Fractals in Biology and Medicine PDF

R C A Sil
L G and
E U E
T. F. Nonnenmacher
G. A. Losa
E. R. Weibel
Editors
Springer Base l A G
Editors
T. F. Nonnenmacher G. A. Losa
Mathematische Physik Laboratorio di Patologia Cellulare
UniversitätUlm Istituto Cantonale di Patologia
Albert-Einstein-Allee 11 Via in Selva 24
D-89069 Ulm CH-6604 Locarno
Germany Switzerland
E. R. Weibel
Anatomisches Institut
der Universität Bern
Bühlstrasse 26
CH-3012Bern
Switzerland
A CIP catalogue record for this book is available from the Library of Congress, Washington D.C., USA
Fractals in biology and medicine / edited by T. F. Nonnenmacher, G. A. Losa, E. R. Weibel

Includes bibliographical references and index.
ISBN 978-3-0348-9652-8 ISBN 978-3-0348-8501-0 (eBook)
DOI 10.1007/978-3-0348-8501-0
1. Medicine-Mathematics-Congress. 2. Biology-Mathematics-Congress. 3. Fractals-Congress.
I. Nonnenmacher, T. F. (Theo F.), 1933-. II. Losa, G. A. (Gabriele A.), 1941-.
III. Weibel, Ewald R.
R853.M3F68 1993
574' .01'51474-dc20
Deutsche Bibliothek Cataloging-in-Publication Data
Fractals in biology and medicine / ed. by T. F. Nonnenmacher

... - Basel; Boston; Berlin : Birkhäuser, 1993
ISBN 978-3-0348-9652-8
NE: Nonnenmacher, Theo F. [Hrsg.]
The publisher and the editors can give no guarantee for the information on drug dosage and administration contained
in this publication. The respective user must check its accuracy by consulting other sources of reference in each
individual case.
The use of registered names, trademarks etc. in this publication, even if not identified as such, does not imply that
they are exempt from the relevant protective laws and regulations or free for general use.
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in other ways, and storage in data banks. For any kind of use permission
of the copyright owner must be obtained.
©1994 S p r i n g e r Base l AG
Originally published by Birkhäuser Verlag, PO Box 133, CH-4010 Basel, Switzerland in 1994
Softcover reprint of the hardcover 1st edition 1994
Camera-ready copy prepared by the editors
Revision and layout by mathScreen online, CH-4123 Allschwil
Printed on acid-free paper produced from chlorine-free pulp
ISBN 978-3-0348-9652-8
98765432 1
Contents
Foreword ...... .. ..................... . ..... . ........ . . . .. .. ........ ix
Preface: Summary of the Symposium . ...... . .. . . ..... ...... . . ... . ... 1

The Significance of Fractals for Biology and Medicine.
An Introduction and Summary
Ewald R. Weibel ..... . ....................................... ... .. ... 2
Fractal Geometry and Biomedical Sciences . .... ... .. . ... ... . ........ 7
A Fractal's Lacunarity, and how it can be Tuned and Measured
Benoit B. Mandelbrot .. . .. .. . .. . ..... . . .. ...... . .. . . .. . ............... 8
Spatial and Temporal Fractal Patterns in Cell and Molecular Biology
Theo F. Nonnenmacher .. . .. . .. ... ........... . ........................ 22
Chaos, Noise and Biological Data
Bruce J. West, W. Zhang and H J. Mackey ... ....... . . . ... . . .... ...... . 38
Fractal Landscapes in Physiology & Medicine: Long-Range
Correlations in DNA Sequences and Heart Rate Intervals
C.-K. Peng, S.v. Buldyrev, J.M. Hausdorff, SHavlin, J.E.M.Mietus ,
M. Simons, H.E. Stanley, and AL Goldberger ..... . . . . . . . . . . ..... .55. . . . .
Fractals in Biological Design and Morphogenesis . .... . . . . . ...... ..... 67

Design of Biological Organisms and Fractal Geometry
Ewald R . Weibel . .... .... .. . . ... .. .. ... . .... .. ... . . . ... . . . ....... .. .. 68
Fractal and Non-Fractal Growth of Biological Cell Systems
Petre Tautu . . .. ..... ..... .. . . ........ .. .. ..... .. . . ....... ...... ...... 86
Evolutionary Meaning, Functions and Morphogenesis of
Branching Structures in Biology
Giuseppe Damiani .. .... . . .. . .... .. .. . ... . . .... ......... . .. ..... ... . . 104
Relationship Between the Branching Pattern of Airways and the
Spatial Arrangement of Pulmonary Acini - A Re-Examination from a
Fractal Point of View
Hiroko Kitaoka and Tohru Takahashi .......................... . .. . . ... 116
Multivariate Characterization of Blood Vessel Morphogenesis in the
Avian Chorioallantoic Membrane (CAM): Cell Proliferation,
Length Density and Fractal Dimension
Haymo Kurz , Jorg Wilting and Bodo Christ ..... .. ..................... 132
Phyllotaxis or Self-Similarity in Plant Morphogenesis
Franf;ois Rothen . . .. ... ..... ...... .... .. ........ .. . . ... . . . ...... . ... . 141
vi Contents
Fractals in Molecular and Cell Biology 159

Evolutionary Interplay Between Spontaneous Mutation and
Selection: Aleatoric Contributions of Molecular Reaction Mechanisms
Werner Arber ........................................................ 160
Error Propagation Theory of Chemically Solid Phase Synthesized
Oligonucleotides and DNA Sequences for Biomedical Application
Zeno Foldes-Papp, Armin Herold, Hartmut Seliger and
Albrecht K. Kleinschmidt ............................................. 165
Fractional Relaxation Equations for Protein Dynamics
Walter G. GlOckle and Theo F. Nonnenmacher .. . ........ ............... 174
Measuring Fractal Dimensions of Cell Contours:
Practical Approaches and their Limitations
Gerd Baumann, Andreas Barth and Theo F. Nonnenmacher .............. 182
Fractal Properties of Pericellular Membrane from
Lymphocytes and Leukemic Cells
Gabriele A. Losa ..................................................... 190
Cellular Sociology: Parametrization of Spatial Relationships
Based on Voronoi Diagram and Ulam Trees
Raphael Marcelpoil, Franck Davoine and Michel Robert-Nicoud 201
A Fractal Analysis of Morphological Differentiation of
Spinal Cord Neurons in Cell Culture
Tom G. Smith, Jr. and E.A. Neale ...................................... 210
Fractal Dimensions and Dendritic Branching of Neurons in the
Somatosensory Thalamus
Klaus-D. Knijfki, Matthias Pawlak and Christiane Vahle-Hinz 221
Fractal Structure and Metabolic Functions 231

Organisms as Open Systems
Manfred Sernetz ..................................................... 232
Transfer to and across Irregular Membranes Modelled
by Fractal Geometry
Bernard Sapoval ................ ... ....................... . .......... 241
Scaling and Active Surface of Fractal Membranes
Ricardo Gutfraind and Bernard Sapoval ....................... . ....... 251
Structure Formation in Excitable Media
Martin Luneburg ..................................................... 266
Colony Morphology of the Fungus Aspergillus Oryzae
Shu Matsuura and Sasuke Miyazima ................................... 274
Contents Vll
Estimation of the Correlation Dimension of All-Night

Sleep EEG Data with a Personal Super Computer
Peter Achermann, Rolf Hartmann, Anton Gunzinger,
Walter Guggenbiihl, and Alexander A. Borbely ......................... 283
Fractals in Pathology ......... . ...................................... 291

Changes in Fractal Dimension of Trabecular Bone in Osteoporosis:
A Preliminary Study
C.L. Benhamou, R. Harba, E. Lespessailles G. Jacquet, D. Tourliere
and R. Jennane ...................................................... 292
Use of the Fractal Dimension to Characterize the Structure of .
Cancellous Bone in Radiographs of the Proximal Femur
Curtis B. Caldwell, John Rosson, James Surowiak and Trevor Hearn 300
Distribution of Local-Connected Fractal Dimension and the
Degree of Liver Fattiness from Ultrasound
Carl J.G. Evertsz, C. Zahlten, H.-O. Peitgen,
I. Zuna, and G. van Kaick .......................................... . . 307
Fractal Dimension as a Characterisation Parameter of Premalignant
and Malignant Epithelial Lesions of the Floor of the Mouth
Gabriel Landini and John W. Rippin ............ . .......... . ........... 315
Modelling ....... . ........................................ .. ......... 323

Modelling HlV/ AIDS Dynamics
Philippe Blanchard... ........................... ..................... 324
Morphological Diagnosis Turns from Gestalt to Geometry
Vittorio Pesce Delfino, Teresa Lettini, Michele Troia, and Eligio Vacca 331
Fluorescence Recovery after Photo bleaching Studied by
Total Internal Reflection Microscopy: An Experimental System
for Studies on Living Cells in Culture
Torsten Mattfeldt, Theo F. Nonnenmacher, Armin Lambacher,
Walter G. GlOckle and Otto Haferkamp ................................ 351
Anomalous Diffusion and Angle-Dependency in the Theory of
Fluorescence Recovery after Photobleaching
Walter G. GlOckle, Torsten Mattfeldt, and Theo F. Nonnenmacher 363
List of Speakers .............................................. . . . .... 373
List of Participants .......... . .. . . . . . . . . . . . . . . . . . . . . . . .... . 376

. . . .......
Index.......... . .................................................... 382

Foreword
This volume contains the scientific presentations given at the International Sym-
posium on fractals in biology and medicine held in Ascona, Switzerland, from
February 1--4, 1993. Scientists from the United States, Austria, Canada, Japan,
France, Germany, Great Britain, Italy and Switzerland came together to present
and discuss their research papers as well as to exchange information on their
experimental findings and theoretical interpretations.
When we started to prepare for this conference, we had the feeling that the
time had come to gather prominent scientists from around the world to discuss
this topical subject. To our knowledge, this symposium has been the first com-
prehensive meeting of scientists on fractals in biology and medicine. Since the
publication of Benoit Mandelbrot's book The Fractal Geometry of Nature, the
fractal concept has been rapidly pushed forward essentially by mathematicians
and physicists alike. However, apart from a few exceptions, it needed still more
time to enter the fields of biology and medicine. One of the reasons for this de-
lay might have been the language barrier between the different fields: fractals,
and mainly their graphical representations, are appealing and even accessible to
intuitive understanding, but to translate the theoretical concepts at their basis into
terms that can be interpreted in a biological context requires a level of mathemat-
ical understanding that is not typical for the group of biologists, for example. In
order to bridge this language barrier, we invited an interdisciplinary group of sci-
entists to Monte Verita in Ascona. The members of this group have been working
in different areas of science but most have focussed some or all of their activities
on biomedical research problems.
Almost all 90 participants could be accommodated in the conference building
at Monte Verita and thus it was easy to make personal contacts. The exchange
of information between different groups of researchers took place in small circles
and within an exciting atmosphere, bringing the family of fractal researchers closer
together.
We are particularly grateful to all the renowned institutions who accepted to
confer their scientific patronage and also to the sponsors who made the achievment
of this symposium possible.
Our thanks are also due to Mrs. Antonella Camponovo and Graziana Pelloni,
Dr. Danilo Merlini, member of the scientific and organizing Committee, the col-
laborators Riccardo Graber, Lorenzo Leoni, and Cristoforo DUrig and to Mr. Luca
Albertini, managing director of the Centro Seminariale Monte Verita, who made
the conference run smoothly, and to Dr. Gerd Baumann for his editorial assistance
while preparing this proceedings volume.
In presenting the different contributions in this volume, we did not follow
the chronological sequence of sessions. Rather, we arranged the proceedings as to
grouping similar topics together.
Ascona 1993 The editors
x Foreword
The symposium took place at the Centro Seminariale Monte Verita, Ascona,
Switzerland, from February 1-4, 1993.
Under the Auspices of
- INTERNATIONAL SOCIETY FOR STEREOLOGY

- SWISS ACADEMY OF SCIENCES
- SWISS NATIONAL SCIENCE FOUNDATION
- SWISS SOCIETY FOR CELL AND MOLECULAR BIOLOGY
- COMMITTEE FOR DIAGNOSTIC QUANTITATIVE PATHOLOGY
- WORKING GROUP OF THE EUROPEAN SOCIETY OF QUANTITATIVE
PATHOLOGY
- DIPARTIMENTO DELL'ISTRUZIONE E DELLA CULTURA AND
DIPARTIMENTO DELLE OPERE SOCIALI, REPUBBLICA DEL
CANTONE TICINO
Financial Support was Received from
- DIPARTIMENTO DELL'ISTRUZIONE E DELLA CULTURA DEL

CANTONE TICINO
- SWISS NATIONAL SCIENCE FOUNDATION
- SWISS ACADEMY OF SCIENCES
- FRATELLI MONZEGLIO, CARROZZERIA SA, LOCARNO
- CITIA DI LOCARNO
- BANCA DEL GOTTARDO, LOCARNO
- SOCIETA ELETIRICA SOPRACENERINA, LOCARNO
- RADIO DELLA SVIZZERA ITALIANA, RETE 2, LUGANO
- PHILIPS AG, ZORICH
- LEICA AG, ZURICH
Preface
Summary of the Symposium
The Significance of Fractals for
Biology and Medicine
An Introduction and Summary
Ewald R. Weibel
Department of Anatomy
University of Bern
Biihlstrasse 26
CH-3000 Bern 9, Switzerland
In his seminal book The Fractal Geometry of Nature Benoit Mandelbrot made the
point that the mathematical construct of a geometry which allows for fractional
dimensions will prove most useful in the characterization of natural phenomena,
structures and processes alike. This should be particularly true in biology for a
number of reasons. First, the complex structure of living creatures, from the whole
organism down to the cells, is notoriously difficult to reduce to simple geomet-
ric descriptions, and functional processes have very often non-linear properties.
Organisms develop and grow from small and simple units to gradually achieve
their size and complexity. In living systems, design and performance commonly
combine strict rules with some random variation which gives each individual its
species characteristics and its individual traits. Furthermore, the wonderful variety
observed in the plant and animal kingdoms is the result of stepwise «variation over
a common theme» - they are similar but not alike with basic features preserved
between related species but expressed in different size and proportions.
All this has some resemblance to the construction principles of fractal geom-
etry where small changes in the algorithms yield a great variety of form so that
one can expect considerable new insights from subjecting biological structures and
processes to fractal analysis. And, finally, diseases appear often as modulations of
basic design patterns so that fractal geometry may even have an impact on the
study of disease in medical sciences. However, to introduce the concepts and tools
of fractal geometry into biological and medical research is not easy. It requires
skills that even a well-educated biologist cannot necessarily master.
*
The goal of the Symposium on Fractals in Biology and Medicine was first to
explore the potential that fractal geometry offers for elucidating and explaining
the complex make-up of biological organisms and then to develop the concepts,
questions and methods required in research on fractal biology. In order to approach
this task the conference brought together mathematicians, engineers, natural sci-
entists, biologists and medical scientists, because to discuss fractal geometry in
Introduction and Summary 3
the perspective of living systems must be an interdisciplinary endeavour. One of

the problems in this kind of enterprise is that interdisciplinary intercourse must
be based on a very particular communication culture to which most scientists are
unaccustomed. Most contributors to this symposium have met the majority of the
other participants for the first time, and they were only partially conversant in the
language of the fields represented by the others. One of the primary aims of this
Conference therefore was an attempt to bridge the communication gap between
various disciplines, chiefly between theoreticians and experimental or comparative
biologists.
There is, of course, a limit to what can be achieved in three days of intense
discussions of this kind. Thus, much of what has been presented - and is brought
together in this volume - still stands side-by-side; it is not yet amalgamated into
a coherent body of knowledge. And yet, some of the gaps could be bridged and the
discussions resulted in a broadened perspective that may well lead to new insights
when the concepts presented are developed further. Some of this process should
be captured in these remarks which are based on a summary prepared at the end
of the conference and should serve as an introduction to the relations between the
different contributions.
In that sense I propose to review the insights on the role of fractals in biology
and medicine, gained at this symposium, under three main topics: morphogene-
sis, structure-function relationships, and the assessment of structural complexity. I
would then conclude with some remarks on the role of the eye and the problems
related to the attempt to reduce shape perception to simple numbers or words.
1 The fractal nature of morphogenesis in biological systems

The formation of structures in living systems is a non-linear dynamic process, in
general, if only because structures form into a limited space. Furthermore, neither
animals nor plants can grow indefinitely but approach their final size asymptoti-
cally. This all imposes limits on the morphogenetic processes, and much of this
becomes translated into structures that appear to have fractal properties. This has
become evident in the morphogenesis of fungal structure, of lung, bone and blood
vessels. It is particularly noteworthy that very similar structures appear in the
course of diffusion limited aggregation (DLA). The rules of proportional growth
also translate into Phyllotaxis, the prevailing pattern of structural order in many
plants, based on related logarithmic spirals.
The considerations of the evolutionary interplay between mutation and selec-
tion also falls under the perspective of morphogenetic principles. The dichotomous
growth of bacterial populations gives rise to tree-like structures within which mu-
tations and gene exchanges become propagated. At first sight, these patterns are
not fractal because the scale is invariant. However, carrying the principle of the
variation generator from the level of the nucleotide to the levels of genes, cells,
4 Weibel
organisms, and populations may well reveal scale invariant variations eventually,
and this may well introduce new views on how biological diversity evolved.
Finally, the development of an epidemic, such as that currently observed
with respect to the human immune deficiency virus infection is related to «mor-
phogenetic principles» since it determines the development of the structure of a
population, but here again it was not readily apparent that fractals must be invoked
in describing these developments.
2 Importance of fractal analysis for understanding

structure-function relationships
A central concern of this Symposium has been to explore the role of fractal struc-
ture in relation to the biological functions served. The attempt to learn from a com-
parison of living systems with engineering constructions, such as the metabolism of
living organisms compared to bioreactors, or the equivalence between electrodes in
a battery and gas exchange in the lung have introduced a clearly mechanistic con-
cept in which fractal design offers distinct advantages. Some of this is expressed
in well known features of living organisms, such as allometry which describes
design and function as non-linear relationships to size or scale.
Fractal design principles occur in a very large number of biological structures
in relation to many functions. The design of the airway and vascular trees, of
the cardiac conduction system and of neurons, and the structure of interfaces at
various scales (cytoplasmic membranes, plasma membranes, gas exchange surface
of the lung), are a few examples only. The question was raised why are fractals
ubiquitous? Why is fractal design in biology important?
In that respect it is of interest to note that fractals occur in nature under a
variety of conditions. They first of all result in many random phenomena where
randomness or disorder shows similar patterns at various scales, such as in random
walks or in percolation. Similarly, irregular objects such as coastlines and trees
will show fractal features. On the other hand, fractal structure occurs in non-
linear systems, such as in turbulence; it appears that vascular trees mimic the
pattern of turbulence. Fractal structure also characterizes the interfaces in efficient
exchangers such as electrodes and the lung. It has been noted that nature does not
build in one scale but that living organisms in fact occur from microbes to the
elephant. Similar design features are evident among species differing in size by
several orders of magnitude, in mammals ranging from two grams to several tons,
but most structural and functional parameters are not varying linearly with body
size but rather follow logarithmic or power law regressions of allometry.
One of the interesting conclusions on the importance of fractals in biological
design is that it imparts the organism with a considerable error tolerance. The lack
of a preferred scale in biological evolution results in an important evolutionary
advantage of fractal progeny. Also, the fractal nature of the heart beat, resulting in
Introduction and Summary 5
chaotic complexity of heart frequency, appears to be a safety measure, since the

occurrence of regular periodicities may lead to cardiac arrest.
To consider the role of fractal design and fractal organization of time struc-
ture in dynamic events holds a great promise for the future. The discovery of
similar patterns between biological structures and fractal models will allow the
derivation of complex models that describe functional processes. This will allow
the formulation of precise hypotheses on structure-function correlation which can
be subjected to experimental test.
3 Assessing complexity of biostructures in health and disease

The fractal dimension of a structure is a measure of its complexity. Thus it is
tempting to use the fractal dimension as a parameter to describe changes as they
occur in pathological alterations of structure. For example, transformation of cells,
such as leucocytes, or of tissues in malignancies in general cause a greater degree
of disorder, particularly at the interface between the malignant cell or tissue and
its surrounding. It would, therefore, appear that this should be expressed in an
increase of the fractal dimension. Several examples were presented where this
indeed seems to be the case; however, the predictive value of the fractal dimension
for the degree of malignant transformation is still ambiguous. It may well be that
the fractal dimension alone is not a sufficiently sensitive parameter. Mandelbrot, in
his concluding paper, has shown that structures with the same fractal dimension can
have very different structure and that additional measures such as that of lacunarity
are required for a better characterization of structure. In general it would, therefore,
appear, that one would have to define a set of parameters to characterize malignant
transformations, one of which would be the fractal dimension.
The same may be true in the attempt to define the structure of neurons and of
neural networks where the estimation of fractal dimension alone cannot provide
the discriminatory parameters of different types of neurons. This again will require
a set of descriptors, fractal and also metric, to arrive at a full picture. Thus one
of the important tasks for the future will be to find such parameter combinations
which make sense, also from a functional perspective.
With respect to using fractal concepts in structural analysis a number of dif-
ficult technical problems must still be overcome. One of them is the rigorous
development of methods of measurement based on the basic principles of stereol-
ogy since microscopic measurements are done on sections, and it must be clearly
established how measures of the fractal dimension or of lacunarity are translated
into the threedimensional space. Also, we must beware that we are not simply
measuring the fractal dimension of artifacts of the method. This is particularly
dangerous when using methods of automatic image analysis where the finite pixel
size and routine methods of smoothing and image segmentation may well have an
important effect on these measurements.
6 Weibel
4 The role of the eye in morphology

One of the important messages that Benoit Mandelbrot conveys in his lectures is
that the development of science has led to an elimination of the eye from science.
The development of biology and biological morphology is no exception to this rule.
Early last century textbooks of anatomy and embryology were written as purely
verbal descriptions without any illustrations. In recent decades the tendency has
evolved to replace pictorial evidence by numbers also in microscopy - and I
must confess guilty of having promoted this approach through my own work.
Quite clearly, to attach numbers and precise size estimates to biological structures
is important particularly in view of interpreting structure-function relationships.
However, we must realize that every attempt at reducing the image of structure to
a few numbers or words results in a wilful loss of over 99 % of the information
contained in the picture. A student has once sprayed a graffito on the wall of my
Institute which says: «One word = one millipicture». This is the only graffito I
have not had erased because I believe it conveys an important message.
Any measurement of biostructures only makes sense if it is interpreted crit-
ically in the framework of knowledge on the integral nature of the structures
concerned.
The lecture by Benoit Mandelbrot at the end of the Symposium made it
very clear that the application of the principles of fractal geometry in biology and
medicine is only at the beginning. He impressively showed that by estimating the
fractal dimension from the slope of a log-log regression one is casting away a vast
and rich amount of information that is contained in the structure. At least two new
parameters must be taken into consideration. One of them is lacunarity which is
a measure of texture of the fractal structures, or of the distribution of local fractal
properties over the entire structure. It is also evident in the study of many natural
structures that the fractal dimension need not be constant throughout the structure,
particularly at different levels of organization, and that the multifractal properties
are a most important feature of structural design.
5 Conclusion
This Symposium has brought together on Monte Verita, the «Mountain of Truth»,
biologists, medical scientists, physicists and mathematicians to explore the po-
tential of fractal geometry in describing biological phenomena in a very broad
perspective. I would conclude that the undertaking was very worth-while and pro-
ductive in the sense that it has not only shown the potentials of this concept, but
also the pitfalls of a too simplistic application of these principles. It has, however,
also become very clear that the further development of the principles of fractal
geometry holds much in store for improving our approaches to understand the or-
ganization of biological organisms and its role in maintaining health and causing
disease.
Fractal Geometry
and
Biomedical Sciences
A Fractal's Lacunarity, and how it
can be Tuned and Measured
Benoit B. Mandelbrot
Yale University
New Haven CT 06520, USA
IBM Fellow Emeritus, IBM Yorktown Heights, NY 10598, USA
Abstract. The main exhibit in this paper is a stack of Cantor dusts that have identical fractal dimensions
but differ violently from each other. Some look clearly fractal, while others look to the unassisted eye
as filled intervals (they are said to be of low lacunarity), and others seem to reduce to the end points
of a hollowed interval (they are said to be of high lacunarity). One of several quantitative measures of
lacunarity is put forward, and the impact of low lacunarity fractals on modeling of nature is discussed.
1 Introduction
This brief paper is mostly devoted to fractal lacunarity, but the introduction itself
begins with brief but rambling general comments, and the last two sections also
digress.
The readers of this Proceedings need not be told that, today, the basic ideas
of fractal geometry are viewed as being astonishingly simple. It is becoming hard
to believe that, until I introduced computer pictures in a broad and deep fashion
as a tool of science, and used old and new fractals to model nature, those shapes
were few in numbers, were known to few scientists, and were viewed as «bizarre»
and «pathological.» Today, to the contrary, the fractals keep increasing in numbers
and variety, and they attract the attention of everyone - including children. This
sharp change due to picturing and applications deserves special mention in this
Proceedings devoted to medicine and biology, because it unavoidably raises a very
important question of anatomy and/or of physiology: Could it be that, somehow,
human vision is «wired for fractals,» as we know it is wired for line and color?
A second look shows, however, that the apparent simplicity of fractal geom-
etry is altogether deceptive in many ways. After a field has sufficiently matured,
an «intuition» becomes formed. One acquires a «feeling» for those cases where
«cookbook» techniques can be trusted blindly, as contrasted with the cases where
those techniques do not work or must be used with caution. Fractal geometry,
however, has not yet moved beyond the early stage where even the most skilled
users often grope for the proper tool.
To continue the preceding comment in somewhat self-serving style, all too
many readers of FGN (The Fractal Geometry of Nature [4]) seem to recall my
book as devoted to the notion of self-similarity, and to Julia and Mandelbrot sets.
But this is an oversimplification. In fact, my book also deals with other topics,
including self-affine fractals, multifractal measures and lacunarity. The first two
A Fractal's Lacunarity 9
of these topics, which were relegated to Chapter 39 of FGN, have by now been
taken up by many scientists, spawning a vast literature, and the best I can do here,
in Sections 7 and 8, to direct the readers among the books and papers that concern
these topics and followed FGN.
Lacunarity, to the contrary, was allowed several chapters in my book. The
stress was on qualitative considerations. True, page 315 of FGN hinted that the
key to a quantitative approach resides in the properties of diverse prefactors -
mostly the average and their variability, but this hint was not developed. Scattered
uses ensued (e.g., [3]); but, to my surprise, lacunarity was slow to develop beyond
the treatment in FGN. Now I have taken the topic up again, using prefactors as
hinted in FGN. Several papers are on the way, and this paper's goal is to provide
an advance «announcements» of material to appear. Figure 1 and the accompany-
ing comments merely present pictorially simplest form of the considerations and
illustrations one can find in FGN, Chapters 34 and 35. But this paper's Section
3 incorporates comments that are new and important. They will make lacunarity
more complicated, but also, I hope, easier to understand and less controversial.
Section 4 is the first presentation in print of one of the basic numerical measures
for lacunarity.
2 Examples that show that, for Cantor dusts, a given fractal

dimension is compatible with a wide range of different
textures that seem to range from an interval to two points
Let us examine Figure 1, beginning with the «midline,» which is the line to which
a hand points. The bottom part of this line represents the generator of a standard
Cantor dust of fractal dimension D = 1/ 2. To construct this dust, one starts with
the full closed interval [0,1]. First, one erases the open middle half, leaving N = 2
closed intervals, each of length r = 1/4 = N - 2 . Just as in FGN, Plate 60 (which
illustrates the original removal of middle thirds), those intervals are shown as bars,
because otherwise they would be invisible. Second, one erases the middle halves
of each of the above remaining intervals. The same process, if continued without
end, erases most of the interval [0, 1], except for the fractal dust that is pictured
in the top part of the midline of Figure 1.
A fractal dust with D = 1/2 can be increasingly «homogenous-looking.»
Consider the line that is second down from the midline on Figure 1. The bottom
part represents a generator made of N = 4 intervals, with r = 1/ 16 = N- 2 . The
novelty is that they are uniformly spaced. Similarly, on the bottom part of the
k-th line from the midline is a generator with uniformly spaced intervals, such
that N = 2k and r = N- 2 . Each of these generators can be used to construct
the Cantor dust pictured on the top of the corresponding line. The corresponding
fractal dimensions are
D _ log N _ log N _ log N _ 1
(1)
- log l/r - 10gN2 - 210gN - 2'
10 Mandelbrot
~
~
<')
§.
• • • • ",
t
<')
l::
I I I I I I I I I I I I I ;:s I
~
...
~.
I IIIIIIIII I IIII I I III II I I II

II I III II i II I 1111. I II II
"I
Fig. 1 A stack of Cantor dusts. By construction, all these sets are self-similar, and have the same fractal dimension D = 1/2. Yet, as one moves from bottom
to top, one seems to move from a full unit interval with no holes, through a number of sets clearly recognizable as fractals, up to a unit interval that has
been hollowed out, so that only the endpoints are left in. Hole is lacuna in Latin, and this paper proposes a measure of a fractal's lacunarity (= tendency
to have holes) that varies from 0 at the bottom to I at the top. -
12 Mandelbrot
In other words, every Cantor dust in the lower half of Figure 1 (midline or
below) has precisely the same fractal dimension. Yet, these dusts differ violently
from each other. In particular, the dusts near the middle of the whole stack are
very far from being homogeneous, i.e., they are far from being translationally
invariant. Despite the fact that few stages of construction can be shown, everyone
acquainted with fractal geometry immediately recognizes the outcome as being
fractal.
But the fractal at the bottom of the stack is sharply different. As a matter
of fact, it is likely that the unavoidable imperfection of printing shows it as a
solid bar; one can say that the bottom line mimics a homogeneous medium; i.e.,
it seems to be translationally invariant. More generally, the shapes one meets as
one goes down starting in the middle of the stack «look less and less fractal.» If
we had plotted all the possible values of N instead of only those of the form 2k,
the transition from clearly fractal to seemingly homogeneous would have become
very progressive.
A fractal dust with D = 1/2 can be increasingly «discrete-looking.» This is
achieved by a second way of playing with the generator in the midline of Figure
1. This new way is illustrated in the upper half of Figure 1. The k -th line above
the midline, just like the k-th line below, involves N = 2k intervals of length
r = N 2 , but the generator they form is altogether different. It divides into 2 equal
groups of N /2 intervals, each group being squeezed together near an end point
of [0, 1]. (One is tempted to squeeze them into a single group - all to the right
or all to the left of [0, 1]. However, it happens that this would be self-defeating).
Now, the unavoidable imperfection of printing has a different effect: the fractals
near the top of the whole stack seem to reduce to 2 isolated points.
Generalization to D -=I=- 1/2. It is obvious that one can modify Figure 1 by
varying N, while r = N-I/D, with 0 < D < 1: it follows that 10gN/ 10g(l/r) =
D for every fractal dust in the stack. Now change D but use the same laser printer
or the same printing press. When D > 1/2 (e.g., in the removal of middle thirds,
which yields D = .6309), apparent homogeneity is achieved higher in the stack,
i.e., for a lower value of k. When D < 1/2, apparent homogeneity is achieved
lower in the stack.
3 Implications of the examples in Section 2; the loose

concept of qualitative lacunarity
Despite the striking differences seen on Figure 1, all the sets it illustrates are per-
fectly self-similar, hence unquestionably fractal by every definition. The existence
of such fractals is a prime illustration of an assertion made in the Introduction,
that the apparent simplicity of fractal geometry is altogether deceptive. To face
the obvious need to distinguish between those fractals, I was led (in the 1970s),
to introduce the new notion of lacunarity. Lacuna is Latin for lack or hole, hence
the term is self-explanatory: Lacunarity concerns the tendency to have holes. It is
as high as can be for the fractals at infinity above the top of Figure 1, and as low
as can be for the fractals at infinity below the bottom.
A troubling conclusion. Figure 1 is extraordinarily troubling. Years ago, while
introducing fractals into the sciences, I noted that their non-integers dimensions
made them geometric «chimeras,» that is, shapes «in between» the Euclidean
shapes like the line and the plane. To this «in-betweenness,» lacunarity adds new
depth (I almost said a new dimension). As already mentioned, lacunarity was
touched upon in FGN, but not completely enough. Lately, I have (at long last)
advanced numerical measures, of which an example will be described in Section
4, and moved forward in the study of the implications of lacunarity, first in theory,
and then in practice. Here is a sketch of some of these implications.
In the theory that holds after infinite interpolation, the answer is simple:
Euclidean shapes differ from all fractals. As k --+ 00, many subtle issues of
mathematics are bound to arise, but this is not the place to even raise this sort of
question.
In practice, the answer is very delicate. Let us proceed from the essential
point of view of physical problems that involve a positive smallest atomic size E
and a largest size L. This double cutoff confirms the point of view of what the
eye can actually see.
In a Euclidean universe cut off at E and L, physical reality is embodied in
L /E atoms. A continuous interval of length L is nothing but an approximation to
those L / E atoms, but continuity deserves attention because it happens to be far
easier to investigate mathematically.
Now let us tum to the fractal universe. Far enough down the stack in Figure
1, the values of k and N are so large that r = N-I /D :::; Eo If so, a further
fractal interpolation of the generator has no physical meaning. It follows that the
opposites meet: a collection of L / E atoms is a common discrete model for a) a low
lacunarity fractal and b) an interval. Up the stack in Figure 1, similarly, one cannot
distinguish between a) a high lacunarity fractal b) the endpoints of an interval.
Two conclusions. On the line a fractal can «mimic» an interval or two points,
or any set of dimension different from its true dimension. The effect of lacunarity
on our perception of dimension may be illustrated by a well-known metaphor: that
of the «wind chill» factor expressing the effect of wind speed on the perception
of temperature.
The distinction between fractal and Euclidean is only obvious for «middling»
values of lacunarity. As lacunarity decreases or increases in a concretely meaning-
ful environment, the distinction between fractal and Euclidean becomes increas-
ingly blurred.
The case of the distribution of galaxies. From the above second conclusion,
one must expect, given a a set of empirical data, that some statistical tests or other
methods of analysis will declare this set to be fractal, while other tests declare it
to be homogeneous.
14 Mandelbrot
Such a source of possible controversy is by no means hypothetical. As a

matter of fact, the context which compelled me (in the 1970s) to develop the
qualitative notion of lacunarity was the study of the large scale structure of the
universe, which continues to this day to be full of controversy. Very recently, thanks
to the comments summarized in the preceding few paragraphs, I have gained a
better understanding of the points in dispute.
The conventional viewpoint [6] derives from the following facts. On the one
hand, there is an overwhelming visual impression that the maps of galaxies are
more or less unifonn. This explains the conventional view that the overall distri-
bution of galaxies is homogeneous. But, on the other hand, there is overwhelming
numerical evidence that the large scale distribution of galaxies is fractal in a «local
range» up to a crossover at 5-20 Mpc (depending on the definition of the crossover;
the abbreviation pc stands for parsec, equal to a few light years). However, I saw
no reasons to believe that the distribution fails to remain fractal throughout. The
fact that this scenario is conceivable, is established by the material in FGN, Chap-
ters 33 to 35. Defenders of the conventional view did not even design to discuss
the low lacunarity scenario, and later they tried to rough up L. Pietronero when
his analysis of the actual data failed to produce any crossover. This analysis is
summarized in [7]. Were the discussion intrinsically limited to galaxy maps pro-
jected on the sky, I would fear that a controversy could continue forever. But
evidence that clinched the fractal model of galaxies came only with the advent
of three-dimensional evidence about the distribution of galaxies. The widely ad-
vertised «big voids» contradict homogeneity but are a standard characteristic of
fractals.
The preceding case story yields a disquieting conclusion: the same difficulties
are likely to be repeated whenever a possible fractal model is «atypical,» i.e., like
those up or down the stack in Figure 1.
From the fact that the value of D does not suffice to characterize a fractal
numerically, an upbeat conclusion is that there is work to do. I have been chal-
lenged to proceed from qualitative to quantitative lacunarity. At long last, I have
responded to this challenge, as will be seen in the following section.
4 A possible measure of lacunarity: via the filling

rate of E-neighborhoods
FGN largely left lacunarity as a loose visual notion. The next vital task was to
make it quantitative. Having now faced this task, I soon discovered that it can
be achieved in several separate ways. Each must be studied on its own, and they
face us with the new task of studying the relation between separate measures of
lacunarity (see Section 6).
It will suffice here to sketch the measure of lacunarity that is based upon a
set's f-neighborhood (or Minkowski sausage). Behind its complicated name, there
is a very simple notion, to which we have already alluded in words. The basic
idea is that infinitely interpolated fractals are neither visible nor physically realistic.
Observational or experimental science necessarily views any set as being a little
out of focus. Thus, «point» is the proper name for «a little ball of radius E.» If each
point in a set is replaced by such a small ball, one obtains the set's E-neighborhood.
On the line, each point is really an interval of length 2E; neighboring points in
a fractal dust smear out together to form intervals, separated by other intervals,
named gaps, which are of length> 2E.
Let us now evaluate the total length L (E) of the E-neighborhood of a Cantor
dusts in Figure 1, and determine the dependence of L(E) on E. Part of the answer
is well-known and reported in FGN, p.358: in the 1920s, Bouligand showed that
(2)
here the symbol ex means that
lim [log L (E) j log E] = 1 - D. (3)

£---.0
On the line, 1 - D, is a set's codimension C. This result of Bouligand holds

for all fractal dusts, and yields one of several alternative paths to introduce D as
a fractal dimension. But the very generality of this result means that (2) does not
take account of the special structure of the Cantor dust. In particular, for some
dusts in Figure 1, the covering by intervals of length 2E may be efficient, i.e., with
relatively little overlap; for other dusts, the covering may be very inefficient, i.e.,
with extreme overlap. In order to compare degrees of overlap, I observed that in
certain cases that include those of Figure 1, one has the following far less general
but far stronger property:
'Circumstantial evidence suggests that in the case of a fractal it is impossible for

L(E)E O - 1 to be a constant.
For the lower half of the stack in Figure 1, the calculation was performed by
J. Klenk. Using (2E) for reasons that will transpire soon,
-0 1)0 0(1 1-0)1-0
Ijmax{L(E)(2E)0-1} = (r - r - r (5)
1- r
One can imagine using max{L(E)(2E)0-1} to define IjAF, and calling AF

the rate of lacunarity. This maximum is attained when E is of the form E = rn (1 -
Nr)j(N - 1) for integer n. Between two successive values of E in this sequence,
the variation of L(E) is linear. Another calculation yields min{L(E)(2E)0-1}. One
may want to introduce this min in the definition of AF . Yet another possibility is
to take a wide range of E'S and define 1 j AF as the average of L (E) (2E )0-1, or
better, define -log AF as the average of log [L(E)(2EP-I]. In the latter case,
10gL(E) - (1 - D) log(2E) (6)

16 Mandelbrot
0.8
0.6
0.4
0.2
0.0 b=:::±=====....L-===±:::::==---==::1I:::::-L...U
0.0 0.2 0.4 0.6 0.8 1.0
o
Fig.2 The variation of l / max{L(c)(2€ )D-I} with 0 and N, in the case of a family of fractal
dusts. This family includes the stack illustrated on Figure 1 and the corresponding stacks for
other O's, and fills them for all values of N. Each vertical slice corresponds to the 0 given
in the abscissa, and each curve corresponds to a given N, with r being obtained as N - 1/ D •
The horizontal line separates the high - and the low - lacunarity cases. The top 4 lines
correspond to high lacunarity dusts; the bottom 4 lines correspond to low lacunarity dusts,
all constructed as in Figure I. In each case, as one moves away from the horizontal line, one
has N = 2, 10, 102 , 104 , 108 , 10 16 and 1032 •
will oscillate around its average - log AF . Before selecting AF among these pos-
sibilities, I feel it is best to await further developments in the field.
The same argument holds for the upper half of the stack in Figure 1. The
calculation (performed by I. Yekutieli) shows that
(1 1-0)1-0(1 0)
l/max{L(E)(2E)0-1}= -r -r . (7)
1- r
The values of both expressions for 1/ max{L(E)(2E)0 - 1} are plotted on Fig-

ure 2. As one goes down the stack on Figure 1, the value of AF decreases from
a maximum of 1, which holds in the most lacunar case (top) to a minimum of
0, which holds in the least lacunar case (bottom). The fact that this scale ranges
from 0 to 1 is most satisfactory and came out without having to be forced upon
the problem, except for the use of (210) instead of Eo (I do not know whether or
not the inequality 0 < 1/ max{L(E)(2E)0-1} < 1 holds in other constructions).
After a choice is made between the possible definitions of AF, I shall propose
AF as a measure of the lacunarity of a Cantor dust. Closely related definitions are
possible in higher-dimensional spaces.
It may also be worthwhile to consider the variability of L(E)(2E)0 - 1, as
defined by max-min, or by the variance of either L(E)(2E)0-1, of its inverse, or
of its logarithm. This variability also contains lacunarity information.
5 Generalization, random dusts, and a connection between

low lacunarity and a form of global statistical dependence
For exactly (recursively) self-similar fractals, the prefactor of the E-neighborhood's
length settles one aspect of the issue of lacunarity. But there are numerous other
aspects: we have studied rates of lacunarity based on the prefactor of scaling
relations other than (6), and began to examine the connections between various
rates.
The original Cantorial example described in Section 2 is extremely over-
simplified, and its systematic character is unrealistic, especially after it has been
generalized in the plane and higher spaces. Hence the need for random fractal
models. When dealing with them, one must face an additional problem; for ex-
ample, normalization is very far from being completely solved. These problems
cannot be studied here, and are only mentioned for an already-discussed reason:
to avoid creating the false impression that the problem of lacunarity is already
under full control.
I have devoted much attention to a special random fractal construction based
on tremas. This simplest form consists in cutting out statistically a collection of
independent filled-in intervals, discs (filled-in circles) or balls (filled-in spheres).
Chapter 33 of FGN illustrates this procedure, and shows that the resulting sets
are of high (but not extreme) lacunarity. An example is provided by Figure 3,
which reproduces part of Plate 307 of FGN. The next obvious step was to replace
spherical by non-spherical tremas. This change turns out to be sufficient to de-
crease lacunarity. In FGN, Chapter 34 and 35, this decrease was assessed visually
and quantitatively. Major progress has recently been reached towards quantitative
description, but a full description of what is already known would go beyond the
bounds of this paper.
The inquisitive reader may wonder which feature of the non-spherical tremas
makes them fit to accomplish the desired task. As I see it now, the basic underlying
feature is an increase in a very original form of global statistical dependence.
6 The tedious issue of how to define fractals, fractal

dimension and fractal lacunarity
At this stage, fractal lacunarity cannot be measured by a single well-defined num-
ber. It is like bigness, a broad notion in the process of being broken down into
distinct strands that may well coincide in some cases and differ in other cases.
Will this process converge and, if it does, how rapidly? I am not optimistic
about convergence, because of the increasing complications that accompany other
broad notions, those of fractal and of fractal dimension. Many years ago, I observed
that structural similarities no one had noticed should bring together several existing
mathematical construction and concepts. They deserved to be put together, given a
common name, and used as a base (I never said foundation) for a new discipline,
18 Mandelbrot
Fig. 3 A slice of Plate 307 of FGN, illustrating a trema set obtained by cutting off filled-in circles
from the plane. (From: Benoit B. Mandelbrot, The Fractal Geometry of Nature; W.H. Freeman
& Co., New York, and also from: Benoit B. Mandelbrot, Die fraktale Geometrie der Natur;
Birkhliuser Verlag, Basel)
which I called fractal geometry. Did I succeed in defining fractals uniquely? Not
only have I failed, but recent work on fractal aggregates (DLA) is revealing that
shapes everyone wishes to call fractal can exhibit deep complexities, for which
none of the previously advanced possible definitions would allow.
What about fractal dimension? Some of these notions I was able to bor-
row from existing mathematics literature were several distinct generalizations of
dimension. The best known was due to Hausdorff and Besicovitch and purely
tactical reasons (as I have kept repeating) led me at one time to give it top billing
in fractal geometry. But this top billing has long since become inappropriate and
not sustainable. It becomes increasingly clear that fractal dimension is necessarily
a multi-faceted concept.
Against this background, I have little hope of finding a unique measure of
lacunarity.
7 Bibliographic guidance concerning self-affinity

To gain an understanding of self-affinity, the best source is the collection of papers
put together by Family and Vicsek [2].
To preface a few words on the topic, observe that, Cartesian coordinate axes
having become second nature to every scientist, there is little awareness of the
extremely fundamental distinction between two kinds of planes: the isotropic Eu-
clidean plane and the affine planes. Let us state this distinction on two sharply
different cases.
For a clearcut example of isotropic plane, consider a map of a piece on Earth's
surface that is so very small that Earth's curvature does not matter. The custom is
to place the coordinate axes along the direction NS and the orthogonal direction
EW. But all other orthogonal lines would do equally well. In particular, distance
can be measured in any direction.
For a clearcut example of self-affinity, begin with a graph of volume versus
pressure. Distances along each axis, i.e., differences of volume bo V or of pressure
bop are meaningful and can be measured. But what about J (bo V)2 + (bop)2, which
is a formal distance along a straight line between two points on the diagram. This
formal expression depends on the units chosen for V and p. But those units can be
chosen independently of each other; therefore, this formal distance is meaningless.
The same command applies to «noises» that are voltages as function of time.
Turning a knob changes their appearance. The same comment also applies to the
Brownian motion.
The reader must be warned that there exist examples that do not fall neatly
in the above contrast. As obvious example is a vertical cut across a mountain.
In that case, vertical and horizontal coordinates are both measured in meters, and
J (boX)2 + (boh)2 is well defined as the «distance as the crow flies» between
the two points. However, horizontal and vertical distanct,!s have distinct meanings
because of gravity.
Be this as it may, self-affine fractals have by now spawned a large literature.
Once again, the early papers are collected in [2].
20 Mandelbrot
8 Bibliographic guidance concerning multifractals

To gain an understanding of multifractals, I recommend a text I wrote with Carl
Evertsz [1].
To preface a few words on this topic note that the bulk of elementary fractal
geometry is devoted to fractal sets. A set's visual expression is a region drawn
in colored ink against white paper or board (or in chalk against a blackboard).
A set's defining relation is an indicator function I(P), which can only take two
values:
I (P) = I, or I (P) = «true,» if the point P belongs to the set 5;
and
I(P) = 0, or I(P) = «false,» if P does not belong to S .
However, most facts about nature cannot be expressed in terms of the contrast
between «black and white», «true and false», or «1 and 0». I faced this issue early
on, and the opening lines of chapter IX of my first book on the subject (written
in French) translate as follows:
»Before we generalize [fractal sets to measures], it may be recalled that,
among our uses of fractal sets [to describe nature], several involve an approxima-
tion. While discussing clusters errors, we repressed our conviction that, between
the errors, the underlying noise weakens, but does not stop. While discussing the
distribution of stars, we repressed our knowledge of the existence of interstellar
matter, which is also likely to have a very irregular distribution. While discussing
turbulence, we approximated it as having [nonfractal] laminar inserts. In addition,
no new concept would have been needed to deal with the distribution of minerals.
Between the regions where the abundance of a metal like copper justifies com-
mercial mining, the density of this metal is low, even very low, but one does not
expect any region of the world to be totally without copper. All these voids [within
fractals sets] must now be filled - without, it is hoped, inordinately modifying
the mental pictures we have achieved thus far.
«Those aspects cannot be illustrated by sets; they demand more general math-
ematical objects that succeed to embody the idea of «shades of grey.» Those more
general objects are called measures.»
It is most fortunate that the idea of self-similarity is readily extended from sets
to measures; self-similar measures are usually called multifractals. The original
approach to multifractals, which is probabilistic, was first described in two papers
I published in 1974. However, multifractals became best known thanks to formal
and heuristic papers that came out in the mid-1980s. Against this background of
my approach, the nature of various heuristic arguments in those papers becomes
clear, their limitations and proneness to error become obvious, and the unavoidable
generalizations demanded by both logic and the data become easy.
Once again, as a simple introduction to multifractals (with a bibliography), I
recommend [1] to the readers.
Acknowledgements
Figure 1 was drawn by Lewis N. Siegel, and Figure 2 was drawn by Juergen
Klenk to illustrate the formulas he had obtained. The details of the theory of
fractal 1acunarity are being worked out with several collaborators, including J.
Klenk and also D. Fracchia, M. Frame, I. Yekutieli, R. Cioczek-Georges, and D.
Gatzouras. The mathematics is related to the so far little-used notion of Minkowski
content as generalized from integer to non-integer dimensions.
References
[1] Evertsz, C.J.c. and Mandelbrot, B.B: Multifractal Measures. Appendix B to
Chaos and Fractals by H-O. Peitgen, H. Jurgens and D. Saupe. New York,
Springer, 1992, pp 849-881.
[2] Family, F. and Vicsek, T.: Dynamics of Fractal Surfaces, Singapore: World
Scientific, 1991.
[3] Gefen, Y., Meir, Y., Mandelbrot, B.B., Aharony, A: Geometric Implementa-
tion of Hypercubic Lattices with Noninteger Dimensionality by Use of Low
Lacunarity Fractal Lattices. Phy. Rev. Lett 50: 145-148, 1983.
[4] Mandelbrot, B.B.: The Fractal Geometry of Nature. New York, W. H. Free-
man, 1982.
[5] Mandelbrot, B.B.: Plane DLA is not self-similar; is it a fractal that becomes
increasingly compact as it grows? Physica A 191 95-107, 1992.
[6] Peebles, P.J.E.: Principles of Physical Cosmology. Princeton University Press,
1983.
[7] Pietronero, L.: The Fractal Structure of the Universe, Physics Reports 213
(No.6) 311-89, 1992.
Spatial and Temporal Fractal
Patterns in Cell and Molecular Biology
Theo F. Nonnenmacher
Department of Mathematical Physics
Albert-Einstein-Allee 11
University of Ulm
D-89069 Ulm
Abstract. When discussing spatial and temporal fractal patterns in cell biology, we critically investigate
various methods for a practical determination of the fractal dimension D . Asymptotic fractal formulas
are presented and applied to describe self-similar cell profiles or - in the temporal case - scale
invariant protein dynamics. However, because cells are of finite size, and due to the existence of a
largest time scale for dynamical processes, such geometrical objects or dynamical processes are not
fractal even in an asymptotic sense, instead they may be only «fractal between limits». A method of
finding these scaling limits (crossover points) and their mathematical characteristics will be proposed
and applied.
1 Introduction
Many objects in biology and medicine like tissues, cells or subcellular organelles
display irregular shapes and discontinuous morphogenetic patterns in support and
in connection with their functional diversity. To capture all this richness of complex
structure and function into a theoretical model is one of the major challenges of
modem theoretical biology. In recent years a wide range of irregular structures
observed in physical and chemical sciences has been quantitatively characterized
by using Mandelbrot's concept of fractal geometry [13]. However, applications of
this concept to study self-similar patterns are rather scarce in cell and tissue biology
despite the fact that it can also be used as a design principle for morphological
complexity and function in living organisms [28].
In this paper, we will first apply and critically discuss different methods of
digital image analysis to measure the fractal dimension D for geometric structures
of cellular profiles. As prototype examples we investigate two samples of cells: (i)
19 human T-lymphocytes from various normal donors and (ii) 18 hairy leukemic
cells from two different patients.
Second, in order to discuss scale invariant protein dynamics, we need to deal
with the concept of time-fractals. Just as geometric fractals do not have a character-
istic length scale, a fractal time process generates fluctuations which do not have
a single or characteristic time scale and consequently, such fluctuations cannot
be adequately designed with standard statistics based on mean and variance. Levy
statistics [11] [14] which can appear both in time and space is an example for such
a non-standard statistics that brings into play power law behavior f(t) '" t-q for
Spatial and Temporal Fractals 23
a fractal interpretation of certain stochastic, chaotic-like processes. More recently,

some interest has been focused on the question of an integral or differential equa-
tion that leads to power-law solutions t-q where q is not necessarily an integer.
Such an analytical approach is based on fractional operators like dlJ!(t)/dt lJ with
o < v < 1. Here, we will present the basic ideas of such a fractional operator
design. We will relate the scaling dimension q to the fractional order v of the
operator dlJ / dtlJ, and we will discuss scale invariant kinetic mechanisms as an
application to protein dynamics in Myoglobin and in ion channel gating kinetics.
2 Quantitative Structural Analysis of Self-Similar Cell Profiles

In this section, we will critically discuss and apply different experimental methods
of digital image analysis to measure the fractal dimension D of cellular profiles.
The pioneering work using the concept of fractal geometry in cell biology was
done by Paumgartner et al [22]. They estimated mitochondrial and endoplasmic
reticulum membrane surface densities stereo logically and found fractal behavior
up to a critical magnification. Using computerized image analysis, Rigaut [23]
observed that most cell contours are ideally fractal only asymptotically at low res-
olutions and at higher resolutions the cell perimeter tends to a maximum. Methods
of digital image analysis are also adapted to measure the fractal dimension for hu-
man granulocytes [15], lymphocytes [10] [12] and for neurons [26], for example.
The mathematical basis for an understanding of fractal geometric structures,
their measurement and interpretation is the Richardson-Mandelbrot equation [13]
L(f) = N(f)f, commonly used in applications, where L(f) is the length (perimeter)
of the closed cell contour in the image plane, f is the unit length to measure the
perimeter L ( f ), and
(1)
is the number of unit lengths f necessary to cover the perimeter L(f). D is the
fractal contour dimension which increases from 1 to 2 at increasing «wiggliness».
Inserting (1) into the Richardson-Mandelbrot equation, one obtains the perimeter
scaling law
(2)
Mathematically generated fractals (Koch curve, Sierpinski triangle, etc.) show the
property of exact self-similarity over an infinite range of f-scales, i.e. f takes all
values between zero and infinity. Self-similar geometric objects observed in nature,
however, are usually random fractals with a limited range of f-values (i.e. fl < f <
f2) in which statistical self-similarity can be observed and to which the scaling laws
(1) and (2) are applicable. Self-similarity is an important defining characteristic of
fractal geometry. Hence, strategies for measuring the fractal dimension D cannot
ignore the scaling limits fl and f2. Natural fractal objects like cells that have been
analyzed up to now manifest self-similarity over a scale length between one [10]
and almost three [12] orders of magnitude depending on the range of resolution,
24 Nonnenmacher
Fig. 1 Nonnal lymphocyte (upper row) and hairy leukemic cell (lower row) in a grey value repre-
sentation and the extracted contours.
on the experimental device involved and on magnification as we will show in this

investigation. Besides these «technical» limits, there exists a natural lower bound
fL, defined by a critical magnification limit beyond which the surface appears as
two physical sheets of finite thickness and finite surface area [28].
2.1 Measurements of the Fractal Dimension

In our analysis of cell boundaries, the contour profiles of the cells visible on
the electron micrographs (see fig.l) were picked up by a scanner. The data set
was stored into the memory of an HP-9000/835 and a NEe SX-3 computer and
subsequently analyzed by different methods: a) yardstick method; b) box counting
method; and c) probabilistic method. The analysis reported here used the entire
perimeter (closed contour) and not just pieces of it; i.e., we determined the contour
dimension D for the entire cell profile. Pieces of it may have different values for
D as pointed out by Mandelbrot [13]. According to the various methods a) to c)
different estimates of D are obtained.
By making use of the Yardstick Method (YM), the computer program selects
a yardstick size of length 10, selects a starting point on the contour and then begins
to count the number N (f) of yardsticks of length 10 necessary to cover the contour
profile completely. This procedure is repeated for each point on the contour being
the starting point and for many different choices of f. The results are plotted in a
log N (f) versus log 10 graph, giving - if self-similarity is detected - a straight line.
According to (1), the slope of the linear region detected by this procedure directly
yields the contour dimension D = - log N (E) / log f. Throughout the paper, we
will denote by log the logarithm to base 10 (log = 10glO).
In order to apply the Box Counting Method (BCM), the image plane is divided
into small square grids, each of edge length E. Counting the number N (f) of
occupied squares, one again obtains a measure for the contour length since the
number N (f) of squares intersected by the curve is roughly proportional to the
number of steps N needed by YM to cover the contour. Starting with a smallest
f-scale (one pixel), the grid-length E was then increased successively up to 1300
pixels depending on the size of the EM-picture. In a log N (f) versus log 10 plot,
one obtains the fractal (box) dimension D for a self-similar structure. Shifting
the grid-lattice (for each f) several times within the plane, an averaged value for
N (f) is obtained. A verification of the BCM is the modified box counting method
(MBCM) which we also used for our analysis and which is described in detail by
Barth et al [1]. Besides YM and BCM, we further applied a probabilistic method
(PM) which is discussed in detail by Baumann et al [2].
2.2 Estimation of Scaling Limits

In order to find the scaling limits E\ and 102, we developed and used an automatic
fitting procedure. In contrast to a manual fit (see, for example, [10]) where that
proportion of the curve which shows a linear dependence of log L (E) upon log E
is selected subjectively, i.e. by visual inspection of curves like those given in fig.
2 or fig. 3, our automatic procedure which is based on a least square fit algorithm
searches for the widest interval [log 10 h log f2J within which the standard deviation
(s.d.) of the estimated slope does not exceed a given limit which we required
arbitrarily to be s.d. = 0.0086. This corresponds to a 95% confidence interval of
the slope estimate with a width of 0.0337. By this method, the dimension D is
estimated at a fixed s.d. = 0.0086 for each contour data set of an individual cell.
2.3 Estimation of the Fractal Dimension by Asymptotic Fractal Formulas

Fitting data points which start from a plateau for small f-values and approach a
straight line for large f-values (in a log L (f) - log E plot, see fig. 3), one can use
asymptotic fractal formulas like
(3)
or
(4)
26 Nonnenmacher
Formula (3) has been suggested by Barth (private communication). Taking g = -1

in eq. (3), one just recovers Rigaut's asymptotic fractal formula being frequently
used for the data fitting procedure. The additional parameter g that is introduced
in (3) essentially guarantees a more accurate fit than Rigaut's formula for E-values
on the plateau and in the intermediate region between the plateau and the straight
line (a, b, g and D are fit parameters).
The asymptotic formula (4) was derived by making use of Bernoulli scaling
ideas and renormalization techniques (see Eq. (10». 'Y is the incomplete gamma
function and f3 is related to the fractal dimension D via the relation D = f3 + 1. In
the limit E -+ 0 one finds lim€---+o L (E) = Lo and, consequently, Lo represents the
plateau-value for E -+ O. For large E-values, one obtains asymptotically L (E) rv
c(3, i.e. inverse power-law scaling.
Some warnings: Using asymptotic formulas for the data fitting procedure, one
will always get some value for D even if the data set is restricted to the plateau.
In such a case, however, one has not observed a fractal structure! To capture a
fractal structure in the data, you need to continue taking measurements into the
asymptotic region (at least over one order of magnitude).
2.4 Results
First, we tested the computer methods (YM and BCM) as to their accuracy and
reliability and applied them to determine the fractal dimension D for an exact
self-similar mathematical construct. YM was tested with a Koch curve. The result
is that at increasing resolution R (number of pixels) YM approaches the exact
value (Dexact = 1.261...) from below. For instance, taking R = 1200 x 1200 pixels,
we obtain D = 1.22. For R = 2400 x 2400 pixels, we get D = 1.24. BCM was
tested with a Sierpinski triangle (Dexact = 1.584 ...). The results are: (i) the higher
the resolution R, the more accurate the value of D; (ii) BCM and YM approach
the exact value for D from below, PM approaches D from above (see Ref [2]).
Figure 2 demonstrates (in a log N( E) versus log( E) plot) how the self-similar
(fractal) behavior indicated by a straight line is approached by using (a) YM (*)
and (b) BCM (+). Fig. 2 showing the data of a hairy leukemic cell contains even
more information: (i) according to eq. (1), the slope of the straight line yields the
fractal contour dimension D = 1.35, and (ii) the data points that follow the straight
line are restricted to a finite E-domain (El < E < E2) which defines self-similar
(fractal) structure of the cell profile. The upper limit E2 where the data points
break out of the linear behavior is found to be approximately E2 ~ 4 microns for
all contours of the same cell line under analysis. We interpret the existence of
this upper limit as an indication of the finite macroscopic size of the cell contour
(L(E2) ~ 24.8J.Lm) if it is approximated by a circle with radius E2. Indeed, using
an Euclidean measure for the cell perimeter L(E2) = 27rE2 = 25.1 microns, one
discovers that E2 is closely related to the classical radius of a circle. Depending
on the method used for the data analysis, the existence of a lower bound El is
observed by inspection of fig. 2. In case of BCM, we estimate Ef ~ 0.02 microns
and by using YM, we find Ej ~ 0.2 microns. In this case, BCM detects self-
.--..
10- 4
10 - 2
shut interval t/ms
Fig. 2 Contour data of a hairy leukemic cell determined by YM (*) and SCM (+). fl and f2 denote
the scaling limits (see text in Sec. 2.4).
similarity over a domain E ranging between Ef = 0.02 pm < E < E2 = 4 f-Lm which
covers a factor E21 Ef = 200, i.e. more than two orders of magnitude for the hairy
cell under consideration. In case of YM, the self-similar domain shows a ratio of
E2/Ei = 20 which is also an acceptable result because it is beyond the «standard
norm» at a factor of 10 which is conventionally required in order to talk about
«fractal structures».
Applications of the fractal concept to stereology are traditionally based on
the perimeter scaling law (2) which yields - besides the fractal dimension D -
additional structural information. Fig. 3 shows eq. (2) in a log L (E) versus log( E)
plot for two different magnifications of one and the same hairy leukemic cell.
The slope 0: of the straight line indicating self-similar structure is related to D
by D = 1 - 0:. The data points present our numerical results for the perimeter
L (E) of one and the same contour of a hairy leukemic cell taken at two different
magnifications (8 700x, 18 400x). The slope 0: of the straight line is found to
be 0: = -0.34 leading to D = 1.34 which is in good agreement with the value
1.35 estimated via fig 2. This dimension is far beyond the dimension of normal T-
lymphocytes as expected: Just by visual inspection of EM-pictures (see fig.!), more
wiggliness or roughness of a hairy leukemic cell contour can be seen with respect
to the cell boundary of a normal T-Iymphocyte. However, computerized image
analysis yields a reproducible and consistent number for the parameter D free of
any subjective influence. Consequently, D can be used as a quantitative measure
28 Nonnenmacher
60
40
Fig. 3 Data points in a log L (f) VS log € plot of one and the same hairy cell detennined by YM at
two different magnifications: M, = 18400 x (0) and M2 = 8700 x (+) . The straight line
indicates ideal scaling behavior (D= 1.34).
for the structural complexity of the cell contour. Besides the dimension D, fig.3 in
addition shows that for small E-values, the perimeter L (E) reaches a saturated value
(plateau) Lmax . The plateau value depends on the magnification M. For M = MI
= 18 400x the measured maximum value is given by Lmax = L(M,) = 60.8 fkm
(upper curve) and for M = M2 = 8700 x we find Lmax = L(M2) = 53 .1fkm (lower
curve). Additionally, we observe that the higher the magnification, the larger the
E-span that follows the straight line. In other words, the higher the magnification,
the more structural details of the contour are observable and thus one observes a
larger domain of self-similar structures.
Following the way just described, we determined the contour dimension D for
19 T-Iymphocytes and for 18 hairy leukemic cells by using BCM. A crucial point
in finding a reliable D-value for each cell may be seen in determining the scaling
limits EI and E2 . We attacked this problem by using an automatic fit procedure as
discussed in subsection 2.2. All data points between the limits E, and E2 were fitted
by a least square fit procedure resulting in a straight line. The slope of this line
gives just an individual value for D with s.d. = 0.0086. In case of T -lymphocytes,
we found individual D-values ranging between 1.12 and 1.23 for all cells of this
type. In case of hairy leukemic cells, we found values for D ranging between 1.29
and 1.37. In all cases the ratio Ez/E, is larger than 10, i.e. the scaling region is
larger than one order of magnitude. Thus we may talk about «fractal structures».
Having the D-values for several members of one and the same cell line (19
T-Iymphocytes) at our disposal, we can prepare a histogram in order to study the
underlying statistics. We found that a Gaussian distribution is compatible with
the individual D-values. Therefore, we were able to calculate a mean value D =

1.15(s.d. = 0.034) representing the fractal dimension of the sample of 19 T-
lymphocytes. In case of 18 hairy leukemic cells, we found D = 1.32(s.d. = 0.023)
to be the mean fractal dimension of this cell line.
We now shall discuss estimates of D by making use of so-called asymptotic
fractal formulas (eqs. (3) and (4» . Here, least square fit procedures take all data
points into account which are shown in fig.3 (apart from the last five points for
large f-values). Thus, one has available a large set of data in order to fit the
curve. However, one should realize that the data points (see fig.3) which enter
into and which form the plateau for small f-values (high resolution) do not fall
into the scaling region represented in fig.3 by the straight line. Nevertheless, such
asymptotic scaling formulas can be used and frequently have been used in order
to fit experimental data sets. To get some feeling for the accuracy of such an
estimation of D, we used formulas (3) and (4) for a fit of the data (magnification
18 400x) shown in fig.3 . Via formula (3), we estimated D = 1.39. From formula
(4), we found D = 1.38. These values for D are in fairly good agreement with
the estimate (D = 1.34) found by the method described in subsection 2.2 which
takes only those data points into account that fall into the scaling region between
f\ and f2. As a final comment on this section we notice that the fractal dimension
D alone does not completely specify the fractal structure because objects having
the same D may differ by other characteristics as, for instance, lacunarity.
3 Scale-Invariant Protein Dynamics

In this section, we will mainly investigate temporal aspects of scale-invariant
dynamical processes. Usually, fractals are being considered as self-similar objects
in space; i.e., the term fractal is intuitively based on a structural (geometric)
concept that applies to a wide class of complex shapes whereby self-similarity is
a key feature of such irregular geometric patterns. Just as a geometric fractal does
not have a characteristic (dominant) scale of length, a scale-invariant dynamical
process generates fluctuations that do not have a characteristic scale of time and
thus cannot be adequately measured by statistics based on mean and variance [11]
[14] [25]. If the average time of an event were finite, then this would provide a time
scale. A scale-invariant process can be regarded to be a self-similar superposition
of many processes with different time scales for which the term «fractal time
process» was coined [25] . Processes which are scale-invariant can be modelled by
Levy statistics [14] [15] [20] [25] [29].
It is generally accepted that according to their biological function, proteins
assume (during functional activity) a number of different (conformational) states.
Transitions which lead from one of these states to another are called «functionally
important motions» [5] and are usually formulated as a kinetic or relaxation process
characterized by kinetic rate constants ki = I / Ti. The number N of rate constants
k i (i = 1, ... , N) indicates how many conformational states are taking part in the
30 Nonnenmacher
relaxation process. If there would exist just a single characteristic or dominant time
scale T; = TO, one would expect an exponential Debye relaxation rv exp (-t /TO).
Experimental observations, however, indicate that measured relaxation data sets
cannot be fitted by a single exponential function . Instead, slow, non-exponential
relaxation processes are observed giving evidence for scale-invariant underlying
protein dynamics. Theoretical concepts to describle slow relaxation phenomena
like Kohlrausch-Williams-Watts relaxation exp { - (t / T)I3} or inverse power-law
relaxation t- a have been developed over the last few years in order to understand
non-standard relaxation processes in glasses [4], ion channel proteins [16] [19]
[20] polymers [6] [17] and other disordered systems. Here, we will demonstrate
how scaling ideas can be applied to self-similar, scale-invariant protein dynamics.
3.1 Bernoulli Scaling Ideas and Renormalization Technique
Patch clamp experiments [24] for studying ion channel gating kinetics provide
important information on the molecular structure and function of ion channel pro-
teins. Generally, ion channels are considered to be large integral membrane proteins
which allow passive flux of ions through cell membranes. Currents recorded from
ion channels show that the channels repeatedly open and close their pores during
normal activity. Counting the number of channel events and plotting such a count
versus duration time, one can construct a histogram which can be approximated
by a probability function f(t) for the life-time of the conducting state [24]. A
typical plot is given in fig.4 showing the data of a fast Cl- channel [3].
When modelling patch clamp data theoretically, one usually assumes that
channel gating is associated with transitions among a number of conformational
protein states with amplitudes a; and time constants T;. Data from the duration
histograms are fitted to an N -component probability density function of the form
N
f(t) = L ~ e- t / Tj
(5)
;=1 T;
where the amplitudes a; and the corresponding time constants T; are usually pre-
sumed (for the fitting procedure) to be independent. The number of free parameters
for such a discrete Markov chain is 2N - 1 because one of the 2N constants (a;, T;)
is fixed by the normalization condition [19] [20].
As already mentioned, proteins may have a large number of conformational
states and thus many terms in eq. (5) should be taken into account. For instance,
in the case of a fast Cl- channel, Blatz & Magleby [3] conclude N to be five.
For this example, one has at least nine parameters which must - and do - fit
the data.
In order to model such a situation with less than (2N - 1) parameters,
we are considering a process where the time between events is a random vari-
able. A prominent random (no memory) process is the Poisson process g(t) =
00
l/Texp(-t/T). The average time between events is defined by (t) = f tg(t)dt.
o
.. -..
10- 2
shut interval t/ms
Fig. 4 Measured data points [3] of a CI - channel are oscillating along the power-law trend (dashed
line, eq. (10» . The full curve (eq. (11) approaches the power-law trend between the limits
71 and 75, and crosses over to exponential decay for t :::: 75 (see text, Sec. 3.1).
If (t) is finite (( t) =T in the Poisson process), then a natural time scale (t) exists
in which to measure time. If one waits long enough, i.e. if the observation time
tobs is much longer than (t), events occur at an average rate k = 1/ (t). How-
ever, if (t) is infinite (actually much longer than the time of observation tobs ),
then no natural time scale exists in which to gauge measurements. In this case,
one expects to find temporal scaling laws and the concept of fractal time emerges
[25]. Now the questions arise: How can we model such a situation and which
are the physical processes behind all that? Several mechanisms can be mentioned
which generate a fractal time distribution of events. A good example is thermally
activated hopping over a distribution of energy barriers: Let G N (t) be a sum of
N Poisson processes with time constants Ti (i = 1, ... , N) each weighted by a
probability Pi of occurrence, i.e.
(6)
This time series looks similar to the standard Markov-chain series (5) in terms
of «amplitudes» Pi = ai and time constants Ti . However, G N (t) is still not in a
«renormalized» form because renormalization implies the concept of scaling. A
frequently used scaling concept is Bernoulli scaling. The basic idea [14] is this: A
32 Nonnenmacher
duration between events which is longer than an order-of-magnitude occurs with

an order-of-magnitude less probability. This scaling condition can be described
mathematically by taking Ti = T)J and Pi = aNPi with 0 < P < 1 and A > 1,
leading to
(7)
where aN [20] is a normalization constant and ~ = f / T.

This way, a distribution G N (t) having no single characteristic time scale
is generated (it incorporates many time scales Ti in form of a discrete geometric
spectrum Ai of which no time-scale is dominant). For N ----t 00 the average time (t)
00
to gauge measurements (f) = J tGN(t)df tends to infinity as required for fractal
o
time processes. Here, we note that any finite number N would lead, of course, to
a finite (f) increasing with N. However, once N and A are large enough, one can
find an average time (f) that becomes much larger than fobs. Hence, the concept
of fractal time can still be applied, but it is restricted by the limits T1 = T A and
TN = TAN (note: TN ----t 00 for N ----t 00).
Our result (7) in which the Bernoulli scaling condition is built into GN(t)
represents a first stage of renormalization. The second stage of renormalization
utilizes the idea of evaluating the discrete series (7) for G N (t) by replacing the
sum with a series of integrals using the Poisson summation formula; i.e., going to
a continuum model that brings into play the power-law representation of GN(t) '"
t-I-I-' [16] [19] [20] [29].
Here, we would like to point out that the discrete series (7) represent no more
(and no less) than the original Markov chain series (5) subjected to the constraint
requiring Bernoulli scaling in terms of Ti <X Ai and ai <X pi. Therefore, our model
does not contradict, but rather is a consequence of the Markov chain model which
incorporates many more possibilities and parameters for modeling patch clamp
data than our renormalized (scaling) model that filters out of the Markov model
just those gating processes which are consistent with Bernoulli scaling.
Applying the Poisson summation formula to the discrete series (7), one is
coming up with the power-law representation [19] [20]
L
00
GN(~) = T~~ A CVm IN(vm, ~/)..) + ~~ (A -voe-~/).. + A-Nvoe-~/)..N) (8)

m= - oo
where Vm = 1 - (In p + 27rim)/ In A is a complex-valued exponent and IN denotes

the integral
J
x
IN(vm, x) = dwwvm-Ie- w = 'Y(vm, x) - 'Y(vm, XA -N+I). (9)

x)..-N+l
10- 3
10- 6
Fig. 5 This plot demonstrates how the scaling region between the limits 7} and TN(N = 4) emerges
out of eq. (8): The dashed curve (- - -) is calculated from eq. (II) taking into account just the
term m = 0 (i.e. neglecting oscillations in eq. (8»; the full curve takes into account many
m-terms of the sum in eq. (8), i.e. m = 0, ±l, ... , ±12. The straight line (- -) indicates
ideal fractal behavior without limits. Parameters are: 7 = 0.1 , N = 4, ..\ = 10 and JL = 0.7.
In the limit N --t 00,100 given by N~rx)N (vm ,x) = ,(vm,x) approaches the
incomplete gamma function. Introducing the positive number J1, = - In pi In >., one
may write Vm = Vo - 27fiml In >. with Vo = 1 + J1" and thus, taking into account
only the term (m = 0) in (8) one obtains in the limit N --t 00 the asymptotic
fractal fonnula
( 10)
A re-interpretation of the parameters aoo I 7 In >. = (3 L0, J1, + 1 = (3, u>. =

E/Eo, and neglecting the last term in (10), brings us back to Eq. (4) which we
already offered as a three-parameter (L o, (3, EO) formula for an estimation of fractal
dimensions.
In fig. 4, Eq. (10) is tested with measurements of ion-channel data which
oscillate around the power-law trend. Similar oscillations following a power-law
trend were also observed in the lateral diffusion coefficient [18], in the QRS-
spectrum [9] and (if ~ is interpreted as generation number) in the dichotomous
branching system of the bronchial tree [15] [29].
Fig. 5 shows an illustrative plot that demonstrates how the property «fractal
between limits» emerges out of the fonnula
GN(~) = aNI(71n>')C V0 1N(VO,U>') + aN (>,-Voe -U..\ + rNvoe-~/>.N) , (11)
27
34 Nonnenmacher
which represents just the first term (m = 0) of eq. (8) and approaches the asymp-
totic formula (10) for N -> 00. Thus, if there is a finite number N of terms in the
sum (7) and consequently in the integral IN, one has two crossover points (TI = TA
and TN = TAN = TIA N- I). For t < TI, the relaxation kinetics is dominated by the
initial value GN(O) and for t > TN, the decay becomes exponential. Scale invari-
ant processes are restricted to the time interval TI < t < TN. When interpreting
measurements, it is essential to get that range, i.e. TN / TI = AN -I, included in the
data. Of course, if you would «see» in the data just a single conformational state
(i.e. N = I), then the scaling range AN - I would collapse to a single point and you
would not have observed a fractal process. In Fig. 4, we have re-analyzed the data
set [3] for a Cl- channel with N = 5; i.e., five shoulders have been identified [3].
Our fit is based on eq. (11) with T = 0.033, N = 5, f.t = 0.82 and A = 7, leading
to a scaling range T5/TI = 74 ~ 2400; i.e., three orders of magnitude before the
curve crosses over to exponential decay (for t > T5)' Thus, the way to tell what is
going on in the dynamics is to look over as many orders of magnitude as possible.
3.2 Fractional Operator Design
Investigating kinetic or relaxation processes in proteins, polymers, glasses or other
disordered systems, one observes slow decay of the relaxation functions. In a series
of experiments, Frauenfelder [5] studied the function and dynamics of Myoglobin
(Mb). It is a typical heme protein and its central function is the storage of oxygen
(02)' Actually, Frauenfelder [5] studied the reaction Mb + CO <----> MbCO. A
laser flash breaks the bond between the iron atom in the heme pocket and the
ligand. Then, the ligand molecule moves away from the binding side and later
rebinds. If N (t) denotes the fraction of Mb molecules which have not yet rebound
a CO molecule at time t after photodissociation, then Frauenfelder's findings
are that the rebinding is not exponential in time but is closer to the asymptotic
fractal formula N(t) = N(O)(l + t/T)-a indicating power-law kinetics for the
underlying processes. Clearly, if t / T » 1, N (t) approaches asymptotically inverse
power-law behavior N(t) rv t- a . Theoretical models borrowed from the theory
of polymers, glasses and other disordered systems have already been applied to
protein dynamics.
In the current investigation, however, we will not go into a discussion of
such well-known theoretical concepts. Instead, we alternatively propose a method
based on fractional integral and differential operators that enables us to generalize
the standard relaxation equation
dN(t)/dt = -(l/r)N(t) (12)
with the exponential solution N(t) = N(O)exp(-t/T) to a fractional differential

equation representing slow relaxation phenomena.
Our procedure is as follows : The integrated form of (12) reads N (t) - N (0) =
-(l/T)oDi l N(t), where N(O) is the initial value N(t = 0) and the integral
t
operator is defined by oDi ' N(t) = f N(t') dt' . Introducing now the fractional
o
Liouville-Riemann integral operator oD-;-{3, defined by [21] [27]
-{3
oDt N(t) =
Jt
(t - t'){3-1 , ,
f(,8) N(t )dt ( 13)
o
with 0 < ,8 < 1 and replacing (I/T)oDt1N(t) by (I/T{3)oD-;-{3N(t), the inte-

grated version of (12) takes the form
(14 )
representing a fractional integral equation with an incorporated initial value of

N(O). Eq. (14) can be inverted by introducing the fractional differential operator
oDt'N(t) = (:t) n oD~-nN(t), O<v<n (15)
into the fractional differential equation [7] [17]
(16)
The solution of either (14) or (16) is given by [6] [7]
where E{3(x) is a Mittag-Leffler function.

One recognizes that in the limit ,8 ---. 1, the exponential solution of the
standard relaxation equation (12) is rediscovered. For large t-values, solution (17)
approaches the asymptotic fractal formula [7]
(_I)k (T){3(k+l)
L
00
N(t) rv N(O) f(1 _ ,8(k + 1)) t . (18)

k=O
Just as a byproduct of our fractional analysis, we comment that the leading order
term of (18), i.e. N(t) rv t-{3 exhibits the same inverse power-law exponent ,8
which defines the fractional order of the Liouville-Riemann integral operator (13).
We notice that a re-interpretation of Frauenfelder's Myoglobin data is given
by GlOckle and Nonnenmacher [8] in terms of such a fractional relaxation model.
My final comments and conclusions are: self-similar geometric structures and
scale-invariant dynamical processes in biomedical systems are not fractal on all
length and time scales. Instead, they are fractal only within limits. Estimating the
36 Nonnenmacher
fractal contour dimension D we observed that contour profiles of hairy leukemic

cells have D-values significantly larger than those of healthy T-Iymphocytes. How-
ever, investigating profiles of blast cells of acute lymphoblastic leukemia (not dis-
cussed here) we estimated D-values that are close to but smaller than those of
normal T-Iymphocytes (see Losa, this Volume). To confirm this and to get a more
complete understanding of this small change in smoothing down the structural con-
tour irregularities when T-Iymphocytes are forced to transform into lymphoblasts
requires further investigations and, possibly, more information than being stored
in the fractal contour dimension D. Currently we are searching for a unifying
principle correlating both spatial and temporal fractal properties.
Acknowledgments:
This work has been supported in part by Deutsche Forschungsgemeinschaft (SFB
239, Project C 8) and by the Swiss National Science Foundation, grant 31-
25702.88. I gratefully acknowledge stimulating discussions with Dr. Gerd Bau-
mann, Andreas Barth and Walter Glockle. I also acknowledge helpful comments
from Professors Albrecht Kleinschmidt and Torsten Mattfeldt while preparing this
manuscript.
References
[1] Barth A, Baumann G & Nonnenmacher TF: 1. Phys. A: Math. Gen. 25,381
(1992).
[2] Baumann G, Barth A & Nonnenmacher TF: This volume.
[3] Blatz AL & Magleby KL: 1. Physiol. (London) 378, 141 (1986).
[4] Blumen A, Klafter 1 & Zumofen G: In: Optical spectroscopy of glasses,
Zschokke I, ed. Reidel, Dordrecht (1986).
[5] Frauenfelder H: Ann NY Acad. Sci. 504, 151 (1987).
[6] Glockle WG & Nonnenmacher TF: Macromolecules 24, 6426 (1991).
[7] Glockle WG & Nonnenmacher TF: 1. Stat. Phys. 71, 741 (1993).
[8] Glockle WG & Nonnenmacher TF: Fractional relaxation equations for protein
dynamics. This volume.
[9] Goldberger AL, Bhargava V, West Bl & Mandell AI: Biophys. 1. 48, 525
(1985).
[10] Keough KMW, Hyam P, Pink DA & Quinn B: 1. Microsc. 163,95 (1991).
[11] Lipsitz LA & Goldberger AL: Loss of complexity and aging. lAMA 267,
1806 (1992).
[12] Losa GA, Baumann G & Nonnenmacher TF: Path. Res. and Pract. 188,680
(1992).
[13] Mandelbrot BB: The Fractal Geometry of Nature. Freeman, San Francisco
(1983).
[14] Montroll EW & Shlesinger MF: Proc. Natl. Acad. Sci. USA 79, 3380 (1982).
[15] Nonnenmacher TF: In: Thermodynamics and Pattern Formation in Biology,
Lamprecht I & Zotin, AI, eds. Walter de Gruyter, Berlin, 371-394 (1988).
[16] Nonnenmacher TF: J. Colloid. Polym. Sci. 267, 753 (1989).
[17] Nonnenmacher TF: In: Rheological Modeling: Thermodynamical and Statis-
tical Approaches, Casas-Vazquez J and Jou D, eds. Springer, Berlin (1991).
[18] Nonnenmacher TF: Eur. Biophys. J. 16,375 (1989).
[19] Nonnenmacher TF & Nonnenmacher DJF: Phys. Lett A 140, 323 (1989).
[20] Nonnenmacher TF & Nonnenmacher DJF: In: Stochastic Processes, Physics
and Geometry, Albeverio S, Casati G, Cattaneo U, Merlini D and Moresi R,
eds. World Scientific, Singapore (1990).
[21] Oldham KB, & Spanier J: The fractional calculus. Academic Press, New York
(1974).
[22] Paumgartner D, Losa G & Weibel ER: 1. Microsc. 121,51-63 (1981).
[23] Rigaut JP: J. Microsc. 133,41-54 (1984).
[24] Sakmann B & Neher E: Single channel recording, Plenum, New York (1983).
[25] Shlesinger MF: Ann Rev. Phys. Chern 39, 629 (1988).
[26] Smith TG, Marks WB, Lange GD, Sheriff WH & Neale EA: 1. Neurosc.
Meth. 27, 173-180 (1989).
[27] Srivastava HM & Bushman RG: Theory and applications of convolution in-
tegral equations. Kluver Academic Publishers, Dordrecht (1992).
[28] Weibel ER: Am. 1. Physiol. 261, L361-L369 (1991).
[29] West BJ, Bhargava V & Goldberger AL: 1. Appl. Physiol. 60, 1089 (1986).
Chaos, Noise and Biological Data
Bruce J. West, W. Zhang and H.J. Mackey

Physics Department
University of North Texas
Denton, Texas 76203
1 Introduction
In the recent past there has been a torrent of papers concerned with data process-
ing, see e.g., Casdagli and Eubank [I]. In large part this flood of interest has been
associated with the realization that irregular time series, long thought to contain
signal and noise, may in fact be chaotic. Noise is produced by the coupling of
a signal to an infinite dimensional environment, e.g., the billions of neurons in
the brain producing the erratic EEG signal. Chaos on the other hand is produced
by the intrinsic deterministic dynamics of a low-dimensional nonlinear system.
Difficulties arise when one attempts to discriminate between colored noise and
chaos in erratic time series. The difficulties are associated with how one interprets
the erratic nature of the time series. If it is noise, the random character of the
.data masks the underlying signal. If it is chaos, the statistical fluctuations contain
information and can not be conveniently suppressed as one would noise. How-
ever, since quantities such as the correlation or information dimension of the time
series can not distinguish between the two, the development of new measures are
important and new methods of data processing required.
Herein we are concerned with the construction of measures that can distin-
guish between chaos and noise in time series. Osborne and Provenzale [2] have
shown that obtaining a finite fractal dimension D from the processing of a time se-
ries is not in itself sufficient to infer the existence of a low-dimensional dynamical
system. Arandom fractal., e.g., a Gaussian random process with an inverse power-
law spectrum P(f) ex: l/r, has a fractal dimension given by D = 2/(ex-I) . This
theoretical result has been tested [2,3] using the attractor reconstruction technique
(ART) that is usually applied to chaotic time series, i.e., «random» time series
generated by deterministic nonlinear dynamical systems. Therefore the fractal di-
mension alone can not distinguish between a chaotic signal and colored noise, i.e.,
D is not an unambiguous measure of chaos.
One measure that has been proposed for discriminating between chaos and
noise is the power-law form of the spectrum itself. Frisch and Morf [4] presented
an intuitive argument that systems of coupled nonlinear rate equations that have
strange attractor solutions give rise to chaotic time series whose power spectra de-
cay more rapidly than algebraic, e.g., exponential in frequency at high frequencies,
r
i.e., P (f) ex: e-)..f I where A and ex are positive constants. This argument was
Chaos, Noise and Biological Data 39
10 0
10 ·1
10 -2
.g 10-3
10'"
ILl
10·S
10~
10.7
10-11
101 Ie! 1cf 10· 105
N(number of data)
Fig. 1 The average prediction error is plotted versus the number of data points in the time series for
the logistic map (.) and the Henon map (6). The local approximation of FS was used with a
linear polynomial approximation and a sample size of 100. The solid curve is the theoretical
prediction from (7).
numerically verified by Greenside, et al. [5] using systems with known strange
attractor solutions. Sigeti and Horsthemke [6] give a general proof establishing
the power-law form of the spectrum for stochastic systems driven by additive or
multiplicative noise. No such proof exists for the exponential decay of spectra
for chaotic system although mathematical investigations have been initiated to
establish such a proof, e.g. see Ruelle [7]. Contrary to this expectation inverse
power-law spectra have been found for certain deterministic dynamical systems
that have chaotic solutions, e.g., the driven Duffing oscillator [8] and the driven
Helmholtz oscillator [9] in which low frequency inverse power-law spectra are
observed in both analog simulations and numerical integration of the equations of
motion. Inverse power-law spectra also appear in discrete dynamical systems with
intermittent chaos, i.e., systems that manifest randomly spaced bursts of activity
that interrupt the normal regular behavior [10). Therefore an inverse power-law
spectrum is also not an unambiguous measure of colored noise nor of chaos.
The technique that we adopt herein is based on the deterministic generator
of chaos, which is to say that since chaos is deterministically generated one ought
to be able to predict its evolution. At least for short intervals of time. In fact even
beyond where one can faithfully forecast a chaotic signal, the deviation of the
forecast from the data (error) behaves in a characteristic way. Therefore one can
even use the growth of error in forecasts to determine if a given time series is
chaos or noise.
40 West et al.
10°
10 ·1
10-2
10-3
....
~ 10 ....
10-5
10-5
10.7
0 2.5 5.0 7.5 10.0 12.5
Prediction Time
Fig. 2 The prediction error is graphed as a function of the prediction time T for the logistic equation
using two versions of the local approximation (LA): no updating of neighbors ( ... ) and
neighbors are updated at each time step (-).
In Section 2 we review some of the latest techniques for predicting chaos from
chaos. Herein we recalculate the predictions made by a number of authors and
test how the prediction-error grows with the prediction time and with the number
of data points. These methods are then applied to the problem of discriminating
chaos from colored noise and a number of valid indicators are obtained. Included
in this review are the results from a neural network (NN) prediction of chaos from
chaos using a back propagation technique. The NN method is determined to be
more robust to contamination by noise in its predictions of chaos from chaos than
are polynomial methods.
2 Forecasting Chaos from Chaos

We have indicated that «random» time series may really be produced by determin-
istic equations and may therefore be chaos rather than noise. When this is the case
the system dynamics may be modelled by: (i) finding a state space with minimum
embedding dimension (maximum determinism), and (ii) fitting a nonlinear func-
tion to the map that transforms the present states into future states. In this way a
data stream (time series) can be used to determine not only the number of variables
necessary to model the underlying dynamical system, but it can also provide an
approximate functional form for nonlinear dynamics governing the evolution of
the system [1,11].
/
10-
,', ... ' - /
to'
lO~
............
~
1.. tO J
"
/
.
0
Jj t.· '
" .
"
10·
.'
I
,,
.....
.I
10-
10 ' .0 '.0 ,
Prediction Time
Fig. 3 The prediction error using the local approximation (LA) for logistic data is graphed versus
prediction time for a one-dimensional data vector (solid curve) and a two-dimensional data
vector (dotted curve).
2.1 Scaling Error

The great promise of chaos is that although it puts limits on long term predictions,
it implies predictability over the short term. How well this can be done in a
biological context remains to be seen.
As Farmer and Sidorowich [12,13] point out, most forecasting is currently
done with linear methods, and since linear dynamics cannot produce chaos, these
methods cannot produce good forecasts for chaotic time series. Although the de-
velopment of nonlinear forecasting techniques is an active area of research, with
few exceptions, the full exploitation of the concept of chaos has not as yet been
made. Preliminary investigations have been done to compare the quality of fore-
casts using linear and nonlinear techniques and we discuss some of these here.
Finding the state space representation is accomplished using the attractor
reconstruction methods discussed by many researchers. To recapitulate let us con-
sider a discrete time series [~(tj)], j = 0, 1, ... ,M where ~(t) is the observable
in a biological system, e.g., an EEG, ECG, blood pressure, etc. We assume that
[~(tj)] is stationary, as it would be if the data points are generated by a mapping
on an attractor. We construct a state vector X(t) = [X I , (t), X2 (t), ... , Xm (t)] by
assigning the coordinates
Xl (t)= ~(t)
X2(t) = ~(t + 7)
(1)
Xm(t) = W + (m - 1)7]
where 7 is a delay time. If the dynamics take place on an attractor of dimensions d,
then a necessary condition for determinism is m ~ d. If D is the dimension of the
..
42 West et al.
, ,------------------------------------------------,
.... ........ .'

