Discrete dynamical systems

and the logistic map

an easy introduction

Daan van den Berg

Hogeschool van Amsterdam
Riken brain science institute
 Contents

Foreword

Preliminaries

1 Discrete dynamical systems

1.1 Orbit and iteration
1.2 Fixed point, periodic point, repellor, attrator
1.3 Exercises
1.4 Hyperbolicity, the meaning of the derivative.
1.4.1. Graphical analysis.
1.4.2 The mean value theorem
1.4.3 Hyperbolicity
1.5 Exercises

2. A particular kind of discrete dynamical system: the logistic map

2.1 Origin of the logistic map
2.2 Behaviour of to logistic map for various parameters of A
2.2.1 A first investigation: 0<A<-1
2.2.2 Investigation for 1<A<3; bifurcation
2.2.3 Investigation for 3<A<3.57; period-doubling-bifurcation, Feigenbaum's constant.
2.2.4 Beyond the accumulation point: Chaos
2.3 The bifurcation diagram: chaos and order.
2.4 Nearly period three, intermittency
2.5 Period three, Sarkovskii’s theorem
2.6 A=4: the how and why of chaos
2.6.1 The tent map is chaotic
2.6.2 Topological conjugacy
2.7 Back to biology

3. Randomness
3.1 Proces & product
3.2 Infinite strings
3.3 Finite strings
3.4 Chaos and randomness
3.5 Exercises

Solutions to exercises
Foreword

This text is intended for non-mathematicians who would like to explore and understand the
mathematical side of dynamical systems, the logistic map, and chaos. In particular, it serves to
support the comprehension of those participating or interested in Professor C. van Leeuwens
research in models such as the coupled logistic maps for a descriptional model of visual
information processing in the human visual cortex.
As reading the average mathematical literature on these subjects is harder than decyphering the
rosetta stone in pitch-dark, it appears that the gates of insight and understanding only open to
those belonging to the die-hard incrowd: mathematicians and those with a strong mathematical
background. The latter group does not contain the average psychologist. Nor does it the
philosopher or the biologist, the economist. Shame, for if science seeks to stride forward in
measures of describing what is happening around us it must at a certain moment find bridges for
the yawning chasm that lurks between mathematics and everything else. This text hopes to toss a
first line and make a modest bit of mathematical theory accesible to the vast world at the other
side of the scientific crevasse.
After each piece of mathematical theory there are some exercises for the reader to test his or her
newly aquired knowledge. Answers to the exercises are included. The author likes to stress that
this is a first draft; despite careful proofreading errors might be present. Feedback on errors, but
also on the difficulty or readibility of this manuscript are welcomed.

Daan van den Berg

[email protected]
[email protected]
Preliminaries

There are a few things we will have to recall from calculus before we get started. Most of them will
be explained in the text as well, but nonetheless a handhold will be nice. Most of these issues are
related to our understanding of ‘infinity’. Most non-mathematicians seem to have much difficulty
with that concept.

Convergent series

If we start of with one and add one infinitely many times, the result is infinity. This is denoted like
this:

∞
Σ 1 = 1+1+1+1+1+1+1+1+1 ....... = ∞
n=1

It says: start by making n one, add the argument (whatever is in front of the sigma), add one to n,
add the argument etcetera until n reaches the value over the sigma. So n functions as a counter
while the argument in front of sigma is constantly added to the sum. But what happens if the term
is getting smaller all the time? For instance, if we take 1/n instead of n. In the long run, 1/n will tend
to zero so less and less is added.
∞
Σ 1/n = 1/1 + 1/2 + 1/3 + 1/4 + 1/5 + 1/6+.... =?
n=1

This summation in fact goes to infinity. This is because every so many terms are bigger than one.
But it is kind of a boundary case. If we take a smaller term

∞
Σ 1/10n = 0.1 + 0.01 + 0.001 + 0.0001 + 0.00001 ..... = 0.11111 ...
n=1

it is easy to see that the result will never exceed 1. There is a critical boundary:

∞
Σ 1/an = 1/a + 1/a2 + 1/a3 + 1/a4 + 1/a5 + 1/a6 +....
n=1

is convergent for all absolute a greater than one. Convergent means: becomes a certain value and
does not go to infinity. This is a very important property for understanding the bifurcation diagram.
Fractions and square roots
Fractions are numbers like ¼ , ½ etcetera. Being either positive or negative, they have an upper
and a lower part, both natural numbers. Square roots are not fractions and this is a very important
thing to understand. Historically too, it is a much-stirred concept. Hepasos lost hist head because
of the proof that a square root is not a fraction. We will show the same thing here without losing
any body parts. The thing about square roots is that their binairy expansion have non-repeating
tails whereas fractions do. For instance, and ever-repeating of 0.545454.... is a fraction and it is
easily shown:

a = 0.54545454.... so 100a = 54.54545454....

100a - a = 99a =54 (because their endlessly repeating tails fall off)

a = 54/99 = 6/11

Open and closed sets

Consider the real line R. These are all square roots and fractions (remember that values like 2 and
-4 can also be written as a fraction). So basically, all numbers we usually use. Now we can take a
certain piece of this line, an interval. We denote such an interval by giving it’s boundaries. As such,
[2,4] is the interval of all values between 2 and 4, including 2 and 4 themselves. This is a closed
set. We could also denote (2,4), the same interval with the boundaries excluded. There is an
important difference between these two and that is that the latter does not have a smallest value.
The smallest value of [2,4] is obviously zero, but of (2,4) what is the smallest value? It is not 1.99
because 1.999 is smaller. But it is also not 1.9999999..... because that is equal to two! Surprising?
By previously mentioned trick:

a= 1.99999..... 100a = 199.99... 99 a = 198 a=2.

Remember that open sets do not have a smallest or largest value.

Density
To be dense for a certain set in another basically means: they are everywhere and infinitely many.
In [0,1], the fractions are dense. They are eveywhere since you can find a new fraction between
every two fractions you have. Between 0.1 and 0.2, there is 0.15. Between 0.15 and 0.2 there is
0.175 etcetera etcetera.
1. Discrete Dynamical systems

1.1 Orbit and iteration

In common everyday nature, there are many processes that are defined over a time span. A rock
falling down, the propagation of a bee colony, an airplane landing. Such processes can be
described by a discrete dynamical system. A discrete dynamical system is a formula which
describes a certain value (for instance: the height of an airplane) through time. After one second
the airplane is at 80 meters height, after four seconds at 41 meters height. Let's expand a little on
this example. If we know that an airplane loses 20% of it's height each second, we could describe
the height at a certain second by looking at the previous second. If at one moment the airplane is
at 100 meters, then the next second it is at 80 metres. Simply take off 20%. This principle, looking
at the previous second to determine the next is to be captured in a formula. If we call the height of
the plane "x" and time "t" we would, for the airplane example, get the following formula:

xt+1= 0.8*xt

This is a discrete dynamical system. It says: "the value of x at time t+1 is 0.8 times the value at
time t". Or, in other words: "if we have the height at a certain second, the height at the next second
will be only 80% of that". If we now know that the plane starts its descend from 100 metres, we
can quite accurately determine its landing process. The next second will be 80 metres, the second
after that will be 64. In such a way we can produce a time series of values, often called the orbit of
a certain value. Below is shown the orbit for value x0 = 100 (which means: the value of x after 0
seconds equals 100, or, after zero seconds the plane is still at 100 meters height) under iteration
of our above mentioned formula the following orbit is produced. Note that some values have been
rounded off for the sake of convenience.

Time X0 x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 ... X19 X20 x21

Height 100 80 64 51.2 41 32.8 26.2 21 16.8 13.4 10.7 ... 1.4 1.2 0.9

The orbit is generated by constantly imposing our formula over a value and such determining the
next. Such a process is called iteration. Thus, by iterating a function over a certain starting value,
we generate its orbit.
A discrete dynamical system differs from a continuous dynamical system in that we can only take
discrete time values. The height of the plane after one, two, four or seven seconds. But not after
2¼ seconds, or after 4¾ seconds. Continuous dynamical systems are given by differential
equations. Discrete dynamical systems, the ones we will use, are given by difference equations,
though we will hardly use that word.
Finally, a point worth noting is that though xt+1 = 0.8*xt is the most formal definition of a discrete
dynamical system, it is also often written like f(x)=0.8x stating that this is an iterative function. This
is due to practical reasons and mathematical laziness in the first place, but also because when
analyzing these systems it is sometimes useful to use merely the graph of the first iteration of the
initial values, which is exactly equal to the f(x) notation. Apart from the notation, there is no
functional difference.
1.2 fixed point, periodic point, repellor, attrator

From a scientific point of view, it is interesting to understand the behaviour of a dynamical system.
For instance, if we look at our previous example,

xt+1= 0.8*xt

we can imagine what happens to any arbitrary value. Any initial value we take gets reduced by
20% in the next time step, and than again reduced by 20%. If we continue to iterate, the initial
value will be more and more reduced. This means that an initial value like 100 or 430 ultimately
gets smaller and smaller and in the long run will tend to zero. This also applies to negative values,
but since we were talking about the height of a plane this fact must be neglected.
So the gross behaviour of the system is a pretty strong element in understanding it. We know from
this system that the orbit of any arbitrary value will tend to zero as time proceeds Except for zero
itself, of course, which remains in place. The zero value is therefore called an attractor: under
iteration, it 'attracts' all other values. We can discern local and global attractors, dividing them by
the criterion of wether they attract all or merely some points. This one is global: it attracts all other
values in the system. The opposite is called a repellor. If a system has a repellor, you can pick a
value arbitrarily close to the repellor and it will nonetheless tend away from it under iteration.
Repellors are often called unstable whereas attractors are sometimes called stable. We’ll see
examples shortly.
Because the dynamics of this system are too easy to be interesting, we now proceed to our next
example. Consider the following discrete dynamical system:

Xt+1 = -(Xt)3

It is not difficult to see that, like in our previous example, zero remains in place under iteration.
After all, if we would iterate zero, we see -03 = 0 and thus zero projects to zero and doesn't change
under iteration. Such points, points invariant under iteration are called fixed points. It is possible for
a system to have one, multiple or even no fixed points at all.
But this still fairly easy system shows more interesting facts. Lets look at what happens if the
starting value 1 is taken, thus x0 =1. If we raise this value to power three and then multiply by -1 as
our dynamical system prescribes, the value becomes -1. This is the next value in the orbit. But if
we iterate -1 in this formula, the result is again 1. After all, -(-1)3 = 1. Such it comes that -1 evolves
to 1, whereas 1 evolves to -1. This said, we can say that choosing either as an inital value, the
orbit will only contain 1 and -1 and thus is a finite set consisting of only two elements.
Points like these are called periodic points. This because their orbit shows an undeniable periodic
structure. The number of iterations in which it cycles through one phase is called the period (more
accurately: prime period). Our system xt+1 = -(xt)3 has two periodic points: 1 and -1. Their orbits
consist of values -1 and 1 only.
Now can we say anything about the behaviour of the system? Most certainly we can. If we take an
irrational number between -1 and 1, we see that under iteration, it will tend to zero. This is because
an irrational number smaller than one will only shrink. It does this by alternating signs because of
the -1 multiplication, but this is a mere detail: all values between 1 and -1 will tend to zero under
iteration. Except for zero itself of course, which is fixed. Thus we could call zero a local attractor. It
attracts all values between -1 and 1 and this set of numbers is called its stable set. A stable set is
sometimes referred to as the bassin of an attractor.
X0 X1 X2 X3 X4 X5 X6
1.2 -1.7 5.2 -137 2592274 <<10-19 >>1057
1 -1 1 -1 1 -1 1
0.8 -0.5 0.13 -0.002 1.4*10-8 Towards 0 Towards 0
3 -27 19683 >1012 <<10-38 >>10114 <<10-342
0 0 0 0 0 0 0
3 -27 19683 >1012 <<10-38 >>10114 <<10-342

The behaviour of various initial values under iteration of xt+1 = -(xt)3.

1.3 exercises

1. Consider the discrete dynamical system xt+1,= 4xt

a) Determine whether this system has fixed points.
b) Determine whether this system has periodic points.
c) What happens to the value -2 under iteration of this system? What happens to 4.1 ?
d) Determine of all fixed or periodic points wether they are attractors or repellors.

2. Consider the discrete dynamical system xt+1 = 2xt(1-xt)

a) Determine the first piece of the orbit of the intial elements 0, 0.5, 0.1 and 1.
b) Try to determine wether this system has fixed or periodic points and wether they
attract or repel.

3 Consider the discrete dynamical system xt+1 = ¼x,

a) Determine wether this system has fixed points.
b) Determine wether this system has periodic points.
c) What happens to the value -2 under iteration of this system? What happens to 4.1?
d) Determine of all fixed or periodic points wether they are attractors or repellors.

4. Consider the discrete dynamical system xt+1 = 3. 1xt - 3.1(xt)2

a) Determine wether this system has fixed points.
b) Determine wether this system has periodic points.
c) What happens to the value -2 under iteration of this system? What happens to 0.9?
d) Determine of all fixed or periodic points wether they are attractors or repellors.

5. A colony of kangaroos inhabits southern Australia. This population doubles each year.
a) Give the formula of the discrete dynamical system modelling the population size over years.
b) If the formula is right, and I know the number of kangaroos within 5% correct, will I, using the
correct formula be able to determine the kangaroo population size in five years within that same
error margin?
1.4 Hyperbolicity, the meaning of the derivative.

1.4.1. Graphical analysis.

Let's go back to our previous example, the discrete dynamical

system xt+1=-(xt)3. To analyse the behaviour of the dynamical
system, we are now going to introduce 'Graphical Analysis', a
method which is rather straightforward and visual. Through this
we hope to understand more about why certain fixed points are
attractors and others are repellors. If we would draw the graph of
this function, that is, the graph of f(x)=-(x)3, we could equivalently
say we had drawn the first-iteration graph of our dynamical
system mentioned above. After all, to each initial value set out on
the x-axis, the related point after one iteration is put on the y-
axis.
We can now easily find where the fixed points in the dynamical
system are. We draw a 45° angle line through the origin. Easy to
see, all the points of the form (a,a) lie on this line, such as (0,0),
(-2,-2), (¾,¾) and (7,7). If our graph intersects this line, it means
that at this intersection lies a fixed point. For instance our graph,
intersects line in (0,0). This is hardly surprising. The point 0
under iteration remains in place, it is its fixed point.
But this line has another great advantage: it can also show us
the behaviour of other points under iteration. How to work? We
take same initial value b, lets say b=0.9 and we wish to know
what happens to this point under iteration of our dynamical
system. The first iteration is relatively simple. Go vertically to the
graph and you will find the corresponding value (for b=0.9 this
value is -0.73, let's call it b’). Now to proceed, we go from this
point horizontally back to the diagonal line. As soon as we hit the
diagonal line we are in point (-0.73, -0.73), because only points
of the form (a,a) lie on the diagonal line. Now we can find the
first iteration of value -0.73 by vertically returning to the graph.
Note that the first iteration of -0.73 is the same as the second
iteration of 0.9. Lets say this value is 0.39. Now we can proceed
by again horizontally striking the diagonal line and vertically the
graph and if we just hold our pen to the paper in this way we
would get a spiraling picture of the orbit of our initial value 0.9
falling into the attractor of zero.
Equivalently, we might try the initial value of -1. From -1 a line
going straight up strikes the graph at 1. Going horizontally to the
diagonal strikes at (1,1). Returning vertically to the graph we
come to -1 and if we proceed to go the diagonal we come to (1,-
1) forming a perfect square. Proceeding with this process would
only lead to repetition of the route we've already been through.
This is understandable, since 1 is a periodic point. Thus, in
graphical analysis the periodic orbits show through closed cycles
like the square we got for our example.
The reader in encouraged to graphically analyze various initial T. t. b.: first, second and fifth step of
3
values of this dynamical system with pen and a piece of paper. the graphical analysis of f(x)=-x
1.4.2 The mean value theorem

The mean value theorem is an important theorem underlying many other concepts. For the need
of understanding hyperbolicity, or, the behaviour of attractors and repellors, it is explained here.
Imagine a function f(x). Just to recap: a function is a prescription that links all values from a certain
domain to a certain range. For instance, the function f(x)=x2 links all values of x to a certain other
value. It links 2 to 4 and -5 to 25. We might write f(-5)=25 or f(4)=16.
Now imagine a bumpy function f(x). It
need not necesserily be bumpy for this
theorem, but it makes things a bit
more visual. From this function, we cut f(b)
a certain piece. We take a piece from
a to b on the x-axis and all f(x)
corresponding function values. On this f(c)
piece of x-axis, the function has a
certain way of behaving. As we look
from left to right on our piece from a to
b, our function might be ascending,
descending, or possibly both. If we
follow the graph from a to b, the f(a)
function must arrive either higher,
lower or on the same value it left from
a. Lets say our function went up from
a to b. In that case, we can measure a c b
the amount it has gone up. If we divide
this amount by the interval [a,b], we get the average climbing rate. Its comparable to a car and its
travelled distance. If a car has driven 180 kilometers in 2 hours, the average speed was 90
kilometers an hour. It could have gone faster or slower on certain moments in the2 hours, but the
average speed was 90. But there's another important consequence. If the average speed was 90
kilometers an hour, there has been a moment on which the exact speed was 90 kilometers an
hour. Consider this: when we started to measure the speed of the car; it was either exactly 90
kilometers an hour or not. If it was, we are done: there is a moment at which the car travels exactly
90 kilometers an hour. If it was not travelling 90 kilometers an hour exactly, its speed was higher or
lower. Say it is was higher. Then it means the speed must go down somewhere in the coming two
hours. Otherwise the average cannot be 90 kilometers an hour. So at some moment the speed
must drop below 90 kilometers an hour, and since it drops from above 90 to below 90, 90 itself has
to be somewhere in between.
This is an informal way of proving the mean value theorem. It says that if the average derivative
over a certain trajectory is d, at some point in this trajectory the derivative must exactly d. more
formally put:

Mean value theorem

If a function f [a,b]→R is continuous on [a,b] and differentiable on (a,b) then there is a point
c∈(a,b) such that

f’(c) = f(b)-f(a)
b-a

In discrete dynamical system theory, we will use this theorem to investigate the hyperbolicity of
fixed points.
1.4.3 Hyperbolicity

