain. Basal species in such a chain are
primary producers, e.g., plants and algae, that convert solar energy and CO2 into biomass.
Primary producers are eaten by grazers. Grazers are eaten by predators which themselves
are eaten by larger predators. At top of the chain sits the top-predator. Most ecosystems

66
are more complicated and form food-webs rather than food chains. Nevertheless, food
chains are a good starting point to investigate multi-species interactions. The Lotka-
Volterra version of a food chain of length n with self-regulation looks as follows:

ẋ1 = x1 (r1 − a11 x1 − a12 x2 ) (163a)

ẋi = xi (−ri + ai,i−1 xi−1 − aii xi − ai,i+1 xi+1 ) for i ∈ {2, . . . , n − 1} (163b)
ẋn = xn (−rn + an,n−1 xn−1 − ann xn ), (163c)

with ri , aij > 0. For this system we recover the dynamics of the two-dimensional Lotka-
Volterra predator-prey model.

Theorem If Eq. (163) has an internal equilibrium point x∗ > 0, then all orbits in Rn>0
converge to x . Thus, x∗ is globally stable and (with the results above) periodic or chaotic
∗

attractors cannot exist.

Proof We will show that the following function is a Lyapunov function:

n
X
V (x) = ci (x∗i log xi − xi ) (164)
i=1

with ci > 0 to be specified later. Note first that V (x) has a unique maximum at x = x∗ .
Then derive n
X ẋi
V̇ (x(t)) = ci (x∗i − xi ). (165)
i=1
xi
Replacing ẋi with corresponding equation from Eq. (163) and using that

r1 = a11 x∗1 + a12 x∗2

ri = ai,i−1 x∗i−1 − aii x∗i − ai,i+1 x∗i+1 for i ∈ {2, . . . , n − 1}
rn = an,n−1 x∗n−1 − ann x∗n ,

we get
ẋ1
= a11 (x1 − x∗1 ) + a12 (x2 − x∗2 )
x1
ẋi
= ai,i−1 (xi−1 − x∗i−1 ) − aii (xi − x∗i ) − ai,i+1 (xi+1 − x∗i+1 )
xi
ẋn
= an,n−1 (xn−1 − x∗n−1 ) − ann (xn − x∗n ).
xn
Eq. (165) then becomes
n
X n−1
X
V̇ (x(t)) = ci aii (xi − x∗i )2 − (xi − x∗i )(xi+1 − x∗i+1 )(ci+i ai+1,i − ci ai,i+1 ).
i=1 i=1

67
If we choose ci > 0 such that ci+1 /ci = ai,i+1 /ai+1,i , then this equation simplifies to
n
X
V̇ (x(t)) = ci aii (xi − x∗i )2 ≥ 0. (166)
i=1

With the Lyapunov theorem follows that for all points x ∈ Rn>0 we have ω(x) = x∗ . Note
that for the proof it is sufficient if only one species has intraspecific density regulation, i.e.,
aii ≥ 0 ∀i and aii > 0 for at least one species.

Three-level food chain

Let us have a look at a specific example. A minimal food chain consists of three trophic
levels: a primary producer x1 (plants), a herbivore x2 and a carnivore x3 . Here, each
level of the food chain need not to consist of only a single species. Instead, each level can
represent the whole functional group. For simplicity, we assume that intraspecific density
regulation on the level of the consumers can be ignored, a22 = a33 = 0 and obtain
ẋ1 = x1 (r1 − a11 x1 − a12 x2 ) (167a)
ẋ2 = x2 (−r2 + a21 x1 − a23 x3 ) (167b)
ẋ3 = x3 (−r3 + a32 x2 ). (167c)

Equilibria and stability The trivial equilibrium at 0 is unstable for any r1 > 0. A
non-zero single-species equilibrium only exists for the producer, at x∗1 = (r1 /a11 , 0, 0). It
is stable for r1 /a11 < r2 /a21 , which is the producer minimal density at which a herbivore
can persist. In the absence of the carnivore, this result follows from the 2-species LV model
(see Eq. 134). The coordinates of the equilibrium with plants and herbivores coexisting
can also be taken from there (135),
 r a r −a r 
∗2 2 21 1 11 2
x = , ,0 .
a21 a12 a21
Obviously, there cannot be any other 2-species equilibria. x∗2 is stable until the herbivore
density reaches the minimal level for carnivores to invade, at x2 = r3 /a32 . Finally, the
equilibrium x∗3 with all three trophic levels present equals
r1 a32 − r3 a12
x∗3
1 = (168a)
a11 a32
r3
x∗3
2 = (168b)
a32
r1 a21 a32 − r2 a11 a32 − r3 a12 a21
x∗3
3 = . (168c)
a11 a23 a32
Figure 14 shows the bifurcation diagram for the equilibrium densities as a function of
system productivity as given by r1 . A long standing question in population ecology is
whether population densities are regulated top down or bottom up. Based on this simple
model the following conclusions can be drawn.

