Arild Wikan

Discrete Dynamical Systems with an

Introduction to Discrete Optimization
Problems

2
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an Introduction to Discrete Optimization Problems
1st edition
2013 Arild Wikan & bookboon.com
ISBN 978-87-403-0327-8

3
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Contents

Contents

Acknowledgements 6

Introduction 7

Part 1 O
 ne-dimensional maps f :RR x f (x) 11
1.1 Preliminaries and definitions 12
1.2 One-parameter family of maps 16
1.3 Fixed points and periodic points of the quadratic map 19
1.4 Stability 24
1.5 Bifurcations 30
1.6 The flip bifurcation sequence 35
1.7 Period 3 implies chaos. Sarkovskiis theorem 38
1.8 The Schwarzian derivative 42
1.9 Symbolic dynamics I 45
1.10 Symbolic dynamics II 50
1.11 Chaos 60
1.12 Superstable orbits and a summary of the dynamics of the quadratic map 64

www.sylvania.com

We do not reinvent
the wheel we reinvent
light.
Fascinating lighting offers an infinite spectrum of
possibilities: Innovative technologies and new
markets provide both opportunities and challenges.
An environment in which your expertise is in high
demand. Enjoy the supportive working atmosphere
within our global group and benefit from international
career paths. Implement sustainable ideas in close
cooperation with other specialists and contribute to
influencing our future. Come and join us in reinventing
light every day.

Light is OSRAM

4 Click on the ad to read more

Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization Contents

Part II n-dimensional maps f : Rn Rn x f (x) 68

2.1 Higher order difference equations 69
2.2 Systems of linear difference equations. Linear maps from R to R 86 n n

2.3 The Leslie matrix 97

2.4 Fixed points and stability of nonlinear systems 106
2.5 The Hopf bifurcation 115
2.6 Symbolic dynamics III (The Horseshoe map) 132
2.7 The center manifold theorem 138
2.8 Beyond the Hopf bifurcation, possible routes to chaos 147
2.9 Difference-Delay equations 173

Part III Discrete Time Optimization Problems 187

3.1 The fundamental equation of discrete dynamic programming
360 188

.
3.2 The maximum principle (Discrete version) 198
3.3 Infinite horizon problems 206
3.4 Discrete stochastic optimization problems
thinking 218

Appendix (Parameter Estimation) 234

References 247

360
thinking . 360
thinking .
Discover the truth at www.deloitte.ca/careers Dis

Deloitte & Touche LLP and affiliated entities.

Discover the truth at www.deloitte.ca/careers Deloitte & Touche LLP and affiliated entities.

Deloitte & Touche LLP and affiliated entities.

Discover the truth

5 at www.deloitte.ca/careers
Click on the ad to read more
Download free eBooks at bookboon.com

Deloitte & Touche LLP and affiliated entities.

Discrete Dynamical Systems with an
Introduction to Discrete Optimization Acknowledgements

Acknowledgements
My special thanks goes to Einar Mjlhus who introduced me to the fascinating world of discrete dynamical
systems. Responses from B. Davidsen, A. Eide, O. Flaaten, A. Seierstad, A. Strm, and K. Sydster are
also gratefully acknowledged.

I also want to thank Liv Larssen for her excellent typing of this manuscript and . Kristensen for his
assistance regarding the figures.

Financial support from Harstad University College is also gratefully acknowledged.

Finally I would like to thank my family for bearing over with me throughout the writing process.

Autumn 2012
Arild Wikan

6
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Introduction

Introduction
In most textbooks on dynamical systems, focus is on continuous systems which leads to the study of
differential equations rather than on discrete systems which results in the study of maps or difference
equations. This fact has in many respects an obvious historical explanation. If we go back to the time of
Newton (16421727), Leibniz (16461716), and some years later to Euler (17091783), many important
aspects of the theory of continuous dynamical systems were established. Newton was interested in
problems within celestial mechanics, in particular problems concerning the computations of planet
motions, and the study of such kind of problems lead to differential equations which he solved mainly
by use of power series method. Leibniz discovered in 1691 how to solve separable differential equations,
and three years later he established a solution method for first order linear equations as well. Euler (1739)
showed how to solve higher order differential equations with constant coefficients. Later on, in fields
such as fluid mechanics, relativity, quantum mechanics, but also in other scientific branches like ecology,
biology and economy, it became clear that important problems could be formulated in an elegant and
often simple way in terms of differential equations. However, to solve these (nonlinear) equations proved
to be very difficult. Therefore, throughout the years, a rich and vast literature on continuous dynamical
systems has been established.

Regarding discrete systems (maps or difference equations), the pioneers made important contributions
here too. Indeed, Newton designed a numerical algorithm, known as Newtons method, for computing
zeros of equations and Euler developed a discrete method, Eulers method (which often is referred to as a
first order RungeKutta method), which was applied in order to solve differential equations numerically.

Modern dynamical system theory (both continuous and discrete) is not that old. It began in the last
part of the nineteenth century, mainly due to the work of Poincar who (among lots of other topics)
introduced the Poincar return map as a powerful tool in his qualitative approach towards the study of
differential equations. Later in the twentieth century Birkhoff (1927) too made important contributions
to the field by showing how discrete maps could be used in order to understand the global behaviour
of differential equation systems. Julia considered complex maps and the outstanding works of Russian
mathematicians like Andronov, Liapunov and Arnold really developed the modern theory further.

In this book we shall concentrate on discrete dynamical systems. There are several reasons for such a
choice. As already metioned, there is a rich and vast literature on continuous dynamical systems, but
there are only a few textbooks which treat discrete systems exclusively.

7
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Introduction

Secondly, while many textbooks take examples from physics, we shall here illustrate large parts of the
theory we present by problems from biology and ecology, in fact, most examples are taken from problems
which arise in population dynamical studies. Regarding such studies, there is a growing understanding
in biological and ecological communities that species which exhibit birth pulse fertilities (species that
reproduce in a short time interval during a year) should be modelled by use of difference equations
(or maps) rather than differential equations, cf. the discussion in Cushing (1998) and Caswell (2001).
Therefore, such studies provide an excellent ground for illuminating important ideas and concepts from
discrete dynamical system theory.

Another important aspect which we also want to stress is the fact that in case of low-dimensional
problems (problems with only one or two state variables) the possible dynamics found in nonlinear
discrete models is much richer than in their continuous counterparts. Indeed, let us briefly illustrate this
aspect through the following example:

Let N = N(t) be the size of a population at time t . In 1837 Verhulst suggested that the change of
N could be described by the differential equation (later known as the Verhulst equation)
 
N
N = rN 1 (I1)
K

where the parameter r ( r > 0 ) is the intrinsic growth rate at low densities and K is the carrying
capacity. Now, define x = N/K . Then (I1) may be rewritten as

x = rx(1 x) (I2)

which (as (I1) too) is nothing but a separable equation. Hence, it is straightforward to show that its
solution becomes

1
x(t) = x0 1 rt
(I3)
1 x0
e

where we also have used the initial condition x(0) = x0 > 0 . From (I3) we conclude that x(t) 1 as
t which means that x = 1 is a stable fixed point of (I2). Moreover, as is true for (I1) we have
proved that the population N will settle at its carrying capacity K .

Next, let us turn to the discrete analogue of (I2). From (I2) it follows that

xt+1 xt
rxt (1 xt ) (I4)
t

8
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Introduction

which implies
 
rt
xt+1 = xt + rtxt rtx2t = (1 + rt)xt 1 xt (I5)
1 + rt
and through the definition y = rt(1 + rt)1 x we easily obtain

yt+1 = yt (1 yt ) (I6)

where = 1 + rt .

The sweet and innocent-looking equation (I6) is often referred to as the quadratic or the logistic map.
Its possible dynamical outcomes were presented by Sir Robert May in an influential review article called
Simple mathematical models with very complicated dynamics in Nature (1976). There, he showed,
depending on the value of the parameter , that the asymptotic behaviour of (I6) could be a stable
fixed point (just as in (I2)), but also periodic solutions of both even and odd periods as well as chaotic
behaviour. Thus the dynamic outcome of (I6) is richer and much more complicated than the behaviour
of the continuous counterpart (I2).

Hence, instead of considering continuous systems where the number of state variables is at least 3 (the
minimum number of state variables for a continuous system to exhibit chaotic behaviour), we find it much
more convenient to concentrate on discrete systems so that we can introduce and discuss important definitions,
ideas and concepts without having to consider more complicated (continuous) models than necessary.

The book is divided into three parts. In Part I, we will develop the necessary qualitative theory which will
enable us to understand the complex nature of one-dimensional maps. Definitions, theorems and proofs
shall be given in a general context, but most examples are taken from biology and ecology. Equation
(I6) will on many occasions serve as a running example throughout the text. In Part II the theory will
be extended to n-dimensional maps (or systems of difference equations). A couple of sections where
we present various solution methods of higher order and systems of linear difference equations are also
included. As in Part I, the theory will be illustrated and exemplified by use of population models from
biology and ecology. In particular, Leslie matrix models and their relatives, stage structured models
shall frequently serve as examples. In PartIII we focus on various aspects of discrete time optimization
problems which include both dynamic programming as well as discrete time control theory. Solution
methods of finite and infinite horizon problems are presented and the problems at hand may be of both
deterministic and stochastic nature. We have also included an Appendix where we briefly discuss how
parameters in models like those presented in Part I and Part II may be estimated by use of time series.
The motivation for this is that several of our population models may or have been applied on concrete
species which brings forward the question of estimation. Hence, instead of referring to the literature we
supply the necessary material here.

9
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Introduction

Finally, we want to repeat and emphasize that although we have used lots of examples and problems taken
from biology and ecology this is a Mathematics text so in order to be well prepared, the potential reader
should have a background from a calculus course and also a prerequisite of topics from linear algebra,
especially some knowledge of real and complex eigenvalues and associated eigenvectors. Regarding
section 2.5 where the Hopf bifurcation is presented, the reader would also benefit from a somewhat
deeper comprehension of complex numbers. This is all that is necessary really in order to establish the
machinery we need in order to study the fascinating behaviour of nonlinear maps.

We will turn your CV into

an opportunity of a lifetime

Do you like cars? Would you like to be a part of a successful brand? Send us your CV on
We will appreciate and reward both your enthusiasm and talent. www.employerforlife.com
Send us your CV. You will be surprised where it can take you.

10 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Part 1
 One-dimensional maps
f :RR x f (x)

11
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

1.1 Preliminaries and definitions

Let I R and J R be two intervals. If f is a map from I to J we will express that as f : I J ,
x f (x) . Sometimes we will also express the map as a difference equation xt+1 = f (xt ) . If the map
f depends on a parameter u we write fu (x) and say that f is a one-parameter family of maps.

For a given x0 , successive iterations of map f (or the difference equation xt+1 = f (xt ) ) give:
x1 = f (x0 ) , x2 = f (x1 ) = f (f (x0 )) = f 2 (x0 ) , x3 = f (x2 ) = f (f 2 (x0 )) = f 3 (x0 ) . . . , so after
n iterations xn+1 = f n (x0 ) . Thus, the orbit of a map is a sequence of points {x0 , f (x0 ), . . . , f n (x0 )}
which we for simplicity will write as {f n (x0 )} . This is in contrast to the continuous case (differential
equation) where the orbit is a curve.

As is true for differential equations it is a well-known fact that most classes of equations may not be
solved explicitly. The same is certainly true for maps. However, the map x f (x) = ax + b where
a and b are constants is solvable.

Theorem 1.1.1. The difference equation

xt+1 = axt + b (1.1.1)

has the solution

 
t b b
xt = a x0 + , a = 1 (1.1.2a)
1a 1a
xt = x0 + bt , a = 1 (1.1.2b)

where x0 is the initial value.

Proof. From (1.1.1) we have x1 = ax0 + b x2 = ax1 + b = a(ax0 + b) + b = a2 x0 + (a

2 + 1)b x3
x = 2ax0 + b x2 = ax1 + b = a(ax0 + b) + b = a x0 + (a + 1)
. . . = a3 x0 +1 (a + a + 1)b
+ b) + b = a x0 + (a + 1)b x3 = ax2 + b = . . . = a3 x0 + (a2 + a + 1)b. Thus assume xk = ak x0 + (ak1 + ak2 +
2

axkk++bb==aa aakkxx00+ k1 k2

xk = ak x0 + (ak1 + ak2 + . . . + a + 1)b . Then by induction: xxk+1 =ax
k+1 = (ak1
+(a
k
ak2
++ak1 ++. .. .. .++aa
xk+1 = ax k + b = a a x 0 + (a + ak2 + .
. . . + a + 1)b
. . . + a + 1)b
= a a x0 + (ak1 + ak2 + . . . + a + 1)b + b = ak+1 x0 + (ak + ak1 + . . . + a + 1)b . If a = 1 : 1 + a + . . . + ak = (1 at
 k 

1 + a + . . . + ak = (1 at )(1 a)1 so the solution becomes

1 at
 
t t b b
xt = a x0 + b = a x0 +
1a 1a 1a

If a = 1 : 1 + a + . . . + at1 = t 1 = t

xt = x0 + bt

12
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Regarding the asymptotic behaviour (long-time behaviour) we have from Theorem 1.1.1: If
|a| < 1 limt xt = b/(1 a) . (If x0 = b/(1 a) this is true for any a = 1 .) If a > 1 and
x0 = b/(1 a) the result is exponential growth or decay, and finally, if a < 1 divergent oscillations
is the outcome.

If b = 0 , (1.1.1) becomes

xt+1 = axt (1.1.2)

which we will refer to as the linear difference equation. The solution is

xt = at x0 (1.1.3)

Hence, whenever |a| < 1 , xt 0 asymptotically (as a convergent oscillation if 1 < a < 0 ). a > 1
or a < 1 gives exponential growth or divergent oscillations respectively.

Exercise 1.1.1. Solve and describe the asymptotic behaviour of the equations:

a) xt+1 = 2xt + 4 , x0 = 1 ,

b) 3xt+1 = xt + 2 , x0 = 2 . 

Exercise 1.1.2. Denote x = b/(1 a) where a = 1 and describe the asymptotic behaviour of
equation (1.1.1) in the following cases:

a) 0 < a < 1 and x0 < x ,

b) 1 < a < 0 and x0 < x ,

c) a > 1 and x0 > x .

Equations of the form xt+1 + axt = f (t) , for example xt+1 2xt = t2 + 1 , may be regarded as
special cases of the more general situation

xt+n + a1 xt+n1 + a2 xt+n2 + + an xt = f (t) , n = 1, 2, ...

Such equations are treated in Section 2.1 (cf. Theorem 2.1.6, see also examples following equation (2.1.6)
and Exercise 2.1.5).

13
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

When the map x f (x) is nonlinear (for example x 2x(1 x) ) there are no solution methods
so information of the asymptotic behaviour must be obtained by use of qualitative theory.

Definition 1.1.1. A fixed point x for the map x f (x) is a point which satisfies the equation
x = f (x ) .

Fixed points are of great importance to us and the following theorem will be very useful.

Theorem 1.1.2.

a) Let I = [a, b] be an interval and let f : I I be continuous. Then f has at least one fixed
point in I .

b) Suppose in addition that |f  (x)| < 1 for all x I . Then there exists a unique fixed point
for f in I , and moreover

|f (x) f (y)| < |x y|

AXA Global
Graduate Program
Find out more and apply

14 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Proof.

a) Define g(x) = f (x) x . Clearly, g(x) too is continuous. Suppose f (a) > a and
f (b) < b . (If f (a) = a or f (b) = b then a and b are fixed points.) Then g(a) > 0 and
g(b) < 0 so the intermediate value theorem from elementary calculus directly gives the
existence of c such that g(c) = 0 . Hence, c = f (c) .

b) From a) we know that there is at least one fixed point. Suppose that both x and y ( x = y )
are fixed points. Then according to the mean value theorem from elementary calculus there
exists c between x and y such that f (x) f (y) = f  (c)(x y) . This yields (since
x = f (x) , y = f (y) ) that

f (x) f (y)
f  (c) = =1
xy

This contradicts |f  (x)| < 1 . Thus x = y so the fixed point is unique. Further from the mean
value theorem:

|f (x) f (y)| = |f  (c)| |x y| < |x y|

Definition 1.1.2. Consider the map x f (x) . The point p is called a periodic point of period
n if p = f n (p) . The least n > 0 for which p = f n (p) is referred to as the prime period of p .

Note that a fixed point may be regarded as a periodic point of period one. 

Exercise 1.1.3. Find the fixed points and the period two points of f (x) = x3 . 

Definition 1.1.3. If f  (c) = 0 , c is called a critical point of f . c is nondegenerate if f  (c) = 0 ,

degenerate if f  (c) = 0 . 

The derivative of the n -th iterate f n (x) is easy to compute by use of the chain rule. Observe that
f n (x) = f (f n1(x)) , f n1 (x) = f (f n2(x)) . . . , f 2 (x) = f (f (x)) . Consequently:

f n (x) = f  (f n1 (x))f  (f n2 (x)) . . . f  (x) (1.1.5)

(1.1.5) enables us to compute the derivative at points on a periodic orbit in an elegant way. Indeed, suppose
the three cycle {p0 , p1 , p2 } where p1 = f (p0 ) , p2 = f (p1 ) = f 2 (p0 ) and f 3 (p0 ) = p0 . . . . Then


f 3 (p0 ) = f  (p2 )f  (p1 )f  (p0 ) (1.1.6)

15
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Obviously, if we have n periodic points {p0 , . . . , pn1 } the corresponding formulae is

n1

f n (p0 ) = f  (pi ) (1.1.7)
i=0

(Later on we shall use the derivative in order to decide whether a periodic orbit is stable or not. (1.1.7)
implies that all points on the orbit is stable (unstable) simultaneously.)

We will now proceed by introducing some maps (difference equations) that have been frequently applied
in population dynamics. Examples that show how to compute fixed points, periodic points, etc., will be
taken from these maps. Some computations are performed in the next section, others are postponed to
Section 1.3.

1.2 One-parameter family of maps

Here we shall briefly present some one-parameter family of maps which have often been applied in
population dynamical studies. Since x is supposed to be the size of a population, x 0 .

The map

x f (x) = x(1 x) (1.2.1)

is often referred to as the quadratic or the logistic map. The parameter is called the intrinsic growth
rate. Clearly x [0, 1] , otherwise xt > 1 xt+1 < 0 . If [0, 4] any iterate of f will remain in
[0, 1] . Further we may notice that f (0) = f (1) = 0 and x = c = 1/2 is the only critical point.

Definition 1.2.1. A map f : I I is said to be unimodal if a) f (0) = f (1) = 0 , and

b) f has a unique critical point c which satisfies 0 < c < 1 . 

Hence (1.2.1) is a unimodal map on the unit interval. Note that unimodal maps are increasing on the
interval [0, c and are decreasing on (c, 1] .

The map

x fr (x) = xer(1x) (1.2.2)

is called the Ricker map. Unlike the quadratic map, x [0,  . The parameter r is positive.

Exercise 1.2.1. Show that the fixed points of (1.2.2) are 0 and 1 and that the critical point
is 1/r .

16
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

The property that x [0,  makes (1.2.2) much more preferable to biologists than (1.2.1). Indeed,
let > 4 in (1.2.1). Then most points contained in [0, 1] will leave [0, 1] after a finite number of
iterations (the point x0 = 1/2 will leave the unit interval after only one iteration), and once xt > 1 ,
xt+1 < 0 which, of course, is unacceptable from a biological point of view. Such problems do not arise
by use of (1.2.2).

The map

ax
x fa,b (x) = (1.2.3)
(1 + x)b

where a > 1 , b > 1 is a two-parameter family of maps and is called the Hassel family.

1/b
Exercise 1.2.2. Show that x = 0 and x = a 1 are the fixed points of (1.2.3) and that
c = 1/(b 1) is the only critical point for x > 0 . 

The map

ax 0 x 1/2
x Ta (x) = (1.2.4)
a(1 x) 1/2 < x 1

e Graduate Programme
I joined MITAS because for Engineers and Geoscientists
I wanted real responsibili www.discovermitas.com
Maersk.com/Mitas e G
I joined MITAS because for Engine
I wanted real responsibili Ma

Month 16
I was a construction Mo
supervisor ina const
I was
the North Sea super
advising and the No
Real work he
helping foremen advis
International
al opportunities
Internationa
ree wo
work
or placements ssolve problems
Real work he
helping fo
International
Internationaal opportunities
ree wo
work
or placements ssolve pr

17 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

where a > 0 is called the tent map for obvious reasons. We will pay special attention to the case
a = 2 . Note that Ta (x) attains its maximum at x = 1/2 but that T  (1/2) does not exist. Since
Ta (0) = Ta (1) = 0 the map is unimodal on the unit interval.

Figure 1: The graphs of the functions: (a) f (x) = 4x(1 x) and (b) the tent function (cf. (1.2.4) where a = 2).

All functions defined in (1.2.1)(1.2.4) have one critical point only. Such functions are often referred to
as one-humped functions. In Figure 1a we show the graph of the quadratic functions (1.2.1) ( = 4 )
and in Figure 1b the tent function (1.2.4) ( a = 2 ). In both figures we have also drawn the line y = x
and we have marked the fixed points of the maps with dots.

As we have seen, maps (1.2.1)(1.2.4) share much of the same properties. Our next goal is to explore
this fact further.

Definition 1.2.2. Let f : U U and g : V V be two maps. If there exists a homeomorphism

h : U V such that h f = g h , then f and g are said to be topological equivalent.

Remark 1.2.1. A function h is a homeomorphism if it is one-to-one, onto and continuous and

that h1 is also continuous.

The important property of topological equivalent maps is that their dynamics is equivalent. Indeed, suppose
that x = f (x) . Then from the definition, h(f (x)) = h(x) = g(h(x)) , so if x is a fixed point of f , h(x) is a
fixed point for g . In a similar way, if p is a periodic point of f of prime period n (i.e. f n (p) = p ) we have
from Definition 1.2.2 that f = h1 g h f 2 = (h1 g h) (h1 g h) = h1 g 2 h
so clearly f n = h1 g n h . Consequently, h(f n (p)) = h(p) = g n (h(p)) so h(p) is a periodic
point of prime period n for g .

18
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Proposition 1.2.1. The quadratic map f : [0, 1] [0, 1] x f (x) = 4x(1 x) is topological
equivalent to the tent map

2x 0 x 1/2
T : [0, 1] [0, 1] x T (x) =
2(1 x) 1/2 < x 1

Proof. We must find a function h such that h f = T h . Note that this implies that we also
have f h1 = h1 T where h1 is the inverse of h .

Now, define h1 (x) = sin2 (x)/2 . Then

 x  x  x 
f h1 = f sin2 = 4 sin2 1 sin2
2 2 2
x x  x x 2
= 4 sin2 cos2 = 2 sin cos = sin2 x
2 2 2 2
1
h1 T = h1 (2x) = sin2 x 0x
2
1
h1 T = h1 (2(1 x)) = sin2 ( x) = sin2 x <x1
2

Thus, f h1 = h1 T which implies h f = T h so f and T are topological

equivalent.

1.3 Fixed points and periodic points of the quadratic map

Most of the theory that we shall develop in the next sections will be illustrated by use of the quadratic
map (1.2.1). In many respects (1.2.1) will serve as a running example. Therefore, in order to prepare the
ground we are here going to list some main properties.

The fixed points are obtained from x = x(1 x) . Thus the fixed points are x = 0 (the trivial fixed
point) and x = ( 1)/ (the nontrivial fixed point). Note that the nontrivial fixed point is positive
whenever > 1 . Assuming that (1.2.1) has periodic points of period two they must be found from
p = f2 (p) and since

2
f (p) = f (p(1 p)) = 2 p[1 ( + 1)p + 2p2 p3 ]

the two nontrivial periodic points must satisfy the cubic equation

3 p3 23 p2 + 2 ( + 1)p + 1 2 = 0 (1.3.1)

19
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Since every periodic point of prime period 1 is also a periodic point of period 2 we know that
p = ( 1)/ is a solution of (1.3.1). Therefore, by use of polynomial division we have

2 p2 (2 + )p + + 1 = 0 (1.3.2)

Thus, the periodic points are


+1 ( + 1)( 3)
p1,2 = (1.3.3)
2

where > 3 is a necessary condition for real solutions.

Period three points are obtained from p = f3 (p) and must be found by means of numerical methods.
(It is possible to show after a somewhat cumbersome calculation that the three periodic points do not

exist unless > 1 + 8 .)

In general, it is a hopeless task to compute periodic points of period n for a given map when n becomes
large. Therefore, it is in many respects a remarkable fact that it is possible when = 4 in the quadratic
map. We shall now demonstrate how such a calculation may be carried out, and in doing so, we find it
convenient to express (1.2.1) as a difference equation rather than a map.
170x115_MIM_ datos_Espv4.pdf 1 17/12/15 18:29

EL 93%
DE LOS ALUMNOS DEL MIM ESTN
TRABAJANDO EN SU SECTOR
C
A LOS 3 MESES DE GRADUARSE
M

MASTER IN MANAGEMENT
CM

MY

CY ESTUDIA EN EL CORAZN DE MADRID Y APROVECHA LAS OPORTUNIDADES Y EXPERIENCIAS NICAS

Duracin: 1O MESES
QUE OFRECE LA CAPITAL DE ESPAA
CMY Exp. Profesional: 1 AO
POTENCIA TU FORMACIN CURSANDO UNA DOBLE TITULACIN QUE MEJOR SE ADAPTE A TUS
Idioma: ESPAOL / INGLS
K
OBJETIVOS PROFESIONALES
Formato: TIEMPO COMPLETO
ESTUDIA UN SEMESTRE EXTRA EN OTRO PAS CON EL PROGRAMA BEYOND BORDERS EXPERIENCE
Convocatorias: SEPT / FEB
Y AUMENTA TU RED DE CONTACTOS

5 ESPECIALIZACIONES 100 DEL MUNDO 55 NACIONALIDADES

MASTERS IN MANAGEMENT
PERSONALIZA TU PROGRAMA FINANCIAL TIMES
EN CLASE

www.ie.edu/master-management [email protected] Sguenos en IE MIM Experience

20 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Thus consider

xt+1 = 4xt (1 xt ) (1.3.4)

Let xt = sin2 t . Then from (1.3.4):

sin2 t+1 = 4 sin2 t cos2 t = sin2 2t

Further:

sin2 t+2 = 4 sin2 t+1 (1 sin2 t+1 )

= 4 sin2 2t cos2 2t = sin2 22 t

Thus, after n iterations

2
sin t+n = sin2 2n t

which implies:

t+n = 2n t + l

Now, if we have a period n orbit ( xt+n = xt )

sin2 t+n = sin2 t
Hence:

t+n = t + m 2n t + l = t + m

(2n 1)t = (m l)

so

k
t =
2n 1

where k = m l . Consequently, the periodic points are given by

k
pi = sin2 (1.3.5)
2n 1
Example 1.3.1. Compute all the period 1, period 2 and period 3 points of f (x) = 4x(1 x) .
The period 1 points (which of course are the same as the fixed points) are

sin2 =0 sin2 = 0.75
21 2+1

21
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

The period 2 points are the period 1 points (which do not have prime period 2) plus the prime
period 2 points.

2 2
sin = 0.34549 sin2 = 0.904508
5 5

(The latter points may of course also be obtained from (1.3.3).)

There are six points of prime period 3. The points

2 2 4
sin = 0.188255 , sin2 = 0.611260 sin2 = 0.950484
7 7 7

are the periodic points in one 3-cycle, while the points

2 2 4
sin = 0.116977 , sin2 = 0.4131759 sin2 = 0.969846
9 9 9

are the periodic points on another orbit. (The reason why it is one 2-cycle but two 3-cycles is
strongly related to how they are created.)

Exercise 1.3.1. Use (1.3.5) to find all the period 4 points of f (x) = 4x(1 x) . How many
periodic points are there? 

Since f (x) = 4x(1 x) is topological equivalent to the tent map we may use (1.3.5)
together with Proposition 1.2.1 to find the periodic points of the tent map. Indeed, since

h1 (x) = sin2 (x/2) h(x) = (2/) arcsin x (cf. the proof of Proposition 1.2.1) the periodic

points p of T (x) may be found from T (h(p)) = T ((2/) arcsin p) . Thus the fixed points of the
tent map are

 
2 4
T arcsin 0 = arcsin 0 = 0

     
2 3 2 3
T arcsin = 2 1 arcsin = 0.6666
4 4

Exercise 1.3.2. Find the period 2 points of the tent map ( a = 2 ).

We shall close this section by computing numerically some orbits of the quadratic map for different
values of the parameter :

= 1.8 and x0 = 0.8 gives the orbit

{0.8 0.2880 0.3691 0.4192 0.4382 0.4431 0.4442 0.4444 0.4444 . . .}

22
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Thus the orbit converges towards the point 0.4444 which is nothing but the fixed point ( 1)/ . In
this case the fixed point is said to be locally asymptotic stable. (A precise definition will be given in the
next section.)

= 3.2 and x0 = 0.6 gives:

{0.6 0.7680 0.5702 0.7842 0.5415 0.7945 0.5225 0.7984 0.5151
0.7993 0.5134 0.7994 0.5131 0.7995 0.5130 0.7995 0.5130 . . .}

Thus in this case the orbit does not converge towards the fixed point. Instead we find that the asymptotic
behaviour is a stable periodic orbit of prime period 2. The points in the two-cycle are given by (1.3.3).

= 4.0 and x0 = 0.30 gives

{0.30 0.84 0.5376 0.9943 0.02249 0.0879 0.3208 0.8716 0.4476 0.9890 . . .}

Although care should be taken by drawing a conclusion after a few iterations only, the last example suggests
that there are no stable periodic orbit when = 4 . (A formal proof of this fact will be given later.)

23 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Exercise 1.3.3. Use a calculator or a computer to repeat the calculations above but use the initial
values 0.6, 0.7 and 0.32 instead of 0.8, 0.6 and 0.3, respectively. Establish the fact that the long-
time behaviour of the map when = 1.8 or = 3.2 is not sensitive to a slightly change of the
initial conditions but that there is a strong sensitivity in the last case.

1.4 Stability
Referring to the last example of the previous section we found that the equation xt+1 = 1.8xt (1 xt )
apparently possessed a stable fixed point and that the equation xt+1 = 3.2xt (1 xt ) did not. Both
these equations are special cases of the quadratic family (1.2.1) so what the example suggests is that by
increasing the parameter in (1.2.1) there exists a threshold value 0 where the fixed point of (1.2.1)
loses its stability.

Now, consider the general first order nonlinear equation

xt+1 = f (xt ) (1.4.1)

where is a parameter. The fixed point x satisfies x = f (x ) .

In order to study the system close to x we write xt = x + ht and expand f in its Taylor series
around x taking only the linear term. Thus:

df
x + ht+1 f (x ) + (x )ht (1.4.2)
dx

which gives

df
ht+1 = (x )ht (1.4.3)
dx

We call (1.4.3) the linearization of (1.4.1). The solution of (1.4.3) is given by (1.1.4). Hence, if
|(df /dx)(x )| < 1 , limt ht = 0 which means that xt will converge towards the fixed point x .

Now, we make the following definitions:

Definition 1.4.1. Let x be a fixed point of equation (1.4.1). If || = |(df /dx)(x)| = 1 then
x is called a hyperbolic fixed point. is called the eigenvalue.

Definition 1.4.2. Let x be a hyperbolic fixed point. If || < 1 then x is called a locally
asymptotic stable hyperbolic fixed point.

Example 1.4.1. Assume that > 1 and find the parameter interval where the fixed point
x = ( 1)/ of the quadratic map is stable.

24
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Solution: f (x) = x(1 x) implies that f  (x) = (1 2x) || = |f  (x )| = |2 | .

Hence from Definition 1.4.2, 1 < < 3 ensures that x is a locally asymptotic stable fixed point
(which is consistent with our finding in the last example in the previous section).

It is clear from Definition 1.4.2 that x is a locally stable fixed point. A formal argument that there exists an
open interval U around x so that whenever |f  (x )| < 1 and x U and that limn f n (x) = x
goes like this:

By the continuity of f ( f is C  ) there exists an > 0 such that |f  (x)| < K < 1 for
x [x , x + ] . Successive use of the mean value theorem then implies

|f n (x) x | = |f n (x) f n (x )| = |f (f n1(x)) f (f n1(x ))|

K|f n1(x) f n1 (x )| K 2 |f n2(x) f n2(x )|

. . . K n |x x | < |x x | <

so f n (x) x as n .

Motivated by the preceding argument we define:

Definition 1.4.3. Let x be a hyperbolic fixed point. We define the local stable and unstable
manifolds of x , Wloc
s
(x ) , Wloc
u
(x ) as

s
Wloc (x ) = {x U | f n (x) x n f n (x) U n 0}
u
Wloc (x ) = {x U | f n (x) x n f n (x) U n 0}

where U is a neighbourhood of the fixed point x .

The definition of a hyperbolic stable fixed point is easily extended to periodic points.

Definition 1.4.4. Let p be a periodic point of (prime) period n so that |f n  (p)| < 1 . Then p
is called an attracting periodic point.

Example 1.4.2. Show that the periodic points 0.5130 and 0.7995 of xt+1 = 3.2xt (1 xt ) are
stable and thereby proving that the difference equation has a stable 2-periodic attractor.

25
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Solution: Since f (x) = 3.2x(1 x) f  (x) = 3.2(1 2x) we have from the chain rule

(1.1.7) that f 2 (0.5130) = f  (0.7995)f  (0.5130) = 0.0615 . Consequently, according to
Definition 1.4.4, the periodic points are stable. 

Exercise 1.4.1. Use formulae (1.3.3) and compute the two-periodic points of the quadratic map
in case of = 3.8 . Is the corresponding two-periodic orbit stable or unstable?

Exercise 1.4.2. When = 3.839 the quadratic map has two 3-cycles. One of the cycles consists
of the points 0.14989, 0.48917 and 0.9593 while the other consists of the points 0.16904, 0.53925
and 0.95384. Show that one of the 3-cycles is stable and that the other one is unstable.

Excellent Economics and Business programmes at:

The perfect start

of a successful,
international career.

CLICK HERE
to discover why both socially
and academically the University
of Groningen is one of the best
places for a student to be
www.rug.nl/feb/education

26 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

An alternative way of investigating the behaviour of a one-dimensional map x f (x) is to use

graphical analysis. The method is illustrated in Figures 2a,b, where we have drawn the graphs of a)
f (x) = 2.7x(1 x) , and b) f (x) = 3.2x(1 x) together with the diagonal(s) y = x . Now,
considering Figure 2a, let x0 (= 0.2) be an initial value. A vertical line from x0 to the diagonal gives the
point (x0 , x0 ) , and if we extend the line to the graph of f we arrive at the point (x0 , f (x0 )) . Then
a horizontal line from the latter point to the diagonal gives the point (f (x0 ), f (x0 )) . Hence, by first
drawing a vertical line from the diagonal to the graph of f and then a horizontal line back to the diagonal
we actually find the image of a point x0 under f on the diagonal. Continuing in this fashion by drawing
line segments vertically from the diagonal to the graph of f and then horizontally from the graph to
the diagonal generate points (x0 , x0 ) , (f (x0 ), f (x0 )) , (f 2 (x0 ), f 2 (x0 )), ..., (f n (x0 ), f n (x0 )) on the
diagonal which is nothing but a geometrical visualization of the orbit of the map x f (x) . Referring to
Figure 2a we clearly see that the orbit converges towards a stable fixed point (cf. Example 1.4.1). On the
other hand, in Figure 2b our graphical analysis shows that the fixed point is a repellor (cf. Exercise 1.3.2),
and if we continue to iterate the map the result is a stable period 2 orbit, which is in accordance with
Example 1.4.2. In Figure 2c all initial transitions have been removed and only the period 2 orbit is plotted.

Exercise 1.4.3. Let x [0, 1] and perform graphical analyses of the maps x 1.8x(1 x),
x 2.5x(1 x) and x 4x(1 x) . In the latter map use both a) x0 = 0.2 , and b) x0 = 0.5
as initial values.




[W






     
[W

 

 

 
[W

[W

 

 

 
           
[W [W

E F

Figure 2: Graphical analyses of a) x 2.7x(1 x) and b), c) x 3.2x(1 x) .

27
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Exercise 1.4.4. Consider the map f : R R x x3 .

a) The map has three fixed points. Find these.

b) Use Definition 1.4.2 and discuss their stability properties.
c) Verify the results in a) and b) by performing a graphical analysis of f .

Let us close this section by discussing the concept structural stability. Roughly speaking, a map f is said
to be structurally stable if a map g which is obtained through a small perturbation of f has essentially
the same dynamics as f , so intuitively this means that the distance between f and g and the distance
between their derivatives should be small.

Definition 1.4.5. The C 1 distance between a map f and another map g is given by

sup(|f (x) g(x)|, |f (x) g  (x)|) (1.4.4)

xR

By use of Definition 1.4.5 we may now define structural stability in the following way:

Definition 1.4.6. The map f is said to be C 1 structurally stable on an interval I if there exists
> 0 such that whenever (1.4.4) < on I , f is topological equivalent to g . 

To prove that a given map is structurally stable may be difficult, especially in higher dimensional systems.
However, our main interest is to focus on cases where a map is not structurally stable. In many respects
maps with nonhyperbolic fixed points are standard examples of such maps as we now will demonstrate.

Example 1.4.3. When = 1 the quadratic map is not structurally stable.

Indeed, consider x f (x) = x(1 x) and the perturbation x g(x) = x(1 x) + .

Obviously, x = 0 is the fixed point of f and since || = |f  (0)| = 1 , x is a nonhyperbolic
fixed point. Moreover, the C 1 distance between f and g is || . Regarding g , the fixed points

are easily found to be x = . Hence, for > 0 there are two fixed points and < 0 gives
no fixed points. Consequently, f is not structurally stable.

Example 1.4.4. When = 3 the quadratic map is not structurally stable.

28
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Let x f (x) = 3x(1 x) and x g(x) = 3x(1 x) + and again we notice that their
C 1 distance is . Regarding f , the fixed points are x1 = 0 and x2 = 2/3 . Further,
|1 | = |f  (0)| = 3 , |2 | = |f  (2/3)| = 1 . Thus x1 is a repelling hyperbolic fixed point

while x2 is nonhyperbolic. Considering g , the fixed points are x1 = (1/3)(1 1 + 3)

and x2 = (1/3)(1 + 1 + 3) . Note that = 0 x1 = x1 , x2 = x2 .) Further,

|1 | = |g (x1 )| = |1 + 2 1 + 3| and |2 | = |g (x2 )| = |1 2 1 + 3| . Whatever the
sign of , x1 is clearly a repelling fixed point (just as x1 ) since 1 > 1 . Regarding x2 it is stable
in case of < 0 and unstable if > 0 .

The equation x = g 2 (x) may be expressed as

27x4 + 54x3 + (18 36)x2 + (8 18)x + 4 32 = 0 (1.4.5)

and since x1 and x2 are solutions of (1.4.5) we may use polynomial division to obtain

9x2 12x 3e + 4 = 0 (1.4.6)

which has the solutions x1,2 = (2/3)(1 3) . Thus there exists a two- periodic orbit in case
of > 0 .

American online
LIGS University
is currently enrolling in the
Interactive Online BBA, MBA, MSc,
DBA and PhD programs:

enroll by September 30th, 2014 and

save up to 16% on the tuition!
pay in 10 installments / 2 years
Interactive Online education
visit www.ligsuniversity.com to
find out more!

Note: LIGS University is not accredited by any

nationally recognized accrediting agency listed
by the US Secretary of Education.
More info here.

29 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps


Moreover, cf. (1.1.7) g 2 = g  (x1 )g  (x2 ) = 9(1 2x1 )(1 2x2 ) = 1 48 which implies
that the two-periodic orbit is stable in case of > 0 , small. Consequently, when > 0 there
is a fundamental structurally difference between f and g so f cannot be structurally stable.
(Note that the problem is the nonhyperbolic fixed point, not the hyperbolic one.)

As suggested by the previous examples a major reason why a map may fail to be structurally stable is
the presence of the nonhyperbolic fixed point. Therefore it is in many respects natural to introduce the
following definition:

Definition 1.4.7. Let x be a hyperbolic fixed point of a map f : R R . If there exists a

neighbourhood U around x and an > 0 such that a map g is C 1 close to f on U
and f is topological equivalent to g whenever (1.4.4) < on this neighbourhood, then f is
said to be C 1 locally structurally stable.

There is a major general result on topological equivalent maps known under the name Hartman and
Grobmans theorem. The one-dimensional formulation of this theorem (cf. Devaney, 1989) is:

Theorem 1.4.1. Let x be a hyperbolic fixed point of a map f : R R and suppose that
= f  (x ) such that || = 0, 1 . Then there is a neighbourhood U around x and a
neighbourhood V of 0 R and a homeomorphism h : U R which conjugates f on U to
the linear map l(x) = x on V .

For a proof, cf. Hartman (1964).

Example 1.4.5. Consider x f (x) = (5/2)x(1 x) . The fixed point is x = 3/5 and is
clearly hyperbolic since = f  (x ) = 1/2 . Therefore, according to Theorem 1.4.1, f (x) on
a neighbourhood about 3/5 is topological equivalent to l(x) = (1/2)x on a neighbourhood
about 0.

1.5 Bifurcations
As we have seen, the map x f (x) = x(1 x) has a stable hyperbolic fixed point x = ( 1)/
provided 1 < < 3 . If = 3 , = f  (x ) = 1 , hence x is no longer hyperbolic. If = 3.2 we
have shown that there exists a stable 2- periodic orbit. Thus x experiences a fundamental change of
structure when it fails to be hyperbolic which in our running example occurs when = 3 . Such a point
will from now on be referred to as a bifurcation point. When = 1 , as in our example, the bifurcation
is called a flip or a period doubling bifurcation. If = 1 it is called a saddle-node bifurcation. Generally,
we will refer to a flip bifurcation as supercritical if the eigenvalue crosses the value 1 outwards
and that the 2-periodic orbit just beyond the bifurcation point is stable. Otherwise the bifurcation is
classified as subcritical.

30
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Theorem 1.5.1. Let f : R R , x f (x) be a one-parameter family of maps and assume

that there is a fixed point (x , 0 ) where the eigenvalue equals 1 . Assume

f 2 f 2 f f 2 f
    2
f f
a= + 2 = 1 = 0 at (x , 0)
x2 x x2 x x

and
  2 
1 2 f 1 3 f

b= + = 0 at (x , 0 )
2 x2 3 x3

Then there is a smooth curve of fixed points of f which is passing through (x , 0 ) and which
changes stability at (x , 0 ) . There is also a curve consisting of hyperbolic period-2 points
passing through (x , 0 ) . If b > 0 the hyperbolic period-2 points are stable, i.e. the bifurcation
is supercritical.

Proof. Through a coordinate transformation it suffices to consider f so that for = 0 = 0

we have f (x , 0) = x and f  (x , 0) = 1 .

First we show that one without loss of generality may assume that x = 0 in some neighbourhood
of = 0 . To this end, define F (x, ) = f (x, ) x . Then F  (x , ) = 2 = 0 and by
use of the implicit function theorem there exists a solution x() of F (x, ) = 0 . Next, define
g(y, ) = f (y + x(), ) x() . Clearly, g(0, ) = 0 for all . Consequently, y = 0
is a fixed point so in the following it suffices to consider x f (x) where x () = 0 and
f  (0, 0) = 1 .

The Taylor expansion around (x , ) = (0, 0) of the latter map is

1 2f 2 2f 1 3f 3
 
f
g () = + + + 2 + +
2 x2 x 6 x3
= + + 2 + c + d 3 +

where the parameter has the same weight as 2 . The composite (g g)() may be
expressed as

g2 () = + + 3 +

Thus, in order to have a system to study we must assume , = 0 which is equivalent to

2f f 1 2 f
 
= (2c + 2ab) = 2 +2  0
=
x 2 x2
 2 
3 2

1 f 1 f
= (2d + 2b2 ) = 2 +2 = 0
6 x3 2 x2

31
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Figure 3: The possible configurations of 2 h() = + + 3

and we recognize the derivative formulaes as nothing but what is stated in the theorem.

32 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Next, consider the truncated map

2 h() = + + 3

Clearly, the fixed points are


1 = 0 , 2,3 =

Further, h () = 1 + + 3 2 so h ( 1 ) = 1 + and h ( 2,3 ) = 1 2 . Thus we have

the following configurations (see Figure 3), and we may conclude that the stable period-2 orbits
corresponds to < 0 , i.e.

2
2f 1 3f

1
+ >0
2 x2 3 x3

Example 1.5.1. Show that the fixed point of the quadratic map undergoes a supercritical flip
bifurcation at the threshold = 3 .

Solution: From the previous section we know that x = 2/3 and f  (x ) = 1 when = 3 .
We must show that the quantities a and b in Theorem 1.5.1 are different from zero and larger
than zero respectively. By computing the various derivatives at (x , 0 ) = (2/3, 3) we obtain:
 
2 1 1 1
a = (6) + 2 = 2 = 0 and b = (6)2 + 0 = 18 > 0
9 3 2 3

Thus the flip bifurcation is supercritical. When x fails to be stable, a stable period-2 orbit is
established. 

Exercise 1.5.1. Show that the Ricker map x x exp[r(1 x)] , cf. (1.2.2), undergoes a
supercritical flip bifurcation at (x , r) = (1, 2) .

Exercise 1.5.2. Consider the two parameter family of maps x (1 )x x2 + x3 . Show

that the map may undergo both a sub- and supercritical flip bifurcation.

As is clear from Definition 1.4.1 a fixed point will also lose its hyperbolicity if the eigenvalue equals
1. The general case then is that x will undergo a saddle-node bifurcation at the threshold where
hyperbolicity fails. We shall now describe the saddle-node bifurcation.

33
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Consider the map

x f (x) = x + x2 (1.5.1)

whose fixed points are x1,2 = . Hence, when > 0 there are two fixed points
which equals when = 0 . If < 0 there are no fixed points. In case of > 0 , small,

we have f (x1 = ) = 1 2 < 1 , hence x1 = is stable. On the other hand:

f (x2 = ) = 1 + 2 > 1 , consequently x2 is unstable. Thus a saddle-node bifurcation is
characterized by that there is no fixed point when the parameter falls below a certain threshold 0 .
When is increased to 0 , = 1 , and two branches of fixed points are born, one stable and one unstable
as displayed in the bifurcation diagram, see Figure 4a.