.... ..' .. ' ." '
.' .... .'
.•.. .... .' .' .'
.
.'
"
"
..........
to· ••••
fl)"
~.' .• ' , . '
-~ o :I
----~-----------------------~
10
------ 12
Prediction Time
Fig. 4 The prediction error using the local approximation (LA) for logistic data is plotted against
the prediction time using two nearest updated neighbor (solid curve) and using four updated
nearest neighbors (dotted curve).
manifold containing the attractor, Takens [14] showed rigorously that m = 2D + 1

is sufficient, in principle, to determine the dimension of the attractor. Note the
ambiguity in determining the delay time used in (1).
Once the phase space representation is found, we construct a model to fit the
data. We assume time to be discrete and express the dynamics by the map
X(t + T) = h[x(t)] (2)
for the present state X(t) and future state X(t + T). Here both and X are h
m-dimensional vectors and h
is the nonlinear map that propagates X T -units
forward in time. To estimate X(t + T), the problem of interest, we approximate
the dynamics of the map by h
)((t, T) = lr[X(t)] (3)
where the carets denote approximants. The accuracy of the forecast is measured
here using
E = I)((t, T) - )((t + T)I (4)
Ux
where (E) = 0 for perfect forecasts and (E) = 1 for forecasts that are no better
than using )((t, T) = (X(t)). Note that (4) measures how well lr approximates
Ir. This error estimate requires a data sample X(t) and its prediction. This error is
..•0'
'
........,......
10"
,.'.'
.0"
.0" i
.o~ ! Ii
./
'U;'
'0"
S....
I
'0" , i
10"
I
~ .0"
.0"
1
I
J
10'"
'0'"
10.1•
'0'"
10 '"
...
10.1'
10 '00 '5O
Time
Fig. 5 The prediction error is graphed as a function of prediction time for single precision (dotted
curve) and double precision (solid curve) for logistic data to determine the effect of rounding
error.
defined relative to the standard deviation of the entire data stream IJx. The average
of E over a sample X(t) approximates (E) in (4).
Of course to construct the nonlinear functions 1r
one must choose a represen-
tation since there are an infinite number of ways to select the nonlinear mapping
function. As Farmer and Sidorowich point out, in the absence of any theoretical
understanding, finding a good representation is largely a matter of trial and error.
They themselves use a polynomial for the nonlinear map. Due to the determinis-
tic nature of chaos, locally a smooth map can be approximated by a polynomial
through a Taylor expansion. The parameters in such a representation can be lin-
early fit to minimize the least square sequence deviations, i.e., yields a linear least
square problem. Nearly 50 years ago Wiener [15] suggested the use of polynomi-
als for forecasting residuals in random time series, i.e., the differences between
the data and the prediction X(t) - )(t) referred to as the moving average (MA)
model. The most general form of the nth degree m-dimensional polynomial is
. xhxh
aJI··,J2 ,···,Jm x jm
I 2··· m' (5)
h=O, .. . ,jm=O
where h + h + ... + jm :::; n. The number of parameters ah,h, ... ,jm is
(m+n)! • m n
-'-----'-.
m!n! ~
~
. (6)
Fitting these many parameters with a data set rapidly becomes impractical when
m and n are large, but is convenient when they are not.
44 West et ai,
...
..
\
• 0 -
o. .
c
0
'.:;l o. .
:
~
0)
t:
0 o.
U
0'
."
•
--- so
' '" 'so 200
Time
Fig. 6 The correlation of the generated logistic data as a function of the prediciton time is graphed
for both the single precision (dotted curve) and double precision (solid curve) data.
A well-behaved function lr
can be easily modelled using any of a number
of representations provided the data stream is sufficiently long. For a complicated
function, one having many variations in a short time interval, it is unclear that any
representation can provide an adequate approximation to The dependence on lr.
representation may be reduced by means of a local approximation in which the do-
main is segmented into local neighbourhoods and the parameters, in a polynomial
representation say, are fit in each neighborhood separately. Local approximations
are found by Farmer and Sidorowich to produce better fits for a given number
of data points than global approximations, particularly for larger data sets. This
result has been verified by Casdagli [16].
Here we applied the local approximation (LA) developed by Farmer and
Sidorowich (FS) [12] to a number of mappings. As shown by FS the scaling
behavior of the prediction error (E) with the number of data points shown in Fig.
1 is predictable. Consider an nth order polynomial approximant. Ideally the error
lr
then depends on the (n+ 1)st derivative of Ux), so that. if E is the spacing between
data points, the error is proportional to En+l. If the data points are uniformly
distributed throughout the volume then E N- 1/ d as they do for a fractal where
r-.J
d is the fractal dimension of the space and N is the number of data points. Thus,
the rms prediction error grows as
(E) r-.J N-q/d (7)
where q = n+ 1. In Fig. II0g(E) is graphed against 10gN, so that the slope of the
curve should be - (n + 1) / d. It is found that this is indeed the case. Thus the scaling
of the prediction error in itself provides a technique for determining whether a
Chaos. Noise and Biological Data 45
10·
to"
10"
10"
10~
~
S. 10"
... •... -_....

10"
~
10 . 7
....
10"
.. . .. . ... -.--.
10"
to-tO
..... .... .. . ,
...•.•.....-. .#
10'"
~O · tt
10'"
10·'4
10 · t5
10 1$ ao
Prediction Time
Fig. 7 The prediction error is graphed as a function of prediction time using the local approximation
with updated neighbors for a linear polynomial (solid curve), a quadratic polynomial (dotted
curve) and a cubic polynomial (dot-dash curve) for logistic data.
given time series is noise or chaos. Again this scaling result was independently
verified [16].
There is a second important dependence of the prediction error on the fore-
casting parameters, this is the prediction time T . The rate at which the errors grow
depends on the manner in which we make predictions; either directly or itera-
tively. A direct forecast is made by fitting a new model for each individual T . An
iterative one is made by fitting a model to T = 1 and iterating to make predictions
for T = 2, 3.... Intuitively one would guess that the direct model would be more
accurate , since each model is «tailored» for the time it is supposed to predict, and
there is no accumulation of errors due to iterations. FS show that this is in fact
not always the case. They estimate that for direct forecasting (E) rv N-q /deqT>.max
whereas for iterative forecasting (E) rv N - q/deTJ.. max where Amax is the largest
Lyapunov exponent. Thus the exponential growth of errors is a factor q larger in
the former case over the latter. They argue that the superiority of iterative esti-
mates comes from the fact that this scheme makes use of the regular structure of
the higher iterates. In addition, they also give a numerical example showing the
superiority of the direct over the iterative forecasting.
2.2 Local Approximation Results

We implement the local approximation (LA) technique and forecast the data se-
quence generated by the logistic map:
(8)
46 West et al.
10·'
'0"
'0" 1--------~-----10~-----'~6- - - -. --1
Prediction Time
Fig. 8 The prediction ·error is graphed as a function of prediction time for the Henon map. The
forecast uses the local approximation with updated neighbors with a linear polynomial in
two dimensions (solid curve), three dimensions (dotted curve) and four dimensions (dot-dash
curve).
We wish to examine a number of properties of the forecasting schemes using the

logistic map. Firstly, we check the prediction (8) and verify in Figure 2 that the
error does grow exponentially with time, when a linear LA is used. We implement
two ways of forecasting the data in time (here time is given by the discrete integers
of the iterations). The first procedure for predicting all Xn+l, X n+2, ... , etc., is
based on the neighbors of X n . The second method is to predict Xn+ j based on
the neighbors of Xn+j-I, i.e., we update the neighbors along with the prediction.
We call the latter the LA with updated neighbors (LAU). Our numerical results
shown in Figure 2 indicate that updating neighbors reduces the prediction error
significantly. This is indicated by the reduction in the slope of the prediction error
curve.
In Figure 3 the error for LA is again determined as a function of prediction
time T, and the growth of error for two different dimensions for the reconstructed
data vector is compared. Here it is clear that increasing the dimension from one to
two increases the rate at which the error grows. If the forecast is updated at each
time step the dependence on the number of nearest neighbors used in update does
not appear to have a strong effect on the rate of error growth [cf. Figure 4]. There
is also a significant dependence of the error growth rate on the precision of the
calculation. In Figure 5 we compare the effects of rounding errors in generating the
data from the map using both single and double precision. This gives quantitatively
the typical length of the prediction time of the logistic map. It is apparent that
double precision increases the predictable time by nearly an order of magnitude.
However, the error is not the only measure available to assess the quality of the
•••
. .
•0' .
.
...
0
';i;'
10· t
..
.
..
! •0"
J 10"'
••
•
•• ,
#
.o~
l
,
.0" ,.
.... '0 .
Prediction Time
Fig. 9 The prediction error is graphed as a function of prediction time for the logistic map. The
forecast uses a local approximation with updated neighbors with a linear polynomial in one
dimension (solid curve), two dimensions (dot-dash curve) and three dimensions (dotted curve).
forecast. The correlation between the data and the forecast is also a good measure:
(9)
This procedure is equivalent to an auto-regressive (AR) model which mimics the

original time series in terms of its mean, variance, and the autocorrelation function
for multiple time delays.
In Figure 6 the correlation function plotted as a function of prediction time for
both the single precision and double precision data. It is clear that the suppression
of numerical error increases the prediction time by nearly a factor of two. The
precipitous drop in the correlation function after a prolonged interval of nearly
perfect correlation (p = 1) is characteristic of chaotic time series. This observation
was also made by Shugihara and May [17], who also point out that the correlation
for white noise is flat as a function of prediction time.
Here again we use the logistic map as the generator of our «experimental data»
and determine the effects of modeling the forecasting function with polynomials
of various degrees. In Figure 7 we show the result of using a linear, a quadratic
and a cubic polynomial for l(Xn ). It is clear that the latter two polynomials do
about 106 better in error level than does the linear forecast. Further the quadratic
function is 102 better than the cubic, as it should be since the logistic map is a
quadratic polynomial. Indeed we observe the error in quadratic forecasting results
is of the same order as that of iterating the logistic map with double precision
48 West et al.
1.1
1.0
0.'
0.0
.g 0.1
'"t::
'0 0.0
0
u 0.0
0.'
.•
0.'
0.2 ,
.,
..
0.1 ...
• 0
--~--~--------~-------
10 1$ 20
Prediction Time
Fig. 10 The correlation between the forecast and the Henon data as a function of the prediction time is
graphed for the reconstructed data vector in two dimensions (solid curve), three dimensions
(dotted curve) and four dimensions (dot-dash curve). A local approximation with updated
neighbors is used for the prediction.
shown in Figure 5. Thus we should be able to reproduce these data exactly with
the quadratic forecast function.
Let us now consider a data stream generated by the Henon map
Yn+1 = 1- ain + bYn-l (10)
with the parameter values a = 1.4 and b = 0.3. Note that the Henon map can
be written either as a two-dimensional map with the mapping function relating
(Yn, xn) to (Yn+l, xn+d or as the one-dimensional map (10) with Yn+l related to
the two values Yn and Yn-l.
In Figure 8 we depict the growth of error for reconstructed Henon data vectors
of two, three and four dimensions. Here we use a linear forecast with the minimum
number of neighbors. It is clear from this figure that a linear forecast with updates
at each prediction time gives the best results for two dimension, compared with
the error for a higher dimension forecast which is greater than unity. A similar
behavior is observed in Figure 9 where the error growth for the logistic data vectors
of one, two and three dimensions. Here it is clear that the higher dimensions
eventually yield arbitrarily large errors, whereas the error for the one-dimensional
vector stays below unity. The correlation function for the predictions using the
Henon data shown in Figure 10 has the sudden decrease observed earlier for the
logistic data. Here the two-dimensional forecast remains above .75 for 20 time
steps into the future, whereas the three and four dimensional forecasts both faU
. '
..'
10.1
'0'
'0'
--..
e'" .. -
.t:- .0'
e;
t::
Ul
.'
10.7
.0'
'0' ..
10'"
10· n ..
10.11
10'"
. . _M
to ••
Prediction Time
.. to ..
Fig. l1 The root mean square prediction error is plotted as a function of prediction time for logistic
data. The most significant curves are those using the NN ( x ), the LA for a quadratic function
using the nonlinear least squares (- /':,. -) and the LA for a quadratic function using linear
least square (- - - - - -). The upper cluster of curves are for various combinations of linear
and nonlinear solvers using the local approximation.
sharply before T = 9, thereby indicating the two-dimensional nature of the Henon

map.
2.3 Neural Network Results

In this section we compare the results obtained by means of determining the
forecasting function using a neural net [18,19] and that using the polynomial
techniques of the preceding section. It is useful to bear in mind that the neural
network technique always involves a least squares determination of parameters in
the numerical algorithm. Although it is true that a linear least square problem can
be solved exactly to within machine precision, a neural net is a nonlinear least
squares problem which can not be solved exactly in general. In Figure 11 the
rms prediction error is compared between forecasts of logistics data using LA and
those using a neural net (NN). We see that all the forecasts using either linear
or nonlinear least square solvers for a linear forecast are about equally bad and
substantially worse than that produced by the NN. However, when the quadratic
approximation is implemented in the forecast function we see that linear least
square solver does substantially better than either the NN or the nonlinear least
square solver, the latter two being about the same with NN or a little worse.
A similar result is obtained for the Henon data as we show in Figure 12. Here
the rrns error is seen to be significantly reduced from the LA linear polynomial
50 West et al.
,.'
10'
'0"
10"
'U;' 10"
E-... 10·
10'
~ 10'
to"
10"
..............................................................~•..........•..
............
.._.........................
10·
10'" ..........
10·1t -......~........-.........................
10· t• •
10
"
Prediction Time
. " It zo
Fig. 12 The root mean square prediction error is plotted as a function of the prediction time for data
from the Henon map using LA and a linear polynomial (-), LA and a quadratic polynomial
( . . .), NN with a tolerance in the nonlinear solver 10- 3 (- X -) and NN with a tolerance
10- 6 in the nonlinear solver ( . . . x ... ).
approximation to the NN result, and reduced again from the NN result to the LA
with a quadratic polynomial. The two neural net curves result from increasing the
required tolerance from 10- 3 to 10-6 in the NN training algorithm, i.e. nonlin-
ear least square solver. This decreases the rms error by an order of magnitude,
however the NN error remains six orders of magnitude above that of the LA
quadratic polynomial. This might suggest that the NN does not work as well as
the polynomial techniques. However such a conclusion would be premature since
we have only tested its utility on polynomial maps and we must broaden the base
of application before deciding.
Another way of determining the utility of a given method is its robustness
in the presence of noise. The question is whether the NN techniques are more
sensitive to noise than are the LA procedure. The figure of merit we use here is
the signal-to-noise ratio (SNR) which is twenty times the logarithm to the base ten
of the ratio of rms signal strength to the rms noise strength i.e., 10 10glO( a; / a;")
where a; and a;" are the variances of the signal and noise, respectively. In Fig-
ure 13 we show the average relative error on an ensemble of ten members for
the logistic data with zero-centered Gaussian white noise added of the indicated
strength. In this figure the forecasts are made with the LA method, whereas in
Figure 14 the same data is forecast using the NN. It is clear that both procedures
are equally poor when the noise has the same rms strength as the signal. How-
ever not only is the rate of growth of the average relative error faster for the LA
method than for the neural net, but the overall level of the error is two orders of
10°
.•
......
&i 10 "
~
..
p
-' ,
10.2 ,
~ ,
,
.~
,."
10 " , "
~
10"
10"
j
0
T
2 4
.
6 8 10 12
-r
14
-.-
16 18 20
Time
Fig. 13 The average relative prediction error is plotted as a function of prediction time for the LA
forecasts of logistic data with Gaussian white noise. The SNR is: 90 (-), 80 (- - -),
70 (- . -); 50 ( ... ); 40 ( . . . x .. .) and 0 (. - . - .). Here we need 10 members in the
ensemble each of which had 103 data points in the time series.
10'
10°
10 "
....... , - .. ' , ,
10 2
... ,
, ,
~
10 .3
' ,
V
~
-
v'
10'
10'S
10"
10 .7
0
.
2 4
. . 6 B
T
10
. . . .
12 14 16 18 20
Time
Fig. 14 Same data as for Figure 13 but using the NN forecasts and SNR's: 85 (- - -); 80 (- - -); 65
(_. -); 60 (- - .), 45 ( ... ) and 0 (_. -).
magnitude smaller in the latter than in the former for high noise. From this and
other calculations we conclude that the NN is less sensitive to the noise than are
the polynomial forecasting techniques.
The correlation dimension of these two dynamic processes can be obtained
using the Grassberger-Procaccia method. In Figure 15a we plot the slope of the
52 West et at.
10
---
1:'
'-'
U
6
'-' 5
.5
'() ..
§. 3
•
:
....
(i) 2 ~
.& .& ~
0
·10 ·9 ·7 ·5 ·3 ·2 ·1 o
In(r)
10
'""'
1:'
'-'
6
U 5
'-'
.5
.....
0
.
§. 3
(i) 2
o
In(r)
Fig. 15 The slope of the logarithm of the correlation function [In C(r)] is plotted versus the logarithm
of the separation between trajectories on the attractor [In r] . (a) logistic data with Gaussian
white noise superimposed having an SNR of 30 dB is given by the crosses; the noise· free
logistic data is given by the filled circles and the noise data processed through a neural net
is given by the triangles. (b) The same as (a) using a Henon data stream.
logarithm of the OP-correlation function versus the logarithm of the separation (r)
between trajectories on the reconstructed attractor. The flat region - 5 :S In r :S -1
indicates the correlation dimension of the noise-free logistic data. The dimension
obtained from the logistic data plus noise with a 30 dB SNR using the neural net
coincides with the noise free case in this region of r. It is clear that one can not
obtain the dimension of the time series directly from the data. This result carries
down to a SNR of 10 dB without change. A SNR below 10 dB begins to restrict
the region of lor over which the NN yields the proper correlation dimension.
We again follow the above procedure using a data stream generated by the
Henon map. In Figure 15b the slope ofln C( r) is plotted versus lor and we again
see a relatively flat region for -5 :::; In r :::; -I for the noise-free data. The noise
contaminated signal again does not have a well defined dimension and the signal
plus noise as processed by the NN yields a correlation dimension slightly above
that obtained from the noise-free data. These results agree with the more detailed
analyses of Albaro et al. [19]
Acknowledgement:
The support of this work by grant number N6-2269-92-C-0548 from the Naval
Air Warfare Center is gratefully acknowledged.
References
[1] M. Casdagli and S. Eubank, editors, Nonlinear Modeling and Forecasting,
Addison-Wesley, Redwood City, CA (1992).
[2] AR. Osborne and A Provenzale, Physica 35D, 357-381 (1989).
[3] B.1. West and H.J. Mackey, J. Appl. Phys. 69(9), 6747-6749 (1991).
[4] U. Frisch and R. Morf, Phys. Rev. 23A, 2673 (1981).
[5] H.S. Greenside, G. Ahlers, P.e. Hohenberg and R.w. Walden, Physica 5D,
322 (1982).
[6] D. Sigeti and W. Horsthemke, Phys. Rev. 35A, 2276 (1987).
[7] D. Ruelle, Phys. Rev. Lett. 56, 405 (1986).
[8] F.T. Arecchi and F. Lisi, Phys. Rev. Lett 49, 94 (1982).
[9] M.A Rubio, M. De La Torre and J.e. Antoranz, Physica 36D, 92 (1989).
[10] H. Frijisaka and T. Yamada, Prog. Theor, Phys. 74, 918 (1985).
[11] B.1. West, H.1. Mackey and D. Chen, «Methods for distinguishing chaos from
colored noise», in Patterns, Information and Chaos in Neuronal Systems, ed.
B.J. West, World Scientific, Singapore (1993).
[12] J.D. Farmer and J.J. Sidorowich, Phys. Rev. Lett. 59, 845 (1987).
[13] J.D. Farmer and J.J. Sidorowich, Los Alamos preprint (1988).
[14] F. Takens, in Lecture Notes in Mathematics 898, eds. D.A. Rand and L.S.
Young, 366-381 (1981).
[15] N. Wiener, Time Series, MIT Press (1949).
[16] M. Casdagli, Physica 35D, 335 (1989).
54 West et al.
[17] G. Sugihara and R.M. May, «Nonlinear Forecasting as a way of distinguishing

chaos from measurement error in time series,» Nature 344, 734 (1990).
[18] AS. Lapedes and R. Farber, Los Alamos Technical Report LA-UR-87-2662.
[19] AM. Albano, A Passamente, T. Hediger and M.E. Farrell, «Using neural
nets to look for chaos,» Physica D 58, 1-9 (1992).
Fractal Landscapes in Physiology &
Medicine: Long-Range Correlations in DNA
Sequences and Heart Rate Intervals
C.-K. Peng l ),2), S.V. Buldyrev 2), J.M. Hausdorffl),
S. Havlin2),3), J.E. Mietus1), M. Simons1),4),
H.E. Stanley2) and A.L. Goldberger l )
I) Cardiovascular Division, Harvard Medical School, Beth Israel Hospital. Boston, MA 02215.
USA
2) Center for Polymer Studies, Boston University, Boston, MA 02215, USA
3) Department of Physics, Bar-Han University, Ramat-Gan 52100, Israel
4) Biology Department, MIT, Cambridge. MA 02139, USA
Abstract. Healthy systems in physiology and medicine are remarkable for their structural variabil-
ity and dynamical complexity. The concept of fractal growth and form offers novel approaches to
understanding morphogenesis and function from the level of the gene to the organism. For example,
scale-invariance and long-range power-law correlations are features of non-coding DNA sequences as
well as of healthy heartbeat dynamics. For cardiac regulation, perturbation of the control mechanisms
by disease or ageing may lead to a breakdown of these long-range correlations that normally extend
over thousands of heartbeats. Quantification of such scaling alterations are providing new approaches
to problems ranging from molecular evolution to monitoring patients at high risk of sudden death.
We briefly review recent work from our laboratory concerning the application of fractals to two
apparently unrelated problems: DNA organization and beat-to-beat heart rate variability. We show how
the measurement of long-range power-law correlations may provide new understanding of nucleotide
organization as well as of the complex fluctuations of the heartbeat under normal and pathologic
conditions.
1 Long-Range Correlations in Nucleotide Sequences

Genomic sequences contain numerous «layers» of information. While the means
of encoding some of these instructions is understood (for example, the codes
directing amino acid assembly and intron/exon splicing, etc.), relatively little is
known about other kinds of information encrypted in the DNA molecule. In higher
eukaryotic organisms, only a small portion of the total genome length is actually
used for protein coding. The role of introns and the intergenomic sequences that
constitute a large portion of these DNA polymers remains unknown.
Recently we [1] proposed a novel method for studying the global organiza-
tional properties of genomic sequences by constructing a 1: I map of the sequence
onto a «DNA walk». Consider a one-dimensional walker [2] dictated by the se-
quential order of nucleotides. The walker steps up [u(i) = +1] if a pyrimidine
occurs at position a linear distance i along the DNA chain, while the walker steps
down [u(i) = -1] if a purine occurs at position i. The question we ask is whether
such a walk displays only short-range correlations (as in an n-step Markov chain)
56 Peng et al.
[3] or long-range correlations (as in critical phenomena and other scale-free «frac-
tal» phenomena).
This DNA walk provides a novel graphical representation for each DNA
sequence and permits the degree of correlation in the nucleotide sequence to be
directly visualized (Fig. 1). A useful quantity that measures the degree of the
correlation is obtained by calculating the «net displacement» y( n) of the walker
after n steps, which is the sum of the unit steps u(i) for each step i,
n
y(n) == L u(i). (1)
i=l
A useful statistical description of any «landscape» can then be derived by consid-

ering a sliding window of size £ through the landscape and measuring the change
of the «altitude» across this window, i.e.,
t:.ye(n) = y(n + £) - y(n), (2)
where n indicates the starting position of the window. We define the fluctuation
measurement, F(£), as the standard deviation of the quantity t:.Ye.
The calculation of F (£) can distinguish three possible types of behavior. (i) If
the nucleotide sequence were random, then the landscape has the same statistical
properties as that generated by a normal random walk, i.e., F(£) '"" £1 /2 (ii) If
there were a local correlation extending up to a characteristic range (such as in
Markov chains), then the behavior F(£) '"" £1/2 would be unchanged from the
purely random case (for £ » 1). (iii) If there is no characteristic length (i.e., if the
correlation is «infinite-range»), then the fluctuations will be described by a power
law
(3)
with 0: =F 1/2 [4] . If 0: > 1/2, it indicates persistent correlation, i.e., one type of
nucleotide (purine or pyrimidine) is likely to be close to another of the same type.
In contrast, 0: < 1/2 indicates that the nucleotides are organized such that purines
and pyrimidines are more likely to alternate (<<anti-correlation») [5].
The power-law form of Eq. (3) implies a self-affine (fractal) property in the
DNA walk landscape. To visualize this finding, one can magnify a segment of
the DNA walk to see if it resembles (in a statistical sense) the overall pattern.
Fig. I(a) shows the DNA walk representation of a gene and Fig. I(b) shows a
magnification of the central portion. Fig. 1(c) is the further magnification of a
sub-region of Fig. I b. Note the similar fluctuation behavior on all three different
length scales.
We calculate 0: from the slope of double logarithmic plots of the mean square
fluctuation F (£) versus £ (Fig. 2). Measurement of this exponent for a broad range
of representative genomic and cDNA sequences across the phylogenetic spec-
trum reveals that long-range correlations (0: > 1/2) are characteristic of intron-
containing genes and non-transcribed genomic regulatory elements [1,6,7]. The
Fractal Landscapes in Physiology & Medicine 57
SELF-SIMILARITY OF DNA WALKS
10200 10400 10600 10800

nucleotide position, n
Fig. 1 The DNA walk representation for the rat embryonic skeletal myosin heavy chain gene. (a)
The entire sequence. (b) The magnification of the solid box in (a). (c) The magnification of the
solid box in (b). The statistical self-affinity of these plots is consistent with the existence of a
scale-free or fractal phenomenon termed a fractal landscape. In order to observe statistically
similar fluctuations within successive enlargement, the magnification factor along the vertical
direction (M.Ll and horizontal direction (Mil) follows a simple relation: 10gM~/ log Mil = Q.
Here Q = 0.63. Note that these DNA walk representations are plotted so that the end point
has the same vertical displacement as the starting point.
finding of long-range correlations in intron-containing genes appears to be inde-

pendent of the particular gene or the encoded protein - it is observed in genomic
sequences as disparate as myosin heavy chain, beta globin, adenovirus and yeast
chromosome III [1,8].
In contrast, for cDNA sequences (i.e., the spliced together coding sequences)
and genes without introns, we find that a ~ 1/ 2, indicating no long-range cor-
relation (Fig. 2). In fact, the lack of long-range correlations in coding regions
is not very surprising considering that an uncorrelated sequence can carry more
information than a correlated sequence [1]. On the other hand, the existence of
long-range correlations in the non-coding regions is paradoxical and suggests a
new organizational role for so-called «junk DNA». Ongoing investigations are
directed at studying the implications of these correlations for DNA structure and
58 Peng et af.
2.0
o Rat embryonic skeletal myosin (0.=.63)
1 .5 '" cDNA (0.=.5)
:::::--
Lt-
~ 1.0
OJ
.3
0.5
0.0
0.0 1.0 2.0 3.0
Fig. 2 Double logarithmic plot of F(f) versus f for rat embryonic skeletal myosin heavy chain gene
shown in Fig. 1(75% non-coding regions) and its cDNA. Note that the slope (0 = 0.63) for
the intron-containing sequence is > 1/ 2 indicating the presence of long-range correlations. In
contrast, the slope is 0.5 for the coding sequence (cDNA) indicating the absence of long-range
correlations.
function, as well as for molecular evolution. Since power-law behavior repre-

sents a scale-invariant (fractal) property of DNA, it cannot be attributed simply
to the occurrence of nucleotide periodicities such as those associated with nucleo-
some packaging. Whether these long-range correlations are related to higher order
DNNchromatin structure or to DNA bending and looping remains speculative.
A complementary approach to interpreting this correlation behavior is to relate
it to the dynamic processes that modify nucleotide sequences over time. Buldyrev
et af. [9] recently proposed a generalized Levy walk model to account for the
genesis of these correlations, as well as a plausible evolutionary mechanism based
on nucleotide insertion and deletion [10].
From a practical viewpoint, the calculation of F (£) for the DNA walk repre-
sentation provides a new, quantitative method to distinguish genes with multiple
introns from intron-Iess genes and cDNAs based solely on their statistical prop-
erties. The fundamental difference in correlation properties between coding and
non-coding sequences also suggests a new approach to rapidly screening long
DNA sequences for the identification of introns and exons [11].
2 Long-Range Correlations in Heart Beat Intervals

The healthy heartbeat is generally thought to be regulated according to the classical
principle of homeostasis whereby physiologic systems operate to reduce variability
and achieve an equilibrium-like state [12]. However, our recent findings [13] indi-
cate that under normal conditions, beat-to-beat fluctuations in heart rate display the
kind of long-range correlations typically exhibited by .dynamical systems far from
equilibrium. Since the heartbeat is under neuroautonornic control, our findings also
imply that this feedback system is operating in a non-equilibrium state. Our results
demonstrate that such power-law correlations extend over thousands of heart beats
in healthy subjects. In contrast, heart rate time series from patients with severe
congestive heart failure show a breakdown of this long-range correlation behavior,
with the emergence of a characteristic short-range time scale. Similar alterations
in correlation behavior may be important in modeling the transition from health
to disease in a wide variety of pathologic condition~.
Clinicians traditionally describe the normal activity of the heart as «regular
sinus rhythm». But in fact, rather than being metronomically regular, cardiac inter-
beat intervals normally fluctuate in a complex, unpredictable manner. Much of the
analysis of heart rate variability has focused on short-term oscillations associated
with respiration (0.15-0.40 Hz) and blood pressure control (0.01-0.15 Hz). Fourier
analysis of lengthy heart rate data sets from healthy individuals typically reveals
a 1/[-like spectrum for lower frequencies « 0.01 Hz), and some alterations in
spectral features have been reported with a variety of pathologies. However, the
long-range correlation properties of physiologic and pathologic heart rate time
series had not been systematically described.
The mechanism underlying complex heart rate variability is related to com-
peting neuroautonomic inputs. Parasympathetic stimulation decreases the firing
rate of pacemaker cells in the heart's sinus node. Sympathetic stimulation has
the opposite effect. The nonlinear interaction between these two branches of the
nervous system is the postulated mechanism for the type of erratic heart rate
variability recorded in healthy subjects (even during resting or sleeping hours),
although non-autonomic factors may also be important.
Our analysis is based on the digitized electrocardiograms of beat-to-beat heart
rate fluctuations over long time intervals (up to 24 h ~ 105 beats) recorded with an
ambulatory (Holter) monitor. The time series obtained by plotting the sequential
intervals between beat n and beat n + 1, denoted by B(n), typically reveals a
complicated type of variability. To quantitatively study these dynamics over large
time scales, we pass the time series through a digital filter that removes fluctuations
offrequencies > 0.005 beac 1 , and plot the result, denoted by Bdn), in Fig. 3. We
observe a more complex pattern of fluctuations for a representative healthy adult
(Fig. 3a) compared to the pattern of interbeat intervals for a subject with severe
heart disease associated with congestive heart failure (Fig. 3b). These heartbeat
time series produce a contour reminiscent of the irregular «landscapes» of DNA
walks (Fig. 1).
To apply the previous fractal landscape analysis, we can make a simple map-
ping such that Bdn) for the heartbeat is equivalent to yen) for DNA. Thus we
can measure the heartbeat fluctuation F (f) the same way as in DNA walk, where
f indicates the size of the observational window (number of beats). Figure 4 is a
log-log plot of F(f) vs f for the data in Figs. 3. This plot is approximately linear
over a broad physiologically-relevant time scale (f '" 200 to 4000 beats) implying
that F (f) rv fO:.
We find that the scaling exponent a is markedly different for the healthy and
diseased states: for the healthy heartbeat data, a is close to 0, while a is close to
60 Peng et al.
CARDIAC INTERBEAT INTERVAL LANDSCAPES
0.9
a Normal
U 0.8
Q)
.!!2.
0.7
c:
co...J 0.6
0.5
0 20000 40000 60000
Beat Number
0.9
U 0.8
Q)
.!!2.
0.7
e-
co...J 0.6
0.5
0 20000 40000 60000
Beat Number
Fig. 3 The interbeat interval Bdn) after low-pass filtering for (a) a healthy subject and (b) a patient
with severe cardiac disease (dilated cardiomyopathy). The healthy heartbeat time series shows
more complex fluctuations compared to the diseased heart rate fluctuation pattern that is close
to random walk «<brown») noise. The low-pass filter removes all Fourier components for
I 2: fc. The results shown here correspond to Ie = 0 .005 beat- I. but similar findings are
obtained for other choices of Ie ::; 0.005. This cut-off frequency Ie is selected to remove
components of heart rate variability associated with physiologic respiration or pathologic
Cheyne-Stokes breathing as well as oscillations associated with baroreflex activation (Mayer
waves). After Ref. 13.
0.5 for the diseased case in this example. As we discussed previously, 0: = 0.5
corresponds to a random walk (Brownian motion). Thus the low-frequency heart-
beat fluctuations for the diseased state can be interpreted as a stochastic process,
in which case the interbeat increments I(n) == B(n + 1) - B(n) (corresponding to
u(n) in the DNA case) are uncorrelated for £ > 200. For the healthy subject, the
interbeat increments are anti-correlated (0: < 0.5).
To study further the correlation properties of the time series, we choose
to study I(n). Since I(n) is stationary, we can apply standard spectral analysis
techniques [10]. Figures 5a and 5b show the power spectra 5I(f), the square of
the Fourier transform amplitudes for I(n), derived from the same data sets (without
filtering) used in Fig. 3. The fact that the log-log plot of 5I(f) vs f is linear implies
(6)
The exponent f3 is related to 0: by f3 = 1 - 20: [5].
a=O Normal
o
a=O.S
ol Heart disease
o
....J
103
L0910 I [beat]
Fig.4 Double logarithmic plot of F(t') vs n. The circles represent F(t') calculated from data in
Fig. 3(a) and the triangles from data in Fig. 3(b). The two best-fit lines have slopes Q = 0.07
and Q = 0.49 (fit from 200 to 4000 beats). The two lines with slopes Q = 0 and Q = 0.5
correspond to <<l l f noise» and «brown noise,» respectively. We observe that F(t') saturates
for large t' (of the order of 5000 beats), because the heartbeat interval are subjected to
physiological constraints that cannot be arbitrarily large or small.
10.8
a Normal
10.9
-
uf
10. 10
10.11
10.4 10.3
f (beat'1)
10.8
b Heart disease
10.9
-
Clf
10'10
~=O +
10'11
10'4 10.3
f (beat'1)
Fig. 5 The power spectra 51 (f) for the interbeat interval increment sequences over ~ 24 hours
for the same subjects in Fig. 3. (a) Data from a healthy adult. The best-fit line for the low
frequency region has a slope f3 = 0.93. The heart rate spectrum is plotted as a function
of «inverse beat number» (beaC I ) rather than frequency (time- I) to obviate the need to
interpolate data points. The spectral data are smoothed by averaging over 50 values. (b)
Data from a patient with severe heart failure. The best-fit line has slope 0.14 for the low
frequency region, f < fe = 0.005 beat- I. The appearance of a pathologic, characteristic
time scale is associated with a spectral peak (arrow) at about 10- 2 beat- I (corresponding to
Cheyne-Stokes respiration). After Ref. 13.
62 Peng et al.
For the data set from the patient with severe heart disease, we observe a
flat spectrum ({3 ~ 0) in the low frequency region (Fig. 5b) confirming that I (n)
are not correlated over long time scales (low frequencies). Therefore, I(n), the
first derivative of B(n), can be interpreted as being analogous to the velocity
of a random walker, which is uncorrelated on long time scales, while B (n) -
corresponding to the position of the random walker - are correlated. However,
this correlation is of a trivial nature since it is simply due to the summation of
uncorrelated random variables.
In contrast, for the data set from the healthy subject (Fig. 5a), we obtain
(3 ~ 1, indicating non-trivial long-range correlations in B (n) - these correlations
are not the consequence of summation over random variables or artifacts of non-
stationarity. Furthermore, the «anti-correlation» properties of I (n), indicated by
the positive {3 value, are consistent with a nonlinear feedback system that «kicks»
the heart rate away from extremely high or low values. This tendency, however,
does not only operate on a beat-to-beat basis (local effect) but over a wide range
of time scales, a fractal property of cardiac regulation.
We [13] analyzed data from two different groups of subjects: 10 adults with-
out clinical evidence of heart disease (age range: 32-64 years, mean 44) and 10
adults with severe heart failure (age range: 22--63 years; mean 54). Data from
patients with heart failure due to severe left ventricular dysfunction are likely to
be particularly informative in analyzing correlations under pathologic conditions
since these individuals have well-defined abnormalities in both the sympathetic and
parasympathetic control mechanisms that regulate beat-to-beat variability. Further-
more, such patients are at very high risk for sudden death. Both exponents (0: and
(3) were significantly different between the diseased and normal groups [13].
Previous studies have demonstrated marked changes in short-range heart rate
dynamics in heart failure compared to healthy function, including the emergence
of intermittent relatively low frequency (rv 1 cycle/minute) heart rate oscillations
associated with the well-described syndrome of periodic (Cheyne-Stokes) respira-
tion, an abnormal waxing and wanning breathing pattern often associated with low
cardiac output. This pathologic, characteristic time scale is indicated by a vertical
arrow in Fig. 5b.
The long-range power-law correlations in healthy heart rate dynamics may be
adaptive for at least two reasons [14--16]: (i) the long-range correlations serve as
an newly described organizing principle for highly complex, non-linear processes
that generate fluctuations on a wide range of time scales, and (ii) the lack of
a characteristic scale helps prevent excessive mode-locking that would restrict
the functional responsiveness (plasticity) of the organism. Support for these two
related conjectures is provided by observations from severely pathologic states
such as heart failure where the breakdown of long-range correlations is often
accompanied by the emergence of a dominant frequency mode (e.g., the Cheyne-
Stokes frequency). Analogous transitions to highly periodic behavior have been
observed in a wide range of other disease states including certain malignancies,
Healthy Dynamics:
Long-Range
Power-Law Correlations
BREAK DOWN
Brown Noise Highly Periodic

(Random Walk) (Single Dominant White Noise
Time Scale)
e.g. Sinus Rhythm Dynamics e.g. Heart Rate Intervals

in Heart Failure in Atrial Fibrillation
Fig. 6 The breakdown of long-range power law correlations may lead to any of three dynamical
states: (i) a random walk «<brown noise») as observed in low frequency heart rate fluctuations
in certain cases of severe heart failure; (ii) highly periodic oscillations, as also observed in
Cheyne-Stokes pathophysiology in heart failure, and (iii) completely uncorrelated behavior
(<<white noise»), exemplified by the heart rate during atrial fibrillation.
sudden cardiac death, epilepsy, fetal distress syndromes and with certain drug
toxicities [15, 16].
Important unanswered questions currently under study include: What are the
physiological mechanisms underlying such long-range correlations in cardiac beat-
to-beat intervals? Are these fluctuations entirely stochastic or do they represent the
interplay of deterministic and stochastic mechanisms [17]? How do these findings
relate to the suggestion that some features of normal heart rate variability are due
to chaotic dynamics [17-20]?
From a practical viewpoint, these findings may have implications for phys-
iological monitoring and in particular for cardiac rhythm analysis. The complete
breakdown of normal long-range correlations in any physiological system could
theoretically lead to three possible dynamical states (Fig. 6): (i) a random walk
(brown noise), (ii) highly periodic behavior, or (iii) completely uncorrelated be-
havior (white noise). Cases (i) and (ii) both indicate only «trivial» long-range cor-
relations of the types observed in severe heart failure. Case (iii) may correspond to
certain cardiac arrythmias such as fibrillation. More subtle or intermittent degrada-
tion of long-range correlation properties may provide an early warning of incipient
pathology, including an increased risk of sudden cardiac death. A breakdown of
long-range correlations may also be an important marker of aging [21]. Finally, we
observe that the long-range power-law correlations present in the healthy heart-
beat imply that the underlying control mechanisms actually drive the system away
from a single steady state. Therefore, the classical theory of homeostasis, according
64 Peng et al.
to which stable physiological processes seek to maintain «constancy» [12], and

its more recently proposed modifications under the rubric of «homeodynamics»,
need to be revised and extended to account explicitly for this far from equilibrium
behavior.
Acknowledgement:
We wish to thank C. DeLisi, F. Sciortino, M.M.E. Matsa and S.M. Ossadnik for
help at various stages of this work, and AHA, CONACYT, Mathers Charitable
Foundation, NIH, NIDA, NIMH, NSF, ONR and the US-Israel Binational Foun-
dation for support.
References
[1] C.-K. Peng, S.Y. Buldyrev, A.L. Goldberger, S. Havlin, F. Sciortino, M. Si-
mons, and H.E. Stanley, Nature 356, 168 (1992).
[2] E.W. Montroll and M.F. Shlesinger, «The Wonderful World of Random
Walks», in Nonequilibrium Phenomena II. From Stochastics to Hydrody-
namics, eds. J.L. Lebowitz and E.W. Montroll, pp. 1-121 (North-Holland,
Amsterdam, 1984).
[3] S. Tavare and B.W. Giddings, in Mathematical Methods for DNA Sequences,
Eds. M.S . Waterman (CRC Press, Boca Raton, 1989), pp. 117-132.
[4] H.E. Stanley, Introduction to Phase Transitions and Critical Phenomena, pp.
120--121 (Oxford Univ. Press, London, 1971).
[5] S. Havlin, R.B. Selinger, M. Schwartz, H.E. Stanley and A. Bunde, Phys.
Rev. Lett. 61, 1438 (1988).
[6] c.-K. Peng, S.Y. Buldyrev, A.L. Goldberger, S. Havlin, F. Sciortino, M. Si-
mons, and H.E. Stanley, Physica 191, 25 (1992).
[7] c.-K. Peng, S.Y. Buldyrev, A.L. Goldberger, S. Havlin, M. Simons, and H.E.
Stanley, Phys. Rev. E 47, 3729 (1993).
[8] P.J. Munson. R.C. Taylor and G.S. Michaels, Nature 360, 636 (1992).
[9] S.Y. Buldyrev, A.L. Goldberger, S. Havlin, C.-K. Peng, H.E. Stanley, and M.
Simons, preprint.
[10] S.Y. Buldyrev, A.L. Goldberger, S. Havlin, C.-K. Peng, M. Simons, and H.E.
Stanley, Phys. Rev. E 47, 4514 (1993).
[11] H.E. Stanley et al., preprint.
[12] w.B. Cannon, Physiol. Rev. 9, 399 (1929).
[13] C.-K. Peng, J. Mietus, I .M. Hausdorff, S. Havlin, H.E. Stanley, and A.L.
Goldberger, Phys. Rev. Lett. 70, 1343 (1993), and references therein.
[14] B.l. West and A.L. Goldberger, 1. Appl. Physiol., 60, 189 (1986).
[15] RI. West and A.L. Goldberger, Am. Sci., 75, 354 (1987), and references
therein.
[16] A.L. Goldberger, D.R. Rigney and B.J. West, Sci. Am. 262,42 (1990).
[17] D.R. Rigney et al., preprint.
[18] D.R. Rigney, I.E. Mietus and A.L. Goldberger, Circulation 82 (Suppl. III),
236 (1990).
[19] J.E. Skinner, C. Carpeggiani, C.E. Landisman and K.W. Fulton, Circ . Res.
68, 966 (1991).
[20] A.L. Goldberger, NIPS (Int. Union Physiol. Sci.lAm. Physiol. Soc .) 6, 87
(1991).
[21] L.A. Lipsitz and A.L. Goldberger, 1. Amer. Med. Assoc. 267, 1806 (1992).
Fractals in
Biological Design
and Morphogenesis
Design of Biological Organisms
and Fractal Geometry
Ewald R. Weibel
Department of Anatomy
University of Bern
Biihlstrasse 26
CH-3OO0 Bern 9, Switzerland
Abstract. This paper explores the question whether fractal geometry forms an important basis for
determining the design of biological organisms during morphogenesis. It is first observed that function-
ally important internal surfaces such as the intracellular membranes of the liver or the gas exchange
surface of the lung have characteristics of fractal surfaces in that the surface area measured depends
on the microscopic resolution. However, these surfaces are fractal within bounds and the sequential
occurrence of different «generators» of surface texture may require several fractal regressions. It is
secondly noted that airways and vascular trees show the properties of fractal trees and that this may
offer significant functional advantages. The question is finally raised whether and to what extent fractal
constructive algorithms may form part of genetic programming.
1 Introduction
It has become commonplace to note that many pictures generated by fractal al-
gorithms are reminiscent of some of the beautiful structures found in biological
organisms: trees, flowers, or animal structures alike. And it is easily noted that
many of these structures are derived by growth from smaller and often simpler
units, just as fractal algorithms generate progressively more complex structures.
The fundamental underlying principle in fractal geometry is self-similarity and
scale invariance of form as the structure develops. And clearly this is very often
also an underlying principle in the development and growth of biological forms,
as noted already by D'Arcy Thompson in 1942. Two well-known examples are
evidence for this notion:
(1) the Nautilus, a large sea mollusc whose progression of chamber size follows
a logarithmic growth law (Fig. I); all chambers are self-similar, i.e. have
the same proportions, because, with every growth step, the animal needs to
add an air-filled floating chamber which is proportional to the weight gain of
the animal. Since growth is proportional to attained size the growth curve is
logarithmic and, in tum, becomes imprinted in the shell structure. This kind
of structure also appears in many fractal patterns, e.g. at the periphery of the
Mandelbrot set;
(2) the growth of trees, such as a beech, starts from small saplings whose stem
grows in length and thickness and becomes enlarged by the addition of a new
generation of twigs with each growth step. Fig. 2 compares a three years old
sapling with a terminal twig taken from a branch of its parent tree, a beech
about 100 years old; the similarity between sapling and twig is evident so
that the growth of such a tree is clearly governed by self-similarity.
Biological Organisms and Fractal Geometry 69
Fig. 1 Mid-plane section of a Nautilus shell showing the logarithmic spiral resulting from propor-
tional growth of the air chambers.
One could easily come up with numerous examples of this kind. However,
this is anecdotal evidence that seems to document the fractal nature of biological
organisms; and one can, as so often, easily come up with anecdotal evidence that
proves the contrary. Can we be more rigorous and more systematic in exploring
whether fractal geometry serves an important template function in determining the
design of living organisms?
2 The hypothesis of fractal morphogenesis of biostructures

Fractal geometry is conceived as a set of constructive algorithms that determine
design by self-similarity of the structure at various scales, in other words by
scale invariance of form. Such constructive algorithms lend themselves to the
development of a theory of morphogenesis by proportional growth from which
one can derive models that permit to test the hypothesis that morphogenesis of
biostructures follows fractal principles. To approach this let us first examine two
cases of morphogenesis in higher organisms which are of central importance for
their good functioning:
(I) the formation of large internal surfaces in confined spaces, such as cellular
membranes or the gas exchange surface of the lung;
(2) the design of branched vessel systems by which such large internal surfaces
are efficiently accessed from central sources, examples being the vascular
trees that distribute blood and blood-borne substances to the cells, or the
airways which distribute oxygen to the large gas exchange area of the lung.
70 Weibel
Fig. 2 Three years. old sapling of a beech tree compared to an end twig of its nearly 100 years old
parent tree.
3 Fractal nature of internal surfaces

3.1 First case study: cellular membranes
The discovery that cellular membrane systems have fractal properties arose from
the uncertainty of observation about the extent of such membranes. In 1968 meth-
ods for measuring the surface area of cellular membranes on electron micrographs
had been established (Weibel et al. 1966; Weibel 1979). The methods were sim-
ple as they were based on statistical sampling procedures derived from geometric
probability of intersection of some test line probe with the membrane images ob-
served on sections. They could be applied in studies in cell biology that wished to
establish a relationship between some enzyme functions and the system of endo-
plasmic membranes, for example in liver cells (Fig. 3). When the first studies on
the morphometry of liver cell membranes were published (Loud 1968; Weibel et
al. 1969) the results did not match as we obtained much higher values than other
groups (Table 1):
Fig. 3 Electron micrographs of liver cell cytoplasm at two different magnifications (a: 18'000 x,
b:58'000 x) show complex of endoplasmic reticulum membranes (arrows) and mitochon-
dria (m) with their membranes. Higher magnifications reveal greater details in the shape of
membrane profiles.
Authors Surface density Magnification at

of endoplasmic which measurements
reticulum membranes were done
(m 2 . em- 3 )
Loud (1968) 5.7 12'500
Weibel et al. (1969) 10.9 90'000
Table 1 Uncertainty in surface area estimates of endoplasmic membranes in rat liver measured by
different authors.
Long debates followed about which of the estimates was correct, whether
the liver cells contained 6 or 11 m2 of membranes per em 3 , a quite significant
difference, and whether the stereological methods used were reliable since it ap-
peared possible that the same method may yield different results if the measure-
ments were done at different magnifications of the electron micrographs (Table
1). The key event towards the resolution of this debate was the publication, in
1977, of Benoit Mandelbrot's book «Fractals: Form, Chance and Dimension».
There he demonstrated, based on the work of Richardson, that the length of a
coast line was indeterminate unless the fractal dimension and the yardstick were
defined. Were we dealing with a «Coast of Wales Effect»? If so, measurements
obtained at a higher magnification (Fig. 3) were bound to yield higher values than
those done at lower magnification. We undertook a systematic study on measur-
72 Weibel
30
[1-0]
0.0 30
20
[1-0] 20
'E
::l.
!;
10
-
'E
::l.
>-
~:!]~~
f-
(i) 8 10 Z
Z .,SVimi.mi
UJ 6 o 0 -__ 8>-
0 f-
0 0 -0.09
o-SVomi.mi 6U)
I Z
f- UJ
<!> 4 0
Z
UJ 4UJ
...J 3 u
>-
0:: 30::
tt
<t ::)
0 2
Z . ,Sver.cy (f)
::)
0 2
co + + + + + +
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
:::
N 0
!:?
0>
0> ....
CD
~ l!2
4 6 8 10 20 40 60 .10'6
RESOLUTION SCALE '1=a.M- 1
Fig.4 Double logarithmic plot of boundary length density estimates (which stereology converts
into surface density estimates) of membranes against resolution scale. Slope of regression is
related to fractal dimension D. [From Paumgartner et al. 1981]
ing liver cell membranes at different magnifications and found that indeed the
estimates of surface density increased with increasing resolution (Paumgartner
et aI., 1981). The test was the log-log plot of the data (Fig. 4) from which it
followed that endoplasmic reticulum membranes were characterized by a fractal
dimension of 2.7 (assuming that the slope for a membrane in 3D space is 2 - D),
and that the fractal dimension of each type of cellular membranes was different.
It was however interesting that in all cases there was an upper cut-off point at
130'000 x magnification beyond which the curve was flat. This suggested that
there was a change in the nature of the membrane image at this resolution, and
indeed this is the magnification at which the membrane becomes visible as a phys-
ical sheet of finite and measurable thickness. At this magnification the membranes
change from fractal to deterministic structures, but if we could further increase
resolution the surface of the sheets would again assume fractal properties and the
surface area estimates would continue to rise, though possibly at a different slope,
as we resolve surface molecules and their structure.
Fig. 5 Increasing microscopic resolution reveals a hierarchy of structures that form the inner lung
surface: alveoli around the airways (A), capillaries in the alveolar walls with their imprints on
the surface (B), and membrane folds of the epithelial cell forming the surface of the air-blood
barrier. A and B are scanning electron micrographs magnified 75 x and I '()()(} x respectively;
C is a thin section electron micrograph magnified 46'000 x.
3.2 Second case study: the gas exchange surface of the lung
A first look at the lung shows it to be similar to a foam made of some 300
million little bubbles, called alveoli, all open to the airways and densely covered
by a network of blood capillaries (Fig. 5a + b). This is in the interest of efficient
uptake of oxygen from the air which requires a large surface of contact to be
established between air and blood. Just how large is this surface? Here again
an uncertainty arose. The first estimates of alveolar surface area in the human
lung, obtained around 1960 by light microscopy, yielded values of about 60 m 2
(Weibel and Gomez 1962; Weibel 1963), but when a few years later the elctron
74 Weibel
Resolution Effect on Lung Surface Estimate

1000
E
Q
.S:
2:-
'iii
-- ------. --.::
c:
Q)
-0 --.~
Q)
U
.~--
---::::. -- ':::-.t.----
CIS
't
;:, ----~
en
100 'I 'I
0.01 0.1 10
- EM Resolution in ~m - LM-
Fig. 6 Log-log plot of surface density estimates of inner lung surface obtained at different resolution
levels in the light (LM) and electron microscope. A single regression represents the data less
well than two (or perhaps even more) regressions.
microscope became available for such measurements I suspected that this must
be an underestimate of the «true» surface (Weibel 1964). Indeed, when in 1978
human lungs could be subjected to a thorough morphometric study we estimated
the alveolar surface area at 130 m2 (Gehr et al. 1978). Is this again due to a fractal
nature of the inner lung surface?
The obvious test was to measure the alveolar surface area at increasing mag-
nifications, and it was found that it increased with a slope corresponding to a
fractal dimension of 2.24 (Keller et al. 1978). However, there was a problem
because the successive structures that were resolved at increasing magnification
were of different nature (Fig. 5): at first we see «smooth» alveoli (Fig. 5a), then
we resolve the ridges that are imprinted by the capillaries in the thin alveolar
walls (Fig. 5b), and finally we see the fine wrinkles of the alveolar epithelial cell
membranes (Fig. 5c). Is it then justified to use a single fractal dimension as a
descriptor of surface complexity? We can hardly claim that alveoli, capillaries and
surface wrinkles are self-similar structures; each has its own «generator», is de-
termined by its own «constructive algorithm» and, accordingly, we should search
for at least two or three self-similarity levels. This is indeed found if the log-log
regression is scrutinized (Fig. 6). We should note that Rigaut (1984) had found
the fractal dimension of the alveolar surface to change when he studied it by light
microscopy; his analysis was however limited to the lower half of the regressions
shown in Fig. 6.
In conclusion, we found that internal surfaces both at the level of cells and
organs have fractal properties, and that their full morphometric description is best
served by estimating their degree of complexity by determining a fractal dimension

in addition to some metric estimates at a given yardstick or resolution. However,
the fractal model must be used with caution and critically if the morphometric
estimates are to be used in a functional context (Weibel 1991b). «Membranes»
are physical sheets so that we will find a cut-off point on the log-log plots, and
this may allow us to indicate a resolution at which a deterministic surface can
be estimated, one which is related to specific structural entities and also makes
functional sense. Furthermore, biological surfaces may be built of a hierarchy of
non-selfsimilar elements, and it is of great functional significance to sort out the
deterministic structures which serve as generators of surface texture. On the whole
we can say that biological surfaces are fractal with bounds, and it may well be
that the bounds are the critical points that the analysis has to reveal.
From the perspective of the cell or the organism we therefore find that func-
tionally important interfaces show fractal structure although locally, in the func-
tional domain, they are generated by the folding of deterministic sheets into a
confined space. In the future, it may be heuristically very productive to ask the
question whether the fractal design of interfaces brings metabolic or morphogenetic
advantages to cells, organs, and organism.
4 Accessing large surfaces from a central source

The liver or the heart muscle contain a large number of cells which occupy the
organ space rather homogeneously. They must all be supplied with oxygen and
substrates to perform their functions and this is served by the blood vessels. Like-
wise, the alveolar surface is densely and quite homogeneously arranged within
the lung, i.e. the area of 130 m2 - nearly the size of a tennis court - is finely
crumpled into a space of about 5 liters, forming some 300 million alveoli (Weibel
1963, 1984). The capillary network which is contained in the walls between alve-
oli (Fig. 5) must be perfused with blood as evenly as possible by pumping the
blood from the heart through the arteries to the capillaries and then collecting it
into veins that lead it back to the heart. Some alveolar surface elements are close
to the heart, others are far away, so the functional problem is to divide the blood
stream efficiently to allow blood cells to reach all points of the large alveolar
surface evenly and in approximately the same time span. In addition, all alveoli
must be efficiently supplied with fresh oxygen from each inspired volume of fresh
air, and the goal must be to achieve a well-matched ventilation and perfusion of
the large number of gas exchange units.
These problems are solved by designing the blood vessels and the airways as
trees that have their roots in the heart or in the upper airways, nose and mouth,
respectively, and extend with their twigs to the units that must be supplied, cells
in the heart muscle or liver, and alveoli with their capillaries in the lung (Fig. 7).
We had noted above that many botanical trees have the properties of fractal
structures and that the underlying self-similarity is the result of morphogenesis:
76 Weibel
Fig. 7 Resin cast of the airways and blood vessels in the human lung (a) and of coronary arteries
in the heart muscle (b) reveal their tree structure based on dichotomous branching.
they grow and branch systematically from the stem to the twigs (Fig. 2). The
morphogenesis of the lung's airway tree follows much the same principles (Weibel
1984). It begins in the young fetus by formation of the stem, the trachea, and
two bronchial buds that extend sideways into a pad of mesenchyme, primitive
connective tissue, which contains a network of early blood vessels. The bronchial
buds are blind-ending epithelial tubes whose wall is a compact layer of cells. They
grow in length, and at a certain point the terminal cap of epithelium forms a Y-
shaped tip which initiates the formation of two branches which grow in length and
then split at their tips, each into two branches, and so on, a morphogenetic process
called dichotomy. Thus with each branching the number of end twigs is doubled:
2,4,8, 16,32 . . . , until, after 23 generations, some 8 million terminal airways are
formed, the total number found in the normal human lung (Weibel 1963). As this
branching goes on, all airway tubes grow proportionately in length and diameter,
just as in a beech tree (Fig. 2), so that in the final tree the dimensions of airways
decrease regularly from the central bronchi to the peripheral branches. Because
branching occurs in all directions the result is a tree whose end twigs fill the entire
space of the lung (Fig. 7a).
Fig. 8a shows a fragment of a cast of a human airway tree; it is quite evident
that, inspite of irregularities, the branching pattern is similar at all levels from
larger to smaller branches. Is this a fractal tree? In his books on Fractal Objects
Benoit Mandelbrot (1977; 1983) has constructed «Koch trees» which have the
properties to fill, with their tips, the space homogeneously. The generator was a T-
or Y-shaped element with fixed proportions in the length and diameter of the legs.
Fig. 8 Peripheral small airways shown in a resin cast (a) and in a scanning electron micrograph (b)
which shows the transition of conducting airways (asterisk) to alveolar ducts (arrows). The
largest airway in (b) corresponds about to the smallest end twigs in (a).
The tree was self-similar because these proportions were maintained throughout
the branching process. These Koch trees resemble the pattern of airway branching.
To test whether the airways of the human lung confonn to such fractal models we
need to test whether the proportions between length and diameter within one gen-
eration, and from generation to generation are maintained. Length and diameter of
airways vary considerably (Fig. 8a); but when such measurements are averaged for
each generation we find that the proportions are constant. The length-to-diameter
ratio is invariant at 3.25, and the branching ratios (the reduction factor with each
successive generation) are 0.86 for diameter and 0.62 for length, irrespective of
the generation (Weibel 1963). These findings support the notion that the airway
tree has the basic properties of self-similarity. The variation in the proportions of
individual branches is the result of the necessity to reach all points in a space that
is detennined by the shape of the chest cavity (Nelson and Manchester 1988), and
this also explains part of the differences in the branching pattern between different
species (Nelson et al. 1990). Along another line of reasoning it has been argued
that the airway tree is fractal if the progression of dimensional change, diameters
or lengths, of the segments follows a power law when plotted against generation
of branching (West et al. 1986). This assumes that the «yardstick» or scale is
reduced by a constant factor from one generation to the next. Fig. 9 shows a plot
of the average airway diameter against generation for the human lung. West et al.
78 Weibel
AIRWAYS HUMAN LUNG
•
1.0
0.7
0.5
E
u
.S 0.3
n::
...,UJ
UJ 0.2
1: d(z) = 2.36' i1.35
<
is
~ [d in em]
~ 0.1
C(
UJ
~ 0.07
n:: 0
UJ
~ 0.05
ACINUS
0.03
2 3 4 5 10 15 20 24
GENERATION Z
Fig. 9 Log-log plot of airway diameters measured in human lungs against the generation of branch-
ing. [From Weibel 1991]
(1986) have concluded from this kind of plot that the airway tree can be described
as a fractal tree, but that a harmonic variation is superimposed which is similar
for all species examined, except that there is a phase shift in the human lung with
respect to that of other species (Nelson et al. 1990).
The dimensional changes in the airway tree can also be interpreted differently.
In our original analysis of the human airway tree (Weibel and Gomez 1962) we
asked the bioengineering question whether the dimensions of the successive gener-
ations were optimized for minimal work of maintaining air flow. The total work is
determined by the flow resistance which increases as the dimensions decrease and
the work required to move the mass of air which increases as dimensions increase
(Wilson 1967). It is well known in hydrodynamics that the total work of flow in
18 0
0
10
8
6 o
4 o
o
E
E 2
~ ACINAR AIRWAYS
cr
w1
tu
:E
.8 BRONCHI
RESPIRATORY
BRONCHIOLES
ALVEOLAR
.6
~
a d{z):do .2-Z/3 -i-i DUCTS SACS
>- 1+
~ TRANSJTI~NAL -i-i-i-t_i
-t-{
I
cr BRONCHIOLE
<i .2
w
(!)
~
ffi1
~
.04
.0
.01
4567 89~n~~K~~n~~~~~n~~~v~
GENERATION
Fig. 10 Exponential plot of the same data as in Fig.9 reveals different regressions for conducting
airways (open circles) and alveolated airways (closed circles) in the acinus as gas exchange
unit (Fig.8b). The plot is extended to show the size of the arterial end branches which are
located in generations 26--28.
a branched system of pipes is minimal if the diameter of the pipes is reduced by

the cube root of one half, or 2- 1/ 3 , at each dichotomous branching. The system
of pipes that distribute gas in a city are designed according to that law. D' Arcy
Thompson (1942) was first to notice that this rule often applied with respect to
biological vascular systems. When we plotted our measurements of airway diam-
eter against the generation of branching we found that the first 14 generations
follow this law (Fig. 10): the diameter in generation z is d(z) = do· 2- z/ 3 . The
last generations do not follow the same line, but we will see that they serve their
function in a different way.
Mandelbrot (1983) discusses this result in terms of fractal geometry and notes
that in a space-filling tree the diameter exponent of the sequence of branches should
be close to 3 and that accordingly the diameter branching ratio should be 2- 1/ 3 ,
80 Weibel
Fig. 11 The relations between pulmonary arteries (A) and bronchi (B) show that they run in parallel
and that their diameters are similar. When arteries penetrate into the acinus (compare Fig.8b)
they branch more frequently than the airways so that their diameter decreases more rapidly,
as shown in Fig. 10. (a) and (b) are resin casts of human lungs magnified 3 x and 8 x,
respectively; in the light micrograph of a rabbit lung (c), magnified 25 x, the blood vessels
have been highlighted by black stain of the blood plasma.
as observed. Thus, even according to this analysis the human airway tree has the
properties of a fractal tree, at least as far as the airways are simple air conducting
tubes.
Beyond generation 14 the nature of the airways changes (Fig. 8b): their wall
is formed by alveoli which accommodate the large gas exchange surface around
the about eight generations of terminal branchings within the gas exchange units
called acinus (Haefeli-Bleuer and Weibel 1988). The volume of these airways
becomes greatly expanded, and even the central air ducts around which the sleeve
of alveoli is arranged takes on a larger cross-section than predicted from the 2- 1/ 3
rule (Fig. 10). The functional explanation for this change is that, within the acinar
airways, air convection is slowed down and the oxygen transfer to the large gas
exchange surface is accomplished to a significant part by molecular diffusion of
oxygen within the air phase. This process is greatly improved if the crossection
of the diffusion front is as large as possible; the observed slower reduction of
airway diameter appears to strike an optimal balance between the conditions for
convective mass flow of air and oxygen diffusion within the air phase. Mandelbrot
(1977) argues that this feature furthermore forms good conditions for what he
calls the «alveolar inner cut-off» of a space-filling tree. Thus we conclude that the
airway system of the human lung is well designed, both in terms of solving its
physiological functions (Wilson 1967) as in terms of its morphogenetic properties,
and that both are linked to the airways' fractal nature.
Finally, a few brief remarks on the design of the pulmonary vascular trees,
arteries and veins, which develop in concert with the airways. Consider the pul-
monary arteries: they follow the course of the airways, and their diameters are
very nearly matched to those of the associated airways (Fig. lla and b). When
arteries penetrate into the acinus additional branches are formed, i.e. the arteries
branch over a total of about 28 generations, on average, before they reach their
endbranches that discharge blood into the capillary network on the alveolar surface
(Fig. llb and c). When arterial diameters are plotted against generations, as in
Fig. 10, we find that they are reduced according to the 2- 1/ 3 rule all the way out
to the endbranches (Weibel and Gomez 1962); their design thus appears governed
by the requirements for optimized flow conditions. The functional importance of
these design principles has recently been established by a very thorough analysis
of the pulmonary arterial tree as a fractal continuum model (Krenz et al. 1992).
Arteries that supply tissues such as heart muscle show a very similar design (Fig.
7b). Mandelbrot (1983, chapter 17) discusses this in terms of fractal design and
notes that the diameter exponent should be somewhat smaller than that of the
airways, and that this is confirmed by empirical data (Suwa and Takahashi 1971).
It has indeed been shown long before the advent of fractal geometry that arteries
and veins form trees along the same principles (Hess 1903).
We therefore arrive at the conclusion that airways as well as blood vessels
are constructed like fractal trees. Their design principle is that of a hierarchy of
self-similar tubes which develop in morphogenesis by a well balanced succession
of branching and growth. The functional result of this design is that all points
on the gas exchange surface of the lung are at nearly the same distance from the
source of fresh air and blood: the main central airways and the heart, respectively.
Furthermore, the design of such a tree allows air and blood flow through the
tubes to remain laminar throughout such that the work of flow is minimal; but
the «mixing» of air and blood at the gas exchange surface is just as efficient as if
the two media were mixed by turbulent flow: it can be said that in the design of
airway and vascular trees the turbulence has become «frozen into structure».
The fundamental question that may need to be addressed in future research is
whether the apparent fractal nature of airways and vessels in the lung is the result of
interactions at the physical level during morphogenesis, that is whether it is simply
the expression of an adaptive malleability of structures to imposed stresses, the
result of an attraction-repulsion equilibrium (Sernetz et al. 1988); or whether the
fractal design is predetermined, programmed in the genetic instructions for design,
including the maintenance of strict proportionality. The latter alternative has a high
degree of likelihood because much of the basic morphogenesis of the lung occurs
before birth, that is before the lung is subjected to the functional stresses imposed
by gas exchange between the organism and the environment. It may be provocative
to postulate that the organism may use «fractal genes» for efficient programming
of some of its many hierarchical structures, such as airways and blood vessels.
82 Weibel
5 The design of bio-organisms

We have looked at cells and their membranes, where most metabolic functions take
place, and found that they have fractal properties, at least within certain bounds.
We have looked at the lung as the vital supplier of oxygen, and have found fractal
properties in the gas exchange surface as well as in the airways and vessels that
supply it with air and blood. We have also noted that the blood vessel systems
that supply the cells with oxygen and substrate are designed in much the same
way; they are fundamental design elements of the organism as a whole, elements
that connect the units, bind them into an integral system in the functional sense.
Such observations provide strong support to the hypothesis that fractal geometry
is a fundamental design principle for biological organisms. Although it may not
be at the basis of all morphogenetic processes fractal geometry still appears to be
a determinant of many of the key elements of animal design.
This, finally, leads to the consideration of whole animal design. Animals
come in all sizes - from the smallest mammal, the Etruscan shrew that weighs
2 grams to the elephant of 5 tons - and, although their basic design is similar,
many or most of the parameters that determine their function and design are
non-linearly related to body size. This is particularly true for metabolic functions
(Kleiber 1961; Schmidt-Nielsen 1984; Weibel and Taylor 1981; Weibel at al. 1992).
Oxygen consumption rate per unit body mass falls with the 1/4 power of body
mass so that a mouse has a six times greater mass-specific metabolic rate than a
cow, and heart rate also falls with the 1/4 power of body mass. The surface area
of mitochondrial inner membranes, on which oxidative energy metabolism takes
place, varies in proportion to oxygen consumption and is thus non-linearly related
to body size: in accordance with functional demand mitochondrial membranes are
more densely packed in the cells of small animals than in those of large animals.
Accordingly, the design of the vascular systems must be adjusted to supply a
denser capillary network with adequate blood flow (Hoppeler and Kayar 1988).
Such observations have led Sernetz et al. (1989) to postulate that metabolism is
organized on fractal principles, and Spatz (1991) to speculate that this is related
to the necessity to design the blood vessels as fractal trees. That this may not be
as simple is seen in the fact that design is generally determined by well balanced
combinations of structural and functional parameters, and that such vital structures
as the pulmonary gas exchanger may be designed with considerable safety factors
(Weibel et al. 1992).
6 Concluding remarks
What can we conclude from this analysis which focussed on internal surfaces and
the tree structures that access them from central sources? In each case we have
found several elements that suggested fractal design. How much of this is fact,
and how much is fiction? Clearly, we «proved» the fractal nature of some of these
design properties by log-log regressions. But is this enough?
With respect to the functional significance of these design properties we

concluded (a) that the packing of surfaces into cells and organs is best achieved
by using fractal structures, and (b) that to build vascular trees, airways, and even
botanical trees, without fractal principles does not lead to intelligent solutions. But
this was simply obtained by looking for analogies or similarities in the structure of
fractal models and in real biostructures. This coincidental evidence is interesting,
but it becomes heuristically productive only if we can take the analysis one step
further to ask whether morphogenesis is, indeed, governed by fractal principles,
be it at the genetic or epigenetic level. This proof has still to be provided.
There is a fundamental likeness between fractals and genes that determine
biostructures: they are both «constructive algorithms» that create structures with
precise connectivity and proportionality, while still allowing for (epigenetic) vari-
ation under external and internal influences, or imposed stresses. This, therefore,
is the fundamental provocative question linking fractals to biological design: do
genes contain fractal algorithms? Do we not have to suspect this to be true?
Consider the ubiquitous occurrence of some basic structures of the Mandelbrot
set, such as the logarithmic spiral: we find it in snails, molluscan shells, or the
developing fern, and this fundamental structure also appears to underlie the model
for airway trees. But this is no more than speculation; it becomes valid only if
it is formulated into a coherent research strategy which can be tested against the
facts of nature.
Acknowledgement:
This work has been supported by grants from the Swiss National Science Foun-
dation Nos. 31-30946.91 and 31-28610.90, and by a grant from the Maurice E.
Muller Foundation. I gratefully acknowledge the help received in the preparation
of this paper by Karl Babl, Elsbeth Hanger, and Barbara Krieger.
References
[1] Gehr P., M. Bachofen, and E.R. Weibel (1978) The normal human lung: ultra-
structure and morphometric estimation of diffusion capacity. Respir. Physiol.
32: 121-140.
[2] Haefeli-Bleuer B., and E.R. Weibel (1988) Morphometry of the human pul-
monary acinus. Anat. Rec. 220: 401-414.
[3] Hess w.R. (1903) Eine mechanisch bedingte Gesetzmiissigkeit im Bau des
Blutgefasssystems. Arch. Entwickl. Mech. Org. 16: 632-641.
[4] Hoppeler H., and S.R. Kayar (1988) Capillarity and oxidative capacity of
muscle. New Physiol. Sci. 3: 113-116.
[5] Keller Hj., H.P. Friedli, P. Gebr, M. Bachofen, and E.R. Weibel (1975) The
effects of optical resolution on the estimation of stereological parameters.
Proc. Fourth Internat. Congr. Stereology. Gaithersburg, 1975. National Bureau
of Standards, Spec. Publ. 431: 409-410.
84 Weibel
[6] Kleiber M. (1961) The Fire of Life. New York, Wiley.

[7] Krenz G.S., J.H. Linehan, and c.A. Dawson (1992) A fractal continuum
model of the pulmonary arterial tree. J. Appl. Physiol. 72-2225-2237.
[8] Loud A. V. (1968) A quantitative stereological description of the ultrastructure
of normal rat liver parenchymal cells. J. Cell BioI. 37:27.
[9] Mandelbrot B. (1977) Form, Chance, and Dimension. New York, Freeman.
[10] Mandelbrot B. (1983) The Fractal Geometry of Nature. New York, Freeman.
[11] Nelson T.R., and D.K. Manchester (1988) Modeling of lung morphogenesis
using fractal geometries. IEEE Trans. Med. Imaging 7: 321-327.
[12] Nelson T.R., B.I. West, and A.L. Goldberger (1990) The fractal lung: univer-
sal and species-related scaling patterns. Experientia Basel 46: 251-254.
[13] Paumgartner D., G. Losa, and E.R. Weibel (1981) Resolution effect on the
stereological estimation of surface and volume and its interpretation in terms
of fractal dimensions. J. Microsc. 121: 51-63.
[14] Rigaut J.P. (1984) An empirical formulation relating boundary lengths to
resolution in specimens showing «non-ideally fractal» dimensions. J. Microsc.
133: 41-54.
[15] Schmidt-Nielsen K. (1984) Scaling: Why is animal size so important? Cam-
bridge UK, Cambridge University Press.
[16] Semetz M., H.R. Bittner, and P. Wlczek (1988) Fraktale biologische Struk-
turen. Spiegel der Forschung 5: 8-11.
[17] Semetz M., H. Willems, and H.R. Bittner (1989) Fractal organization of
metabolism. In: Energy Transformations in Cells and Organisms. Eds: W.
Wieser and E. Gnaiger. Stuttgart-New York, Thieme, 82-90.
[18] Spatz H.-C. (1991) Circulation, metabolic rate, and body size in mammals.
J. Compo Physiol. B 161: 231-236.
[19] Suwa N., and T. Takahashi (1971) Morphological and Morphometrical Analy-
sis of Circulation in Hypertension and Ischemic Kidney. Munich, FRG: Urban
and Schwarzenberg.
[20] Thoma R. (1901) Ueber den Verzweigungsmodus der Arterien. Arch. Ent-
wicklungsmech. 12: 352--413.
[21] Thompson, D'A.W. (1942) On Growth and Form. Cambridge UK, Cambridge
University Press.
[22] Weibel E.R. (1963) Morphometry of the Human Lung. Heidelberg, Springer-
Verlag; New York, Academic Press.
[23] Weibel E.R. (1964) Morphometrics of the lung. In: The Handbook of Physi-
ology. American Physiological Society, Respiration Section, Vol. 1, Chapter
7,285-307.
[24] Weibel E.R., G.S. Kistler, and W.E Scherle (1966) Practical stereological
methods for morphometric cytology. J. Cell BioI. 30: 23-38.
[25] Weibel E.R. (1979) Stereological Methods. Vol. I: Practical Methods for Bi-
ological Morphometry. London-New York-Toronto, Academic Press.
[26] Weibel E.R. (1984) The Pathway for Oxygen. Cambridge MA, Harvard Uni-
versity Press.
[27] Weibel E.R. (1991a) Design of airways and blood vessels considered as
branching trees. In: The Lung: Scientific Foundations. Eds: R. G. Crystal,
J.B. West, PJ. Barnes, N.S. Chemiack and E.R. Weibel. New York, Raven,
711-720.
[28] Weibel E.R. (1991b) Fractal geometry: a design principle for living organisms.
Am. J. Physiol. 261: L361-L369.
[29] Weibel E.R., and D.M. Gomez (1962) Architecture of the human lung. Sci-
ence 137: 577-585.
[30] Weibel E.R., W. Staubli, H.R. Gnagi, and EA. Hess (1969) Correlated mor-
phometric and biochemical studies on the liver cell. I. Morphometric model,
stereologic methods and normal morphometric data for rat liver. J. Cell BioI.
42: 68-91.
[31] Weibel E.R., and c.R. Taylor (1981) Design of the mammalian respiratory
system. Respir. Physiol. 44: 1-164.
[32] Weibel E.R., c.R. Taylor, and H. Hoppeler (1992) Variations in function and
design: Testing symmorphosis in the respiratory system. Respir. Physiol. 87:
325-348.
[33] West BJ., V. Barghava, and A.L.Goldberger (1986) Beyond the principle of
similitude: renormalization in the bronchial tree. J. Appl. Physiol. 60: 1089-
1097.
[34] Wilson T.A. (1967) Design of the bronchial tree. Nature 213: 668-669.
Fractal and Non-Fractal Growth of
Biological Cell Systems
Petre Tautu
Department of Mathematical Models
Research Program Bioinformatics
German Cancer Research Center
D-69009 Heidelberg
Abstract. The paper gives a general view of the microscopic growth rules and the behaviour of the
spatial cell systems which generate fractal and non-fractal structures, with the purpose of finding their
mathematical characteristics.
1 Introduction
The present paper finds its motivation in the following questions: (A) Which
are those random spatial models that would be candidates for describing and
explaining the fractal growth of (large) biological cell systems? (B) Under the
modest assumption that fractals are not everywhere, are there random processes
which generate non fractal morphologies? (C) Or, is there a distinct «regime»
in the growth dynamics geometrically represented by a random fractal transient
structure?
The questions above have arisen out of the analysis of a stochastic spatial
growth process with local interactions as a model for carcinogenesis, originally
called the Markov configuration model [57]. Numerical experiments have shown
(under the caution that numerical experiments are never rigorous proofs) the for-
mation of compact and quasi-circular patterns with fluctuating irregular shape. The
conjecture made by the authors (Conjecture 3.2 in [57]) follows the well-known
Richardson theorem (1972) which is set in a first-passage percolation framework.
In [59], this Markovian model is associated with the original Eden models (1958,
1961) and the Williams-Bjerknes models (1972) in the family of interacting bi-
ological cell systems. They are characterized by an asymptotic shape which is
non-fractal.
In the present paper, the term «fractal» is strictly employed only for those
structures which are invariant under contraction or dilation, and possess the prop-
erty of self-similarity. According to the broad definition given by B.B. Mandelbrot
[40], «mathematical and natural fractals are shapes whose roughness and fragmen-
tation neither tend to vanish, nor fluctuate up and down, but remain essentially un-
changed as one zooms in continually and examination is refined.» Scale-invariant
fractal objects are «open and diffuse», containing large empty regions (<<holes»)
on every length scale: this configuration is the origin of non-trivial scale invari-
ance [65, p. 15] and occurs if there are long-range correlations in the pattern.
Fractal and Non-Fractal Growth 87
However, objects having the same fractal dimension may have different «lacunar-
ity»; also, there are non-lacunar fractals. Recently, it is shown that lacunar DLA
clusters become compact (and not self-similar) if the aggregate enlarges up to
30 million particles [41]. According to the strictly limited definition adopted in
this paper, such objects are no more ordinary fractals. It appears that the fractal
dimensionality - and its corresponding geometry - associated with simple ag-
gregation models often varies continuously with some parameters used to define
the model [45]. After all, fractal structures represent a geometrical expression of
critical phenomena.
Then, the term «non-fractal» will be used here for any other irregular, not
self-similar morphology. The reasons for this sharp distinction will be given in the
next paragraph.
2 On fractals and non-fractals

2.1 A fundamental distinction
In the present paper, scale invariant and statistically self-similar structures are
considered stricto sensu as genuine fractals. Accordingly, other irregular-looking
structures are denoted on the whole as «non-fractals». This discrimination appears
as unusual but avoids a certain ambiguity and its consequences.
The scaling hypothesis in statistical mechanics states that a physical system at
its critical point is invariant under scale transformations, and near its critical point
this invariance is broken and a «preferred» length scale - the correlation length -
appears. In addition, the scaling hypothesis requires that small scale details must be
ignored in order that scale invariance can be possible. For a mathematical review
see [34]. The concept of non-integral (fractal) dimension is strongly connected to
uniform (self-similar) scaling, so that ideal fractals are self-similar on all length
scales [l]. Nevertheless, statistical self-similarity reflects the fact that a structure
may look statistically similar while being different in detail at different scales.
Habitually, the generic term «fractal»conceals structures already defined as not
identical, like self-similar and self-affine. Indeed, self-affinity describes a nonuni-
form scaling where shapes are statistically invariant under transformations that
scale different coordinates by different amounts. The extension of fractal dimen-
sion to self-affine surfaces appears as «tricky» [63, p. 59]. Except for the singular
case when the fractal dimension D, the surface «fractal» dimension Ds and the
Euclidean dimension d are equal, objects cannot be «mass-fractal» and «surface-
fractal» on the same length scale [54]. In addition, it is likely to be difficult to
obtain a general expression of self-affine sets: this is an intractable problem [34,
pp. 128-129].
If under the label «fractal» one collects geometrical objects that are self-
similar, self-affine, piecewise self-affine, «fat» fractals and so on, one would come
to the idea that the classical geometrical forms are only particular cases in the
88 Tautu
fractal geometry (see the note in [16, p. 179]), that is, «familiar objects such as
the line, square, and cube are also fractals» [70, p. 666].
Finally, one must notice that in some cases self-similarity is «a side-effect of
computational universality», since the behaviour of the computer on a subset of
its input space may replicate its behaviour on the whole input space [4].
2.2 Scale in variance and morphogenesis
The actual paradigm in developmental biology postulates that morphogenesis and
morphologic evolution depend upon a special set of molecular regulatory mech-
anisms mediated by the cell surface as it interacts at particular places in the
embryo with other surfaces both cellular and noncellular [15, p. 4]. Although it
was often indicated that reaction-diffusion (RD) models are inadequate to explain
morphogenesis, there are some interesting RD models where diffusion is driven by
nonlinear collective interactions (<<cooperative topobiological interactions») which
are characteristic to the interacting cell biological systems. For instance, in [50] it
is assumed that the morphogen diffusivity may be controlled by the concentration
of a diffusible regulatory chemical that is produced at a constant rate by all cells.
The desired «perfect» invariance can be obtained by adjusting the production rate
as well as the rate at which the regulatory chemical leaks into the surroundings.
One may presume that in morphogenesis there exist dynamical mechanisms which
drive a cell system into a spatially and temporally scale invariant state. The sys-
tem might show self-organized criticality: it describes the tendency of dissipative
systems to drive themselves to a critical state with a wide range of length and time
scales. In such self-organizing systems may occur simultaneously global criticality
and local complexity.
Analyzing the aspects of dimensionality and the size of a cell system devel-
oping on a plane lattice, M.D. Hatlee and J.J. Kozak [29] considered that there are
changes of structures determined by changes in the nature of boundary conditions
(<<passive» to «active»). The most interesting hypothesis is that there may be two
stages of morphogenesis: the early stage showing biases toward normal growth
in integral dimensions, and the late stage showing «volume-preserving growth in
fractal dimensions». This makes explicit an optimality principle, namely the main-
tenance of the area (volume) constant while increasing the edge length (surface
area) in order to have the reactive sites distributed on the boundary of the system.
The hypothesis in [29] might suggest a partial answer to Question C above and
also to a question posed by H.E. Stanley [58]: «what evolutionary advantage does
a fractal morphology convey?».
2.3 Irregular cell surfaces
In principle, a surface is the boundary of an object embedded in the three-
dimensional Euclidean space. If d = 2, this boundary can be called «contour.»
As a matter of fact, many biological proliferative cell systems may be studied as
surface phenomena. The most interesting instance is surface morphogenesis [28]
which presents the changes in shape of a thin sheet of cells due to intrinsic growth
processes generated from inside the sheet itself. If one takes into consideration the
isotropic growth of a curved surface from a flat sheet, the «Gaussian curvature»
of the final structure determines whether the growth rate of the surface is sub- or
superharmonic [61]. By definition, a «rough» surface is an irregular one on which
there are no projections (<<overhanging regions») or where at least these protru-
sions do not dominate the scaling properties [23]. This justifies the introduction of
a function hex, t), x E Rd, d > 1, t ?: 0, which gives the «height» of the surface
at position x on the original d - 1 dimensional flat surface, and at time t. For
details the reader is referred to the indispensable survey [18].
Irregular surface is a subtle subject in itself, but as H.E. Stanley repeatedly
pointed out [58], there are many different surfaces, depending on the considered
physical process. It is evident that, in a biological context, the surface of a growing
cell system must be distinguished from the surface (membrane) of a single living
cell. The cytoskeleton governs the form of the cell surface which is locally and
globally modulated [15]. Unicellular organisms (e.g., algae) show fractal and non-
fractal structures that can be classified according to the values of the coefficients
a, j3, and 'Y of a deterministic system of differential equations [51, Eqs. 2 and 4]
governing the evolution of the cell boundary.
Growing cell systems, and particularly interacting cell systems, show a vari-
able irregular surface of their resulting compact clusters. Numerical experiments
with Eden, Williams-Bjerkness or Markov configuration models make visible
changing boundaries from one step to another (see, e.g., [10, pp. 20 and 49]
and [57, 52]).
2.4 A note of caution

In his investigations on mosaic patterns in chimaeric rats, P.M. Iannaccone [31]
showed that characteristic cell clusters (<<patches») formed in the mosaic liver are
fractals (Fig. 2 and Table 1 in [31]). Interpreting his observations, he claimed
that (i) self-similar processes are responsible for the generation of these fractal
cell patterns, and that (ii) mosaic pattern «may be definable as equations with
relatively few elements known as iterating functions sets». These suppositions are
reproduced in [47] and sustained by computer simulations [48]. The conclusion
is that «iterating self-similar division rules without focal centers of growth or cell
movement were responsible for increase in parenchymal mass». Because these
claims reveal a certain language confusion some brief definitions are necessary.
(i) Rigorously speaking, self-similar processes are characterized as limits in
distribution of some stochastic processes {~( t), t ?: O} under joint rescaling of
space and time. The law P of the new process {X (t) , -00 < t < oo} must verify
for all t the relation
P { X (a t) ::::: x} = P {a H X ( t) ::::: x} , for all real a > 0, HER.
In distribution, a dilation by a of the time axis and a dilation by aH of the x-

axis does not change the representation of the trajectories of this process. The
90 Tautu
common example of a self-similar process is the Brownian motion with H = 1/2

and more generally, a stable process is self-similar with parameter H = l/ex, 0 <
ex S; 2. As limits, self-similar processes have to generalize the role of stable laws,
e.g., the Cauchy distribution, and processes within the limiting theory of sums of
independent random variables. Birth-and-death processes converge to self-similar
diffusions satisfying Eq. 5.2 in [36]. Undoubtedly, these limit processes are not
the processes P.M. Iannaccone meant.
(ii) Iterated function systems were introduced in [3] as a unified way of
generating a broad class of fractals. Formally, an iterated function system is a
triple {(K,d), w,p} where (K,d) is a compact metric space with distance function
d(x,y), x,y E K, and W = {Wi, 1 S; i S; n} is a finite collection of Lipschitz
functions Wi : K -+ K. Clearly, p = {Pi, 1 S; i S; n} is a set of probabilities.
In a natural way, an associated Markov chain {~n, n E N} arises by iteratively
applying the maps Wi: its initial distribution is concentrated at an arbitrarily given
x E K and its random transition mechanisms can be described as follows: given ~n,
one has ~n+l = Wi(~n) with probability Pi(~n), n E N, 1 S; i S; r. The transition
operator U associated with ~n is given by
Uf(x) = Lpi/[Wi(X)], x E K,
i=l
for any f E C(K), where C(K) is the space of all bounded measurable real-valued
functions defined on K [26].
Actually, the iterated function systems represent the basis of a probabilistic
algorithm for the computation of images but not the description of a (biological)
growth process! One must remind the term «growth of figures» introduced in 1962
by S.M. Ulam: by numerical experiments many different (fractal) patterns can be
produced, i.e., figures «growing» according to certain recursive rules. Yet, we were
warned: the amount of information contained in these objects is «therefore quite
small, despite of their apparent complexity and unpredictability» [55]. It is not
surprising that iterated function systems generate fractal forms: both K. Falconer
[17, p. 113 and Th. 9.1] and G.A. Edgar [16, p. 105] begin to specify invariant sets
by introducing iterated function schemes (or systems). In a beautiful dissertation
on modelling marine organisms by iterative geometric constructions [32], some
realistic restrictions are introduced, e.g., the exclusion of multiple occupancy of a
site and the influence of the environment on the generator rules. Such restrictions
will obviously disturb the self-similarity.
A particular form of iteration is the stochastic iteration of stable processes, in-
troduced in [2], as a generalization of the notion of branching process. For instance,
self-annihilating branching processes, continuous state space branching processes,
state dependent branching processes, etc. can be constructed as stochastic iterates
of a sequence of i.i.d. random walks with particular properties.
3 Spatial growth models of cell systems

