NPS52-86-008

NAVAL POSTGRADUATE SCHOOL

Monterey, California
DTIC
*...LECTE

MIAR209MB

' The Fractal Geometry of Nature:

3asis and Application to
Its Mathematical
Computer Graphics

Michael E. Gaddis

Michael J. Zyda

C-2
,Jantuary 1986

Jqiproved for public release; distribuLion unlimited

Prepared for;

Chit' of Naval Research

Arlington, VA 22217

86 3 19 06c"
 NAVAL POSTGRADUATE SCHOOL
Monterey, California

Rear Admiral R. H. Shumaker D. A. Schrady

Superintendent Provost

The work reported herein was supported in part by the Foundation Research
Program of the Naval Postgraduate School with funds provided by the Clief of
Naval Research.

Reproduction of all or part of this report is authzrized.

This repor" was prepared by:

Mlichael J y
Assistant Professor
CompuLer Science

Reviewed by: Released by:

INCEnT Y.•NXEALE T. MARS

Chairman Denof Inforn-vaion a
Department of Computer Science Policy Science

1A
 UNCLASSIFIED
SECURITY CLASSIFICATION Of THIS PAGE (ften DoE. Entoted)________________

AGEREAD
REPOT DCUMNTATON INSTRUCTIONS
DCUMNTATON
_______REPOT AGEBEFORE COMPLETING FORM
1. REPRT NUMEK 12;GOVT ACCESSION 3. RECIPICNT'S CATALOG NUMIIER
NPS52-86-OOC
4. TITLE (and Sublifti) 5. TYPE OF REPORT & PERIOD COVERED

The Fractal Geometry of Nature: Its

Mathematical Basis and Application to 6. PERFORMING ORG. REPORT NUMIIER
ComputerGraphics ______________

7. AUTI4OP(e) I. CONTRACT O1 -GRANTNUUUIERWa)

Michael E. Gaddis
Michael J. Zyda ______________

S. PERFORMING ORGANIZATION NAME AND ADDRESS 1. PROGRAM ELEMENT. PROJECT. TASK

Nava Potgrduat
ScoolARIA A WORK UNIT NUMBtERS

Mnavaly
PotgAdut School0 61152N; 1ROO0-01-MP
MontreyCA
9943-100N0001485WR41005

It. CONTROLLING OFFICE NAME AND ADDRESS 12. REPORT DATE

Chief of Naval Research January 1986
Arlington, V.A 22217 13. NUMIIER OF PAGES
128
1I. MONITORING AGENCY NAME & AOORESS(*I 011oftat boo conuWja.. 0111 ce) hopect~u)
IS.SECURITY CLASS. (of04

I50. OECL A131 VIC ATION/ DOWWNAO

SCMEOUI.!

1G. DISTRISUTION STATEMENT (of ALW

Report)

Approved for public release; distribution unlimited

OS.TRiDUTION STATEMENT
01 W 00A deeW *DE0ufat
mEtoro# ta st'Ch30 to 6"e *006s)

18 SUPOLEMCNtANY NOTES

~istK
14V90301 tcCoJS 00040#4 eld. Its-*4ve."p od Id.4omo Iy Woe$' atw"E)

General Term~s; Algorithats, techniqueu;

fr,*ctals, fractal mountaias, Koch curve

40 £gSYeaCT feeuINK0 ** tmvdeep .d. 14

C *~4(~vt b sbieIe
Fractal Geomet~ry is a recent synthesis of old matematical cotstructr3. It
was first popularized by complex renderings of terrain onl a conputer graphic.ý.
medium. Fractal geometry has since spawned research in many diverse sclentifle
disciplines. It rapid acceptance has been achieved due to its ability to mode-
phenomena that defy discrete computation due to roughness and discontintiltits.
With its quick acceptance has come pioble~s. Fractal geometry is a misundc-*-
stj~d id-a that. is quickly becoming buried under grandiose terninology that
Iserves no purpose. Its ess;enge Is Induction ýusni t-sir~ie geeretric coni:Lrcts~.,-
00 .'"j 1473 roITIow OP 1 OV Isois OBSOLETe US'CLA SS11 ED)
S N 0102. LF-C0 d.A6601 ECUNWTV CL&WFtCAYIO4WOrI *bilAG-E fIee. Dte NMoodd
 UNCLASSIFIED
iECUmiTv CLAPIFICATIOW OF THiS PAGl fi3m DOm&

to transform initiating objects. The fractal objects that we create with this
process often resemble natural phenomenon. The purpose of this work is to
present fractal geometry to the graphics programmer as a simple workable tech-
nique. We hope to demystify the concepts of fractal geometry and make it
available to all who are interested.

SeA"CUIM~"O V "tl a

;•"•¢::~ 'iCL.AS
S IFIED
 The Fractal Geometry of Nature: Its Mathematical
Basis and Application to Computer Graphics
Michael E. Gaddi. and Michael J. Zyda
Naval Postgraduate School,
Code 52, Dept. of Computer Science,
Monterey, California 93943

ABSTRACT

Fractal Geometry is a recent synthesis of old mathematical

constructs. It was first popularized by complex renderings of terrain
on a computer graphics medium. Fractal geometry has since
spawned research in many diverse scientific disciplines. Its rapid
acceptance has been achieved due to its ability to model phenomena
that defy discrete computation due to roughneas and discontinuities.
With its quick acceptance has come problems. Fractal geometry is
a misunderstood idea that is quickly becoming buried under grandi-
ose terminology that serves no purpose. Its essence is induction
using simple geometric constructs to transform initiating objects.
The fractal objects that we create with this process often resemble
natural phenomenon. The purpose of this work is to present fractal
geometry to the graphics programmer as a simple workable tech-
nique. We hope to demystify the concepts of fractal geometry and
make it available to all who are interested.
Categories and Subject Descriptors: 1.3.3 [Picture/Image Gen-
erationi: surface visualization. 1.3.5 [Computationaal Geometry
and Object Modelinig): Bezier surfaces, fractals; 1.7 JThree-
Diwensiosal Graphics and Realism]: fractal surfaces-

General Terms: Algorithts-, techtiques;

Additional Key Words mid Phrases.: fractals, fractal mountains.
Koch curve:

n1s CRAW
,~~) raa- 0n
OTIC TAB E

3 Thbw voeb ku tme asvpned by the NPS VowdeaFa Uk,.wO VtwptActi __" v'. i$--=...

SAvad andlo&
 TABLE OF CONTENTS

I. AN INTRODUCTION TO FRACTAL GEOMETRY ........................ 7

A. MATHEMATICS AS A MODEL FOR OUR UNIVERSE ........... 7

B. FRACTAL GEOMETRY ............................................................. 9

C. GOALS OF THIS RESEARCH ..................................................... 9

IL THE MATHEMATICAL BASIS OF FRACTAL

GEOMETRY ...................................... 10

A. PRELIM INARIES ........................................................................ 10

B. DIM ENSION ................................................................................. 15

C. FRACTAL CURVES AND SETS ............................................. 24

Ill. THE IMPLEMENTATION OF FRACTALS IN COMPUTER

GRAPHICS ....................................... .. 35

A. THE IMPLEMENTATION PROBLEM ................... 35

B. MAPPING FRACTALS TO A BOUNDED SPACE .................... 38

IV. FRACTAL COMPUTATION IN RI .......................... . ... ... .. ....... .... 45

A. THE GEOMETRY OF INITIATOR -. GENERATOR .......... 45

B. THE MiD-POINT DISPLACEMENT TECHNIQUE .................... 49

C. A KOCPI-LIKE FRACTAL ALGORITHM ................................... 51

D. IMPLEMENTATION - ~~ .......
.~ ............ I.....5-5

. ... s UMM A RY ................................................................................... 55

V. FRACG$kA GEOMETRY FOR GRAPHICS TERRAIN ............ 61
MODELING MOUNTAINOUS TERRAIN .................
'A 61

"" ... FRA.CAL TOOLS FOR TERRAIN MODELING ...................... 68

.4.-.- -.----
VI. SHORT CUTS TO MOUNTAIN SHAPES ......................................... 87

A. RECTANGULAR MIDPOINT TECHNIQUE .............................. 87

B. PARAMETRIC CUBIC SURFACES ........................................... 92

VII. CONCLUSIONS ............................................................................. 100

A. DIRECTIONS FOR FURTHER STUDY ..................................... 100

B. CONCLUSIONS ........................................................................ 102

APPENDIX A: FRACTAL COMPUTATION IN R2 ................................ 103

APPENDIX B; RANDOM NUMBER GENERATORS ............................. 112

APPENDIX C: THE TRIANGULAR MOUNTAIN ................................. 116

APPENDIX D: THE RECTANGULAR MOUNTAIN .............................. 121

APPENDIX E: GEOMETRIC SUPPORT ........................ 126

LIST OF REFERENCES ............................................................................ 128

INITIAL DISTRIBUTION LIST ................................................................ 129

 I. AN INTRODUCTION TO FRACTAL GEOMETRY

A. MATHEMATICS AS A MODEL FOR OUR UNIVERSE

The various branches of mathematics have through time developed as a
response to the need for more detailed models to describe new developments,
both technological and philosophical. This was true when Newton developed
calculus and also true during the late 1800's through the 1920's when a schism
developed between the classical mathematicians and some brilliant innovative
thinkers.
1. The Mathematical Crises of the Early 19th Century
One of man's greatest strengths is his ability to question his surroundings
and beliefs and through this questioning develop new insight and innovation.
Most mathematical systems are developed for use in applications. Man's natural
inquisitiveness often leads him to develop his systems beyond the application and
into abstract theory. This theory drives him to investigate the applications and
often yields direction for new discoveries that were not previously foreseen or that
defy intuition.
Georg Cantor (1845-1018) was the most notable of a number of
mathematicians who questioned the basic precepts of mathematics and developed
the modern met theort. Some of Cantor's discoveries seemed to invalidate many
of the long held beliefs of mathematics. Cantor and his peers became deeply
involved in cotrovasy ovw th6i ftadiars. Their discovery of functiorw which
seemed to violate the basc rules of geometry and calculus wene deemed as
monalers and unworthy of consideration by reasonable men because they lacked
usefulness to any applicetion then known Jlef. 1 pp. 9J. These new concepts
would ieinaIn in the arena of pure theoretical ruathematics until science
developed to a point where the old models could no longer adequately describe its
processes and would look to the new mathematics for a new perspective.
It was from these discoveries that FracWa Gcamdrt was born Iftef.
1:Chap. 21. It will be seen in the following chapters that fractal geometry

7
IS *LANK
 is a synthesis of many of the concepts which developed from the mathematical
schism of the 19th century, most notably set theory and topology.
2. What is a Mathematical Model
"4 Reference 2 defines a mathematical model in the following fashion [Ref.
. 2:pp 1-3]:

A matkematical model is a .p*thematical characteritation of a phenomenon or

proess. It has three essenti parts: a pjrocess or phenomenon which ;a to be
moeled a mathematical structure capable of expressing the important proper-
ties of the object to be modeled, and an explicit correspondence between the
two.
Although the phenomenon of interest need not be taken from the real worl,
they usually are. The real world component s described quantitatively by such
things as parameter values and at which time things occur.
The second component of a model is an abstract mathematical structure. In it-
self, the structure is abstract and has no intrinsic relation to the real world.
However because of its abstractness it can be used to model many different
phenomena. Every mathematical structure has an associated language for mak-
ing assertions. It the mathematical model ir successful the language of its
mathematical structure can be used to make assertions about the object being
modeled.
The third component of a model is a specification of the way in which the real
world is represented by the mathematical structure, that is, a correspondence
between the elements ol the first component and those of the second.

3. The Euclidean Model

When using mathematics to describe man-made objects, the Euclidean
model (standard Euclidean geometry) is usually satisfactory. Its structure is
simple and pure, which appeals to an engineer's nature. But as technolo•,
expands and we need to describe procesfes that are not twel bekaved, we need to
develop a geometry that can adequately model oui proccss within a certain
closeness of scale.
No model can completely dtscribe a natural object because nature does
not follow the man-made rules that we impose on our model. But at a given
scale, the model (if it is accurate) can describe the object with enough precision
to be of help in constructing it. Engineers use the geometry of a straight line to
describe a wall but this wall, when viewed closely enough, is not straight at a1.
This is of no matter to tLe engineer because his model is accurate for his scale of
reference.
 B. FRACTAL GEOMETRY
One man who saw a need for a new geometry was Benoit B. Mandlebrot. He
felt that Euclidean geometry was not satisfactory as a model for natural objects.
To anyone who has tried to draw a picture of a nonregular object (such as a tree)
on a computer graphics screen, using the Euclidean drawing primitives usually
provided, this is an obvious statement. The strength of Mai dlebrot's finding was
his research ;nto the findings of the earlier mathematicianb and the development
of a practical application of their theory. Mandlebrot coined the term Frie"h to
describe a class of functions first discovered by Cantor (Cantor's dust), Koch (the
Koch curve) and Peano. He showed how these functions yield valuable insight
into the creation of models for natural objects such as coastlines and mountains.
Mandlebrot popularized the notion of a fractal geometry for these types of
objects. Although he did not invent the ideas he presents, Mandlebrot must be
considered important because of nis synthesis of the theory at a time when
science was reaching out for new more accurate models to describe its processes.

C. GOALS OF THIS RESEARCH

There are two approaches that can be taken in the investigatiou of fractal
geometry and computer graphics.

- To view the computer as a tool to enhance the investgatioa of fractal

geometry.

- To view fractal geometry as a tool to enhauce the realisi of computer

graphics.

This research will take the later approachl. It is designed to investigate the
mathematics of fractal geometry and to show its application to computer
graphics I hope to be able to tome the subject of fractal geometry by waking it%
oathematics and technique accessible to the average computer scientist.

Whm Mau took

•" the -foirta.

m9
*1
 U. THE MATHEMATICAL BASIS OF FRACTAL GEOMETRY

This chapter is a brief introduction to the mathematical foundations that

underlie the theory of fractals. Little technique currently exists for the practical
application to attain complete mathematical rigor when using fractal functions
(i.e. it is very difficult to prove that a set is fractal). This causes the non-
mathematician to accept much of what he does with fractals on faith. It is
instead important to understand the theory intuitively. This can be gained by a
cursory look at the mathematical foundations for fractals.

A. PRELIMINARIES
A complete definition of fractals is given later in this chapter but before we
can understand that definition, we must establish a foundation in set theory.
Fractals were discovered in set theory and topology. They can be considered as
an outgrowth of investigations into these related fields.
1. What is a Set
A set is defined in IRef $:pp. II1 in the following fashion:
A set is fored bv the grouping .t..oether of single objects into a whole. A set is
a plurality thoujlzt of as a unit. We can consider thes-e as r
as teference to a primitive Clrnept, familiar to us all, whose resolu ion into
-re fundamental conrepts would perhaps be neither competent nor nee, y,
will content ourwlves with this construction and will assume a. an axiom
that anibject M inherently determines otrtain other obifets o, 4,c, ..- in sOn
undefined way, and vic versa. We eptmhs relation by the worda Thea t M
consists of QZýOd d,6, C
othi
This definition is intentionally vague to allow the set to become the bask
building block for all mathematical consumas.
,: ,. 2. Some Set Theoretic Concepts
"Thi section prewets some backrnd definitios ror rocepts used in
the body of this chapter. The reader is directed to the references for a detailed
explanation or proof (Ref. 2:Chap 51.

I. Cardluality
Two sets S and T ae said to have the sanm number of rkrnwtntl. or to have the
same wdimraws, if there is a on"" frutioa I from S to T.

V•. 10
 lFhmt. ana Infixite Sets
A Set Sis eald to be•fi.*e iftS hs the samne cardinality as #,or if there is.po-
uitive ipte•re n such that S has the same cardinality as (1, ,SD,$...n ) Other-
wise S i swid to be imn, ile.
Countabilty
A set S is ieid to be cowstaikf S has the same crdinality as a subset of N, the
set of positive integem. Otherwise, S is said to be

a). Any subset of a finite set is finite.

b). Any subset of any countable aet S is countable.
c). The set of natural numbers N is countable.
d). The set of rational numbers Q is countable.
d). The set of irrationul numbers is countable.
e). The union of a countable collection of countable sets is coatable.
Q. The set of read numbers R is uncountable.

3. Some Topoi ogical Coneps

This section presents some topological background concepts used in the
body of this chapter. The reader is directed to the referces Zor a detailed
explanation IRef. 3).

Metric Space
Ar ,e -14 #Pistt is a set in which we have a r e of 1he doseness oW proximity
o two elements of thf set, that is, we have a distance defined op t4e .et For ex-
ample, a moeic on R? wo*4d be the pytbusor•a metric:

CovrIng a S-4
Let X be a totpogira $wae- atad S a subsoetof X. A c~ d the W S is ex•tly
what its naqw- i=pis coludetao of subsets of X whbcb corer S. tWa Lk, w7os

., What is a Function
A function ia defined in Ref. 2:pp. 103-191] as

A functio-P from stsot A W &astH s aggi*whi-b spedla anekemeuof Bfor

whb element of A.

leI A and I be se. A 1xvitiOc (Ic mop. .or Up*&fot*oioa) I from A to 1),
denoted VA - B, is aretatin From A to B such tell.am
for every e which 6 an
Ielem, t of A thee exists uaique in Buuch thm <*,*- is au
neeat ofJ.
We write fYuJ - 4.
 If f is function from A to1 B, then A is called the domain of the function f and B
iscalled the codomain of f
To completely define a function we must specify the domain, the codomain, and
the value 1(x) for each possible argument z.

Functions can be viewed as a specification of a method to describe the

creation of a set from other sets using some agreed upon mathematical
symbolism. The functions can yield powerful results when the target set (co-
domain) is complex and not easily described by set theoretical constructs. This is
especially true in fractal functions when the domain is RN and the object created
(set, co-domain) is a nonregular shape. This is one reason why the computer
graphics system is useful in the investigation of fractal functions. The computer
can model the infinite function and display a finihe approximation of the created
fractal set.
5. Useful Functional Concepts
It is often helpful to clearly understand the universe of discourse within
which a function exists. The function can be rigorously defined within the above
constructs but lack intuitive appeal due to its complexity. Mathematicians have
defined many useful concepts to describe functions. The concepts applicable to
this study are described below.
a. Partial Functions
Most of the fractal functions in this study have as their domain some
undefined subset of RN. It is useful then to consider them as partial functions
and not concern ourselves with a rigorous description of the domain of the
fractal. We take our definition from [Ref. 2:pp. 201-2021.
It is often convenient to consider a function from a subset AIof A to a set B
with.Jut exactly specifying the domain A 'of the function. Alternativel , we can
view such a situation as one where a function has a domain A and .codrain ai,•
but the value of the function does not exist for some arguments of A. This is
called a partial ctin.
Deflnitiemw
Let A and B be sets, A partial fucttion f with domain A and todomain B is any
function from A' to B wbere A' is a subset of A. For any x which is an element
of A A', the value o? f (z) is said to be undef ined.

b. Bijectivity
It is often useful to know to what extent a function maps from the
uomain to the codomain. If a function is not a partial function and every point of

12
 the domain maps to a point in the codomain then we want to know if all points
of the domain A in the mapping /(A) mar to distinct points in the codomain B.
We may also want to know to what extent the mapping 1(A) covers the set B.
The definition of bijective, surjective end irijective functions is from [Ref 2:pp.
2041.

Let f be a function /: A-,B.

(a) I is turiective (onto) if f (A) = B,
(b) I is injective (one.to-one) if a o a implies f (a) - t/(a'),
(c) / is bijective (one-to-one and onto) if I is both surjective aM J injective.

6. Functions From RN -* RN
A point in RN space is specified by an n-tuple of the form
(z'z
(Z 24z 31...... Zn). To completely specify a function from RN -_ RN each point in
tY, d-'-main must map to a point in the codomain. An example is:

2 2
:R-,R
2
f ((Z1, 2)) (Z, 2 )

This function is well defined. For each point in the domain of the function we
have specified a unique point in the codomain.
Most of the functions that are covered in this study are mappings within
R2 or R3. Fractal sets exist in all finite dimensions but it is impractical at this
point to use fractal functions beyond the fourth dimension in view of the graphics
display medium's limitation to two dimensions. The use of fractal functions
whose dimension is between 3 and 4 is currently being investigated by allowing
the function to roam the fourth dimension and then taking time slices which
yield three dimension approximations of the seut [Ref. 4J.
"7. Inductive Definitions of Sets
Functional constructs do not always provide a convenient we.ans of
charactering an infinite set. It is sometimes more eloquent and powerful to use
the inductive method to characterize a set.

13

:4
 Our definition of inductive definitions of sets is from [Ref. 2:pp. 199201J.

An inductive 2 definition of a set always consists of three distinct components.

1. The basis or basic clause, of the definition establishes that certain obj~ct.,
are in the set. The basic clause estalishes that a set is not empty and charac-
terizes the "building blocks" (the seeds of the induction) which are used to con-
struct the set from the inductive clause.
2. The induction. or indvctive clause, of an inductive definition establishes
in which elements of the set can be combined to obtain new elements.
rThe wyays
tWe inductive clause always asserts that if oblects z .z are elements of the
set then they can be combined in certain spcrdwy ocet ther elements
'of Ih et(tus from the basic clause (or seeds) of the induction we induce the
remaining elements of the set).
3. The extremal clause asserts that unless an object can be shown to be a
member of a set b~y applying the basis and inductive clauses a finite number of
times, then the object is not a member of the set.
AP -4pinple of &ninductively defined set is:
(Bauis)
0t A
(Induction)
If n cA,then (it+2) cA
(Extremal)
No intlger is an element of A unless it can be shown to be so in a finite number
of applications of clauses I and 2 above.
The set that we defined is the set of all even nonnegative integers.

