2/22/2013

1
IES-01 Fractals and Application
Attendance: 10 marks
Assignments: 10 marks
Class Performance: 5 marks
(Weightage: Absent : 0, Late : , Present : 1)
Geometry : developed as a collection of tools for understanding
the shapes of nature.
For millenia, Symmetry has been recognized as a powerful principle in geometry,
and in art.
We begin by reviewing the familiar forms of symmetry, then show that fractals
reveal a new kind of symmetry, symmetry under magnification.
Many shapes that at first appear complicated reveal an underlying simplicity when
viewed with an awareness of symmetry under magnification.
We begin by reviewing the familiar symmetries of nature: symmetry under
translation, reflection, and rotation.
We are familiar with three forms of symmetry, exhibited approximately in many
natural and manufactured situations. They are translational, reflection, and
rotational
Less familiar is symmetry under
magnification:
zooming in on an object leaves the shape
approximately unaltered.
Here we introduce some basic geometry of fractals, with emphasis on the Iterated
Function System (IFS) formalism for generating fractals.
In addition, we explore the application of IFS to detect patterns, and also several
examples of architectural fractals.
First, though, we review familiar symmetries of nature, preparing us for the new
kind of symmetry that fractals exhibit.
The geometric characterization of the simplest fractals is self-similarity: the
shape is made of smaller copies of itself. The copies are similar to the whole:
same shape but different size.
The simplest fractals are constructed by iteration. For example, start with a filled-
in triangle and iterate this process:
For every filled-in triangle, connect the midpoints of the sides and remove the
middle triangle. Iterating this process produces, in the limit, the Sierpinski
Gasket.
The gasket is self-similar. That is, it is made up of smaller copies of
itself.
We can describe the gasket as made of three copies, each 1/2 as tall and 1/2 as
wide as the original. But note a consequence of self-similarity:
each of these copies is made of three still smaller copies, so we can say the
gasket is made of nine copies each 1/4 by 1/4 of the original, or 27 copies
each 1/8 by 1/8, or ... . Usually, we prefer the simplest description.
This implies fractals possess a scale invariance.
2/22/2013
2
More Examples of Self-Similarity
The gasket is made of three copies of itself, each scaled by 1/2, and two copies
translated. With slightly more complicated rules, we can build fractals that are
reasonable, if crude, approximations of natural objects.
Later we will find the rules to make these fractals.
For now, to help train your eye to find fractal decompositions of objects, try to
find smaller copies of each shape within the shape.
The tree is not so hard, except for the trunk.
The Mandelbrot set: a different nonlinear transformation gives the most famous of
all fractals.
Fractal landscapes: With more sophistication (and computing power), fractals
can produce convincing forgeries of realistic scenes.
Making realistic-looking landscapes is
difficult enough, but doing this so they can
be stored in small files is remarkable.
Fractals in nature: after looking at so many geometrical and computer-generated
examples, here is a short gallery of examples from Nature
Fractals found in nature differ from our first mathematical examples in two
important ways:
the self-similarity of natural fractals is approximate or statistical and
this self-similarity extends over only a limited range of scales.
To understand the first point, note that many forces scuplt and grow natural
fractals, while mathematical fractals are built by a single process.
For the second point, the forces responsible for a natural fractal structure are
effective over only a limited range of distances.
The waves carving a fractal coastline are altogether different from the forces
holding together the atoms of the coastline.
One way to guarantee self-similarity is to build a shape by applying the same
process over smaller and smaller scales. This idea can be realized with a process
called initiators and generators.
The initiator is the starting shape.
The generator is a collection of scaled copies of the initiator.
The rule is this: in the generator, replace each copy of the initiator with a scaled
copy of the generator (specifying orientations where necessary).
The initiator is a filled-in triangle, the generator the shape on the right.
Sierpinski Gasket How can we turn "connect the midpoints
and remove the middle triangle" into
initiators and generators?
2/22/2013
3
Koch curve
Tents upon tents upon tents ... makes a shape we shall see is very strange, a
curve enclosed in a small box and yet that is infinitely long.
Take as initiator the line segment of length 1, and as generator the shape on
the right.
Though its construction is so simple, the Koch curve has some properties that
appear counterintuitive.
For example, we shall see that it is infinitely long, and that every piece of it, no
matter how small it appears, also is infinitely long.
Cantor set
Cut all the tents out of the Koch curve and we are left with something that
appears to be little more than holes. But we can be fooled by appearances.
Again, take as initiator the line segment of length 1, but now the generator
is the shape shown below.
Here is a picture of the Cantor set resolved to the level of single pixels.
Although so much has been removed that the Cantor set is hardly present at
all, we shall find this fractal in many mathematical, and some physical and even
literary, applications.
Fractals in the Kitchen
Cauliflower is a wonderful example of a natural fractal. A small piece of a
cauliflower looks like a whole cauliflower.
Pieces of the pieces look like the whole cauliflower, and so on for several more
subdivisions.
Here is a picture of a cauliflower and a piece broken from it.
As -------- cook, the boiling batter forms bubbles of many different sizes, giving rise
to a fractal distribution of rings.
Some big rings, more middle-size rings, still more smaller rings, and so on.
2/22/2013
4
Some breads are natural fractals. Bread dough rises because yeast produces
bubbles of carbon dioxide.
Many bubbles are small, some a middle-size, a few are large, typical of the
distribution of gaps in a fractal.
So bread dough is a foam; bread is that foam baked solid.
Kneading the dough too much breaks up the larger bubbles and gives
bread of much more uniform (non-fractal) texture
Do fractals have practical applications?
How about an invisibility cloak?
On Tuesday, August 28, 2012, U.S. patent
number 8,253,639 was issued to Nathan
Cohen and his group at FracTenna, for a
wide-band microwave invisibility cloak,
based on fractal antenna geometry
The antenna consists of an inner
ring, the boundary layer, that
prevents microwaves from being
transmitted across the inside of
this ring. This is the region that
will be invisible to outside
observers. Surrounding the
boundary layer are six
concentric rings that guide
microwaves around the
boundary layer, to reconverge at
the point antipodal to where they
entered the cloak.
On the left is a magnification of one of the outer rings of the cloak. On the right is
the boundary layer fractal.
If fabricated at the sub-micron scale, instead of the current mm scale, this
technology should act as an optical invisibility cloak.
In late August, 2012, Cohen's group cloaked a person. Interesting times ahead.
www.fractenna.com
Now down to work. We learn to grow fractal images, but first must build up the
mechanics of plane transformations.
Geometry of plane transformations is the mechanics of transformations that
produce more general fractals by Iterated Function Systems
To generate all but the simplest fractals, we need to understand the geometry
of plane transformations. Here we describe and illustrate the four features of
plane transformations
Affine transformations of the plane are composed
of scalings, reflections, rotations, and translations.
Scalings
The scaling factor in the x-direction is denoted r.
The scaling factor in the y-direction is denoted s.
Assume there are no rotations. Then if r = s, the
transformation is a similarity
otherwise it is an affinity
Note the scalings
are always toward
the origin. That is,
the origin is
the fixed point of
all scalings.
Reflections
Negative r reflects across the y-axis.
Negative s reflects across the x-axis.
Reflection across both the x- and y-axes is equivalent to rotation by 180about the origin
Rotations
The angle measures rotations of horizontal lines
The angle measures rotations of vertical lines
The condition = gives a rigid
rotation about the origin.
Positive angles are
counterclockwise
2/22/2013
5
Translations
Horizontal translation is measured by e
Vertical translation is measured by f.
The matrix formulation of an affine transformation that involves scaling by r in the
x-direction, by s in the y-direction, rotations by and , and translations by e and
f.
We adopt this convention:
scalings first, reflections second, rotations third, and translations last.
This order is imposed by the matrix formulation.
Emphasizing this order, the components of a transformation are
encoded in tables of this form
With this encoding of transformations of the plane, we can make fractals using
the method called Iterated Function Systems (IFS)
Generating fractals by iterating a collection of transformations is the Iterated