3.1 Definition and examples
Let us begin with a brief definition of interacting cell systems on a regular lattice
Zd, d ? 1. For technical details the reader is referred to [39]. Let 5 be the state
space 5 = {all subsets of Zd} and set :=: = 5 Zd which is the state space of the
stochastic process. The elements ~ of :=: are called configurations and the elements
x of Zd are called sites. The evolution of the interacting system {~(t) , t ? O} is
determined by its jump (flip) rates c(x, 0 which possess the properties of (i) finite
range, (ii) translation invariance, and (iii) attractiveness.
Definition
A family {~(t)} of S-valued Markov processes is called a stochastic growth model
if 0 is an absorbing state and the family is a FTA system (i.e. a finite range, trans-
lation invariant, attractive system of infinitely many interacting particle system).
The interacting biological cell systems discussed in this papers are generally
one-site (<<seed») growth processes which also are permanent (irreversible, immor-
tal, or supercritical), i.e. P{ T = oo} > 0, where T = min{t : ~? =I- 0} (see also
[52]).
In [13], R. Durrett and D. Griffeath considered the following models as ex-
amples of stochastic spatial growth processes: (1) the Richardson model, (2) the
coalescing random walks with nearest neighbour births, (3) the Williams-Bjerknes
models, and (4) the basic contact process. Lately, an epidemic model [11], the
oriented site percolation and a process with sexual reproduction [12] are included.
3.2 Microscopic growth rules

This paragraph will deal with the analysis of the primary differences between the
above cell interacting systems on Zd, d ? 1 (and other cell systems) and the
known growth models generating fractal structures. We will refer to the essential
microscopic growth rules (for details see [59]).
(i) The initial configuration. Most of cell interacting systems are one-site
growth models: the process starts with one cell at site 0 E Zd, d ? 1, or with a
small group of adjacent cells (nearest neighbours). In the Markov configuration
model [57] generating cells of different types, the process begins with one normal
cell. On the contrary, in the DLA model, particles are added one at a time to a
growing aggregate of particles via random walk (Brownian) trajectories, but these
particles start outside of the region occupied by the initial aggregate. Actually, they
are launched from a r'P1domly selected point on a circle that just enclose the cluster
[45]. However, DLA-like bacterial colonies (B. subtilis) can be obtained if the
bacteries are inoculated at the centre of a Petri dish and the peptone concentration
is up to Ig/l [42, 22]. In other experiments [66, 43], the «seed» is a line: the
inoculation of some microorganisms (E.coli, B. subtilis, Aspergillus oryzae) is
made along a straight line (<<strip geometry»).
92 Tautu
(ii) The multiplication process. Biological growth is characterized by cell

division. At each site on the lattice a birth process (or a birth-and-death process)
of a particular type takes place (remind that the jump rate is dependent on the
configuration D. In [57] the cell cycle is explicitly introduced. As a consequence
of binary fission, a daughter cell must look for a vacant site. This search is governed
by the following
(iii) Occupation rules. (10) Multiple occupancy is excluded. (2°) Occupation

of a nearest neighbour vacant site (short range): the choice can be made with
(a) equal or (b) unequal probabilities. If the configuration is compact, the cell
proliferation is restricted at the boundary. The most familiar example of an asym-
metrical cell growth system on Z2 was given by M. Eden where the probability
PI of adjoining a cell at a vertical edge can be an integral multiple of the prob-
ability P2 of adjoining to a horizontal edge. One can assume, for instance, that
an anisotropy in the nutrient medium of a tissue culture may determine config-
urations with elliptic shape. It is known that lattice anisotropy has an important
effect on the growth of DLA clusters, and that this effect is very much smaller for
hexagonal and triangular lattices than for the square lattice [45]. (3°) Competitive
occupation of a nearest neighbour site (if it is occupied). This kind of short-range
interaction is assumed in [57]; in this model a newly born cell can choose (with
equal or unequal probabilities) an already occupied site and eliminate the «old»
cell. In an interacting cell system this competition of the «fittest» may be inter-
preted as the general dynamics of «cell loss». Another kind of local competition
is assumed in the Williams-Bjerknes model, where the malignant cell possess a
«carcinogenic advantage»; in the theory of interacting particle systems, the in-
finite Williams-Bjerknes cell system is known as the «biased voter model» [39,
10]. (4°) Occupation of a distant vacant site (long range). A new cell must go on
to search for a vacant site until it finds one (spending no time at the intervening
occupied sites). The resulting migration process is called long range exclusion
process because cells can migrate very long distances in short times when a large
configuration exists [38]. In addition, the rate at which a cell at lattice site x at-
tempts to move to site y may depend on the configuration at other sites (<<speed
change»).
Except for the random walk, DLA is perhaps the simplest process (Laplacian
growth model [1]) that generates a fractal structure. The more interesting is then
the «internal» variant of DLA that generates a non-fractal structure [37]. In this
model, particles are repeatedly dropped at the origin 0 E Zd, d ?: 2; one by one,
each particle performs an independent simple symmetric d-dimensional random
walk until it «sticks» at the first site not previously occupied. The asymptotic
shape of an internal DLA configuration is an Euclidean ball [37, Th. 1].
The modest conclusion is that the above microscopic biological growth rules
suggest fine distinctions between fractal and non-fractal growth processes; how-
ever, two supplementary conditions, namely (a) the size of the observed con-
figuration and (b) the external growth circumstances, must also be taken into
consideration.
(a) For example, numerical experiments have shown that the shape of a DLA
configuration evolves from a (more or less) circular shape, characteristic of very
small clusters, to a cross-like shape of a cluster containing more than 106 occupied
sites, with an intermediate passage to a diamond-like shape if the configuration
consists of about 105 sites [45]. Also, the simulated Eden cluster shows the size
dependence of its shape: even for d = 2 very large clusters are not exactly circular.
(b) Bacteriological experiments show that the morphology of growing colo-

nies depends on the nutrient concentration, e.g. DLA-like structures at small con-
centrations of peptone (lg/l), round and smooth forms, «similar to petals which are
fused together into one», at intermediate concentrations (4g/l), and round colonies
at high nutrient concentrations (64g/l). Obviously, the structure of the support is
also important: at a low concentration of agar, the bacterial colonies have a «dense
branching morphology» [22], and, simulated on a continuum, the Eden cluster is
no more compact. The compactness of the morphology is a function of the viscos-
ity ratio TJ in the general gradient governed growth (GGG) model: for rJ = 0.0001
the realized structure is fractal, for TJ = 0.01 the structure is irregular but not
fractal, and one obtains an Eden-like structure if TJ = 10.0 [49].
3.3 The asymptotic shape (I)

Strictly speaking, the shape of a geometrical object represents those geometrical
attributes that remain invariant under any translation, rotation and scaling. The
reader is referred to [7] and [62] for applications in morphometrics (= measurement
of shapes, their variation and changes). Particularly interesting in [62, Ch. 3] is
the application of Dirichlet principle which may supply a quantitative definition of
the term «smooth». (See the definitions of smoothness and roughness in [7] with
the analytic approximation.)
This paragraph is motivated by the following simple question: If the random
walk is the simplest example of a random fractal, what about its shape? An an-
swer is given in [53] where the results were obtained in the form of a power
series expansion, called lid expansion. A different, probabilistic approach will be
presented in the sequel.
Let {Xi , I :::; i :::; n} be a sequence of independent, identically distributed,
d-dimensional random variables (d > I) with a common distribution F and, as
usual Sn = 2:£= 1 Xi. We say that Sn is the position at instant n of a particle
performing a d-dimensional random walk. In order to obtain asymptotic estimates
for the probability that Sn is contained in a cube (or a ball) centered at the origin,
another random variable X with the same distribution as Xl is introduced. If X
satisfies two special geometrical conditions, namely the direction and the cone
conditions, and if it is, in addition, genuinely d-dimensional (i.e. its distribution
is not supported on a (d - I)-dimensional hyperplane), then there exist positive
94 Tautu
constants Cl, C2 and AO such that for all A ~ AO and all n,
(1)
where C(O, A) is the cube of side length 2A centered at the origin [25, Th. 2.1].
This is a consequence of a local limit theorem in the case that the additional
random variable X is in the domain of attraction of a stable law.
According to W. Feller [21], the variables Xi are said to belong to the domain
of attraction of a nondegenerate distribution G if there exist real constants an > 0
and bn such that P { (Sn - bn)/ an ::; x} -+ G (x), as n grows to infinity. It is
important to point out that the localization of the variable X in the domain of
attraction of a stable law with index a, 0 < a ::; 2, means that there is a centering
sequence {b n} such that {( Sn - bn)/ an} is stochastically compact [21], that is, it is
a tight sequence and all limit laws are nondegenerate. Typical local limit theorems
depend on compactness. The sequence {an} is used, when suitably chosen, to
normalize {Sn} for weak convergence.
Two additional remarks must be made: (a) the problem is dependent on di-
mensionality because in d ~ 2 the spread of {Sn} may vary greatly in different
directions and the centering task becomes more complicated, and (b) the theorem
(1) does not apply to all variables in the domain of attraction of a stable law [25].
Remind that the normal distribution (a = 2) is the only member of the class
of stable laws that has a finite second moment: for all other members of the class,
the calculations using characteristic functions, show infinite variance (see, e.g.,
[24D. The characteristic exponent a measures the «thinness» of the tails of the
distribution and also indicates the existence and nature of the moments. A small
value of a will mean the setting of considerable probability mass in the tails
of the distribution and thus, a long-tailed distribution implies a distribution with
infinite variance. It is known that if a ! 0, the Levy distribution has the lognormal
distribution as limiting form.
It is very important to point out that special random walks which generate self-
similar clusters [30] possess an infinite mean-squared displacement per jump and
their distribution is long-tailed. This might be the mathematical characterization
of typical random fractal growth processes.
A similar problem can be formulated for the Galton-Watson branching process
{Zn, n E N}. If the process is supercritical and the basic distributions {Pi}, i EN,
are normally attracted to a particular stable law, then there exist constants an , bn, Cn,
for which the limit
. P {Zn - bn ::; x Zo =
hm Cn } = G(x) (2)
n-+oo an
holds with G as the corresponding stable law [35, Coroll. 3].
More adequate in a biological context is the question about the asymptotic
shape of a branching random walk. If I(n) is the set of positions of the n-th
generation of individuals generated by a supercritical branching random walk on

Rd, scaled by 1In, then I(n) looks like a convex set I for large n, provided that the
process survives. Then I is the asymptotic shape of this branching random walk,
and according to Theorem A in [6], it is a smooth rounded shape. Moreover, J.D.
Biggins proved that the upper bound C on the asymptotic shape r of Mollison's
contact birth process equals this shape, r = C. This leads to the conclusion that the
propagation of a monotone continuous time Markov contact process in a particular
direction is bounded: the limits for individual direction can be «stuck together»
into a convex body r such that the convex hull Ht of inhabited points converges
in shape and velocity (Eq. 3.29 in [46]). This result can be correlated with the
observation of M. Eden that in his spatial model for embryogenesis configurations
with many short branches are more probable than those with a few long branches.
3.4 The asymptotic shape (II)

This paragraph is devoted to the presentation of the asymptotic geometrical be-
haviour of known interacting cell systems by introducing the essential arguments
derived from the theory of subadditive stochastic processes. Indeed, the main idea
which dominates this approach is that the spread times (occupation of a site, in-
fection, etc.) are subadditive. The interesting result is that the set of occupied sites
grows linearly with t (possibly at an infinite rate) and has an asymptotic shape
which is not random [34] (see also [10, llc]). It must be pointed out that the
argument behind the proof of the asymptotic shape theorem is the evidence of the
linear growth in a fixed direction; this implies the right growth rate in any finite
number of directions simultaneously, and therefore the requirement of some uni-
form estimate on different growth rates in two directions which are close together
(Lemmas 3.5 and 3.6 in [33]). The problem of self-affinity for such interacting
systems is then solved.
The asymptotic shape of all interacting cell systems is investigated in the same
framework, e.g., the Richardson model [14], the Williams-Bjerknes model [8], and
the Markov configuration model [57]: they may be considered as asymptotically
non/ractal cell growth models. Actually, D. Richardson was the first author who
compared the Eden model (his Example 9) with the Williams-Bjerknes model with
infinite carcinogenic advantage (his Example 6). In [56] K. Schiirger introduced
two generalized Eden models and demonstrated a similar asymptotic geometric
behaviour. This generalized Eden model is a discrete time interacting cell system
with state space 2 the set of all configurations; 20 is the set of all configurations
rJ E 2 having a finite non-void support. Because this model was frequently in-
vestigated especially for its nontrivial surface (see a collection of papers in [19]),
we include here the probabilistic approach of its asymptotic geometric behaviour.
The Eden growth process rJ~k) E 20 chooses at random one of the sites (say y) on
its k-boundary fh(rJ~k)). This choice is equiprobable. The process makes a jump
from state rJ~k) to state rJ~~ 1 = rJ~ k) U {y}. In order to formulate the shape theorem,
one must take into account the monotonicity condition, the formal definition of a
norm, and the hypothesis of a finite span of the process. One obtains the following
96 Tautu
Theorem
(Schiirger, 1981). Let k2: 1. Then there exists a nonn N* (.) on Rd such that, for
all ~ E 30 and 0 < E < 1, we have for the generalized Eden process {7J~k)} for
which 7Jb k ) = ~, that
{xix E Zd ,N*(x) :::; (1 - E)nl /d} C 7J~k) C {xix E Zd, N*(x) :::; (1 + E)nl /d}
(3)
almost surely for all sufficiently large n.
Such results might be thought of as a strong law of large configurations
because for all ~ E 30, the set (l / t){xIT(x) :::; t} (pointwise scalar multiplication)
converges in some sense - which is common in measure theory - a.s. (P~) to
the set {xIN(x) :::; I} as t --+ 00 . Obviously, T(X) is the first instant at which site
x is occupied [56].
The result above fonnalizes the original observation made by M. Eden that
«the colony is essentially circular in outline». The biological analogue of this
model can be, for example, the two-dimensional growth of a cell clone: «wherever
there is a nutrient medium we can be quite sure that the colony will have the largely
circular morphology exhibited by the model».
This is just the moment for two important remarks:
(1) In Richardson's paper, the probability that a site will be occupied is

denoted by p. His first remark is that as p varies from 1 to 0, the unit ball of the
associated norm N varies from a diamond to a circle (<<however, no proof of this
is available», he added). It is now possible to prove this conjecture in tenns of
F, the distribution of the passage time. One assumes that F is exponential, i.e.
F(x) = l-e- x , if x 2: O. Let A(F) = inf{x: F(x) > O} be the left endpoint of the
support of F. Also, denote by Bo C Rd a nonrandom convex set having nonempty
interior. According to a proposition given in [33, Prop. 6.15], if A(F) > 0 and F is
not concentrated on A, then this convex set is strictly contained in the «diamond»
{x E Rd : Ixl = I / A} . Moreover, for high dimensions and exponential F, it is not
a ball [33, Coroll. 8.4].
The reader should peer at Durrett's simulation of an epidemic model (or
«forest fire») in order to realize the structure changes (black burnt sites): a diamond
form (with a «flat edge» as in [14]) if P = 0.75, a tendency to circularity at
p = 0.60, and a clear fractal structure at p = 0.50 [10, pp. 239-241].
(2) D. Richardson pointed out an essential fact: as p decreases (p=O.1), the
roughness of the boundary seems to increase. In consequence, the natural norm
imposed by the lattice breaks down «and the Euclidean norm appears to assert
itself».
3.5 Crinkliness
Re-analyzing the Williams-Bjerknes model (believed by its authors to be a fractal
model), D. Mollison introduced a measure of surface irregularity
1 .
C(~) = IiCT #{ (x ,y)lx E ~ , y ~~, x and y adJacent}, d = 2, (4)
4y I~I
called crinkliness. The formula (4) implies that C(~) = 1 for the square arrays, all
other structures having C(~) > 1. It appears that for I~I = n small, the crinkliness
has a great variability, while for increasing n, C(~) tends to a limit which de-
pends upon K- , the carcinogenic advantage. It is conjectured in [57, Conj. 3.1] that
limt->oo C(~t) = a, almost everywhere (P~), in the case of the Markov configura-
tion model, where a > 0 depends on cell cycle and cell differentiation intensities.
According to the numerical experiments [57], C(O varies form 1.38 to 1.82 (the
total length of the boundary increases and the growth quotient t 2 In slowly de-
creases). There is a close connection between crinkliness and spread velocity [46].
4 The diffusion-injection fractal process

Based on his data on agricultural uniformity trials, H. Fairfield Smith suggested
in 1938 that the covariance r(r) of yield at points a distance r apart falls off as
a power of r at large distance. The same behaviour was reported by flood height,
roughness of roads and runways, yam diameter, and response from population
samples. These data stimulated P. Whittle [67] to invent a diffusion model with
particular characteristics of its moments (variance, autocovariance). Lately, the
author called it a diffusion-injection model [68] which actually is ad-dimensional
anisotropic random field (see also [64]).
Let us consider a random variable ~(x, t), x = (Xl, ... , xd), d > 1, which
may be interpreted in the agricultural example as the concentration of a nutrient
in a field (or the height of the runway, the nutrient concentration in a bacterial
culture, or the density gradient of a growth factor in a tissue culture, etc.). ~(x, t)
obeys the stochastic differential equation
(5)
This is a diffusion equation driven by a Wiener process E(X, t) which is explained

as follows: variation is «injected» continuously by the noise process and is spread
out in space by the diffusion term \7 2 ~. The strength of the field tends to be
damped by the term -a 2 ( In the original paper ~(x , t) is called a «fertility»
function and \7 2 ~ = L j [-Ix; ]
2 ~, is the physical diffusion of nutrient laterally in
the soil, tending to equalize concentration. {~( x, t)} is ad-dimensional random
field.
98 Tautu
The following characteristics of the model are of importance:
(i) The process {~(x,t)}, x E Rd, d 2: 1, t 2: 0, is stationary with respect to

n
both x and t; the input process {f( x, t is stationary in its arguments, too.
(ii) If one defines the spatial autocovariance function by
e-,,/2a
Cov[~(x, t), ~(x + y, t)] = r(r) = -4--' x,y E Rd , d 2: 1, Iy - xl = r, (6)
rrr
then, if d > 2 and Q; = 0,
r(r)=cr- d +2 , c=const, (7)
[67, Eq. 7], where r is called the lag.
(iii) The spectral density of the process {~( x, tn, i.e. the Fourier transform
of the autocorrelation A(r), can be related with the covariance r(r). Let w be the
frequency and v the wave-number vector, v = (vi, ... ,vd ), d 2: 1, these being
complementary to t and x, respectively. If {~(x, tn has spectral density function
f (v, w), then its autocovariance becomes
r(r) = (2:)d J eivr f(v)dv. (8)
Clearly, the covariance is dimension sensitive: if d 2: 2, the integral does not

converge absolutely, and if d < 2, it will diverge as Q; ! 0 due to the singularity
of f(v) for small v. This divergence corresponds to genuine non-stationarity of
the process for d < 3 (Th. 20.3.1 in [68]).
One might say that the driven diffusion process possesses a covariance r(r)
but not a variance: actually, the variance is infinite due to the fact that the diffusion
mechanism does not smooth f sufficiently powerfully. However, a mild degree of
spatial correlation in the input will also ensure that ~ (x, t) has finite variance [67].
More generally, a spectral density f(v) ex: Ivl- 2 corresponds to an autoco-
variance
r(r) ex: IrI 2- d , d > 2 (9)
[68, Eq. 20.4.3]. In this case, one demonstrates the existence of a self-similar power
law r(r) '"" cr->' with 1/2 :::; A :::; 3/2, with strong peaking in the frequency at
A = 1, and lesser peakings at the extreme values AC(I/2,3/2). Indeed, the data
give D = 1.6-1.8, estimated from block variance [9].
The above scrutiny of Whittle's model pinpoints the intrinsic conditions a
stochastic process which generates fractal structures must satisfy. A noteworthy
remark is that Whittle's process is a zero-mean generalized Gaussian process and,
in the terminology of D. Williams [69], a X harness process. Accordingly, it is
claimed that P. Whittle arrived at his process by using a serial harness approach
[27, Sect. 7].
Recently, the diffusion-injection model was mentioned as an example of spa-
tiallong memory [5]. Formally, a stationary process is said to exhibit long-range
dependence if the spectral density is
f(v) '" L 1(v)jvjl-2H, HE (I/2, 1), (10)
where L 1(-) is slowly varying for jvj ~ O.

Finally, one must mention the diffusion model reported by P. Meakin [44]: it is
assumed that the nutrient is supplied by the boundary conditions which maintain a
fixed concentration x of nutrient on a circle of lattice sites surrounding the growing
cluster. The growth probability p is proportional to some power A of x, so that if
p '" x2 , the model becomes equivalent to the dielectric breakdown model, known
as Laplacian growth model [1].
References
[1] Aharony, A.: Fractal growth. In: Fractals and Disordered Systems (A. Bunde,
S. Havlin eds), pp. 151-173. Berlin: Springer (1991).
[2] Athreya, K.B.: Stochastic iteration of stable processes. In: Stochastic Pro-
cesses and Related Topics (M.L. Puri ed.), Vol. 1, pp. 239-247. New York:
Academic Press (1975).
[3] Bamsley, M.F., Demko, S.: Iteratedfunction systems and the global construc-
tion of fractals, Proc. Roy. Soc. (London) A 399, 243-275 (1985).
[4] Bennett, C.H.: Dissipation, information, computational complexity and the
definition of organization. In: Emerging Syntheses in Sciences (D. Pines ed.),
pp. 215-233. Redwood: Addison-Wesley (1988).
[5] Beran, 1.: Statistical methods for data with long-range dependence, Statist.
Sci. 7,404-427 (1992).
[6] Biggins, J.D.: The asymptotic shape of the branching random walk, Adv.
Appl. Probab. 10,62-84 (1978).
[7] Bookstein, F.L.: The Measurement of Biological Shape and Shape Change,
Lecture Notes Biomath. 24. Berlin: Springer (1978).
[8] Bramson, M., Griffeath, D.: On the Williams-Bjerknes tumour growth model,
I, II, Ann. Probab. 9, 173-185 (1981); Math. Proc. Camb. Phil. Soc. 88,
339-357 (1980).
[9] Burrough, P.A.: Fractal dimensions of landscapes and other environmental
data, Nature 294, 240--242 (1981).
100 Tautu
[10] Durrett, R.: Lecture Notes on Particle Systems and Percolation. Pacific Grove
(CA): Wadsworth & Brooks/Cole (1988).
[11] Durrett, R. : Stochastic growth models: Recent results and open problems. In:
Mathematical Approaches to Problems in Resource Management and Epi-
demiology (C. Castillo-Chaves et al. eds) [Lecture Notes Biomath. 81], pp.
308-312. Berlin: Springer (1989).
[12] Durrett, R.: Stochastic growth models: Bounds on critical values, 1. Appl.
Probab. 29, 11-20 (1992).
[13] Durrett, R., Griffeath, D.: Contact processes in several dimensions, Z. Wahr-
scheinlichkeitstheorie verw. Gebiete 59, 535-552 (1982).
[14] Durrett, R, Liggett, T.M.: The shape of the limit set in Richardson's growth
model, Ann. Probab. 9, 186-193 (1981).
[15] Edelman, G.M.: Topobiology. An Introduction to Molecular Embryology. New
York: Basic Books (1988).
[16] Edgar, G.A.: Measure, Topology, and Fractal Geometry. New York: Springer
(1990).
[17] Falconer, K.: Fractal Geometry. Mathematical Foundations and Applications.
Chichester: Wiley (1990).
[18] Family, E: Dynamic scaling and phase transitions in interface growth, Phys-
ica A 168, 561-580 (1990).
[19] Family, E, Vicsek, T. (eds): Dynamics of Fractal Surfaces. Singapore: World
Scientific (1991).
[20] Feller, W.: The asymptotic distribution of the range of independent random
variables, Ann. Math. Statist. 22,427-432 (1951).
[21] Feller, W.: On regular variation and local limit theorems, Proc. 5th Berkeley
Symp. Math. Statist. Probab., Vol. II, Part 1, pp. 373-388. Berkeley: Univ.
California Press (1967).
[22] Fujikawa, H., Matsushita, M.: Bacterial fractal growth in the concentration
field of nutrient, J. Phys. Soc. Japan 60, 88-94 (1991).
[23] Gouyet, J.-E, Rosso, M., Sapoval, B.: Fractal surfaces and interfaces. In:
Fractals and Disordered Systems (A. Bunde, S. Havlin eds), pp. 229-261.
Berlin: Springer (1991).
[24] Granger, C.W.J., Orr, D.: «Infinite variance» and research strategy in time
series analysis, J. Amer. Statist. Assoc. 67, 275-285 (1972).
[25] Griffin, P.S.: An integral test for the rate of escape of d-dimensional random
walk, Ann. Probab. 11, 953-961 (1983).
[26] Grigorescu, S., Popescu, G.: Random systems with complete connections as
a framework for fractals, Stud. Cerc. Mat. 41,481-489 (1989).
[27] Hammersley, J.M.: Harnesses, Proc. 5th Berkeley Symp. Math. Statist. Prob-
ab., Vol. llJ, pp. 89-117. Berkeley: Univ. California Press (1967). (Reprinted
in [38])
[28] Hart, T.N., Trainor, L.E.H.: Geometrical aspects of surface morphogenesis,
J. Theor. BioI. 138, 271-296 (1989).
[29] Hatlee, M.D., Kozak, 1.1.: Stochastic flows in integral and fractal dimensions
and morphogenesis, Proc. Natl. Acad. Sci. USA 78, 972-975 (1981).
[30] Hughes, B.D., Shlesinger, M.F., Montroll, E.W.: Random walks with self-
similar clusters, Proc. Natl. Acad. Sci. USA 78, 3287-3291 (1981).
[31] Iannacone, P.M.: Fractal geometry in mosaic organs: a new interpretation of
mosaic pattern, FASEB J. 4, 1508-1512 (1990).
[32] Kaandorp, J.A.: Modelling growth forms of biological objects using fractals,
Ph. D. Thesis, Univ. of Amsterdam (1992).
[33] Kesten, H.: Aspects of first passage percolation. In: Ecole d'Ete de Proba-
bilites de Saint-Flour XIV-1984 [Lecture Notes Math. 1180], pp. 125-264.
Berlin: Springer (1986).
[34] Kesten H.: Percolation theory and first-passage percolation, Ann. Probab.
15, 1231-1271 (1987).
[35] Lamperti, J.: Limiting distributions for branching processes, Proc. 5th Berke-
ley Symp. Math. Statist. Probab., Vol. II, Part 2, pp. 225-241, Berkeley: Univ.
California Press (1967).
[36] Lamperti, J.: Semi-stable Markov processes, Z. Wahrscheinlichkeitstheorie
verw. Gebiete 22, 205-225 (1972).
[37] Lawler, G.F., Bramson, M., Griffeath, D.: Internal diffusion limited aggrega-
tion , Ann. Probab. 20, 2117-2140 (1992).
[38] Liggett, T.M.: Long range exclusion processes, Ann. Probab. 8, 861-889
(1980).
[39] Liggett, T.M.: Interacting Particle Systems. New York: Springer (1985).
[40] Mandelbrot, B.B.: Fractal geometry: what is it, and what does it do?, Proc.
Roy. Soc. (London) A 423, 3-16 (1989).
[41] Mandelbrot, B.B.: Plane DLA is not self-similar; is it a fractal that becomes
increasingly compact as it grows?, Physica A 191, 95-107 (1992).
[42] Matsushita, M., Fujikawa, H.: Diffusion-limited growth in bacterial colony
formation , Physica A 168,489-506 (1990).
[43] Matsuura, S., Miyazima, S.: Self-affine fractal growth of Aspergillus oryzae,
Physica A 191, 30-34 (1992).
[44] Meakin, P. : A new model for biological pattern formation, J. Theor. BioI.
118, 101-113 (1986).
102 Tautu
[45] Meakin, P.: Models for colloidal aggregation, Ann. Rev. Phys. Chern. 39,
237-267 (1988).
[46] Mollison, D.: Spatial contact models for ecological and epidemic spread, J.
Roy. Statist. Soc. B 39, 283-326 (1977).
[47] Ng, Y.-K., Iannaccone, P.M.: Fractal geometry of mosaic pattern demon-
strates liver regeneration is a self-similar process, Develop. BioI. 151, 419-
430 (1992).
[48] Ng, Y.-K., Iannaccone, P.M.: Experimental chimeras: Current concepts and
controversies in normal development and pathogenesis, Current Topics De-
velop. BioI. 27,235-274 (1992)
[49] Nittmann, J., Daccord, G., Stanley, H.E.: When are viscous fingers fractal? In:
Fractals in Physics (L. Pietronero, E. Tosatti eds), pp. 193-202. Amsterdam:
North Holland (1986).
[50] Othmer, H.G., Pate, E.: Scale-invariance in reaction-diffusion models of spa-
tial pattern formation, Proc. Nati. Acad. Sci USA 77, 4180-4184 (1980).
[51] Pelce, P., Sun, J.: Geometrical models for the growth of unicellular algae, J.
Theor. BioI. 160, 375-386 (1993).
[52] R6thinger, B., Tautu, P.: On the genealogy of large cell populations. In:
Stochastic Modelling in Biology (P. Tautu ed.), pp. 166-235. Singapore:
World Scientific (1990).
[53] Rudnick, J., Gaspari, G.: The shapes of random walks, Science 237, 384--389
(1987).
[54] Schaefer, D.W., Bunker, B.C., Wilcoxon, J.P. : Fractals and phase separation.
Proc. Roy. Soc. (London) A 423, 35-53 (1989).
[55] Schrandt, R.G., Ulam, S.M.: On recursively defined geometrical objects and
patterns of growth. In: Essays on Cellular Automata (A.W. Burks ed.), pp.
232-243. Urbana: Univ. Illinois Press (1970).
[56] Schiirger, K.: On a class of branching processes on a lattice with interactions,
Adv. Appi. Probab. 13, 14--39 (1981).
[57] Schiirger, K., Tautu, P.: A Markov configuration model for carcinogenesis.
In: Mathematical Models in Medicine (J. Berger et ai. eds) [Lecture Notes in
Biomath. 11], pp. 92-108. Berlin: Springer (1976).
[58] Stanley, H.E.: Fractals and multifractals: The interplay of physics and ge-
ometry. In: Fractals and Disordered Systems (A. Bunde, S. Havlin eds), pp.
1-49. Berlin: Springer (1991).
[59] Tautu, P.: On the qualitative behaviour of interacting biological cell systems.
In: Stochastic Processes in Physics and Engineering (S. Albeverio et ai. eds),
pp. 381-402. Dordrecht: Reidel (1988).
[60] Tautu, P.: Interacting biological cell systems. In: Stochastic Modelling in
Biology (P. Tautu ed.), pp. 50-90. Singapore: World Scientific (1990).
[61] Todd, P.H.: Gaussian curvature as a parameter of biological surface growth,
J. Theor. BioI. 113,63-68 (1985).
[62] Todd, P.H.: Intrinsic Geometry of Biological Surface Growth. Lecture Notes
Biomath. 67. Berlin: Springer (1986).
[63] Tsonis, AA: Chaos. From Theory to Applications. New York: Plenum Press
(1992).
[64] Vecchia, A Y.: A general class of models for stationary two-dimensional ran-
dom processes, Biometrika 72, 281-291 (1985).
[65] Vicsek, T.: Fractal Growth Phenomena. Singapore: World Scientific (1989).
[66] Vicsek, T., Cserz6, M., Horvath, Y.K.: Self-affine growth of bacterial colonies,
Physica A 167, 315-321 (1990).
[67] Whittle, P.: Topographic correlation, power-law covariance functions, and
diffusion, Biometrika 49,305-314 (1962).
[68] Whittle, P.: Systems in Stochastic Equilibrium. Chichester: Wiley (1986).
[69] Williams, D.: Some basic theorems on harnesses. In: Stochastic Analysis
(D.G. Kendall, E.F. Harding eds), pp. 349-363. London: Wiley (1973).
[70] Yates, F.E. : Fractal applications in biology: Scaling time in biochemical net-
works. In: Numerical Computer Methods (L. Brand and M.L. Johnson eds)
[Methods in Enzymology 210], pp. 636-675. San Diego: Academic Press
(1992).
Evolutionary Meaning,
Functions and Morphogenesis of
Branching Structures in Biology
Giuseppe Damiani
I.D.Y.G.A.-C.N.R.
c/o Dipartimento di Genetica e Microbiologia
Universita di Pavia
via Abbiategrasso 207
27100 Pavia, Italy
Abstract. A causal explanation of the regularity and invariance of biological forms must relate to
morphogenetic generative laws. to allometric constraints of growth processes, to biunique relations
of an organism with environment and other organisms, to the pattern of evolutionary processes, and
to physical causes as the forces present in the environment and the mechanical properties of the
available building materials. The evaluation of the relative importance of these different causes is not
always easy. Are evolutionary trends of organic forms the results of functional adaptation or of random
processes? A general model to explain the ontogenetic and phylogenetic upward trend of complexity
of biological branching structures is proposed. The function of these structures in living organisms is
to distribute or to gather biological material and physical entities. Optimization processes produced
by natural selection led to minimization of the networks length (the Steiner problem) and of the
amount of information needed for the construction of these structures. What is the optimization method
chosen by nature? Experimental data and computer simulations suggest a simple way to generate
minimized network: sensitive entities spread or concentrate in a surface or in a volume responding to
morphogenetic gradients of physical forces or of chemical substances. Recursive local rules of repulsion
or attraction leads to the formation of fractal , branching, vascular and global networks. Slight changes of
morphogenetic fields and of responsiveness of sensitive entities influence shapes of produced patterns.
Natural selection allows survival of the biological structures more suitable for their physiological
functions: to distribute or to gather something in an economic (cost functions minimization), uniform
(space-filling) and size-independent (self-similar scaling) way. Therefore biological branching structures
are the results of functional adaptation processes. Computer simulations by means of genetic algorithms
based on the same optimization method chosen by nature may be applied successfully to solve a wide
variety of scientific and engineering problems.
1 Introduction
The aim of scientists is to discover and to explain regularities in observations
of natural phenomena. These regularities can be described by natural laws which
are expressed in mathematical terms. Euclidean geometry and analytical calculus,
typically in the form of linear differential equations, are the standard way of mod-
elling the laws of nature. These mathematical models allowed the formulation of
laws describing simple physical mechanic systems, but complex systems, as the
biological ones, often remained unexplained. Moreover many patterns and shapes
of natural objects and phenomena cannot be represented by Euclidean geometry.
Recently, with the advent of computers, a new kind of mathematics was devel-
Evolutionary Meaning, Functions and Morphogenesis 105
oped which deals with discrete entities rather than continuos sets of points and
with approximate numerical solutions of both linear and non linear equations rather
than exact analytical solutions of linear equations. In particular B.B. Mandelbrot
developed the fractal geometry which describes shapes with similar properties
at different scales (Mandelbrot, 1975). The mathematical community debates the
value of these new discrete models which are mainly deduced through computer
simulations instead of a series of theorems and proofs. Some traditional mathe-
maticians believe fractals are just pretty pictures. Mandelbrot replies that «pretty
pictures in the appropriate minds lead to pretty problems and entire new fields».
Undoubtedly computer simulation and fractal geometry are very useful tools to
understand structures and dynamics of complex systems generated by simple it-
erative rules. A wide range of natural structures have been quantitatively char-
acterised using the idea of a fractal dimension (Mandelbrot, 1982). Our universe
is a fractal (Damiani, 1984; Maddox, 1987) with fractals everywhere (Barnsley,
1988). In particular many biological structures have fractal, branching and vascu-
lar morphologies: trees, venation of leaves, lungs and in general most circulatory
and transport systems in living organisms, suture lines of ammonoids, aggregator
structures produced by myxobacteria and myxomyceta, cromatophores of some
animals, nervous cells and many others. As written recently, «the mathematical
concept of fractal scaling brings an elegant new logic to the irregular structure,
growth and function of complex biological forms» (West & Goldberger, 1987).
2 Structure and function

The formation of structures and shapes in living organisms has long been consid-
ered a mysterious and supernatural phenomenon. When the problem was examined
by Darwin (1859) in the context of the evolutionary theory, the forms of the living
organisms has been explained largely in terms of adaptation produced by natural
selection. A natural or artificial selection process requires the following steps: (1)
reproduction of discrete units; (2) occurrence of heritable differences among these
units; (3) different rate of reproduction of the different units. Individuals with
the best fitness to a certain environment have the greatest probability to generate
progeny. Biotic and physical factors present in a certain environments are respon-
sible for selective reproduction of living organisms. Selection is cumulative: each
generation is build on the successful adaptations of the generation before. There-
fore a structural and functional optimization of biological structures is achieved
by evolutionary processes. The strong correspondence between form and function
supports the neo-Darwinian adaptationist explanation for the biological shapes.
However, forms are not only moulded by functional selection, but are subjected to
genetic, developmental and physical constraints (Russel, 1916). If functionalism
presupposes the primacy of function over form, on the other hand, structuralism
emphasises the primacy of form over function (Rieppel, 1988). Internal factors
might be more important than external selection. Only few morphological trans-
formation are permitted by the available genetic information and developmental
106 Damiani
pathway. Only few types of genomic reorganization are viable and economic solu-
tions. Moreover relationships between different structures of an organism tend to
vary during its growth. These genetic and structural constraints limit the number
of the possible morphogenetic changes which are the substrate for adaptive se-
lection. A compromise solution between the functionalistic and the structuralistic
viewpoints was proposed: environment may create form, but only within the limits
of the types created by ontogeny (Alberch, 1980).
3 Functional adaptation or random choices?

Some researchers reject the adaptationist program of neo-Darwinian biology
(Gould & Lewontin, 1979). For Lewontin evolution is a very complex phenomenon
and adaptation is insufficient to explain most of the patterns of change: «Genes,
organisms, and environments are in reciprocal interaction with each other in such a
way that each is both cause and effect in a quite complex way» (Lewontin, 1983).
For Stephen Jay Gould non-adaptive processes, acting by differential origin and
extinction of species, direct evolutionary trends within clades (macroevolution),
while natural selection, acting by differential birth and death of individual, di-
rects evolutionary change only within populations (rnicroevolution): «Two classi-
cal trends are the complexification of suture lines between chambers in ammonite
shells and increasing symmetry of the cup in Palaeozoic crinoids. If we explore
the speculative explanations offered to account for them we find that these ex-
planations presuppose a direct advantage for some particular morphology under a
phyletic regime of natural selection. The complex sutures strengthen the ammonite
shell, for example, or perfection of radial symmetry permit sessile organisms like
crinoids to collect food from all directions. We also find that such explanations
have been notoriously unsuccessful, a plethora of contradictory speculations that
never brings resolution. Perhaps the problem lies deeper than our failure to devise
a good adaptive story. The geometry of punctuated equilibrium means that we
cannot extrapolate natural selection within populations to produce evolutionary
trends because trends are a product of the differential fate of species considered
as stable entities. Trends may occur simply because some kinds of species speci-
ate more often than others, not because the morphologies so produced have any
advantages under natural selection (indeed, such a trend will occur even if extinc-
tion is completely random).» (Gould, 1982). Raup and Gould (1974) showed with
stochastic simulation of evolutionary processes that many morphological patterns
of apparent order arise at unexpectedly high frequency in random models. Are
macroevolutionary trends simply extrapolation of microevolutionary events? Are
internal factors more important than external selection? Are symmetry and regu-
larity of natural forms and patterns the results of adaptive processes or of random
fluctuations? Are there simple universal laws that allow us to comprehend them?
It is difficult to answer these general questions but in particular cases it is possible
to search a solution. One of these cases is the evolution of biological branching
structures.
4 The optimal form

D'Arcy Thompson (1942) tried to identify the most important mathematical and
physical forces which moulds the form of organisms so as to assure their func-
tional adaptation: «Few had asked whether all the patterns might be reduced to
a single system of generating forces, and few seemed to sense what significance
such a proof of unity might possess for the science of organic form.» If some
general laws of organic form exist we must search for them in universal patterns
which are evolved independently in different groups of organisms. Many structures
constructed by growing organisms converge repeatedly on a limited number of ar-
chitectural designs (Hildebrant & Tromba, 1985). For example the analysis of the
organizational properties of animal skeletons in a theoretical morphospace suggest
that «organic structure must necessarily approach recurrent elements of design.»
(Thomas and Reif, 1993). Undoubtedly the fractal, branching and vascular mor-
phologies are very common biological structures. Many researchers have studied
the geometrical and topological properties of these structures which were usually
described by centrifugal or centripetal schemes grouping the different bifurca-
tion into different orders (Weibel, 1963; MacDonald, 1983). Quantitative analysis
of several parameters, as branch length, branch diameter, and bifurcation angle,
demonstrates the presence of general laws and regularities. For example in human
lung the diameter of a branch and the length to the branch from the base of the
trunk are correlated according to an approximate power law (Weibel, 1963). Some
of these laws indicate the geometrical self-similarity of the biological branch-
ing structures. Three main processes need to explain regularity and invariance of
biological structures: (1) compatibility with organism growth, (2) physical opti-
mization in relation to their physiological functions, and (3) minimization of the
energetic cost for their construction and maintenance.
5 Scale invariance during growth

Galileo was one of the first to describe the allometric problem. The change of
scale during growth is one of the most important structural problems of pluricel-
lular organisms (McMahon, 1973). Many physiological processes are influenced
by the size. For example the mass and the nutritional requirements of an organism
increases as the cube of length but surface area only as the square. The problem
is that exchanges of matter and energy between a living organism and its envi-
ronments depend on its surface area. Evolution solved this scaling problems by
means of fractal geometry. A fractal structure is endowed with scale-invariance:
it reveals the same pattern, more or less ordered, at different scales when exam-
ined with magnifying lenses of different strength. The exponent of the allometric
equations of Tessier (1931) and Huxley (1932) is the fractal dimension D of Man-
delbrot (1975) and describes the self-similarity at different scales of the biological
structures and processes. The dimensional analysis explains the meaning of the
value of this exponent. In biological branching and vascular structures this value
108 Damiani
has a spatial meaning. It is between 1 and 3 indicating how a fractal pattern fills a
plane (1 <D<2) or a space (2<D<3). For example the fractal dimension of human
lung is approximately 2,2. An extensive review of the studies on the relationships
between size and shape in biology was written by McMahon and Bonner (1983).
6 Physiology of transport systems

The function of branching structures in living organisms is to distribute or to
gather biological material (as lymph, blood, etc.) or physical entities (as light, air,
bending stresses, etc.). Therefore optimization processes produced by natural se-
lection led to minimization of their length and of their resistance to the flow of the
transported entity and to maximizatation of their resistance to mechanical stress.
Gutfraind and Sapoval (1993) studying the Laplacian potential around irregular
object show that «the fractal geometry can be the most efficient for a membrane
or electrode that has to work under very variable conditions». Generally the prob-
lem is to transport something from a plane (D = 2) or a volume (D = 3) to a
point (D = 0) or viceversa. What is the shortest network of line segment inter-
connecting an arbitrary set of points? This mathematical problem, known as the
Steiner problem, became popular in 1941, when Courant and Robbins included it
in their book «What Is Mathematics?». The solution to this problem «has eluded
the fastest computers and the sharpest mathematical minds» (Bern and Graham,
1989). An application of the Steiner problem is the construction of minimum-
spanning-tree representing the evolutionary relationships. Another variants of the
Steiner problem is the construction of a rooted tree to connect a fixed point to
a set of points uniformly spread in a limited plane or volume. The biological
branching structures are good solutions of this problem. Lungs and plants are two
paradigms of the biological transport systems showing their physiological mean-
ing. As Leonardo da Vinci observed: «The total amount of air that enters the
trachea is equal to that in the number of stages generated from its branches, like
a plant in which each year the total estimated size of its branches, when added
together, equals the size of the trunk.» Weibel (1963) described the morphology
of the lung airways as a fractal binary tree using a centrifugal scheme. McMahon
(1975) and McMahon and Kronauer (1976) have highlight the mechanical prin-
ciples underlying the branching patterns of trees. Moreover the evolution of land
plants has been computer-simulated considering some characteristics of branching
patterns and several parameters for quantifying the physiological advantages of
the different morphologies (Niklas, 1986). Three characteristics define a universe
of possible branching patterns: probability of branching, branching angle and rota-
tion angle. The selective parameters considered are: the light gathering ability, the
resistance to mechanical stresses and the capacity to shade neighbouring plants or
part of itself. The simulated evolutionary trends are in good agreement with trends
found in the fossil record (Niklas, 1986).
7 The ammonoid shell

When a special geometry is required to solve particular physiological functions it is
difficult to understand the relationship between function and form. For example the
meaning of the evolution of ammonoid septa and their associated suture lines is not
obvious. The ammonoid phragmocone was a rigid floating structure functioning
like a small submarine. The functional problem is to have maximum volume with
minimum weight. The analysis of the mechanical principles involved in the static
of ammonoid shell (Seilacher, 1975; Damiani, 1986; Hewitt & Westermann, 1986;
Hewitt & Westermann, 1987) leads to the conclusion that continuous increase
in septum complexity serves to increase shell resistance to hydrostatic pressure
applied via the external shell wall (Westerman, 1975) and the body chamber (Pfaff,
1911). The problem is to gather the bending stress on a limited surface of the
external shell (D = 2) and to transmit these stresses to the perpendicular surface
of the internal septum (2<D<3). Moreover the septum is generally subjected to
different perpendicular pressure on the two sides. The ontogenetic and phylogenetic
development of septum and suture lines has improved a balanced distribution of the
deformation energy in every element of the shell. The septum is an equipotential
fractal surface. The junction of the external shell with the internal septum is
named suture line (1 <D<2) and it has a fractal geometry. Everybody can observe
the morphological analogy between the septum and the plots of equipotential lines
around a fractal cluster (Mandelbrot et aI., 1990) and between complex suture lines
and the tree-like fractals structures (Damiani, 1989). In particular, the ontogenetic
and phylogenetic development of Mesozoic ammonoid suture lines is analogous
to the development of a line discovered in the 1904 by the mathematician Koch
(Damiani, 1987; Damlani, 1989).
8 Minimization of energetic cost

Structures construction and storage of genetic information for the morphogenetic
processes have an energetic cost. Therefore optimization processes produced by
natural selection led to minimization of morphogenetic changes and of the infor-
mation amount needed for construction of the biological structures. Iteration of
a simple rule can generate very complex system. The somatic development of
an organism requires the following steps: (1) reproduction of discrete units; (2)
occurrence of differences among these units; (3) different rate of differentiation
and reproduction of the different units (Eigen and Schuster, 1979). These rules
are similar to those required for an evolutionary process: evolution works at in-
dividual level and development at cellular level. Even if ontogenesis is often a
recapitulation of phylogenesis, clearly the spatial and temporal self-organisation
of a pluricellular organism is a process more deterministic than the evolutionary
dynamics of a population. Undoubtedly computer simulations are very useful tools
to understand structures and dynamics of complex systems generated by simple
iterative rules. A branching structures is produced by the repetition of a set of
110 Damiani
simple rules called generator: every branch splits in two other branches in each
generation. The number of the branches is duplicated in each generation cycle.
Other generators producing fractal structure have some characteristics in common
with the generators producing the tree-like branching structures: the new tips are
located at the free extremity of the old tip in the tree-like lines, at an asymmetric
position in the Mandelbrot line and at the middle of the old tip in the Cesaro
line (and also in most of the other Koch-like lines). A great variety of different
structures is produced by slight changes of generation rules. It is not surprising
that recursive branching is a good metaphor for organisms evolution and for em-
bryonic development. Dawkins (1986) uses simple branching rules to simulate the
spectacular evolution and development of an unlimited number of Biomorphs in
the computer world.
9 Morphogenetic rules
The description of a biological branching structure as a tree has some limita-
tions. Asymmetrical branching and cross-linking of branches, technically known
as anastomosis, are not considered in these schemes. Moreover the structural anal-
ogy between biological branching forms and Koch-like lines do not explain the
mechanism of morphogenetic processes. The question of how cells estimate their
location within the body has preoccupied embryologists for the past century. The
concept of gradients of unknown diffusible substances, first propounded by Mor-
gan in the 1905, have been invoked as determinants of polarity in a wide variety
of organisms. In the 1952 Turing realised that a particular level of the postu-
lated gradients might determine the developmental patterns. At any given position
cells are able to feel the local concentration of a morphogenetic gradient and they
may move and/or differentiate in reply to it. In some cases cells are sensitive
to physical stresses and therefore the ontogenetic development is determined by
interaction between genetic and environmental elements. This is the case of am-
monoid septum construction: successive septa of the same shell vary their patterns
in response to the corrugations of the shell wall which produce different distri-
bution of the bending stresses (Seilacher, 1975). Oster et al. (1980) suggest that
mechanical deformation influencing both the viscosity and the elasticity of the
cellular cytogel are the major factor determining morphogenetic transformation.
The morphogenetic mechanisms suggested by Turing can explain the formation of
a wide range of patterns (Gierer and Meinhardt, 1972).
10 Cellular automata and viscous fingers

Simple computer simulations of the Turing reaction-diffusion model are based on
the so called «cellular automata». A cellular automaton is a set of interacting enti-
ties that can replicate themselves (von Neumann, 1966). The evolution of a cellular
automaton depends on the initial configuration of the cells and on the rules for the
Evolutionary Meaning, Functions and Morphogenesis III
calculation of the next state of each cell in each generation (Hayes, 1984; Wolfran,
1984). An important programme of this type is the diffusion limited aggregation
(DLA) model developed by Witten and Sander in the 1981 to simulate the den-
dritic growth (Sander, 1986; Sander, 1987). A programme similar to the DLA is
the dielectric breakdown model (DBM) by Niemeyer, Pietronero and Wiesmann
(1984) in which the tip splitting and growth are regulated by different param-
eters. These models produces branched patterns similar to the «viscous finger»
formed by a fluid when it is forced under pressure into another immiscible fluid
of higher viscosity (Nittmann, Daccord & Stanley, 1985; Nittmann, & Stanley,
1986). The experimental study of these phenomena involves an apparatus called
the Hele-Shaw cell which consists of viscous fluid confined between two parallel
plates. If a less viscous fluid is injected into the middle of the cell, it breaks up
into many branched fractal structures displacing the higher viscous fluid. Another
simple experimental apparatus for studying the viscous finger patterns is obtained
by squeezing a drop of a viscous substance between two glasses which are then
pulled off slowly (Damiani, 1984; Damiani, 1986). The attractive force among the
molecules of the viscous substance and the tendency to go toward the interface
between the substance and the air produce convergent vascular structures.
11 How evolution solved the Steiner problem

I have developed and studied many cellular automata producing different kinds of
divergent or contracting many-branched, fractal and vascular structures (Damiani,
1984). A general characteristic of these programmes is that when uniformly dis-
tributed entities are diffused or concentrated by a repulsive or attractive force, they
respectively produce divergent or contracting fractal and vascular structures. For
example the DBM simulate a diffusion process with tip splitting and DLA describe
the fractal growth by means of random aggregation. A very important aspect in
these simulated processes is that recursion of few rules produces complex multidi-
mensional lattices beginning with simple monodimensional entities or viceversa.
Similar processes and rules produce a variety of inorganic and organic branched
patterns. What are the recursive rules chosen by nature to generate minimal net-
works? Experimental data and computer simulations suggest two complementary
ways to construct Steiner networks: punctiform entities are spread or concentrate
in a surface or in a volume responding to gradients of physical forces (as bend-
ing stresses) or of chemical morphogenetic elements. In the diffusion model the
sensitive entities go away according to a repulsive force and when one entity is
too isolate it splits into two parts. In the contraction model uniformly distributed
entities converge according to an attractive force and when two entities are very
close they stick together. These simple recursive local rules leads to the formation
of fractal, branching, vascular and global networks. Slight changes of morpho-
genetic gradients and of responsiveness of sensitive entities influence the shapes
of produced patterns. Natural selection allow survival of the biological structures
more suitable for their physiological functions. Computer simulations by means
112 Damiani
of genetic algorithms (Bound, 1987) reconstruct morphological changes found in

the fossil record with reasonable accuracy.
12 Conclusion
The results of computer and mechanical simulations strongly support the hypothe-
sis that some biological forms, as the branching structures, are the most functional
design in relation to the material properties of the universe. These forms «are
topological attractors that evolution cannot avoid.» (Thomas and Reif, 1993). A
paradigm of this functional design is the complex structure of ammonoid septum
which increases the shell resistance to mechanical stresses. The adaptionist pro-
gram of neo-Darwinian biology is often sufficient to explain evolutionary trends.
When it fails perhaps the problem is our failure to devise a good adaptive story.
The optimized structures and procedures developed by natural selection may be
applied successfully to solve a wide variety of scientific and engineering problems
(Otto, 1982).
References
[1] Alberch P., Ontogenesis and morphological diversification, Amer. Zoo!., 20,
653-667 (1980)
[2] Barnsley M., Fractals everywhere (Academic Press, Inc.) (1988)
[3] Bern and Graham R.L., The shortest-network problem, Scientific American,
60, 66--71 (1989)
[4] Bound D.G., New optimization methods from physics and biology, Nature,
329,215-219 (1987)
[5] Courant R. and H. Robbins, What is mathematics?, (Oxford University Press),
(1941 )
[6] Damiani G., II gioco della vita, (E.I.A.ed., Roma) (1984)
[7] Damiani G., Significato funzionale dell'evoluzione dei sette e delle Ii nee di
sutura dei nauliloidi e degli ammonoidi, In: Fossili, Evoluzione, Ambiente
(Pergola, Italy) 1, 123-130 (1984)
[8] Damiani G., Simulations of some ammonoid suture lines, In: Fossili, Evolu-
zione, Ambiente (Pergola, Italy) 2, 221-228 (1987)
[9] Damiani G., I frattali e Ie linee suturali delle ammoniti, Le Scienze, 245,
60-68 (1989)
[10] Darwin C., On the origin of species (Murray J. ed., London) (1859)
[11] Dawkins R., The blind Watchmaker (Longman) (1986)
[12] Eigen M. & Schuster P., The Hypercyc1e, (Springer Verlag) (1979)
[13] Gierer A & Meinhardt H., A theory of biological pattern formation, Kyber-
netik, 12,30-39 (1972).
[14] Gould S.J., The meaning of punctuated equilibrium and its role in validat-
ing a hierarchical approch to macroevolution. In: Prespectives on evolution,
(Milkman R. ed., Sinauer) 83-104 (1982)
[15] Gould S.J. & Lewontin RC., The spandrels of San Marco and the Panglossian
paradigm: a critique to the adaptationist programme, Proc. Roy. Soc. Lond.
Ser. B., 205, 581-598 (1979)
[16] Gutfraind Rand Sapoval B., Active surface and adaptability of fractal mem-
branes and electrodes, submitted to Journal de Physique I (1993)
[17] Hayes B., The cellular automaton offers a model of the world and a world
unto itself, Scient. Am., 250, 10-16 (1984)
[18] Hewitt, R.A & Westermann G.E.G., Function of complexly fluted septa in
ammonoid shells. I. Mechanical principle and functional models, N. Jb. GeoI.
Palaont., Abh., 172, 47-69 (1986).
[19] Hewitt, RA. & Westermann G.E.G., Function of complexly fluted septa in
ammonoid shells. II. Septal evolution and conclusions, N. Jb. GeoI. Palaont.,
Abh., 174, 135-163 (1987)
[20] Hildebrant S. & Tromba AI., Mathematics and optimal form (Freeman, New
York) (1985)
[21] Huxley, J.S., Problems in Relative Growth (Methuen, London) (1932)
[22] Lewontin RC., Gene, organism and environment, In: Evolution from mole-
cules to men (Bendall D.S. ed., Cambridge University Press) (1983)
[23] Koch H. von, Sur une courbe continue sans tangente, obtenue par une con-
struction geometrique elementaire, Arkiv for Matematik, Astronornioch Fysik,
1,681-704 (1904)
[24] MacDonald N., Trees and networks in biological models, (Wiley J. and sons
ed.) (1983)
[25] McMahon T.A, Size and shape in biology, Science, 173, 1201-1204 (1973)
[26] McMahon T.A, The mechanical design of trees, Scientific American, 233,
92-102 (1975)
[27] McMahon T.A & Kronauer R.E., Tree structures: deducing principles of
mechanical design, J. theor. BioI., 59,443-466 (1976)
[28] McMahon T.A & Bonner J.T., On size and life (Freeman, New York) (1983)
[29] Maddox J. The universe as a fractal structure, Nature, 329, 195 (1987).
[30] Mandelbrot B., Les objets fractals: forme, hasard et dimension, Flammarion,
Paris (1975)
114 Damiani
[31] Mandelbrot B., The Fractal Geometry of Nature (Freeman, San Francisco)
(1982)
[32] Mandelbrot B. & Evertsz C.J.G., The potential distribution around growing
fractal clusters, Nature, 348, 143-145 (1980)
[33] Morgan T.H., The pysical basis of heredity (Lippincott Co., Philadelphia)
(1919)
[34] Neumann von 1., Theory of self-reproducing automata, (University of Illinois
Press, Urbana) (1966)
[35] Niemeyer, Pietronero & Wiesmann, Fractal dimension of dielectric break-
down, Phys. Rev. Lett., 52, 1033-1036 (1984)
[36] Nittmann 1., Daccord G. & Stanley H.E., Fractal growth of viscous fingers:
quantitative characterization of a fluid instability phenomenon, Nature, 314,
141-144 (1985)
[37] Nittmann J. & Stanley H.E., Tip splitting without interfacial tension and
dentritic growth patterns arising from molecular anisotropy, Nature, 321,663-
668 (1986)
[38] Niklas KJ., Computer-simulated plant evolution, Scientific American, 254,
68-75 (1986)
[39] Oster G.F., Odell G.M., Alberch P. & Burnside B., The mechanical basis of
morphogenesis, Devei. BioI., 85, 446-462 (1981)
[40] Otto F., Natiirliche Konstruktionen, (Deutsche Verlags, Stuttgart) (1982)
[41] Pfaff, E., Uber Form und Bau der Ammonitensepten und ihre Beziehungen
zur Suturlinie, Iber. nieders. geol. Vers., 4, 207-233 (1911)
[42] Raup D.M. & Gould. SJ., Stochastic simulation and evolution of morphol-
ogy - towards a nomothetic paleontology, Systematic Zoology, 23, 305-322
(1974)
[43] Rippel O.C., Fundamentals of comparative biology (Birkhauser Verlag, Basel)
(1988)
[44] Russel E.S., Form and function (Murray J. ed., London) (1916)
[45] Sander L.M., Fractal growth processes, Nature, 322, 789-793 (1986)
[46] Sander L.M., Fractal growth, Scientific American, 256, 82-89 (1987)
[47] Seilacher A., Mechanische Simulation und funktionelle Evolution des Am-
monitenseptums, Paleont. Z., 49, 268-286 (1975)
[48] Tessier G., Recherches morphologiques et physiologiques sur la croissance
des insectes, Travaux de la station biologique de Roscoff, 9, 27-238 (1931)
[49] Thomas R.D.K. & Reif W.E., The skeleton space: a finite set of organic
designs, Evolution, 47, 341-359 Thompson D' Arcy W., On Growth and Form,
(Cambridge University Press) (1917)
[50] Turing AM., The chemical basis of morphogenesis, Philosophical Transac-

tions of the Royal Society, B237 (1952)
[51] West B.J. & Goldberger AL., Physiology in fractal dimensions, American
Scientist, 75, 354-365 (1987)
[52] Weibel E.R. (1963) Morphometry of the human lung (Springer, Berlin) (1963)
[53] Wetermann G.E.G., Model for origin, function and fabrication of fluted cepha-
lopod septa, Paleont. Z., 49, 235-253 (1975)
[54] Witten T.A & Sander L.M., Diffusion limited aggregation, a kinetic critical
phenomenon, Phys. Rev. Lett., 47, 1400-1403 (1981)
[55] Wolfran S., Cellular automata as models of complexity, Nature, 311,419-424
(1984)
Relationship Between the Branching Pattern
of Airways and the Spatial Arrangement of
Pulmonary Acini - A Re-Examination from
a Fractal Point of View
Hiroko Kitaoka 1) and Tohru Takahashi2)
I) Department of Internal Medicine, Kitaoka Hospital, Kurayoshi 682, Japan

2) Department of Pathology, Institute of Development, Aging and Cancer, Tohoku University,
Sendai 980, Japan
Abstract. The upper lobe of the left lung, surgically removed from an adult for a small carcinoma,
was subjected to serial slicing and computer assisted 3-D reconstructionof airways and acini. A fractal
dimension of the airways of about 1.74 was obtained applying a 3-D box-counting to the graphics data.
At the same time, volumetry was perfonned of the acinus, the structural and functional unit of the lung,
using the same software. The volumes obtained from a total of 130 acini proved to be comparatively
unifonn, showing a normal type distribution with a mean of 173 ± 38 mm 3 • The volume of acini had
no correlation with their location in the lung or with the generation number of the tenninal bronchioles
supplying them, and this was considered to be reflecting a homogeneous 3-D arrangement of acini
in the organ. Furthermore, a simple 3-D model of lung was introduced to analyze the relationship
between the branching pattern of airways and the volume of acini supplied by them. It was shown that
the self-similarity of the spatial distribution of airways is consistent with a volumetrically homogenized
space division of the lung.
1 Introduction
The structure of the human lung can be simplified as a finely branching tree, with
the principal bronchus corresponding to its stem. Two parts are discriminated in
this tree: the airways and the air spaces. The airways are thick segments composing
the conducting part of the tree, starting at the principal bronchus and gradually
thinning as the segments divide several times over the bronchi and bronchioli
until they end at the terminal bronchioles (TBs). Division further continues, but
the area of lung tissue belonging to one TB is called an acinus (Fletcher et al .
1957). The acinus forms a respiratory part of the lung and serves as a structural
and functional unit. The lung consists of a vast number of acini that are packed
in the space according to some rule. Thus, there are two aspects in the structure
of the lung: a tree, and at the same time, an aggregation of unitary bodies, both
closely correlated in the organ formation.
The concept of the diameter exponent was proposed about one century ago by
Thoma (1901), and recently referred to by Mandelbrot (1982) as a fractal property.
In a branching ductal structure, this allows one to estimate the flow Q through a
branch of d in diameter by
(1)
Relationship Between the Branching Pattern of Airways 117
where n is diameter exponent and C is a constant. When the branching pattern

is assumed to be a symmetric regular dichotomy: a parent branch divides into
two daughter branches, each having the same length and diameter, this equation
is equivalent to the following:
(2)
where d z is the diameter of the branch in generation z. Weibel (1963, 1991) stated
that the airway diameter was accordant with the above equation in which n was
3, in the range of z from 2 to 10, with the trachea being the Oth generation.
The diameter exponent of human systemic arteries n was estimated at 2.6-2.7
by Suwa and Takahashi (1971). Suwa (1981) attempted to induce this value of
n from a polyhedral space division point of view. Here he assumes that all the
supracellular structures of organs are subordinate to a particular space division
which he designated as «equilibrium space division», implying that it minimizes
the potential energy of the constituent units in a field of mechanical force. Aside
from his statements, the equation concerning the diameter exponent is likely to
express at least the equivolumic space division by the branching structure. It may
be reasonable to assume that when terminal branches have an equal diameter, the
flow of air or blood is equal and, therefore, also the volume of tissues supplied
by the branches must be equal.
In this article, we present an attempt at correlating the branching pattern of
human airways with the volume of acini, by performing 3-D reconstruction from
serial slices of a human lung. With an additional analysis of 3-D airway models,
discussion will be extended to the relationship between the branching pattern and
the space division of lung.
2 The spatial distribution of the airways
2.1 Preparation of a lung specimen

Our investigation of the 3-D structure of lung was greatly facilitated by introducing
Heitzman's fixation of lung (Heitzman, 1973). This technique allows a fixed lung
specimen to retain an inflated state without being collapsed even in dry conditions.
The lung specimen thus fixed was cut into serial thin slices using a microslicer, a
devise for slicing soft tissues at a constant thickness of several hundred microns.
A left upper lobe was obtained from a female, aged 51, in whom lobectomy
was performed for small lung cancer of the apicoposterior segment. Besides the
cancer, there were morphologically no abnormalities in airways or air spaces.
The resected lobe was fixed by Heitzman's method we modified, using a fixative
containing polyethylene glycol (Kitaoka and ltoh 1991, 1992). After fixing for two
days at a constant intrabronchial pressure of 25 cm H20, a block, about 5.5 x 4.5
cm, was taken from the anterior segment to study the 3-D properties of airways
118 Kitaoka and Takahashi
and acini contained. This block was sliced serially at a thickness of 0.5 mm with
a microslicer and 48 serial slices were obtained (Figure 1).
The other part of the lobe was sliced serially at a thickness of 2 mm, in
order to reproduce the airway tree of the lobe. All the slices were submitted to
radiography. The radiographs presented the internal structure of the slices as in
Figure 2. The entire airway tree from the left upper bronchus down to TBs was
followed in the slices, and the generation numbers were determined for all the TBs,
according to Weibel's definition (Weibel 1963). In the next place, the contours of
airways in each slice were inputted into a microcomputer (PC-9801; NEC, Tokyo)
by digitization, using a software designed for 3-D visualization and morphometry.
Three-D images of airways were then integrated in a computer display as in
Figure 3.
2.2 Measurement of the fractal dimension for the conductive airways

The software was also designed so as to apply the 3-D box-counting method to the
data of airways thus inputted. The size of cubic box for sampling was variable, and
had a multiple of 0.5 mm, the thickness of a single slice. The software allowed a
computer to count the number of boxes containing the reconstructed airways. Three
places in the specimen were selected for the measurement of fractal dimension
(Figure 4). At one of them, a cube with an edge of 24 mm was set (Space A),
and at each of the other two, a cube with an edge of 18 mm (Spaces B and C). In
Space A, most branches proved to rise from a bronchus of sixth generation. Space
B contained a tree branching from a bronchus of 9th generation. Space C had two
airway trees.
Figure 5 shows log-log plots where a, the edge length of the sampling box
in the abscissa, is related with N(a), the number of boxes containing the airways
shown in the ordinate. There are close linear correlations in the range of edge
length from 0.5 to 6 mm. The slopes were almost the same in the three samplings,
being 1.74, 1.73 and 1.74 for Spaces A, B and C, respectively. The branching
generations of the airways contained in Space A ranged from 6 to 14. Those in
Space B from 9 to 16, larger by about three generations than in Space A. Spaces B
and C, having the same volumes (18 3 mm 3 ), significantly differed in the volume
of the airways contained: the volume in Space C was about three-fourths of that
in Space B. This is reflected in the different y - intercept in Figure 5. However,
the slopes of the regression lines for the three spaces were almost the same. Thus,
despite such local differences as revealed with conventional metric quantities like
the number of branching generation or the volume density, the airways have a
constant fractal dimension, which gives an integrated description for the spatial
distribution of airways.
It has been reported several times that the cluster-cluster aggregation in the
3-D space gives rise to having a fractal dimension of about 1.75. Meakin (1987)
estimated the fractal dimension by computer simulation. According to Weitz and
Oliviera (1984), who experimentally estimated the fractal dimension of the ag-
gregating gold colloids in the 3-D space by transmission electron microscopy, the
Fig. 1 Forty-eight serial slices from a fixed lung specimen. Thickness of each slice is 0.5 mm.
Fig. 2 Radiograph of a slice. Bar represents 5 mm.

Fig. 3 Three-D reconstructionof the airways in the computer display. The upper figure presents an
upward view from the bottom. The lower one presents a frontal view.
_~~___ t.
I'
Fig. 4 Spaces for measuring of fractal dimensions. Space A is a cube with an edge of 24 mm, and
Spaces Band C are cubes of with edges of 18 mm.
5000 A
B
C
1000
......
~
~ 100
10
0.5 1.0 4.0 9.0

a
Fig. 5 Log-log plots of the edge length of the box a, versus the number of boxes containing the
airways N(a).
fractal dimension also proved to be about 1.75. The proximity of the airway frac-
tal dimension to these suggests that there is a principle common to living and
non-living matters.
3 Volumetry of the pulmonary acini

It was confirmed in a 3-D scanning through the 48 serial slices, that there were
altogether 171 TBs. Of these, 130 had their acini wholly contained in the volume.
For all these 130 acini, the contour was defined on the slices and inputted into the
computer using the 3-D reconstructionsoftware. The volume of individual acini
was calculated on Cavalieri principle.
n 130
min 86
30 mal( 258
mean 172.7
S.D. 37.9
c
u
c
20
'0
..,...
.0
E
:J
Z 10
o~----~----~~~--~--~~~~--~--~--~~--~
o 40 80· 120 160 200 240 280
Acinar volume (mm 3 )
Fig. 6 The histogram of acinar volumes estimated.
The estimated volumes of the 130 acini showed an approximately normal

distribution with a mean of 173 ± 38 mm 3 (Figure 6). There was no significant
correlation between the volumes of acini and the generation numbers of the TBs
supplying them (Figure 7). Furthermore, in another investigation, it was proved
that the location of an acinus had nothing to do with its volume (Kitaoka and Itoh
1992). The acini were classified into three groups according to their location. Group
A comprised those facing the costal surface, Group B those facing the mediastinal
surface and Group C those having no pleural surface. As shown in Table 1, Group
A proved to differ significantly from the other groups in the generation number
of the supplying TB. The mean in Group A was larger than the other groups by
about 2 generations. This difference suggests that the TBs in the outer zone of the
lung like Group A were more distant from the central airways than in the inner
zone like Groups B and C. In the acinar volume, however, there was no significant
difference among the three groups.
These results suggest that the airway tree is designed so as to assign a constant
volume to the acini they supply. On the other hand, the branching pattern itself
has been proved to be statistically self-similar. It appears that the self-similar
branching of the airways serves, somehow, to realize the equivolumic division of
the air spaces.
240 ·•
••
··• .•
..
oS
...E •
•
E • i If
·•
T 1• • J
<II
E
:;,
160
. + 1•
0
>
'-
0
· ·· I
c:
u
<{ 80
o I
10 15 20
Generation number of T8
Fig. 7 Correlation between the number of generation of TBs and the acinar volume. The correlation
coefficient is -0.24, which is not significant.
Volume of
Group n Generation number of TBs
acini (mm 3 )
A 43 174 ± 35 16.7 ± 1.2*

B 23 176 ± 38 14.8 ± 1.6
C 61 170 ± 40 14.6 ± 1.7
Group A, acini facing the costal surface; Group B, those facing the mediastinal
surface; Group C, those not facing any pleural surfaces.
* p < 0.01 between A and B and also between A and C.
Table 1 Correlation among the location of acini, the volume of acini and the generation number of
TBs.
4 Three-D models of the branching tree and the space division

In order to give a ground for this presumption, 3-D models of airways are to be
introduced. Modell, as in Figure 8a, has a symmetric dichotomous branching.
The figure presents a view from a Z direction. Suppose that the whole space is a
cube with an edge of 24 mm and divided into 64 small cubic cells, each with an
I Generation I Diameter (mm) I Length (mm)

0 18 120
1 12.2 47.6
2 8.3 19.0
3 5.6 17.6
4 4.5 12.7
5 3.5 10.7
6 2.8 9.0
7 2.3 7.6
8 1.86 6.4
9 1.54 5.4
10 1.30 4.6
11 1.09 3.9
12 0.95 3.3
13 0.82 2.7
14 0.74 2.3
15 0.66 2.0
16 0.60 1.65
Table 2 Human airway Model A (Weibel 1963).
Generation number Diameter (mm) Length (mm)
567 4 12
8 9 10 2 6
11 12 13 3
Table 3 Diameters and lengths in the airway model.
edge of 6 mm. Let us assume that these small cubic cells correspond to pulmonary
acini. The end points of the tree are located at the centers of the cubes. Referring
to the data of the specimen used in this study and those of Model A of Weibel
(1963, Table 2), the generation numbers, diameters and lengths of the branches
were determined as in Table 3. Both the diameter and length of branches were
assumed to decrease to half after branching three times. This is nearly synonymous
to say that the diameter exponent of the airway is 3. Strictly speaking, the volumes
of cubic acini differs to a certain extent because they contain different volume of
airways. The volume of the largest acinus was calculated at 212 mm 3 , that of the
smallest at 172 mm 3 , with a mean of 198 mm 3 . However, the difference is so
little as to be disregarded. Figure 8b shows the frontal view of the space. Now, we
imagine that we cut the whole space into 48 serial slices at a thickness of 0.5 mm
and follow the same steps we did in measuring the fractal dimension in the actual
lung specimen. However, the edge length of the sampling box was limited to three
levels of 0.5 mm, 1.5 mm and 3.0 mm, because all airways in this model were
running along the edges of acini or penetrating toward their centers. As shown in
Figure 8c, the log-log plot shows a linear relation with a slope of 1.74, the same
as the fractal dimension' in the actual lung.
Model 2 was constructed so that the volumes of the acini might not be equal.
Figure 9a shows a symmetric dichotomous branching pattern except for the pairs
of terminal branches whose lengths are not equal, with the larger one three times
longer than the other. Correspondingly, different volumes were assigned to the
acini at a ratio of 1:3. In this model, log-log plot was not linear (Figure 9b).
Model 3 was proposed as an asymmetric branching model (Figure lOa). In this
model, each acinus has the same volume as in Model I, but the last three branches
were modified as in Figure lOb. The generation numbers of the terminal branches
were counted from 11 to 15. The branching pattern of airways in the actual lung
is more similar to this model than to Model 1. The result of 3-D box-counting in
Model 3 was almost the same as in Model I, giving a fractal dimension of 1.74.
In addition, another model (Modell') was proposed, changing the diameters
of branches in Modell (Fig. 11a). All the branches were given the same diameter
of 1 mm so that the diameter exponent was 00. In this model, log-log plot was
not linear (Fig. lIb). N(a) calculated in the four models are presented in Table 4.
These results seem to be suggesting that self-similarity of a branching tree is
realized when the following two conditions are satisfied: the equivolumic space
division and a diameter exponent of near by 3. The results obtained in Models 1
and 3 indicates that whether the branching routes to the end points are symmetric
or not has nothing to do with the statistical self-similarity if the two conditions are
met. In Model 3, the generation numbers of the terminal branches ranges between
11 and 15 because of asymmetric branching, though their diameters are the same.
In an asymmetric branching system, which is commonly seen in living organisms,
the generation number seems to bear little significance. This may explain why
relationship between the generation number and the diameter of airways defined
by Weibel (1963) does not apply to the range beyond the 11th generation.
Of course, the acinus in the actual lung is not cubic, and the end points of
TBs are not always located at their centers. However, these models indicate at
least the relationship between the branching pattern and the space division in a
simplified way.
1
b:
[! ,. l!J
t--
r~~ ~~
- i
... L J ....
I O~
.---- ........
- - I"'" I"'" I"'" -
11
............
"--
,....-- --
-
- -
N (0 )
10 .000
r. 0-1
Model I
1.000
0=1.74
100~ _______ ~ ______ ~ ____ ~_~
0.1 0.5 1.5 3.0 0 ("")
Fig.8 A symmetric dichotomous branching (Model I). A view from a Z direction (a) and a frontal
view (b). The whole space, a cube with an edge of 24 mm, is cut into 48 serial slices at a
thickness of 0.5 mm. The fractal dimension is estimated at 1.74 (c).
N(O.5) NO.5) N(3.0) D
Model 1 7359 1048 325 1.74

2 6563 993 397 x
3 7363 1030 3-28 1.74
l' 3339 921 325 x
Table 4 Calculated N(a) and fractal dimension D in the four airway tree models.
--
I ..
.. C!.
J..
~
L.J :-1 0_
I I
0 ...
I
,. (a)
\
s. \
\
, Model 11
,,
\
,
" \
\
\
,,
IOO~ ________ ~ ____ ~ ____ ~ ____

0.1 o.s I 5 l.O ._
Fig. 9 Model 2 is constructed so that the volumes of acini might not be equal. The pairs of terminal
branches are different in length (a). Log-log plot is not linear (b).
5 Conclusion
Though as yet unable to give a coherent explanation, we may assume that there
is a principle which closely relates the self-similarity of airways with the space
division of lung tissue into unitary acini of an equal volume. Room is left for the
10
14 14-
Fig. 10 Model 3 is an asymmetric branching in which acini have the same volumes (a). The branching
pattern of the last three branches is modified (b).
concept «equilibrium space division» by Suwa to be re-examined from a statistical

physics point of view. However, this concept seems to provide an important key
to disclose a fractal-related design principle for living organisms.
References
[1] Heitzman, E.R. (1973) The lung. Radiologic pathologic correlation. Mosby,
St. Louis.
[2] Kitaoka, H. and Itoh, H. (1991) Spatial distribution of the peripheral airways
- application of fractal geometry. Forma 6, 181-191.
[3] Kitaoka, H. and Itoh, H. (1992) Computer-assisted three-dimensional volume-
try of the human pulmonary acini. Tohoku 1. Exp. Med. 167, 1-12.
[4] Mandelbrot, B.B.(1982) The fractal geometry of nature. Freeman, San Fran-
cisco.
- -
,
~ (.)
5.(lr() ,
Modl'I I
,,
\
I 1 ,
,,
,
,,
100 ~
01 0.5 15 3.0 .1.-1
Fig. 11 Modell' consists of branches ~hose diameters are all equal (a). The branching pattern is the
same as in Model 1. Log-log plot is not linear (b).
[5] Meakin, P. (1987) Diffusion-limited cluster-cluster aggregation. In: Domb, C.

and Lebowitz, J.L. eds. Phase Transitions and Critical Phenomena, Academic
Press, New York, pp. 432-439.
[6] Suwa, N. and Takahashi, T. (1971) Morphological and morphometrica1 anal-
ysis of circulation in hypertension and ischemic kidney. Urban & Schwarzen-
berg, Miinchen-Berlin-Wien.
[7] Suwa, N. (1981) Supracellular structural principle and geometry of blood
vessels. Virchows Archiv A 390, 161-179.
[8] Thoma, R. (1901) Uber den Verzweigungsmodus der Arterien. Arch. Ent-
wicklungsmechanik, 12, 352-414.
[9] Weibel, E.R. (1963) Morphometry of the human lung. Springer, Heidelberg.
[10] Weibel, E.R. (1991) Fractal geometry: a design principle for living organisms.
Am. J. Physiol. 261, L361-369.
[11] Weitz, D.A. and Oliveria, M. (1984) Fractal structures formed by kinetic
aggregation of gold colloids. Phys. Rev. Lett. 52, 1433-1436.
Multivariate Characterization of Blood Vessel
Morphogenesis in the Avian Chorioallantoic
Membrane (CAM): Cell Proliferation, Length
Density and Fractal Dimension
Haymo Kurz, J6rg Wilting and Bodo Christ
Anatomisches Institut II der UniversiUit
D-79001 Freiburg
Abstract. The development of blood vessels in the CAM of chicks between 6 and 19 days of
development was characterized as a time series of their fractal dimension D. The growth rate of
endothelial and other cells that constitute the vessel walls, was estimated in terms of numerical density
NA , and the arterial vessel length density LA was also assessed. Image analysis and stereology were
used for this quantitative study. D changed between day 10 and 14 from the initial 1.3 to 1.5, whereas
NA dropped from 2500 mm - 2 to 250 mm - 2 . LA showed only an increase from 4 mm - I to 6 mm - I ,
and the total vessel length was estimated to expand from 2.5 m to 50 m. The effect of Vascular
Endothelial Growth Factor on the bifurcation frequency was documented by an increased D of 1.8.
Multiple scatter plots were used for visualization of the data, which suggest mainly «intussusceptive»
(i.e. by mesh formation without sprouting) capillary growth. This explains the increased density of
hemodynamic sources and sinks in the capillary layer, which is the anatomical correlate of a postulated
functional shunt. The very rich vascularization of the older CAM rather is a side-effect of angiogenic
factors acting in the embryo than optimal metabolic adaptation.
1 Introduction
Living structures may be described as being in self-organizing fluctuating steady-
state far from equilibrium [1] . Fractal geometry, on the other hand, has been
applied for physical phenomena far from equilibrium, like turbulence (cf. chpt. 10
in [2]). With the cardio-vascular system, which amounts to approximately 50% the
weight of early developmental stages of birds, the first fully functioning system
is organized, and it is just this system, which establishes steady-state and far-
from-equilibrium conditions over several orders of magnitude. It therefore is not
too astonishing that vessel systems have been characterized with fractal dimen-
sions (cf. chpts. 15-17 in [2], [3]), and that they have been modeled as fractals
with branching algorithms [4]. While the significance of «fractal organization» is
fairly accepted with respect to the limiting physical conditions of blood circulation
[5,6,7], this concept seems rather speculative in view of the genetic and devel-
opmental realization in the living organism. We therefore attempted for the first
time to describe the pattern formation of a real vessel system by determining its
fractal dimension D. Since D is a number without physical dimension, additional
information is needed to more precisely describe the actual biological generator
and the physiology of the system,and to combine this knowledge with exist-
ing morphological and physiological findings. Hence, the density of proliferating
Multivariate Characterization of Blood Vessel Morphogenesis 133
Fig. 1 Binary image of arterial vessels in the l4-day CAM. The cycloidal waves are superimposed
with the XOR function to show the counted intersections in gray.
cells that contribute to the developing vessel pattern, NA, and the observed vessel
length density L A were determined with stereological methods, and combined by
explorative multivariate statistics.
The CAM was chosen for methodological and practical reasons: Being the
largest extra-embryonic and the major respiratory organ of the embryo, it grows
in intimate neighborhood to the egg shell and is easily accessible. It displays a
comparatively simple, layered structure, and the vessel «tree» hence grows in a -
topologically two-dimensional - tissue sheet. Its fractal dimension consequently
lies between 1.0 and 2.0 and thus may be treated with standard image analysis.
Moreover, the CAM assay has become a standard for the evaluation of angiogenetic
substances [8], and for the grafting of e.g. tumors [9], and the physiology of avian
gas exchange is known since long [10).
Two major questions were to answer with the work presented here: Does
the quantitative analysis of blood vessel formation give evidence for a possible
mechanism of angiogenesis [11] in the CAM? Can the concept of an arterio-
venous shunt, which was derived from respiratory data [12,13], be supported by
the structural data?
134 Kurz et al.
Fig. 2 Box counting method with the SIS system. The reduced pattern indicates the scaling. Three
regions of different slopes are distinguished in the log-log plot.
2 Materials and Methods

After removal of ca. 1 ml albumen, White Leghorn chicken eggs were fenestrated
at day 3 or 4 of incubation at 310 K and 80% humidity. CAMs were obtained
at days 6 through 19 (n = 10 each); after fixation in situ, rectangular pieces of
about 1 cm2 were transferred to a petri dish with fixative on an illuminated box.
The native contrast of the erythrocyte-filled arteries in these whole mounts was
viewed via a B/W CCD with Macro objective 1:2, with one randomly oriented
specimen as one sampling unit. The SIS (Miinster) system was used for 8 bit gray
value, 5122 pixel image processing at a spatial resolution of 15 11m/pixel. The
images were treated with automatic shading correction, gray value standardization
and binarization at a fixed threshold; isolated pixels were eliminated with the
connectivity operation. For the dimensional analysis with the box counting method
(BCM, cf. [2,14]), D was taken as the negative slope of the log-log plot of box
edge length (1,3,5, ... 25 pixels) vs. number of boxes hitting the border of the
binary vessel image. In a separate series of preparations [15], vascular endothelial
growth factor (VEGF) was locally applied to the CAM.
The binary image was also used for stereological estimates of LA = n(Ix +
Iy)/4c by automatically counting the result of an AND operation with a cycloidal
grid, with Ix: intersection count in a random orientation x, Iy: count at right angles
to x, c: length of test lines in one sampling frame (modified from [16]).
'1.10 D 9.0..
2.00 ~'hE~~.
(Pi"..l)
~ A
a
0 7 .. 15
...
lea ....... 1.. 6 .0..
.0..
liD ~
.0..
170 c>
.0.. 0
teo 0 c> 0
/ " .. i
0 Q
160
0 0
L2:,. 0
140
130 ~ 0 ><
120 0
+=t= ++ ~
>< ... *
liD -+- VEGF
Vc>>< daY'"
100 I
" 1
I I I I I I I I I
10 11 12 13 14 16 18 17 18 Ie
Fig. 3 Dependence of D from incubation time and box size. The application of VEGF leads to an
elevated D mostly for medium sized, but also small boxes.
For the estimation of NA, 0.3 ml of 40 mM bromo-deoxyuridine (BrdU,

Sigma) was applied to the CAM for 45 minutes before fixation. The incorporation
of BrdU into the nuclei of proliferating cells was detected with monoclonal anti-
BrdU antibody (Dakopatts) after acid hydrolysis. The diamino-benzidine (DAB)
reaction indicated the labelled nuclei with brown precipitate. With whole mounts
embedded in glycerol on a clean slide, nuclei in the capillary layer of the CAM
were counted at 16 x microscopic magnification in stripes of 0.04 mm 2 in a
systematic random sample.
Whole mounts were also stained in DAB without pretreatment to obtain strong
positive contrast of the arterial and capillary vessels, filled with erythrocytes. This
allowed for microscopic registration of the vessel «endpoints» in the capillary layer
at 16 x magnification. Tessellations around these were generated by mathemat-
ical morphology (skeletonization, dilation), and arterio-venous distances (AVD)
measured in two-dimensional random samples.
3 Results
The projected image of a CAM arterial tree is shown in Fig. 1. Statistical self-
similarity appears to be present in the branching pattern. The dimensional analysis
of Fig. 1 with the BCM is shown in Fig. 2. The results for the first time-series
of experiments are combined in Fig. 3. As this natural «fractal» has upper and
lower limits, the local dimensions D of the branching vessels were determined for
three regions: for box lengths below 90 j.lm, D = 1.1 varied only slightly with
development, and for box lengths over 230 j.lm, D approached 2.0. For intermediate
box lengths, a considerable rise of D with development was measured between
day 7 (1.3) and 13 (1.5), with no further increase later. The application of the
136 Kurz et al.
LA (mm- 1 )
64 L (rn) CJ CJ
P (mrn Hg) CJ
6 32
& 18
r7
/.
8 9
.
10
.'11 12 13 14 1& 18 17 •
18
daYI!I
•
.19
Fig. 4 Change of LA (/quares), L (thick line) and p (thin line. from [17]) with development. The
logarithmic ordinate applies to Land p.
potent angiogenetic substance VEGF [15] led to a D = 1.8 (day 13 .. 16) under the
carrier, as compared to the D = 1.5 outside (Wilcoxon p < 0.05).
The length density of arteries with apparent diameters larger than 15 J..lm,
together with the calculated total arterial vessel length and the mean blood pressure
[17] is plotted against incubation time in Fig. 4. LA shows approximately linear
growth with only two-fold increase between day 7 and 13; as the CAM itself
continues to expand during this period, L, however, grew exponentially from 2.5
m to about 50 m.
Multiple scatter plots combine the three numerical descriptors of vascular
morphogenesis in Fig. 5. NA declined from about 2500 mm- 2 (day 8) to 1200
mm- 2 (day 14) and then dropped to 250 mm- 2 from day 16 on. Again, the sudden
jump of D and the gradual increase of LA is found as before.
A microscopic image of the terminal arterial and venous vessels, together
with their pattern of hemodynamic sources and sinks is shown in Fig. 6.
The histogram of AVDs is given in Fig. 7 for the 10 and 16 day CAM. The
large variability is obvious in both, but the occurrence of AVDs shorter than 100
J..lm is typical only for the older CAM.
4 Discussion
The CAM arterial vascular pattern has been portrayed as growing in a planar, tree-
like fashion. The thickness of the CAM is known to be about 100--150 J..lm:, with
all the larger vessels remaining in a plane parallel to and underneath the capillary
layer, i.e. the chorionic surface. Only the immediate pre- and post-capillary vessels
can deviate considerably from this plane, but their diameters are below 15 J..lm and
9 :itt, 1. 999 2 999 L 3 1.. 4 L 5 L 6
••
V V
Vx V
~1tx
;px V
x •••
xxx ••
••
L5
L 4
0
~~#q§l
0
~
1..3 +<1"
0 0
+ + 0() 0
•• • •
7 ~~V
~
~ Vx
V
•
• • • •
~
0
•
0
0 tfI
0
0
•
•x
. ~
£x
~~
~ V
6 Xx 0
• 0 0
0
XXx .
00 b
5 00 00
3
o a~
<lr>
~
~~
@"
·~t
++ + +*~.p~ f ~
Fig. 5 Scatter plots of NA vs. D (top). NA vs. LA (bottom left), and D vs. LA (bottom right). Age
in days (d) is indicated by symbols: 6d +; 7d 0; 8d 0; IOd 0 ; l2d 6; l4d .; l6d '\7; 18d x .
Note the well separated 12d and 14d clusters.
they remain invisible to the CCD with macro objective, which guarantees that no
detail is lost by evaluating the projection. A certain amount of «roughness» of the
apparent vessel border is lost, however, by eliminating isolated pixels, which were
mostly due to noise. This may account for the rather low value of D = 1.1 for
small boxes. The projection through a random fractal, though, should not reduce
its dimension (p. 378 in [2]), if D is smaller than the topological dimension, which
is equal to two in our case.
Vessel «trees» are characterized by two different parameters, D and the diam-
eter exponent ~ [2,7]; in our case, self-similarity and thus D = ~ can be assumed
[2]. The estimation of fractal dimensions also depends on the chosen method [14].
As the Sierpinski triangle test with our BCM shows a difference in D of only
0.01, we are confident that D is determined precisely in view of the biological
variability. The BCM is found to be a reproducible measure for box edge lengths
138 Kurz et al.
Fig.6 Patterning of the CAM capillary bed (granulated grey) by arterial and venous vessel endpoints
(dots). The sources are connected to the arteries (lines along and branching off the edges),
the sinks drain to the central vein (oblique middle line).
between 5 and 19 pixels. The best linearity, however, and the most striking differ-
ences are seen in the region 7 ... 15 pixels. We use this «local» D as a measure of
complexity, one interpretation being that vessel bifurcation intensity is enhanced
during the CAM growth period, just when the CAM covers the entire shell from
inside after day 12. The elevatedD after application of VEGF corresponds to the
even more enhanced bifurcation frequency and to the many irregular vessels found
in histological preparations [15].
The stereological estimation of LA by counting intersections of the vessel
image with reference lines has the advantage of working around the problems of
defining length in the quadratic pixel raster. It thus is rather independent from the
BCM, which counts pixels directly. The cyc10idal lines were found to be very
effective in that the variation coefficient of LA between different orientations on
one pattern always was below 2%. This regularity also suggests that the vessels
in any single CAM represent a stationary and isotropic fibre process, although the
variability between the CAMs of one age group becomes large after day 10 (cf.
Fig. 5).
From the triple scatter plot it is obvious that the proliferation rate in the
capillary layer is highest, as long as larger vessels grow, while D stays at 1.3 during
20 n
10
Fig. 7 Distribution of AVDs from random samples of IO-day CAM (hatched, n = 160), and 16-day
CAM (white, n = 169), in which the proportion of AVDs larger than 250 /lm is reduced in
favor of rather short AVDs below 100 /lm.
this period. It jumps to values above 1.5, when LA has reached its maximum,
and NA has dropped to half the earlier value. This suggests a sequence of first
purely intussusceptive capillary growth, which later appears as «sprouting» of
precapillary vessels [9,11,15]. With 0, the conductivity also increases [7], which
can compensate the still low blood pressure.
The measurement of arterio-venous endpoint distances, and their distribution
in the capillary layer may be easily related to the great variability of blood flow
seen in the living CAM, which appears as a non-stationary temporal pattern. The
small AVDs appear as a side-effect of angiogenic factors needed for the embryo
proper, and explain the functional shunt, so that the structure of the older CAM
seems to be not optimally adapted to metabolic needs. For a transient organ, which
is needed for less than 14 days, this seems acceptable in view of e.g. increasing
hematocrit etc. [12].
In summary, our introductory questions could be answered: Explorative data
analysis of cell proliferation, vessel length, and fractal dimension supports the
hypothesis of intussusceptive growth in the capillary plexus, which in turn forms
the major conducting vessels. The functional shunt is most likely realized by the
diminished AVDs, and is caused by the conducting vessels forming more and more
connections to the capillary bed.
The strength of the fractal concept of dimension lies in the description of
complexity, which is generated from scarce information. Structure may thus be
related to both molecular and macroscopic function, and biological and synthetical
pattern formation may be compared. Following this notion, one can speculate,
whether fractal geometry can be useful in extending the «hypercycle» [1] concept.
140 Kurz et al.
Acknowledgement:
This work was supported by DFG grant Ch 44/9-2.
References
[1] M. Eigen, Die Naturwissenschaften 58/10, 465-523 (1971).
[2] B.B. Mandelbrot, The Fractal Geometry of Nature, Freeman, San Francisco
(1983).
[3] T. Matsuo, R Okeda, M. Takahashi, and M. Funata, Forma 5, 19-27 (1990).
[4] H.R. Bittner, in: H.O. Peitgen et al. (Eds.), FRACTAL 90, Elsevier, Amster-
dam (1990).
[5] M. Semetz, B. Gelleri, and 1. Hofmann, J. Theor. BioI. 117,209-230 (1985).
[6] J.B. Bassingthwaighte, RB. King, and S.A. Roger, Circ. Res. 65, 578-590
(1989).
[7] H.C. Spatz, 1. Compo Physiol. B 161,231-236 (1991).
[8] 1. Wilting, B. Christ, and M. Bokeloh, Anat. Embryol. 183, 259-271 (1991).
[9] D.H. Ausprunk, D.R Knighton, and J. Folkman, Devel. BioI. 38, 237-248
(1974).
[10] H. Rahn, c.v. Paganelli, and A. Ar, Resp. Physiol. 22, 297-309 (1974).
[11] S. Patan, B. Haenni, and P.H. Burri, Anat. Embryol. 187, 121-130 (1993).
[12] H. Tazawa, Am. Zool. 20, 395-404 (1980).
[13] D. Wangen steen, and E.R. Weibel, Resp. Physiol. 47, 1-20 (1982).
[14] D. Stoyan, and H. Stoyan, Fraktale - Formen - Punktfelder, Akademie
Verlag, Berlin (1992).
[15] J. Wilting, B. Christ, M. Bokeloh, and H.A. Weich, Cell Tiss. Res. (1993).
[16] E.R Weibel, Stereological Methods I, Academic Press, New York (1979).
[17] L.H.S. Van Mierop, and C.J. Bertuch, Am. J. Physiol. 212,43-48 (1967).
Phyllotaxis or Self-Similarity In
.
Plant Morphogenesis
Fran<;:ois Rothen
Universite de Lausanne, BSP-Dorigny
CH-1015 Lausanne, Switzerland
1 Phyllotaxis or an exotic form of two-dimensional crystallography

Phyllotaxis refers to the geometry governing the arrangement of inner florets of
a sunflower (Fig. 1), of the scales of a pineapple (Fig. 2), of the leaves around a
stem and so on. The florets align with spiral whorls in the case of spiral phyllotaxis
(daisy) or with helices in the case of cylindrical phyllotaxis (pineapple or fir-cone).
In Fig. 2 showing the picture of a pineapple with its hexagonal scales, the eye
is at once attracted by the helices on which the neighbouring scales are aligned. A
given hexagonal (rectangular) scale belongs to three (two) different helices, called
parastichies; they bind each scale with its six (four) different neighbours.
Parastichies are grouped into families, a parastichy family being the set of all
parastichies parallel to each other (see Fig. 3). Each family consists of a number
k of parastichies which completely characterizes the family. Depending on the
number of parastichies an arbitrary scale belonging to a given phyllotaxis will
then be specified by a pair (k , 1) or a triplet (k , 1, m) characteristic of the two or
three families of parastichies through the pattern of scales (this is also true for
plane phyllotaxis whose parastichies are spirals rather than helices).
It was soon recognized that k, 1, m are very often (but not always) successive
members of the Fibonacci sequence [1,2]
Un} = 1, 1,2, 3, 5,8 , 13, 21,34, ... (1)
defined by the recurrence law
(fl = h = 1). (2)
The Fibonacci sequence is related to the golden section T:
lim /k + 1 = T =: 1 + y'5 = T -1 +1 (3)

k -->oo /k 2
The golden section is an irrational number whose remarkable properties cannot

all be listed here; one of them however is worth mentioning. Every irrational x
between 0 and 1 can be expressed in a unique way as an infinite sequence of
142 Rothen
Fig. 1 Sunflower: a hidden self-similarity.
stacking fractions:
x = - - - - - - ; - - - =: [a\,a2,a3 , ... J (4)

a\ + - - - ' 1-
a2+---
a3 + ..
Phyllotaxis or Self-Similarity 143
Fig. 2 Pineapple: a further hidden self-similarity.