In previous sections we have seen that fixed points are important factors in describing the
behaviour of a dynamical system. Some are attractors, drawing all values in the neighbourhood
towards it. Others are repellors. Initial values close to a repellor will tend away from it under
iteration. In this section we will examine a little analysis on the attractiveness of fixed points.
The behaviour of a system around a fixed point generally depends on wether the derivative of the
system in the fixed point is greater or smaller than one. Since also the negative values must be
included, we are better off by saying that the absolute value of the derivative determines wether a
fixed point is a repellor or an attractor. If the absolute derivative in that point is smaller than one,
the point is an attractor. If it is greater than one, it is a repellor. We'll start with the analytical side of
statement and subsequently add some graphical analysis to make it a little more visual, a little
more intuitive.
Let f(x) be a discrete dynamical system and let p be a fixed point of f(x), with |f(p)|<1. Thus, p is a
fixed point of the system in which the derivative is lower than 1. Then p will attract all nearby
values under iteration. Nearby means a neighbourhood around p in which the absolute derivative
of f(x) is smaller than 1 everywhere. Because we only examine continuous functions with
continuous derivatives, such a neigbourhood always exists. It might be very small, but it exists. We
might say this is the stable set or the bassin of the attractor. Now if we take any x out this bassin, it
will under iteration unstoppable approach the attractor. Why?
If we would measure the difference quotient between f(x) and f(p) it will be smaller than one. After
all, we decided to call the bassin a stretch on which the derivative is lower than one and therefore,
the average can never be higher than one. The mean value theorem then garantees us that at
some point between x and p, the derivative of the function is exactly equal to the difference
quotient mentioned above. We will call this value A. Thus,

|f(x) - f(p)| = A and now by multiplying both sides of the equation by Ix - p|, we get
|x-p|

|f(x) - f(p)| = A|x-p|

But look to what the formula has resulted now. It says, the difference between f(x) and f(p) equals
A times the difference between x and p. And since A is smaller than one we can conclude: the
difference between f(x) and f(p) is smaller than the difference between x and p. Or, iterating brings
x and p closer and since p is a fixed point and does not move, it is x moving towards p under
iteration.
The 'attractiveness' of the fixed point also shows by performing graphical analysis. As we have
seen earlier, graphical analysis is a way of visualizing the orbit of an initial value. If we would
perform this trick on our value x, the value from the bassin, the results become intuitively clearer.
From iterating x, striking the graph, we can only
go rightwards, towards the attractor. There is no
other way to draw, you can only go towards the
attractor.
Repellors and attractors share the common
lable hyperbolic. Any point or orbit is called
hyperbolic if its derivative is unequal to one. To
formalize: a hyperbolic fixed point is a fixed
point of which the derivative is unequal to one.
If the derivative of the point is exactly one the
point is labeled 'nonhyperbolic'.

x
P- ε P P+ε
1.5 Exercises

1. Consider the discrete dynamical system xt+1 = 2xt on R→R

a) Decide wether this system has fixed points.
b) Decide of all fixed points whether they are stable or unstable

2. Consider the discrete dynamical system xt+1 = Xt2 on R→R

a) Decide whether this system has fixed points.
b) Decide of all fixed points wether they are stable or unstable

3. Consider the discrete dynamical system xt+1,= 2+√xt on [0, ∞) →[0, ∞)

a) Decide whether this system has fixed points.
b) Decide of all fixed points wether they are stable or unstable
2. A particular kind of discrete dynamical system: the logistic map

2.1 Origin of the logistic map

Imagine a certain tropical island. This island is inhabited by two lizards. Of lizards it is known that
they repoduce exponentially. That is, each month the number of lizards is doubled in respect to the
previous month. This is a non realistic model. As time proceeds, the island will get overpopulated
and many lizards will die. Therefore, if the number of lizards is very high, the population will drop
rapidly. If the population is small however, it will grow through the abundant availability of food.
This principle of competing forces, proliferation on the one hand and food shortage on the other,
was modeled (originally by the Belgian mathematician Pierre Verhulst) to a discrete dynamcal
system called the Verhulst-function, or more commonly, the logistic map:

xt+1 = 2xt(1-xt), or equivalently: f(x) = 2x(1-x).

This model describes the population size (x) relative to the time (t). A few formal remarks are to be
made. First of all, the function is a normalized parabola (it intersects the x-axis at 0 and 1). This is
by convention and makes the mathematics a little easier.
As long as A remains smaller than 4, all points from [0,1] map back into [0,1] under iteration This
means that if we take an initial value from between 0 and 1 and iterate it, the result will also be a
value somewhere between 0 and 1. And as all values between 0 and 1 result in a value between 0
and 1, any initial value of between 0 and 1 will under iteration produce an orbit which in entirely
confined within the [0,1] region. All values outside this interval, that is, values smaller than 0 or
greater than 1 are in the stable set of minus infinity.
This does mean, however, that all values on the orbit of some initial value lie between 0 and 1.
Note therefore, that we are (if talking about population magnitudes) talking about thousands,
tenthousands or some other multiplicative factor of the original values. Thus if an orbit rends the
values 0.22, 0.84 and 0.11 we would be speaking of populations of 220, 840 and 110 lizards in the
previous example.
The number 2 in the above example represents the growth rate of the lizards. If there was no
constraining factor like food shortage, the population would double each year. We will in this
chapter examine the behaviour of the logistic map for various growth rates, and therefore call it A.
Parameter A can be any number between 0 and 4 for the behaviour we will study. Thus, the
general notation will be:

f(x) = Ax(1-x)

A brief survey will ensure us the logistic map has at least two fixed points. First of all 0 because
the parabola is normalized and furthermore the intersection with the function y=x. Recall from
1.4.1 that fixed points lie on the diagonal of 45°, which is equivalent to the function y=x. Thus,
solving the equation Ax(1-x) = x results in finding the fixed points 0 and 1-1/A. For most parameter
values, the second fixed point will lie in our area of investigation. Only for A<1 it does not.
The logistic map is a nice function to model population growth. But it has also found applications in
chemistry, physics, economy and even the science of psychology. This is due to the rich dynamics
of the map we will describe shortly.
2.2 Behaviour of to logistic map for various parameters of A

2.2.1 A first investigation: 0<A<-1

If A is smaller than one, this means disaster for the lizards. Since A represented their growth rate,
the population will drop and the species will become extinct. This is a nice intuitive guess of what
happens to orbits under iteration of this dynamical system, but since this text is intended
explanatory mathematical, we will have to formalize as well.
If A is smaller than one, the value of the second fixed point, 1-1/A is smaller than zero and not in
our area of investigation. Therefore we ignore it and focus on the other fixed point, 0. If we derive
the function f(x) = Ax(1-x) we would get f’(x) = A-2Ax. If we now investigate the fixed point for
hyperbolicity, thus determining the derivative in that point, we get f (0) = A - 2*A*0. This is equal to
f (0) = A and since A is smaller than one the fixed points derivative is smaller than one and it is an
attractor. This matches our intuition perfectly. If the reproduction rate is smaller than one, the
population will drop to zero.
Now left to be made is a remark about what happens to A=1. In other words, when the system is
non-hyperbolic. Is this case it is weakly attracting. Non hyperbolic points can also be weakly
repelling. We will not pay too much attention to non-hyperbolicity, it is uninteresting in respect to
the rest of the dynamics of the logistic map.

2.2.2 Investigation for 1<A<3; bifurcation

As mentioned earlier, for A<1 the logistic map has no fixed points apart from 0 in the interval [0,1].
This interval, all points between 0 and 1, is our area of investigation, because this is the place
where all the action happens. As soon as A becomes greater than one, two things happen. First of
all, the derivative of the fixed point 0 passes the critical 'greater than 1 boundary'. As we have
seen, a fixed point with a derivative smaller than one is an attractor, and with greater than one is a
repellor. So if A grows from smaller than one to greater than one, the stability of the 0 fixed point
changes radically. But something else happens. From the moment A is greater than one, our
equation 1-1/A results in a value between 0 and 1. In other words, our function gains a fixed point
within our area of investigation.
Thus, as A grows greater than 1 the dynamics of the system changes drastically. Such a sudden
change in the system's dynamics when a parameter is varied is called a bifurcation. The
parameter change for which the bifurcation occurs is called the bifurcation point. A bifurcation of
this type is called a transcritical bifurcation. Other types of bifurcations are for instance the saddle
node bifurcation and the period doubling bifurcation. The latter is also present in the logistic map.
We will investigate it shortly.
2.2.3 Investigation for 3<A<3.57; period-doubling-bifurcation, Feigenbaum's constant.