Function System (IFS) method, popularized by Barnsley, based on theoretical work
by Hutchinson and Dekking. We use a simple example to see how it works
Iterated Function Systems
To illustrate the IFS method, we show how a specific set of IFS
rules generates a Sierpinski gasket
We begin with a right isosceles Sierpinski gasket. Certainly, the gasket can be
viewed as made up of three copies of itself, each scaled by a factor of 1/2 in
both the x- and y-directions
To determine the translation amount of each piece, take
some point of the whole fractal (the lower left corner, for
example) and observe where that point goes in each
piece.
Here we derive the rules for the right isosceles Sierpinski
gasket
Invariance of the Gasket
Note that applying all three of these transformations to the gasket gives the gasket
again
That is, the gasket is invariant under the simultaneous application of these three
transformations.
What happens if we apply these transformations to some shape other than
the gasket?
What happens if we apply these transformations to the resulting shape?
What happens if we iterate this process?
Here is an instance of this idea applied to a sketch of a cat
We observe a sequence of pictures that
converges to the gasket, independently of
the starting shape.
For concreteness we illustrate this converge using the gasket rules. Because
all the transformations are applied at each iteration, this is called
the determinisitc algorithm.
Specifically, suppose T
1
, ..., T
n
are contractions, and P
0
is any picture.
For example,
T1(x,y) = (x/2, y/2),
T2(x,y) = (x/2, y/2) + (1/2, 0),
T3(x,y) = (x/2, y/2) + (0, 1/2),
and P
0
=
Generate a sequence of pictures
P1 = T1(P0) ... Tn(P0)
P2 = T1(P1) ... Tn(P1)
...
Pk+1 = T1(Pk) ... Tn(Pk)
This sequence converges to a unique shape, P, the only (compact)
shape invariant under the simultaneous application of T1, ..., Tn:
P = T
1
(P) ... T
n
(P) That is,
Because of this convergence property, P is
called the attractor of the IFS {T
1
, ... , T
n
}.
2/22/2013
6
Inverse problems
finding the transformations to produce a given fractal
Given a fractal F, the Inverse problem is to find affine
transformations T
1
, ..., T
n
for which
F = T
1
(F) ... T
n
(F)
Here we present a method to solve this problem
Solving the inverse problem takes just two steps.
1. Using the self-similarity (or self-affinity) of F, decompose F as F = F
1
... F
n
, where
each F
i
is a scaled copy of F.
2. For each piece F
i
, find an affine transformation T
i
for which T
i
(F) = F
i
. By "find an affine
transformation" we mean find the r, s, , , e, and f values.
Remarkably, solving the inverse problem has only two steps:
Because the transformations can involve rotations, reflections, and scalings
by different factors in different directions, decomposition is not always as
simple a task as it may seem at first. Here are some examples of more
complicated decompositions.
Decomposition
This fractal is an instructive example for people who have
seen the gasket and a few of its relatives.
The primacy of the gasket in early examples of fractals
makes this shape one of the easiest to recognize.
The most common response to first seeing this picture is,
"It's half a gasket."
But we don't have rules for making half of a fractal.
The main lesson here is that we're looking for scaled copies of the
whole shape, and the whole shape is not a gasket.
Tracing small copies of the outline of the whole shape, perhaps
cutting them out of paper, is a good way to build up intuition for this
process. Here's a decomposition.
Note the bottom left piece
is a reflected copy of the
whole shape
In the x-direction we see the familiar
Cantor middle thirds set; in the y-direction
just a line segment. Again, look for scaled
copies of the whole shape. Here's
a decomposition.
An additional problem is that decompositions never are unique. Here are
some examples of different decompositions of the same fractal.
Find a decomposition of this fractal into
smaller copies of itself.
Here's one decomposition, and
here's another.
Usually, we try to find a decomposition into
the smallest number of pieces, keeping in
mind that each piece must be a contracted
copy of the whole shape.
We have already seen one decomposition
of this fractal.
When we give up the requirement that the
pieces be similar to the whole, new
possibilities appear.
2/22/2013
7
1. Using the self-similarity (or self-affinity) of F, decompose F as F = F
1
... F
n
, where
each F
i
is a scaled copy of F.
2. For each piece F
i
, find an affine transformation T
i
for which T
i
(F) = F
i
. By "find an
affine transformation" we mean find the r, s, , , e, and f values.
Solving the inverse problem has only two steps:
(a) Trace the main features of the fractal and cut out
smaller copies of the tracing.
(b) To allow for reflections, flip the small copies and on the
back trace over the lines on the front. Label the front
image with a small F, to distinguish it from its reflection,
and to indicate the original orientation.
(c) Place the small copies, perhaps rotating or reflecting
them, to make a copy of the original fractal.
Examples
This fractal can be decomposed into three
pieces:
Note the top and bottom left pieces have
the same orientation as the entire fractal,
while the bottom right piece is reflected
across a vertical line
Keeping in mind that our transformation rules allow only reflections across the
x- and y-axes, some care must be taken with the translation after the reflection
- 0.5 0.5
0.5 0.5
0
0 0
0 1.0
0.0 0.5
0.0
This fractal can be decomposed into three
pieces:
Note the top and bottom left pieces have the same orientation as the entire
fractal, while the bottom right piece is rotated.
Keeping in mind that our transformation rules allow only rotations fixing the
origin, some care must be taken with the translation after the rotation
Find IFS rules to generate each of these fractals
r s e f
.333 .333 0 0 0 0
.333 .333 0 0 .667 0
.333 .333 0 0 0 .667
.333 .333 0 0 .667 .667
.333 .333 0 0 .333 .333
r s e f
.333 .333 0 0 0 0
.333 .333 0 0 .333 .333
.333 .333 0 0 .667 .667
.333 .333 0 0 0 .667
r s e f
.5 .5 0 0 0 0
.5 .5 0 0 .5 0
.5 .5 0 0 0 .5
.25 .25 0 0 .75 .75
r s e f
.5 .5 0 0 0 0
-.5 .5 0 0 1 0
-.5 .5 0 0 .5 .5
2/22/2013
8
r s e f
-.5 .5 90 90 .5 .5
.5 .5 180 180 1 .5
.5 .5 -90 -90 0 1
r s e f
.5 .5 -90 -90 0 .5
.5 .5 180 180 1 .5
.5 .5 0 0 .5 .5
When the pieces are not scaled by such obvious amounts, we can find scalings
and rotations by measuring distances and angles
Splitting the fractal into three pieces is not difficult.
Often it is convenient to specify a point as the origin of the coordinate system.
The lower left corner can be a good choice, but in general, use any symmetry
available, unless there is a compelling reason to do differently
First, find the scalings and any reflections, that is, the r and s values.
Second, Find the rotations.
Third, find the translations.
Here's the IFS table that
generated this fractal.
With a bit of thought, now we can find an IFS to generate the tree
First, it is easy to see the four main branches of the tree are scaled copies of the
whole tree.
The pieces have been pulled apart slightly to emphasize the decomposition.
The trunk is more complicated.
Simply shrinking the tree a lot horizontally works for the top of the trunk, but
makes the bottom of the trunk too thin.
Two shrunken copies of the tree are needed to make the trunk.
Here are the IFS rules, color coded to match each transformation to the
corresponding piece of the tree.
Here is the picture generated by the tree rules, leaving
out the second part of the trunk.
Common Mistakes in Finding Fractals:
Things that look like fractals but aren't
First, recalling that no physical fractal can exhibit scaling over
infinitely many levels, nevertheless to make a plausible claim
of fractality, a pattern must be repeated on at least a few
levels.
The "skulls within skulls" of Dali's Visage of War are repeated three
(maybe four) times.
In Dali's Visage of War (1940) note the eyes
and mouth each contain a face, whose eyes
and mouth each contain a face, many of whose
eyes and mouth each contain a face, an
obvious, if gruesome, example of fractals in art.
Dali thought of the Spanish Civil War, in 1940 a
source of frightening images to
him. Descharnes described the painting as
having "eyes filled with infinite death," referring
to the recursive effect set in motion by self-
similarity.
Note also the handprint in the lower right of the painting. This is Dali's handprint; this
is the only painting signed with his handprint, a testimony to the power this painting
had for Dali.
By contrast, two fists do not make a covincing Cantor set.
The decomposition of this picture into two pieces, the two fists, does not
continue to even one more level.
The fists are not split into smaller pieces. The more levels of the pattern,
the more convincing the fractality of the picture.
Here is an analogous example based on
the Sierpinski tetrahedron.
This is not plausibly fractal: it is a shape made of four
tetrahedra, but the tetrahedra have no substructure.
This is more believably fractal:
the structure has four levels of
substructure.
2/22/2013
9
Second, a repeating pattern alone is not sufficient to guarantee fractality.
A checkerboard or a brick wall has a repeating pattern, but the repetition
is with respect to translation, whereas for fractals the appropriate
transformation is magnification