144 Rothen
Fig. 3 Looking at the pineapple, one immediately realizes that the scales are regularly disposed
along spirals, called parastichies. The three ones going through the hatched scale have been
marked here A, B, C. Parallel spirals are grouped in a given family: so are A and A'.
where ai (i = 1,2,3, . . .) is a positive integer [3]. The semi-infinite sequence is

called a continued fraction. Now T- 1 = T - 1 belongs to the interval [0,1] and
its continued fraction writes
T- 1 = [1,1,1, ... ]. (5)
Moreover all numbers whose continued fraction exhibits only ones from some
stage are called noble numbers [4]. For instance
1
-2-=[3,1,1,1, ... ] (6)
+T
is called Lucas~ number. Noble numbers share with T-l, «the noblest of all noble
numbers», some interesting properties we shall discuss below.
Throughout history, there is an impressive number of people associated with

phyllotaxis (see for instance [1] or [5]), among whom one could quote, somewhat
arbitrarily: D' Arcy Thompson, the famous author of «On Growth and Form» [6],
H.S.M. Coxeter, who devoted one chapter of his «Introduction to Geometry» to
phyllotaxis [7] and recently Levitov who made an outstanding analysis of an
equivalent phenomenon in physics [8].
2 Self-similarity in phyllotaxis
Self-similarity arises in an obvious way in the case of a nautilus (Fig. 4) as it merely
reflects a growth process without any change of relative size (the proportions
remain constant during the process, there are no superimposed constraints). The
case of a phyllotactic pattern is however more complicated. Scales and florets have
to expand on a tissue whose proper growth rate and shape influence not only the
final geometry of the pattern but also its composing units (scales, florets).
We shall see in the next section that, according to the value of the parastichy
numbers, some hidden but important self-similarity may be present in the pattern.
On the other hand, the parastichy numbers can change in a given pattern
during the growth, giving rise inside the same pattern to different domains, each
one characterized by a particular pairl) of parastichy numbers. This transition is
called a bifurcation. During such a process, one pair can be replaced by some other
pair (sect. 5). As a consequence, all possible phyllotactic patterns can be classified
by a bifurcation tree. Interestingly enough, the tree itself shows characteristic self-
similarity properties.
In the next section, we will introduce two important new geometrical pa-
rameters, i.e. divergence and plastochrone ratio, in order to discuss these various
expressions of self-similarity. Sect. 4 and 5 will be devoted to the discussion of
self-similarity itself.
3 Divergence and plastochrone ratio

It is useful to replace the parastichy numbers by two other parameters, the di-
vergence and the plastochrone ratio, which allow a more precise description and
better describe a phyllotactic pattern.
Let us consider Fig. 5 in which the scales of a pineapple are numbered
according to the order of their birth, the older ones having the lowest numbers. The
divergence x is the angular distance between two successive scales, i.e. between
scales k and k + 1. Notice that x is expressed as a fraction of whole turns:
O::;x<l.
1) In the following, we shall forget the fact that parastichy numbers can also occur as a triplet.
146 Rothen
Fig. 4 Nautilus: an obvious self·similarity.
The pineapple can be schematically represented by replacing each scale by a

point on a cylinder with circular cross-section and a circumference of unit length.
One could even unwind the cylinder on a plane (Fig. 6a, 6c and 6d) producing a
periodical two-dimensional lattice. In the case of plane phyllotaxis (sunflower), the
successive florets can be represented by points aligned on a spiral (Fig. 6b) with
a divergence equal to 1/27r times the angular distance between two successive
Fig. 5 The scales of pineapple have been numbered from the bottom upward. Notice that the numbers
increase regularly along parastichies: 5 to 5 along parastichies belonging to the same familiy
as A, which is therefore called a S-family. Notice also the presence of an 8-family (B) as
well as a 13-family (C). A. Band C refer to Fig. 3.
points. Hereafter, we shall speak of a periodical lattice (spiral lattice) to describe

the pattern of Fig. 5d (5b).
To determine completely the geometry of both lattices, we have to add a
further parameter, namely the plastochrone ratio 2 ) z. In the case of the periodical
lattice, if one chooses the origin a~ a point of the lattice (Fig. 6d), the plastochrone
ratio is equal to the positive ordinate zp of the point with abscissa x . In other
words, (0,0) and (x, zp) are the coordinates of the points corresponding to two
successive scales. In the case of the spiral lattice, the plastochrone ratio is defined
as the ratio Zs > 1 of the distances of two successive points to the origin. This
2) Divergence as well as plastochrone ratio are terms coined by botanists.

148 Rothen
•
•
•
•
• v• 0 -+----------+--
•
U• 0 U• 1
© Cd)
Fig. 6 The pineapple (a) is idealized as a cylinder (c). On the corresponding cylindrical lattice,
each point represents the center of a scale. Unrolling the cylinder on a plane, one gets the
periodical lattice (d). The spiral lattice (b) is an idealization of the pattern of a sunflower or
a daisy. In the spiral lattice, the angular distance of two successive points aligning with the
spiral is equal to 211'x, x being the divergence of the pattern. In the periodical lattice, the point
having the lowest positive vertical coordinate is specified by the pair (x , zp) where zp is the
plastochrone ratio. The divergence is still given by x.
definition implies that the spiral is equiangular3) (or logarithmic), the origin of
the coordinates coinciding with the point to which the spiral converges inwards.
In polar coordinates, two successive points coincide with (r, q) and (zsr, e+ 27rx).
4 Symmetry properties of a periodical lattice under compression

Assume for a while that the pineapple can be identified with a cylinder with
circular cross-section and that its rate of growth is a constant so that the interval
between the birth of two successive scales is al)Vays the same. Then the pineapple
can be compared in a realistic way with the above mentioned periodic lattice.
We shall now put forward a hidden symmetry of the periodic lattice which
is highest when the divergence is equal to the golden number 7; we then speak of
golden divergence.
In the vegetable kingdom, the chirality4) of most patterns seems to be arbitrary
(there are as many patterns turning right than left). As a consequence, for x between
o and 1, both divergences x and 1 - x can be considered equivalent. For a golden
divergence, this means t~at 7- 1 and 7- 2 = 1 - 7- 1 are equivalent to 7. Botanists
prefer to use degrees to measure divergences and to avoid angles greater than 180°,
so that botanists fix the value of the golden divergence to 7 - 2 .360° = 137° .507 ...
Arithmetic gives a well-known recipe to approximate irrational numbers by
rational fractions. The best approximations (to be specified [3]) are given by the
successive troncations of the continued fraction corresponding to a given irrational
(7)
7- 1 and 7- 2 are respectively approximated by the following sequences:
(8)
(9)
These rational fractions involve only members of the Fibonacci sequence (1).
As a consequence, if the divergence of a pattern almost equals 5/13, this
means that approximately 13 points align with the spiral while it describes 5 turns
(this is the case of the spiral of Fig. 6b). Florets number k, k + 13, k + 26,
... almost align with a straight line (actually a very steep spiral) issued from the
3) As will be clear in the next section, this choice of a particular shape of the spiral does not restrict
the range of validity of the model.
4) In the case of an enantiomorphic object such as a cork-screw, the chirality can be identified with
the direction of winding.
150 Rothen
,,
,
---- G - .. __
t,
I
--t....... -- -----..-- S'

,,
--t-- -- .
--------f-+-+- .
•,
,
•
••,
a b c
Fig. 7 If the divergence of the periodical lattice is rational (here x = 1/3). the points align with
vertical lines. The polygons shown here in each case are the Voronoi polygons (VP), i.e. the
closed curves which surround the set of all points of the plane which are nearer to a given
element of the lattice than to any other element. As can be seen, there is no shape invariance
of the VP when the plastochrone ratio Zp is reduced, i.e. when the lattice is compressed along
the vertical direction. Dashed lines indicate two of the most visible parastichies.
origin. This remark is valid for k = 0, 1, 2, ... , 12; the steep spiral is a parastichy
belonging to a family of just 13 members. Notice that 13 is the value of the
denominator of 5/13. The same argument is also valid for 3/8 and 2/5 which
are (less good) rational approximations of T - 2 . It follows that 5, 8 and 13 are
the parastichy numbers corresponding to the pattern of Fig. 6b. Using the same
argument, one can determine the parastichy numbers of the pineapple of Fig. 5 and
they happen to be the same. Now to the central point of this section. Notice that
we shall restrict ourselves to the case of the periodic lattice (recall for the present
discussion that the periodic lattice is equivalent to the cylindrical lattice 5 ) but
the case of the spiral lattice is perfectly analogous.
Obviously, the pineapple is neither a cylinder nor is its growth rate a constant.
The question therefore arises: what are the consequences of the important shape
gap between botanical objects and their geometrical idealization?
Observation tells us that some hidden mechanism leads to an almost constant
divergence (we shall come to this point later). The plastochrone ratio, however,
very frequently changes throughout a phyllotactic pattern. First of all the rate of
5) A helix with constant pitch winding around the cylinder corresponds to the straight line con-
necting (0,0) and (x , Zp) in the periodic lattice.
,. •• •• • ••••••• j' •
" ............. " ..
.... ., " .. ,., ....... ..........:...
.. .'. .
•• • • 4
"- 'W,.. /
....... I ...
. .... ... ., .........

.......... .. . . I .
.... , .....
.. .. .. . .. . ..... ~/.
~'liJ:..
. , "- ... ., ..•• ......

/'
,~
/'
•• •• I •••••
,
r •• • • 4
. .......... .
•• ........
a b c
Fig.8 If x is irrational, parastichy transitions occur during the compression (here x = 21 /2 - I) . (a)
Parastichies belong to the 2- and 5-families. (b) Due to vertical compression of the lattice,
a transition has occured. Parastichies now belong to the 5- and 7-families and the Voronoi
polygon has been distorted and rotated. (c) Further compressing of the lattice leads to a new
transition: 5- and 12-families become visible and the Voronoi polygon recovers the shape it
had in (a).
growth is not always constant. On the other hand, putting some helicoidal lattice
onto a surface differing markedly from a cylinder is more or less equivalent to
change the plastochrone ratio of a cylindrical lattice from point to point.
The consequence of the fact that x can be considered as constant while z
can vary throughout the pattern is easy to express through the use of the paras-
tichy numbers (we consider here only parastichies occuring as pairs rather than
triplets): there will be separate regions of the pattern, each one being characterized
by a different pair. However, it follows from the constancy of x that each pair
must correspond to the sequence of the rational approximants of x. More pre-
cisely, both members of the pair will generally be denominators of two successive
approximants of x [5].
There is now a very interesting point. Let us investigate the shape of the
Voronoi-polygons 6 ) (VP) defined around each lattice site. In principle, their shape
depends on x and zp (or only on zp if x is constant). VP tile the space occupied
by the lattice as well as the scales of the pineapple do. One therefore expects the
6) The Voronoi-polygon (or Wigner-Seitz cell, or Dirichlet domain) is the set of all points (here of
the cylinder) which are nearer to a given element of a discrete point lattice than to any other
element.
152 Rothen
, .,, ... .. ..... .. • •• ••• •• •• 1'- °
.....
, ··............... ••..... . '. .
• • 4- •
""'.
•• •• •••
• •••• • I ••
•• *1 .. .
. "." "
.. ·...........
./ °
..o...
~
. ... ., " .......

; .......... .
....
0
. . ~"o
° - ........ ~.
..
~ ~
., . . ..I . .
° ~'.
, " --- •• • • 4 •• .......
,
0
/ • ••* ••* • .* ••* ••
,~
a b c
Fig. 9 If x is equal to the golden mean. compressing the lattice along the vertical axis leads to normal
transitions which do not distort the shape of the Voronoi polygon. The parastichy numbers
are always successive members of the Fibonacci sequence (I): one successively notices (a)
(2,3); (b) (3,5) and (c) (5 ,8).
shape of the botanic units to be modified in the same way as the Voronoi-polygons
as soon as the value of 2p changes from region to region.
In order to understand what could then happen, consider a periodical lattice
with a fixed value of the divergence x but with successive values 21 > 22 > 23
of 2p. Decreasing 2p from 21 to 22, then to 23 is equivalent to a compression
of the lattice along a direction orthogonal to the x-axis. The possible results of
this compression can be classified according to the arithmetical properties of the
divergence:
• If x is a rational number the VP get thinner and thinner as 2p decreases (see
Fig. 7). The shape of a scale would greatly change throughout the pattern.
• When x takes some irrational value, a new phenomenon occurs during a
compression, namely the appearance of parastichy transitions: for special
values of 2 , the pair of parastichy numbers jumps to another value 7 ) (Fig.
8). In general, during such a transition, the shape and orientation of the VP
abruptly change.
However, after a finite number of transitions, the VP always recovers its
original shape. In the case of the golden divergence (Fig. 9), the shape remains
7) If x is rational, parastichy transitions only occur at an early stage of the compression.

a b
c
Fig. 10 Starting from a given configuration (a) with parastichy numbers (k , /) (here k = 5, / = 8),
the plastochrone ratio Zs is slowly reduced. The lattice becomes more and more compact, up
to close-packing (b) where the three parastichy families (k, /, k +/) = (5 , 8,13) are apparent.
Further decrease of Zs can lead either to (/, k + /) = (8,13) through a normal transition (c)
or to (k, k + 1) = (5, 13) through a singular transition (d).
invariant during a transition. The whole succession of shape changes of the VP is

directly linked to the development of x as a continued fraction [5].
The behaviour of the periodical lattice under compression has an important
consequence for a real botanical pattern. If the plastochrone ratio is non uniform
throughout the lattice, separate domains appear corresponding to different pairs of
parastichy numbers. If the divergence is equal to T (or to a noble number [5]),
the shape of the scales will be the same in all domains. However, in the case of a
rational divergence, or if x is very «far» from a noble number, either the botanical
154 Rothen
units cannot fill equally well each domain or they have to change their shape
from domain to domain. Now, observation rules out these latter cases: one notices
only noble divergences. Interestingly enough, this shape invariance of scales is the
consequence of a seemingly pure arithmetic property, namely the occurence of an
infinite sequence of ones in their continued fraction!
5 Self-similarity of the bifurcation-tree

What are the contraints ruling the parastichy transitions introduced in the preceed-
ing section? In order to answer this question, it is necessary to add something to
the point lattices considered so far, either by replacing points by non overlapping,
tangent circles [9,10], or by introducing well-defined forces between the points
[8,11], i.e. by transforming the system into an interacting point system.
Consider the pattern consisting of points aligned with an equiangular spiral so
that the divergence and the plastochrone ratio are constant everywhere (see sect.
3). Assume that we number the points along the spiral, successive points corre-
sponding to successive numbers (the numbers are assumed to increase outward).
Number 0 is given to an arbitrary fixed point of the pattern. Assume further that
the distance of 0 to the center n of the spiral is equal to 1. The distance between
point k and n is equal to z: (-00 < k < +00). If we draw a circle of radius
PI = az: around k, we can adjust [9,10] the (positive) number a so that (see Fig.
lOa):
• any circle Co is tangent with exactly 4 other circles, say C 1, C 1', C2, C2';
• two tangent circles, say C 1 and C 1', belong to the same parastichy going
through Co; the second parastichy going through Co involves C2 and C2';
• there is no overlapping;
• except for scaling and rotation, any circle of the pattern is fully equivalent to
any other (there is complete self-similarity of the pattern).
Now the whole system depends on a continuous way of Zs: if we slightly
change zs, we have to let x vary in order that the tangent condition remains
fulfilled: x must be equal to some function Xt(zs) of Zs for the circles to remain
tangent.
Assume now that, starting from some value Zo > 1 of zs, we steadily reduce
Zs. For some z\ such that Zo > z\ > 1, the lattice of circles will be close-packed:
each circle has now two new neighbours C3 and C3' which are also tangent to
Co and define a third parastichy going through it (Fig. lOb). If we further reduce
Zs by preserving the whole set of tangency rules, we have to let either (Cl, C1')
or (C2, C2') move away. One can then prove [9,10] that, if k and 1 are relative
primes (k < 1) and if (k , 1) represent the parastichy numbers for Zo > Zs > ZJ,
then the parastichy transition corresponds to a bifurcation:
(k 1) {(l,k+l) normal transition

(10)
: --., (k,k+l) singular transition
oqo
(D
N
(D x = (;,-1
0
X
<=) { 2 .3 }
(D
OJ <=) .
u
a:::
OJ ee ••
C>
t...
OJ
.....
>
Cl
CD
10
0
(D
....
m
0
oqo
10
0
1.0 1.1 1.2 1.3 1.4 1.5

Plastochrone ratio z
Fig. 11 The curves put forward the values of (x, zs) for which the construction of a lattice of tangent
circles is possible. Starting with a system characterized by parastichy numbers (k, I) =
(2,3), a reduction of Zs leads to close-packing of circles. A further decrease of Zs may
then produce two different situations: either the k-parastichies disappear (nonnal transition
(k , I) --> (I , k + I)) or the I-parastichies vanish (singular transition (k, I) --> (k , k + I)).
Further reduction of Zc leads to new transitions.
Further details can be found in reference [10]. Anyhow notice that starting from a
pair (!r , fr+ 1) of consecutive Fibonacci numbers, a sequence of n normal transitions
leads to <tr+n, fr+n+ 1) and the divergence remains a golden one.
On the other hand, it is easy to show that a normal transition always leads to
a sign change of the curve x = Xt(zp). After a cascade of normal transitions, the
divergence oscillates around an asymptotic value. Starting from any pair (k, I),
such a cascade leads to new pairs (k(n),I(n» such that limk--+oo k(n)/I(n) is
equal to some noble number. Among others, this fact leads to the self-similarity
of the bifurcation tree of Fig. 11. Self-similarity is actually a consequence of the
form of (9) which controls all bifurcations [11].
156 Rothen
Levitov has recently investigated the compression along the zp-axis of a

periodical lattice of point particles [8]. The particles were interacting through
an isotropic but otherwise arbitrary potential (no kinetic energy was assumed).
He found a tree analogous to the bifurcation tree of tangent circles, except for an
important point: all bifurcations were replaced by quasi-bifurcations. To perform
a singular transition, the system had to jump across a gap in the (x , zp) space
while the normal bifurcations were still continuous. As a consequence, from a
loose state where all interparticle distances are large, there is a unique connected
way leading to a dense lattice; this way leads, through an infinite sequence of
normal bifurcations, to a lattice with golden divergence! An experiment involving
ferroftuid drops leads to a similar conclusion [12].
Now it is too early to say that the mechanism hidden behind the quasi-
universal emergence of the golden mean has been solved. First of all, following
an idea originated by Thomley [13], another investigation [14] was based on a
very simple model of hormones reacting around a circle8 ). It also led to a tree
of quasi-bifurcations, although the plastochrone ratio was replaced by quite a
different parameter, namely the degradation and diffusion rate ,\ of the hormone.
The self-similarity of the tree is baffling in this case. Moreover, using the same
model, Guerreiro [15] was able to show that the unique connected way in the
(x, ,\) tree does not universally lead to a golden divergence (the asymptotic value
of the divergence depends on rather unimportant details of the model).
Acknowledgement:
Knowledge about self-similarity in phyllotaxis has been gathered by many people:
it is impossible to acknowledge all contributors. I would like to quote those people
with whom I had special contacts, and among them G. Bernasconi, J. Boechat, D.
Bonnaz, Y. Couder, S. Douady, A. Joyet, J. Guerreiro, A.-J. Koch, M. Kunz, L.S.
Levitov, Piotr Pieranski and N. Rivier. R. Huguenin was kind enough to read the
text; I wish to thank M. Ede, F. Pretzsch and C. Rossier for technical help. Last
but not least, I am very indebted to the organizers of the symposium «Fractals in
Biology and Physics», namely G. Losa, D. Merlini and T. Nonnenmacher and to
the Swiss National Foundation for Scientific Research for its material support.
References
[1] Adler I., J. Theor. BioI. 45 (1974), 1.
[2] Jean R.Y., Phytomathematique, Presses de I'Universite du Quebec, Montreal
(1978).
[3] Hardy G.H., Wright E.M., An Introduction to the Theory of Numbers, Fifth
Edition, Clarendon Press, Oxford (1979).
8) In the neighborhood of the tip of a growing stem, new sprouts are born along a ring-shaped
region (the meristem); this region surrounds the apex.
[4] Schroeder M.R., Number Theory in Science and Communication, 2nd Edition,
Springer Verlag, Berlin (1986).
[5] Rothen F., Koch A.-J., J. Phys. France 50 (1989), 633.
[6] Thompson D' Arcy W., On Growth and Form, 2nd Edition, Cambridge Uni-
versity Press, Cambridge (1942).
[7] Coxeter H.S.M., Introduction to Geometry, 2nd Edition, Wiley and Sons,
New-York (1969).
[8] Levitov L.S., Phys. Rev. Lett. 66 (1991), 224; Europhys. Lett. 14 (1991),
533.
[9] van Iterson G., Mathematische und Mikroskopische Anatomische Studien
tiber Blattstellungen, Gustav Fischer Verlag, Jena (1907).
[10] Rothen F., Koch A.-J., J. Phys. France 50 (1989), 1603.
[11] Rivier N., Koch A.-J., Rothen F., in Biology Inspired Physics, L. Peliti Ed.,
Plenum Press, New-York and London (1991).
[12] Douady S., Couder Y., Phys. Rev. Lett. 68 (1992) 2098.
[13] Thomley J.H.M., Ann. Bot. Fenn. 39 (1975),493.
[14] Koch A.-J., Guerreiro J., Bernasconi G., Sadik J., submitted to J. Phys. I,
France.
[15] Guerreiro J., Thesis, Universite de Lausanne (1993).
Fractals in Molecular
and Cell Biology
Evolutionary Interplay Between
Spontaneous Mutation and Selection:
Aleatoric Contributions of
Molecular Reaction Mechanisms
Werner Arber
Abt. Mikrobiologie
Biozentrum, Universitat Basel
Klingelbergstr. 70
CHA056 Basel, Switzerland
The concept of evolutionary change does not assume biological evolution

to progress into a specific direction. Rather, single steps of change are at least
in part aleatoric. In this view evolutionary change may be considered to be of
fractal nature. However, in living systems, only a small fraction of mutants, i.e. of
individuals having suffered a change in their genetic information, are maintained
and propagated. Many other mutants can not compete with the successful ones and
with their parents and are thus eliminated. This phenomenon is called selection.
Molecular genetics has revealed that spontaneous mutagenesis is the result of
many different mechanisms, all affecting in parallel the genetic stability of living
beings. This has most successfully been studied with microorganisms such as
bacteria and their viruses. Theses studies have revealed that in many cases enzymes
are involved in the formation of mutations. Nevertheless, mutagenic reactions
usually do not reproducibly yield predictable products. Rather, the implied enzymes
act as variation generators and the resulting mutations show a certain degree of
randomness both with regard to their location on genomes and with regard to the
time of their appearance. Probably, the intrinsic structural flexibility of biological
molecules also helps to give to mutagenesis an aleatoric nature.
Let us recall that genetic information is contained in filamentous DNA mo-
lecules, in which it is encoded by the linear sequences of nucleotide pairs. Four
different nucleotide pairs are available as letters to compose the linear script of
the genetic information. For convenience and simplicity, we will call here a mu-
tation any change occurring to the inherited nucleotide sequence, independent of
whether such change results in an alteration of the properties, i.e. the phenotype,
of the organism. Bacteria are unicellular organisms and they carry about 5 x 106
nucleotide pairs in their genome which is defined as the total genetic information
of an organism. Mutational changes can affect one or more nucleotide pairs at
once and can be substitutions, deletions, additions or structural rearrangements.
Upon propagation of cells, DNA of course also reproduces. In bacteria, the
DNA replication rate is quite high, and the replication machinery progresses with a
speed of about 1000 nucleotide pairs per second. Once in a while replication lacks
Evolutionary Interplay . .. 161
fidelity, most likely due to the structural flexibility of the involved molecules. This
can for example result in the substitution of one nucleotide by another. Although
some sites on DNA molecules can be affected by this type of mutation somewhat
more often than others, in general practically any location on a genome can be hit
by a mutation due to replication infidelity. This statement applies also to a number
of other mutagenesis mechanisms, so that in the course of long periods of time
many different sites may suffer a mutation.
Mutations are quite often lethal, i.e. the organism is not able to propagate any
longer. Other mutations can give the organism a selective disadvantage. Positive
selection, in contrast, occasionally provides to a mutated organism an advantage
as compared to its parent. The mutant may then eventually overgrow other forms
of life. This kind of selection has been shown generally not to exert a direct
influence on the mutagenesis per se. Rather it acts by the preference given to
the particular mutant form of spontaneous origin. Finally, many alterations in the
nucleotide sequences remain without noticable influence to the organism. These
are called neutral mutations. Some of them may have a longterm implication on
genetic variation.
From these statements we can see that mutagenesis is essential for a steady
progress of biological evolution. However, mutation rates should be lower than
one new mutation per genome and generation. Otherwise the stability of a species
would be endangered, particularly since many randomly occurring mutations are
lethal or partly detrimental. Interestingly, haploid microorganisms with different
genome sizes show often a different mutation rate per nucleotide pair, but have
their mutation probability limited to in between 1% and 1%0 new mutations per
genome and generation. This is usually the result of intrinsic antimutagenic factors
which reduce the genetic instability of the cell.
A number of different enzyme-mediated processes provide means for rear-
rangements of the nucleotide sequences of a genome. This can involve the duplica-
tion or the deletion of a given DNA segment, or also its inversion. Much attention
has been given in recent years to DNA inversion systems, the enzymes of which
interact with specific DNA sequences. They thereby cut the DNA at these sites
and rejoin the free ends in different ways so as to produce a DNA inversion. In
rare cases, rather than to use a specific site on the DNA, the enzymes may also
act on sequences widely deviating from the efficiently used sequences. This then
can give rise to novel gene fusions. Such fusions can be within the sequences of
two different genes and thereby produce a gene encoding a product composed of
functional parts determined by the participating genes. This kind of mechanism
can explain how the genetic information for a particular functional domain of a
protein can become fused with information for other functional domains. As a
matter of fact, many genes are known to be composed of several different func-
tional domains. DNA inversion could thus be an evolutionary strategy to produce
once in a while gene fusions. As long as such rearrangements affect only a few
individuals in large populations and each of them in a different way, the population
as such would not suffer from detrimental consequences but could still profit from
162 Arher
very rare successful mutations providing improved fitness to the organism con-
cerned. Although enzymes are involved in the generation of the mutations, none
of the mutations is specifically programmed but rather appears to some degree at
random.
Mobile genetic elements are defined as DNA segments able to change their
location on DNA molecules. This phenomenon is called transposition and is catal-
ysed by a specific enzyme produced by the element itself. The genomes of bacteria
and higher organisms may each contain several different such elements. Generally
speaking, mobile genetic elements have properties similar to those just described
of DNA inversion systems. Indeed, each mobile genetic element has its char-
acteristic strategy to select insertion target sites on DNA. Some elements prefer
particular sites determined by a specific nucleotide sequence, but they can also use
other targets although they do it with much lower probability. In contrast, other
mobile genetic elements prefer certain regions on DNA molecules for insertion,
but they may integrate practically at random into many different sites within the
preferred region. This allows for a great number of structurally different transposi-
tional DNA rearrangements. For the fitness of an individual suffering transposition,
phenotypic alterations may be either negative or positive, depending on both the
element involved and the target site used in transposition.
Mobile genetic elements which can also contain genes of chromosomal origin
can transpose to so-called natural gene vectors which happen to be in a cell.
A good example for a gene vector is the genome of a virus. Having suffered
transposition, the virus can spread the mobilized genes of its host to other host
strains. This is an example for another important mechanism that contributes to
spontaneous mutagenesis of bacteria. The phenomenon is called horizontal gene
transfer or acquisition of genetic information. Gene acquisition is also possible in
direct contact between two genetically distinct cells in a process called conjugation.
In addition, under particular conditions some bacteria are able to take up free
DNA molecules from the environment upon transformation. All of these processes
facilitate gene acquistion. Again, acquisition is to some degree random both in
time and with regard to the informational content of the acquired DNA. But
DNA acquisition does represent an important source of genetic information for
bacterial strains which did not have a chance themselves to develop in the past an
appropriate genetic function in need.
An impressive lesson on the ecological and evolutionary importance of DNA
acquisition comes from studies of the reasons for the wide appearance of antibiotic
resistances since antibiotic therapies have been introduced in the middle of our
century. It has become clear that many resistance determinants which are today
carried in enterobacteria had been acquired recently from other organisms. Hori-
zontal gene transfer had of course always occurred at low rates and it still does
today. What has changed, however, are the selection conditions for the benefit of
rare cells having acquired resistance genes. This is a good example for our un-
Evolutionary Interplay ... 163
derstanding the natural interplay between mutagenesis and selection. Upon drastic
changes in the selection conditions, rare mutant forms in large populations can
rapidly overgrow the large majority of nonselected cells.
As there is practically no action without counteraction in nature, gene ac-
quisition also encounters natural barriers. These act at various levels, such as by
inhibiting the penetration of foreign DNA or its establishment after uptake. In
addition, functional compatibility between the products of the acquired genes and
the residential biological activities of the recipient cell will be decisive for the
success of an acquisition. Interestingly, practically all bacterial strains have one
or several restriction systems able to specifically recognize if DNA penetrating
into the cell is of foreign origin. In this case, the filamentous DNA is cleaved into
fragments which are then rapidly degraded. But fragments may have some chance
in their short lifetime to become incorporated into the cellular genome. This can
represent an important evolutionary step for the involved bacterium. This type of
acquisition in small steps (since relatively short DNA fragments are involved) is
again to some degree aleatoric, and selection phenomena will eventually decide
on the success of the process.
A tree has become the classical representation of biological evolution. Branch-
ing and the divergence of branches represents the increase in diversity and the ac-
cumulated differences between organisms of two branches. In this view, evolution
progresses steadily as a result of the interplay between mutation and selection.
In view of the possibilities of gene acquisition which results from the hor-
izontal transfer of bits of genetic information between two different branches,
horizontal shunts should be drawn here and there between different branches of
the evolutionary tree. In this way, it would symbolically become clear that future
evolution, as in the past, will also depend not only on further internal development
of existing branches, but also on genetic information which had become available
in any other branch of the tree. The possibility of its horizontal transfer increases
the chance for any type of living organism to make an important evolutionary
step by sharing in the success of others. Indeed, as testified by the spreading of
antibiotic resistances, gene acquisition seems to be a quite efficient evolutionary
strategy, particularly if its efficiency is evaluated per event of transfer. This is
related to the universality of the genetic language and to the fact that a biological
function developed in one kind of organism may have a good chance also to serve
similarly to another organism. As compared with gene acquisition, internal DNA
rearrangement may be of lesser efficiency if again considered per event. This is
due to both the relatively high chance of lethality of a DNA rearrangement and
to the small probability to improve the fitness of the organism by the internal
recombination. Finally, nucleotide substitution may be the least efficient of the
three mechanisms of genetic variation which we have here specifically discussed.
Single steps of nucleotide substitution are, however, important contributions to a
longterm process of developing new biological functions. But it takes many con-
secutive steps to reach from a given DNA sequence a novel biological function.
164 Arber
We are aware that there exist many more molecular mutagenesis mechanisms
than those described here and that each contributes in its specific way to the
production of mutants. In most cases, the generation of mutants appears to be of
fractal nature. However, the potential richness of fractal development is drastically
limited by the effect of natural selection. Interestingly, selection is not constant.
Rather, it is a function both of the physico-chemical environment and the activ-
ities of all living organisms present in the concerned ecological niche. Selection
therefore varies with time and it of course also varies in space. Thus selection may
favor a particular genetic constitution under some circumstances but disfavors it
under other circumstances. This renders ecological studies so difficult. It is to be
hoped, however, that progress in the fractal theorey will also open new approaches
to better understand ecological equilibria and their evolutionary development.
Further readings:
- W. Arber: Mechanisms in microbial evolution. 1. Struct. BioI. 104, 107-111
(1990)
- W. Arber: Elements in microbial evolution. 1. Mol. Evol. 33,4-12 (1991)
- W. Arber: Evolution of prokaryotic genomes. Gene (in press)(1993)
Error Propagation Theory of Chemically Solid
Phase Synthesized Oligonucleotides and DNA
Sequences for Biomedical Application*)
Zeno Foldes-Papp, Armin Herold l ),

Hartmut Seliger and Albrecht K. Kleinschmidt 2)
Sektion Polymere
Universitat Ulm
D-89069 Ulm (Donau)
I) Schering AG, Zentrale Biologische Forschung, Berlin,

2) Universitat Ulm, Germany
Abstract. Our interest was focused on error propagation during the synthesis of single stranded
oligonucleotides and DNA sequences. We have modeled the assumed degenerative effects of such
3'-->5 '-driven syntheses, or vice versa 5'-->3'. Quantification in scaling the theoretically synthesized
yields provided the rationale for our approaches to error propagation of synthesized oligonucleotides
and DNAss sequences. Our homeodynamical model is governed by nonlinear equations; they possess
the relevant features of fractal dimensions. This is significant for implication of a new scale-invariant
property of oligonucleotide synthesis. An inverse power law of driven multi-cycle synthesis on fixed
starting sites is described with a model of growing oligonucleotides in fractal measures. It equates
the constant coupling efficiency do and the constant capping efficiency Po to the length distribution
of sequences produced in the preparation set of oligonucleotides. Attractors in the dynamical model
are included. Each sequence is produced randomly with the probability coupling function d and/or
randomly with the probability capping function p; d and p are independent from one another. For this
general problem we construct effective and computable recursive functions of the relation described
by the inverse power law of driven multi-cycle synthesis. Nonlinear fractal dimensions D(N) are
computed for theoretical deviations from the constant coupling efficiency do. They offer a new access
to the growth dependency on the nucleotides A, G, C, T.
1 Introduction
The chemical synthesis of oligonucleotides and single stranded DNA sequences is
a powerful technique applied in genetic engineering, molecular biology and molec-
ular medicine. Oligonucleotides as primers in polymerase chain reaction, for exam-
ple, are essential for analytic, diagnostic and therapeutic purposes. The chemical
synthesis of oligonucleotides and single stranded DNA using nucleoside-3' -phos-
phoramidites is a cyclic repetition of detritylation, coupling, capping, and oxidation
in organic solvents [4] . We used mathematical formulations of monomers to build
up a polymer structure as a model for directional growth. Multiplicative coupled,
nonlinear equations for which the parameter search is made by experimental ob-
servations were a way to formalize sequence errors. We studied the equivalent
*) In memory of Severo Ochoa 1905-1993

166 Foldes-Papp et al.
of an algebraic instead of a geometric notation [13]. The mathematical model of

a multi-cycle synthesis on fixed starting sites was tested by computer algorithms
for experimental verification.
2 Chemical synthesis
Any mathematical formulation of chemical oligonucleotide synthesis on solid sup-
port is based on the following prerequisites (briefly described here, see also [4]):
It is a driven growth process starting from 3' to 5' direction, or vice versa. The
growing oligonucleotide or single stranded DNA is covalently bound to the sup-
port. We assume average constant conditions of fully tritylated nucleoside loading
of the support. The cyclic repetition of coupling and capping steps during the
growth process is considered as a linear undisturbed flow in first approximation.
Error sequences are shorter than the target sequence, ending with a nucleoside;
they can be separated according to their length. The error sequences are somehow
truncated versions of a target sequence.
3 Mathematical formulation of a multi-cycle synthesis

The system described above acts as a variation generator. «1» represents an arbi-
trary (A, G, T, C) nucleotide/nucleoside in a sequence of oligonucleotides. As a
coupling step either succeeds or fails, we use a second symbol for such missing
nucleotides: symbol «0». Furthermore, we use a symbol for the capping group:
symbol «2». Capping can only be done after a failed coupling step. Hence, symbol
«2» is located only after symbol «0» as the last element of a sequence.
It is convenient to formulate the problem as follows: We consider the variation
V with repetition w of two elements (m = 2) «1» and «0» to (N - l)th power.
N is the integer number of nucleotides/nucleoside of the target sequence. N - 1
gives us the integer number of the last reaction cycle.
(1)
Let us ask how many of these sequences for a given N - 1 obtain either «1» or
«0» as the last element. The answer is
(2)
From eqns. (1) and (2) follows that the number of sequences aN-I for which the
three elements «1, 0, 2» are mixed, is
aN-I = (wVmN-I)
=2
1 (w N
+2 -I)
Vm=2 (3)
= 3 (2 N - 2 ) (4)
(5)
Error Propagation Theory 167
Here (Eqn. (5», I + I is the integer number of nucleotides/nucleoside of target

or error sequences. Since the first element of arbitrary sequences is «I», in fact,
N - I nucleotides/nucleoside are varied. In other words, aN _I is the number
of nucleotides/nucleoside sequences of the last reaction cycle. They have been
included in coupling and capping reactions during the last reaction cycle together
with those that are left with reactive ends after the last reaction cycle.
Now we consider the number of truncated nucleotides/nucleoside sequences
aN-2,E (errors, E) that have been left capped during all previous reaction cycles i
aN-2,E = ~ (~(WV~=2)) - ~ (wV~=21) (6)

1=1
= 2N - 2 - 1 (7)
(8)
Here (Eqn. (8», I is the integer number of nucleotides/nucleoside of error se-

quences.
Hence (by Eqn. (6», the number of all truncated, capped nucleotides/nucleo-
side sequences aN - I ,E is
N
aN-I,E = L (~ (WV~=2)) - ~ (wV:=2) (9)
1=1
= 2N - 1 -1 (10)
(11 )
including those of the last reaction cycle.

In view of eqns. (3) and (6), the total number of sequences aN,S (sum of error
and target sequences, S) is presented
(12)
( 13)
=t(~)
1=1
(14)
Because of identical serial arrangements of nucleotides/nucleoside, the total num-

ber of sequences aN,S is corrected to a lower value aN,S(corr) after cleaving off the
capping group
aN,S(corr) = (WV:=21) + ~ (~(WV~=Z)) (15)

1=1
- (~ (~ (WV~=Z)) - ~ (WV:=Z) )
=2N - 1 (16)
(17)
These formulations make it possible to generate all integer numbers of sequences

that are available in the synthesis. The nonlinear equations are typical for scale
invariance [12] formally controlled by means of the scale exponent.
A system of equations is used for yields of sequences in dependence on
the constant coupling efficiency do as well as the constant capping efficiency Po
in order to show the self-similarity of serial nuc1eotides/nuc1eoside arrangements.
Here is the rule: Write all yields of sequences (A/N) in dependence on the constant
coupling (do) as well as capping (Po) efficiencies until you get the yield of the
target sequence (ANN). Hence, we have eqns. (18) and (19)
(18)
where 1 = {I,2, ... ,N}, N = {2,3,···} and
(N
1-
-2)1
(19)
The equations for yields of sequences are converted into a vectorial matrix format
(eqns. (20) to (24))
(20)
(1 - do)N-1 . (1 - PO)N-I
(1 - do)N-1 . (1 - PO)N-Z . Po
(1 - dot- z . (1 - PO)N-3 . Po
(21 )
(1 - do)Z . (1 - Po) . po
(1 - do) . po
(22)
(23)
(24)
Each vector element a/-IN-I corresponds to a unique element a (a/-IN-I),

where a is a linear operator [5]. A model is obtained which is scale-invariant
by vectorial definition. The constructed vector spaces have Ndimensions. Each
vector can be written according to Eqn. (25)
-
a (-
a/-IN - I ) -
= a/-IN-I . --
=
do alN (25)
By serial nucleotides/nucleoside arrangements of target and error sequences this

system embodies a higher-order fractal structure characterized by nonlinear fractal
dimensions.
The moment of probability [7][11] of serial nucleotides/nucleoside arrange-
ments (M/, q=l) is derived from the total number of sequences aN,S presented
after a given number of coupling cycles, and from the proportion of target and
error sequences. The proportion of sequences depends on the constant coupling
efficiency (do) and constant capping efficiency (Po). The moment of probability of
serial nucleotides/nucleoside arrangements during chemical synthesis of a target
sequence is
(N
1-
-2) 1
(26)
where 1 = {1,2, .. ·,N}, N = {2,3, .. ·}. P(l, N) is a probability matrix (not

shown). Hence, the frequency distribution of micleotides/nucleoside arrangements
is described in the synthesis of a target sequence

Mt,q=1 = f(l,N) (27-a)
=(~)-(I-I){(
do
)-1 1
(1- do)po
+ (~) - I . N-(I+I) ((N _(k + 1)) (_1_) -(N-(I-J+k)) (_1_) -(N-(l+k)))

po L
k=1
1-1 I-do I-po
+
(N-I)(
1-1 1
(I-do)(I-po)
)-(N-l)}
MN-I,q=1 =f(N -I,N)

1 ) -(N-2) { (
= ( do
1
(I-do)po
)-1 (27-b)
+(
N-I)(
N-2
1
(I-do)(I-po)
)-I}
MN,q=1 = f(N,N) = ( do
1) -(N-I) (27-c)
where
N = number of nucleotides/nucleoside of a target sequence
do = constant coupling efficiency
Po = constant capping efficiency
An inverse power law of driven multi-cycle synthesis on fixed starting sites
(IPLCS) is found (Eqn. 27-a,b,c). The linear operator a of vector space SN-I
corresponds to the function F, where F is equal to
F = f (1, N) - f (1- I, N - 1) (28)
It is improbable that d and p are constants; d and p depend on internal conditions
(e.g., concentrations of reactants, clustering, conformer variability, depurination,
artificial branching, side reactions) and on external conditions (e.g., support fixed
starting sites, temperature, pressure, time). We previously defined i as the integer
number of reaction cycles (i = {I, 2, ... ,N - I}), u and v are special internal or
external conditions. Let d = f (i, u) and p = f (i, v) in the Euclidean plane 1R2, the
coordinate plane of calculus, then {1R2, t} is a dynamical system of the iterated
function system theory [1]. A simple example is given: Multivariate nonlinear
analysis of factors do and po is equal to the «sums of exponentials» [6][3]
d=d o (taje-f3i(ii-I)) (29)

J=I
P = Po (t Tje-Oi(ii-I))
J=I
(30)
where i l = {I, 2,'" ,g}, i2 = {g + 1"" ,h}, i3 = {h + 1,'" ,N - I}. Sum

coefficients are rand s; Clj, {3j' 'Yj, OJ represent scaling factors. Parameters do
and po are any constants of coupling and capping efficiency, respectively, in first
approximation. Into the original relation described by the IPLCS, d and p are
introduced as new functions. The analysis of this general case shows that the
numbers of any nucleotides/nucleoside arrangements and the exponents of their
mathematical expressions do not change. The inverse power law of driven multi-
cycle synthesis on fixed starting sites has only to be modified in terms of integer
numbers of reaction cycles and scaling factors. Assuming that «blocks in error» or
«error clusters» influence d and p, phenomena of a multi-cycle synthesis on fixed
starting sites may be related to the self-similar error cluster model of Mandelbrot
[8], though for other purposes.
A priori, the regime of a multi-cycle synthesis on fixed starting sites is limited
to values of d, or d and p less than 1.0 by acting of attractors. A fractal set is a
distribution function (see [13]) which gives an inverse power law (see [13][1][9]).
The IPLCS shows a logarithmic progression, which can be expressed in terms
of fractal dimension D. The fractal dimension D depends on the errors of the
growing chains.
Setting In (L~ll MI, q=l) = Z + B In (N) and according to [2] B = 2 - D
then we have
N-I
LM,q=1 - C
_ (1N )D-2 (31 )
1=1
Z is an intercept and B is the slope of a log-log plot L~ll M,q=1 versus N
[10]; C is given by eZ . The values of fractal dimension D can be computed from
Eqn. (31). The fractal dimension obtained from measuring data should allow to
recognize all deviations from constant coupling efficiency do. This position can
be made into a common expression. Substituting from Eqn. (29) into Eqn. (31)
gives us
In {I - (dO)N-I (tl 1.Cl7- exp (-!!t (N - 2) (N - 1)) ) }+ In (C)

D = 2- In(N)
(32)
Clj (0 < Cljd o < 1), {3j are given by scaling procedures and i j = {I, 2, ... , N -I}.
When the distribution shown by Eqn. (31) is worked out for increasing values of
N, the steps in the curve will become smaller and smaller until finally, with infinite
N, a continuous curve is obtained. In continuous distribution we differentiate Eqn.
(31) to eliminate C. This relation yields a general expression of D
(33)
The fractal dimension D(N) we discuss measures properties of sets of points

determined by Eqn. (33) in a log-log plot 1 - MN,q=1 versus N. Hence, we
discuss the functions which are defined by the independent variable N. The fractal
dimension D(N) is thus generalized. 2:::1:11 ~,q=1 can be directly measured
by high performance liquid chromatography or capillary electrophoresis. Eqn. (33)
gives us a new theoretical access to the growth dependency on the nucleotides A,
G, C, T. Without substantial change we may consider sets of points that are defined
by functions of 1 at N = canst.
4 Conclusions
Error propagation of a multi-cycle synthesis of oligonucleotides on fixed start-
ing sites can be theoretically regarded as a directed growth model with fractal
measures. We use a model in which target and error sequences are formed by
repetition of building blocks. Each sequence is produced randomly with the prob-
ability coupling function d and/or randomly with the probability capping function
p; d and p are independent from one another. The inverse power law of driven
multi-cycle synthesis on fixed starting sites has only to be modified in terms of
integer numbers of reaction cycles and scaling factors. For this general problem
we construct effective and computable recursive functions of the IPLCS. The the-
oretical yields of serial arrangements of sequences are indispensible for planning
of experiments; the real growth process of synthesized short and long oligonu-
cleotides is still arcane. The paper shows that the mathematical formulation of
fractalistic error propagation in fractal dimensions D is a basis for new theoreti-
cal and experimental qualification (in preparation: I Theoretical considerations; II
Application to experiments).
Acknowledgment:
Supported by DFG, Graduiertenkolleg «Biomolekulare Medizin» and Stiftung zur
F6rderung der molekularbiologischen Forschung, Univ. Ulm. For advice in math-
ematical symbols we thank Dr. Gerd Baumann, Dept. of Mathematical Physics,
Univ. Ulm.
References
[1] Barnsley M: Fractals Everywhere. Academic Press, New York (1988)
[2] Batty M: Cities as fractals: Simulating growth and form. In: Fractals and
Chaos. AJ Crilly, RA Earnshaw, H Jones (eds.), Springer, New York, pp.
43-69 (1991)
[3] Beechem JM: Global analysis of biochemical and biophysical data. In: Meth-
ods in Enzymol. 210, 37-54 (1992)
[4] Caruthers MH, Beaton G, Wu JV, Wiesler W: Chemical synthesis of deoxy-

oligonucleotides and deoxyoligonucleotide analogs. In: Methods in Enzymo!.
211 , 3-20 (1992)
[5] Curtis ML: Abstract Linear Algebra. Springer, New York, pp. 9-46 (1990)
[6] Johnson ML, Faunt LM: Parameter estimation by least-squares methods. In:
Methods in Enzymo!. 210, 1-37 (1992)
[7] Losa GA, Baumann G, Nonnenmacher TF: The fractal dimension of pericel-
lular membrane from lymphocytes and lymphoblastic leukemic cells. Acta
Stereo!. 11, Supp!. 1, 335-341 (1992)
[8] Mandelbrot B: Self-similar error clusters in communication systems and the
concept of conditional stationarity. IEEE Trans. on Communication Technol-
ogy, vo!' COM-13, 71-90 (1965)
[9] Marek M, Schreiber J: Chaotic Behaviour of Deterministic Dissipative Sys-
tems. Cambridge Univ. Press, Cambridge (1991)
[10] Parker TS, Chua LO: Practical Numerical Algorithms for Chaotic Systems.
Springer, New York, pp. 167-199 (1989)
[11] Small EW: Method of moments and treatment of nonrandom error. In: Meth-
ods in Enzymo!. 210, 237-279 (1992)
[12] Vicsek T: Fractal Growth Phenomena. World Scientific Pub!., Singapore, pp.
9-46 (1989)
[13] West BJ: Fractal Physiology and Chaos in Medicine. World Scientific Pub!.,
Singapore (1990)
Fractional Relaxation Equations for
Protein Dynamics
Walter G. GlOckle and Theo F. Nonnenmacher
University of Ulm
Albert-Einstein-Allee 11
D-89069 Ulm, Germany
Abstract. Due to a large amount of conformational substates, relaxation processes in proteins are
governed by many time constants and therefore, they decay more slowly than a Debye relaxation. For
processes occurring on different time scales in a self-similar manner, we derive and solve a fractional
order differential equation for the relaxation function. Solutions of this well-posed initial value problem
are given in terms of a Mittag-Leffler function. Applications to ligand rebinding data of myoglobin
are presented leading to a 3-parameter fractional model.
1 Introduction
In a large variety of biological systems, non-exponential relaxations are observed.
Their relaxation function deviates from the Debye relaxation
¢(t) = ¢oexp (-~) (1)
and it shows a decay which is slower than the exponential, e.g. in terms of a
stretched exponential
(2)
or an inverse power law
(3)
for large t. In the last decades, among others, proteins were intensively studied
from a physical point of view. According to their biological function, two states of
a protein are distinguished, e.g. an ion channel can be open or closed, a hemoglobin
or myoglobin protein can have bound oxygen or not. The transition between the
two states is considered to be a relaxation process where the relaxation function
is the probability of a protein to be in the «excited state».
The slow relaxations are often modeled by a superposition of Debye relax-
ations. If the two observable states consist of many substates, so-called conforma-
tional substates, with different relaxation times, the superposition of these single
exponential processes leads to a slow relaxation. To model a power law decay.
however, a large amount of such Markov processes is necessary. Therefore, in the
last years concepts known from the description of polymers or glasses [1] have
been applied to proteins. Doster, Cusack a. Petry [2] e.g. used the mode coupling
Fractional Relaxation Equations 175
N(t)
60K
80K
10- 2
lOOK
10 - 6 10- 3
t [ ]
Fig. 1 Rebinding of CO to myoglobin after photodissociation [5]. The solid lines correspond to a
fit with the solution (25) of the fractional equation (26) with (3(T) = 0.41 T / 120K and ro(T)
given in fig.3.
theory of the glass transition to interpret inelastic neutron scattering data of pro-
teins. Millhauser [3] applied the reptation model to estimate time constants of the
dynamics of ion channels. Nonnenmacher [4] applied renormalization group ideas
leading to the algebraic decay of the distribution function for open times of ion
channels and to the observed oscillations around the power law trend.
Here we will consider the ligand rebinding in myoglobin after flash photo-
dissociation [5]. Because of the well-known structure of this relatively simple
protein, myoglobin serves as a model system to study protein motions and reac-
tions. After the dissociation, the ligands are rebound to the heme iron and the
fraction of the free ligands N (t) decays. In fig.l measurements at different tem-
peratures [6] are shown. The decay is much more slowly than the exponential and
the function
N(t) = (1 + t/r) - n (4)
is often used to fit the data with temperature dependent parameters r and n [5].
Alternatively the deviation from the exponential law is regarded to be a
consequence of the large number of conformational substates leading to many
different reaction rates or time constants [6]. Therefore, the reaction kinetics is
described by using a distribution g(E) of energy barrier heights. Then N(t) is
176 GlOckle and Nonnenmacher
1
given by
N(t) =
00
g(E) exp( -k(E)t)dE (5)
where for k (E) an Arrhenius law
k(E) = Aexp(-EjRT) (6)
is assumed. Instead of g( E) the distribution b( k) of reaction rates k
N(t) = Jor RT
b(k)Texp(-kt)dk (7)
can be used. The two distributions are connected by g(E) = b(Aexp(-EjRT)).

If we consider observation times t much greater than the time scale of molecular
motions 1j A, A can be substituted by <Xl in (7). Thus, N (t) is the Laplace transform
of b(k)jk:
(8)
Here we will use fractional order differential equations to describe the dy-
namics of proteins. Utilizing ideas of self-similar dynamical processes we will
derive a fractional order initial value problem for the relaxation process.
2 Self-similar dynamics
Since many conformational substates are accessible to a protein, the protein system
is not relaxing with a single time constant T but with many time constants Tn
leading to
N(t) = Lwnexp(-tjTn) . (9)
n
Self-similar dynamics means that the weights Wn and the time constants Tn are
not independent but correlated. If the time scale is changed by t - t At (,\ > 1),
the behavior of the system should be the same apart from a renormalization of the
statistical weight (w - t w).
In case of the Bernoulli scaling
r
Tn';" ,\nTO (,\ > 1) (10)
Wn = pnpO = (,\-/3 Po ({3 > 0, p < 1) (11)
with infinitely many processes, one obtains in the continuum limit the relaxation
function [4]
Po
Pn
o 1
Fig. 2 Statistical weights Wn = an / b~ of the fractional scaling model in comparison to pn of the

Bernoulli scaling model for 1/ 2 < (3 < I.
expressible by the incomplete i-function. For large t an algebraic decay
N(t) '" Pof({3) (~)-(3 (13)

In.x. TO
is found. This continuous Bernoulli scaling can be translated to the description by

energy barrier heights leading to
g(E) = poexp (-~~) (14)
which is independent of the temperature T if the parameter {3 is proportional to T.

In order to give a connection to the fractional relaxation equation, we consider
a slightly different self-similar process, here called fractional scaling. Contrary to
the Bernoulli scaling, the weight Wn is not proportional to pn = .x. -(3n but modified
in the form
Tn = .x.nTO (15)
pn sin(7r{3)
Wn = po p2n + 2pn cos (7r{3) + 1 (16)
In fig.2 the weights Wn = an/b~ of the fractional scaling model are geometrically
compared with those of the Bernoulli scaling model (w n = pn). For large values
of n, bn tends to 1 so that Wn '" pn sin( 7r (3), i.e. for large n the two models are
equivalent. In terms of the renormalization group theory, the two models belong
to the same «class of universality» where the universality in these time processes
is the asymptotic power law decay with the power -{3.
178 GI6ckle and Nonnenmacher
For the fractional scaling, the energy barrier distribution is given by
e-f3E/RT sin(7r,6)
g(E) = po e-2f3E/RT + 2e-f3E/RT cos(7r,B) + 1 (17)
and the corresponding distribution of reaction rates reads as
b(k) = 0 ( kTo)f3 sin(7r,B)

(18)
p (kTo)2f3 + 2( kTo)f3 cos( 7r,B) +1
3 Fractional equation
In order to derive the fractional order differential equation connected with the
self-similar process, we calculate the Laplace transform
Q(p)
-p-= C(N(t) , p) (19)
of the relaxation function N(t) . Because of (8), we can use the Stieltjes transform
S = CC given by
S(f(k), p) = 10 :~~ dk
00
(20)
to calculate Q(p) from the rate distribution b( k). With (18),
1
Q(p) = No (TOp)-f3 + 1 (21 )
is found with No = RT7rpo, where the fractional power ,6 indicates a fractional

order differentiation or integration.
A proper initial value problem is obtained by considering the type of the
relaxation process. With respect to the photodissociation experiment, there is a
constant amount of free ligands because of an external force E (t) for t < O. At
t = 0 the force is switched off and the number N (t) decays. For this type of
relaxation, (21) leads to the fractional integral equation
(22)
by inverse Laplace transform. In (22), oD;f3 is the Liouville-Riemann fractional

integral operator [7] with the order ,6 being equal to the parameter ,6 of the
self-similar process (15,16). By applying the inverse operator, i.e. the fractional
differential operator oDf, on (22), the fractional differential equation
f3 t- f3 _ -f3
oDt N(t) - NOr(1 _ ,6) - -TO (N(t) - xoE(t)) (23)
To(T) [8]
10- 3
50 100 150
T[K]
Fig. 3 Dependency of To(T) used in fig. J. The solid line is an Arrhenius curve (28) with E* = 1470K
and Tm = 3.4.10- 105 .
is attained from which for j3 -+ 1 the ordinary relaxation

d 1
-d N(t) = --(N(t) - XoE(t)) (24)
t TO
results.
The solution [8]
(25)
for times t > 0 follows from the initial value problem

TO(3 oD;-(3N(t) + N(t) - No = 0 . (26)
In (25), E(3(x) is a Mittag-Leffler function [9] from which the asymptotic decay
N(t) rv (t/T)-(3 for 0 < j3 < 1 can be derived. For j3 = 1 we recover the
exponential function.
In fig. 1 the fractional relaxation function (25) is applied to the ligand rebind-
ing to myoglobin. The fractional function fits the data better than the power law
(4). The temperature dependency of j3 is taken proportional to T
j3(T) = 0.41 T (27)

120K
180 GlOckle and Nonnenmacher
60K
SOK
10- 2 -
lOOK
10- 3
t [8]
Fig. 4 3-parameter model (29) for the rebinding of CO to myoglobin with Tm = 8.4 . 10- 108,
Q = 3.5 · 1O- 3K- 1, and k = 130.
and To(T) is determined by a least square fit. As demonstrated in fig.3, To(T)

agrees well with an Arrhenius law
To(T) = Tmexp (~) (28)
The activation energy E* = 1470K=12.3kJ /mol is of the order of a characteristic

energy barrier height Epeak ~ 1O .. 11kJ/mol [10] and Tm = 3.4 .10- 10 5 is a time
constant of molecular motion.
Because of these temperature dependencies, the fractional equation (26) can
be transformed to
k oD;!3N(z) + N(z) - No = 0 (29)
with the dimensionless time z = t/Tm, the fractional order (3 = aT, and a constant
k which is independent of the temperature. In fig.4 this 3-parameter model is
applied to the Mb-CO rebinding with Tm = 8.4 . 10- 10 5, a = 3.5· 1O- 3 K- i ,
and k = 130( =E* = 11.5kJ /mol). For temperatures T lower than the glass
temperature Tg ~ 180K [11] the dynamics and the temperature dependency are
well-described by the 3-parameter model.
In this paper we considered self-similar processes in which each time scale
occurs in a self-similar manner and therefore, no internal time scale exists. These
processes form an analogue in the time domain to objects with fractal geom-
etry in space. We demonstrated that fractional relaxation can be deduced from
a self-similar process. Especially a connection between the statistical weight of
the processes and the fractional order of the differential equation has been es-
tablished. The fractional order initial value problem was applied to the kinetics
of ligand rebinding to myoglobin after photodissociation. It turned out that for
low temperatures the time and the temperature dependency is reproduced by a
3-parameter fractional model.
Acknowledgement:
This work has been supported by Deutsche Forschungsgemeinschaft (SFB 239,
Project C8).
References
[1] A. Blumen, J. Klafter, a. G. Zumofen, in Optical Spectroscopy of Glasses,
ed. by I. Zschokke, Reidel, Dordrecht (1986) 199.
[2] w. Doster, S. Cusack, a. W. Petry, J. Non-Cryst. Sol. 131-133 (1991) 357.
[3] G.L. Millhauser, Biophys. J. 57 (1990) 857.
[4] T.F. Nonnenmacher a. DJ.F. Nonnenmacher, Phys. Lett. A 140, 323 (1989)
and in Stochastic Processes, Physics and Geometry ed. by S. Albeverio,
G. Casati, U. Cattaneo, D. Merlini, a. R. Moresi, World Scientific, Singapore
(1990) 627 .
[5] R.H. Austin, K.W. Beeson, L. Eisenstein, H. Frauenfelder, a. I.e. Gunsalus,
Biochemistry 14 (1975) 5355.
[6] H. Frauenfelder, F. Parak, a. R.D. Young, Ann. Rev. Biophys. Biophys. Chern.
17 (1988) 451.
[7] K.B . Oldham a. J. Spanier, The Fractional Calculus, Academic Press, New
York (1974).
[8] w.G. Glockle a. T.F. Nonnenmacher, Macromolecules, 24 (1991) 6426.

[9] A. ErMlyi (edt.), Bateman Manuscript Project, Higher Transcendental Func-
tions I-III, McGraw-Hill, New York (1953).
[10] R.D. Young a. S.F. Bowne, l Chern. Phys. 81 (1984) 3730.
[11] I.E.T. Iben, D. Braunstein, W. Doster, H. Frauenfelder, M.K. Hong, lB. John-
son, S. Luck, P. Ormos, A. Schulte, PJ. Steinbach, A.H. Xie, a. R.D. Young,
Phys. Rev. Lett. 62 (1989) 1916.
Measuring Fractal Dimensions
of Cell Contours:
Practical Approaches and their Limitations
Gerd Baumann, Andreas Barth and Theo F. Nonnenmacher
University of Ulm
D-89069 DIm
Abstract. We discuss practical methods to examine fractal properties of electron micrographs and
outline a procedure to extract information from a gray scale image (EM-image) by using different
filtering methods. The data obtained from this reduction of information is used as a basis for numerical
calculations of fractal dimensions D. To determine D of a given object, we use the yardstick, the box
counting and the probabilistic methods. These three approaches will be critically discussed and the
finite scaling range for natural objects such as cells will be examined. The method of digital image
analysis as discussed here incorporates an algorithm that detects self-similar domains in the structure
of cells. We apply this method to determine the fractal dimension of cell contours and we discuss
advantages and limitations of these methods. As prototype examples, we investigate cellprofiles of
lymphocytes and lymphocyte leukemic cells.
1 Introduction
In close analogy to Richardson's examination of borders for different European
countries [1], we discuss practical methods to examine fractal properties of electron
micrographs and outline a procedure to extract information from a gray scale
image (EM-image) by using different filtering methods. The data obtained from
this reduction of information is used as a basis for numerical calculations of fractal
dimensions D. To determine D of a given object, we use the yardstick, the box
counting and the probabilistic methods. These three approaches will be critically
discussed and the finite scaling range for natural objects such as cells will be
examined. The method of digital image analysis as discussed here incorporates
an algorithm that detects self-similar domains in the structure of cells. We apply
this method to determine the fractal dimension of cell contours and we discuss
advantages and limitations of these methods. As prototye examples, we investigate
cell-profiles of lymphocytes and lymphocyte leukemic cells, as a first step into
fractal analysis of cell profiles [2][3]. The characteristic property of boundaries
Richardson and later on Mandelbrot found is the self-similarity of the border. For
true fractals this means scaling with a common exponent on different length
scales ranging from 0 to infinity. The corresponding scaling exponent, the fractal
dimension D describes the deviation of the object from the euclidean dimension in
which it is embeded. For some irregular and wrinkled curves in the plane one will
determine the fractal dimension greater than 1 and less than 2. In our studies [5],
we will restrict our considerations to sets in the plane because our experimental
data are points of a two dimensional picture.
Measuring Fractal Dimensions of Cell Contours 183
2 Image Extraction Procedure

Our examination is based on EM-pictures with a magnification of 8000 x. The
negative is turned into a positive at a magnification of 18400 x. This print is
used to extract the contour of a cell. Two different methods are taken to get the
boundary of the cell: first, a computer based procedure and second, a manually
based one.
Both procedures use a scanner and a contour tracing algorithm. The computer
procedure is based on EM-prints. The print is first discretized at a resolution of
300 dpi in X and y directions. The amplitude of each picture element (pixel) is
represented by a 8 bit data word. This representation discretizes the gray value of
the original picture in 256 gray levels. The amplitude resolution produces about
4 MB of data depending on the size of the picture. The next step in deriving the
cell's contour is the application of a filter. Two types of filter were used in our
examinations: i) gradient filters and ii) band limited filters.
In a first test, we applied a standard gradient and a Laplace filter on the
discretized picture and detected edges of steep, gray value changes. The results
from these two filters were rather poor and contained many errors. By using a
different filter operator, the so-called Sobel operator, we managed to get good
results for our work. Depending on the print quality, the contours derived from
this filter type may be broken in which case a further processing of the picture
becomes impossible. Yet, a Sobel operator acting in x and y directions gives a
good representation of the cell contour. A main problem with this filter which
works above a certain gray value threshold is that the cell contour is not closed.
This defect is the main obstacle in isolating the cell contour from a gray scale
picture.
The second filter we used was a Marr-Hildreth or Mexican hat filter [4]. This
type of filter uses a specific gray value distribution located around a pixel to detect
edges in the picture. The results produced by this filter are as good as the results
obtained by a Sobel filter. Again, the main problem is a broken contour line and
a broadening of the edges. The broadening of the edges results in a small loss
of information on the contour structure. This error originates from the assumed
GauB-distribution of gray values around a pixel.
Once a closed contour line is successfully extracted with one of the above
filters, we can go on to the next step in processing the image. In the computer
based procedure the object now needs to be isolated from its background. There is
no standard method with which to isolate the cell under consideration. Normally,
we auto fill the objects by hand. The largest filled object is then selected by the
computer program and used by the contour tracing algorithm. At times, it may
be necessary to remove parts of the picture by using the delete functions of our
program. If the contour is closed and if we succeeded in deleting the neighbour-
ing cells from the picture, we can apply the last step of the contour extraction.
This step is fully automatic and uses the contour tracing algorithm to isolate the
(x, y)-coordinates from the binary picture. This step is necessary since both filter
184 Baumann et al.
procedures widen the contours and consequently may cause errors in the ensuing
analysis.
The second method which is in part a manually executed method is in compar-
ison with the computer-based method a very simple, reliable yet time consuming
method. In this procedure, we extract the contour from the EM-prints by means of
a hand drawing. Parchment paper and a sharp Faber 2H pencil suffice to extract
the contour from the print. Many different «artists» tested this method with good
results. If the cell contours are drawn accurately, the results are as good as or bet-
ter than from the computer-based method. EM-pictures always show some faded
regions which a computer cannot interprete correctly. One of the benefits of the
«artistic» method is a closed contour at the end of the drawing. Once the contour
of a cell is drawn, we scan the drawing and apply the contour tracing algorithm
on the picture to get the binary representation of the cell border. At this stage of
the analysis, we extract the contour of the cell in binary form and the coordinates
of the contour are known. Different methods can now be used to detect scaling
behavior in the geometric structure of the cell contour.
3 Calculation of Fractal Dimensions

There are two main methods to find scaling behaviour in natural objects: first, the
two metric methods and second, the probabilistic method.
The metric methods are the yardstick method and the box counting method.
Both methods measure the perimeter of a cell by using the geometric properties of
the contour. Each element in this measuring process has the same weight. Contrary
to the property of equal weight, the probabilistic method allocates a certain weight
to each measure. However, all three methods give approximations of the perimeter
on the length scale of the yardsticks used.
3.1 Yardstick Method