But the previous bifurcation is not the only one in the parameter range 0<A<4. There is another
bifurcation point at A=3. Thus as A changes from smaller than three to greater than three, there is
another dramatic change in the dynamics of the system.
As A grows, the height of the parabola changes. It gets higher as A gets higher. But another thing
happens as A grows: the derivative of the fixed point other than zero changes. The fixed point is
given by 1-1/A and the derivative is given by A-2Ax. As we are investigating the fixed point’s
derivative, we might rewrite A-2Ax to a form in which x is the fixed point. Thus, x=1-1/A.
Substitution rends A-2A(1-1/A) = A - (2A - 2A/A) = A-(2A -2) = -A+2 = 2-A. So the derivative of the
fixed point relates to the A-parameter as 2-A. If A=1½, the derivative of the fixed point is ½, thus it
is stable. If A=2.1, the fixed point is stable, for its derivative is -0.1, in absolute value smaller than
one. If A=3 its derivative is exactly one and we are at a bifurcation point because it goes from
being stable to being unstable. But there's another surprise.
Lets have a look at the second-iteration
function for a change. To visualize,
check the illustrations on this page. Top-
left is the firstiteration-graph of the
logistic map, top-right is the second-
iteration map. As the first-iteration graph
is given by Ax(1-x), the second is given
by the same formula iterated twice. This
is A(Ax(1-x))(1-(Ax(1-x)) which equals
A2x-A2x2-A3x2+2A3x3-A3x4. The exact
formula matters not. What is important is
to realize that a twice iterated logistic
map equals a once iterated fourth-
degree polynomial. This second-
iteration-map has a graph with two
humps instead of one. Now if the A-
parameter gets higher and higher, the
graph gets more eccentric. In case of
A=2.9 (top left and right) the fixed point
shows to be stable in the left picture.
The right picture, the second-iteration-
graph, also shows one stable fixed
point and if we look closely, we see it is Top-left: first iteration graph for A<3 Top-right: same for A>3
Bottom-left:2nd-iteration graph for A<3 Bottom-right: same for A>3
the same point as in the leftmost graph.
This doesn't surprise us, because what
is fixed after one iteration is also fixed after two iterations. But what happens as the A-parameter
gets bigger? As we have seen, the top of the first-iterationgraph gets higher, but also the second-
iteration-graph gets more eccentric. The maxima grow higher and the minimum deepens and as
soon as the A -parameter passes the value 3, the minimum suddenly crosses the 45° line. This
means it suddenly own two more periodic points of period 2 (after all, these are fixed points in the
second-iteration-graph). These periodic points are stable and thus we can draw a remarkable
conclusion:

As A passes the value 3 and the fixed point changes from being stable to being unstable, the
function gains a stable period 2 periodic point.
To transfer to our lizard-island concept: if the rate of lizard reproduction grows beyond 3 (for
instance, because of a fertility drug) the population doesn't grow towards a singular equilibrium.
Rather it keeps alternating each year. One year 760 lizards, the other year 240, then 760 again.
Remarkable but true, this periodicity seems to be a property of the system itself rather than
induced from an external source.
Because the function behaves extremer as the A-parameter increases there is a point for A where
in the first-iteration graph the fixed point becomes unstable and the second iteration graph gains
two fixed points which manifest themselves in the first iteration function as a stable orbit of period
two. This transition from stable to unstable is an issue in bifurcation theory. Notice how when A
was smaller than one, the zero fixed point was stable and as A increased became unstable. With it
came a new stable fixed point. The stability of this fixed point became questionable as well as A
approached 3. After A passed 3, this point became unstable as well and made way for a stable
period 2 orbit. We could see this in the second-iteration graph.
Now, wonders the curious spirit: "as A slowly increases from 3 to 4, and the graph of first- and
second-iteration behave more extreme, could the fixed points in the second iteration graph also
make a transition from stable to unstable?". And indeed this happens. At a value real close to
A=3.45, another period doubling bifurcation occurs. This because, as one might expect, the fourth
iteration graph is a polynomial of degree 16 with even more humps (maxima and minima) than the
seconditeration graph. Result is that at some point some of these minima cross the 45° line and a
stable period four occurs. Very similar to the period two. And as A increases, so does the period of
the stable orbit.
After the first period doubling bifurcation (for A=3) the stable attractor changes from period one to
period two. At the second bifurcation (for A = almost 3.45) it increases to period four and as A
increases towards 3.55, another period doubling bifurcation occurs and the stable attractor's
period increases to eight.
An interesting fact is that these period-douling bifurcation occur at a fixed ratio. As the parameter
distance between the first and second period doubling bifurcation is 3.45-3 = 0.45 and the distance
between the second and the third is 3.55-3.45 = 0. l. The second distance is therefore about 4.669
times shorter than the first distance. The distance between the third and the fourth is again 4.669
times shorter and for all subsequent period-doubling bifurcations the distance gets shorter by the
factor 4.669. This result was aquired by Mitchell Feigenbaum. This distance reduction factor has
ever since been known as Feigenbaum's constant:

δ = 4.669 20...

This constant is universal. This means that this factor goes for all unimodal functions. Technically
spoken, these are all continuous functions with one zero-deravtive, 'single hump-functions' one
might say. The fact that all these kinds of iterated functions are subject to Feigenbaum's constant
is known as the principle of universality.
But, as the curious mind continues, "if distance between the bifurcation points gets about five
times shorter at each bifurcation point, this sequence of bifurcations converges to a certain value;
this sequence is not infinitely long". Indeed it is not, and the series converges to a certain point
called the accumulation point. The accumulation point, written Aacc is about 3.57, thus Aacc ≈ 3.57.
"But what happens BEYOND this accumulation point?" The answer is... chaos, to which we will
dedicate the next paragraph.
2.2.4 Beyond the accumulation point: Chaos

It is a good thing for us to know what happens for parameter values in the range 0 to 3.57. All
initial values are either a point of the repelling orbit, a point of the attracting orbit or a point in its
bassin. But what happens beyond this accumulation point?
On the one hand, since the Feigenbaum constant is a convergent series, we know that the all
periodic points of period 2' have become unstable. Therefore we can expect all periodic orbits of 2'
when A is greater the the accumulation point. Thus there are infintely many periodic points. But on
the other hand we must consider the fact they are all repelling. Repelling towards what? The
answer is that for values of A greater than the accumulation point, the function Ax(1-x) behaves in
a chaotical manner. We'll take A=4 as an example.
What is a chaotical function anyway? There are several definitions of chaos, non of them
universally accepted. We will adopt the very common definition by Robert L Devaney.

Definition (Devaney).
A function f. V→V is chaotic if
• periodic points are dense in V
• f is topologically transitive
• f has sensitive dependence on initial conditions

Consider our function f(x) = 4x(1-x). This function maps [0,1] to [0,1] and matches all above
criteria. A set being dense means that between all two numbers, there is another number. For
instance, between 2 and 2.1 there is 2.05. But between 2 and 2.05 is 2.005. Between 2 and 2.005
again, there is again a number. It follows therefore that a dense set has infinitely many points. Now
if we look at the way periodic points are spread, we see that there are infinitely many and between
any two lies another third. We can- imagine this if we recall how the stable attractor was
transformed during the perioddoubling bifurcation cascade. After one bifurcation, when the stable
period two emerged, there was an unstable fixed point. After two bifurcations, the period of the
stable attractor is four, the unstable period 1 still exists but an unstable period two is added. Thus
two more periodic points. In the next bifurcation, more unstable periodic points are added. And this
all the way to infinity. In other words: in [0,1] under iteration of 4x(1-x) there are infinitely many
periodic points, and they are so densely packed that between every two, we can find another third.
Periodic points are said to be dense in [0,1]
Toplogical transitivity is different matter. If we pick an open interval of any size in [0,1], there is a
point in this interval that under iteration will go to any other interval. Thus, is we pick any small
neighbourhood in [0,1], it has a point in it that under iteration goes anywhere else in the domain.
We might say this function 4x(1 -x) well-mixes the domain [0,1]. Like a blender in a bowl of brown
and white sugar, we can intuitively feel why chaotic means well-mixed. If the brown and the white
sugar remain side by side we would say it is quite orderly. Whereas if it is well-mixed, we would
say it is a chaotic disorganization of sugar crystals. The well-mixedness is an intuitive obligation for
chaos.
But the most intuitive hallmark is the sensitive dependence on initial conditions. This means that if
we take two initial values very close to each oter, the will under iteration become more and more
separate. We can compare this to a river. Put two pieces of cork very close to each other in the
water. If the river is nice and calm, both of them will flow to the middle, towards the mainstream of
the river but if the river is wild and turbulent, the cork pieces soon move their own way, which is
not structural and fully unpredictable. The same goes for the chaotic logistic map. Take some
initial value and on forehand there is no predicting on where it is going to be.
2.3 The bifurcation diagram: chaos and order.

What more can we say for the logistic when A>Aacc, apart from the fact that for A=4, it is chaotic.
To investigate, we shall make use of the very important theorem of Fatou:

Fatou's theorem: If a quadratic function has a stable periodic orbit, then the critical point is in the
stable set of a point from this orbit.

So what this theorem says is: "If you iterate the critical point, it will fall into the attractor (given it
exists). The critical point is the point where the derivative is zero. In case of the logistic map,
therefore, the critical value is always 0.5. This is a fact we can make very good use of, for we are
now able to draw a so-called bifurcation diagram. This goes as follows. On the the x-axis, we put
the range of parameter A, form 0 to 4 that is. On the y-axis, we will place the value of the stable
point. Note therefore that for A<1 there is no value drawn. The stable point is 0 and therefore when
drawn falls exactly on the x-axis. Between 1 and 3, the function has one stable point. After 3, the
function has two stable fixed points: a stable periodic orbit, one of the points somewhere between
0 and 0.5, the other one somewhere between 0.5 and 1. After another bifurcation it has a stable
period four. Note that distance between bifurcationpoints gets shorter and shorter. The period
eight attractor is still wellvisible but one needs to look real close for the period sixteen and the
period 32 seems to disappear into the black mass behind it.