S. The Path To Fractals

* The path to fractals by the non-mathematician is not through theory but
through the investigation of their functions and methods of construction. This
investigation (&ad experimentation) yields. considerable insight into the nature of
fractals.
The choice of which set-descriptive methodology to use (functional or
inductive) in describing a set is often a matter of style but cani be dictated by
necessity if one method is inordinately tedious.
Most of the fract..e functions that are introduced in this study use the
inductive method as the primary functional tool. In fact, these functio.i uce a
hybrid of the functional and inductive constructs- described above.

Often called a reurrence definition.

14
 B. DIMENSION
The classification of fractal sets from non-fractal sets is based on the
dimensional qualities of the set. To understand fractals you must have an
* appreciation for these differences.
The concept of dimension is one rife with difficulties. Many of the great
mathematicians have attempted to define dimension as a rigorous concept
consistent with the known mathematical systems. For each of their attempts
however, the concept becomes more prone to contradiction and paradoxes.
There currently exist five definitions of dimension that date back to the late
1800's 3. The classification of Fractal sets into a class of sets is the result of the
discovery of functions that created sets which did not fit comfortably into the
topological definition of dimension (which was the accepted definition at the
* ti-.e). Fractal sets are rigorously classified as those sets that demonstrate a
difference between the Hausdorff-Besicovitch dimension and the standard
topclogical dimension.
1. An Inaiitive Apprcach to Dimension
Dimension is a concept that seems intuitive when it is first introduced in
Euclidean geometrv ai the standard three dimensions. Long after Euclid made
the first attempts at defining dimension and concurrent with the discovery of
atomic particle pbysi:s the concept of dimension was rethought by the prominent
mathematicians of the umsn 1 his was necessary to realign the mathematical
model for the geoiet'ry of objicts with the new view of what those objects were
made ul. Our increasing -bility to focus on We nature of matter inevitably
causes the models we use to change.
The dilemma that arose from the new concepts of dimension quickly
developed into a theoretiý 4 debate that left intuition behind. When human
untuition fails, we must rely upon well founded models that are based on axioms
of basic mathematical truth. It is only through the rigorous ipvestigation of our
"V mathematical models that allows us to gn beyond ir•uitioa and investigate the

'11. Cattow and Minkowski; 2). Bouligean ane, Minkowski; 3). Pontraigin, Schnirehrnan and
Kolomozorov, Tihornirov; 4). Hausdorff-Pekitovitch and 5). the topological dimension (there are
others). Most of these definitions ,Ptconcerned with the moat efficient method of covering a set
(i.e. am topological Coae"%".
 true physical nature of the objects we model. The debate still rages today and
borders on the philosophical. Two examples should suffice to demonstrate the
complexity and possible paradoxes that can arise from dimension theory.
The first is from [Ref. 5:pp. 323-344].
Consider the wa, in which we define the densityof air at a given point and at a
Eiven moment. We picture a sphere of volume V centered at that point and in-
luding the mass M. The quotient M/V is the mean density within the sphere,
and by the true density we denote some limiting value of this quotient. This no-
tion, however, implies that at the given moment the mean density is practically
constant for spheres below a certain volume. This mean density may be notably
different for spheres containing 1,000 cubic meters and 1 cubic centimeter
respectively.
Suppose the volume becomes continually smaller. Instead of becoming less and
less important, these fluctuations come to increase. For scales at which the
brownian motion shows great activity fluctuations may attain I part in 1,8O0
and they become of the order of I part in 5 when the radius of the hypothetical
spherule becomes of the order of a hundredth of a micron.
One steri further and our spherule becomes of the order of *a molecule radius. I~n
a gas it will generally lie in intermolecular space, where its mean density will
henceforth vanish. At our point the true density will also vanish. But about
once in a thousand times that point will lie within a molecule, and the mean
density will be a thousand times higher than the value we usually take to be
the true density of the gas.
Let our spherule grow steadily smaller. Soon, except under exceptional cir-
cumstances, it will become empty and remain so henceforth owing to the intra-
atomic space; the true density vanishes almost everywhere, except at an infinite
number of isolated points, where it reaches an infinite value.

The second is from [Ref. h:pp. 17-18].

Consider a ball of 10 cm diameter made of a thick thread of I mm diameter that
(in latent fashion) possesses several distinct effective dimensions.
To an observer placed far away, the ball appears as a zero-dimensional figure: a
*point. As seen from a distance of 10 cm resolution, the ball of thread is a three-
dimensional figure. At 10 mm it is a mess of one-dimensional threads. At 0.1
mm, each thread becomes a column and the whole becomes a three-dimensional
figure again. At 0.01 mm, each column dissolves into fibers, and the ball aain
bicomes one-dimensional, and so on, with the dimension crossin over repaof
ly from one value to another. When the ball is represented by anfuite number of
atomlike pinpoints, it becomes zero-dimensional again.

It is interesting to note that each of these examples demonstrate

dimension as a reflection of physical properties dependent on the observers point
of reference. Each ends with reference to the paradox of atomic particles. That
paradox is, for any collection of finite (or countably infinite) points, the
*• dimension is zero [Ref. 6:pp. 1-81. Since the earth and sun each have a finite
collection of atoms then accordingly their dimension is zern. Dimension exists

116
 only for a mathematical continuum and as such lacks application to the physical
universe as we currently know it 4 .
The two properties (continuity and dimension) cannot be separated.
Before the advent of atomic theory, matter was viewed as continuous and
composed of basic elements that were indivisible. While the debate raged over the
practical and philosophkal aspects of the nature of matter it became apparent
that the mathematical models which represent matter would have to change. It is
not practical to represent objects by representing each atom and its position
relative to the entire set. The power of modeling would thus be lost; that is, the
ability to model complex objects and their interaction by relatively simple
constructs. Thus the fact is reinforced that models can only represent objects
through gross approximations and that the model is only effective for a restricted
frame of reference. Without this realization, dimension would have very little
application.
2. Topological Dimension

a. An Intuitive Approach
The concept of dimension is very old. It is based on the algebraic
concepts of Euclidean n space and the notion that a set has dimension n if the
least number of real parameters needed to describe its points was n. This fuzzy
definition was accepted for a very long time until the advent of Cantor and set
theory. Cantor showed that dimension can be changed by a 1-1 tranformation
from an interval to a planar object. The fuzzy notin of dimension, as defined,
was challenged and required rethinking,
The mathematicians who did not accept many of the findings of set
theory at the time (but who could not disregard Cantors findings) began to
consider ways of explicitly defining dimension. The new definition would have to
be applicable to the bizarre functions of Cantor, Koch and Peano as well as the

The set theoretical concepts of finitenes, countably infinite and uncountably infinite (con.
tinuotus) sets carry with them very ptofound implications. It is premature to view matter as mere-
ly collcthions of finite atoms. Science may yet find true continuity (in the mathematical sense) in
atomic matter and the universe. For now, matter is what it is and ouw proau•ciatiom. upon it will
not change its true texture.

17
:-6
 relatively simple objects that had previously fit into the old definition without
contradiction.
There was one crucial problem that Cantor's findings raised
(Ref. 6:pp. 4-6]:
An extremely important question was left opn: Is it possible to establish a
correspondence bitween Euclidean n space and Euclidean m space combinin the
features of both Cantor's and Peano's constructions, i.e. a correspondence which
is both 1:1 and continuous? The question is crucial since the existence of a
transformation~ of the stated type b~*ween Euclidean n sace and* Euclidean m
space would signify that dimension tin the natural sense that Euclidean n space
has dimension n ) has no topological meaning whatsoever.

This fundamental problem was answered in 1911 by Brouwer. He

proved that Euclidean n space and Euclidean m space were not homeomorphic
unless n equals m. To say that two spaces A and B are homeomorphic means
that a mapping f :A -+B exists, such that f is continuous over A and bijective.
Additionally, the inverse of this mapping f - l:B -+A, is continuous over B and
*• bijective. If two spaces are homeomorphic then it is analogous to saying that they
"aretopologically equivalent.
Further research was done and a precise definition of topological
dimension of a set was developed. This definition assigned an integer value as the
dimension of any set based on its topological properties.
b. Definition of Topological Dimension n
The rigorous investigation of dimension is beyond the scope of this
study but the following definition is included for completeness [Ref. 6:pp. 24J:

Roozhly speaking, we may say that a space has dimension < n if an arbitrarily
small piece or the sa ce surrunding each point moy be delimited by subsets of
dimension 4 n - 1. his meth of definition is inductiv, and an eleant start-
:pg point for the induction is given by prescibig the null set as the (-I) dimeo.
Miond space.

The empty set and only the empty st has dirension -L.
A space X has dimension <, n ( n ;t 0) at a point p it p ha~s arbitratily small
neig borhoods whose boundaries have dimenmion i a -
X has dimension 4 it, dim X .ý i, if X has dimension i n at each of its points.
X h*s dim nsion n at poinip if it is true that X has dimension ( n at p and it is
false that rhas dimenaion a - I at p.
X ha- dimension n ifdim X 4 n is true and dim X• n - I is (alse.
X has diimeniom oc if dim X, a is fase for each a.

18
 The topological dimension is rigorous and consistent for all sets that
exist within a metric space. The problem that arises with fractal sets and its
topological dimension is not that the topological dimension is wrong. Fractal sets
like all sets in a metric space exhibit a topological dimension. The question is
then, is the topological dimension an accurate description of the dimension of the
set or can we find a better way to characterize the dimensional qualities of the
set? This question can be extended; is the topological definition of dimension
useful and consistent with the iw;ý&n of dimension and space? Can we devise a
better concept which can further refine dimension and make it more useful'.
3. The Hausdorff-Besicovitch Dimension
This section is intentionally brief due to the subject's complexity and to
the lack of practical technique that it yields. The Hausdorff measure of a set is a
complex characterization of a method for covering a set. Hausdorff's theorem is
proved using the ezistential qualities of infinite sets in a metric space 0. While
the theorem may be important to mathematical theory, it proves unfortunate
that there is no straightforward practical method for determining the Hausdorff
measure of a set.
a. An Intuitive Approach to the Hausdorff Dimension
The acceptance of the Hausdorff method for covering a set as a
measute of dimension is not universal [Ref. 6:pp. 102] and [Ref. l:pp. 363.365I.
The debate is between the dieciplines of topology and metrics and is not wholly
germane to this study. It is beneficial to divorce ourselves from the debate and
consider both the topological dimension DT and the Haundorff-Besicovitch
dimension (HB) as merely measures of 4psaesii of a set's structure. Certainly
sets exist that have a topological dimension equal to I but in no way resemble a
simple Euclidean curve If the Hausdorff dimension yields a better measure of a
set's structure that provides a mathematical and intuitive difference that is useful
to us, it would be beneficial for us to investigate it.

'Try not to tsribe grandiote implications to a set's dimension (the fourth dimension as time

or some such) as this is premature at best. lather, view a set's dimension au merely descriptive
terminology much like the terminology of bijectivity describes a function's characteristics. The
problem most people have with this mental abstrattion is the visual reinforcement that they re.
ceived from the notion of te standard tL'ee dimenasios.

*V!
 The Hausdorff measure of a set was developed during the same
period that the new topological dimension was invented to solve the paradoxes of
Cantor. The topological dimension was based on the idea of a neighborhood of a
point within a Euclidean space of RN. The connection to metric spaces and the
idea of measure is obvious when you consider that the Hausdorff measure of a set
is also based on this notion of a spherical neighborhood and what Hausdorff calls
A the test jundiom of a set. The test function of a set denoted h(p ) is a function
that characterizes the "Ust " method of covering a point with the spherical ball
of radius p that covers points of the set, which in their union, cover the entire
set.
Consider for example the test function for a surface within R 2. A
surface can be covered by discs (circles). The formula for the area of a circle
becomes the test function for the surface. The formula for the area of a circle
always contains the constant factor ir multiplied by the square of the radius r.
This radius is the measure p as above. This leaves us with a test function for a
planar shape in Rofh(p) = rp 2 .
You might expect that the test function for a spherical neighborhood
in a Euclidean space above IR3 would be very difficult to imagine, and indeed it
is. Hausdorff further complicated the idea of test functions (even within the lower
dimensions) by allowing a test function to assume a non-integer parameter d so
that the test function h(p) - *(d)p d could have a real-valued parameter d.
"Thisfurther refinement of the test function allowed Hausdorff to make assertions
about how this test function h(p) behaved when the parameters p and d were
allowed to vary.
SHausdorff imagined the parameter p reducing in size until it
approached zero. The effect of this on our disc example is increasingly smaller
and smaller discs around points of the planar set. As the disc size is decreased,
fewer points of the set are contained in each disc neighborhood. ThLs requires
more discs to cover the set. As the parameter p becomes infinitely small the
number of discs required to cover the planar set approaches oM. We allow the
parameter to grow arbitrarily small. It is interesting to study the test function

20
 and see what happens to the total area when the areas of the collection of discs
which cover the planar set are summed.
Let's reflect upon the mathematical process that we are developing.
When we attempt to approximate the area of the planar s-t by the union of the
discs which make up its cover, we are essentialiy observing small patches of the
surface and approximating the area of the set by making assertions about the
intrinsic qualities of the patch. The notion that this mieaure is merely an
approximation is important. As the size of our patch grows intrensingly small, we
can expect that we will get a better fit with our patches and hence a better
approximation of the area. The notions of approximation and fit become
increasingly helpful when you consider functions which describe sets of infinitely
rough tezdure as we find in fractal sets.
The importance of Hausdorff's discovery lies in the fact that for a
test function of a set, the parameter d is special. As p-*O he discovered that
there existed a unique real number d such that for d*<d, the infuium 6 defined
by the test function using d* approaches oo (for any countably infinite set). And
for a d'>d the corresponding infinum approaches zero. This means that e-•
set has a parameter which can be associated with it that is closely relate, .hm,
f
amount of spate that it occupies. This number is the Hausdorff-Besicovi,ýh
dimension (measure) of the set.
These results only tell us that a number and function exist. They do
not tell us how to compute them in the general case. This gives us the quandary
of dealing with the HO dimension as a known concept that we can make
allusions to, but can rarely compute (at least at the present time).
b. Definition of the Hausdorff Dimension
The formal definition of the Hausdor•f dimension requires that we
formalize the intuitive discussion above, We first define what a p-measure of a set
is (analogous to the disc above) jRef. 6:pp. 102-1031.
SP-dinmen!siontJ measure for each non-negative real number p w&s defined by
ausdorfl for arbitrary metric soace, This measure is a meltal concept, whire
dimension is purely lopolofw•f. Nevertheless there is a strong connection
between the two concept&. tr it turns out that a sp fe(topdo oicgi dimen-
sion a must have positive .iumioital nmeasure (Wh a-o. me~asre).

SFor our exaple this would be the sum of the ares of the discs.

4 21
 "'Let X be a space and p an arbitrary real number, 0 Ep <oo. Given E >0 let

where X = A' + A2 + A' + is any decompmsition of X into a countable

nuwber of subsets of diameter less than c, and the superscript p denotes ex-
ponent.Waion. Let
ml (X) = OUR m (X).
•nra (X) is called the p-measure or (p-dimensional) measure of X.

If p <q then n, (X) >, mf (X); in fart p <9 and in (X) <oo imply m, (X) = 0.
Conversely, if p> then nieX) < out(X); and if p>9 and
np (X) <00 implie•e P( = 00.

c. Mandelbrot's Misgivings about the HB Measure

It is clear from the previous definition that for practical applications
the Hausdorff dimension is difficult to compute directly. In iRef. 1:pp. 14-19] and
[Ref 71, Mandelbrot expressed a disaste for the focusing of attention on the HB
dimension. He states:
"I developed the definition of fractals using the topological and Hausdorff di-
mensions in response to colleagues who urge me to do so. They felt that it was
pecessary to rigorously define the concept within firm mathematical criteria. I
have come to believe that an empirical definition would be more beneficial at
this time because the present definition denies the inclusiou of some shapes that
could best be described as fractals."

Mandelbrot believes that the definition of the Hausdorff dimension is

too difficult to deal with and is perhaps too restrictive. He prefers to focus on the
behatioraJ aspects of the recurrence relatioship involved in fractals and the
empirical results from these equations.
A practical application of fractal functions in computer graphics
does, by necessity, bend to this same paradigm. This realizaLion should not blind
us to the fundamental nature of a fractal equation's uniqueness, however. It is
important to rnderstaad the dimeasioaud aspect of the fractal discourse to
appreciate the patential importance of fractals.
It is not preordained that fractal equations model nature with a
greater degree of accuracy then does the Euclidean model. The future way prove
the fractal model the superior method, however. The dimenaiorwl qualities of
fractal funct i ons may be the aspect that proves this to be so.

22
 4. Why Cc:-sider Dimension
For the purpose of this study, it is not important that a rigorous feel for
the mathematical properties of dimension theovy be grasped. In fact one needs no
knowledge of dimension to use fractal tech,•ques in the generation ot computer
graphics terrain. The literature is rife with articles about fractal objects in
computer graphics and it seems de rigue,'r to include an approximate fractal
dimension as put of its description. The techniques used to approximate
dimension as presented ia' [Ref. 1:pp. 5&-471 are mathematically unproven7 . More
importantly, the dimension yields little intuitive insight; one is hard pressed to
describe the differences between an object with an approximate dimension of 2.37
and another with a dimension of 2.45. The pictures are much more descriptive.
One use of approximate fractal dimension is to describe an object's
relative roughness. It is beneficial to view a fractal dimension as degrees of
roughness between the standard three dimensions. If the dimension is between 1
and 2 then the object should be a very irregular curve. If the dimension of that
curve approaches 1 then the curve is probably not very rough and would lack any
interesting diversion from an ordinary plane curve. If the dimension of the curve
approaches 2 then the curve becomes like . plane or filled polygon and again
lacks appeal. The most interesting fractal curves axe those which demonstrate
dimensioa never the center of the scale between the standrxd Euclidean
dimensions. A similar argument cau be made for #oisd objects with di
between 2 and I.
This i'* not to say that fractal geometry is not a powerful tool for the
graphics programmer. The evidence of the power of fractals to model objects of
considerable complexity is clearly demonstrated. To date, this power has not
been matched by other standard methods.
The graphics prorammer should not concern himself grerdy with the
dimension that is demonstrated at different levels of object construction. He must
concern himself with the techiaiques of construction and the realism that is
achieved.
7 At.i ft's method tos estimating the I&c.-tff.Ik ,itoc dimentaio for nonAftdQM
set built thwoIt self sisilar dwpes will be iuatduced in sectiou C afler the Koch tw, is
des-'bed.

23
 C. FRACTAL CURVES AND SETS
1. Definition of Fractal Sets
We take our definition of fractal sets from (Ref 1:Chap 3 and pp. 361].

A fr.au set is a set for which the HEcu wtBcsi&.uitc dimension strictly
exceeds the .P--: u dimunsion.

As we have established and is emphasized by Mandelbrot, this definition

is not very useful. The definitions of topological and Hausdorff dimensions are
very involved. It is a gargantuan effort to prove that a set has a topological or
Hausdorff dimension (if one desires complete rigor, typically the topological
dimension is derived by the least parameter approach (section 2)). When using
A fractal functions then, it is practical to uaum that because the functional
* method you use is tractal-like that the dimension is fractal.
The functional techniques to be introduced have a certain methodology
that creates fractal sets with a behavior that is disciplined and predictable. The
assumption is, since these methodologies are well behaved, that any set created
by these methods (with some careful restrictions) will itself be a fractal set.
We are left with a practical methodology whereby we discover fractal
functional methods, prove that the set created is fractai, characterize the frac.al
part of the functional method (carefully) and then cwhriw that method a a
fractal method. If one's purpose is a practical application of fractal techniques'
and not a rigorous matheuatical investigation then this is a reasonable and
practical approach. This approach is taken in the remainder of this study.
2. Constructing Frutal $ets
In order to describe the construction of the Koch curve, it is cessary to
present terminology introduced by Mindelbrot [Ref. Upp. 34-351.
We use a geometrical shape (at first a straight line) and call this shape
an INITIAMO& We create another shape that is constructed with shapes
similar to the initiator and call this a GENERA TO. We define a sequence of

4 A woqhing modal or eqtatiou whm you Lm only co.weemd abo th.w behavior and tat tke
exact amtheauaical pike.i
24
 transformations upon all current initiators (suppose there are m such initiaor8)
in the construction by applying the generator to all initiators. This creates a new
construction that consists of - x r (where r is the number of distinct parts of the
generator) sides where each side is a shape that is similar to the initiator. The
next step is to again apply the generator to all initiators.
This recursive definition has no terminating event but is continued ad
infinitum. This functional process is well suited for recursion because the
•¶ 3 application of the generator to the initiator is constant with respect to method
and varies only to scale. It is also well suited for parallel processing (in the
computer science sense) because each application of a generator on all current
initiators is independent.
These concepts are probably confusing at this point and were especially
difficult to visualize when they were first envisioned because the authors had few
tools beyond mental imagery to convey their point. This is probably why they
were largely ignored for 70 years. It is much easier to visualize these functions
when they are shown on a computer graphics display.
3. The Koch curve.
The mathematicians Koch and Cantor developed functions which
attempted to challenge the mathematical models of continuity and
"differentiability. These equations were developed during the great debate on set
theory and were used by Cantor in arguing for his theory. These functions were
"like none before, using constructions which played upon natural geonaetric
€onstrucs but when combined with the power of infinite recursion becasue sets
which defied intuition. It was not until much later that mathematicians wvere able
to reconcile these functions with algebra and set theory. The function to be
introduced has been proven to have a Hausdorlf-D.eicotCtc dimension which
exceeds its topotogical diaesion'9.
The Koch curie is a very beautiful curve that at firt gives the observer
: ,,the impression of a snowflake or a coattline (Fig 2.1). Mandelbrot uses this type
of construction (with variation - i.e. randoomwe or less behaved generators) to
draw coastlines that look very realistic. To understand this consttuction is to

'Thus keetab do •exa uJ it is pmibc

to oe iimaoroy,
t*•be so.