One of the simplest and fastest methods is the yardstick method. The scaling or
fractal dimension 0 for an object with a closed contour is derivable from the
numbers N (€) of yardsticks with a length € needed to cover a cell border. Already
Richardson applied this method for estimating border lengths between various
countries [1]. The length is given by
(1)
This perimeter tends to increase as € decreases. For natural objects there exist
lower and upper limits of the yardstick length €. The lower limit in our computer
studies is given by one pixel. One possible characteristic for the upper limit is
given by the radius of the cell. If scaling behaviour is located between these two
limits, such behavior is characterized by the fractal exponent D. For true fractals
such upper and lower limits do not exist as € ranges ideally from 0 to infinity. The
-. I
1-dlm approximation
~~~
** ~~
¥
x
I
non averaged
+ averaged
30 >f.¥
~;y:
*~ x
*x+ macro-
+
- -----_.-- scopic
plateau
20
I
10- 1
Fig. I Typical measunnents with the yardstick method.
term true fractal also means that N (E) diverges as L -+ 00 if E -+ O. For fractals,
we thus get
(2)
and consequently
(3)
In order to estimate the scaling exponent for a true fractal, it suffices to count the
number of yardsticks at different yardstick lengths E and to find the limit
D = lim 10gN(E). (4)

HO log(l/E)
In practical applications, the number of yardsticks N( E) at a fixed length E ranging

between lower and upper limits has to be counted. Then the logarithm of both
quantities N (E) and E is taken and their values are plotted in an (x, y)-plot. If the
natural object possesses the scaling property, a straight line in the log-log-plot
can be detected. The slope of the scaling region is directly related to the scaling
exponent or fractal dimension D. It is clear from figure 1 that this procedure is
only reliable if a scaling region is detected in the plot. To find the scaling region,
we used an automatic fit procedure in our studies to check the confidence level
within which this region is being located. Checking the confidence level guarantees
the reproducability of the results. Using a standard linear regression program, all
186 Baumann et al.
points of the calculation have the same weight in the fit procedure and as a result,
the total 1'0 range is assumed to possess scaling behavior. A second source of errors
in this method is the closing gap at the end of the tracing. The closing error is a
typical error of the yardstick method. The error occurs if the yardstick length does
not match the contour length exactly. This error is very large for large yardstick
lengths and shows up in large fluctuations in the number of yardsticks for larger E.
To eliminate this effect, we use all discrete points of the contour as starting points
when measuring the length. The calculation of an averaged number N (E) by
(5)
will suppress these fluctuations. Nmax is the total number of contour points. Nj (E)
corresponds to the number of yardsticks for each starting point i. The scaling
exponent D is then calculated by
D - · 10gN(E)
I1m ---"--:---7--'- (6)
- 10g(I/E) ,
f--+fl
with 1'0\ the lower limit of the scaling region. While the yardstick procedure is
a very fast method, its problems are the detection of the scaling region and the
closing error at the end of the contour. However, the closing error can be avoided
and the scaling range extended by using the box counting method.
3.2 Box Counting Method

The box counting method is based on the division of a plane into squares of
edge length E. The box counting method also delivers an estimate of the length
of a contour by counting the number of boxes N (E) of a given size Eo Each
box containing at least one pixel is counted in N(E). This number is roughly
proportional to the number of N needed by the yardstick method. Starting with
the smallest 1'0 scale (one pixel), the grid length 1'0 is increased successively to about
1300 pixels depending on the size of the EM-print. Again in a log-log-plot of N(E)
versus 1'0, a scaling range for self-similar structures is obtained (see Fig. 2). The
scaling dimension D is again given by (4). Since the covering of the plane with
a grid of edge length 1'0 is somewhat arbitrary, we have to use several locations of
the grid's origin to eliminate this arbitrariness. For each edge length 1'0, which is
an integer value, we use 10 2 points as starting points to calculate a mean number
N(f) by
(7)
whereby Nj(f) counts the number of occupied boxes of the grid for each starting
point i. Thus the dimension D can be estimated by using the described fit procedure
in the sense of equation (6). It is clear from this procedure that after selecting the
L(e:)/I-\m
* geometric
limit x non averaged
lit + averaged
x
macro-
~* ++- SCOPIC
,...--r-TT--r----,-t--r-r---r-.........-r----ir---r-x-r-rT Ii mit
10- 2 10 1
e:/j.lm
Fig. 2 Measunnents taken by the box counting method for a contour line with a finite width. If the
width of the boundary is reduced to a one pixel line, we observe the same behavior as shown
in figure 1.
scaling range, only reliable estimates for L are obtained. The box counting method
does not show any closing errors and thus the scaling range is larger than in the
yardstick method. The upper scaling limit E2 is again given by the macroscopic
extension of the cell. The lower scaling limit EJ depends on the method of analysis
and the details of preparation.
As seen in figures 1 and 2 both methods possess for large and small yardsticks
E limiting values. As mentioned above the plateau value for large E'S is determined
by the macroscopic extension of the object (cell). The limit for small E depends
on the method used. For the yardstick method one observes a plateau which is
typical for the one dimensional approximation of the contour line. In case of the
box counting method the limiting behaviour for small E'S is also a plateau if the
width of the contour line is one pixel (not shown in fig. 2). This plateau changes to
a straigth line with a slope near one if the width of the contour line is some pixels.
Such a situation is shown in figure 2. The line with slope around one is infact a
consequence of the two dimensional approximation of the contour line. The limits
for small E'S occur since the structural variance of the boundary disappears for
these yardstick lengths. The scaling range between EJ and E2 is thus a function of
the roughness of the contour line too.
188 Baumann et al.
3.3 Probabilistic Method

The third method used in our experiments is the probabilistic method. In deter-
mining the scaling dimension D by the two methods discussed so far, all squares
or yardsticks needed to cover the cell border are equally weighted even if the
number of pixels with which a border visits a box may be different. Lacking other
information, any point of the contour is assumed to be equivalent and is an equally
probable origin for our analysis. The spatial arrangement of the points in the dis-
cretized image along the cell border is used to determine the probability P(m, E).
The distribution P (m, E) gives the probability to locate m points within a square
of size E which is centered on an arbitrary point along the contour line. Again, the
origin of the grid which covers the picture is arbitrary and thus we determine the
number of boxes of edge length E containing m pixels as an average over different
locations of the origin. The scaling behaviour of the moments
N
Mq = L mqP(m , E) (8)
m=1
of the probability distribution P(m, E) can be used to determine scaling exponents

Dq which are given by
D = ~ log Mq(E) (9)
q q lOgE .
Different Dq values for the moments indicate multi-fractal behaviour of the cell
profile.
4 Results
First, we tested the three methods as to their accuracy and reliability in determining
the fractal dimension D of an exact self-similar construct. In these checks, we
changed the resolution of representation on a grid of the fractal. The yardstick
method was tested with a Koch curve. The result is that with increasing resolution
R (number of pixels) this method approaches the exact scaling dimension Dexact =
1.261 ... from below. For instance, taking R = 1200 x 1200 pixels, we get D =
1.22 and for R = 2400 x 2400, we get 1.238 as scaling exponents. The box
counting method and the probabilistic method were tested with a Sierpinski triangle
(D exact = 1.5849 . .. ). The results are collected in table 1. Two points become
obvious: i) the higher the resolution R, the more accurate the value D, ii) the
probabilistic method approaches Dexact from above while the box counting method
approches Dexact from below. The results listed in table 1 are clearly in favour of
the probabilistic method since the errors are about a factor of 10 smaller than for
the box counting method. However, the computer time for one run is larger by a
factor of three.
To test the scaling hypothesis for cell boundaries, we examined lymphocytes
in four different states: stimulated, unstimulated, healthy, and disease states. The
R error(PM) D BeM error (BCM)
1024 x 1024 1.63 2.84% 1.49 -5.7%

1124 x 1124 1.63 2.84% 1.51 -4.6%
1526 x 1526 1.62 2.21% 1.53 -3.3%
1824 x 1824 1.60 0.95% 1.54 -3.0%
2048 x 2048 1.59 0.321 % 1.54 -3.1%
2524 x 2524 1.58 0.309% 1.54 -2.6%
Table 1 Detennination of the fractal dimension by the probabilistic method (PM) and the box counting
method (BCM) for a Sierpinski triangle for different resolutions R. R counts the number of
pixels on a square lattice in x and y directions. The Sierpinski triangle was constructed with
an iterated function system with n = 106 iterations. The exact theoretical value for this fractal
is Dexact = l.5849 . ...
fractal dimensions collected in the work [3] are results from the probabilistic
method. As a system of reference we took 52 normal T-Iymphocytes. The mean
value of all individual fractal dimensions for this cell type is D = 1.20 ± 0.05.
The same value within the standard deviation is obtained for integer moments of
q up to q = 5. In [3] there are also some other types of cells with which various
influences like preparation, cell line etc. on the scaling exponents are discussed.
Acknowledgement:
The work has been supported by the Swiss National Science Foundation under the
grant No. 31-25702.88.
References
[I] B. Mandelbrot, The fractal geometry of nature, Freeman, San Francisco,
(1983).
[2] G.A. Losa, G. Baumann, and TF. Nonnenmacher, Path. Res. Pract. 188,680-
686 (1992).
[3] G.A. Losa, G. Baumann, and TF. Nonnenmacher, Acta Ster. l1/Suppl I, 335-
341 (1992).
[4] T.G. Smith Jr., W.B. Marks, G.D. Lange, W.H. Sheriff Jr., and E.A. Neale, 1.
Neuro. Sci. Meth. 26, 75-82 (1988).
[5] TF. Nonnenmacher, G. Baumann, and G. Losa, Trends in Biological Cyber-
netics 1, 65-73 (1990).
Fractal Properties of Pericellular Membrane
from Lymphocytes and Leukemic Cells
Gabriele A. Losa
Laboratorio di Patologia Cellulare
Istituto Cantonale di Patologia, 6600 Locamo
and Faculte des Sciences, Universite de Lausanne
1000 Lausanne, Switzerland
Abstract. Blood mononuclear cells examined on two-dimensional electron microscopy pictures dis-
play plasmalemmal membranes with irregular contour due to the altemance of microprotrusions and
invaginations at the cell surface. Surface irregularity can be quantified in terms of the fractal dimension
D because cell contours manifested the property of statistical self-similarity. A probabilistic method
which defined the scaling behaviour through five scaling exponents or moments of order q was ap-
plied for evaluating the fractal dimension D of the plasma membrane of distinct cell populations.
Resting CD8 suppressor T-lymphocytes showed a monofractal dimension D (the five moments of D
had a similar value) of their plasma membrane higher but than the D of plasma membrane from CD4
helper T-lymphocytes. In contrast, blast cells of acute lymphoblastic leukemia of Band T cell lineage
were characterized by a smaller fractal dimension D. This indicates that the examined leukemic lym-
phoblasts were proliferating cells with surface pattern smoother (i.e. less irregular) than those of mature
immunocompetent lymphocytes. Similar smooth surface patterns showed T-blasts raised by culturing
mononuclear cells with mitogenic phytohaemagglutinin. In contrast, concanavalin A-triggered T-cells
displayed a plasma membrane of more irregular contour and consistently higher fractal dimension,
close to that of control T lymphocytes. Finally, established lymphoblastoid cell lines of B lineage
with progressive stage of differentation showed a corresponding increase of their surface fractional
dimension.
1 Introduction
The immune system is constituted of several populations of circulating and resi-
dent cells, including lymphocytes, granulocytes, mononuclear phagocytes, which
generate in the bone marrow from cellular precursors. Through successive stages
of differentation and selection final cells reach the immunologic competence. The
current working hypothesis postulated that cells derived from precursors may not
proceed through maturation steps but remain in effect frozen in a particular point
of the differentiation process. Those elements that fail to become differentiated
into mature cells due to the neoplastic transformation, can eventually expand into
a dominant population of undifferentiated or partially differentiated cells [with a
phenotype close to the corresponding normal counterparts] and thus give rise to
leukemias of various type. The quantitative evaluation of morphological patterns
of surface and nuclear membranes is of great importance not only to establish
correlations with biochemical functional mechanisms but also in histopathology in
order to identify the types of leukemia and to assess grading systems wich may
help in predict cancer progression, invasiveness, and malignancy [1]. A certain
number of morphometric shape descriptors integrated by a series of mathematical
transformations [2] are used to quantitate structural and morphological features of
Fractal Properties of Pericellular Membrane 191
cells and tissues [3]. The application of conventional morphometric methods to

measure size, shape and area reposes on the assumption that biological structures
approximate geometrical objects and conventional dimensions adequately describe
morphological features and peculiar changes of cellular and nuclear patterns. In
practice, this implies to disregard the real form and shape of cells and cellular
organelles. It however offers an explanation why ambiguous and divergent re-
sults may often occur from morphometric estimates recorded at a single scale [8].
Even computer-assisted quantitative analysis which utilizes the ratio of investi-
gated area to the area of the surrounding rectangular figure recorded at a unique
scale might not be adequate to quantify nuclear irregularity [4]. In fact, cellu-
lar structures are almost irregular and therefore reveal more details about their
morphological configuration by closer inspection at increasing scale of magnifica-
tion. Hence the degree of morphological irregularity and the entity of geometrical
parameters cannot be evaluated at a single scale of length in reason of the prop-
erty of the statistical self-similarity on multiple scales, a characteristic of fractal
objects [5], which connotated also subcellular ultrastructures and organelles of
many physiological competent and neoplastic cells as well [6,7,8]. In the present
study, a probabilistic method was applied [9,10,11,12] to evaluate the degree of
irregularity of cell surface contour in terms of fractal dimension D of several cell
populations, namely immunocompetent human blood lymphocytes, T-Iymphocytes
stimulated by mitogenic lectins in vitro, lymphoid blasts isolated from patients with
acute lymphoblastic leukemia of different phenotype and finally, established Iym-
poblastoid cell lines which constituted a neoplastic system at progressive stages
of differentation along the B lineage.
2 Material and methods

2.1 Immunophenotyping of mononuclear cells
Resting mononuclear cells isolated by gradient centrifugation from blood samples
of normal donors [21], mononuclear cells cultured in vitro in presence of mitogenic
lectins Phytohaemagglutinin or Concanavalin A, established Iymphoblastoid cell
lines and blasts cells isolated from patients with acute lymphocytic leukemia were
identified on the basis of the membrane surface imrnunophenotype.
Cells were stained with fluorescent conjugated monoclonal antibodies directed
against surface membrane antigens CD2, CD3, CD4, CD7, CD8, CD1O, CD19,
CD20, CD33, CD34, and positivity assessed by flow cytometry (FACScan, Becton
Dickinson).
2.2 Lectins stimulation and cell cultures

Aliquots of peripheral blood mononuclear cells (2 x 106 jml) were incubated at
37°C for five days in a humidified 5% C02 atmosphere with phytohaemagglutinin
(PHA) and Concanavalin A (ConA) at the concentration of 10 jLgjml.
192 Losa
Stimulated cells were then collected, washed with 0.05 M Tris-K-Mg buffer
solution (TKM), counted and processed for both conventional and immunoelectron
microscopy.
Cultures of human lymphoblastoid cell lines, REH-6, NaIrn-I, Raji and 6410,
were maintained at 37°C in a humidifed 5% C02 atmosphere in plastic flasks.
The medium consisted of RPMI 1640 supplemented with 2 mM L-Olutamin, 40
mg/l folic acid, 2 gIl sodium bicarbonate, 50 mg/l gentamycin and 10% FCS. Cells
were examined for surface markers phenotype by flow cytometry and processed
for electron microscopy.
2.3 Conventional and immunoelectron microscopy
Cell Samples (1 - 5 x 106 ) were fixed during 5-6 hours at 4° with a 30 gIl of
glutaraldehyde, calcium chloride in O.IM sodium cacodylate containing solution,
pH 7.4. After washing, cell samples were postfixed for 1-2 hours at room temper-
ature with a 2g/100ml of osmium tetroxyde in O.IM sodium cacodylate containing
solution. Dehydratation with ethanol and propylene oxyde preceded embedding in
Epon.
Thin sections having an interference color of gray to silver were cut with
an LKB 2128 microtome, mounted on 200-mesh grids coated with a formwar
film and finally stained with uranyl acetate (5g/l00ml) for 1 hr and successively
with lead citrate for 30 min. Micrographs were recorded on plastic plates with
a Philiphs 400-ET microscope at the magnification of 8000 x and 12500 x in
order to visualize completely a single cell. The cell phenotype was identified in
EM preparations by the peroxidase-antiperoxidase technique after staining cell
sample for 30 min at 4°C, with unconjugated monoclonal antibodies directed to
surface antigens previously mentioned. After washing with phosphate buffered
NaCl solution, cells were fixed for 30 min at 4°C in a 0 .1g/100 ml glutaraldehyde
solution, thereafter incubated for 30 min at room temperature with rabbit anti-
mouse IgO-antibody (diluition 1/100), rinsed and finally reacted for 30 min with a
mouse peroxidase-antiperoxidase complex. The enzymatic activity was activated
by adding diaminobenzidine-H202 as substrate which transformed in a electron
dense complex.
3 Fractal Dimension of the Plasmalemma

The fractal dimension Dq of cell surface perimeter was estimated by applying a
probabilistic method exhaustively described in previous reports [10,11] and in the
present volume [13,14]. The boundary profile of a cell projected into a plane was
picked up by a scanner and the data stored into a memory of a computer in the
form of a two dimensional array. Some twenty EM pictures of each cell popUlation
have been analyzed with a NEC SX-3 of the Swiss Scientific Computing Center
(CSCS), Manno (Ti), using a frame of 2048 x 2048 pixels. Experimental image
analysis is dealing with a geometrical object defined by a set S of points at position
x = (XI, X2 , ... x€) in a f-dimensional Euclidean space. The spatial arrangement
of pixels along the cell border was then used to define a probability P (m, E) which
counts the number of points m within a square of size E centered about an arbitrary
point of the image S. P(m, E) is directly related to other probability measures as
introduced by Mandelbrod [5]. The moment Mq(E) of the probability P(m, E) is
defined by
N
Mq(E) = L mqP(m , E) (1)
m=!
where N counts the number of boxes which at least contain one pixel. These
Mq(E) are related to the fractal dimension Dq by logarithmic derivatives
D = ~ (dIOgMq(E)) (2)
q q d logE
This particular definition of the probabilty P(m, E) has been introduced by [15]
and it is very efficient to implement on a computer [14]. The range of E used in our
study was established in three orders of magnitude, namely from Emin = 0.00311m
to Emax = 311m. If all moments give the same value of Dq , then the fractal set is
considered uniform (monofractal) while for a non-uniform fractal set (multifractal
object) the fractal dimension Dq would take different values. Our findings indicated
that distinct cell populations [20 elements for each population were examined]
had Dq values for the five moments q almost similar within a maximum standard
deviation of SD = 0.06, what excludes the presence of multi-fractal structures.
4 Results
Differentiated resting lymphocytes isolated from peripheral blood of normal donors
share a plasmalemma with a definite degree of irregularity as documented by the
dimension Dq of 1.21 ± 0.04 (table 1). Among T-Iymphocytes distinct subsets
could even be identified by the fractal dimension: namely, CD4 T-helper lym-
phocytes with a Dq = 1.17 ± 0.03 which revealed a less irregular surface bor-
der than CD8 T-suppressor lymphocytes (D q = 1.23 ± 0.03). In another series
of experiments, isolated mononuclear cells were cultured for five days with T-
cell mitogenic lectins, namely PHA and ConA at the unique concentration of
lOfJgjml. The fractal dimension Dq was found reduced in PHA rather than in
ConA transformed T-cells, i.e Dq = 1.11 ± 0.02 versus 1.27 ± 0.02 respectively,
indicating that the former proliferating T-cells lose morphological irregularity of
their plasma membrane (table 1). We have also examined peripheral blood cells
isolated from patients affected by acute lymphoblastic leukemia (ALL) of distinct
immunophenotype, identified by the expression of surface markers specific of B
and T cell lineage. Seven out of nine cases of acute leukemia of B cell lineage,
i.e four cases of common acute lymphoblastic leukemia (c-ALL) (no 1,7,8,9) and
three cases of pre-B-ALL (no 2,4,5), both undifferentiated subtypes of leukemia,
expressed D values (D = 1.12-1.17 and 1.17-1.20 respectively) lower than the
194 Losa
I Cells Ds I ±SD
Monomolecular cells 1.19 1.20 1.21 1.21 1.22 0.04
CD3 T-Iymphocytes 1.20 1.20 1.20 1.20 1.21 0.03
CD4 T-Iymphocytes 1.17 1.16 1.16 1.16 1.17 0.03
CD8 T-Iymphocytes 1.22 1.23 1.23 1.23 1.23 0.03
CD19 B-Iymphocytes 1.19 1.20 1.20 1.19 1.20 0.03
PHA -stimulated
1.11 1.11 1.12 1.12 1.12 0.03
mononuclear cells
ConA-stimulated
1.25 1.26 1.26 1.27 1.27 0.02
mononuclear cells
Table 1 Fractal dimension Dq for five moments of plasmalemma from human lymphocytes.
SD: standard deviation.
fractal dimension value which characterized the plasma membrane of mature B

lymphocytes. (fig. 1, table 1). Two other cases, namely 3 and 6, were character-
ized by a plasmalemmal contour more irregular with a value of Dq = 1.23, i.e.
higher than the Dq of normal B cell counterpart. Both cases were recognized as
acute lymphoblastic leukemias of the early B-immunophenotype (B-ALL), which
corresponds to a more differentiated stage of acute leukemias according to the
hypothetical scheme of B cell differentiation [16]. Accordingly, these cells were
found devoid of CD 10 and slightly positive for CD34 antigen, which are usually
expressed at the surface of undifferentiated lymphoid cells and on leukemic blasts
of c-ALL and pre-B-ALL (figure 1).
As a matter of comparison cells of Hairy cell leukemia (case 10), a chronic
type of leukemia, were characterized by a plasma membrane with numerous
microvilli and protrusions wich consistently yielded an high fractal dimension,
namely Dq = 1.33 ± 0.03.
Our investigation has also included four cases of acute leukemia of the T-cell
lineage (T-ALL).
Blasts with an immature T phenotype (pre-TALL), i.e. expressing CD7 and
CD3 thymic antigens, were found to have a very low fractal dimension of 1.10 ±
0.03, with an order of magnitude of Dq close to that evaluated on proliferating
cells triggered in vitro by mitogenic PHA (table 2).
Two other cases ofT-ALL with cells of thymic medullar origin bearing surface
markers of more mature T-phenotype (CDS) have an higher fractal dimension of
1.20 ± 0.02, whilst the highest Dq value of 1.23 ± 0.03 was recorded on T-ALL
blasts with a phenotype similar to the one characterizing normal thymus cells of
the cortical zone.
100
8. ~~
.4
•
Z
70 1. 04
Z
e-60 10
0
:~ 9~
'Vi
&.
~40
3
20 86
8 .36 10 •
0 n
1.10 1.15 1.20 1.25 1.30
fractal dimension D
Fig. 1 Common Acute Lymphoblastic Leukemia (c-ALL): cases no 1,7,8,9. Surface immunophe-
notype: TdT, HLA-DR, CDlO, CD19, (CD34). Pre-B-Acute Lymphoblastic Leukemia (pre-
B-AlI): cases no 2,4,5. Surface immunophenotype: Tdt, HLA-DR, CDlO, CDl9, (CD34),
Cytopasmic immunoglobulin /1 (CY/1). B-Acute lymphoblastic Leukemia (B-ALL): cases no.
3,6. Surface immunophenotype: HLA-DR, CDI9, CD20, (CDlO), surface membrane im-
munoglobulin (SmIg). Hairy cell leukemia: case no 10. TdT, deoxynucleotidyl terminal trans-
ferase; HLA-DR, human leukocyte antigen class II. CD = Antigen cluster differentation. Cells
positive to fluorescent anti CD lO( 0) and CD34(.) monoclonal antibodies were evaluated by
flow cytometry as indicated in the Material and Methods section.
ICells D5 I±SD I
pre T-ALL (CD7; CD3;TdT-) 1.10 1.11 1.11 1.11 1.11 0.02
T-ALL medullar
1.21 1.20 1.20 1.20 1.20 0.02
(CD2; CD3; CD5; TdT-)
T-ALL medullar (CD3; CD5; TdT+) 1.20 1.20 1.20 1.20 1.20 0.02
T-ALL cortical (CD2; CD5; TdT+) 1.25 1.23 1.24 1.24 1.24 0.03
Table 2 Fractal dimension Dq for five moments of plasmalemma from Acute Lymphoblastic Leuke-
mias (ALL) expressing thymic antigens. SD: standard deviation.
196 Losa
1.28
1.26 - ..
0
1.24-
•
0
[Dq1
1.22 -
1.20 - II
..
0
• •
1.18 -
1.1 6 -
..
0
0 0
1.14
• I I I
REH-,6 NALM-1 RAJI 6-6410
Fig, 2 Fractal dimension Dq for q: 1(0), 2(e), 3 (0), 4 (6) and 5 (\7) of plasmalemma from
Iymphoblastoid cell lines; REH-6 (HLA-DR, CDIO); NALM-I (HLA-DR, CDlO, CYIl); RAJ!
(HLA-DR, CDIO(+ /- ), SmIg) and B-641O (HLA-DR, SmIg, K).
Human lymphoblastoid leukemic cell lines with an immunophenotype corre-

sponding to a progressive stage of differentation of the B lineage were examined
at the end of the growth exponential phase (fig. 2).
The lowest Dq value (1.16) characterized the cell membrane of REH-6 cells,
the less differentiated among the four cell lines examined, while the highest D
(1.25) pertained to the plasma membrane of the most differentiated B-cellline 6410
(fig 3). Nalm-1 and Raji cells could not be distinguished from each other on the
basis of the fractal dimension; interestingly, they presented a similar intermediate
surface immunophenotype.
Indeed, the range of Dq was 1.18-1.20 and 1.18-1.21 for Nalm-1 and Raji
cells respectively. This indicates that the plasma membrane of these cell lines has
a fractional border with a similar degree of irregularity (fig.3) although inferior to
that of the differentiated 6410 B cell line.
5 Discussion
In the present study we could demonstrate that mature blood mononuclear cells
like lymphocytes, immature cells isolated from patients with acute lymphoblastic
leukemia and cultured human lymphoblastoid cells as well, have plasma mem-
brane boundaries of irregular pattern. Cell surface borders manifest the property
Fig. 3 Electron micrograph of plasmalemma portion from Iymphoblastoid cell of a) REH-6; b)

NALM-l; c) RAJ!; and d) B-64IO. Final magnification 18400 x; bar: Icm= 540 nm.
of statistical self-similarity within a definite scaling range and could be portrayed

as fractal objects properly described in terms of fractal rather than topological di-
mensions. According to fractal dimension values of the plasma membrane it tums
out that neoplastic lymphoid cells from patients with acute leukemia of the undif-
ferentiated type (c-ALL, pre-B-ALL and T-ALL) were in general smoother than the
corresponding mature T or B lymphocytes [20]. Similar average fractional dimen-
sions have been reported for human lymphocytes from whole blood [17]. Contour
smoothness characterized also the plasma membrane of mononuclear cells bear-
ing T-antigens which were triggered for proliferation with PHA mitogenic lectin.
In contrast, the plasma membrane of cells proliferating with ConA displayed an
higher degree of irregularity confirmed by the fractal dimension value.
198 Losa
Recently, we could demonstrate that mononuclear cells stimulated with either

PHA or mitogenic anti-CD4 monoclonal antibodies underwent a partial transloca-
tion of signal transducing enzymes (PIP2 phospholipase C) bound to the plasma
membrane toward nuclei, whereas in ConA cultured cells a reversed transloca-
tion toward the plasma membrane occurred [18]. Such a transmembrane process
which enabled the translocation of membrane fragments from the surface into the
cell interior might explain the decreased roughness of cell membrane in PHA
treated cells and provide a functional-structural link with the apparent reduction
of the PLC activity in CD4 helper T-Iymphocytes triggered by PHA. The true bio-
chemical mechanisms inherent the observed changes of irregularity on the plasma
membrane of leukemic cells and of cells proliferating in vitro are however far to
be understood.
One could evoke modifications in the configuration and composition of mem-
brane lipids induced by mitogens and/or neoplastic factors and cytokines. As con-
sequence, the bilayer fluidity could change and create zones within the membrane
with an increased order of lipids but a decreased structural complexity [19].
Alternatively, proliferating cells might undergo an enhanced shedding of
membrane fragments from cell surface as still observed with undifferentiated cells
of common acute lymphoblastic leukemia [21]. All these membrane phenomena
may intervene in modifying surface complexity and structural richness of prolifer-
ating lymphoid cells and leukemic cells. Cancer cells might eventually exploit one
or the other mechanism to escape the defense barrier exerted in vivo by the reticu-
loendothelial system. Does, however, the degree of morphogeometrical irregularity
of cell surface quantitated by the fractal dimension Dq reflect the differentiation
level, in terms of surface immunophenotype, of cells and tissues? On the whole,
undifferentiated cells from acute lymphoblastic leukemia displayed a plasma mem-
brane less irregular in comparison to differentiated T and B resting lymphocytes.
Such a relationship was also noticed in four established lymphoblastoid cell lines
which showed a similar high proliferation rate or S-phase but a specific degree
of differentiation. Consistently, the higher fractal dimension was measured on cell
line with the most differentiated surface immunophenotype.
The fractal dimension Dq constitutes a quantitative descriptor of the real
cell surface morphology and other cellular organelles as well. So far, the level
of cellular differentiation expressed in numerical term might help, together with
other quantitative cellular parameters, in view of a comprehensive description of
the biological profile of neoplastic tissues. The closer the fractal dimension to
the reference Dq of membranes from structural and functional mature cells the
more differentiated appear to be the neoplastic cells. Finally, this study aimed
to underline that by taking into account the property of self-similarity and the
scaling domain beyond which morphological measures remain invariable one could
estimate the true length, size, surface and volume of irregular living structures [22].
Acknowledgments:
The present study has been supported by grant no. 31-25702.88 of the Swiss
National Science Foundation.
References
[1] Baak J.P.A., Chin D., Van Diest PJ., Ortiz R., Matze-Cok p., Bacus S.S. 1991.
Comparative long-term prognostic value of quantitative HER /neu protein
expression, DNA ploidy, and morphometric and clinical features in paraffin
embedded invasive Breast cancer. Lab.lnv. 64,215-223.
[2] Pienta K.J.Coffey D.S. 1991 Correlation of Nuclear morphometry progression
of Breast Cancer. Cancer 68, 2012-2016
[3] Sorensen F.B. 1992. Quantitative analysis of nuclear size for objective malig-
nancy grading: areview with emphasis on new, unbiased stereologic methods.
Lab. Inv. 66, 4-23.
[4] Wydner K.S, Godyn J.J, Lee M.L, Sciorra LJ, 1991. A new approach to the
computer-assisted quantitative analysis of nuclear shape. Modem Pathology
4, 154-160.
[5] Mandelbrot BB. 1982. The fractal geometry of the nature. Freeman, San
Francisco.
[6] Nonnenmacher TF. 1987. A scaling model for dichotomous branching pro-
cesses. BioI Cyber 56, 155-159.
[7] Nonnenmacher TF. 1988. Fractal shapes of cell membranes and pattern for-
mation by dichotomous branching processes. In: Lamprecht I, Zotin AI,eds.
Thermodynamics and Pattern Formation in Biology, Walter de Gruyter, Ber-
lin-New York, 1988,371-394
[8] Paumgartner D., Losa GA., Weibel ER. 1981. Resolution effect on stereolog-
ical estimation of surface and volume and its interpretation in terms of fractal
dimensions. J Micros 121,51-63.
[9] Baumann G., Nonnenmacher,T.F. 1989. Determination of fractal dimensions.
In Losa GA, Merlini D, Moresi R, eds. Gli oggettti frattali in astrofisica,
biologia, fisica e matematica. Cerfim, Locarno, 93-104.
[10] Nonnenmacher T. 1989. Fractal structures in biomedical systems and mor-
phogenetic scaling processes. In: Losa GA, Merlini D., Moresi R., eds. Gli
Oggetti Frattali in Astrofisica, Biologia, Fisica e Matematica, Cerfim, Locano,
64-92.
[11] Nonnenmacher TF., Baumann G., Losa GA. 1990. Self-organization and frac-
tal scaling patterns in biological systems. In Trends in Biological Cybernetics,
Menon J.,ed., Publications Manager, Research Trends, Council of Scientific
Research Integration, Trivandrum India, 1, 65-73.
200 Losa
[12] Nonnenmacher TF,Baumann G., Barth A, Losa G.A 1993 A quantitative

structural analysis of self-similar cell profile. Proceedings of the Royal Society
(London) Series B (Biological Sciences) (in press).
[13] Nonnenmacher TF. 1993. Spatial and temporal fractal patterns in cell and
molecular biology. In Fractals in Biology and Medicine. Eds. Nonnenmacher
TF., Losa GA, Weibel ER., Birkhauser Verlag, Basel
[14] Baumann G, Barth A , Nonnenmacher T.F. 1993. Measuring Fractal dimen-
sions of cell contours: practical approaches and their limitations. 1993 Fractals
in Biology and Medicine. Eds. Nonnenmacher TF., Losa G.A, Weibel ER.,
Birkhauser Verlag Basel
[15] Voss RF. 1985. Random Fractals: Characterization and Measurements. In:
Pynn R., Skjeltrop A, eds. Scaling Phenomena in Disordered Systems, Plen-
um Press, New York, 1-11.
[16] Van Dongen J.J.M., Adriaansen H.J., Hooijkaas H. 1988. Immunophenotyping
of leukemias and non-Hodgkin Lymphomas. Neth.J.Med. 33,298-314
[17] Keough K.M.W., Hyam P., Pink D.A, Quinn B. 1991. Cell Surfaces and
fractal dimensions. J.Microscopy 163,95-160.
[18] Graber R., Leoni L., Carrel S., Losa G.A 1993 Lectins and anti-monoclonal
antibodies induced changes of second messengers generating enzymes in hu-
man peripheral blood mononuclear cells. Cellular Molecular Biology. 39,
45-54.
[19] Graber R., Losa G.A. 1993 Subcellular localization of inositide enzymes in
established T-cell lines and activated lymphocytes. Analytical Cell. Pathology,
5, 1-16.
[20] Losa G.A, Baumann G.,Nonnenmacher TF. 1992. Fractal dimension of peri-
cellular membrane in human lymphocytes and lymphoblastic leukemic cells.
Path.Res.Practice 188, 680-686.
[21] Losa G.A, Heumann D, Carrel S, Von Fliedner V, Mach JP.1986. Characteri-
zation of membrane vesicles circulating in the serum of patients with common
acute lymphoblastic leukemia. Lab. Investigation 55, 573-579.
[22] Weibel ER. 1991. Fractal geometry: a design principle for living organisms.
Am J Physiol, 261, L361-L369.
Cellular Sociology: Parametrization of Spatial
Relationships Based on Voronoi Diagram
and Ulam Trees
Raphael Marcelpoil, Franck Davoine and
Michel Robert-Nicoud
Equipe de Reconnaissance des Formes et Microscopie Quantitative
Universite Joseph Fourier
CERMO, BP 53X. 38041
Grenoble Cedex, France
1 Introduction
At all levels (molecules, cells, organisms, ... ), biological systems are made of
structural and functional units. Those units are interacting with each other and,
tend to use space in an optimal way with respect to their specific function and
environmental constraints. Hence, cells can be defined as being the smallest struc-
tural and functional units capable of auto-reproduction. The rapid technological
advances during the 20th century have made possible to describe the inner ar-
chitecture of cells at different levels of organization. Our approach is an attempt
to study cellular populations at the «sociological» level, i.e. at the level of their
spatial organization and interrelationship in a given tissue. The topographies of
cellular population have to be considered linked to morphogenesis, structural sta-
bilities and functional state of a given tissue. This approach makes use of the
relations that links form to disorder and is based on space partition constructed
from a set of points defining the position of cells. This spatial partition consisting
of a set of individual forms (Voronoi paving) associated to its dual (Delaunay
triangulation) permits the calculation of Ulam trees which are descriptors of the
local surrounding of a given cell and the search for characteristic features of the
normal and pathological state.
The invariance by dilation, exact or statistic, is a characteristic of all fractals
assumed figures. For an object statistically self-similar like an aggregate, we can
demonstrate this invariance by tracing a circle of radius R around a point X of
the aggregate. We note that the mass of the object M(R) within the circle is
proportional to RD(X), D(X) being the local scale exponent at the point X. In the
case of the Ulam tree (characteristic of the local topography around a given cell),
the figure growths with the iteration number 1. We observe that the number of
segments (branches and leaves of the tree) is proportional to 1D, where D is the
fractal dimension (D :::; 2). The Ulam tree can thus be used to give a consistent
description of the area covered by the figure.
202 Marcelpoil et al.
• • •
• •
V(i)
'01
/. .
• .p
•
•
•
•
•
Fig. 1 Bidimensional distribution of points and the Voronoi polygon V(i) associated to Pi .
This method requires three main steps in its development, (i) a step of con-
struction of the Voronoi diagram and Ulam trees which associates a polygonal form
and a Ulam tree to each point of the population, (ii) a step of elimination of points
whose associated polygonal form or Ulam tree has been altered due to border
effects, and (iii) a final step of parametrisation and quantitation of topographical
informations.
2 Construction of the Voronoi diagram

The Voronoi diagram is the space partition containing the most information (Tou-
ssaint, 1980). Let us briefly recall the basic definitions of this space partition. Let
5, be a set of N points in the plane, i.e. the nucleus barycenters of cells. For each
point pi in S what is the locus of points (x, y) in the plane that are closer to pi
than to any other point of 5?
The solution to the above problem is to partition the plane into regions (each
region being the locus of points (x, y) closer to a point of 5 than to any other
point of 5). Given two points Pi and Pj, the locus of points closer to pi than to
pj is the half-plane containing Pi that is defined by the perpendicular bisector of
PiP j. Let us denote this half-plane by H (Pi, Pj ). The locus of points closer to Pi
than to any other point, which we denote by V(i) (Fig. I), is the intersection of
N - 1 half-planes, and is a convex polygonal region having no more than N - 1
sides, that is,
(1)
Cellular Sociology 203
Where n denotes the intersection and V(i) is the Voronoi polygon associated with
Pi (Preparata & Shamos, 1985). The construction of the Voronoi partition associ-
ated to S, denoted Vor(S), follows an incremental method by local modification
of the diagram after each insertion of a point of S. (Bowyer, 1981).
3 Construction of the Ulam trees

After the study of the tissular structure at the level of its global topography, we
now present a method for the analysis of the local environment of each cell. This
method is based on the analogy between a cell in a tissue, and a tree in a forest.
Ulam trees are objects growing in time and space under given induction rules
(Ulam, 1962). One of the aims of Ulam's work is to throw light on the question
of how much «information» is necessary to describe the seemingly enormously
elaborate structures of living objects.
Let us briefly recall the basic definitions used to construct those trees.
The basic idea is a fixed division of the plane (or space) into regular elemen-
tary figures. For example, the plane may be divided into squares or into equilateral
triangles, or into regular hexagons, the space into cubes. An initial configuration
will be a finite number of elements of such a subdivision and the induction rule
will define successive accretions to the starting configuration.
Our purpose is to quantify the environment of every cell. The initial config-
uration we use is the Voronoi graph of the population, and a particular point of
the population on which the tree is to be constructed.
A large variety of Ulam trees can be generated depending on the choice of
the induction rule used to grow the tree. We chose an Ulam rule which respects
the following conditions: (i) the rule does not favor any direction in the graph, (ii)
the tree covers as much space as possible to increase its statistical significance,
(iii) a strong physical significance can be given to the resulting process.
Given a number of Voronoi regions in the nth generation, the regions of the
(n + l)th generation will be all those which are adjacent to the existing ones but
with the following proviso: the regions which are adjacent to more than one region
of the preceding generations will not be taken into account. (Fig. 2)
Remark: The neighborhood notion is the Delaunay neighborhood notion (dual
to the Voronoi diagram). This Ulam rule may be compared to a wave which
originates on one point in the graph, propagates from node to node, and dies
everywhere the wave intersects itself. The propagation rate is one graph distance
per unit time.
4 Elimination of marginal points

Due to the properties of the Voronoi partition, some regions of the paving should
be modified if other points outside the analysis window were acquired. The points
Fig. 2 Part of a Voronoi diagram and the Ulam tree (bold lines) of depth 3 associated to the gray
polygon. A true like tree can be generated from the Ulam tree. A given population of cells
thus can be transformed to a forest.
of 5 which belong to such regions are considered to be marginal. The subset of

all the marginal points is called the Marginal Subset (M5). It can be demonstrated
that these points are associated to regions having one or more summits which
are outside or closer to the analysis window than to the considered point. All the
marginal points of S are not taken into account in the subsequent calculations.
For example, let us consider a bidimensional distribution of points and the asso-
ciated Voronoi diagram. Border-effect-free polygons (in white) are conserved for
further calculations and marginal polygons (in gray) are eliminated according to
the previous rule (Fig. 3).
Selection of the subset of points which are mathematically significant for the
Ulam tree construction :
This step consists in the selection of points for which the associated Ulam
trees are strictly included in the set of points which are not marginal. This step
is performed with the help of mathematical morphology. Let us consider an Ulam
tree depth of seven (i.e., a length of seven in terms of graph, separates the furthest
leaf of the tree from its root). To be mathematically significant, this tree must not
intersect the subset of marginal points (M5). If a marginal point is included in a
tree, this tree can be biased due to border effects of points that would be outside
the analysis window. Therefore, a morphological dilatation of seven (equal to the
Ulam tree depth) of the M5 subset is performed before any Ulam tree is calculated.
The points of 5 which are eliminated during the dilatation of M5 make up the
Intermediary Subset (IS). This dilatation ensures mathematical significance of all
the trees associated to the final subset of points (F 5P), and the following relation
can be derived:
F5P = 5 - (M5 +15) (2)
Fig. 3 Bidimensional distribution of points and the associated Voronoi diagram. Polygons in white
are conserved for further calculations and marginal polygons, in dark gray, are eliminated
according to the marginal rule and the analysis window AW.
5 Topographical parametrisation and quantitation

5.1 Quantition of the global topography
This attempt to quantitate cellular topography is based on the strong relationship
which obviously links form to disorder. Therefore, we defined parameters that
are descriptors of both form and disorder. These parameters are calculated on the
polygonal form which has been associated to each cell of the population during
the construction of the Voronoi partition.
Determination of the average type of the spatial occupation: for a convex set,
X, where A(X) is the area and L(X) is the perimeter, it is possible to demonstrate
the following isoperimetric inequality:
L(X)2 - 47rA(X) ~ 0 (3)

Since Voronoi polygons are convex, the average type of S population spatial
occupation is well characterized by the average roundness factor (RFav)
N
RF = ~ '"' 47rA(Xi) (0 < RFav ::; 1) (4)
av N L L(K)2
i=l I
Example: the roundness factor (RF=47rA(X)jL(X)2) of a n sided Reuleaux poly-

gon (Regular polygon with all angles and sides equals) is RF (n) =7r j (ntan( 7r j n)).
Thus, this formula makes it possible to predict the RF value of simulated polygons.
The RF of a circle is 1, the RF of a line is O.
The intrinsic disorder of the popUlation has been expressed as two prime
factors, the disorder concerning the area heterogeneity of the population and the
disorder of the geometrical properties.
Determination of area heterogeneity and geometrical homogeneity of the spa-
tial occupation inside S can be quantified by the following two parameters, area
disorder AD, and roundness factor homogeneity RF H
AD=I_(I+~)-l (5)
Aav
where OA denotes the area standard deviation, O"RF the roundness factor standard
deviation, Aav and RFav the mean area and the mean roundness factor. Using
two types of invariant, geometrical one and area one, it can be demonstrated that
RFav, AD and RFH are uncorrelated. Thus RFav, RFH and AD define a three
axis graph which represents topographical informations. We tested this model with
different theoretical populations.
5.2 Quantitation of the Ulam trees

The first step in the quantitation of the cell neighborhood from the Ulam tree is to
transform the tree into a mathematical object. It is possible to transpose the tree
into a matrix M . This matrix is composed of two orthogonal properties which are
respectively: the integration and the topological properties of the tree.
The integration property is the number of points added to the tree at a given
expansion level. The topological property of the tree is the number of simple,
double, triple, etc., junctions in the tree.
Thus the M (i, j) cell of the matrix is set to the number of junctions of jth
order at the ith level in the tree.
For example, in a popUlation ordered at the nodes of a triangular mesh, the
root of any tree is the level 0 of the tree and is a 6 order junction node. Thus the
M(O, 6) cell of the matrix is set to 1. The first level of the tree is made of 6 nodes
which are all 2 order junctions. Thus the M(l, 2) cell of the matrix is set to 6.
The tree depth is 7 and the maximal junction order is 6. Thus this tree matrix is
a (7+1) by 6 matrix.
.L.-_ _ _ AD
Fig. 4 Each prostatic tissue section can be located in the topographical space on a particular way
which shows the topographic evolution of the prostatic tissue from the normal homeostasy
(N) to the cancer (C) or to the hyperplasia (H). RFav = average Roundness Factor. RF H =
Roudness factor homogeneity and AD = area disorder.
From each Ulam tree, we determine 2 parameters which are characteristic of

the cell environment.
The first factor, DORT is expressed as the integrated absolute difference be-
tween the matrix of an ordered reference tree constructed on a perfectly triangular
ordered popUlation, and the matrix of the given tree. The second one is expressed
as the branch length heterogeneity, ELH of the tree.
6 Results
6.1 Global topography quantitation
The topography of different prostatic tissue sections of various pathological grades
..... " .
' ..
.'.:.,;
Fig. 5 Prostatic carcinoma on the left side and hyperplasia on the right side. The form of the Ulam
trees grown on the cancer and hyperplasic parts of the tissue are very different and thus can
help diagnosis.
have been mesured from their Voronoi diagram and located in the topographical
space as shown in figure 4.
6.2 Local topography quantitation

Using the Ulam trees, it is possible to caracterize the environment of each cell in
a prostatic tissue for example (Fig. 5).
7 Conclusion
This approach should make it possible to describe and quantify abnormalities of
the cellular topography characteristic of pathological states, and to objectively
grade tumours by measuring the amount of perturbations compared with normal
population, and in a more general way, to study cellular interactions and sociology.
Finally, this model appears to be useful for analyzing the effects of the cell
surrounding on a given cellular function and vice-versa.
References
[1] Bowyer A: Computing Dirichlet tesselation. Comput J, 24: 162-166, 1981.
[2] Preparata FP, Shamos MI: Computational geometry. Springer-Verlag, 1985.
[3] Toussaint GT: Pattern recognition and geometrical complexity. In: Proceed-
ings 5th International Conference on Pattern Recognition, IEEE Catalog NO.
80CHI499-3, Miami Beach 1980, pp. 1324-1347.
[4] Ulam S: Patterns of growth of figures: mathematical aspects. Proceedings
of symposia in applied mathematics, XIV: Mathematical problems in the
biological sciences, American Mathematical Society. 64-75 , 1962.
A Fractal Analysis of
Morphological Differentiation of
Spinal Cord Neurons in Cell Culture
Tom G. Smith, Jr.
Laboratory of Neurophysiology, NINDS and
Laboratory of Developmental Neurobiology, NICHD, NIH
Bethesda, MD., U.S.A.
Abstract. Cell cultures of murine spinal cord neurons, selected at discrete intervals after plating over a
5-7 day period, were fixed, stained and classified on the basis of the number of large, primary dendrites
emanating from the cell body (2; 3/4; 5 or greater) and dendritic branching patterns (5A; 5B). Images
of individual cells were captured from a light microscope via a video camera to an image processor.
Gray scale images were converted to binary ones with the Marr-Hildreth convolution algorithm. The
fractal dimension (D) of individual images was determined by Flook's dilation logic method. The two
main features that contribute to D are the profuseness of branching and the ruggedness of the border.
Plots of mean D vs. time for all of the four cell types (2; 3/4; 5A; 5B) could be well fitted to a model
of the form: D(t) = A + B exp(l/t), where A and B are constants, t is time (hours) in culture and t
is a time-constant (hours). Each cell type had a distinct final, plateau D (ascending rank order: 2 =
1.28; 3/4 = 1.32; 5B = 1.37; 5A = 1.41) and a different time course as measured by t (ascending rank
order: 5B = 12.6; 5A = 14.5; 3/4 = 16.4; 2 = 20.4). The time-courses of the changes in D for all
four groups were significantly different (p < 0.05) from one another (ANOVA). We conclude that the
fractal dimension is a useful, quantitative and unbiased measure of the complexity of neuronal borders
and branching patterns and that its time course of development can be described by a simple equation
with a characteristic time constant and final plateau value. We suggest that D is also a useful measure
of morphological cellular differentiation.
1 Introduction
As the Proceedings of First International Symposium on Fractals in Biology and
Medicine [14] adequately and convincingly demonstrate, Mandelbrot's fractal di-
mension of fractal geometry [8] is a useful measure or statistic of images and other
types of data. The definition and application of fractal concepts were demonstrated
abundantly in the various papers to be found in these proceedings. In this paper
we show how the fractal dimension can be employed as a useful, quantitative and
unbiased measure of the complexity of neuronal borders and branching patterns
and that its time course of development can be described by a simple equation
with a characteristic time constant and final plateau value. We also suggest that it
is a useful measure of morphological cellular differentiation.
2 Methods
The raw materials for our research are grey scale images of individual neurons,
mainly grown in dissociated cell culture. Figure 1 is a photograph of a mouse
A Fractal Analysis of Morphological Differentiation 211
Fig. 1 A spinal cord neuron in dissociated cell culture for four days and stained immunohistochem-
ically using Fragment C of tetanus toxin and a monoclonal antibody against Fragment C.
Neuronal surface membranes are stained intensely with this procedure. The stained neuron is
a pyramidal Type 3/4 neuron. Bar = 100 jlm. Reprinted with permission from Neale, Bowers,
and Smith, 1993.
spinal cord neuron stained after 4 days in culture. One particularly important tech-
nical point is that it is necessary to have an original image with as much contrast
as possible between the object of interest (cell) and the background. In addition,
since the initial goal is to obtain the border of the object, no intracellular structures
should be stained or visible, elsewise they produce unwanted borders. Finally, the
analog TV signal used to produce the digitized image should match the dynamic
range of the AID converter of the image processor's frame grabber (e.g., 8 or
16 bits). These objectives are achieved by proper histological preparation (high
contrast) and/or adequate illumination. In addition, an analog video amplifier, with
gain (contrast) and DC (background) controls placed between the video camera
and the frame-grabber can help considerably in obtaining the desired high-contrast
digitized image [21]. To measure the fractal dimension (D), we need a binary im-
age. To that end we employ the Marr-Hildreth algorithm, which is a convolution
operation employing a large (7 x 7 to 13 x 13 pixels), two-dimensional, circularly
symmetrical Laplacian of a Gaussian kernel (<<Mexican Hat» filter) [21]. The prin-
cipal merit of this technique is that it almost always produces unbroken borders
which are ideal for filling to obtain a cell silhouette. Most cells illustrated are
silhouettes since they are easier to visualize than borders, particularly where den-
dritic branches are numerous and crossing. But the measurements for the fractal
dimension are made on borders-only images. The main drawback of the method
212 Smith, Jr. and Neale
is that, with large kernels, there is some blurring of detail; however, this is not se-
rious if, as here, the images and their measures are used for comparative purposes
only. In our hands, the size of the kernel needed is inversely related to the degree
of contrast in the original image.
3 Measuring the Fractal Dimension

We have employed three different methods of measuring the fractal dimension of
the borders of individual cells [22]. Two of these, the yardstick and box-counting
methods, have been discussed fully by the VIm group in these proceedings [2].
The third method, called the dilation algorithm, was developed by Flook [4]. It is
a convolution operation, where the kernels are single-valued (Boolean 1) circular
discs of differing diameters. These kernels are applied, one at a time, to a border
image and each resultant area is divided by the relevant disc diameter to give an
equivalent perimeter. Then, the log of that result is plotted against the log of the
diameter of the disc. A straight line, with slope = -5, is plotted and D calculated
from: D = 1 - 5. For images lying in a plane, 1 < D < 2 [8].
All of these three measurements are related to lengths, and the filtering action
of the larger measuring elements remove the higher spatial frequencies and produce
smaller equivalent perimeter lengths. We calibrated our D-measuring techniques
against borders of known fractal dimension (e.g., Koch islands, snowflakes, etc.)
and found that our measurements consistently underestimated the true value of
the known fractal dimension by about 5%. This may be due to the fact that
our digitized images (512 x 512 pixels) do not contain the degree of resolution
inherent in a true fractal. That was of little consequence to us, since we used the
numbers for comparative purposes only.
As has been discussed elsewhere [2,22], the yardstick and box-counting meth-
ods can have technical problems in securing unambiguous data. For example, with
short yardsticks and images with relatively smooth borders (straight lines), the log-
log curves tend to deviate from a straight line and approach a Euclidian slope of
zero. On the other hand, with the box-counting method and large boxes, the box
count of a poorly centered image is underestimated and the log-log slope steepens.
Both these cases make for difficulty in drawing an objective, unambiguous straight
line. The dilation method, being a convolution operation rather than a counting
one, is not as sensitive to these problems over the range of discs of 2-128 pixels in
diameter - we consistently get unambiguous log-log straight lines over the entire
range (6 octaves or 1.26 decades, the range required to obtain useful results). Thus,
all of the data illustrated were obtained by the dilation method.
4 Measuring Neuronal D's

We obtained our original biological measurements from spinal cord neurons which
had been grown in cell culture for more than 6 weeks [22]. Some selected results
Fig. 2 A set of neuronal images ordered by increasing fractal dimension. The value of D for each
image is indicated by its position on the D axis. Visually perceived morphologic complexity
correlates well with estimated fractal dimension. Reprinted with permission from Smith et
al..1989.
are shown in Fig. 2 as a one-dimensional plot in D, where a line is drawn from each
cell to its corresponding measured D. This illustrates that the fractal dimension of a
cell's border is a measure of the complexity of that border. As cellular morphology
becomes more complex, D increases. We find that two characteristics are important
in determining D: 1) the ruggedness of the border and 2) the profuseness or degree
of dendritic branching. In addition, a particular D does not uniquely define a border.
This is illustrated in Fig. 3, where two cells have approximately the same D but
very different morphological characteristics. Cell A has few branches but a very
rugged border, while B has smooth borders but lots of branches. Having developed
a «tool» for quantifying the complexity of an image's border, we next attempted
to apply this tool to an experimental situation to determine its utility.
Fig. 3 Two neuronal images of differing morphology but with similar high values for D. The high
D for cell A is related to its spiny surface, whereas the extent of branching accounts for the
high D of cell B. Reprinted with permission from Smith et aI., 1989.
TYPE 2 TYPE 3/1 TYPE 5A TYPE 5B

z I.S
o
Q
B en.
Z
1.4 or 0 - - - - 0- - - - - - - ~
/ _--------- Il -------------
t.Ll 6,'- - --- -0 FD (2)
-o
0
~ 1.3 I - - . ~- FD (3/4)
~I• .,? - 0 - e- FD (SA)
~
1.2 (4
~ o -- . -- FD (56)
E-
U 1.1
~ D(t) = A+ 8exp{ tIt}
M
u.. 1~~~~~_r~~~~~~~,
o 40 60 120 160 200
C TIME (HR)
9 HR
--r
21 HR ~8HR ~~:~
Fig_ 4 Change in fractal dimension and morphology of spinal cord neurons with time in culture.
Spinal cord neurons were grouped (see text) into four types, all illustrated in (A). Fractal
dimensions were analyzed at various times during the !irst week in culture and plotted for
each morphologic type in (B). The development of morphology over this same interval is
shown for Type 3/4 neurons in (C). Reprinted with permission from Neale, Bowers, and
Smith, 1993.
5 Results and Discussion

To provide a frame of reference, we grouped 258 spinal cord neurons on the basis
of the number of primary dendrites growing directly from the cell body (Fig. 4A).
The neurons can be recognized in this way even when the cells are quite immature_
We grouped our population as Type 2 (two primary dendrites), Type 3/4 (three or
four primary dendrites) and Type 5 (five or more primary dendrites). This latter
group breaks down to Type 5A or 5B, depending upon the pattern of dendritic
branching [12,13] _
We start a batch of sister cultures at time zero and halt growth of individual
cultures at given times by fixing and staining the cells. These cultures are then
examined under the microscope and discrete cells of each cell type are chosen
for analysis. After digitization, the fractal dimension of each individual neuronal
image is measured by the dilation method [4,22].
100
'/
~ /
;:J
?
.
~
<
;;;.-
80 /
~
? /
;:J
....~
60 0 ?
/
• /?
~
<
~ 40
~
,% TYPE 5B
/
Z
~
U 20
0-
if
---e--- Fractal Dimension
-+- Total Dendritic Length
~
~ -0- Area
~ 0
0
0 40 80 120 160 200
TIME IN CULTURE (HR)
Fig. 5 Change in neuronal growth compared with change in D. Measures of total dendrite length
and area covered by the dendrite arbor (territory) are shown along with changes in D for Type
5B neurons. Whereas fractal dimension increases rapidly during the first 2-3 days in culture
and then plateaus, other measures of dendrite growth continue to show linear increases for
the first week after plating. Reprinted with permission from Neale, Bowers, and Smith, 1993.
What is clear, as shown in Fig. 4C, for Type 3/4 cells, is that the cell borders
become progressively more complex, over time, during growth and differentiation
in culture. Figure 4B shows the time courses for all 4 cell types, starting with a
D of 1 for spherical cells plated at time zero. As dendrites emerge and branch,
D increases progressively and plateaus at a maximum value for each cell type.
The curves that best fit the time course of the mean value for each of the cell
types is the model shown in Fig. 4B, center. The correlation coefficients for the
fit of all the curves shown in Fig. 4B to the data were greater than .95. This
equation-model, which describes how D varies with time, t, has two constants, A
and B, and is characterized by a time constant, t. We should emphasize that we
find it quite remarkable that so complex a process, with its many-fold changes in
complexity, can be described so well by such a simple expression. It suggests to
us that complexity is an important aspect of cellular morphological growth and
differentiation. Table I shows the actual values for the maximum D and the t's
for the 4 cell types. For our population of cells, the maximum D rises from a low
value of 1.28 for Type 2 cells to a high value of 1.41 for Type SA, while the time
constants run from a low value of 12.6 hours for Type 5B to a maximum of 20.4
hours for Type 2. All these curves are significantly different from one another
(p < 0.05; ANOVA).
We would note in passing that similar developmental data have been obtained
from glial cells grown in culture, and the same D(t) model provides the best fit
for those data also [20].
Figure 5 illustrates how the change in D with time (t), for Type 5B cells,
relates to other more conventional or Euclidian forms of cellular measurements
(growth area, total dendritic length), all normalized to the maximum value for
each variable. It is immediately apparent that, while the Euclidian measures all
increase nearly linearly, D changes non-linearly with time. We would note that if
one takes the conventional definition of «fractal growth» as growth with constant
D, then the change in area, etc. from time zero to about 80 hours is not fractal
growth, but that the period after about 80 hours, when D has reached its stable
value, does represent fractal growth. We would note the changes in D with time
are clearly not fractal growth.
Some authors have proposed that neuronal growth and differentiation repre-
sent a special case of diffusion limited aggregation-like growth (DLA) [3]. The
evidence cited for such a conclusion is that some neurons have a fractal dimen-
sion of l.70, the hallmark of DLA. The results from our group [13,20,21] and
from others [7,9,10,11] suggest that such an assignment of DLA properties may
be premature since most neurons, either in vivo or in vitro, do not in fact have a
D = 1.70.
We have some observations that we would like to mention that may relate
to the meaning of «fractal growth». In order to produce images of known fractal
dimension to calibrate our methods (like the Koch islands), we used commercial
software that employs the so-called L-systems method, named after Landenmayer,
a Dutch botanist, who devised it [15]. It employs a system much like that discussed
in Mandelbrot's book, «The Fractal Geometry of Nature» [8], except Mandelbrot's
«initiator» and «generator» are called «axiom» and «production», respectively,
in L-systems. We undertook an exercise [8; p.5l], beginning with a Euclidian
square (initiator, axiom) and applied the generator or production through successive
interations to generate progressively more complex images. Then, we took the
result of each iteration of a number of deterministic fractals and calculated the
fractal dimension for that image. The results are shown in Fig. 6 for six different
deterministic fractals, listed in the figure, where the fractal dimension is plotted
as a function of iteration. Then we fitted these points with the same equation we
had used for neurons demonstrated in this paper and, as mentioned, for glia.
It is clear that this fractal generation begins with what are Euclidian objects
(initiator, axiom) like lines, squares or triangles and have a D of one. As they
progress, they become more complex and D increases to reach some plateau value
like our neurons and glia. But is this fractal growth? And, how do these lattermost
results relate to the similar changes found with cultured neurons and glia?
The notion that biological tree-like, fractal structures represent an optimal
design for a particular function is generally accepted. For example, such ideas
have been proposed for the lung for the flow of air [1] and the vascular bed for
---KOCH 1.5
-0- KOCH 1.26.
- 1- . KOCH ISLAND1
2 - - KOCH ISLAND2 o
.0'
• ·0· • PEANO CURVE
o.
-1.- SIERPENSKI GASKET
1.8 o
--- _e----- ...

... --.-- --.---.
:.. ~ _ - -A
----- ....
1 234 5 6
L-SYSTEM INCREMENT
Fig. 6 A plot of the fractal dimension as a function of iteration number for a number of known fractal
objects listed in the upper, left part of the figure. The curves of the points fit an expression
of the form: D(i) = A + Bexp (iln ) (cf. equation given in Fig. 4B).
the flow of blood [23]. It may well be that the fractal, dendritic trees of neurons
are also optimally designed, but in this case for the flow of their most important
commodity: information.
Indeed, there is a considerable corpus of theory and data which indicates
that some neurons are optimally structured for the integration of their graded
input signals (synaptic potentials) which results in a pulsatile output signal (action
potential) [5,6,16,17,18,19]. In cases where the organization is known, excitatory
synaptic inputs are located mainly on distal dendritic branches whereas inhibitory
synaptic inputs are located on the cell body and proximal dendrites. Since the
action potential generating membrane is located at the axon hillock or the axon
itself, the excitatory synaptic potentials (EPSPs) must depolarize that membrane to
a threshold value in order to generate an action potential. If however, prevention
of such action potentials is desired, then an increase in inhibitory conductances
can shunt the excitatory currents flowing from the distal dendrites toward the axon
hillock and block the generation of an action potential.
6 Conclusion
Finally, we conclude that the fractal dimension is a useful, quantitative and unbi-
ased measure of the complexity of neuronal borders and branching patterns and
that it is often more useful and insightful than conventional Euclidian measures.
Moreover, its time course of development can be described by a simple equa-
tion with a characteristic time constant and final plateau value. We suggest that
o is also a useful measure of morphological cellular differentiation, in the sense
that Webster 's Collegiate American Dictionary defines differentiation as «a change
from the simple to the complex» [24].
References
[1] Bassingthwaighte JB : News Physiol. Sci. 3, 5-10 (1988).
[2] Bauman G, Barth A and Nonnenmacher TF: In Proceedings of the First Inter-
national Symposium on Fractals in Biology and Medicine, (ed. Nonnenmacher
TF, Losa GA and Weibel ER), Birkhauser-Verlag, Basel, pp. (1993).
[3] Caserta F, Stanley HE, Eldred WD, Daccord G, Hausman RE and Nittmann J:
Phys. Rev. Lett. 64, 95-98 (1990).
[4] Flook AG: Powder Technol. 21, 295-298 (1978).
[5] Jack HB, Nobel 0 and Tsein RW: Electric Current Flow in Excitable Cells,
Clarendon, Oxford, U.K., (1975).
[6] Jack HB and Redman SJ: J. Physiol. (Lond.) 215, 321-352 (1971).
[7] Kniffki KD, Pawlak M and Vahla-Hinz C: In Proceedings of the First Interna-
tional Symposium on Fractals in Biology and Medicine, (ed. Nonnenmacher
TF), Birkhauser-Verlag, Basel, pp. (1993).
[8] Mandelbrot BB: The Fractal Geometry of Nature, W.H. Freeman, New York,
(1982).
[9] Montague PR and Friedlander MJ: Proc. Natl. Acad. Sci. (U.S.A.) 86, 7223-
7227 (1989).
[10] Montague PR and Friedlander MJ: 1. Neurosci. II, 1440-1457 (1991).
[11] Morigiwa K, Tauchi M and Fukuda Y: Neurosci. Res. Suppl. 10, S131-S140
(1989).
[12] Neale EA, Bowers LM and Smith TG, Jr.: Soc. Neurosci. Abst. 17,36 (1991).
[13] Neale EA, Bowers LM and Smith TG, Jr.: J. Neurosci. Res. 34,54--66 (1993).
[14] Nonnenmacher TF, ed.: Proceedings of the First International Symposium on
Fractals in Biology and Medicine, (Basel: Birkhauser-Verlag) (1993).
[15] Prusinkiewicz P and Landenmayer A: The Algorithmic Beauty of Plants,
Springer Verlag, New York, (1990).
[16] Rall W: In Handbook of Physiology, Vol. 1, Pt. 1, The Nervous System,

Cellular Biology of Neurons, (ed. Brookhart JM, Mountcastle VB, Kandel
ER and Geiger SR), American Physiological Society, Bethesda, Maryland,
pp. 39-97, (1977).
[17] Rall W: Biophys. J. 9, 1483-1508 (1985).
[18] Rall W, Burke RE, Smith TG, Nelson PG and Frank K: J. Neurophysiol. 30,
1169-1190 (1967).
[19] Redman S and Walmsley B: J. Physiol. (Lond) 343, 117-133 (1983).
[20] Smith TG, Jr., Behar TN, Lange GD, Marks WB and Sheriff WH, Jr.: Neu-
roscience 41, 159-169 (1991).
[21] Smith TG, Jr., Marks WB, Lange GD, Sheriff WH, Jr. and Neale EA: J. Neu-
rosci. Methods 26, 75-82 (1988).
[22] Smith TG , Jr., Marks WB, Lange GD, Sheriff WH, Jr. and Neale EA: J. Neu-
rosci. Methods 27, 173-180 (1989).
[23] West BJ and Goldberger AL: American Scientist 75, 354-365 (1987).
[24] Woolf HB, ed.: Webster's New Collegiate Dictionary, (Springfield, Mas-
sachusetts, U. S. A.: G & C Merriam Company) (1976).
Fractal Dimensions and Dendritic Branching
of Neurons in the Somatosensory Thalamus
Klaus-D. Kniffki, Matthias Pawlak and
Christiane Vahle-Hinz 1)
Physiologisches Institut
Universitat Wiirzburg
D-97070 Wiirzburg, Germany
1) present address: Physio\ogisches Institut. Universitats-Krankenhaus Eppendorf, Universitat Ham-

burg, D-20246 Hamburg
1 Introduction
The investigation and modelling of irreversible growth phenomena has become a
topic of considerable interest in the last decade, stimulated by the introduction of
the concept of fractality by B.B. Mandelbrot. This concept provides a quantitative
framework to study in particular biological growth phenomena of complex shapes,
such as the branching structures of trees, of bronchial trees and of blood vessels.
Recently, it was shown that the shapes of 2-dimensional retinal neurons are fractal
objects, and hence may be characterized by their fractal dimension D = 1.68±0.IS
(Caserta et aI., 1990). The authors proposed an explanation of certain stages of
neuronal development by a diffusion-limited particle-cluster aggregation (DLA)
model, which predicts in 2-dimensional space D = 1.70 ± 0.1.
2 Fractal dimension of thalamic neurons

The fractal dimension of neurons stained with a Golgi impregnation method located
in the eat's thalamic ventrobasal complex (VB) and its ventral periphery (VBvp)
was examined using 2-dimensional projections of the fractal structures embedded
in Euclidean space E3 (Kniffki et aI. , 1991; Kniffki et aI., 1993).
Fractal structures embedded in Ed usually do not have a finite d-dimensional
volume and a D-dimensional Hausdorff measure has to be used in order to assign
well-defined contents to such objects. D E [0, d] is the Hausdorff dimension of a
fractal point set X C Ed. A very good estimate can be achieved by computing the
so-called box counting dimension DB (Mandelbrot, 1983). It is defined as
. InN(E)
DB = £11m - ( I/ )' (1 )
->0 n 1 E
222 Kniffki et al.
7 .0
60
5.0
r-.....
r-.....
W 4 .0
'--'"
Z 3.0
'--'"
~ DB = 1.61
......-1
2 .0 r = O.997
1.0
0.0
1.0 2.0 3.0 4. 0 5.0 6.0
In (E)
Fig. 1 Fractal analysis (box-counting method) of the dendritic branching of a neuron (inset) located
in the cat's ventrobasal complex. A box covering the dendritic tree is successively dissected
into smaller and smaller boxes, where ~ is the edge length of the boxes. N is the number of
boxes of a particular edge length, which are necessary to cover the whole dendritic field: N
vs ~ in In-In plot. The logarithmic straight line implies the indicated fractal dimension DB
(Eq. I).
where E is the lattice constant of d-cubic covers of X and N (E) is the number of
cubes contained in the minimal cover. Numerically, DB is calculated by counting
the number N(E) of boxes of edge length E, which are necessary to cover all
points of the data set and calculating the slope of In N (E) versus In Efor decreasing
values of E. Two methods of analysis were employed in the present study. Firstly,
the drawings of the thalamic cells were placed under grids of different sizes E.
These grids of different scales were constructed by successively dissecting the
largest grid which covered the whole cell. A typical example is shown in Fig. 1.
As indicated by the straight line (least square fit), this particular cell exhibited
DB = 1.61 with a correlation coefficient r = 0.997. Secondly, the images of
the thalamic neurons were digitized by a scanner with a resolution of 200 dpi
in 2 grayscales and the obtained data were automatically processed using a box
counting algorithm implemented on an AT personal computer (Block et aI., 1990).
A main result of new study (Kniffki et aI., 1993) is the fact that dendritic
branching patterns of thalamic cells have fractal structures. The mean fractal di-
mension «(DB) of neurons within VB in general is larger than that of neurons
within VB vp ' In particular, neuron of the ventral posteromedial nucleus (VPM)
of VB have (DB) = 1.51 ± 0.05, those in the ventral posterolateral nucleus
Fractal Dimensions and Dendritic Branching of Neurons 223
(VPL) of VB have (DB) = 1.44 ± 0.05. The corresponding values of neurons

located within the ventral peripheries of these nuclei are: ventral layer of VP M vp ,
(DB) = 1.39 ± 0.04; VPL vp , (DB) = 1.39 ± 0.08.
It is important to emphasize that we have determined the fractal dimension
of the camera lucida drawings of thalamic cells (see inset of Fig. 1) which are
projections of the original 3-dimensional objects. Nevertheless, this fact does not
affect the estimation of the fractal dimension, if the fractal dimension of the 3-
dimensional object is smaller than 2. In general, for the dimension of a fractal
structure $ of dimension D projected upon an Euclidean subspace of dimension
d the projection $* satisfies:
dimension $* = min [d,D] (2)
(Mandelbrot, 1983; Takayasu, 1990). In our case this means D* = D, i.e. the
fractal dimensions which we have estimated from the 2-dimensional camera lucida
drawings are the true fractal dimensions of the neurons embedded in E 3. As a
consequence, simple DLA models are not suitable for describing the fractal pattern
formation of dendritic outgrowth of thalamic neurons (Kniffki et al., 1993). Since
the estimated fractal dimensions are significantly smaller than predicted for neurons
described by a DLA model in 3-dimensional space, which would give D = 2.5,
it is suggested that the neuron's morphology is not a result of purely stochastic
processes but of strong interactions between the cell and its environment. This is
reminiscent of, e.g., aggregation with interaction, which leads to growth structures
that are similar to DLA products at first glance, yet possess fractal dimensions
significantly smaller than 2.5 (Block et al., 1991).
3 Thalamic neuronal branching

When a biological tree structure is a fractal object and exhibits self-similarity over
a certain range of length scales, in addition to its fractal dimension D another
parameter is involved, termed the diameter exponent n. This exponent n is part of
the scaling law for a bifurcation:
(3)
where do is the diameter of the parent branch and d 1 and d 2 are the diameters of the
two daughter branches, the thicker and the thinner, respectively. For the bronchial
tree, e.g., Mandelbrot (1983) assumed a simple self-similar growth process, which
does not require the branching ratio do/di(i = 1, 2) to be encoded genetically.
The branching pattern of the bronchial tree, which at first sight appears to be
very complex, is fully determined by two parameters: the width/length ratio of the
branches and the diameter exponent n = 3.
Assuming a similar simple rule for the growth of neurons, it was the aim
to quantify the dendritic branches of neurons being located within VB and VB vp
224 Kniffki et al.
A B
~f~-:
SO.,m
50J,lm
.>" ~l
J 1 'r
Fig. 2 Dendritic morphology of a thalamic neuron located in the ventrobasal complex (A). and of a
neuron of its ventral periphery (B), as well as their schematic reconstructions. The diagrams
represent the topological branching pattern. The lengths of the branches have been corrected
for the travel through the planes of section. The diameters of the individual dendritic branches
were measured directly from the histological sections and were not drawn to scale in this
figure.
(Kniffki et aI., 1993). The morphology of neurons in these two thalamic regions
has been described previously (Vahle-Hinz et aI., 1993). Here we wanted to test
in particular, whether the dendritic arborizations of the neurons in both regions of
the lateral thalamus may be describable by the scaling law [Eq. (3)] with a single
value of the diameter exponent n.
The quantitative analysis was carried out on representative neurons from VB
and VB vp (Kniffki et aI., 1993). The diameters and the lengths of the primary
dendrites and the dendritic branches of the neurons were measured directly from
the histological sections using a 1000 x magnification and an eyepiece graticule.
Two-dimensional schematic drawings were made for each neuron representing
the topological pattern of dendritic branching and the lengths of the individual
branches. In Fig. 2 the dendritic morphology of a thalamic neuron located within
VB (A) and of a VB vp (B) neuron as well as their schematic reconstructions are
shown.
1.0
'-. .--:',-'-'-'-
'-'-'- '-.
'"
".
d 1 / do -'-'-'-
0.8
0
'ij 0.6
"-...
'ij
0.4 -- n=3/2
--- n=2.0
........ n=5/2
0.2 .-.- n=3 .0
o.n-jL--~-.---~----r----.--~---'---~---'
0.0 0.2 0.4 0.6 0.8 1.0
Fig.3 Representation of Eqs. (4) and (5) with the indicated diameter exponents n.
To test the scaling law [Eq. (3)], some transformations were carried out.
After division of Eq. (3) by do and with the introduction of the asymmetry ratio
a = d2 /d] ::; 1, Eq. (3) is equivalent to:
(4)
and
(5)
When both interbranching segments have the same diameter, i.e. a = dd d 1 = 1,
the di / do results in
(6)
For n -+ 00 this leads to an increasing value of 2- l / n up to the value of 1.
This method of representing the scaling law [Eq. (3)] has the advantage that
dIldo for n > 1 shows only a weak dependency as a function of a = d2/dl, while
ddd o shows a stronger relationship. The functional behaviour of dildo for four
different values of n is shown in Fig. 3. For the experimental data, a nonlinear
least square fit routine was used for Eqs. (4) and (5) to determine the exponent n
for dIldo and d 2/d o separately.
The diameters of the dendritic segments of neurons within VB and VB vp
were pooled according to the order of branching with respect to the soma and the
mean dildo were plotted against the asymmetry ratio d2/dl after the ddd l values
226 Kniffki et al.
1.0
d1/d o
0.8
0
"d 0.6
~
"d 0.4 Branch 1

d z / do n=1.76
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0
dz / d1
Fig. 4 Scaling behaviour for thalamic dendritic bifurcations for the I st order of branching. The
filled circles show the averaged measured values, whereas the continuous lines represent Eqs.
(4) and (5), respectively, with the indicated values of the diameter exponent resulting from a
nonlinear least square fit procedure. The error bars larger than the size of the symbols indicate
the standard error of the mean.
had been grouped into classes with a width of 0.05. No statistically significant
difference was found for the diameter exponent n for d 1/do and d 2 / do of VB
and VB vp neurons (Kniffki, et aI., 1993a). Therefore, the data of both groups of
neurons were pooled and in Figs. 4 and 5 the filled circles represent the pooled
data for branch order 1 and 3. For branch order 1, fl = 1.76 and for branch order
3 the corresponding value fl = 2.93 was found. Near the soma, i.e. for branch
order 1 the lowest value of fl was obtained. For increasing order of branching the
diameter exponent increased up to its highest value of fl = 3.92 for the 7th (and
last) order of branching. The corresponding values of fl are plotted against the
order of branching in Fig. 6.
In the thalamic neurons studied, the scaling relation [Eq. (3)] was fulfilled for
the ramifications of the dendritic trees; the diameter exponent n, however, was not
constant throughout the dendritic trees, but varied systematically with the order of
branching.
4 Conclusion
During their outgrowth, neuronal dendrites exhibit a morphology typical for the
kind of neuron within a particular nucleus in the central nervous system. To char-
1.0
d 1 / do
0.8 ~
0
'"d 0.6 J
'"d 0.4 1 Branch 3

n=2.93
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0
d 2 / d1
Fig. 5 Scaling behaviour of thalamic dendritic bifurcations for the 3rd order of branching (filled
circles) with the indicated diameter exponent. The continuous lines represent Eqs. (4) and
(5); the error bars larger than the size of the symbols indicate the standard error of the mean.
acterize the space-filling properties of the neurons' dendrites, the fractal dimension
D might be a useful parameter (Smith et aI., 1989). In addition, the fractal di-
mension might be a useful parameter to describe differences in the dendritic trees
of neurons in different nuclei. Differences between VB and VB vp neurons with
respect to their fractal dimension were shown, D of VB vp neurons being smaller
or equal to D of VB neurons (Kniffki et aI., 1993). Since D of both types of
neurons is significantly smaller than predicted for DLA model neurons, which
would give D = 2.5, the morphology of thalamic neurons is not a result of purely
stochastic processes but of strong interactions between the outgrowing cell and
its local environment, e.g., by specific chemical gradient fields (Hamilton, 1993).
Models of aggregation with interaction, which leads to growth structures similar
to DLA structures, exhibit fractal dimensions significantly smaller than 2.5 (Block
et aI., 1991). However, the functional significance of D for neurons is currently
unknown. It is speculated to describe some aspects of the neuronal integrative
properties, including aspects of the synaptic connectivity.
In order to study the dendritic integrative properties of synaptic signal con-
duction, i.e. the passive spread of depolarization towards the axon hillock, RaIl
(1959) introduced an ideal neuron with an equivalent-cylinder dendrite. At each
dendritic bifurcation this ideal equivalent-cylinder dendrite requires a diameter ex-
ponent n = 3/2 in the scaling law [Eq. (3)]. Agreement with and deviations from
228 Kniffki et al.
• •
4
,.......,
• •
Id
~
3
•
.....
• •
d
Q) 2
d
0
Po
><
rz:I 1
0
1 2 3 4 5 6 7
Order of branch
Fig. 6 Values of the diameter exponent as a function of the order of dendritic branching of thalamic
neurons. The small dots represent n = 3/2, the value for the ideal equivalent-cylinder model
dendrite.
the ideal equivalent-cylinder model have been noted in neurons of several species
(for reviews, see Schierwagen and Grantyn, 1986; Kernell and Zwaagstra, 1989)
resulting in values for n ranging from 1.0 to 3.0. The present results also show
clear deviations from n = 3/2. In dendritic trees of thalamic neurons, up to the
7th order of branching n increased systematically from 1.76 for the 1st order of
branching near the soma to 3.92 for the 7th order of branching (Kniffki et aI.,
1993a). That means, that the description of the detailed integrative properties of
these neurons requires a segmental non-uniform equivalent cable model instead of
RaIl's equivalent-cylinder approximation (Schierwagen, 1989).
The outgrowth of the branching dendrites should be studied by modeling
the dynamic behaviour of the growth cones, which are the active structures of
the outgrowing dendritic tips. The basic actions of these growth cones comprise
elongation, branching and guidance as a result of the cell's local environment (van
Veen and van Pelt, 1992). The dependence of the diameter exponent n on the order
of the dendritic branching is judged as an indication for the involvement of more
than one intrinsic rule in the neuronal growth process, assuming that the dendritic
branching ratio for bifurcations is not required to be encoded genetically within
the DNA-sequences.
References
[1] Block, A, von Bloh, W., ScheIlnhuber, H.J. Efficient box-counting determi-
nation of generalized fractal dimensions. Phys. Rev. A. 42: 1869-1874 (1990).
[2] Block, A., von Bloh, W., ScheIlnhuber, H.J. Aggregation by attractive particle-
cluster interaction. 1. Phys. A. Math . Gen. 24: Ll037-Ll044 (1991).
[3] Caserta, F., Stanley, H.E., Eldred, W.D. , Daccord, G., Hausman, R.E., Nitt-
mann, J. Physical mechanisms underlying neurite outgrowth: A quantitative
analysis of neuronal shape. Phys. Rev. Lett. 64: 95-98 (1990).
[4] Hamilton, P. A language to describe the growth of neurites. Bioi. Cybern. 68:
559-565 (1993).
[5] KerneIl, D., Zwaagstra, B. Dendrites of cat's spinal motoneurons: Relation-
ship between stem diameter and predicted input conductance. J.Physiol. 413:
255-269 (1989).
[6] Kniffki, K.-D., Chialvo, D., Vahle-Hinz, c., Apkarian, AV. Fractal dimen-
sions of neurons located in the cat's thalamic ventrobasal complex (VB) and
its ventral periphery (VBvp). Soc. Neurosci. Abstr. 16: 622 (1991).
[7] Kniffki, K.-D., Pawlak, M., Vahle-Hinz, C. Scaling behavior of the dendritic
branches of thalamic neurons. Fractals 1/2: 171-178 (1993a).
[8] Kniffki, K.-D., Vahle-Hinz, c., Block, A, von Bloh, w., ScheIlnhuber, H.J.
Fractal scaling characteristics of 3-dimensional thalamic neurons. In prepa-
ration (1993).
[9] Mandelbrot, B.B. The fractal geometry of nature. Freeman, New York (1983).
[10] RaIl, W. Branching dendritic trees and motoneuron membrane resistivity. Exp.
Neurol. 1: 491-527 (1959).
[11] Schierwagen, A, Grantyn, R. Quantitative morphological analysis of deep su-
perior colliculus neurons stained intracellularly with HRP in the cat. J. Hirn-
forsch. 27: 611-623 (1986).
[12] Schierwagen, A A non-uniform cable model of membrane voltage changes
in a passive dendritic tree. J . theor. BioI. 141: 159-179 (1989).
[13] Smith, T.G., Marks, W.B., Lange, G.D., Sheriff Jr., W.H. , Neale, E.A.
A fractal analysis of cell images. J. Neurosci. Methods 7: 173-180 (1989).
[14] Takayasu, H. Fractals in the physical sciences. Manchester University Press,
Manchester and New York (1990).
[15] Vahle-Hinz, c., Pawlak, M., Kniffki, K.-D. Morphology of neurons in the
ventral periphery of the thalamic ventrobasal complex (VBvp) of the cat.
1. Compo Neurol. (in press, 1993).
[16] van Veen, M. van Pelt, J. A model of outgrowth of branching neurites. 1. theor.
Bioi. 159: 1-23 (1992).
Fractal Structure
and Metabolic Functions
Organisms as Open Systems
Manfred Semetz
Institut fUr Biochemie und Endokrinologie
Justus-Liebig-Universitat Giessen
D-35392 Giessen, Gennany
1 Introduction
Organisms are open, energy-dissipative driven systems which build up and sustain
their complex structure and functional organization in a multiplicity of steady states
far from thennodynarnic equilibrium by continuous exchange of mass and energy.
Both from theoretical reasoning and from experimental evidence they are as living
systems comparable with continuous bioreactors as technological equivalents (Fig.
1). This applies to scale dependence of reaction kinetics and of structure and
can be demonstrated on a broad variety of common features with far reaching
correspondences (Table I). Application of the concepts of heterogeneous catalysis
and of fractal structure helps to understand scaling phenomena and the organization
of metabolism on the level of single cells as well as of metazoa.
2 The Reduction Law of Metabolism

It is a well known phenomenon in biology that absolute metabolic rates S of
similar organisms of different size, say mammals, are not directly proportional
to their volume V (or body mass, at constant average density ~ 1), but scale
according to an empirical power law
S = a· Vb (1)
with an non-integer exponent b ~ 0.75 . This relation is experimentally well doc-

umented over a range of 1 : 106 , that is from about 3 ml to 3 m 3 (mouse to
elephant). Thus the specific rates SIV (in the conventional definition) decrease
with body volume
SIV = a· Vb- 1 ~ a. V- O.25 (2)
leading to the allometric reduction law of metabolism. This means that small
animals metabolize at appreciably higher volume-specific rates than bigger ones.
We can achieve, however, constant specific rates a, if we relate the abso-
lute metabolic rates S to the intuitively strange, but proper fractal entity Vb of
organisms, namely
$ IV b = a = $ IL 3b ~ SIL 2.25 = constant (3)

Organisms as Open Systems 233
s
substrates products
V volume, 'i/ flow rate
5 p
'Hi.OOOc;o
5
o 0
O® 0
V, 'i/
Fig. 1 Correspondence between organisms and continuous bioreactors as open systems. S substrates,
P products, E constraint enzymes, C concentrations, V volume, V dilution rate.
This indicates that the body is not adequately represented by its mere mass or
its unstructured volume V = L3 , but behaves as a whole or with its rate limiting
parts rather like a 2.25·dimensional object (Semetz et al. 1985, 1986, 1989a). It is
noteworthy that isolated cells under identical experimental conditions exhibit equal
turnover rates irrespective of the species of origin. Thus allometry of metabolism
is not due to genetic differences, as previously suspected, but is a «technical»
consequence of transport limitation of turnover rates in scaling volumes.
Allometric scaling can be observed for clearances CI (e.g. of kidneys) and
many other first order elimination processes with rate constants 1 k or relaxation
times T =1 k- I .
Cl Vv .1 k
Vb - V . Vb-I = a (4)
Since the relative volume of distribution Vv/V of a compound in the body can be
regarded as a constant for similar organisms of different size, rearrangement gives
1k = a'. V b- 1
T = l/a' . V I - b (5)
and shows that these cases reduce to the scaling merely of rate constants 1 k,
frequencies or time constants T in bodies of different size. Only by reference to
the proper fractal entity V l-b the so defined specific time constants Tspec become
independent of size
Tspec = T /V I - b = constant (6)
234 Sernetz
Organisms Bioreactors
open, energy-dissipative, driven, continuous, catalytic and structured systems
living animal, experimental analysis of technical device, analysis of reaction
physiological functions kinetics and energy balance
catalytic, compartimented, multiphase systems (solid, liquid, gaseous). Station-
ary phase as the carrier of catalytic activity (enzymes)
Cells as basic, catalytic units. Tissues, Suspensions of porous enzyme carrier
parenchyma with intracellular con- particles or enzyme membranes with
straint enzymes immobilized enzymes
Liquid phase, driven for transport of reactants
Blood, intercellular liquid driven by Bulk solution of substrates and prod-
the heart ucts driven by a stirrer
Kinetics: heterogenous catalysis, turnover determined by enzymatic reaction
and convective and conductive transport, transport limited turnover
Metabolic rates Turnover rates in steady state
Size dependent volume specific rates and frequencies
Allometry of metabolism Scaling of turnover
Kinetics of mixing between stationary and liquid phase
«Almost turbulent» mixing via a frac- Turbulent mixing by stirring Dilution
tal vessel system. Clearances and rate rates, exponential residence time dis-
constants of first order processes. AI- tributions. Scaling of characteristic
lometry of characteristic times and times
frequencies
Optimization of efficiencies E by reduction of transport resistances
Fractal structure
Fractal branching of transport ves- Turbulence as the dynamic equivalent
sels, fractal folding of tissues, fractal of fractal structure of stirred bioreac-
borders between stationary and liquid tors. Ergodic trajectories of carrier gel
phases Ergodic trajectories of blood particles. Percolation of stationary and
cells in the vascular system. Perco- liquid phase in the open porosity of
lation of stationary and liquid phase polymer enzyme carrier gels
from macroscopic to molecular, intra-
cellular scale
Table 1 Correspondence between organisms and continuous bioreactors (CSTR)

3 Scaling of Turnover in Continuous Bioreactors

We can apply the same arguments to describe the turnover ofbioreactors, especially
the type of continuous stirred tank reactors (CSTR) (Fig. Ic). Under turbulent
stirring its characteristics are again the dilution rates D
(7)
and clearances Cl
Cl =1 k· V (8)
Here too the clearances or relative turnover rates follow scaling power functions
with non-integer exponents of the reactor volume v,..
Its proportional part is the catalytically active stationary phase, the gel volume
Vg with immobilized enzymes
. b
Cl = ShedS = k . Yg = f· "vb (9)
Turnover in bioreactors is governed by the laws of heterogeneous catalysis, where

the catalytic activity is constrained to the Volume Vg of the stationary phase. The
efficiency E is defined as the ratio of the actual turnover rate of the substrate in the
heterogeneous system Shet to the maximum possible rate Shorn in the homogeneous
case of solute enzymes.
(10)
The efficiency depends on the influence of two modules of rate limitation, namely
the module I.L of external transport resistance between stationary and liquid phase,
and the module a of internal diffusion of the substrates within the porous stationary
phase. The module I.L describes the influence of a quasi un stirred Nernst-Iayer 8
at the surface of the stationary phase. Its influence can be reduced by rigorous,
turbulent stirring. The module a represents the influence of diffusional transport
limitation within the stationary phase. It can be minimized by reducing the radius
R of the carrier gel particles or the thickness of catalytic membranes.
Thus the efficiency E of a bioreactor can be optimized by reducing these
two transport resistances, that is of the distance 8 between liquid and stationary
phase by stirring and of the distances R within the stationary phase by using small
particles or thin layers.
By transferring this terminology from bioreactor kinetics, we can define or-
ganisms of increasing size as bioreactors with constant resistance of internal dif-
fusion, since the size R and the density of cells remain constant, independent of
the size of an animal, but with increasing external transport resistance. The re-
sulting decrease of efficiency with size must be counterbalanced by optimizing
the distribution characteristics of a growing vessel system and by maximizing
the exchange surfaces between liquid phase and the tissue or parenchyma as the
stationary catalytic phase.
236 Sernetz
The scaling of turnover rates has been demonstrated with CST-bioreactors on

the basis of suspensions of mammalian cells and bacteria in continuous fermenta-
tions (Sernetz et al. 1990). The turnover rates of bioreactors as purely technical,
analytical systems scale with reactor volume and with the particle density as ex-
perimental parameter according to power functions with non-integer exponents in
complete analogy to the allometric reduction law of the metabolism of organisms.
The value of the exponent however depends very much on the cell or particle
number densities in the reactor. At low cell densities the total turnover of the
reactor is merely the sum of the individual turnover of independent cells of the
suspension. In this case the specific turnover S/V is independent of size and the
system behaves isometrically. At high cell densities however, the Nernst-layers of
cells overlap, the total turnover now is the sum of concurrent, interacting cells and
the specific turnover decreases allometric ally with the size of the system (Sernetz
et al. 1990).
A comparison between the value of the scaling exponents of bioreactors and
the allometric exponent of organisms leads to an important qualitative difference.