This black mess is where chaos enters the system. Just beyond the accumulation point, iterating
the critical point doesn't get it into a nice stable attractor anymore. It moves around like a
bumblebee, without showing any particular pattern so it seems. This is the region of the chaotic
orbits. Chaotic orbits are non-periodic (they never repeat) nor are they in a bassin of a periodic
attractor. But the most significant fact is that they are sensitive to initial conditions. This means that
an orbit nearby, even an orbit very nearby, will eventually move away from it and go it's own path
under iteration.
A striking fact are the white bars which run cross the chaotic region. If we look closely, we see a
periodic orbit. And this is true. All of a sudden, at varous parameters of A, stability occurs. A stable
periodic orbit resides amidst the chaos. If you look closely at the bifurcation diagram depicted on
this page you might find more of them. In fact, there is a stable periodic attractor arbitrarily close to
any chaotic attractor.
2.4 Nearly period three, intermittency

Chaos is stunning. It is surprising to see how a relatively simple formula like the logistic map
shows such complex behaviour in the chaotic region. Another strange thing is the periodic
windows inside the chaotic region; sudden areas of order which seem to emerge from nowhere.
Let's have a closer look at the period three window, because more interesting behaviour is
showing there.
If we want to find the values for which a period three exists, we have to solve f(t(f(x))) = x in which
f(x) stands for Ax(1-x), the logistic map. That is, we look for values of A for which the thirditeration-
graph of f(x) strikes the 45° diagonal. Two of these points are obvious; the fixed points 0 and 1-
1/A. These have prime period one, so they also have period three (what is fixed after one iteration
remains fixed after three as well).
But at A = 3.828... another period three orbit exists, and this one, remarkably enough, is. stable.
But an even more interesting phenomenon displays in the regio just before the period three
window. If we look closely at the graph depicted on- the. page we can. see that it nearly touches
the diagonal.- And the keyword is nearly. Graphical analysis should make things a bit clearer. If we
describe the orbit of a certain initial value with graphical analysis and we get near the point where
the stable period three is about to appear, the graphical analysis squeezes through the small
space left between the graph and the diagonal.
This is a very interesting phenomenon. Because the distance between the graph and the diagonal
is so small, the steps taken in our graphical analysis are also very small. This means that two
subsequent steps in the analysis are very close to each other in numerical value. And since we are
analyzing the third-iteration graph, we might conclude that all third values of the orbit are very
close together. We could call this quasi periodicity. Because every third value is alike, so are the
second and the first in some way, because these are projections of the third and since the logistic
map is not too bumpy, the projections of the points close together are also rather close together.
This quasiperiodicity is more commonly known as intermittency or sometimes called the laminar
phase. Intermittency is a period of quasi periodicity followed by a chaotic burst. The latter follows
when the orbit escapes from the narrow channel again. In physical systems, intermittency is a
well-known route to chaos.

(above) Graphical analysis of the third-iteration function of the logistic map in the intermittent phase (A=3.828). Note
the little box in the left picture is enlarged in the right picture.
2.5 Period three, Sarkovskii’s theorem

The intermittent regime just before the period three attractor shows surprising behaviour. At the
moment the period three attractor is established though, another strange phenomenon occurs.
Behold the Sarkovskii theorem:

Theorem (Sarkovskii):
Let f be a continuous function. If f has a periodic point of period three, it has periodic points of all
other periods as well.

It looks like we have quite a strong statement on our hands here. The proof is coming up, again in
an informal setting since this text is intended for non-mathematicians. It is supported by two pillars,
the first them being the intermediate value theorem (not to be confused with the mean value
theorem).
Imagine two intervals, I and J such that I⊂J. A function f is defined
such that J ⊂ f(I). Then f has at least one fixed point in I. First, I⊂J
means J is a subset of I. Second, f(I) are all the points in interval I
after one iteration. For the sake of the example, we’ll take the
interval of J to range from 0.6 to 0.8 and the interval I to range from
0.55 to 0.85. We’ll take the range of our function even a little bigger
than that, from 0.5 to 0.9, also for the sake of example. Since f is
continuous (i.e.: it has no sudden breaks, we can draw it without
taking our pencil off the paper) it has to cross the diagonal y = x at
least once, because its range is encompassing its domain (see
illustration). This guarantees us that if the condition I ⊂ J ⊂ f(I) is
true, the function has at least one fixed point in f.
The second pillar is the following: suppose we have two intervals called A0 and A1. We have
chosen these intervals such that A1 ⊂ f(A0). That is, if we would iterate all points in A0 we would get
the area of A1 and a little more, we say f(A0) covers A1. This ’little more’ is important. We could
take a somewhat smaller piece of A0, such that its points would under iteration exactly make up A1
and call this A0-piece J0. Now imagine we take another A called A2 such that A2 ⊂ f(A1). Since J0
was the exact precursor of A1 and f(A1) is a larger interval than A2, there must be a piece of J0
which exactly makes up A2. We will call this piece J1. We could do this again, Adding an A3 and a
J2 and continue in this fashion until An and Jn-1, for as long as we wish.
Now look at the following. We have three periodic points, because our function has period three.
We will call them a, b and c such that a is smaller than b and b is smaller than c. In mathematical
terms: a<b<c and f(a)=b, f(b)=c and f(c)=a. The interval between a and b will be given the name I0
and the interval between b en c will be called I1. Since the borders of I0 project to the borders of I1
under iteration of f, it follows that f(I0) covers I1. But for the similar reason, f(I1) covers I0∪I1. If
you’re not too sure of this, note the projection of b and c onto c and a (i.e. I0 and I1 and maybe
even a little more).
This last fact is utmost important for the theorem. Since f(I1) also covers I1, there must be part of I1
which exactly makes up I1. We’ll call this A0. As such, f(A0) =I1. But moreover, we can find an
interval A1 such that f(A1) covers A0, or better, such that f(A1) = A0. See the inductiveness: we can
find an interval A2 which exactly maps to A1. We can find an A3 for A2 and an A4 for that one. In
fact, we can go on an find an An as large as we wish. From this it follows that in An there lies a
period n periodic point. But how do we make sure this is its prime period? After all, a fixed point is
also a point of period 3,4 and 6 because it returns after each iteration. We would falsely be
claiming we found a point of period 3, 4 or 6 whereas it is actually a period 1 point returning,
returning and returning.
But here comes the trick. Let’s have a look at our period 2 point. Since f(I1) covers I0 and f(I0)
covers I1, it is possible to find a period two point which lies in I1 but who’s first iterate lies in I0.
Result: it must be period 2, and cannot be period one.
But this also works a point with period 5. Let’s choose the following intervals: A0, A1, A2, A3, A4.
Each interval covers the previous one under iteration of f. Thus A3⊂f(A4) and A2⊂f(A3) and so on.
Second A0 = I1. So this is a sequence of nested subintervals, all in I1. But I1 covers I0 and I0 again
covers I1. As a consequence, A0 = I0 ⊂ f5(A4). In words: there exists an interval such that it’s fifth
iteration covers I0. But the last iteration of this interval is actually We choose our A0 such, that it
exactly covers I0 instead of I1. If we now go looking for subsequent A1, A2 A3 etcetera and find their
periodic points, all their iterations will be in I1 except for the last. Result: it cannot have a period
smaller than the prime period we were looking for, otherwise it would totally be contained within I1.
Convince yourself that this also works for any other period. Just choose any period n and there will
be intervals A0...An-1 such that An-1 contains a point p which under iteration resides the first n-2
iterations result in a I1-point and its n-1th iteration results in a I0-point. Iteration number n has it
returning to its starting point, since it is periodic.

2.6 A=4: the how and why of chaos

In this section we will try to uncover the deeper mathematical meaning of the chaotic behaviour of
the logistic map when A=4. We will do this by proving the tent-map is chaotic and subsequentially
showing that its behaviour is in some essential way equivalent to the behaviour of the logistic map.
The tent map is much simpler and given by the formula f(x) = 1 -2|x-½|

2.6.1 The tent map is chaotic

1
The chaotic behaviour of the logistic map is a
profoundly studied phenomenon. There are several
definitions for it and we will stick to a very famous
one, the one as stated by Robert L. Devaney.

Defnition: a discrete dynamical system h(x) on

→D) is chaotic if:
(D→
• the system is topologically transitive on D.
• the system is sensitive to initial conditions.
• periodic points are dense in D, the domain of
the system.
0 1
Proposition: the tent map is chaotic
The tent map.
Well, that gives us something to chew on. The definition of a chaotic system. How can we actually
prove that the tent map is chaotic? That sensitive dependence was one of the prerequisites was
clear to us, but what about the other two? Imagine the domain of the function as an interval. For
the logistic map, this is the interval [0,1]. If we take any point from this line and it turns out not to be
a periodic point, and therefore a point of a chaotic orbit it should travel trough the entire domain
under iteration of the function. After all, if its travel permit was restricted to a certain piece of the
domain it wouldn’t correspond to our intuition of chaos. “Why should it not come in this particular
area?” Formally spoken the definition is as follows: if a system is topologically transitive, we can
choose any open interval. This interval then contains a point that under iteration becomes an
element of every other interval of the domain. The last requirement is one of regularity. Amidst the
chaos there are periodic points as well, and even infinitely many and everywhere in the domain.
Propostition: periodic points are dense in [0,1] under iteration of the tent map

Lets recap on density. We call a set ‘dense’ if there’s a point to be found between every other two
points. Think of the set of fractions. Now to address the question why periodic points are dense in
[0,1] under iteration of the tent map.
Considering this context our question means that 1
between any two periodic points on the interval [0,1],
a third one exists We know of the existence of two
fixed points, namely zero and two-third in the first-
iteration-graph. Now if we proceed to constructing
the second-iteration-graph, the interval on which it
projects from 0 to 1 gains a fixed point. This interval
is [0,½], but also [½,0]. After all, if some area projects
to [0,1] after one iteration, then the second iteration is
again a tent map. It follows that in the second-
iteration graph two ‘small tents’ be visible. These
small tents also go from zero to one, but on half the
domain. You can imagine that for the third iteration
graph, two more tents become visible. And these 0 1
tents, whose number increase as the iteration graph
increases, all go from zero to one on a certain
domain and thus cross the y=x line. 1

Subproposition: The set of (eventually) periodic

points is equal to the set of fractions.