68
 x∗3

x∗2

x∗1

r2 a11 a11 a32 r2 − a12 a21 r3 r1

a21 a21 a32

Figure 14: Bifurcation diagram for the equilibrium densities of Eq. (167) with bifurcation
parameter r1 .

1. System productivity controls total biomass and the number of trophic levels.

2. In a tritrophic food chain, herbivore density is controlled top down. Plants and
carnivores benefit from increased system productivity.

3. Increased death rate in the carnivore increases herbivore abundance and thereby
suppresses plant abundance (trophic cascade).

4. Adding more trophic levels has the following effect: systems with an odd number
of trophic levels release plant growth while systems with an even number of trophic
levels control plant growth.

Empirical support for this model comes mainly from aquatic system.

• At the east coast of North America different kelp species, invertebrate grazers such
as sea urchins, and sea-otters form a tritrophic food chain. Data from Alaska show
that as sea otters became rare, sea urchins became more common and many kelp
forests disappeared (and with them a very rich ecosystem of organisms living in the
kelp forests). Sea otters apparently became rarer when orca whales increased their
hunting pressure on them, possibly due to decreased fish stocks.

69
 • The food web in many fresh water lakes essentially consists of phytoplankton as
primary producers, zooplankton, planktivorous fish, and possibly piscivorous fish.
Whole-lake manipulation experiments show that lakes that lack piscivorous fish tend
to be “green” (tritrophic food chain resulting in abundant phytoplankton), whereas
“blue” lakes result where piscivorous fish are added as fourth trophic level. This
mechanism forms the basis for many lake restoration projects.

6.3 Coexistence and Competitive Exclusion

In many natural systems a set of consumer species shares the same resources. Look at the
different species of birds at your feeder. But how many consumer species can coexist on
a given set of resources? In the Lotka-Volterra competition model we did not model the
cause of competition explicitly but only indirectly via the interaction terms aij . Two causes
of competition can be distinguished. First, competition can be due to direct interactions
such as fighting. This is know as interference competition. Second, competition can be
due to the depletion of shared resources or exploitation competition. In the latter case one
might want to model the dynamics of the resources explicitly. A minimum model consist
of three species: one resource (or prey) and two competing consumers (or predators) and
can be modeled as follows,

ẋ = rx x − axx x2 − axy1 xy1 − axy2 xy2 (169a)

ẏ1 = −ry1 y1 + ay1 x xy1 (169b)
ẏ2 = −ry2 y2 + ay2 x xy2 , (169c)

where all parameters are positive. There is no direct interaction among the two types
of predators and no direct intraspecific competition in the predators, either. For only a
single predator, the system reduces to the Lotka-Volterra predator-prey model that we
have discussed above (with intraspecific exploitation competition among predators). For
the two predator species to coexist at equilibrium we require

ẏ1 = 0 ⇐⇒ ry1 = ay1 x x∗

ẏ2 = 0 ⇐⇒ ry2 = ay2 x x∗ .

This is a system of two linear equations in one unknown and thus overdetermined. In our
case, coexistence is only possible if ry1 /ry2 = ay1 x /ay2 x . Thus, coexistence is not robust with
respect to perturbations in the parameter values: two consumer species cannot robustly
coexist on one resource. This results can be stated much more generally, also beyond the
Lotka-Volterra framework.

Theorem: Competitive Exclusion In a system of exploitation competition with n

consumer species and m < n resources, robust coexistence at an equilibrium point is not
possible.

70
Proof Let the dynamics of m resource and n consumer species be given by

ẋ = f (x, y)

and
ẏ = y ◦ g(x),
respectively, where ◦ denotes the Hadamard product. At a point equilibrium y ∗ ∈ Rn>0 it
is required that
0 = g(x∗ ).
This is a system of n equations in m unknowns. It can only have a robust solution when
m ≤ n.