The specific turnover rates of bioreactors reach scaling exponents in the range of
b - 1 = -0.3 at cell densities of about 109 /i. In contrast the allometric exponent
of specific metabolic rates of organisms is only b - 1 ~ -0.25, yet for average
cell densities of about 10 12 /i. Therefore we postulate that our body exhibits ex-
cellent mixing characteristics in spite of its extremely high packing density of the
stationary phase by virtue of the peculiar properties of its fractal vessel systems.
4 Fractal Structure
Over a broad range of scales organisms are characterized by fractal structure
of tissues and vessels or transport systems. It is the intricate interlacing of a
stationary, catalytic phase with a liquid, driven phase, which minimizes distances
and resistances of transport down to subcellular resolution and enables optimum
turnover through all scales and sizes of organisms. In fact we can define the
structural organization of the body as that of an effective percolation of two phases,
solid and liquid, with their common fractal border.
The comparison with a percolation can already be applied to the gel-state of
polymers in the subcellular and molecular range. The fractal properties have been
established experimentally e.g. for the gel matrix of hyaline cartilage and for the
intercellular permeabilities of corneal tissue by gelfiltration (Sernetz et al. 1989b,
Sernetz 1992). Biological gels and tissues are characterized by the open porosity
of the interlacing phases. Open porosity means that both in the solid phase and
the liquid phase a wanderer can reach any point of the entire space.
The broad pore size distributions of gels can be determined experimentally
by size exclusion or permeation equilibrium techniques. Classical size exclusion
chromatography or gelfiltration, however, is a typical fractal analytical approach,
since it measures the scaling of the gel pore volume accessible to testmolecules
of different size as yardsticks, thus corresponding to the mass-radius- or box-

counting-method in image and structure analysis (Wlczek et al. 1992, Sernetz et
al. 1992). The fractal dimension of the border between both phases of the gel
can be determined from the scaling of the complementary, namely non-accessible
volume of the polymer network, which includes the fractal border (Sernetz et al.
1989b).
Within the body the cells are only constant, autarc units of almost common
size, above which the fractal architecture of tissues and transport systems continues
in an even more conspicuous manner up to macroscopic scales. Thus e.g. for the
lung the scaling of bronchial branching (Kitaoka, this volume) and that of the
alveolar surface as a phase border (Weibel, this volume) has been measured by
serial sectioning and image analysis.
In my group we devote special attention to the fractal properties of the kidney
arterial tree, which we derived from arterial cast preparations according to the
corrosion cast technique and by 3-dimensional, scale-dependent image analysis of
3D-data sets. These data sets were both gained from NMR-computertomography
and from fine serial sectioning of the entire organ with an isotropic resolution of
0.1 mm (Sernetz et al. 1992). The mass-radius-analysis yielded a scaling of the
mass with about D = 2.2 to 2.5, which corresponds well with the overall scaling
exponent of organisms for metabolic rates as a function of body size (see above).
The scaling of the surface of the kidney arterial tree varies from about Ds :::::;
2.5 at coarse resolution (> 3 mm) to about Ds :::::; 2.0 at fine resolution (0.3
mm). Due to natural limits exponents in the body represent scale dependent, local
dimensions and an average fractal behaviour in a certain range and change towards
approaching these limits. It is possible that at the scale of diffusing molecules
the exchange surfaces (say in alveoli or capillaries) may become two-dimensional
again. Thus non-limiting structures may therefore exhibit isometric, i.e. non-scaling
behaviour in a proper range of resolution.
With the maximum resolution achieved experimentally in the kidney arterial
tree (0.1 mm) we are still within the transport and thus turnover limiting part of
arteriolae and far from the capillary bed. A technical consequence of mass-radius
analysis of arterial surfaces is, that at high resolution it is not necessarily the fine
structure of the smallest vessels, but rather the «flat» surface of the bigger ones
that is tested (Semetz et al. 1992). Non-integer exponents can be interpreted as
direct evidence of the fractal organization and in comparison with the allometric
exponents of specific turnover rates as a measure of the effective transport-limiting
border in a catalytic system (Sernetz et al. 1985).
Now what does fractal organization mean and what are the fractal borders
in the body? It can easily be demonstrated with any two arterial and venous
casts of the vessel tree of an organ (e.g. of kidneys, Sernetz et al. 1985) or even
of the entire body that each of them has a definite volume (blood) and that in
its finest branchings the vessels reach almost any point and depict the form of
the entire organ or body. Moreover, the two systems of defined volume fit exactly
238 Sernetz
between each other, yet are still merely negatives of the tissue. Only the remaining
inter«space» as a highly folded fractal «area» and border between the two volumina
of the vessels (liquid phase) actually represents the tissue, literally the parenchyma
or stationary phase of the organ or body.
Thus from both our present knowledge on the fractal exponents of vessels and
from the allometric reduction law of metabolism, organisms should be represented
in fractal terms as 2.25-dimensional rather than voluminous objects.
5 Turbulence in the Body

The determination of the fractal dimension of tissues or vessels by means of 3D-
image analysis techniques and by 3D-computer simulation (Bittner and Sernetz
1991, Bittner 1991) serves not only for the morphological description as a fractal
structure, but also yields the key to the dynamics of turnover and turbulent mixing
by means of the vessel system. In bioreactors high efficiency E (eq. 9) is achieved
by turbulent mixing, which keeps Nernst-Iayers {j small and resistance of external
diffusion p, low. Turbulence itself is a dynamic self-similar phenomenon, namely a
recursive regress of eddies in eddies through a wide range of scales. Turbulent sys-
tems continuously dissipate energy and must be driven (periodically) to maintain
their structure. In turbulent flow the paths of two initially neighboured particles
or volume elements diverge exponentially and unpredictably into entirely different
regions, after sufficiently long, periodical or stroboscopic observation their paths
however fill the entire region and depict its internal structure.
There is ample kinetic evidence, although it is apparently paradoxical, of
turbulent distribution and transport in the organism as the basis of the high ef-
ficiencies of metabolic reactions. Although all the flow in the blood vessels is
locally and everywhere laminar, as known from the low Reynolds-numbers, the
mixing characteristics of the organism are practically turbulent, as expressed by
the exponential residence time distributions which are the basis of the definition of
clearances in physiology as first order distribution processes. The fractal organiza-
tion of tissues and transport systems is the structural equivalent and morphological
support of turbulence.
Nevertheless the turnover remains to some extent transport limited as can be
derived from actual deviations of ideal clearance or pure first order distribution.
Traditionally this is taken into account e.g. in pharmacokinetics by defining so-
called superficial and deep compartments of tissues. In our context by way of
fractal folding and branching of vessels all the cells within the body share an
average reactive surface of dimension 2.25.
Although blood flow is locally laminar within the entire blood vessel system,
due to its fractal structure it imposes turbulent mixing characteristics. The path of
single cells, say erythrocytes, within the vessels is unpredictable and ergodic. Af-
ter sufficiently long, stroboscopic observation the pathway of a single cell would
however fill the entire volume of the vessels and with the resolution of its diameter
depict its structure and the anatomy of the entire body as its complement. The ves-
sel system behaves as a «strange attractor», and the tissue as a «strange repellor»
of a cells trajectory. Any real anatomical cross-section x, y of the body is also a
two-dimensional x, y-section of the complete Poincare-map of the ergodic pathway
of an erythrocyte. The coordinates of its repetitive transitions through this plane of
section after sufficiently long iterations (or simultaneously for all erythrocytes the
instantaneous «bleeding» of the section) fill and depict with increasing resolution
the anatomical hierarchy, the resolution dependent interlaced architecture both of
the liquid (vessels) and stationary (tissue) phase of the body and also the areas of
nested common residence time distributions (Sernetz et al. 1989a).
6 Limited selfsimilarity
up to now allometric relations in biology have usually been formulated as power
law equations without setting limits of validity. Instead, as in any natural fractal
system, it is however realistic to assume limited ranges of allometric or fractal
behaviour. This can be achieved by two different approaches: One can either define
a fractal or selfsimilar region by truncation of the power law within upper and
lower bounds (cut off) Emax and Emin of the range of resolution Eo
As opposed, selfsimilarity can also be limited by upper and lower limits Ymax
and Ymin of the measured quantity y, with asymptotically fading fractal properties
of the measured system towards approaching these limits as a function of the
resolution E. By introducing such limits, the selfsimilar power function, which is
a straight line in a graph with logarithmic coordinates, modifies into a log-logistic
distribution function with sigmoidal course in logarithmic coordinates (Semetz et
al. 1985, Sernetz and Bittner 1991) and with the fractal dimension related to the
parameter of dispersion of the distribution function.
Viewing the body as a fractally structured system allows a new, functional
interpretation of the allometric power law of metabolism. The relation of turnover
rates to the proper fractal entity, namely the 2.25 dimensionality instead of a
volume, yields equal efficiency for small and large animals and thus explains the
broad size range of realization of similar metazoa.
Acknowledgement:
This work was supported by grant of Deutsche Forschungsgemeinschaft (Se 315/
13-1 to 4).
240 Sernetz
References
[1] M. Sernetz, B. Gelleri and J. Hofmann: The organism as bioreactor. Interpre-
tation of the reduction law of metabolism in terms of heterogeneous catalysis
and fractal structure. 1. Theor. BioI. 117 (1985) 209-239.
[2] M. Sernetz, H.R. Bittner and H. Willems: Organismen als Bioreaktoren: Frak-
tale Struktur und heterogene Katalyse, Umschau 86 (1986) 582-587.
[3] M. Sernetz, H Willems and H.R. Bittner: Fractal organization of metabolism,
in: Energy Transformations in Cells and Organisms, W. Wieser, E. Gnaiger
(eds.), Thieme, Stuttgart 1989a, pp. 82-90.
[4] M. Sernetz, H. Willems and K. Keiner: Dispersive analysis of turnover rates of
a CST-reactor by flow-through micro fluorometry under conditions of growth.
Ann. N.Y. Acad. Sci. 613 (1990) 333-337.
[5] M. Sernetz, H.R. Bittner, H. Willems and C. Baumhoer: Chromatography,
in: The Fractal Approach to Heterogeneous Chemistry, D. Avnir (ed.), Wiley,
Chichester 1989b, pp. 361-379.
[6] M. Sernetz: Fractal structure of gels determined by size exclusion chromatog-
raphy or equilibrium techniques, in: Proc. IUPAC-Symp. Macromolecules
1992,1. Kahovec (ed.), VSP, Zeist 1993 pp. 423-429.
[7] P. W1czek, A. Odgaard and M. Sernetz: Fractal 3D analysis of blood vessels
and bones, in: Fractal Geometry and Computer Graphics, 1.L. Encama~ao et
aI. (eds.), Springer, Berlin 1992, pp. 240.-248.
[8] M. Sernetz, 1. Wtibbeke and P. W1czek: Three-dimensional image analysis
and fractal characterization of kidney arterial vessels. Physica A 191 (1992)
13-16.
[9] H.R. Bittner and M. Semetz: Selfsimilaritiy within limits: Description with the
log-logistic function, in: Fractals in the Fundamental and Applied Sciences,
H.-O. Peitgen et aI. (eds.), Elsevier, Amsterdam 1991, pp 47-58.
[10] H.R. Bittner: Modelling of fractal vessel systems, in: Fractals in the Funda-
mental and Applied Sciences, H.-O. Peitgen et aI. (eds.), Elsevier, Amsterdam
1991, pp. 59-71.
Transfer to and across Irregular Membranes
Modelled by Fractal Geometry
Bernard Sapoval
Laboratoire de Physique de la Matiere Condensee
C.N.R.S. Ecole Polytechnique
91128 Palaiseau, France
Abstract. We apply results from the study of fractal electrodes to diffusion across an irregular passive
membrane with finite permeability. Diffusion efficiency is more specially discussed and seems to
explain semi-quantitatively the final step of respiratory physiology i.e. the diffusion and capture of
oxygen in the acinus region of the lung. We define a «best possible acinus» for which the acinus cut
by a plane has a length equal to the ratio of the diffusivity to the membrane permeability. The observed
anatomy of the acinus of several animals corresponds roughly to this optimized geometry. This close
relation between morphometry and transport parameters like diffusivities and permeabilities could help
to understand the design of biological organisms.
1 Introduction
Transport across rough or porous surfaces is a basic problem in the study of
several natural or industrial processes. For instance, in the design of high current
batteries it is natural to consider porous electrodes as a means of increasing the
net output current. Any process that is limited by transport across a surface or
interface can be enhanced in this way. Many systems like plant roots, villi in the
human intestine or lung alveoli are found that have a very irregular or ramified
geometry that can be considered as approximate examples of large area «fractal»
structures [1-3]. In the case of a membrane, neutral reacting species are brought
to the membrane surface by diffusion currents whereas in a battery the ions are
transported by the electric field. The transport across the membrane plays the same
role as the redox reaction on the battery electrode [4]. The problem is best stated
in the electrochemical case and was studied extensively in this context because
frequency dependent transport can easily be studied experimentally by impedance
spectroscopy. A variety of studies have indicated that very often the impedance of
electrochemical cells behave as
Z = Ro + k(jw)-1) (1)
where Ro represents the electrolyte resistance, w is the frequency of the a.c. ap-
plied voltage, j is (-1) 1/2. This behavior is called «constant phase angle» (CPA)
response and 7] is the CPA exponent which depends on surface electrode rough-
ness (k is a constant). A number of groups have worked on the fractal impedance
problem and a number of papers have been published on this subject by A. Le
Mehaute, L. Nyikos and T. Pajkossy, T. Kaplan, L.J. Gray and S.H. Liu, M. Blunt,
242 Sapoval
R. de Levie, W. Geertsma, J.E. Gols and L. Pietronero, A.M. Marvin, F. Toigo and
A. Maritan, W.H. Mulder and J.H. Sluyters [5-8]. A complete list of references
up to 1990 is given in the review given in reference [5].
Up to very recently, there existed apparent contradictions in the literature on
that subject but we believe that these contradictions can now be lifted and a single
quantitative model, con finned by numerical simulations and by experiments, can
now be considered as giving a proper understanding of the transfer across self-
similar interfaces [9]. The scaling form of this theory is presented in this volume
by R. Gutfraind and B. Sapoval [10]. Our purpose here is to apply what we
know from electrodes to the problem of diffusion to membrane in a biological or
physiological context. Here diffusion could be diffusion of oxygen or C02 in air
at the level of the pulmonary acinus or diffusion of Na+ or other ions or drugs in
the kidney or of some nutrient in the intestine.
It was shown that an exact equivalence exists between the response of an
electrode under d.c. conditions and the net diffusion across a membrane of the
same geometry. There exists also a si mple and direct connection between,the a.c.
and the d.c. response of an electrode [4,5,8]. The electrode results that we 'use here
have been obtained by several methods: numerical simulation, scaling arguments
and a continuous fraction expansion study of the problem. They allow to explain
quantitatively the impedance measurements on model electrode in d = 2 [9].
We present briefly the 2d results and their extension to more real 3d systems
and uses these results to understand why and to what extent fractal exchangers
can be more efficient. The reader is refered to [10] to find the scaling arguments
that justify the following results.
2 The Diffusion Results in D = 2

We consider here the self-similar fractal membrane shown in Fig.I. The simplified
diffusion system that we consider has a diffusion source at the bottom of the figure
where the concentration of diffusing particles is constant. We then consider systems
in which particles have to diffuse before crossing the membrane. The problem that
we consider is then a problem of diffusion to and across the membrane. Diffusing
particles (think of oxygen in air) emitted at the diffusion source execute random
walks. If the random walker collides with the membrane it is absorbed with a
finite probability a (also known as the sticking probability) which corresponds to
the existence of a finite penneability of the membrane. In a diffusion situation
most of the particles which leave the diffusion source return to it. This is due
to the fact that the net flux, from Fick's law, is proportional to the gradient of
the concentration and not to the concentration itself. The net diffusion transfer
between the source and the membrane is due to the very few particles that are
absorbed before returning to the source. There are then two limitations to the
transfer efficiency. First the particles have to diffuse to the membrane and second
they have to be absorbed by the membrane. In much the same situation, in an
Transfer to and across Irregular Membranes 243
Fig. 1 Schematic representation of a fractal membrane system in d = 2. Here the dimension is

Of = log 5/ log 3. The membrane has a smaller scale Lo below which it can be considered
as smooth. The size of the system is L. We assume that an external mechanism maintains a
constant concentration Co at the diffusion source. The diffusion source is situated at a distance
of the order of L. In our discussion of the lung the diffusion source is supposedly situated
near the entrance of the acinus.
electrochemical cell, ions have to be transported through the electrolyte and then
have to undergo a redox reaction on the active electrode to finally contribute to
the current. The total system then presents a macroscopic diffusion admittance Yo
defined by
<I> = Yo Co (2)
where <I> is the total flux and Co is the concentration at the source. The admittance
must then be written, in analogy with Eq.(l) as
(3)
where RB is the resistance of the bulk (the gas or the liquid region that diffus-
ing particles have to cross) and YM the admittance of the membrane. The two
244 Sapova/
transport parameters that characterise the problem are the diffusion coefficient D
and the coefficient of transport across the surface or permeability W. (If c is the
concentration at the surface of a linear membrane of length L the net flux across
the membrane is cWL). The membrane here is supposed to be passive and we
neglect back diffusion coming from the other side of the membrane where the
concentration is supposed to be zero.
It is useful to keep in mind a lattice picture of the diffusion process. If
diffusion results from random walks on a square lattice with lattice spacing a and
jump rate per unit time (1/7) we have D = a2 /47 and W = M / 47. Note that
nature gives D and W but not CY. The absorption probability cy is not determined
uniquely from D and W but depends on the lattice spacing a in this picture of
diffusion. The true physical variable in that problem is the length [8]
A=D/W=a/cy. (4)
which is defined without a necessary lattice representation.

There exists two regimes for the diffusional transfer. In the small permeability
regime the resistance to the transfer is limited by the membrane resistance itself
and not by the difficulty for the particles to reach the membrane or bulk resistance.
The diffusion admittance is then the admittance of the developed fractal line
(5)
where L is the macroscopic size of the electrode, La is the size of the smaller linear
part of the electrode and Df the fractal dimension. This is the classical regime.
For higher values of the permeability one reaches a fractal regime in which the
membrane admittance is
(6)
The crossover between the high and low permeability regime is obtained when
these expressions are equal or
D/WLo = L/Lo = (L/Lo)Df (7)
In the fractal (high permeability) regime it is useful to consider the euclidian

distance that a random walker that hits the membrane with absorption probability
s travels in space before being finally absorbed. We call this distance Lc(A). The
above lattice representation of the diffusion process is useful to compute Le (A).
In the picture of a walk on a lattice, the walk is terminated (by final absorption
on the membrane) when the number of fractal sites within the euclidian distance
Le(A) (equal to (Lo/a)(Le(A)/Lo)Df) is equal to the average number of collisions
needed by the walker to be absorbed. This number is of order l/CY. Writing the
equality gives the relation
(8)
Fig. 2 The generator of an hypothetic detenninistic fractal membrane in d = 3 with a fractal

dimension Dr = log 13/ log 3.
When this length is equal to the size of the system L, then the random walker has
the same probability to return to the source (and then be lost for the transfer) than
to be absorbed. Then we expect that when relation (7) is satisfied we have
(9)
In d = 2, RB is typically the resistance to diffusion of a square of side L. Its value

is then RB = D - I and one can verify that this condition is satisfied by replacing
(7) in (5).
3 Extension to d = 3 Membranes
The generator of a possible deterministic fractal membrane in d = 3 is shown in
Fig.2. In the case of self similar membranes embedded in d = 3 space there exits
also a low permeability regime where [9]
(10)
and a high permeability regime where
YM = L2(D/ Lo)(Dr-2)/ (Dr-l)wl / (Dr -l)

(11 )
= (A / Lo) (Dr- 2 )/(Dr- I )L 2W
The crossover is obtained in that case when
(12)
For this crossover value the value of the membrane admittance is equal to the value
of the admittance to reach the surface, which, in d = 3 is typically the admittance
of a cube of size L : YB = DL. Note that the statement that, at the crossover point,
the value of the classical admittance of the fractal membrane is equal to the access
resistance is very simple. It may then be of more general value and may apply to
246 Sapova/
SIZE L (log scale)
Fig. 3 Schematic behavior of the admittance of the total cell as a function of the size of the fractal
membrane. For small sizes the admittance is proportional to LOf while for larger size the
admittance is limited by the access resistance inversely proportional to the size. The optimum
for the acinus for which Dr = 3 should be found at the crossover between the two regimes
because the specific admittance YO / L3 in that case would decrease above the crossover.
many irregular geometries. This is probably why the above considerations seem
to apply to the acinus of several animals, as discussed below, although their real
geometry is not a simple fractal [11].
If one considers systems of small sizes (with L > Lo but not too large) the
resistance of the system is dominated by the membrane admittance which varies
as L Dr. Considering larger and larger systems the admittance decreases with this
power law of the size. This is true up to the crossover point after which the fractal
membrane admittance varies as L2 but the access resistance, which dominates
now the total admittance, varies only with the power one of the size L. This is
shown schematically in Fig. 3. The cross over condition, inverting relation (12)
gives A = Lo(L/Lo)(Dr- I ). This corresponds to the fact that the length of the
perimeter of a cut of the fractal membrane by a plane, which is a fractal with a
fractal dimension equal to (Df - 1), has a length equal to A.
4 Comparison between fractal Membrane Geometry

and ordinary Geometry
Before comparing fractal and non fractal exchangers it is useful to realize that d =
2 and d = 3 fractal systems behave differently. In d = 2 the system admittance is
YD = (D -I + YM ) -I . If one considers larger and larger exchangers, the admittance
increases with the size up to the situation where the resistance of the system is
dominated by the access resistance to diffusion D- 1 which do not depend on the

size. In consequence in d = 2 a system with a size larger than Le(A) is not a
better exchanger than a system of size L e .
In contradistinction, in d = 3 the bulk admittance is equal to LD and the
admittance is Yo = ([ L D j- 1 + YM1) -I which increases for larger and larger
fractal systems (see Fig. 3). One can compare the admittances given by relations
(10) and (11) with that of a flat membrane of same area L2 which is simply equal
to Yfl. = WL2. The admittance ratio YM/Yf 1. is equal to (A/L o)Df-2 in the low
permeability regime and to (A/L o)(Df- 2)/(Df- l ) in the high permeability regime.
In both cases there is a net increase of the admittance due to the fractal geometry.
Note that the smaller Lo and the larger the fractal dimension Df, the larger the
fractal admittance. These results give some insight to what extent fractal geometry
increases pratically the effectiveness of transfer through irregular membranes.
The robustness or adaptability of fractal exchangers is related to the fact that
the membrane admittance in the fractal regime is a power law of the transport
coefficients with exponent (Df - 2)/(Df - 1). This number is smaller than one.
This means that the effective flux through fractal membranes is less sensitive to the
physical conditions or that a same membrane can absorb different species which
have different permeabilities and/or diffusivities.
Also, in the fractal regime, the regions of the fractal surface which are really
active for the transport are different for different species. This is discussed in
[10] . For a multiple species transfer there could exist a space selection or filtering
along the fractal surface. (In contrast, in the low permeabilty regime the transfer
is uniform.)
5 Application to Airways
The above discussion indicates that the admittance per unit volume or specific
admittance defined as YD/L 3 = ([LDj-1 + yM1)-I/L 3 varies as L Df-3 for small
Land L -2 for large L. If the structure is dense Df = 3 and the specific admittance
does not depend on the size of the fractal up to the critical size discussed above
where it decreases with the power L2. From this discussion a geometry with a
dense (Df = 3) arrangement of cells of smaller size Lo, up to a scale L satisfing
condition (12) should be optimum because lower losses in the higher airways ask
for the larger size compatible with a large specific admittance. In that sense the
«best possible acinus» size should be that for which the perimeter of a cut of the
acinus by a plane should have a length of the order of A.
The data that was obtained by E. Weibel [11] allows to compare the above
considerations with what is observed in natural systems. The transport coefficients
governing diffusion of oxygen in air and its capture by the alveolar membrane
are D = 0.2 cm 2 sec-I and W = 10- 2 cm sec-I leading to L = 20 cm.
In the lung the transport process is due to hydrodynamic flow in the superior
248 Sapoval
Animal Mouse Rat Rabbit Human

Va (l0-3 cm 3) 0.4 1.76 3.46 187
Sa (cm 2 ) 0.42 1.36 1.55 69
Computed sac
9.5 13 22 27
diameter (l0-4cm)
Computed perimeter
5.7 11.2 10.3 120
Sa/Va1/3 (cm)
Table 1 Dense model parameters of the acinus computed from the data of reference [11-14] listed
on the two top lines. The third line gives the diameter of a sac and the last line gives the
perimeter deduced in the dense model.
airways and to diffusion within the alveoli. The acinus region is the region where
there is a transition between these processes so that it is not sure that the above
calculations which takes into account only diffusion can really apply. Nevertheless
in a simplified picture we consider the acinus as a dense [11] arrangement of alveoli
of size Lo up to a size L. In this simplified picture Lo is of the order of the diameter
of an alveolar sac while L is of the order of the macroscopic size of the acinus.
The length of the perimeter of a cut is then of the order of Lo(L/Lo)2.
Following these guidelines one can make a comparison for several animals:
the mouse, the rat, the rabbit and the human lung using the data of references
[11-14] In these papers, the quantity of interest which have been determined are
the acinus volume Va together with the alveolar surface per acinus Sa . In terms
of the above quantities Va = L3 and Sa = (L/Lo)3L5 = L3/Lo. The quantity Lo
can then be found as Va/Sa and the perimeter Lo(L/Lo)2 is equal to Sa/VaI / 3 . If
our interpretation is correct the length Va/Sa should be found of the same order of
magnitude as the typical small alveola or sac diameter and the quantity Sa/VaI / 3
should be found of the same order of magnitude as A that is of the order of 20
cm. The data that we compute from Va and Sa is given in Table I.
The values of Lo found from these calculations are a little smaller but have
the same order of magnitude as the sizes of the sacs observed in the pictures of
references [11-14]. Concerning the perimeters, although there exist discrepancies,
the general agreement can be considered as satisfactory if one takes into account
the oversimplified model that we use and the fact that the value A = 20 cm is
found from transport measurements only and not from geometrical determinations.
It seems that these considerations can be extended to the gills of fishes in
which the geometry is regular. Because the diffusion coefficient of oxygen in water
is of the order of 10- 5 cm 2 sec-I the value of A should be of the order of 20
micrometers if the permeabilty keeps the same value. No conclusive results exist
in this case but the size of the gills is comparable to that value and the gills have
a regular structure which corresponds to Df = 2 and not to Df = 3 [15]. In that

case, equation (12) indicates A = L as approximately observed for several fishes.
6 Conclusion
The overall agreement between the anatomical and physiological data and the
expressions that we predict for the «best possible acinus» seems to indicate that
the design for the oxygen exchanger in the lung of several animals is not far
to be the best: The morphometry seems to be in satisfactory agreement with an
optimized geometry which takes into account diffusion only in the acinus region.
A better model should take care of the possible remaining effect of air flow in this
region. In our restricted situation, we have shown that the best fractal membrane
should have a fractal dimension equal to 3 and a size such that the perimeter of
a cut is of the order of A = D / W. As quoted above, the statement that at the
crossover point which represents an optimized situation, the value of the classical
admittance of the fractal membrane is equal to the access resistance is very simple.
It may then apply to the acinus although the real geometry of the acinus is not
«simply» fractal.
One should note that the lung is a simple gas exchanger in the sense that only
two gases are exchanged. If one considers the transfer of several species with very
different transport parameters, the simultaneous optimisation of the transfer should
imply that different A's should correspond to the same morphometry. This cannot
be realized by an homogeneous membrane. In an optimized mUlti-species filter
or exchanger membrane system this could be nevertheless realized by a suitable
distribution on the membrane surface of cells permeable to specific species. In that
sense it is possible that the micro geometry of the distribution of specific cells (or
of cells with active transfer) in inhomogeneous membranes permits to optimize
the transfer of very different species in the same organ. That could be the case
of the filtration by the kidney. In such a case the above simplified picture of the
transfer can help to understand the relation between anatomy, morphology and
physiological function .
Acknowledgement:
The author gratefully acknowledges helpful discussions with R. Gutfraind and
E.R. Weibel. Laboratoire de Physique de la Matiere Condensee is U.R.A. D1254
of C.N.R.S.
250 Sapova/
References
[I] B. Mandelbrot, The fractal geometry of nature (W.H.Freeman and company,
San Francisco, 1982)
[2] B. Sapoval, Fractals (Aditech, Paris,1990)
[3] G.H. Bell, D. Emslie-Smith and C.R. Paterson, Physiology and Biochemistry,
(9th Ed.Churchill Livinstone:Edimburgh, 1976)
[4] B. Sapoval, Acta Stereologica, 6/III, 785 (1987).
[5] B. Sapoval, «Fractal electrodes, fractal membranes and fractal catalysts» in
Fractals and disordered systems, (Ed. A. Bunde and S. Havlin, Springer-
Verlag, Heidelberg,1991) p.207.
[6] T.c. Halsey and M. Liebig, Europhys. Lett.,14,815 (1991) and Phys. Rev.A
43, 7087 (1991).
[7] R. Ball, in «Surface Disordering, Growth, Roughening and Phase Transitions»
ed. by R. Julien, P. Meakin and D. Wolf (Nova Science Publisher,1993)
[8] B. Sapoval, J. Electrochem. Soc.l37,144C (1990) and Extended Abstracts,
Spring Meeting of the Electrochemical Society, Montreal, Canada, 90-1, 772,
(1990) and P. Meakin and B. Sapoval, Phys. Rev.A,43,2993 (1991).
[9] B. Sapoval, R. Gutfraind, P. Meakin, M. Keddam and H. Takenouti, Phys.
Rev.E,(1993)
[10] R. Gutfraind and B. Sapoval, «Scaling and active surface of fractal mem-
branes», this volume and Journal de Physique 1,1993.
[11] E.R. Weibel in «Respiratory Physiology, an Analytical Approach» ed. by
H.K.' Chang and M. Paiva. (M. Dekker, Inc., 1989)p.l and E.R. Weibel,
private communication.
[12] E.R. Weibel, «Design of biological organisms and fractal geometry», this
volulI)e.
[13] M. Rodriguez, S. Bur, A. Favre and E.R. Weibel, Amer. Journ. Anat.,155,143
(1987).
[14] B. Haefeli-Bleuer and E.R.Weibel, Anatom. Rec.,220,401, (1988).
[15] A. Beaumont and P. Cassier, Biologie animale (Dunod Universite, Paris 1978)
p.443.
Scaling and Active Surface
of Fractal Membranes
Ricardo Gutfraind 1) and Bernard Sapoval
Laboratoire de Physique de la Matiere Condensee, C.N.R.S.
Ecole Poly technique
91128 Palaiseau, France
1) present address: Groupe Matiere Condensee et Materiaux. Universite de Rennes I. Bat. lIB.
Campus de Beaulieu, 35042 Rennes Cedex. France.
Abstract. We study the properties of a Laplacian potential around an irregular object of finite surface
resistance. This describes the concentration in a problem of diffusion towards an irregular membrane
of finite permeability. We show that using a simple fractal generator one can approximately predict
the localization of the active zones of a fractal membrane of infinite permeability. When the the
permeability of the membrane is finite there exists a crossover length Le: In pores of size smaller than
Le the flux is homogeneously distributed. In pores of size larger than Le the same behavior as in the
case of infinite permeability is observed. namely the flux concentrates at the entrance of the pore. From
this consideration one can predict the active surface localization in the case of finite permeability. We
then show that a coarse-graining procedure. which maps the problem of finite permeability into that of
infinite permeability, permits to obtain the dependence of the resistance and of the active surface on
the surface and bulk properties. Finally. we show that the fractal geometry can be the most efficient
for a membrane that has to work under very variable conditions.
1 Introduction
Many natural as well as industrial processes take place in the environment of
surfaces or at the interface between two media. The surfaces that one encounters in
various of these processes are complex and irregular. Examples of natural processes
are the exchange of water and inorganic salts between the roots of a tree and its
surrounding environment [1] or the transport of oxygen to the blood flow through
the surface of the pulmonary alveoli [2]. Any process that is limited by transport
across a surface or interface can be enhanced using large-surface objects. This is
probably the reason why so many natural systems are found that have ramified
structures. Many of these surfaces can also be approximately described using
fractal geometry [3-6]. Therefore, using the tools provided by the fractal approach
can help in analyzing such processes.
In the simplest physical situation the mathematical problem is to find the
solution of the Laplace equation with mixed boundary conditions on the irregular
surface. This problem has been first considered in the case of electrochemistry
where the Laplace equation describes the electric potential in the electrolyte. In
electrochemistry, frequency dependent transport experiments can be performed
252 Gutfraind and Sapo va/
which have shown that in the presence of an irregular electrode many electrolytic
cells have an impedance which behaves as [7-9]:
(1)
where Rb can be considered as the bulk or electrolyte resistance, k is a constant,

rs is the Faradaic resistance which describes the finite rate of the electrochemical
reaction, "( is the specific capacitance describing the charge accumulation at the
surface and TJ is an exponent which satisfies 0 < TJ < 1. This behavior is known
as constant phase angle (CPA). For a smooth electrode TJ = 1 and TJ decreases
with the degree of roughness of the electrode. In direct current conditions the
surface contribution to the admittance will be of the form Ysurf. = k-1(rs)-T/ [
the inverse of the second term in Eq. (1) under direct current conditions]. It can
be easily shown that this problem can be mapped into that of the response of
an irregular membrane [10-11]. This is the case we study here. In the problem
of diffusion towards and across a membrane the Laplace equation describes the
concentration field which satisfies the steady-state diffusion equation:
(2a)
with the boundary condition C = Co on a source of molecules at certain distance

of the irregular membrane and
-V\lC = rs -IC (2b)
on the irregular membrane (see Fig. la), where V is the bulk diffusivity and the
term rs is the surface resistance which represents the resistance due to the finite
permeability (W) of the membrane itself. Eq. (2b) arises from conservation of
matter which imposes that transport to the membrane is equal to transport across
the membrane.
In the case of self-similar fractals like DLA clusters, diffusion fronts or deter-
ministic fractals as the one shown in Fig.(l b) the following result can be written
[12,13]:
(3)
where RT is the total resistance to transport, p is the bulk resistivity (which is
equivalent to 1IV), L is the size of the system, Lo is the smallest feature size of
the self-similar membrane and b is the thickness of the system. Thus TJ = 1I D,
where D is the fractal dimension. This result has been obtained by four different
methods. First a dimensional analysis with a scaling argument [14], second an
iteration method [12,13], third a scaling argument based on a simplified picture
of the working regions of a self-similar interface [15], and extensive numerical
simulations based on the exact analogy between electrical and diffusion Laplacian
fields [13,14].
Scaling and Active Surface 253
I II III
GAS OR UQU10
<1>..1.= we
(0)
(b)
(1)
(2)
Fig. 1 a) Schematic representation of the system. The rough surface is a fractal generator. b) Gen-
eration of a two-recursion level fractal. At each generation the middle third of each segment
is substituted by a square; the fractal dimension is D = log(5)j log(3).
254 Gutfraind and Sapoval
In this paper we discuss the concept of active surface which is the subset
of an irregular interface where the flux concentrates [12,15]. Although this no-
tion is an approximation, it provides a simple way to visualize the active zones
of fractal membranes. This concept also permits to obtain equation (3) through a
coarse-graining which maps the problem of rs -=I- 0 into that of rs = 0 (Section 2).
We also show that an important property arises from the physical picture of the
studied process: this is how the fractal surface adapts to environment conditions
(Section 3). In Section 4 we compare the numerical calculations of the flux dis-
tribution with the geometrical interpretation given in Section 3.
2 Active zone
In this section we show a practical way of visualizing the regions of a membrane
that can be considered as the active zones for the case rs = 0 (in this case the
boundary condition C = 0 substitutes the boundary condition 2b on the fractal
membrane). In this situation we search for the regions of the membrane which
receive the flux of molecules coming from the bulk (called the information set in
Ref. [12]).
Consider first the problem of the flux distribution onto the fractal generator
shown in figure lao If the source of diffusing molecules is not too close to the
fractal membrane the concentration map for a single pore is expected to behave as
shown in figure 2 [16]. In the representation of the picture the boundary between
the stripes are lines of equal concentration and the concentration varies with a
factor two from one line to the next one. The flux is perpendicular to the lines
of equal concentration and inversely proportional to their separation. One can see
that at the entrance of the pore, the density of equipotential lines next to the pore
walls is much larger than the density in the central part of the pore. This shows
that there is little penetration through the central part of the pore. Therefore, one
can see qualitatively that there is a zone of high flux (active zone) and a zone of
small flux.
Based on this and according to the scaling properties that the active zone must
have (it has to scale as L 1 [17]), we construct a generator as shown in figure 3
(top). From this generator we build the active zone by iteration (figure 3) but to be
consistent with the above requirement, it is necessary for the size of the generator
to be exactly equal to L. This is the only possibility if one wants to build the set
from a single generator. If the generator would be equal to (1 + o:)L, 0: being any
number different from zero, the active zone will not scale as L 1 in the iterative
process used to build the object. Thus the dashed line in the generator must extend
to exactly half of the depth of the well in figure 3 (top).
In summary, if we use the (approximate) notion of an active zone, for this
notion to be consistent with the known properties of the Laplacian field, the size
of its generator has to be taken equal to L. This idea is approximative in two
aspects: first not all the molecules coming from the bulk arrive in this fraction of
Fig. 2 Laplacian field around the fractal generator in the case rs = 0 [16]. Each stripe represents a
drop in the concentration by a factor of 2. The concentration C= I at a distance of the fractal
object equal to the width of the central pore.
the surface and second the flux is not homogeneously distributed on it. However,
it is a good approximation when one calculates the exponents that describe the
behavior of the system as it is shown in Section 3 and 4.
3 Scaling argument
3.1 Scaling form of the surface contribution to the total resistance
In this section we present the scaling argument which permits to obtain Eq. (3).
This argument is based on a coarse-graining procedure which maps the problem
with boundary condition (2b) on the above situation with rs = O.
For a real system with finite p and rs the first problem is to compare the sur-
face resistance Rsurf. with the resistance to access the surface Race .. For very large
values of r s , the resistance in the bulk (determined by the value of the resistivity
p) becomes negligible in comparison to the resistance of the surface. In this case,
the entire surface which can be considered as exposed to the same concentration.
The total resistance, which is determined only by the surface contribution, is then
proportional to rs and inversely proportional to the total area of the fractal surface:
rs
Rsurf. = D (4)
b(L/Lo) Lo
--------, r--------
I r,:--- I r,:---
I I I I
I I
I I
Fig. 3 Iterative process for the generation of the active surface. The solid line is the fractal object, the
dashed line is its associated active surface. At each recursion level each segment is substituted
by a version of the generator (top figure) rescaled by a factor(I / 3)n, where n is the recursion
level. This process assures that independently of the surface irregularities the total length of
the active surface is the linear size of the object.
For a two-dimensional system, independently of the system size, the access resis-
tance is of the order:
Race. '" p/b (5)
For given values of p and rs and when rs/Lo > p/b, namely the resistance of
the smallest feature of the fractal is larger than the access resistance, the geo-
metrical features can be separated depending on their size: Small parts have a
surface resistance larger than the resistance to access them, whereas large features
have a negligible surface resistance. Consequently, there exists a crossover length,
Lc(rs/ p), which separates two different geometrical behaviors: a) In pores smaller
Fig.4 The Laplacian potential around a four-iteration object: a) rs = 0 and b) for rs/(pLo) = 3.
The smallest pores are covered only by one stripe. In these pores the surface resistance is
larger than the resistance to access to the pore walls. They are approximated as exposed to a
constant concentration. In the largest pores the access resistance is much larger than the wall
resistance and the concentration map practically recovers the behavior of rs = 0 [ see (a)],
the only difference is a local effect near the walls.
than Le, the resistance is dominated by the surrace and the pore walls can be
approximated as exposed to the same concentration. b) In pores larger than Le the
system recovers the behavior of rs = 0 because the surrace impedance is smaller
than the access resistance in the bulk (see figure 4). Then the crossover length is
determined by the equality between the surrace resistance [Eq. (4)] and the access
resistance [Eq. (5)]:
rs p
rv - (6a)
b(LciLoFLo b
or,
Le rv Lo ( Lrs riD (6b)
p 0
This means that if one considers a coarse graining of the initial membrane to
the size Le (see figure 5), any larger feature will behave with a surrace resistance
small as compared to the access resistance plb. Once the object is coarse grained
to a new one where the size of the elementary unit is Le, the resistance of the
generated macrosites is small as compared to the access resistance and then the
Laplacian field recovers the behavior of rs = O. This implies the existence of an
active surrace of the linear size of the object L (Section 2). The difference now is
that the bulk resistance Rb (Eq. 1) is in series with a contribution of the surrace.
The surrace contribution to the resistance is then equivalent to (LILe) elements
of order (plb) or
Rsur. rv (plb)(LclL) (7)
which is equivalent to Eq. (3). Notice that the coarse graining does not affect
the geometry of the cell (the height-length relation), so the bulk resistance [Rb in
Eq. (1)], being independent of the surrace irregularities, will be the same in both
the original and the coarse-grained object.
We would like to point out that the scaling argument developed in this section
is general, it is based on the existence of an active zone of dimension one, but this
does not have to be exactly located as it was shown in Section 2. However, the
iterative procedure that we proposed allows to visualize this set in a very simple
manner.
3.2 Active surface and adaptability

The scaling form dependence of the active surrace on the resistance, rs , can
also be estimated using the preceeding argument. For a given value of rsf p,
the surrace which corresponds to each macrosite is rsl p itself and the number
of these macrosites is of the order LILe. The total surrace of the electrode is
ST = Lo(LILo)D so that the fraction of active surrace SAIST is:
(8)
An important property arises from the physical picture developed in this section:
This is how the fractal surrace adapts to environment conditions. One can easily
Rb
- -- ;l r.: - - -I r.: - - -I r.: ---

Lc 1 I
1 I I I
I I
I I
- -
.. ~
Fig. 5 A three-iteration object (top) is coarse-grained using the characteristic length shown in the
figure. In the coarse-grained object (bottom) the active surface recovers the behavior of Ts = 0,
namely its total length is L.
260 Gutjraind and Sapoval
see that this is because the fractal symmetry provides new active surface when
Ts increases and this part compensates for the larger value of Ts. Then a fractal
membrane has the capability of providing new active surface whenever it has
to absorb molecules having a lower permeability constant, avoiding in part a
reduction in the transfer rate. This property can be crucial in the life of many
natural systems, where a rough membrane can provide a solution to keep a suitable
supply rate for nutrients being in a broad range of permeability values.
4 Numerical Results
In this section numerical results are compared to the theoretical predictions of
the previous sections. To verify that the response of the studied object behaves
according to Eq. (3), we used a relaxation method to numerically compute the re-
sponse of a five-iteration object of the type shown in figure 1b. In this calculation
one obtains the total resistance, RT which is the sum of the bulk and membrane
resistances. The resistance of the bulk Rb is the resistance RT(Ts = 0). The expo-
nent 'rJ is then obtained from the slope of the log-log plot of [(RT - RT(Ts = 0)]
against Ts. The value 'rJ = 0.64 was obtained in good agreement with the value
'rJ = 1/0 = 0.68 as shown in figure 6 [15]; the same result was obtained for
several other fractal objects [14].
The idea of the theoretical active surface presented in Section 2 for the case
Ts = 0 can be tested by comparing its theoretical localization with actual numerical
observations. The first thing one can ask is what is the fraction of the total flux
that indeed reaches it. When Ts = 0 the SA/ST ratio is equal to (L/Lo)1-0. In the
case of a five-iteration object of the type shown in figures 1-3 SA/ST = 7.8%.
From the solution of Laplace equation the flux that arrives in each surface site (jo)
is computed. For comparison purposes one can look at the 7.8% of the surface
associated with the largest jo values. First, one finds that 86% of the flux arrives
in this fraction of the surface. Second, the localization of these sites, as shown
by the black points in figure 7, is close to that predicted from the theoretical
active surface as suggested in figure 3 [15]. Notice then that the simple procedure
suggested in Section 2 gives a good approximation of the localization of the most
active zones of the fractal membrane.
We present now the visualization of the working regions of the membrane
for different values of the surface resistance (Ts). To build the active surface we
take the sites associated to the largest jo values up to the accumulation of a given
percentage of the total flux. The case of 86% of the flux for Ts/ pLo = 5 is
shown in figure 8a. One can distinguish two classes of pores, those that behave
homogeneously (a continuous line of active zone extends to the bottom of the
pores) and those in which the behavior characteristic of Ts = 0 is recovered (the
active zone penetrates one half of the pore width). It is on this fact that is based
the idea of the coarse-graining procedure: The existence of a crossover length
upon which one can coarse-grain the object (figure 8b) and recover the behavior
of Ts = O.
.Q
a:
0.1
I-
a:
. I
0.01
10 100 1000
rs
Fig. 6 Dependence of (RT - Rb), in computer units, on Ys for a five-iteration fractal object of
the type shown in Fig. (lb). 1) = 0.64 is obtained from the slope in good agreement with
1) = 1/0 = 0.68.
Fig. 7 Numerical estimation of the active surface (Ys = 0) for a five-iteration object (the active sites
are in black). The linear size of the object is L = 729 lattice units. The active surface are
the L = 729 boundary sites associated with the largest values of flux per site; 86% of the
total flux really arrives in this zone. Notice that the part of the surface considered as active
is approximately one half of the pore width at each pore wall.
(a)
LD Lln
.,
[
(b)
Fig.8 Numerical estimations of the active surface for a five-iteration object when ys / {pLol = 5. a)
In the original object and b) in the coarse-grained object. Notice that the localization of the
active surface in the coarse-grained object is close to its theoretical localization.
I
(b) L JJ [IL J
I
Fig.9 Numerical estimations of the active surface for a five-iteration object when rs / (pLo = 25).
a) In the original object and b) in the coarse-grained object. Notice that there is a shift in the
value of Lc by approximately a factor of three with respect to that of figure 8, in agreement
with Eq. (6).
The case of (rs/ pLo) = 25 is shown in figure 9. The picture shows again
where arrives 86% of the total flux associated with the largest jo values. The same
type of behavior as in the case (rs/ pLo) = 5 is observed, but with an increase
of the crossover length. Comparing figure 8 and figure 9, one can see that the
visual crossover length has been shifted approximately by the factor (LdL2)1/0 =
(25/5)1 /0 = 3 as predicted by Eq. (6).
5 Conclusions
The concept of active surface can be a powerful tool in the study of the macro-
scopic behavior of an irregular membrane. It gives first a very simple picture of
the working regions of a fractal membrane in the case rs = O. Then a simple
coarse-graining procedure permits to find the active surface when rs -=I=- O. This
procedure provides a simple visualization of the process and a simple explanation
of the transfer equation (3). We have also shown that a fractal membrane is self-
adapted to variable working conditions. We think that a fractal membrane may be
the optimal solution to keep a suitable rate of transfer for absorbing simultaneously
molecules having different permeability values. This property can be crucial in the
life of many natural systems, where a surface showing irregularities at all length
scales can provide sufficient surface for the transfer at an adequate rate of all the
nutrients necessary for the metabolism.
References
[1] Hall E, Oldeman R.A.A., Tomlinson P., Tropical Trees and Forests (Springer-
Verlag, New York, 1978).
[2] a) Weibel E.R,The Pathway for Oxygen (Harvard University Press, Cam-
bridge Massachusetts, 1984). b) Weibel E.R., Respiratory Physiology, an An-
alytical Approach, ed: Chang H.K. and Paiva M. (Dekker, Basel, 1989), p.l.
[3] Mandelbrot B.B., Fractal Geometry of Nature, 3rd ed. (Freeman, New York,
1983).
[4] Sapoval B., Fractals (Aditech, Paris, 1990).
[5] a)The Fractal Approach to Heterogeneous Chemistry: Surfaces, Colloids,
Polymers, ed: Avnir D. (Wiley, Chichester, 1989).
[6] Feder 1., Fractals (Plenum, New York, 1988).
[7] Scheider w., J. Phys. Chern. 79 (1975) 127.
[8] Armstrong D. and Burnham R.A., J. Electroanal. Chern. 72 (1976) 257.
[9] Rammelt U. and Reinhard G., Electrochimica Acta 35 (1990) 1045.
[10] Sapoval B., Acta Stereologica, 6/111 (1987) 785.
[11] Sapoval B., Fractal electrodes, fractal membranes and fractal catalysts in
Fractals and disordered systems, ed: Bunde A. and Havlin S. (Springer-Verlag,
Heidelberg, 1991) p. 207.
[12] a) Sapoval B. and Gutfraind R., Workshop on Surface Disordering, Growth,
Roughening and Phase Transitions, Les Houches April 1992. ed: Julien R.,
Kertesz 1., Meakin P. and Wolf D. (Nova Science Publisher, 1993, page
285). b) Sapoval B., Spring meeting of the Electrochemical Society St. Louis,
Missouri 92-1 (1992) 514.
[13] Sapoval B., Gutfraind R., Meakin P., Keddam M. and Takenouti H. (phys.
Rev. E, accepted)
[14] Meakin P. and Sapoval B., Phys. Rev. A 43 (1991) 2993.
[15] R. Gutfraind and B. Sapoval (Journal de Physique I, to be published)
[16] Evertsz C., Mandelbrot B.B. and Normant F., 1. Phys. A: Math Gen. (1992)
1781.
[17] Jones P. and Wolff T., Acta Math., 161 (1988) 131.
Structure Formation in Excitable Media
Martin Llineburg
Bielefeld University,
Fakulllit flir Mathematik
UniversitatsstraBe
D-33615 Bielefeld
Excitable media are widespread in nature. They play important roles in physics and
chemistry as well as in biology and medicine. Applications comprise phenomena
as diverse as the pigmentation patterns of vertebrate skins or of shells of molluscs,
cardiac arrythmia, formation of galaxies, energy metabolism, aggregation of slime
mold amoebae, spatio-temporal EEG-pattems and circadian rhythms in physiology
and biological populations (cf. [6], [10], [11]). The most prominent feature of an
excitable medium, of course, is its ability to receive and distribute excitation. The
main phenomenon observed in such media, therefore, is that of waves formed by
the propagation of excitation gradients. As the examples above might suggest, the
supporting medium need not be a continuum but can also have discrete structure.
The former, as a rule, is modeled by means of partial differential equations, the
latter more often by cellular automata. Obviously, many reaction-diffusion and
aggregation processes are intimately connected with the phenomena encountered
in excitable media.
Starting point for our own investigations in Bielefeld on excitable media
was the self-similar curve depicted in figure 1. It shows the varying degree of
overall activity of certain catalytic metal surfaces which has been measured by
N. Jaeger et al. at Bremen university (cf. [7] and the literature quoted there). In
their experiments, a flow of carbon monoxide was sent from below through a
heap of granular pieces of catalysts. Theses catalysts were supposed to oxidize
continuously at a constant rate the carbon monoxide to carbon dioxide. Instead,
measurements showed perplexing patterns of spontaneous activity reduction. Now,
compare figure 1 with figure 2 which has been gained from a completely differ-
ent situation. The latter shows for 0 :::; n :::; 120 the number K(n) of binomial
coefficients (I:) which are odd. If you visualize these coefficients in the Pascal
triangle modulo 2, where black stands for odd and white for even (fig. 3), K(n)
is simply the number of black spots in the nth column. As is well known, the
Pascal triangle modulo 2 for n ---+ 00 is equivalent to the Sierpinski gasket, so
that the self-similarity of figure 2 is no surprise. Furthermore, a fact which will
be of importance later on, this triangle can be interpreted as the temporal record
of a I-dimensional cellular automaton, Pascal's parity automaton: starting with
exactly one nonzero entry in an infinite vertical line of cells (the leftmost column)
it develops by the rule that the state of a cell at time t + 1 is defined to be the
sum of its own state and that of its lower neighbour at time t.
Structure Formation in Excitable Media 267
%C Oscillation Pattern With Selfslmilanly
100
30
30 min
Fig. 1
K(n)
60
40
20
O~~~~~~~~-r~~~~~~-r-'~
10 20 30 40 SO 60 70 80 90 100 110 120 n
Fig. 2
Is there any reason for the resemblance of the activity record and the Pascal
histogram? There is, if you are willing to believe in the following first approxima-
tion model [1] : Below a certain temperature threshold each single catalyst oscillates
periodically between its active and its passive state, while above this threshold it
exhibits bistability. Bistability means that it remains active when active and passive
when passive. Moreover, we may assume that activity of a catalyst raises the tem-
perature of the catalyst above it, so the upper one will remain or become bistable,
preserving its state of activity or passivity, respectively. In contrast, passivity may
cool the catalyst just above - so in this case the upper one remains or becomes
oscillating and immediately starts changing its state of activity. If you look more
closely to these rules, you see that finally you can forget about bistability and
oscillations since only activity and passivity are really relevant for deciding how
the process will continue. Namely, imagine two catalysts lying above each other.
Then the upper one will be active after the next time step if and only if it and
its lower neighbour are both passive or both active. If you identify active with 0
and passive with 1 this coincides with addition modulo 2. Hence, the catalyzing
rules correspond precisely to the rules of Pascal's parity automaton. Thus, the
model would explain the experimental observations and provide a rather simple
description of a seemingly highly complex phenomenon.
268 Laneburg
-II
Fig. 3
Elaborating on Pascal's parity automaton in order to develop a chemically

more realistic 2-dimensional variant, Martin Gerhardt and Heike Schuster from
Bremen and Bielefeld university designed what they later came to call the hodge-
podge machine (cf. [2], [4]): This is a 2-dimensional cellular automaton consisting
of cells all of which would perform simple sawtooth oscillations if left alone. By
diffusion coupling of neighbouring cells the speed at which they run through their
cycles is either enhanced or reduced, depending on the phase of their neighbours.
These simple rules produce moving wavefronts which organize themselves into
an intricate scrollwork of spiral and concentric patterns (fig. 4) resembling to an
amazing degree those known from the Belousov-Zhabotinsky reaction. Moreover,
recently a group of chemists in Berlin has developed new experimental methods
which enabled them to observe, in fact, such patterns on a real catalytic metal
surface [8].
Of course, coupled sawtooth oscillations as in the hodge-podge machine can-
not explain but in a roughly qualitative way what happens on a catalyst's ~urface.
Therefore, A. Dress in close cooperation with the aforementioned Bremen chemists
group developed a more realistic model, called «ideal storage», for the behaviour
of a single catalyst [3]. The discrete time steps it employs are only for the sake
of convenience and simplicity. The following are the rules for a single catalytic
crystallite:
Qn+ I : = Qn + S - Rn . Qn
Rn+l : = F(R n . Qn)
Thereby, Q is the quantity of reactive substance, R is the reaction rate, S is

the supply of reactive substance sent «continuously» to the catalyst and F is
Fig. 4
a monotonously increasing function defined on the non-negative real axis with

values in the intervall (0,1].
An analysis of this discrete dynamical system and its unique equilibrium
(Ra, Qa) = (F(S), SjF(S)) showed that the system can indeed exhibit oscillatory
behaviour. More precisely, for standard choices of F and with increasing values
of the control parameter S, the system undergoes two discrete Hopf bifurcations,
first from a low level stable equilibrium to an intermediate periodic attractor Ks
and then back to another stable equilibrium, now of high reactivity. Since this
coincided well enough with what was known from experiments, the question was:
What behaviour do coupled catalysts exhibit, do they, for instance, produce spiral
waves?
More precisely, we asked for the evolution of a certain cellular automaton. Its
cells were arranged like the junctions of a regular rectangular grid, the neighbours
of each cell being its 3, 5 or 8 surrounding companions depending on whether the
cell lied on a comer, on the boundary or in the inner part of the rectangle. After
random initialization, the (Q, R)-values of the cells were simultaneously updated
by applying the above mentioned equations and thereafter averaging the reaction
rates of neighbouring cells as follows:
rn+l := W· rn + (1 - w) . an
Thereby, r n and r n+ I are the reaction rates of a single cell at time n or n + 1,

respectively, an is the arithmetic mean of the reaction rates of its neighbouring
270 Liineburg
cells at time n and w is a weight with 0 :::; W :::; I . Hence, for w = I no coupling
is present and every cell evolves without being influenced by others, and for
w = 1/9 = 0.1 all reaction rates of a cell's neighbourhood play the same role for
the dynamics of the automaton. A coupling via the reaction rates is reasonable,
since the reaction rate of a single catalyst depends monotonously on the amount
of heat produced during the reaction process (cf. [3]) and heat gradients move and
dissolve by conduction and diffusion. An interpretation of the coupling parameter
w will be given later.
Before presenting our results (for details and references see [9]) we have to
introduce some notions: Let I(F, 5) be an «ideal storage» built on the monotonous
function F and with supply 5, and A(F, 5, w) a 2-dimensional cellular automaton
the cells of which are such ideal storages coupled as described above with cou-
pling parameter w. Typically, A(F, 5, w) consists of at least 100 times 100 cells.
Furthermore, let
n(F) : = {51 0 <5 E IR, I(F,5) is a global oscillator},

~(F,w): = {5 E IRIA(F,5 , w) permits spirals stable in time}
and
a(F,w): = {5 E IRI A(F,5 , w) permits spontaneous formation of spirals} .
«Stable in time» means that such a spiral does not vanish as long as the parameters
F and 5 are not changed. The fact that after random initialization the automaton
produces spiral waves is called «spontaneous formation of spirals». If w = 0.1,
which will be the case for most of the time, we simply omit it, i.e. A(F,5) =
A(F, 5, 0.1), etc.
The occurrence of spontaneous formation of spirals and the adaptability of
spiral waves to variations of the supply 5 are the two principal phenomena we
have studied. For this purpose the following experiments were carried out:
EI) After initialization at random, A(F, 5, w) was run with 5 E n(F)

E2) After spirals stable in time had evolved during EI with w = 0.1, the supply
was changed.
EI proved that spontaneous formation of spirals is possible with suitable choices of
F, w and 5. The combination of Eland E2 showed a strong discrepancy between
the attractivity and the stability of spirals: a(F) is a relatively small subintervall
of ~(F) n n(F). It consists of those 5 in whose neighbourhood the size of the
cycle Ks changes drastically. In contrast, ~(F) is an intervall of similar size as
n(F) but not a subinterval of n(F); even for values 5 not in n(F) spirals can
remain stable. Hence, coupling of cells which would tend to the same equilibrium,
if left alone, can give rise to highly non-trivial forms of symbiosis, manifesting
itself by endlessly turning spiral waves.
Starting with spirals stable in time, variation of the supply in general changes
the «spectrum» of the spirals which adapt to the new conditions, this spectrum es-
sentially being the states every cell belonging to the spiral runs through cyclically.
Increase or decrease of 5 in general induces an increase or decrease, respectively,

of the spirals ' curvature. Figures 5-7 show snapshots of a typical example (the two
variables Q and R of each cell are depicted on the left and the right, respectively):
A(F, 5, oj) was started with F defined as a function of x by
and D = 86, and with 5 = 91 E a(F) . After spirals had fully developed (fig. 5),
5 was set first to 140 and then to 200, the process after every increase running
long enough for complete adaptation of the spirals (fig. 6-7).
Variation of the coupling parameter w between 0.1 and 1 showed that even
quite weak a coupling induces strong correlations between the phases of neigh-
bouring cells. With weaker coupling only the size of the single spirals decreases
while their total number increases. This to a certain degree can be interpreted as a
scaling phenomenon: Consider the catalyst as an ensemble of coupled subcatalysts
by subdividing it into parts of equal size. Then, of course, the coupling intensity
must depend on the size of the boundaries these parts have in common. In a
coarser subdivision these boundaries in comparison to the area of the subcatalysts
are smaller than in a finer one. Since a weaker coupling of the automaton 's cells,
as stated above, corresponds to a smaller degree of heat exchange, such a coupling
could be interpreted as a coarser subdivision of the catalyst in question, i.e. as a
change of scale implying that the same number of cells represents a larger catalyst,
though of similar character, than before. The scaling problem we have touched
hereby is one of the major problems if not the main one which has to be solved
for a quantitative application of cellular automata.
Spontaneous formation of spirals stable in time is a non-trivial example of
self-organization. The temporal stability of these spirals is due to replication: a
situation develops which is non-stationary itself but the characteristics of which
are invariant in time as they replicate themselves in any time step.
Another important property of these spirals is their tendency to spread by
growth over all the space not yet occupied by other spirals. It implies that after
some time the whole automaton 's surface is tiled by spirals. This final situation
coincides with minimal temporal variance, though not constance, of the histogram
representing the number of cells in a certain state. Hence, spiral waves make the
automaton run as steadily as possible without resting. This could be most desirable
for a catalyst but maybe fatal for the heart since the working rhythm of the latter,
in contrast, must be fractal in case of health (cf. [5]).
272 Liineburg
Fig. 5
Fig. 6
Fig. 7
References
[1] A. Dress: Cellular automata: a simplistic approach to complexity. To appear
in: Neurobionics (eds. Bothe et al.), Elsevier Science Publ. 1993
[2] A Dress, M. Gerhardt, H. Schuster: Cellular automata simulating the evolu-
tion of structure through the synchronization of oscillators. In: From Chemi-
cal to Biological Order (eds. L. Rensing and N.!. Jaeger), Springer Series in
Synergetics 39 (1988), 134-145
[3] A Dress, N.I. Jaeger, PJ. Plath: Zur Dynamik idealer Speicher. Ein einfaches
mathematisches Modell. Theoret. Chim. Acta 61 (1982),437-460
[4] M. Gerhardt, H. Schuster: A cellular automaton describing the formation of
spatially ordered structures in chemical systems. Physica D 36 (1989), 209-
221
[5] A.L. Goldberger: Fractal mechanisms in the electrophysiology of the heart.
IEEE Engineering in Medicine and Biology, June 1992, 47-52
[6] A.V. Holden, M. Markus, H.G. Othmer (eds.): Nonlinear Wave Processes
in Excitable Media, Proceedings of a NATO Advanced Research Workshop.
Plenum Press, New York and London 1991
[7] N.!. Jaeger, K. Moller, P.J. Plath: On the classification of local disorder in
globally regular spatial patterns. In: Temporal Order (eds. L. Rensing and
N.I. Jaeger), Springer Series in Synergetics 29 (1985), 96-100
[8] S. Jakubith, H.H. Roterrnund, W. Engel, A von Oertzen, G. Ertl: Spatiotem-
poral concentration patterns in a surface reaction: propagating and standing
waves, rotating spirals, and turbulances. Phys. Rev. Lett. 65 (1990), 3013-
3016
[9] M. Liineburg, in preparation
[10] AT. Winfree: The Geometry of Biological Time. Springer-Verlag Berlin Hei-
delberg New York 1990
[11] D.A Young, A local activator-inhibitor model of vertebrate skin patterns,
Math. Biosciences 72 (1984), 51-58
Colony Morphology of the Fungus
Aspergillus Oryzae
Shu Matsuura and Sasuke Miyazima 1)

School of High-Technology for Human Welfare
Tokai University
Nishino 317, Numazu
Shizuoka 410-03, Japan
1) Department of Engineering Physics, Chubu University, Kasugai, Aichi 487, Japan
Abstract, Thick and homogeneous colonies of fungus Aspergillus oryzae with connected growth fronts
have been found to grow on solid agar media or on the media with high nutrient concentration, and
the roughness of their growth front have been characterized by self-affine fractals. On the other hand,
fairly ramified colonies with low growth rate have been seen on the media of low stiffness with very
low glucose concentrations.
1 Introduction
Most filamentous fungi are found in the soil, where the mineral matters, bacterial
colonies, waste products of living organisms, gaseous pores, water, etc., form ran-
dom crumb structures. The branched structure of filamentous fungi is advantageous
as a strategy to grow in such inhomogeneous substrates.
Mycelium is one giant cell consisting of the basal hypha, the aerial hypha,
and other organs for reproduction such as conidia. Leading hypha can continue to
extend even through the unfavorable field using nutrients transported from other
part of the mycelium, and branch out when they reach the nutritious areas. Different
physiological tasks are assigned to each part of the thallus through the active
transport of materials within the thallus. Such adaptation to the inhomogeneous
environments will contribute to the pattern formation of the fungal colony.
On the other hand, growth of thin filaments of fungi must be largely subject to
the environmental conditions, such as the nutrient content, pH, water, temperature,
and gaseous conditions. Also, several physical conditions such as the attachment
to substrate, osmotic pressure on the cell wall, the degree of complexity of the
medium, etc., might influence the regulation of colony formation.
In this study, we present the morphology of the 2-dimensional colonies of
fungus Aspergillus oryzae cultivated at 24°C from line inoculum on the synthetic
agar media with various glucose and agar contents. The strain forms thick colony
on the solid agar medium. Dependence of the roughness of growth fronts of these
colonies on the glucose concentrations is analyzed in terms of self-affine fractal.
The roughness of the growth front of I-dimensionally growing colony is
determined in the following way. Let us choose a point on the growth front and
Colony Morphology of the Fungus Aspergillus Oryzae 275
indicate as the position i. Then, the height hi of the point is defined by the distance
from the inoculation line to the front point i. Next, let us define a horizontal range
of 1 parallel to the inoculation line. When there are N digitized front points in this
range, the mean height h(l) of the colony within 1 is written as,
h(l) = L h;/N. (1)
Now, the roughness of front within 1 is estimated by,
(2)
Generally, (J increases with increasing h in the growth process. When the horizontal
range 1 is broadened, (J includes longer wave length fluctuations of front.
As for the random growth processes, it is known that (J is scaled with 1 and
h, in the fonn of so called self-affinity relations as l ),
(J(h , 1) rv 1° for 1« h, (3)

(J(h, 1) rv hi3 for 1» h.
Here, a is within the range of 0 < a < 1. A typical random growth is represented
by Eden model2) generated from a line seed at initial. In this model, every sur-
face point grows with an equivalent probability, and a and f3 values have been
detennined as a = 1/2 and f3 = 1/3.
Several examples are known to satisfy the above self-affinity relations such
as the fungal colonies 3), bacterial colonies4), burning fronts 5), and rupture lines
in paper sheets6 ) . In this paper, we present the dependence of h, (J, and a on the
glucose concentrations for the thick colonies cultivated on the solid agar media.
2 Experimental methods
The strain used is Aspergillus oryzae (supplied by the Institute for Fennenta-
tion Osaka), cultivated on the Czapek synthetic agar medium containing NaN03
(O.3wt%), K2HP04 (O.lwt%), MgS04·7H20 (0.05wt%), KCI (0.05wt%), FeS04 '-
7H20 (0.001 wt%), glucose (varied from 0.01 wt% to 5wt%), and Difco Bact-Agar
(1.5wt%, O.3wt%, and 0.15wt% for «solid», «soft», and «liquid-like» medium,
respectively). Stiffness of the medium is controlled by the amount of agar. 25
ml of sterile medium is poured into each petri plate of 9 cm in diameter. Line
inoculation is carried out by the method previously described in reference 3, and
the strain is cultivated at 24°C.
Photographs of colonies are taken with a 35mm camera. Colonies are illu-
minated from the bottom side by a halogen lamp. The black-and-white prints of
colony images are digitized using an image-scanner. The final resolution of images
is approximately 600 pixels in length of the total inoculation line.
If)
•
o
•
o
(%lM.) u01lUllu8JUOJ 8soJnI~
,....--...
tR
~
~
'--""
~
.~
o
-t-:l
cO
H
-t-:l
~ 0.01
<J.)
U
~
o
u
<J.)
rJJ 0.15 1.5 0.3
o
U
;:j
r---I
agar concentration ( wt% )
b.()
Fig. 1 Aspergillus oryzae colonies cultivated on the media with various glucose and agar concentrations. The cultivation periods for colonies on liquid-like
media (0.15wt% agar) are 10 days for 5wt% glucose, 20 days for 1wt% glu., 20 days for 0.5wt% glu., 20 days for 0.1 wt% glu., 30 days for 0.05wt%
glu., and 30 days for O.Olwt% glu., for colonies on soft media (O.3wt% agar) are 10 days for 5wt% glu., 10 days for Iwt% glu., 10 days for 0.5wt%
glu., 15 days for 0.1wt% glu., 30 days for 0.05wt% glu., and 30 days for 0.01 wt% glu., and for colonies on solid media (1.5wt% agar) are 8 days for
5wt% glu., 8 days for Iwt% glu., 8 days for 0.5wt% glu., 8 days for O.lwt% glu., 8 days for 0.05wt% glu., and 10 days for O.Olwt% glu., respectively. N
-...j
The length of the inoculation line is approximately 55 mm for soft and solid media. -...j
278 Matsuura and Miyazima
3 Results and discussion

3.1 Shapes of colonies
Photographs of A. oryzae colonies grown with various glucose and agar concen-
trations are shown in Fig.1. On the solid agar media, colonies become thick and
compact. The band shaped colony is an assembly of sub-colonies as clearly seen
in the photos of low glucose cases. The front shape of the whole colony is the
connection of the front lines of the sub-colonies. These sub-colonies have the most
advancing portion in the middle of their fronts, exhibiting smooth arch shapes.
At high glucose concentration (2': I wt%), thick colonies cover the medium
homogeneously, independent on the medium stiffness. The colony growth rate
is, however, higher, and the front is rougher for the colonies on solid medium,
indicating active growth at the front.
Stiffness of the agar medium causes remarkable morphological change of
the colony. Roughening of front and ramification of whole colonies appear on
the soft and liquid-like media (Fig. 1). On the liquid-like media, localization of
growth points occurs with decreasing glucose concentration. At very low glucose
condition of 0.01 wt%, only strong leading hyphae continue to extend, and most
secondary hyphae seem to cease growing at some physiological age. The inactive
hyphae are seen to form the chains of conidia at their apical portions.
Hypha secretes various enzymes to decompose and absorb nutrient materials.
If the nutrients are distributed uniformly in the substrates, the diffusion of nutrient
to the mycelium or the individual hypha must be one of the important factors
which determine the mycelial pattern formation. In our experimental conditions,
diffusion of glucose must be easier both in the medium of low agar and high
glucose concentration.
However, at a fixed glucose concentration, low stiffness of the substrate seems
to be unfavorable condition for mycelial growth. Selection of hyphal filaments pro-
ceeds, and the active regions are localized at the top of a few leading branches. At
present, there is no concrete explanation for this phenomenon. The attachment of
vegetative hypha to the substrate might be a significant factor for branch creation.
The box-counting fractal dimensions D box of individual branch colonies form-
ed on the liquid-like medium with 0.01 wt% glucose take the values between rv
1.6 and rv 1.8 at 20 days since inoculation. Then, Dbox'S are between rv 1.55 and
rv 1.75 at 40 days since inoculation. In many cases, D box decreases with time. This
suggests that the growth process of branch colony is not perfectly self-similar, and
then there is a high singularity of growth probability only at the apical portion.
Many fungi are known to produce peculiar root-like aggregation of hyphae,
called rhizomorph, in the soil. In this kind of structure, the strand hyphae are
thought to take part mainly in the transport of nutrients etc. Our branched colony
might be an example of the prototype of such differentiated structures, and might
be a positive activity of exploiting an unfavorable substrate.
0.4
0.3 -
~
'-.......
..c:
0.2
~ f ·f
0.1 I
0.01
I
"
0.1
I I
1
I
glucose concentration (wt% )

Fig. 2 Dependence of the mean height of colonies hlL at 8 days since inoculation on the glucose
concentrations. Points indicate average values over 10 '" 20 samples. Bars indicate the stan-
dard deviations. The colony images for 0.1 and 0.01 wt% glucose were processed by a slightly
modified method (see ref. 3).
3.2 Self-affine growth fronts formed on solid media

Let us now turn to the dependence of the roughness of growth front on the glucose
concentration for colonies formed on the solid media. Figure 2 shows the mean
height of colonies h(L), calculated over the total length of the inoculation line L.
As seen in the figure, h increases with the glucose concentration. The roughness
a(L) calculated over the total length of the inoculation line L is plotted against
the glucose concentration in Fig.3. a(L) seems to take minimum values at around
0.5wt% glucose and then increases with the glucose concentration.
By visual inspection of colony images in Fig. I ,it is clearly seen that colony
is sparse rough at low nutrient (0.01 wt% glucose) and become homogeneous with
flat front at 0.05wt% glucose. Then, as the glucose concentration is increased,
colony produces dense hyphal filaments and forms an uneven growth front again.
The roughened surface is the connection of the growth fronts of sub-colonies
described in the previous section.
Formation processes of the sub-colony fronts seem not to be identical for
various nutrient conditions. At low nutrient, competition of colonies will be con-
trolled by the limiting nutrient. Low hyphal density interface between neighboring
sub-colonies seen in the case of 0.01 wt% glucose is thought to be exhibiting an
15
10
5
t t
o 0.01 0.1 1
glucose concentration (wt% )
Fig, 3 Dependence of the roughness O"(L) of growth fronts at 8 days since inoculation on the glucose
concentrations.
inhibitory interactions between colonies. The branches formed in the interface

tend to cease growing. At high nutrient, mycelium can get sufficient resources to
produce hyphae. Then, the colony growth may be controlled mainly by the inher-
ent growth ability of the individual hypha. In this case, very strong hyphae will
branch out vigorously to the outer fresh medium space, so that the front rough-
ness is thought to exhibit a positive exploitation. At 0.5wt% glucose, colony front
becomes remarkably fiat and smooth. This is probably a balanced state between
2 limiting states mentioned above. The front hyphae have enough resources to
sustain an approximately equivalent growth rate to form a fiat front.
Figure 4 shows the self-affinity exponents a for colonies on the solid media
plotted against the glucose concentration. a is larger than or near to 0.5, except
for the fiat colonies formed on the medium with 0.5wt% glucose.
Although the biological meaning of a has not been established as yet, an
analogical conceiving with the case of the fractional Brownian motion (fBm)l)
may be possible. The fractional Brownian function VH (t) is a function of time t,
0.8
0.7
t t
0.6
0.5
t
0.4
0.3
0.2
0.1
t
0.01 0.1 1
glucose concentration ( wt% )
Fig. 4 Dependence of the self-affinity exponent a of growth fronts on the glucose concentrations.
The data for 0.1 and 0.01 wt% glucose are of colonies at 10 days, and others are those at 8
days since inoculation.
with its increments VH (t2) - VH (t I) obeying a Gaussian distribution, as,
(4)
where the brackets denote ensemble averages and H is 0 < H < 1. The case
H = 1/2 is the Brownian motion and the increments are uncorrelated. Also,
the increments are positively correlated for H > 1/2, and they are negatively
correlated for H < 1/2. If we plot the increments against the time, then a self-
affine fluctuation is obtained, and H corresponds to a.
An arch-shaped growth front of each sub-colony may corresponds to the case
of positively correlated fBm, where the increase (or the decrease) of increments
often persists relatively long time period. In this way, large a-value suggests larger
sized arches formed at the front.
At O.5wt% glucose, a is found to be lower than 1/2. The long wave length
components which produce the arch-shaped contours are reduced, and the short
wave length components which are due to the fluctuation in the local hyphal
distribution remain in the front roughness. This corresponds to the negatively cor-
related increments, where the long wave length fluctuations are relatively reduced
and increments change very frequently.
Finally, it seems worth notable that we find a fairly wide variation of a and a
for colonies under the equivalent conditions. Range of the variations of a(L) seems
to be broadened as the glucose concentration is increased (Fig.3). This suggests
that much dominant branch systems with high growth rates can be created from
time to time, and the fluctuations of front shape are amplified. Further, a varies
significantly from sample to sample, and this fact is found in other experimental
works5),6) . Variation of a is found even for the flat colonies with 0.5wt% glucose,
where a(L) 's are relatively close to each other (Fig.3). Thus, the self-affinity
relation (3) is considered to be very sensitive to the geometry of the individual
colony, in comparison with the parameters such as the colony size.
4 Concluding remarks
Morphology of the fungal colony is thought to be a direct appearance of the
flexible adaptation and, sometimes, positive strategies to exploit the environment.
The mycelial controls governing the complex hyphal growth may gain an enough
efficiency when the formation of whole colony becomes appropriate to the com-
plex environment. Even on artificially prepared homogeneous media, the mycelial
growth will bring about complex conditions to their environment through physio-
logical activities. In this sense, it seems to us intriguing to consider the mycelial
control from the colony shapes formed and the relationships between them.
References
[1] F. Family and T. Vicsek, Dynamics of Fractal Surfaces, (World Scientific,
Singapore, 1991).
[2] M. Eden, Proc. 4-th Berkeley Symp. on Math. Statistics and Probability, vol.4,
ed. F. Neyman, (University of California Press, Berkeley, 1961).
[3] S. Matsuura and S. Miyazima, Fractals, 1 (1993) 11.
[4] T. Vicsek, M. Cserzo, and V. K. Horvath, Physica A, 167 (1990) 315.
[5] J. Zhang, Y.-C. Zhang, P. Alstrom, and M. T. Levinsen, in press.
[6] J. Kertesz, V. K. Horvath and F. Weber, Fractals, 1 (1993) 67.
Estimation of the Correlation Dimension of
All-Night Sleep EEG Data with a Personal
Super Computer
Peter Achermann, Rolf Hartmann l ), Anton Gunzingerl),
Walter Guggenbuhl l ), and Alexander A. Borbely
Institute of Pharmacology
University of Zurich
I) Electronics Laboratory, Swiss Federal Institute of Technology (ETH), Ziirich, Switzerland
Abstract. The correlation dimension (CD) is a gauge of the complexity of the recorded signal. An
algorithm proposed by Takens (1985) was implemented on the personal super computer MUSIC (multi
processor system with intelligent communication) which allowed to analyze for the first time the data
of entire sleep episodes. The all-night sleep EEG of a subject recorded under three different conditions
(baseline (BL), sleep in a sitting condition after intake of placebo (PL) or triazolam (TR» served as
the data base. The CD of the EEG was modulated by the sleep cycle. Median values decreased from
nonREM (non rapid eye movement) sleep stage I to 4. For REM (rapid eye movement) sleep the
values were between those of stage I and 2. For sleep in a sitting position, the CD in REM sleep (PL
and TR) and stage 2 (PL) was increased compared with BL. In stage 4 the CD was lower for TR than
for BL and PL, and higher for PL than BL.
1 Introduction
In addition to the conventional scoring of sleep stages (Rechtschaffen and Kales
1968), methods based on quantitative EEG analysis are increasingly used (i.e.
spectral analysis e.g. Borbely et al. 1981; period-amplitude analysis, e.g. Feinberg
et al. 1978). Various mathematical models of processes underlying the regulation
of sleep and circadian rhythms have been established since the early 1980s (for
an overview see Borbely and Achermann 1992). EEG slow-wave activity (SWA;
power of the sleep EEG in the 0.75 to 4.5 Hz range) in particular has been shown
to be a useful measure for describing sleep homeostasis, since it is a function of
the duration of prior waking and sleep. SWA exhibits a global declining trend as
well as an ultradian variation reflecting the nonREM-REM sleep cycle (Borbely et
al. 1981, Achermann and Borbely 1987, Dijk et al. 1990, Achermann et al. 1993).
The study of the dynamics of non-linear systems opened new ways for char-
acterizing signal properties of biological processes. Simple and deterministic non-
linear systems can display a very regular behavior as well as a highly disordered,
turbulent (chaotic) one. Within chaos order is observed. Sensitive dependence on
initial conditions is a further property of such systems. Various procedures have
been developed to investigate the dynamics of a nonlinear system. Usually only a
few variables can be measured and little is known about the complex mechanisms
generating their temporal evolution (see e.g. Schuster 1988).
284 Achermann et al.
The correlation dimension (CD) is a gauge of the complexity of a recorded

signal. In recent years the CD was estimated for various sleep states. The results of
the early studies were summarized by Mayer-Kress and Layne (1987). The recent
(Ehlers et al. 1991 , Roschke and Aldenhoff 1991, 1992, 1993) studies provided
statistics and were based on a sample of 8 to 12 subjects. However, only one,
short, artifact-free segment of the EEG (5 to 164 s) was selected and analyzed
for each sleep stage. The CD of the EEG was reported to decrease from nonREM
sleep stage 1 to 4, while for REM sleep the values were between stage 1 and 2.
The complexity parameter has not yet been evaluated for the data of an entire
sleep episode and its changes as a function of time have not been analyzed. The
present preliminary report is the first to provide such data. Computation of the
CD is very time consuming. By using the personal super computer MUSIC it was
possible to perform the analysis almost in real time.
2 Methods
2.1 Recording and data processing
The present analysis is based on the data of one individual who participated in a
study in which 8 healthy male subjects were recorded for a baseline night (BL),
a night in a sitting position after intake of placebo (PL) or triazolam (TR 0.25
mg), and a recovery night in bed (Aeschbach et aI., in prep.). The recordings
were preceded by an adaptation night. The data were subjected to a dimension
analysis and a spectral analysis for the three conditions BL, PL, and TR. The
C3-A2 EEG derivation was amplified by a Grass 7P511 amplifier (time constant:
0.9 s). The combined action of the amplifier's 50-Hz notch filter and an analog
low-pass filter served to attenuate high frequency components (-3 dB at 27 Hz).
After AD-conversion (sampling rate 128 Hz, resolution 12 bit) and digital low-
pass filtering (4th order Butterworth with 3-dB attenuation at 25 Hz) the data were
stored on optical disk for further processing. Sleep stages were scored for 20-s
epochs according to Rechtschaffen and Kales (1968). Power spectra were on-line
calculated for consecutive 4-s epochs. Artifacts were eliminated on a 4-s basis.
SWA was averaged for 1 min using a moving window in 20-s steps (BL) or 40-s
steps (PL and TR).
2.2 Algorithm for the estimation of the correlation dimension

Usually, the Grassberger-Procaccia (GP) algorithm (Grassberger and Procaccia
1983) or modifications thereof had been used in previous studies to estimate the
CD. Since the GP algorithm has certain shortcomings (Gershenfeld 1992, Theiler
1990), we applied a more elaborated algorithm that has been proposed by Takens
(1985), and further examined by Cawley and Licht (1987). This algorithm requires
a much larger number of computations than the GP algorithm, and has therefore
been little used. We implemented the algorithm on the personal super computer
MUSIC (multi processor system with intelligent communication) (Gunzinger et
al. 1992a,b) designed and built by the Electronics Laboratory of the ETH Zurich.
Estimation of the Correlation Dimension of All-Night Sleep EEG 285
Up to 63 processing elements operate in parallel on infonnation partitioned in the

data space, resulting in a peak perfonnance of 3.7 . 109 floating point operations
per second 3.7 GFlops. The MUSIC achieves super computer perfOJmance in a
desktop unit with low cost and low power dissipation.
The state vector was reconstructed on the basis of time-delay embedding
(Takens 1981). The estimator is based on Fischer's maximum likelihood rule using
the interpoint distances in the calculation of the CD d2 (Takens 1985). The method
uses all distances less than an upper boundary EO:
d2 = -1/ < 10g(E/Eo) > .

The upper boundary EO was calculated as
EO = Emin + a( < E > -Emin).

Angle brackets indicate mean values. For calculating distances the L I-nonn (Man-
hattan-nonn) was used. Since a fit of a straight line in a log-log plot was not
required, a user interaction was not needed. The CD was estimated for embedding
dimensions of 25, 30, and 35. The other parameters used in the algorithm were:
Time delay (7) 39.06 ms (5 sample points); number of data points 7680 (per
analysis sequence of 1 min); a = 0.55. The algorithm was applied to I-min
epochs of sleep EEG data. A I-min window was moved through the data of an
entire night by shifting the window in 20-s (BL) or 40-s (PL, TR) steps. For the
further analysis only the dimension values with an embedding dimension 30 were
used. Furthennore, only I-min epochs were included which (1) contained three
identical 20-s sleep scores; (2) were without artifacts; and (3) showed a rise rate
of the CD lower than 0.1 per embedding dimension for the embedding range of
25-35. The chosen time delay (7 ~ 40ms), the number of data points (7680), and
the sampling rate (128 Hz) are in accordance with the literature (e.g. Mayer-Kress
and Layne 1987, Ehlers et al. 1991, R6schke and Aldenhoff 1991). Since a robust
algorithm for the estimation of 7 does not exist, and to avoid additional variability,
we have kept 7 constant throughout the analysis. Two recent papers (Gershenfeld
1992, Theiler 1990) give an excellent overview of the theoretical background and
the problems encountered in the numerical estimation of the CD.
The calculation speed depended on the number of processors used; the MU-
SIC can be extended in steps of three processors (i.e. one board). For a I-min
epoch the algorithm as specified (see above) needed 374 s of computer time with
3 processors, 58 s with 21 processors, and 32 s with 45 processors. To perfonn
the analysis in real time, 21 processors (7 boards) are sufficient if the analyzed
I-min epochs do not overlap.
2.3 Statistics
Non-parametric statistics (notched box plot; Velleman and Hoaglin, 1981) were
used to establish significant differences (p < 0.05) between stages and between
conditions.
BASEUNE
3000
2000
1000
w
"R
••
I
•
12
,0
~ I I . . • I
'''-,
...,........
<Il 8 "
:.t.•..'},-:\.
: •
~:'"
..,.... ~ •• ~ 1\ )'f , .. ,. '''' '4'\' ."'"
• ~ " ., ".
. ...... .
po -.... , . , ' J"
ffi ,$.. •• '
:::.
a
6
,
~
- • r"
0--'.~.:''.
r , • f
",~ #
.... ~
... ,'"
... ",:o,-!
,
•
'.
"
a:
a:
8 2
0
0 180 360 540 no 900 ,080 1200
TIME 120-1 EPOCHS)
PLACEBO
3000
,2
,0
~
..1.A, ,.,, \:" ....,
.
<Il
8 "
). .,':r..... ;......
ffi:::. ':~'t : ;~
is
6
', ·'''''!.o '" "
a: ~
a:
8 2
0
0 ,80 360 540 720 900 1080 1260 1440
TIME 120-. EPOCHSI
Fig. 1 Slow-wave activity (SWA; power in the 0.75 to 4.5 Hz range), vigilance states (W: Waking,
M: movement time, R: REM sleep, 1-4: nonREM sleep stages) and correlation dimension
for a subject under baseline (BL) condition (sleep in bed) and placebo (PL) condition (sleep
in a sitting position). Sleep stages are plotted on a 20·s basis (overlaps occur due to plotting
resolution). SWA and dimension values were calculated for I-min epochs by shifting a I-min
window by 20 s (BL) or 40 s (PL).
3 Results
The algorithm was first tested with a calibration signal and a theoretical test signal.
For the calibration signal (lO-Hz sine wave) the mean dimension estimate was
1.03 (sd 0.00086, n = 8). The analysis of the test signal (seven sine waves with
incommensurate frequencies) resulted in a CD of 6.77 (sd 0.017, n = 8) with a
rise rate of the dimension of 0.011 per embedding dimension (sd 0.002, n = 8).
For both signals the estimated dimension was close to the theoretically expected
value of 1 and 7, respectively.
Figure 1 illustrates the time course of SWA, of the vigilance states, and of
the CD for the baseline condition and for sleep in a sitting position after placebo
intake. For the drug condition (TR) the data (not shown) were very similar to those
of the PL-condition. In the PL-condition the number of plotted dimension values
is low because moving the I-min window by 40 s and not by 20 s increased the
relative number of state transitions and reduced the relative number of artifact free
I-min epochs with identical sleep scores. Moreover, due to the sitting position,
more state changes and artifacts occurred in PL than in BL. In all conditions the
CD was clearly modulated by the sleep cycles. Low values with low variability
were observed during consolidated episodes of slow wave sleep. Sleep in a sitting