Proof: each fraction can be written as a/b with a,b∈Z.

But the number of fractions with b in the lower part is
limited. In this sense only five fractions of four exist in
[0,1], 0/4, 1/4, 2/4, 3/4, 4/4. And since iteration only
incorporates doubling with or without a substraction
of either 2 or ½, and in this sense remains a fraction
or at most is turned into a fraction in which a and b
are multiplied by two (for the ½ compatibility). 0 1
Considering that the is only a limited number of a
fraction with a certain b it follows that if we iterate a The first and second-iteration graph of
fraction it must repeat itself under iteration. In other the tent map.
words: all fractions are periodic points. Is the converse
also true? That is, are all periodic points fractions, or are there also periodic points that are not
fractions?
The truth reveals itself by looking at the nth-iteration graph of the tent map. As we have seen, each
iterations more tents emerge, the same height but on ever smaller domain pieces. This means the
derivative of each linepiece after n iterations is ±2n. Since each of these linepieces goes from 0 to
1 they each cross the y=x line (guaranteed by the intermediate value theorem). So solving all
these line formulas to y=x gives all the fixed points. Because of their form y = ±(2nx) ± c (with
c∈N) these values are all fractions. Concluding we may confidently say that all fractions are
periodic points and all periodic points are fractions. The sets are equal. Since fractions are dense
in D, so are periodic points.
Proposition: the tent map is topologically transitive

Definition: a function f: D→D is topologically transitive if for two open intervals U and V in D there
is a point z in U such that fn(z) ∈V for some n.

So a function from and to the same interval (in our case [0,1]) is topologically transitive if from any
open interval you pick a point it travels to any other open interval under iteration of f. We could say
such a function ‘well- mixes the domain’. You might wonder: “why not say: pick any point and
under iteration it goes everywhere in the domain?”. This would be a reasonable question but
wrong in the details. We cannot pick just any point, for if we pick a fraction it does not go
everywhere since it’s periodic. Therefore, we need the open interval.
We have seen that the set of fractions is equal to the set of periodic points. From this it follows
that any point which is no fraction is no periodic point. Lucky us, for there’s the square roots.
Square roots are not fractions as we’ve seen in the preliminaries and second: they’re dense. So if
we pick any open interval, however small, there’s a square root in it. So if we can prove that every
arbitrarily chosen square root travels around in the entire interval [0,1] we have proven that the
tent map is topologically transitive. Here goes.
Take a square root w. For instance w = √(1/6). This square root has two precursors, one on the
left and one on the right of ½. Remind yourself that this is true because the graph of the tent map
goes from 0 to 1 on the interval [0, ½] but also on the interval [½,1]. So on both these intervals
there must be a value (let’s call them w1 and w2) which after iteration takes on the value √(1/6).
And here’s the recursive trick again. These values w1 and w2 of course also have two
precursors, both on one side of ½. Let’s give these precursors of w1 the name w1a and w1b and
similarly w2a and w2b for w2. Because they are precursors f(w2a)=w2 and f(w2b) = w2. And since
iteration of w2 yields w, it follows that f2(w2a) = w. The further we’re looking at precursors the more
we find. There are two points for one iteration (which we called w1 and w2), there are four for two
iterations (which we labeled w1a, w1b, w2a and w2b) and similarly eight points for third-iteration-
precursors of w.
Since the two points of a precursors are located on either side of ½, that is, one in the first half
of the domain and one in the second half of the domain, we can imagine that second-iteration-
precursors all lie in their respective quarters of the domain. This is not strange to imagine because
the second-iteration-graph of the tent map assumes all values, including w, four times (check out
the earlier shown pictures if you’re not too sure).
Continuing is this fashion, there are eight points in [0,1] which project to w after three iterations
and 16 points for four iterations. The further we look the more we find. For 30 iterations we find
1073741824 (=230) different points that eventually yield √(1/6), and all nice and equally distributed
on [0,1]. In other words, precursors of √(1/6) are dense in [0,1]. This again means that for any
open interval we choose, there is a precursor of √(1/6) in it. You might get the point right now, but
for the sake of completeness we’ll have a look at the proposition again.

Proposition (restated): The tent map is topologically transitive if for any two open intervals U and
V in [0,1] there is a point in U that under iteration goes to V.

Take some open interval V. Take some square root value. This is possible because square roots
are dense in [0,1]. Since it’s precursors are also dense there is a precursor in U and therefore
there is a point in U which under iteration reaches V (eventually). Ergo: the tent map is
topologically transitive.
Proposition: the tent map is sensitive to initial conditions

This last requirement might just be the most characteristic for a chaotic dynamical system:
sensitivity for initial conditions. This is a very easy requirement. Of you take any two values really
close together two things can be the case:
Take two initial conditions, x0 and y0, which are very at at a very small distance (let’s call this
distance δ (delta)). Now two situations are possible:

1) x0 and y0 are on the same side of ½ (both under the same diagonal)
2) x0 and y0 are on different sides of ½,

Note that if 2) is the case they lie very close to ½ just because we chose δ so small. Because of
this, they might even be closer together after one iteration, but at least then they’re on the same
half of the domain by which situation 1) is the case. As soon as they are, they each iteration, their
difference will get twice as big because they project to a line that, due to its steepness, covers
twice the distance vertically as it does horizontally.

2.6.2 Topological conjugacy

If you would ever pick up a dutch bible, the first words you would read would be “in den
beginne...”. If you would ever read the first words of an english bible it would say “In the
beginning...”. The exact same idea though in a different language. You could prove this by
constructing a mapping from english to dutch {(in,in), (the,den), (beginning,beginne), ... }. Note that
this mapping also works the other way around, from dutch to english. The essence, the structure,
is the same though the language is different.
We have reached the final stage of our proof. We have just shown that the tent map matches all
criteria for being a chaotical function and now we will show, by a mapping called a topological
conjugacy that the structure of the tent map and the logistic map for A=4 are the same though their
appearance is different. This topological conjugacy could be seen as the language mapping for our
maps.

Definition: We have two functions on their domains: f:D→D and g:E→E. A topological conjugacy
is a homeomorphism τ:D→E such that τ(f(x)) = g(τ(x)). If such a topological conjugacy exists
between f and g we call them topologically conjugated.

This definition ircorporates the word homeomorphism

Definition: Een homeomorphisme is een function that is surjective, injective, continuous and has a
continuous iverse.

In an injective function, every projection is unique. The function f(x)=x2 for example, is not injective,
because both -2 and 2 project to 4. A surjective function uses the whole codomain. The function
f(x)= √x is surjective on [0,4]→[0,2] because for every value in codomain [0,2] there is an original
in [0,2] that projects to it. For the same reason [4,16]→[0,4] is not surjective. If we take the value 1
from [0,4], there is no original in [4,16].
Example homeomorphism: f(x)=tan(x) is a homeomorphism from (-π/2, π/2) to R. For all values
in (-π/2, π/2) there is a uniquely corresponding value from R. is er een unieke bijbehorende waarde
uit R. Besides, the entire R is used. Tan(x), and arctan(x) are both continuous and therefore, tan(x)
is a homeomorphism from (-π/2, π/2) to R.

Example topological conjugacy: Two functions f(x) = ½x :[0,2]→[0,2] and g(x)= ¼x:[0,4]→[0,4]
are topologically conjugate. The topological conjugacy used is τ(x)=x2. Look what happens if we
take the highest initial value:

f(x) = ½x :[0,2]→[0,2] 2 1 0.5 0.25 0.125

g(x) = ¼x:[0,4]→[0,4] 4 1 0.25 0.0625 0.015625

Clearly to be seen is that both system’s initial values tend towards the 0 attractor. The vertical
quadratical relation is clearly visible. For explanatory values, we give a commutative diagram In
which the arrows indicate the appliances of f(x),g(x) en τ(x). Check the diagram to convince
yourself, that improves understanding of the concept.