• Note that the theorem only excludes fixed points. For a general model, there could be
closed orbits – we have previously seen that predator-prey systems have a tendency
to cycle – or chaotic attractors. Orbits could also diverge to infinity, although this
should not occur for a realistic population dynamical model.

• For Lotka-Volterra systems, absence of an internal fixed point implies that other
bounded ω-limits are also impossible (see our general results above). If the dynamics
is bounded, we thus can conclude that all orbits converge to the boundary and at
least one species will go extinct.

• A crucial assumption of the theorem is that there is no direct intraspecific competi-

tion for the consumer species. It is quite obvious that sufficiently strong intraspecific
competition will enable coexistence on even just a single (abundant) resource. How-
ever, intraspecific competition usually indicates that there is some other limiting
resource (space, nesting places, some other prey species), even if this resource is not
explicitly included in the dynamics.

• The exclusion principle is an influential concept in ecology: coexistence requires

exclusive resources, which can not (or less easily) be exploited by competing species.
However, it is not necessarily easy to apply it in practice. What should be counted
as different resource types? E.g., if birds feed on the seeds of a certain plant species
this may be counted as a single resource or as infinitely many different resources if
the size distribution of the seeds is continuous.

Cyclic competition
Consider the Lotka-Volterra competition model with two species, following

ẋ1 = x1 (1 − x1 − αx2 ) (170)

ẋ2 = x2 (1 − βx1 − x2 ) , (171)

71
with α, β > 0. For simplicity, we have set r1 = r2 = a11 = a22 = 1. The single species
thus have carrying capacities K1 = K2 = 1. For α, β < 1 both species can coexist, but
for α, β > 1 they are mutually exclusive, and for α < 1 and β > 1 (α > 1, β < 1)
species 1 (species 2) dominates and drives the other one to extinction. Consider a case of
dominance, e.g., species 2 dominating over species 1. One can ask whether it is possible
to introduce a third competing species such that the system of all three species allows
for stable coexistence. Intuitively, if such a species 3 exists, it should “beat” (dominate)
species 2, but should be dominated by species 1. An easy symmetric example is provided
by the system

ẋ1 = x1 (1 − x1 − αx2 − βx3 ) (172)

ẋ2 = x2 (1 − βx1 − x2 − αx3 ) (173)
ẋ3 = x3 (1 − αx1 − βx2 − x3 ) (174)

with 0 < β < 1 < α. Clearly, we find circular dominance on the boundary of the state
space: in the x1 − x2 plane x2 dominates and e2 = (0, 1, 0) is globally attracting. Similarly,
e3 = (0, 0, 1) and e1 = (1, 0, 0) are the global attractor on the x2 − x3 plane and the x3 − x1
plane, respectively. In the full three species system, however, the unit vectors ei , i = 1, 2, 3
are no longer stable, but are saddle points. The only other fixed point on the boundary is
the trivial one (0, 0, 0), which is always a source. Finally, there is a unique internal fixed
point at
1
m= (1, 1, 1) . (175)
1+α+β
The Jacobian at m reads
 
−1 −α −β
1 −β −1 −α .
1+α+β
−α −β −1

This is a so-called circulant matrix with equal entries in each row (and column), with
a circular shift by one from one row to the next. The eigenvalues and eigenvectors of
circulants are easy to derive.

Lemma The eigenvalues λk and eigenvectors v k , k = 0, . . . , n − 1 of an n-dimensional

circulant matrix  
c0 c1 . . . cn−1
cn−1 c0 . . . cn−2 
J =  ..
 
.. 
 . . 
c1 c2 . . . c0
are given by
n−1
X
λk = cj µj·k , v k = (1, µk , µ2k , . . . , µ(n−1)k ) (176)
j=0

72
where  2πi 
µ = exp
n
is the nth root of unity. (The proof follows immediately by inserting λk and v k into the
eigenvalue equation.)
We thus obtain for the Jacobian J ,

λ0 = −1 , v 1 = (1, 1, 1), (177)

1  
λ1 = λ̄2 = − 1 − α exp[2πi/3] − β exp[4πi/3] . (178)
1+α+β
We observe the following

• Since
1  α + β
Re[λ1 ] = Re[λ2 ] = −1+
1+α+β 2
m is a sink (stable) if α + β < 2 and that m is a saddle (unstable) if α + β > 2.