position (in both the PL- and the TR-condition) was impaired in comparison to the
BL-condition (sleep in bed). Sleep efficiency was decreased, waking after sleep
onset was increased, and SWA was decreased (Aeschbach et aI., in prep.).
Median and quartile values of the CD and SWA are summarized in Table
1. In all conditions, SWA increased with the deepening of sleep whereas the CD
decreased. The CD was significantly different between the sleep stages, except be-
tween stage 2 and REM sleep. SWA was significantly different between all sleep
stages. Computed for all 8 subjects, SWA in the TR-condition was significantly
reduced in comparison to the PL-condition and both conditions differed signifi-
cantly from BL (Aeschbach et al., in prep.). This is also reflected in the median
values of SWA of the single subject analyzed in this study (Table 1). On the other
hand, the CD in stage 4 was lower for TR than for PL and BL, and higher for PL
than BL. The PL and TR values in REM sleep and the PL value in stage 2, were
higher than the corresponding values in BL.
4 Discussion
The present study demonstrates the feasibility of computing the CD for an entire
sleep episode with a personal super computer. The analysis showed that the sleep
cycles are reflected in changes of the CD, particularly when SWA peaks were
present. The CD showed a decline from stage I to stage 4 with REM sleep values
being intermediate between stage 1 and 2. These results correspond to those of
previous authors who had based their analysis on short EEG segments (e.g. Mayer-
Kress and Layne 1987, Ehlers et al. 1991, Roschke and Aldenhoff 1991). However,
it is also clear that a considerable range of values occur within a single sleep stage
Stage Parameter Corr. Dimension SWA

BL PL TR Bl PL TR
Ql 8.34 8.48 8.70 108.5 70.4 68.4

1 MED 8.85 9.00 8.75 122.0 81.9* 89.7
Q3 9.24 9.39 8.86 145.1 99.8 132.5
Q1 6.39 7.49 6.40 310.2 220.3 215.2
2 MED 7.48 8.06* 7.71 451.1 268.6* 324.4*
Q3 8.10 8.25 8.27 640.7 356.9 480.1
Q1 5.57 6.44 5.92 1085.8 694.3 732.4
3 MED 6.25 6.84 6.37 1244.5 762.2* 851.2*
Q3 6.83 7.35 6.82 1478.7 884.4 1149.1
Ql 5.36 5.63 5.03 1880.5 1352.3 1269.8
4 MED 5.63 5.81* 5.36*$ 2070.6 1470.5* 1389.9*
Q3 5.86 6.02 5.68 2250.4 1592.3 1580.8
Q1 7.17 7.70 7.81 149.7 111.7 123.1
REMS MED 7.70 8.15* 8.17* 176.0 128.0* 139.5*
Q3 7.99 8.38 8.38 222.9 172.7 167.0
Table 1 Correlation dimension and slow-wave activity (SWA; QI: first quartile, MED: median, Q3:
third quartile) for the different sleep stages of one subject. BL: baseline, PL: placebo, and TR:
triazolam. SWA is in JL V 2 . * indicates a significant difference with respect to BL; $ indicates
a significant difference between TR and PL (p < 0.05; see text for method). The number of
data points per stage ranged from 7 to 273.
(e.g. see stage 2 baseline record in Fig. 1). The selection of single short segments
for analysis is therefore unlikely to provide a reliable estimat<t of the CD. This
conclusion is supported by the differences between the values for sleep in bed
(BL) and sleep in a sitting position (PL, TR). The more restless sleep in the
sitting position was evident from the frequent sleep disruptions and the reduced
consolidation of stage 2 sleep (Fig. 1). The increased CD in stage 2 in PL as well
as the reduction of SWA (Tab. 1) indicate that a more superficial type of stage
2 prevailed under this condition. The increase in the REM sleep value in both
PL and TR is a comparable finding. Of course, the data from a single subject
cannot be generalized and analysis of more nights are required to substantiate
these observations. It will be also interesting to see whether the CD in REM sleep
is more sensitive to pharmacology and pathology than the CD in other sleep stages
as some studies seem to indicate (Roschke and Aldenhoff 1992, 1993).
The present study demonstrates that the use of a personal super computer
opens new possibilities for performing complex EEG analyses in real time.
Acknowledgement:
We thank Daniel Aeschbach and Christian Cajochen for providing the data (record-
ing and scoring), Markus HUrzeler (Institute of Mathematics, ETH ZUrich) for cal-
culating the boxplots, Bernhard Biiumle, Peter Kohler, Urs MUller, Walter Scott,
Hans-Ruedi vonder MUhll for their support (MUSIC group), and Dr. Irene Tobler
for comments on the manuscript. The study was supported by the Swiss National
Science Foundation grant nr. 31.32574.91.
References
[1] Achermann, P. and Borbely A.A. Dynamics of EEG slow wave activity dur-
ing physiological sleep and after administration of benzodiazepine hypnotics.
Hum. Neurobiol., 1987,6: 203-210.
[2] Achermann, P., Dijk, DJ., Brunner, D.P. and Borbely A.A. A model of sleep
homeostasis based on EEG slow-wave activity: quantitative comparisons of
data and simulations. Brain Res. Bull., 1993, 31 : 97-113.
[3] Borbely, A.A. and Achermann, P. Concepts and models of sleep regulation:
an overview. 1. Sleep Res., 1992, 1: 63-79.
[4] Borbely, A.A., Baumann, F., Brandeis, D., Strauch, I. and Lehmann, D. Sleep
deprivation: effect on sleep stages and EEG power density in man. Electroen-
cephalogr. Clin . Neurophysiol. , 1981,51: 483-493.
[5] Cawley, R. and Licht, A.L. Maximum likelihood method for evaluating cor-
relation dimension. Lect. Notes in Phys., 1987, 278: 90-103.
[6] Dijk, DJ., Brunner, D.P. and Borbely, A.A. Time course of EEG power
density during long sleep in humans. Am. 1. Physiol., 1990,258: R650-R661.
[7] Ehlers, C.L., Havstad, J.W., Garfinkel, A. and Kupfer, DJ. Nonlinear analysis
of EEG sleep states. Neuropsychopharmacology, 1991,5: 167-176.
[8] Feinberg, I., March, J.D., Fein, G., Floyd, T.c., Walker, J.M. and Price, L.
Period and amplitude analysis of 0.5-3 Hz activity in NREM sleep of young
adults. Electroencephalogr. Clin. Neurophysiol., 1978,44: 202-213.
[9] Gershenfeld, N.A. Dimension measurement on high-dimensional systems.
Physica D, 1992,55: 135-154.
[10] Grassberger, P. and Procaccia, I. Measuring the strangeness of strange attrac-
tors. Physica D, 1983,9: 189-208.
[11] Gunzinger, A., MUller, U., Scott, w., Biiumle, B., Kohler, P. and Guggen-
bUhI, W. Architecture and realization of a multi signalprocessor system. In:
J. Fortes, E. Lee and T. Meng (Eds) Proc. Int. Con! on application specific
array processors. IEEE Computer Society Press, Los Alamitos, CA, 1992a:
327-340.
[12] Gunzinger, A., Milller, U.A., Scott, w., Bliumle, B., Kohler, P., Yonder Milhll,
H.R. Milller-Plathe, E, Van Gunsteren, W.E and Guggenbilhl, W. Achiev-
ing super computer performance with a DSP array processor. In: R. Werner
(Ed) Supercomputing '92. IEEE Computer Society Press, Los Alamitos, CA,
1992b: 543-550.
[13] Mayer-Kress, G. and Layne, S.P. Dimensionality of the human electroen-
cephalogram. In: A.S. Mandell and S. Koslow (Eds) Perspectives in biolo-
gical dynamics and theoretical medicine. Annals of the New York Academy
of Sciences, New York, 1987,504: 62-87.
[14] Rechtschaffen, A. and Kales, A. A manual of standardized terminology, tech-
niques and scoring system for sleep stages of human subjects. Bethesda, MD,
U.S. Department of Health, Education, and Welfare, 1968.
[15] Roschke, J. and Aldenhoff, J. The dimensionality of human's electroen-
cephalogram during sleep. Bioi. Cybern., 1991,64: 307-313.
[16] Roschke, J. and Aldenhoff, J.B. A nonlinear approach to brain function: de-
terministic chaos and sleep EEG. Sleep, 1992, 15: 95-101.
[17] Roschke, J. and Aldenhoff, lB. Estimation of the dimensionality of sleep
EEG data in schizophrenics. Eur. Arch. Psychiatry Clin. Neurosci., 1993,
242: 191-196.
[18] Schuster, H.G. Deterministic chaos. An introduction. VCH Verlagsgesell-
schaft mbH, Weinheim, 1988.
[19] Takens, E Detecting strange attractors in turbulence. Lect. Notes in Math.,
1981,898: 366-381.
[20] Takens, E On the numerical determination of the dimension of an attractor.
Lect. Notes in Math., 1985, 1125: 99-106.
[21] Theiler, J. Estimating fractal dimension. 1. Opt. Soc. Am. A, 1990, 7: 1055-
1073.
[22] Velleman P.E and Hoaglin D.C. Applications, basics, and computing of ex-
ploratory data analysis. Duxbury Press, Boston MA, 1981.
Fractals
in Pathology
Changes in Fractal Dimension of Trabecular
Bone in Osteoporosis: A Preliminary Study
C. L. Benhamou I) , R. Harba2), E. Lespessailles I) ,

G. Jacquet2) , D. Tourliere l ) and R. Jennane 2 )
I) Pole d'Activite Rhumatologie, CHR Orleans La Source, France.

2) Laboratoire d'Electronique Signaux Images, Universite d'Orleans, France.
1 Introduction
The trabecular bone network represents a complex structure, either in a macro-
scopic examination, in a microscopic histological study, or in a radiological projec-
tion [1, 2]. Particularly for radiological images, euclidean geometry does not lend
itself to the description of such a complex structure. Fractal geometry is much
better suited to this analysis, allowing a characterization of this type of rough
irregular texture, or of its projection [3,4].
The choice of the fractal model is dependant on the characteristics of the
images. Some models necessitate a binarisation of the image [5]. This binarisa-
tion may be simple for some textures. On radiological images of bones, which
represent a 2 D projection of a 3 D structure, the binarisation implies a definition
of a threshold between trabeculae and intertrabecular spaces. Unfortunately, this
definition of a threshold remains arbitrary, and the threshold may vary in different
parts of the image due to the variations of bone thickness. So it seems better for
our study to use a fractal model taking into account the grey level variations. The
Fractional Brownian Motion (FBM), and its increment the Fractional Gaussian
Noise (FGN) respond to this condition [6]. It has been used to undergo an anal-
ysis of radiological bone images [3,4]. FBM, noted BH, is characterized by the
H parameter (0 < H < 1). H is related to the fractal dimension D: for a one
dimensional signal D = 2 - H. The smaller is H, the bigger is the roughness of
the texture.
The trabecular bone microarchitecture organization is nowadays considered
as a very important factor in biomechanical competence of bone [7,8,9] accounting
in a great part for the properties of resistance and elasticity of this tissue. But bone
is not only an amorphous framework, its trabecular microarchitecture plays a great
role in the calcium homeostasis function: this structure offers a very large surface
for ions exchanges between trabecular bone itself and the highly vascular bone
marrow which fills the intertrabecular spaces.
Bone ageing and osteoporosis may alter cortical or trabecular bone. In trabec-
ular bone, the main changes consist in a decrease of bone mass [10]. In addition
Changes in Fractal Dimension of Trabecular Bone in Osteoporosis 293
to this loss of bone tissue, qualitative changes are considered as a very important
component of the biomechanical alterations of bone [11]. They are mainly linked
to abnormalities outcorning in the trabecular network, for instance perforations,
dysconnections, loss of trabeculae [12,13]. These abnormalities would be much
more deleterious in terms of biomechanical competence than a simple thinning of
trabeculae [14]. Their analysis requires at the present time an histological study,
implying an invasive procedure of bone biopsy [15].
We report the results of a new method of trabecular microarchitecture fractal
analysis on radiological images of calcaneus, in a study comparing osteoporosis
cases and control cases.
2 Material and Methods

2.1 Patients
We have studied 31 X ray-views from 17 osteoporotic patients and 24 X ray-views
from 12 control subjects.
Osteoporosis was defined by fractures occurring after a fall from a standing
position or spontaneus fractures: mainly hip fracture or vertebral crush fracture
except in two cases (rib and sacral recurrent fractures in one, ankle and clavicle
for the other one). The 17 osteoporotic patients, 3 men and 14 women, were all
Caucasian (age 74.4± 14.3 years). They had no history of metabolic bone diseases
except osteoporosis.
In the control group, 5 men and 7 women, all subjects were Caucasian (age
77 .3± 12.1 years), they had no history of fracture, of diseases interfering with bone
metabolism and they were not taking medication known to affect bone metabolism
(e.g., fluoride, calcitonin, estrogen, bisphosphonates, thiazide diuretics, corticos-
teroids).
2.2 Acquisition of images

Os calc is X-ray images have been performed with the same standardized condi-
tions. Focal-calcaneus distance was fixed at one meter. Gridless cassettes, Plastilix
with a thickness of 1 cm were used. Calcaneus was placed in contact with the
film. The same X-ray tube, voltage (48 KV), and exposure conditions (18 mA.s
for 0.08s) were used. Radiographs were taken on a Kodak min R film (18 x 24
cm) usually devoted to mammography. Development was made at 34° Celsius in
a Kodak developing bath. Tested in the Kodak laboratory conditions, this film has
given a resolution of fifteen pairs of lines a millimeter.
Radiographs were put on a lighting table and images were acquired with a
CCD camera. The lens aperture was always at its optimum: 8. The numerisation
process is performed with a matrox frame grabber at a format of 512 x 512 pixels of
8 bits. The gain and offset of the frame grabber were set in order to have a correct
histogram with the greatest dynamic without saturation. The acquisition field was
kept constant which led to a pixel size of 1051lm. A specific acquisition software
294 Benhamou et al.
Fig. 1 A typical radiograph of calcaneus. The drawn square delimitates the region of interest.
ran on an IBM PC-AT computer which allowed to select an area of 256 x 256
pixels corresponding to the central part of the trabecular bone. Figure 1 shows
a typical radiograph of a calcaneum and the selected region of analysis. Then,
data files were sent through an ethemet network to IBM RS/6000 workstations for
numerical computations.
2.3 Estimation of the fractal dimension

Because of the anisotropy of the image, a one dimensional analysis is performed.
400 lines of 100 samples each are taken from images in a given direction as seen
in figure 2. The H value is computed on each line and the final value for a given
direction is the average on these 400 lines. 36 different angles are computed. A
polar diagram represents the H value as a function of the direction as seen on
figure 3. The anisotropy of the images can be seen with this kind ofrepresentation.
For this study, we only use the global value for the whole image which is the aver-
age on the 36 H values for the different angles. The H value has to be estimated
for each line. Among the numerous methods, we choose the Maximum Likeli-
hood Estimation method because it gives the smallest theoretical variance for the
estimation of H. This method consists in maximizing the conditional probability
density function p(x; H). x is a FGN vector, increment of the data set modelized
by the FBM BH:
(1)
Analysis angle = 0 degres Analysis angle = 45 degres
Fig. 2 Lines selection for the oriented analysis.
x is stationary, zero mean and has Gaussian distributed samples as:
p(x;H) =
1 (T ~l )
27r)N/2IRII /2 exp -x Rx x/2 . (2)
N is the length of the data set, Rx is the covariance matrix which elements depends
on the autocorrelation function of the FGN such as [Rxlij = Tx[i - jl with:
(3)
(J"2 is the variance of the FGN vector.

No analytical solution is available to find the maximum of p(x; H). The
shape of this function with H from 0 to 1 is unimodal wich allows to implement
numerical methods. Since the logarithm is a monotonic function, the log of p(x; H)
is maximized:
N I l T ~l
10gp(x;H) = -2"log27r - "2 log IRI- "2x R x. (4)
Two parameters have to be estimated: Hand (J"2. The R matrix can be decomposed
as:
(5)
Using this in the log likelihood function and letting the derivative go to zero
with respect to (J"2 gives the final function to be maximized:
296 Benhamou et al.
90
120
, ).,. --- -'.....
.... - .I .. .. ..
....., ,
60
,,,"
~ ~
.- .......... "
~' " ' ,
150 , , , .. .. ' ~ , 30
~"..
· - .... _.. ..
. .. .."
'\.
"
' ,I, .'.,..- ",'
#,~......
I
I ..- ...... •
",
,~, "
, ., ~.. til
, ....... ••• • .. ".. '. to, ' ,
"" . : , •.• '. • , t . . '
iB nv-,I •••••• :
• : " " :. l······.: '.• .. • ·• •~ o
•
...............
• I 'I
......
•
,:'..
• '" ••••••••• .: I' : :
...'
,.,
. .
," ... ...."
. .." ... . ,
.. "..
\
210"" , , " " ,

"........ .. . .. .. . .. .". , , ,
'" .."
.. .. .. ".. .
- , , Co
, , , 330
.. I
, , ,
. .., ,~
" .... , # ,
240" ...... - - _ ........ < 300

- 2fo - ..
H (THETA)
Fig. 3 Polar representation of the H value for an image.
R' has to be inverted and its detenninant computed. R' is definite positive and can
be decomposed using the Levinson algorithm as
R,-I = LD- 1LT (7)
where L is a lower triangular matrix with ones on the diagonal and D is diagonal.
Then, the detenninant of R' is equal to the detenninant of D. Both detenninant
and inverse are computed in one step. The golden search technique has been
implemented to find the maximum of the log likelihood function. The total .time
to compute the global H value for an image is about 40 minutes on an IBM
RS/6000 workstation.
3 Results
Mean results of H mean value were 0.401 ± 0.065 in the osteoporotic group
and 0.500 ± 0.052 in the control group. Student's paired t test showed a statistical
significant difference p < 0.0001 between these two populations as shown in table
1. There was no statistical difference between these two groups for the mean age
(p = 0.42). We have found no statistical significant correlation between age and
H mean. Figure 4 represents the histogram of the two populations and allows a
visual appreciation of the overlap.
H age
Osteoporotic cases 0.401 ± 0.065 74.4 ± 14.3

Control cases 0.500 ± 0.052 n.3±12.1
Statistical significance p < 0.0001 P = 0.42
Table 1 Mean result for the H parameter for the two groups
REPARTITION OF TWO POPULATIONS

0.6
P < 0 . 0001
00
o a
----..
o a
c: a
a a
lJJ a a
t-
lJJ
0.5 a a
::E a a
a a
« u6
c: aa
--
« a
0
00
00
a
a.. a
a
a
a
::c 0.4 a a
a a
1:1
a a
a
a a
a a
0.3
0 2 3 4 5
CONTROLS OSTEOPOROSIS
CONTROLS (MEAN AGE 77.3) HMEAN VALUE 0 . 5005
STANDARD DEVIATION 0.0509
OSTEOPOROSIS (MEAN AGE 74.5) HMEAN VALUE 0.4006
STANDARD DEVIATION 0.0640
Fig. 4 The histogram of H mean in osteoporosis cases and control cases.
4 Discussion
These data indicate that our fractal evaluation allows to discriminate an osteo-
porotic elderly population from an age-matched control group. We have to under-
line the high statistical significant difference between these two populations, in
spite of a limited number of subjects. This fractal model has been previously used
in preliminary experiences. In one immobilizated patient, the H value has shown
the same trend to a progressive lowering [3]. Acid dissolution of bone samples
also results in a reduction of H Value [4]. Other models have been tested by
other groups in such preliminary studies: Van der Stelt and Geraets also report a
decrease of H value following menopausis [5]. Ageing is generally considered to
result in a loss of complexity, implying a lowering of the fractal dimension [16],
particularly in dynamic processes. Our decrease of the H value corresponds to an
298 Benhamou et al.
increase of the fractal dimension, probably accounting for the disorganization of

the trabecular network.
One of the advantages of our method consists in the ability to describe an
image with a various range of grey. This method allows to analyze a lot of infor-
mations, coming from a 2 D projection of a 3 D structure.
The study of calcaneus radiological images offers the advantages of a rapid,
simple and safe non invasive procedure. It is not expensive. The irradiation is
limited, allowing epidemiological studies, and successive evaluations in the same
subject. Spine, hip and wrist are the preferential sites of osteoporotic fractures [17].
However, spine and hip are surrounded by large soft tissues, so that the results
may differ widely with morphological variations. We have tested some evaluations
on the radius, but this site offers a limited area of analysis. Our calcaneus study
does not involve these difficulties. Furthermore, in terms of bone mineral density,
it has been shown to be the best site for the determination of the relative risk of
osteoporotic fractures [18].
5 Conclusion
In our experience, this fractal model based on the FBM, has been very interesting
for trabecular bone microarchitecture characterization on X-Ray images. Before
to start large epidemiological studies, we have searched to assess its discriminant
power between controls and osteoporosis cases. Our preliminary study has shown
that this model could valuably separate these two populations, with a limited
degree of overlap. We have not found a statistical correlation whith age, but this
was not the aim of the study, so that both groups were mainly constituted of elderly
subjects, with limited variations in age.
Bone mineral density measurements are not able to fully explain the mech-
anisms of osteoporosis. We think that our model is able to account for structural
qualitative changes.
Acknowledgements:
We thank Rhone-Poulenc Roser IntI . for its support, Kodak-Pathe Division Sante
for its technical advice, and Mrs. Nadege Marchant for her secretarial assistance.
References
[1] Rockoff S.D., Scandrett J. and R. Zacher: Quantitation of relevant image
information: automated radiographic bone trabecular characterization. Radi-
ology, 101,435-439 (1971).
[2] Aggarwal N.D., Singh G.D., Aggarwal R. et al.: A survey of osteoporosis
using the calcaneum as an index. Int. Orthop., 10, 147-153 (1986).
[3] Lundahl T., Ohley W.J., Kay S.M., Siffert R.: Fractional Brownian motion:
a maximum likelihood estimation and its application to image texture. IEEE
Trans. med. imaging, MI-5, 152-160 (1986).
[4] Jacquet G., Ohley W.1., Mont M.A., Siffert R, Schmukler R: Measure-
ment of bone structure by use of fractal dimension. Proc. Annual Conf. of
IEEEIEMBS, 1402-1403 (1990).
[5] Van der Stelt P.E, Geraets: The fractal dimension of the trabecular pattern in
patients with increased risk of alveolar ridge resorption. Proc. Annual Conf.
of IEEEIEMBS, 12, 2071-2072 (1990).
[6] Mandelbrot B.B., van Ness J.: Fractional brownian motion, fractional noise
and applications. SIAM 10, 422--438 (1968).
[7] Kleerekoper M., Villanueva A.R, Stanciu 1., Sudhaker Rao D., Parfitt A.M.:
The role of three dimensional trabecular microstructure in the pathogenesis
of vertebral compression fractures. Calcif. Tissue Int., 37, 594-597 (1985).
[8] Parfitt A.M.: Trabecular bone architecture in the pathogenesis and prevention
of fracture. Am. 1. Med., 82 (suppl IB), 68-72 (1987).
[9] Mosekilde Li.: Age-related changes in vertebral trabecular bone architecture.
Assessed by a new method. Bone, 9, 247-250 (1988).
[to] Dempster D.W., Lindsay R : Pathogenesis of osteoporosis. LANCET, 341,
797-801 (1993).
[11] Parfitt A.M.: Implications of architecture for the pathogenesis and prevention
of vertebral fracture. Bone, 13, S41--47 (1992).
[12] Vernon-Roberts B., Pirie C.J.: Healing trabecular micro-fractures in the bodies
of lumbar vertebrae. Ann. Rheum.Dis. 32, 406 (1973).
[13] Parfitt A.M.: Age related structural changes in trabecular and cortical bone:
cellular mechanisms and biomechanical consequences. Calcif. Tissue Int., 36,
SI23-S128 (1984).
[14] Weinstein A.S., Hutson M.S.: Decreased trabecular width and increased tra-
becular spacing contribute to bone loss with aging. Bone, 8, 137-142 (1987).
[15] Birkenhager-Freukel D.H., Courpron P., Hupscher E.A., Clermonts E.,
Coutinho M.E, Schmitz P.I.P.M. and Meunier P.J.: Age related changes in
cancellous bone structure. Bone and Mineral 4, 197-216 (1988).
[16] Lipsitz L.A., Goldberger A.L.: Loss of complexity and aging. JAMA, 267,
1806-1809 (1992).
[17] Nordin B.E.C.: The definition and diagnosis of osteoporosis. Calcif. Tissue
Int. 40, 57-58 (1987).
[18] Black D.M., Cummings S.R, Genant A.K. et al. : Axial and appendicular bone
density predict fractures in older women. J. Bone Miner. Res., 7, 633-638
(1992).
Use of the Fractal Dimension to Characterize
the Structure of Cancellous Bone in
Radiographs of the Proximal Femur
Curtis B. Caldwell I , John Rosson2 ,
James Surowiak2 and Trevor Heam 2
I) Departments of Medical Biophysics and Radiology, University of Toronto, and Department of

Radiological Sciences, Sunnybrook Health Science Centre, 2075 Bayview Avenue, North York,
Ontario, Canada M4N 3M5
2) Orthopaedic Biomechanics Research Laboratory, Sunny brook Health Science Centre, 2075 Bay-
view Avenue, North York, Ontario, Canada M4N 3M5
Abstract. Preliminary work on the development of a quantitative method of characterizing the ra-
diographic appearance of cancellous bone is described. The method is based upon calculation of the
«fractal dimensiOn» in regions of interest of digitized radiographs. The method was applied to standard
pre-operative radiographs of the hip and proximal femur of fifteen patients who were to undergo total
hip arthroplasty. Cancellous bone samples were retained at arthroplasty and their compressive strength
and ash density assessed. The fractal dimension was found to correlate more strongly with compres-
sive strength than did a conventional method of assessing bone quality (the Singh Index). Among the
potential benefits of the new technique is the possibility of providing guidance in the pre-operative
selection of the optimal method of implant fixation for an individual patient.
1 Introduction
Fracture of the hip is one of the major clinical risks associated with osteoarthritis.
If a hip replacement becomes necessary, the possibility of implant failure poses
an additional risk to the patient. Although these risks are well known, reliable,
non-invasive diagnostic methods which accurately discriminate among patients at
risk for hip fracture or implant failure are not currently available. The majority
of current methods attempt to measure gross bone density, while ignoring the
influence of bone structure. While bone density is clearly an important parameter,
assessing risk by density alone neglects any possible influence of the geometrical
distribution or structure of cortical and cancellous bone. Some would contend that
bone density alone is sufficient to predict fracture risk (Hui et ai. 1988). This is
equivalent to assuming that the only change in diseased or damaged bone is loss of
mass, with no change in structure. In addition, there appears to be an age-related
bone loss of about 1 % a year in normal postmenopausal women. This implies that
if bone mass measurements are to be meaningful for individual patient follow-up ,
they must have very good precision. In a study of four methods of bone mass
measurement in 49 osteoporotic women obtained at time zero and at one year of
treatment, it was found that none of the methods could be used to predict changes
as measured by any of the other techniques (Ott et aI., 1986). This result occurred
To Characterize the Structure of Cancellous . .. 301
despite changes in some patients which were as high as 5 % by one or other of

the techniques.
The most common current diagnostic techniques for assessing the bone qual-
ity of the proximal femur include quantitative bone densitometry (e.g., using the
dual energy x-ray technique - DEXA or the dual photon radioisotope technique
- DPA) and qualitative evaluation of plane film radiographs. Quantitative x-ray
computed tomography (Lotz et al. 1990) has been suggested as an alternative
method of assessing bone quality, but is a relatively expensive technique which
is often not available. Singh et al. (1970) developed a method of subjectively
evaluating changes in the trabecular pattern of the femur (as seen in plane film
radiographs of the hip), as an aid to distinguishing the degree of osteoporosis.
While Singh's work addresses the problem of assessing trabecular bone struc-
ture, the Singh Index is, unfortunately, subject to high inter- and intra-observer
variability and has not proved to correlate well with fracture risk.
We have been working on the development of a method of quantitatively
evaluating bone structure from digitized radiographs of the hip. The potential ben-
efit of such a method is that it may provide structural information complementary
to the bone density information already available. Since we propose to derive the
information from conventional radiographs of the hip, the patient is exposed to no
more invasive procedure or to any further radiation dose. Together, information
on structure and density should provide a better means of risk stratification, when
considering treatment for degenerative bone disease. The combination of structural
and density information may provide a more precise method of assessing longi-
tudinal changes in patients undergoing drug or dietary treatment for osteoporosis.
A method of evaluating the structure of cancellous bone may be particularly im-
portant when deciding upon the best method of fixing an implant in place in an
area such as the femoral head.
Fractals are a powerful tool for modelling biological objects. While no uni-
versal definition of a «fractal» has been agreed upon, fractal shapes and patterns
are often identified by their properties of deterministic or statistical self-similarity
and independence to scaling. The use of the fractal concept is based on the work
of Mandelbrot (1983), who showed that many physical systems in nature are well
described by sets of numbers which show elements of self-similarity and random-
ness over some range of scale. Such a set may be considered to have a particular
«fractal dimension» characteristic of the physical system or biological structure
in question. Fractals are especially useful in providing a means of quantitatively
characterizing complex natural structures which are not well described by classical
geometry (i.e., the Euclidean geometry of straight lines, circles, spheres, etc which
describes most man-made objects very well).
Image analysis using fractal mathematics has been investigated by a num-
ber of researchers. It has been shown that image textures which are subjectively
rough or highly disorganized have higher fractal dimensions than images that are
subjectively more smooth or more regularly organized (Pentland 1984). Fractals
302 Caldwell et al.
have been used in medical imaging to discriminate between nonnal and abnonnal
Nuclear Medicine liver scans (Cargill et al. 1989), mammograms (Caldwell et al.
1990), and images of retinal vessels (Mainster MA 1990).
While it is known that fractals are useful in describing many complex nat-
ural structures, why would one propose that the structure of cancellous bone in
particular would be «Fractal»? We were lead to this hypothesis through the work
of Kaandorf (1991) who applied fractal methods to modelling the structure of sea
sponges. Cancellous or «spongy» bone has a subjectively similar structure to that
of sea sponges. Moreover, other researchers have found the fractal dimension to
be a useful descriptor of radiographic bone quality in the knee (Lynch et al. 1991)
and in the peridental alveolar bone (Ruttimann et al. 1992). Ruttimann contends
that a type of fractal scaling may be seen in bone as the texture of bone tends
to change from coarse to fine as one moves across the boundary from cortex to
spongiosa.
2 Methods
Fifteen patients undergoing total hip arthroplasty were studied. The underlying
pathology was osteoarthritis in all patients. Standard pre-operative radiographs of
the hip and proximal femur were taken with the proximal femur rotated fifteen
degrees to the plane of the x-ray beam. These radiographs were digitized using
a Konica laser scanning microdensitometer with a pixel size of 0.175 mm and
a ten-bit grey-scale resolution. Digital images were subsequently analyzed using
computer programs written in «C» on a Sun workstation (Sun Microsystems,
California, USA).
Each of the patients was assigned a Singh Index value (one to six in ac-
cordance with the original method) by an experienced radiologist who had no
knowledge of the patients' age or sex. Note that the radiologist used the original
radiographs (i.e., not the digitized images of these radiographs).
At surgery, the femoral neck resection specimens were retained and stored
at - 20 degrees Celsius. Cubes of cancellous bone measuring one centimetre on
each side were cut from the specimens while still frozen using a water cooled
saw equipped with a diamond tipped blade and a micrometer. The anatomical
orientation of the cubes was recorded and compression was applied in a consistent
direction in all samples during mechanical testing.
Prior to testing the specimens were warmed to 37 degrees Celsius and the
dimensions of each cube were measured to within 0.1 mm. The specimens were
then compressed in a servo-hydraulic materials-testing machine (MTS, Minnesota,
USA) at a rate of 0.1 mm per second (strain rate 1 % per second). The speci-
mens were not otherwise constrained. The force displacement history was recorded
digitally for 75 seconds at a sampling rate of 20 Hz.
Following mechanical testing the specimens were ashed in a box furnace for
24 hours at 580 degrees Celsius (S 1800 model, Lindberg Unit of General Signal,
To Characterize the Structure of Cancellous . . . 303
Wisconsin, USA). Ash mass of each specimen was then determined allowing
calculation of ash density from the previously measured, pre-compression volume.
Irregular regions of interest (ROIs) corresponding to the projected area of the
femoral head were traced on the digital images displayed on a computer screen.
The pixels within these ROIs were used in the subsequent digital analysis. To
obtain an estimate of the «fractal dimension» of the ROJ, a method similar to
that of Lundahl et al. (1985) was applied. Conceptionally, the radiograph was
considered as a topologically three dimensional object, with the optical density at
each point in the radiograph (x, y) plane treated as a distance measure representing
the third topological dimension. Thus, the digital image was taken as representing
a «mountainous» surface. The problem of assessing the fractal dimension of a
mountainous surface can be solved by fitting an equation of form :
(1)
where
(AE) is the area of the surface measured with a square of side E
A is a scaling constant
D is a constant characteristic of the surface (i.e., the «fractal dimension»)
The surface area of the radiograph was calculated using different «square
sides». In computing the area as a function of square size, the surface of the
image was considered to be a collection of adjacent «skyscrapers», that is, a set
of rectangular columns of differing heights but with square «roofs» having side
length E. The area of the surface is given by the sum of the area of the «roofs»
plus the sum of the areas of the exposed sides of the columns:
A(E) = L E2 + LEI I(x,y) - I(x + l,y) I + I I(x,y) - I(x, Y + 1) I

x,y x ,y
where I (x , y) is the height of a column, found for a particular value of E by

averaging the values stored in adjacent image array elements to produce pixels
with a side length of E.
A(E) was first calculated with E equal to single pixel width (0.175 mm).
Subsequently, the grey levels in squares of 4, 9, 16, 25, and 36 contiguous pixels
were averaged to form new image arrays. A(E) was re-calculated for each E.
As expected for a fractal surface, the measured area was dependent on the
ruler size. From Equation 1, the fractal dimension for the «mountainous» surface
is directly related to the slope of a plot of 10g(A( E)) versus log( E). A more rugged
or complex surface results in a higher fractal dimension. A gentler or smoother
surface results in a lower fractal dimension.
In addition to the fractal dimension, mean pixel grey level and grey level
standard deviation were derived for the pixels in each ROJ.
304 Caldwell et al.
3 Results
The data derived from the 15 patients are presented in Table 1. Ash density,
Singh Index, fractal dimension, mean pixel grey level, pixel standard deviation and
compression strength are listed. The value of the correlation coefficient relating to
the degree of linear correlation between each parameter and compressive strength
is found in Table 2. Note that the highest correlation is between ash density and
compressive strength. Of the remaining parameters, only the fractal dimension
displayed a high degree of correlation with compressive strength.
Compress Density Singh Fractal D Mean Grey Standard

Strength Index Level Deviation
(MPa) (mg/cm3) in Grey Level
0.57 83 2 2 6. 4 639 54
0.75 89 3 2 6. 4 678 55
0.80 78 4 2.69 586 57
0.86 109 5 2.66 633 31
0.96 95 2 2.65 773 57
1.0 124 3 2.61 792 27
1.3 103 6 2.66 648 50
1.6 143 6 2.74 667 60
1.6 108 2 2.65 669 47
1.7 128 3 2.86 773 42
1.7 116 5 2.85 709 44
1.7 103 5 2.77 741 60
1.9 128 3 2.81 813 41
2.3 144 4 2 7. 9 772 54
3.5 183 5 2.78 676 54
Table 1 Measurements for 15 Subjects
Parameter Tested Linear Correlation p-value

versus Compressive Coefficient
Strength
Density 0.89 < 0.001
Fractal Dimension 0.65 < 0.01
Singh Index 0.33 > 0.1
Mean Grey Level 0.27 > 0.1
Standard Deviation
0.15 > 0.1
in Grey Level
Table 2 Correlations with Compressive Strength

To Characterize the Structure of Cancellous . .. 305
4 Discussion
In this preliminary study, we found that patients who had relatively strong can-
cellous bone also had radiographs with higher fractal dimension than did patients
who had relatively weak cancellous bone. While the number of patients studied
was small (N = 15) this may indicate that normal cancellous bone has a sharply
defined structure of high fractal dimension which may be «smoothed out» to a
lower fractal dimension by the action of osteoarthritic disease. Two simpler mea-
sures (the mean and standard deviation in pixel grey level) did not appear to be
useful predictors of bone strength. The Singh Index was also not strongly corre-
lated with bone strength. Bone density, as assessed by ash mass measurements, was
highly correlated with bone strength, as expected. However, conventional methods
of measuring bone density in vivo are unlikely to correlate with strength to this
degree. Bone density measurements made by conventional (i.e., non-CT) methods
in vivo are, in effect, averages of the density of both cortical and trabecular bone.
Variations in the amount of fat in tissue surrounding the bone can also disturb the
estimate of bone density by these techniques. In addition, even if knowledge of
cancellous bone density were sufficient to predict cancellous bone strength, can-
cellous bone density is not measured directly by either DPA or DEXA. It may be
that the quality of cancellous bone is of great importance in some clinical appli-
cations. For example, the role of cancellous bone in the support of an un cemented
femoral component remains unknown. It has been postulated that the quality of
cancellous bone is important in resisting rotational forces at the prothesis bone
interface (Spotomo and Rognoli 1988). A parameter based on a combination of
bone density and structural measurements (possibly using the fractal dimension)
may prove to be useful in the evaluation of cancellous bone using only digitized
conventional plane film radiographs. This would be preferable to the current status,
where neither bone density measurements nor subjective evaluations of plane films
provide adequate information to predict the quality of cancellous bone. It may also
be preferable to extensive use of quantitative x-ray CT, a relatively expensive and
not widely available procedure.
Acknowledgements:
The authors are grateful to Dr. Brian Howard for his help in assigning a Singh
Index value and to Dr. Keith Willett for helpful discussions.
References
[1] Browne MA, Gaydecki PA, Gough RF, Grennan DM Khalil SI, Mamtora
H Radiographic image analysis in the study of bone morphology Clin Phys
Physiol Meas 1987 8: 105-121.
[2] Cargill EB, Donohoe K, Kolodny G, Parker JA Zimmerman RE Analysis of
lung scans using fractals Proc SPIE 1989 1092: 2-9.
[3] Hui SL, Slemenda C, Johnston CC Age and bone mass as predictors of
fracture in a prospective study J Clin Invest 1988 81: 1804-1809.
306 Caldwell et al.
[4] Kaandorp JA Modelling growth fonns of sponges with fractal techniques in:
Fractals and Chaos New York: Springer Verlag 1991.
[5] Lotz JC, Gerhart TN, Hayes WC Mechanical properties of trabecular bone
from the proximal femur: a quantitative CT study J Comput Assist Tomogr
1990, 14: 107-114.
[6] Lotz JC, Hayes WC The use of quantitative computed tomography to estimate
risk of fracture of the hip from falls J Bone Joint Surgery 1990 72-A: 689-
700.
[7] Lundahl T, Ohley WJ, Kuklinski WS, Williams DO, Gerwitz H, Most AS
Analysis and interpolation of angiographic images by use of fractals Com put
in Cardiol 1985 24: 355-358.
[8] Lynch JA, Hawkes DJ, Buckland-Wright JC Analysis of texture in macrora-
diographs of osteoarthritic knees using the fractal signature Phys Med BioI
1991 36: 709-722.
[9] Mainster MA The fractal properties of retinal vessels: embryological and
clinical implications. Eye 19904: 235-241.
[10] Mandelbrot BB The Fractal Geometry of Nature New York: W.H. Freeman
and Company 1983.
[11] Martens M, Van Audekercke R, Delport P, De Meester P, Mulier JC The
mechanical characteristics of cancellous bone at the upper femoral region J
Biomechanics 1983 16: 971-983.
[12] Ott SM, Kilcoyne RF, Chesnut CH III Longitudinal changes in bone mass af-
ter one year as measured by different techniques in patients with osteoporosis
Calcif Tissue lnt 198639:133-138.
[13] Pentland AP Fractal-based descriptions of natural scenes IEEE Trans Pattern
Anal and Machine Intelligence 1984 PAMI-6: 661-674.
[14] Rosson J, Surowiak J, Schatzker J, Hearn T The relationship between the
radiographic appearances and the structural properties of proximal femoral
bone in patients undergoing total hip arthroplasty (submitted for publication)
[15] Ruttimann VE, Webber RL, Hazelrig JB Fractal dimension from radiographs
of peridental alveolar bone: A possible diagnostic indicator of osteoporosis
Oral Surg Oral Med Oral Pathol 1992 74: 98-110.
[16] Singh M, Nagrath AR, Maini PS Changes in trabecular pattern of the upper
end of the femur as an index of osteoporosis J Bone Joint Surg 1970 52-A:
457-467.
[17] Spotorno, Rogmagnoli. Indications for the CLS stem in: Cementless Total
Hip Replacement System CLS. Monograph published by Protek AG, 1988.
[18] Stulberg BN, Bauer TW, Watson JT, Richmond B Roentgenographic versus
histologic assessment of hip bone structure Clinical Orthopaedics and Related
Research 1989 240: 200-205.
Distribution of Local-Connected
Fractal Dimension and the Degree of
Liver Fattiness from Ultrasound
Carl J.G. Evertsz 1), C. Zahlten 1), H.-a. Peitgen 1) ,
I. Zuna2 ), and G. van Kaick2 )
I) Center for Complex Systems and Visualization, University of Bremen (FB III), Box 330 440,
0-28334 Bremen, Germany
2) Oeutsches Krebsforschungszentrum, 1m Neuenheimer Feld 280, 0-69120 Heidelberg, Germany
1 Introduction
Central to fractal geometry [3] is the concept of self-similarity. Fractal and mul-
tifractal [1] analysis provide tools for the quantitative analysis and description of
self-similarity in natural and mathematical sets and distributions. The arsenal of
tools for fractal analysis which is growing both in diversity and sophisication pro-
vides fresh new ways to analyse geometries, even those which are non-generically
fractal. This contribution to Fractals in Medicine discusses an application of a
hybrid of fractal and multifractal analysis to a problem in medical ultrasound
imaging. Our method is a continuation of work by Richard Voss [7] which he
applied to the X-ray detection of malignant breast tumors and the classification of
chineese landscape paintings.
Ultrasound imaging of the liver is a standard part of a checkup by the spe-
cialist in internal medicine. Besides being a harmless tool for diagnosing several
life threathening diseases ultrasound is know to provide some information on the
degree of liver fattiness. It is also known to be pretty much imprecise and to a
large degree subjective when it comes to diagnosing the latter. Quantitative image
analysis could play a role in an assessment of the limitations of ultrasound as a
diagnostic tool for liver fattiness .
The right unprocessed liver ultrasound in Figure 1 is from a healthy patient
and the left one from a patient with a fatty liver. Our aim is to quantify the
qualitative strategy an experienced medical doctor would apply when asked to
diagnose the fat condition of a patient's liver given only such ultrasound images.
Depending on this person's experience a varying number of visual aspects are
taken into account; amongst others the shape of certain parts of the liver, the
overall intensity of the image, the liver/bloodvessel -contrast, and the intensity
difference between the liver and kidney tissue, etc.
308 Evertsz et al.
Fig. 1 Ultrasound images of a fatty (left) and a healthy (right) liver.
Ultrasound images of homogeous tissue are not smooth, but have a speckled
stucture, which in our images have typical sizes of about 30 pixels. One quanti-
tative approach is based on texture analysis of a small region of interest (ROI) in
the ultrasound image containing only liver tissue [5]. The basic assumption is that
the speckle structure contains information about the fattiness of the liver tissue.
The other, new approach, is much more close to the qualitative analysis of the
medical specialist and tries to take as much as possible the impression conveyed
by the whole image. It concentrates on the overall contrast between bloodvessels
and liver tissue and that between liver tissue and liver boundary. The microscopic
speckle structure is of little relevance.
We first review the relevant basic tools from fractal geometry [3][4], namely
mass dimension and local mass dimension, and then move on to the specifics of
the method.
2 Mass dimension
Figure 2a shows the three stages of a «growing» Sierpinski triangle. The building
blocks are little triangles of size I and mass 1. Thus, the respective sizes L of the
three stages depicted are 1, 2 and 4. The corresponding masses are M(I) = 1,
M(2) = 3 and M(4) = 9 ... and so on for increasing sizes. Therefore, the mass
and length have a scaling relationship of the form M (L) "-' L D where the exponent
D = log 3/ log 2 is called the mass dimension.
3 Local mass dimension

The local mass dimension D]oc also characterizes a mass/size relationship. Let
x be a point in the set considered, and let Bdx) be a box or disc of size L
centered around x. When the mass M(Bdx)) of such boxes of different sizes
scales like M(Bdx)) "-' LDloc then D]oc is called the local mass dimension in
Distribution of Local-Connected Fractal Dimension 309
o
Fig. 2 a) Left: the first three stages of a Sierpinski triangle. b) Right: A black and white image
containing two objects; a line and a surface.
point x. In general this dimension varies with the location x. For example in
fig. 2b Dloc(X) = I and Dloc(z) = 2.
In medical images this scaling behaviour can only be tested with box-sizes
between I and the size of the image. When the 10gM(BL(x)) versus IogL plot
yields a straight line, or oscillates around a line, then the concepts of dimension
make sense and one can use linear regression to determine the exponents D.
4 Local-connected mass dimension in image analysis

Most often an image contains several objects which can be differentiated by means
of preprocessing steps. When such a segementation of the image into distinct
objects is possible it can be desirable to only consider the local mass dimension in
those points x belonging to a particular object. However, in general boxes BL(x)
will also cover parts of objects different from the one to which x belongs: BL (y)
in figure 2b covers both the line and the square. When one does not want to mix
the geometric properties of the object point x is part of with those of other objects,
it is desirable to only count the mass contributed to the box by the object to which
x belongs. For example, let the line in figure 2b be separated from the square by
a distance 8, and let us take Lmin larger than 8, say Lmin = 28 and Lmax = 88.
The local mass dimension D loc in point x will be approximately Dloc (x) ::::; I
while Dloc(y) ::::; Dloc(z) ::::; 2. The reason for Dloc(Y) ::::; 2 is that the boxes BL(y)
will not only cover the I-dimensional line, but also part of the 2-dimensional
square. If the line and the square are recognized as two different objects then one
could exclude the contribution of the square to the mass in the boxes BL (y) and
only keep the contribution of the line. This would yield a local-connected mass
dimension [7] Dconn(Y) ::::; 1. So the local-connected mass dimension is an object
oriented quantitative characterization of a geometry.
310 Evertsz et al.
There are many techniques to segment images into distinct objects. The image
in figure 2b consists of a white (l = 1) background and a black (I = 0) geometry.
Therefore the first step in segmenting this image is simply to take all sites x with
intensity I(x) = 0 to belong to objects and the rest of the sites to be background.
This is segmentation by means of intensity thresholding. The object to which, say
point x belongs, consists of all other sites with intensity 0 which can be connected
to x by a string of nearest neighbor sites. The application to liver utrasound
uses such segmentation by thresholding combined with a further segmentation
in connected objects, called clusters. The mass Mconn used for the estimation
of the local-connected mass dimension Dconn by means of the scaling relation
Mconn (B d x)) rv L Dconn, only incorporates the mass of those points which are also
connected to the center point x.
5 Distribution of local-connected mass dimension

The local-connected mass dimension Dconn(x) contains quantitative information
about the local geometry of the cluster to which x belongs. In general Dconn(x) will
differ with the location x. One convenient way of representing this large amount
of data is to consider the probability density p(~)d~, which is the probability that
a randomly picked point x on any of the segmented objects has a local-connected
mass dimension with ~ < Dconn(x) < ~ + d~ . In such an analysis of figure 2b,
the «line-cluster» will contribute a peak in p( ~) at ~ = 1 and the «square-cluster»
a peak at ~ = 2. So by only looking at the distribution p(~)d~, one can conclude
that the original picture consists of linear structures (the line) and surfaces (the
square).
6 Preprocessing and segmentation of the liver ultrasound image

In order to get rid of possible noise in the image due to the lower cutoff in
the resolution of the ultrasound machine or due to the digitization procedure, the
images are smoothed on length-scales of the order of the pixelsize. This can be
done in several ways. We chose to replace the intensity I(x) in each pixel x by the
average intensity Is(x) over a square box of sides 7 x 7. For x near the boundaries
of the image one only averages over the image points within the box.
Since the speckle structure, the vein/liver and liver boundaries seemed im-
portant, we decided to enhance the edges by taking the normalized gradient of the
image, i.e., the gradient of the logarithm of the image
I s.g -'r71 J( )_lgradIs(x)1

- v og s X - Is(x) .
The new image Is.g has the advantages that it is independent of the setting of
the intensity button on the machine or the variable layer of body fat between
Fig. 3 The nonnalized gradients of the smoothed images depicted in figure I.
Fig. 4 The white regions have a nonnalized gradient larger than the threshold value T. The rest is
black. The value T used here was the one giving the largest differentiation between images
of typically fatty and healthy livers.
utrasound head and liver. Here it is assumed that the ultrasound intensity output is
linear in the relevant intensity domain. Also, the speckle structures which usually
have a weaker intensity than the vein/liver and liver boundaries, are put on a par.
The processed images ls.g corresponding to the images in figure 1 are shown in
figure 3. The small wormy structures are the edges of the speckles, and the large
structures are the edges of the veins and liver boundaries.
In order to segment the image ls.g one chooses a threshold T and makes
black all pixels with ls.g(x) < T, and white all others. An example is shown in
figure 4. All connected objects can be found using an algorithm developed for
cluster detection in percolation theory [6]. This yields a new image lobj in which
all clusters are numbered and all pixels x in a cluster have lobj (x) equal to the
cluster number. This is a considerable aid when computing the local-connected
mass dimensions.
312 Evertsz et al.
1. 0,....-~Iii :----.---.---~---,
Prob ( D c.... ~ 1.5 )

0.8
~ 0.6
o
A me.hen
o
.g 0.4
ct
10
0.0 o
0.0 0.5 1.0 1.5 2.0 2.5 00 05 10 15 20 25
o
Dc....
Fig. 5 Cumulative distributions of the local· Fig. 6 The dimension probabil-

connected mass dimension for twenty ity densities for the 20
patients with different degrees of liver patients. The dots mark
fattiness. the median.
7 Quantitative analysis of the liver images

The analysis starts by determining the local-connected mass dimensions of all
points belonging to all objects in the image Iobj. This is done using square boxes
BL(x) of sides 2,4 and 8 and then doing a least square fit of 10g.Mconn(BL(x))
versus log L to find Dconn (x). In order to insure that one is doing something sensi-
ble, only those points with a small error in the least square fit are considered. The
probability density of the local-connected mass dimension provides a quantitative
description of the geometry of the objects.
The optimal value for T is determined experimentally by doing the analysis
for different values of T for both a typical ultrasound image of a healthy and fatty
liver. The optimal value of T is the one yielding the largest difference among
the two probability densities p(b.)db. (see the marked curves in figure 6). Such
optimal threshold values T were determined for both the analysis of ROJ's and
the overall analysis, and are subsequently used in the analysis of liver ultrasounds
of new patients.
8 Results for 20 patients

For each of twenty patients we estimated the p(b.)db., combining results from
3 ultrasound liver images slices of each. The settings of the ultrasound machine
were the same for all patients.
Figure 5 shows the cumulative distribution of local-connected mass dimen-

sion for the 20 patients. Two numbers seem to be relevant: The first is the median,
i.e., the value of the dimension where the different curves intersect with the hor-
izontal line through 0.5. The second is the probability Prob{D1oc > 1.5} for a
local-connected mass dimension larger than 1.5, i.e., the value of the cumulative
probability at which the different curves intersect with the vertical line through
D 10c = 1.5. Both the curves corresponding to the typically healthy and fatty liver
intersect this vertical line at very different values: For the healthy liver (squares)
half of the points have dimension larger than 1.5. For the fatty (triangle) one
that's only 10%. In a similar fashion also the median provides such a quantitative
differentiation between healthy and fatty.
Figure 6 shows the corresponding densities p(6.)d6. ordered from top to
bottom by increasing median, i.e., increasing content of higher dimensional objects.
Since the boundaries between the different tissues is not so clearly demarkated in
fatty livers, their thresholded Is.g images contain less 2 dimensional structures (see
figure 4) as those of healthy ones. This explains the potential relevance of both
the median and Prob{D1oc > 1.5} as measures of the degree of liver fattiness.
Similar plots were also investigated for ROI's containing the speckle texture
of pure liver tissue. The cumulative distributions obtained from that analysis were
all very much the same, making them apparently useless for quantitative diagnostic
purposes.
9 Comparison with CT
As a test our results were compared with CT scan fat estimates for the same livers:
the intensity of CT scans is known to correlate well with tissue's fat content. The
Spearman correlation coefficient [2] between the CT ordening and the ordening
according to increasing Prob{D 1oc > 1.5} for these 20 patients was 0.74, while
that between the CT ordening and median ordening was 0.63. In absolute sense
these numbers may not seem impressive. Nevertheless they are very much better
than anything obtained in previous textures analysis.
Showing the ultrasound images to a specialist, asking this person to order
them in 4 categories, we found remarkably close agreement with our fractal image
analysis. In a sense this is not surprising, since the method of analysis discussed
here was aimed at quantifying what the specialist's eye was capturing.
Two ingredients are new in this context: First the fractal analysis based on the
distribution of local dimensions and second, perhaps more important, our departure
from the traditional methodology of looking at small regions of interest containing
only smooth liver tissue.
Acknowledgements:
We are very grateful to Richard Voss for many discussions and for his plot program
VP. We also like to thank Dr. W. Hofer for his diagnoses of the unprocessed
ultrasound liver images.
314 Evertsz et al.
References
[1] Evertsz, C.J.G., Mandelbrot, B.B. Multifractal measures In: Ref. 4, 921-953.
[2] Hays, w.L. Statistics Holt, Rinehart and Winston, London, 1963.
[3] Mandelbrot, B.B. The fractal geometry of nature W.H.Freeman, San Fran-
cisco, 1982.
[4] Peitgen, H.-O., Jurgens, H., Saupe, D. Chaos and Fractals Springer-Verlag,
New York, 1992.
[5] Raeth, U., Schlaps, D., et al. Diagnostic accuracy of computerized B-scan
texture analysis and conventional ultrasonography in diffuse parenchymal
and malignant disease J. Clin. Ultrasound 13,87-99, 1985.
[6] Stauffer, D., Aharony, A. Introduction to Percolation Theory Taylor & Fran-
cis, London, 1985.
[7] Voss, Richard F. The local connected fractal dimension: multifractal classifi-
cation of early Chinese landscapes drawings In: R.A.Earnshaw (ed.) Appli-
cation of Chaos and Fractals, Springer-Verlag, London, 1992.
[8] ZahIten, c., Evertsz, C.J.G., et al. Fraktale in der Analyse van Ultraschall-
bildern der Leber preprint
Fractal Dimension as a Characterisation
Parameter of Premalignant and Malignant
Epithelial Lesions of the Floor of the Mouth
Gabriel Landini and John W. Rippin
Oral Pathology Unit, School of Dentistry
University of Birmingham
St. Chad's Queensway
Birmingham B4 6NN, England, U.K.
Abstract. Irregularity of shape of epithelial-connective tissue interfaces is a well-recognised feature of

malignant and premalignant epithelial lesions, yet few attempts have been made to assess it objectively.
The fractal dimension (as a measure of irregularity of shape) of the epithelial-connective tissue interface
of normal , premalignant and malignant epithelial tissues of the floor of the mouth was measured using
box counting and boundary trace methods.
The lowest estimated value was 0.97 (a normal mucosa using the box method), and the highest
1.61 (a carcinoma using the trace method). Analysis of the values against the histopathological diag-
noses (normal, keratosis with mild dysplasia, keratosis with moderate/severe dysplasia and squamous
cell carcinoma) showed no significant difference between normal epithelium and that from kerato-
sis with mild dysplasia, but these were significantly different from the two. other diagnoses, which
were also significantly different from each other. The study illustrates the potential of fractal analysis
for providing objective diagnostic information about irregular shapes in histopathology and objective
descriptors for tumour growth.
1 Introduction
The oral cavity is covered by mucosa which is composed of two different tissues:
a stratified squamous cell and underlying connective (lamina propria). Malignant
transformation of the epithelial tissue into a squamous cell carcinoma may occur,
and its occurrence is thought to depend on a multitude of exogenous and endoge-
nous factors [II]. Oral cancer accounts for approximately 5% of all diagnosed
tumours in western countries and as much as 60% in some South Asian regions.
The shape of the junction between the basal epithelial cells and the under-
lying lamina propria, the epithelial-connective tissue interface (ECTI) of the oral
mucosa is regular: the shapes of rete pegs and interdigitating papillae repeat. In
the (normal) floor of the mouth, which is the most susceptible area of the oral
cavity to cancer transformation, the shape of this interface is relatively flat. In pre-
malignant lesions (epithelial dysplasia), the ECTI often becomes irregular due to
increased cell proliferation. In frank malignancy (squamous cell carcinoma), this
effect is more marked and detached islands of cells penetrate the lamina propria.
At this stage, as seen ultrastructurally, the basement membrane can be defective
[8] or even absent [I] . One of the pathologist's major difficulties is classifying
premalignant mucosal lesions. Most often this is done in a rather arbitrary (or at
316 Landini and Rippin
a
b
c
d
Fig. 1 Four profiles of epithelial-connective tissue interfaces form (a) normal tissues, (b) mild ep-
ithelial dysplasia, (c) moderate/severe epithelial dysplasia and (d) squamous cell carcinoma
of the floor of the mouth.
least, subjective) way based on a varying number of cytological features such as

number of mitosis, cellular and nuclear atypia and pleomorphism, nuclear hyper-
chromatism and also irregularity of the Een [4][5].
It has been suggested that premalignancy and malignancy progress by the
proliferation of some transformed stem cells at the expense of neighbouring normal
cells [15]. Such proliferation may be balanced by increased cell exfoliation at the
free surface of the mucosa, but it also helps to account for the irregularity of the
EeTL Since fractal geometry can be regarded, for present purposes, as a numerical
method of describing the irregularity of shapes [7], fractal analysis may provide an
objective means of classifying the irregularity of the EeTI and eventually become
a diagnostic parameter in histopathology.
2 Materials and methods

Hematoxylin-eosin stained histological sections of lesions from the floor of the
mouth, presenting to the routine histopathology service of the University of Birm-
ingham School of Dentistry, were selected. Sections from lesions of each of the
following were examined and re-diagnosed: 10 normal mucosae, 10 keratoses
with mild dysplasia, 10 keratoses with moderate or severe dysplasia and 10 well
differentiated squamous cell carcinomas.
Premalignant and Malignant Epithelial Lesions . .. 317
10000 ,-- -.=:

,- ,-
-
-
4·-
.-
f-
-·4 f-- - ,- f- f--
4
1000 r......
- - -1-
(~
.. --I-
"-
'"
1-0 _
en (
>()
w IIlr..
.
X 100
o :>J
ID
I'h -
- "<.ii"h !II
~~ III ,-
~~
1"11
I
10
- -:-=:
,-
--= - - ~
-
- --I -
I I
10 100 1000
BOX SIZE
Fig. 2 The fractal dimension of the epithelial-connective tissue interface by the box method is given
by the negative of the slope of the linear part of the log-log plot. Empty circles represent
data from figure la (normal tissue, box dimension = 0,98); filled circles: data from figure ld
(squamous cell carcinoma, box dimension = 1.47). Note the tendency to flatten of the upper
part of the plot. Box size is in pixel units,
The classification of dysplasia into mild, moderate and severe categories, at

least as far as the oral mucosa is concerned, is largely dependent on the pathol-
ogist's subjective interpretation [11], but it is recognised practice. The dysplastic
lesions were grouped into two as «mild» and «moderate/severe», since it has been
found that the former are much less likely to transform into carcinoma than the
latter. The distinction between moderate and severe dysplasias is less marked in
this respect, and hence it is recommended that, for the floor of the mouth, mod-
erate/severely dysplastic lesions be fully excised, while mildly dysplastic lesions
may be treated conservatively [11].
Projected images of the junction between the basal epithelial cells and the
adjacent lamina propria were enlarged 23.7 times, hand traced and then scanned
into a personal computer (Fig. 1) achieving a final resolution of 1 pixel = 11.72
micrometres.
The fractal dimensions of the shapes were calculated using the box counting
and boundary trace methods [7][2]. The box counting method measures the space
OBOX .TRACE
1.9
1.8
1.7
Z
0
enz
1.6 •
•
••
W
:i!:
1.5
0 I
0
...J
1.4
• 0
•
c{
1.3
I
~
~
u
~
• §I
1.2
u.
0
• 6•
~I
1.1
~I 0
0.9
NORMAL MILD MOD/SEVERE CARCINOMA
Fig. 3 Fractal dimensions of the epithelial-connective tissue interfaces of the 40 cases estimated by
the box and trace methods.
filling properties of an object at different resolutions by superimposing a grid of

increasing size on the planar space of the figure being measured, and counting the
number of «boxes» (N (r)) of linear size r that contain any part of the figure. If a
plot oflog(r) against 10g(N(r)) tends to a straight line in some range ofr (usually
larger than one order of magnitude), then the image can be regarded as fractal,
and its fractal dimension (D) is estimated as D = -$, where $ is the slope of
the regression line in that range. In the present study, 58 box sizes of side length
ranging from 1 to 128 were used (Fig. 2).
With the boundary trace or yardstick method [7], the length of the image (1)
is measured with, rulers of increasing size (r). The value of D is calculated as
D = 1 - $, where $ is the slope of the linear part of the regression of log(r)
on log(l). In this study the ruler sizes used were from 1 to 200 pixels (unit steps
from 1 to 10, steps of 5 from 10 to 30, steps of 10 from 30 to 200). Mean ruler
length values were used after each figure measurement; this is necessary because
of unavoidable errors associated with the square shape of the computer pixels. For
example, the possible distances in a boundary for a ruler r = 2 are: 2, SQR(5) and
SQR(8). The methods were validated by measuring computer generated shapes of
known fractal dimension. For the experimental data a two-way analysis of variance,
with diagnosis and method of fractal dimension estimation as the variables, was
carried out.
Premalignant and Malignant Epithelial Lesions . .. 319
ONORMAL
1.7 .MILD
llMODISEVERE
.CARCINOMA
1.6
•••
..
1.5
z
D 1.4
(ij
• •
z
w ~
::E 1.3
C
w II
• II
()
II
~ 1.2
4l
0: 8
~
1.1
o
~o
0.9
0.9 1.1 1.2 1.3 1.4 1.5 1.6 1.7
BOX DIMENSION
Fig. 4 Box dimension versus trace dimension. Note the clear separation of the normal from the
malignant cases.
3 Results
Estimated fractal dimensions of standard figures agree very closely with their
known fractal dimensions (Table 1). For the experimental data, the mean ranges
of scales for which the log-log plots were linear were between 40.4 and 884.6 mi-
crometres for the box counting method, and between 44.6 and 1134.3 micrometres
for the trace method. For ranges smaller than these, the plots tended. to flatten to
o = 1 (Fig. 2). A two-way analysis of variance showed significant differences
between the methods of fractal dimension estimation (P < 0.001) and the diag-
noses (P < 0.001), but no significant interaction between method and diagnosis
(P > 0.4). Further examination showed that for both methods, the data for normal
mucosa was not significantly different from that for keratosis with mild dysplasia
(P > 0.7), but that these were both significantly different from the two other
diagnoses (moderate/severe dysplasia and squamous cell carcinoma), which were
in tum significantly different from each other (for all, P < 0.05). The two meth-
ods gave significantly different results for each diagnosis (P < 0.05); as can be
Object Ideal D BOX SD TRACE SD

Line 1.0000 0.99 0.004 1.00 0.000
Circle 1.0000 1.00 0.013 1.00 0.000
Square (perimeter) 1.0000 1.00 0.000 1.01 0.009
Square (filled) 2.0000 1.97 0.013 *
Koch island coastline 1.2618 1.26 0.018 1.26 0.004
Koch island coastline 1.5000 1.49 0.018 1.48 0.018
Table 1 Ideal and mean estimated fractal dimension of known objects. The mean values were obtained
from 5 measurements starting at random initial positions of the measurements. SD: standard
deviation. * not applicable
seen (Figs. 3 & 4), the trace method gave fractal dimensions consistently higher
than the box method by a factor of about 1.06. The minimum fractal dimension
obtained was 0.97 for a normal mucosa (box method), and the maximum 1.61 for
a carcinoma (trace method).
4 Discussion
Mathematical fractals show scaling at all orders of magnitude, from the infinitely
small to the infinitely large, but in natural structures there are usually both up-
per and lower bounds. In the present study, the log-log plots showed a linear
trend at relatively coarse resolutions, but asymptotic behaviour tending to D = 1
at relatively high ones, in accord with Rigaut's concept of an asymptotic fractal
[12][13]. It may be that as the measurements approximate the highest light mi-
croscope resolutions, the size of the individual cells is reached, whence the shape
of the ECTI would depend on individual cell factors (such as the variability in
number of pseudopodia and hemi-desmosomes) [9][16] rather than on the lesion
as a whole.
Interestingly the changes of the ECTI reveal in some way the dynamics of the
tissue in its different stages and this prompts speculations about tumour growth
as a self-similar process. Self-similar processes have already been proposed as
being responsible for pattern formation in normal chimaeric liver of rat [3] and
in simulated «cell pushing» replicative systems [6] revealing their characteristic
fractal measures. However in carcinogenesis and tumour growth, it may help to
understand local aggressivity, tumour infiltration and metasatatic spread.
It is not surprising that, as measured by fractal dimension, the shape of the
ECTI is significantly more complex in squamous cell carcinomas than normal mu-
cosa, and dysplastic lesions are somewhere between the two. Many pathologists
would regard such statements as facile, and yet few if any would be prepared to
make their diagnoses based on the shape of the ECTI alone (Fig. 1). Instead diag-
nosis is done by taking into account a variable number of features, most often in a
non-systematic way. By contrast, the methods of fractal analysis are systematic and
Premalignant and Malignant Epithelial Lesions . . . 321
statistically reliable. Other morphometric methods such a~ the quantification of the

number of basal cell pseudopodia [9][16] have been described, but these methods
involve electron microscopy and intensive user intervention on what structures to
analyse. Further attempts to standardise the shape of the epithelial buds (and con-
sequently the irregularity of the ECTI) in the esophagus yielded a classification of
epithelial buds of type I, II or III, depending if they are regular and same sized
(type I), regular but differing in size (type II) or irregular (type III) [14]. However
this represents another subjective categorical variable.
Scale-bounded measurements (such as maximum length of the ECTI at a de-
termined resolution) are of use only if all the measurements are done with the same
resolution [10]; at the moment there is no such standardisation in histopathology.
This approach may also lead to wrong assumptions about the assessment of «ir-
regularity» in lesions such as pillomatrixomas, inverted papillomas and basal cell
carcinomas (lesions with large and smooth epithelial buds) in which the length of
the ECTI is increased but there is no «irregularity». The technique presented here,
after the isolation of the epithelial-connective tissue interface (now being investi-
gated using selective staining methods and a digitizer camera), is fast, straightfor-
ward and performed at a relatively low magnification. Of course, this is not to say
that every single case can be classified on the basis of the fractal dimension of
its ECTI. The usual implication of designating a lesion as showing «mild epithe-
lial dysplasia» or «severe epithelial dysplasia» is that the latter is more likely to
transform than the former, yet such studies as have been carried out have demon-
strated the poor reliability of such a process [17]. Thus, it may be that those
lesions that in the present study have been designated as «mild dysplasia» but had
a high fractal dimension (the highest in this group was l.29), or those designated
as «moderate/severe dysplasia» but with a high fractal dimension (the highest in
this group was l.37), have in fact a greater risk of transformation than the histo-
logical designation would imply. This contention can be appraised only when a
large number of cases have been analysed and followed for a number of years,
when their ultimate prognosis will be known. The fractal dimensions can then be
statistically analysed against that prognosis in order to determine the predictive
value of fractal dimension of the ECTI in histopathology.
Acknowledgement:
This study was supported by The Leverhulme Trust grant No. F.94AO.
References
[1] Ashworth CT, Stembridge VA and Luibel FJ: Astudy of basement membranes
of normal epithelium, carcinoma in situ and invasive carcinoma of the uterine
cervix utilizing electron microscopy and histochemical methods. Acta Cytol
5: 369-384, 1961.
[2] Barnsley M: Fractals everywhere. San Diego, Academic Press, 1988.
[3] Iannaccone PM. Fractal geometry in mosaic organs: a new interpretation of

mosaic pattern. FASEB J 4: 1508-1512, 1990.
[4] Kramer IRH: Computer-aided analyses in diagnostic histopathology. Postgrad
Med J 51: 690-694, 1975.
[5] Kramer, IRH, EI-Labban N, Sonkodi S: Further studies on lesions of the oral
mucosa using computer-aided analyses of histological features. Br J Cancer
29: 223-231, 1974.
[6] Landini G, Rippin JW: Fractal fragmentation in replicative systems. FRAC-
TALS (In press), 1993.
[7] Mandelbrot BB: The fractal geometry of nature. San Francisco, Freeman;
1982.
[8] McKinney RV, Singh BB: Basement membrane changes under neoplastic oral
mucous membrane. Ultrastructural observations, review of the literature, and
a unifying concept. Oral Surg 44: 875-888, 1977.
[9] Okagaki T, Clark BA, Twiggs LB: Measurement of number and cross-section-
al area of basal cell pseudopodia: a new morphometric method. J Cell BioI
91: 629-636, 1981.
[10] Paumgartner D, Losa G, Weibel ER. Resolution effect on the stereological
estimation of surface and volume and its interpretation in terms of fractal
dimensions. J Microsc 121(1): 51-63, 1981
[11] Pindborg 11: Oral cancer and precancer. Bristol, John Wright & sons Ltd.;
1980.
[12] Rigaut JP: An empirical formulation relating boundary lengths to resolution
in specimens showing «non-ideally fractal» dimensions. J Microsc 133 Ptl:
41-54, 1984.
[13] Rigaut JP: Fractals, semi-fractals et biometrie. In: Fractals, dimensions non
entieres et applications. G. Cherbit ed. Paris, Masson, 1987.
[14] Rubio CA, Liu F-S, Zhao H-Z: Histological classification of intraepithelial
neoplasias and microinvasive squamous carcinoma of the esophagus. Am J
Surg Pathol 13: 685-690, 1989.
[15] Selby P, Buick RN, Tannock I: A critical appraisal of the «Human Tumor
Stem-cell Assay». N Engl J Med 308, 129-134, 1983.
[16] Twiggs LB, Clark BA, Okagaki T: Basal cell pseudopodia in cervical intraep-
ithelial neoplasia; progressive reduction of number with severity: a morpho-
metric quantification. Am J Obstet Gynecol 139: 64~4, 1981.
[17] Wied GL, Bartels PH, Bibbo M, Dytch HE: Image analysis in quantitative
cytopathology and histopathology. Hum Pathol 20: 549-571, 1989.
Modelling
Modelling HIV/ AIDS Dynamics
Philippe Blanchard
Theoretische Physik and BiBoS
Universitat Bielefeld
D-33615 Bielefeld
1 Introduction
The main purpose of this paper is to give a short account of two mathematical
models dealing with the HIV dynamics. Details and proofs of these results are
presented in [1], [2], [3], [8]. As we~l known AIDS epidemics has particulari-
ties which made it quite different from previous epidemics. First there are wide
variations in human sexual behaviour (heterogeneity both of the individuals and
the contacts they make). Secondly the time-lag between infection by HIV and
the appearance of overt AIDS is both long and variable (median roughly about
10 years for homosexual men). Thirdly the degree of infectiousness is also time
dependent. To include all these facts discrete stochastic models are really neces-
sary and needed. For this reason we have proposed and developed in recent years
in Bielefeld a new type of stochastic modelling based on Random Graphs which
allows both precise mathematical predictions and better adaptation to the available
behavioural, epidemiological and biological data. This model will be presented in
Section II.
Despite intensive study a lot of uncertainty still surrounds the interaction
mechanism between immune system andHIV-viruses in infected patients. Many
different hypotheses have been discussed and suggested to explain the decay of the
immune system and the induced increased susceptibility of HIV infected patients
to opportunistic infections. In [4] M. Nowak, R. Anderson and R. May use a simple
differential equation model to try to explain the long latent period on the basis of
the extreme sequence diversity of HIY. According to the model, AIDS is the result
of an increase in the number of genetically distinct virus strains throughout the
incubation period. The authors show that when the number n of variants exceeds
a critical value nc one observes a dramatic increase of the total number of viruses
and the collapse of the host immune system. In other words according to this
model, HIV infection is an evolutionary process that determines the time scale
from infection to disease.
A mathematical model describing how infection agents can persist in reser-
voirs of latently infected cells, the larger of which being the reservoir of follicular
dendritic cells (FDC) and showi!1g why progression to AIDS occurs in a sudden
transition will be presented in section III. With hindsight it is not surprising that
the lymph nodes should be sites at which the presence of HIV is most important;
Modelling HIVIAlDS Dynamics 325
that, after all, is where most T-cells usually reside. The conclusion must be that
the apparently latent period between infection with HIV and the overt symptomes
of AIDS is not clinical latency but rather a period in which the replication of the
virus proceeds apace in the lymph nodes or some other tissues of the immune sys-
tem (the spleen, adenoid glands, . .. ). It is clear that each model of the interaction
of the HIV with parts of the immune system is a very simple caricature of the true
tremenduous complexity but it can help to reduce the ignorance about how HIV
causes the diseases and for this reason mathematical modelling can playa serious
role in the advancement of the knowledge of the pathogenesis of HIV infection.
Let us mention briefly a third very interesting aspect of HIV dynamics namely
the study of the evolutionary history of HIV and AIDS . How old is HIV? Where
does the HIV virus come from? Using the new method of statistical geometry in
sequence space developed by M. Eigen, R. Winkler-Oswatisch and A. Dress [5],
M. Eigen and K. Nieselt-Struwe obtain answers to questions of age and kinship
of HIV by comparative sequence analysis of the viral genomes or parts thereof,
concluding that the present form of AIDS is probably not older than 50-100
years [6].
2 The Spread of HIV on Random Graphs

The main use of mathematics in epidemiology is to understand the relative impor-
tance of factors influencing the spread of a disease. A topical and central question
is to obtain insight into the relation between mechanics operating on the individual
level and phenomena resulting on the level of population. The last decade has seen
enormous interest in the mathematical modelling of the spread of AIDS. The wast
majority of papers have been concerned with deterministic models.
The basic probability space we consider in our model is a space of random
graphs encoding mathematically a set of behavioural and demographical properties
of a given community. Let V = {Xl , ... ,XN} be the set of vertices (individuals)
which split into several relevant groups Vi (expressing e.g. the risk behaviour and
the heterogenity of the population), V = UiE1 Vi. Since edges between vertices
describe a sexual contact, the basic given data generating the underlying graphs are
the degree distributions inside the Vi as well as the coupling between pairs (Vi , \.-})
of subgroups of the vertices. All graphs G which satisfy these given distributions
of contact rates are the elements of the random graph space 9 which become a
probability space by giving each element G E 9 the same probability peG) = Ibl'
where I ... I denotes the total number of elements.
To describe the epidemics dynamics we introduce on each graph with vertex
set V = {Xi hEI a sequence {<p }nEN, <Pn : V -+ V of pairing (matching) functions.
The <Pn tells us at time n the individuals which are potentially able to propagate
an «infection» by having a sexual contact i.e. <pn(x) represents if <Pn(x) =I- X the
sexual partner of X at time n. Now we tum to the description of the spread of the
326 Blanchard
virus. First of all for all n E N we define a mapping
Xn : V--+{O,I, ·· · , k} kEN
Xn(Xi) = £ being the function which assigns to individual xi its state of health £
at time n. We consider for instance the following progression: Xn(x) = 0 {:} x is
healthy at time n, Xn(x) = I +-+ x is HIV-infected at time n and not infectious,
Xn(x) = 2 +-+ x is HIV-infected at time n and infectious, .. . . . .
With this in mind the dynamics of the model can be described for each x
using the set of transition probabilities
P . [Xn+1 = ilXn = j]
for i, j = 0, . .. , k. These conditional probabilities express the progression of the

disease.
The simplest model of this class is generated by independent matchings and
can be interpreted as describing a homosexual community V of constant size and
divided in p age groups Vi of size N . At each iteration step n = 1,2,· · · we
allow complete independent matchings between the individuals. Furthermore each
iteration is connected with a shift of the elements of Vi into Vi+1 (growing old
of the population) and with the choice of a sexual partner. The model has only
two states: 0 = healthy, 1 = infected. The probability of infection per partnership
(i.e. per contact) is assumed to be , , 0 ::; , ::; 1. At each time n = 0, 1,2, .. . the
process will be in one of the (N + l)P possible states corresponding to the number
of infected individuals in each group, i.e. by a random vector (Ir, ... , Ip) = In and
the matrix of all transition probabilities P [In+ 1 = . .. II" = ... ] can be explicitely
given. Moreover the asymptotic behaviour of this Markov chain can be analysed
[9] . For this model we can introduce a discrete dynamical system for the iterated
expectations of the number of infected individuals, which leads for large N to
a good approximation. The type of the globally attractive stationary solutions
depends on the value of a critical parameter R which is nearly equal to ,),(p; I) . This
threshold parameter R ~ 1P;1 has a epidemiological as well as a graph theoretical
interpretation. R is on the one side nothing else than the basic reproduction ratio
giving the expected number of secondary cases generated by a typical individual
during its infectious period.
For R > 1 there exists therefore a unique non trivial globally attractive
endemic state. On the other side the initial degree of our graphs was p. Let us
know introduce an effective degree by
p-l
dett = '-2- + 1.
On the graphs the condition dett > 2 can be interpreted as the connectIVIty
threshold condition for the existence of a «giant» component corresponding to a
non trivial endemic state.
Modelling HIVIAIDS Dynamics 327
For independent matching models in the limit where the population size N ---+
+00 it can be shown that the effective graph restricted to any finite subset of
vertices is almost surely a tree and that the epidemic dynamics can be described
by a branching process with time replaced by graph distance. We studied also
in the framework of independent matching models an age structured population
with age dependent choice of the partners and non constant infectivity. In the limit
N ---+ +00 the critical parameters are obtained by studying a (p-l )-type branching
process (p being the number of age groups). The overcritical case corresponds to
the situation where Amax > 1, Amax being the largest eigenvalue of the transition
matrix for the corresponding (p - 1)-type branching process. Amax = 1 defines
the critical manifold in the parameters space (transmission probability "(, length
of partnership, age dependent coupling matrix, time dependent infectivity). Our
result shows that time dependent infectivity can indeed correspond to the worst
case in spread of HIY. We refer to [3], [13] for a more detailed discussion.
The principal advantage of this modelling approach is, of course, to have
incorporated in the mathematical framework an object representing the network
of sexual partnerships in its important structural properties. This property distin-
guishes the random network approach to epidemiological modelling and standard
differential equation models and explain their differing predictions for the course
of the epidemics at least for certain sUbpopulations. The model simulates single
individuals, making it particularly well adapted to study certain prevention sce-
narios (condom use, contact tracing, change of behaviour), to include different
characteristic properties for single individuals (like predisposition, cofactors) as
well as time dependent infectivity and different virus mutants.
In [2] we studied a reference scenario for which the graph consists of 100.000
vertices, divided into 5 groups, corresponding respectively to homosexual males,
bisexual males, heterosexual males, heterosexual females without and with contact
to bisexual males. 4% of the males are homo/bisexual. Partner numbers integrated
over a period of one year are Poisson distributed with an average of 5 for the
homosexuals and 1.22 for heterosexuals, corresponding to 83% living in monog-
amous relationships with an average of 8 life time partners.
We use for our discrete process a time step of length one day and in each
time step only one edge is activated per vertex. The epidemic dynamics inside the
heterosexual population follows a different pattern inside the homosexual popula-
tion, and this difference persists over time. The homosexual epidemics shows the
familiar pattern of an exponential increase turning into saturation. The number of
heterosexual cases, apart from a very short transient period, raises linearly with
time. Due to the low number of partner the graph inside the effective heterosexual
population consists almost exclusively of linear or almost linear structures like
chains or small trees with few branchings, which implies that the epidemics in
the heterosexual part of the population is undercritical. It is very important to
realize that HIV-infection cannot persist inside the heterosexual popUlation as a
self-sustaining epidemics for sets of parameters typical for North European so-
cieties without constant influx from those subgroups where it is overcritical or
328 Blanchard
endemic. But on the other hand even if the heterosexual epidemic is undercritical
we should be aware of the possibility of large number of cases from purely het-
erosexual transmission, since slight modifications in the assumptions about trans-
mission probabilities will increase the absolute number of heterosexual cases to
or beyond the level of homosexual cases in the long term. In comparison with
differential equation models the spread of HIV from the highly promiscuous part
of the homosexual population into the heterosexual population with low partner
numbers is considerably slowed down by the discrete properties of the interface
between bisexual man and the heterosexual population. For more details the reader
is referred to [1], [2].
3 Follicular Dentritic Cells as Major HIV Reservoir

HIV infection causes progressive depletion of CD4/HIV receptor positive T helper
lymphocytes, ultimatively leading to AIDS. The major HIV reservoir of T-helper
cells infection in lymphoid tissues however has remained poorly defined. In [7]
H. Spiegel et al. have identified the FDC cells as a major reservoir for HIV-l in
lymphoid tissues facilitating infection of C D4 + -T-helper cells. Since T-helper cells
travel through the FDC meshwork during the migration within lymphoreticular
tissues, it appears likely that HIV-replicating T cells may infect FDC, which then
infect now T cells, thus causing a gradual dissemination of the virus to cell FDC
and thereby a steadily increasing infection of T-helper/memory cells. This results
in C D4 + -T cell depletion and ultimatively in immunodeficiency. Two other groups
have also shown that the long and variable latency period after HIV infection is
explained by replication of virus in the lymph nodes [11], [12].
The FDC cells can be directly infected by HIV and can stay infected for a
long time, corresponding to a immunological silent way to transmit the virus. We
start by formulating a mathematical model in the framework of interacting particle
systems. Consider an finite lattice 0 C Zd , d = 2, 3, whose sites w E 0 are
occupied by cells (representing the FDC). At each discrete time tEN i particles
(representing the T4 lymphocytes) are emitted by the source with probability Pi
and each particle is supposed to undergo a random walk in 0 up to the (random)
time it reaches the boundary 80 of 0 and leaves O. Let T be the mean sojourn
time of a particle in O. Each site of 0 is therefore occupied by 1 or 2 particles;
so if two particles simultaneously make the site occupied a cell-particle contact
takes place. We consider only two possible states {O, I} for the particles and
the cells. The states 0 and 1 have the following meanings: 0 = healthy, 1 =
infected. With Ao we denote the set of cells infected at time 0 and with At the
set of infected cells at time t. Infection can only take place during a contact cell-
particle. Let a be the infection probability by a contact infected cell-particle and (3
the infection probability by a contact infected particle-cell. Our attention focuses
on the following problems:
Modelling HIVIAlDS Dynamics 329
1) Describe the asymptotic behaviour of IAI I for t --) +00 starting from a finite
set IAol of infected cells.
2) Study the behaviour of PI, PI being the conditional probability that a particle
entering healthy in n at time t will leave n infected.
The main results says that there is a critical time te == te ((X, f3, T) so that each
particle entering healthy in n will leave n with probability one
healthy for t - T < te
infected for t > te.
In other words PI obeys almost a 0 - 1 law and for t = te «percolation
occurs» and the system undergoes a phase transition. The occurence of a «critical
phenomena» is central to the appeal of percolation. The transition time t.t for
(X ~ f3 is much smaller that te i.e. t.t « te. For (X > f3 if we suppose that te ~
years then t.t ~ months. At percolation time te the critical density Pc = 'f~,' of
infected cells is nearly zero and for t > te PI can remain low. These results can
give an explanation for the very long incubation time by HIV infection. Indeed
intuitively the results above say that the FDC meshwork behaves like a switch:
In the subcritical phase when t < te the T4 lymphocytes leave the lymphe nodes
n with very probability non infected while in the supercritical phase when t > tc
a very high proportion of the T4 lymphocytes becomes infected in the meshwork
of FDC cells.
Acknowledgements:
This work has profited from collaborations with G. Bolz, T. KrUger and B. Voigt
and discussions with S. Albeverio, K. Dietz, A. Dress and H. Heesterbeck. I wish
to thank them for their invaluable contributions.
References
[1] Ph. Blanchard, G.F. Bolz, T. KrUger, Modelling AIDS-Epidemics or any
venereal diseases on random graphs, in «Stochastic Processes in Epidemic
Theory», Eds. J.P. Gabriel, C. Lefevre, P. Picard, Lecture Notes in Biomath-
ematics 86, Springer 1990.
[2] G. Bolz, Ph. Blanchard, T. KrUger, Stochastic Modelling on Random Graphs
of the Spread of Sexually transmitted diseases, in «Progresses in AIDS-
Research in the Federal Republic of Germany», Proceedings of the 2nd
Statusseminar, BMFT Research Program on AIDS, Ed. M. Schauzu, MMV-
Verlag MUnchen 1990.
[3] Ph. Blanchard, G. Bolz, T. KrUger, Modelling contact structure and the dy-
namics of HIV/AIDS on random graphs, to appear in Mathematical Bio-
sciences.
330 Blanchard
[4] M. Nowak, R. Anderson, R. May, The evolutionary dynamics of HIV-l qua-

sispecies and the development of immunodeficiency diseases, AIDS (1990)
4 1095-1103.
[5] M. Eigen, R. Winkler-Oswatisch, A. Dress, Statistical geometry in sequence
space: A method of comparative sequence analysis, Proc. Nat!. Acad. Sci.
USA 85 5913-5917 (1988).
[6] M. Eigen, K. Nieselt-Struwe, How old is the immunodeficiency virus? AIDS
19904 (Supp!. 1) 85-93.
[7] H. Spiegel, H. Herbst, G. Niedobitek, H.D. Foss, H. Stein, Follicular Dendritic
Cells Are a Major Reservoir for Human Immunodeficiency Virus Type 1 in
Lymphoid Tissues Facilitating Infection of CD4+T -helper cells, American
Journal of Pathology 140 15-22 (1992).
[8] Ph. Blanchard, T. KrUger, B. Voigt, A stochastic model for the HIV-infection
of follicular dendritic cells, BiBoS-Preprint 1993.
[9] S. Albeverio, Th. Stahlmann, Discrete Stochastic Processes on Random
Graphs and the Spread of HIV, BiBoS-Preprint 525/92.
[10] G. Bolz, Analyzing the Structure of Contact Graphs used in Epidemiolog-
ical Models of AIDS, in «Dynamics of Complex and Irregular Systems»,
Eds. Ph. Blanchard, L. Streit, M. Sirugue-Collin, D. Testard, World Scientific
(1993) 264-277.
[11] G. Pantaleo, C. Graziosi, J. F. Demarest, L. Butini, M. Moutroni, C. H. Fox,
J. M. Orenstein, D. P. Kotler, A. S. Fanci, HIV infection is active and progres-
sive in lymphoid tissues during the clinically latent stage of disease, Nature
1993, 362, 355-358.
[12] J. Embretson, M. Zupanzic, J. L. Ribas, A. Burke, P. Racz, K. Tenner-
Racz, A. T. Haase, Massive covert infection of helper T lymphocites and
macrophages by HIV during the incubation period of AIDS, Nature 1993,
362, 359-362.
[13] Ph. Blanchard, T. KrUger, Independent matching models: The worst case in
a model for HIV-infection with age structured population, pair formation and
non-constant infectivity, BiBoS Preprint 1993.
Morphological Diagnosis Turns from
Gestalt to Geometry
Vittorio Pesce Delfino, Teresa Lettini,

Michele Troia, and Eligio Vacca
Consorzio di Ricerca Digamma
Universita di Bari, Italy
In 1917, W. D'Arcy Thompson had posed the question of size-shape rela-

tionship in terms of «Growth and Form» by linking the size-shape problem to
the following two fundamental concepts: on the one hand, that is possible to base
morphological classifications on a non dimensional type of mathematics (analyt-
ical functions), on the other hand, it can be postulated that this approach would
allow the verification of the «principle of discontinuity», with particular relevance
if this principle occurs in morphogenesis. It is by a Sistema Naturae rethinking
D' Arcy Thompson (the «great» D' Arcy Thompson, in our opinion), and not by
more popular D' Arcy Thompson's proposal of coordinate trasformation method
which is, actually, umealable and uneffective because it uses discrete landmarks
that require to assign an arbitrary meaning of biological homology to the given
landmarks, so loosing information on intervals among landmarks.
On this legacy, the problem of allometric finding and shape measurement
became particularly relevant. There was a full awareness that differential growths,
although per se purely quantitative variations, because they are responsible for
allometric changes, are still the fundamental mechanism for shape differentiation
in biological structures. The problem became so relevant that R. Holloway (1981)
strongly but efficaciously said «measurements such as lenght, width, height, wheter
in chords or arcs, only describe space ... and further run into the abyss of allometric
correction ... If additional information (shape?) to size is expected, some methods
of allometric correction must be used». .
Measures of development length, both along straight line segments and along
boundaries however irregular, together with the definition of areas and angles,
represent the primary data in size evaluation. We must fear redundancy and other
elements of confusion which may be introduced by use of derived fractions, usually
calculated in a great number, but may be submitted to criticism:
in the first place any index between the two primary measures does not
obviously increase the information already available;
moreover, the derived fractions, typically indexes or percentage, and which
are the results of a division, tend to make uncontrollable and hardly inter-
pretable the behaviour of the relative numerators and denominators;
332 Pesce Delfino et al.
finally, the derived fractions cannot be considered as independent on the

measures used to calculate them and, therefore, their overall use for statistical
comparisons or discriminations is not advisable.
Attempts to obtain deterministic valuations of volume beginning from sin-
gle determinations of surface are frequently misleading; in fact in these cases a
reference to regular solids considered as sufficiently approximating to the biolog-
ical structure examined, whose real irregular shape cannot be submitted to such
a reduction, is inevitable. It is also necessary to pay attention to the fact that, in
general terms, the difference between the linear and surface measurements is not
merely computational, but can be conditioned by intrinsic characteristics of the
investigated components.
As overall example: the sum of the perimeters of the transverse sections of
two blood vessels considered after a bifurcation does not allow one to go back to
the perimeter from the original trunk section, while this is possible when the sur-