2 →f(x) 1 →f(x)→ 0.5 →f(x)→ 0.25 →f(x)→ 0.125

→
↓ ↓ ↓ ↓ ↓
τ(x) = x2 τ(x) τ(x) τ(x) τ(x)
↓ ↓ ↓ ↓ ↓
4 →g(x) 1 →g(x)→ 0.25 →g(x)→ 0.0625 →g(x)→ 0.015625
→

But analytically it also makes sense: The paramount prerequisite for a topological conjugacy is:

τ(f(x))=g(τ(x)), In our case: (½x)2= ¼(x2), written out: ¼x2 = ¼x2

In other words: since this equation is firm like a rock and τ(x) [0,2] -> [0,4] is a homeomorphism,
f(x) and g(x) are topologically conjugate.
The tent and it’s logistics. There is a topological conjugacy from the tent map to the logistic
map, though it’s split up in two parts because of the tent’s sharp top. This is a mere detail. The
topological conjugacy for x≤½ is τ(x)= sin2(½π x). It relies on some mathematical formulas and
there’s no very easy way of explaining them in the proof, so we will just apply them:

1) sin2(πx) = 4sin2(½πx) cos2(½πx)

2) cos2(½πx) = 1- sin2(½πx)

τ(T(x)) = h(τ(x)), filled out:

sin2(½π 2x) = 4sin2(½π x) (1-sin2(½π x)) by expanding on 2:

sin2(½π 2x) = 4 sin2(½π x)cos2(½πx) and then by expanding on 1:

4 sin2(½π x)cos2(½πx) = 4 sin2(½π x)cos2(½πx)

The function sin2(½π x) is continuous, injective, invertible and has a continuous inversion on [0,1].
Therefore it is a homeomorphism and so the tent map and the logistic map are topologically
conjugate. They have similar dynamics and since the tent map was proven chaotic, so is the
logistic map. By now, you should be aware of the powerful forgery that a homeomorphism is.

2.7 Back to biology

What implications do all these results have for the real world? Just imagine our model for
population growth, Ax(1-x) in which A is the fertility rate, is correct. If the biologist measures the
number of lizards on the island and wants to make estimates for the future, he could be in serious
trouble As we have seen, for low A-parameter values the system simply propagates towards its
fixed point. This would mean for the biologist that he could predict very accurately what the
population goes to. If the fertility parameter is slightly higher, he must realize that the population
size will fluctuate between certain values, or , the population fluctuates over a period of two, four or
eight years. But what happens if the fertility rate grows beyond 3.57? All predictability ceases to
exist and there is no way of saying what next year's number of lizards will be.
This is quite a fundamental argument. Of certain phenomena of nature, such as gravity exerting its
influence on a rock, we can very accurately predict what will happen; we have very well-defined
formula's and the only chance incorporated is the small perturbation we cannot measure of choose
to neglect. But certain aspects of nature so it seems, are fundamentally unpredictable. One such
an example is the logistic map. Another such example might be the weather. Fully governed by a
large set of differential equations, it shows some of the same characteristics including
unpredicatbility. And the most remarkable is: the unpredictability is not due to minor external
fluctuations, it is a property of the system itself.
3. Randomness

3.1 Proces & product

It’s a rainy day in Amsterdam. You’re sitting in your room, bored because it’s raining, flipping a
guilder coin through the air and seeing which side it lands on. From the outcome, you decide it’s
totally random whether the heads or tails comes up. As on all rainy days, you’re kind of
philosophical and wonder whether the chaos of the universe reflects the unpredictability of the coin
flip. After quite a few flips, you decide the coin flip is a real random generator. There’s no saying
which side comes up. They both have the probability ½. You recall a famous statement by Joseph
Ford: “Chaos is merely a synonym for randomness.” “Is this true?, Is chaos the same as
randomness?” you wonder. To this rainy day, we’ll devote a chapter.
After getting your umbrella and raincoat, you go out to visit a friend. Bad weather as it is, you want
to take a taxi. You walk up to Amsterdam central station where lots of taxis are parked waiting for
passengers. You skim the numbers briefly to see if your favourite number 71 is in but as you look,
you only see even numbers. This strikes you as extremely odd. Only taxi’s with numbers
2,4,6,8,10,12.... Why is this? Are only even numbers going on rainy days? It could be a
coincidence but that doesn’t seem very likely to you. Still it could be just a coincidence, every
combination of taxinumbers is equally probable but you refuse to believe it is. It just can’t be.

"In our minds, we organize all events in classes. We regard as special classes those which are
very rare. As such, we regard one hunderd consecutive heads from a coin flip as special,
because it has much more structure and structured sequences are rare as opposed to
unstructured sequences. That is, sequences with repeative patterns which we can conceive
and understand. " [freely translated after Laplace]

A statement by Pierre-Simon Laplace (1749-1827).

What Laplace means is further illustrated once you
reach your friend. Since the bad weather, he’s just
as bored as you and you decide to play some
scrabble. You both take seven letters from the box
and put them on your stand. H-R-N-I-C-E-K. Not very
good. You can make ‘hire’ or ‘neck’ but it’s a nasty
combination. Your friend starts and lays down the
word ‘chaotic’. All letters gone in one go. He’s won in
the first turn. How is this possible? First the taxi
numbers and now this! He must be cheating! How
else can he draw the letter C-H-A-O-T-I-C and you a
worthless combination like H-R-N-I-C-E-K?
Still, it just might be a coincidence and this exactly
emphasizes laplace’s point. Such an event you
would classify as rare, because it’s structured and
there’s more unstructured than structured
sequences. It’s all about the structure in your
interpretation. If the same game would be played in
Hungaria the opposite could be the case. The guy
with H-R-N-I-C-E-K on his stand lays it all down in
one go because ‘hrnicek’ happens to be a perfectly Pierre-Simon Laplace (1749-1827)
constructed hungarian word.
The guy with ‘chaotic’ would be perplexed, for he doesn’t speak english. (translated from
hungarian)“This can not be a coincidence! How can you get ‘hrnicek’ in one go while I have a
rubbish combination like C-H-A-O-T-I-C on my stand? You must be cheating!” Aim of the story is
that all combinations of letters have the same probability but although a combination like
“CHAOTIC” has the same probability as "HRNICEK", "AKNLWMB" or "ENIARKU", it is very clear
that our intuition of randomness of objects (like taxinumbers or scrabble-sequences) is dependend
on whether is has interpretable structure.
Back to our more universal, language-independent example. We take a fair coin (assume that
probabilities are fifty-fifty). If we now flip a number of times adn recors a ‘1’ each time heads
comes up, and a ‘0’ for each tails. The result will be a random string, no matter what it be. It might
be 100 zeros or 100 ones but since we assumed that our coin is fair, it is random. In such cases
you could wonder if the coin really is fair, but that would trap us in a circle-statement: “We flip our
fair coin 100 times, but if the outcome is really structured the coin wasn’t fair after all.” But what if
we only get a string presented and then have to decide wether the generation of this string was
done by a random process?
And here’s our problem presented on a silver plate. It depends on the process wether the
product is random. But what if we only have the product? Can we say anything about its
randomness without knowing how it was formed?

3.2 Infinite strings

Let’s examine a string of which we don’t know whether the process was random or not. An infinte
string of zeros and ones. The chance process guarantees us that the distribution of zeros and
ones should be about fifty-fifty if the string is random. Imagine we do heads or tails a ‘large number
of times’. Then half that ‘large number of times’ should be heads, the other half be tails. But is this
enough to decide for its randomness? We flip a number of times and get :

1100110011 0011001100 1100110011 0011001100 1100110011 0011001100 1100110011 00....

Now that doesn’t look very random. If this continues in this fashion, you would be able to predict
exactly the 1021st outcome. Just a repeating pattern. If the first is ‘1’, then so is the 21st, the 41st,
the 81st and so the 1021st. And isn’t it a hallmark of randomness not to be predictable? But yet, the
distribution of heads and tails exactly matches the fifty-fifty criterion. Obviously, that requirement is
not enough. For instance, the subsequence ‘000’ is to be found nowhere. But also five
consecuitive tails do not exist in this sequence.
The mathematician E. Borel called a string of zeros and ones ‘normal’ if every block of length n
had frequency 2-n in that string. Would you count the number of blocks of length two (00,01,10,11)
than every block would have equal probability and thus account for 25%. Would you count all
blocks of length four (0000,000... 1111) then each of these should have 1/16 = 6.25%. That seems
a likely measure for a random string but there came Champernowne.
Champernowne has show that numbers exist that are normal in Borel’s sense but still not
correspond to our intuition of randomness.:

1234567891011121314151617181920212223... (Champernowne's number)

Intuitively very clear that this is not a random string (be it on base 10 except of 2). Still every block
of every length has the same frequence. It follows that being-normal according to Borel is not a
sufficient measure for randomness. On the other hand, if a string is random it is also normal,
otherwise we would have an unequal disribution of certain subsequences and that would not be
the result of a fair coin.
Valid is: "x is random ⇒ x is normal"
And also: "x is not normal ⇒ x is not random",

But not: "x is normaal ⇒ x is random" (because of Champernowne's number, which

Is normal but still not random)

The randomness of an individual infinite string is quite hard to pinpoint. Even be there a statistical
measure, like some advanced-Borel-normality there would be a practical problem. The string
would in the long run converge to match this measure. But what is the long run? A thousand coin
flips? A million? 10600 coin flips? None. Because even after 10600 coinflips there could come 10700
zeros messing up our statistical measure. As Keynes told us “The long run? In the long run we’ll all
be dead!” by which he meant you can never trust a sequence to converge up to a certain point.
There is always a subsequent-sequence that trows a rock in the pond. By now it might be clear
that the determination of randomness of an individual string is no Swiss cottage cheese.