Third, repeating a pattern under magnification is not sufficient to guarantee
fractality.
For example, a spiral is symmetric under magnification about its center
point, but about no other point.
Iterating this process, the limiting shape is just a single point.
In order to produce a fractal, at each level the decomposition must involve at
least two scaled copies.
Nested dolls are another example of a non-fractal involving a single scaling
transformation,
as is the cat bottle.
The bottle label includes a cat and a
bottle,
with a label that includes a cat and a
bottle,
with a lable that includes a cat and a
bottle,
and so on
The limit of this process is a single
point, not a fractal.
Inside the largest doll nestles a smaller doll,
inside that doll nestles a still smaller doll,
inside that doll nestles an even smaller doll,
and so on.
The limit of this process is a single point, not a fractal.
Both earrings have pictures
of the cow with two earrings,
both of which have pictures
of the cow with two earrings,
both of which have pictures
of the cow with two earrings,
and so on.
The limit of the cow pictures
is a Cantor set.
On the other hand, the cow picture is fractal, as would be more
obvious if the cow's left earring were turned toward us.
Here we study the random IFS algorithm, another way to render IFS images. This
includes a careful look at what random means.
To motivate the Random IFS algorithm, we begin with
the Chaos Game.
Here we observe the apparent effect of randomness is to
guarantee the points dance across the picture to be
generated.
Definition Movement, in random order, toward a collection of points can
fill in a polygon, or grow a fractal
The Chaos Game is played by specifying a number of vertices (a
1
, b
1
), (a
2
, b
2
),
..., and (a
N
, b
N
), and a scaling factor r < 1.
To play the game, start with the point (x
0
, y
0
) and pick one of the vertices,
say (a
i
, b
i
), randomly.
The point (x
1
, y
1
) is the fraction r of the distance between (a
i
, b
i
) and (x
0
, y
0
).
That is,
(x
1
, y
1
) = r(x
0
, y
0
) + (1 - r)(a
i
, b
i
)
For example, with four vertices, r = 1/3, and (a
2
, b
2
) is the first randomly
selected vertex, we obtain
(If r = 1, the point (x
1
, y
1
) is the same as the initial point (x
0
, y
0
); if r = 0, the
point (x
1
, y
1
) is the same as selected vertex (a
i
, b
i
).)
Now pick another vertex, (a
j
, b
j
), randomly.
The point (x
2
, y
2
) is given by
(x
2
, y
2
) = r(x
1
, y
1
) + (1 - r)(a
j
, b
j
)
and so on.
The Chaos Game Plot is the sequence of points (x
0
, y
0
), (x
1
, y
1
), ...
generated this way.
2/22/2013
10
For example of the Chaos Game, take four vertices, the
corners of the unit square, and take r = 1/2
(a
3
, b
3
) = (0, 1) (a
4
, b
4
) = (1, 1)
(a
1
, b
1
) = (0, 0) (a
2
, b
2
) = (1, 0)
Suppose the random number generator begins by selecting the vertices in
this order: 1, 3, 4, 3, 2.
See the first five points generated by this run of the chaos game.
If we continue, the points will fill in the square.
This should be plausible: we start with a point inside the
unit square, and each move is half-way between where
we are and a corner of the square, so we never leave the
square.
Because we select the corners randomly, no part of the
square is preferred over any other.
So since some parts of the square fill in, all parts must fill
in.
Do you believe this argument?
What would happen if we used just three vertices (a
1
, b
1
), (a
2
, b
2
), and (a
3
, b
3
)?
As with the square, we start with a point in the triangle. (In this example, it's on the
edge of the triangle, but that's still in the triangle.)
Each move is half-way between where we are and a corner of the triangle, so we
never leave the triangle.
Because we select the corners randomly, no part of the triangle is preferred over
any other.
So since some parts of the triangle fill in, all parts must fill in.
Thus played with three vertices of a triangle, the chaos game should fill in the
triangle. Right?
Here is the answer.
Here are more Chaos Game examples. Try to determine the shape
vertices the corners of a square, r = 1/3
move the top right vertex to the left, r = 1/2
five vertices, four the corners of a square,
one at the center of the square, r = 1/2
five vertices, four the corners of a square,
one at the center of the square, r = 1/3.
The chaos game often is used as an introduction to the more general Random IFS.
To illustrate its simplicity, frequently the chaos game is performed manually.
While this does convince of the simplicity of the chaos game, it is less effective in
showing the chaos game will generate fractals. For example, generating 30 points
manually requires some patience, but does the picture give much hint of a gasket?
The right picture, consisting of 300 points,
is much more convincingly a gasket.
No one would play the chaos game manually for 300 points, but this is not such
a problem: software can produce a 300 point chaos game in milliseconds (or
less).
This is to show that combining the results of 10 people generating 30 points
each produces as good a picture of the gasket as that of a single person
generating 300 points
Purpose To use Random IFS to illustrate an interesting property of random
sequences of numbers
Material Triangle template, about 12 overhead transparencies, a die (singular of
dice), adhesive tape, a ruler with cm scale, a permanent marking pen, an
overhead projector to display the data.
Here are 10 samples of 30 points each
2/22/2013
11
Below on the left is the aggregate of these 10 pictures; on the right is the picture of
300 iterates of one point. Both are reasonable representations of the gasket.
Among other things, the 30 point pictures illustrate two less-familiar
aspects of the chaos game.
(1) Manual chaos game experiments require a great deal of patience to
produce a recognizable image.
(2) The evident variability between short chaos game runs is
considerable.
Conclusion The superposition of several
short runs of the chaos game is at least
visually indistinguishable from a longer run.
The Random IFS Algorithm
Given IFS rules, the Deterministic Algorithm renders a picture of the
fractal by
1. applying all the rules to any (compact) initial picture,
2. then applying all the rules to the resulting picture,
3. and continuing this process.
Regardless of the starting shape, this sequence of pictures
converges to a unique limiting shape, the only (compact) set
invariant under simultaneous application of all the rules.
The Random Algorithm is another method of rendering the fractal determined by a given
set of rules, T
1
, ..., T
N
.
Definition and illustration of the random algorithm
Start with a point (x
0
, y
0
) belonging to the fractal, for example, take (x
0
,
y
0
) the fixed point of one of the T
i
.
A fixed point of a transformation T(x,y) is a point left unchanged by the
transformation. That is,
T(x, y) = (x, y)
Example 1 T(x, y) = (x/2, y/2) has fixed point (0, 0).
(x, y) = T(x, y) = (x/2, y/2). So x = x/2 and y = y/2, hence x = 0 and y = 0.
Example 2 T(x, y) = (x/2, y/2) + (1/2, 0) has fixed point (1, 0).
(x, y) = T(x, y) = (x/2, y/2) + (1/2, 0). So x = x/2 + 1/2 and y = y/2, hence x = 1 and y = 0.
If T(x,y) is a contraction, then it has exactly one fixed point.
A contraction is a transformation T that reduces the distance between every
pair of points.
That is, there is a number r < 1 with
dist(T(x, y), T(x', y')) rdist((x, y), (x', y'))
for all pairs of points (x, y) and (x', y').
Here dist denotes the Euclidean distance between points:
dist((x, y), (x', y')) = ((x - x')
2
+ (y - y')
2
)
1/2
The contraction factor of T is the smallest r satisfying
d(T(x, y), T(x', y')) rd((x, y), (x', y'))
for all pairs of points (x, y), (x', y').
In general, contractions can reduce distances between points by different amounts,
depending on the position of the points.
Here are some special kinds of contractions.
A similarity reduces all distances by the same number, r < 1. That is,
d(T(x, y), T(x', y')) = rd((x, y), (x', y'))
for all pairs of points (x, y), (x', y').
The transformation T(x, y) = (rx, ry) is an example; its contraction factor is r.
An affinity reduces distances by different amounts in different directions. For example,
T(x, y) = (rx, sy),
where both r < 1 and s < 1, and r and s are different.
If all the transformations of an IFS are contractions, then iterating the IFS is guaranteed to
converge to a unique shape.
Let {n
1
, n
2
, ... } be a random sequence of numbers, each from {1, ..., N}.
Generate a sequence of points
(x
1
, y
1
) = T
n1
(x
0
, y
0
),
(x
2
, y
2
) = T
n2
(x
1
, y
1
),
...
We shall see this sequence of points eventually will fill up the fractal to any
prescribed accuracy. For example, here are pictures of the Random Algorithm
applied to the gasket rules.
500 points
5000 points
(x
1
, y
1
) = r(x
0
, y
0
) + (1 - r)(a
i
, b
i
)
2/22/2013
12
This leads to the question What is random?
What makes an infinite sequence of digits random?
Are these random sequences?
1111111111111111111111111111...
1212121212121212121212121212...
1010010001000010000010000001...
1415926535897932384626433832...
What makes these
sequences nonrandom? We
can give short descriptions
that exactly specify the entire
sequence. This motivates our
definition of randomness.
This sequence is the first 28 digits in the decimal
expansion of pi.
While the digis of pi may pass many satistical tests for
randomness, we would not call pi random.