• The diagonal x1 = x2 = x3 is invariant under the ODE flow. If m is a saddle, it is

the stable manifold of the equilibrium.

• All orbits in the interior of a boundary plane have one of the edges ei as ω-limit.
Each of these points has an unstable manifold on the “‘next” boundary (e.g., e2 is the
ω-limit for all orbits in the interior of the x1 /x2 plane, but has an unstable manifold
on the x2 /x3 plane). The three unstable manifolds connect the three edges. Together
with these they form a closed so-called heteroclinic cycle.

To analyze the dynamics in the interior of the state space, we consider two functions

S = x1 + x2 + x3 ; P = x1 · x2 · x 3 . (179)

Since we have
(
S(1 − S) α+β >2
Ṡ = x1 +x2 +x3 − x21 +x22 +x23 +(α+β)(x1 x2 +x2 x3 +x3 x1 ) ≤


S(1 − α+β
2
S) α+β ≤2

the ω-limits of all orbits are confined to a bounded region S ≤ max[1, 2/(α + β)] and no
population can explode. We further derive


Ṗ = ẋ1 x2 x3 + x1 ẋ2 x3 + x1 x2 ẋ3 = P 3 − (1 + α + β)S .

We will now show that P/S 3 is a Lyapunov function in the interior of the state space.
Indeed, we have
d P P (S − 3xi )
= = 0 for x1 = x2 = x3 ,
dxi S 3 xi S 4

73
where it assumes a maximum. We also have P/S 3 ≥ 0 and P/S 3 = 0 (minimum) at the
boundary where P = 0. For the time derivative, we derive

d P P α + β  2 2 2

= 1 − (x 1 − x 2 ) + (x 2 − x 3 ) + (x 3 − x 1 )
dt S 3 S4 2

and conclude the following

• For α+β < 2, the Lyapunov function is monotonically increasing and we converges to
the diagonal x1 = x2 = x3 . We have already seen that on this diagonal convergence
to m follows, which is thus the global attractor for every orbit in the interior.

• For α + β < 2, the Lyapunov function is monotonically decreasing. Every orbit not
on the diagonal x1 = x2 = x3 thus converges to the boundary. at the boundary it
must follow the heteroclinic cycle, which is the (almost) globally attracting ω-limit
in this case.

• Our symmetric example of cyclic competition does not have a real biological mo-
tivation. It mainly demonstrates what kind of phenomena can occur in higher di-
mensional systems. Indeed in Lotka-Volterra models with two consumers and two
resources heteroclinic cycles appear quite naturally (see Hofbauer & Sigmund, chap-
ter 16).

• Heteroclinic cycles can also occur in two-dimensional dynamical systems (but not in
2d Lotka-Volterra systems or other typical population dynamical models). This is
why we needed to account for a third category for ω-limits in the Pointcaré-Bendixson
Theorem. See the book by Teschl (chapter 7) for an example.

6.4 Coexistence concepts

At this point we should make precise what we mean by coexistence. In systems with
(potentially) closed orbits, heteroclinic cycles, or chaotic motion this can be a subtle issue.
We can rarely expect to obtain a full description of all ω-limits together with their attraction
domains. For many purposes, however, the precise asymptotics is less important than the
question of extinction, i.e., whether all species in the system will survive in the long term
or not. We are particularly interested in global concepts, in the sense that survival does
not depend on initial conditions (given that all species are initially present). Various
mathematical concepts exist, with different biological implications. Consider a dynamical
system
ẋi = xi fi (x) i = 1, . . . , n

that is defined on Rn≥0 .

74
Defintion
1. The system is persistent if for every orbit with starting value x(0) in the interior of
the state space
lim sup xi (t) > 0 ∀i . (180)
t→∞

2. We say that a system is invasible if every attracting ω-limit on the boundary of the
reduced system (with one or several species removed) can be invaded by a missing
species when added in very low density. Formally, we require that averaged per
captita growth rate for some missing species i is positive,
1 T
Z
lim fi (x(t)) dt > 0 .
T →∞ T 0

(For chaotic attractors we need to specify whether we require this for every orbit or
almost every orbit x(t).)
3. The dynamical system is said to be permanent if there exists δ, D > 0 such that
xi (0) > 0 for i ∈ {1, . . . , n} implies

lim inf xi (t) > δ for i ∈ {1, . . . , n}. (181)

t→+∞

and
lim sup xi (t) ≤ D for i ∈ {1, . . . , n}. (182)
t→+∞

Note that δ and D do not depend on the initial condition xi (0). Condition (182)
(uniform boundedness) holds whenever populations cannot grow to infinity, thus, for
any well-formulated population dynamical model.