3.3 Finite strings

Let’s have a look at finite strings. If the randomness of an infinite string is hard to pinpoint, what to
say about finite strings now that even convergence is not something to rely on? What makes a
finite string random or not?
Back to our heads or tails game. We decide to flip our coin twenty times and denote the result.
We repeat this experiment five times which rends the following data:

1. 11011100101010001010
2. 11111111111111111111
3. 00011011100000101110
4. 01101010001010001010
5. 10101010101010101010

Some of these strings look very random. Some don’t. Even if the outcome seems to match a fifty-
fifty criterion in heads-tails it still doesn’t need to be random. Borel’s normality does not apply for
infinte strings, but even apart from that, look at string number 5. Look at string number 2. No
sensible human being would call either of those a random string. There is a gap between our
intuition and science. Probability theory says each string has the same chance though 2 and 5 feel
very unrandom. We can clearly see structure. Laplace, appear!
What is this structure we observe? In any case, repeating patterns exhibit a ‘rule’, a generalized
way of its apperance. String two is just 20 times “1”, string 5 just 10 times “10”. The nice thing is
that we can formalize this. Repeativeness, structure means we can give an ‘easy formula’ (20 x
"1") that describes the original object (1111111111111111111). Obviously, the randomness we
feel to haunt over certain strings have something to do with it’s lack of pattern whereas orderly
strings seem to be describable in an easy formula. An appeal to intuition: which telephone number
is easier to remember:

1) 0031 - 37501 - 683992 -1143893

2) 4444 - 55555 - 666666 -7777777

Obviously the one with the most structure. Obviously the one for which we can, in our minds,
design an easy formula to remember. Is there a way of formalizing this intuition? The answer is of
course yes, but we will have to rely on computational theory. You will have to know a little it about
computer programming for a thorough understanding but references to ‘method’ and ‘algortihm’
will also be given. Check strings A and B:

A: 1111111111 1111111111 1111111111 1111111111 1111111111 1111111111

B: 1010100011 1011111101 1011111010 1000110110 0011111011 1010101001

String B is random, String A is not. String A can be written as:

A2: For a=1 to 60; print"1"; next a

As string B can be written as:

B2: print"1010100011 1011111101 1011111010 1000110110 0011111011 1010101001"

And now we have caught our ‘structural intuition’ in a computer program. String A can be given by
a very short description, a repeative algorithm whereas string B cannot. It is exactly this ‘being
describable in a short formula’ which indicates the randomness of a finite string. In computer
science we would say: String A is compressible, string b is not. The program A2 required to
produce A is shorter than A itself. The shorter this description, the less random it is.
We bump our heads into a number of problems at once. As a practical one, this method is
obviously not valid for strings of all lengths. If we have a string of only ten bits, it is impossible to
find a computer program which is shorter. Computer programs of less then ten bits do not exist.
Second, it obviously depends on the programming language we use. If we use a cumbersome
programming language with long instructions the description will obviously be longer.
Third, how do we know which program is actually the shortest we can find. Obviously, we can add
a huge amount of useless instructions to any program, so there are many programs that can
produce the string we want to compress. It is also possible to device two sets of instructions totally
equivalent in output. Which we use is of course of critical importance to the length of our program.
Is there any saying if there is an even shorter equivalent set?
But all these problems can easily be avoided by referring to a Turing Machine. Alan Turing, an
englishman and pure genius that decyphered the German enigma-code in the second world war
was also foremost concerned with computation theory and artificial intelligence. He designed a
machine, later called the Turing Machine, which can be seen as a purely mathematical model of a
computer. Using this abstraction we can avoid or solve some of the issues. In essence, all
computer-processors are able to perform the same operations. These operations are described
by the Turing Machine and as such diffuses the question of different programming languages.
We can very easily make some general remarks about structureless strings now. Imagine we have
a 100-bit long string. We call such a string ‘random’ if it is compressible by 20%, that is, if there is
a program of 80 bits or less that exactly reproduces our string . This is a very modest choice. If you
WINZIP a text file you can reach compression rates of 50% easily. How many of these strings do
exist anyways? Binairy calculation shows there are 2100 ≈ 1.26*1030 such strings. But there are
only 281-1 ≈ 2.42*1024 strings of length 80 or shorter. So how many strings are incompressible
according to our 20% standard? At least 1.26*1030 - 2.42*1024! More than 99% that is. A simple
counting argument gives a striking result. With a very modest compression rate on a relatively
short string, the number of random strings very far exceeds the non-random strings. And just how
neetly does this fit Laplace’s intuition: “...and structured sequences are rare as opposed to
unstructured sequences ...”.
The notion of randomness as stated here is the invention of A.N. Kolmogorov and closely related
to Kolmogorovcomplexity. His original notion was far more complicated and thorough, but the main
idea is that randomness, lack of structure, is conceivable of as incompressibility.
3.4 Chaos and randomness

Then the main question. To what extend is chaotic behaviour comparable to randomness? We ask
ourselves this question step-by-step, always related to the logistic map.
• As a process: A random process is a process for which a certain probability distribution
governs the outcome. The logistic map is different. It is deterministic. Each starting value
projects only to one subsequent value. Therefore, the logistic can never be regarded of as a
random generator.
• As a product. If we take a product, like the binary expanded orbit of initial value ½, is the
product it yields by any chance random. No. There is an easy-to-construct computer program
that does the iteration for us and therefore the whole string would be compressible.
• So what do they have in common? Answer is unpredictability. For 2000 coinflips there is no
saying what the outcome is and so it is for the logistic map. It just follows a certain pattern of
probability and so does the logistic map. When A=4, orbits dance around In a negative-
exponential manner always showing the same statistical characteristics, never the same
pattern.

3.5 Exercises

1. Below depictuted are a few samples of the Japanese board game GO. In all of them, fifty
squares are covered with white chips and fifty are covered with black chips. Which look the
most randomly covered? Why?

1 2 3

4 5 6
2. Which of the following phonenumbers is easiest to remember? Which is the intuitively most
compressible and how is this related?

A: 0041-4572-33819
B: 0011-9999-99999
C: 0057-5757-57162

3. Put these 54-bits strings according to increasing randomness.

A: 111110011010100001110001010000110010101010110001110101
B: 110011001100110011001100110011001100110011001100110011
C: 101001000100001000001000000100000001000000001000000000
D: 001010011101011010010101000011101010001111010111001101
E: 111111111111111111111111111111111111111111111111111111
Solutions to exercises

Paragraph 1.3

1. Consider the discrete dynamical system: xt+1 = 4xt

a) The system's only fixed pint is 0.
b) It has no periodic points.
c) The orbit of-2: {-2,-8,-32,-128,-512,...}.
The orbit of 4.1: (4.1, 16.4, 65.6, 262.4, 1049.6,...)
d) 0 is a repellor

2. Consider the discrete dynamical system: xt+1= 2xt(1-xt)

a) The orbit of 0: {0}
The orbit of 0.5: {0.5}
The orbit of 0.1: {0.1, 0.18, 0.2952, 0.416.... and further towards 0.5)
The orbit of 1: {1,0}
b) This system has one repellor at 0 and one attractor at 0.5. In fact, this is a special case of
attractor; in this case, the attractor is the critical point and convergence of all orbits towards
0.5 is much faster than in other cases. It is therefore sometimes called superstable.

3. Consider the discrete dynamical system: xt+1= ¼ xt

a) It has one fixed point at 0.
b) It has no periodic points other than zero (which we call fixed, because its period is 1)
c) They both tend towards zero.
d) 0 is an attractor.

4. Consider the discrete dynamical system. xt+1= 3.1xt - 3.1(xt)2

a) It has two fixed points: one at 0 andthe other at 21/31
b) This is best done by graphical analysis
c) -2 goes to -∞, 0.9 tends to a period two orbit.
d) Determine of all fixed points are repellors, the periodic orbit is attracting.

5. A colony of kangaroos inhabits southern Australia. This population doubles each year.
a) xt+1 =2xt, in which x is the population size and t the time.
b) Because your (inaccurate) estimate will also double under iteration the margin remains
within 5% correct. In fact, it will stay precisely the same value.
Paragraph 1.5

1. Consider the discrete dynamical system xt+1 = 2xt on R->R

a) It has one fixed point at 0, found by solving 2x=x
b) It is a repellor (unstable), since the derivative is 2 which is greater than 1.

2. Consider the discrete dynamical system xt+1 = (xt)2

a) It has two fixed points: 0 and 1
b) The derivative of f(x)=x2 is f(x)=2x. We can determine the hyperbolicity by determining f’(0)
and f’(1).In this case, f ’(0)=0, so this one is stable whereas f’(1)=2 which implies it is
unstable.

3. Consider the discrete dynamical system xt+1,= 2+√xt on [0, ∞) →[0, ∞)

a) It has two fixed points: one in 0 en and one in 4.
b) The derivative of f(x) = 4x is given by f (x) = 1/(2√xt). In 0, the derivative is not defined, but
any value arbitrarily close to zero is very high and thus we may conclude that it repels.
In 4 the derivative is 0.25 and it is therefore an attractor.

Paragraph 3.5

1. The least random are obviously boards 5,3 and 6. They are all comprised of repeating patterns.
Intuitively, we can remember them easily “Oh, it’s just half-black half-white.” Or we could easily
construe a computer program for them. The most random are probably boards one and two and
board four is somewhere in between. It could be recreated by a computer program which lines
1,2,3 are generated by a repeating loop, 4 is literally printed, 5,6,7,8 are again looped, 9 is again
printed and 10 is a loop again.

2. Telephone number B is most easily remembered and most easily reconstructable from an easy
algorithm. C has some repetition but a small irregularity on the end and is therefore more random
whereas A is the most patternless. Note that this is an appeal to intuition more than to computation
theory.

3. E,B,C,A,D