It is a specific number, the ratio of the circumference to
the diameter of any circle, always the same.
How do you know this is the first 28 decimal digits of pi?
If you didn't know the first 28 decimal digits of pi, would
you think this sequence is random?
Does our perception of randomness depend on how much
we know?
In case you're interested, here are the first 1000 decimal digits of pi
141592653589793238462643383279502884197169399375105820974
944592307816406286208998628034825342117067982148086513282
306647093844609550582231725359408128481117450284102701938
521105559644622948954930381964428810975665933446128475648
233786783165271201909145648566923460348610454326648213393
607260249141273724587006606315588174881520920962829254091
715364367892590360011330530548820466521384146951941511609
433057270365759591953092186117381932611793105118548074462
379962749567351885752724891227938183011949129833673362440
656643086021394946395224737190702179860943702770539217176
293176752384674818467669405132000568127145263560827785771
342757789609173637178721468440901224953430146549585371050
792279689258923542019956112129021960864034418159813629774
771309960518707211349999998372978049951059731732816096318
595024459455346908302642522308253344685035261931188171010
003137838752886587533208381420617177669147303598253490428
Motivated by our thoughts about these examples, we define an infinite
sequence of digits to be random if it cannot be specified completely in any way
shorter than listing the whole sequence.
How can we generate a random sequence?.
To make an infinite random sequence of 0s and 1s, toss a coin infinitely many
times.
Each time a heads comes up, put a 1 in the sequence. Each times tails comes
up, put a 0 in the sequence.
So if the coin toss started out
heads, heads, tails, heads, tails, tails, tails, heads
the random sequence would start
11010001
Of course, computers don't toss coins, but rather generate pseudorandom
numbers.
One method starts with the time on your computer's clock, multiplies by a large
number, divides by another large number, and takes the remainder.
Random sequences have this property:
Every infinite sequence of random numbers contains all finite sequences.
Suppose we have an infinite sequence of 0s and 1s, and nowhere in the sequence do we find
000.
By the definition given, this infinite sequence is not random. Why?
Suppose we're describing the sequence by listing all its terms.
Whenever we get to a 00 pair, we don't have to say what the next number is.
It MUST be 1, because otherwise the infinite sequence would contain 000.
Consequently, we don't have to list the entire infinite sequence to specify it completely. We say
only once that the sequence does not contain the triple 000, and then whenever the pair 00
occurs, we know the next number must be 1.
Similar arguments show that all finite sequences must occur somewhere (in fact, infinitely
often) in an infinite random sequence. If any one is missing, we can use this missing sequence
to describe the infinite sequence without listing all its entriety.
To understand why the Random and Deterministic algorithms generate the
same pictures, we also need to understand the notion of the address of parts
of a fractal.
Addresses are the main tool for relating fractals and
dynamics.
The order of the elements of an address is important, and
to some counterintuitive,
The notion of addresses is familiar in one dimension from the decimal expansion of
real numbers in the unit interval, [0, 1].
The left-most digit of the decimal expansion of x tells into which 10
th
of [0, 1] x falls.
The second digit tells into which hundredth - that is, which 10
th
of the 10
th
- x falls.
And so on. Here is an illustration
Addresses in fractals
To relate this to IFS, we need IFS rules to generate the unit interval. There are
(infinitely) many families of such rules, but for ease of interpretation with the
decimal expansion, we use
T
i
(x) = x/10 + i/10
for i = 0, ..., 9. Then
T
i
(I) is the i
th
10
th
,
T
i
T
j
(I) is the j
th
100
th
of the i
th
10
th
,
and so on. Note the order of the subscripts. This is the tricky part of
understanding addresses. We say
the digit i is the address of the i
th
10
th
,
the pair ij is the address of the j
th
100
th
of the i
th
10
th
,
and so on. Notice from left to right the address digits specify smaller intervals.
Addresses are unique.
2/22/2013
13
Addresses of a square
For concreteness in the two-dimensional case, we
consider the transformations
T
3
(x, y) = (x/2, y/2) + (0, 1/2) T
4
(x, y) = (x/2, y/2) + (1/2, 1/2)
T
1
(x, y) = (x/2, y/2) T
2
(x, y) = (x/2, y/2) + (1/2, 0)
These generate the filled-in unit square S. That is,
S = T
1
(S) T
2
(S) T
3
(S) T
4
(S),
with overlaps only along edges.
To each of the 1/2 1/2 squares T
i
(S) we associate the length 1 address i.
Each of these squares can be subdivided by iterating this decomposition process. For
example,
T
1
(S) = T
1
T
1
(S) T
1
T
2
(S) T
1
T
3
(S) T
1
T
4
(S).
To each of the 1/4 1/4 squares T
i
T
j
(S) we associate the length 2 address ij,
and so on.
In order of application, addresses are read right to left: the left-most digit is the
index of most recent transformation applied.
Because this seems confusing sometimes, we emphasize the order of
addresses is consistent with the order of composition of functions:
ij is the address of T
i
T
j
(S).
Another way to think of addresses is as relative coordinates. The 1/4 1/4
squares with addresses 11, 12, 13, and 14 are the 1, 2, 3, and 4 parts of 1.
T
3
(x, y) = (x/2, y/2) + (0, 1/2) T
4
(x, y) = (x/2, y/2) + (1/2, 1/2)
T
1
(x, y) = (x/2, y/2) T
2
(x, y) = (x/2, y/2) + (1/2, 0)
Longer addresses
Because each T
i
is a contraction, longer addresses specify smaller portions of S.
For example, here are the length 3 addresses for the square transformations.
T
3
(x, y) = (x/2, y/2) + (0, 1/2) T
4
(x, y) = (x/2, y/2) + (1/2, 1/2)
T
1
(x, y) = (x/2, y/2) T
2
(x, y) = (x/2, y/2) + (1/2, 0)
That longer addreses specify locations with greater accuracy is part of our
common experience. Let's abandon geometrical abstraction and turn to our own
sense of place.
Where are you?
At some fairly crude level, you are on the earth.
More precisely, you are in Asia, on the earth.
Still more precisely, you are in India, in Asia, on the earth
You are in Uttarakhand, in India, in Asia, on the earth.
You are in Roorkee, in Uttarakhand, in India, in Asia, on the earth.
You are at IIT Roorkee, in Roorkee, in Uttarakhand, in India, in Asia, on the earth.
You are in LHC 003, at IIT Roorkee, in Roorkee, in Uttarakhand, in India, in Asia, on the earth.
You are in the second row, third seat, in LHC 003, at IIT Roorkee, in Roorkee, in Uttarakhand, in
India, in Asia, on the earth.
This general kind of description has been familiar to us from childhood, so
we have known for years that a longer address specifies location more
precisely.
For fractals with symmetries, different rules can generate the same shape, but
with different addresses for the same region.
Addresses and Symmetries
Both IFS tables generate the equilateral Sierpinski gasket.
For the first table, it is not difficult to
see the length 2 addresses (left).
For the second table, the reflection in rule
2 has an effect on the addresses, for
example, the length 2 addresses
To see how address 21 winds up in the indicated position, start with the solid triangle S
This red triangle is T
2
(T
1
(S)),
hence has address 21
apply transformation T
1
,
obtaining T
1
(S).
apply T
2
, shrinking and
reflecting across the y-axis.
and translating by 1 horizontally.
2/22/2013
14
To understand why the Random and
Deterministic algorithms generate the
same pictures, we tried to understand the
notion of the address of parts of a fractal.
Now we can show why the Random and Deterministic algorithms
generate the same picture
First, fix a resolution, usually one pixel, to which the picture is to be
rendered.
Then we show
long enough addresses specify regions smaller than a
pixel,
that randomness guarantees all finite addresses are visited
by the points generated by the random IFS algorithm, and
that consequently every pixel of the attractor is visited.
Diameter goes to 0 under iteration of any contraction map
The diameter of a set is the maximum distance between any pair of
points in the set.
For example, the diameter of a circle is just the common notion of diameter; the
diameter of a square is the diagonal length of the square.
Some diameters
Because all the IFS rules are contractions, the diameter of a region of
address length N goes to 0 as N goes to infinity. We illustrate this with the
four transformations
T
3
(x, y) = (x/2, y/2) + (0, 1/2) T
4
(x, y) = (x/2, y/2) + (1/2, 1/2)
T
1
(x, y) = (x/2, y/2) T
2
(x, y) = (x/2, y/2) + (1/2, 0)
diam(S) = 2
diam(T
i
(S)) = (2)/2
diam(T
j
T
i
(S)) = (2)/4
and in general
diam(T
iN
...T
i1
(S)) = (2)/(2
N
)
Consequently, diam(T
iN
...T
i1
(S)) 0 as N .
how the Random IFS algorithm fills, to given resolution, the
attractor of the Deterministic IFS algorithm.
Because
diam(T
iN
...T
i1
(S)) = (2)/(2
N
),
if we take N large enough that
(2)/(2
N
) < resolution,
then the Random Algorithm will fill in the picture to the desired resolution if
all regions of address length N are visited.
Why should the Random Algorithm do this?
First, recall the Random Algorithm starts with the fixed point of one of the T
i
.
For definiteness, say we start with the fixed point (x
0
, y
0
) of T
1
.
The address of this fixed point is 111... = 1