VS 25
 understari one method of obtaining a fractal set from a well defined non-fractal
set (R 2 ) using non-random techniques. This method of construction (in the
general sense) is very powerful and is used throughout this thesis.
To construct the Koch curve, we use three initiators (line segments) of
equal length and join them to form an equilateral triangle 10 . To construct the

generator we use four line segments that are each 1 the length of the initiators
3
and apply these to each initiator (Fig 2.1). This yields a new geometric figure
with 12 sides versus the original 3 and a total perimeter length of 4 units of
length versus the original 3 units.
Figure 2.1 demonstrates the first and second recursive iterations of
building the Koch curve. Observe how the progression develops to yield the final
figurn. in Figure 2.1. Imagine this progression occurring indefinitely.

-With each iteration

1 of applying the generator- the total perimeter length
increases by - over the previous perimeter.

The length of the curve begins to increase without bound even though the
length of the initiator decreases to an infinitely small length. Hence the
curve's length is unbounded with no point intersecting but yet is contained
in a small bounded two dimensional area.
- The points of the curve are by construction only the end-points of each
initiator and each point is clearly distinct from the other ( no two points are
connected).
-Although each point is distinct at any one level of the curve construction, it
can be proven that the curve when viewed in the limit is continuous at every
point.
- That due to the above qualities the curve is not differentiable at ANY point.

•1 It is important to realize that the endpoints of the lines (initiators) are the only

points of the curve. The line only serves as a vehicle by which the points may be
easily determined. The exact same set could be built using 180 arcs as initiators.
An algorithm and computer graphics program for the construction of the
Koch curve is presented in Chapter 4 and Appendix A.

10 The choice of an equilateral triangle was arbitrary. We could have chosen any shape as
long as it was made up of Initiators a.:d avoided intersecting lines during recursion.

26
NZ
 SINITIATOR
GENERATORA N =4
1 Initiating
r3 Structure

st
I" Recursive2n Re u s v
i
Iteration
Ierton•

THE KOCH CURVE

Recursive term- Output medium:

ination distance Laser printer
is .05 inch. Resolution:
[DT - 1,9J3 1.281 300 dote/inch

p Figure 2.1 The Kcch Curve.

27

==-%
 4. Mandelbrot's Dimension Approximation Function
In the above section, one functional construction technique for building a
fractal set has been introduced. It is possible to approximate the dimension of a
fractal curve that is built using these constructs. In [Ref. 1:pp 56-57], Mandelbrot
introduces a function that is based on the similarity properties of the above
technique. This function has a real exponent D that is t11 approximate
dimension of the fractal set"1 .
Consider a method of paving (covering?) a Euclidean shape. Divide a line
segment into N segments with each segment a part of the original segment such
that the sum of the lengths of the N segments equals the length of the original
segment. It follows then, that the sum of the ratios of the divided segments
SDivided~na segment ms

lengths to the original segment length e.g. rs = Divided aegment m

equal 1.

N
Er.= I
-1I

We know that the dimension of a line is equal to 1. If we raise each of the above
ratios to the power D (where D = 1) the equivalence still holds.
w ~N

6-1
Let's allow the Koch function to assume a similar dimensional relationship but
treat D as a real valued unknown. Refer once again to Figure 2.1. Notice that
the length of the four line segments that make up the generator have a length

ratio of -to the initiator. Call this ratio P.. Notice that the generator is made
3
up of four initiator shapes. Call this number N.

'~ (r.)' 1
.M11

' CAUTION: This technique does not mwccuwtp apply to other frutal functional
methods,

"28
 Substituting and Solving for D:

4 ( =)

log3

12
Which is equal to the Hausdorff-Besicovitch dimension of the Koch curve .
This is not a proof of a general equation for the fractal (Hausdorff)
dimension of a self similar fractal set but implies that a general dimension
generating function is possible:

r D -
G(D) = r
rn-1

Where N = number of sides of the generator.

Where Rm = ratio of side m to the initiator.

Mandelbrot claims that experimental evidence suggests that this equation holds
whenever this functional method is used.
When each segment of the generator is a fixed ratio to the initiator (as is
the case with the Koch curve) then the solution to this equation is trivial:
!og(N)
log(.--)
'mm

5. Functional Characterization of the Koch Method

A complete and rigorous inductive definition of most fractal functions
can stretch the notational capabilities of the symbolic aspects of the inductive
and functional methods. It is thus generally impractical to use these methods

"The Koch curve was proven to have a Housdorf-Besicovitch dimension equal

Los- su 1.2818 and a topological dimension equal to 1, IHausdodrf Dimension uad auueres Mas].
og
V1

29
 exccept in a verbose and non-rigorous manner. It can be insightful however, to
diasect the beaut (oneel) and hopefully gain further intuitive insight.
To simplify the process, we define a Koch half-line as in the above fractal
but with a single line segment as the Initiator (versus the equilaterai '.iaugk1.
We consider the line interval of [(0,0),(1,0)] within the set R 2 . This restricts the
fractal shape that is drawn to a partial function on: [0,1] X (0,1] -+ [0,11 X [0,1].
Using the inductive process to define the essence of the fractal
sequencing, we have the following definition:

(Eaaae)

Step 0

110 = {(o.0'0.0),(I.o,0.o))

(Induction)

Step k

Label the ordered set of 2-tuple points from Step k-1 as:
Ilk-1 ff Pk-11,Pk-" 2,Pk-1V,..'Pk-l• }

where
n ; 4 (0-) + I

Determine the new set of ordered points for Step k as:

filk={ p P ,P3,
3 ...-,P* )

where
m =4- +1

where pk-1
Pk =
pts p -2
ks= P 12
ph 9 = ph-I 3

"II:e.= i
30

S
 where

(pkh2 ph 3 ,Ph4) = pk-i [1Ii..]pk-12

(pk 6 ,pk 7 ,pks) = pk-I 2 [ 2R 1]pk-1 3

(ph10 pk 11,p 12) =p-i 3 3R4jp*-i 4

(Pkm-, p -3 in
m ,P-2
k _) =f p k-1n [.-iRn ]pk-1n

and where the relation [,... Rj] is the geometrical relationship between the two
end points of the initiator line segment from step k-1 and the five points of the
* generator that compose the ordered subset for step k as above. For purposes of
brevity, the full functional definition for this geometrical relationship is explained
in detail in Chapter 4.

(Ezitemaal)
No point is an element of fl unless it can be shown to be so in a finite
number of applications of clauses I and 2 above.
It is thus possible (but tedious) to rigorously characterize the Koch
fractal set within the well defined constructs of induction and the functional
technique.
6. Anether Fractal Set, Cantor's Dust
Cantor developed a function' 3 which used the same functional technique
as the Koch curve but with a reverse twist. Cantor "s function takes an initiator
(the unit interval 10,11) and dissolvee it into a discontinuous set which is as rich
in points as the interval 10,11 but contains no interval itself (Ref. 6:pp. 22-231.
The points of the set are all distinct but the set has the same cardinality as the
4 •!unit interval. The best description for this set is that it resembles a dust.

'Aso refented to as Cantor's diacoutiauum or Cantor's triadic set.

"31
 Cantor's Dust is difficult to demonstrate on a graphics display because it
quickly dissolves below the resolution of the display. Refer to Figure 2.2 to
visualize the initiator and the generator. The initiator is a line interval [a , b I
1
where 0 ( a < b < 1 and the generator is t.wvo intevals each I the leagth of
3
the initiator such that interval 1 is [a, a + ((b-a) x 3 J] and interval 2 is

[b- (b-a) xlb]. After the initial application of the generator to the unit

interval there will be two intervals in the current construction [0,11 and 13,1].
3
The second iteration of the recursive routine will yield four intervals
1,1, 12,1] and I ,J].
10I]

Every initiator that is created by the generator has an infinite sequence

that is begun at the next application of the generator. This cruses a series of
convergent sequences to each end point of each initiator. Thus each initiator

-INITIATOR 2
CANTOR'S DUST
I, ~T ~ Initiating Structure: D
___________
The Interval (0,13) B= 60

1st Iteration - - Initiator Length 1

2nd Iteration . Initiator Length1

3rd Iteration - Initiator Length T

4th Iteration Initiator Length 1

Figure 2.2 Cantor's Dust.

"!I
 e;awns a convergent sequence toward its endpoints. But for each initiator there
are also two spawned sequences toward the center. It is possible to imagine this
as an infinitely dividing organism which leaves behind four points (eggs) each
time it divides and then each egg itself replicates ... ad infinitum.
This functional method of diaslving a line is very powerful as a tool in
computer graphics because it can be used to cause many special effects from
ordinary objects. Mandelbrot has used variations of this method to create images
of star systems for example.
The topological dimension of Cantor's dust is 0 and by Mandelbrot's
dimension generating function we have the fractal (Hausdorff) dimension equal
to:

D)
G N Ir.)D

Where N = 2

Where R..

Substituting and aohving:

Al ()D

D = log 2 : .6309
log 3
Cantor's dust was proven to have a Hausdorff-Besicovitch dimension of
log 2 .', .6309 (Ref. 1:pp. 77-78].
log 3
7. A Note on the Concept of Similarity and Fractals Sets
The use of the mathematical concept of similarity continually shows up
in the investigation of fractal functions. This can be expected because of the type
of functional building tools that are used for most of these sets. The relationship
of the generator to the initiator has similarity built in. Thus Mandelbrot is able
to use similarity properties in formulating a dimension generating function and in
making claims about the set's inherent structure.
•3

I•.
 The functional method invented by Koch and Cantor is but one method
of determining r, fractal set. Similarity fractals may eventually be grouped into a
class of fractals (impeortant but restricted). This type of functional method
provides a vehicle for the creation of disciplined fractal sets but makes you
wonder about the rest of the space that fills the gaps between the standard
Euclidean shapes' 4, the self-similar fractals and the infinitude of RN.

4>

, Much like the rIte4MJCftd&tle numbers (* and o ra example ) in all their uncoun•ably rich
expn I gap
M the P betwe the failar (and well behaved) sets in R.

3
 III. THE IMPLEMENTATION OF FRACTALS IN COMPUTER GRAPHICS

This chapter introduces the reader to the practical aspects of implementing

Fractals on a two dimensional computer graphics display. We are forced to
confront the issues of economy of scale between the infmite fractal function and
the finite computational environment of the computer. Understanding these
compromises is a necessary bridge to successful fractal programming.

A. THE IMPLEMENTATION PROBLEM

1. Infinite Recursion, Stacks and Data sets
It is naive to view the computer as a truly infinite abstract machine
which is capable of any binary computation of any length. Infinity is a concept
0 that when applied to physical objects quickly breaks down as soon as that object
is bounded in any way.
It is possible to model infinite behavior in computers though mathematics
(automata theory) and gain useful insight into possible capabilities of the
computer (the use of push down automata in compilers for instance). But in
order for automata theory to make assertions it is often necessary to make the
assumption that the automata (computer or an abstract machine) is in fact
infinite. When the assumption of infinity is made, there are many powerful
mathematical tools which can be brought to bear upon non-intuitive abstract
problems that would otherwise be functionally"8 intractable if the automata was
considered to have an arbitrarily large but boanded space. The question arises,
can we consider the implementation of these constructs (insights) as valid? The
answer is yes, but only in the context of some bounded space (for instance, the
maximum stack space for a push down compiler).
The question of how many valid programs can be recognized by such a
compiler is usually too dimcult to determine. The compiler's solution to the
problem is to use a passive sensor to detect stack overflow and notify the user

Functionally in the serm that the problem mwl be decidable but the solution *Wecs is so
large &adundefined that it could not be determined in'
to aabaonae amount of time.

35

.1. IN% 11.L.t*N2 INI !."N

 NU that his problem is to large for the current stack space. Although the compiler is
theoretically a recognizer for an infinite set of programs, the finiteness of the
computer is the grounding factor.
A similar paradigm exists for implementing infinite functions like fractals
incompu~ters. Fractal functions use recursion, randomness, massive floating point
12 computations and large amounts of primary memory. In order to use the fractal
function productively, we must manage the methods we use and produce a finite
approximation of the fractal set we create. This can be done by a passive or
active means.
2. The Bounded Stack Limits Recursion
The Koch-like fractal functional method is by definition a recurrence
equation. All fractal methods M..'oduced in this thesis have a similar recurrence
definition. Thus it is impossible to avoid the use of recursion in the production of
fractal graphics'
Recursion on most Von Neuman type computers is implemented on the
system stack (which may be hardware (fixed) jr software (in primary memory
r and therefore expandable)). Such a stack is always bounded by a fixed upper
limit of allocatable memory space. The formal definition of the Koch-like fractal
method has no recursive terminating event. In order to use these methods, we
must determine the precision that we require and develop a termination event to
aigual the beginning of the recursive ascent when this precision is met.
Faiure to manage the recursive descent inevitably causes the exhaustion
of the system stack which stores the program saita on each recursive call"'. The
programmer should beewaeful to keep his local variables at a minimum in the
recursive subroutine. By doing so, the total data stored durin echb recursive
call is reduced.
.1 4

SWhy should you try 'The recumisje pan of fricts.s fucio is the r uteaically beau-
tiful aspect which mtakes them so eloquent and powtiful.
'~This.could be the r'ecursive temitnaaiozn event if you amc nok careful, It might be a good
idea to try tibs experimentally on your computer to determia, &his maxiimum mreusive descent
distaure.

.46
 3. The Computer Graphics Set Paradigm
The paradigm that is used for displaying objects on the raster graphics
screen is inherently discrete and two-dimensional. A typical system consites of
prim'ifive that allow the user a view of his modeling world as a largely
unrestricted three dimensional space. These primitives are limited in their power
and are usually bused on the Euclidean geometry (lines, points, polygons etc.).
The user is required to supply viewing puranclars that def'me a limited three
dimensional viewing space. These commands are processed by the graphics
system which proj.ct., the objects contained in the viewing space onto a two-
dimensional space th-.t is the display screen. The display screen can be divided
into discrete entities called pizda, each of which represents one point on the
screen. The number of pixels defines the resolution of the screen.
Given the graphics paradigm, it is natural to model objects that are
* defimed on the display as a collection of discrete points. When viewed on the
screen, these pictures can accurately model objects of considerable complexity
[Ref. I]. This view is a departure from the normal view of the graphics
application programmer. The application programmer views the objects he
creates through a collection of Euclidean geometric primitives (lines polygons
etc.) that are abstracted from the actual display". Both views are useful models
in describing objects but for the fractal graphics programmer the furmer is the
more powerful because it neatly maps to the fractal method of object
construction.

lFor this study, we view the world and the objects it contains as
collections (sets) of discrete three dimensional points. When an object is built
through fractal techniques, the length of the pixel is a natural termination point
for the fractal recursive process. This yields an attractive bijective mapping from
our object to the display screen.

' All computer Craphics to an bistraction; realism is achieved through deception ot the eye
by a very small colltetioa of colored points. Our goal then, is to fiad the most eflicitt method of
dtfiams tho. powoi.

~' %~,37
A
4. Summary
From our previous discussion, it should be obvious that fractal sets do
not exist in our computer. The sets that we create are finite approximations of
the actual set. The same is true of the fractal pictures that we create and to
which we ascribe fractal dimensions. It is an illusion of the eye that we create by
taking the fractal computations below the resolution of the screen.

B. MAPPING FRACTALS TO A BOUNDED SPACE

In view of the graphics set paradigm, it behooves u. to develop a completely
bounded set space which has a one to one correspondence between the points
computed by the fractal equations and the entities of the screen. This abstraction
is appropriate because these entities called pixels (which are nothing more than
colored points) are the only components of the picture' 9 . The only reason for not
using this methodology before is that it was functionally difficult to compute
(without some simplifying abstractions) the large number of pixels in the dispay
act,
1. Fractal Recursive Termination Event
Most graphics software packages provide v. means by which the user can
define the spAce on the display screen onto which he wishes to m~ip, This is done
by providing to the graphics system, viewing parmeters, which define a bounded
three dimensional space which is then projected onto the two dimensional screen.
Many mappings are possible within the above constructs but for Jhe
fractal programmer it is convenient to establish a mapping that provides for a
one to one (or bijective) relationship between the 3 D pixei set and the real-valued
coordinate space in which we compute a fractal object. We develop this bijective
relationship by defining a fratal paA/ to be determined when the distace
between generating points (in real coordiuates) is less than the dhtance of one
pixel in the mapping of real to screen coordinates (SCR), Figure 3.1. Thi
mapping is explicit when the uindow, tiewporf and :-clipping are defined. This
mapping limits the numaber of pixels that can be Lsocited with a given fractal
Sinitialof to a one to one mapping of fracutl part to the pixd set. The sie of the

'nla fact. this is the way the Eldi'e" grapbhk ab.tutt sath ua a w am mapped to the
scrs. At some poa•t in thk display peam t"is mappme mugt Lae Place.

58
 pixel is a cubic space which clumps together the infinite fractal space into a
discrete number of cubic spaces equal to the number of pixels in the pixel set.
The fractal programmer should choose his viewport and window carefully
so as to elicit the most attractive mapping possible between his object and the
screen. The viewport (SCR) rectangle should be defined such that the ratio of
the X distance (horizontal) to the Y distance (vertical) is equal to the ratio of the
X-Y distances of his (real-valued) window. The Z-distance (front and back
clipping plane distance) should be equal to the distance in the window
coordinates that defines the length of the pixel multiplied by the desired Z-depth.
Figure 3.2. Formally:

XSCR XWLD
Ysca Y WLo

Lecursive Ter Ination Event

Recursive Iteration i I Recursive Iteration i At
The POINT PLOT Routine

I
Z•IbI

I.'
nori'Ctefor L*04t Zo Loop rho Pis*e Wr* Wsppcd To
rhn Tno PZJU• $X[S, ,Sa $*P 7%i r C'orrewpandisS Powaicioa

Th•e Poiats. To Ih* PiX*1 Sot. Z= rho Pixvi Sie•.

S~Figure 3.1. The Fractal Recursive Termination Event,.

s30
 o
Terminh,i pal IThe Pixel Set]
lrw ------
9 .U.er Defined Viewport

r Dist. :-Oraphic& Display Space

SCR.

User Detined Window

In World Coord. System- \..
YDistace

S~-"

X Distance
WZVD

The Cubic Pixel Space

To Pixel Size X& Defined By The Z Depth In Pixel, Ia De fined By

1 PELSIZE X WLD Z-Depth1 ,x Z-Depth ,D

-- Xs PIXEL SIZE

Figure 3.2. Normalizing the Pixel Space.

40
If the above relationship holds then the 8ize of the pixel in world coordinates is
equal to the X (or Y) world distance divided by the number of pixels, Figure 3.2.
XWLD
PIX SIZE = -

XSCR

The front and back Z-clipping planes should be established such that the distance
from front to back is equal to the desired Z depth (expressed in pixels) multiplied
by the pixel size (in world coordinates), Figure 3.2.
If the above viewing relationships hold, then the inclusion of a simple
check:

IF (InitiatorDstance < PIXSIZE) THEN

return;
ELSE
continue;

terminates the current recursive descent and begins backtracking.

2. Memory Requirements
The amount of main memory required by the fractal programmer is
directly proportional to the size of the pixel space that is being used to display a
fractal figure and the sophisication of the desired display.
a. Data Locality During Recursive Descent
If the fractal in question is a 1-2 aimensional display that is
displayed as a collection of points on a two dimensional plane then the
requirement to store the 3-tuple points is eliminated ((z ,p ,z) coordinates which
represent a point in R3 ). The points can be computed by the function and
displayed on the screen in the immediate mode and then destroyed. The only
requirement for memory is the data that is germane to the program ("globals"
and thus fixed at run time) and the amount of data pushed onto the run-time
stack for the recursive calls up to the point of the deepest recursive level. This
requirement is minimal and does not present any significant limitations.
An example cf such a frartal is the Koch curve The algorithm
presented in chapter 4 uses the data that is local to the subroutine that is
computing the current generator with the input coordinates of the initiator. The

41

U- A__

--%
 inductive nature of the fractal method provides data independence so that the
programmer does not have to have available to his subroutine all points
computed thus fur.
b. Memory Requirements for the Fractal Set Paradigm
If the programmer requires a more sophisticated display that includes
a requirement for hidden surface removal and (or) lighting enhancement then the
fractal method becomes juat a part of the overall display process. The
programmer has the fractal process compute the points of the picture and these
points are stored in a memory structure. This structure is processed by a hidden
surface algorithm and a lighting algorithm and is then projected onto the screen.
The ideal Fractal computer would have an exhaustive memory and
unlimited romputing power to be able to allow us to store the entire pixel set in
memory. This however, can stretch the capabilities of most present day
computers. For example; if the pixel set is defined such that its dimensions are
100Oxl00Ox1000 (which is certainly reasonable), the amount of storage required
would be an array of II? x 24 bits20 or 24 billion bits.
Current techniques exist where this storage can be minimized. A Z-
buffer array can be used to store the current Z coordinate of the forward most
displayed pixel. By using this method the fractal computation can be made and
then a hidden surface computation can be immediately invoked to check the
tdiibility of the point by checking it against any other point in the same position
on the X-Y plane. This technique reduces the above space requirements to
106 x 24 bits + (Z- buffer = 106 x 32 bits)2 1 or 56 million bits.
A fractal programmer must manage his memory resources carefully
in order to maximize the computational resources available to him. This usually
requires a tradeoff between efficiency, realism and the memory space utilized.
The algorithms for hidden surface and lighting are well covered Iin the literature
and although they are integral aspects of the fractal realism issue they are not
germane to this study. Throughout the remainder of this study, we are not

0 24 bits for a machine ausuming the RGB color system with 24 bit planes. •ach color; red.
green and blue would then have 8 bits of precision.
I' Assuming the Z coordinate is stored as a 32 bit floatina point number; this number can be
further reduced if we frtrict the Z precision.