Figure 4: (a) The bifurcation diagram (saddle node) for the map x x+ x2
(b) The bifurcation diagram (transcritical) for the map x x(1 x)

The other possibilities at = 1 are the pitchfork and the transcritical bifurcations. The various
configurations for the pitchfork are given at the end of the proof of Theorem 1.5.1 (see Figure 3). A
typical configuration in the transcritical case is shown in Figure 4b as a result of considering the quadratic
map at (x , 0 ) = (0, 1) .

Exercise 1.5.3. Do the necessary calculations which leads to Figure 4b.

Exercise 1.5.4.

a) Show that the map x x2 undergoes a supercritical flip bifurcation at

(x , 0 ) = (1/2, 3/4) .

b) Perform a graphical analysis of the map in the cases = 1/2 and = 1 . 

Exercise 1.5.5. Find possible bifurcation points of the map x + x2 . If you detect a flip
bifurcation decide whether it is sub- or supercritical. 

Exercise 1.5.6. Analyze the map x x x3 .

34
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

1.6 The flip bifurcation sequence

We shall now return to the flip bifurcation. First we consider the quadratic map. In the previous section
we used Theorem 1.5.1 to prove that the quadratic map x x(1 x) undergoes a supercritical flip
bifurcation at the threshold = 0 = 3 . This means that in case of > 0 , | 0 | small, there
exists a stable 2-periodic orbit and according to our findings in Section 1.3 the periodic points are given
by (1.3.3), namely

+1 ( + 1)( 3)
p1,2 =
2

The period 2 orbit will remain stable as long as


|f (p1 )f  (p2 )| < 1

cf. Section 1.4. Thus, in our example,

|(1 2p1 )(1 2p2 )| < 1

i.e.

|1 ( + 1)( 3)| < 1 (1.6.1)

Join the best at Top masters programmes

3
 3rd place Financial Times worldwide ranking: MSc
the Maastricht University International Business
1st place: MSc International Business
School of Business and 1st place: MSc Financial Economics
2nd place: MSc Management of Learning

Economics! 2nd place: MSc Economics

2nd place: MSc Econometrics and Operations Research
2nd place: MSc Global Supply Chain Management and
Change
Sources: Keuzegids Master ranking 2013; Elsevier Beste Studies ranking 2012;
Financial Times Global Masters in Management ranking 2012

Maastricht
University is
the best specialist
university in the
Visit us and find out why we are the best! Netherlands
(Elsevier)
Masters Open Day: 22 February 2014

www.mastersopenday.nl

35 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

from which we conclude that the 2-periodic orbit is stable as long as

3 < < 1 + 6 (1.6.2)



Since = f 2 = f  (p1 )f  (p2 ) = 1 when 1 = 1 + 6 there is a new flip bifurcation taking place at
1 which in turn leads to a 4-periodic orbit. We also notice that while the fixed point x = ( 1)/
is stable in the open interval I = (2, 3) , the length of the interval where the 2-periodic orbit is stable is
roughly (1/2)I . In Figure 5a we show the graphs of the quadratic map in the cases = 2.7 (curve a)
and = 3.4 (curve b) respectively, together with the straight line xt+1 = xt . = 2.7 gives a stable
fixed point x while = 3.4 gives an unstable fixed point. These facts are emphasized in the figure by
drawing the slopes (indicated by dashed lines). The steepness of the slope at the fixed point of curve a
is less than 45 , || < 1 , while < 1 at the unstable fixed point located on curve b. In general,
if f (x) is a single hump function (just as the quadratic map displayed in Figure 5a) the second iterate
f2 (x) will be a two-hump function. In Figures 5b and 5c we show the relation between xt+2 and xt .
Figure 5b corresponds to = 2.7 , Figure 5c corresponds to = 3.4 . Regarding 5b the steepness of
the slope is still less than 45 so the fixed point is stable. However, in 5c the slope at the fixed point is
steeper than 45 , the fixed point is unstable and we see two new solutions of period 2.

Figure 5: (a) The quadratic map in the cases = 2.7 and = 3.4. (b) and (c)
The second iterate of the quadratic map in the cases = 2.7 and = 3.4, respectively.

Let us now explore this mechanism analytically: Suppose that we have an n -periodic orbit consisting
of the points p0 , p1 . . . pn1 such that

pi = fn (pi ) (1.6.3)

36
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Then by the chain rule (cf. (1.1.7))

n1

fn  (p0 ) = f (pi ) = n (p0 ) (1.6.4)
i=0

Hence, if |n (p0 )| < 1 the n -periodic orbit is stable, if |n (p0 )| > 1 the orbit is unstable.

Next, consider the 2n -periodic orbit

p
i
= f2n (pi ) = fn (fn (pi ))

By appealing once more to the chain rule we obtain

n1 2



f2n (p0 ) = f (pi ) = 2n (p0 ) (1.6.5)
i=0

This allows us to conclude that if the n -point cycle is stable (i.e. |n | < 1 ) then 2n < 1 too. On
the other hand, when the n -cycle becomes unstable (i.e. |n | > 1 ) then 2n > 1 too. So what this
argument shows is that when a periodic point of prime period n becomes unstable it bifurcates into
two new points which are initially stable points of period 2 n and obviously there are 2n such points.
This is the situation displayed in Figure 5c. So what the argument presented above really says is that as
the parameter of the map x f (x) is increased periodic orbits of period 2, 22 , 23 , . . . and so on
are created through successive flip bifurcations. This is often referred to as the flip bifurcation sequence.
Initially, all the 2k cycles are stable but they become unstable as is further increased.

As already mentioned, if f (x) is a single-hump function, then f2 (x) is a two-hump function. In the
same way, f3 (x) is a four-hump function and in general fp will have 2p1 humps. This means that the
parameter range where the period 2p cycles are stable shrinks through further increase of . Indeed,
the values at successive bifurcation points act more or less as terms in a geometric series. In fact,
Feigenbaum (1978) demonstrated the existence of a universal constant (known as the Feigenbaum
number or the Feigenbaum geometric ratio) such that

n+1 n
lim = = 4.66920 (1.6.6)
n n+2 n+1

where n , n+1 and n+2 are the parameter values at three consecutive flip bifurcations. From this
we may conclude that there must exist an accumulation value a where the series of flip bifurcations
converge. (Geometrically, this may happen as a valley of some iterate of f deepens and eventually
touches the 45 line (cf. Figure 5c), then a saddle-node bifurcation ( = 1 ) will occur.)

37
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

As is true for our running example x x(1 x) we have proved that the first flip bifurcation occurs

at = 3 and the second at = 1 + 6 . The point of accumulation for the flip bifurcations a is
found to be a = 3.56994 .

Exercise 1.6.1. Identify numerically the flip bifurcation sequence for the Ricker map (1.2.2).

In the next sections we will describe the dynamics beyond the point of accumulation a for the flip
bifurcations.

1.7 Period 3 implies chaos. Sarkovskiis theorem

Referring to our running example (1.2.1), x x(1 x) we found in the previous section that the
point of accumulation for the flip bifurcation sequence a 3.56994 . We urge the reader to use a
computer or a calculator to identify numerically some of the findings presented below. [a , 4] .

38 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

When > a , a small, there are periodic orbits of even period as well as aperiodic orbits.
Regarding the periodic orbits, the periods may be very large, sometimes several thousands which make
them indistinguishable from aperiodic orbits. Through further increase of odd period cycles are
detected too. The first odd cycle is established at = 3.6786 . At first these cycles have long periods
but eventually a cycle of period 3 appears. In case of (1.2.1) the period-3 cycle occurs for the first time
at = 3.8284 . This is displayed in Figure 6. (The point marked with a cross is the initially fixed point
x = ( 1)/ which became unstable at = 3 . It is also clear from the figure that the 3-cycle is
established as the third iterate of (1.2.1) undergoes a saddle-node bifurcation.

Figure 6: A 3-cycle generated by the quadratic map.

An excellent way in order to present the dynamics of a map is to draw a bifurcation diagram. In such a
diagram one plots the asymptotic behaviour of the map as a function of the bifurcation parameter. If we
consider the quadratic map one plots the asymptotic behaviour as a function of . If a map contains several
parameters we fix all of them except one and use it as bifurcation parameter. In somewhat more detail a
bifurcation diagram is generated in the following way: (A) Let be the bifurcation parameter. Specify
consecutive parameter values 1 , 2 , ..., n where the distance |i i+1 | should be very small. (B)
Starting with 1 , iterate the map from an initial condition x0 until the orbit of the map is close to the
attractor and then remove initial transients. (C) Proceed the iteration and save many points of the orbit
on the attractor. (D) Plot the orbit over 1 in the diagram. (E) Repeat the procedure for 2 , 3 , ..., n .

Now, if the attractor is an equilibrium point for a given bifurcation value i there will be one point only
over i in the bifurcation diagram. If the attractor is a two-period orbit there will be two points over
i , and if the attractor is a k period orbit there are k points over i . Later on we shall see that an
attractor may be an invariant curve as well as being chaotic. On such attractors there are quasiperiodic
orbits and if either of these two types of attractors exist we will recognize them as line segments provided
a sufficiently number of iteration points. The same is also true for periodic orbits when the period k
becomes large. (In this context one may in fact think of quasiperiodic and chaotic orbits as periodic
orbits where k .) Hence, it may be a hopeless task to distinguish these types of attractors from
each other by use of the bifurcation diagram alone.

39
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Figure 7: The bifurcation diagram of the quadratic map in the parameter range 2.9 4

In the bifurcation diagram, Figure 7, we display the dynamics of the quadratic map in the interval
2.9 4 . The stable fixed point ( < 3 ) as well as the flip bifurcation sequence is clearly identified.
Also the period-3 window is clearly visible. Our goal in this and in the next sections is to give a
thorough description of the dynamics beyond a .

We start by presenting the Li and Yorke theorem (Li and Yorke, 1975).

Theorem 1.7.1. Let f : R R , x f (x) be continuous. Suppose that f has a periodic

point of period 3. Then f has periodic points of all other periods. 

Remark 1.7.1: Theorem 1.7.1 was first proved in 1975 by Li and Yorke under the title Period
three implies chaos. Since there is no unique definition of the concept chaos many authors today
prefer to use the concept Li and Yorke chaos when they refer to Theorem 1.7.1. The essence of
Theorem 1.7.1 is that once a period-3 orbit is established it implies periodic orbits of all other
periods. Note, however, that Theorem 1.7.1 does not address the question of stability. We shall
deal with that in the next section. 

We will now prove Theorem 1.7.1. Our proof is based upon the proof in Devaney (1989), not so much
upon the original proof by Li and Yorke (1975).

40
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Proof. First, note that (1): If I and J are two compact intervals so that I J and J f (I)
then f has a fixed point in I . (2): Suppose that A0 , A1 , . . . , An are closed intervals and that
Ai+1 f (Ai ) for i = 0, . . . , n 1 . Then there is at least one subinterval J0 of A0 which
is mapped onto A1 . There is also a similar subinterval in A1 which is mapped onto A2 so
consequently there is a J1 J0 so that f (J1 ) A1 and f2 (J1 ) A2 . Continuing is this
fashion we find a nested sequence of intervals which map into the various Ai in order. Therefore
there exists a point x A0 such that fi (x) Ai for each i . We say that f (Ai ) covers Ai+1 .

Now, let a , b and c R and suppose f (a) = b , f (b) = c and f (c) = a . We further
assume that a < b < c . Let I0 = [a, b] and I1 = [b, c] , cf. Figure 6. Then from our assumptions
I1 f (I0 ) and I0 I1 f (I1 ) . The graph of f , cf. Figure 6, shows that there must be a
fixed point of f between b and c . Similarly, f2 must have fixed points between a and b
and at least one of them must have period 2. Therefore we let n 2 . Our goal is to produce
a periodic point of prime period n > 3 . Inductively, we define a nested sequence of intervals
A0 , A1 , . . . , An2 I1 as follows. Let A0 = I1 . Since I1 f (I1 ) there is a subinterval
A1 A0 such that f (A1 ) = A0 = I1 . Then there is also a subinterval A2 A1 such that
f (A2 ) = A1 which implies f2 (A2 ) = f (f (A2 )) = f (A1 ) = A0 = I1 . Continuing in
this way there exists An2 An3 such that f (An2 ) = f (An3 ) so according to (2), if
x An2 then f (x), f2 (x), . . . , fn1 (x) A0 and indeed fn2 (An2 ) = A0 = I1 .

Need help with your

dissertation?
Get in-depth feedback & advice from experts in your
topic area. Find out what you can do to improve
the quality of your dissertation!

Get Help Now

Go to www.helpmyassignment.co.uk for more info

41 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Now, since I0 f (I1 ) there exists a subinterval An1 An2 such that fn1 (An1 ) = I0 .
Finally, since I1 f (I0 ) we have I1 fn (An1 ) so that fn (An1 ) covers An1 . Therefore,
according to (1) fn has a fixed point p in An1 .

Finally, we claim that p has prime period n . Indeed, the first n 2 interations of p is in I1 , the
(n 1) st lies in I0 and the n -th is p again. If fn1 (p) lies in the interior of I0 it follows that p
has prime period n . If fn1 (p) lies on the boundary, then n = 2 or 3 and again we are done. 

Theorem 1.7.1 is a special case of Sarkovskiis theorem which came in 1964. However, it was written in
Russian and published in an Ukrainian mathematical journal so it was not discovered and recognized
in Western Europe and the U.S. prior to the work of Li and Yorke. We now state Sarkovskiis theorem:

Theorem 1.7.2. We order the positive integers as follows:

1  2  22  . . .  2m  2k (2n + 1)  . . .  2k 3  . . . 2 3  2n 1  . . .  9  7  5  3

Let f : I I be a continuous map of the compact interval I into itself. If f has a periodic
point of prime period p , then it also has periodic points for any prime period q  p . 

Proof. Cf. Devaney (1989) or Katok and Hasselblatt (1995). 

Clearly, Theorem 1.7.1 is a special case of Theorem 1.7.2. Also note that the first part in the Sarkovskii
ordering (1  2  22 . . .  2m ) corresponds to the flip bifurcation sequence as demonstrated through
our treatment of the quadratic map. As the parameter in (1.2.1) is increased beyond the point of
accumulation for the flip bifurcations. Sarkovskiis theorem says that we approach a situation where there
are an infinite number of periodic orbits.

1.8 The Schwarzian derivative

In the previous section we established through Theorems 1.7.1 and 1.7.2 that a map may have an infinite
number of periodic orbits. Our goal in this section is to prove that in fact only a few of them are attracting
(or stable) periodic orbits.

Definition 1.8.1. Let f : I I be a C 3 function. The Schwarzian derivative Sf of f is defined

as
2
f  (x) 3 f  (x)

Sf (x) =  (1.8.1)
f (x) 2 f (x)

42
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Considering f (x) = x(1 x) we easily find that Sf (x) = 6/(1 2x)2 . Note that Sf < 0
everywhere except at the critical point c = 1/2 . (However, we may define Sf (1/2) = .)

The main result in this section is the following theorem which is due to Singer (1978):

Theorem 1.8.1. Lef f be a C 3 function with negative Schwarzian derivative. Suppose that f has
one critical point c . Then f has at most three attracting periodic orbits. 

Proof. The proof consists of three steps.

(1) First we prove that if f has negative Schwarzian derivative then all f n iterates also have
negative Schwarzian derivatives.

To this end, assume Sf < 0 and Sg < 0 . Our goal is to show that S(f g) < 0 . Successive
use of the chain rule gives:

(f g)(x) = f  (g(x))g (x)

(f g)(x) = f  (g(x))(g (x))2 + f  (g(x))g (x)

(f g) (x) = f  (g(x))(g (x))3 + 3f  (g(x))g (x)g  (x) + f  (g(x))g  (x)

Then (omitting function arguments) Definition 1.8.1 gives

2
f  g 3 + 3f  g  g  + f  g  3 f  g 2 + f  g 

S(f g) =
f g 2 f g

which after some rearrangements may be written as

 2  2
f  3 f  g  3 g 
 
2
g +  = Sf (g(x))(g (x))2 + Sg(x)
f 2 f g 2 g

Thus S(f g)(x) < 0 which again implies Sf n < 0 .

(2) Next we show that if Sf < 0 then f  (x) cannot have a positive local minimum.

43
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

To this end, assume that d is a critical point of f  (x) . Then f  (d) = 0 , and since Sf < 0
it follows from Definition 1.8.1 that f  /f  < 0 so f  (d) and f  (d) have opposite signs.
Graphically, it is then obvious that f  (x) cannot have a positive local minimum, and in the same
way it is also clear that f  (x) cannot have a negative local maximum. Consequently, between any
two consecutive critical points d1 and d2 of f  there must be a critical point c of f such that
f  (c) = 0 and moreover, (1) and (2) together imply that between any two consecutive critical

points of f n there must be a critical point of f n .

(3) By considering f n  (x) = 0 it follows directly from the chain rule that if f (x) has a critical
point then f n (x) will have a critical point too. Finally, let p be a point of period k on the
attracting orbit and let I = (a, b) be the largest open interval around p where all points approach
p asymptotically. Then f (I) I and f k (I) I . Regarding the end points a and b we have:
If f (a) = f (b) then of course there exists a critical point. If f (a) = a and f (b) = b (i.e. that
the end points are fixed points) it is easy to see graphically that there exist points u and v such
that a < u < p < v < b with properties f  (u) = f  (v) = 1 . Then from (2) and the fact that
f  (p) < 1 there must be a critical point in (u, v) . In the last case f (a) = b and f (b) = a we
arrive at the same conclusion by considering the second iterate f 2 . Thus in the neighbourhood
of any stable periodic point there must be either a pre-image of a critical point or an end point
of the interval and we are done. 

Brain power By 2020, wind could provide one-tenth of our planets

electricity needs. Already today, SKFs innovative know-
how is crucial to running a large proportion of the
worlds wind turbines.
Up to 25 % of the generating costs relate to mainte-
nance. These can be reduced dramatically thanks to our
systems for on-line condition monitoring and automatic
lubrication. We help make it more economical to create
cleaner, cheaper energy out of thin air.
By sharing our experience, expertise, and creativity,
industries can boost performance beyond expectations.
Therefore we need the best employees who can
meet this challenge!

The Power of Knowledge Engineering

Plug into The Power of Knowledge Engineering.

Visit us at www.skf.com/knowledge

44 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Example 1.8.1. Assume x [0, 1] and let us apply Theorem 1.8.1 on the quadratic map
x f (x) = x(1 x) . For a fixed (1, 3) the fixed point x = ( 1)/ is stable,
and since f (0) = f (1) = 0 and the fact that 0 is repelling there is one periodic attractor,
namely the period-1 attractor x which attracts the critical point c = 1/2 .

When [3, 4] both x and 0 are unstable fixed points. Thus according to Theorem 1.8.1 there
is at most one attracting periodic orbit in this case. (Prior to a there is exactly one periodic
attractor.) When = 4 the critical point is mapped on the origin through two iterations so there
are no attracting periodic orbits in the case. 

Example 1.8.2. Let us close this section by giving an example which shows that Theorem 1.8.1
fails if the Schwarzian derivative is not negative. The following example is due to Singer (1978).
Consider the map

x g(x) = 13.30x4 + 28.75x3 23.31x2 + 7.86x (1.8.2)

The map has one fixed point x = 0.7263986 , and by considering g 2 (x) = x there is also one
2-periodic orbit which consists of the points p1 = 0.3217591 and p2 = 0.9309168 .

Moreover: 1 = g  (x ) = 0.8854 and = g  (p1 )g  (p2 ) = 0.06236 . Thus both the fixed
point and the 2-periodic orbit are attracting.

The critical point of g is c = 0.3239799 and is attracted to the period-2 orbit so it does not belong
to Wloc
s
(x ) , cf. Definition 1.4.3. The reason that x is not attracting c is that Sg(x ) = 8.56 > 0
thus the assumption Sg(x) < 0 in Theorem 1.8.1 is violated. 

Exercise 1.8.1. Compute the Schwarzian derivative when f (x) = xn . 

Exercise 1.8.2. Show that Sf (x) < 0 when f is given by (1.2.2) (the Ricker case). 

1.9 Symbolic dynamics I

Up to this point we have mainly been concerned with fixed points and periodic orbits. The main goal
of this section is to introduce a useful tool called symbolic dynamics which will help us to describe
and understand dynamics of other types than we have discussed previously. To be more concrete, we

shall in this section analyse the quadratic map x x(1 x) where > 2 + 5 on the interval
I = [0, 1] , and as it will become clear, although almost all points in I eventually will escape I , there
exists an invariant set of points which will remain in I . We shall use symbolic dynamics to describe
the behaviour of these points.

45
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

First we need some definitions. Consider x f (x) . Suppose that f (x) can take its values on two
disconnected intervals I1 and I2 only. Define an infinite forward-going sequence of 0s and 1s {ak }
k=0
so that

ak = 0 if f k (x0 ) I1 (1.9.1a)

ak = 1 if f k (x0 ) I2 (1.9.1b)

Thus what we really do here is to represent an orbit of a map by an infinite sequence of 0s and 1s.

Definition 1.9.1.

2 = {a = (a0 a1 a2 . . .)|ak = 0 or 1} (1.9.2)

We shall refer to 2 as the sequence space.

Definition 1.9.2. The itinerary of x is a sequence (x) = a0 a1 . . . where ak is given by (1.9.1).

We now define one of the cornerstones of the theory of symbolic dynamics.

Definition 1.9.3. The shift map : 2 2 is given by

(a0 a1 a2 a3 . . .) = a1 a2 a3 . . . (1.9.3

Hence the shift map deletes the first entry in a sequence and moves all the other entries one place to
the left.

Example 1.9.1. a = (1111 . . .) represents a fixed point under since (a) = n (a) = (111 . . .) .
Suppose a = (001, 001, 001, . . .) . Then (a) = (010, 010, 010, . . .) , 2 (a) = (100, 100, 100, . . .)
2 (a) = (100, 100, 100, . . .) and 3 (a) = (001, 001, 001, . . .) = a . Thus a = (001, 001, 001, . . .) represents a
periodic point of period 3 under the shift map. 

The previous example may obviously be generalized. Indeed, if a = (a0 a1 . . . an1 , a0 a1 . . . an1 , . . .)
there are 2n periodic points of period n under the shift map since each entry in the sequence may have
two entries 0 or 1.

Definition 1.9.4. Let U be a subset of a set S . U is dense in S if the closure U = S . 

46
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Definition 1.9.5. If a set S is closed, contains no intervals and no isolated points it is called a
Cantor set. 

Proposition 1.9.1. The set of all periodic orbits Per () = 2n is dense in 2 . 

Figure 8: The quadratic map in the case > 2 + 5 Note the subintervals I1 and I2 where f (x) = x(1 x) 1

Proof. Let a = (a0 a1 a2 . . .) be in 2 and suppose that b = (a0 . . . an1 , a0 . . . an1 . . .)

represent the 2n periodic points. Our goal is to prove that b converges to a . By use of the
usual distance function in a sequence space, d[a, b] = (|ai bi |/2i ) we easily find that
d[a, b] < 1/2n . Hence b a . 

47 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

We now have the necessary machinery we need in order to analyse the quadratic map in case of

>2+ 5.

Let x f (x) = x(1 x) where > 2 + 5 . From the equation x(1 x) = 1 we

find x = 1/2 + 1/2 1 4 . Hence in the intervals I1 = [0, 1/2 1/2 1 4] and

I2 = [1/2 + 1/2 1 4] , f 1 , cf. Figure 8. Moreover, |f  (x)| = | 2x| and whenever

> 2 + 5 we find that |f  (x)| > 1 .

Denote I = [0, 1] . Then I f 1 (I) = I1 I2 so if x I (I f 1 (I)) we have f > 1 (cf.

Figure 8) which implies f 2 < 0 and consequently f n . All the other points will remain in I
after one iteration. The second observation is that f (I1 ) = f (I2 ) = I so there must be a pair of open
intervals, one in I1 and one in I2 , which is mapped into I (I1 I2 ) such that all points in these two
intervals will leave I after two iterations. Continuing in this way by removing pairs of open intervals
(i.e. first the interval I (I1 I2 ) , then two intervals, one in I1 ( J1 ) and one in I2 ( J2 ), then 22 open
intervals, two from I1 J1 , two from I2 J2 . . . and finally 2n intervals) from closed intervals we are
left with a closed set which is I minus the union of all the 2n+1 1 open sets. Hence consists of
the points that remain in I after n iterations, I f 1 (I) and consists of 2n+1 closed intervals.

Now, associate to each x a symbol sequence {ai }

i=1 of 0s and 1s such that ak = 0 if f (x) I1
k

and ak = 1 if f k (x) I2 .

Next, define

Ia0 ...an = {x I/x Ia0 , f (x) Ia1 . . . f n (x) Ian } (1.9.4)

as one of the 2n+1 closed subintervals in . Our first goal is to show that Ia0 ...an is non-empty when
n . Indeed,

Ia0 ...an = Ia0 f 1 (Ia1 ) . . . f n (Ian )

(1.9.5)
= Ia0 f 1 (Ia1 ...an )

Ia1 is nonempty. Then by induction Ia1 ...an is non-empty, and moreover, since f 1 (Ia1 ...an ) consists
of two closed subintervals it follows that Ia0 f 1 (Ia1 ...an ) consists of one closed interval. A final
observation is that

Ia0 ...an = Ia0 . . . f (n1) (Ian1 ) f n (Ian )

= Ia0 ...an1 f n (Ian ) Ia0 ...an1

Consequently, Ia0 ...an is non-empty. Clearly the length of all sets Ia0 ...an approaches zero as n
which allows us to conclude that the itinerary (x) = a0 a1 . . . is unique.

48
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

We now proceed by showing that is a Cantor set. Assume that contains an interval [a, b] where
a = b . For x [a, b] we have |f  (x)| > > 1 and by the chain rule |f n  x)| > n . Let n be so
large that n |b a| > 1 . Then from the mean value theorem |f n (b) f n (a)| n |b a| > 1
which means that f n (b) or f n (a) (or both) are located outside I . This is of course a contradiction so
contains no intervals.

To see that contains no isolated points it suffices to note that any end point of the 2n+1 1 open
intervals eventually goes to 0 and since 0 these end points are in too. Now, if y is isolated
all points in a neighbourhood of y eventually will leave I which means that they must be elements
of one of the 2n+1 1 open sets which are removed from I . Therefore, the only possibility such that
y is that there is a sequence of end points converging towards y so y cannot be isolated.

From the discussion above we conclude that the quadratic map where > 2 + 5 possesses an invariant
set , a Cantor set, of points that never leave I under iteration. is a repelling set. Our final goal is to
show that the shift map defined on 2 is topological equivalent to f defined on .

Let f : , f (x) = x(1 x) , : 2 2 , (a0 a1 a2 . . .) = a1 a2 . . . and : 2 ,

(x) = a0 a1 a2 . . . . We want to prove that f = .

Observe that

(x) = a0 a1 a2 . . . = Ia0 a1 a2 ...an ...
n0

Further

Ia0 a1 ...an = Ia0 f 1 (Ia1 ) . . . f n (Ian )

so

f (Ia0 a1 ...an ) = f (Ia0 ) (Ia1 ) . . . f n+1 (Ian ) = Ia1 . . . f n+1 (Ian ) = Ia1 ...an

This implies that

    
 
(f (x)) = f Ia0 ...an = Ia1 ...an = ((x))
n0 n1

Thus, f and are topological equivalent maps.

49
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

1.10 Symbolic dynamics II

In Section 1.8 we proved that if a map f : I I with negative Schwarzian derivative possessed an
attracting periodic orbit then there was a trajectory from the critical point c to the periodic orbit. Our
goal here is to extend the theory of symbolic dynamics by assigning a symbol sequence to c or more
precisely to f (c) . We will assume that f is unimodal. The theory will mainly be applied on periodic
orbits.

Note, however, that the purpose of this section is somewhat different than the others so readers who are
not too interested in symbolic dynamics may skip this section and proceed directly to the next where
chaos is treated.

Definition 1.10.1. Let x I . Define the itinerary of x as (x) = a0 a1 a2 . . . where

0 f j (x) < c
aj = 1 f j (x) > c (1.10.1)
C f j (x) = c

What is new here really is that we associate a symbol C to the critical point c . Also note that we may
define two intervals I0 = [0, c and I1 = c, 1] such that f is increasing on I0 and decreasing on I1 .

50 Click on the ad to read more

Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Definition 1.10.2. The kneading sequence is defined as the itinerary of f (c) , i.e.

K(f ) = (f (c)) (1.10.2)

Example 1.10.1.

1) Suppose that x f (x) = 2x(1 x) . Then c = 1/2 and f (c) = 1/2 , f 2 (c) = 1/2 . . . f j (c) = 1/2
f 2 (c) = 1/2 . . . f j (c) = 1/2 so the kneading sequence becomes K(f ) = (C C C C . . .) which also may
be written as (C C C . . .) where the bar refers to repetition.

2) Suppose that x f (x) = 4x(1 x) . c = 1/2 , f (c) = 1 , f 2 (c) . . . = f j (c) = 0 so

K(f ) = (1 0 0 0 . . .) . 

An unimodal map may of course have several itineraries.

Example 1.10.2. By use of a calculator we easily find that the possible itineraries of x 2x(1 x)
are

(0 0 . . . 0 C C C . . .) (C C C . . .) (1 0 . . . 0 C C C . . .) (0 0 0 . . .) (1 0 0 0 . . .)

(The last two itineraries correspond to the orbits of x0 = 0 and x0 = 1 respectively. Note that
the critical point is the same as the stable fixed point x in this example.

In case of x 3x(1 x) we obtain the sequences

(0 0 . . . 0 1 1 1 . . .) (C 1 1 1 . . .)(1 1 1 . . .)

(1 0 . . . 0 1 1 1 . . .) (0 0 0 . . .) (1 0 0 0 . . .)

(0 C 1 1 1 . . .) (1 C 1 1 1 . . .)

where the last two itineraries correspond to the orbits of x0 = (1/6)(3 3) and

x0 = (1/6)(3 + 3) respectively. 

The reader should also have in mind that periodic orbits with different periods may share the same
itinerary.

51
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Indeed, consider x 3.1x(1 x) . Then x = 0.6774 > c = 1/2 so the itinerary of the fixed point
becomes (x ) = (1 1 1 . . .) . However, there is also a two-periodic orbit whose periodic points are (cf.
formulae (1.3.3)) p1 = 0.7645 , p2 = 0.5581 . Again we observe that pi > c so the itinerary of any of
the two-periodic points is also (1 1 1 . . .) . (When becomes larger than 3.1 one of the periodic points
eventually will become smaller than c which results in the itinerary (1 0 1 0 1 0 . . .) or (0 1 0 1 0 1 . . .) .)

Our next goal is to establish an ordering principle of the possible itineraries of a given map. Let
a = (a0 a1 a2 . . .) and b = (b0 b1 b2 . . .) . If ai = bi for 0 i < n and an = bn we say that the
sequences have discrepancy n . Let Sn (a) be the number of 1s among a0 a1 . . . an and assume
0 < C < 1.

Definition 1.10.3. Suppose that a and b have discrepancy n . We say that a b if

Sn1 (a) an < bn (1.10.3a)

Sn1 (a) an > bn (1.10.3b)

Example 1.10.3. Due to a) we have the following order:

(1 1 0 . . .) (1 1 C . . .) (1 1 1 . . .)

Due to b) we have

(1 1 0 . . .) (1 0 1 . . .) (1 0 0 . . .)

Also note that any two sequences with discrepancy 0 are ordered such that the sequence which has 0 as
the first entry is of lower order than the one with C or 1 as the first entry. Thus:

(0 1 . . .) (C 1 . . .) (1 1 . . .)

Exercise 1.10.1. Let a = (0 1 1 0 1 1 . . .) be a repeating sequence. Compute (a) and 2 (a)

and verify the ordering a (a) 2 (a) . 

The following theorem (due to Milner and Thurston) relates the ordering of two symbol sequences to
the values of two points in an interval.

52
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Theorem 1.10.1. Let x, y I

(x) (y) x<y

x<y (x)  (y)

Proof. Suppose that (x) = (a0 a1 a2 . . .) and (y) = (b0 b1 b2 . . .) and let n be the discrepancy
of (x) and (y) . First, suppose n = 0 . Then x < y since 0 < C < 1 . Next, suppose that a) is
true with discrepancy n 1 . Our goal is to show that a) also is true with discrepancy n . By use
of the shift we have (f (x)) = (a1 a2 a3 . . .) and (f (y)) = (b1 b2 b3 . . .) . Suppose a0 = 0 .
Then (f (x)) (f (y)) since the number of 1s before the discrepancy is as before. Therefore
f (x) < f (y) but since f is increasing on [0, c it follows that x < y .

Next, assume a0 = 1 . Then (f (x))  (f (y)) since the number of 1s among the ai s ( i 1 )
has been reduced by one. Therefore f (x) > f (y) which implies that x < y since f decreases on
(c, 1]. If a0 = C we have x = y = c .

Challenge the way we run

EXPERIENCE THE POWER OF

FULL ENGAGEMENT

RUN FASTER.
RUN LONGER.. READ MORE & PRE-ORDER TODAY
RUN EASIER WWW.GAITEYE.COM

1349906_A6_4+0.indd 1 22-08-2014 12:56:57

53 Click on the ad to read more
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Regarding b) suppose x < y and assume that (x) and (y) has discrepancy n . First, note
that if x < c < y we have directly (x) < (y) . Otherwise (i.e. x < y < c or c < x < y )
note that f i is monotone in [x, y] for i n . Since the number of 1s (cf. the chain rule) directly
says if f n is increasing or decreasing it is easily verified that (x) (y) .

Theorem 1.10.2. Let x = (a) = a0 a1 a2 . . . and suppose that x f (x) unimodal. Then
( n (a))  K(f (c)) for n 1 . 

Proof. Since the maximum of f is f (c) we have f (x) < f (c) and f n (x) f (c) . Moreover,
x = ((a)) = a1 a2 . . . = (f (x)) so inductively n x = (f n (x)) . Therefore, according
to Theorem 1.10.1

n
( (a))  (f (c)) = K(f (c))

The essence of Theorem 1.10.2 is that any sequence a such that (x) = a has lower order than the
kneading sequence.

Now, consider periodic orbits. In order to simplify notation, repeating sequences (corresponding to
periodic points) of the form a = (a0 a1 . . . an a0 a1 . . . an a0 a1 . . . an . . .) = (a0 a1 . . . an a1 a1 . . . an )
will from now on be written as a = (a0 a1 . . . an ) .

We also define a sequence a = (a0 . . . an1 an ) where an = 1 if an = 0 or an = 0 if an = 1 . If

b = (b0 b1 . . . bm ) , a b = (a0 a1 . . . an b0 b1 . . . bm ) .

Suppose that there exists a parameter value such that there are two periodic orbits 1 and 2 of the
same prime period. We say that the orbit 1 is larger than the orbit 2 if 1 contains a point pm which
is larger than all the points of 2 . Note that, according to Theorem 1.10.1, the itinerary of pm satisfies
(pi )  (pm ) where pi are any of the other periodic points contained in 1 .

Our main interest is the ordering of itineraries of periodic points p which satisfy:

(A) The periodic point p shall be the largest point contained in the orbit.

(B) Every other periodic orbit of the same prime period must have a periodic point which is
larger than p .

Before we continue the discussion of (A) and (B) let us state a useful lemma.

54
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Lemma 1.10.1. Given two symbol sequences a = (a0 a1 a2 . . .) and b = (b0 b1 b2 . . .) .

Suppose that a0 = b0 = 1 and a1 = b1 = 0 . aj = bj = 1 for 2 j l , al = 0 , bl = 1 .

If l is even then b a . If l is odd then a b . 

Proof. Assume l even. Then the number of 1s before the discrepancy is odd and since bl > al
Definition 1.10.3 gives that b a .

If l is odd the number of 1s before the discrepancy is even and since al = 0 < bl = 1 , a b
according to the definition. 

A consequence of this theorem is that sequences than begin with 1 0 are of larger order than sequences
which begin with 1 1 . In the same way, a sequence which first entries are 1 0 0 is larger than one which
begins with 1 0 1 .

Now, consider the quadratic map x x(1 x) . Whenever > 2 the fixed point
x = ( 1)/ > c = 1/2 so the (repeating) itinerary becomes (x ) = (1) . When x bifurcates

at the threshold = 3 , the largest point p1 contained in the 2-cycle is always larger than c , hence
the itinerary of p1 starts with 1 in the first entry. Therefore, when > 3 , there may be two possible
itineraries (1 0) and (1 1) and clearly (1 1) (1 0) . We are interested in (1 0) . Considering the 4-cycle
which is created through another flip bifurcation the itinerary of the largest point contained in the cycle
which we seek is (1 0 1 1) which is of larger order than the other alternatives.

Turning to odd periodic orbits, remember that they are established through saddle-node bifurcations,
thus two periodic orbits, one stable and one unstable, are established at the bifurcation. Considering the
stable 3-cycle at = 3.839 (see Exercise 1.4.2 or the bifurcation diagram, Figure 7) two of the points
in the cycle 0.14989 and 0.48917 are smaller than c while the third one 0.95943 is larger. Hence the
itinerary of largest order of 0.95493 is (1 0 0) . Referring to Exercise 1.4.2 the largest point contained
in the unstable 3-cycle is 0.95384 and the other points are 0.16904 and 0.53392. Hence the itinerary of
0.95384 of largest order is (1 0 1) and according to (A) and (B) this is the itinerary we are looking for,
not the itinerary (1 0 0) .

Therefore, the itineraries we seek are the ones that satisfy (A) and (B) and correspond to periodic points
which are established through flip or saddle-node bifurcations as the parameter in the actual family is
increased. (A final observation is that sequences which contain the symbol C are out of interest since
they violate (B).)

55
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Now, cf. our previous discussion, define the repeating sequences:

S0 = (1) S1 = (1 0) S2 = (1 0 1 1) S3 = (1 0 1 1 1 0 1 0)
and

Sj+1 = Sj Sj (1.10.4)

Clearly, the sequence Sj has prime period 2j so it represents a periodic point with the same prime period.

Another important property is that Sj has an odd number of 1s. To see this, note that S0 = (1)
has an odd number of 1s. Next, assume that Sk = (S0 . . . Sk1 1) has an odd number of 1s. Then
Sk = (S0 . . . Sk1 0) has an even number of 1s so the concatenation Sk+1 = Sk Sk clearly has an
odd number of 1s. (If Sk has a 0 at entry Sk we arrive at the same conclusion.) We have also that

Sj+1 = Sj Sj = Sj (1.10.5)

Indeed, suppose Sk = (S0 . . . Sk ) . Then Sk+1 = Sk Sk = (S0 . . . Sk S0 . . . Sk ) so Sk+1 = (S0 . . . Sk S0 . .

Sk+1 = (S0 . . . Sk S0 . . . Sk ) = Sk Sk = Sk .

This e-book
is made with SETASIGN
SetaPDF

PDF components for PHP developers

www.setasign.com

56 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Lemma 1.10.2. The sequences defined through (1.10.4) have the ordering

S 0 S1 S2 S3 . . .


Proof. Assume that Sj = (S0 . . . Sj1 Sj ) . If Sj = 1 there must be an even number of 1s among
(S0 . . . Sj1 ) so according to Definition 1.10.3a Sj Sj . If Sj = 0 there is an odd number
of 1s among (S0 . . . Sj1 ) so according to Definition 1.10.3b Sj Sj also here. Therefore, by
use of (1.10.5), we have Sj  Sj = Sj1 Sj1 = Sj1 . 

Let us now turn to periodic orbits of odd period. The following lemma is due to Guckenheimer.

Lemma 1.10.3. The largest point pm in the smallest periodic orbit of odd period n has itinerary
(pm ) = a such that ai = 0 if i = 1(mod n) and ai = 1 otherwise. 

Example 1.10.4. If n = 3 , (pm ) = (1 0 1 1 0 1 1 0 1 . . .) = (1 0 1) which is in accordance

with our previous discussion of 3-cycles. 

Proof. Suppose that we have a sequence a and that there exists a number k such that ak = 1
and ak+1 = ak+2 = 0 . Then by applying the shift map k times we arrive at k (a) = (1 0 0 . . .)
which according to Lemma 1.10.1 has larger order than any sequence with isolated 0s. Hence the
sequence k (a) violates (A) and (B).

Therefore, the argument above shows that the sequence we are looking for in this lemma must
satisfy that if ak = 0 then both ak1 and ak+1 must equal 1. Consequently there are blocks in
a of even length where the first and last entry of the blocks consist of 0 and the intermediate
elements of 1s. As a consequence of Lemma 1.10.1 the longer these blocks are the smaller is the
order of the sequence. Note that the blocks in this lemma have maximum length n + 1 for a
periodic sequence of period n . 

Example 1.10.5. (1 0
 11 0 1 1 0 1) is a 3-cycle where the length of the block is 4.

(1 0 1 11 1 0 1 1 1) is a 5-cycle where the length of the block is 6. Clearly, the order of the 5-cycle
is smaller than the order of the 3-cycle. 

Lemma 1.10.4. Let n > 1 be an odd number. Then there is a periodic orbit of period n + 2
which is smaller than all periodic orbits of period n . 

57
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Proof. The lemma is an immediate consequence of how the itinerary in Lemma 1.10.3 is defined
combined with the results of Lemma 1.10.1. 

We now turn to orbits of even period where the period is 2n m where m > 1 is an odd number.
The fundamental observation regarding the associated symbol sequences is that they may be written as
Sj+1 Sj . . . Sj or Sj Sj Sj . . . Sj where the number of Sj blocks following Sj+1 (or Sj ) is m 2 .
(See Guckenheimer (1977) for further details.)

Example 1.10.6. If n = 2 (cf. 1.10.4) and m = 3 we have the sequence (1 0 1 1 1 0 1 0 1 0 1 1)

and if n = 1 and m = 5 we arrive at (1 0 1 1 1 0 1 0 1 0). 

Lemma 1.10.5. Let P be a periodic orbit of odd period k . Then there exists a periodic orbit of
even period l = 2n m where m > 1 is odd which is smaller than any odd period orbit. 

Proof. From Lemma 1.10.4 we have that the longer the odd period is the smaller is the ordering of
the associated symbol sequence. From Lemma 1.10.3 it follows that such a symbol sequence may
be written as (1 0 1 1 1 . . . 1 1 1 0 1 1 1 . . .) . Therefore by comparing an even period sequence
with the odd one above it is clear that the even period sequence has 0 as entry at the discrepancy.
If the even period is 2 it is two 1s before the discrepancy. If the even period is larger there are
three consecutive 1s just prior to the 0 and since the first entry of the sequence is 1 there is an
even number of 1s before the discrepancy also here and the result of the lemma follows. 

We need one more lemma which deals with periodic orbits of even period.

Lemma 1.10.6. Let u = 2n l , v = 2n k and w = 2m r where l , k and r are odd numbers.

a) Provided 1 < k < l there are repeating symbol sequences of period u which has smaller
order than any repeating symbol sequence of period v.

b) Provided m > n there are repeating symbol sequences of period w which has smaller
order than any repeating symbol sequence of period v . 

Sketch of proof. Regarding a) consider Sj such that j is odd. Then by carefully examining the
various sequences we find that the discrepancy occurs at entry 2j (k + 2) in the repeating sequence
of the 2n k periodic point and it happens as the last entry of the Sj block (which of course
is 1 since j is odd) differs from the same entry in the 2n l sequence. Now, since Sj Sj has an
odd number of 1s the number of 1s before the discrepancy is even, so according to Definition
1.10.3a we have that sequences of period 2n l are smaller than any sequence of period 2n k .
(The case that j is even is left to the reader.)

58
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Turning to b) and scrutinizing sequences a of period 2m k it is clear that all of them have
1 0 1 1 as the first entries and that ai = 1 if i is even and ai = 0 if i = 1(mod 4) . Moreover,
assuming k > r whenever m > n we find that at discrepancy the sequence of period w has 1
as its element and in fact it is the last 1 in 1 0 1 1 . Now, since Sj Sj Sj . . . Sj has an even number
of 1s the observation above implies that the sequence of period 2n k must have an even number
of 1s before the discrepancy so the result follows. 

Now at last, combining the results from Lemmas 1.10.11.10.6 we have established the following ordering
for the itineraries of periodic points that satisfy (A) and (B):

2 22 23 . . . 2n 2n (2l +1) 2n (2l 1) . . . 2n 5 2n 3 2n1 (2l +1)

. . . 2n1 3 . . . (2l + 1) (2l 1) . . . 5 3

which is nothing but the ordering we find in Sarkovskiis theorem.

We do not claim that we actually have proved the theorem in all its details, our main purpose here have
been to show that symbolic dynamics is a powerful tool when dealing with periodic orbits. For further
reading, also of other aspects of symbolic dynamics we refer to Guckenheimer and Holmes (1990),
Devaney (1989) and Collet and Eckmann (1980).

www.sylvania.com

We do not reinvent
the wheel we reinvent
light.
Fascinating lighting offers an infinite spectrum of
possibilities: Innovative technologies and new
markets provide both opportunities and challenges.
An environment in which your expertise is in high
demand. Enjoy the supportive working atmosphere
within our global group and benefit from international
career paths. Implement sustainable ideas in close
cooperation with other specialists and contribute to
influencing our future. Come and join us in reinventing
light every day.

Light is OSRAM

59 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

1.11 Chaos
As we have seen, the dynamics of x x(1 x) differs substantially depending on the value of
the parameter . For 2 < < 3 there is a stable nontrivial fixed point, and in case of larger values

of we have detected periodic orbits both of even and odd period. If > 2 + 5 the dynamics
is aperiodic and irregular and occurs on a Cantor set and points x (I  ) approaches .
( I is the unit interval.)

In this section we shall deal with the concept chaos. Chaos may and has been defined in several ways.
We have already used the concept when we stated Period three implies chaos.