.
Iteration and address shift: how iteration
affects the address of the points generated
by the Random IFS algorithm.
Observe
(x
0
, y
0
) = T
1
(x
0
, y
0
) = T
1
(T
1
(x
0
, y
0
)) = T
1
(T
1
(T
1
(x
0
, y
0
))) = ...
So (x
0
, y
0
) belongs to the square with address 1, the square with address 11, the
square with address 111, and so on.
Because the diameters of these squares are shrinking to 0, the address 111... =
1

corresponds to the single point(x

0
, y
0
).
Alternately, note that the only point left unchanged by repeated application of
T
1
is the point with address 111... = 1

Suppose the first transformation applied is T

i1
, the next T
i2
, and so on.
What is the effect of these transformations on the address of the point, and on
the address length N region in which the point lies?
point address of the point
address length N region
containing the point
(x
0
, y
0
) 1

1
N
(x
1
, y
1
) = T
i1
(x
0
, y
0
) i
1
111... = i
1
(1

) i
1
1
N-1
(x
2
, y
2
) = T
i2
(x
1
, y
1
) i
2
i
1
(1

) i
2
i
1
(1
N-2
)
(x
3
, y
3
) = T
i3
(x
2
, y
2
) i
3
i
2
i
1
(1

) i
3
i
2
i
1
(1
N-3
)
... ... ...
(x
N
, y
N
) = T
iN
(x
N-1
, y
N-1
) i
N
...i
3
i
2
i
1
(1

) i
N
...i
3
i
2
i
1
(x
N+1
, y
N+1
) = T
iN+1
(x
N
, y
N
) i
N+1
i
N
...i
3
i
2
i
1
(1

) i
N+1
i
N
...i
3
i
2
... ... ...
So we see each new transformation applied has this effect on the N-digit
address: discard the right-most digit, shift the remaining N-1 digits one place to
the right, and insert the new address on the left.
If the transformations are applied randomly, then eventually
every finite sequence of transformations will occur,
and in particular, every sequence of length N will be applied.
Consequently,
every region with address length N will be visited by the (x
ik
, y
ik
).
To the specified resolution, the Random Algorithm will generate the same picture
as the Deterministic Algorithm.
2/22/2013
15
For example, suppose we specify the resolution corresponding to addresses of length N =
3 and we start with the point (x
0
, y
0
) with address 1
infinity
.
To the specified resolution, (x
0
, y
0
) lies in the region with address 111.
If T
2
is the first transformation applied, then resulting point (x
1
, y
1
) = T
2
(x
0
, y
0
) lies in the
region with address 211.
If T
3
is the next transformation applied, then resulting point (x
2
, y
2
) = T
3
(x
1
, y
1
) lies in the
region with address 321.
If T
4
is the next transformation applied, then resulting point (x
3
, y
3
) = T
4
(x
2
, y
2
) lies in the
region with address 432.
Continuing will fill in all the 4
3
regions of address length 3.
Example
Probability and the Random IFS Algorithm
In the Random IFS Algorithm the transformations T
i
are applied in random order, but
they need not be applied equally often.
Associated with each T
i
is a probability p
i
, 0 < p
i
< 1, representing how often each
transformation is applied. That is,
when N points are generated, each T
i
is applied about Np
i
times.
To illustrate the effect of changing the probabilities, we use the IFS
T
3
(x, y) = (x/2, y/2) + (0, 1/2) T
4
(x, y) = (x/2, y/2) + (1/2, 1/2)
T
1
(x, y) = (x/2, y/2) T
2
(x, y) = (x/2, y/2) + (1/2, 0)
We take
p
4
to range from 0 to 1 in steps of .05,
and
p
1
= p
2
= p
3
= (1 - p
4
)/3.
Starting with p
4
= 0, the first picture is the gasket. Do you see why?
Very roughly speaking, the probability p
i
of applying T
i
is the fraction of the
whole attractor A occupied by T
i
(A).
The area contraction factor of T
i
is
A
i
= r
i
s
i
(cos(
i
)cos(
i
) + sin(
i
)sin(
i
))
(This is just the determinant of the matrix formulation of T
i
.)
Note that if
i
=
i
, the area contraction factor simplifies to
A
i
= r
i
s
i
Then the probability is given by
p
i
= A
i
/(A
1
+ ... + A
N
)
Here is a way to find the probabilities that give approximately uniform fill of the
attractor.
Driven IFS
What happens if we use a non-random sequence in the random IFS algorithm?
In particular, can we run the random IFS algorithm with a sequence of data -
daily closing prices of a stock, or the intervals between your heartbeats, for
example?
Will patterns in the IFS picture reveal patterns in the data?
Because we use a data sequence to select the order in which the
transformations are applied, we call this approach driven IFS. The
data drive the order in which the IFS rules are applied.
Referring to the chaos game, Ian Stewart asked "What
happens if the computer's random number generator is
replaced by some more familiar dynamical system?"
For most of his tests, Stewart used the chaos game fixed points the
vertices (0, 0), ((3)/2, 1/2), and (0, 1) of an equilateral triangle. The
corresponding IFS rules are
Stewart's experiments
T
1
(x, y) = (x/2, y/2)
T
2
(x, y) = (x/2, y/2) + (0, 1/2)
T
3
(x, y) = (x/2, y/2) + ((3)/4, 1/4)
For reference, a random number generator (1000 points) gives
2/22/2013
16
is produced by iterating the logistic map x
n+1
= 4x
n
(1 - x
n
).
is produced from sin(t) + sin(t2).
t = 1, 2, ..., 1000.
is produced from sin(t) + sin(t3) + sin(t5)
Pictures are
generated
sequences of
numbers, coarse-
grained into three
equal-size bins.
Coarse-Graining the Data
We turn the data y
1
, y
2
, ..., y
N
into a sequence i
1
, i
2
, ..., i
N
of 1s, 2s, 3s, and 4s.
This sequence is called the symbol string associated with the data.
The data values y
i
often are measured as decimals and because we are
converting these to only four values, the process of turning the y
k
into i
k
is
called coarse-graining.
The range of y values for corresponding to a symbol is the bin of that symbol.
equal-size bins Divide the range of values into four intervals of equal length.
equal weight bins Arrange the bin boundaries so (approximately) the same number of points lie in each
bin.
zero-centered bins For data whose sign is important, take 0 as the boundary between bins 2 and 3; place
the other boundaries symmetrically above and below 0. Unlike the first two cases, this is a family of
coarse-grainings depending on the placements of the other two bin boundaries.
mean-centered bins Take the mean of the data to be the boundary between bins 2 and 3; place the other
boundaries symmetrically above and below the mean, usually expressed as a multiple of the standard
deviation.
median-centered bins Take the median of the data to be the boundary between bins 2 and 3; place the
other boundaries symmetrically above and below the median, usually expressed as a multiple of the
range. Note the equal-weight bins are a special case of this.
Though there are others, we use five kinds of coarse-graining:
To illustrate the different kinds of coarse-graining, we use a
data set consisting of successive differences of 1000 numbers
generated by iterating the logistic map.
Here is the time series y
1
, y
2
, ..., y
1000
generated by 1000 iterates of the logistic
map with equal-size bin lines drawn, and the corresponding driven IFS
Here is the time series y
2
- y
1
, y
3
- y
2
, ..., y
1000
- y
999
generated by successive
differences of 1000 iterates of the logistic map with equal-size bin lines drawn,
and the corresponding driven IFS
Here is the time series y
1
, y
2
, ..., y
1000
generated by 1000 iterates of the logistic
map with equal-weight bin lines drawn, and the corresponding driven IFS
Here is the time series y
2
- y
1
, y
3
- y
2
, ..., y
1000
- y
999
generated by successive
differences of 1000 iterates of the logistic map with equal-weight bin lines
drawn, and the corresponding driven IFS
Driven IFS Rules (Square functions)
T
3
(x, y) = (x/2, y/2) + (0, 1/2) T
4
(x, y) = (x/2, y/2) + (1/2, 1/2)
T
1
(x, y) = (x/2, y/2) T
2
(x, y) = (x/2, y/2) + (1/2, 0)
This IFS generates the filled-in unit square. Consequently, any departure from
uniform randomness will be visible through departures from uniform fill of the
square.
Here are some examples, all with 10000 points.
uniform random
sequence
p
1
= p
4
= 0.1;
p
2
= p
3
= 0.4
p
1
= 0.1;
p
2
= p
3
= p
4
= 0.3
p
1
= p
3
= 0.1;
p
2
= p
4
= 0.4
2/22/2013
17
IFS Driven by DNA Sequences
An example first explored by H. Joel Jeffrey. The genetic code is written in an
alphabet of four characters: C, A, T, and G.
A sequence of several billion of these makes each of us. A sequence of 3957
symbols is needed to encode the formation of the enzyme amylase.
T G A A T T C A A G T T T G G T G C A A A A C T T G G C A C A G T T A T C C G
C A A G T G G A A T G G A G A G A A G A T G T C C T A T T T A A A G T A A A T A
T A T A C G A T T T T G T C A T T T G T T C T G T C A T A C A T C T G T T G T C
A T T T T C T T A A A T A T T G T A A C T T A A A T T G T T G A T T A T T A G T T
A G G C T T A T T G T T C A T T T A T C C T T A A T T A A T T A T G T T T T T C A
T T T G A T A C A T C A G T C A C C T G A T A A C A G C T G A A A T C T A A A G
T A T C A C T T A G T G A G T T T T G T T G G G T T G T G T T A A G T C C A T
T A G A G T C T A A G A A T G T T T G C T T A T G G C C T A C T A A A A A T A T
G G T A G C A T C C T A A G A A T A G T T A T A C T A A A A A G T G A T C C C T
A T A A T A T G A C T A C A C T A G G G A A T T T A T T T A T G C T A C A T T A
G G G A A T T A A T T G A A A T T T A A A A G T G A A T G T A A A A G C A G A G
T T A T A A A T T A A T T T C C A T T C T G T A T T A T A T A A C A T G G A T G T
C T T A A T T C T C A A G T C C A T A A T G T T A C A A T A A A A T T T T A A A A
A T C T A A A A T A A A A T C A A A A C A A A A G A T T A T A T A G T A A A A C C
T A A A T C T G G A T A A A A T T C C C A T G T A T G T T A T C A T A G A A A A T
T G T A C A T A T G T G T A C A T A T A C A A T.
How can we convert a DNA sequence into an IFS
picture?
Read the sequence in order, and
apply T
1
whenever C is encountered,
apply T
2
whenever A is encountered,
apply T
3
whenever T is encountered, and
apply T
4
whenever G is encountered,
Of course, any assignment of
transformations to C, A, T, and
G can be used.
On the left is the picture that results from applying these rules to the amylase
sequence. Note there are very few points in the region with address GA.
(Remember the ordering of the addresses.)
On the right is a picture that results when 3957 points are generated randomly,
except that T
4
never immediately follows T
1
.
Not an exact match, but certainly suggestive.
some more examples of IFS driven by DNA sequences.
First, two examples from ion channels in human cardiac cells.
potassium channel sodium channel
Here are four examples from chromosome 7 from the University of
Washington human genome site http://www.genome.washington.edu/UWGC/
g1346a094 g0771a003 G1564A209 gap3
The interpretations of these pictures is still being investigated. Stay tuned.
Cyclic Driven IFS
What happens if we drive an IFS in a
repeating pattern?
Here we will learn to recognize the visual signature of IFS driven by
cyclic data, that is, numbers that repeat a particular pattern.
The simplest repeated sequence is constant, just repeat the same
number, for example, 11111... = 1