• It is clear that permanence implies persistence, but the opposite is not true, even
if we consider only uniformly bounded systems. A system with ẋi = 0 for xi < D
is persistent, but not invasible or permanent. The same holds for the degenerate
Lotka-Volterra predator-prey model with center-type behavior.
• For purposes of biological application, persistence is usually too weak. There is
no force inherent to the system that would protect a species from going extinct.
Coexistence is thus not stable. This is different for permanent systems, where the
boundary is a repeller. Indeed, a system is permanent if a compact set K in the
interior of the state space exists such that all orbits in the interior end up in K.
• Invasibility guarantees that species can grow from low densities. The concept thus
implies better stability than persistence. However, invasibility does not imply persis-
tence, an example is given in the model below. It can also be “tricked” by heteroclinic
cycles, where invasion of one species means that another one will eventually die out.
The only concept to exclude an attracting heteroclinic cycle like in the 3-species
Lotka-Volterra system discussed above is permanence.

75
 • An advantage of invasibility over permanence as defined here is that it is robust with
respect to small perturbations of the system. Indeed, a system can be permanent, but
not invasible, if orbits close to the boundary are repelled with zero speed. However,
we can extend the definition of permanence to require robustness. A more relevant
advantage of invasibility is that the conditions are often easier to check.

6.5 Dynamic coexistence of two predators on a single resource

As an instructive application of coexistence concepts, we study the potential for coexistence
of two predator species y1 and y2 on one resource species x. Consider the following model
due to Armstrong and McGehee:
 x  y1 xe1
ẋ = xr 1 − − − y2 xe2 (183a)
K  a+x
xc1 
ẏ1 = y1 f1 (x) = y1 − d1 (183b)
a + x 
ẏ2 = y2 f2 (x) = y2 xc2 − d2 (183c)

We have have previously studied the two-species systems (183a) & (183b) and (183a) &
(183c) in isolation. For the first system (Rosenzweig-MacArthur, RMA), we have seen that
the equilibrium point is either globally stable or, when it is unstable, prey and predator
coexist on a limit cycle. For the second system (Lotka-Volterra, LV), the internal equilib-
rium point is globally stable whenever it exists. We denote the internal equilibrium points
of RMA system (183a & 183b) and the LV system (183a & 183c) by (x∗ , y1∗ ) and (x∗∗ , y2∗∗ ),
respectively. To study coexistence in the three-species system, we first focus on invasibility.
It is easy to see that no predator can survive in the presence of the other one if it cannot
even survive without its competitor. We then need to distinguish two cases.
1. Assume first that both the RMA and the LV two-species systems have a stable
equilibrium point. It is easy to see that the system is not mutually invasible in this
case. Indeed, the second predator can invade the stable equilibrium (x∗ , y1∗ ) of the
RMA system if and only if
x∗ c2 − d2 > 0. (184)
Since x∗∗ c2 −d2 = 0, this is possible if and only if x∗ > x∗∗ , in other words, if predator
2 can survive on fewer resources than predator 1. Vice versa, the first predator can
invade the equilibrium (x∗∗ , y2∗∗ ) of the LV system if and only if
x∗∗ c1
− d1 > 0. (185)
a + x∗∗
Since x∗ c1 /(a+x∗ )−d1 = 0 and the per-capita growth rate f1 (x) is strictly monotonic
in x, this is possible if and only if x∗∗ > x∗ . These two conditions cannot be satisfied
simultaneously: there is no mutual invasibility. Since we have no interior fixed points
in the (y1 , y2 )-plane (we have two equations in one unknown) coexistence is indeed
impossible and invasion implies fixation.