42
 concerned with the methods of minimizing the the pixel set. We assume that we
have the full three dimensional set.
3. Concurrency
"Aquestion arises, is it possible to generate fractals in Red Time 22? We
have enough knowledge at this juncture to discuss the possibility for concurrent
operations during the fractal recursive descent. There are two basic strategies
that can used if concurrent operations are available on the computer. We could
*: assign a process to each of the initial Initiators and process each to its desired
*
-- precision. We could also allow the imagined processor to have a means by which
at any level of the recursive descent, we could spawn a process to apply the
2
generator to the local initiator 3.
The local data independence of the fractal recursive function allows such
a spawning because the application of the generator to an initiator is local to the
initiator. This application is a result of the inductive process of the K-i steps
that led up to the Kth step. The mathematical process of applying the generator
at the K'h level needs only the initiator data from the K-I" step and requires no
knowledge of the entire fractal set. The power of induction is then
computationally realized. The number of calculations before the entire pixel set is
determined is not greater than the number of computations required to apply a
generator times the maximum recursive descent distance from any initiator to the
recursive termination event. A short algorithm describing the processing aspects
of the second method is shown in Figure 3.3.

"* . u To beat the human eye and achieve the ultimate graphics illusion, typically by completing
all computations and screen I/0 prior to the -.Leo. refresh rate (for a 30HZ display) of most
raster graphics displays.
"This of course would require an enormous computational power that does not currently ex-
uit, so in reality the spawning would have to be bounded in some way.

S '.43

-4

.q. 511 11 4"-- *01.Xtý ;-;

 LOAD INITIATOR AND GENERATOR DATA

concurrent;
generate(INITIATOR 1);
generate(INITIATOR 2);
generate(INITIATOR 3);

generate(INITIATOR n);
end concurrent;
end main;

generate(INITIATOR I)

BUILD GENERATOR FROM INITIATOR

if (Distance < PIX size) then

PLOT POINT
return;
endif

concurrent;
generate(INITIATOR 1);
generate(INITIATOR 2);
generate(INITIATOR 3);

generate(INITIATOR m);
end concurrent;
end generate;

Figure 3.3. An Algorithm Which Utilizes Concurrent Operations.

44
 2
IV. FRACTAL COMPUTATION IN R

The algorithm introduced in the following section is capable of computing a

very broad class of Koch-like fractal curves within R2. It provides the fractal
graphics programmer with the basic tools for fractal computation and an
algorithmic template that can be used for many applications. The graphics
programmer needs to fully understand fractal programming within R2 before he
attempts the more complicated concepts of fractal terrain modeling in R 3 .

A. THE GEOMETRY OF INITIATOR -+ GENERATOR

The geometric relationship between the Koch curve initiator and generator
can be described through a very simple set of data that captures the essence of
the Koch curve. The method introduced also allows us to vary the data which
defines the generator and compute many different fractal shapes2 4 .
The general strategy for computing generator points from a set of initiator
points is to determine two lines that intersect at an unknown generator point,
Figure 4.1.a. The unknown generator point is defined by constant relationships
r between the initiator and generator. The first line is the perpendicular from the

generator point to the initiator line. The second line is formed between the first
point of the initiator and the unknown generator point. All data to compute the
line equations are derivable from the two endpoints of the initiator or from
constant initiator/generator ratios. For simplkity~s sake, we ignore the divide.by.
zero problems that are encountered when any of the lines are parallel to the X or
Y axis. These situations only simplify the computations and their solution is
demonstrated in Appendix A.
If we label the endpoints of the initiator as P1 =(X1 ,Y1 ) and P2 =(X2 ,Y2 )
(Figure 4.l.a) then the slope of the initiator line is:
Y2 - Y,
Slope.hiat = X2- Al

• We must be ca'reful in our terminoloay because the relationship described can draw shapes
that do not avoid self interection sad thus must be coatidered quui-ractal.

- -V•
 The slope of the perpendicular intercept line is the negative inverse of the
slope of the initiator, hence the slope of this linm is:

Slope.perp -1
Flope.init

The intercept point on the initiator can be determined by using a constant

distance ratio as shown by the following equations:

SX.pexp X 1 + (Generator.ratio. constant X X 2)

1 + Geeratorratio.contant

x Y,)
Y~per Y + (Generator.vatio.•onstant
1 + Gcnerator.ratio.constant

The value of "Gcuerdnordio.consad"is a fixed constant (although it might be

-@ interesting to randomize it) where you determine at what point the generator
intercept point intersects the initiator •nd express it as demonstrated in Figure
4.1.b. This ratio can be determined graphically, through hand calculation or via
an interactive automated means.
With the slope and a point of the line (X.perp,Y.perp) we can determine the
- line equation for the perpendicular line by determining the Y intercept:

Yintercept.perp - Y.perp - (Slope.perp x X.perp)

M This yields the line equation for the perpendicular line:

Y = (Slope.perp x X) + Y.intercept.perp

To detea .irme the generator line equation for the line segment which connects
the first point of the initiator and the unknown generator point, we need
constant information about the angle between the initiator and this line. This
&ngle is always constant with respect to the initiator and like the ratio
information above, it is recorded as a constant at run time (see Figure 4.1.c),
note that the angle may be positive or negative.

40
A
2*
 0(
son,
Y -- Unknown
0• " Generator Point
Generator Line--
'

CPOW r,1-'4---Perpendicular Intercept

Fig 4.1.a Intersecting Lines Determine a Generator Point.

Initi t r Generator Intercept Point

., G, pa
dist. a .. i dist. b

Ratio-constant - =
G, pa b

Fig 4.1.b. The Generator Ratio Constant

Point with positive angle Point with negative angle

s# 0 "',

-:, II7
Fig 4.1.c. The AngleO ",Go

i• Figure 4.1 Building the Generator with Intoersirting Lines.

I N

V
 We record the data about the angle 9 as the tangent of the angle. With this
information, the slope of the generator line can be determined with the following
equation:

Slopgen TaniU + Slope.init

peg (1 - Tani9) x Slope.init

The Y intercept for the generator line can now be determined:

Y.intercept.gen = Yj- (Slope.gen x X1)

This yields the line equation for the generator line:

Y = (Slope.gen x X) + Y.intercept.gen

The Cartesian points of the unknown generator point can now be determined
by intersecting the two line equations and solving for X... (Figure 4.1.a) then
0 substituting Xse2 into one of the line equations and solving for Y ...
3 The
equations follow:

Xgn=Y.intercept.perp - Y.intercept.gen
Slope.gen - Slope.perp

y~n=(Slope.gen x X...) + Y~intercept..gen

The constant data for the Koch curve that corresponds to this geometric
method is illustrated in Figure 4.2.
There are many different ways to build a generator given an initiator using
standard geometric constructs but the method introduced allows experimentation
with the Koch function to diecovcr new shapes. By varying the tangent of 0 and
the rixed ratio of Figure 4.1 we can describe new generator constructions. These
new constructions can be used in the same algorithms that compute the Koch
curve. We Initiate the recursion on a generating structure built of line segment
initiators and allow the recursion to progress until a terminating event~s, creating
many diverse shapes. Figures 4.8 through 4.9 demonstrate some of these shapes.

'The length of a pixel or an aLrbazy line lenagth.

48
 The Koch Generator 1

Point 1 Point 2 Point 8

I
- --

2 1 .
Ratio - 0.5 Ratio - 1.0 Ratio - 2.0

0- 0.0 6n 0.4710 reds. 8, 0.0

Figure 4.2 Ratio Constants for the Koch Curve Generator.

B. THE MID-POINT DISPLACEMENT TECHNIQUE

Mandelbrot [Ref. l:pp. 43 and pp. 233-234] uses an alternate technique to
draw the Koch-like curves that is equally valid. He calls his technique mid-point
displacement because he determines the generator via fixed relationships that
displace a point from the mid-point of the'initiator. This method uses many of
the same geometric relationships that are used above but provides a different
progression to the method of building the generator. This new view allows us to
look at the relationship between the initiator and generator in a slightly different
light. By so doing we are provided new insight as to how we might alter the
relationship to create new images anzd sets.
The mid-point displacement technique can be useful for two other reasons. It
is the best known method for fractal set building and because of this, it facilitates
communication between fractal programmers. Most of the terrain models that
have been developed use the mid-point technique. It also provides an easier
method to avoid line intersection.

49
 The Midpoint Displacement
A. Technique

Figure 4.3.a First Midpoint Displacement.

Figure 4.3.b Second Midpoint Displacement.

THE KOCH GBNERATOR

Figure 4.3.c Third Midpoint Displacement.

Figure 4.3. The Koch Generat or Using Midpoint Displacement.

The mid-point displacement method is demonstrated in Figure 4.3. The

method progresses by taking the initial initiator and applying the first midpoint
displacement. This yields the figure demonstrated in Figure 4.3.a. The next step
performs a mid-point displacement on the left initiator created by step 1. This
yields the figure demonstrated in Figure 4.3.b. The third and final step is to
replace the right initiator created by step I as demonstrated in Figure 4.3.c.

so

. ;ýI V w.
-* The inversion of the direction of the mid-point application (in Figure 4.3.&
the displacement is above the initial initiator where in Figure 4.3.b it is below)
can be accomplished with a single computational procedure. We only need to
invert the position of point I and point 2 in the parameters of that procedure.
The parameter inversion changes the orientation of the initiator in space with
respect to the computation cf the midpoint displacement. The procedure blindly
computes the wid-point displacement relative to a fixed relationship to the
initiator input points. This operation implicitly defines the orientation of the
generator in space.
The geometry for computing the midpoint given two initiator points can be
computed in precisely the same manner as the intersecting line algorithm (The
Koch midpoint ratios are 1.0 and the angle 9 is .437 radians). An alternative
method is to use the equations of a line (or a plane) normal. The second method
(utilizing the normal) provides a geometric relationship which is intuitively
appealing. Its appeal comes from the desire to modify the length of the
displacement relative to the initiator with a random scaling factor.
The mid-point displacement technique has some advantages over -the line
intereection algorithm. Random modification of the length of the displacement
along the normal (from the computed generator point to the initiator using the
line intersection algorithm) is not intuitively appealing. It requires a translation
of the desired displacement into an angle (the angle 0 (Figure 4.1.c) between the
initiator line and the unknown generator intercept line). The control of that
angle is less intuitive than the control of a displacement length. The geometry
for mid-point displacement using the initiator normal is introduced in Chapter 5.

C. A KOCH-LIKE FRACTAL ALGORITHM

Implementing the above function is a relatively easy process that ie
demonstrated in this section and Appendix A via a gradual unorapping of the
-.
lyers of complexity that are required to successfully implement the algorithm.
-
k The algorithm roughly follows the template used in Chapter 3 to demonstrate
concurrent processing. A C-like language is used for the algorithmis.
The first algorithm (Figure 4.4) is a template that delineates the basic
- •processing steps. This recursive process is typical of fractal functions and can be
 used as a template for many fractal programs. The second algorithm (Figure 4.5)
is an expansion of the first and demonstrates the replacement of a given initiator
using the line intersection method.
Appendix A is a complete Fractal program. This program was used to
produce the data for Figure 2.1 and Figures 4.6 through 4.9. This program
demonstrates the precautions that must be taken to avoid dvimd-by-zero when
lines are parallel to the X or Y axis.

main()

Load Initiator Coordinates;

V Load Generator Relationship Values;

(W 1= 1; 1<= Number of Initiators; 1=1+1;

4 generate(X(I)Y
1 (1)( 1,X(1) 2 ,Y(1)2 )
S~)

generate(X1 ,Y1 ,X2 Yy)

Determine Distance Between the Endpoints of the Initiator

if (DIST < Pixel.length) retumn;

Replace Initiator with the Computed Generator;

Load the New Initiator Data into Local Generator Array

for J•-1; J<- Number of Generator Segments; J3J+1;

generawte(Xgen(J)I, Ygea(J),, Xgen(J),, Ygen(J)•),

S}
.1

Figure 4.4. High-level View of the Koch AIgorithm.

452

1
11jj
 Load Initiator Coordinates;
Load Generator Relationship Values;

for 1=1; I<= Num~ber of Initiat~ors; 1=1+1;

generate(X(I) 11Y(I) 1 X(I)2,Y(I) 2)

geneWate(X 3 X1X
1 2 ty2)

/* Determine Disatnce Between the Endpoints of the Initiator '

DIST =soWt((X2-X 1)**2 + (2Y)*)

/If Distance is teas, than Aid Length; Solo and Return.

if (DIST <~ Pixel-length~)

plot powuto; /* Your Grap-ics Po int Plotting Routine/

return; 1' Point I and'29 P149 the game p.Zel

1Load The Endpoints of the faitiator into the. Generaor Array /

Generator.XIJ)X1
GeneratorX(Oj '4;
-

Getterator-XINurogeatrpns + 1
Gecrator.YfNuwieroL~enerator.poinbs + 1I YI

/Detrrnifte Slope of -:At Iitialor /

Figure 4.5. Detaited View of the Koch Algonitihm.

 /* Calculate the Unknown Generator Point via Intersecting Lines */
for J= 1; J<= Number.of.generator.points; J-J+1;
{
/* Determine the Generator Point Intercept on the Initiator *1
X.perp = (X 1+ (Generator.ratio.constant [J] * X 2)) /
(1+ Generator.ratio.constant [J]);
Y.perp = (Y 1+(Generator.ratio.constant [J] * Y2)) /
(1+ Generator.ratio.constant[J]);

"/*Determine the Slope of the Perpendicular Line */

Slope.perp = (-1 / Slope.init);

/* Determine the Y-intercept of the Perpendicular Line *1

Y.intercept.perp = Y.perp - (Slope.perp * X.perp);
/* Determine the Slope of the Generator Line */
Slope.gen = Generator.tan.theda [JJ + Slope.init) /
(1-Generator.tan.theda[J] * Slope.init);

/* Determine the Y -intercept of the Generator Line */

Y.intercept.gen = Y, - (Slope.gen * X!);

* /* Determine the Unknown Generator Point */

Generator.X[J] = (Y.intercept.perp - Y.intercept.gen) /
(Slope.gen - Slope.perp);
Generator.Y[Jj - Slope.gen * Generator.X[J] + Y.intercept.gen;

for K=O; K<= Number.of.generator.points + 1; K=K+l;

generate(Generator.X(KJGenerator.YIKI,
Generator.X[K+I],Generator.YfK+1l);

/} End Generate

Figure 4.5. Detailed View of the Koch Algorithm (continued).

54

- --- =
 D. IMPLEMENTATION STRATEGIES
There are numerous ways to display the fractal shapes that the above
algorithm is capable of computing. The graphics primitives required are limited
to the standard initiation and termination commands coupled with the ability to
plot a point (or alternatively a line). Any raster graphics system, plotter or
similar technology suffices.
The algorithm can be extended to include:

- Online generator drawing to compose a generator relationship visually.

- Rotation in 3 dimensions (if your system has this capability).
- Variation of the inductive application of the generator by the inclusion of
randomness with respect to the generator constants.
Figures 4.6 through 4.9 represent a few of the shapes that this algorithm can
compute. Each figure has the generator data used to compute the shape and a
progression that shows the first two recursive iterations.

E. SUMMARY
The line intersection algorithm as it stands is not very useful for the
production of graphics images of realistically textured terrain. Its importance
results from its encapsulation of the eatence of the non-random inductive fractal
method. This algorithm demonstrates the idea sad intent of fractal functions and
their implementation within computer graphics. The potential fractal
programmer must throughly understand the salient parts of this chapter before
successfully attempting fractal image. in three dimensions (i.e. before climbing the
mouvntin Chapter 5).

' 55
 A Cloud-like Shape
,nitisatin Structures

GE RATOR INIT ATOR HquIlA cral Trianglo

"lot •teration * ,
S........o.....

GBNERATOR DATA

Number of pointa& =
PaIntl t
Angle 0.4M23 reds.
Ratio = 3.4223 2ad Zteration

.ecursive
Termination
At 0.01 Inch

Figure 4.6. A Cloud-like Shape.

~w.~..--"~ ~- ~ ,- , - - * --. ~- *. --- ~- .* * *. * .6-

- L•~~. .*~*

DAN
... •*,--•.~. *- 'VA ~ \ ~*
 Boxes Ad Infinitum
Initiating Structures

GEN RATOR Squilateral Triangle

I.teration lot

~I~i'i1 Iteation

GENERATOR DATA
Number of points = 3
Pointi *
Angle = 0. 0000 rade
Ratio = 1.0000
Pointa:

. _Angle =- 0.
Ratio 1.0000
7854 rd 2nd Ztertion .... ..

Point3t = 0.4036 Padsw

'An~le
___•••'Ratio = 2Wo.0o

•ARecul ve

",Terminaaiom
0.075 inch

Figure 4.7. Boxes Ad Infinitum.

67

'.l

U=
 An Exaggerated Koch Curve
rimitiatime Structure:
GS RIATOR 5quilato~ rigl

lot Xteration

GBNERATOR DATA 3
Number of points

Angl e = 0.0000 reds

Ratio = 0.8607
Point3: r a tion
Ite
7 8 54r ed. 2n d
nA4al.= 01..O0000
Ra~to
,.• PointS:
•'1

Anal.e = 0.0000 red.

Ratio = 1.5000

,~.B

V3

t'

igure 4.8. An Exa

IJ g•ger ated Koch Curve.
- I

:68

UN
 A Plane Filling Curve
This plane filling curve requires
a slight modification to the dir-
ection of the generator applica-
tion. This reversal is explained
on the next pa*e.

INIT. GENERATOR

GENERATOR DATA
Number of point-s -
Pointl1
An&le = 1.0535 radi
Ratio = 0.a000
Poin t2
Aaple - 1.0535 rads
Ratio - 0.5000

Poin t4a
Angle 0.7092 rade
Ratio 3.0000

Anrle a 0.3289 rads

Ratio = 4.0000
Poiotat
Awl.@ 0.000 reds
Ratio 0,500
Poin ts#
Angle - 0.0000 aadv
Ratio 0 2. 0W

-aontlnuod-

.... -
Figure 4-9. A Plane Filling Ct~rve.J
.. . .. - *

5o0
 Orientation
of points After Second Iteration
P1 o P2

Angle.s neg

Orientation Orientation
of point& of points
P.1 -4PP2 Pi go P2
PoeAngle.
Angles neg

Figure 4.9. A Plane Filling Curve (continued).

80
 V. FRACTAL GEOMETRY FOR GRAPHICS TERRAIN

One of the most widely recognized fractal images found in the literature is of
the mountain scene. This type of terrain modeling is perfectly attuned to the
fractal technique. The reason for this is that mountains are highly irregular
shapes, with a rough but consistent texture when viewed from a distant vantage
point. It is appropriate then, to introduce graphics terrain simulation techniques
through this model.
This chapter describes the theory and techniques of simulating mountainous
terrain with computer graphics. It provides the blueprint for fractal graphics
programming within R 3 by providing general tools and a methodology that is
easily adapted to many other modeling needs.

A. MODELING MOUNTAINOUS TERRAIN

For the programmer who fully understands the essence of the method of
fractal programmning introduced in Chapter 4, the movement into programming
,J*Iin R 3 is not difficult. The primary differences lie in the quantum
jump in
computing resources that are required and the requirement to perform the
generator geometry in R3 versus R 2 . The theory and technique of fractals does
not change substantially.
Chapter 3 provided a rough framework to begin the coalescence of fractal
programming into a workable technique. We need to develop a number of tools
from that chapter and use standard computer graphics techniques to manage
those tools. To this framework, we add new fractal functions which provide the
texture of realism for our simulated mountain.
1. The Artist's Model
One way for an artist to build a physical relief model of a mountain is to
use a frainwwork to provide structure to the model and a texturizing clay to
provide realism. The artist might use chicken wire on top of small boxes as the
frame with modeling clay as the texturizing element. His choice of clay is

predicated by the type of look that he wants to achieve. The chicken wire

081

S
provides an inexpensive and disguised method to quickly build the mountainous
shape and structure. This method minimizes the cost and time to build up the
clay.
The artist continues the modeling process after the development of the
basic mountain shape to achieve hues and contrast in the coloration. He might
achieve this by the use of natural lighting to cast shadows or by a careful
painting of prominent features.
2. The Fractal Programmer's Model
There is very little in science that is truly new or innovative. We borrow
the essence of the above idea to guide us in developing a model for the discrete
computation of our two dimensional picture of the mountain. This section
describes the process intuitively and leaves the implementation details to later
sections.
a. The Lattice Control Structure
The pixel space that we developed in previous chapters can be
divided into discrete cubic units by use of a concept from mathematics called a

latUiee (in our case we can view it as three dimensional graph paper). This lattice
serves as our controlling structure, the equivalent of the chicken wire structure
above. It is beneficial to build the lattice as a structure with well-formed
relationships, where the number of lines evenly divides the boundaries of the pixel
space and each line is a constant distance from its neighbors. By this method, we
do not have to store the lattice but can express it as a mathematical function of
the pixel space.
The lattice can be very useful in developing a rough approximation
of the mountain that we wish to model. This can be done in many different ways
but should resuit in a sick frame model of the mountain (a connected polygon
nmesh like that of Figure 5.1).
The frame can be developed through an online graphics interface
that allows the prograinuer to select a ground ietd plane of the lattice and
provide a means to visually select points for the rough outline (eusentially draw
the framework). This approach 6-useful when a particular shape is desired.