Referring to the examples and exercises at the end of Section 1.3 we found that whenever the long-
time behaviour of a system was a stable fixed point or a stable periodic orbit there was no sensitive
dependence on the initial condition x0 . However, when x f (x) = 4x(1 x) we have proved that
there is no stable periodic orbit and moreover, we found a strong sensitivity on the initial condition.
Assuming x [0, 1] and that x0 = 0.30 is one initial condition and x00 = 0.32 is another we
have |x0 x00 | = 0.02 but most terms |f k (x0 ) f k (x00 )| > 0.02 and for some k ( k = 9 )
|f k (x0 ) f k (x00 )| 1 which indeed shows a strong sensitivity.

Motivated by the example above, if an orbit of a map f : I I shall be denoted as chaotic it is natural
to include that f has sensitive dependence on the initial condition in the definition. It is also natural to
claim that there is no convergence to any periodic orbit which is equivalent to, say, that periodic orbits
must be dense in I . Our goal is to establish a precise definition of the concept chaos but before we do
that let us first illustrate what we have discussed above by two examples.

Example 1.11.1. This is a standard example which may be found in many textbooks. Consider
the map h : S  S  , h() = 2 . ( h is a map from the circle to the circle.) Clearly, h
is sensitive to initial conditions since the arc length between nearby points is doubled under
h . Regarding the dense property, observe that hn () = 2n so any periodic points must be
obtained from the relation 2n = + 2k or = 2k/(2n 1) where the integer k satisfies
0 k 2n . Hence in any neighbourhood of a point in S there is a periodic point so the periodic
points are dense so h does not converge to any stable periodic orbit. Consequently, h is chaotic
on S  . 

Example 1.11.2. Consider x f (x) = x(1 x) where > 2 + 5 . We claim that f is
chaotic on the Cantor set . In order to show sensitive dependence on the initial condition let
be less than the distance between the intervals I0 and I1 (cf. Figure 7). Next, assume x, y
where x = y . Then the itineraries (x) = (y) so after, say, k iterations f k (x) is in I0 (I1 )
and f k (y) is in I1 (I0 ) . Thus |f k (x) f k (y)| > which establishes the sensitive dependence.

60
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Since f : is topological equivalent to the shift map : 2 2 it suffices to show that

the periodic points of are dense in 2 . Let a = (a1 . . . an ) be a repeating sequence of a periodic
point and let b = (a1 a2 a3 . . .) be the sequence of an arbitrary point and note that n (a) = a .
By use of the distance d between two symbol sequences one easily obtains d[a, b] < 1/2n so in
any neighbourhood of an arbitrary sequence (point) there is a periodic sequence (periodic point).
Hence periodic points of f are dense (and unstable). 

In our work towards a definition of chaos we will now focus on the sensitive dependence on the initial
condition.

If a map f : R R has a fixed point we know from Section 1.4 that if the eigenvalue of the
linearized system satisfies 1 < < 1 the fixed point is stable and not sensitive to changes of the initial
condition. If || > 1 one may measure the degree of sensitivity by the size of || . We may use the
same argument if we deal with periodic orbits of period k except that we on this occasion consider the
eigenvalue of every periodic point contained on the orbit. If a system is chaotic it is natural to consider
the case k since we may think of a chaotic orbit as one having an infinite period. Therefore, define
 1/k
d k
= lim  f (x)x=x0  (1.11.1)

k dx

where we have used the k th root in order to avoid problems in order to obtain a well defined limit. If
x0 is a fixed point = |(df /dx)(x = x0 )| . For a general orbit starting at x0 we may think of as
an average measure of sensitivity (or insensitivity) over the whole orbit. Let L = ln , that is
 1/k k1
d k 1
ln |f  (x = xn )| (1.11.2)

L = lim ln  f (x0 ) = lim
 
k dx k k
n=0

The number L is called the Lyapunov exponent and if L > 0 (which is equivalent to || > 1 ) we have
sensitive dependence on the initial condition. By use of L we may now define chaos.

Definition 1.11.1. The orbit of a map x f (x) is called chaotic if

1) It possesses a positive Lyapunov exponent, and

2) it does not converge to a periodic orbit (that is, there does not exist a periodic orbit
yt = yt+T such that limt |xt yt | = 0 .) 

Note that 2) is equivalent to, say, that periodic orbits are dense.

In most cases the Lyapunov exponent must be computed numerically and in cases where L is slightly
larger than zero such computations have to be performed by some care due to accumulation effects
of round-off errors. Note, however, that there exists a theorem saying that L is stable under small
perturbations of an orbit.

61
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Example 1.11.3. Compute L for the map h : S  S  , h() = 2 . In this case h = 2 for all
points on the orbit so

k1
1 1
L = lim ln |h (x = xn )| = lim k ln 2 = ln 2 > 0
k k k k
n=0

and since the periodic orbits are dense h is chaotic. 

Example 1.11.4. Compute L for the two periodic orbit of x f (x) = x(1 x) where

3<<1+ 6 . Referring to formulae (1.3.3) the periodic points are


+1 ( + 1)( 3)

360
p1,2 =
Thus, 2

L = lim

= lim
1
k k
1 k
k k


2
 k 
ln |f (x = p1 )| + ln |f (x = p2 )|
2
 thinking .
{ln |f  (x = p1 )| + ln |f  (x = p2 )| + ln |f (x = p1 )| + . . . + ln |f (x = p2 )|}

1
= ln |f (x = p1 )f  (x = p2 )|
2

360
thinking . 360
thinking .
Discover the truth at www.deloitte.ca/careers Dis

Deloitte & Touche LLP and affiliated entities.

Discover the truth at www.deloitte.ca/careers Deloitte & Touche LLP and affiliated entities.

Deloitte & Touche LLP and affiliated entities.

Discover the truth

62 at www.deloitte.ca/careers
Click on the ad to read more
Download free eBooks at bookboon.com

Deloitte & Touche LLP and affiliated entities.

Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Since

f  (x = p1 )f  (x = p2 ) = (1 2p1 )(1 2p2 ) = 1 ( + 1)( 3)

it follows that L = (1/2) ln |1 ( + 1)( 3)| and as expected L < 0 whenever

3<<1+ 6 . (Note that if > 1 + 6 then L > 0 but the map is of course not chaotic

since there in this case (provided | (1 + 6)| small) exists a stable 4-periodic orbit with
negative L .) 

Example 1.11.5. Show that the Lyapunov exponents of almost all orbits of the map
f : [0, 1] [0, 1] , x f (x) = 4x(1 x) is ln 2 .

Solution: From Proposition 1.2.1 we know that f (x) is topological equivalent to the tent map
T (x) . The nice property of T (x) which we shall use is that T  (x) = 2 for all x = c = 1/2 .
Moreover, h f = T h implies that h (f (x))f  (x) = T  (h(x))h (x) so

T  (h(x))h (x)
f  (x) =
h (f (x))

We are now ready to compute the Lyapunov exponent:

n1
1
L = lim ln |f  (x = xi )|
n n
i=0
n1  
1   T  (h(xi ))h (xi ) 
= lim ln 
n n h (f (xi )) 
i=0
n1 n1
1 1
= lim ln |T  (h(xi ))| + lim {ln |h (xi )| ln |h (f (xi ))|}
n n n n
i=0 i=0

Since xi+1 = f (xi ) the latter sum may be written as

1
{ln |h (x0 )| ln |h (xn )|}
lim
n n

which is equal to zero for almost all orbits. Thus, for almost all orbits:

n1
1 1
L = lim ln |T  (h(xi ))| = lim n ln 2 = ln 2
n n n n
i=0

63
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

For comparison reasons we have also computed L numerically with initial value x0 = 0.30 in the
example above. Denoting the Lyapunov exponent of n iterations for Ln we find L100 = 0.67547 ,
L1000 = 0.69227 and L5000 = 0.69308 so in this example we do not need too many terms in order
to show that L > 0 .

A final comment is that since we have proved earlier (cf. Example 1.8.1) that the quadratic map does
not possess any stable orbits in case of = 4 , Definition 1.11.1 directly gives that almost all orbits of
the map are chaotic. Other properties of Lyapunov exponents may be obtained in the literature. See for
example Tsujii (1993) and Thieullen (1994).

1.12 Superstable orbits and a summary of the dynamics of the quadratic map
The quadratic map has two fixed points. One is the trivial one x = 0 which is stable if < 1 and
unstable if > 1 . If > 1 the nontrivial fixed point is x = ( 1)/ and as we have shown this
fixed point is stable whenever 1 < < 3 . Whenever > 2 the fixed point is larger than the critical point

c . At = 3 the map undergoes a supercritical flip bifurcation and in the interval 3 < < 1 + 6 the
quadratic map possesses a stable period-2 orbit which has a negative Lyapunov exponent. The periodic
points are given by formulae (1.3.3).

At the threshold = 1 + 6 there is a new (supercritical) flip bifurcation which creates a stable orbit
of period 2 and through further increase of stable orbits of period 2k are established. However, the
2

parameter intervals where the period 2k cycles are stable shrinks as is enlarged so the values at the
bifurcation points act more or less as terms in a geometric series. By use of the Feigenbaum geometric ratio
one can argue that there exists an accumulation value a for the series of flip bifurcations. Regarding the
quadratic map, a = 3.56994 . In the parameter interval a < 4 we have seen that the dynamics
is much more complicated.

Still considering periodic orbits, Sarkovskiis theorem tells us that periodic orbits occur in a definite order
so beyond a there are periodic orbits of periods given by Theorem 1.7.2 (see also Section 1.10). Even
in cases where such orbits are stable they may be difficult to distinguish from non-periodic orbits due

to the long period. In many respects the ultimate event occurs at the threshold = 1 + 8 where a
3-periodic orbit is created because period 3 implies orbits of all other periods which is the content both
in Li and Yorke and in Sarkovskiis theorem.

64
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Chaotic orbits may be captured by use of Lyapunov exponents. In Figure 9 we show the value of the
Lyapunov exponent L for [a , 4] . L < 0 corresponds to stable periodic orbits, L > 0 corresponds
to chaotic orbits. (Figure 9 should be compared to the bifurcation diagram, Figure 7.) The regions where
we have periodic orbits are often referred to as windows. The largest window found in Figure 7 (or 9)
is the period 3 window. The periodic orbits in the interval 3 < < a are created through a series of
flip bifurcations. However, the period-3 orbit is created through a saddle-node bifurcation. In fact, every
window of periodic orbits beyond a is created in this way so just beyond the bifurcation value there is

one stable and one unstable orbit of the same period. (If is slightly larger than 1 + 8 there is one
stable and one unstable orbit of period 3.) Within a window there may be flip bifurcations before chaos
is established again, cf. Figure 7. Since the quadratic map has negative Schwarzian derivative there is at
most one stable periodic orbit for each value of .

We will turn your CV into

an opportunity of a lifetime

Do you like cars? Would you like to be a part of a successful brand? Send us your CV on
We will appreciate and reward both your enthusiasm and talent. www.employerforlife.com
Send us your CV. You will be surprised where it can take you.

65 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

Figure 9: The value of the Lyapunov exponent for [a , 4] . L < 0 corresponds to stable periodic orbits.
L > 0 corresponds to chaotic orbits.

There is a way to locate the periodic windows. The vital observation is that at the critical point c ,
f  (c) = 0 , so accordingly ln |f  (c)| = which implies L < 0 and consequently a stable periodic
orbit. Also, confer Singers theorem (Theorem 1.8.1).

Definition 1.12.1. Given a map f : I I with one critical point c . Any periodic orbit passing
through c is called a superstable orbit. 

Hence, by searching for superstable orbits one may obtain a representative value of the location of a
periodic window. Indeed, any superstable orbit of period n must satisfy the equation

fn (c) = c (1.12.1)

Example 1.12.1. Consider the quadratic map and let us find the value of such that
f3 (1/2) = 1/2 .

We have

1 1 1 1 3
c= f (c) = f2 (c) = 2
2 4
 4 16
 
3 1 3 1 4 1 2 1 3
f (c) = 1
4 16 4 16

Hence, the equation f3 (1/2) = 1/2 becomes

7 86 + 165 + 164 643 + 128 = 0 (1.12.2)

66
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization One-dimensional maps

By inspection, = 2 is a solution of (1.12.2) so after dividing by 2 we arrive at

6 65 + 44 + 243 162 32 64 = 0 (1.12.3)

This equation may be solved numerically by use of Newtons method and if we do that we find
that the only solution in the interval a 4 is = 3.83187 . Therefore, there is only one
period-3 window and the location clearly agrees both with the bifurcation diagram, Figure 7 and
Figure 9. In the same way, by solving f4 (1/2) = 1/2 one finds that the only solution which
satisfies a < < 4 is = 3.963 which shows that there is also only one period-4 window.
However, if one solves f5 (1/2) = 1/2 one obtains three values which means that there exists
three period-5 windows. The first one occurs around 1 = 3.739 and is visible in the bifurcation
diagram, Figure 7. The others have almost no widths, the values that correspond to the superstable
orbits are 2 = 3.9057 and 3 = 3.9903 . 

Referring to the numerical examples given at the end of Section 1.3 where < a we observed a rapid
convergence towards the 2-period orbit independent on the choice of initial value. Within a periodic
window in the interval [a , 4] the dynamics may be much more complicated. Indeed, still considering
the period-3 window, we have according to the Li and Yorke theorem that there are also periodic orbits
of any period, although invisible to a computer. (The latter is a consequence of Singers theorem.) If we
consider an initial point which is not on the 3-periodic orbit we may see that it behaves irregularly through
lots of iterations before it starts to converge, and moreover, if we change the initial point somewhat it may
happen that it is necessary to perform an even larger amount of iterations before we are able to detect
any convergence towards the 3-cycle. Hence, the dynamics within a periodic window in the interval
[a , 4] is in general much more complex than in the case of periodic orbits in the interval [3, a ] due
to the presence of an (infinite) number of unstable periodic points.

By carefully scrutinizing the periodic windows one may find numerically that the sum of the widths of
all the windows is roughly 10% of the length of the interval [a , 4] . In the remaining part of the interval
the dynamics is chaotic. If we want to give a thorough description of chaotic orbits we may use symbolic
dynamics in much of a similar way as we did in Sections 1.9 and 1.10. Here we shall give a more heuristic
approach only. If is not close to a periodic window, orbits are irregular and there is almost no sign of

periodicity. However, if is close to a window, for example, if is smaller but close to 1 + 8 (the
threshold value for the period-3 window) one finds that an orbit seems to consist of two parts, one part
with appears to be almost 3-periodic and another irregular part where the point x may take almost any
value in (0, 1) . The almost 3-periodic part of the orbit is established when the orbit becomes close to

the diagonal line xt+1 = xt . Then, since is close to 1 + 8 the orbit may stay close to the diagonal
for several iterations before it moves away. Therefore, a typical orbit close to a periodic window consists
of an irregular part which after a finite number of iterations becomes almost periodic and again turns
irregular in a repeating fashion. For further reading on this topic we refer to Nagashima and Baba (1999),
Thunberg (2001), and Jost (2005). We also recommend the books by Iooss (1979), Berg et al. (1984),
Barnsley (1988), Devaney (1989), Saber et al. (1998), and Iooss and Adelmeyer (1999).

67
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Part II
 n-dimensional maps
f : Rn Rn x f (x)

68
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

2.1 Higher order difference equations

Consider the second order difference equation

xt+2 + at xt+1 + bt xt = f (t)  (2.1.1)

If f (t) = 0 , (2.1.1) is called a nonhomogeneous difference equation. If f (t) = 0 , that is

xt+2 + at xt+1 + bt xt = 0  (2.1.2)

we have the associated homogeneous equation.

Theorem 2.1.1. The homogeneous equation (2.1.2) has the general solution

xt = C1 ut + C2 vt

where ut and vt are two linear independent solutions and C1 , C2 arbitrary constants.

Proof. Let xt = C1 ut + C2 vt . Then xt+1 = C1 ut+1 + C2 vt+1 and xt+2 = C1 ut+2 + C2 vt+2
and if we substitute into (2.1.2) we obtain

C1 (ut+2 + at ut+1 + bt ut ) + C2 (vt+2 + at vt+1 + bt vt ) = 0

which clearly is correct since ut and vt are linear independent solutions. 

Regarding (2.1.1) we obviously have:

Theorem 2.1.2. The nonhomogeneous equation (2.1.1) has the general solution

xt = C1 ut + C2 vt + ut

where C1 ut + C2 vt is the general solution of the associated homogeneous equation (2.1.2) and
ut is any particular solution of (2.1.1).

Just as in case of differential equations there is no general method of how to find two linear independent
solutions of a second order difference equation. However, if the coefficients at and bt are constants then
it is possible.

Indeed, consider

xt+2 + axt+1 + bxt = 0  (2.1.3)

69
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

where a and b are constants. Suppose that there exists a solution of the form xt = mt where m = 0 .

Then xt+1 = mt+1 = mmt and xt+2 = m2 mt so (2.1.3) may be expressed as

2
(m + am + b)mt = 0

which again implies that

m2 + am + b = 0  (2.1.4)

(2.1.4) is called the characteristic equation and its solution is easily found to be

a a2
m1,2 = b  (2.1.5)
2 4
Now we have the following result regarding the solution of (2.1.3) which we state as a theorem:

Theorem 2.1.3.

1) If (a2 /4) b > 0 , the characteristic equation have two real solutions m1 and m2 . Moreover,
mt1 and mt2 are linear independent so according to Theorem 2.1.1 the general solution of
(2.1.3) is

a a2
xt = C1 mt1 + C2 mt2 where m1,2 = b
2 4

2) The case (a2 /4) b = 0 implies that m = a/2 . Then mt and tmt are two linear

independent solutions of (2.1.3) so the general solution becomes:

xt = C1 mt + C2 tmt = (C1 + C2 t)mt where m = a/2

(In order to see that tmt really is a solution of (2.1.3) note that if a2 /4 = b , then (2.1.3) may

be expressed as (*) xt+2 + axt+1 + (a2 /4)xt = 0 . Now, assuming that xt = t(a/2)t we
have xt+1 = (a/2)(t + 1)(a/2)t , xt+2 = (a2 /4)(t + 2)(a/2)t and by inserting into

(*) we obtain (a2 /4)[t + 2 2(t + 1) + t](a/2)t = 0 which proves what we want.)

3) Finally, if (a2 /4) b < 0 we have

a  a 
m= (b (a2 /4) = b (a2 /4) i = + i
2 2

From the theory of complex numbers we know that

+ i = r(cos + i sin )

70
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

where


  2
r = 2 + 2 = (a/2)2 + b (a2 /4) = b

and

a/2 b (a2 /4)
cos = sin =
b b
which implies that

mt = [r(cos + i sin )]t = r t (cos + i sin )t = r t (cos t + i sin t)

where we have used Moivres formulae (cf. Exercise 2.1.2) in the last step. Since the real and
imaginary parts of mt are linear independent functions we express the general solution of
(2.1.3) as

xt = C1 r t cos t + C2 r t sin t

AXA Global
Graduate Program
Find out more and apply

71 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Example 2.1.1. Find the general solution of the following equations:

a) xt+2 7xt+1 + 12xt = 0 ,

b) xt+2 6xt+1 + 9xt = 0 ,
c) xt+2 xt+1 + xt = 0 .

Solutions:

a) Assuming xt = mt the characteristic equation becomes m2 7m + 12 = 0 m1 = 4 ,

m2 = 3 so according to Theorem 2.1.3 the general solution is xt = C1 4t + C2 3t .
b) The characteristic equation is m2 6m + 9 = 0 m1 = m2 = 3 . Thus

xt = C1 3t + C2 t 3t = (C1 + C2 t)3t .

c) The characteristic equation becomes m2 m + 1 = 0 m = (1 3)/2 = 1
2
1
2
3 i .

Further
 
2 2
1

1
r = + 3 =1
2 2

cos =
1 1
2
1
2
3 1
= sin = = 3=
1 2 1 2 3
Thus

xt = C1 1t cos t + C2 1t sin t = C1 cos t + C2 sin t
3 3 3 3

Exercise 2.1.1. Find the general solution of the homogeneous equations:

a) xt+2 12xt+1 + 36xt = 0 ,

b) xt+2 + xt = 0 ,
c) xt+2 + 6xt+1 16xt = 0 . 

Exercise 2.1.2. Prove Moivres formulae: (cos + i sin )t = cos t + i sin t .

(Hint: Use induction and trigonometric identities.) 

72
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Definition 2.1.1. The equation xt+2 + axt+1 + bxt = 0 is said to be globally asymptotic stable if the

solution xt satisfies limt xt = 0 . 

Referring to Example 2.1.1 it is clear that none of the equations considered there are globally asymptotic
stable. The solutions of the equations (a) and (b) tend to infinity as t and the solution of (c) does
not tend to zero either.

However, consider the equation xt+2 (1/6)xt+1 (1/6)xt = 0 . The characteristic equation is

m2 (1/6)m (1/6)m = 0 m1 = 1/2 , m2 = (1/3) so the general solution becomes

xt = C1 (1/2)t + C2 (1/3)t .

Here, we obviously have limt xt = 0 so according to Definition 2.1.1 the equation

xt+2 + (1/6)xt+1 (1/6)xt = 0 is globally asymptotic stable.

Theorem 2.1.4. The equation xt+2 + axt+1 + bxt = 0 with associated characteristic equation
m2 + am + b = 0 is globally asymptotic stable if and only if all the roots of the characteristic
equation have moduli strictly less than 1. 

Proof. Referring to Theorem 2.1.3, the cases (1) and (3) are clear (remember |m| = r in (3)).

Considering (2): If |m| < 1

t
lim tmt = lim
t t st

where s = 1/|m| and s > 1 . Then by Lhopitals rule

t 1
lim t
= lim t 0
t s t s ln s

and the results of Theorem 2.1.4 follows.

As we shall see later on, Theorem 2.1.4 will be useful for us when we discuss stability of nonlinear systems.

We close this section by considering the nonhomogeneous equation

xt+2 + axt+1 + bxt = f (t)  (2.1.6)

73
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

According to Theorem 2.1.2 the general solution of (2.1.6) is the sum of the general solution of the
homogeneous equation (2.1.3) and a particular solution ut of (2.1.6).

If f (t) is a polynomial, say f (t) = 2t2 + 4t it is natural to assume a particular solution of the form
ut = At2 + Bt + C .

If f (t) is a trigonometric function, for example f (t) = cos u t we assume that

ut = A cos u t + B sin u t .

If ft = ct , assume ut = Act (but see the comment following (2.1.7)).

Example 2.1.2. Solve the following equations:

a) xt+2 + xt+1 + 2xt = t2 ,

b) xt+2 2xt+1 + xt = 2 sin(/2)t ,

e Graduate Programme
I joined MITAS because for Engineers and Geoscientists
I wanted real responsibili www.discovermitas.com
Maersk.com/Mitas e G
I joined MITAS because for Engine
I wanted real responsibili Ma

Month 16
I was a construction Mo
supervisor ina const
I was
the North Sea super
advising and the No
Real work he
helping foremen advis
International
al opportunities
Internationa
ree wo
work
or placements ssolve problems
Real work he
helping fo
International
Internationaal opportunities
ree wo
work
or placements ssolve pr

74 Click on the ad to read more

Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Solutions:

a) The characteristic equation of the homogeneous equation becomes

m2 m 2 = 0 m1 = 2 and m2 = 1 so the general solution of the homogeneous
equation is xt = C1 2t + C2 (1)t . Assume ut = At2 + Bt + C . Then
2
ut+1 = A(t + 1)2 + B(t + 1) + C , ut+2 = A(t + 2) + B(t + 2) + C which inserted
into the original equation gives

A(t + 2)2 + B(t + 2) + C [A(t + 1)2 + B(t + 1) + C] 2 [At2 + Bt + C] = t2

2At2 + (2A 2B)t + (3A + B 2C) = t2 + 0t + 0

and by equating terms of equal powers of t we have (1) 2A = 1 , (2) 2A 2B = 0 , and (3)
3A + B 2C = 0 from which we easily obtain A = 1/2 , B = 1/2 and C = 1 . Thus
ut = (1/2)t2 (1/2)t 1 and the general solution is xt = C1 2t + C2 (1)t (1/2)t2 (1/2)t 1
2t + C2 (1)t (1/2)t2 (1/2)t 1 .

b) The solution of the characteristic equation becomes m1 = m2 = 1 homogeneous

solution (C1 + C2 t)1t = C1 + C2 t . Assume ut = A cos(/2)t + B sin(/2)t . Then,

ut+1 = A cos[(/2)(t+1)]+B sin[(/2)(t+1)] = A[cos(/2)t cos(/2)sin(/2)t sin(/2)]+

B[sin(/2)t cos(/2)+ ut+1 = A cos[(/2)(t+1)]+B sin[(/2)(t+1)] = A[cos(/2)t cos(/2)s
A[cos(/2)t cos(/2)sin(/2)t sin(/2)]+ B[sin(/2)t cos(/2)+ sin(/2) cos(/2)] = A sin(/2)t + B cos(/2)t

/2)] = A sin(/2)t + B cos(/2)t . In the same way, ut+2 = A cos(/2)t B sin(/2)t so after inserting

ut+2 , ut+1 and ut into the original equation we arrive at

2B cos t + 2A sin t = 0 cos t + 2 sin t
2 2 2 2

Thus 2B = 0 and 2A = 2 A = 1 and B = 0 so ut = cos(/2)t . Hence, the general

solution is xt = C1 + C2 t + cos(/2)t .

Finally, if xt+2 + axt+1 + bxt = ct we assume a particular solution of the form ut = Act . Then

ut+1 = Acct and ut+2 = Ac2 ct which inserted into the original equation yields
2
A(c + ac + b)ct = ct

75
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Thus, whenever c2 + ac + b = 0 the particular solution becomes

1
u = ct  (2.1.7)
c2 + ac + b

Note, however, that if c is a simple root of the characteristic equation, i.e. c2 + ac + b = 0 , then we

try a solution of the form ut = Btct and if c is a double root, assume ut = Dt2 ct .

Example 2.1.3. Solve the equations:

a) xt+2 4xt = 3t ,
b) xt+2 4xt = 2t ,

Solutions:

a) The characteristic equation is m2 4 = 0 m1 = 2 , m2 = 2 thus the homogeneous

solution is C1 2t + C2 (2)t . Since 3 is not a root of m2 4 = 0 we have directlyfrom(2.1.7)

that ut = (1/5)3t so the general solution becomes xt = C1 2t + C2 (2)t + (1/5)3t .

b) The homogeneous solution is of course C1 2t + C2 (2)t but since 2 is a simple root of

m2 4 = 0 we try a particular solution of the form ut = Bt2t . Then

ut+2 = 4B(t + 2)2t and by inserting into the original equation we arrive at

4B(t + 2)2t 4Bt 2t = 2t

which gives B = 1/8 . Thus xt = C1 2t + C2 (2)t + (1/8)t 2t . 

Exercise 2.1.3. Solve the problems:

a) xt+2 + 2xt+1 3xt = 2t + 5

b) xt+2 10xt+1 + 25xt = 5t c)

xt+2 xt+1 + xt = 2t

t
d) xt+2 + 9xt = 2 e)
xt+2 5xt+1 6xt = t 2t

(Hint: Assume a particular solution of the form (At + B) 2t .) 

76
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

In the examples and exercises presented above we found a particular solution ut of the nonhomogeneous
equation in a way which at best may be called heuristic. We shall now focus on a general method
(sometimes referred to as variation of parameters) which enables us to find ut of any nonhomogeneous

equation provided the general solution of the associated homogeneous equation is known.

Theorem 2.1.5 (Variation of parameters). Let x1,t and x2,t be two linear independent solutions
of (2.1.3) and let
 
 x1,t x2,t 
w t = 
 x1,t1 x2,t1



Then a particular solution ut of the nonhomogeneous equation

xt+2 + axt+1 + bxt = ft

may be calculated through

 
 x1,t x2,t 
t
 
  x1,m1 x2,m1 
ut = fm2 t0
m=0
wm


170x115_MIM_ datos_Espv4.pdf 1 17/12/15 18:29

EL 93%
DE LOS ALUMNOS DEL MIM ESTN
TRABAJANDO EN SU SECTOR
C
A LOS 3 MESES DE GRADUARSE
M

MASTER IN MANAGEMENT
CM

MY

CY ESTUDIA EN EL CORAZN DE MADRID Y APROVECHA LAS OPORTUNIDADES Y EXPERIENCIAS NICAS

Duracin: 1O MESES
QUE OFRECE LA CAPITAL DE ESPAA
CMY Exp. Profesional: 1 AO
POTENCIA TU FORMACIN CURSANDO UNA DOBLE TITULACIN QUE MEJOR SE ADAPTE A TUS
Idioma: ESPAOL / INGLS
K
OBJETIVOS PROFESIONALES
Formato: TIEMPO COMPLETO
ESTUDIA UN SEMESTRE EXTRA EN OTRO PAS CON EL PROGRAMA BEYOND BORDERS EXPERIENCE
Convocatorias: SEPT / FEB
Y AUMENTA TU RED DE CONTACTOS

5 ESPECIALIZACIONES 100 DEL MUNDO 55 NACIONALIDADES

MASTERS IN MANAGEMENT
PERSONALIZA TU PROGRAMA FINANCIAL TIMES
EN CLASE

www.ie.edu/master-management [email protected] Sguenos en IE MIM Experience

77 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Proof. The elements ut must be linear functions of the preceding elements of the sequence {ft } .

Hence, t

ut = dt,m fm2

m=0

which inserted into the nonhomogeneous equation gives

t
(dt+2,m + adt+1,m + bdt,m )fm2 + (dt+2,t+1 + adt+1,t+1 )ft1 + dt+2,t+2 ft = ft
m=0

The equation above must hold for any t > 0 . Consequently, for each m, the coefficients of fm

on both sides of the equation must be equal. Therefore,

dt+2,m + adt+1,m + bdt,m = 0 t>m1

dt+2,t+1 + adt+1,t+1 = 0
dt+2,t+2 = 1
The first of the three equations above expresses that the sequence {dt,m } is a solution of the

homogeneous equation in case of t > m 1 . Moreover, by imposing the initial condition

dm,m1 = 0 the second equation may be replaced by the first if t m 1 and we have the
initial conditions dm,m1 = 0 , dm,m = 1 . Now, since x1,t and x2,t are two linear independent
solutions of the homogeneous equations there are constants c such that

d = c1,m x1,t + c2,m x2,t

t,m

and the initial conditions are satisfied whenever

c1,m x1,m + c2,m x2,m = 1

c1,m x1,m1 + c2,m x2,m1 = 0
from which we easily obtain

x2,m1 x1,m1
c1,m = c2,m =
wm wm

Consequently,

x1,t x2,m1 x2,t x1,m1

dt,m =
wm

and the formulae in the theorem follows. 

Example 2.1.4. Use Theorem 2.1.5 and find a particular solution of

xt+2 5xt+1 + 6xt = 2t

78
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Solution. Clearly, two linear independent solutions of the associated homogeneous equation are
x1,t = 2t and x2,t = 3t . Moreover,
 t   

 2 3t   x1,t x2,t
wt =  t1 t1  = 6t1  = 6m1 [2t(m1) 3t(m1) ]

 
2 3  x1,m1 x2,m1 

Thus

t
 6m1 [2t(m1) 3t(m1) ] m2
ut = m1
2
m=0
6
t   m   

t1 1 t+1 2 t1 3 t+1 t+1
= 2 3 = (t + 1)2 (3 2 )
m=0
4 3 4
9 (t + 4) t
= 3t 2
4 2

Note that the particular solution found here is not the same as ut = 12 t 2t which would be

the result by use of a heuristic method (see Example 2.1.3b). However, the general solutions match.
Indeed,
 
9 t t+4 t
ut
xt = C1 x1,t + C2 x2,t + = C1 2 + C2 3 + 3 t
2 t
4 2
 
t 9 t 1 1
= (C1 2)2 + C2 + 3 t 2t = D1 2t + D2 3t t 2t
4 2 2

which is in accordance with the heuristic method. 

Example 2.1.5. Find the general solution of

xt+2 5xt+1 + 6xt = ln(t + 3)

Solution. By use of the findings from the previous example:

t

ut
 t(m1)
3t(m1) ln(m + 1)

= 2
m=0
t t
 
m
=3 t+1
3 ln(m + 1) 2 t+1
2m ln(m + 1)
m=0 m=0

Hence, the general solution becomes

t
 t

m
xt
t
= C1 2 + C2 3 + 3 t t+1
3 ln(m + 1) 2 t+1
2m ln(m + 1)
m=0 m=0

79
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Note that the solution above (in contrast to all our previous examples and exercises) contains
sums which may not be expressed in any simple forms. However, in a somewhat more cumbersome
way, we have obtained the general solution for any t 0 . Moreover, the constants C1 and C2

may be determined in the usual way if we know the initial conditions. Indeed, assuming x0 = 0
and x1 = 1 we arrive at the equations (A) C1 + C2 = 0 and (B) 2C1 + 3C2 + ln 2 = 1 from

which we obtain C1 = ln 2 1 and C2 = 1 ln 2 . 

Exercise 2.1.4. Use Theorem 2.1.5 and find the general solution of the equations

a) xt+2 7xt+1 + 10xt = 5t

b) xt+2 (a + b)x1,t + abxt = at

(Hint: distinguish between the cases a =

 b and a = b .)

Exercise 2.1.5. Consider the equation xt+2 = xt+1 + xt with initial conditions x0 = 0 , x1 = 1 .

a) Solve the equation.

b) Use a) and induction to prove that xt xt+2 x2t+1 = (1)t+1 , t = 0, 1, 2, ... .

80 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Let us now turn to equations of order n , i.e. equations of the form

xt+n + a1 (t)xt+n1 + a2 (t)xt+n2 + + an1 (t)xt+1 + an (t)xt = f (t)  (2.1.8)

In the homogeneous case we have the following result:

Theorem 2.1.6. Assuming an (t) = 0 , the general solution of

xt+n + a1 (t)xt+n1 + + an (t)xt = 0  (2.1.9)

is xt = C1 u1,t + + Cn un,t where u1,t . . . un,t are linear independent solutions of the
equation and C1 . . . Cn arbitrary constants.

Proof. Easy extension of the proof of Theorem 2.1.1. We leave the details to the reader. 

Regarding the nonhomogeneous equation (2.1.8) we have

Theorem 2.1.7. The solution of the nonhomogeneous equation (2.1.8) is

x
t
= C1 u1,t + + Cn un,t + ut

where ut is a particular solution of (2.1.8) and C1 u1,t + + Cn un,t is the general solution of
(2.1.9).

If a1 (t) = a1 , ..., an (t) = an constants we arrive at

xt+n + a1 xt+n1 + + an xt = f (t)  (2.1.10)

and as in the second order case we may assume a solution xt = mt of the homogeneous equation. This
yields the n -th order characteristic equation

mn + a1 mn1 + + an1 m + an = 0  (2.1.11)

Appealing to the fundamental theorem of algebra we know that (2.1.11) has n roots. If a root is real
with multiplicity 1 or complex we form linear independent solutions in exactly the same way as explained
in Theorem 2.1.3. In case of real roots with multiplicity p , linear independent solutions are

mt , tmt , ..., tp1 mt .

81
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Example 2.1.5. Solve the equations:

a) xt+3 2xt+2 + xt+1 2xt = 2t 4 ,

b) xt+3 6xt+2 + 12xt+1 8xt = 0 .

Solutions:

a) The characteristic equation is m3 2m2 + m 2 = 0 . Clearly, m1 = 2 is a solution and

m3 2m2 + m 2 = (m 2)(m2 + 1) = 0 . Hence the other roots are complex,

m2,3 = i . Following Theorem 2.1.3 r = 02 + 12 = 1 , cos = 0/1 = 0 ,

sin = 1/1 = 1 = /2 which implies the homogeneous solution

C1 2t + C2 cos(/2)t + C3 sin(/2)t . Assuming a particular solution ut = At + B
we find after inserting into the original equation, 2At 2B = 2t 4 so A = 1 and

B = 2 . Consequently, according to Theorem 2.1.7, the general solution is

xt = C1 2t + C2 cos(/2)t + C3 sin(/2)t t + 2 .

Excellent Economics and Business programmes at:

The perfect start

of a successful,
international career.

CLICK HERE
to discover why both socially
and academically the University
of Groningen is one of the best
places for a student to be
www.rug.nl/feb/education

82 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

b) The characteristic equation becomes m3 6m2 + 12m 8 = 0 (m 2)3 = 0 .

Hence, there is only one root, m = 2 , with multiplicity 3. Consequently,

xt = C1 2t + C2 t 2t + C3 t2 2t .

Exercise 2.1.6. Find the general solution of the equations:

xt+3 2xt+2 5xt+1 + 6xt = 0 xt+1 2xt = 1 + t2

xt+4 xt = 2t xt+1 2xt = 2t + 3t

Definition 2.1.2. The equation xt+n + a1 xt+n1 + + an xt = 0 is said to be globally

asymptotic stable if the solution xt satisfies limt xt = 0 .

Theorem 2.1.8. The equation xt+n + a1 xt+n1 + + an xt = 0 is globally asymptotic stable

if all solutions of the characteristic equation (2.1.11) have moduli less than 1.

It may be a difficult task to decide whether all roots of a given polynomial equation have moduli less
than unity or not. However, there are methods and one of the most frequently used is the Jury criteria
which we now describe.

Let

P (x) = xn + a1 xn1 + a2 xn2 + + an  (2.1.12)

be a polynomial with real coefficients a1 . . . an . Define

bn = 1 a2n , bn1 = a1 an an1 , bnj = aj an anj , b1 = an1 an a1

cn = b2n b21 , cn1 = bn bn1 b1 b2 , cnj = bn bnj b1 bj+1 , c2 = bn b2 b1 bn1

dn = c2n c22 , dnj = cn cnj c2 cj+2 . . . d3 = cn c3 c2 cn1

and proceed in this way until we have only three elements of the type

wn = vn2 vn3
2
, wn1 = vn vn1 vn3 vn2 , wn2 = vn vn2 vn3 vn1

83
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Theorem 2.1.9 (The Jury criteria). All roots of the polynomial equation P (x) = 0 where P (x)

is defined through (2.1.12) have moduli less than 1 provided:

P (1) > 0 (1)n P (1) > 0

|an | < 1 , |bn | > |b1 | , |cn | > |c2 | , |dn | > |d3| , |wn | > |wn2 | .

Remark 2.1.1. Instead of saying that all roots have moduli less than 1, an alternative formulation
is to say that all roots are located inside the unit circle in the complex plane.

Regarding the second order equation

x2 + a1 x + a2 = 0  (2.1.13)

the Jury criteria become

1 + a1 + a2 > 0

1 a1 + a2 > 0  (2.1.14)

1 |a2 | > 0

If we have a polynomial equation of order 3

x3 + a1 x2 + a2 x + a3 = 0  (2.15)

the Jury criteria may be cast in the form

1 + a1 + a2 + a3 > 0

1 a1 + a2 a3 > 0  (2.16)

1 |a3 | > 0

|1 a23 | |a2 a3 a1 | > 0

Evidently, the higher the order, the more complicated are the Jury criteria. Therefore, unless the coefficients
are very simple or on a special form the method does not work is the order of the polynomial becomes
large.

Later, when we shall focus on stability problems of nonlinear maps (which often leads to a study of
polynomial equations), we will also face the fact that the coefficients a1 . . . an do not consist of numbers

only but a mixture of numbers and parameters. In such cases, even (2.1.16) may be difficult to apply.

84
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

However, let us give one simple example of how the Jury criteria works.

Example 2.1.6. Show that xt+3 (2/3)xt+2 + (1/4)xt+1 (1/6)xt = 0 is globally

asymptotic stable.

Solution: According to Theorem 2.1.8 we must show that the roots of the associated characteristic
equation m3 (2/3)m2 + (1/4)m (1/6) = 0 are located inside the unit circle. Defining

a1 = (2/3) , a2 = 1/4 , a3 = (1/6) the four left-hand sides of (2.1.16) become 1/12, 25/12,
5/6 and 5/6, respectively. Consequently, all the roots are located inside the unit circle so the
difference equation is globally asymptotic stable.

Another theorem (from complex function theory) that may be useful and which applies not only to
polynomial equations is Rouches theorem. (In the theorem below, z = + i is a complex number.)

Theorem 2.1.10 (Rouches theorem). If f (z) and g(z) are analytic inside and on a simple closed

curve C and if |g(z)| < |f (z)| on C then f (z) + g(z) and f (z) and the same number of

zeros inside C .

Remark 2.1.2. If we take the simple closed curve C to be the unit circle |z| = 1 , then we may
use Theorem 2.1.10 in order to decide if all the roots of a given equation have moduli less than
one or not.

Example 2.1.7. Suppose that a > e and show that the equation az n ez = 0 has n roots

located inside the unit circle |z| = 1 .

Solution: Define f (z) = az n , g(z) = ez and consider f (z) + g(z) = 0 . Clearly, the

equation f (z) = 0 has n roots located inside the unit circle. On the boundary of the unit circle

we have |g(z)| = | ez | e < a = |f (z)| . Thus, according to Theorem 2.1.10, f (z) and

f (z) + g(z) have the same number of zeros inside the unit circle, i.e. n zeros.

85
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

2.2 Systems of linear difference equations. Linear maps from Rn to Rn

In this section our purpose is to analyse linear systems. There are several alternatives when one tries
to find the general solution of such systems. One possible method is to transform a system into one
higher order equation and use the theory that we developed in the previous section. Other methods
are based upon topics from linear algebra, and of particular relevance is the theory of eigenvalues and
eigenvectors. Later when we turn to nonlinear systems and stability problems it will be useful for us
to have a broad knowledge of linear systems so therefore we shall deal with several possible solution
methods in this section.

Consider the system

x1,t+1 = a11 x1,t + a12 x2,t + + a1n xn,t + b1 (t)

x2,t+1 = a21 x1,t + a22 x2,t + + a2n xn,t + b2 (t)  (2.2.1)

xn,t+1 = an1 x1,t + an2 x2,t + + ann xn,t + bn (t)

Here, all coefficients a11 . . . ann are constants and if bi (t) = 0 for all 1 i n we call (2.2.1) a
linear autonomous system.

American online
LIGS University
is currently enrolling in the
Interactive Online BBA, MBA, MSc,
DBA and PhD programs:

enroll by September 30th, 2014 and

save up to 16% on the tuition!
pay in 10 installments / 2 years
Interactive Online education
visit www.ligsuniversity.com to
find out more!

Note: LIGS University is not accredited by any

nationally recognized accrediting agency listed
by the US Secretary of Education.
More info here.

86 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

It is often convenient to express (2.2.1) in terms of vectors and matrices. Indeed, let
x = (x1 , ..., xn )T , b = (b1 , ..., bn )T and

a11 a1n
a21 a2n
A=  (2.2.2)


an1 ann

Then, (2.2.1) may be written as

xt+1 = Axt + bt  (2.2.3)

or in map notation

x Ax + b  (2.2.4)

First, let us show how one may solve a system by use of the theory from the previous section.

Example 2.2.1. Solve the system

(1) xt+1 = 2yt + t

(2) yt+1 = xt + yt

Replacing t by t + 1 in (1) gives

xt+2 = 2yt+1 + t + 1 = 2(xt + yt ) + t + 1 = 2xt + 2yt + t + 1

(2)

Further, from (1): 2yt = xt+1 t . Hence

xt+2 xt+1 2xt = 1

Thus, we have transformed a system of two first order equations into one second order equation,
and by use of the theory from the previous section the general solution of the latter equation is
easily found to be

xt = C1 2t + C2 (1)t 1/2

87
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

yt may be obtained from (1):

 
1 1 t+1 t+1 1
yt = (xt+1 t) = C1 2 + C2 (1) t
2 2 2
1 1 1
= C1 2t C2 (1)t t
2 2 4

The constants C1 and C2 may be determined if we know the initial values x0 and y0 . For example,

if x0 = y0 = 1 we have from the general solution above that

1 = C1 + C2 1/2
1
1 = C1 C2 1/4
2

which implies that C1 = 4/3 and C2 = 1/6 so the solution becomes

4 t 1 1 4 t 1 1 1
xt = 2 + (1)t yt = 2 + (1)t t
3 6 2 3 12 2 4


Exercise 2.2.1. Find the general solution of the systems

xt+1 = 2yt + t xt+1 = xt + 2yt

yt+1 = xt + 3yt yt+1 = 3xt



Another way to find the solution of a system is to use the matrix formulation (2.2.3). Indeed, suppose
that the initial vector x0 is known. Then:

x1 = Ax0 + b(0)

x2 = Ax1 + b(1) = A(Ax0 + b(0)) + b(1) = A2 x0 + Ab(0) + b(1)

and by induction (we leave the details to the reader)

xt = At x0 + At1 b(0) + At2 b(1) + + b(t 1)  (2.2.5)

In the important special case b = 0 we have the result:

xt+1 = Axt xt = At x0  (2.2.6)

88
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

where A0 is equal to the identity matrix I .

Exercise 2.2.2. Consider the matrix

 
1 1
A=
0 1

a) Compute A2 and A3 .

b) Let t be a positive integer and use induction to find a formulae for At .

c) Let x = (x1 , x2 )T and solve the difference equation xt+1 = Axt where x0 = (a, b)T .

Our next goal is to solve the linear system

xt+1 = Axt  (2.2.7)

89 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

in terms of eigenvalues and (generalized) eigenvectors. Recall that if there exists a scalar such that
Au = u , u = 0 , is said to be an eigenvalue of A, and u is called the associated eigenvector. Moreover,
we call a vector v satisfying (A I)v = u a generalized eigenvector of A. (Note that the definitions

above imply (A I)u = 0 and (A I)2 v = 0 .) Thus, consider (2.2.7) and assume a solution of

the form xt = t u where = 0 . Then

t+1 u At u = 0

 (2.2.8)

(A I)u = 0

so is nothing but an eigenvalue belonging to A and u is the associated eigenvector. As is well known,

the eigenvalues may be computed from the relation

|A I| = 0  (2.2.9)

There are two cases to consider.

(A) If the n n matrix A is diagonalizable over the complex numbers, then A has n distinct
eigenvalues 1 , ..., n and moreover, the associated eigenvectors u1 , ..., un are linear

independent. Consequently, the general solution of the linear system (2.2.7) may be cast in
the form

xt = C1 t1 u1 + C2 t2 u2 + + Cn tn un  (2.2.10)

(B) If A is not diagonalizable (which may occur when A has multiple eigenvalues) we may proceed
in much of the same way as in the corresponding theory for continuous systems, see Grimshaw
(1990) and express the general solution in terms of eigenvalues and (generalized) eigenvectors.
Suppose that is an eigenvalue with multiplicity m and let u1 , ..., up be a basis for the

eigenspace of . If p = m we are done. If p < m we seek a solution of the form

xt = t (v + tu) where u is one of the ui s.