.
Starting with (1/2, 1/2), applying T
1
(x, y) = (x/2, y/2) repeatedly produces the
sequence of points
T
1
(1/2, 1/2) = (1/4, 1/4),
T
1
(1/4, 1/4) = (1/8, 1/8),
T
1
(1/8, 1/8) = (1/16, 1/16),
...
converging to the lower left corner of the unit square.
Because it is gotten by applying T
1
infinitely many times, the address of this point
is 11111... = 1

.
T
1
(x, y) = (x/2, y/2) = ((x+0)/2, (y+0)/2), the midpoint of (x, y) and (0, 0)
T
2
(x, y) = (x/2, y/2) + (1/2, 0) = ((x+1)/2, (y+0)/2), the midpoint of (x, y) and (1, 0)
T
3
(x, y) = (x/2, y/2) + (0, 1/2) = ((x+0)/2, (y+1)/2), the midpoint of (x, y) and (0, 1)
T
4
(x, y) = (x/2, y/2) + (1/2, 1/2) = ((x+1)/2, (y+1)/2), the midpoint of (x, y) and (1, 1)
Recall
we see the cycles 2

, 3

, and 4

generate sequences of points that converge

to (1, 0), (0, 1), and (1,1), respectively
Constant sequences generate sequences of points that converge to
the fixed point of the corresponding transformation.
Constant sequences generate sequences of points that converge to
the fixed point of the corresponding transformation.
The sequence of points generated by applying T
1
repeatedly converges to the point with
address 1111... = 1

.
We show this is the fixed point of T
1
, and find its coordinates.
Fixed point. Say (x*, y*) is the point with address 1

. Then
T
1
(x*, y*) has address 1(1

) = 1

.
Because T
1
(x*, y*) and (x*, y*) have the same (infinite) address, they must be the same point.
That is,
T
1
(x*, y*) = (x*, y*)
and (x*, y*) is the fixed point of T
1
.
Coordinates. We see
(x*, y*) = T
1
(x*, y*) = (x*/2, y*/2),
and so (x*, y*) = (0, 0).
Similar arguments show 2

, 3

, and 4

are the fixed points of T

2
, T
3
, and T
4
, respectively.
These points have coordinates (1, 0), (0, 1), and (1, 1), respectively. For example,
(x*, y*) = T
2
(x*, y*) = ((x* + 1)/2, y*/2).
So x* = x*/2 + 1/2 and y* = y*/2, so x* = 1 and y* = 0.
2/22/2013
18
The next simplest repeated sequence is a 2-cycle
it alternates between two values.
The next simplest repeated sequence is 121212... = (12)

.
Applying T
1
and T
2
alternately produces a sequence of points converging to two points
along the x-axis
Some more sequences of points generated by other 2-cycles.
(13)

(14)

limit points of (14)

What cycle is this?

2-Cycle Addresses
The limiting points in the last example [121212... ] have addresses (12)

and (21)

. To
see this, recall the relation between the address and the order in which the transformations
are applied. The sequence 121212... gives points in regions with addresses
1
21
121
2121
12121
212121
and so on.
Alternate
entries in the
sequence are
1
121
12121
1212121
...
(12)

And
21
2121
212121
21212121
...
(21)

2-cycle Coordinates
The limiting points in the last example have coordinates (1/3, 0) and (2/3, 0).
To see this, say (x
1
, y
1
) is the point with address (12)

and (x
2
, y
2
) is the point with
address (21)

. Then notice
T
2
(x
1
, y
1
) has address 2(12)

=
2(12)(12)(12)(12)... = (21)(21)(21)(21)... = (21)

Because T
2
(x
1
, y
1
) and (x
2
, y
2
) have the same (infinite) address,
T
2
(x
1
, y
1
) = (x
2
, y
2
)
By a similar argument,
T
1
(x
2
, y
2
) = (x
1
, y
1
).
combining these two, we see
T
1
T
2
(x
1
, y
1
) = (x
1
, y
1
)
and
T
2
T
1
(x
2
, y
2
) = (x
2
, y
2
).
From the first we obtain
(x
1
, y
1
) = T
1
T
2
(x
1
, y
1
) = T
1
(x
1
/2 + 1/2, y
1
/2) = (x
1
/4 + 1/4, y
1
/4),
so
x
1
= x
1
/4 + 1/4 and y
1
= y
1
/4
Solving for x
1
and y
1
, we obtain (x
1
, y
1
) = (1/3, 0). A similar argument gives (x
2
, y
2
) = (2/3, 0).
Choice of Initial Point
How do the limiting points depend on the choice of (0.5, 0.5) as starting point?
Not at all.
This should be no surprise: recall the attractor of an IFS does not depend on
the initial picture. (This is a bit different than the situation at hand, however, but
certainly makes the result plausible.)
Here is an illustration. We follow the sequences generated from three different
starting points, all with vertices selected in the order 121212... = (12)

. Blue
lines for one starting point, green for the second, red for the third.
3-Cycles
The repeating pattern
T
1
followed by T
2
followed by T
3
produces a sequence converging to the three points with addresses
(321)

, (132)

, and (213)

.
the three points to which these
sequences converge, along
with their addresses
2/22/2013
19
3-Cycle Addresses
The point with address (321)

is the fixed point of T

3
T
2
T
1
.
To see this, say (x, y) is the point with address (321)

. Then
T
1
(x, y) has address 1(321)

,
T
2
T
1
(x, y) has address 21(321)

, and
T
3
T
2
T
1
(x, y) has address 321(321)
infinity
= (321)