02

,Y.
 A frame can also be developed using fractal functions to pervert the
lattice into a controlled random shape from a given plane of the lattice. This is a
powerful method that can be controlled via bounds on the random tools,
heuristics or discrete functional bounding of the fractal function. This approach is
most useful when a class of mountain shapes are required but no particular
mountain needs to be modeled, i.e. when random landscapes suffice.
Alternatively, the stick frame of the w'-untain can be determined via manual
(hand computation) means. This approach is tedious and limited in its flexibility,
and is not recommended.
b. Surface Texture via Fractal Functions
The next step in the creation of a mountain is to provide the
gr..p ua day to cover our stick frame model. This clay is a fractal function which
closes the polygons of the frame model with an inductive process that provides a
continuous pixel surface for the entire structure of the mountain.
The initiator/generator paradigm is used. The initial set of initiators
is the frame described above and the generator is a similar geometrical shape that
reduces in size continually until it becomes the size of a pixel and is mapped.
After the stick frame of the mountain is developed, this texturizing of
the surface becomes an automatic process that terminates when each geometrical
shape that makes up the framework is reduced to a continuous set of pixels in the
pixel set. At this point, the mountain exists in the pixel set (memory) but must
be provided color and light to bring it to life.
c. Hidden Surface Eliwination
The pixel set has the entire structure of the mountain in memory,
but we can project only a two dimensional image of one phlae onto the screen.
The fractal function which texturizes the surface does not concern itself with
local computations so many overlapping pixels are mapped to the pixel set.
"There are two reasons then, why we need hidden surface removal (in this case
better referred to as hidden pixel elimination). First we have to eliminate the
back or hidden sides of the mountain by projecting only those pixels which are
visible along the axis of sight to the perpendicular planar surface of the display
screen. The second kind of hidden pixel removal is caused by mapped pixels

3 -

V! Ii tR
which were covered up by other recursive fractal descents either before or after
the pixel was mapped.
The removal of hidden pixels is greatly facilitated by the use of the
concept of the pixel set. In standard computer graphics hidden surface
elimination, the programmer is confronted with graphics primitives which are
functionaUll continuous Euclidean shapes. To effectively remove the hidden parts
of these shapes is in general very tedious and mathematically complicated. Since
the graphics programmer is shielded from the primitive -, pixel mapping, he is
functionally denied access past the simplifying abstraction 26 of graphics Euclidean
primitives. The fractal programmer must have access to this level of the graphics
mapping and thus can use simple techniques to determine if a pixel is hidden or
visible.
The simplest and most economical means available to provide hidden
pixel elimination is through the use of Z-bvfer algorithms. With this method, the
"determination of whether a pixel is hidden can be appended to the pixel set
mapping process. The Z coordinate of a pixel that is to be mapped is checked
against the Z coordinate of the pixel currently in the pixel set at the same row
and column of the three dimensional array used to store the pixel set. If the pixel
is closev to the planar surface of the display screen, then the Z coordinate is
changed to reflect the position of the newly mapped pixel.
The Z-buffer approach, while powerful, does limit the fractal
programmer's flexibility. The axis of sight toward the mountain must be
determined prior to the fractal recursive procesm so that the determination of the
line through the (now two dimensional) pixel set is known. The Z-buffer becomes
an adjacency matrix to the pixel set and can retain information about forwardly
displayed pixels only. All information is lost about other pixels that were
computed in the fractal proces. If another view of the mountain is required thent
the entire pixel set has to be re'couiputed with a new axis of sight. If the fractal
function uses (oon-tabular) random tecuiiques then the mountain varies with
each view.
• The abutraction protvid( by"Euclidean pimitives is a powerful one whtn the alternative od
pixel rnappitng is onsidetred. Without roite powerful mapping tool (such as fraital funct.44"), the
pixel te%-*I modeliag proctS i to graftal vety ddiicult.

64
 Most fractal pictures consume such vast computing resources that
only one view is computed for a given picture. As more requirements for graphics
terrain are determined, a more powerful method has to be used to retain all of
the computed pixels in the three dimensional pixel set. This method requires
that all pixels be stored in the three dimensional array previously described. The
hidden surface calculation can then be performed during the pixel mapping
operation or as a separate calculation that is performed after all fractal recursion
has terminated.
As specified in Chapter 3, the full array approach requires large
amounts of memory. This method, however, allows the computed fractal
mountain to become an entity that can be manipulated versus an instance of the
fractal mountain as above.
Both methods are viable but the latter approach provides more
flexibility for the programmer whereas the first approach is a response to the
economies of scale of data processing. As new architectures are developed27 with
capacities geared toward fractal image computatit- the first method can be
eliminated.
d. Illuminating the Mountain
If you stood on the dark side of the moon without illumination, the
•'S. mountains and craters of the moon would not be visible. They still exist however,
just as our imaginary mountain exists in memory. In order to visualize them, we
must illuminate them.
x Illumination in computer graphics is achieved by varying the light
intevisities of pixels displayed on the screen. The color mixture of these discrete
points determines the lighting effect that a viewer perceives. This perception 's
not reality but another deception caused by scale and composition. A lighting
11o0del then, is one which is able to abstract the csesuce of color from a real world

- An Ideal athartur# is one with a largt main memory and pa•allel pwoeuins capabilti"
*a% K ptoceuinag tiwuats (wmhmt K is gmtcth aa the matrmum itcursite descent dattnce).
 object and transform that essence into a set of color values (intensities) that
accurately deceive the human eye via the graphics medium.
The literature on computer graphics contains many lighting models
with diverse approaches to the same problem. Many of these models (like those of
hidden surface) concern themselves with illumination of continuous Euclidean
surfaces and as such, are not directly germane to our studyzs.
An object i4 space is a composition of basic elements. These elements
interact with the physics of light reflection to create the spectrum of light that
our eyes decode. In a graphics image, this process has to be simulated with
discrete lighting intensity values for each pixel. Thus, the illumination of the
mountain is a two step process; the fractal entities that are mapped to the pixel
set must be provided with a basic color, and these colors have to be highlighted
and dimmed by the lighting algorithm.
The basicS.color can be determined during the pixel mapping event of
the fractal recursion process or as a separate process prior to or in conjunction
with the lighting algorithm. This color can add realism to the picture through
heuristics which the programmer defines. Most mountains are composed of
different types of rocks and flora and these elements change at different altitudes.
This type of heuristic combined with some random control structure (i.e. to vary
the snow peak) can provide for improved realism (versus umaking the whole
mountain brown). The process of dettrmining the basic color must be
accomplished prior to applying the lighting algorithm since the lighting algorithm
can only vary the intensities of an existing color". Developing the process rif
basic color determination is best accomplished through trial and error. It is the
artistic aspect of developing fractal mountains.
The general process of computer graphics illumination concerns itself
with casting Aadowis from one object to another given a direction from an
imaginary light source and willh higllighting surfaces which are directly expmsed
to the source. A surface is highlighted relative to the angle at which the Igght

k,÷: u Thr Goturad model (taunkty intatp"Iatiae # shading) tUw •stan•o e

%.nt $0 0MAy diveM Cokoi mo kh tut. 4 the dtakil Of Co0 rePrMets o;00 staOct

~W,,
 source's rays strike the surface. This poses special problems for fractal surfaces
due to their discontinuity at every point.
The process of illuminating a fractal surface is best aided by
divorcing the lighting process from the fractal computation process (except as
noted above). It is beneficial to view the pixel set as a collection of pebbles which
have size and position. This abstraction allows us to view the pixel as a
continuous space that can block light (cast shadows) and for which an angle of
illumination can be determined (usually in conjunction with neighboring pixels).
A well formed fractal mountain surface is completely connected (no
space between adjacent pixels in the pixel set). Thus the surface can also be
viewed as a continuous (while very rough) surface where reflected light can be
cast from or to adjacent pixels.
One lighting model which fits the fractal process is the Towrwnm
Sparrow model [Ref. 8:pp. 578-579j.
This model views an obiect as collection of fqcets which is each a perfect reflec-
tor (i.e. does not absoi.b light). The orientation of each facet is given by the
Gaussian probability distri ution functior (i.e. the smoolh surfaie of -;%e Eu-
clidean object is roughed by the Gaussian relationship). The geometry of the
facet and the dirtr.ti6rp of light (assumed to be from an infinitely distant source,
so all rays are parallel) determines the intensity and direction of specular reflec-
tion as a function of Ihe light source intepsity, the normal to the average sur-
face, the direction to the light source and the dirct1ion to the viewpoint.

"Thismodel has to be modified to adapt to the fractal s. method. In the fractal

method, there is no need to rough the .surface to provide reflection because the
-surface 6 by design roughly textured. A method of assigning planar front, to
each pixel space has to be determined and the geometry of coneatingZ i
frowitN identified. With these modifications to the lighting model, each individual
pixel's color intensity can be wvdifed for the increase in intensity associated with
the light which falls upon it.
The model xlso allows diffuse reflection (light reflected from one
object to another) which is ciritirni to bring out clarity of the fractal i0a-ge. For
further information on the indel the reader ii referred to the rf•-ýroce.
"A4

67
 e. Summary of the Fractal Mountain Paradigm
To summarize the methodology we can view the process as a five
step process:

- Build the initiatorframework or stick frame model of the mountain.

- Give the frame's surface te.iure with fractel functions.
- Remove hidden surfaces (pixels) from the display.
- Illuminate the surface with lighting algorithms.
- Project the surface to the screen.

B. FRACTAL TOOLS FOR TERRAIN MODELING

The tools presented in this section can be used in the creation of fractal
images within R3 . The list provides a basic set of programnming tools to guide the
creation process.
1. Equations of the Lattice
The lattice (or controlling structure) can be very useful to the graphics
programmer to implement heuristics or bounding functions on the eCs,'ntially
random progression of the fractal figure. The graphics programmer may wish to
limit the growth of the mountain by implementing a ground level plane of the
lattice and a maximum height that the mountain can obtain. He accomplishes
this by arbitrarily assigning another plane of the lattice as the upper bounding
plane. The height of the mountain can then be checked during any level of the
fractal recursive descent against this fixed plane. The programmer can then clip
the height by adjusting the random equation that controls the upward trend to
tend towards the ground again. This is an example of a heuristic applied to the
fractal recursion that controls the external qualities of the function.
A fractal programmer can use the lattice to assign the initial colors to
the mountain via a user designed seý of rules. The lattice aids the user il the
implementation of the rulks by giving reference points for inclusion of branching
conditions (tree line to snow line etc.) and can be used in conjunction with

of the initial color of a mapped pixel is usually a controlled raindom piocess, one
of the primary mi...hods of control being the lattice or some derivative thereof. As

68
 the height of the mountain increases (lattice level), it becomes increasingly more
likely that it will transition to another texture. This can be controlled by adding
the lattice level as a factor to the rule that decides color.
An example of a potential lattice equation and how it might relate to the
pixel space is demonstrated in Figure 5.1. The actual lattice has been extended
from the pixel space in order to visually demonstrate how it relates to the pixel
set. In actuality, this is not the case. The lattice coincides with the boundaries of
the pixel space. Although the lattice can have a one-to-one relationship with the
pixel space, this defeats the purpose of the lattice (macro control). By grouping
cubic sets of pixels into a well-formed relationship, we can better implement
heuristics and bounding functions.
As a lattice example, consider a pixel space that is created by abstracting
the real world coordinate space for our mountain as described below. We desire a
real world space to be a cubic area established by the box 20,000 ft.
(z coordinate) by 15,000 ft. (y) by 20,000 ft. (z) This can be seetioned into a
lattice by establishing the increment of distance between adjacent lattice points
to be 1000 ft. and establishing the corner lattice point as (0,0,0)30.
The mapping function between the lattice and pixel space is then
straightforward. The size of the pixel (recall equation from Chapter 3) is
20,000 ft. =20 ft. and the lattice cubic sections contain 108 cubic feet or
1,000 ft.
equivalently 75,000 cubic pixels.
The ground level car&then be identified as the 20b0 foot level and the
bounding height can be assigned a level of 10500 feet. If we wish, we can make
the bounding heuristic more realistic by sectioning the lattice into mountainous
areas, each having different bounding levels.

J1.

SA compkuit|y arbitrary set of dimensions, iacremetas and powint.

Il ..,16[
= _. • 4
 The Lattice Control Structure

-- 3:

LATTICE
FRACTAL lIDUNTAIN STRUCTURE
PIMk SET

Fig 5.1. The L1 attice Control Structure.

Z4 70
 2. A Fractal Function for Contouring Mountains
The usual method for contouring mountains uses a randomized variation
of the mid-point displacement method introduced in Chapter 4. The planar
the triangle31 imbedded in R 3 . The basic methodology is
is typicallySstructure
demonstrated in Figure 5.2. Figure 5.2.a shows a triangle with its first iteration of
mid-point displacement. This process continues until all triangles have reached
the desir,-d level of precision. One completed structure is demonstrated in Figure
5.2.b. TL, precision is typically lower (pixel level) than that demonstrated in
Figure 5.2.b but it was terminated at a higher level to better demonstrate the
idea. Random techniques (described below) are used to produce the relatively
accurate picture of a mountain frame as depicted in Figure 5.2.c.
In praitice, the random techniques are implemented with the mid-point
displacement function during the fractal recursive descent. The random
* techniques provide local disorder to the fractal function which provides the
computationk, structure. Results have shown that very little randomness needs to
be applied to the regular structure of Figure 5.2.b to achieve satisfactory results.
The mountains created for the film Star Trek: The Search For Spock used a
limited random number look-up table consisting of fewer than 300 entries [Ref.
71.
3. The Geom Zry for Mid-Point Displacement
The general approach to building a fractal shape as illustrated in Figure
5.2.c is to use the algorithm of midpoint triangle displacement combined with a
randomized displacement along the normal to the X-Z plane of a cartesian three
space coordinate system. A recursive procedure which computes this relationship
requires as inputs the points of the triangle. It comnputes the midpoints of each
line of the triangle wad inscribes a triangle inside of the initiating triangle by
connecting each midpoint, Figure 5.3.a. This process yieds four triangles
coincident with the plane of the initiating trisngle. When we fix the X-Z normal
at any of the midpoints, we can displace the midpoint by a discrete distance

""Anyregular structure sufficts; the triangle is emy to use aad yields veqr s&isfactory
mutts.

71

4W...* N -TtT
 The Triangular Midpoint
Displacement Technique

Det.ermine Connect Connect Connect

Midpoints MA to M2 Ma to M3 M3 to MI

Fig 5.2.a The Ist Iteration of Midpoint Displacement.

The Random Structure was rotated

-30 degrees around the X 4xLis to
accentuate its texture.

4Fig 5,2.b Completed Structure Fig 5.2.c Randomized Version.

Figure 5.2. The Triangular Midpoint Displacement Technique.

-S~ .72

.I ,
 along the normal and determine a point, Figure 5.3.b. Since the normal is to the
X-Z plane, it is sufficient to simply modify the Y coordinate according to a
'positive or negative value. This is equivalent to displacing the midpoint along the
X-Z normal up or down. We perform this displacement to each midpoint normal
and replace the midpoint with these new points. This yields a new structure that
still consists of four triangles but with each coincident with a different plane,
Figure 5.3.c.
3
a. Midpoint of a Line in R
The determination of the midpoints of the lines of the initiating
triangle is a simple process that uses the equation of Chapter 4, and fixes the
generator ratio constant at 1. This simplifies the general equation of:
Xi + (Generator.ratio.constant x X2 )
Xmld =,
1 + Generator.ratio.constant

to the well-known midpoint relationships of:

X1 f X2
Xmid = 2
Ymld = Y1 2+ Y2

~Znid ZI .+ Z2
2
• 2
The above equations completely determine the midpoints of the lines formed by
each endpoint of the initiating triangle.
b. Displacement along the X ýZ Normal
The process of displacing the midpoint along the X-Z normal is a
simple one. We need a factor such that the displacement can obtain a varied
magnitude. This is best aided by the inclusion of a random variable as a multiple
of sonie scaling factor that is added to the Y coordinate of each computed
midpoint. This process is demonstratcd in the following code seguient:

Randvar getrand(Seod);
Pointl~y] Poiatily] + (Scale° Randvar);

44 '73
 P3
are
Four Triangles &
3

for each
created
M Ma
triangle initiator m

P, Uj Pa

Figure 5.3.a. Triangular Midpoint Displacement.

y axisComputed isplaceAent

z i sAfter lot displaeseaea

... . . . .. . . . . .

Figure
3 c cmplete Rando placneTiage

Figure 5.3. h-b dm ipit

Displacement
AlogchniqueNrma .

7'
 A valid question is, why the normal to the X-Z plane? There are
three good answers to this question. Using the normal to a fixed plane simplifies
the computation (eliminates the need to perform planar computations at each
recursive division). It also is generally the direction that we want the mountain
to grow. The most important reason however, is related to the gapping problem
(described below). With a fixed direction for displacement, there is no need to
communicate the direction of displacement along the, normal between adjacent
side computations. The recursive levels that compute adjacent sides are
functionally discordant. It is demonstrated below that the solution to the gapping
problem (inconsistent random numbers) which creates the need to communicate
along discordant recursive levels is algorithmically difficult to solve and thus
should be avoided.
c. The Gapping Problem
One problem exists for the midpoint displacement procedure which
utilizes a random displacement along the X-Z normal line. It is indirectly caused
by the data locality aspect of the inductive process of the recursive fractal
descent. The problem exists when two adjacent sides of two adjacent triangles are
not displaced with the same value. Each side is computed during independent
levels of the recursive descent so there is no practical method to communicate the
random numbers for the displacement.
The gapping problem is illustrated in Figure 5.4. For the two
triangles that are extrapolated from the structure, there is an unknown
relationship that is the random variable used to displace the coniuon midpoint.
If triangle A uses Rand - 0.3 and triangle D uses Rand = -1.07, then the
displacement for each adjacent midpoint (whirh are at the start coincident) is
a -,skewed in the opposite d'rection. This creates a pp in the fractal landscape that
will (in all likelyhood) not be filled by other fractal shapes front neighboring
triangles. We need an algorithm which can insure that each midpoint (which is
always sharted by two triangles) has the saame displacement along the normal to
the plane.

i75

N
 d. Solving the Gapping Problem via Random Tables
The solution to the gapping problem is straightforward if the
programmer adopts the random number table as his random function
implenentation. The goal is to match adjacent triangles with a seed or
displacement within the random table so that the random number returned is
equivalent for each coincident midpoint.
There exists a symmetry within the triangle of Figure 5.5.a that
allows such an approach. Ideally we want the point Ma to be displaced by the
same magnitude when triangles T 2 and T 4 (highlighted by textures) compute
their random numbers for M.. This can be facilitated by the inclusion of a table
seed for each recursive call to the midpoint displacement routine and by rotating
the orientation of the midpoint triangle (the triangle created by the three
computed midpoints) labeled T 4 in Figure 5.5.a. This rotation is performed in
- relation to the random table and not in relation to the Cartesian space. It is
accomplished by adjusting the order of the points in the recursive call.

The order of the points for triangles T, T 2 and T 3 are as described

in Figure 5.5.b and for T 4 as described in Figure 5.5.c. All four triangles generate
a recursive sequence and use the same seed to the random number table. The
random numbers retrieved from the table must observe the order of asignment
that is demonstrated in Figure 5.5.d. For example, the line segment formed by
the first two points (Pi and P7) input to the midpoint displacement routine
determine the midpoint RI. This midpoint is assigned the rarndom displacement
from the table correspooding to the eatryuseed. The next midpoint retrieves the
table entry correspoidisig to -eed + I and so on.
If this technique is followed the %quezace of randomh numbers will
nmoeh-. up az denontrated in Figure S5,e. Tie recurive cals correspond to the
code sMeiumt in lFigure 5.6.

To
 The Gapping Problem
Generated random structure

Ini ri1t

. . . . ..... ........ ..... ......

ADJACENT TRIANGLES SHARING

A COMMON MIDPOINT

Sh common midpoiant

AFTER MIDPOINT DISPLACEMENT

Cap creoted by diacoanti'.uouw

r&ando. displaco"..pn oaloa;g

.•Shrod common aidpoioc

Figure 5.4 The Gapping Problem.

S77

*L
V

Solving the Gapping Problem

. Shred c a midpoint

i ~~~~ .......... :.... ......... :...:.

Figure 5.5. a

p1 P2 PP

¶4Figure 5.5.b Z.R.I Figure 5.5.c

£
2 P2
"-i- Given that the input to the random tebJi io seedai

A Figure 5.5.d

ti

.3 Figure 5.5.e

"Figure 5.5. A Solution to the Gapping Problec.