Then from (2.2.7) one easily obtains

(A I)v = u  (2.2.11a)

(A I)u = 0  (2.2.11b)

90
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

and after multiplying (2.2.11a) with (A I) from the left we arrive at

(A I)2 v = 0  (2.2.12)

Now suppose that we can find v1 , ..., vq such that v1 , ..., vq , u1 , ..., up are linear independent.
Now, if p + q = m we are done. If p + q < m we continue in the same fashion by seeking a
solution of the form xt = t (w + tv + (1/2!)t2 u) . In this case (2.2.7) implies
 
1
(A I)w = v + u  (2.2.13a)
2!

(A I)v = u  (2.2.13b)

(A I)u = 0  (2.2.13c)

which again leads to

(A I)3 w = 0  (2.2.14)

and we proceed in the same way as before. Either we are done or we keep on seeking solutions
where cubic terms of t are included. Sooner or later we will obtain the necessary number of linear

independent eigenvectors, cf. Meyer (2000).

Exercise 2.2.3.

a) Referring to the procedure outlined above suppose a cubic solution of the form
xt = t (y + tw + (1/2!)t2 v + (1/3!)t3 u) . Use (2.2.7) and deduce the following
relations: (A I)y = (w + (1/2!)v + (1/3!)u) , (A I)w = (v + (1/2!)u) ,

(A I)v = u , (A I)u = 0 , and moreover that (A I)4 y = 0 .

b) In general, assume a solution of degree m 1 on the form

m
t
 1
xt = tmi vi
(m i)!
i=1

and show that vi may be obtained from

(A I)v1 = 0

91
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

and

i
 1
(A I)vi+1 = vk , i = 1, 2, ..., m 1 .
k=1
(i (k 1))!


Remark 2.2.1. A complete treatment of case (B) should include a proof of linear independence of
the set of eigenvectors and generalized eigenvectors. However, such a proof requires a somewhat
deeper insight of linear algebra than assumed here and is therefore omitted.

Let us now illustrate the theory presented above through three examples. In Example 2.2.2 we deal with
the easiest case where the coefficient matrix A has distinct real eigenvalues. In Example 2.2.3 we consider
eigenvalues with multiplicity larger than one, and finally, in Example 2.2.4, we analyse the case where
the eigenvalues are complex conjugated.

Join the best at Top masters programmes

3
 3rd place Financial Times worldwide ranking: MSc
the Maastricht University International Business
1st place: MSc International Business
School of Business and 1st place: MSc Financial Economics
2nd place: MSc Management of Learning

Economics! 2nd place: MSc Economics

2nd place: MSc Econometrics and Operations Research
2nd place: MSc Global Supply Chain Management and
Change
Sources: Keuzegids Master ranking 2013; Elsevier Beste Studies ranking 2012;
Financial Times Global Masters in Management ranking 2012

Maastricht
University is
the best specialist
university in the
Visit us and find out why we are the best! Netherlands
(Elsevier)
Masters Open Day: 22 February 2014

www.mastersopenday.nl

92 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Example 2.2.2. Let

 
T 2 1
x = (x1 , x2 ) , A=

3 6

and solve xt+1 = Axt .

Assuming x = t u the eigenvalue equation (2.2.9) becomes

 
 2 1  = 0 2 8 + 15 = 0 1 = 5 , 2 = 3


 3 6 

The eigenvector u1 = (u1 , u2 )T belonging to 1 = 5 satisfies (cf. (2.2.8))

    
25 1 u1 0
=

3 6 5 u2 0
 
1
Hence, we choose u1 = .
3

In the same way, the eigenvector u2 = (u1 , u2 )T belonging to 2 = 3 satisfies

    
1 1 u1 0
=

3 3 u2 0
 
1
Thus u2 = . Therefore, according to (2.2.10), the general solution is
1
     
x1 t 1 t 1
xt = = C1 5 + C2 3
x2 t 3 1


Example 2.2.3. Let

2 1 1
x = (x1 , x2 , x3 )T , A= 0 2 2

0 0 2

and solve xt+1 = Axt .

Assuming xt = t u , we arrive at the eigenvalue equation

 
 2 1 1 
 = 0 (2 )3 = 0
 
 0
 2 2 
 0 0 2 

93
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

so we conclude that = 2 is the only eigenvalue and that it has multiplicity 3. Therefore, according

to (B) the general solution of the problem is

 
t t 1 2 t
xt = C1 u + C2 (v + tu) + C3 w + tv + t u
2

where = 2 and u , v and w must be found from (2.2.13a,b,c). Let u = (u1 , u2, u3 )T ,

v = (v1 , v2 , v3 )T and w = (w1 , w2 , w3 )T . (2.2.13c) implies

0 1 1 u1 0
u + u3 = 0
0 0 2 u2 = 0 2
2u3 = 0
0 0 0 u3 0

so u3 = 0 u2 = 0 and u1 is arbitrary so let u1 = 1 . Therefore u = (1, 0, 0)T . (2.2.13b)

implies

0 1 1 v1 1
0 0 2 v2 = 2 0 v2 + v3 = 2

2v3 = 0
0 0 0 v3 0

thus, v3 = 0 , v2 = 2 and v1 may be chosen arbitrary so we let v1 = 0 . This yields v = (0, 2, 0)T .

Finally, from (2.2.13a):

0 1 1 w1   1
0 0 2 w2 = 2 v + 1 u = 4 w2 + w3 = 1
2 2w3 = 4
0 0 0 w3 0

Hence, w3 = 2 , w2 = 1 and we may choose w1 = 0 so w = (0, 1, 2)T . Consequently, the

general solution may be written as

x1 1 0 1
t t
xt = x2 = C1 2 0 + C2 2 2 + t 0

x3 t 0 0 0

0 0 1
t 1 2
+ C3 2 1 +t 2
+ t 0
2
2 0 0


94
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Example 2.2.4. Let

 
T 2 1
x = (x1 , x2 ) , A=
1 2

and solve xt+1 = Axt .

Suppose xt = t v . (2.2.9) implies

 
 2 1  = 0 2 + 4 + 5 = 0


 1 2 

1 = 2 + i , 2 = 2 i (distinct complex eigenvalues).

Further: |1 | = 5 cos = (2)/ 5 sin = 1/ 5 so

(2)2 + 12 =

1 = 5(cos + i sin ) .

The eigenvector u = (u1 , u2)T corresponding to 1 may be found from

    
2 (2 + i) 1 u1 0 iu1 + u2 = 0
=
1 2 (2 + i) u2 0 u1 iu2 = 0

95 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

       
u1 i u1 i
Let u2 = t , u1 = it so =t , so we choose = as
u2 1 u2 1
eigenvector. Therefore (by use of Moivres formulae), the solution in complex form becomes
 t 
t
 
i 5 {i cos t + sin t}
xt = 5 (cos t + i sin t) = t
1 5 {cos t + i sin t}

Two linear independent real solutions are found by taking the real and imaginary parts
respectively:

   
x1r sin t t
= 5
x2r t
cos t
t cos t
   
x1i

= 5
x2i t sin t

Thus, the general solution may be written as

       t 
x1 x1r x1i 5 {C1 sin t C2 cos t}
xt = = C1 + C2 = t
x2 t
x2r t
x2i t 5 {C1 cos t + C2 sin t}


   
1 2 1 1
Exercise 2.2.4. Let x = (x1 , x2 )T , A= , B= and find the
3 2 2 1
general solution of

a) xt+1 = Axt ,

b) xt+1 = Bxt ,

c) Let x = (x1 , x2 , x3 )T and find the general solution of xt+1 = Cxt where

3 1 1
C = 7 5 1

6 6 2


We close this section by a definition and an important theorem about stability of linear systems.

Definition 2.2.1. The linear system (2.2.7) is globally asymptotic stable if limt xt = 0 . 

96
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Theorem 2.2.1. The linear system (2.2.7) is globally asymptotic stable if and only if all the
eigenvalues of A are located inside the unit circle |z| = 1 in the complex plane.

Proof: In case of distinct eigenvalues the result follows immediately from (2.2.10).

Eigenvalues with multiplicity m lead according to our previous discussion to terms in the solution
of form tq t where q m 1 .

Now, if || < 1 , let || = 1/s where s > 1 . Then by LHopitals rule: limt (tq /st ) = 0 so
the result follows here too. 

2.3 The Leslie matrix

In Part I of this book we illustrated many aspects of the theory which we established by use of the
quadratic map. Here in Part II we will use Leslie matrix models which are nothing but maps on the form
f : Rn Rn or f : Rn+1 Rn+1 .

Leslie matrix models are age-structured population models. They were independently developed in
the 1940s by Bernardelli (1941), Lewis (1942) and Leslie (1945, 1948) but were not widely adopted by
human demographers until the late 1960s and by ecologists until the 1970s. Some frequently quoted
papers where the use of such models plays an important role are: Guckenheimer et al. (1977), Levin and
Goodyear (1980), Silva and Hallam (1993), Wikan and Mjlhus (1996), Behncke (2000), Davydova et al.
(2003), Mjlhus et al. (2005), and Kon (2005). The ultimate book on matrix population models which
we refer to is Matrix population models by Hal Caswell (2001). Here we will deal with only a limited
number of aspects of these models.

Let xt = (x0,t , ..., xn,t )T be a population with n + 1 nonoverlapping age classes at time t .

x = x0 + + xn is the total population.

Next, introduce the Leslie matrix

f0 f1 fn

p0
0 0  (2.3.1)
A= 0




0 pn1 0

97
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

The meaning of the entries in (2.3.1) is as follows: fi is the average fecundity (the average number of

daughters born per female) of a member located in the i th age class. pi may be interpreted as the

survival probability from age class i to age class i + 1 and clearly 0 pi 1 . The relation between
x at two consecutive time steps (years) may then be expressed as

xt+1 = Axt  (2.3.2)

or in map notation

h : Rn+1 Rn+1 , x Ax  (2.3.3)

Hence, what (2.3.2) really says is that all individuals xi ( i > 1 ) in age class i at time t + 1 are the
survivors of the members of the previous age class xi1 at time t (i.e. xi,t+1 = pi1 xi1,t ), and since

the individuals in the lowest age class cannot be survivors of any other age class they must have originated
from reproduction (i.e. x0,t+1 = f0 x0,t + + fn xn,t ).

Need help with your

dissertation?
Get in-depth feedback & advice from experts in your
topic area. Find out what you can do to improve
the quality of your dissertation!

Get Help Now

Go to www.helpmyassignment.co.uk for more info

98 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Depending on the species under consideration, nonlinearities may show up on different entries in the
matrix. For example, in fishery models it is often assumed that density effects occur mainly through the
first year of life so one may assume fi = fi (x) . It is also customary to write fi (x) as a product of a
density independent part Fi and a density dependent part fi (x) so fi (x) = Fi fi (x) . In the following

we shall assume that every fertile age class has the same fecundity. Thus, we may drop the subscript i

and write fi (x) = f (x) . Frequently used fecundity functions are:

f (x) = F ex  (2.3.4)

which is often referred to as the overcompensatory Ricker relation and

F
f (x) =  (2.3.5)
1 + x

the compensatory Beverton and Holt relation.

Instead of assuming f = f (x) one may alternatively suppose f = f (y) where y = 0 x0 + + n xn

is the weighted sum of the age classes. If only one age class, say xi , contributes to density effects one
writes f = f (xi ) . In the case where an age class xi is not fertile we simply write Fi = 0 . (Species

where most age classes are fertile are called iteroparous. Species where fecundity is restricted to the last
age class only are called semelparous.)

The survival probabilities may of course also be density dependent so in such cases we adopt the same
strategy as in the fecundity case and write p() = P p() where P is a constant.

A final but important comment is that one in most biological relevant situations supposes p () 0
and f  () 0 . The standard counter example is when the All effect (cf. Caswell, 2001) is modelled.
Then one may use f  (x) 0 and/or p (x) 0 in case of small populations x . (All effects will not

be considered here.)

In the subsequent sections we shall analyse nonlinear maps and as already mentioned the theory will
be illustrated by use of (2.3.2), (2.3.3). However, if both fi = Fi and pi = Pi the Leslie matrix is linear

and we let

F0 Fn

P0 0 0
M =
0 

(2.3.6)


0 0 Pn1 0

99
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

We close this section by a study of the linear case

h : Rn+1 Rn+1 , x Mx  (2.3.7)

The eigenvalues of M may be obtained from |M I| = 0 .

Exercise 2.3.1.

a) Assume that M is 3 3 and show that the eigenvalue equation becomes

3 F0 2 P0 F1 P0 P1 F2 = 0
b) Generalize and show that if M is a (n + 1) (n + 1) matrix then the eigenvalue equation

may be written

n+1 F0 n P0 F1 n1 P0 P1 Pn1 Fn = 0  (2.3.8)

Next, we need some definitions:

Definition 2.3.1. A matrix A is nonnegative if all its elements are greater or equal to zero. It is

positive if all elements are positive.

Clearly, the Leslie matrix is nonnegative.

Definition 2.3.2. Let N0 , ..., Nn be nodes representing the n + 1 age classes in a population

model. Draw a directed path from Ni to Nj if individuals in age class i at time t contribute to
individuals of age j at time t + 1 including the case that a path may go from Ni to itself. A

diagram where all such nodes and paths are drawn is called a life cycle graph.

Definition 2.3.3. A nonnegative matrix A and its associated life cycle graph is irreducible if its
life cycle graph is strongly connected (i.e. if between every pair of distinct nodes Ni , Nj in the

graph there is a directed path of finite length that begins at Ni and ends at Nj ).

Definition 2.3.4. A reducible life cycle graph contains at least one age group that cannot contribute
by any developmental path to some other age group.

Examples of two irreducible Leslie matrices and one reducible one with associated life cycle graphs are
given in Figure 10.

100
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Exercise 2.3.2. Referring to Figure 10 write down the matrix and associated life cycle graph in
the case of four age classes where only the two in the middle are fertile.

F0 F1 F2
P0 0 0
0 P1 0

0 0 F2
P0 0 0
0 P1 0

F0 F1 0
P0 0 0
0 P1 0

Figure 10: Two irreduible and one reduible matries with corresponding life cycle graphs.

Brain power By 2020, wind could provide one-tenth of our planets

electricity needs. Already today, SKFs innovative know-
how is crucial to running a large proportion of the
worlds wind turbines.
Up to 25 % of the generating costs relate to mainte-
nance. These can be reduced dramatically thanks to our
systems for on-line condition monitoring and automatic
lubrication. We help make it more economical to create
cleaner, cheaper energy out of thin air.
By sharing our experience, expertise, and creativity,
industries can boost performance beyond expectations.
Therefore we need the best employees who can
meet this challenge!

The Power of Knowledge Engineering

Plug into The Power of Knowledge Engineering.

Visit us at www.skf.com/knowledge

101 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Definition 2.3.5. An irreducible matrix A is said to be primitive if it becomes positive when raised
to sufficiently high powers. Otherwise A is imprimitive (cyclic) with index of imprimity equal to
the greatest common divisor of the loop lengths in the life cycle graph.

Exercise 2.3.3. Show by direct calculation that the first irreducible Leslie matrix in Figure 10 is
primitive and that the second one is imprimitive (cyclic) with index of imprimity equal to 3.

Regarding nonnegative matrices the main results may be summarized in the following theorem which
is often referred to as the Perron-Frobenius theorem.

Theorem 2.3.1 (Perron-Frobenius).

1) If A is positive or nonnegative and primitive, then there exists a real eigenvalue 0 > 0 which

is a simple root of the characteristic equation |A I| = 0 . Moreover, the eigenvalue is

strictly greater than the magnitude of any other eigenvalue, 0 > |i | for i = 0 . The

eigenvector u0 corresponding to 0 is real and strictly positive. 0 may not be the only positive

eigenvalue but if there are others they do not have nonnegative eigenvectors.

2) If A is irreducible but imprimitive (cyclic) with index of imprivity d + 1 there exists a real
eigenvalue 0 > 0 which is a simple root of |A I| = 0 with associated eigenvector
u0 > 0 . The eigenvalues i satisfy 0 |i | for i = 0 but there are d complex eigenvalues
equal in magnitude to 0 whose values are 0 exp(2ki/(d + 1)) , k = 1, 2, ..., d .

For a general proof of Theorem 2.3.1 we refer to the literature. See for example Horn and Johnson (1985).

Concerning the Leslie matrix M (2.3.6) we shall study two cases in somewhat more detail: (I) the case
where all fecundities Fi > 0 , and (II) the semelparous case where Fi = 0 , i = 0, ..., n 1 but

Fn > 0 . In both cases it is assumed that 0 < Pi 1 for all i .

Let us prove Theorem 2.3.1 assuming (I):

Since Fn > 0 and 0 < Pi 1 it follows directly from (2.3.8) that = 0 is impossible. Therefore, we

may divide (2.3.8) by n+1 to obtain

F0 P0 F1 P0 P1 Pn1 Fn
f () = + 2 ++ = 1 (2.3.9)
n+1

102
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Clearly, lim0 f () = , lim f () = 0 , and since f  () < 0 for > 0 it follows that there

exists a unique positive 0 which satisfies f (0 ) = 1 . Therefore, assume 1

0 = e and rewrite (2.3.9)

as

f () = F0 e + P0 F1 e2 + + P0 P1 Pn1 Fn e(n+1) = 1  (2.43)

j = exp( + i) = e (cos + i sin ) for j = 1, ..., n and since 0 is unique and

Next, let 1

positive we may assume real and positive and = 2k , k = 1, 2, . . . . Then

p
j = e (cos p + i sin p) which inserted into f () , considering the real part only, gives
p

F0 e cos + P0 F1 e2 cos 2 + + P0 P1 Pn1 Fn e(n+1) cos(n + 1) = 1  (2.3.11)

Now, since is not a multiple of 2 it follows that cos j and cos(j + 1) cannot both be equal

to unity. Consequently, by comparing (2.3.10) and (2.3.11), we have e > e |j | < 0 for
j = 1, ..., n .

Finally, in order to see that the eigenvector u0 corresponding to 0 has only positive elements, recall that
u0 must be computed from Mu0 = 0 u0 , and in order to avoid u0 = 0 we must choose one of the
components of u0 = (u00 , ..., un0 )T free, so let u00 = 1 . Then from Mu0 = 0 u0 : P0 1 = 0 u10 ,

P1 u10 = 0 u20 , ..., Pn1un10 = 0 un0 which implies

P0 P1 u10 P0 P1 P0 Pn1
u10 = , u20 = = 2 un0 =
0 0 0 n0

which proves what we want.

(This proof is based upon Frauenthal (1986).) The proof of Theorem 2.3.1 under the assumption (II) is
left to the reader.

Let us now turn to the asymptotic behaviour of the linear map (2.3.7) in light of the results of Theorem
2.3.1.

In the case where all Fi > 0 we may express the solution of (2.3.7) (cf. (2.2.10)) as

xt = co t0 u0 + c1 t1 u1 + + cn tn un  (2.3.12)

103
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

where i (real or complex, 0 real) are the eigenvalues of M numbered in order of decreasing magnitude

and ui are the corresponding eigenvectors. Further,

 t  t
xt 1 n
t = c0 u0 + c1 u1 + + cn un
0 0 0

and since 0 > |i | , i = 0

xt
lim = c0 u0  (2.3.13)
t t0

Consequently, if M is nonnegative and primitive, the long term dynamics of the population are described
by the growth rate 0 and the stable population structure u0 . Thus 0 > 1 implies an exponential

increasing population, 0 < 0 < 1 an exponential decreasing population, where we in all cases have

the stable age distribution u0 .

If M is irreducible but imprimitive with index of imprimity d + 1 it follows from part 2 of the Perron-

Frobenius theorem that the limit (2.3.13) may be expressed as

d
xt 
lim t = c0 u0 + ck e(2k/(d+1))it ui  (2.3.14)
t 0
k=1

104 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

As opposed to the dynamical consequences of 1) in the Perron-Frobenius theorem we now conclude

from (2.3.14) that u0 is not stable in the sense that an initial population not proportional to u0 will

converge to it. Instead, the limit (2.3.14) is periodic with period d + 1 .

Figure 11: The hypothetical beetle population of Bernardelli as function of time. is the total population
, + and correspond to the zeroth, first and second age classes respectively. Clearly, there is no stable age distribution.

Example 2.3.1 (Bernardelli 1941). The first paper where the matrix M was considered came in
1941. There, Bernardelli considered a hypothetical beetle population obeying the equation

0 0 6
xt+1 = Bxt where B = 1/2 0 0
0 1/3 0

Clearly, B is irreducible and imprimitive with index of imprimity equal to 3 (cf. Exercise 2.3.2).
Moreover, the eigenvalues of B are easily found to be 1 = 1 and 2,3 = exp(2i/3) and it
is straightforward to show that B 3 = I so each initial age distribution will repeat itself in a regular
manner every third year as predicted by (2.3.14). In Figure 11 we show the total hypothetic beetle
population together with the three age classes as function of time, and clearly there is no stable
age distribution.

105
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

2.4 Fixed points and stability of nonlinear systems

In this section we turn to the nonlinear case x f (x) which in difference equation notation may
be cast in the form

x1,t+1 = f1 (x1,t , ..., xn,t )

 (2.4.1)
xn,t+1 = fn (x1,t , ..., xn,t )

Definition 2.4.1. A point x = (x1 , ..., xn ) which satisfies x = f (x ) is called a fixed point

for (2.4.1). 

Example 2.4.1. Assume that F0 + P0 F1 > 1 , x = x0 + x1 and find the nontrivial fixed point
(x0 , x1 ) of the two-dimensional Leslie matrix model (the Ricker model)

F0 ex F1 ex
    
x0 x0
 (2.4.2)
x1 P0 0 x1

According to Definition 2.4.1 the fixed point satisfies

x0 = F0 ex x0 + F1 ex x1  (2.4.3a)

x1 = P0 x0  (2.4.3b)

and if we insert (2.4.3b) into (2.4.3a) we obtain 1 = ex (F0 + P0 F1 ), hence the total equilibrium
population becomes x = ln(F0 + P0 F1 ) . Further, since x = x0 + x1 and x1 = P0 x0 we

easily find
 
1 P0
(x0 , x1 ) = x , x  (2.4.4)
1 + P0 1 + P0

(Note that F0 + P0 F1 > 1 is necessary in order to obtain a biological acceptable solution.)

Exercise 2.4.1. Still assuming F0 + P0 F1 > 1 , show that the fixed point (x0 , x1 ) of the

two-dimensional Beverton and Holt model

F0 F1
    
x0 x0
1+x 1+x  (2.4.5)
x1 P0 0 x1

106
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

becomes
 
1 P0
(x0 , x1 ) =
x,
x  (2.4.6)
1 + P0 1 + P0

where x = F0 + P0 F1 1 . 

Example 2.4.2. Find the nontrivial fixed point of the general Ricker model:

F0 ex Fn ex

P0 0 0

x0
x0

 (2.4.7)



x1 xn
0 0 Pn1 0

The fixed point x = (x0 , ..., xn ) obeys

x0 = ex (F0 x0 + + Fn xn )

x1 = P0 x0

xn = Pn1 xn1

107 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

From the last n equations we have x1 = P0 x0 , x2 = P1 x1 = P0 P1 x0 , xn = P0 Pn1 x0

which inserted into the first equation give

1 = ex (F0 + P0 F1 + P0 P1 F2 + + P0 Pn1 Fn )  (2.4.8)

Hence,
 n


x = ln(F0 + P0 F1 + + P0 Pn1 Fn ) = ln Fi Li

i=0

where Li = P0 P1 Pi1 and by convention L0 = 1 . From xi = x and



x1 = P0 x0 = L1 x0 , x2 = P0 P1 x0 = L2 x0 and xi = Li x0 we obtain
 
L1 Li Ln
(x0 , ..., xn ) = n x , n x , n x  (2.4.9)
i=0 Li i=0 Li i=0 Li
n
Again, i=0 Fi Li > 1 is required in order to have an acceptable biological equilibrium. 

Exercise 2.4.2. Generalize Exercise 2.4.1 in the same way as in Example 2.4.2 and obtain a formulae
for the fixed point of the n + 1 dimensional Beverton and Holt model. A detailed analysis of the
Beverton and Holt model may be obtained in Silva and Hallam (1992).

In order to reveal the stability properties of the fixed point x of (2.4.1) we follow the same pattern as
we did in Section 1.4. Let x = x + , then expand fi (x) in its Taylor series about x , taking the
linear terms only in order to obtain

f1 f1
x1,t+1 + 1,t+1 fi (xt ) + 1,t + + n,t
x1 xn

fn fn
xn,t+1 + n,t+1 fn (xt ) + 1,t + + n,t

x1 xn

where all derivatives are evaluated at x . Moreover, xi,t+1 = fi (xt ) . Consequently, the linearized

map (or linearization) of (2.4.1) becomes

f1 f1
(x ) (x )

1 x1
xn 1
 (2.4.10)


fn fn
n x1
(x ) xn
(x ) n

where the matrix is called the Jacobian.

108
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

If the fixed point x of (2.4.1) shall be locally asymptotic stable we clearly must have

lim t 0  (2.59)
t

and according to Theorem 2.2.1 this is equivalent to say:

Theorem 2.4.1. The fixed point x of the nonlinear system (2.4.1) is locally asymptotic stable if
and only if all the eigenvalues of the Jacobian matrix are located inside the unit circle |z| = 1
in the complex plane. 

Example 2.4.3.

a) Define F x = F0 x0 + F1 x1 and show that the fixed point (2.4.4) of the Ricker map (2.4.2)

is locally asymptotic stable provided

F x(1 + P0 ) > 0  (2.4.12a)

2F0 + F x(P0 1) > 0  (2.4.12b)

2P0 F1 + F0 P0 F x > 0  (2.4.12c)

b) Assume that F0 = F1 = F (same fecundity in both age classes) and show that (2.4.12b),
(2.4.12c) may be expressed as

1
F < e2/(1P0 )  (2.4.13b)
1 + P0
1
F < e(1+2P0 )/P0  (2.4.13c)
1 + P0

Solution:

a) Rewrite (2.4.2) as
x0 f1 (x0 , x1 ) = F0 ex x0 + F1 ex x1
x1 f2 (x0 , x1 ) = P0 x0

Then the Jacobian becomes

 
ex (F0 F x) ex (F1 F x)
J=  (2.4.14)
P0 0

109
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

and the eigenvalue equation |J I| = 0 may be cast in the form

F0 F x F1 F x
2 P0 =0 (2.4.15)
F0 + P0 F1 F0 + P0 F1

where we have used ex = (F0 + P0 F1 )1 .

(2.4.15) is a second order polynomial and || < 1 if the corresponding Jury criteria (2.1.14)
are satisfied. Therefore, by defining

F0 F x F1 F x
a1 = a2 = P0
F0 + P0 F1 F0 + P0 F1

we easily obtain from (2.1.14) that the fixed point is locally asymptotic stable provided the
inequalities (2.4.12a)-(2.4.12c) hold.

Challenge the way we run

EXPERIENCE THE POWER OF

FULL ENGAGEMENT

RUN FASTER.
RUN LONGER.. READ MORE & PRE-ORDER TODAY
RUN EASIER WWW.GAITEYE.COM

1349906_A6_4+0.indd 1 22-08-2014 12:56:57

110 Click on the ad to read more
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Remark 2.4.1: Scrutinizing the criteria, it is obvious that (2.4.12a) holds for any (positive)
equilibrium population x . It is also clear that in case of F x sufficiently small the same is true

for both (2.4.12b,c) as well which allow us to conclude that (x0 , x1 ) is stable in case of small

equilibrium population x . However, if F x becomes large, both (2.4.12b) and (2.4.12c) contain
a large negative term so evidently there are regions in parameter space where (2.4.12b) or (2.4.12c)
or both are violated and consequently regions where (x0 , x1 ) is no longer stable.

b) If F0 = F1 = F , then F x = F x , thus (2.4.15) may be expressed as

1 x
2 1 x
P0 =0 (2.4.16)
1 + P0 1 + P0

and the criteria (2.4.12b), (2.4.12c) simplify to

2 + x (P0 1) > 0

2P0 + 1 P x > 0

(2.4.13b) and (2.4.13c) are now established by use of x = ln[F (1 + P0 )] .

A final but important observation is that whenever 0 < P0 < 1/2 , (2.4.13b) will be violated
prior to (2.4.13c) if F is increased. On the other hand, if 1/2 < P0 1 , (2.4.13c) will be

violated first through an increase of F. (As we shall see later, this fact has a crucial impact of
the possible dynamics in the unstable parameter region.) 

Example 2.4.4 (Example 2.4.2 continued). Let the fecundities be equal (i.e. F0 = = Fn = F )

in the general n + 1 dimensional Ricker model that we considered in Example 2.4.2. Then,
n
x = ln(F D) , D = i=0 Li and the fixed point x may be written as x = (x0 , ..., xi , ..., xn )
where xi = (Li /D)x .

The eigenvalue equation (cf. (2.4.16)) may be cast in the form

n
1 
n+1
(1 x ) Li ni = 0  (2.4.17)
D i=0

Our goal is to show that the fixed point x is locally asymptotic stable whenever x < 2 (i.e.
that all the eigenvalues of (2.4.17) are located inside the unit circle.)

111
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

In contrast to Example 2.4.3, Theorem 2.1.9 obviously does not work here so instead we appeal
to Theorem 2.1.10 (Rouchs theorem). Therefore, assume |1 x | < 1 , let f () = n+1 ,
n
g() = (1/D)(1 x ) i=0 Li ni and rewirite (2.4.17) as f () + g() = 0 . Clearly, f
and g are analytic functions on and inside the unit circle C and the equation f () = 0 has
n + 1 roots inside C.

On the boundary we have

 
 1 n 
|g()| =  (1 x ) Li ni 
 
 D 
i=0
     
 L0 n
  L1 n1 
  Ln 

 (1 x )  +  (1 x )  + +  (1 x )
  
D D D

|1 x | < |f ()|

Thus, according to Theorem 2.1.10, f () + g() and f () have the same number of zeros inside

C, hence (2.4.17) has n + 1 zeros inside the unit circle which proves that x < 2 is sufficient to
guarantee a stable fixed point. Other properties of the Ricker model (2.4.7) may be obtained in
Wikan and Mjlhus (1996). 

Exercise 2.4.2 (Exercise 2.4.1 continued).

a) Consider the two-dimensional Beverton and Holt model (see Exercise 2.4.1) and show that
the fixed point (x0 , x1 ) is always stable. ( F0 = F1 = F .)

b) Generalize to n + 1 age classes. ( F0 = = Fn = F .) 

Exercise 2.4.3: Assume P0 < 1 and consider the two-dimensional semelparous Ricker model:

x0,t+1 = F1 ext x1
(2.4.18)
x1,t+1 = P0 x0
a) Compute the nontrivial fixed point (x0 , x1 ) .

b) Show that the eigenvalue equation may be written as

x1
2 + (1 x1 ) = 0
P0

and use the Jury criteria to conclude that (x0 , x1 ) is always unstable.

112
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

c) Show that

x0,t+2 = (P0 F1 )ext+1 x0,t  (2.4.19)

x1,t+2 = (P0 F1 )ext x1,t

d) Assume that there exists a two-cycle where the points in the cycle are on the form (A, 0) ,
(0, B) and show that the cycle is ((1/P0 ) ln(P0 F1 ), 0) , (0, ln(P0 F1 )) .

e) Show that the two cycle in d) is stable provided 0 < P0 F1 < e2 . 

Next, consider the general system (2.4.1) and its linearization (2.4.10) and let be the eigenvalues of

the Jacobian. We now define the following decompositions of Rn .

This e-book
is made with SETASIGN
SetaPDF

PDF components for PHP developers

www.setasign.com

113 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Definition 2.4.2.

E s is the subspace which is spanned by the (generalized) eigenvectors whose eigenvalues satisfy
|| < 1 .

E c is the subspace which is spanned by the (generalized) eigenvectors whose eigenvalues satisfy
|| = 1 .

E u is the subspace which is spanned by the (generalized) eigenvectors whose corresponding
eigenvalues satisfy || > 1 .

Rn = E s E c E u and the subspaces E s , E c and E u are called the stable, the center and
the unstable subspace respectively. 

By use of the definition above, the stability result stated in Theorem 2.4.1 may be reformulated as follows:

x = (x0 , ..., xn ) is locally asymptotic stable if E u = {0} and E c = {0} .

x is unstable if E u = {0} .

x = (x0 , ..., xn ) is called a hyperbolic fixed point if E c = {0} (cf. Section 1.4). ( x is attracting if
|| < 1 , repelling if || > 1 .)

We close this section by stating two general theorems which link the nonlinear behaviour close to a
fixed point to the linear behaviour.

Theorem 2.4.2 (Hartman-Grobman). Let f : Rn Rn be a C 1 diffeomorphism with a

hyperbolic fixed point x and let Df be the linearization. Then there exists a homeomorphism
h defined on some neighbourhood U on x such that

(h f )() = Df (x ) h()  (2.4.20)

for U . 

Theorem 2.4.3. There exists a stable manifold Wloc

s u
(x ) and an unstable manifold Wloc (x )
which are a) invariant, and b) is tangent to E s and E u at x and have the same dimension as

E s and E u . 

114
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

2.5 The Hopf bifurcation

There are three ways in which the fixed point x = (x0 , ..., x1 ) of a nonlinear map, f : Rn Rn

may fail to be hyperbolic. One way is that an eigenvalue of the linearization crosses the unit circle
(sphere) through 1. Then, in the generic case, a saddle-node bifurcation occurs. Another possibility is
that crosses the unit circle at 1 which in turn leads generically to a flip bifurcation. The third

possibility is that a pair of complex eigenvalues , cross the unit circle. In this case the fixed point

will undergo a Hopf bifurcation which we will now describe. Note that the saddle-node and the flip
bifurcations may occur in one-dimensional maps, f : R R . The Hopf bifurcation may take place

when the dimension n of the map is equal or larger than two. In this section we will restrict the analysis
to the case n = 2 only. Later on in section 2.7 we will show how both the flip and the Hopf bifurcation
may be analysed in case of n > 2 .

Theorem 2.5.1. Let f : R2 R2 be a C 3 two-dimensional one-parameter family of maps

whose fixed point is x = (x0 , x1 ) . Moreover, assume that the eigenvalues () , () of the
linearization are complex conjugates. Suppose that

|(0 )| = 1 but i (0 ) = 1 for i = 1, 2, 3, 4  (2.5.1)

and

d|(0 )|
= d = 0  (2.5.2)
d

Then, there is a sequence of near identity transformations h such that hf h1 in polar coordinates
may be written as

hf h1 (r, ) = ((1 + d)r + ar 3 , + c + br 2 ) +  (2.5.3)

Moreover, if a = 0 there is an > 0 and a closed curve of the form r = r () for 0 < <

which is invariant under f . 

Before we sketch a proof of the theorem let us give a few remarks.

Remark 2.5.1. Performing near identity transformations as stated in the theorem is also called
normal form calculations. Hence, formulae (2.5.3) is nothing but the original map written in
normal form.

115
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Remark 2.5.2. If d > 0 (cf. (2.5.2)) then the complex conjugated eigenvalues cross the unit circle
outwards which of course means that (x0 , x1 ) loses its stability at bifurcation threshold = 0 .

If d < 0 the eigenvalues move inside the unit circle. 

Remark 2.5.3. (0 ) = 1 or 2 (0 ) = 1 (cf. 2.5.1)) correspond to the well known saddle-node

or flip bifurcations respectively. 3 (0 ) = 1 and 4 (0 ) = 1 are special and are referred to as

the strong resonant cases. If is third or fourth root of unity there will be additional resonant
terms in formulae (2.5.3).

Remark 2.5.4. As is well known, if a saddle node bifurcation occurs at = 0 it means that in
case of < 0 there are no fixed points but when passes through 0 two branches of fixed

points are born, one branch of stable points, one branch of unstable points.

If the fixed point undergoes a flip bifurcation at = 0 we have (in the supercritical case) that
the fixed point loses its stability at = 0 and that a stable period 2 orbit is created.

Theorem 2.5.1 says that when (x0 , x1 ) undergoes a Hopf bifurcation at = 0 a closed invariant

curve surrounding (x0 , x1 ) is established whenever > 0 , | 0 | small. 

www.sylvania.com

We do not reinvent
the wheel we reinvent
light.
Fascinating lighting offers an infinite spectrum of
possibilities: Innovative technologies and new
markets provide both opportunities and challenges.
An environment in which your expertise is in high
demand. Enjoy the supportive working atmosphere
within our global group and benefit from international
career paths. Implement sustainable ideas in close
cooperation with other specialists and contribute to
influencing our future. Come and join us in reinventing
light every day.

Light is OSRAM

116 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Remark 2.5.5. Much of the theory of Hopf bifurcations for maps have been established by
Neimark and Sacker, cf. Sacker (1964, 1965) and Neimark and Landa (1992). Therefore, following
Kuznetsov (2004), the Hopf bifurcation is often referred to as the Neimark-Sacker bifurcation, see
for example Van Dooren and Metz (1998), King and Schaffer (1999), Kuznetsov (2004), Zhang
and Tian (2008), and Moore (2008). 

Sketch of proof, Theorem 2.5.1. Let (x0 , x1 ) be the fixed point of the two-dimensional map
x f (x) ( x = (x0 , x1 )T ) and assume that the eigenvalues of the Jacobian Df (x0 , x1 ) are

, = a1 a2 i . Next, define the 2 2 matrix T which columns are the real and imaginary
parts of the eigenvectors corresponding to the eigenvalues at the bifurcation. Then, after expanding
the right-hand side of the map in a Taylor series, applying the change of coordinates
(x0 , x1 ) = (x0 x0 , x1 x1 ) (in order to bring the bifurcation to the origin) together with
the transformations
       
x0 x x 1 x0

=T =T
x1 y y x1

our original map may be cast into standard form at the bifurcation as
      
x cos 2 sin 2 x R1 (x, y))
+  (2.5.4)
y sin 2 cos 2 y R2 (x, y)

where , equal exp(2i) , exp(2i) respectively, and = arctan(a2 /a1 ) . Our next

goal is to simplify the higher order terms R1 and R2 . This will be done by use of normal form
calculations (near identity transformations). The calculations are simplified if they first are
complexified. Thus we introduce

x = cos 2x sin 2y + R1 (x, y)

y  = sin 2x + cos 2y + R2 (x, y)
z = x + yi z  = x + y  i  R = R1 + R2 i
z = x + yi z =x +yi R = R1 + R2 i

and rewrite (2.5.4) as

f : C C, z f (z, z) = e2i z + R(z, z)  (2.5.5)

where the remainder is on the form

R(z, z) = R(k) (z, z) + O(|z|k+1)

117
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

(k) (k) (k)

Here, R(k) = r1 z k + r2 z k1 z + + rk+1 z k .

Next, define

z = Z(w) w = W (z) = Z 1 (z)  (2.5.6)

Then

z  = f (z))f (Z(w))  (2.5.7)

which in turn implies

w  = f(w) = Z 1 (z  ) = (Z 1 f Z)(w)  (2.5.8)

Now, we introduce the near identity transformation

z = Z(w) = w + P (k) (w)  (2.5.9)

and claim that

w = z P (k) (z) + O(|z|k+1 ) = W (z)  (2.5.10)

This is nothing but a consequence of (2.5.9). Indeed we have

w = z P (k) (w) = z P (k) (W (z))

= w + P (k) (w) P (k) (w + P (k) (w))

=w+ k

Thus, we may now by use of the relations

f (z) = e2i z + R(k) (z) + h.o.

Z(w) = w + P (k) (w)

Z 1 (z  ) = z  P (k) (z  ) + h.o.

(where h.o. means higher order) compute f(w) . This is done in two steps.

First,


z = (f Z)(w) = e2i w + e2i P (k) (w) + R(k) (w + ...)

118
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Then

f(w) = (Z 1 f Z)(w) = z  P (k) (z  ) + h.o.  (2.5.11)

= e2i w + e2i P (k) (w) + R(k) (w + ...) P (k) (e2i w) + h.o.

Next, we want to choose constants in order to remove as many terms in R(k) (w) as possible. To

this end let H k be polynomials of homogeneous degree k in w, w and consider the map

K : Hk Hk K(P ) = e2i P (w) P (e2i w)  (2.5.12)

Clearly, w l wkl is a basis for H k and we have

360
K(w l wkl ) = e2i w l wkl e2il w l e2i(kl) w kl

= e2i e2i(2lk) w l w kl

thinking .
 

= w l w kl

where k = 2, 3, 4, ... , 0 l k .

360
thinking . 360
thinking .
Discover the truth at www.deloitte.ca/careers Dis

Deloitte & Touche LLP and affiliated entities.

Discover the truth at www.deloitte.ca/careers Deloitte & Touche LLP and affiliated entities.

Deloitte & Touche LLP and affiliated entities.

Discover the truth

119 at www.deloitte.ca/careers
Click on the ad to read more
Download free eBooks at bookboon.com

Deloitte & Touche LLP and affiliated entities.

Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

From this we conclude that terms in R(k) (w) of the form w l wkl such that (, k, l) = 0

cannot be removed by near identity transformations. There are two cases to consider: (A)
irrational, and (B) rational.

(A) Assume irrational. Then = 0 2l = k + 1 thus k is an odd number. Here k = 1

corresponds to the linear term and the next unremoval terms are proportional to w 2 w and w|w|4
(i.e. third and fifth order terms).

(B) Supppose = /r rational, , r N , /r . Then = 0 (2l (k + 1))/r = m

where m Z . This implies (2l (k + 1)) = mr . Therefore r must be a factor in

(2l (k + 1)) . Thus the smallest k ( l = 0 ), equals r 1 which means that the first unremoval
terms are proportional to w r1 . When r = 2 the flip occurs. The cases r = 3, 4 which
corresponds to eigenvalues of third and fourth root of unity respectively are special (cf. Remark
2.5.3 after Theorem 2.5.1.)

Now, considering the generic case, irrational, we may through normal form calculations remove
all terms in R(k) except from those which are proportional to w 2 w and w|w|4 , hence (2.5.5)

may be cast into normal form as

z  = f (z) = e2i z(1 + + |z|2 ) + O(5)  (2.5.13)

where and are given complex numbers. Now introducing polar coordinates (r, ) , (2.5.13)
may after first neglecting terms of O(5) and higher and then neglecting terms of O(2 , r 2 , r 4 )
be expressed as

r  = r(1 + d + ar2 )  (2.5.14a)

 = + c + br 2  (2.5.14b)

which is nothing but formulae (2.5.3) in the theorem.

Finally, observe that the fixed point r of (2.5.14a) is


d

r =  (2.5.15)
a

120
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Figure 12: The outcome of a supercritical Hopf bifurcation. A point close to the unstable fixed point x moves away from x and
approaches the attracting curve (indicated by a solid line). In the same way an initial point located outside the curve is also attracted.

Thus, if a and d have opposite signs we obtain an invariant curve for > 0 . In case of equal

signs the curve exists for < 0 . Hence, the truncated map (2.5.14a) possesses an invariant curve.
Moreover, the eigenvalue of the linearization of (2.5.14a) is = 1 2d . Consequently,

whenever a < 0, > 0, d > 0 and d small, r is an attracting curve which corresponds to a

supercritical bifurcation. This is displayed in Figure 12. 

Remark 2.5.6. To complete the proof of Theorem 2.5.1 we must show that the full system (2.5.13)
possesses an invariant closed curve too. The basic idea here is to set up a graph transform of any
closed curve (containing higher order terms) near r and show that this graph transform has a
fixed graph close to r . However, in this procedure there are technical difficulties involved which
are beyond the scope of this book, cf. the original work by Sacker (1964). 

Referring to section 1.5 where we treated the flip bifurcation we stated and proved a theorem (Theorem
1.5.1) where we gave conditions for the flip to be supercritical. Regarding the Hopf bifurcation there
exists a similar theorem which was first proved by Wan (1978).

Theorem 2.5.2 (Wan). Consider the C 3 map K : R2 R2 on standard form

      
x cos sin x f (x, y)
+  (2.5.16)
y sin cos y g(x, y)

with eigenvalues , = ei . Then the Hopf bifurcation is supercritical whenever the quantity
d (cf. (2.5.2)) in Theorem 2.5.2 is positive and the quantity a (cf. (2.5.14a)) is negative. a may
be expressed as
2
 
(1 2) 1
a = Re 11 20 |11 |2 |02 |2 + Re(21 )  (2.5.17)
1 2

121
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

where

1
20 = [(fxx fyy + 2gxy ) + i(gxx gyy 2fxy )]
8
1
11 = [(fxx + fyy ) + i(gxx + gyy )]
4
1
02 = [(fxx fyy 2gxy ) + i(gxx gyy + 2fxy )]
8
1
21 = [(fxxx + fxyy + gxxy + gyyy ) + i(gxxx + gxyy fxxy fyyy )]
16

For a formal proof we refer to Wans original paper (Wan, 1978).

(The idea of the proof is simple enough: we start with the original map, write it on standard form (i.e.
(2.5.16)) and for each of the near identity transformations we then perform we express the new variables
in terms of the original ones, thereby obtaining a in (2.5.14a) expressed in terms of the original quantities.
The problem of course is that the calculations involved are indeed cumbersome and time-consuming as
formulae (2.5.17) suggests.)

We will turn your CV into

an opportunity of a lifetime

Do you like cars? Would you like to be a part of a successful brand? Send us your CV on
We will appreciate and reward both your enthusiasm and talent. www.employerforlife.com
Send us your CV. You will be surprised where it can take you.

122 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Example 2.5.1. Consider the stage-structured cod model proposed by Wikan and Eide (2004).

x1,t+1 = F ex2 ,t x2,t + (1 1 )x1,t  (2.5.18)

x2,t+1 = P x1,t + (1 2 )x2,t

Here the cod stock x is split into one immature part x1 and one mature part x2 . F is the density

independent fecundity of the mature part while measures the strength of cannibalism from
the mature population upon the immature population. P is the survival probability from the
immature stage to the mature stage and 1 , 2 are natural death rates. We further assume:

0 < P 1 , 0 < 1 , 2 < 1 , > 0 , F > 0 and F P > 1 2 .