Because (x, y) and T

3
T
2
T
1
(x, y) have the same (infinite) addresses, they are the
same point.
Similarly,
the point with address (132)

is the fixed point of T

1
T
3
T
2
, and
the point with address (213)

is the fixed point of T

2
T
1
T
3
.
3-Cycle Coordinates
Solving the fixed point equations gives the coordinates of each
point. For example,
(x, y) = T
3
T
2
T
1
(x, y) = T
3
T
2
(x/2, y/2)
= T
3
(x/4 + 1/2, y/4) = (x/8 + 1/4, y/8 + 1/2)
so
x = x/8 + 1/4 and y = y/8 + 1/2
Solving for x and y gives (x, y) = (2/7, 4/7)
Similar calculations show
(1/7, 2/7) is the fixed point of T
1
T
3
T
2
and
(4/7, 1/7) is the fixed point of T
2
T
1
T
3
3-Cycle in a different order What happens if we apply the 3-cycle
transformations in a different order?
If we repeat the pattern
T
3
followed by T
1
followed by T
2
instead of
T
1
followed by T
2
followed by T
3
?
Limiting points
Same as produced
by this
To understand why this is so, write the first several terms of both sequences
123123123123123123123123123123...
312312312312312312312312312312...
The second sequence is just the first shifted two terms to the left.
123123123123123123123123123123...
312312312312312312312312312312...
The first sequence is the same as the second, but starting from
T
2
T
1
(0.5, 0.5) instead of from (0.5, 0.5).
As with the fractals generated by regular IFS, here the final pattern does not
depend on the starting point. We prefer to start with (0.5, 0.5) because this
is the most neutral choice.
Can you find an order of cycling through 1,2, and 3 that converges
to a different triple of points?
The sequence
T
2
followed by T
1
followed by T
3
?
produces sequences of points converging to the points
(1/7, 4/7), (4/7, 2/7), and (2/7, 1/7)
shown below on the left.
For comparison, the previous 3-cycle is shown on the right.
The same points are produced by any cyclic permutation of the original sequence.
For example, the cyclic permutations of
T
2
followed by T
1
followed by T
3
,
are
T
1
followed by T
3
followed by T
2
, and T
3
followed by T
2
followed by T
1
.
N-Cycles
The number of points to which the sequence converges is the
length of the (shortest) repeated pattern.
The reason for shortest is that the
sequences (123)

and (123123)