78

Wi7
 LOAD THE INITIATING TRIANGLE
Seed = 1;

frac tniangle(P 11P2,P3,Seed)

frac. trlangle(P3 ,P
2,P3,Seed)
f
DETERMINE DISTANCE BETWEEN ENDPOINTS OF AN INITIATOR

If (DIST < Pixel.length)

@1 PlotypixeIO;,
return;

COMPUTE THE MIDPOINTS (MI.M..M 3 )

ADJUST THE Y COORDINATE FOR M, Using Randtable(Seed)

ADJUST THE Y COORDINATE FOR M2 Using Randtable(seed+i)
ADJUST THE Y COORDINATE FOR M3 Using Randtable(Seed+2)

Seed Seed +3;

/* Triangle T, '
frac triangkv(M,,P2,Mj.Sccd)
P* Trimigle T2
fractrian lc(NIM,3MP 3 .Seed)
/* Tris.igl T:,~
frac triwirlv(P 1 .M,,M3,,Sccd)
/* Triazigle T,4 /
frat-triazagk(NI2,M 3 ,M1,SeCd)

Figure 5.6. An Algorithm for the Midpoint Displacen-ent Techniquc.

 4. Random (Stocastic) Fractals
One common complaint about computer graphics images and animations
is the artificial perfection of the displayed shapes. Our mind eubconsciously rebels
against the order that is displayed, our expectations about the rough reality of
nature are not satisfied. The use of randomness in generating fractal images is
necessary to approximate the observed disorder of nature. An example is the
Koch curve. Although it resembles a snowflake, it lacks the realistic look that
experience trains our eyes to see. In a mathematical sense, the Koch curve is
beautiful; as an approximate to nature it lacks appeal.
To approximate the rough texture of nature, we are forced to modify the
well-behaved mathematical relationship of the initiator-,generator in a controlled
manner to add variety to our computed image. This modification is usually by
the inclusion of a random variable into the control structure widhin the fractal
equation. The random variable must exhibit restraint. It cannot be allowed to
vary wildly without structure.
One of the most appealing random functions which prowides very
satisfactory results in frrctal images is the normal distribution3 The normal
distribution (as opposed to a uniform distribution) approximates the -xpected
local disorder in nature (at least experimentally).
a, The Normal (Gaussian) Distribution
The normal distribution is used throughout the natural scitce fos
, many applications. It was first derivei as an empirical result of the observed
error aboit a true value that normally occurs when teaseuraw4aslt are taken 4" a

natural event. The symmetry that was observed fron error measurement and
"sanpling svgeated that there was a natural order to such observations. These
Cetpirical results spurred natural scientists and watheinaticiims to try to fit a
curve to the observed graph that behavec as probability requires (i.e. the sum of
the area under the curve equals unity). Many of the early scientists

- Often referd to " the CGattutan doinbutooa. the aoamal d-trinbt.t60 W tke st4.drd b1*0
curve to whidh every utudeat tis &cUOM@ad

80

so.-.. . ,

V~****
 referred to the normal distribution as the law of error ia deference to its roots in
experimental natural .,ience.
Many functional charactvrl - s6 of the normal distribution were
developed33, but credit is usually attriwuted to Carl Frederic Gauss (Ref. 9:pp. 1-
11i who formulated a least squars approach, published in 1809 in Theona Motua
Corporum Coelestium. The form of the normal distribution was not finalized
until the early 20 'h century.
We take our definition of the normal distribution from [Ref. 9:pp 181.
refer to Figure 5.7.
Definition:
The probability density function of a normal random variable X is given by:
Sis the mean, a is the standard deviation
and 03 is the variance.

ex4 tx-012
cPV21 2oz
tAhere co<x< c -oo< p <c andv>0

The Normal Distribution

SCALBE
'hors kSer.t /

pL-o ,p jp4

Figuire 5.7. The Normal Distribution.

Many tthfotl~tlt caq lay daiem 10 Fourw tht ial datonbut, moo n~oiblb. P1.
trt S4mat de Lapl"@ and Abralam 4e Mosvre R#( 9pp I I

':' 81
 B• * This definition is the general case of the normal distribution. We are
interested in the behavior of the function and need a practical way to deter-nine
a random number that we car use in the parameter of the normal to the plane in
the geometrical relationship described above. To facilitate this, we simplify the
general normal distribution to the well known standard normal distribution,
illustrated in Figure 5.8. The standard normal distribution function is the special
case where # = 0 and or I. This reduces the general equation to the simplified
equation:

f8(x) - exp 21

The above functions describe the behavior of a normal random

variable. We need a function that returns values from that function which will
observe the period of the normal distribution. This means we need a string of
real numbers , an assigned range about a mean that will observe the
frequency of the normal distributioii.

SThe Standard
.Normal (Gau.sian)
SCALE: Distribution
01* 1 horz a 5vert
-
0
0.0
c=-1.02
- 0.
N

N.(x) 2

' -3 -3 -1 0 1 3 3

j- 09. 37X-

Figure 5.8. The Standa rd Normal Distribution.

82
 b. Standard Computer Random Functions
Some computer systems provide a random number generating
function which observes the normal distribution. If this is provided, then it can
be used directly (after scaling) as a parameter to displace the Y coordinate in the
geometry of the normal to the midpoint displacement as described above.
Many computer systems only provide a random number generating
function which is uniformly distributed over an interval of integers. This is a
p8cudo-random number. Such a function, when given a seed, will produce a
sequence of numbers distributed over the fixed interval defined by that system.
The interval is typically proportional to the maximum integer defined in the
compilers of the system. A normal distribution routine must then be defined that
transforms the uniform random numbers into random numbers which behave
according to the standard ncrmal distributLon function.

There exist transformation functions that take a unifoirm random

0 variable distributed over the interval [0,1] into an approximate normal random
variable over -oo < x < 0014. This requires the uniform random variable to be

mapped into the interval [0,11 and then transformed by the normal
approximating function.
To transform a uniform random variable distributed over an interval
r1 [O,maz int] while maintaining the distribution density, requires the following step:
UNF 0,mazjnt I
*UNF 1 0, = maz int

One commonly used function that transforms uniform random

* variables into normal random variables is found in [Ref. 9:pp 49]. This function
uses two uniform variables from [0,11, denoted UNF1 and UNF 2 , and computes
two normal random variables, denoted NORM, and NORM 2 .

NORM, = ,/--2logeUNF, cos(2irUNF 2 )

NORM 2 = •(-21ogeUNF, sin(27rUNF 2 )

"34 This is how most standard system provided computer subroutines perform the operation.

: -S." 83

- -. .
 Appendix B contains a C UNIX routine that implements an algorithm to
compute the uniform[O,1] -- normal [-oo,+oo] transformation.
A programmer must be very careful when dealing with random
number generators from standard system subroutines. These routines vary widely
and can provide good to barely adequate results. When the normal
transformation routine is written, the programmer must verify experimentally
that his function adequately models the normal distribution. This process is
illustrated by Figure 5.9. Appendix B also contains experimental results which
verify the transformation.
The purist may not accept the results displayed in Figure 5.9 as an
accurate transformation (there appears to be a skew to the negative direction).
We must remind ourselves that we are trying to approximate the roughness of
nature and minor random skewness will not deter us. If the programmer demands
a better approximation, it is a simple process to expand the sample space of the
0,L test and build a tablk with exact proportions by selective deletion of skew
density.
c. Random Functions versus Table Driven Methods
The application of a random modifier in the midpoint displacement
technique can be achieved via two methods.

- By invoking the above function iteratively as a variable.

- Or by a variable returned from a table lookup operation from a random
"table.

The choice of which method to use depends on the programmer's application but
"each has its ramifications.
In general, the table lookup operation is considerably faster than the
functional method but must by its definition limit the amount of randomness it
contains. The major issue however is the need to reproduce a figure under some
requirement for fixed terrain. This issue was the driving force for Loren Carpenter
from Lucas Film in determining that he needed to use a table driven method to
produce the planet images for the film Star Trek: The Search for Spock [Ref. 7J.
He had to be able to fix a space where the images of the actors could be imposed

84
 Experimental Results for the
Computer Generated Normal
Distribution
Sample Space:
.500 Random Events,22

IxPerimentsl Results
With a Normalised
Transformation Equations Uniform Distribution
Over [0.0,1.0]
From Uniform Distribution
Over [0,I] to the Normal88 8
Distr~ibution Over -w(x(~~~w.:
. 8.. ...

Sace

202

-3.0 0 03.
~ ~ Exe.ena.esl.
Figure ~ ~ ~~.. ... fo. omue
generated norma
~ ...
d .trbuio ~
ove..

..85..
. .
onto the fractal images and coul)d not allow the fixed space to change with each
frame computed. This is the major advantage of the table method. By retaining
a 8eCd to a table of random numbers, you can reproduce the sequence of
displacements along thew normal during the fractal recursive descent.
When you consider the existential quahi4ies of randomness you are
cowEfronted with basic questions about determinisip antt order in the universe. It
i: not. at all clear which rulec chance. IV siijr rse, we can deceive perception
with a relatively small table of random numnicrs.
The question of how much rantdzm,-.,ewz is enough to provide for a
visually appealing texture is not completely clear, In [Ref. 71 Smith demonstrates
a variety uf shapes computed with the same algorithm of Figure 5.6 using
random n;xmbei, tables of different oizes. He demonstrated that as few as five
numboers can &uffice to provide enough local disorder to give the viewer the
acceptable textthr of a mountain. If the mountain segments are viewed at the
correct perspective .si thig perception is clearly felt.
scaje, A trained
aiatheinatician would find the five element mountain siatistically unappealing
however. A true st-,wlhatic consi•,u-tion requires a continuous random function
irather than a discrete table method. As wiog as the goal of our computations is
mnerely to deceive ihe graphies viewer, it suticka to use the random number table.
The table must be large enough ,'Zprovide for an appealing textlural perception.
A complete C program that c'omputes a triangular mountain segment uisip. the
random displacement midpoint technique is contained in appendix C.

86
 VI. SHORT CUTS TO MOUNTAIN SHAPES

Since the fractal mountain computation (the full approach with hidden
surfaces etc.) is so costly in terms of resources, it is important for us to consider
shortcuts that can lessen this burden. This is best realized by utilizing the hidden
surface and curve fitting capabilities that are provided on some advanced
graphics systems.
Our goal is to match the well known bicubic surface procedures with the
structure computed by the simple fractal algorithms. This is best accomplished
by modifying the triangular midpoint displacement technique and using a
rectangle 35 as the basic geometric building block. Most of the cubic surface
algorithms use the rectangular structure as their basis, so it is easier to adapt
them to our fractal structure.
When the fractal algorithm of Figure 6.2 has its computations terminated
before reaching the level of pixel size, it yields a connected rectangle structure
like the one shown in "tiguwe 6.3. This structure is a connected Euclidean
structtore that can be used & a Dse on which other algorithms can be applied.
Cubic equations cati fill the polygons to an arbitrary precision and standard
hidden stvirfate aOorithms caa eliminate the hidden sides of the computed
surfaces Simple lighting algorlZ'Ims can be applied to the computed surface to
achieve f realsl.ic iitiyig effect'. This is holw Voo and Carpenter created their
fracwal ,urfac'es in IRJES. 7).

A. RECTANGULAR WDPOINT TEi'CHNMQPGk

Molifying the triaagufar ingpoint algithin •f chapter 5 is a straigh•forward
process that i-ttroduces no iiew niath&enatic' or 'difficulties. It consists of a
proce•,re to splii the midpoint.t of euwi side nf tho 'ectangle and g procedure to
find the ceter of the re•tanyk1•", From these lfvQ points, we construct four scaled

35 We actually uxqe a non-plantar fgo gided polygm.- We refer to the basic structure as a rec-
tigle Lo sittplify the termimiogy8
"36Gourau%,. shading fo~r exatuple.
 rectangles, as demonstrated in Figure 6.1. The five shared midpoints of the
generated rectangles are then displaced along the normal to the X-Z plane
according to a random Gaussian value. This process is exactly the same as for the
triangular algorithm of chapter 5. The gapping problem still exists and this
requires an algorithm to rotate the rectangle relative to the random number table
and the starting seed to insure that adjacent midpoints are displaced relative to
the same random number. The basic methodology is displayed in Figure 6.1, the
algorithm is contained in Figure 6.2 with sample results in Figure 6.3.

IN

88
 Computed
Four Rectangles Midpoints
are created for
each rectangu-
lar initiator.

Computed displacement

Y axist

a x a..x.......

Y.
v ......
... .. . ..
l.. .......
.. .....
.

..
.. .
...
 Main()
{
LOAD THE INITIATING RECTANGLE
Seed = 1;

frac rectangle(P1 ,P 2,P 3 ,P 4,Seed)

frac rectangle(P3 ,P2 ,P 3,P 4,Seed)

DETERMINE DISTANCE BETWEEN ENDPOINTS OF AN INITIATOR

If (DIST < Pixel.length)

{
Plot point();
return;

COMPUTE THE MIDPOINTS (M 1,M 2 ,M 3 ,M 41Mc)

ADJUST THE Y COORDINATE FOR M, Using Randtable(Seed)

ADJUST THE Y COORDINATE FOR M 2 Using Randtable(Seed+l)
ADJUST THE Y COORDINATE FOR M 3 Using Randtable(Seed+2)
ADJUST THE Y COORDINATE FOR M, Using Randtable(Seed+3)
ADJUST THE Y COORDINATE FOR Mc Using Randtable(Seed+4')

,S•ed =Seed+5;
/* Rectangle R, '/
frac rectangle(P,,M 1 ,McNM 41 Seed)
/* Rectangle R2 VI
frac rectangle(M 2 ,MCM 1 ,P2,Seed)
/* Rectangl R3 R/
frac rectangle(M 2,MCM 3,P 3,Seed)
/* Rectangle R4 V/
frac rectangle(POM 31,MCM 4 Seed)

Figure 6.2. An Algorithm for the (Rect.) Midpoint Displacement Technique.

14

4rNgo
 Rectangular Mountain Fractal

CenrTz rectce~a2

The recursive tefrmonation eveao ws, J/2 iocb.

"Figure 6.3. An Example of the Rectangular Wo-:- m-,tctal.

A.9
 B. PARAMETRIC CUBIC SURFACES
A complete description of parametric cubic surfaces is too involved to be
'I described in * study. The theoretical basis of cubic curves is not directly
applicable to &;'.ýctal geometry. For a complete description refer to [Ref. 8:pp.
S514-536]. If the reader is already familiar with cubic curves and their derivations,
he can skip by the section on cubic curves to the section that details the
application of cJ-)ic surfaces. For apy reader who has not been exposed to the
derivations of -a'anmetric equations which yields cubic curve computational
engines, it is recommended that he read the following section so that he may gain
insight into the mathematics of cubic surfaces. Detailed knowledge of cubic
curves is not a prerequisite to the successful use of cubic surface fitting engines
with respect to fractal surfaces. It is helpful, however, to understand the
underlying mathemnatics whenever conned equations are used.
1. Cubic Curves
The general method of cubic curves has as its basis that any continuous
curve in R 3 can be expressed in parametric form. This form relates the points
z,y,: with a parameter t such that as I varies within some range of values 37 the
equations solve for unique points on the curve. Specifying two endpoints and two
control points of a segment of the curve allows us to define certain constraints to
be applied to the parametric equations. These constraints allow us to mlauipi4mte
the parametric form of"the equations to yidd a simple vector produc't definition
of that segment. Once this vector product is established, we caan solve for points
on the curve by picking discrete values of I and solving for z-,z in turn. This
yields a discrete approximation of the curve that can bW as prtcise as needed.
a. Parametric Cubic Equatimis of a Curve
A parametric cubic curve is one for which the points in R 3 (x ,y.:)
are each represented as a third-order (cubic) polynomial of NoGIV paraWeter t.
IBecause we deal with a finite swgiznt of a curve, we limit the range of the

•I •: I may vary between 0 and I for exarnplt

49 02

e.l
parameter t to the range, 0 •<, <*1. This yields the equations:

x(t) = at 3 + bxt 2 + Cet + d'

3 + byt 2 + eYt +
y(t) = a0t
£(t) = a~t 3 + b~tZ + ct + ds
Each equation can be expressed as a vector product as x(t) is below:

ax
bx
x(t) = It 3 tl t 1] C

This vector product separates the distinct parameters of the parametric equation
into the unknown coefficients of x(t); faxb~cd, and the parameter t that we
wish to manipulate. Through this separation, we are able to manipulate them as
algebraic entities. If you multiply the vector product out, you find that the
vector product is equivalent to the parametric equation that precedes it. Denote
this product as x(t) = TC,, where

T ! 3 t3
t ' t ii
aMid

ia|-x

dx

The vector T iMthe same for x(t). y(t) and a(t).

We now estabhl'. constraints (as a set of control poinjt) for the
equation. (t) evaluated at the bounds of the range of the araueter *, (i.e. t 0
and ti). We considrr four equations of x(t) and its first dcrivative x (t) wherv
thrse boundary conditions yield four known points.

x(t)--T(C.; when evaluated atam0. --. x(O) - 0 00 JC

x(t)- TC( ; when evaluated it-• 1. x(l) j1u 1
11 Cx

93
 and since the first derivative of x(t) is:
x'(t) = 3t 2 2t I O]Cx = T'Cx

x'(t)=T'C.; when evaluated at=0, -. x'(0) = 10001 0]Cx

x'(t)= T'C.; when evaluated at= 1, -o x'(1) = 13 2 1 0]Cx

We now have four equations that can be grouped into a vector product:

x(o) 1 0 01
x (1) 1 1 1 1
x'(o) = 00 1 X
x'(1) 3 2 1 0

We recognize that x(0) and x(l) are lihe endpoints of the curve
segment and x'(0) and x'(1) are componeints of the tangent vector at the
endpoints (y'(t) and 2 (t) are the other components). With this knowledge we
are able to solve the left hand side of the equation above. These points (that we
call P, through P4 ) are the control points that we establish for curve fitting'.
For a given curve segment the control points are fixed. We rewrite the equations
above with respect to these known control points:

P , 10 0 0 1
P2• I I 1 1
IP31
P. 1 000 0 1 0 1 'X

P4 3210

Delloto this equatiuo as:

SGz M.MCx

The matrix (GO% is often rteerred to as the geomwtry of the cubic curve and M as
the b~s
This equation hex the 4 by I tow vector Ca a the only unkilown.
The eliment,% of thi( CS vector are the pranwtkrs ('fbg~c,,dx) from the

" WC UW',t&A opJV*O 1 0b4

•-en iht tbe four poaa UT th t sa
Wiput to oua rou-
""P-

V%
~9
 parametric equations. We can solve this equation for these parameters and
establish the parametric equations with the only unknown being the parameter t.
The parameter t can be discretely varied over its range of 0 < t < 1, providing a
set of points on the curve. It is through these constraints that the control points
conlrol the parametric equations and produce an equation thrt can produce a

discretely sainpleable curve segment in three space. Solving the equation for C. is
straightforward:

Cx = M-'Gx

Substituting CS into the equation for x(t) yields39:

x(t) = TM'-'(;

!ziaiilar arguments yield the equations for y(t) and z(t):

y(t) TM- Gy
Z~t) Im c

"The "iratrix M is constant for all three equations and is usually

Sdvý&otuA by- tl., type of surfaze that it relates to Bezier -. Mb, Hermite ". Mh
etc. It is-through the control lwints and their interaction with the constraints
that the models l3kzier, B-ý4pi.., Cwratdrul Spliuw, Fergus.oi (Herwite or Coon"')
surfacee Ott. uodify tiwe praiaetric equationu and provide diferent 'CurVe fitthig-

For-each model. the mitrix N-t,,dt M €O ASMt throughout PI1

Comsutat-iot1o. To use the model retquirc% the deerintu f the control, point,%
(in otnjunction with how thoy, reoate to til 'Curve) and a vitor multiplication
""nlgiiw. Sior- voctor piltwine t¢o!pUt4ns are h-idally $Uitfd to Computers, th•.
-i.iio hbecomns a fut tochnotogy for curve fitting with an i*UWlViwC appeal for a
4i}. pro~raz~wnr.

We-€ have juim 4prmanmn Ah ihtrmct vnodd equal mao wia).

I05
 b. An Example: Bezier Cubic Curves
We consider the model called Bezier [Ref. 8:pp. 514-536]. The Bezier
model defines the position of the curve's endpoints and uses two other points (not
on the curve) which define tangents at the curve's endpoints (by the line segment
joining the tangent points to the endpoints).
The matrix M is derived by setting the following constraints (see
Figure 6.4). One endpoint of the segment is located at PI:

x(0)= P1

The other endpoint is located at P 4 :

x(M)= P 4

The line segment from P, to Pý defines a tangent at P1 such that x'(0) relates
to the points PP: as below:

x'(0) = 3(P2-Pl)

And similarly for the tangent at P 4 defined by P 3 ,P4:

x'(1) - 3(P 4 -P 3 )

|ii Solving for C. in terms of Mb yields the cubic Bezier niatrix as:

- CX - 33 -63 3 00 GX
()
Hec h orxt s -1to 3 3
1 0 0 0
-1 3 -31
Hence th~e equation for x(t) is:

- 3 -3 1 IP

3 -6 3 0 [P,
xt)[, t €zj_3 3 0 0 IP31
1 0 0 0 p4

100
 Bezier Curve
.rqent,defined by
P1 P 2 •line segment. /12

SP4

at defimed by

PP3. saeat
gie

Figure 6.4. An Example of a Bezier Curve.