Assuming x1 = x1,t+1 = x1,t and x2 = x2,t+1 = x2,t the fixed point of (2.5.18) is found to be
    
2 FP 1 FP
(x1 , x2 ) = ln , ln  (2.5.19)
P 1 2 1 2

The eigenvalue equation of the linearized map becomes (we urge the reader to work through the
details)

2 (2 1 2 ) + (1 1 )(1 2 ) 1 2 (1 x2 ) = 0  (2.5.20)

Now, defining a1 = (2 1 2 ) , a2 = (1 1 )(1 2 ) 1 2 (1 x2 ) and appealing

to the Jury criteria (2.1.14) it is straightforward to show that the fixed point is stable as long as
the inequalities

1 2 x2 > 0  (2.5.21a)

2(2 1 2 ) + 1 2 x2 > 0  (2.5.21b)

1 + 2 1 2 x2 > 0  (2.5.21c)

hold. Clearly, (2.5.21a) and (2.5.21b) hold for any positive x2 . Thus, there will never be a transfer
from stability to instability through a saddle-node or a flip bifurcation. (2.5.21c) is valid in case
of x2 sufficiently small. Hence, the fixed point is stable in case of small equilibrium populations.

However, if x2 is increased, as a result of increasing F which we from now on will use as our

bifurcation parameter, it is clear that (x1 , x2 ) will lose its stability at the threshold

1 + 2
x2 =
1 2  (2.5.22a)

123
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

or alternatively when

1 2 (1 +2 )/1 2
F = e  (2.5.22b)
P

Consequently, the fixed point will undergo a Hopf bifurcation at instability threshold and the
complex modulus 1 eigenvalues become

2 1 2 b
, = i  (2.5.23)
2 2

where b = 4(1 + 2 ) (1 + 2 )2 .


In order to show that the Hopf bifurcation is supercritical we have to compute d (defined through
(2.5.2)) and a (defined through (2.5.17)) and verify that d > 0 and a < 0 .

By first computing from (2.5.20) we find

(1 1 )(1 2 ) 1 2 (1 x2 )  (2.5.24)

|| =

which implies

d 1 1 2
|| = 
dF
2 (1 1 )(1 2 ) 1 2 (1 x2 ) F

and since the square root is equal to 1 at bifurcation and F is given by (2.5.22b) we obtain

d 1
|| = P e(1 +2 )/1 2 = d > 0  (2.5.25)
dF 2

which proves that the eigenvalues leave the unit circle through an enlargement of the bifurcation
parameter F.

In order to compute a we first have to express (2.5.18) on standard form (2.5.16). At bifurcation

the Jacobian may be written as

1
 
1 1 [(1 2 (1 + 2 )]
J= P  (2.5.26)
P 1 1

so by use of standard techniques the eigenvector (z1 , z2 )T belonging to is found to be

T
1 2

T b
(z1 , z2 ) = + i , 1 + 0i  (2.5.27)
2P 2P

124
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

and the transformation matrix T and its inverse may be cast in the form
2 1 b
   
2P 0 1
T = 2P T 1
= 2 1  (2.5.28)
1 0 2Pb b

The next step is to expand f (x2 ) = F ex2 up to third order. Then (2.5.18) becomes

1
x1,t+1 = f (x2 ) + f  (x2 )(x2,t x2 ) + f  (x2 )(x2,t x2 )2
2

1
+ f  (x2 )(x2,t x2 )3 x2,t + (1 1 )x1,t
6
x2,t+1 = Px,t + (1 2 )x2,t

and by introducing the change of coordinates (x1 , x2 ) = (x1 x1 , x2 x2 ) , in order to

bring the bifurcation to the origin, the result is

 
1
x1,t+1 = (1 1 )x1,t + 1 2 (1 x2 )x2,t 1 2 1 x2 x22,t
P P 2
2
 
1 3
+ 1 2 x x
P 2 6 2 2,t  (2.5.29a)

x2,t+1 = P x1,t + (1 2 )x2,t  (2.5.29b)

where all terms of higher order than three have been neglected.

AXA Global
Graduate Program
Find out more and apply

125 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Finally, by applying the transformations

       
x1 u u 1 x1
=T =T  (2.5.30)
x2 v v x2

we obtain after some algebra that the original map (2.5.18) may be cast into standard form as

2 1 2 b
ut+1 = ut vt
2 2
b 2 1 2
ut+1 = ut + vt + g(ut , vt )  (2.5.31)
2 2

where

2 2
   
2 2 1
g(u, v) = 1 2 1 x2 u 1 2 x u3
b 2 b 2 6 2

Now at last, we are ready to compute the terms in formulae (2.5.17)

4 12 2
guu = 1 2 A guuu = 1 2 B
b b

where A = 1 (/2)x2 , B = (1/2) (/6)x2 . This yields:

1 1 1 1
20 = iguu 11 = iguu 02 = iguu 21 = iguuu
8 4 8 16

and

2
 
2
(1 2) guu
Re 11 20 =
1 256(1 + 2 )
3(1 + 2 ) (2 u1 u2 )2 b2 2(2 1 2 )b2
   

so finally, by computing |11 |2 = (1/16)guu

2
, |02 |2 = (1/64)guu
2
, Re(21 ) = (1/32)bguuu

and inserting into (2.5.17) we eventually arrive at

2
(21 2 )2 + (1 + 2 ) (21 2 (1 + 2 ))2 1 2  (2.5.32)
  
a=
16(1 + 2 )

126
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

which is negative for all 0 < 1 , 2 < 1 . Consequently, the fixed point (2.5.19) undergoes a
supercritical Hopf bifurcation at the threshold (2.5.22a,b) (i.e. when (x1 , x2 ) fails to be stable
through an increase of F, a closed invariant attracting curve surrounding (x1 , x2 ) is established).

For further analysis of (2.5.18) we refer to the original paper by Wikan and Eide (2004) but also
confer Govaerts and Ghaziani (2006) where a numerical study of the model may be obtained.

In the next exercise most of the cumbersome and time-consuming calculations we had to perform in
Example 1.5.1 are avoided.

Exercise 2.5.1. Assume that the parameter > 1 and consider the map
   
x y
 (2.5.33)
y y(1 x)

a) Show that the nontrivial fixed point

1 1
 

(x , y ) = ,

b) Compute the Jacobian and show that the eigenvalue equation may be expressed as

2
+1 =0

c) Use the Jury criteria (2.1.14) and show that the fixed point is stable whenever 1 < < 2

and that a Hopf bifurcation occurs at the threshold = 2 .

d) Show that || = 1 and moreover that

d
||=2 > 0
d

which proves that the eigenvalues leave the unit circle at bifurcation threshold.

e) Assuming = 2 , apply the change of coordinates (x, y) = (x (1/2), y (1/2))

together with the transformations
       
x u u 1 x
=T =T

y v v y

127
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

where
 
1 3
T = 2 2

1 0

(verify that the columns in T are the real and imaginary parts of the eigenvectors belonging to
the eigenvalues of the Jacobian respectively) and show that (2.5.33) may be written on standard
form at bifurcation threshold as
 
1 3
    
u u f (u, v)
2 2 +  (2.110)
v 3 1 v g(u, v)
2 2

where f (u, v) = u2 3 uv and g(u, v) = (1/ 3)u2 + uv .

f) Referring to Theorem 2.5.2 show that the quantity a defined in (2.5.17) is negative, hence that
in case of > 2 , | 2| small, there exists an attracting curve surrounding the unstable
fixed point (x , y ) . 

e Graduate Programme
I joined MITAS because for Engineers and Geoscientists
I wanted real responsibili www.discovermitas.com
Maersk.com/Mitas e G
I joined MITAS because for Engine
I wanted real responsibili Ma

Month 16
I was a construction Mo
supervisor ina const
I was
the North Sea super
advising and the No
Real work he
helping foremen advis
International
al opportunities
Internationa
ree wo
work
or placements ssolve problems
Real work he
helping fo
International
Internationaal opportunities
ree wo
work
or placements ssolve pr

128 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Exercise 2.5.2 (Strong resonant case I). Consider the two-age structured population model

(x1 , x2 ) (F2 x2 , P ex1 x1 )  (2.5.35)

where 0 < P 1 , F2 > 0 and P F2 > 1 .

a) Show that the fixed point (x1 , x2 ) = (ln(P F2 ), (1/F2 ) ln(P F2 )) .

b) Show that the eigenvalue equation may be cast in the form 2 + x1 1 = 0 and further that
a Hopf bifurcation takes place at the threshold x1 = 2 (or equivalently when
F2 = (1/P ) exp(2)) .

c) Show that equals fourth root of unity at bifurcation threshold.

Note that the result obtained in c) violates assumption (2.5.1) in Theorem 2.5.1 which of course
means that neither Theorem 2.5.1 nor Theorem 2.5.2 applies on map (2.5.35). We urge the reader
to perform numerical experiments where F2 > (1/P ) exp(2) in order to show that when
(x1 , x2 ) fails to be stable, an exact 4-periodic orbit with small amplitude is established. (For
further reading, cf. Wikan (1997.) 

Exercise 2.5.3 (Strong resonant case II). Repeat the analysis from the previous exercise on the map

(x1 , x2 ) F e(x1 +x2 ) (x1 + x2 ), x1  (2.5.36)

 

Hint: equals third root of unity at bifurcation threshold. 

As is shown in the sketch of proof of Theorem 2.5.1 most terms in (2.5.5) may be removed by a series
of near identity transformations. In the next exercise the reader is actually asked to perform such
transformations.

Exercise 2.5.4. Let j = 1 , j = 1, 2, 3, 4, 5 and consider

(i) zt+1 = zt + 1 z2t + 2 zt zt + 3 z2t + O(3)

a) Apply the near identity transformation (cf. (2.5.9))

z = w + 1 w 2 + 2 ww + 3 w2

129
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

together with (cf. (2.5.10))

w = z (1 z 2 + 2 zz + 3 z 2 )

and show that (i) may be written as

(ii) wt+1 = wt + (1 + 1 1 2 )wt2

2
+ (2 + 2 2 )wt w t + (3 + 3 3 )w2t + O(3)

b) Show that if we choose

1 2 3
1 = 2 = 3 = 2
(1 ) (1 )

then all second order terms in (ii) will disappear. Thus, after one near identity transformation
we have a system on the form (where we for notational convenience still use z as variable)

(iii) zt+1 = zt + 1 z3t + 2 z2t zt + 3 zt z2t + 4 z3t + O(4)

c) Apply

z = w + a1 w 3 + a2 w 2 w + a3 ww2 + a4 w 3

w = z (a1 z 3 + a2 z 2 z + a3 zz 2 + a4 z 3 )

on (iii) and show that if we choose

1 2 4
a1 = a3 = 2 a4 = 3
(1 2 ) (1 )

then the w 3 , ww2 and w 3 terms will disappear. Note that we cannot use

2
a2 =
(1 )

because 1 = 0 for any located on the boundary of the unit circle.

d) After two near identity transformations our system is on the form

(iv) zt+1 = zt + 2 z2t zt + O(4)

130
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Write out all fourth order elements and perform a new near identity transformation in the
same way as in a) and c) and show that all fourth order terms may be removed, hence that
our system may be cast in the form (normal form!)

(v) zt+1 = zt + 2 z2t zt + O(5)

Remark 2.5.7. Note that Exercise 2.5.4 in many respects offers an equivalent way of establishing the
normal form (2.5.13). Moreover, if 3 = 1 , the denominator in the expression for 3 becomes zero,

hence the terms w 2t in (ii) is not removable. Consequently, there will be an additional resonant term
4
on the form z 2t in (v). In case of 4 = 1 or 5 = 1 the additional terms are z 3t and z t respectively.

For further reading we refer to Kuznetsov (2004) and Kuznetsov and Meijer (2005). 

We close this section by once again emphasizing that the outcome of a supercritical Hopf bifurcation is
that when the fixed point fails to be stable an attracting invariant curve which surrounds the fixed point
is established. In section 2.8 we shall focus on the nonstationary dynamics on such a curve as well as
possible routes to chaos. However, before we turn to those questions we shall in section 2.6 present an
analysis of the Horseshoe map where we once again invoke symbolic dynamics and in section 2.7 we
170x115_MIM_ datos_Espv4.pdf 1 17/12/15 18:29
shall explain how we may analyse the nature of bifurcations in higher dimensional problems.

EL 93%
DE LOS ALUMNOS DEL MIM ESTN
TRABAJANDO EN SU SECTOR
C
A LOS 3 MESES DE GRADUARSE
M

MASTER IN MANAGEMENT
CM

MY

CY ESTUDIA EN EL CORAZN DE MADRID Y APROVECHA LAS OPORTUNIDADES Y EXPERIENCIAS NICAS

Duracin: 1O MESES
QUE OFRECE LA CAPITAL DE ESPAA
CMY Exp. Profesional: 1 AO
POTENCIA TU FORMACIN CURSANDO UNA DOBLE TITULACIN QUE MEJOR SE ADAPTE A TUS
Idioma: ESPAOL / INGLS
K
OBJETIVOS PROFESIONALES
Formato: TIEMPO COMPLETO
ESTUDIA UN SEMESTRE EXTRA EN OTRO PAS CON EL PROGRAMA BEYOND BORDERS EXPERIENCE
Convocatorias: SEPT / FEB
Y AUMENTA TU RED DE CONTACTOS

5 ESPECIALIZACIONES 100 DEL MUNDO 55 NACIONALIDADES

MASTERS IN MANAGEMENT
PERSONALIZA TU PROGRAMA FINANCIAL TIMES
EN CLASE

www.ie.edu/master-management [email protected] Sguenos en IE MIM Experience

131 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

2.6 Symbolic dynamics III (The Horseshoe map)

As we have seen maps may possess both fixed points and periodic points, and through Theorem 2.5.1 we
have established that the dynamics may be restricted to invariant curves as well. However, in Part I our
analysis also revealed other types of invariant hyperbolic sets. To be more concrete we showed in Section

1.9 that whenever > 2 + 5 the quadratic map possessed an invariant set of points (a Cantor set)
that never left the unit interval through iterations. Our next goal is to discuss a similar phenomenon in
case of a two- dimensional map, the Horseshoe map, which is due to Smale (1963, 1967). There are several
ways of visualizing the Horseshoe. We prefer the way presented in Guckenheimer and Holmes (1990),

(a)

H f(H) H
U

H0 1
1
(b) f

H1 1

1
f(H) H f ( f(H) H )
U U

H0
2
(c) f

H1

Figure 13: a) The Horseshoe map f . b) The inverse f 1. c) The image of four thin horizontal strips under f 2.

132
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Example 2.6.1 (The Horseshoe map). Consider the unit square H = [0, 1] [0, 1] , see Figure

13a, and assume that we perform two operations on H: (1) a linear expansion of H by a factor
, > 2 , in the vertical direction and a horizontal contraction by a factor , 0 < < 1/2 .
(2) A folding in such a way that the folding part falls outside H. The whole process is displayed
in Figure 13a. We call this a map f : H R2 and restricted to H we may express the two vertical

strips as f (H) H .

If we reverse the process (folding, stretching and contracting) we see from Figure 13b that we
obtain two horizontal strips H0 and H1 and each of them has thickness 1 . Also note that the
inverse image may be expressed as f 1 (f (H) H) = f 1 (H) H . Thus we conclude that
on each of the horizontal strips H0 and H1 , f stretches by a factor in the vertical direction
and contracts by a factor in the horizontal direction.

As is clear from Figures 13a,b and the text above, when f is iterated most points will leave H
after a finite number of iterations. However, as we shall see (just as we did in the corresponding
one-dimensional example in Section 1.9) there is a set

= {x | f i (x) H} iZ

133
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

which never leaves H. Now, let us describe the structure of . First, observe that f stretches both
H0 and H1 vertically by so that f (H0 ) and f (H1 ) both intersect H0 and H1 (Figure 13b).
Therefore, points in H0 must have been mapped into H0 by f from two thinner strips, each of

width 2 contained in H0 . The same is of course true for points in H1 , so after applying f
twice on the four horizontal strips of widths 2 in Figure 13c the result is four thin vertical

strips each of width 2 as displayed in Figure 13c. (Note, that after only one iteration of f the
result is four rectangles, each of height 1 and width .) Moreover, since
H0 H1 = f 1 (H f (H)) the union of the four thinner strips may be written as
f 2 (H f (H) f 2 (H)) , and proceeding in this way f n (H f (H) f n (H)) must
n n
be the union of 2 such strips where each strip has a thickness of . Since > 2 the thickness
of each of the 2n strips goes to zero when n . Now, consider one of the 2n horizontal strips.
Each time f is applied on the strip it is stretched by in the vertical direction and contracted
by in the horizontal direction so the image under f n must be a strip of length 1 in the vertical
direction and width n in the horizontal direction, and since 0 < < 1/2 the latter tends to

zero as n . Thus, the 2n horizontal strips are mapped into 2n vertical strips. The points

that will remain in H forever are those points which are located both in the horizontal and the
vertical strips, hence is nothing but the intersection of the horizontal and vertical strips.
Moreover, is a Cantor set. Indeed, when n , contains just points (no intervals) and

these points are not isolated but they are accumulation points of (cf. Definition 1.9.5).

In order to describe the dynamics on let us invoke symbolic dynamics in much of the same
way as we did in Section 1.9 and assign a sequence a = {ai }
i= to every point x . We

also define another sequence b through bi = ai+1 . Thus : 2 2 . (a) = b is the shift
map. The itinerary of x , : 2 is defined as (x) = . . . a2 a1 a0 a1 a2 . . . and we let
ai = 0 if f i (x) H0 and ai = 1 if f i (x) H1 which means that x if and only if
f i (x) Hai for every i .

134
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

First, observe that since f i+1 (x) = f i (f (x)) then (f (x)) = b which proves that acts in

the same way as . Consequently, if we are able to prove that is a homeomorphism then
(according to Definition 1.2.1), f and are topological equivalent maps on . The 1 1 and
continuity properties of may be proved along the following line. Let
SV = SV (bm , bm+1 , ..., b1 ) be the central set of x s such that f i (x) is contained in one of
the 2m vertical strips in H f (H) f m (H) and let SH (b0 , ..., bn ) be the central set of
x s contained in a horizontal strip. Then S = S(bm , ..., b0 , ...bn ) = SV SH is the set of x s
such that f i (x) Hbi and clearly S must be a rectangle with height (n+1) and width m .
When n, m all areas 0 . Consequently, is both continuous and 1 1 .

Regarding the onto property, following Guckenheimer and Holmes (1990), it suffices to prove that
n+1
S is nonempty. To see this, observe that f (SH (b0 , ..., bn )) fills the entire S in the vertical
direction. In particular it intersects both S0 and S1 so that SH (b0 , ..., bn , bn+1 ) must be a
nonempty horizontal strip. Moreover, every vertical strip SV intersects every SH which

immediately implies that S = SV SH is nonempty. Consequently,

f = (2.6.1)

whenever f is restricted to . 

Before we leave the Horseshoe map let us emphasize and comment on a few more topics. First, note the
difference between the symbol sequence {ai }
i= defined for the horseshoe and the sequence {ai }i=0

we used in our study of the quadratic map in case of > 2 + 5 (see Section 1.9). Unlike the quadratic
map (1.2.1), the two-dimensional horseshoe map is invertible (Figure 13b) so it makes sense to consider
backward iteration. Therefore we may use negative indices in order to say which vertical strip f (x) is

located in and positive indices in order to say which horizontal strip f (x) is contained in. If we glue

together {ai }1
i= and {ai }i=0 we have a description of the whole orbit.

The shift map which in this context often is referred to as the two-sided shift, may be defined as in
Example 2.6.1 or in the usual manner as

(...a2 a1 a0 a1 a2 ...) = (...a2 a1 a0 a1 a2 ...)  (2.6.2)

135
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

(cf. Definition 1.9.3). The inverse is defined through

1 (...a2 a1 a0 a1 a2 ...) = (...a2 a1 a0 a1 a2 ...)  (2.6.3)

Periodic points of period N for may be expressed as before. For example, a 3-period orbit may be
expressed by the sequence c = {...010010010...} and 3 (c) = c . Moreover, since each element of
{ai } may take two values (0 or 1) a period n orbit for corresponds to 2n periodic points. From this
we may conclude that if n has 2n periodic points in 2 , then from (3.6.1) f n = 1 n has

2n periodic points in . Actually, these periodic points are unstable points of the saddle type. In order
to see this, observe that segments contained in H0 and H1 are compressed horizontally by
( 0 < < 1/2 ) and stretched by ( > 2 ) in the vertical direction. This means that f

restricted to H f 1 (H) is linear so the Jacobian becomes J = diag(, ) and if we apply f n on

one of the 2n horizontal strips described in Example 2.5.2 the resulting Jacobian may be expressed as
Df n = diag( n , n ) . Consequently, the eigenvalues are 1 = and 2 = , and since 1,2 are real
and 1 is located on the inside of the unit circle and 2 on the outside the periodic points are saddle

points.

136 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

The distance function (cf. Proposition 1.9.1) between two sequences a and b in 2 is defined as

 |ai bi |
d[a, b] =  (2.6.4)
i=
2|i|

where |ai bi | = 0 if ai = bi and |ai bi | = 1 if ai = bi . The fact that periodic points for are
dense in 2 may be obtained from (2.6.4) and by use of the same method as in the proof of Proposition
1.9.1. We leave the details to the reader. There are also nonperiodic points for in 2 which are dense
in 2 . In order to show this we must prove that the orbit of such a point comes arbitrarily close to any

given sequence in 2 . Thus, let a = (...ak ...a0 ...ak ...) be a given sequence and let b be a sequence
whose central block equals the central block of a (i.e. ak = bk , ...a0 = b0 , ...ak = bk ) . Then, from

(2.6.4):

k1
 |ai bi |  |ai bi |  |ai bi | 1 1
d[a, b] = |i|
= |i|
+ i
k + k = 21k
i=
2 i=
2 i=k+1
2 2 2

Hence, when k becomes large, b a so according to Definition 1.9.4, b represents a dense orbit in
2 .

Finally, let us give a few comments on stable and unstable sets of points in . In general, two points x1
and x2 are said to be forward asymptotic in a set S if f n (x1 ) S , f n (x2 ) S for all n and

lim |f n (x1 ) f n (x2 )| = 0  (2.6.5a)

n

If f n (x1 ) S , f n (x2 ) S for all n and

lim |f n (x1 ) f n (x2 )| = 0  (2.6.5b)

n

then x1 , x2 are said to be backward asymptotic in S . By use of (2.6.5a,b) we may define the stable set
of a point x in S as

W S (x) = {y || f n (x) f n (y)| 0 n }  (2.6.6a)

and the unstable set as

W U (x) = {z || f n (x) f n (z)| 0 n }  (2.6.6b)

137
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

The shift map makes it easy to describe W S (x) and W U (x) . For example, if x is a fixed point of f
and (x ) = (...a2 a1 a0 a1 a2 ...) then any point y whose itinerary is the same as the itinerary of
x to the right of an entry ai is contained in W S (x ) . (2.6.6a) allows us to describe the stable set of
points in . Indeed, let x be a fixed point of f which lies in H0 . Then (x ) = {...0000...} . Then,

since f contracts in the horizontal direction, any point which is located in a horizontal segment through
x must be in W S (x ) . But there are also additional points in W S (x ) . In fact, any point p which
eventually is mapped into the horizontal segment through x after a finite number of iterations k is

also contained in W S (x ) because |f k+n (p) x | < n . This implies that the union of all horizontal
intervals given by f n (l) , n = 1, 2, 3, ... , (where l is a horizontal segment) lies in W S (x ) . We leave
to the reader to describe the set W U (x ) .

2.7 The center manifold theorem

Recall that in our treatment of the flip bifurcation (cf. section 1.5) we considered one-dimensional maps
of the form f : R R and when we studied the Hopf bifurcation in section 2.5 the main theorems

were stated for two-dimensional maps f : R2 R2 . Let us now turn to higher-dimensional maps,
f : Rn Rn . Of course, || = 1 at bifurcation in these cases too but how do we determine the nature
of the bifurcation involved when the fixed point fails to be hyperbolic?

The main conclusion is that there exists a method which applied to a map on the form f : Rn Rn
reduces the bifurcation problem to a study of a map g : R2 R2 (Hopf), or g : R R (flip). The

cornerstone in the theory which allows this conclusion is the center manifold theorem for maps which
we now state.

Theorem 2.7.1 (Center manifold theorem). Let f : Rn Rn be a C k , k 2 map and assume

that the Jacobian Df (0) has a modulus 1 eigenvalue and, moreover, that all eigenvalues of Df (0)

splits into two parts c , s such that


1 if c
|| =
< 1 if s

138
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Further, let Ec be the (generalized) eigenspace of c , dim Ec = d < . Then there exists a

domain V about 0 in R and a C k submanifold W c of V of dimension d passing through 0

n

which is tangent to Ec at 0 which satisfies:

I) If x W c and f (x) V then f (x) W c .

II) If f (n) (x) V for all n = 0, 1, 2, ... then the distance from f (n) (x) to W c approaches
zero as n . 

For a proof of Theorem 2.7.1, cf. Marsden and McCracken (1976, p. 28 43). Also cf. the book by
Iooss (1979) and the paper by Vanderbauwhede (1987).

When Ec has dimension two, as it does for the Neimark-Sacker case at criticality, the essence of Theorem
2.7.1 is that there exists an invariant manifold of dimension 2 Rn which has the eigenspace belonging
to the complex eigenvalues as tangent space at the bifurcating nonhyperbolic fixed points. In case of flip
bifurcation problems, dim W C = 1 . Thus close to the bifurcation, our goal is to restrict the original
map to the invariant center manifold W C and then proceed with the analysis by using the results in
Theorems 2.5.1 and 2.5.2 in case of Hopf bifurcation problems and Theorem 1.5.1 in the flip case.

Excellent Economics and Business programmes at:

The perfect start

of a successful,
international career.

CLICK HERE
to discover why both socially
and academically the University
of Groningen is one of the best
places for a student to be
www.rug.nl/feb/education

139 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Let us now in general terms describe how such a restriction may be carried out. To this end, consider
our discrete system written in the form

xt+1 = Axt + F (xt , yt )  (2.7.1)

yt+1 = Byt + G(xt , yt )

where all the eigenvalues of A are on the boundary of the unit circle and those of B within the unit circle.
(If the system we want to study is not on the form as in (2.7.1) we first apply the procedure in Example
2.5.1, see also the proof of Theorem 2.5.1.)

Now, since the center manifold W C is tangent to the (generalized) eigenspace Ec , we may represent

it as a local graph

W C = {(x, y)/y = h(x)} h(0) = Dh(0) = 0  (2.7.2)

and by substituting (2.7.2) into (2.7.1) we have

yt+1 = h(xt+1 ) = h(Axt + F (xt , h(xt )) = Bh(xt ) + G(xt , h(xt ))

or equivalently

h(Ax + F (x, h(x))) Bh(x) G(x, h(x)) = 0  (2.7.3)

An explicit expression of h(x) is out of reach in most cases, but one can approximate h by its Taylor

series at the bifurcation as

h(x) = ax2 + bx3 + O(x4 )  (2.7.4)

where the coefficients a, b are determined through (2.7.3), and finally the restricted map is obtained
by inserting the series of h into (2.7.1).

Example 2.7.1. Consider the Leslie matrix model

F (1 x)1/ F (1 x)1/
    
x1 x1
f :R R
2 2
 (2.125)
x2 P 0 x2

where x = x1 + x2 is the total population. 

140
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

(2.7.5) is often referred to as the Deriso-Schnute population model. Note that if 0 , (2.7.5) is nothing

but the Ricker model (see (2.3.4) and Examples 2.4.1 and 2.4.3). If = 1 we are left with the Beverton

and Holt model (see (2.3.5) and Exercise 2.4.1).

Our goal is to show that under the assumptions F (1 + P ) > 1 , 0 < P < 1/2 , > (1 P )/2
the fixed point (x1 , x2 ) of (2.7.5) will undergo a supercritical flip bifurcation at instability threshold.

We urge the reader to verify the following properties:

 
1 1
(x1 , x2 ) = x , x  (2.7.6)
1+P 1+P

where x = (1/)[1 (P + P F ) ] . Defining f (x) = F (1 x)1/ the Jacobian becomes

f  x + f f  x + f
 

P 0

where f = f (x ) = 1/(1 + P ) and f  = f  (x ) .

Show by use of the Jury criteria (2.1.14) that whenever 0 < P < 1/2 , > (1 P )/2 the fixed

point (2.7.6) will undergo a flip bifurcation when f  x = 2/(1 P 2 ) and that the Jacobian at

bifurcation threshold equals

1 1
 
1P 1P
 (2.7.7)
P 0

and moreover, that the eigenvalues of (2.7.7) are 1 = 1 and 2 = P/(1 P ) .

Now, in order to show that the flip bifurcation is of supercritical nature we must appeal to Theorem 1.5.1
but since that theorem deals with one-dimensional maps, we first have to express (2.7.5) on the appropriate
form (2.7.1) and then perform a center manifold restriction as explained through (2.7.2)-(2.7.4).

The form (2.7.1) is achieved by performing the same kind of calculations as in Example 2.5.1. The
eigenvectors belonging to 1 and 2 are easily found to be (1/P, 1)T and (1/(1 P ), 1)T

respectively so the transformation matrix T and its inverse become

 
P (1P ) P
P1 1
 
1P
T = T 1 = 2P 1 2P 1
 (2.7.8)
1 1 P2P(1P )
1
1P
2P 1

141
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Further, by expanding f up to third order, i.e.

1 1
f (x) f (x ) + f  (x)(x x ) + f  (x )(x x )2 + f  (x )(x x )3
2 6

and applying the change of coordinates (x1 , x2 ) = (x1 x1 , x2 x2 ) , using the fact that
f  x = 2/(1 P 2 ) at bifurcation threshold gives

1 1
x1,t+1 = x1,t x2,t + {1}x2t + {2}x3t  (2.7.9)
1P 1P

x2,t+1 = P x1,t

where all terms of higher order than 3 have been neglected and {1} and {2} are defined through

1  1  1 
{1} = f  + f x {2} = f + f x
2 2 6

Now, performing the transformations

       
x1 u u 1 x1

=T =T
x2 v v x2

American online
LIGS University
is currently enrolling in the
Interactive Online BBA, MBA, MSc,
DBA and PhD programs:

enroll by September 30th, 2014 and

save up to 16% on the tuition!
pay in 10 installments / 2 years
Interactive Online education
visit www.ligsuniversity.com to
find out more!

Note: LIGS University is not accredited by any

nationally recognized accrediting agency listed
by the US Secretary of Education.
More info here.

142 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

on (2.7.9) we arrive at

ut+1 = ut + g(ut, vt )  (2.7.10)

P
vt+1 = vt g(ut, vt )
1P

where g(u, v) = A[(1 P )2 u + P 2 v]2 + B[(1 P )2 + P 2 v]3

1 1
A= {1} B= {2}

P (2P 1)(1 P ) P 2 (2P 1)(1 P )2

and we observe that (2.7.10) is nothing but the original map (2.7.5) written on the desired form (2.7.1).

The next step is to restrict (2.7.10) to the center manifold. Thus, assume

v = i(u) = Ku2 + Lu3  (2.7.11)

By use of (2.7.3) we now have

P
i(ut + g(ut , vt )) + i(ut ) + g(ut, i(ut )) = 0
1P

which is equivalent to
 
PK
K+ + (1 P ) A u2 +
4
1P
 
PL
2KA(1 P ) L + 2AP (1 P ) K + B(1 P ) u3 = 0
4 2 2 6

1 P

from which we obtain

K = (1 P )5 A L = (1 P )7 [B + 2A2 (1 P )(1 2P )]

Finally, by inserting v = Ku2 + Lu3 into the first component of (2.7.10) the restricted map may be

cast in the form

ut+1 = h(ut ) = ut + A(1 P )4 u2t

+ (1 P )6 [B 2A2 P 2 (1 P )]u3t + O(u4)  (2.7.12)

143
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Since u h(u) is a one-dimensional map we may now proceed by using Theorem 1.5.1 in order to

show that the flip bifurcation is supercritical. A time-consuming but straightforward calculation now
yields that the quantity b defined in Theorem 1.5.1 becomes

2
2h 1 3h

1
b= +
2 u2 3 u3

2
2(1 P )3
  
2 1
= +1 (P ) + (1 )(4 3P + 1)  (2.7.13)
2
1P P 2 (1 + P )(1 2P ) 6

at bifurcation. Here we may observe that W () = { } attains its minimum when = (9/4)P 3/4
and that W ((9/4)P 3/4) > 0 whenever 0 < P < 1/2 . Hence b > 0 .

Regarding the nondegeneracy condition a defined in Theorem 1.5.1, it may be expressed as

h 2 h 2h
 
h
a= 1 = 0 (u, v) = (0, 0)

F u2 u uF

Now, since the bifurcation is transformed to the origin it follows that h/u = 1 and h/F = 0 .

Therefore the condition a = 0 simplifies to

2h
a=2 = 0 2 = 0
uF F

since in general h/u = . From the Jacobian:

1 
= w w 2 + 4P w
2
where  
 1 1
w =f x +f = [(F + F P ) 1] + 1
1+P

it follows that
 
dw 1 dw dw
2 = 2w + 4P
F dF 2 w 2 + 4P w dF dF
 
dw 1
= 1 (w + 2P )
dF 2
w + 4P w

144
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

At bifurcation, w = (1 P )1 which inserted into the expression above gives

1(1/)
(1 P )2

2
2 = 2 +1  (2.7.14)
F 1P 1 2P

and clearly, (2.7.14) is nonzero whenever 0 < P < 1/2 . Consequently, the flip bifurcation is
supercritical, which means that when the fixed point fails to be stable, a stable two-periodic orbit is
established.

We close this section by showing the dynamics beyond the flip bifurcation threshold for the Ricker
map

(x0 , x1 ) (F ex (x0 + x1 ), P x0 )  (2.7.15)

which is a special case of map (2.7.5) (the case 0 ). Assuming F (1 + P ) > 1 the nontrivial
fixed point of (2.7.15) is
 
1 P
(x0 , x1 ) = ln(F (1 + P )) , ln(F (1 + P ))

1+P 1+P

145 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Figure 14: The bifurcation diagram of map (2.7.14) in the case P = 0.2. For small F values we see the stable xed point of (2.7.14) which
undergoes a supercritical flip bifurcation when F = 10.152. Through further increase of F stable orbits of period 2k are created until an
accumulation value Fa for the flip bifurcations is reached. Beyond Fa the dynamics is chaotic.

and whenever 0 < P < 1/2 we have according to the preceding example that the fixed point undergoes

a supercritical flip bifurcation at the threshold F = (1/(1 + P )) exp(2/(1 P )) .

Now, consider the value P = 0.2 . Under this choice the fixed point is stable in the F interval
0.834 < F < 10.152 and in Figure 14 we have plotted the bifurcation diagram of (2.7.15) in the range
5 < F < 80 . We clearly identify the supercritical flip at the threshold F = 10.152 and beyond that
stable periodic orbits of period 2k are established through further increase of F so what we recognize
is essentially the same kind of dynamical behaviour as we found when we considered one-dimensional
maps. Beyond the point of accumulation for the flip bifurcation sequence the dynamics becomes chaotic
as displayed in Figure 15. Note that the chaotic attractor consists of 4 disjoint subsets (branches) that
are visited once every fourth iteration so a certain kind of four periodicity is preserved in the chaotic
regime. In case of higher F values the branches merge together.

146
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Figure 15: The chaotic attractor consisting of four separate branches just beyond the point of accumulation
for the ip bifurcations in the case P = 0.2, F = 58.5. The dynamics goes in the direction A B C D

2.8 Beyond the Hopf bifurcation, possible routes to chaos

As we proved in section 2.5, the outcome of a supercritical Hopf bifurcation is that when the fixed
point of a discrete map fails to be stable, an attracting invariant curve which surrounds the fixed point
is created. Our goal in this section is to describe the dynamics on such an invariant curve. We will also
discuss possible routes to chaos and as it will become clear, the dynamics may be much richer than in
the one-dimensional cases discussed in Part I.

In general terms, the dynamics on an invariant curve (circle) created by a Hopf bifurcation may be
analysed by use of equation (2.5.14b). Indeed, if we substitute the fixed point r of (2.5.14a) into (2.5.14b)
we arrive at

bd
+c = + ()  (2.8.1)
a

where c = arg . Also recall that when we derived (2.5.14a,b) we first transformed the bifurcation to
the origin. If the Hopf bifurcation occurs at a threshold 0 = 0 , () = c + (bd/a)(0 ) .

Now, the essential feature is that successive iterations of (2.8.1) simply move or rotate points from one
location to another on the invariant curve. Hence, the original map f : R2 R2 may be regarded as

being topological equivalent to a circle map g : S  S  once the invariant curve is established.

Moreover, considering g , one may define its rotation number as the average amount that points are
rotated by an iteration of the map. Therefore, we may (to leading order, recall that (2.8.1) is a truncated
map) regard (2.8.1) as a circle map with rotation number () .

147
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Remark 2.8.1. A more precise definition of the rotation number may be achieved along the
following line: Given a circle map g : S S we first lift g to the real line R by use of
: R S , (x) = cos(2x) + i sin(2x) and then define the lift F as F : R R ,
F = g . Next, let 0 (F ) = limn F n (x)/x and finally define the rotation number
of g, (g) , as the unique number in [0, 1 such that 0 (F ) (g) is an integer. In Devaneys
book there is an excellent introduction to circle maps.

Join the best at Top masters programmes

3
 3rd place Financial Times worldwide ranking: MSc
the Maastricht University International Business
1st place: MSc International Business
School of Business and 1st place: MSc Financial Economics
2nd place: MSc Management of Learning

Economics! 2nd place: MSc Economics

2nd place: MSc Econometrics and Operations Research
2nd place: MSc Global Supply Chain Management and
Change
Sources: Keuzegids Master ranking 2013; Elsevier Beste Studies ranking 2012;
Financial Times Global Masters in Management ranking 2012

Maastricht
University is
the best specialist
university in the
Visit us and find out why we are the best! Netherlands
(Elsevier)
Masters Open Day: 22 February 2014

www.mastersopenday.nl

148 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Returning to map (2.8.1) the rotation number may be irrational or rational. In the former case this
means that as the number of iterations of the map tends to infinity, the invariant curve will be filled with
points. Whenever irrational, an orbit of a point is often referred to as a quasistationary orbit. If

= 1/n , rational, the dynamic outcome is an n -period orbit. It is of great importance to realize that
whenever the rotation number is rational for a given parameter value = r , it follows from the implicit
function theorem that there exists an open interval about r where the periodicity is maintained. This
phenomenon is known as frequency locking of periodic orbits. Consequently, periodic dynamics will
occur in parameter regions, not at isolated parameter values only. As we shall see, such regions (or
intervals) may in fact be large. So in order to summarize: beyond the Hopf bifurcation (and outside the
strongly resonant cases where is third or fourth root of unity) there are quasistationary orbits restricted
to an invariant curve and there may also be orbits of finite period established through frequency locking
as the value of the parameter in the model is increased.

Our next goal is by way of examples to study in more detail the interplay between these cases as well as
studying possible routes to chaos. We start by scrutinizing a population model first presented in Wikan
and Mjlhus (1995).

Example 2.8.1. First, consider the two-age class population model

(x0 , x1 ) (F x1 , P ex x0 )  (2.8.2)

which is a semelparous species model where the fecundity F is constant while the survival
probability p(x) = P exp(x) is density dependent. is a positive number (scaling constant)

and we assume that P F > 1 .

It is easy to verify that (2.8.2) possesses the following properties: The fixed point may be expressed as
 
F 1
(x0 , x1 ) = x , x  (2.8.3)
1+F 1+F

where x = x0 + x1 = 1 ln(P F ) . Moreover, the eigenvalue equation may be cast in the form

ln(P F ) F ln(P F )
2 + + 1 =0  (2.8.4)
1+F 1+F

and from the Jury criteria one obtains that the fixed point is stable in case of PF small but undergoes
a Hopf bifurcation at the threshold

149
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

1 2(1+F )/F
e) P = Pc = e  (2.8.5)
F

Figure 16: The dynamics of map (2.8.2) (a quasistationary orbit), just beyond the Hopfbifurcation threshold.

Note that drops out of (2.8.4), (2.8.5) which simply means that stability properties are
independent of . At bifurcation threshold (2.8.5) the solution of the eigenvalue equation becomes

1 1
= 1 i  (2.8.6)
F F2

A final observation is that by rewriting (2.8.2) on standard form (as in Example 2.5.1) and then
apply Theorem 2.5.2, it is possible to prove that the bifurcation is supercritical.

Now, let us scrutinize a numerical example somewhat closer. Assume P = 0.6 . Then from (2.8.5)
the F value at bifurcation threshold is numerically found to be F = Fc = 14.1805 . We want to

investigate the dynamics when F > Fc . In Figure 16 we show the dynamics just beyond the
instability threshold in the case (, P, F ) = (0.02, 0.6, 15) . From an initial state (x00 , x10 ) 500

iterations have been computed and the last 20 together with the (unstable!) fixed point are plotted.
The invariant curve is indicated by the dashed line so clearly the original map (2.8.2) does nothing
but rotate points around that curve, i.e. (2.8.2) acts as a circle map.

150
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Figure 17: A 4-periodi orbit generated by map (2.8.2).

Moreover, Figure 16 demonstrates a clear tendency towards 4-periodic dynamics. This is as

expected due to the location of the eigenvalues. Indeed, when Fc = 14.1805 it follows from
(2.7.6) that the eigenvalues are located very close to the imaginary axis

0.0750 0.9975 i ( 1,2 = 0.0750 0.9975 i ), and since the rotation number (up to leading order! ) has the
(F ) = c form (F ) =c
+ (bd/a)(F c+F )(bd/a)(Fc F ) where c = arg it follows that must be close to 1/4 in
case ofFF close to Fc

Fa 25.07

151 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

If we increase F beyond 15 we observe (due to frequency locking! ) that an exact 4-periodic orbit
is established. This is shown in Figure 17 in the case (, P, F ) = (0.02, 0.6, 20) and further, it
is possible to verify numerically that the exact 4-periodicity is maintained as long as F does not
exceed the value 21.190.

At F = 21.190 the fourth iterate of (2.8.2) undergoes a flip bifurcation, thus an 8-periodic orbit

is established, and through further enlargement of F we find that new flip bifurcations take place
at the parameter values 24.232 and 24.883 which again result in orbits of period 16 and 32
respectively. Hence we oberve nothing but the flip bifurcation sequence which we discussed in
Part I. The point of accumulation for the flip bifurcation is found to be Fa 25.07 and in case
of F > Fa the dynamics becomes chaotic.

These findings are shown in Figures 18, 19 and 20. In Figures 18 and 19 periodic orbits of period 8
and 32 are displayed. In Figure 20 we show the chaotic attractor. Note that the attractor is divided
in 4 disjoint subsets and that each of the subsets are visited once every fourth iteration so there
is a kind of 4-periodicity preserved, even in the chaotic regime. 

Figure 18: An 8-periodic orbit generated by map (2.8.2).

152
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Figure 19: A 32-periodic orbit generated by map (2.8.2).

Figure 20: Map (2.8.2) in the chaotic regime.

Example 2.8.2. The next example (which rests upon the findings in Wikan (1997)) is basically
the same as the previous one but the dimension of the map has been extended by 1 and we consider
a general survival probability p(x) , 0 < p(x) 1 , p (x) 0 , instead of p(x) = P exp(x).

Hence we consider the problem

(x1 , x2 , x3 ) (F3 x3 , p(x)x1 , p(x)x2 )  (2.8.7)

Skipping computational details (which are much more cumbersome here than in our previous
example) we find that the nontrivial fixed point is

x x x
 
(x1 , x2 , x3 ) = , p(x ) , p2 (x )  (2.8.8)
K K K

153
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

3 1/(n1)
where K = i=1 pi1 (x ) and x = p1 (F3 ) . ( p1 denotes the inverse of p .)

Moreover, by first computing the Jacobian and then use the Jury criteria, it is possible to show
that (2.8.8) is stable as long as

x 1 + p(x ) 2p2 (x )
p (x ) < p(x )  (2.8.9)
K (1 + p(x ))(1 p2 (x ))

(2.8.8) becomes unstable when F3 is increased to a level FH1 where (2.8.9) becomes an equality.
At that level a (supercritical) Hopf bifurcation occurs and the complex modulus 1 eigenvalues
may be expressed as


p2 (x ) p4 (x )
1,2 = 1 i  (2.8.10)
1 + p(x ) (1 + p(x ))2

Need help with your

dissertation?
Get in-depth feedback & advice from experts in your
topic area. Find out what you can do to improve
the quality of your dissertation!

Get Help Now

Go to www.helpmyassignment.co.uk for more info

154 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Now, for comparison reasons, assume that p(x) = P exp(x) just as in Example 2.8.1. Then it
easily follows that F3 is a large number at bifurcation threshold FH1 and further that p(x )  1.

Consequently, 1,2 are located very close to the imaginary axis, in fact even closer than the
eigenvalues from Example 2.8.1. When we increase F3 beyond FH1 we observe the following
dynamics: In case of F3 FH1 small we find an almost 4-periodic orbit restricted on an invariant
curve and through further enlargement of F3 we once again find (through frequency locking)
that an exact 4-periodic orbit is the outcome. Thus the dynamics is qualitatively similar to what
we found in Example 2.8.1. However, if we continue to increase F3 we do not experience the flip
bifurcation sequence. Instead we find that the fourth iterate of map (2.8.7) undergoes a (supercritical)
Hopf bifurcation at a threshold F3 = FH2 . Therefore, beyond that threshold, and in case of
F3 FH2 small, the dynamics is restricted on 4 disjoint invariant attracting curves which are
visited once every fourth iteration. This is displayed in Figure 21. At an even higher value, F3 = Fs
, map (2.8.7) undergoes a subcritical bifurcation which implies that whenever F3 > Fs there is
no attractor at all so in this part of parameter space we simply find that points (x1 , x2 , x3 ) are
randomly distributed in state space. 