76
2. Since the per-capita growth rates fi (x) of both predators are monotonic, there are
unique minimal prey densities f1 (x∗ ) = 0 and f2 (x∗∗ ) = 0 at which the ith predator
can maintain itself in a constant environment (with constant x). Let us now drop
the assumption of a stable fixed point in the two-species systems. In particular, we
allow that the predator and prey density, yi (t) and x(t), vary in time but return to
their initial value after some time τ . The mean prey density over one cycle of period
τ is given by
1 τ
Z
x̄ := x(t)dt. (186)
τ 0
The mean per-capita predator growth rate over one cycle is
1 τ
Z
fi (x) := fi (x(t)) dt. (187)
τ 0
After one period the population size is unchanged: yi (t) = yi (t + τ ). Therefore,
Z τ Z τ
1 dyi
fi (x(t)) dt = dt = ln yi (τ ) − ln yi (0) = 0 (188)
0 0 yi dt

and thus
fi (x) = 0 (189)
has to hold.
A species with a linear functional response (LV system) requires the same average
resource density to maintain itself in an environment with cycling x as it does in a
constant environment, x̄ = x∗∗ . Indeed, for f2 (x) = c2 x − d2 we have
c2 τ 1 τ 1 τ
Z Z Z
f2 (x̄) = x(t)dt − d2 = (c2 x(t) − d2 )dt = f2 (x(t))dt = f2 (x) = 0,
τ 0 τ 0 τ 0
(190)
where the last equality follows from Eq. (189). It follows that x̄ = x∗∗ .
Next, consider a species with a saturating functional response (RMA system), f1 =
xc1 /(a + x) − d1 . In this case, the average resource density required for maintenance
in a time-varying environment is higher than in a constant environment. This makes
intuitive sense: the predator profits only little from times with abundant prey, but
suffers a lot from times with rare prey. To verify this, denote with l(x) the tangent
line to f1 (x) at x∗ (see Fig. 15). Then l(x∗ ) = f1 (x∗ ) = 0, but l(x) > f1 (x) for x 6= x∗
and therefore l(x̄) = l(x) > f1 (x) = 0. Thus, l(x̄) > l(x∗ ), which implies x̄ > x∗ .
Consider now mutual invasibility for parameter values with unstable fixed point
(x∗ , y1∗ ) for the RMA system, where predator and the prey coexist along a peri-
odic orbit γ with period τ . Predator 2 can invade the periodic attractor of predator
1 and the resource if and only if its average per capita growth rate over one cycle
(while rare) is positive:
1 τ 1 τ
Z Z
f2 (x(t))dt = (c2 x(t) − d2 )dt = c2 x̄ − d2 > 0. (191)
τ 0 τ 0

77
 [

f1(x)

Figure 15: Graphs of the functions f1 (x) and l(x). l(x) is the tangent line to f1 (x) at x∗ .
(From Armstrong and McGehee 1980.)

Hence, invasion is possible if and only if x̄ > x∗∗ . We can conclude that the condition
for mutual invadability equals
x∗ < x∗∗ < x̄, (192)

which can be satisfied for appropriately chosen parameter values. One can show that
the three-species community is stable and cycles along a periodic orbit (McGehee &
Armstrong, 1977, J.Diff.Eq. 23:30-52; Armstrong & McGehee, 1980, AmNat 115:151-
170).

• Why is coexistence in this model possible, seemingly violating the principal of com-
petitive exclusion? Figure 16(a) shows the per capita growth rates for the two preda-
tors for the different prey densities that occur during one cycle. We can see that
predator 1 is superior at low resource densities while predator 2 is superior at high
resource densities. Thus, which predator is superior regularly changes along the
periodic orbit, preventing that one predator excludes the other (Figure 16b).

• One can argue that different densities of the same resource species have to be in-
terpreted as different resources. In this sense, the present model contains infinitely
many resources. Accordingly, in principle any number of consumer species can co-
exist on a single cycling resource. The principle of competitive exclusion can be
generalized in a way that it still holds. Of course, more consumers can only coexist
if the parameters of their functional response are suitably fine-tuned, making their
coexistence less robust to perturbations.

78
 (a) (b)
f(x)

Figure 16: (a) Growth rates of predators of Eq. (183) as a function of prey density x. The
region C represents the approximate range of variation over x over one cycle in the three
species system (from Armstrong and McGehee 1980). (b) Predator 1 has a higher per
capita growth rate at low resource densities whereas Predator 2 has a higher per capita
growth rate at high resource densities (from Yodzis 1989).

• For stable coexistence in the sense of mutual invasibility we need x∗ < x∗∗ . Note that
the unstable fixed point (x∗ , y1∗ ) of the RMA 2-species system can not be invaded by
the second predator under this condition. Indeed, (x∗ , y1∗ ) is a saddle with stable
manifold pointing into the interior of the state space. As a consequence, there is
a single orbit (along the stable manifold) that converges to the boundary and the
system is neither permanent nor persistent!

79