The process of creating a Bezier curve given t0 above parainetric

cubic engine is a !imnple process of rotapting discrete points on the curve by
substituting values sAlong the rdnge of t and fitting the curve by connating each
po0,1t wilh a line Tegme'•.
This pro.'ides an approximation to the curve that ican
be proceshed at an arbitrary precision by incrementing 6t with smaller and
l ~ ~,•maller '.eg |s

The process of shaping a curve is accomplished by increasing or

decreasing the two endpoint tangents formed by the four control points. It can be
viewed intuiiveq by thinking about each tangent as a force which pulls the
curve in the diriction of the tangent until the force from the other endpoint
overcornes ih. original at tlie midpoint. The two endpoint tangerits work against
one another proomrtional to the distance of 61 fron each ei dpoint.

Exteundian 1he above method to cubic surface sections is accomplished by

adding a new parameter s that we vary from 0 < j 1I as we did with the

4 07
A',
 parameter t in cubic curves. The connection between cubic curves and surfaces
can be made by fixing one parameter and varying the other over its range. This
yields a cubic curve. The equation is of the form x(s,t) and is written as:

2 3
x(s,t) = allS3 t 3 + a 12 s3t + al 3s t + a14s3

+ a 2 1s2 t 3 + a 2 2s2t 2 + a 23s2 t + a24s2

3 2
+ a 3 1 st + a 3 2 st + a 33 st + a34s

+ a 4 1 t 3 + a 4 2t 2 + a 43 t + a44

Written in the algebraic form:

x(s,t) = SCxTt

where S = [s3,s 2,s,11, T = [t3 ,t 2 ,t,l] and TV is the transpose of the matrix T.
The complete alge'braic manipulation of the equation to arrive at the
equation below is similar to the curve process as described in the previous
section. Its details are covered in [Ref. 8:pp. 524-536]. The equation for a Bezier
surface patch is:

X(Gt) = SMbQMb t T t

where Mb is the same matrix as in the curve equation, Mbt is its transpose and
Q. is the x component of sixteen control points of a surface patch. Bezier
surfaces are intuitive in their appeal and serve the fractal rectangular mountain
well. To apply the technique to the motmntain of Figure 6.3 requires the
application of a routine that takes the non-planer four sided shape of a computed
iaitiato*- and develops a connected sixteen point figure as illustrated in Figure 6.5.
The inclusion of a l1ezier subroutine at the recursive termination event after this
figure is developed matches the sixteen point figure with a smooth cur've. To
achieve ,Age continuity requires that adjacent tides have the saute four points in
proper juxtal"oition in the sixteen point matrix. This is also denmonstrated in
Figure 6.5. I1czier surfaces guarantee such continiity.

' P

U
 Bezier Surface Patches

Shared ecdge points

I B . . .. . .. ,

SB , B9 2•
"A.
R,ltIfor
= B. Bii B9 B1--0E

P 3 B, Bl: P,
P.P , H- P 83

Buu B
H. 4 U, B8

QforR,= Btu 3. B1
. P3 B11 nil P,

Figure 6.5. Matching Bezier Surface Patches to the Fractal

Rect~angular Mountain Structure.

00

"4'
 VII. CONCLUSIONS

A. DIRECTIONS FOR FURTHER STUDY

Fractal geometry as an area of research is very new. Because of this, there is
a great need for refinement and exploration. What is known needs to be refined
into a set of workable techniques with reasonable, simple terminology as its root.
The areas that are unknown need to be explored intrepidly. With this goal in
mind, the following paragraphs quickly review some areas of prospective research.
The reader is invited to explore their potential.
1. Development of New Fractal Functional Methods
The current tools of fractal functions are tentative and limited in their
ability to yield insight. New applications of the recursive initiator- generator
paradigm are waiting to be discovered. This area of research is especially good for
the graphics programmer since the graphics medium is currently the best method
for fractal experimentation. As these new functions are developed, they can be
shared, yielding a glossary of modeling functions that can be molded into a
cohesive theory 40 . Related to this is the need to develop a functional language
(within the language of mathematics) of fractal geometry to aid in the
communication of ideas and in the eventual coalescence of the theory.
2. Fractal Lighting Model
The current state of .he art in computer graphics lighting models lacks a
complete model for the pixel sat paradigm that was introduced in Chapter 5.
There are a great many practical applicatiomias which demonstrate successful
lighting techniques but no published model exists. This indicates a piecemeal
undisciplined adaptationi of the Euclidean b-ased lighting models. Ray tracing
techniques look promising, "- does an adaptation of the Torvance-Siprrow
lighting model that was discussed in chapter 5. A good pixel set lighting model
would open the avenue of complex rtrain utdefing to a much wider audience.
••"~~ Xturv's (firat|a map?
\a
SAs evidenced b tbv fatal pittutws that have been pubiwshed.

100
 3. Fractal Music
In [Ref. 71 Voss demonstrates the application of fractal recursive
1
techniques to I- noise and has produced interesting if not pleasing tonal results.

It is safe to surmise that sound is a roughly textured physical phenomenon and

that it may be possible to create or decipher sound using a fractal model. Such a
discovery would aid science in the area of (rapid) speech recognition.
4. Fractal Conmputer Graphics Architectures
It is clear from our discussion that new computer architectures need to
be developed to support the pixel set paradigm and the computational aspects of
fractal functions. Such special architectures require parallel processing capabilities
coupled with vast memory resources. A real-time fractal terrain image generator
is one such architectural possibility.
5. A Better Fractal Definition
* Fractal geometry is currently attaining a wide audience. Because of that,
it is time that trained mathematicians tackle the problems associated with the

imprecise and unworkable current definition of fractal sets 42 . That definition uses
competing definitions of dimension, each of which is somewhat difficult. A new

i definition could be based on a fractal set's functional or statistical qualities. Such

a definition scheme must provide tools to further its workability.

,••

Sadl.y. tbert ha bftn lnitle attention from the mathemnaticl• Community, although tbat is
-bansiamg It is with grtt ttmidity that one &ec" fractal gCOmetr, without such scrutiay.

101
 B. CONCLUSIONS
FractaJ geometry is an old idea that has found a new application with the
advent of computer imaging techniques. Its acceptance, has spawned a great deal
of research and has provided a new tool to observe nature through a different
perspective. We must be careful to insure that our findings are in fact valid. We
also must begin the coalescence of the many techniques that have been developed
in order to control the growth of this concept and to attain true scientific
acceptance. Without this acceptance the theory will be criticized (validly) as an
imprecise and unproven idea 43 . This would be an Unfortunate occurrence because
of the potential that fractal geometry possesses.
It is the hope of the author that this work has illuminated the subject of
fractal geometry and that it will aid others in their research. The purpose and
essence of fractal geometry is based on simple concepts. The reader must not be
overawed by the current literature and should retain his perspective with a mild
dose of skepticism. He must not be blinded by skepticism though as the potential
of fractal geometry has not yet been realized. In the final analysis, we expect
that even the skeptical reader will discover the mathematical beauty and
applicative power that fractal geometry possesses.

SThtt of ui the cmt stat of afTa with frtal ga y

: • 102
 2
APPENDIX A: FRACTAL COMPUTATION IN R

The first routine is the main routine which initializes the data for the Koch curve
generator and initiates the recursive process on each side of the initiator triangle.
The second routine is the recursive subroutine which performs the generator
replacement until the recursive termination event is reached. The termination
event is defined by the precision of the desired output medium.

KOCH.C
1*
This is the main program which controls the initialization of
the koch generator parameters and initiates recursive operations
on each side of the initiator triangle.
'/

P Global generator and initiator data */

int Generator._points;
/ The number of points in the GENERATOR I
double Genangle~luj;
/* The angle formed betweev init-Pointl
and genyoint 1
double GenratioflOj;
/ The between initiointl to genpoint and
gen [oint to initypoint2
Q/
double Tan theda(lOI;
/ The tangent of the wngle formed between
init -tintl and genapoint "/
double Curjyoint[201[21;
/" Vertices of initiator Structume ./
int Object points nmb;
/* The number of vertices of the initiating structure "/

* include <iuath.h> /" Standard UNIX inluce file for math library "/

#derine x 0
#define y I

105
/* BEGIN MAIN PROGRAM /

mnain()
{
/* Local variables /
int I;

/* Initialize global variables /

/* Initial points of the INITIATORS for demo */

Curointloljx] = 4.0;
Curypoint[0[y] = 3.0 + sqrt(3.0);

Curjyoint[l][x] = 5.0;
Cur.joint[lilyj - 3.0;

Cur7_oint[2(xl = 3.0;
Cur_.point[2][yI = 3.0;

/' Remember to close the side of the triangle '1

Cur_.pointI3I[xI = 4.0;
"Cur-point(311yj = 3.0 + sqrt(3.0);

Objectjpnts nmb = 3;
Generatorjoiit.s - 3;

/ Angle (in radians

./ formed between initpoiutl and gen izt
for demo)

Gensangletil - U.0;
Gen angle[21 - 0.4712M88;
Gen s 0gle31 - 0.0;
/* Ratio of dli•tance between initjyoint! and gen poiat(i) and
distauce between geanpoint(i) and iaitypoint2 "/

Gen ratloill 0.5;

G~en ratw'3 2.0;

10.
 /* Tangent of angle between initjpointl and ge__point(i) si
for (1=1; I <= Generatorpoints; I++)
{
- )Tan thedallj = tan(Gen ang•eill);

/* BEGIN RECURSIVE BUILD OF ALL INITIATORS INTO KOCH CURVES 1

/* The Koch curve is defined in the infinite bat our recursion

will terminate after the distance between points becomes less
then the length of the precision. */

for (I-0; I < Obectt_pnts nmb; I++)

4 {
generate( Cur pointilll(x,Cur.point~l[yl,
Curjpoin4tl+ 1I x1,CurJ.ointjI+ illyl);
I
/f END MAIN 'I
1

q++.

4++.*.-

.4-'
 GENERATE.C
/.
This subroutine computes the generator from a given set
of points in R 2 that define a line segment which is the
initiator. The routine is recursive and terainates at a predefined
precision that is input to the subroutine.

j* External global generator data; defined in main subroutine /

extern int GeneratorUoints;
/ The number of points in the GENERATOR */
extern double Gen angle10]j;
/* The angle formed between init pointl
and gen. point 5/

exterm double Genratio[10l;

/* The between inityointl to genpoint and
gen yint to init.oint2 "/
Sextern double Tanthedal I0l;
"*" /* The tangent of the angle formed between
initj ointi and gen&point

*include <mwathh> / Standard math include file for UNIX lib /

/0 BEGIN' RECVRSIVE PROCESS 'I

/ • Parai•,et~'r vaviabls */
doube XL.YIX2.Y2,pcision;
/* Local var..... "I
long N,
double P*frrfavJ312-
double GxiwI•ItOf120J.DIST;
dotil Soe nh , lr k n
diabte X jp.Yj~rpjkrp~b enTP;
Sdouble ton thowmad~one.ia-o•wiuone;

tva
en tt u . 100U),O.0; otoe
-'and j.0; zero 0.0,
0. minus one - -1.0;

10

U,- It
 /* The Koch curve is defined in the infinite but our recursion
will terminate after the distance between points becomes less
then the length of a pixel. */

/ Determine distance between point I and point 2 /

TEMP = (X2 - XI)*(X2 - X1) + (Y2 - Yl)*(Y2 - Y,')"

DIST = sqrt( TEMP );

/* IF DIST less than the precision then terminate this

recursion and begin backtracking /

if (DIST < precision)

{
/* Put your Point plotting routine here
printf("polyline 2");
printf("%f %f O.O0OOOO"X ,Y1);
printf("%f %f0.000000",X2,Y2);
return;

/* Put INITIATOR points one and two into the first and last
points of the GENERATOR points array as they are always
part of the generated structure
• €:G•pont~lli~ll = XI;
-3m[enrtr•o G4 + ,]1 2
il " " -"+ 21['21 Y2;

,•*s

*llI

V107
 /* Determine the slope of the line formed by the init_pointl
and init.point2. This is the slope of the INITIATOR

if (X2 != Xl)
{
if (Y2 != YI)
{
Slopeinit = (Y2 - Y1)/(X2 - XI);
}
else
{
Slope init = 0.0;
I
else

/* We can't have infinity in a register

so settle with 10k */
Slope init = ten thousand;

/* For each GENERATOR point (except end points as they are equal
to the INITIATOR end points) find the X,Y values. This is
accomplished by using the data from the global external variables.
The constant data about the ratios and angles between the
INITIATOR and GENERATOR remain the same regardless of the
INITIATORS length or position in EUCLIDIAN space */

for (I=I; I <= Generatorjpoints; I++)

/* Using the ratios of the generator perpendicular intercept

points on the INITIATOR determine the X,Y values of the
point of intersection of the perpendicular from the
GENERATOR point to the INITIATOR line.

X-perp = (XI + Gen ratio[I] * X2)/(l.O + Cen ratio[l]);

Y_perp = (Y1 + Gen ratio[I] * Y2)/(I.O + Genratio[IJ);

l0

108
 /* If the ar-ge of the INITIATOR point 1 and the GENERATOR
Zpoint in question is zero the:L the GENERATOR point is
coincident with the INITIATOR line and no further
calculations are necessary *1
if ( Gen angle[I] == .ero )
G pit[I-4l][1'_oi
, I _perp;
G_point[I-+ 1[212 Yperp;

else

1* There are three STATES possible at this time. STATE 1

where the slope of the initiator line is parallel
to the X or Y axis (which causes havoc with the line
equations). STATE 2 where the slope of the line formed
by the initiator point 1 and the unknown ge ierator point
is parallel to the X or Y axis. Or STATE 3 where no lines
are parallel to any axis. */

/* Determine the slope of the line through the INITIATOR

point I and the unknown GENERATOR point using the
tangent of the Gen angle in Init,h *1
SlopeZen = (Tan theda[I] + Slope init)/
(one - Tan theda[I] * Slope init);

if ((Slope gen != zerc) &&

(Slopejen < ten thousand))

/* Condition one of STATE 3 */

/* Determine Y-intercept for the generator line /

ben YI - (Slope eu * Xl);

if ((Slope init • reo)tl

, Or (Slope.,it ten thousand))

/ STATE I

4:
 if (Slopeinit == ten thousand)
{
/* STATE 1 condition 1; INITIATOR is parallel
to the Y axis */

G.pointfI+ 1][2] = Y..erp;

G.point[I+I1[11 = (G..point[I+1][2] - bý_en)/
Slope..en;
}
else
{
/* STATE 1 condition 2; INITIATOR is parallei
to the X axis *1

G_point[I+1J[1] = X.perp;
G.point[I+1]•J[ - Slope.gen
G.point[I+l1]l] + bSen;
}
}/* END STATE 1 /
else

/* STATE 3 #/

/* Determine slope of perpendicular line through the

INITIATOR perpendicular intercept. */

Slopeerp = (minus, one)/Slope--nit;

/* Determine Y-intercoept for perpendicular line */

b perp YJperp - (Slopeerp * X.perp);

/* Determine the XY values of the unknown GENERATOR

point.

G;oimntIlt1!l= (byperp - bgen )/

(Slopejgen - Slopejyerp);
11121 - Slope en
Gpoit,-+
G_pointII+1j[11 + bjen;

} / END STATE 3 cond. I if "/

110
 else

/* STATE 2 *
Slope~perp = (minus one)/Slope init;
b~perp YYyerp - (slopeer Xpp)

if (Slope~gen == one)

Gypoint[l+i]f1] = xi;
G~point[I+11[2) = Slopeyerp *G~point[I+iJ[1J

+ byperp;

else

Cpointjl+11j2J = YI;
GQpointll+l~t1] = (G~pointjl+iJ[2] - byperp)/
Slope~perp;

}/* END IF *
}/* END FOR ~
P* Start recursion on each line formed by the generator fromi
right to left

for (P=1;J <= Generatorjoints + ;J++)

generate (G joint JJJ [lJG..point(JJ (21,
G~point(J+ (IJ ]Gjoint [J + 1] (21,precision);

P'END gene-ate '

 APPENDIX B: RANDOM NUMBER GENERATORS

The routine below is a C UNIX UCB implementation of the

uniform distribution 10,11 -+ standard normal [-ooc transformation. It
generates a 500 entry table of random numbers that observes the period of the
standard normal distribution. Following this routine are statistics that verify the
transformation.

RAND OM.TABLE.GENERATOR.C
/,
This subroutine will build a table in memory that contains 500 random
numbers that observe the period of a standard normal variable
,/

#include <math.h> /* Standard UNIX include file for math library */

/* External global variables /

extern double RAND[500];

/' BEGIN MAIN PROGRAM d/

rand tablegen()

/ Local Variables ,I

int lA
double UNFI, UNF2;
double rangepi;
int factor;

pi 3.1415926535;

112

Sq
 /* Determine the range for the random numbers of UNIX UCB */

range = 2;
for (J=l; J<=30; J++)
{
range = range * 2;

range = range- 1;
/* Set the random number generator seed */

srandom(475836);

/* Create a Table for 500 entries */

for (I=O; k 500; I = I + 2)

/* Get a uniform random number through the Unix C subroutine 1

UNFI = randomn0;
UNF2 = randomo;

/* Normalize the uniform random number to the interval [0,11

UNFi = UNFi/ range;

UNF2 = UNF2 / range;

/" Mold the uniform random variable into the approximate normal
•
-. 4 distribution */

factor = 1.0;
if (log(UNFI) < 0.0) factor = -1.0;
RAND(I] = sqrt(factor * (2.0 * log(UNFI)))
cos ((2'pi'UNF2));
SRAND[11 = RANDII) ' factor;
factor = 1.0;
if (Iog(UNF2) < 0.0) factor = -1.0;
RAND(I+ij - sqrt(factor ' (2.0 log(UNFI)))
,-0 sin ((2"pi*UNF2));
RANDjI+1j RAND(1+i] factor;

return;

113
 VERIFYING STATISTICS
The UNIX UCB operating system's uniform distribution random number
generating function spans the interval defined by its integer range. For a VAX
11/780 implementation this is equivalent to 23 - 1 or [0,21474836471.
The random number seed was assigned the value of 475836. The UNIX UCB
random number generator with a fixed seed yields a fixed sequence of numbers
returned from the function, uniformly distributed over the range. This yields a
valuable function if the table needs to be reproduced with the same sequence
after transformation.
The table below shows the results of the uniform distribution sequence after
it was sapezeud into the interval 10,1]. These results show that the uniform
distribution has an acceptable distribution over its range. The transformation
Sinto 10,1] preserves the distribution from the original range ([0,231 - 1]).

Analysis of the normalized uniform random numbers

S0.0 0.1 = 52
0.1 - 0.2 = 47
0.2 -. 0.3 = 44
0.3 - 0.4 = 49
0.4 0.5 = 49
0.5-. 0.6 = 47
0.6 0.7 = 51
0.7 0.8 = 57
0.8 -. 0.9 = 48
0.9 -. 1.0 56
The table below shows the distribution after the
t•uniform distribution O,11 standard normal j-oo,aco transformation given the
numbers asmdescribed in the above table. This is the data which was used to
build Figure 5.9. The transformation is acceptable for the purpose intended, that
J..: is., to simulate nature's perceived disorder in a fractal function.

114

aI
Analysis of the normal (Gaussian) random numbers

X <= -2.75= 5
-2.75 < X <= -2.25 = 8
-2.25 < X <= -1.75 16
-1.75 < X <= -1.25 29
-1.25 < X <= -0.75 = 56
-0.75 < X <= -0.25 = 88
-0.25 < X <= 0.25 = 115
0.25 < X <= 0.75 = 87
0.75< X <= 1.25=59
,-< X <= 1.75 = 21
1.75 < X <= 2.25 = 12
2.25< X<= 2.75=4
2.75< X =0

115
 APPENDIX C: THE TRIANGULAR MOUNTAIN

The first routine is the main routine which initializes the generator data for
the initiating triangle and initiates the recursive process. The second routine is
the recursive subroutine which performs the generator replacement until the
recursive termination event is reached, which is defined by the precision
parameter.

MOUNTAIN.C
/,
This is the main program that controls the initialization of the triangle
initiating structure and initiates the recursion on that triangle. The
recursion will proceed until the recursive termination event (defined
by the precision global parameter)
'/

#include <math.h> /* Standard UNIX include file for math library */

/* Global Defines */

idefine x 0
#define y I
#define z 2

/' Global Tables '/

double RANDj500I;
double Preciskin;
double Scale;

N
 /* BEGIN MAIN PROGRAM */

inain()

f
/* Local Variables */

int I,J,K;
double PI[31,P2[3],P313];
int Seed;
/* Create the initiating triangle structure /

Plx) = 4.5;
Pl[y] = 3.25;
Plz) = 0.0;

P2[x) = 7.0;
P2(y- = 3.25;
P2[z] - 0.0;

P3(x] = 5.75;
P3fy) = 3.25 + sqrt(((2.5 2.5) - (1.25 1.25)));
P3(z] = 0.0;

/* Build the random number table (appendix B) 0/

rand table gen();

/* Frattalize until desired precision /

Seed = 0; /* Entry secd to the randont number table "/

Precision - 0.3; /' Recursive termination distance 'I
Scale = 0.2; /* Scaling factor for vertical Y displacement
in the uiountaihugenerate subroutine /

ui ountain_rienerate(P lP2,P3,Seed);

I' END MAIN

-•.; )

* 117

44•
 GENERATE.MOUNTAIN.C
/,
This is the subroutine that computes the four generated triangles fromn
an initiating triangle. The routine is recursivc and terminates at a
predefined precision defined in the global parameter Precision
/P
#include <math.h> /* Standard math include file for UNIX lib /

#define x 0
#define y I
#define z?