So far we have demonstrated that although the dynamics is a quasistationary orbit just beyond the
original Hopf bifurcation threshold, the dynamical outcome may be a periodic orbit as we penetrate
deeper into the unstable parameter region. Such a phenomenon may happen when | arg | is close to
/4 at bifurcation threshold (4-periodicity). Another possibility (among others! ) is that | arg | is
close to 2/3 (3-periodicity).

Figure 21: Map (2.8.7) after the secondary Hopf bifurcation.

155
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Note, however, that if arg is close to a critical value, say /2 , at bifurcation it does not necessarily
imply that a periodic orbit is created when we continue to increase the bifurcation parameter. In fact,
when the parameter is enlarged the location of the eigenvalues may move away from the imaginary axis,
hence the periodicity will be less pronounced as the bifurcation parameter growths. In our next example
there is no periodicity at all.

Example 2.8.3. Consider the two-dimensional population map

(x1 , x2 ) (F ex2 x1 + F ex2 x2 , P x1 )  (2.8.11)

Hence, only the second age class x2 contributes to density effects. As before, F > 0 ,
0 < P 1 and F (1 + P ) > 1 .

We urge the reader to verify that the fixed point (x1 , x2 ) may be written as
 
1
(x1 , x2 ) = x , ln[(1 + P )F ]  (2.8.12)
P 2

Figure 22: Dynamics generates by map (2.8.11). Parameter values: (a) (P, F) = (0.6, 2.5); (b) (P, F) = (0.6, 5.0).

and further that a (supercritical) Hopf bifurcation occurs at the threshold

1
F = FH = e(1+2P )/(1+P )  (2.8.13)
1+P

156
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

and finally that the solution of the eigenvalue equation at threshold (2.7.13) becomes

1   
= 1 4(1 + P ) 1 i 
2
(2.8.14)
2(1 + P )

Now, assume that P is not close to zero. Then, the location of clearly suggests that frequency
locking into an orbit of finite period will not take place. In Figure 22a we show the invariant curve
just beyond the bifurcation threshold (P, F ) = (0.6, 2.5) and on that curve we find no tendency

towards periodic dynamics.

As we continue to increase F (P fixed) the radius of the invariant curve becomes larger. Eventually,
the invariant curve becomes kinked and signals that the attractor is not topological equivalent to a
circle anymore and finally the curve breaks up and a chaotic attractor is born. This is exemplified
in Figure 22b. 

In our final example (cf. Wikan and Mjlhus (1996) or Wikan (2012b)) all bifurcations that we have
previously discussed are present.

Brain power By 2020, wind could provide one-tenth of our planets

electricity needs. Already today, SKFs innovative know-
how is crucial to running a large proportion of the
worlds wind turbines.
Up to 25 % of the generating costs relate to mainte-
nance. These can be reduced dramatically thanks to our
systems for on-line condition monitoring and automatic
lubrication. We help make it more economical to create
cleaner, cheaper energy out of thin air.
By sharing our experience, expertise, and creativity,
industries can boost performance beyond expectations.
Therefore we need the best employees who can
meet this challenge!

The Power of Knowledge Engineering

Plug into The Power of Knowledge Engineering.

Visit us at www.skf.com/knowledge

157 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Example 2.8.4. Referring to section 2.4, Examples 2.4.1 and 2.4.3 we showed that the fixed point
(x0 , x1 ) of map (2.4.2), i.e.

(x , x )
0 1 (F0 ex x0 + F1 ex x1 , P0 x0 )

is stable in case of small equilibrium populations x = x0 + x1 but eventually will undergo a

supercritical Hopf bifurcation at the threshold

1
F = FH = e(1+2P0 )/P0
1 + P0

provided 1/2 < P0 < 1 and equal fecundities F0 = F1 = F . In Figure 23 we have generated

the bifurcation of the map in the case P0 = 0.9 , = 0.01 . The bifurcation parameter F is along

the horizontal axis, the total population x along the vertical. Omitting computational details
(which may be obtained in Wikan and Mjlhus (1996)) we shall now use Figure 23 in order to
reveal the dynamics of (2.4.2).

In case of 5.263 < F < 10.036 there is one attractor, namely the stable fixed point (x0 , x1 ) .
(The lower limit 5.263 is a result of the requirement F (1 + P ) > 1 .) At the threshold Fs = 10.036

a 3-cyclic attractor with large amplitude is created. Thus beyond Fs there exists a parameter (F )

interval where there are two coexisting attractors and the ultimate fate of an orbit depends on the
initial condition. It is a well known fact that multiple attractors indeed may occur in nonlinear
systems. What happens in our case is that the third iterate of the original map (2.4.2) undergoes
a saddle-node bifurcation at Fs .

Figure 23: The bifurcation diagram generated by map (2.4.2).

158
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

This may be verified numerically by computing the Jacobian of the third iterate and show that the
dominant eigenvalue of the Jacobian equals unity. Moreover (referring to section 1.5, see also
Exercise 1.4.2 in section 1.4), a 3-cycle consisting of unstable points is also created through the
saddle node at threshold Fs . This repelling 3-cycle is of course invisible to the computer.

In the interval 10.036 < F < 11.81 the large amplitude 3-cycle and the fixed point are coexisting
attractors. At FH = 11.81 the fixed point undergoes a supercritical Hopf bifurcation (for a proof,
cf. Wikan and Mjlhus (1996)), thus in case of F > FH , F FH small, there is coexistence

between the 3-cylic attractor and a quasistationary orbit restricted to an invariant curve. The
coexistence takes place in the interval 11.81 < F < 12.20 . In somewhat more detail we also

find that since arg (where is the eigenvalue of the Jacobian of (2.4.2)) is close to 2/3 at
FH there is a clear tendency towards 3-periodic dynamics on the invariant curve but there is no
frequency locking into an exact 3-periodic orbit.

At FK = 12.20 the invariant curve disappears. Consequently, in case of F > FK , there is again
only one attractor, namely the attracting 3-cycle. The reason that the invariant curve disappears
at threshold FK is that it is hit by the three branches of the repelling 3-cycle. This phenomenon
is somewhat akin to what is called a crisis in the chaos literature.

As we continue to increase F successive flip bifurcations occur, creating orbits of period 3 2k ,

k = 1, 2, ... , in much of the same way as we have seen in earlier examples. Eventually an
accumulation value Fa for the flip bifurcations is reached, and beyond that value the dynamics
becomes chaotic. At first the chaotic attractor consists of three separate branches which are visited
once every third iteration. When F is even more increased the branches merge together. 

Through our previous examples, which all share the common feature that the original (first) bifurcation
is a Hopf bifurcation, we have experienced that the nonstationary dynamics beyond the instability
threshold may indeed be different from map to map. In the following exercises even more possible
dynamical outcomes are demonstrated.

Exercise 2.8.1. Consider the map (cf. Wikan (1998))

(x0 , x1 ) (F1 x1 , P0 (1 x)1/ x0 )

where > 0 , 0 .

159
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

a) Compute the nontrivial fixed point (x0 , x1 ) .

b) Assume that > c = (F1 /2(1 + F1 ) and show that the fixed point undergoes a Hopf
bifurcation at the threshold

 1/
1 2(1 + F1 )
P0 = 1+
F1 F1

c) Assume that > c but c small. Investigate numerically the dynamical outcomes when
P0 is fixed and F1 is increased beyond the bifurcation threshold.

d) (difficult!) Show that the Hopf bifurcation is supercritical. 

Exercise 2.8.2. Consider the semelparous population model

0 0 F2 ex

x0 x0
x1 = P0 0 0 x1
x2 t+1 0 P1 0 x2 t

160 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

a) Show that the fixed point is

 
1 P0 P0 P1
(x0 , x1 , x2 ) = x , x , x
1 + P 0 + P0 P 1 1 + P 0 + P0 P 1 1 + P 0 + P0 P 1

where x = ln(P0 P1 F2 ) .

b) Compute the Jacobian and show that the eigenvalue equation may be cast in the form

3 + 2 + P0 + P0 P1 1 = 0

where = x /(1 + P0 + P0 P1 ) .

c) Use the Jury criteria (2.1.16) and show that the fixed point is stable whenever

4 < < 2

where
1 + P0 2P0 P1 2
4 = and 2 =
P0 P1 (1 P0 P1 ) 1 P 0 + P0 P 1

d) Use the result in c) and show that the fixed point is stable provided

1 1 + P0
< P0 < 1 P1 >
2 3P0

e) The results from c) and d) are special in the sense that they imply that the fixed point is unstable
in case of x (or F2 ) small, becomes stable for larger values of x (or F2 ) and then becomes
unstable again through further enlargement of x (or F2 ). Note that 4 and 2 are Hopf and
flip bifurcation thresholds respectively. Investigate (numerically the dynamics in case of < 4
(i.e. x small) and < 2 (i.e. x large). (Hint: cf. Exercise 2.4.3.) Other properties of this
model as well as properties of more general semelparous population models may be obtained
in Mjlhus et al. (2005). 

Exercise 2.8.3 (Coexistence of age classes). Consider the two age class map (Wikan 2012a)

(x1 , x2 ) (F ex x2 , P ex x1 )

cf. (2.3.1) where x = x1 + x2 , 0 < P 1 , F > 0 and , > 0 .

161
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

a) Show that the nontrivial fixed point of the map is

 
1 aP

(x1 , x2 ) = x , x
1 + aP 1 + aP

where a = R0(/(+)) , x = ( + )1 ln R0 and R0 = P F > 1 .

b) Use the Jury criteria and show that if > then there exists a parameter region where
(x1 , x2 ) is stable and, moreover, that when R0 increases there will occur a Hopf bifurcation
at the threshold
 
2( + )(1 + aP )
R0 = exp
+ aP

c) Investigate numerically the behaviour of the map beyond instability threshold. (Hint: the cases
small, large should be treated separately.)

d) The parameters and may be interpreted as strength of density dependence. Show that
if the strength of density dependence in the fecundity is equal or larger than the strength
of density dependence in the survival then (x1 , x2 ) will always be unstable.

e) What kind of dynamic outcome do you find in the case < ? 

Exercise 2.8.4 (Permanence in stage-structured models). In Example 2.5.1 we analysed a stage-

structured cod model. A slightly more general form of such a model is

(i) xt+1 = Ax xt

where x = (x1 , x2 )T and

 
(1 1 )S(x) f (x)

Ax =
p(x) (1 2 )

Here, x1,t and x2,t are the immature and mature part of the population respectively and just as
in the age-structured case f (x) is the fecundity. p(x) is the fraction of the immature population
that survives to become mature, and 1 and 2 are (natural) death rates. Finally, it is also assumed
that the remaining part of the immature population (1 1 )x1 is reduced by a nonlinear factor
s(x) .

162
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Further, let s(x) = S s(x) , f (x) = F f(x) , p(x) = P p(x) where 0 S 1 , 0 < P 1 ,
F > 0 , 0 1 , 2 < 1 , 0 < s(x), p(x), f(x) 1 , s(0) = p(0) = f(0) = 1 . A final but
important restriction in such models is (1 1 )S + P 1 . Otherwise, the fraction of juveniles

that survives to become adults plus the fraction that survives but remain juveniles may be larger
than 1 even in case of zero fecundity which of course is unacceptable from a biological point of
view.

Definition. Let xt = x1,t + x2,t be the total population at time t . Model (i) is said to be permanent

if there exists > 0 and D > 0 such that

< lim inf xt lim sup xt < D
t t

Thus, if a population model is permanent, the total population density neither explodes nor goes
to zero (see Kon et al. (2004)). Define the net reproductive number R0 as

PF
R0 =
2 [1 (1 1 )S]

163 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Our goal is to prove the following theorem:

Theorem: Suppose that model (i) is continuous and that one of p(x)x1 or f(x)x2 is bounded
from above. Further assume that the matrix A0 is irreducible and R2+ \ {0} forward invariant
(i.e. that Ax x R2+ \ {0} for all x R2+ \ {0} ). Then model (i) is permanent provided

R0 > 1 . 

a) Clearly, (x1 , x2 ) = (0, 0) is a fixed point of (i). Use the Jury criteria and show that (0, 0) is
unstable provided R0 > 1 .

b) Explain why A0 is irreducible and R2+ \ {0} forward invariant.

It remains to prove that the population density does not explode, i.e. that (i) is a dissipative model.
From Kon et al. (2004), see also Cushing (1998), we apply the following definition of dissipativeness:

Definition: Model (i) is said to be dissipative if there exists a compact set X R2+ such that for

all xt R2+ there exists a tM = tM (x0 ) satisfying xt X for all t tM . 

c) Assume p(x)x1 K0 where K0 is a constant. Use (i) and induction to establish the relations

x2,t+1 P K0 + (1 2 )x2,t

and
P K0
x2,t (1 2 )t x2,0 +
2

d) Use c) to conclude that there exists tA = tA (x2,0 ) such that for t > tA
2P K0
x2,t = K1
2

e) Use the previous result together with (i) and induction to show that

x1,t+1 (1 1 )Sx1,t + F K1

F K1
x1,t (1 1 )t S t x1,0 +
1 (1 1 )S


164
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

f) Show that there exists tB = tB (x1,0 ) such that for t > tB (x1,0 )

2F K1
x1,t = K2
1 (1 1 )S

g) Take tM = max{tA , tB } and K = max{K1 , K2 } and conclude that x1,t K and

x2,t K , hence (i) is dissipative if p(x)x1 is bounded from above.

h) Assume f(x)x2 K0 and show in a similar manner that (i) is dissipative in this case too.

Remark 2.8.2. In Leslie matrix models nonoverlapping age classes are assumed. This is not the
case in the stage-structured model from the previous exercise (or the model presented in Example
2.5.1). Moreover, while Leslie matrix models are maps from Rn Rn (or Rn+1 Rn+1 )
where n may be a large integer, stage-structured models are mainly maps from R2 R2 where
we do not have the possibility to study the dynamic behaviour of age classes in detail. Some stage-
structured models are maps from R3 R3 . Typically, they are insect models where the population
is divided into three stages: larvae (L), pupae (P), and adult insects (A). In fact, such models are
fully capable of describing and even predicting nonstationary and chaotic behaviour in laboratory
insect poulations, see Cushing et al. (1996), Costantino et al. (1997), Dennis et al. (1997), and
Cushing et al. (1998). 

Exercise 2.8.5 (Prey-Predator systems). In 1920 Lotka introduced a system of differential

equations which described the interaction between a prey species x and a predator species y .
These equations were rediscovered by Volterra in 1926 and today they are often referred to as the
Lotka-Volterra equations. A discrete version of the equations (written as a map) is

(i) (x, y) [((1 + r) ay)x, (c + bx)y]

The first component of the map expresses that the growth rate of the prey is a constant (1 + r)
due to the species itself minus a term proportional to the number of predators. In the same way,
the growth rate of the predator is proportional to the number of prey minus a term c which is
due to the predator species itself. All constants are assumed to be positive.

a) Find the nontrivial fixed point of the map and show that it is always unstable.

b) Consider the prey-predator map

(ii) (x, y) (f (y)x, g(x)y)

165
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

where f /y < 0 and g/x > 0 . Show that || > 1 where is the solution of the

eigenvalue equation. (See Maynard Smith (1979) for computational details.) What is the
qualitative dynamic behaviour of maps like (i) and (ii)?

c) Next, consider the two parameter family prey-predator maps

(iii) (x, y) [((1 + r) rx ay)x, axy]

where r > 0 , a > 0 (Maynard Smith, 1968). Show that (iii) has three fixed points,
(x, y) = (0, 0) , (x, y) = (1, 0) and (x , y ) = (1/a, r(a 1)/a2 ) .

d) Following Neubert and Kot (1992) who perform a detailed analysis of (iii) show that 1) (x, y)
is always unstable, 2) (x, y) is stable whenever 0 < r < 1 and 0 < a < 1 , and 3) (x , y )
is stable provided 1 < a < 2 and 0 < r < 4a/(3 a) .

e) Still referring to Neubert and Kot (1992), show that (iii) undergoes a transcritical bifurcation
when a = 1 and draw a bifurcation diagram similar to Figure 4b in Section 1.5.

Challenge the way we run

EXPERIENCE THE POWER OF

FULL ENGAGEMENT

RUN FASTER.
RUN LONGER.. READ MORE & PRE-ORDER TODAY
RUN EASIER WWW.GAITEYE.COM

1349906_A6_4+0.indd 1 22-08-2014 12:56:57

166 Click on the ad to read more
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

((iii) has several other interesting properties. It should be easy for the reader to verify that in
case of 1 < a < 2 , r = 4a/(3 a) gives birth to a flip bifurcation, but unlike most of the
cases treated so far (however, see Exercise 1.5.2), this bifurcation is of the subcritical type and
the predator goes extinct at instability threshold. (Formally, this may be proved by using the
same procedure as in Example 2.7.1.) Moreover, when a = 2 and r =
 4, r = 6 a Hopf
bifurcation occurs and whenever a > 2 , |a 2| small the dynamics is restricted on an
invariant curve. In the strong resonant cases r = 4, r = 6 we find the same qualitative picture

as we did in Exercises 2.5.1 and 2.5.2. For further reading of this fascinating map we refer to
the original paper by Neubert and Kot (1992).)

f) Finally, consider the age-structured prey-predator map

 
(x+1 y) Q 2 x
(x1 , x2 , y1 , y2 ) F2 x2 , P e x1 , G2 x2 , y1
1 + y 1 + 2 x

where F2 and G2 are the fecundities of the second age classes of the prey and predator

respectively. P and Q are survival probabilities from the first to the second age classes, 1

and 2 are positive interaction parameters and x = x1 + x2 , y = y1 + y2 . Find the nontrivial

fixed point (x1 , x2 , y1, y2) and show that it may not undergo a saddle node or a flip bifurcation

at instability threshold. Thus stability or dynamics governed by Hopf bifurcations are the only
possible dynamic outcomes.

g) If P = 0.6 and F2 = 25 then the prey in absence of the predator exhibits chaotic oscillations.

Now, suppose Q = 0.6 , G2 = 12 and assume = 1 = 2 . Investigate numerically how

the prey-predator system behaves in the following cases: [0.1, 0.22] (weak interaction),
[0.4, 0.6] (normal interaction), [0.85, 1.00] (strong interaction) (see Wikan
(2001)). 

Exercise 2.8.6 (Host-Parasitoid models). Following Kot (2001), see also the original work by
Nicholson (1933), Nicholson and Bailey (1935), the books of Hassel (1978), and Edelstein-Keshet
(1988), most host-parasitoid models are on the form

xt+1 = af (xt , yt)xt

yt+1 = c[1 f (xt , yt )]xt

167
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Here xt and yt are the number of hosts and parasitoids at time t respectively. f (x, y) is the

fraction of hosts that avoids parasitoids at time t and a is the net reproductive rate of hosts. c

may be interpreted as the product of the number of eggs laid per female which survive to pupate
times the probability that a pupae will survive the winter and give rise to an adult next year
(Maynard Smith, 1979). Kot (2001) simply refers to c as the clutch size of parasitoids.

a) Assume that f (x, y) = f (y) = ey where > 0 and find the nontrivial fixed point of the
map. Use the Jury criteria and discuss its stability properties. What are the possible dynamic
outcomes of this model?

b) A slightly modified version of the Nicholson and Bailey model in a) which also contains a
self-regulatory prey term was proposed by Beddington et al. (1975)

x1,t+1 = er(1xt )yt xt

yt+1 = c[1 eyt ]xt

Denoting the nontrivial fixed point for (x , y ) , show that

r
x =1 y 0 < y <
r

and that y is the unique solution of

ry

c[1 ey ] = 0
r y

Moreover, show (numerically) that there exists a parameter region where (x , y ) is stable.

Remark 2.8.3. As is clear from Exercises 2.8.5a,b and 2.8.6a, if prey-predator models or host-
parasitoid models shall possess a stable nontrivial equilibrium where both species exist we may
not assume that one of the species is a function of only the other species. Thus, the function f

in the exercises above should be on the form f = f (x, y) with properties f /x < 0 ,
f /y < 0 . In prey-predator systems self-limitational effects are often assumed to be crowdening
or cannibalistic effects (the latter is typically the case in fish populations). However, what the
self-regulatory effects in parasitoid species are, is far from obvious, cf. the discussion in
Beddington et al. (1975), Hassel (1978), Edelstein-Keshet (1988), Murdoch (1994), and Mills
and Getz (1996). 

168
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Exercise 2.8.7 (Competition models). Suppose that two species x and y compete on the same
resource. From a biological point of view the competitive interaction between the two species
would be that an increase of one of the species should reduce the growth of the other and vice
versa. Hence, in a model of the form

xt+1 = (xt , yt )xt

(i)
yt+1 = (xt , yt)yt
where also self-regulatory effects are included, we should regard all partial derivatives of the
functions and as negative. (Note that these sign restrictions differ from the prey-predator

models we studied in Exercise 2.8.5.)

a) Consider the competition model

xt+1 = (a bxt c1 yt )xt

(ii)
yt+1 = (d eyt c2 xt )yt
where all constants are positive and a > 1 , d > 1 . Find all the fixed points of (ii). (There are
four of them.)

This e-book
is made with SETASIGN
SetaPDF

PDF components for PHP developers

www.setasign.com

169 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

b) (x, y) = ((a 1)/b, 0) is one of the fixed points. Use the Jury criteria and find conditions
for (x, y) to be stable.

c) (i) has a nontrivial fixed point (x , y ) , (x > 0, y > 0) which is a solution of the equations

(x , y ) = 1 and (x , y ) = 1 . Show that the solutions 1,2 of the linearization of (i)

may be expressed as (Maynard Smith, 1979)

1  
1,2 = 2 (a + d) (a + d)2 4(ad bc)
2
where

a = x d = y
x y

and
 

ad bc = x y
x y y x

Note that since all partial derivatives are supposed to be negative, a , b , c and d are positive.

d) Explain that (/x)(/y) > (/y)(/x) (i.e. that the product of changes in
and due to self-regulatory effects are larger than the product of changes in and due
to the competitive species) is necessary in order for (x , y ) to be stable.

e) Discuss the possibility of having oscillatory behaviour in model (i).

For further reading of discrete competition models we refer to Adler (1990). 

Exercise 2.8.8 (The Hnon map). Consider the two parameter family of maps (the Hnon map)

Ha,b : R 2 R2 (x, y) (y, 1 + bx ay 2 )

where 0 < b < 1 .

Ha,b (in a slightly different version), was constructed and analysed by Hnon (1976), and is one
of the first two-dimensional maps where there was found numerical evidence of a chaotic attractor.
(Hnons paper may also be obtained in Cvitanovi (1996) where several classical papers on
dynamical systems are collected.)

170
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

a) Let V be a region elongated along the y -axis in the R2 plane and consider the following

maps:

h1 : (x, y) (bx, y) V y

h2 : (x, y) (1 + x ay 2 , y) x

h3 : (x, y) (y, x)

Show that Ha,b = h3 h2 h1 .

b) Let a0 = ((1 b)/2)2 and show that Ha,b has two fixed points if a > a0 , one fixed

point if a = a0 and no fixed points if a < a0 .

c) Show that Ha,b undergoes a saddle node bifurcation at the threshold a = a0 .

d) Let a1 = 3a0 and show that in the interval a0 < a < a1 there is one stable fixed point
(x+ , y+

) and one unstable fixed point (x , y

).

e) Show by use of the Jury criteria that (x+ , y+

) undergoes a flip bifurcation at a = a1 .

f) Show that the second iterate of Ha,b may be written as

xt+2 = 1 + bxt ayt2

yt+2 = 1 + byt a(1 + bxt ayt2 )2

and verify that whenever a > a1 there is a two-period orbit where the points are
  
1 ay12 1 b + 4a 3(1 b)2
(x1 , y1 ) = ,
1b 2a

  
(x , y )
1 ay22 1 b 4a 3(1 b)2
2 2 = ,
1b 2a

g) Show that
 
2 2
lim (x1 , y1 ) = lim (x2 , y2 ) = lim (x+ , y+

) = ,
aa1 aa1 aa1 3(1 b) 3(1 b)

171
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

h) Assume b = 1/2 and let a > a1 . Investigate numerically if Ha,1/2 possesses a chaotic

attractor.

i) Still assuming b = 1/2 , generate a bifurcation diagram in case of a > a0 .

1
j) Show that Ha,b has an inverse and compute Ha,b .

Next, let b = 0 . Then Ha,0 contracts the entire R2 plane onto the curve fa (y) = 1 ay 2 and

since the value of Ha,0 is independent of the x coordinate we may study the dynamics through

the one-dimensional map

y fa (y) = 1 ay 2

k) Show that the map undergoes a saddle node bifurcation when a = 1/4 and find a parameter

interval where the map possesses a unique nontrivial fixed point.

www.sylvania.com

We do not reinvent
the wheel we reinvent
light.
Fascinating lighting offers an infinite spectrum of
possibilities: Innovative technologies and new
markets provide both opportunities and challenges.
An environment in which your expertise is in high
demand. Enjoy the supportive working atmosphere
within our global group and benefit from international
career paths. Implement sustainable ideas in close
cooperation with other specialists and contribute to
influencing our future. Come and join us in reinventing
light every day.

Light is OSRAM

172 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

l) Show that fa (y) is topologically equivalent to the quadratic map g (y) = y(1 y) .)

(Hint: Use Definition 1.2.2 and assume that h is a linear function of y . Moreover, show that

the relation between a and is given through 2 2 = 4a and > 1 .)

(The case b = 1 will be considered in the next exercise.) 

Exercise 2.8.9 (Area preserving maps). Consider the map (x, y) f (x, y) . If the area of a region

in R2 is preserved under f we say that f is an area preserving map. In order to decide whether
a map is area preserving or not we may apply the following theorem:

Theorem. Let f : R2 R2 be a two-dimensional map. f is area preserving if and only if

|J| = 1 where J is the Jacobian corresponding to f .

A formal proof may be obtained in Stuart and Humphries (1998).

a) Let b = 1 in the Hnon map (cf. Exercise 2.8.8), and show that Ha,1 is area preserving.

b) Show that the map (x, y) (xy, ln x) is area preserving too.

c) Compute all nontrivial fixed points of the maps in a) and b) and decide whether the fixed
points are hyperbolic or not.

d) In general, what can you say about the eigenvalues of the linearization of an area preserving
map? 

2.9 Difference-Delay equations

Difference-Delay equations are equations of the form

xt+1 = f (xt , xtT )  (2.9.1)

where T is called the delay.

Referring to population dynamical studies, equation (2.9.1) is often used when one considers species
where there is a substantial time T from birth to sexual maturity. Hence, instead of using a detailed Leslie
matrix model where the fecundities Fi = 0 for several age classes, the more aggregated form (2.9.1) is
often preferred.

173
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

One frequently quoted example is Colin Clarks Baleen whale model (Clark, 1976)

xt+1 = uxt + F (xtT )  (2.9.2)

where xt is the adult breeding population. u ( 0 u 1 ) may be interpreted as a survival coefficient

and the term F (xtT ) is the recruitment which takes place with a delay of T years. In case of the Baleen

whale, 5 T 10 .

A slightly modified version of (2.9.2) was presented by the International Whaling Commission (IWC) as

xt+1 = (1 u)xt + R(xtT )  (2.9.3)

Here (just as in (2.9.2)), (1 u)xt , 0 < u < 1 , is the fraction of the adult whales that survives at
time t and enters the population one time step later.

1    x z 
tT
R(xtT ) = (1 u)T xtT P + Q 1  (2.9.4)
2 K

and regarding the parameters in (2.9.4) we refer to IWC report no. 29, Cambridge (1979). Other models
where a variety of different species are considered may be obtained in Botsford (1986), Tuljapurkar et
al. (1994), Higgins et al. (1997), see also Kot (2001) and references therein.

Now, returning to the general nonlinear equation (2.9.1), the fixed point x is found by letting

xt+1 = xt = xtT = x . The stability analysis follows the same pattern as in section 2.4. Let
xt = x + t where |t |  1 . Then from (2.9.1)

f f
x + t+1 f (x , x ) + (x )t + (x )tT  (2.9.5)
xt xtT

Thus the linearization becomes

t+1 = at + btT  (2.9.6)

where a and b are f /xt , f /xtT evaluated at equilibrium respectively.

174
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

The solution of (2.9.6) is found by letting t = t which after some rearrangements result in the

eigenvalue equation

T +1 aT b = 0  (2.9.7)

which we recognize as a polynomial equation of degree T + 1 . As before, || < 1 guarantees that x

is locally asymptotic stable. The transfer from stability to instability occurs when x fails to be hyperbolic
which means that crosses the unit circle through 1, through 1 or crosses the unit circle at the
location exp(i) .

360
thinking .

360
thinking . 360
thinking .
Discover the truth at www.deloitte.ca/careers Dis

Deloitte & Touche LLP and affiliated entities.

Discover the truth at www.deloitte.ca/careers Deloitte & Touche LLP and affiliated entities.

Deloitte & Touche LLP and affiliated entities.

Discover the truth

175 at www.deloitte.ca/careers
Click on the ad to read more
Download free eBooks at bookboon.com

Deloitte & Touche LLP and affiliated entities.

Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Example 2.9.1. Compute the nontrivial fixed point x and derive the eigenvalue equation of the
model
  xtT 
xt+1 = xt exp r 1  (2.9.8)
K

where r and K both are positive. ((2.9.8) is often called the delayed Ricker model and the
parameters may be interpreted as the intrinsic growth rate ( r ) and the carrying capacity ( K ).)

The fixed point obeys

x
  

x = x exp r 1
K

so clearly, x = K .

The coefficients a and b in (2.9.7) become

  
f K
a= (x ) = 1 exp r 1 =1
xt K
  
f
 r K
b= (x ) = K exp r 1 = r

xtT K K

Hence, the eigenvalue equation may be cast in the form

T +1 T + r = 0  (2.9.9)


Exercise 2.9.1. Consider the difference-delay equation

  xtT 
xt+1 = xt 1 + r 1  (2.9.10)
K

and repeat the calculations from the previous example. 

Exercise 2.9.2. Repeat the calculations in Exercise 2.9.1 for the equation

xt
xt+1 =
1 + xtT  (2.9.11)

176
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Let us now turn back to the general eigenvalue equation (2.9.7). Although it is a polynomial equation
of degree T + 1 , its structure is simpler than most of the equations which we studied in Part II. Therefore,
unless the delay T becomes too large, the Jury criteria work excellent when one tries to reveal stability
properties. (The Baleen whale equations (2.9.4), (2.9.5) were analyzed by use of Theorem 2.1.9.) It is
also possible to use (2.9.7) in order to give a thorough description of the dynamics in parameter regions
where the fixed point is stable.

Our next goal is to demonstrate this by use of the difference-delay equation (2.9.8) and its associated
eigenvalue equation (2.9.9).

As a prelude to the general situation, suppose that T = 0 (no delay) in (2.9.8), (2.9.9). Then, from

(2.9.9), = 1 r from which we may draw the following conclusions: (i) If 0 < r < 1 , then

0 < < 1 , hence from a given initial condition we will experience a monotonic damping towards the
fixed point x = K . (ii) 1 < r < 2 implies that 1 < < 0 , thus in this case there will be oscillatory
damping towards x . (iii) At instability threshold r = 2 it follows that = 1 and a supercritical flip

bifurcation occurs (cf. Exercise 1.5.1). Consequently, in case of r > 2 but |r 2| small, the dynamics
is a stable period-2 orbit.

Next, consider the small delay T = 1 . Then (2.9.9) becomes 2 + r = 0 and by use of (2.1.14)

stability of x = K is ensured whenever the inequalities r > 0 , r + 2 > 0 and r < 1 are satisfied.

Hence, at instability threshold r = 1 but in contrast to the case T = 0 it also follows from (2.1.14) that

is a complex number at bifurcation threshold.

If T = 2 the eigenvalue equation may be written as 3 2 + r = 0 and the four Jury criteria (2.1.16)

simplify to r > 0 , 2 r > 0 , r < 1 and r < (1/2)( 5 1) 0.6180 respectively. Clearly,

r = (1/2)( 5 1) at bifurcation threshold and again we observe that is a complex number.

Now, consider the general case T 1 . From our findings above it is natural to assume that = exp(i)

when the fixed point x = K loses its hyperbolicity. Moreover, the value of r at instability threshold

becomes smaller as T increases which suggests that an increase of T acts as a destabilizing effect.

Substituting = exp(i) into (2.9.9) gives

ei(T +1) = eiT r  (2.9.12)

177
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

which after multiplication by exp(i(T + 1)) may be written as

1 = ei r ei(T +1)  (2.9.13)

Therefore

1 = cos i sin r cos(T + 1) + ir sin(T + 1)

and by separating into real and imaginary parts we arrive at

1 = cos r cos(T + 1)  (2.9.14a)

0 = sin + r sin(T + 1)  (2.9.14b)

Finally, by squaring both equations (2.9.14) and then add we obtain the relation between r and as

r = 2 [cos cos(T + 1) + sin sin(T + 1)] = 2 cos T  (2.9.15)

We will turn your CV into

an opportunity of a lifetime

Do you like cars? Would you like to be a part of a successful brand? Send us your CV on
We will appreciate and reward both your enthusiasm and talent. www.employerforlife.com
Send us your CV. You will be surprised where it can take you.

178 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Substituting back into (2.9.13) then implies

eiT + eiT
 
i
1 =e 2 ei(T +1)
2

and through multiplication by exp(i) we get

ei = ei2T = ei(2T )  (2.9.16)

Thus

= 2T + 2k  (2.9.17)

Hence

(2k + 1)
=  (2.9.18)
2T + 1

From this we may draw the following conclusion. Since r = 2 cos T there are several values of r
which result in modulus 1 solutions of the eigenvalue equation (2.9.9). The smallest r which results in

a modulus 1 solution is clearly when k = 0 , i.e.

T
r2 = 2 cos
2T + 1  (2.9.19)

Let us now focus on possible real solutions of the eigenvalue equation (2.9.9). Assume = R ( R
real). Then from (2.9.9):

r = RT RT +1  (2.9.20)

and since r > 0 , T > 0 it follows that R < 1 . Moreover,

dr
= RT 1 [T (T + 1)R]
dR
such that the maximum value of r occurs when

T
R=  (2.9.21)
T +1

Hence, R is a positive number and the corresponding maximum value of the intrinsic growth rate r
is

TT
r1 =  (2.9.22)
(T + 1)T +1

179
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Exercise 2.9.3. Show that limT r1 = 1/T e .  

Figure 24: The graph of r(T ) = RT RT +1 T = 2

In Figure 24 we have drawn the graph of (2.9.20) in the case T = 2 . (The graph has a similar form for
other T 1 values.) Thus, when R is increasing from 0 to T /(T + 1) , r will increase from 0 to r1

and when R increases from T /(T + 1) to 1, r will decrease from r1 to 0. Clearly, if 0 < r < r1 ,

(2.9.20) has two positive roots. If r > r1 , there are no positive roots.

Following Levin and May (1976) we now have

TT 1 T
r1 = T +1
r2 = cos  (2.9.23)
(T + 1) 2 2T + 1

Indeed, first observe that

 
T T 2 1
cos = sin = sin > =
2T + 1 2T + 1 2 2(2T + 1) 2(2T + 1) 2T + 1

Next, by rewriting r1 :
1 1 1
r1 = 1 T <
(T + 1)(1 + T ) (T + 1)2 2T + 1

which establishes (2.9.23).

180
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Figure 25: 30 iterations of xt+1 = xt exp[r(1 xt1 )] . Monotonic orbit, r = 0.24. Oscillatory orbit, r = 0.90.

AXA Global
Graduate Program
Find out more and apply

181 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Now, considering an orbit starting from an initial value x0 = K , we may from the findings above

conclude that in case of 0 < r < r1 the orbit may approach x = K monotonically. If r1 < r < r2

the orbit will always approach x as a convergent oscillation. If r > r2 , x is unstable and an orbit will

act as a divergent oscillation towards a limit cycle (provided the bifurcation is supercritical). These cases
are demonstrated in Figure 25 and Figure 26. In Figure 25 we show the behaviour of the map
xt+1 = xt exp[r(1 xt1 )] (i.e. K = T = 1 in (2.9.8)) in case of r = 0.24 ( < r1 ) and r = 0.90
( r1 < r < r2 ) respectively and clearly, one orbit ( r = 0.24 ) approaches the fixed point x = 1

monotonically while the other orbit ( r = 0.90 ) approaches x in an oscillatory way. In Figure 26

r = 1.02 and r = 1.10 ( r > r2 ) and there is no convergence towards x . Note that the orbit with
small amplitude ( r = 1.02 ) is almost 6-periodic.

Figure 26: 30 iterations of xt+1 = xt exp[r(1 xt1 )] . Small amplitude orbit, r = 1.02. Large amplitude orbit, r = 1.10.

Remark 2.9.1. Note that in the case 0 < r < r1 we have not actually proved that an orbit must

approach x monotonically. After all (2.9.9) may have complex solutions with magnitudes larger
than = R = T /(T + 1) . However, this is not the case as is proved in Levin and May (1976).

(The proof is not difficult, it involves the same kind of computations as we did when (2.9.19) was
derived.) 

182
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Let us now comment on possible periodic dynamics. Referring to section 2.8 Beyond the Hopf
bifurcation we learned that although the dynamics was a quasistationary orbit just beyond the Hopf
bifurcation threshold, the dynamics could be periodic (exact or approximate) as we penetrated deeper
into the unstable parameter region. Periodic phenomena may of course also occur in difference-delay
equations. Indeed, consider

xt+1 = xt exp[1 r(1 xt1 )]  (2.9.24)

which is nothing but (2.9.8) where T = K = 1 . At bifurcation threshold the dominant eigenvalue
becomes (see (2.9.18))
 
D = exp(i) = exp i  (2.9.25)
3

and since 6D = exp(2i) = 1 , D is equal to 6th root of unity at bifurcation threshold. Therefore, in

case of > D but | D | small, arg is still close to /3 which definitely signals 6-periodic
dynamics. That the dynamics is almost 6-periodic is clearly demonstrated in Figure 26 ( r = 1.02 ).
Through an enlargement of r ( r = 1.10 ) the periodicity is not so profound as the other orbit in Figure

26 shows. More about periodic phenomena in difference delay equations may be obtained in Diekmann
and Gils (2000).

In one way the results presented above are somewhat special in the sense that we were able to find the
complex eigenvalues at bifurcation threshold on closed form (cf. (2.9.18)). Typically, this is not the case.
However, the method we used may still be fruitful in order to bring equations where it is difficult to
locate modulus 1 solutions numerically to a form where it is much more simple. This fact will now be
demonstrated through one example and one exercise.

Example 2.9.2. In Example 2.4.4 (section 2.4) we studied a (n 1) (n 1) Leslie matrix

model with equal fecundities F . If we in addition assume that the year to year survival probabilities

are equal, i.e. P0 = P1 = ... = Pn1 = P , 0 < P < 1 , the eigenvalue equation (2.4.17) may

be cast in the form

n
1 
n+1 (1 x ) P i i = 0  (2.9.26)
D i=0

where

1 P n+1
x = ln(F D) and D = 1 + P + P 2 + ... + P n =
1P

183
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

Our goal is to locate complex modulus 1 solutions of (2.9.26) for given values of P. Using the fact
that P i i is nothing but a geometric series, it is straightforward to rewrite (2.9.26) as


n+2 + An+1 B = 0  (2.9.27)

where
(1 P )(x 1) (1 P )(x 1) n+1
A = P and B= P
1 P n+1 1 P n+1

By inspection, (2.9.27) has a root = P which is located inside the unit circle. The n + 1 other

roots of (2.9.27) are the same roots as of (2.9.26). Hence, assume that = exp(i) in (2.9.27).

Then (we urge the reader to perform the necessary calculations), by using the same method as
we did when we derived (2.9.14) from (2.9.12) we find that

sin = B sin(n + 1)  (2.9.28a)

B 2 A2 1  (2.9.28b)
cos =
2A

e Graduate Programme
I joined MITAS because for Engineers and Geoscientists
I wanted real responsibili www.discovermitas.com
Maersk.com/Mitas e G
I joined MITAS because for Engine
I wanted real responsibili Ma

Month 16
I was a construction Mo
supervisor ina const
I was
the North Sea super
advising and the No
Real work he
helping foremen advis
International
al opportunities
Internationa
ree wo
work
or placements ssolve problems
Real work he
helping fo
International
Internationaal opportunities
ree wo
work
or placements ssolve pr

184 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

We know from Example 2.4.4 that the fixed point of (2.4.7) is stable in case of small equilibrium
populations x . Therefore, numerically it is now easy to find the solutions of (2.9.27) (and (2.9.26))

at bifurcation threshold for given values of n and P by simply increasing F which means that x

is increased too and compute B up to the point where (2.9.28a) is satisfied. Then we compute the
corresponding value of A and finally through (2.9.28b) as

B 2 A2 1
 
= arccos  (2.9.29)
2A


Exercise 2.9.4. Consider the eigenvalue equation

T (1 + a)b T 1 + ab2 T 2 = D  (2.9.30)

where D is real, 0 < a 1 , a < 1/b .

a) Show that (2.9.30) may be written as

T 2 ( ab)( b) = D

b) Assume that D = 0 and conclude that the dominant root of the eigenvalue equation is
= b if 0 < a < 1 or = ab if 1 < a < 1/b .

c) Suppose D = 0 and assume that = R is real and positive. Show that the maximum
value of D is

RT (1 + a)b RT 1 + ab2 RT 2

where

R =
(1 + a)b(T 1) (1 + a)2 b2 (T 1)2 ab2 (T 2)
+
2T 4T 2 T
d) Assume that = exp(i) and separate (2.9.30) in its real and imaginary parts respectively.

Explain how and D may be found numerically in case of given values of a , b and T.

(Equation (2.9.30) arises in an analysis of the general Deriso-Schute model. A thorough

discussion of the model may be obtained in Bergh and Getz (1988).) 

185
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization n-dimensional maps

In this section we have used a variety of different techniques in order to find the roots of polynomial
equations. We close by stating Descartes rule of signs which is a theorem that also may give valuable
insight of location of the roots.

Theorem 2.9.1 (Descartes rule of signs). Consider the polynomial equation

a n + a1 n1 + a2 n2 + ... + an1 + an = 0
0

where an > 0 .

Let k be the number of sign changes between the coefficients an , an1 , ..., a0 disregarding any
which are zero. Then there are at most k roots which are real and positive and, moreover, there

are either k or k 2 or k 4 ... real positive roots. 

Example 2.9.3. Consider

T +1 T + r = 0

where r > 0 . Here k = 2 , hence there are at most 2 real positive roots and, moreover, there are

either 2 or 0 such roots.

Next, suppose that = .

1) If T is an even number, the equation may be written as T +1 T + r = 0 . Thus there

is only one change of sign, consequently there is exactly 1 negative root of T +1 T + r = 0 .
(From our previous analysis of (2.9.9) this means that if 0 < r < r1 , there are 2 positive roots,
1 negative root and T 2 complex roots. If r1 < r , there are T complex roots and 1 negative

root.)

2) If T is an odd number, the equation may be cast in the form T +1 + T + r = 0 . Hence,

there are no sign changes so there are no negative roots . (Thus 0 < r < r1 implies 2 positive
roots and T 1 complex roots. If r1 < r , all T + 1 roots are complex.) 

186
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Part III
Discrete Time
Optimization Problems

187
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

3.1 The fundamental equation of discrete dynamic programming

In the following sections we shall give a brief introduction to discrete dynamic optimization. When one
wants to solve problems within this field there are mainly two methods (together with several numerical
alternatives which we will not treat here) available. Here, in section 3.1, we shall state and prove the
fundamental equation of discrete dynamic programming which perhaps is the most frequently used
method. In section3.2 we shall solve optimization problems by use of a discrete version of the maximum
principle.

Dynamic optimization is widely used within several scientific branches like economy, physics and biology.
As an introduction to the kind of problems that we want to study, let us consider the following example:

Example 3.1.1. Let xt be the size of a population at time t . Further, assume that x is a species

of commercial interest so let ht [0, 1] be the fraction of the population that we harvest at each

time. Therefore, instead of expressing the relation between x at two consecutive time steps as
xt+1 = f (xt ) or (if the system is nonautonomous) xt+1 = f (t, xt ) , we shall from now on
assume that

xt+1 = f (t, xt , ht ) (3.1.1)

If the function f is the quadratic or the Ricker function which we studied in Part I, (3.1.1) may
be written as

xt+1 = r(1 ht )xt [1 (1 ht )xt ] (3.1.2)

or

xt+1 = (1 ht )xt exp[r(1 (1 ht )xt )] (3.1.3)

respectively. In case of an age-structured population model (cf. the various examples treated in
part II) the equation xt+1 = f (t, xt , ht ) may be expressed as

x1,t+1 = F1 ext x1,t (1 h1,t ) + F2 ext x2,t (1 h1,t ) (3.1.4)

x2,t+1 = P x1,t (1 h2,t )

(For simplicity, it is often assumed that ht = h and hi,t = hi which means that the population

or the age classes are exposed to harvest with constant harvest rate(s).)

188
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Now, returning to equation (3.1.1), assume that t = f0 (t, xt , ht ) is the profit we can make of

the harvested part of the population at time t . Our ultimate goal is to maximize the profit over

a time period from t = 0 to t = T , i.e. we want to maximize the sum of the profits at times
t = 0, 1, ..., T . This leads to the problem

T
f0 (t, xt , ht ) (3.1.5)

maximize
h0 ,h1 ,...,hT
t=0

subject to equation (3.1.1) given the initial condition x0 and ht [0, 1] .

To be somewhat more precise, we have arrived at the following situation: Suppose that we at time
t = 0 apply the harvest rate h0 . Then, according to (3.1.1) x1 = f (0, x0 , h0 ) is known at time
t = 1 . Further, assume that we at time t = 1 choose the harvest h1 . Then x2 = f (1, x1 , h1 ) is
known and continuing in this fashion, applying (different) harvest rates ht at each time we also
know the value of xt at each time. Consequently, we also know the profit t = f0 (t, xt , ht ) at
each time. As stated in (3.1.5) our goal is to choose h0 , h1 , ..., hT in such a way that
T
t=0 f0 (t, xt , ht ) is maximized. 