converge to the same three

points.
2/22/2013
20
Driven IFS with Forbidden Combinations
Gaps in the driven IFS picture indicate combinations
of transformations that do not occur.
To understand the effect of forbidden combinations, we use
the addresses of regions of the unit square S.
We begin with three simple examples.
Forbid a transformation: T
4
never occurs
If T
4
is never applied, then no points land in the subsquare S
4
with
address 4. That is, the upper right subsquare is completely empty.
Because the subsquare S
4
contains no points, the subsquares
T
1
(S
4
) = S
14
, T
2
(S
4
) = S
24
, and T
3
(S
4
) = S
34
contain no points.
Continuing, the subsquares T
1
(S
14
) = S
114
, T
1
(S
24
) = S
124
, T
1
(S
34
) = S
134
, ...,
and T
3
(S
34
) = S
334
contain no points. That is,
if i=4, j=4, or k=4, the subsquare S
ijk
contains no points (below left).
Similarly,
the subsquare S
ijkm
contains no points if any of i, j, k, or m is 4 (below right).
The result of continuing this process is clear:
if T
4
is never applied, every square whose address contains a 4 is empty.
With this restrction, we see the IFS generates a right isosceles Sierpinski gasket.
This is no surprise, because the IFS {T
1
, T
2
, T
3
} generates a right isosceles
Sierpinski gasket.
Forbid a pair: T
4
never immediately follows T
1
If T
4
never immediately follows T
1
, then no points land in the square T
4
(S
1
) = S
41
.
Because S
41
contains no points, the subsquares T
1
(S
41
) = S
141
, T
2
(S
41
) = S
241
, T
3
(S
41
) = S
341
,
and T
4
(S
41
) = S
441
contain no points
Continuing, here are the pictures showing the subsquares of address length 4
(left) and 5 (right) containing no points.
The pattern should be clear:
no points lie in any subsquare with address containing 41.
Another Example of forbidden pair
We impose these restrictions:
T
1
never immediately follows T
4
T
2
never immediately follows T
3
T
3
never immediately follows T
2
T
4
never immediately follows T
1
What are the first few generations of the IFS with these restrictions?
First, note these restrictions imply no points land in the squares with addresses
14, 23, 32, and 41. That is, the shaded squares will contain no points
Continuing on to the length 3 address squares, we see
every square with address including 14, 23, 32, and 41 will
contain no points. For example,
Here is a movie showing the first few iterates of the driven IFS with these
restrictions. In contrast to the diagrams above, here the nonempty regions
are shaded
These examples illustrate the general situation: if some combination of T
in
...T
i1
of
transformations never occurs, then all subsquares with addresses containing the
sequence i
n
...i
1
must be empty.
The transformation combinations that never occur are called forbidden
combinations. For example, if T
4
never immediately follows T
1
, we say 41 is
a forbidden pair.
Graphical Representation
For those Driven IFS determined completely by forbidden pairs, a compact
representation of the IFS can be given by a graph showing the allowed pairs.
The graph has four vertices, one for each T
i
,
and an edge from vertex i to vertex j if T
i
can be be followed immediately by T
j
.
For example, the Driven IFS with the single forbidden pair 41 has this graph:
Because T
1
cannot be immediately followed by T
4
, the graph has no arrow
from vertex 1 to vertex 4.
All other combinations are allowed, so all other pairs of edges are connected
by arrows.
2/22/2013
21
Examples of Driven IFS with Forbidden Pairs
Different combinations of forbidden pairs can generate some very interesting
pictures.
Finally we move on to driving IFS by numerical data series
Not all data are written in an alphabet of four symbols. Can we adapt the driven
IFS method to study a wider variety of data, daily closing stock prices, for
example? The answer is yes. As a first step, we must coarse-grain the data.
Driven IFS and Data Analysis
Coarse-Graining the Data
We turn the data y
1
, y
2
, ..., y
N
into a sequence i
1
, i
2
, ..., i
N
of 1s, 2s, 3s, and 4s,
called the symbol string associated with the data.
Often the data values y
i
are measured as decimals and because we are
converting these to only four values, the process of turning the y
k
into i
k
is
called coarse-graining.
The range of y values for corresponding to a symbol is the bin of that symbol.
Though there are others, we use five kinds of coarse-graining:
equal-size bins Divide the range of values into four intervals of equal length.
equal weight bins Arrange the bin boundaries so (approximately) the same number of points lie in each bin.
zero-centered bins For data whose sign is important, take 0 as the boundary between bins 2 and 3; place the other
boundaries symmetrically above and below 0. Unlike the first two cases, this is a family of coarse-grainings depending
on the placements of the other two bin boundaries.
mean-centered bins Take the mean of the data to be the boundary between bins 2 and 3; place the other boundaries
symmetrically above and below the mean, usually expressed as a multiple of the standard deviation.
median-centered bins Take the median of the data to be the boundary between bins 2 and 3; place the other
boundaries symmetrically above and below the median, usually expressed as a multiple of the range. Note the equal-
weight bins are a special case of this.
Equal-Size Bins
One method of converting the measured data y
1
, y
2
, ..., y
N
into a symbol string i
1
, i
2
,
..., i
N
is first to find the range of the values, that is, the interval between the maximum
(max) and the minimun (min) of the y
i
.
Next, divide the range (min, max) into four equal-size bins:
y
k
lies in bin
1
if min y
k
< min + .25(max - min)
y
k
lies in bin
2
if min + .25(max - min) y
k
< min + .50(max - min)
y
k
lies in bin
3
if min + .50(max - min) y
k
< min + .75(max - min)
y
k
lies in bin
4
if min + .75(max - min) y
k
max
The numbers separating the bins are called bin boundaries, B
1
, B
2
, and B
3
:
B
3
= min + .75(max - min)
B
2
= min + .50(max - min)
B
1
= min + .25(max - min)
Each y
k
lies in one of these bins, so we can
convert the sequence y
1
, y
2
, y
3
... into the
symbol string i
1
, i
2
, i
3
... of 1s, 2s, 3s, and
4s associated with the data.
i
k
= 4 if y
k
lies in bin
4
i
k
= 3 if y
k
lies in bin
3
i
k
= 2 if y
k
lies in bin
2
i
k
= 1 if y
k
lies in bin
1
2/22/2013
22
Equal Weight Bins
Another method of converting the measured data y
1
, y
2
, ..., y
N
into a symbol
string i
1
, i
2
, ..., i
N
is to select the bin boundaries B
1
, B
2
, and B
3
so
one quarter of the y
k
satisfy y
k
B
1
,
one quarter of the y
k
satisfy B
1
< y
k
B
2
,
one quarter of the y
k
satisfy B
2
< y
k
B
3
, and
one quarter of the y
k
satisfy B
3
< y
k
.
Because each bin contains the same number of points, this is called an equal-
weight binning.
One way to produce an equal-weight binning is to sort the list of y
k
in increasing
order. Then take
B
3
the element one-quarters of the way from the top of the sorted list,
B
2
the element of one-half of the way from the top of the sorted list, and
B
1
to be the element three-quarter of the way from the top of the sorted list.
Equal-weight bins can be called
a maximum entropy partition.
Zero-Centered Bins
While there is no reason to think 0 holds any special importance for the data set
y
1
, y
2
, ..., y
N
,
it may for the first differences
y
2
- y
1
, y
3
- y
2
, ..., y
N
- y
N-1
.
For instance, a positive first difference means the data values are increasing, a negative
first difference means the data values are decreasing. This has clear significance for
financial data.
Take B
2
= 0, and in the absence of other information, take B
1
and B
3
symmetrically
spaced about B
2
= 0.
Usually express B
1
and B
3
as a fraction of the range of the differences.
Mean-Centered Bins
Another way to bin the data is to set B
2
= m, the mean of the data values, and set B
1
and
B
3
to some fraction of or multiple of the standard deviation.
Median-Centered Bins
Another way to bin the data is to set B
2
= m, the median of the data values, and set
B
1
and B
3
to some fraction of the range.
Comparing Stocks and Indices by Driven IFS
In his project, Joseph Thornton investigated some variants on driving IFS by
financial data.
He use daily closing prices for a two year period, about 500 data points.
Daily Percentage Changes
First we look at two years' data for several stocks, using daily percentage
changes to set the bin boundaries. Observing that the market tends to
fluctuate most often between +2.5%and -2.5%, as a first experiment
Thornton looked at daily changes in stock prices and used zero-centered
bins with these bin boundaries
B
3
: the price increases by > 2.5% its previous value
B
2
: the price remains unchanged
B
1
: the price decreases by < -2.5% its previous value
Citigroup
American
International
General Electric
Dell Sonus Networks Qwest
Tyson Foods
Colgate-
Palmolive
Lucent
Here are the driven IFS
Note that for Citigroup, AI, GE, Tyson, and Colgate-Palmolive, the predominant feature is strong
motion along the 2-3 diagonal, indcating most changes are within 2.5% of the previous price.
Another feature, especially strong in the GE graph, is the absence of points along the 1-
4 diagonal. Although there are points in address 1 and 4 (even in address 11 and 44), there are
very few consecutive moves of > +2.5%followed by < -2.5%, and vice versa.
Contrast this with Sonus, where most of the activity in along the 1-4 diagonal, indicating relatively
wild swings in closing price. The heavy cluster of points in corner 1 does not speak of a successful
stock.
Roughly speaking, the older economy companies - Citigroup, Tyson, Colgate-Palmolive, GE, and
AIG - have stronger 2-3 diagonals, indicating less volatile behavior.
2/22/2013
23
Next we experiment with uniformizing the driven IFS by scaling the bin
boundaries with the stock's value.
scaling
The factor of a stock is the volatility of the stock relative to that of the market.
> 1 means the stock is more volatile than the market.
= 1 means the stock and the market are equally volatile.
< 1 means the stock is less volatile than the market.
The 1-4 diagonals of the tech stocks reflects larger daily percentage changes, so
we would expect higher volatility.
As a quantitative test of this, Thornton scaled the bin boundaries with each stock's
. For example, Qwest has = 2.15, so the first and third bin boundaries are set
at2.15x2.5% = 5.38%above and below 0.
Here are the rescaled driven IFS (left), each grouped with its original (right) for
comparison.
Citigroup: = 1.33
American International: = 0.85
General Electric: = 1.08
Dell: = 1.83
Sonus: = 4.95
Tyson: = 0.47
Note that in general rescaling the bin boundaries by the stock's makes the
driven IFS look much alike. Note particularly the change in the Sonus IFS.
The obvious exception is Tyson, whose is so small that the rescaling puts many
more points into bins 1 and 4.
IFS Driven by Texts
Driven IFS is a potentially interesting tool for analyzing patterns in data, but we
need to be careful with how the data are binned.
A text is a string, but in an alphabet of more than four symbols. How can we
convert this into a string in an alphabet of four symbols?
One possibility is to treat words as the fundamental units
of the text, and assign bins by parts of speech. This has
an obvious problem: distinguishing how much of the
driven IFS structure is due to the author's style, and how
much to grammatical constraints.
Another choice is to treat letters as the fundamental
units and assign bins by ignoring some letters, or grouping
the letters together. Of course, any choices must be
justified in a way reflecting properties of the text. This is
not an easy problem.
Phonological analysis
As we have seen, suitable ways of binning texts are elusive.
Binning by phonemes, individual sounds making up the words of a spoken language, is a
promising direction, especially because phomemes can be divided into a small collection of
categories.
In her spring 2003 project, Emily Runde studied phonological properties in two of T. S. Eliot's
poems, The Love Song of J. Alfred Prufrock and Hollow Men.
Do these poems exhibit similar phonological patterns? What about patterns in Lewis
Carroll's Jabberwocky, which contains many fabricated words?
Will Elliot's texts, Tradition and the Individual Talent for example, reveal different patterns?
Category Examples Description
vowels a, e no vocal tract friction, most
sonorous
glides y, w consonantal forms of vowels i
and u, slightly less sonorous
liquids l, r friction caused by the tongue
nasals m, n, ng slightly more friction,
articulated through the nose
obstruents all other consonants varying degrees of vocal tract
constriction
fricatives s, f, th partial vocal tract constriction
plosives p, t, k air buildup with complete
vocal tract closure, then
released in a short burst
affricatives ch, dg fricative and plosive sound
combined
syllabic boundary
word boundary word boundaries supersede
syllable boundaries
Here are the phonological categories used.
2/22/2013
24
First she analyzed these pieces using a four bin IFS with the familiar transformations
Love Song of J. Alfred Prufrock
Hollow Men
Jabberwocky Tradition and the Individual Talent
The order of the transformations is determined by the order in which
the phonological categories occur in the text, according to this correspondence.
T
3
: liquids and nasals T
4
: all obstruents
T
1
: vowels T
2
: glides
Some analysis
In all four the squares 22, 32, and 42 are empty, indicating only a vowel can follow a glide.
In all four the squares 111, 234, 334, and 434 are empty.
111: three successive vowels do not occur.
234, 334, 434: a vowel must follow the combination obstruent then liquid or nasal.
Prufrock, Jabberwock, and Tradition all have 212 empty, while Hollow does not.
Consequently, glide-vowel-glide is possible, but uncommon, at least among these samples.
Tradition has 433 empty, while Prufrock, Jabberwock, and Hollow do not.
(In Jabberwock why does the point on the boundary between 433 and 344 belong to 433?) Is
this combination a phonological construction that distinguishes poetry from prose?
Next she analyzed these pieces using a nine bin IFS.
Love Song of J. Alfred Prufrock
Hollow Men
Jabberwocky
Tradition and the Individual Talent
T
7
:
plosives
T
8
: syllabic
boundaries
T
9
: word
boundaries
T
4
: nasals
T
5
:
fricatives
T
6
:
affricatives
T
1
: vowels T
2
: glides T
3
: liquids
Some analysis
In all four the squares 22, 32, 42, 52, 62, 72, 82, and 92 are empty: only a vowel can follow a glide.
In all four, 23 and 26 are empty: a glide cannot follow a liquid or an affricate.
In all four, among *8 (for * = 1, ..., 7) the most filled in are 58 and 78: syllables begin with fricatives or
plosives more often than with other phonemes.
Among *9 the most filled are 19, 59, and 79: words begin with vowels as often as with fricatives or plosives.
In all four, among 8* the most filled is 81: syllables are more likely to end with vowels.
The common sonority profile is an arc: often a syllable begins with a low sonority phoneme, followed by one
of higher sonority, then lower.
831 is more densely filled than 871: after a vowel a syllable is more likely to end in a nasal than in a plosive.
781 is densely filled: after a syllable ends with a vowel the next syllable can start with a plosive.
The rules governing nasals and liquids are similar: in each of the four plots, the pattern in square 3 is similar
to that in square 4.
In each of the four plots, the pattern in 5 is similar to that in 7 (6 also is similar, though much less filled),
indicating that the three subcategories of obstruents behave similarly.
Although 57 and 75 are fairly populated, 55 and 77 are not, and in general the pairs ii are (nearly) indicating
that only rarely do two phomemes of the same phonemic category occur adjacently within a syllable.
Finally, the patterns in 8 and 9 are similar, thought 9 is more filled, indicating many one-syllable words.
Although the sonority profile refers to a syllable rather than to a word, this shows the same pattern of sonority
near a syllable boundary is repeated at word boundaries.
Nine Bin IFS Rules
T
7
(x,y) = (x/3, y/3) + (0, 2/3) T
8
(x,y) = (x/3, y/3) + (1/3, 2/3) T
9
(x,y) = (x/3, y/3) + (2/3, 2/3)
T
4
(x,y) = (x/3, y/3) + (0, 1/3) T
5
(x,y) = (x/3, y/3) + (1/3, 1/3) T
6
(x,y) = (x/3, y/3) + (1/3, 2/3)
T
1
(x,y) = (x/3, y/3) T
2
(x,y) = (x/3, y/3) + (1/3, 0) T
3
(x,y) = (x/3, y/3) + (2/3, 0)
These transformations generate the unit square by dividing it into nine
subsquares, each of 1/3 the width and height of the unit square.
Here are the length 1 addresses.
Here are the length 2 addresses.