/* Global Structures /

extern double RAND1500];

extern double Precision;
extern double Scale;

/* BEGIN RECURSIVE PROCESS */

mountain_..enerate(PIP2,P3,Seed)

P Parameter variables */
double P1131,P2131,P314];
int Seed;

/* Local variables /

int IAJ;
double Pgen 1[3j,Pgen2j3jPgen3j3j;
".4•:• double Pwid1131, Pmid2(3J,Pxuid3(3I;
double TEMP,DIST,TWO;

TWO - 2.0;

118
 /4 Determine distance between point I and point 2 */

TEMP -(P2[xI - PllxI)*(P2IxJ - Plixi) +

(P2[yJ - Plyl)*(P2'y] - Pilyl) +
(P2[z] - P1IzI)*(P2'ji .. P1(zj);
DIST = sqrt( TEMP);

/* IF DIST less than one tnen terminate this recursion and

begin backtracking *1

if (DIST < Precision)

{
/' Put your polygon plotting routine here */
printf("polygon3"';
printf("%f %f %f",PIl[xj,Pl yj,P1lzI);
printf("•f %f %f" ,P2[xl,P21y],P2[zI);
printf("%f %f %f',P3[x,,P3[y],P3[zl);
return;
}
/ Manage the Seed number for a 500 entry table '/

if (Seed > 496) Seed = 0,

/ Find the midpoints of each triangle leg '/

for (1=0; 1<=2; I++) /. 0 thru 2 => x~y~z /

1.) {
Pwidl(l (PI11 + P2[I1) / TWO;

for (I=0; i<=2; 1++) /* 0 thru 2 => xy,z '/

{
(P2111 + P3(I) / TWO;
)Pmid2jlj -

for (iU-0; I<-2; I++) /. 0 thru 2 x> 'z/

Pmid3 ll- (P3111 + Pill]) / TWO;

vv

iAl
 /* Adjust the Y coordinate => normal from Z-X plane /

Pmid~yJ = (Scale * RANDISeedl) + Pmidlly);

Pmid2[y] = (Scale * RAND[Seed+1]) + Pmid2tyl;
Fmid3jy] = (Scale * RAND(.Seed+2]) + Pmid3JyJ;
Seed =Seed+ 1;
/* Recurse on the triangles according to the reverse order rule
for the interior triangle to preserve seed order */
mountain.generate(Pmidl,P2, Pmid2,Seed);
mountainjenerate(Pmid3,P2id2,P3, Seed);
mountain..enerate(PmI, Pmidl,Pmid3,Seed);
mountain..enerate(Pmid2,Pmid3,Pmidl,Seed);
/* END geWerate .1
)

~ I

V.

0120
 APPENDIX D: THE RECTANGULAR MOUNTAIN

The first routine is the main routine which initializes the generator data for
the initiating rectangular shape and initiates the recursive process. The second
routine is the recursive subroutine which performs the generator replacement
*.until the recursive termination event is reached, which is defined by the precision
parameter.

RECTANGULAR.MOUNTAIN.C

i.This is the main program that controls the initialization of the rectangular
initiating structure and initiates the recursion on that rectangle. The
recursion will proceed until the recursive termination event (defined-
by the precision global parameter)
4../
finclude <math~h> /' Standard UNIX include file for math library /
*'Global Defines /
idefine x 0
#define Z

/ Global Tables /
double RANDI500);
double P-ecision;
double c*ale;

121
/* BEGIN MAIN PROGRAM */
mnaiL()

/* Local Variables */
int I,J,K;
double P1l3],P2[3],P3[3],F413];
int Seed;
/* Create the four sided polygon initiating structure */
P1[x] = 2.0;
Pi[y] = 5.0;
P1[z] = 1.0;

P2[x] = 6.0;
P2[y] = 5.0;
P2[z] = 1.0;

P3[xJ = 7.5;
P3[y] = 6.5;
P3[z] = 3.0;

P4[x] = 3.5;
P4[y] = 6.5;
P4[z] = 3.0;

/* Build the random number table (appendix B) */

randtable_geno;

/* Fractalize until desired precision *1

Seed = 100;
Precisioii = 0.5;
Scale = 0.07;

mountaingenerate(PI,P2,P3,P4,Seed);

/* END MAIN */
}

122

A"~ 1
 RECTANGULAR.GENERATE.MOUNTAIN.C

This is the subroutine that computes the four generated rettangles from
an initiating rectangle. The routine is recursive Rnd terminates at a
predefined precision defined in the global parameter Precision

#include <math.h> /* Standard math include file for UNIX lib *,

#define x 0
#define y 1
. #define z 2

/* Global Structures *1

extern double RAND[500];

extern double Precision;
extern double Scale;

*
-- 1
/* BEGIN RECURSIVE PROCESS */

mountaingenerate(PI,P2,P3,P4,Seed)

/* Parameter variables *l
double P1[3],P2[3],P313j,P413];
jut Sced;

/* Local variables '/

int I,J;
double Pmid 1[g],Pmid2[3],Pmid3[3],Pmid4[3],Center[3];
double TEMP,DISTTWO,FOUR;

TWO 2.0; FOUR = 4.0;

/ Determine distance between point 1 and point 2

TEMP = (P2[xJ - PI[x)"(P2[xj - PIjx]) +
(P22yJ - PI[y])*(P2[y] - Pity]) +
(P2{zI - PI(z])'(P2[zj - PI{zj);
DIST sqrt( TEMP),

123

-7ý%t. ` ý,O .'1

 /* If DIST less than one then terminate this recursion and
begin backtracking */
if (DIST < Precision)

/* Put your Polygon output routine here */

printf(I polygon4');
printf("%f %f %f`,,P1 [x],P1 [y],P1 [z]);
printf( df %1'
%f•',P2(x],P2[y],P2[zJ);
printf(" %f %f %f',P3[x],P3[y],P3[z]);
printf("%f %f %vf",P4[x],P4[y],P4[z]);
return;
}
/* Manage the Seed number for a 500 entry table */
if (Seed > 496) Seed = 0;
K

/* Find the midpoints of each rectangle leg */

for (1=0; 1<=2; I++) /* 0 thru 2 => x,y,z */

PmnidilI] = (PI[Ij + P21I]) / TWO;

for (1=0; I<=2; I++) /* 0 thru 2 => x~yz /

Pmid2[Ij = (P2[1] + P3[I]) / TWO;
}
for (I=0; I<-2; 1++) /*.0 thru 2 => x•y,z */

Prnid3[I1 = (P31I1 + P411]) / TWO;

for (4=0; 1<=2; 1++) /" 0 thru 2 => x•yz */

{
Prmid4[1J= (PIll] + P4111) / TWO;
)

•• 124
 /* The four sided polygon is non-planar to average the xyz-displacement
for a best fit approach */
for (1=0; I<=2; I++) /* 0 thru 2 => x,y,z */
{
Center[I] = (P3[I] + PI[I] + P4[I] + P2[1]) / FOUR;
}
/* Adjust the Y coordinate => normal from Z-X plane */
Pmidl[y] = (Scale * RAND[Seed]) + Pmidl[y];
Pmid2[y] = (Scale * RAND[Seed+1]) + Pmid2[y];
Pmid3[y] = (Scale * RAND[Seed+2J) + Pmid3[y];
Prnid4[y] = (Scale * RAND[Seed+3]) + Pmid4[y];
Center[y] (Scale * RAND[Seed+4J) + Center[y];
Seed = Seed + 4;

/* Recurse on the rectangles according to the reverse order rule

for the interior rectangles to preserve seed order */

mountainjenerate(P1, Pmidl, Center,Pmid4,Seed);

mountainjgenerate(Pmid2,Center,Pmidl, P2, Seed);
mountain _enerate(Pmid2,Center,Pmid3, P3, Seed);
mountainZenerate(P4, Pmid3, Center,Pmid4,Seed);

/* END generate */
1

•"•)•125

4g,-'

'c
 APPENDIX E: GEOMETRIC SUPPORT

Many fractal applications and computer graphics models use the normal to a
plane as a computational reference point. For this reason, this appendix is
devoted to two tools for determining the plane equation of a polygon and the
equation of the normal to the computed plane.
Determinant Approach to the Planar Equation

One of the most common forms of a planar equation is the general form. This
form uniquely describes a plane through four coefficients A,B,C and D:

Az +By +Cz = D

With three points on a plane, you can determine the planar equation by
computing the coefficients. This approach utilizes the determinant form of the
planar equation. Given the points P1 = (zi,yU,zi), P 2 = (z 2 ,Y2 ,z 2 ) and
P3 = (z 3 ,y 3 ,z 3) such that P 1 o P 2 # P 3, these points determine a unique plane
in space through the determinant equation:

X2 -X 1 Y2 -Y Z2 =1 0

x2 - XI Y2
Y - YI Z2 Z1 1

To simplify the equation, we replace the constant differences by the expressions:

Cix X2 - XI

C2= x 3 - xI
Cly x2 - Y1
CZ,1 X - Y-

cis x2 - Zt
C22 X3 - ZI

120
 Evaluating the determinant using the diagonal approach yields:

[(x-
x)C1C2,- (x - xi)C15 C2y1 +
[(y - yl)C1.C2. - (y - yl)Cl.C2,] +
Rz - z,)C11C2y - (z - zd)ClC21 ] = 0

Solving the equations for x,y and z in terms of the constant expressions:

A= ClYC22 - CIC2y
B = C1IC2x- ClXC23
C = ClXC2, - C1YC2X
D= -[Ax +By, +Cz1 J

The Normal to the Plane

Once the parameters A,B and C have been determined, the solution of the
linear equation for any normal to the plane is straightforward. Using the plane
parameters in the parametric equation for the normal line to the plane and using
any known point on the plane (xkwn,ykwzkwn) (the midpoint of the fractal
triangle for example) determines a normal line as:

I X = xkw + cA
V.• Y = Yk + eB
•X•Z Z= k.,n + CC

where c is a parameter such that c is an element of R. By varying the parameter

•:4 cwe can solve for unique points on the normal line to the plane.

127
 LIST OF REFERENCES

1. Benoit B. Mandelbrot, The Fractal Geometry of Nature, W. H.

Freeman and Company, 1983.

2. Donald F. Stanet and David F. McAllister Discrete Mather-Aatics

in Computer Science, Prentice-Hall, Inc., 1977.

3. Felix Hausdorff, Set Theory, Chelsea Publishing Company, 1957.

4. Alan Norton, Generation and Display of Geometric Fractals in 3-D,

Computer Graphics, vol 18, no. 13, July 1984.

5. J. Perrin, La Discontinuite's de la Matiere, Revue du Mois, vol 1, 1906;

quoted by Mandelbrot in Ref. l:pp 7-9.

6. Witold Hurewicz and Henry Wallman, Dimension Theory, Princeton

University Press, 1941.

7. Special Interest Group on Computer Graphics of the Association for

Computing Machinery (SIGGRAPH), Fractals: Basic Concepts,
Computation and Rendering, Course on Fractals July 23, 1985.

8. James D. Foley and Andries Van Dam, Fundamentals of Interactive

Computer Graphics, Addison-Wesley, 1982.

9. Jagdish K. Patel and Campbell B. Read, Handbook of the Normal

Distribution, Marcel Dekker, Inc, 1982.

128

-. L.t S
 DistributionList for Papers Written by Michael J. Zyda

Defense Technical Information Center,

Cameron Station,
Alexandria, VA 22314 2 copies

Library, Code 0142

Naval Postgraduate School,
Monterey. CA 93943 2 copies
Center for Naval Analyses,
2000 N. Beauregard Street,
Alexandria, VA 22311
Director of Research Administration,
Code 012.
Naval Postgraduate School.
Monterey. CA 93943

Dr. Henry Fuchs.

208 New West Hall (035A).
University of North Carolina.
Chapel Hill. NC 27514

Dr. Kent R. Wilson.

University of California. San Diego
B-014.
Dept. of Chemistry,
La Jolla. CA 92093
Dr. Guy L. Tribble. III
"900 \Vaverly St.
Palo Alto. California 94301

- Bill Atkinson.
Apple Computer.
20525 Mariani Ave.
Culwrtino. CA 95014

Dr. Victor Lesser.

Uni'ers.ity of Massahusetts. Amherst
Dept. of Computer and Information Science.
Ainherst. MA 01003

Dr. Gunther Schrack.

Dept. of Electrical Engineering.
University of British Columbia.
Vancouver. B.C.. Canada VOT 1\\5

Dr. R. Daniel Bergeron.

Dept. of Comnputer Science.
Uriivcr-ity of New Hanpshire.
Durham. NH 03824

OI
 -2-

Dr. Ed Wegman,
Division Head,
Mathematical Sciences Division,
Office of Naval Research,
800 N. Quincy Street.
Arlington, VA 22217-5000

Dr. Gregory B. Smith,

ATT Information Systems,
190 River Road.
Summit. NJ 07901

Dr. Lynn Conway.

University of Michigan,
263 Chrysler Center.
Ann Arbor, MI 48109
Dr. John Lowrance.
SRI International.
333 Ravenswood Ave.
Menlo Park. CA 94025

Dr. David Mizell.

Office of Naval Research,
1030 E. Green St.
Pasadena, CA 91106

Dr. Richard Lau,

Office of Naval Research.
Code 411.
800 N. Quincy St.
Arlington. VA 22217-5000

Dr. Y.S. Wu.

Naval Research Laboratory.
Code 7007.
Washington. D.C. 20375

Dr. Joel Trimble,

Office of Naval Research.
Code 251.
Arlington. VA 22217-5000
Robert A. Ellis,,
CalhiCn C.on1paly.
R &-I) Engineering.
525 Sycatnore Dr.. MIS C510
Milpitas. CA 95035.7489
Dr..-latt- Iti. Clarki.
Silicon (Graphic,. hlir.
2011 Stierlin Road.
\Mouatain View. CA N4043
 -.3-

Edward R. McCracken,
Silicon Graphics, Inc.
2011 Stierlin Road.
Mountain View, CA 94043

Shinji Tomita,
Dept. of Information Science,
Kyoto University,
Sakyo-ku. Kyoto, 606, Japan

Hiroshi Hagiwara,
Dept. of Information Science,
Kyoto University,
Sakyo-ku, Kyoto, 606, Japan

Dr. Alain Fournier.

Dept. of Computer Science.
University of Toronto.
Toronto. Ontario. Canada
M5S 1A4

Dr. Andries Van Dam.

Dept. of Computer Science.
Brown University.
Providence, RI 02912

Dr. Brian A. Barsky,

Berkeley Computer Graphics Laboratory.
Computer Sciences Division.
Dept. of Electrical Engineering and Computer Sciences,
University of California.
Berkeley. CA 94720

Dr. Ivan E. Sutherland.

Carnegie Mellon University,
Pittsburg. PA 15213

Dr. Turnrr Whitted.

New West Ha!! (035A).
University of North Carolina,
Chapel Hill. NC 27514

Dr. llotwrt )). (;rafton.

Office of Naval Research.
Codte 433.
Arlington. Virginia 22217-5000
Professor Eihachiro Nakamae.
Electric Machinery Laboratory.
Hjiroshima University.
Higasliihiroshinaa 724. Japan

:-•, ~k .....
 -4.

Carl Machover,
Machover Associates,
199 Main Street,
White Plains. New York 10601

Dr. Buddy Dean,

Naval Postgraduate School,
Code 52, Dept. of Computer Science,
Monterey, California 93943
Earl Billingsley.
43 Fort Hill Terrace,
Northhampton, MA 01060

Dr. Jan Cuny,

University of Massachusetts, Amherst
Dept. of Computer and Information Science,
Amherst. MA 01003

Robert Luni.
Silicon Graphics. Inc.
2011 Stierlin Road.
Mountain View. CA 94043

Jeff Hausch.
Silicon Graphics. Inc.
2011 Stierlin Road.
Mountain View, CA 94043

Robert A. Walker.
7657 Northern Oaks Court,
Springfield. VA 22153
Dr. Barry L. Kalman.
Washington V'niversity.
Departnient of Computer Science.
iPox 1045.
St. Louis, Missouri 63130

Dr. Win. Randolph Franklin.

Etletrieal. Computer. and Systeian Enginvering Deprtuiaent.
Rensselaehr Polytt-lhnoc Institute,
Troy. New York 12180-3590

Dr. ;ershon Kedtkn.

Microelctronics Center ef North Carolina.
PO Box 12689,
3021 Cornwallis Road.
0Research Triangle IPark.
North Carolina 27709

A
 Dr. Brankf. J. Gerovac.
Digital Equipment Corporation,
150 Locke Drive LMO4/H4, Box 1015
Marlboro, Massachusetts 01752-9115
Rober! A. Schumacker,
Evans and Sutherland,
PO Box 8700,
580 Arapeen Drive.
Salt Lake City, Utah 84108

R. A. Dammkoeh!er,
Washington Unitet•.,ty,
Department of Computer Science.
Box 1045.
St. Louis, Missouri 03130
Dr. Lynn Ten Eyck.
Interface Software.
79521 Highway 99N.
Cottage Grove. Oregon 97424

Toshiaki Yoshinaga.
Hitachi Works. Hitachi Ltd.
1-1. Saiwaicho 3 Choine.
Hitachi-shi. lbaraki-ken.
317 Japan

Takatoshi Kodaira.
Omika Works. Hitachi Ltd.
2-1. Omika-cho 5-chotwe.
Hitachi-shi. Ibaraki-ken.
319-12 Japan

Atu,%hi Suzuki.
Hit•chi Engineering. Co. Ltd.
2-1. saiwai-cho 3-Chionle.
Hitachi-shi. Ibaraki-ken.
317 Japan
Toshiro Nisirnura.
Hitiachi Engineering. Co. Ltd.
2.1.',iwa-rho 3-Chn-

317 Japan
D)r. John :tandliamrnir.
-* " Dept. of Ehlerical Enginetring.
t'viversi.t of Florida.
(Gain','ilte. Florida 32611
zt

Ir
 Dr. Lewis E. Hitchner.
Computer and Information Science Dept.
237 Applied Science Building,
University of California at Santa Cruz,
Santa Cruz. California 95064

Dr. Pat Mantey,

Computer Engineering Department.
University of California at Santa Cruz,
Santa Cruz, California 950M4
Dr. Walter A. Burkhardt.
University of California. San Diego
Dept. of Computer Science.
La Jolla. California 92093
P. K. Rustagi.
Silicon Graphics. Inc.
2011 Stierlin Road.
Mountain View. CA 94043

Peter Broadwell.
Silicon Graphics, Inc.
2011 Stierlin Road.
Mcuntain View, CA 94043

"Norm Miller.
"Silicon ;raphirs. Inc.
2011 Stierlin Road,
Mountain View. CA 94043

Dr. Tosiyasu L. Kunii.

Department of hnformtionsciencec.
Facuhty of Science.
The tVniver:iity of Tokyo.
7-3-1 |iongo. B:kakyo-ku. Tokyo 113.
Japan
Dr. Kainhiro Furlhi.
Institute for New Generation Comsputer Technology.
SMit4-Kok-tta) Building '21FL.
1-4-2S Mira. Minatu-ku. Tokyo 10S. .Japan

Tony%LowI'.
S•.,, Silicon rp hir•I .
• + 1901 A\'entu of the Starn.
*+ ,%uit, 1774.
Los Angeles. CA %000-

Kevij Uatnmon5.
NASA AME -DDrden Flight RCfearch Facility.
PO Box 273.
Mail Stop O(I.
Edward%. California 93523
 -7-

Sherman Gee.
Code 221,
Office of Naval Technology.
800 N. Quincy St.
Arlington, VA 22217

Dr. J.A. Adams.

Department of Mechanical Engineering,
ITS Naval Academy.
Annapolis. MD 21402

Dr. David F. Rogers.

Dept. of Aerospace Engineering.
US Naval Academy.
Annapolis, MD 21402
Dr. Robert F. Franklin.
Environ mental Research Institute of Michigan.
PO Box 8618.
Ann Arbor. MI 48107

LT Mark W. Hartorag.
900 Cambridge Dr 17.
Bkenicia, CA 94510
Ca.pt. Mike Gaddis.
DCA /JDSSC/C720.
1860 \Vichle Ave
Rcstou. VA 22090

Lt. Cdr. Patrick G. Hogan. USN

102 Borden Aveiue.
WVilmington. North Carolina 28403

Dr. Edwin Catwuli.

P0 Box 200W.
Sanl Rafarl, CA 9491-

Dr. John Upatty.

Compvmier S-rientce ,patact
4' •-rrmitV of \\vtrlor|.
•rritro.
oti On1t ario.
0anada N2L 3(;1

Dr jar;19 Folw. ai .m trSrc

• George \War-hingon |Vnivor-ItA%
,'Dept, of Elerirical Engineering and Computer Sci .
ws'ahinato D.C. .ZW5u
,.*" •Dr. Dnnald Grr*-itwrg.
*" (in-n.'ll I 'nirrsty.
Itbrograw (if Compulvr Graphirs.
Ithaca. NY 14853

4l
 -8-

Dr. Leo J. Guibas,

Systems Research Center.
Digital Equipment Corporation,
* 130 Lytton Avenue,
Palo Alto, CA 94301

Dr. S. Ganapathy,
Ultrasonic Imaging Laboratory,
Dept. of Electrical and Computer Engineering.
University of Michigan.
Ann Arbor, MI 48109

Dr. Hank Christiansen.

Brigham Young University,
Dept. of Civil Engineering,
368 Clyde Bldg.
Provo. Utah 84602

Dr. Thomas A. DeFanti.

Dept. of Electrical Engineering & Computer Science,
University of Illinois at Chicago.
Box 4348,
Chicago. IL 60680

Dr. Lansing Hatfield,

Lawrence Livermore National Laboratory,
7000 East Avenue,
PO Box 5504. L-156.
Livermore. CA 94550