Let us now formulate the situation described in Example 3.1.1 in a more general context. Suppose that the
state variable x evolves according to the equation xt+1 = f (t, xt , ut ) where x0 is known. At each time
t the path that x follows depends on discrete control variables u0 , u1 , ..., uT . (In Example 3.1.1 we used
harvest rates as control variables.) We assume that ut U where U is called the control region. The sum
T
t=0 f0 (t, xt , ut ) where f0 is the quantity we wish to maximize is called the objective function.

Definition 3.1.1. Suppose that xs = x . Then we define the value function as

T

Js (x) = maximize f0 (t, xt , ut ) (3.1.6)
us ,us+1 ,...,uT
t=s

Hence, a more general formulation of the problem we considered in Example 3.1.1 is: maximize Js (x)
subject to xt+1 = f (t, xt , ut ) , xs = x and ut U .

We now turn to the question of how to solve the problem.

189
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Suppose that we know the optimal control (optimal with respect to maximizing (3.1.6)) us at s = 0 .
Then, according to the findings presented in Example 3.1.1, we find the corresponding x1 as
x1 = f (0, x0 , u0 (x0 )) and if we succeed in finding the optimal control u1 (x1 ) at time t = 1 we have
x2 = f (1, x1 , u1 (x1 )) and so on. Thus, suppose that xs = x at time t = s , how shall we choose us
in the best optimal way? Clearly, if we choose us = u as the optimal control we achieve the immediate
benefit f0 (s, x, u) and also xs+1 = f (s, x, u) . This consideration simply means that the highest total
benefit which is possible to get from time s + 1 to T is Js+1 (xs+1 ) = Js+1 (f (s, x, u)) . Hence, the
best choice of us = u at time s is the one that maximizes f0 (s, x, u) + Js+1 (f (s, x, u)) . Consequently,
we have the following theorem:

Theorem 3.1.1. Let Js (x) defined through (3.1.6) be the value function for the problem

T

maximize f0 (t, xt , ut ) xt+1 = f (t, xt , ut )
u
t=0

where ut U and x0 are given. Then

Js (x) = max [f0 (s, x, u) + Js+1 (f (s, x, u))] , s = 0, 1, ..., T 1 (3.1.7)

uU

JT (x) = max f0 (T, x, u) (3.1.8)

uU

Theorem 3.1.1 is often referred to as the fundamental equation(s) of dynamical programming and serves
as one of the basic tools for solving the kind of problems that we considered in Example 3.1.1. As we
shall demonstrate through several examples, the theorem works backwards in the sense that we start
to find uT (x) and JT (x) from (3.1.8). Then we use (3.1.7) in order to find JT 1 (x) together with
uT 1 (x) and so on. Hence, all value functions and optimal controls are found recursively.

Example 3.1.2.

T

maximize (xt + ut ) xt+1 = xt 2ut , ut [0, 1] , x0
u
t=0

Solution: From (3.1.8), JT (x) = maxu (x + u) so clearly, the optimal value of u is u = 1 .

Hence at time t = T , JT (x) = x + 1 and uT (x) = 1 .

190
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Further, from (3.1.7):

JT 1 (x)JT=1max u [x
(x) = max + JT (x
+ uu [x++JTu (x 2u)] = max
2u)] u [x u+
= max [xu++u (x
+ (x 2u2u++1)]
1)]==max
maxuu[2x +1]1] .
[2x uu+
Consequently, u = 0 is the optimal choice, thus at t = T 1 we have JT 1 (x) = 2x + 1

and uT 1 (x) = 0 .

This implies: JT 2 (x) = maxu [x + u + JT 1 (x 2u)] = maxu [3x 3u + 1] so again

u = 0 is the best choice and JT 2 (x) = 3x + 1 and uT 2 (x) = 0 .

From the findings above it is natural to suspect that in general

J (x)
T k
= (k + 1)x + 1 , uT k (x) = 0 , k = 1, 2, ..., T

The formulae is obviously correct in case of k = 1 and by induction we have from (3.1.7) that

JT (k+1) = max[x + u + JT k (x 2u)]

u
= max[x + u + (k + 1)(x 2u) + 1] = max[(k + 2)x 2(k + 1)u + 1]
u u

= (k + 2)x + 1 = [(k + 1) + 1]x + 1

hence the formulae is correct at time T (k + 1) as well. Therefore

JT k (x) = (k + 1)x + 1 , uT k (x) = 0 , k = 1, 2, ..., T

J (x) = x + 1
T
uT (x) = 1

Example 3.1.3.
T
 T
2 2
maximize
maximize (u
(u +t
t +t u ut xt )xt ) xt+1
xt+1
= x=t + t ,
xt u+t ,ut , ut u ,

, x, 0 x0
t=0t=0

Solution: From (3.1.8), JT (x) = maxu (u2 x + u) and since the function
u22
=u
h(u) =
h(u) +uu clearly is concave in u the optimal choice of u must be the solution of
xx+
h (u) = 0 , i.e. u = 1/2 . Hence, at time t = T , uT (x) = 1/2 and
JT (x) = (1/4) x + (1/2) = x + (1/4) .

191
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Further, (3.1.7) gives

JT 1 (x) = maxu [u2 x + u + JT (x + u)] 2= maxu [u2 x + u (x + u) + (1/4)] =
2 JT 1 (x) = maxu [u x + u + JT (x + u)] = maxu [u2 x + u (x + u)
max u [u 2x + (1/4)]
maxu [u2 x + u (x + u) + (1/4)] = maxu [u2 2x + (1/4)] and again since h1 (u) = u2 2x + (1/4)
is concave in u we find that u = 0 is the optimal choice. Thus JT 1 (x) = 2x + (1/4) and
uT 1 (x) = 0 .

Proceeding in the same way (we urge the reader to work through the details) we find that
JT 2 (x) = 3x + (1/2) , uT 2 (x) = (1/2) and JT 3 (x) = 4x + (3/2) , uT 3 (x) = 1
uT 3 (x) = 1 .

Therefore, it is natural to suppose that

JT k (x) = (k + 1)x + bk

where b0 = 1/4 and uT k (x) = k1

2 ,
k = 1, 2, ..., T . The formulae is obviously correct
when k = 0 and by induction

JT (k+1) = max[u2 + u x + JT k (x + u)]

u
= max[(k + 2)x u2 ku + bk ]
u

Again, we observe that the function inside the bracket is concave in u so its maximum occurs at
u = (k/2) which means that the corresponding value function becomes

J
T (k+1)
(x) = [(k + 1) + 1]x + bk + k 2 /4 = [(k + 1) + 1]x + bk+1

It remains to find bk . The equation bk+1 bk = k 2 /4 has the homogeneous solution C 1k = C .

Referring to the remark following Example 3.1.4 we assume a particular solution of the form
pk = (A + Bk + Dk 2 )k . Hence, after inserting into the equation and equating terms of equal
power of k we find that A = 1/24 , B = (1/8) and D = 1/12 so the general solution

becomes bk = C + (1/24)k (1/8)k 2 + (1/12)k 3 . Finally, using the fact that b0 = 1/4

which implies that C = 1/4 , we obtain

1 k1
JT k (x) = (k + 1)x + (6 + k 3k 2 + 2k 3 ) uT k (x) =
24 2

192
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Example 3.1.4 (Exam exercise, UiO).

T

maximize (xt ut ) xt+1 = ut xt , ut [0, 2] , x0
u
t=0

Solution: JT (x) = maxu (x u) . Clearly, u = 0 is the optimal choice so JT (x) = x and

uT (x) = 0 , JT 1 (x) = maxu [x u + JT (ux)] = maxu [x + (x 1)u] . Thus, if x 1
we choose u = 2 and if x < 1 we choose u = 0 . Consequently,

x + (x 1)2 = 3x 2 if x 1 and uT 1 (x) = 2


JT 1 (x) =

x + (x 1)0 = x if x < 1 and uT 1 (x) = 0

(Note that JT 1 (x) is a convex function which is continuous at x = 1 .)

In order to compute JT 2 (x) we must consider the cases JT 1 (x) = 3x 2 and

JT 1 (x) = x separately.

170x115_MIM_ datos_Espv4.pdf 1 17/12/15 18:29

EL 93%
DE LOS ALUMNOS DEL MIM ESTN
TRABAJANDO EN SU SECTOR
C
A LOS 3 MESES DE GRADUARSE
M

MASTER IN MANAGEMENT
CM

MY

CY ESTUDIA EN EL CORAZN DE MADRID Y APROVECHA LAS OPORTUNIDADES Y EXPERIENCIAS NICAS

Duracin: 1O MESES
QUE OFRECE LA CAPITAL DE ESPAA
CMY Exp. Profesional: 1 AO
POTENCIA TU FORMACIN CURSANDO UNA DOBLE TITULACIN QUE MEJOR SE ADAPTE A TUS
Idioma: ESPAOL / INGLS
K
OBJETIVOS PROFESIONALES
Formato: TIEMPO COMPLETO
ESTUDIA UN SEMESTRE EXTRA EN OTRO PAS CON EL PROGRAMA BEYOND BORDERS EXPERIENCE
Convocatorias: SEPT / FEB
Y AUMENTA TU RED DE CONTACTOS

5 ESPECIALIZACIONES 100 DEL MUNDO 55 NACIONALIDADES

MASTERS IN MANAGEMENT
PERSONALIZA TU PROGRAMA FINANCIAL TIMES
EN CLASE

www.ie.edu/master-management [email protected] Sguenos en IE MIM Experience

193 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Assuming JT 1 (x) = 3x 2 we obtain

JT 2 (x) = max[x u + 3ux 2] = max[x + (3x 1)u 2]

u u

Figure 27: JT 2 (x) possibilities.

so if x 1/3 our optimal choice is u = 2 and if x < 1/3 we choose u = 0 .

In the same way, using JT 1 (x) = x , we find

JT 2 (x) = max[x u + ux] = max[x + (x 1)u]

u u

so whenever x 1 , u = 2 and if x < 1 our best choice is u = 0 .

Hence, the possibilities are

x + (3x 1) 2 2 = h1 (x) = 7x 4 if x 1/3
x + (3x 1) 0 2 = h2 (x) = x 2 if x < 1/3

JT 2 (x) =

x + (x 1) 2 = h3 (x) = 3x 2 if x1
x + (x 1) 0 = h4 (x) = x if x<1

In Figure 27 we have drawn the graphs of the hi functions in their respective domains. The point

of intersection between h1 (x) and h4 (x) is x = 2/3 so clearly, if x 2/3 , h1 (x) is the largest

function. If x < 2/3 , h4 (x) is the largest function.

194
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Consequently, we conclude that

7x 4 if x 2/3 and uT 2 (x) = 2


JT 2 (x) =

x if x < 2/3 and uT 2 (x) = 0

and again we notice that JT 2 (x) is a convex function which is continuous at x = 2/3 .

Now at last, let us try to find the general expression JT k (x) . The formulaes for JT 1 and JT 2
suggest that our best assumption is

bk
ak x + bk x 1ak
=c
JT k (x) = bk
x x< 1ak
=c

k = 1, 2, ..., T and that uT k (x) = 2 if x c and uT k (x) = 0 if x < c .

The formulae is certainly correct in case of k = 1 . Further, by using the same kind of considerations
as in the computation of JT 2 (x) and induction there are two separate cases.

JT (k+1) (x) = max[x u + ak ux + bk ] = max[x + (ak x 1)u + bk ]

u u

Hence x 1/ak u = 2 and x < 1/ak u = 0 , and

JT (k+1) (x) = max[x u + ux] = max[x + (x 1)u]

u u

Thus x 1 u = 2 and x < 1 u = 0 .

This yields (just as in the JT 2 (x) case) the following

(2ak + 1)x + bk 2 = ak+1 x + bk+1 = g1 (x) x 1/ak
x + bk = g2 (x) x < 1/ak

JT (k+1) (x) =

3x 2 = g3 (x) x1
x = g4 (x) x<1

and we recognize that the forms of g1 (x) and g4 (x) are in accordance with our assumption and

moreover that the point of intersection between g1 (x) and g4 (x) is bk (1 ak )1 which also

is consistent with the assumption.

195
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Further, ak obeys the difference equation ak+1 = 2ak + 1 . Therefore, the general solution is
ak = D 2k 1 and since a1 = 3 D = 2 we have ak = 2k+1 1 . In the same way,
bk+1 = bk 2 (see the remark following this example, see also (1.1.2b)) has the general solution
bk = K 2k and since b1 = 2 K = 0 we obtain bk = 2k .

Finally, since (1) g1 (1) g3 (1) and ak+1 > 3 . (2) g4 (x) > g2 (x) and (3) g1 (x) > g4 (x)

when x > bk+1 (1 ak+1 )1 (recall that ak+1 > 3 ) we obtain the general solution

k
(2k+1 1)x 2k x uT k = 2

2k 1
JT k (x) = k
x x< 2k 1
uT k = 0

196 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Remark 3.1.1: Referring to section 2.1, Exercise 2.1.3, the difference equation
xt+2 5xt+1 6xt = t 2 has the homogeneous solution C1 (1) + C2 6 and since the
t t t

exponential function 2t on the right-hand side of the equation is different from both exponential

functions contained in the homogeneous solution it suffices to assume a particular solution of the
form (At + B)2t in this case. In Example 3.1.3 we had to solve an equation of the form
xt+1 xt = at2 . The homogeneous solution is C 1t = C but since at2 = at2 1t we have
the same exponential function on both sides of the equation. Therefore, we must in this case
assume a particular solution of the form (A + Bt + Dt2 )t . In the same way, if xt+1 xt = bt

we assume a particular solution (A + Bt)t and finally, in the case xt+1 xt = K , assume a

particular solution A + Bt (cf. (1.1.2b)). 

Exercise 3.1.1. Let a be a positive constant and solve the problem

T

max (xt + ut ) xt+1 = xt aut , ut [0, 1] , x0
u
t=0

Exercise 3.1.2. Solve the problem (Exam Exercise, UiO):

T

max (xt ut ) xt+1 = xt + ut , ut [0, 1] , x0
u
t=0

(Hint: Use Remark 3.1.1.) 

Exercise 3.1.3. Solve the problem:

T

max (xt + 1) xt+1 = ut xt , ut [0, 1] , x0
u
t=0

197
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

3.2 The maximum principle (Discrete version)

When t is a continuous variable, most optimization problems are formulated and solved by use of the
maximum principle which was developed by Russian mathematicians about 60 years ago. The maximum
principle, sometimes referred to as Pontryagins maximum principle, is the cornerstone in the discipline
called optimal control theory which may be regarded as an extension of the classical calculus of variation.
An excellent treatment of various aspects of control theory may be found in Seierstad and Sydster (1987),
see also Sydster et al. (2005). In this section we shall briefly discuss a discrete version of the maximum
principle which offers an alternative way of dealing with the kind of problems presented in section 3.1.

Consider the problem

T

maximize f0 (t, xt , ut ) , ut U , U (3.2.1)
t=0

xt+1 = f (t, xt , ut ) t = 0, 1, ..., T 1 x0

together with one of the following terminal conditions

xT xT XT xT = XT (3.2.2)

Thus, the problem that we consider here is somewhat more general than the one presented in section
3.1 due to the terminal conditions (3.2.2b,c).

Next, define the Hamiltonian by


f0 (t, x, u) + pf (t, x, u) t < T
H(t, x, u, p) = (3.2.3)
f0 (t, x, u) t=T

where p is called the adjoint function.

Then we have the following:

Theorem 3.2.1 (The maximum principle, discrete version). Suppose that (xt , ut ) is an optimal
sequence for problem (3.2.1), (3.2.2). Then there are numbers p0 , ..., pT such that

ut Hu (t, xt , ut , pt )u u U (3.2.4)

Moreover,

pt1 = Hx (t, xt , ut , pt ) , t = 1, ..., T 1 (3.2.5a)


pT 1 = f0x (T, xT , uT ) + pT (3.2.5b)

198
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

and to each of the terminal conditions (3.2.2) we have the following transversal conditions

a) pT = 0. (3.2.6a)

b) pT 0 xT > XT (3.2.6b)

c) pT (3.2.6c)

Theorem 3.2.1 gives necessary conditions for optimality. Regarding sufficient conditions we have:

Theorem 3.2.2. Suppose that (xt , ut ) satisfies all the conditions in Theorem 3.2.1 and in addition

that H(t, x, u, p) is concave in (x, u) for every t . Then (xt , ut ) is optimal. 

Proof. Our goal is to show that

T
 T

K= f0 (t, xt , ut ) f0 (t, xt , ut ) 0
t=0 t=0

Excellent Economics and Business programmes at:

The perfect start

of a successful,
international career.

CLICK HERE
to discover why both socially
and academically the University
of Groningen is one of the best
places for a student to be
www.rug.nl/feb/education

199 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Introducing the notation f0 = f0 (t, x, u) , f0 = f0 (t, x , u ) and so on, it follows from

(3.2.3) that

T
 T

K= (Ht Ht ) + pt (ft ft )
t=0 t=0

Now, since H is concave in (x, u) we also have that H H Hx (x x ) + Hu (u u ).

Thus

T T T 1

  
K Hu (ut ut ) + Hx (xt xt ) + pt (ft ft )
t=0 t=0 t=0

Due to (3.2.4) and the concavity of H the first of the three sums above are equal or larger than

zero. Indeed, suppose ut [u0 , u1 ] . If ut (u0 , u1 ) then Hu = 0 . If ut = u0 , then Hu 0

and ut ut 0 and finally, if ut = u1 , Hu 0 and ut ut 0 , hence in all cases

Hu (ut ut ) 0 .

Regarding the second and the third sum they may by use of (3.2.5a), (3.2.5b) and (3.2.1) be written as

T 1
 T 1

pt1 (xt xt ) + (pT 1 pT )(xT xT ) + pt (xt+1 xt+1 )
t=0 t=0

= pT (xT xT ) = K1

Next, assume xT free. Then from (3.2.6a), pT = 0 which implies K1 = 0 . If xT XT , (3.2.6b)

gives pT 0 and since xT XT we must have K1 0 if xT = XT . If xT > XT , pT = 0 ,

thus in either case K1 0 . Finally, if xT = XT , K1 = 0 . Therefore, whatever terminal condition

(3.2.2), K1 0 which implies K 0 so we are done. 

Example 3.2.1. Solve the problem given in Example 3.1.2 by use of Theorems 3.2.1 and 3.2.2.

Solution: From (3.2.3) it follows


x + u + p(x 2u) t < T
H(t, x, u, p) =

x+u t=T

Consequently, whenever t < T , Hx = 1 + p and Hu = 2p and if t = T , Hx = Hu = 1 .

200
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

By use of the results above, (3.2.5a,b) gives

pt1 = 1 + pt t<T, pT 1 = 1 + pT

and since xT is free, (3.2.6a) implies that pT = 0 so pT 1 = 1 .

The equation pt1 = 1 + pt may be rewritten as pt+1 pt = 1 and its general solution is
easily found to be pt = C t . Further, since PT 1 = 1 it follows that 1 = C (T 1) . Thus
C = T so pt = T t and we observe that pt > 0 for every t < T .

From the preceding findings, (3.2.4) may be formulated as

u = ut 2(T t)u t < T

u = uT 1u t=T

Accordingly, we make the following choices: If t = T , choose uT = 1 . If t < T (recall that

2(T t) < 0 ), choose ut = 0 for every t . Hence, we have arrived at the same conclusion
as we did in Example 3.1.2.

A final observation is that the Hamiltonian is linear in (x, u) so H is also concave in (x, u) .
Consequently, (xt , ut ) solves the problem ( xt is found at each t from the equation
xt+1 = xt 2ut and x0 is given). 

Example 3.2.2. Solve the problem

T

maximize (xt ut ) xt+1 = xt + ut
u
t=0

x0 = 1 , xT = XT , 1 < XT < T + 1 , ut [0, 1] .

Solution:

x u + p(x + u) t < T
H(t, x, u, p) =

xu t=T

Therefore, whenever t < T , Hx = 1 + p , Hu = 1 + p and if t = T , Hx = 1 and Hu = 1 .

201
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Further, (3.2.5b) gives pT 1 = 1 + pT and (3.2.5a) gives pt1 = 1 + pt if t < T . Clearly (cf.

our previous example), the latter difference equation has the general solution pt = C t so pt

is a decreasing sequence of points.

From (3.2.4) it follows

u = ut (1 + pt )u t < T

u = uT 1u t=T

Thus at t = T the optimal control is uT = 0 . In the case t < T we have that if pt 1 0 ,

then u = ut = 1 and if pt 1 < 0 , we choose ut = 0 .

First, assume pt 1 0 for all t < T . Then ut = 1 and xt+1 = xt + 1 which has the general
solution xt = K + t . x0 = 1 K = 1 , which means that xt = t + 1 . This implies that
xT = T + 1 but this is a contradiction since XT < T + 1 . Next, assume pt 1 < 0 for all
t T . Then ut = 0 . Thus, xt+1 = xt which has the constant solution xt = M . Again we
have reached a contradiction since 1 < XT .

American online
LIGS University
is currently enrolling in the
Interactive Online BBA, MBA, MSc,
DBA and PhD programs:

enroll by September 30th, 2014 and

save up to 16% on the tuition!
pay in 10 installments / 2 years
Interactive Online education
visit www.ligsuniversity.com to
find out more!

Note: LIGS University is not accredited by any

nationally recognized accrediting agency listed
by the US Secretary of Education.
More info here.

202 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Finally, let us suppose that there exists a time tc such that whenever t tc , then pt 1 0

and in case of tc < t T , pt 1 < 0 .

First, consider the case t tc . Then xt+1 = xt + 1 so xt = K + t . x0 = 1 K = 1 , hence

xt = t + 1 . If t > tc we have xt+1 = xt . Hence, xt is a constant, say xt = M , and since
xt = XT it follows that xt = XT .

Thus,

t tc , pt 1 = C t 1 0 xt = t + 1 ut = 1

t > tc , pt 1 = C t 1 < 0 xt = XT ut = 0

It remains to determine tc and the constant C . At time tc , C tc 1 = 0 so C = tc + 1 .

Therefore, pt = tc t . Further, from xtc +1 = xtc + utc we obtain XT = tc + 1 + 1 so

tc = XT 2 . Consequently, by use of the conditions in the maximum principle we have
arrived at

xt = t + 1 ut = 1 0 t XT 2

xt = XT ut = 0 XT 2 < t T

and pt = XT 2 t for every t . Finally, since H is linear and concave in (x, u) it follows

from Theorem 3.2.2 that we have obtained the solution. 

We close this section by looking at one extension only.

If we have a problem which involves several state variables x1 , ..., xn and several controls u1 , ..., um
we may organize them in vectors, say x = (x1 , ..., xn ) , u = (u1 , ..., um) and reformulate problem

(3.2.1), (3.2.2) as
T

f0 (t, xt , ut ) (3.2.7)
t=0

subject to xt+1 = f(t, xt , ut ) , x0 given, ut U , and terminal conditions on the form

xi,T xi,T Xi,T xi,T = Xi,T (3.2.8)

The associated Hamiltonian may in case of so-called normal problems be defined as

 n
f0 (t, x, u) + pi fi (t, x, u) t < T
H(t, x, u, p) = i=1 (3.2.9)
f0 (t, x, u) t=T

203
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

where p = (p1 , ..., pn ) is the adjoint function.

Then we may formulate necessary and sufficiently conditions for an optimal solution in the same way
as we did in the one-dimensional case.

Theorem 3.2.3. Suppose that (xt , ut ) is an optimal sequence for problem (3.2.7), (3.2.8) with

Hamiltonian defined as in (3.2.9). Then there exists p such that

m
 H
u = ut (t, xt , ut , pt )ui (3.2.10)
i=1
ui

Moreover

pi,t1 = Hx i (t, xt , ut , pt ) , t = 1, ..., T 1 (3.2.11a)

f0
pi,T 1 = (T, xt , ut ) + pi,T (3.2.11b)
xi
and

a) pi,T = 0 if the terminal condition is (3.2.9a).

b) pi,T 0 (= 0 if xi,T > Xi,T ) (3.2.12)

if the condition is (3.2.9b).

c) pi,T free if condition (3.2.9c) applies.

Finally, if H is concave in (x, u) for each t then (xt , ut ) solves problem (3.2.7), (3.2.8). 

As usual, we end with an example.

Example 3.2.3. Solve the problem

T
 1
max (u2t 2xt ) xt+1 = yt , yt+1 = ut + yt
2
t=0

x0 = 2 , y0 = 1 , ut R , xT free, yT free.

Solution: Denoting the adjoint functions by p and q respectively, the Hamiltonian becomes

204
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

u2 2x + 12 py + q(u + y) t < T

H(t, x, y, u, p, q) =

u2 2x t=T

which implies

Hx = 2 Hy = 12 p + q Hu = 2u + q t < T

H  = 2 Hy = 0
x
Hu = 2u t=T

Then, from (3.2.11a) it follows that pt1 = 2 , qt1 = (1/2)pt + qt and since xT , yT is free,

(3.2.12a) implies pT = qT = 0 . Thus (3.2.11b) reduces to pT 1 = 2 and qT 1 = 0 .

Consequently, pt = 2 for each t and if we insert this result into the difference equation for q
we easily obtain the general solution qt = C + t . Moreover, since qT 1 = 0 it follows that

0 = C + T 1 so C = 1 T which means that qt = t T + 1 .

Now, since the control region is open, it follows from (3.2.10) that Hu = 0 , thus 2ut + qt = 0
if t < T and 2uT = 0 whenever t = T . Hence at time t = T , uT = 0 and in case of t < T ,
ut = (1/2)qt = 1/2(t T + 1) .

205 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Therefore, the problem is in many respects already solved. Indeed, yt is now uniquely determined

from the relation yt+1

= ut + yt (recall that y0 = 1 ) and xt is subsequently found from
xt+1 = (1/2)yt . We leave the details to the reader.

Finally, observe that the Hesse determinant of H ( t < T ) may be written as

 
 0 0 0 
 
 0 0 0 
 
 0 0 2 

so clearly (1)1 1 0 , (1)2 2 = 0 , (1)3 3 = 0 where i is all possible principal

minors of order i respectively. Consequently H is concave in (x, y, u) . (At time t = T the

result is clear.) 

Exercise 3.2.1. Solve Exercises 3.1.2 and 3.1.3 by use of the maximum principle. 

3.3 Infinite horizon problems

In the previous two sections we considered discrete dynamic optimization problems where the planning
period T was finite. Our goal here is to study problems where T . Such problems are called

infinite horizon problems. Note that the extension from the finite to the infinite case is by no means
straightforward. Indeed, since the sum we want to maximize now consists of an infinite number of terms
we may obviously face convergence problems which were absent in sections 3.1 and 3.2.

There are mainly two different solution methods available (along with some numerical alternatives) when
we deal with infinite horizon problems. The first method which we will describe is based upon Theorem
3.1.1 (The fundamental equation of discrete dynamic programming).

Consider the problem


maximize t f0 (xt , ut ) (3.3.1)
u
t=0

subject to xt+1 = f (xt , ut) , (0, 1) , x0 given, ut U . Clearly (3.3.1) is an autonomous system
and it serves in many respects as a standard problem in the infinite horizon case. Especially economists
study systems like (3.3.1). Indeed, they often assume that = 1/(1 + r) is a discount factor where r

is the interest rate. Under this assumption, (3.3.1) may be interpreted as maximizing the present value
of a quantity like a profit or a utility function f0 (x, u) subject to xt+1 = f (xt , ut) over all times

regardless of any terminal conditions.

206
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Now, returning to (3.3.1), in order to ensure convergence of the series, we impose the restriction:

K1 f0 (x, u) K2 (3.3.2a)

where K1 and K2 are constants, or

f0 (xt , ut ) c t (3.3.2b)

where (0, 1 ) and 0 < c < .

Next (compare with section 3.1), define the (optimal) value function at time t = s as


Js (x) = max t f0 (xt , ut ) = s J s (x) (3.3.3)
u
t=s

where


s
J (x) = max ts f0 (xt , ut ) (3.3.4)
u
t=s

Denoting J0 (x) = J(x) we now have the following result:

Theorem 3.3.1 (Bellmans equation). Consider problem (3.3.1) under the restriction(s) (3.3.2).
Then the (optimal) value function J0 (x) = J(x) defined through (3.3.3) satisfies

J(x) = max [f0 (x, u) + J(f (x, u))] (3.3.5)

u

Proof. Since the horizon is infinite, J s+1 (x) = J s (x) . Hence,

Js+1 (x) = s+1 (x)J s+1 (x) = s J s (x) = Js (x)

Now, using the same argument as we did in the last paragraph before Theorem 3.3.1 was
established it now follows:



J(x) = J0 (x) = max f0 (x0 , u0 ) + max t f0 (xt , ut )
u u1 ,...
t=1
= max [f0 (x0 , u0 ) + J1 (x)] = max [f0 (x0 , u0 ) + J1 (f (x0 , u0 ))]
u u

= max [f0 (x0 , u0 ) + J0 (f (x0 , u0 ))]

u

207
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Remark 3.3.1. Note the fundamental difference between equations (3.1.7), (3.1.8) in Theorem
3.1.1 and equation (3.3.5) in Theorem 3.3.1. (3.1.7) relates the value function J at different times
T, T 1, ... and as we have demonstrated, the (finite) optimization problem could then be solved
recursively. Regarding (3.3.5), this is not the case. Bellmans equation is a functional equation and
there are no general solution methods for such equations. Therefore, often the best one can do is
to guess the appropriate form of J(x) for a given problem. 

Remark 3.3.2. In the proof of Theorem 3.3.1 it is implicitly assumed that the maximum exists at
each time step. This is not necessarily true but (3.3.5) still holds if we use the supremum notation
instead of the max notation. 

Let us now by way of examples show how Theorem 3.3.1 applies.

Example 3.3.1. Solve the problem


max t xt ut
u
t=0

Join the best at Top masters programmes

3
 3rd place Financial Times worldwide ranking: MSc
the Maastricht University International Business
1st place: MSc International Business
School of Business and 1st place: MSc Financial Economics
2nd place: MSc Management of Learning

Economics! 2nd place: MSc Economics

2nd place: MSc Econometrics and Operations Research
2nd place: MSc Global Supply Chain Management and
Change
Sources: Keuzegids Master ranking 2013; Elsevier Beste Studies ranking 2012;
Financial Times Global Masters in Management ranking 2012

Maastricht
University is
the best specialist
university in the
Visit us and find out why we are the best! Netherlands
(Elsevier)
Masters Open Day: 22 February 2014

www.mastersopenday.nl

208 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

subject to xt+1 = (1 ut )xt , (0, 1) , u (0, 1) , x0 given. 

Solution. First, consider f0 (xt , ut ) = xt ut . Clearly, 0 < ut < 1 and since xt+1 < xt it
follows that 0 < f0 (xt , ut ) < x0 . Hence, (3.3.2a) is satisfied.

Next, from Theorem 3.3.1:

 
J(x) = max xu + J((1 u)x)
u

Assume that J(x) = x . > 0 . Then
 
x = max x u + 1 u x
u

Thus

 
= max u + 1 u (3.3.6)
u

Defining g(u) = u + 1 u , the maximum of [ ] occurs when g  (u) = 0 , i.e. when

1
u= (3.3.7)
1 + ()2

and by inserting into (3.3.6) we eventually arrive at

(1 2 )1 (3.3.8)

=

Finally, by substituting (3.3.8) back into (3.3.7) we obtain u = 1 2 so consequently the

solution is

x

J(x) = x = (3.3.9)
1 2

with associated optimal control u = 1 2 .

For comparison reasons let us also compute the maximum value of the infinite series in
another way. From the constraint it follows that

x
t+1
= (1 ut )xt = 2 xt

Thus xt = 2t x0 . Consequently, the series becomes

209
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

x0
    
2t 2 2 2t
1 x0 = (1 )x0 =
1 2
t=0 t=0

in accordance with (3.3.9) ( x0 = x ). 

Example 3.3.2. Assuming J(x) = x2 , > 0 , solve the problem:


max t (x2t u2t )
u
0

subject to xt+1 = xt + ut , (0, 1) , u (, ) , x0 > 0 given. 

Solution. From Theorem 3.3.1

J(x) = max x2 u2 + J(x + u)

 
u

Thus (due to the assumption)

x2 = max x2 u2 (x + u)2 = max x2 u2 x2 2xu u2

   
u u

The function g(u) = u2 2xu u2 is clearly concave in u , hence [ ] attains its

maximum where g  (u) = 0 which gives

x
u= (3.3.10)
1 +

Consequently,

2 2 22 2 2 3 3
x2 = x2 2
x2
x2
+ x 2
x2

(1 + ) 1 + (1 + )

so after cancelling by x2 and rearranging we eventually arrive at

(1 + ) 2 + (2 1) + 1 = 0 (3.3.11)
 

Now, since > 0 , the only acceptable solution of (3.3.11) is



2 1 1 + 4 2 (3.3.12)
= +
2 2

210
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Hence,

1   
J(x) = 2 1 + 1 + 4 x2 (3.3.13)
2
2

It is still no yet clear that we have solved the problem. We must check if (3.3.2a) is satisfied. Clearly,
f0 (xt , yt ) = x2t yt2 0 so if the sum shall be maximized it is natural to assume that
|xt+1 | |xt | . Hence |xt | x0 . In the same way |ut+1 | |xt | x0 . Under this assumption
(3.3.13) will solve the problem. 

Exercise 3.3.1. Solve the problem


max t (xt ut )
u
0

subject to xt+1 = 12 xt + 12 ut , (0, 1) , u 0 , x0 given. 

Exercise 3.3.2. Consider the problem

211 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems


max t (u2t xt )
u
0

subject to xt+1 = (1 ut )xt , 0, 1 , u ,  , x0 given.

a) Suppose that J(x) = x and use Bellmans equation to show that

4(1 )
J(x) = x
2

with associated optimal control

2( 1)
u =

b) Try to evaluate the sum of the series in the same way as we did at the end of Example 3.3.1
and conclude whether the found J(x) solves the problem or not. 

Exercise 3.3.3. Find J(x) and ut for the problem


max t (e2xt )
u
0

subject to xt+1 = xt 2ut , (0, 1) , u [1, 1] , x0 given. 

Our next goal is to show how infinite horizon problems may be solved by use of the maximum
principle.

Consider the problem


maximize f0 (t, xt , ut)dt (3.3.14)
u
t=0

subject to xt+1 = f (t, xt , ut ) , x0 given together with one of the following terminal conditions:

lim x(T ) = x (3.3.15a)

T

limT x(T ) x (3.3.15b)

212
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

limT x(T ) free (3.3.15c)

Remark 3.3.3. Recall the definition of limt :

limt f (t) = lim inf {f (s)|s [t, )}

t

which means that limt f (t) a implies that for each > 0 there exists a t such that
t > t implies that f (t) a . 

Remark 3.3.4. Note that both f0 and f may depend explicitly on t in problem (3.3.14), (3.3.15)
which is in contrast to the case covered by Bellmans equation. Also note the more general terminal
conditions (3.3.15a,b,c). 

Let the Hamiltonian H be defined just as in section 3.2. Then we have the following:

Theorem 3.3.2 (Maximum principle, infinite horizon). Suppose that ({xt }, {ut }) is an optimal

sequence for problem (3.3.14), (3.3.15). Then there exist numbers pt such that for t = 0, 1, 2, ...

Hu (t, xt , ut , pt )(ut ut ) 0 (3.3.16)

pt1 = Hx (t, xt , ut , pt ) (3.3.17)

Theorem 3.3.3. Assume that all conditions in Theorem 3.3.2 are satisfied and moreover, that
H(t, x, u, p) is concave in (x, u) for every t and that

limt pt (xt xt ) 0 (3.3.18)

Then ({xt }, {ut }) is optimal. 

Example 3.3.3.


max t xt ut
u
t=0

subject to xt+1 = (1 ut )xt , x0 given, limt xt = x where 0 < x < x0 , (0, 1) and
u (0, 1) . 

213
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Solution. Let H(t, xt , ut , pt ) = t xt ut + pt (1 ut )xt . Since u (0, 1) is an interior point,

(3.3.16) simplifies to Hu (t, xt , ut , pt ) = 0 , thus


1 t xt
= pt xt (3.3.19)
2 ut

Further, from (3.3.17) it follows that


1 t ut
= pt1 pt (1 ut ) (3.3.20)
2 xt

and through division

xt pt xt

=
ut pt1 pt (1 ut )

which again implies that pt1 pt = 0 . Hence, pt = K and clearly K > 0 (cf. (3.3.19)).

Further from (3.3.19)

1
ut = 2
2t

4K xt

Need help with your

dissertation?
Get in-depth feedback & advice from experts in your
topic area. Find out what you can do to improve
the quality of your dissertation!

Get Help Now

Go to www.helpmyassignment.co.uk for more info

214 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Thus
 
1
xt+1 = 1 2
2t
xt

4K xt

which gives

1
xt+1 xt = 2t
4K 2

The general solution becomes

1
xt = C 2t

4K 2 ( 2 1)

and moreover (since x0 is given)

1
xt = x0 1 2t
 

4K 2 (1 2)

Finally, from the terminal condition limT xT = x it follows that

1
x0 =x
4K 2 (1 2 )
so

2 1
K =
4(x0 x)(1 2 )

Consequently,

(x0 x)(1 2 ) 2t
xt = x + (x0 x) 2t ut =

x + (x0 x) 2t

Note that if we substitute these solutions back into the original series we obtain


    x0 x
(x0 x)(1 2 ) t = x0 x 1 2
t
2t =
1 2
t=0 t=0

(This example should be compared with Example 3.3.1.) 

215
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Example 3.3.4.


max t (e2xt )
u
t=0

subject to xt+1 = xt + 2ut , x0 given, limT xT free, u [0, 1] , (0, 1) . 

Solution. The Hamiltonian becomes H = t e2xt + pt (xt + 2ut ) and evidently H is concave
in (x, u) . Moreover, (3.3.16), (3.3.17) may be expressed as

2pt (ut ut ) 0 (3.3.21)

and

pt1 = 2 t e2xt + pt (3.3.22)

Consequently (from (3.3.21)) we conclude that ut = 1 whenever pt 0 and ut = 0 if pt < 0 .

First, suppose ut = 1 . Then xt+1 = xt + 2 . Thus xt = C + 2t and the corresponding pt may

be obtained from (3.3.22) as

2e2(C+2)
pt = K + (e4 )t (3.3.23)
1 e4

and we observe that pt is a decreasing sequence of points.

Next, assume ut = 0 . Then xt+1 = xt xt = M and (3.3.22) implies that

2e2M t
pt = W + (3.3.24)
1

and again we recognize that pt is a decreasing sequence.

We are now left with three possibilities: (A) ut = 1 for every t , (B) ut = 0 for every t , or (C)
there exists t = t such that ut takes the value 1 (or 0) if t < t and the value 0 (or 1) if t t .

Suppose (A). Then ut = 1 , xt = x0 + 2t and (3.3.23) may be expressed as

2e2(x0 +2)
pt = K + 4
(e4 )t
1 e

216
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Further, since limT xT is free, (3.3.18) implies that limT pT = 0 . Thus K = 0 so clearly
pt > 0 . Hence, possibility (A) satisfies both Theorem 3.3.2 and Theorem 3.3.3.

Next, consider (B). Then ut = 0 , xt = x0 and (3.3.24) becomes

2e2x0 t
pt = W +
1

and just as in the treatment of (A), (3.3.18) implies that W = 0 , hence pt > 0 which contradicts

(3.3.21).

Finally, assume (C), i.e. that there exists a t = t such that for t = 0, 1, ..., t 1 we have ut = 0 ,
xt = x0 and for t = t , t + 1, ... we have ut = 1 , xt = C + 2t . The relation
xt = xt 1 + ut 1 now implies C + 2t = x0 + 0. Thus C = x0 2t . But then (from
(3.3.22))

pt 1
= 2 t e2xt + pt > 0

Brain power By 2020, wind could provide one-tenth of our planets

electricity needs. Already today, SKFs innovative know-
how is crucial to running a large proportion of the
worlds wind turbines.
Up to 25 % of the generating costs relate to mainte-
nance. These can be reduced dramatically thanks to our
systems for on-line condition monitoring and automatic
lubrication. We help make it more economical to create
cleaner, cheaper energy out of thin air.
By sharing our experience, expertise, and creativity,
industries can boost performance beyond expectations.
Therefore we need the best employees who can
meet this challenge!

The Power of Knowledge Engineering

Plug into The Power of Knowledge Engineering.

Visit us at www.skf.com/knowledge

217 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

(recall that pt > 0 ) which contradicts ut = 0 .

Consequently, xt = x0 + 2t , ut = 1 and

2e2(x0 +2)
pt = (e4 )t
1 e4

solves the problem. The maximum value becomes

  e2x0
t (e2(x0 +2t) ) = e2x0 (e4 )t =
1 e4
0 0

3.4 Discrete stochastic optimization problems

In sections 3.13.3 we discussed various aspects of discrete deterministic optimization problems. The
theme in this section is to include stochasticity in such problems, so, instead of assuming a relation of
the form xt+1 = f (t, xt , ut ) between the deterministic state variable x and the control u (cf. (3.1.1),

see also Theorem 3.1.1), we shall from now on suppose that

Xt+1 = f (t, Xt , ut , Vt+1 ) X0 = x0 , V0 = v0 (3.4.1)

where x0 and v0 are given, and ut U .

The use of capital letters X and V indicates that they are stochastic variables. Indeed, X will in general
depend on the values of V . V is a random variable which may be interpreted as environmental noise

or some other kind of disturbances. Regarding V , we may in some cases know the distribution of V

explicitly, for example that Vt+1 , is identically normal distributed with expected value E(Vt+1 ) = .
Alternatively, we may know the probability P (Vt+1 = v) , for example P (Vt+1 = 1) = p , and a third

possibility is that we have a knowledge of the conditional probability P (Vt+1 | Vt ) . (Later, when we

turn to examples, all cases mentioned above will be considered.) A final comment is that the control u

may depend on both X and V , thus ut = ut (Xt , Vt ) and from now on we shall refer to ut as a Markov

control. We further assume that we actually can observe the value of Xt before we choose ut . (If we

have to choose ut before observing the value of Xt , that may lead to a different value of the optimal

Markov control.)

218
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Now, referring to section 3.1, in the deterministic case we studied optimization problems of the form

T

max f0 (t, xt , ut )
u
t=0

subject to xt+1 = f (t, xt , ut ) where ut U , x0 given. In the stochastic approach which we consider
here, it does not make sense to maximize f0 at each time t , so instead we have to maximize the
expected value of f0 at each time. Consequently, we study the problem
T
 

maximizeE f0 (t, Xt , ut , Vt ) (3.4.2)
u0 ,u1 ,...,uT
t=0

subject to Xt+1 = f (t, Xt , ut , Vt+1 ) where X0 = x0 , V0 = v0 and ut U .

Define
T
 

Js (t, xt , vt ) = max E f0 (s, Xs , us (Xs , Vs ) | xt , vt (3.4.3)
u
t=s

Then, somewhat roughly, we have by the same argument that eventually lead to Theorem 3.1.1 the
following:

Theorem 3.4.1. Let Js (t, xt , vt ) defined through (3.4.3) be the value function for problem (3.4.2).

Then

J(t 1, xt1 , vt1 ) = max {f0 (t 1, xt1 , ut1 ) + E [J(t, Xt , Vt )]}

ut1

= max {f0 (t 1, xt1 , ut1 )
ut1

+E [J(t, f (t 1, xt1 , ut1 , Vt ), Vt )]}(3.4.4a)

and

J(T, xT , vT ) = J(T, xT ) = max f0 (T, xT , uT ) (3.4.4b)

uT

219
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Remark 3.4.1. Note that Theorem 3.4.1 works backwards in the same way as Theorem 3.1.1. First,
we find the optimal Markov control uT (xT , vT ) and the associated value function J(T, xT )

from (3.4.4b). Then, through (3.4.4a) the Markov controls and corresponding optimal value
functions at times T 1, T 2, ... are found recursively. 

Example 3.4.1. Solve the problem

T
 

max E (ut + Xt )
u
t=0

subject to Xt+1 = Xt 2ut + Vt+1 , where ut [0, 1] , x0 given, and Vt+1 0 is Rayleigh

distributed with probability density h(v) = (v/2 ) exp[v 2 /22 ] and > 0 .

Solution: From (3.4.4b): J(T, xT ) = maxu (x + u) so obviously we choose u = 1 . Hence:

J(T, xT ) = xT + 1 and uT = 1 .

220 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Now, using the fact that E(Vt+1 ) = /2 = K it follows from (3.4.4a):



J(T 1, x) = max{x + u + E(Xt + 1)}

u

= max{x + u + x 2u + K + 1} = max{2x u + K + 1}
u u

so clearly, the optimal Markov control is 0 which implies

J(T 1, xT 1 ) = 2xT 1 + 1 + K and uT 1 = 0

Proceeding in the same way, (3.4.4a) gives

J(T 2, x) = max{u + x + E(2XT + K + 1)}

u

= max{u + x + 2(x 2u + K) + K + 1}
u

= max{3x 3u + 3K + 1}
u

Again, the optimal Markov control is u = 0 , so consequently:

J(T 2, xT 2 ) = 3xT 2 + 3K + 1 uT 2 = 0

From the findings above it is natural to suspect that in general:

J(T k, x) = (k + 1)x + k K + 1 0 = 0

The formulae is certainly correct in case of k = 0 and by induction

J(T k, x) = (k + 1)x + k K + 1 0 = 0

Clearly, the optimal Markov control is u = 0 so

J(T (k + 1), x) = (k + 2)x + (k + k + 1)K + 1 = (k + 2)x + k+1 K + 1

which proves what we want. k obeys the difference equation k+1 k = k + 1 . The

homogeneous solution is C , and by assuming a particular solution of the form (Ak + B)k

together with the fact that 0 = 0 it follows that k = (k/2)(k + 1) . Thus

J(T, xT ) = xT + 1 uT = 1

1
J(T k, xT k ) = (k + 1)xT k + (k 2 + k)K + 1 uT k = 0, k 1
2

221
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

or alternatively:

J(T, x) =x+1 uT = 1

1
J(t, x) = (T t + 1)x + (T t)(T t + 1)K + 1 ut = 0, t < T
2

Example 3.4.2. Solve the problem:

T 1 

max E u2t XT2
u
t=0

subject to Xt+1 = (Xt + ut )Vt+1 . Vt+1 {0, 1} , P (Vt+1 = 1) = 12 , P (Vt+1 = 0) = 12 ,

xt > 0 , x0 given and u R . (Note that an alternative way of expressing the probabilities above
is to say that Xt+1 = Xt + ut with probability 1/2 and Xt+1 = 0 with probability 1/2.)

Solution:

J(T, xT ) = max(x2T ) = x2T uT arbitrary

u

 
2 1 2 1
 2 2
 2
J(T 1, x) = max u + E(XT ) = max u (x + u) + O
u u 2 2
 
1
= max u2 (x + u)2

u 2

Denoting g1 (u) = u2 (1/2)(x + u)2 , the equation g1 (u) = 0 gives u = (1/3)x .

(Note that g1 is concave.) Thus

 2  2
1 1 2 1 1
J(T 1, x) = x x = x2 and uT 1 = x
3 2 3 3 3

In the same way:

  
2 1 2
J(T 2, x) = max u + E XT 1
u 3
    
2 1 2 1 2 1 2 1 2
= max u (x + u) + O = max u (x + u)

u 3 2 2 u 6

222
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Letting g2 (u) = u2 (1/6)(x + u)2 , we easily obtain the solution of g2 (u) = 0 as
u = (1/7)x , hence
 2  2
1 1 1 1 1
J(T 2, x) = x x x = x2 uT 2 = x
7 6 7 7 7

Now, assume that J(T k, x) = k x2 where 0 = 1 . Then:

J(T (k + 1), x) = max u2 + E k XT2 k

  
u
  
2 2 1 2 1
= max u k (x + u) + O
u 2 2
 
1
= max u2 k (x + u)2

u 2

223 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

The equation g  (u) = 0 (where g(u) is the concave function inside the { } bracket) has the

solution u = k (2 + k )1 x . Thus,

 2  2
k 1 k
J(T (k + 1), x) = x k x x
2 + k 2 2 + k
k
= x2 = k+1 x2

2 + k

which is in accordance with the assumption. Consequently,

J(T, xT ) = x2T uT arbitrary

k
J(T k, xT k ) = k x2T k uT k = x k1
2 + k

where

k
k+1 =
2 + k

In the previous examples we have considered the cases that Vt+1 is from a known distribution (Example

3.4.1) and P (Vt+1 = v) is known (Example 3.4.2). In the next example we present the solution of a

problem found in Sydster et al. (2005), which incorporates conditional probabilities.

Example 3.4.3. Solve the problem

T 1 

max E u2t XT2
t=0

subject to Xt+1 = Xt Vt+1 + ut , x0 > 0 given, ut R , Vt+1 {0, 1} ,

P (Vt+1 = 1 | Vt = 1) = 3
4 , P (Vt+1 = 1 | Vt = 0) = 14 .

Solution: First, note that the conditional probabilities above also imply
P (Vt+1 = 0 | Vt = 1) = 1/4 and P (Vt+1 = 0 | Vt = 0) = 3/4 . Clearly:

J(T, xT ) = max(x2T ) = x2T uT arbitrary

u

224
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Regarding J(T 1, xT 1 , vT 1 ) there are two cases to consider, the case vT 1 = 1 and the

case vT 1 = 0 . The former yields:

J(T 1, x, 1) = max u2 + E(x2T )

 
u
 
2 3 2 1 2
= max u (x 1 + u) (x 0 + u)

u 4 4
 
5 3
= max u2 (x + u)2

u 4 4

Defining g1 (u) = (5/4)u2 (3/4)(x + u)2 , the solution of g1 (u) = 0 is u = (3/8)x

which after some algebra gives

15 2 3
J(T 1, xT 1 , 1) = x uT 1 = xT 1
32 T 1 8

In the same way, the latter yields

7 2 1
J(T 1, xT 1 , 0) = x uT 1 = xT 1
32 T 1 8

Now, assume:

J(T k, x, 1) = k x2 J(T k, x, 0) = k x2 (3.4.5)

Then, by induction:
    
2 3 2 2 1
J(T (k + 1), x, 1) = max u k (x 1 + u) k (x 0 + u)
u 4 4
 
3 1
= max u2 k (x + u)2 k u2

u 4 4

Letting g(u) = u2 (3/4)k (x + u)2 (1/4)k u2 , the equation g  (u) = 0 implies

u = 3k (3k + k + 4)1 x . Substituting back into J(T (k + 1), x, 1) then gives after
some algebra

3 k
J(T (k + 1), x, 1) = (k + 4)x2 = k+1 x2
4 3k + k + 4

225
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

and

3k
uT (k+1) (1) = xT (k+1)
3k + k + 4

By applying the same technique as above:

 
1 3
J(T (k + 1), x, 0) = max u2 k (x 1 + u)2 k (x 0 + u)2
u 4 4
 
1 3
= max u2 2
k (x + u) k u 2

u 4 4

and we easily conclude that u = k (k + 3k + 4)1 x is the optimal Markov control.

Inserting back into J(T (k + 1), x, 0) gives

k (3k + 4) 2
J(T (k + 1), x, 0) = x = k+1 x2
4(k + 3k + 4)

and

k
uT (k+1) (0) = xT (k+1)
k + 3k + 4

Challenge the way we run

EXPERIENCE THE POWER OF

FULL ENGAGEMENT

RUN FASTER.
RUN LONGER.. READ MORE & PRE-ORDER TODAY
RUN EASIER WWW.GAITEYE.COM

1349906_A6_4+0.indd 1 22-08-2014 12:56:57

226 Click on the ad to read more
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Finally, since 0 = 0 = 1 , we may by iteration find k , k for any k < T through the equations

3k (k + 4) k (3k + 4)
k+1 = k+1 =
4(3k + k + 4) 4(k + 3k + 4)

so the solution is given by (3.4.5) and associated optimal Markov controls. 

Exercise 3.4.1. Solve the problem

T
 

max E Xt
u
t=0

subject to Xt+1 = ut Xt Vt+1 , where ut [0, 1] , x0 given, P (Vt+1 = 1) = 1/3 ,

P (Vt+1 = 0) = 2/3 . 

Exercise 3.4.2. Solve the problem

T
 

max E t (u2t Xt2 )
u
t=0

subject to Xt+1 = Xt + ut + Vt+1 , 0, 1 , ut R , x0 given, Vt+1 is normal distributed

where E(Vt+1 ) = = 0 and Var (Vt+1 ) = 2 > 0 .

Hint: referring to Remark 3.4.3, E(Vt+1

2
) = v.

Exercise 3.4.3. Solve the problem

T
 

max E (Xt ut )
u
t=0

subject to Xt+1 = Xt + ut + Vt+1 , ut [0, 1] , x0 given, Vt+1 0 is exponential distributed

and E(Vt+1 ) = 1/ , > 0 for all t . 

Exercise 3.4.4. Show that the solution of the problem

T
 

max E Xt
u
t=0

227
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

subject to Xt+1 = ut Xt Vt+1 , ut [0, 1] , Vt+1 {0, 1} , P (Vt+1 = 1) | Vt = 1) = 2/3 ,

P (Vt+1 = 1 | Vt = 0) = 1/3 may be written as
  T t    T t 
2 1 1 3
J(t, x, 1) = 2 +3 x J(t, x, 0) = + x
3 2 3 2

Next, let us briefly comment on the case T , i.e. the infinite horizon case. As explained in

section 3.3, the extension from T finite to T infinite is by no means straightforward (mainly due to

convergence problems). Therefore, adopting the same strategy as in section 3.3 we now restrict the
analysis to the autonomous problem

 

max E t f0 (Xt , ut (Xt , Vt ) (3.4.6)
u
t=0

subject to Xt+1 = f (Xt , ut(Xt , Vt ), Vt+1 ) , x0 given. (0, 1) , ut R and where all probabilities
P (Vt+1 = v) are time independent. Moreover, cf. (3.3.2a), we also impose the boundedness condition
K1 f0 (x, u) K2 .

Now, define (se Remark 3.3.2)

 

J(s, xs , vs ) = sup E t f0 (Xt , u(Xt , Vt ) (3.4.7)
t=s

Then, (roughly) by using the same kind of arguments that lead to Theorem 3.3.1 we may formulate the
stochastic version of the Bellman equation as:

Theorem 3.4.2. Consider problem (3.4.6) and let J(s, xs , vs ) be defined through (3.4.7). Then

J(x, v) = max {f0 (x, u) + E(J(X1 , V1 ))} (3.4.8)

u

where J(x, v) = J(t = 0, x, v) and X1 = f (X, u, V1) . 

Remark 3.4.2. Just as in section 3.3, note the fundamental difference between (3.4.4a,b) and (3.4.8).
The latter is a functional equation which may not be solved recursively. Therefore, often the best
we can do is to guess the appropriate form of J(x, v) in (3.4.8). 

228
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Remark 3.4.3. Before we turn to an example, let us briefly state a useful result. Suppose that V is
a continuous stochastic variable with expected value

E(V ) = vf (v)dv =

where f (v) is the probability density. Then:


Var(V ) = (v )2 f (v)dv

 
2 2
= v f (v)dv 2 vf (v)dv + f (v)dv


= v 2 f (v)dv 2 = E(V 2 ) 2

Thus,

E(V 2 ) = Var(V ) + 2 (3.4.9)

Example 3.4.4 (Stochastic extension of Example 3.3.2). Find J(x) for the problem

This e-book
is made with SETASIGN
SetaPDF

PDF components for PHP developers

www.setasign.com

229 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

 

max E t (u2t Xt2 )
u
t=0

subject to Xt+1 = Xt + ut + Vt+1 , where (0, 1) , x0 given, ut R and Vt+1 is normal

distributed with expected value E(Vt+1 ) = = 0 and variance Var (Vt+1 ) = 2 = v .

Solution: Referring to the deterministic case (Example 3.3.2), we supposed a solution on the form
J(x) = x2 . Regarding our problem here, we shall assume that J(x) is on the form
J(x) = ax2 + b since E(Xt+1
2
) will contain terms where neither X nor u will occur. Thus,
from (3.4.8):

ax2 + b = max u2 x2 + E a(X + u + V1 )2 + b

  
u

= max u2 x2 aE (X + u)2 2(X + u)V1 + V12 + b

   
u

Now, since E(V1 ) = 0 , it follows from (3.4.9) that E(V12 ) = v . Hence,

2
+ b = max u2 x2 a(x + u)2 av + b
 
ax
u

Clearly, u = a(1 + a)1 x maximizes the expression within the bracket, so

1 + 2a 2
ax2 + b = x av + v
1 + a

Equating terms of equal powers yields

a(1 + a) = (1 + 2a) (3.4.10a)

b = av + b (3.4.10b)

The solution of (3.7.10a) is easily found to be


(1 2) + 1 + 4 2
a=
2

which implies

av
b=
1

230
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Consequently
     
(1 2) + 1 + 4 2 v (1 2) + 1 + 4 2
J(x) = x2
2 1 2

with associated optimal Markov control u = a(1 + a)1 x . 

Exercise 3.4.5. Find J(x, v) for the problem:

 

max E t (e2Xt )
u
t=0

subject to Xt+1 = Xt 2ut + Vt+1 . (0, 1) , x0 given, ut [1, 1] , Vt+1 0 is

tV
identically distributed with E(e2 t+1
)<.

Now, referring to Example 3.4.4 as well as Exercise 3.4.5, it is still not clear if the optimal value functions
J(x) which we found really solve the given optimization problems. The problem is the boundedness
condition. Neither of the f0 (x, u) functions from the example nor the exercise satisfy
K1 f0 (x, u) K2 (cf. Theorem 3.4.2). Still, there exists a few ways to show that J(x) can solve a
given problem even if the boundedness condition fails (Bertsekas, 1976; HernndezLerma, 1996;
Sydster et al., 2005). One way to proceed is to argue along the following line:

Suppose that f0 (x, u) 0 and (0, 1) (which is the case both in Example 3.4.4 and Exercise 3.4.5).

Moreover, assume that we have succeeded in solving the corresponding finite horizon problem (i.e. T

finite), and that U is compact and f0 (x, u), f (x, u) are continuous functions of (x, u) . Denote the

optimal value function in the finite case by J(0, x, v, T ) . Then limT J(0, x, v, T ) (if it exists! ) is

the optimal function which solves the infinite horizon problem. We shall now demonstrate (partly as
an exercise) that J(x) found in Exercise 3.4.5 really solves the given optimization problem.

The optimal value function of the infinite horizon problem given in Exercise 3.4.5 is found to be

1
J(x) = ae2x = e2x
1 Ke4

where K = E(e2Vt+1 ) .

231
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Now, consider the corresponding finite horizon problem

 T 

max E t (e2Xt )
u
t=0

We leave it as an exercise to the reader to show that the solution of this problem is:

J(T, x) = T e2x

uT arbitrary and J(T k, x) = T k k e2x , uT k = 1 , where k+1 = 1 + Ke4 k and

T = 1 , or alternatively

J(t, x) = t t e2x

where t1 = 1 + Ke4 t . Clearly, J(0, x, T ) = 0 e2x (and 0 = 0 (T ) ). Our goal is to

find limT J(0, x, T ) which is the same as finding limT (0 (T )) which again is the same

as finding limt (t ) when T is fixed.

Note that (t1 ) < (t ) and when t , = 1 + Ke4 , thus

www.sylvania.com

We do not reinvent
the wheel we reinvent
light.
Fascinating lighting offers an infinite spectrum of
possibilities: Innovative technologies and new
markets provide both opportunities and challenges.
An environment in which your expertise is in high
demand. Enjoy the supportive working atmosphere
within our global group and benefit from international
career paths. Implement sustainable ideas in close
cooperation with other specialists and contribute to
influencing our future. Come and join us in reinventing
light every day.

Light is OSRAM

232 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

1
=
1 Ke4

which is nothing but the quantity a in J(x) obtained in the infinite horizon problem. Consequently,

the optimal value function found in Exercise 3.4.5 really solves the optimization problem.

233
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization Discrete Time Optimization Problems

Appendix
(Parameter Estimation)

360
thinking .

360
thinking . 360
thinking .
Discover the truth at www.deloitte.ca/careers Dis

Deloitte & Touche LLP and affiliated entities.

Discover the truth at www.deloitte.ca/careers Deloitte & Touche LLP and affiliated entities.

Deloitte & Touche LLP and affiliated entities.

Discover the truth

234 at www.deloitte.ca/careers
Click on the ad to read more
Download free eBooks at bookboon.com

Deloitte & Touche LLP and affiliated entities.

Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

Referring to both the linear and nonlinear population models presented in Part I and Part II, most of them
share the common feature that they contain one or several parameters. Hence, if we want to apply such
models on a concrete species (for example a fish stock) we have to use available data in order to estimate
these parameters. In this appendix we shall briefly discuss how such estimations may be carried out.

Suppose that we know the size of a population x at times t = 0, 1, 2, . . . , n , i.e. that x0 , x1 , . . . , xn

is known, how do we for example estimate the growth rate r if the population obeys the difference

equation

xt+1 = xt er(1xt ) (A.1)

(the Ricker model)? The usual way to perform such an estimation is first to convert the deterministic
model like (A.1) into a stochastic model. Now, following Dennis et al. (1995), ecologists draw a major
distinction between different classes of factors which may influence the values of vital parameters
and thereby impose stochastic variations in ecological models. Demographic factors such as intrinsic
chance of variation of birth and death processes among population inhabitants are factors that occur
at an individual level. Environmental factors, chance variations from extrinsic factors occur mainly at
population (or age or stage class) level.

Moreover, it appears as a general ecological principle that stochastic fluctuations due to the latter type
of factors seem to affect population persistence in a much more serious way than those of demographic
type (Dennis et al., 1991).

Now, as is true for the analysis of almost all population models in Part I and Part II, we typically were
interested in the population as a whole, not at individual levels. Thus, for our purposes we want to build
stochasticity into models like (A.1) of the environmental type. Therefore, we consider the stochastic
version of (A.1)

xt+1 = xt er(1xt ) eEt (A.2)

where Et is a normal distributed stochastic variable with expected value = 0 and variance 2 .

(Recall that if Z is normal distributed with expected value and variance 2 the probability density
is given by
 2 
1 z

1 1
f (z) = exp (A.3)
2 2

235
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

and if Z1 , . . . , Zn are all normal distributed stochastic variables with expected values and variances
1 , . . . , n and 12 , . . . , n2 respectively we may express the joint probability density function as

(z )T 1 (z )
 
1 1
f (z1 , . . . , zn ) = n  exp (A.4)
2 || 2

where (z ) = (z1 1 , . . . , zn n ) and the variance covariance matrix is given by

12 Cov(Z1 , Z2 ) Cov(Z1 , Zn )

Cov(Z1 , Z2 ) 22 Cov(Z2 , Zn ) (A.5)

=




Cov(Z1 , Zn ) Cov(Zn , Z2) n2
Now, before we turn to (A.1), (A.2) let us first study the estimation problem in a more general
context.

Consider

x1,t+1 = f1 (x1,t , ..., xn,t , 1 , ..., q )eE1,t

x2,t+1 = f2 (x1,t , ..., xn,t , 1 , ..., q )eE2,t

(A.6)

xn,t+1 = fn (x1,t , ..., xn,t , 1 , ..., q )eEn,t

Hence, there are n state variables x = (x1 , ..., xn )T , q parameters = (1 , ..., q ) and
Et = (E1,t , ..., En,t )T is a stochastic environmental noise vector which is multivariate normal
distributed with expected value 0 and variance, covariance matrix . (If there is one variable only, all


covariance terms vanish so we are left with = 0 and variance v = 2 .)

Now, defining

Mt+1 = (ln x1,t+1 , ..., ln xn,t+1 )T Mt = (ln x1,t , ..., ln xn,t )T

we may reformulate (A.6) on a logarithmic scale as

Mt+1 = h(Mt ) + Et (A.7)

236
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

where h(Mt ) = (ln f1 (x1,t , ..., xn,t , 1 , ..., q ), ..., ln fn (x1,t , ..., xn,t , 1 , ..., q ))T and we may

observe that the environmental noise is added to the original model on a logarithmic scale.

Next, assuming that yt , t = 0, ..., k is k + 1 consecutive time observations of xt , it follows that the
conditional expected value E(Mt+1 | xt = yt ) may be expressed as

E(ln xt+1 | xt = yt ) = ln f(yt , ) = h(mt ) = ht (A.8)

Hence, referring to Tong (1995), (A.8) expresses that the nonlinear deterministic skeleton xt+1 = f(xt , )
is preserved on a logarithmic scale.

The likelihood function for our problem now becomes

k
p(mt | mt1 ) (A.9)

I(, ) =
t=1

(where m (as in (A.8)) contains the observation values y ) and we may interpret I as a measure of the
likelihood of the observations at each time as functions of the unknown parameters.

We will turn your CV into

an opportunity of a lifetime

Do you like cars? Would you like to be a part of a successful brand? Send us your CV on
We will appreciate and reward both your enthusiasm and talent. www.employerforlife.com
Send us your CV. You will be surprised where it can take you.

237 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

Now, following Dennis et al. (1995), the probability p(mt | mt1 ) is the joint probability density for
Mt conditional of Mt1 = mt1 . It is a multivariate normal probability density which according to
(A.8) possesses the expected value E(Mt ) = h(mt1 ) and variance, covariance matrix given by .

Therefore, by use of (A.8) and (A.4) we may express the joint probability distribution as

(mt ht1 )T 1 (mt ht1 )

 
1
p(mt | mt1 ) = exp (A.10
2

(2)n/2 ||

The maximum likelihood parameters are now obtained by computing zeros of derivatives of (A.9) with
respect to 1 , ..., q and . Moreover, calculations are simplified if we first apply the logarithm. Thus,

instead of computing the derivatives directly from (A.9) we compute the derivatives of
k
ln p(mt | mt1 ) (A.11)

ln I(, ) =
t=1
k
nk k 1
= ln 2 ln || (mt | ht1 )T 1 (mt | ht1 )
2 2 2 t=1

Estimates obtained from (A.9), (A.11) are often referred to as maximum likelihood estimates. Evidently,
the log-likelihood function (A.11) is complicated in case of several state variables x1 , ..., xn . Therefore,

most estimations must be done by use of numerical algorithms. One such frequently used algorithm
which has several desired statistical properties is the NelderMead simplex algorithm which is described
in Press et al. (1992). Here, we shall concentrate on cases where it is possible to estimate parameters
without using numerical methods.

To this end, consider the stochastic difference equation with one state variable

xt+1 = f (xt , )eEt (A.12)

which we may interpret as the stochastic version of almost all nonlinear maps considered in Part I. Now,
since n = 1 , the variance, covariance matrix degenerates to only one term, namely the variance v .

(We prefer v instead of 2 for notation convenience.) If we in addition have k + 1 observation points
yt of xt at times 0, 1, ..., k , the log- likelihood function (A.11) may be cast in the form
k
k k 1  2
ln I(1 , ..., q , v) = ln 2 ln v u (1 , ..., q ) (A.13)
2 2 2v t=1 t

where the log-residuals ut = ln yt ln f (yt1 , 1 , ..., q ) .

238
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

The maximum likelihood parameter estimators are then obtained from

k
ln I 1 ut
= ut (1 , ..., q ) (1 , ..., q ) = 0 (A.14a)
i v t=1 i
k
ln I k 1  2
= + u (1 , ..., q ) = 0 (A.14b)
v 2v 2v 2 t=1 t

or equivalently (by use of the definition of ut ) from

k f
 i
(yt1 , 1 , ..., q )
ut (1 , ..., q ) = 0 (A.15a)
t=1
f (yt1 , 1 , ..., q )

where i = 1, 2, ..., q and

k
1 2
v= u (1 , ..., q ) (A.15b)
k t=1 t

Example A.1. Suppose that we have observations yt of xt at times t = 0, 1, ..., k (i.e. k + 1

observations yt ) and estimate r in the nonlinear equation (A.1).

Solution: Consider the stochastic version of (A.1)

xt+1 = f (xt , r)eEt = xt er(1xt ) eEt

(which is nothing but (A.2)). The log-residuals become

ut = ln yt ln(yt1 er(1yt1 ) ) = ln yt ln yt1 r(1 yt1 )

Thus, according to (A.15a)

k
 yt1 (1 yt1 )er(1yt1 )
{ln yt ln yt1 r(1 yt1 )} =0
t=1
yt1 er(1yt1 )
or

k

{ln yt ln yt1 r(1 yt1 )} (1 yt1 ) = 0
t=1

from which we obtain

k  
yt
(1 y t1 ) ln
r=
t=1

yt1 (A.16)

(1 yt1 )2

239
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

The variance v may be estimated from (A.15b) as

k k    2
1 2 1 yt
v= u (1 , ..., q ) = ln r(1 yt1 )
k t=1 t k t=1 yt1

after we have first estimated r from (A.16). 

Example A.2. Suppose that we have observations yt of xt at consecutive times t = 0, ..., k

and estimate the parameters F and r in the equation

xt+1 = f (xt , F, r) = F xt erxt (A.17)

Solution: The stochastic version of (A.17) becomes

xt+1 = F xt erxt eEt

so the log-residuals may be expressed as

 
ryt1 yt
ut = ln yt ln(F yt1 e ) = ln ln F + ryt1

yt1

AXA Global
Graduate Program
Find out more and apply

240 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

Hence equation(s) (A.15a) may be cast in the form

k  
yt1 eryt1
 
 yt
ln ln F + ryt1 ryt1
=0
y t1 F y t1 e
t=1

k   2
(F )yt1 eryt1
 
 yt
ln ln F + ryt1 =0
yt1 F yt1 eryt1
t=1

or equivalently

k ln F Ar = B (A.18a)

A ln F Cr = D (A.18b)

where

k k  
  yt
A= yt1 , B= ln
t=1 t=1
yt1
k k  

2
 yt
C= yt1 , D= yt1 ln
t=1 t=1
yt1

Consequently, from (A.18)

AD BC kD AB
ln F = r=
A2 kC A2 kC

Finally, (A.15b) implies

k    2
1 yt
v= ln ln F + ryt1
k i=1 yt1
1
G 2B ln F + 2rD k(ln F )2 2rA ln F + r2 C

=
k

where

k  

2 yt
G= ln
yt1
i=1

241
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

Remark A.1. Cushing (1998) considers a similar model as (A.17) where he generates data points
at 60 consecutive times. There, he obtains estimates of parameters b and c (corresponding to F

and r in (A.17)) which accurately recover the correct parameters used in the generation of data

with seven significant digits. For further details, see Cushing (1998). 

Exercise A.1. Suppose that we have observations yt of xt at k + 1 consecutive times t = 0, ..., k

and estimate in equation (1.2.1) (the quadratic map). 

Exercise A.2. Use (A.15) and estimate parameters a and b in the Hassel family

axt
xt+1 = a > 1, b>1
(1 + xt )b

by use of observation points yt of xt at times t = 0, ..., k . 

In the previous examples (and exercises) the estimations have been carried out by use of the log-likelihood
function (A.11). Another possibility is to apply conditional least squares and we close this appendix by
giving a brief overview of the method. (We still denote state variables by x , observations by y and

parameters by .)

Now, suppose that we have k + 1 consecutive time observations y0 , . . . , yk , the purpose of the method
is to minimize log-residuals (recall that environmental noise is additive on a logarithmic scale, cf. (A.7))
so if we are dealing with a map x f(x, ) we search for parameter estimates that minimize
k

D= (ln yt ln f(yt1 , ))2 (A.19)
i=1

Estimates found through (A.19) are often referred to as conditional least squares estimates because they
are found (on a logarithmic scale) through a minimization of conditional sums of squares. We shall now
by way of examples show how the method works.

Example A.3. Assuming k + 1 time observations y0 , ..., yk , estimate parameter r in map (A.1).

Solution: In this case (A.19) becomes

k
    yt  2
r(1yt1 ) 2
D= (ln yt ln(yt1 e )) = ln r(1 yt1 )
yt1
i=1

242
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

Hence,

D    yt  
=0 2 ln r(1 yt1 ) (yt1 1) = 0

r yt1

which yields
 
 yt
(1 yt1 ) ln yt1
r= 
(1 yt1 )2

in accordance with the result we obtained by use of (A.15). (Also note that
2 D/r 2 = (1 yt1 )2 > 0 , hence the r estimate really corresponds to a minimum.) 


Exercise A.3. Given k + 1 time observations y0 , ..., yk of xt , estimate F and r in equation

(A.17). (Compare with the results of Example A.2,) 

e Graduate Programme
I joined MITAS because for Engineers and Geoscientists
I wanted real responsibili www.discovermitas.com
Maersk.com/Mitas e G
I joined MITAS because for Engine
I wanted real responsibili Ma

Month 16
I was a construction Mo
supervisor ina const
I was
the North Sea super
advising and the No
Real work he
helping foremen advis
International
al opportunities
Internationa
ree wo
work
or placements ssolve problems
Real work he
helping fo
International
Internationaal opportunities
ree wo
work
or placements ssolve pr

243 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

Referring to the n-dimensional nonlinear population models considered in Part II, most of them have

from an estimation point of view a desired property, namely that the various parameters in the model
occur in one equation only. (See for example the three- dimensional model presented in Exercise 2.8.2.
Here the fecundity F2 is in the first equation, parameter P0 is in the second equation only and P1 shows

up in the third equation only.) In such cases the method of conditional least squares is particularly
convenient to use because we may apply the method on each equation in the model separately. As an
illustration, consider the following example:

Example A.4. Consider the nonlinear map or difference equation model

x1,t+1 = F x2,t (A.20)

x2,t+1 = P e(x1,t +x2,t ) x1,t

Note that (A.20) is a special case of (2.8.2), ( = 1) , which was extensively studied in Example

2.8.1. Since acts as a scaling factor only, (A.20) possesses the same dynamics as (2.8.2). In case

of small values of F the dynamics is a stable nontrivial equilibrium. Nonstationary dynamics

is introduced through a supercritical Hopf bifurcation and when F is increased beyond instability

threshold, the various dynamical outcomes are displayed in Figures 1620 (cf. Example 2.8.1).

Now, suppose a time series of k + 1 observation points (y1,0 , y2,0 ), ..., (y1,k , y2,k ) of (x1,t , x2,t ).
Our goal is to use these points in order to estimate F and P by applying conditional least squares.

To this end (cf. (A.19)), define

k
    y1,t  2
2
D1 = [ln y1,t ln(F y2,t1)] = ln ln F
y2,t1
t=1

k    2
 y2,t
D2 = ln ln P + y1,t1 + y2,t1
y1,t1
t=1

The equations D1 /F = 0 , D2 /P = 0 give respectively

 
 y1,t
ln k ln F = 0

y2,t1

   y2,t  
ln + y1,t1 + y2,t1 k ln P = 0

y1,t1

244
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

Consequently, we may estimate F and P through

k  
1 y1,t
ln F = ln (A.21a)
k t=1 y2,t1
k   
1 y2,t
ln P = y1,t1 + y2,t1 + ln (A.21b)
k t=1 y1,t1

In order to investigate how good the estimates really are we have performed the following
experiment. Let F = 27.0 and P = 0.6 . Then, from (A.20) we have generated a time series of

50 observation points (y1,t , y2,t ) . The points are located on a chaotic attractor as displayed in

Figure 20. Next, pretending that F and P are unknown we have used the observations in (A.21)
in order to estimate F and P. The result is, F = 27.00003065 and P = 0.6000000143 so the

estimation appears to be excellent. 

Still considering the map (A.20) let us for comparison reasons also find the maximum likelihood estimates
of F and P. Suppose that

12 c
 
2 =

c 22

Then, by use of (A.4), (A.5) we may express (A.11) as

k
ln I(F, P, 2 ) = k ln 2 ln | 2 2 c2 |
2 1 2
 y  2  y  2
1,t 2,t
1 k
ln y2,t1
ln F ln y1,t1
+ y t1 ln P
+
2(1 2 ) t=1
1 2
      
y1,t y2,t
ln y2,t1 ln F ln y1,t1 + yt1 ln P
2
1 2

where yt1 = y1,t1 + y2,t1 and = c/1 2 .

The equations (ln I)/F = 0 and (ln I)/P = 0 may be cast in the forms

k k A B
ln F ln P = (A.22a)
1 2 1 2
k k B A
ln F + ln P = (A.22b)
1 2 2 1

245
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

where
     y2,t  
 y1,t
A= ln B= ln + yt1

y2,t1 y1,t1

The solution of (A.22a,b) is easily found to be

1 1
ln F = A and ln P = B
k k

which is the same as we obtained by use of conditional least squares.

Exercise A.4. Given k + 1 time observations (y1,0 , y2,0 ), ..., (y1,k , y2,k ) of (x1,t , x2,t ) and
find the conditional least squares estimates of F , P and in the age structured Ricker model

x1,t+1 = F x1,t ext + F x2,t ext

x2,t+1 = P x1,t

where x = x1 + x2 . 

Exercise A.5. Given k + 1 consecutive time observations, find the conditional least squares
estimates of all parameters in the map

(x1 , x2 ) (F ex x2 , P ex x1 )

246
Download free eBooks at bookboon.com
 Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

References
Adler F. (1990) Coexistence of two types on a single resource in discrete time. J. Math. Biol. 28, 695713.

Barnsley M. (1988) Fractals Everywhere. Academic Press Inc.

Beddington J.R., Free C.A. and Lawton J.H. (1975) Dynamic complexity in predatorprey models framed
in difference equations. Nature, 255, 5860.

Behncke H. (2000) Periodical cicadas. J. Math. Biol., 40, 413431.

Berg P., Pomeau Y. and Vidal C. (1984) Order within Chaos. John Wiley & Sons.

Bergh M.O. and Getz W.M. (1988) Stability of discrete age-structured and aggregated delay-difference
population models. J. Math. Biol., 26, 551581.

Bertsekas D.P. (1976) Dynamic Programming and Stochastic Control. Academic Press.

170x115_MIM_ datos_Espv4.pdf 1 17/12/15 18:29

Bernardelli H. (1941) Population waves. Journal of Burma Research Society, 31, 118.

EL 93%
DE LOS ALUMNOS DEL MIM ESTN
TRABAJANDO EN SU SECTOR
C
A LOS 3 MESES DE GRADUARSE
M

MASTER IN MANAGEMENT
CM

MY

CY ESTUDIA EN EL CORAZN DE MADRID Y APROVECHA LAS OPORTUNIDADES Y EXPERIENCIAS NICAS

Duracin: 1O MESES
QUE OFRECE LA CAPITAL DE ESPAA
CMY Exp. Profesional: 1 AO
POTENCIA TU FORMACIN CURSANDO UNA DOBLE TITULACIN QUE MEJOR SE ADAPTE A TUS
Idioma: ESPAOL / INGLS
K
OBJETIVOS PROFESIONALES
Formato: TIEMPO COMPLETO
ESTUDIA UN SEMESTRE EXTRA EN OTRO PAS CON EL PROGRAMA BEYOND BORDERS EXPERIENCE
Convocatorias: SEPT / FEB
Y AUMENTA TU RED DE CONTACTOS

5 ESPECIALIZACIONES 100 DEL MUNDO 55 NACIONALIDADES

MASTERS IN MANAGEMENT
PERSONALIZA TU PROGRAMA FINANCIAL TIMES
EN CLASE

www.ie.edu/master-management [email protected] Sguenos en IE MIM Experience

247 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization Appendix (Parameter Estimation)

Botsford L.W. (1986) Population dynamics of the Dungeness crab (Cancer magister). Can. Spec. Publ.
Fish. Aquat. Sci., 92, 140153.

Caswell H. (2001) Matrix Population Models. Sinauer Ass., Inc. Sunderland, Massachusetts.

Clark C.W. (1976) A delayed-recruitment model of population dynamics with an application to baleen
whale populations. J. Math. Biol., 3, 381391.

Collet P. and Eckmann J.P. (1980) Iterated maps on the interval as dynamical systems. Progress on Physics,
Vol. 1, Birkhuser Boston, Boston.

Costantino R.F., Desharnais R.A., Cushing J.M., and Dennis B. (1997) Chaotic dynamics in an insect
population. Science, 275, 389391.

Cushing J.M. (1998) An introduction to structured population dynamics. SIAM, Philadelphia.

Cushing J.M., Dennis B., Desharnais R.A. and Costantino R.F. (1996) An interdisciplinary approach to
understanding nonlinear ecological dynamics. Ecol. Model., 92, 111119.

Cushing J.M., Costantino R.F., Dennis B., Desharnais R. A. and Henson S. M. (1998) Nonlinear population
dynamics: models, experiments and data. J. Theor. Biol., 194, 19.

Cvitanovi P. (1996) Universality of Chaos. Institute of Physics Publishing.

Davydova N.V., Diekmann O. and van Gils S.A. (2003) Year class coexistence or competitive exclusion
for strict biennials. J. Math. Biol., 46, 95131.

Dennis B., Munholland P.L. and Scott J.M. (1991) Estimation of growth and extinction parameters for
endangered species. Ecological Monographs, 61, 115143.

Dennis B., Desharnais R.A., Cushing J. M. and Costantino R.F. (1995) Nonlinear Demographic Dynamics:
Mathematical Models, Statistical Methods, and Biological Experiments. Ecological Monographs, 65,
261281.

Dennis B., Desharnais R.A., Cushing J.M. and Costantino R.F. (1997) Transition in population dynamics:
equilibria to periodic cycles to aperiodic cycles. J. Anim. Ecol., 6b, 704729.

Devaney R.L. (1989) An Introduction to Chaotic Dynamical Systems. AddisonWesley Publishing

Company, Inc.

248
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization References

Diekmann O. and van Gils S. (2000) Difference Equations with Delay. Japan J. Indust. Appl. Math., 17,
7384.

EdelsteinKeshet L. (1988) Mathematical Models in Biology. Random House, New York.

Feigenbaum M. J. (1978) Quantitative universality for a class of nonlinear transformations. J. Stat. Phys.,
19, 2552.

Frauenthal J. (1986) Analysis of Age-Structure Models. Biomathematics, Vol. 17, Mathematical Ecology,
eds. Hallan T.G. and Levin S.A. Springer Verlag, Berlin, Heidelberg.

Govaerts W. and R. Khoshsiar Ghazani (2006) Numerical bifurcation analysis of a nonlinear stage
structured cannibalism model. J. difference eqn. and applications, 12, 10691085.

Grimshaw R. (1990) Nonlinear Ordinary Differential Equations. Blackwell Scientific Publications.

Guckenheimer J. (1977) On the bifurcation of maps of the interval. Invent. Math., 39, 165178.

Guckenheimer J. and Holmes P. (1983) Nonlinear Oscillations, Dynamical Systems, and Bifurcation of
Vector Fields. Springer Verlag.

Guckenheimer J., Oster G.F., and Ipaktchi A. (1977) The dynamics of density dependent population
models. J. Math. Biol., 4, 101147.

Hartman P. (1964) Ordinary Differential Equations. Wiley, New York.

Hassel M.P. (1978) The Dynamics of Arthropod PredatorPrey Systems. Princeton University Press.

Hnon M. (1976) A two dimensional mapping with a stange attractor. Comm. Math. Phys., 50, 6977.

HernndezLerma O. and Lasserre J.B. (1996) Discrete-Time Markov Control Processes. Springer Verlag.

Higgins K., Hastings A. and Botsford L.W. (1997) Density dependence and age structure: nonlinear
dynamics and population behaviour. Am. Nat., 149, 247269.

Horn R.A. and Johnson C. R. (1985) Matrix Analysis. Cambrigde University Press.

International Whaling Commission (1979) Report No. 29. International Whaling Commission,
Cambridge.

249
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization References

Iooss G. (1979) Bifurcation of maps and applications. North Holland Math. Studies, 36.

Iooss G. and Adelmeyer M. (1999) Topics in Bifurcation Theory. Adv. Series in Nonlinear Dynamics, 3,
World Sci. 2nd ed.

Jost J. (2005) Dynamical Systems. Springer Verlag, Berlin, Heidelberg, New York.

Katok A. and Hasselblatt B. (1995) Introduction to the Modern Theory of Dynamical Systems. Cambridge
University Press.

King A. and Schaffer W. (1999) The rainbow bridge: Hamiltonian limits and resonance in Hamiltonian
limits and resonance in predatorprey dynamics. J. Math. Biol., 39, 439469.

Kon R. (2005) Nonexistence of synchronous orbits and class coexistence in matrix population models.
SIAM J. Appl. Math., 66, 616626.

Kon R., Saito Y. and Takeuchi T. (2004) Permanence of single-species stage-structured models. J. Math.
Biol., 48, 515528.

Kot M. (2001) Elements of Mathematical Ecology. Cambridge University Press.

250 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization References

Kuznetsov Y.A. (2004) Elements of Applied Bifurcation Theory. 3rd ed., New York, Springer-Verlag.

Kuznetsov Y.A. and Meijer H.G.E. (2005) Numerical normal forms for codim 2 bifurcations of maps
with at most two critical eigenvalues. SISC, 26, 19321954.

Leslie P.H. (1945) On the use of matrices in certain population mathematics. Biometrika, 33, 183212.

Leslie P.H. (1948) Some further notes on the use of matrices in population mathematics. Biometrika,
35, 213245.

Levin S.A. and May R.M. (1976) A note on difference-delay equations. Theor. pop. biol., 9, 178187.

Levin S.A. and Goodyear P.H. (1980) Analysis of an age-structured fishery model. J. Math. Biol., 9,
245274.

Lewis E. G. (1942) On the generation and growth of a population. Sankhya: The Indian Journal of
Statistics, 6, 9396.

Li T.Y. and Yorke J.A. (1975) Period three implies chaos. Am. Math. Monthly, 82, 985992.

Marsden J.E. and McCracken M. (1976) The Hopf Bifurcation and its Applications. Springer Verlag, New
York, Heidelberg.

May R.M. (1976) Simple mathematical models with very complicated dynamics. Nature, 261, 459467.

Maynard Smith J. (1968) Mathematical Ideas in Biology. Cambridge University Press.

Maynard Smith J. (1979) Models in Ecology. Cambridge University Press.

Meyer C.D. (2000) Matrix Analysis and Applied Linear Algebra. SIAM, Philadelphia.

Mills N.J. and Getz W.M. (1996) Modelling the biological control of insect pests: a review of host
parasitoid models. Ecological Modelling, 92, 121143.

Mjlhus E., Wikan A., and Solberg, T. (2005) On synchronization in semelparous populations. J. Math.
Biol., 50, 121.

Moore G. (2008) From Hopf to Neimark Sacker bifurcation: a computational algorithm. Int. J. Comp.
Sci. Math., 2, 132180.

251
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization References

Murdoch W.W. (1994) Population regulation in theory and practice. Ecology, 75, 271287.

Murray J.D. (1989) Mathematical Biology. Springer, Berlin, Heidelberg.

Nagashima H. and Baba Y. (1999) Introduction to Chaos. Institute of Physics Publishing, Bristol and
Philadelphia.

Neimark Y.I. and Landa P.S. (1992) Stochastic and Chaotic Oscillations. Dordrecht Kluwer Acc. Publ.

Neubert M.G. and Kot M. (1992) The subcritical collapse of predator populations in discrete-time
predatorprey models. Math. Biosciences, 110, 4566.

Nicholson A.J. (1933) The balance of animal populations. Journal of Animal Ecology 2, 132178.

Nicolson A.J. and Bailey V.A. (1935) The balance of animal populations. Part I. Proceedings of Zoological
Society of London, 3, 551598.

Saber E., Gyri I. and Ladas G. (1998) Advances in Difference Equations. Proceedings of the Second
International Conference on Difference Equations. CRC Press.

Sacker R.J. (1964) On invariant surfaces and bifurcation of periodic solutions or ordinary differential
equations. IMMNYC, Courant Inst. Math. Sci., New York University.

Sacker R.J. (1965) A new approach to the perturbation theory of invariant surfaces. Comm. Pure and
Appl. Math., 18, 717732.

Seierstad A. and Sydster K. (1987) Optimal Control Theory with Economic Applications. North-Holland.

Silva J.A.L. and Hallam T. (1992) Compensation and Stability in Nonlinear Matrix Models. Math. Biosci.,
110, 67101.

Silva J.A. and Hallam T.G. (1993) Effects of delay, truncation and density dependence in reproduction
schedules on stability of nonlinear Leslie matrix models. J. Math. Biol., 31, 367395.

Singer D. (1978) Stable orbits and bifurcation of maps of the interval. SIAM, J. Appl. Math., 35, 260267.

Smale S. (1963) Diffeomorphisms with many periodic points. Differential and Combinatory Topology, S.
S. Cairns (ed.), 6380, Princeton University Press, Princeton.

Smale S. (1967) Differential Dynamical Systems. Bull. Amer. Math. Soc., 73, 747817.

252
Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization References

Spiegel M. (1974) Complex Variables. Schaums Outline Series, McGrawHill.

Stuart A.M. and Humphries A.R. (1998) Dynamical Systems and Numerical Analysis. Cambridge
University Press.

Sydster K. (2002) Matematisk analyse, Bind II, Universitetsforlaget (in Norwegian).

Sydster K., Hammond P., Seierstad A., and Strm A. (2005) Further Mathematics for Economic Analysis.
Prentice Hall.

Thieullen P.H., Tresser C. and Young L.S. (1994) Positive Lyapunov exponent for generic one-parameter
families of unimodel maps. Journal dAnalyse Mathematique, 64, 121172.

Thunberg H. (2001) Periodicity versus Chaos in One-Dimensional Dynamics, SIAM Rev., 43, 330.

Tong H. (1995) Non-linear Time Series. Oxford Science Publications.

Tsujii M. (1993) Positive Lyapunov exponents in families of one dimensional dynamical systems. Invent.
math., 111, 113137.

Excellent Economics and Business programmes at:

The perfect start

of a successful,
international career.

CLICK HERE
to discover why both socially
and academically the University
of Groningen is one of the best
places for a student to be
www.rug.nl/feb/education

253 Click on the ad to read more

Download free eBooks at bookboon.com
Discrete Dynamical Systems with an
Introduction to Discrete Optimization References

Tuljapurkar S., Boe C. and Wachter K.W. (1994) Nonlinear Feedback Dynamics in Fisheries. Analysis
of the DerisoSchnute Model. Can. J. Fish. Aquat. Sci., 51, 14621472.

Vanderbauwhede A. (1987) Invariant manifolds in infinite dimensions. Dynamics of Infinite Dimensional

Systems, ed. S.N. Chow and J.K. Hale, Speinger Verlag, Berlin.

Van Dooren T.J.M. and Metz J.A.J. (1998) Delayed maturation in temporally structured populations with
non-equilibrium dynamics. J. Evol. Biol., 11, 4162.

Wan Y.H. (1978) Computations of the stability condition for the Hopf bifurcation of diffeomorphisms
on R2 . SIAM, J. Appl. Math., 34, 167175.

Wikan A. (1994) Bifurcations, Nonlinear Dynamics, and Qualitative Behaviour in a Certain Class of
Discrete Age-Structured Population Models. University of Troms, Norway.

Wikan A. (1997) Dynamic consequences of reproductive delay in Leslie matrix models with nonlinear
survival probabilities. Math. Biosci., 146, 3762.

Wikan A. (1998) Four-periodicity in Leslie matrix models with density dependent survival probabilities.
Theor. Popul. Biol., 53, 8597.

Wikan A. (2001) From chaos to chaos. An analysis of a discrete age-structured preypredator model.
J. Math. Biol., 43, 471500.

Wikan A. (2012a) Age or stage structure? A comparison of dynamic outcomes from discrete age- and
stage-structured population models. Bull. Math. Biol., 74(6), 13541378.

Wikan A. (2012b) On nonlinear age- and stage-structured population models. J. Math. & Stat., 8(2),
311322.

Wikan A. and Mjlhus E. (1995) Periodicity of 4 in age-structured population models with density
dependence. J. Theor. Biol., 173, 109119.

Wikan A. and Mjlhus E. (1996) Overcompensatory recruitment and generation delay in discrete age-
structured population models. J. Math. Biol., 35, 195239.

Wikan A. and Eide A. (2004) An analysis of a nonlinear stage-structured cannibalism model with
application to the northeast Arctic cod stock. Bull. Math. Biol., 66, 16851704.

Zhang Q. and Tian R. (2008) Calculation of Coefficients of Simplest Normal Forms of NeimarkSacker
and Generalized NeimarkSacker Bifurcations. Journal of Physics: Conference Series, 96, 012152.

254
Download free eBooks at bookboon.com