faces are being considered; such behaviour can clearly produce misleading effects
on the statistical evaluations. Also the probabilistic evaluation of volumes, starting
from a mean value of the measured object areas in a section, must be submitted
to some critical considerations: the statement, very important in stereology, that a
volume fraction may be calculated from a mean area fraction, is based on the prin-
ciple of Delesse who applied it in geology to an analogous problem of components
of sedimentary layers. The principle of Delesse assumes that in a defined object the
different components have an isotropic distribution, i.e. that they are distributed
randomly and with the same density in each part of the examined object. When this
is not the situation, as Aherne and Dunnill correctly say «morphometric methods
which have been devised for randomly disposed or isotropic elements cannot be
validly applied to anisotropic element or tissue components ... ». From a practical
point of view the inadequacy of this approach is revealed in another passage, of
the same Authors, about the anisotropy of biological structures: «a cursory knowl-
edge of anatomy will at once reveal that this (isotropy) is a property possessed
by very few structures and tissues». It means that each time that a fundamental
stereological principle is applied, the hypothesis that in different compounds the
component distribution is istropic must be tested in a convincing way. This state-
ment, apart from the difficulties in correct data handling, runs each time into a
general reflection. In geology the «morphogenetic factors» of a sediment are cer-
tainly less numerous and more simple than in biological structures; for example
a water sediment formed in undisturbed conditions gets effects only of the regu-
lating vector which is represented by gravity. For this reason different granules,
hypothetically present during the sedimentation, will be ranged according to a gra-
dient which is the function of their mass. Therefore, for the same mean value of
the mass one can expect an isotropic distribution of the granules. The distribution
of the biological structure components, on the other hand, arises from numerous
morphogenetic factors which are of different types (structural, functional, active,
passive, stable, provisory) and oppose isotropic distribution. In fact, stereology
uses widely different types of reference grid to perform the calculus of surface
From Gestalt to Geometry 333
function; some of these grids are conceived to count the oriented components (and
so not the isotropic ones). But this refers only to the optimization of the counting
procedure to define the surface fraction and does not modify in any way the limits
of the validity of valuation of the relative volume fraction.
The techniques of analytical morphometry, in our opinion, overcome the lim-
itations of discrete measurements, and are effective in describing shapes.
This approach approximates the natural logic with which we usually identify
a person known to us. We base recognition on the physiognomy of this person
- his or her dimensions, color, and, above all, the shape of constituent elements
(nose, forehead, or any other profile, narrow areas or wide regions). The trend
of anatomical profile (its shape) undoubtedly contains more information than any
other evaluation made in the same region. For example, any dimension measur-
able between landmarks on the glabella, the nasal spine, and the extremity of
the nasal bones will not tell us anything about the shape of the nose. However,
the profile between those points represents the characteristic with physiognomic
value - this forms the basis for any identification operation. It is a case of the
classic morphological problem of the relationship between «size» and «shape».
Actually, our visual perception is very effective in understanding and using the
information of «shape». Furthermore our mind is very effective in shape recog-
nition but very poor in shape description. Yet, in contrast, methods that quantify
morphology have already produced relatively simple and effective procedures for
dimensional evaluations by calling for specific solutions for numerical descriptions
of the «shape» in order to make them good parameters for multivariate statistics.
Since the «shape» of an anatomical region must be reduced to the characteristics
of the profiles that compose it and since these profiles can be fully considered as
curves, the solutions that can describe them must revert to analytical geometry.
Once we have progressed from the anatomical term «profile» to the analytical
term «curve», many descriptive quantifications are possible.
It is, however, necessary to include the algorithms in a strictly defined logical
architecture. The logical architecture of the S.A.M.(Shape Analytical Morphom-
etry) ® system, (Pesce Delfino et aI., 1990) originally developed to address the
anthropological problems of evolutionary morphology, and subsequently expanded
to numerous domains of medical morphological diagnosis, assumes that each irreg-
ular shape contains elements of two distinct logical domains: gross distortions that
interest the contour as a whole, and its minor local variations. In fact, any irregular
object is a displaying of a complex mixing of such two kinds of characters.
Thus, the information is processed separately by appropriate procedures so
as to acquire independent parameters both on the logical level and the numerical
one; the best prerequisite for multivariate statistics.
Any review of morphological reports shows the frequent occurence of terms
such as «irregular shape», «more or less regular shape», «dysmorphic», «poly-
morphic», «pleomorphic», «distorted», «asymmetric» etc. These are all terms and
statements related to the biological object shape (cells, nuclei, particles, tissues,
organs, etc.) and are considered so relevant as to be expressly mentioned in support

of the diagnosis.
However, this particular aspect at present lacks non only of adequate and eas-
ily available instruments, but also of an organised methodologic reference frame-
work. Moreover, it appears to suffer from limited, approximate and sometimes
incorrect solutions.
The S.A.M. (Shape Analytical Morphometry) approach is based on the extrac-
tions of shape-descriptive parameters, by using procedures which do not require
any previous information about the object to be studied in an image and which
supply numerical parameters with deterministic significance, i.e. characterized by
a not ambiguous relation between the obtained numerical value, the described
morphological characteristic and the adopted algorithm. Such parameters, used for
a description able to reduce the share of subjectivity, moreover overcome the im-
possibility of teaching the logic of the diagnostic thinking and may be easily used
in multivariate statistics.
The logic of the S.A.M. architecture assumes that the relevant objects from
a histological feature are considered as closed figures to be studied in terms of
less or more stressed distortions of the whole figure, combined with more or less
marked local irregularities of the outline itself; these characteristics may reach
extreme degrees by following one another or by variously combining indentations,
extroversions, and offshoots which may have a high degree of recursiveness.
When the morphologist uses the terms as above in the diagnosis, he tries
to distinguish somehow between the two components and to evaluate them. The
aim is thus to quantify the two features independently. As already mentioned, the
methods which are traditionally available are not very effective and merely give
one very simple parameter called Form Factor (contour index, shape index or
roundness factor) that relates the perimeter length (P) to the area value (S) of the
corresponding domain by the relationship:
F F = (47rS)/p2 (1)
This parameter is quite insufficient because there are many different shapes for
which the value of this ratio is the same, and, moreover, because it does not
express in any way what proportion of the profile length is supported by great
distortions of the whole figure and what is affected by the fine contour irregu-
larities. The form factor frequently turns out to be misleading because, when the
morphometric machines produce it under this name, the user is led to insert it in
his statistical matches and he is convinced that it is just that information relative
to «shape» which he, when carrying out the observation and the non morphome-
tric evaluation of the investigated object, extracts by a logical procedure which
is practically its opposite; because the observer tries to distinguish between the
overall object shape and the local contour perturbations, while the form factor
evaluates, with all the numerical limits of a two measure fraction, the result of
their combination. As already written just above, the architecture of the software
3 6
Fig. 1
system S.A.M. supplies parameters with which it is possible to describe indepen-

dently the morphological features carrying the information about figure symmetry
and about local perturbations.
The step in which the separation of these two class scales of information oc-
curs corresponds to the synthesis of the so-called «fundamental shape». Afterwards
two separate working channels are defined to parametrize the corresponding char-
acteristics. Before the fundamental shape synthesis the dimensional normalization
(dimensional evaluations are performed separateley since the shape parameters are
typically independent of the dimensions) and standardization operations are per-
formed. The standardization concerns the detection of the point to be assumed as
8
Fig. 2
the starting point for the analytical processing of the object contour. In fact, when
investigating a nucleus or a cell contour there is no intrinsic rule for finding a
unique point which could be taken as standard; however the numerical analysis,
using equations, requires the validity of the equation itself to be considered as
definite.
Actually, since the «shape» of a district must be referred to the characteristics
of the performance of the profiles which compose it, and since these profiles can be
fully considered as curves, the solutions able to describe them must be referred to
the mathematical domain of analytical geometry (Pesce Delfino and Ricco, 1983).
The starting point is determined In an arbitrary way which is nevertheless
strictly bounded by a procedure that will be described later and that, moreover,
produces a group of parameters describing some characteristics.
This procedure requires the detection of the figure's centre of gravity since it
is necessary to apply the algorithm for the following rotation steps in relation to
that point. By the term «fundamental shape» we mean a new curve approximating
the original profile by introducing a smoothing effect.
The need for such an operation derives from the need to obtain, on the one
hand, a simplified model of the investigated shape which is easier to submit to
classificative schemes and, on the other hand, from the need to have at our disposal
a reference curve for further computation.
This procedure supplies some synthetic parameters of contour irregularities
in terms of reciprocal punctual error between the function curve and the original
profile.
The procedure operates as follows (tab. 1, figs. 1-2-3): the abscissa and or-
dinate profile point values are considered separately (tab.I, figs. 1-2-3: b,c), as
dependent variables from which to calculate the coefficients of an upper degree
polynomial (according to the least square method), type:
D
f(x) = bo + L bk Xk (2)
k= 1
where D is equation degree, x is the independent variable (each time the abscissa
and the ordinate value), bo is a constant. For each profile the method is thus
applied twice and the independent variable is, in both cases, the series of positive
integers with unitary incremental stepping and which is included between 1 and
N, ie number of points into which the profile has been subdivided (tab.I, figs.
1-2-3: d,e). The two polynomials are then merged so to obtain a new smoothed
closed curve - fundamental curve - (tab. 1, figs. 1-2-3: f) that can be matched
with the original curve (tab. 1, figs. 1-2-3: a).
This procedure has been adopted because, whatever the original profile com-
plication entity may be, two values series without any ambiguity relative to the
point coordinates and consequently the analytical procedure applicability are al-
ways obtained.
Concerning the gross distortions of whole figure, we adopted a procedure we
named S.A.E. (Shape Asymmetry Evaluator); it consists of the interpolation of the
figure with a parabola, type:
f(x) = a + bx + cx 2 (3)
and in the construction of a set consisting of a parabola arc defined each time
(which may vary from a straight line segment to a more or less convex arc) and of
a straight line segment joining its ends (chord). A barycentric rotation for a 180 0
total range with constant angular step of a sufficiently small value (lOa represents
a typical value) is being carried out. This way figure exploration is carried out and
the place of the arc/chord complex for the maximum difference of development
between the arc and the chord (maximum distortion condition) and the minimum
difference is found.
For every rotation step, development of the parabola, length, chord intercept
and angular coefficient, surface given by the arc and the chord (allometric area)
3
Fig. 3
and surface of the domain between the chord and the profile of the figure on
the side opposite to the one to which the parabola convexity is turned (isometric
area) are computed. The parabola convexity expresses, for each rotation step, the
relative prevalence of that part of the profile towards which the convexity itself is
orientated.
For such a prevalence a vectorial representation is given. If the chord and the
arc coincide, the allometric area will be equal to zero (in fact the distortions are
lacking) and the domains on the two sides of the chord will be equal.
It, typically, happens for any rotation step of the circle. Tab. 2 shows an
irregular figure (fig. 1) fitted by the parabola, its fundamental curve (fig. 3) fitted
the same way. Fig. 5 refers to a circle and fig. 6 to a circle bearing a distortion.
The isometric area represents the fraction that could be duplicated in a spec-
ular way on the opposite side of the chord and which will give a figure that may
be irregular but without distortions (figs. 4, 7).
At the end of the whole rotation the mean values, standard deviations, vari-
ability, minimum, maximum and the interval for the relationship arc/chord and for
the surface values are given.
This procedure, applied to the fundamental shape, is also used to define the
starting point of the profile on the original shape; this point is located in the
intersection of the profile itself with the chord in the condition of minimum value
of the length difference between the arc and the chord; of the two intersections
the one whose distance from the barycentre is smaller is chosen.
Concerning local contour perturbations, we used Fourier harmonic analysis.
The Fourier harmonic analysis has been put forward (Lestrel, 1974; Rhole
and Archie, 1984; Diaz et aI., 1989) as an effective procedure in form analysis.
The values used as the dependent variable in the trigonometric polynomial type:
N / 2-1
[(x) = a + L bk sin(kx) + ck cos(kx) (4)
k=1
are represented relative to a series of angular values which are ranged at equal
distance in the interval 27r, by distance of profile single points from an original
(centre of gravity or mid point of maximum diameter) and which may be decreased
by the minimum found distance. The result is represented by the Fourier sine-
cosine coefficient couple series and by values which may be obtained from these
coefficients, of amplitude and phase of each single harmonic contributor.
The procedure gives the exact matching of the investigated figure. It also
enables the synthesis of the figure itself or of its relevant patterns simply using
subsets of the harmonics.
The reported version, may, however, be applied only to the figures without
recursive trends of the profile parts in relation to the other one and without con-
vexities towards the barycentre. In different cases the procedure cannot be applied
340 Pesce Delfino et at.
Fig. 4
because it will obviously occur that the same vector radius will intersect the profile
more than once and, thus, different values of the dependent variable will result
for the same angular value of the independent variable. Recursiveness situations
are very frequent in biological objects. The S.A.M. system not only provides for
the application to figures which do not present such ambiguous situations, but
also for a solution of such conditions. As seen above, the polynomials calculated
separately for the abscissa and ordinate values, to define the fundamental shape,
eliminate any ambiguous situation; the function approximating the curve obtained
from the polynomials is being used as reference line for the description, by means
of harmonic analysis, of trends which the original values curve expresses in re-
lation to the function curve. To do that it is necessary to rectify the polynomial
function curve which becomes the reference line (zero line) and to perform the
necessary transformations of the corresponding original curve.
In this way a series of values with an irregular but periodic trend is obtained
that presents a particularly suitable format (difference graph) for submission to
harmonic analysis (tab. 1, figs. 1-2-3: g).
Because the analysis is made on the contour distances from the fundamental
shape and not on the distances of the radius vector from the barycentre, the sinu-
soidal components which describe the contour irregularities are typically stressed,
while the amplitude value of the fundamental (1st degree) harmonic is very small.
Most of the information is contained in the low degree harmonics.
The result of the harmonic analysis is typically represented by the Fourier
spectrum: a bar graph where, for all the harmonic contributors disposed in a rising
order from left to right, the sine-cosine coefficient amplitude with positive values
up and negative values down are reported.
It is typical for the Fourier harmonics to be numerically independent and
therefore the phase and amplitude values calculable by coefficients represent very
effective variables from the statistical point of view.
It is easy to draw (tab. 1) three regular shaped figures - a triangle (fig. 1),
a rectangle (fig. 2), an ellipse (fig. 3) - characterized by the same area, but with
very dissimilar shapes.
The harmonic which represents the maximum amplitude is different for per-
centage values and above all for the order for each of the analyzed shapes. There-
fore, there is no possible confusion between the graphic of Fourier coefficients
which describes the triangle and which is characterized by the maximum am-
plitude value in correspondence with the third order harmonic (fig. 4), and the
spectrum which describes the rectangle with the maximum amplitude values for
the fourth harmonic (fig.5) or for the ellipse with the maximum amplitude on the
second harmonic (fig. 6).
The information on shape differences due to the different ratios between
sides or diameters is expressed by the amplitude and phase values of the other
harmonics, so that only in the case of the equilateral triangle, the analysis of
Fig. 5
(\
I I
.o>l
U 2
Fig. 6
a)
b)
Fig. 7
Fourier recognizes the actual presence of a 3th order harmonic and only in the
case of the square, the analysis of Fourier recognizes the actual presence of a 4th
order harmonic.
Concerning the symmetry conditions, in the triangle, where this peculiarity
is extreme, there is a considerably high allometric average rate in respect to the
rectangle and to the ellipse which have similar fundamental shapes and which in
any case have the same allometric rate for the same rotation step.
It's also easy to draw figures with different shapes (<<L», «I», «T», shaped
objects), but with the same perimeter, area and so called form factor; three example
are reported (tabs. 3-4-5-6).
S.A.M. system gives the fingerprints of the three figures by combination of
parameters given from different fundamental shape, symmetry analysis and Fourier
analysis.
- «L» shaped figure (tab. 3, fig.!) shows a Fourier spectrum driven by
the second order harmonic (tab. 4, fig .I) and very pronounced asymmetry
for both the minimum value (tab. 5, fig 1) and the maximum one (tab. 6,
fig. 1).
Fig. 8
- «1» shaped figure (tab. 3, fig. 2) shows a Fourier spectrum driven by the
second order harmonic (tab. 4, fig. 2) but shows a very low asymmetry
condition in orthogonal examination (tab. 5, fig. 2; tab. 6, fig . 2).
- «T» shaped figure (tab. 3, fig. 3) shows a Fourier spectrum driven by the third
order harmonic (tab. 4, fig. 3) and symmetry exploration reveals absence of
asymmetry (tab. 5, fig. 3) for a positioning and very high asymmetry for a
90 degrees rotation (tab. 6, fig. 3).
The S.A.M. software system does solve M.e. Escher's «Circle limit IV» (tab.
7, fig. I) and «Metamorphose» (tab. 7, fig.2) tables.
In the «Circle limit IV» table, our perception is able to discriminate Angels
from Devils only by a judgement on the profile shape. Values of perimeter (450),
area (3050), abscissa projection (87), ordinate projection (81), roundness factor
Fig. 9
(.18) are equal for the two figures and, thus, completely unable to distinguish
Angels from Devils in spite of well evident shape differences.
By a gestaltic approach, sharp-cut, pointed, hooked and hard features are
referred to Devils (takete, malignant), on the contrary rounded, soft, smooth, gently
sloped features are referred to Angels (maluma, benign). This should be considered
a popular example of our current image culture; also morphologist works with
these cultural suggestions and in his judgments on the shape of biological objects
widely uses gestaltic terminology in diagnosis as already reported.
S.A.M. system discriminated very well the angel (tab. 8, fig. 1) from devil
(tab. 9, fig. 1) by order and amplitude of maximum Fourier harmonic: third order
with 6822 amplitude for angel (tab. 8, fig. 2) and eighth order with 4731 amplitude
for devil (tab. 9, fig. 2).
[J
VI XI
o
II
[]
TIl
o
IV
o
V x XV
Fig. 10
Moreover an index computed by the match of the original profiles and their
fundamental curves gives .43 value for angel and .96 for devil. The problem of
«Circle limit IV», an example for a global view of the morpho metrical analytical
procedures, applied on artificial but intriguing figures, consisted in discriminating
Angel from Devil: a «taxonomic» goal. The problem of «Metamorphose» consisted
in monitoring step by step shape changes from a regular pattern (square) to another
one (hexagon) through irregular forms: a «morphogenetic» goal. In this strip the
morphological continuous transformation ends where a wasp flies out from the
hexagon of a comb (where a morphological discontinuity is overcome by a logic
continuity).
The S.A.M. system monitors the transformation giving non ambiguous fin-
gerprint for each figure (tab. 8, figs. I-XV).
For the square (fig. 1), Fourier analysis gave a pure fourth order harmonic
spectrum, whereas for the hexagon (fig.xV) Fourier analysis gave a pure sixth
order harmonic spectrum, with smaller amplitude.
First four intermediate irregular forms (figs.I1-V) were driven by the fourth
order harmonic: modulated square patterns. Last eight intermediate irregular forms
(figs. VII-XIV) were driven by the sixth order harmonic: modulated hexagon
patterns. The turning point was located at the sixth intermediate irregular form
(fig. VI) where maximum amplitude Fourier harmonic was of fifth order.
B. Mandelbrot (1983) suggests a «Fractal Geometry of the Nature» where
fractal, among other specifications, is a shape so made that, at smaller scales, the
form exhibits self-similarity.
Whereas for artificial fractals, for instance Koch curves, the self-similarity
occurs in a deterministic and exact way, for several natural objects at smaller
scale of observation, we observe a statistical self-similarity; in other terms not the
same pattern is repeated, but the degree of the complexity (Voss, 1988).
But, subtle differences whitin the same degree of complexity could be con-
sidered the marks of the individuality and if a specific recognition is required, at
that level the crucial information lies.
It occurs, commonly, in medical morphology diagnosis where an individual
case has to be attributed to a class (the patterns related to a disease).
The geometry of nature may be or may be not fractal but, anyway, for many
purposes we need of a method to measure the geometry of nature, in most cases
trying to discover specific characters in order to describe, but also to classify and
discriminate, the objects.
Tab. 11 shows two closed Kock curves less (fig. I) and more (fig.2) iterated;
they have the same fractal dimension.
They were worked by S.A.M. system giving, form the left to right, symmetry
evaluation, fundamentral shape and Fourier spectrum; fig. 3 shows the same figure
as in fig. 1 after distortion.
$0
eo 2
3
Fig. II
Figs. 1 and 2 are symmetrical whereas in fig. 3 the parabola fitting finds
asymmetry, Fourier spectrum is driven by sixth order harmonic for all three figures,
but for fig. 1 the sixth order harmonic is together with second order and fourth
order; for fig. 2 the sixth order harmonic is together with harmonic contributors
of order twelve, eighteen and twenty-four; for the fig . 3 the driving sixth order
harmonic is together with a composite subset that expresses the modulation due
to the loss of symmetry.
Fractal dimension measures fractal objects alone; we try to give a tool able
to work all objects, including fractal ones, even distorted (but is it still a fractal?).
Acknowledgement:
The Authors wish to thank Dr. E. Ferrara, E Potente, P. Ragone and Mr. D. Vinci
for their valuable contributions to this work.
® S.A.M. (Shape Analytical Morphometry) is registered trade mark of Metamor-
phosis s.r.l. - Bari, Italy.
References
[1] Aherne W.A. and Dunnill M.S. Morphometry. Edward Arnold, 1982.
[2] Diaz G., Zuccarelli A., Pelligra I., Ghiani A., Elliptic Fourier analysis of cell
and nuclear shapes. Compo Biom. Reas 22: 405-414, 1989.
[3] Holloway, R.L., Exploring the dorsal surface of Hominoid brain endocasts
by stereoplotter and discriminant analysis. Phil. Trans. R. Soc. Lond. 292,
155-166, 1981.
[4] Lestrel P.E., Some problems in the assessment of morphological size and
shape differences. Yearbook of physical anthropology, 18, 140-162, 1974.
[5] Mandelbrot B.B., The Fractal Geometry of Nature. 2nd edition. W.H. Freeman
and Company, New York, 1983.
[6] Pesce Delfino v., Ricco R., Remarks on analytical morphometry in biology:
procedure and software illustration. Acta Stereol. 2,458-464, 1983.
[7] Pesce Delfino v., Potente E, Vacca E., Lettini T., Ragone P., Ricco R., Shape
Evaluation in Medical Image Analysis. European Microscopy and Analysis,
21-24, September, 1990.
[8] Rhole EJ. & Archie J.W., A comparison of Fourier method for the description
of wing shape in Mosquitoes (Diptera: Culicidae). Syst. Zool. 33, 302-317,
1984.
[9] Voss R.E, fractals in Nature: from characterization to simulation. In the Sci-
ence of Fractal Images. H. Peitgen and D. Saupe, Eds. Springer-Verlag, New
Yor, 1988.
Fluorescence Recovery after Photobleaching
Studied by Total Internal Reflection
Microscopy: An Experimental System for
Studies on Living Cells in Culture
Torsten Mattfeldt l ), Theo F. Nonnenmacher2),
Armin Lambacher l ), Walter G. Gli:ickle2 ) and Otto Haferkampl)
I) Institute of Pathology, Oberer Eselsberg M23, University of Ulm , 0-89069 Ulm, Germany
2) Department of Mathematical Physics, Oberer Eselsberg 025, University of Ulm, 0-89069 Ulm,
Germany
Abstract. After bleaching of cultured cells labelled with a fluorescent dye with laser light, one ob-
serves fluorescence recovery due to lateral motion of dye molecules from adjacent unbleached domains
(Fluorescence Recovery After Photobleaching, FRAP). In the usual FRAP systems, the laser beam is
directed vertically onto the cells, hence one measures the average fluorescence kinetics of all mem-
branes throughout the full depth of the cell. This experimental design was modified by an inclination of
the laser beam at an oblique angle, which was selected in such a manner that total reflection of the laser
light near the cell surface resulted (Total Internal Reflection Fluorescence Microscopy, TIRF). Despite
of T1RF, the irradiated region is bleached because a small fraction of the beam energy is transferred
into the cell, which exponentially decays to values near zero already within a penetration depth in the
order of magnitude of 50-150 nm. This combination of FRAP and TIRF for the study of living cells
is new and permits a largely selective measurement of the fluorescence kinetics in the outer cell mem-
brane, due to the selective superficial bleaching process. As the theory of the classical design is not
fully appropriate in the present context, a generalized theoretical model of FRAP was developed which
considers angular-dependent effects. Experiments on lateral mobility of the low molecular weight lipid
analogon Nile Red (9-Diethylamino-5H-benzo[a ]phenoxazin-5-on) on confluent cultures of HeLa cells
were performed at bleaching periods of 200-1000 msec with an Argon laser of wavelength 488 nm.
From the resulting data, we could quantify the speed of diffusion in terms of time for half recovery
of the fluorescence intensity, determine whether the investigated molecule was completely mobile or
partially fixed to the membrane, and explore whether its lateral mobility is due to diffusion only or
also to active transport mechanisms.
1 Introduction
According to the well-known «fluid mosaic model» of membrane structure, the
fundamental component of biomembranes is a double layer of phospholipids, in
which lipids and proteins are either freely mobile or firmly anchored [10]. In ad-
dition to the transmembranous mobility perpendicular to the membrane surface,
it was shown that intramembranous lipids and proteins can move parallel to it,
travelling within the membrane plane. This phenomenon is denoted as lateral mo-
bility, in the special case of a purely passive movement following a concentration
gradient we have lateral diffusion. The speed of lateral diffusion is closely related
to membrane fluidity . This important concept describes the viscous properties of
the membranes, indicating whether the membrane is in a more fluid or in a more
352 Mattfeldt et al.
rigid state [6,9]. Several studies have shown that lateral diffusion and membrane
fluidity play a key role in the modulation of important biological processes at the
cellular level. Receptor-mediated membrane interactions, the immunogenity of the
membrane, cellular temperature adaptation, membrane permeability, and activity
of intramembranous enzymes may be mentioned in this context [4]. Studies on
lateral diffusion are therefore of paramount interest for the investigation of the
effects of drugs and toxins at the cellular level. We have developed an experi-
mental system which allows to explore lateral diffusion largely selectively in the
plasma membrane (i.e. the outer cell membrane) in living cells in culture. The
system consists only of simple, commercially available building blocks (Argon
laser, inverse microscope, PC), which makes its assembly easily reproducible in
other laboratories. In the present contribution we describe the practical realization
of the method. The theoretical principles are given in the companion paper [5].
2 Methods
2.1 Biophysical foundations
When living cells are bleached with a laser beam after incubation with a fluorescent
dye, one observes subsequently a recovery of fluorescence in the bleached domain
of the culture plate, which results from transport of fluorescent molecules from
neighbouring regions into the irradiated area. This phenomenon is denoted as Flu-
orescence Recovery After Photobleaching (FRAP) [3]. The fluorescent molecules
are irreversibly lost due to chemical reactions during the bleaching process. There-
fore a gradient of fluorescent dye concentration to the adjacent domains is in-
duced, which can be totally or partially reversed by diffusion of intact fluorescent
molecules from unbleached regions. A mathematical analysis of the observed flu-
orescence kinetics in such experiments provides useful information on the lateral
mobility of the investigated molecule. In a classical FRAP-experiment, a vertical
laser beam traverses the full depth of the cell. The resulting fluorescence kinetics
represent the global recovery of all membranes irradiated by the laser light, i.e.
the plasma membrane, the nuclear membrane, mitochondrial membranes, and so
on. We have tried to construct an apparatus for the selective investigation of the
plasma membrane, which is the first target of drugs and toxins when these reach
the cell. This purpose was achieved by the combination of Total Internal Reflection
Fluorescence Microscopy (TIRF) with the FRAP principle, which was applied to
living cells for the first time (Fig. la,b). In a TIRF-design, the laser beam does
not traverse the whole cell, but there is total reflection at the glass/cell interface
[1,2,11]. At the area of reflection, the laser beam transfers a small amount of en-
ergy to the cell, which is propagated as an electromagnetic wave whose intensity
decays exponentially. Because of this property, the latter is denoted as the evanes-
cent wave. The energy of the evanescent wave is nevertheless sufficient to bleach
the fluorescent molecules irreversibly. As the penetration depth of the evanescent
wave is in the order of magnitude of 0.1 to 0.3 of the wavelength of the laser
light (~ 50--150 nm; see eq. (9) in the companion paper [5]), a largely selective
Fluorescence Recovery after Photobleaching 353
flcr Pholobleachin
xperimenlal y lem
Prial r
J. [I
DI.phn.m
IUNC'h n(t 100"\
I\Ju~urin. 1«:
Fig. 1 a) Schematic overview of the complete experimental system. b) Schematic drawing of the
construction by which T1RF is achieved at the interface of the coverslip and the cells.
354 Matt/eldt et al.
h'l('u,;;I~ len . .
. . . . . . . . . . ...... '" ............ '" . . . . . . . .~. . . . . . . . . . . . .~

,uhnm _ _
.......................I ..- .................................
. ~::.:~= "'!:;llon
-, ............ ,- .,,:::...",:.:::
- ........<....... r------___,
....~~ ...
<i·"
/'i'7
////
//
Fig. Ie
OI!J Ih.
t <,.
bleaching effect on the cell constituents immediately adjacent to the glass plate,
i.e. the cell membrane, is achieved.
2.2 Experimental design

A 7 mW Siemens Argon laser LGK 7800 with wavelength >'0 = 488 nm was
used as light source. This device was equipped with a diaphragm connected to an
IBM-compatible personal computer AT-286. According to the diaphragm position,
the intensity of the emitted laser beam is switched between 100% and I %. The
full intensity is used for bleaching, the weak intensity is used for measuring the
cellular fluorescence before bleaching and in the recovery period; the bleaching
effect of the attenuated beam is negligible. After passing a focusing lens, the beam
reaches a movable mirror and thereafter traverses a glass cube (Fig. la,b). This
cube is positioned on the coverslip of the sample chamber which contains the cells
incubated with the fluorescent dye [7] . The cube is separated from the coverslip
only by a thin layer of immersion oil and can be moved horizontally. The laser
beam is adjusted in such a manner that total reflection takes place at the inter-
face of the coverslip and the cells. These are bathing in culture medium, thereby
attached to the coverslip from below, hanging downwards like bats (Fig. la,b).
An inverse microscope Zeiss IM35 is focused with a 40x objective lens on the
plane where the cells adhere. The microscope transfers the fluorescence of the cells
to a photodiode, which converts the photons into electric signals. The resulting
voltage is amplified and fed into an AID-converter board which is located within
Fig. 2 a) A group of 4 HeLa cells. after incubation with NileRed, shows specific fluorescence
labelling. The cells appear as 4 bright spots whereas the adjacent medium is dark. b) In
this sample, more than 90% of the area is occupied with HeLa cells, hence nearly the whole
illuminated area shows fluorescence. The elliptical profile of the obliquely incident laser beam
is clearly visible.
the aforementioned computer. Using self-written programs in Turbo-Pascal, the

PC thus controls the laser intensity by positioning the diaphragm, measures the
fluorescence of the cells according to the voltage from the photodiode, and stores
the data on the harddisk. An intensified video camera is attached to the micro-
scope. This camera is connected to the image analysis system Kontron IBAS 2000,
where the stored microscopical image of the fluorescent cells remains available
for subsequent quantitative stereological studies.
2.3 Evaluation of measurements

When irreversible bleaching of the irradiated molecules, vertical direction of the
laser beam, pure two-dimensional diffusion of the dye, and Gaussian energy dis-
tribution within the beam profile may be surmised, the fluorescence intensity F (t)
at time t after bleaching is given by the well-known equation [3]:
(1)
where 1/ = (1 + 2t/TO )-1, F represents the initial intensity of the fluorescence

before bleaching, and K and TO are parameters which have to be estimated for
each experiment by nonlinear regression of F(t) on t. In a combined FRAPITlRF
design, however, the requirement of vertical direction of the laser beam is not met,
hence a generalized theory allowing for arbitrary oblique incident angles of the
beam is needed. The corresponding theory, presented as eq. (18) in the companion
paper [5], shows that the following equation holds for general values of"( = cos ():
(_K)n
= F L -,-{(1 + n + 2n,,(2 t / Td )(1 + n + 2nt/Td)} -1/2
00
F(t) (2)
n.
n=O
The generalized approach reduces to eq. (1) for vertical incidence as a special
case, when the angle () between the laser beam and the vertical axis is 0°. To
determine () in our experimental apparatus, the fluorescent profile of the laser beam
was stored by the image analyzer. As the true laser beam has a circular profile
of radius r, the eccentricity of the visible elliptical beam profile with semiaxes
b, r is related to () by cos () = (r /b), according to elementary trigonometry. The
actual values are r = 291Lm, b = 691Lm, hence () :::::: 65°. The area of the beam
profile at the level of the cells amounts to :::::: 62861Lm2 . Hence the irradiated area
is considerably larger than the projected area of an average cultured cell; from this
reason a bleaching experiment represents an average value for a sample of some
cells hit by the laser beam (Fig. 2a,b). The speed of diffusion can be indicated
by the time for half recovery T1/ 2 • This is the time after which one half of the
bleached fluorescence has recovered according to the fitted curve. The percentage
offluorescence recovery was determined from the ratio of the asymptotic end value
of the fitted curve to the mean intensity value before the bleaching impulse. In
addition, the initial voltage decline immediately after bleaching was recorded in
each experiment.
Volta e (mV)
4600
4000 Bleachinl .. .~
J.
3600
!
3000 !F(mln)+F(end)1!2
2600
2000
1600
1000
600
0
-6 0 6 10 16 20 26 30
T Time (sec)
1/2
Fig. 3 The diagram shows the fluorescence intensity (here measured as voltage of the photodiode)
as a function of time t. Negative values of t represent the time before bleaching, t = 0 is the
moment of bleaching. The curve was fitted by nonlinear least square regression of voltage on
t, according to eq. (2) with () = 65° .
2.4 Cell cultures and protocols

Confluent cultures of HeLa cells in RPMI 1640 medium on coverslips were used
for the studies. Nile Red, a lipid analogue of low molecular weight (9-Diethyl-
amino-5H-benzo[a]phenoxazin-5-on), which localizes largely specifically to the
hydrophobic inner domains of biomembranes [8], was used as fluorescent dye.
Incubation with the dye was begun at 37° C on the coverslips I hour before the
bleaching experiments at a concentration of 6.25 f.LI Nile Red per ml medium. The
chambers were inserted into the experimental system at room temperature (Fig.
la,b). The fluorescence was measured under steady state conditions for 3.75 sec;
after bleaching, the fluorescence recovery was measured for 26.25 sec. Thus a
single bleaching experiment lasts 30 seconds. The voltage is recorded every 75
msec, which provides a series of 400 measurements per experiment. It is possible
to bleach a constant domain of the culture repetitively for 3-4 times without
detectable changes of the fluorescence kinetics, thereafter a new domain of the
cell culture should be investigated. To obtain reliable data, at least 6, but more
often 7-10 positions were tested, each with 3 consecutive bleaching experiments.
Therefore 18-30 FRAP-curves are available in a typical study. Mean values of
initial voltage decline, % recovery, and time for half recovery were determined
between replications within positions. Finally, means and standard deviations of

these average values per position were determined.
2.5 Experiments
First the effect of the areal density of the cells, i.e. the ratio of the area covered
by cells to the total area of the laser beam profile, on the fluorescence kinetics
was explored. At different time points of culture growth the areal density of the
cells shows considerable variations, in addition there is spatial variability within a
culture at a given moment. Therefore FRAP-experiments were performed at 9 dif-
ferent positions on untreated HeLa cells. After a series of 3 replicated experiments
at each position, the image of the elliptical profile of the laser beam with the cells
inside was stored each time on disk. The areal density was evaluated interactively
by tracing the cell profiles and the laser beam profile. This was done after the
whole series of bleaching studies on a culture was completed, hence there was
no need to keep the cells longer in the sample chambers than usual. Next, we in-
tended to explore the effect of increasing bleaching periods on the FRAP kinetics.
For this purpose, 3 positions on untreated HeLa cell cultures with bleaching times
of 200, 400, 600, 800, and 1000 msec were evaluated with 3 replications each.
Bleaching times below 200 msec or above 1000 msec duration did not provide
additional information at the selected, fixed voltage amplification, as the initial
voltage decline was too small, or the initial voltage decline was so large that the
resulting voltage difference exceeded the range of values that can be measured by
the NO converter board.
3 Results
The key parameter TI/2 could be determined with good accuracy within a cell
culture; typical coefficients of variation were below 5%. The estimate of TI/2
in the order of magnitude of a few seconds corresponded to the quick lateral
mobility, which could be expected in a low molecular weight lipid analogon (Fig.
3). The kinetics of fluorescence recovery after photobleaching could be fitted by
a pure 2D diffusion process in very good approximation according to eq. (2)
(Fig. 3). Furthermore, all experiments showed 100% recovery of fluorescence
after bleaching. Increasing bleaching periods from 200 to 1000 msec caused an
increasing initial voltage decline, i.e. the initial voltage jump at t = 0 was enhanced
(correlation coefficient r = +0.71, p < 0.005) and a concomitant increase of
TI/2 (Fig. 4a,b). The positive correlation between bleaching time and TI/2 is
also statistically significant (r = +0.84, p < 0.(001). Furthermore, a positive
correlation could be detected between the areal density of the HeLa cells and the
initial voltage decline (r = +0.84, p < 0.05) (Fig. 5).
Initial voltage decline (mV)

2000r---------------------------~
1500
1000 I
I
500 I I
200 400 600 800 1000 1200

Bleaching time (msec)
Time for half recovery (msec)

3000~----------------------------~
2500
2000
I
I.
1500
1000
I
I I
500
200 400 600 600 1000 1200

Bleaching time (msec)
Fig. 4 a) The initial voltage decline (voltage decay immediately after bleaching) increases with
bleaching time. b) Mean time for half recovery also increases with bleaching time. Both plots
represent mean values with standard errors.
Initial voltage decline (m V)

3000~----------------------------~
2500
2000
I
1500
I I
1000
500
O~----~--~--~~--~----~~----~
0.0 0.2 0.4 0.6 0.8 1.0
Area of cell~/ Area of laser beam
Fig. 5 The initial voltage decline increases with the areal density of cells within the laser beam
profile. The plot represents mean values with standard errors.
4 Discussion
The finding of 100% fluorescence recovery indicates complete lateral mobility of
the lipid molecules. An «immobile fraction», which would lead to an incomplete
fluorescence recovery, could not be detected. Considerable immobile fractions
have been described for membrane proteins (e.g. [12]). Excellent agreement was
obtained when curves were fitted to the experimental data according to eq. (2).
Therefore, the whole observed lateral lipid mobility may be explained in terms of
diffusion. In other words, the data do not indicate active transport mechanisms for
the lipid in a relevant order of magnitude. An interesting finding is the dependence
of T 1/ 2 on the duration of bleaching. The effect is apparently caused by the fact
that more molecules are bleached when the laser beam is longer exposed. When
lipid fluidity and the coefficient of lateral lipid diffusion remain constant, a longer
time will subsequently be necessary until the induced gradient of concentration is
fully reversed, and thus T1/ 2 will tend to increase, too. Presumably, the increased
initial voltage jump with increasing areal density of HeLa cells is due to the same
reason: when more cell surface area containing dye molecules is exposed to the
laser beam, more molecules will be bleached, which enhances the initial voltage
decay. A potential drawback of the method, as it stands now, is the fact that
single cells are not studied selectively. Instead, global FRAP curves are obtained
for clusters of cells. On the other hand, the theory as given above is strictly
correct only for studies where only a small spot on the surface of a single cell is
bleached, although it shows such a good agreement with the experimental data.
To make further progress here, the following modifications are possible. In order
to examine the behaviour of single cells, the laser beam could be focused more
narrowly, or a central segment of the emitted beam could be evaluated selectively
with a photomultiplier after insertion of a diaphragm at the detection side of the
beam path. Alternatively, the experimental apparatus might remain as it stands,
and the theory could be further generalized by allowing for multi-cell effects
with multiple recovery compartments. From our point view, such refinements are
necessary to estimate the parameters K and T[) in absolute terms; in fact, this is
why we have restricted ourselves to a presentation of T1/2' initial voltage decline
and percentage of recovery, i.e. parameters which merely require a good curve
fitting for their estimation. In summary, we have described an experimental device
linking the principles of TlRF and FRAP, which allows accurate and reproducible
measurements of lateral mobility of membrane components in living cells. From
the resulting data, the experimenter can quantify the speed of diffusion, determine
whether the investigated molecule is completely mobile or partially fixed to the
membrane, and explore whether its lateral mobility is due to diffusion only or
also to active transport mechanisms. It should be possible to extend the method
to other interesting molecules, e.g. long-chain lipids, tunnel proteins and receptors
labelled with fluorescent dyes.
References
[1] Axelrod, D. (1981) Cell-substrate contacts illuminated by total internal re-
flection fluorescence. 1. Cell BioI. 89, 141-145.
[2] Axelrod, D. (1989) Total internal reflection fluorescence microscopy. Methods
Cell BioI. 30, 245-270.
[3] Axelrod, D., Koppel, D.E., Schlessinger, I., Elson, E., Webb, WW (1976)
Mobility measurement by analysis of fluorescence photobleaching recovery
kinetics. Biophys. 1. 16, 1055-1069.
[4] Boullier, 1.A., Melnykovych, G., Barisas, B.G. (1982) A photobleaching re-
covery study of glucocorticoid effects on lateral mobilities of a lipid analog
in S3G HeLa cell membranes. Biochim. Biophys. Acta 692, 278-286.
[5] G16ckle, WG., Mattfeldt, T., Nonnenmacher, T.F. (1993) Anomalous dif-
fusion and angle-dependency in the theory of fluorescence recovery after
photobleaching. Proceedings of the First Symposium on Fractals in Biology
and Medicine, ed. Nonnenmacher, T.F., Losa, G. and Weibel, E.R. Basel:
Birkhtiuser (this volume, in press).
[6] Lenaz, G., Castelli, G.P. (1985) Membrane fluidity: molecular basis and phys-
iological significance. In: Structure and Properties of Cell Membranes, Vol.
I, ed. Benga, G. Boca Raton: CRC Press, pp. 93-136.
[7] Pentz, S., Horler, H. (1992) A variable cell culture chamber for «open» and
«closed» cultivation, perfusion and high microscopic resolution of living cells.
J. Microsc. 167,97-103.
[8] Sackett, D.L., Wolff, J. (1987) Nile Red as a polarity-sensitive fluorescent
probe of hydrophobic protein surfaces. Analyt. Biochem. 167,228-234.
[9] Shinitzky, M. (1984) Membrane fluidity and cellular functions. In: Physiology
of Membrane Fluidity, Vol. I, ed. Shinitzky, M. Boca Raton, CRC Press, pp.
1-51.
[10] Singer, S., Nicolson, S. (1972) The fluid mosaic model of the structure of
cell membranes. Science 175,720-731.
[11] Thompson, N.L., Burghardt, T.P., Axelrod, D. (1981) Measuring surface dy-
namics of biomolecules by total internal reflection fluorescence with photo-
bleaching recovery or correlation spectroscopy. Biophys. J. 33,435-454.
[12] Vaz, W.L.C., Goodsaid-Zalduondo, F., Jacobson, K. (1984) Lateral diffusion
of lipids and proteins in bilayer membranes. FEBS Letters 174, 199-207.
Anomalous Diffusion and Angle-Dependency
in the Theory of Fluorescence Recovery
after Photobleaching
Walter G. GI6ckle l ), Torsten Mattfeldt2) and Theo F. Nonnenmacher l )
I) Department of Mathematical Physics, University of Ulm, Albert-Einstein-Allee 11, 0-89069

Ulm, Germany
2) Institute of Pathology, University of Ulm, Albert-Einstein-Allee 11,0-89069 Ulm, Germany
Abstract. Fluorescence recovery after photobleaching (FRAP) is an experimental method for the
investigation of transport processes of fluorescent molecules. Under total reflection conditions a selec-
tive measurement of processes in thin surface layers is achieved. We present theoretical calculations
demonstrating the influence of the angle under which the detection beam is directed to the probe on
the fluorescence signal. Apart from the angle dependency, we consider anomalous diffusion processes
with the mean square displacement (r2) ex: ta. Modeling the process by a fractional order diffusion
equation, we find that the relaxation of the fluorescence to the equilibrium value is asymptotically
ex: t- a for large t.
1 Introduction
The study of fluorescence recovery after photobleaching (FRAP) provides a me-
thod for measuring transport and reaction processes of fluorescent molecules in
biological systems like living cells [1][5]. In the experiment, a short, intense laser
pulse irreversibly bleaches the fluorophore in a small region. Transport coefficients
can be determined by measuring the rate of fluorescence recovery resulting from
the transport of fluorescent molecules from non-illuminated regions of the system
into the bleached spot. Under total reflection conditions, only the fluorescence
of molecules in a thin superficial layer is excited by the exponentially decaying
evanescent wave [7]. Thereby a largely selective monitoring of protein or lipid
molecules labelled with fluorescent dyes in cell membranes is achieved. By ana-
lyzing the fluorescence recovery after photobleaching, the lateral diffusion of the
molecules in the membrane can be studied.
Theoretical calculations were performed for several intensity profiles of the
bleaching and probing beam and for different reaction and diffusion processes
[1][7]. Here, we present results of the dependency of the fluorescence recovery
on the angle of illumination for a Gaussian beam profile. Furthermore we discuss
the influence of cooperative effects if the fluorescent molecules do not behave like
Brownian particles but show correlated motion. In this case, the transport process
with long-range correlations is modeled by a fractional order diffusion equation.
364 GlOckle et al.
2 General Fonnula for FRAP

The formula for the fluorescence recovery after photobleaching has to take three
processes into account: (i) the bleaching by a short laser pulse, (ii) the transport
process by which unbleached fluorescent molecules migrate into the bleached area,
and (iii) the detection of the fluorescence signal.
(i) According to ref. [1] we assume that photobleaching of the fluorescent
molecules is a simple first-order reaction process with a rate constant proportional
to the intensity of the bleaching beam /b. Then the concentration of the unbleached
fluorophore C(f, t) is governed by
:t C(f, t) = -a/b(f)C(f, t) . (1)
With the equilibrium concentration C, the concentration after a short bleaching

time T is given by
(2)
which defines the initial state, at which the fluorescence recovery starts.
(ii) By bleaching, fluorescent molecules are destroyed. Because of the con-
centration gradient, molecules from unbleached areas migrate into the bleached
spot. The most simple case of such a transport process is diffusion
:t C(f, t) = D6.C(f, t) (3)
with a constant diffusion coefficient D. If G(f, t) is the Green's function of the

transport process, the concentration of unbleached particles is given by the con-
J
volution integral
C(f, t) = df' C(f', O)G(f - 1"', t) . (4)
(iii) The property which is detected is the total fluorescence in the detection
area. It is proportional to the intensity of the probing beam and the fluorophore
concentration and, thus, given by
F(t) =Q J dfIP(f)C(f, t) (5)
where integration is performed over the whole detection area. In eq. (5), Q is a
constant containing the quantum efficiencies of the light absorption and emission.
If F (t) is considered to be the electrical signal of the detector, attenuation and
amplification of the optical and electrical systems are also included in Q.
Collecting the three components (2), (4), and (5), we arrive at
F(t) = QC J Jdf' df exp ( -aT/b(f')) G(f - 1"', t)/P(f) (6)
which is valid for arbitrary transport processes.

3 Fluorescence recovery and relaxation function

In total internal reflection measurements the laser beam is directed from a medium
2 (glass) with refraction index n2 to a medium 1 (cell) with the index nl < n2. If
the angle () between the direction of the beam and the optical axis (z direction) is
greater than the angle of total reflection
()e = arcsin (:~) (7)
the beam is totally reflected. In medium 1 only an exponentially decaying evanes-

cent wave with the intensity
g18I(z) = 1(0) exp( -zjd) (8)
penetrates. The penetration depth
>'0 ( 2 sm
. 2 2) -1 /2
d = 47r n2 (()) - nl (9)
depends on the angle () and the vacuum wavelength >'0. For typical values nl =
1.33 (H20) and n2 = 1.5 (glass) the angle of total reflection is ()e ~ 62.5°. For
values of () more than about 5 degrees above ()e, the value of the penetration
depth d is nearly constant between 0.1>'0 and 0.2>'0. Fluorescence is therefore
only excited in a thin surface layer. If the fluorescent molecules are solved in lipid
membranes, molecules in the outer cell membrane are largely specifically detected.
Furthermore the mobility of the molecules in the lipid membrane is restricted to
lateral processes, i.e. G(f, t) = G (x, y, t), and thus the method provides a selective
tool to study lateral transport in cell membranes. In the thin layer the dependency
of the beam intensity (8) on z can be neglected.
The intensity of a laser beam is well described by a Gaussian beam profile
J. In the (x, g)-plane perpendicular to the beam direction, the intensity is given by
I(x ,y) = 10 exp ( -

2(x2 +
w2
f)) = IoJ(x,y) . (10)
In the (x, y)-system of the membrane we have
2(-lx2 + yZ))
I(x,y) = '"Ylo exp ( - w2 = '"Ylolbx , y) (11)
with the geometric factor '"Y = cos(()). Hence, if lateral processes in the outer cell
membrane are measured by FRAP under total internal reflection conditions, the
fluorescence recovery function (6) takes the form
F(t) = QClb'"Y J J
dxdy dx'dy'e-KJhxl ,y')G(x-x',y-y',t)lbx,y). (12)
366 GLOckle et al.
In eq. (12), the bleaching parameter K = aT-yI8 is introduced which is a measure

for the strength of bleaching.
An alternative representation considers the relaxation function G(t) = F -
F(t) describing the approach of F(t) to the equilibrium value F = F(t ---+ 00) .
Using the Fourier transform
G(kx , ky, t) = Jdxdyei(kxx+kYY)G(x , y, t) (13)
of the Green's function G(x,y), then
G(t) = -QCIg-y (2:)2 J dkxdkyG((k x , ky) JdxdyJ(-yx ,y)e-ikxx-ikyy

(14)
is obtained from eq. (12). For a Gaussian beam profile J, we find
G(t) = -QIPC~ ~ (_K)n ~

o 16-y~ n! n
n=1
(15)
x J (
dkxdkyG(kx, ky , t)exp - (n+8n l)w2(k;-y2 + ky2))
by series expansion ofexp(-KJbx',y')) and integration.
4 Ordinary diffusion
Now we can use models of the lateral transport process in the membrane. The
normal diffusion equation (3) in the (x,y)-plane has the Green's function
(16)
in the Fourier space and
7r + f)
(X24Dt
G(x , y,t) = Dt exp - (17)
in the ordinary (physical) space, resp. From eq. (12) the fluorescence recovery
(_K)n
F(t) = F L ----;t!((1
00
+ n + 2n-y2tj d)(1 + n + 2ntj d))-1 /2
T T (18)
n=O
I I
4000
F(t) ... .. .....
3000 -
2000 -
1000 -
o I I
o 5 10
t []
Fig. 1 Data points from a typical FRAP experiment [5) fitted by (18) with K = 0.47, Td = 0.41s,
F = 3464, I = 1 (fJ = 0°) as well as by (18) with K = 0.49, Td = 0.21s, F = 3470, I = 0.5
(fJ = 60°); the curves are not distinguishable.
is found in terms of a convergent series expansion. In eq. (18) .
- 1 2 P -
F= -7TW IoQC (19)
2
is the equilibrium fluorescence value, Td = w 2/(4D) is a characteristic time con-

stant, and K describes the strength of bleaching. In , = cos( 0) the dependency
on the angle of incidence of the beam is included. For a perpendicular incident
beam (, = 1), the sum can be expressed in closed form, e.g. by the x2-probability
function
F(t) = FvK- lI f(v)X 2 (2KI2v) (20)
with v = w 2 /(w 2 + 8Dt) [1]. The relaxation function is given by
G(t) = -F f (_;)n
n=l
((1 + n + 2n,2 t / Td )(1 + n + 2nt/Td)) -1/2 (21)
where for K < 1 the first term dominates.

A typical experimental result [5] is shown in Fig. 1. For negative times we
have constant fluorescence. By the bleaching at time zero the fluorescence becomes
smaller and afterwards it tends to the original value because of the diffusion of
368 GlOckle et al.
G(t)/G o
I
8 = 60°
45°
o.
30°
0°
10
t/Td
Fig. 2 Relaxation function (22) for various values of the angle (J under which the laser beam is
directed to the probe h = cos (J).
unbleached molecules into the detection area. Experimental data sets of FRAP
measurements under total internal reflection conditions are well represented by
eq. (18) with the given angle () ~ 65°. In Fig. 1 the data points were fitted with
eq. (18) corresponding to () = 0° and () = 60°. The two curves are scarcely to
distinguish. They fit the data points within the accuracy of the measurement, but
the time constants and thus the diffusion coefficients differ by a factor 2. Fig. 2
shows the dependency on the angle () of the normalized relaxation function. One
recognizes that with increasing () the relaxation decreases.
5 Anomalous diffusion
Diffusion processes in systems with long-range correlations or in 'disordered me-
dia' deviate from the standard diffusion (3) [3][4]. For anomalous diffusion the
mean square displacement is given by
2 Ej2 -
(r ) = - 8pG(k,t)lk=o ex: t
Q
(22)
(0: f=. 1) instead of the result (r2) = 2Dt of eq. (3). An equation leading to such
a behavior is the fractional diffusion equation [6]
C(f, t) - C(f,O) = D oD;q 6.C(f, t) (23)

q = 1.0
F(t)/ F 0.7
0.3
0.8
0 . 75
0.7-1------t------:"1':-0-----:1.,..5-----::<~0
Fig. 3 F10urescence recovery for anomalous diffusion (q = 0.3 and q = 0.7) compared with normal
diffusion (q = \.0) for K = 0.5 (only the n = 1 term is considered).
where the operator oD;q is a fractional Liouville-Riemann integral operator of

the order q (0 < q < 1) defined by
1
aD;q = r(q) it
a (t - t,)q- If(t')dt' . (24)
The Green's function, Fourier transformed in the space coordinates and Laplace
transformed in the time, takes the form
-I
G(kx, ky,p) = 1 + Dp!q(k; + k~) (25)
leading to the exponent ex = q in eq. (22).

With eq. (25), the relaxation function
G(t) = -F ~ ~ (_f<)n ~H2, 1 (Tn ,q I (1, l/q), (1, 1)) (26)

"( ~ (n + I)! q 2,2 t (1, l/q), (1, l/q)
n=1
is obtained in terms of the Fox function Hi'i

by applying Laplace-Mellin transform
techniques [2]. The time constants Tn ,q ar~ given by
_ (w2(n +
Tn ,q - SDn{3
1)) l/q (27)
370 GLOckle et al.
log(G(t)jGo)
Fig. 4 Asymptotic behavior of the fractional relaxation function (27) for various values of q in the
case of K < I (i.e. only the n = 1 term is considered).
and (3 is a numerical constant «(3 E [1, Ih 2 ]) which is (3 = 1 for "I = 1. For the
H-functions series and asymptotic expansions can be derived [2]. Especially in
the case (26) we find
G(t)",/(3f (_K)n ('Ye+QW(I-q)+qlOg(rn,qjt)

"I n=l (n+ I)! r(I-Q)
c n,qr+ o (t-2q))
t
(28)
for large t. Here, "Ie is the Euler constant and W is the logarithmic derivation of
the r-function. In Fig. 3, the recovery function F(t)jF in the cases of anomalous
diffusion (Q = 0.3 and Q = 0.7) is compared with the result for the ordinary
diffusion (Q = 1.0). It demonstrates that the value of Q determines the velocity
with which the plateau value F is reached. The relaxation function (26) is plotted
for various values of the fractal parameter Q in Fig. 4. In the log-log plot one
recognizes the power law decay G(t) rv t-q for anomalous diffusion and G(t) rv
t- I for normal diffusion (Q = 1). Especially for large protein molecules we expect
the occurrence of diffusive transport with long-range correlations leading to a
relaxation function of FRAP of the type (26).
References
[1] Axelrod, D., Koppel, D.E., Shlessinger, 1., Elson, E., Webb, W.W. (1976)
Mobility measurement by analysis of fluorescence photobleaching recovery
kinetics. Biophys. J. 16, 1055-1069.
[2] Gli:ickle W.G., Nonnenmacher, T.F. (1993) Fox function representation on
non-Debye relaxation processes. J. Stat. Phys. 71, 741-757.
[3] Havlin, S., Ben-Avraham, D. (1987) Diffusion in disordered media. Adv.
Phys. 36,695-798.
[4] M. Isichenko, B. (1992) Percolation, statistical topography, and transport in
random media. Rev. Mod. Phys. 64, 961-1043.
[5] Mattfeldt, T., Nonnenmacher, T.F., Lambacher, A., Gli:ickle, W.G., Haferkamp,
0. (1993) Fluorescence recovery after photobleaching studied by total internal
reflection microscopy, in this volume.
[6] Schneider, W.R., Wyss, W. (1989) Fractional diffusion and wave equations.
J. Math. Phys. 30, 134-144.
[7] Thompson, N.L., Burghardt, N.P., Axelrod, D. (1981) Measuring surface dy-
namics of biomolecules by total internal reflection fluorescence with photo-
bleaching recovery or correlation spectroscopy. Biophys. J. 33, 435-454.
List of Speakers
Peter Achermann Curtis B. Caldwell

Institute of Pharmacology Departments of Medical
University of ZUrich Biophysics and Radiology
Winterthurstr. 190 University of Toronto and
CH-8057 ZURICH Department of Radiological Sciences
Switzerland Sunnybrook Health Science Centre
2075 Bayview Avenue
North York, Ontario
Werner Arber Canada M4N 3M5
Abt. Mikrobiologie
Biozentrum Giuseppe Damiani
Universitat Basel Universita di Pavia
KlingelbergstraBe 70 Dipartimento di
CH-4056 BASEL Genetica e Microbiologia
Switzerland Via Abbiategrasso 207
1-27100 PAVIA
Italy
Gerd Baumann
Carl 1. G. Evertsz
Abteilung flir
Center for Complex
Mathematische Physik
Systems & Visualisation
Universitat Ulm
University Bremen
Albert Einstein Allee 11
BibliothekstraBe I
D-89069 ULM
Postfach 330440
Germany
D-28334 BREMEN
Germany
C.L. Benhamou Zeno F61des-Papp

CHR Orleans-La Source Section for Polymers
B.P.6709 Universitat Ulm
F-4507 ORLEANS-Cedex 2 Albert-Einstein-Allee 11
France D-89069 ULM
Germany
Philippe Blanchard Walter GlOckle

Theoretical Physics and BiBoS Department of
Universitat Bielefeld Mathematical Physics
UniversitatsstraBe 25 University of Ulm
Postfach 8640 Albert-Einstein-Allee 11
D-33615 BIELEFELD D-89069 ULM
Germany Germany
374 List of Speakers
Ary L. Goldberger Gabriele A. Losa

Department of Cardiology Laboratorio Patologia Cellulare
Harvard Medical School Istituto Cantonale di Patologia
Beth Israel Hospital Via in Selva 4
330 Brookline Ave 6604 LOCARNO
BOSTON, MA 02215 Switzerland
USA
Martin Liineburg
Ricardo Gutfraind Universitat Bielefeld
Group Matiere Fakultat fUr Mathematik
Condensee et Materiaux UniversitatsstraBe 25
Batiment 11 B 0-33615 BIELEFELD 1
Universite Rennes I Germany
35042 RENNES Cedex
France Benoit B. Mandelbrot
IBM Research Division
Hiroko Kitaoka Thomas J. Watson Research Center
Department of Internal Medicine P.O.Box 218
Kitaoka Hospital YORKTOWN HEIGHTS
1031 Meijimachi N.Y. 10598 USA
Kurayoshi Tottori 632
Japan Raphael Marcelpoil
Equipe de Reconnaissance des
Klaus-D. Kniffki Formes et Microscopie Quantitative
Institute of Physiology Universite Joseph Fourier
University of Wiirzburg CERMO
Rontgenring 9 BP 53X.38041
D-97070 WORZBURG GRENOBLE Cedex
Germany France
Haymo Kurz Shu Matsuura

Image Analysis and Tokai University
Stereology Laboratory School of High Technology
Anatomisches Institut II for Human Welfare
AlbertstraBe 17 Numazu
0-79001 FREIBURG Nishino 317
Germany SHIZUOKA 410-03
Japan
Gabriele Landini
Oral Pathology Unit Torsten Mattfeldt
The Dental School Institute of Pathology
University of Birmingham University VIm
St. Chad's Queensway Oberer Eselberg M23
BIRMINHGHAM B4 6NN 0-89069 ULM
England UK Germany
List of Speakers 375
Theo F. Nonnenmacher Manfred Sernetz

Abteilung fUr Institut fUr Biochernie
Mathematische Physik und Endokrinologie
Albert-Einstein-Allee 11 Justus-Liebig-Universiilit
UniversiHit Ulm FrankfurterstraBe 100
D-89069 ULM D-35392 GIESSEN
Germany Germany
Tom G. Smith, Jr.

Laboratory of Neurophysiology
Vittorio Pesce Delfino National Institute of Health
Consortio di Ricera DIGAMMA BETHESDA, MD 20892
C.So.A. de Gasperi 449/a USA
1-70125 BARI
Italy Petre Tautu
German Cancer Research Center
Department of Mathematical Models
Neuenheimer Feld 280
Fan~ois Rothen D-69009 HEIDELBERG
Institut de Physique Experimentale Germany
Universite de Lausanne
BSP Dorigny Ewald R. Weibel
CH-1015 LAUSANNE Anatomisches Institut
Switzerland Universitat Bern
BiihlstraBe 26
CH-3000 BERN 9
Switzerland
Bernard Sapoval
CNRS Ecole Poly technique Bruce J. West
Laboratoire de Physique de Physics Department
la Matiere Condensee University of North Texas
F-91128 PALAISEAU Cedex DENTON, TX 76203
France USA
List of Participants
Dr. Peter Achennann Prof. Philippe Blanchard

Institute of Phannacology Theoretical Physics and BiBoS
University of ZUrich Universitat Bielefeld
Winterthurstr. 190 Universitatsstrasse 25
CH-8057 ZURICH Postfach 8640
Switzerland D-33615 BIELEFELD 1
Gennany
Prof. Gianluigi Agnoli
Dipartimento di Matematica Dr. Bonnaz Didier
Universita di Bologna Institut de Physique Experimentale
40138 BOLOGNA UNIL, Bet. Sci. Physiques
Italy CH-1015 LAUSANNE-DORIGNY
Switzerland
Prof. Werner Arber
Abt. Mikrobiologie Dr. Urs Braschler
Biozentrum Rehetoblstrasse 5
Universitat Basel 9000 ST GALLEN
Klingelbergstrasse 70 Switzerland
CH-4056 BASEL Dr. Hans-Jiirgen Bubenzer
Swizerland Diabetes Forschung
Biochem. Abteilung
Dr. Gerd Baumann Auf'm Hennekamp 65
Abteilung flir
D-40225 DUSSELDORF
Mathematische Physik Gennany
Universitat Ulm
Albert Einstein Allee 11 Dr. Curtis B. Caldwell
D-89069 ULM Departments of Medical
Gennany Biophysics and Radiology
University of Toronto and
Dr. C.L. Benhamou Department of Radiological Sciences
CHR Orleans-La Source Sunny brook Health Science Centre
B.P.6709 2075 Bayview Avenue
45067 ORLEANS-Cedex 2 North York, Ontario
France Canada M4N 3M5
Dr. Bernasconi Giuliano Dr. Giovanna Codipietro

lust. de Physique Exp. Dip. Sci. Ambiente Forestale
Biit. Sciences Physiques Universita della Tuscia
Universite Lausanne Via S. Camillo de Lellis
CH-1015 LAUSANNE-DORIGNY 1-01100 VITERBO
Switzerland Italy
List of Participants 377
Dr. Giovanni Creton Dr. George Feron

Clinica Villa Flaminia GSF
Via Bodio 58 Ingolstadter Landstrasse 1
1-191-ROMA 0-91465 NEUHERBERG
Italy Germany
Prof. Giuseppe Damiani

Universita di Pavia Dr. Josef Feyertag
Dip. Genetica e Microbiologia Dept. for Medical Physiology
Via Abbiategrasso 207 University of Vienna
27100 PAVIA Schwarzspanierstrasse 17
Italy A-1090 VIENNA
Austria
Dr. Angelo Destradis
Dip. di Chimica Dr. Zeno Faldes-Papp
Universita della Basilicata Section for Polymers
Via Nazario Sauro 85 Universitat Ulm
8510 POTENZA Albert Einstein Allee 11
Italy 0-89069 ULM
Germany
Dr. Marc Dupuis
Institut de Biochimie
Universite de Lausanne Dr. Elio Giroletti
Ch des Boveresses 155 Servizio di Radioprotezione
CH-1066 EPALINGES Palazzo dell' Universita
Switzerland C.so Strada Nuova, 65
1-27100 PAVIA
Cristoforo Durig Italy
Contrada Maggiore 30
CH-6613 LOSONE
Switzerland Dr. Walter Glackle
Department of
Dr. Carl Eugster Mathematical Physics
Marigen S. A. 40 University of VIm
Hackbergstrasse Albert Einstein Allee 11
CH-4125 RIEHEN 0-89069 ULM
Switzerland Germany
Dr. Carl J.G. Evertsz
Center for Complex Prof. Ary L. Goldberger
Systems & Visualisation Dept. of Cardiology
University of Bremen Harvard Medical School
Bibliothekstrasse 1 Beth Israel Hospital
Postfach 330440 330 Brookline Av.
0-28334 BREMEN BOSTON MA 02215
Germany USA
378 List of Participants
Dr. Riccardo Graber Dr. Hubert Jacot-Guillarmod

Laboratorio di Patologia Cellulare Hoffmann La Roche
Istituto Cantonale di Patologia Grezucherstrasse 124
CH-6600 LOCARNO CH-4000 BASEL
Switzerland Switzerland
Dr. Guerreiro Joaquim Dr. Hiroko Kitaoka

Institut de Physique Experimentale Dept. of Internal medecine
UNIL, Bat. Sci. Physiques Kitaoka Hospital
CH-IOI5 LAUSANNE-DORIGNY 1031 Meijimachi, Kurayoshi
Switzerland TOTTORI682
Japan
Dr. Massimo Gulisano
DIBIT-Istituto HS Raffaele Prof. A.K. Kleinschmidt
Via Olgettinago University of Ulm
1-20123 MILANO Albert Einstein Allee 11
Italy D-89069 ULM
Germany
Dr. Ricardo Gutfraind
Prof. Dr. Klaus-D. Kniffki
Group Matiere
Institute of Physiology
Condensee et Materiaux
University of Wiirzburg
Batiment 11 B
Rontgenring 9
Universite Rennes I
D-97070 WURZBURG
35042 RENNES Cedex
Germany
France
Dr. Christian Kocourek
Dr. Rachid Harba
Dept. for Medical Physiology
CHR Orleans-La Source
University of Vienna
B.P.6709
Schwarzspanierstrasse 17
45067 ORLEANS-Cedex 2
A- I090 VIENNA
France
Austria
Dr. Achakri Hassan Dr. Martin Kunz
Laboratoire de Genie Medical Institut de Physique Experimentale
Dept. Physique EPFL Bat. Physique UNIL
Champs-Courbes 1 CH-1015 DORIGNY-LAUSANNE
CH-1024 ECUBLENS Switzerland
Switzerland
Dr. Haymo Kurz
Dr. Armin Herold Image Analysis and
Schering Company Stereology Laboratory
Central Biological Research Anatomisches Institut II
Miillerstrasse 171 Albertstrasse 17
D-13353 BERLIN 65 D-79001 FREIBURG
Germany Germany
Dr. Gabriele Landini Dr. Martin Ltineburg

Oral Pathology Unit Universitat Bielefeld
The Dental School Falcultlit ftir Mathematik
Binningham University Universitatsstrasse
St. Chad's Queensway D-33615 BIELEFELD
BIRMINHGHAM B46NN Germany
UK
Prof. Benoit B. Mandelbrot

Dr. Leoni Lorenzo IBM Research Division
Laboratorio di Patologia Cellulare Thomas J. Watson Research Center
Istituto Cantonale di Patologia P.O.Box 218
CH-6600 LOCARNO YORKTOWN HEIGHTS, N.Y. 10598
Switzerland U.S.A
Dr. E. Lespessailles
Dr. Raphael Marcelpoil
CHR Orleans-La Source
Eq. de Reconnaissance des Formes
B.P. 6709
et Microscopie Quantitative
45067 ORLEANS-Cedex 2
CERMO-Universite Joseph Fourier
France
F-BP53X.38041 GRENOBLE Cedex
France
Dr. Petra Leuchtenberg
Falc. Biologie
Zoologie Neurobiologie Prof. Claudio Marone
Postfach 102148 Primario Medicina Intema
Universitatstr. 150 Ospedale San Giovanni
D-44799 BOCHUM-l CH-6500 BELLINZONA
Germany Switzerland
Dr. Vehel Levy Dr. Corrado Mascia

Institut National de Recherche Clinica di Pediatria
en Informatique et Automatique Universita la Sapienza
INRIA B. P. 105 00100 ROMA
78153 LE CHESNAY Cedex Italy
France
Dr. Shu Matsuura

Prof. Gabriele A. Losa Tokai University
Laboratorio Patologia Cellulare School of High Technology
Istituto Canton ale di Patologia for Human Welfare
Via in Selva 4 Nishino 317, Numazu
6604 LOCARNO SHlZUOKA 410-03
Switzerland Japan
380 List of Participants
Prof. Dr. Torsten Mattfeldt Prof. Vittorio Pesce Delfino

Institute of Pathology Consortio di Ricera DIGAMMA
University VIm C.So.A. de Gasperi 449/a
Oberer Eselsberg M23 1-70125 BARI
W-89069 ULM Italy
Germany
Dr. Frederic Pythoud
Prof. Danilo Merlini Laboratoire de Genie Medical
CERFIRM Dept. Physique EPFL
Via Rusca 1 Champs-Courbes 1
CH-6600 LOCARNO CH-1024 ECUBLENS
Dr. Silvia Metzeltin Prof. Rosalia Ricco

RTSI Istituto di Zoologia
Besso e Antropologra
CH-6900 LUGANO Universita di Bari
Switzerland 1-70124 BARI
Italy
Dr. Christian MUller
Institut de Biochimie Dr. Daniel Robert
Universite de Lausanne Mikrobiologisches Institut
Ch des Boveresses 155 Frohbergstrasse 3
CH-I066 EPALINGES CH-9000 ST GALLEN
Prof. Theo F. Nonnenmacher Prof. Fran~ois Rothen
Abteilung fUr Mathematische Physik Institut de Physique Experimentale
Albert Einstein Allee 11 Universite de Lausanne
Universitat Ulm BSP Dorigny
D-89069 ULM CH-I015 LAUSANNE
Germany Switzerland
Dr. Frederic Paycha Prof. Bernard Sapoval

Service d 'Explorations Fonctionelles CNRS Ecole Poly technique
Hopital Luis Mourier 178 Laboratoire de Physique de
Rue Renouillers la Matiere Condensee
F-92700 COLOMBES CEDEX F-91128 PALAISEAU Cedex
France France
Dr. G .P. Pescarmona Dr. Anne Schulz

Dipartimento di Genetica Projekt Inhaltion
Bilogia e Chimica Medica GSF
Via Santerna 5 bis Ingolstadter Landstrasse 1
10126 TORINO D-91465 NEUHERBERG
Italy Germany
Prof. Manfred Sernetz Dr. Ernst Wehrli

Justus-Liebig-Universitat Lab. for Electron Microscopy
Inst. Biochernie u. Endokrinologie ETH Institut Zellbiologie
Frankfurterstrasse 100 Schmelzbergstrasse7
0-35392 GIESSEN CH-8029 ZURICH
Gennany Switzerland
Prof. Ewald R. Weibel

Anatomisches Institut
Universitat Bern
Dr. Tom G. Smith, Jr. Biihlstrasse 26
Laboratory of Neurophysiology CH-3000 BERN 9
National Institute of Health Switzerland
BETHESDA, MD 20892
U.S.A Prof. Bruce J. West
Dept. Physics
University of North Texas
DENTON TX 76203
USA
Prof. Petre Tautu
Gennan Cancer Research Center Dr. Tiziano Zanin
Dept. Mathematical Models Servizio Istologia
Research Program Bioinfonnatics Ospedali Galliera
Neuenheimer Feld 280 Via Mura delle Cappuccine 14
0-69120 HEIDELBERG 1-16128-GENOVA
Gennany Italy
Index
acinus, 79-81, 116, 123, 126, 241, Aspergillus oryzae, 274

242, 246-249 attractors, 165
- volume, 248
active zone, 254, 258, 260 biological
adaptability, 258, 270 - and synthetical pattern, 139
adaptation, 104, 274 - branching, 104
admittance, 243-247, 249, 252 - branching structure, 104, 106,
108, 110
ageing, 55
- cell system, 86,91
aggregation processes, 266 - context, IX, 41, 89, 94
AIDS, 324 - design, 4, 83
AIDS epidemics, 324 - evolution, 4, 160, 161
air spaces, 116 - forms, 68, 104, 105, 112
airway, 4, 68, 69, 73, 75-78, 80, - fractal, 217
81, 108, 116-118, 126, 247 - function, 4, 29, 163, 174
- branching, 77 - generator, 132
- growth, 221
- diameter, 77, 79,80, 117
- growth process, 90
- fractal dimension, 122
- growth rules, 92
- tree, 76-78, 80, 83, 118, 123 - homology, 331
allometric area, 337 - material, 104, 108
alveolar - measurements, 212
- bone, 302 - morphology, 6
- ducts, 77 - objects, 301
- organisms, 2, 6, 68, 69, 82,
- inner cut-off, 80
241
- membrane, 247 - phenomena, 6
- sac, 248 - processes, 283, 352
- surface, 75, 81, 237, 248 - profile, 198
- surface area, 73, 74 - shapes, 105
- walls, 73, 74 - structure, 4, 6, 104, 105, 107,
ammonoid septa, 109 109, 111, 191,301,331
- surfaces, 75
angiogenesis, 133
- system, 41, 88, 174,201,363
animal, 3, 68, 82, 105, 107, 232, - transport systems, 108
234, 235, 241, 246, 248, 249
- tree structure, 223
- design, 82 - variability, 137
anti-correlation, 56, 62 biological organisms, 68
arterio-venous shunt, 133 biomedical application, 165
Index 383
blood. 75.81.82. 108. 117. 191. - pattern. 76. 77. 108. 116. 117.
218 123. 126. 135.210.219.222.
- capillaries. 73 224
- cell. 75. 193. 234 - process. 77. 90. 94. 327
- circulation. 132 - random walk. 94
- flow. 82.139.238.251 - ratio. 77. 79. 223. 228
- plasma. 80. 193 - structure. 104. 106-110. 112.
117.221
- pressure. 41. 59. 136
- tree. 116
- stream. 75
bronchial branching. 237
- substances. 69
brown noise. 60. 61. 63
- vessel. 3. 75. 76. 8~82. 132.
Brownian motion. 60.90.91.280.
221. 238. 332
281. 292. 363
- vessel formation. 133
bone. 300 capillary
- ageing. 292 - electrophoresis. 172
- structure. 300 - growth. 132. 139
border effect. 202 - layer. 135. 136. 139
Bouligand. 15 - network. 75. 81. 82
box - plexus. 139
- counting. 118. 182 - vessel. 135
- counting 3-D. 118. 126 carcinoma. 116.208.315.317.
- counting method. 24. 25. 116. 319-321
134. 182. 184. 186-189. 212. cardiac
221. 222. 237. 278. 310. 315. - arrythmia. 63. 266
317-320 - beat. 63
- dimension. 25 - conduction system. 4
- edge. 134. 137 - death. 63
- lengths. 135 - death sudden. 63
- size. 135 - disease. 60
box-counting fractal dimensions. - rhythm. 63
278 cell. 3. 5. 22. 75. 83. 110. 162. 182.
201.222.315
branching. 76. 104. 105. Ill. 126.
- biology. 22. 23. 70
170. 222. 225. 228
- blood. 190
- algorithms. 132
- body. 210
- bronchial. 237
- border. 184. 188
- dendrites. 228 - boundary. 24. 27. 89
- dendritic. 213. 224 - clone. 96
- dichotomous. 33. 76. 79. 127 - clusters. 89
- generation. 118 - contour. 23. 26. 182-184. 190.
- model. 126 336
- of thalamic neurons. 228 - culture. 191.212.357
- of vessels. 238 - cycle. 92. 97
384 Index
- cytoplasm, 71 correlation, 47, 56, 116,271

- dendritic, 324 - coefficient, 124, 216, 222, 304,
- electrochemical, 241 313,358
- epithelial, 73 - dimension, 38, 51-53, 283,
- growth,95 284, 286
- HeLa, 357 - function, 47, 52
- leukemic, 24, 26, 27, 195 - length, 87
- line, 26, 28, 189 - long-range, 56, 57, 62, 368
- membranes, 30, 72, 74 ~ structure-function, 5
- metabolism, 232 crinkliness, 97
- movement, 89 critical phenomena, 56, 329
- perimeter, 23
- proliferation, 139, 315 D' Arcy Thompson, 68, 79
- surface, 88, 89, 190, 191, 198, Delaunay triangulation, 201
351 design, 68
- system, 86, 92 - principle for biological organ-
- thalamic, 223 isms, 82
- Wigner-Seitz, 151 development
cellular - at cellular level, 109
- automata, 110, 111, 266, 268, - fractal, 164
270 - neuronal, 221
- differentiation, 198, 210, 219 - of an epidemic, 4
- genome, 163 - of biological forms, 68
- membrane, 69, 70, 72 - of biology, 6
chaotic, 5, 23, 38,39,41,47,63, - of blood vessels, 132
173,283 - of Mesozoic ammonoid suture,
Cheyne-Stokes respiration, 61 109
chorioallantoic membrane, 132 - of morphology, 215
circadian rhythms, 266 - of new measures, 38
cluster, 49, 89, 91, 99, 195,310, - of nonlinear forecasting, 41
311 - of population, 4
- aggregation, 118, 221 - of septum, 109
- DLA, 87, 92, 252 diagnosis, 208, 318, 320, 333, 334,
- fractal, 109 346
- of cells, 361 diameter, 212, 248, 275, 339
- self-similar, 94 - arterial, 81
Coast of Wales Effect, 71 - branching ratio, 79
colored noise, 38--40, 53 - exponent, 79, 81, 116, 117,
complexity, 3, 5, 22, 28, 55, 75, 90, 126, 137, 223-226, 228
109, 138, 139, 198, 210, 213, - of a branch, 107
219, 274, 283, 284, 297, 325, - of airways, 77, 126
348 - of an alveolar sac, 248
conduction, 270 - of dendritic branches, 224
configuration model, 86, 89, 91, 97 - of the branch, 117
Index 385
- of the legs, 76 digital image analysis, 182

- tubes, 76 discrete
diameters - cells, 215
- in human lungs, 78 - dynamical system, 39, 269,
dichotomous 326
- branching, 33, 76, 79, 124, - geometric spectrum, 32
126, 127 - Hopf bifurcations, 269
- growth,3 - Markov chain, 30
dichotomy, 76, 117 - model, 105
diffusion, 80, 88, 235, 238, 241, - point lattice, 151
242, 245, 249, 270, 351, 352, - points of the contour, 186
364 - process, 327
- across a membrane, 242 - series, 32
- admittance, 244 - stocliastic models, 324
- and capture of oxygen, 241 - structure, 266
- anomalous, 363, 368, 369 - time, 41, 95, 268, 328
- coefficient, 33, 244, 248, 364, disorder, 4, 34,201,205-207,283,
368 368
- coupling, 268 distribution, 6, 32, 89, 93, 116, 310
- currents, 241 - bidimensional, 202
- equation, 97, 252, 363, 366, - Cauchy, 90
368 - cumulative, 313
- front, 80, 252 - f points, 204
- injection model, 97, 99 - function, 171, 175
- internal, 235 - GauB, 183
- lateral, 351 - GauBian, 281
- limited aggregation (DLA), 3, - isotropic, 332
111,217 - log-logistic, 239
- macroscopic, 243 - lognormal, 94
- mechanism, 98 - nondegenerate, 94
- model, 97, 99, 111 - of airways, 116, 118
- normal, 370 - of energy barriers, 31, 175
- of Na+, 242 - of events, 31
- of nutrient, 278 - of local dimensions, 313
- of oxygen, 80, 242, 247 - of nucleotides/nucleoside, 169
- of unbleached molecules, 368 - of reaction rates, 176
- process, 98, 111, 244, 358 - of the bending stresses, 110
- rate, 156 - of the deformation energy, 109
- self-similar, 90 - of the passage time, 96
- source, 242, 243 - probability, 188
- system, 242 - residence time, 238
- two-dimensional, 356 - turbulent, 238
- within the alveoli, 248 distributions
diffusion of glucose, 278 - of pore size, 236
386 Index
DNA, 59, 160, 165 - far-from, 132

- acquisition, 162 - like state, 58
- chain, 55 - punctuated, 106
- fragments, 163 - space division, 117, 129
- inversion, 161 - thermodynamic, 232
- junk, 57 - value, 363, 366
- molecule, 55, 160 error
- polymers, 55 - accumulation, 45
- rearrangement, 162 - clusters, 171
- replication, 160 - curve, 46
- segment, 161, 162 - estimate, 42
- sequence, 56, 163, 165 - in forecasts, 39
- sequences, 55 - numerical, 47
- stranded, 166 - of closing, 186
- structure, 57 - propagation, 165, 172
- walk, 55-59 - sequences, 166, 167, 172
driven multi-cycle synthesis, 165 - tolerance, 4
drugs, 242, 352 errors, 183
evolutionary
Eden model, 86, 89, 93, 95, 275 - advantage, 88
EEG, 41, 283 - advantage of fractal progeny, 4
- analyses, 288 - change, 160
- data, 285 - interplay, 3
- patterns, 266 - mechanism, 58
- signal,38 - morphology, 333
- sleep, 283 - process, 104, 105, 324
- slow-wave activity, 283 - strategy, 161
electrodes, 4,241,242 - theory, 105
electron micrographs, 182 - tree, 163
embryo, 88, 132, 133 - trends, 104, 106, 108, 112
energy, 107, 156 excitable media, 266
- activation, 180 explorative multivariate statistics,
- balance, 234 133
- barrier, 31, 175, 178
- dissipative, 232 fibrillation, 63
- metabolism, 82, 266 filtering methods, 182
- of the deformation, 109 fixed starting sites, 165, 166, 170-
epidemic, 4,324 172
- dynamic, 325, 327 fluorescence
- homosexual,327 - after bleaching, 358
- model, 91, 96 - intensity, 351, 356, 357
equilibrium, 269 - kinetics, 351, 352, 358
- attraction-repulsion, 81 - microscopy, 351
- concentration, 364 - recovery, 351, 357, 363
Index 387
- signal, 363, 364 gas

follicular dendritic cells, 324, 330 - exchange, 4, 69, 75, 79,133
forecast, 39,41-43,45-49 - exchange surface, 80
Fourier - exchanger, 82, 249
- analysis, 59, 348 generators
- coefficients, 341 - of surface texture, 68, 75
- components, 60 genetic
- harmonic analysis, 339 - algorithms, 104, 112
- constitution, 164
- space, 366
- differences, 233
- spectrum, 341, 349
- element, 162
- transform, 98, 366
- engineering, 165
- transform amplitudes, 60
- function, 162
fractal, 55-57, 59, 62, 132,292
- information, 105, 109, 160,
- between limits, 22
162, 163
- landscape, 59 - instability, 161
- time process, 29 - instructions for design, 81
- dimension, 5, 6, 22, 23, 26, 27, - programming, 68
29,38,44,71,72,74,87,116, - variation, 161, 163
118, 122, 126, 127, 132, 133, gradient, 110, 242, 332
139, 165, 171, 172, 182, 184, - centrifugation, 191
185, 188-194, 197, 198,210- - fields, 227
213, 215-219, 221-223,227, - filters, 183
237,238,244,247,249,253, - governed growth, 93
292, 297, 300, 302-305, 315, - morphogenetic, 104, 111
318-321
- normalized, 310
- geometry, 2, 6, 22, 68, 69, 79, - of a growth factor, 97
81,82,88, 105, 108, 181, 199, - of concentration, 351, 360
210, 247, 251, 307
- of heat, 270
- properties, 6, 72, 82, 182 growth, 3, 165,271
- structure, 4, 26, 29, 92, 96, - and differentiation, 216
223, 232, 236, 238 - and form, 55
- tree, 80-82 - and function, 105
fractals and genes, 83 - and invariance, 107
fractional - area, 217
- Brownian motion, 280 - by means of random aggrega-
- operator design, 23 tion, 111
fractional Brownian motion, 292 - capillary, 132
fundamental shape, 335, 336, 341, - circumstances, 93
344 - curve, 68
fungal - dendritic, 111, 216
- colony, 274, 282 - dichotomous, 3
- structure, 3 - directional, 165
fungi, 274, 278 - dynamics, 86
388 Index
- factor, 97, 134 - rate, 58, 59, 62, 82

- fractal, 86, 217 - rate dynamics, 62
- front, 274, 279, 281 - rate fluctuation, 59, 60
- gradient governed, 93 - rate oscillations, 62
- in fractal dimensions, 88 - rate spectrum, 61
- irreversible, 221 - rate variability, 55, 60, 63
- logarithmic law, 68 - regular sinus rhythm, 59
- model, 91, 99, 172 Henonmap, 39,46,48-50, 53
- model Laplacian, 92 Heitzman's fixation, 117
- mycelial, 278, 282 homeodynamical model, 165
- neuronal, 216, 217 homeostasis, 58, 63, 283, 292
- nonfractal cell models, 95 homogeneous
- of biological forms, 68 - 3-D arrangement, 116
- of DLA clusters, 92 - case of solute enzymes, 235
- of error in forecasts, 39 - colonies, 274
- of figures, 90 - media, 282
- of individual cultures, 215 - membrane, 249
- of neurons, 223 Hopf bifurcations, 269
- of trees, 68 human
- of tumour, 315,320 - airway tree, 76
- period, 138 - airways, 117
- probability, 99 - blood, 191
- process, 86, 88,90,91,94,95,
- granulocytes, 23
104, 145, 166, 172, 223, 275,
- immune deficiency, 4
278
- intestine, 241
- quotient, 97
- leukocyte antigen, 195
- random, 275
- lung, 73, 76-78,80, 107, 108,
- rate, 89,95, 145, 150, 274,
116, 117, 248
278, 280, 282
- Iymphoblastoid cell, 192, 196
- rate of cells, 132
- lymphocytes, 197
- rules, 86, 91, 92
- sexual behaviour, 324
- step, 68
- systemic arteries, 117
- two-dimensional, 96
- T-Iymphocytes, 22
- vision wired for fractals, 8
hairy cell, 22
hyperplasia, 207, 208
hairy cells, 22, 24, 26, 27, 29, 36
harness process, 98, 103
heart, 75, 76, 234,271 ideal
- beat, 5, 58-60 - fractal behavior, 33
- beat dynamics, 55 - fractals, 87
- disease, 59, 62 - neuron, 227
- failure, 59, 61, 62 - scaling behavior, 28
- frequency, 5 - storage, 268, 270
- muscle, 75,81 image
Index 389
- analysis, 5, 22, 23, 27, 132, - reflection microscopy, 371

133, 182, 192,237,307,313, - septum, 109
356 - structure, 118, 238
- analysis 3D, 238 - surfaces, 68, 69, 74, 82
- analyzer, 356 - time scale, 180
- anisotropy, 294 internal surfaces, 70
- digital, 302, 303 inverse
- EM-image, 182 - beat number, 61
- neuronal, 213, 215 - Laplace transform, 178
- of airways, 118 - microscope, 352, 354
- of arterial vessels, 133 - operator, 178
- of the liver, 313 - power law, 165, 170-172, 174
- plane, 23, 25 - power law behavior, 34
- processing, 134 - power law exponent, 35
- processor, 210 - power law relaxation, 30
- scanner, 275 - power law scaling, 26
- segmentation, 5 - power law spectrum, 38, 39
- textures, 301 irregular
- X-ray, 293 - contour, 190
image arrays, 303 - electrode, 252
images - figure, 339
- of calcaneus, 293 - forms, 348
- radiological, 292 - geometric patterns, 29
impedance, 241, 242, 252 - geometries, 246
inhomogeneous - interface, 254
- environment, 274 - landscapes, 59
- membrane, 249 - membrane, 247, 251, 252, 264
- substrates, 274 - object, 4, 251, 333
interacting - pattern, 196
- biological cell systems, 86, 91 - shape, 22, 86, 315
- cell biological systems, 88 - shape, 332
- cell system, 89,91,92,95 - structure, 22, 105
- cells, 236 - surface, 89, 193,251
- particle system, 91 , 92, 328 - texture, 292
- point system, 154 - time series, 38
interaction of the HIV, 325 - vessels, 138
internal isometric, 237
- diffusion, 235 - area, 339
- DLA configuration, 92 iterated function systems, 90, 170,
- DNA rearrangement, 163 189
- influences, 83 iterative
- medicine, 307 - estimates, 45
- recombination, 163 - forecasting, 45
- reflection, 351, 365, 368 - geometric constructions, 90
390 Index
- procedure, 258 - distribution, 165

- process, 254 - fluctuations, 282
- rules, 105, 109 - of a coast line, 71
- of a grid, 25
kidne~ 233, 242, 249 - of the active surface, 256
- arterial tree, 237 - of the boundary, 97
- tissue, 307 - of the box, 122, 222
Koch,109 - of the branches, 125, 224
- curve, 23, 26, 188 - of the closed cell contour, 23
- islands, 212 - of the perimeter, 248
- trees, 76 - of the primary dendrites, 224
- of the sampling box, 118, 126
lacunarity, 5, 6, 8, 9, 12-14, 16, 17, - of yardstick size, 25
21,29,87 - scrue, 22, 56, 86, 87, 264, 310
Laplace - scale of the yardsticks, 184
- equation, 251, 252, 260 - scales, 223
- filter, 183 - to-diameter ratio, 77
- Mellin transform, 369 Levy wruk, 58
- transform, 176, 178, 369 liver, 68, 71, 75, 89, 307
laser, 351, 354 - and kidney tissue, 307
- beam, 351, 352 - boundaries, 311
- beam profile, 358 - boundary, 308
- flash,34 - cell, 70, 71
- intensity, 356 - cell cytoplasm, 71
- light, 351, 352 - cell membranes, 72
- pulse, 363 - fattiness, 307, 313
- scanning microdensitometer, - images, 312
302 - of rat, 320
lateral - scans, 302
- diffusion, 351,352,363 - tissue, 308, 313
- diffusion coefficient, 33 local
- lipid diffusion, 360 - aggressivity, 320
- lipid mobility, 360 - approximation, 39, 40, 42, 44-
- mobility, 351, 352, 358, 361 46, 48
- motion, 351 - competition, 92
- processes, 365 - complexity, 88
- thalamus, 224 - concentration, 11 0
- transport process, 366 - connected fractal dimension,
length 314
- and diameter, 76, 117 - contour perturbations, 334
- and time scales, 35 - correlation, 56
- characteristic, 56 - differences, 118
- correlation, 87 - dimensions, 237, 313
- density of arteries, 136 - effect, 62, 257
Index 391
- fractal properties, 6 - processes, 91, 174

- geometry, 310 mass dimension, 308, 309
- hyphal distribution, 281 maximum
- interactions, 86 I - amplitude, 341
- limit theorem, 94 - D, 216
- mass dimension, 308, 309 - determinism, 40
- rules, 104, 111 - diameter, 339
- scale, 201 - distortion condition, 337
- topography, 201 - Fourier harmonic, 346
logarithmic - length, 321
- coordinates, 239 - likelihood estimation, 294
- derivation, 370 - likelihood rule, 285
- derivatives, 193 - resolution, 237
- growth law, 68 - standard deviation, 193
- plots, 56, 61, 72 - volume, 109
- progression, 171 membrane
- spiral, 3, 69, 83 - admittance, 244, 246
logistic - antigens, 191
- data, 41, 52 - catalytic, 235
- equation, 40 - cellular, 70
- log-distribution function, 239 - endoplasmic, 70
- map, 39, 45-47 - fluidity, 351
long-range - folds, 73
- correlations, 55, 56, 58, 62, 63, - fractal, 243, 246, 247, 249,
86, 368 260,264
- dependence, 99 - fragments, 198
- power-law, 55, 62, 63 - image, 72
lung, 3, 4, 68, 69, 73-78, 80-82, - immunoglobulin, 195
105, 107, 108, 116, 117, 119, - interactions, 352
123, 126, 128, 129, 217, 237 - lipids, 198
lymphoblasts, 36, 190 - mitochondrial, 82
lymphocytes, 22, 182, 188, 190, - of cells, 197
193, 194, 197, 198, 328, 329 - of endoplasmic reticulum, 72
- T, 22, 27, 28, 36, 189, 190 - of leukemic cells, 198
- T4, 328 - of mature B lymphocytes, 194
- of REH-6 cells, 196
marginal - of the liver, 68
- points, 203, 204 - permeability, 241, 352
- polygons, 204, 205 - plasma, 190
Markov - proteins, 30, 360
- chain, 30-32, 56, 90, 326 - resistance, 244, 260
- configuration model, 86,91,97 - structure, 351
- contact process, 95 - surface, 23, 241, 249
- model, 86, 89 - surface immunophenotype, 191
392 Index
- with finite penneability, 241 - processes, 3, 109

method - properties, 80
- box counting, 182 - transformation, 110
- probabilistic, 182 morphometric
- yardstick, 182 - description, 74
- based on fractional integral, 34 - estimates, 75, 191
- boundary trace, 318 - evaluation, 334
- box counting, 134, 186 - machines, 334
- dilation, 215 - methods, 191, 321
- of fractal dimension, 318 -' shape, 190
- probabilistic, 25 - study, 74
- yardstick, 184, 318 morphometry, 118, 241, 249, 333
micro architecture, 292, 293, 298 - of liver cell, 70
microscopic mosaic patterns, 89
- biological growth, 92 mouse, 82, 192,210,232,248
- growth rules, 86, 91 multi-species filter, 249
- histological study, 292 multifractal, 6, 8, 20, 102, 193,307,
- image, 136 314
- magnification, 135 myoglobin, 23, 34, 174, 175, 179-
- measurements, 5 181
- resolution, 68, 73
- speckle structure, 308 natural
microscopic growth rules, 91 - fractal, 86, 135, 239
Minkowski sausage, 14 - gene vectors, 162
molecular - logic, 333
- biology, 165 - objects, 184
- evolution, 55, 58 - phenomena, 104
- mutagenesis, 164 - processes, 251
- regulatory mechanisms, 88 - scientists, 2
- structure, 30 - selection, 104, 106
- weight, 358 - structures, 6, 105, 301
morphogenesis, 55,68,69, 75, 81, - time scale, 31
83, 88, 132, 201 Nautilus, 68, 69, 146
- of fungal structure, 3 neural
- of the lung, 76 - net, 49, 50, 52, 54
morphogenetic - network, 5,40
- advantages, 75 - network technique, 49
- changes, 106 neuron, 5, 23, 38, 211, 212, 215,
- factors, 332 217,218,221-229
- fields, 104 noise, 38,40,45,47,50-53,97,
- goal, 348 299,310
- gradient, 104, 110, 111 noise
- patterns, 22 - 1If, 61
- principles, 3, 4 nonREM sleep, 284, 286
Index 393
nonREM-REM sleep cycle, 283 - law, 4, 32, 55, 56, 58, 62, 63,
normalized 98, 107, 179, 232, 239, 246
- gradient, 310, 311 - law behavior, 22
- relaxation function, 368 - law decay, 174, 177, 370
nucleotides, 55, 56, 165-168 - law kinetics, 34
- law relaxation, 30
oligonucleotides, 165, 166, 172 - law scaling, 26
optimization, 104, 105, 107-109, - law solutions, 23
333 - law spectra, 39
organization, 5, 6, 55, 132, 201, - law trend, 31, 33, 175
218, 232, 236, 237, 292 - of body mass, 82
os calcis, 293 - series expansion, 93
osteoarthritis, 300, 302 - spectra, 61
osteoporosis, 292,293,297,298, prediction
301 - error, 39-45, 47
- of chaos, 40
pattern - time, 40, 42, 45, 48
- formation, 132, 199, 223, 274, probabilistic method, 24, 25, 182,
278, 320 184, 188-192
- in cell biology, 22 prostatic
- of airway branching, 77 - carcinoma, 208
- of airways, 126 - tissue, 207
- of dendritic branching, 224 protein dynamics, 22, 23, 30, 34
- of evolutionary processes, 104 pulmonary
- of fluctuations, 59 - acinus, 125, 242
- of hemodynamic sources, 136 - alveoli, 251
- of structural order, 3 - arterial tree, 81
- of the bronchial tree, 223 - arteries, 80, 81
- of turbulence, 4 - gas exchanger, 82
- smoother, 190 - vascular trees, 80
percolation, 4, 86, 91, 236, 311,
329 rabbit, 248
permeability, 241,244, 247, 251, radiographs, 118, 300-302, 305
260, 352 radiological
- of the membrane, 242 - bone images, 292
- values, 260 - images, 292, 298
personal super computer, 283, 284, - images of calcaneus, 293
287 - projection, 292
photobleaching, 351, 358, 363, 364 random
pigmentation patterns, 266 - aggregation, 111
plasma membrane, 4, 190, 193, - character, 38
196-198, 352 - field, 97
power - fluctuations, 106
- dissipation, 285 - fractal, 23, 38, 86, 93, 94, 137
394 Index
- graphs, 325 - behavior, 28, 44, 184, 186,

- growth processes, 275 190,309
- network, 327 - Bernoulli, 177
- orientation, 134 - concept, 31
- phenomena, 4 - condition, 32
- process, 38, 86, 104 - dimension, 23, 184, 188
- samples, 135 - domain, 198
- spatial models, 86 - exponent, 59, 182, 185, 186,
- time series, 43 236
- transition mechanisms, 90 - factors, 171, 172
- variable, 30 - form, 258
- walk, 4, 56, 60, 62, 63, 91-94, - hypothesis, 87, 188
242, 244, 328 - ideas, 30
Random Graphs, 324 - law, 23, 27, 31, 223, 225
rat, 248 - limit, 22, 23, 25, 28, 187
reaction - model, 177
- cycle, 166, 167, 172 - of bronchial branching, 237
- diffusion, 266 - of the mass, 237
- diffusion model, 88, 110 - of the prediction error, 44
- kinetics, 175, 232 - patterns, 199
- processes, 363 - power functions, 235
- rate, 176, 269 - problem, 107, 271
reconstruction, 38,41, 116, 120, - procedures, 171
122 - properties, 254
REM sleep, 283, 284, 287, 288 - range, 34, 182, 186, 197
- region, 28, 29, 33, 185
respiratory
- relation, 226
- data, 133
- relationship, 308
- organ, 133
- result, 45
- part of the lung, 116
- volumes, 233
- physiology, 241
scaling formulas, 29
- system, 85
self
rhizomorph, 278
- adapted, 264
roughness, 27, 86,93, 96, 97, 187,
- affine, 56, 87
198, 241, 252, 274, 275, 279-
- affine fluctuation, 281
281
- affine fractals, 274
- affinity, 19,57,87,95,280,
S.A.M.(Shape Analytical Morphom- 282
etry), 333 - affinity exponent, 281
scale - affinity relations, 275
- invariance, 68 - annihilating branching pro-
scaling cesses,90
- alterations, 55 - organization, 109, 271
- argument, 242, 252, 255 - organized criticality, 88
Index 395
- organizing, 132 - biological, 105

- organizing systems, 88 - description, 333
- similar, 68, 77, 81, 87, 174, - differentiation in biology, 331
177,201,239,278 - fractal, 301
- similar - fundamental, 335
- branching, 123 - gap, 150
- cell profiles, 22 - in biology, 108
- clusters, 94 - invariance, 150, 154
- construct, 188 - irregular, 86
- curve, 266 - measurement, 331
- diffusions, 90 - of a branching random walk,
- division rules, 89 94
- domains, 182 - of a cluster, 93
- dynamical processes, 176 - of biological objects, 346
- error cluster, 171 - of botanic units, 152
- fractal membrane, 242 - of cells, 191
- fractals, 252 - of epithelial-connective tissue,
- geometric structures, 35 315
- growth process, 223 - of membrane profiles, 71
- interface, 252 - of the chest cavity, 77
- interfaces, 242 - of the nose, 333
- mathematical construct, 26 - of the spiral, 149
- membrane, 245, 252 - of Voronoi-polygons, 151
- object, 29 - perception, 3
- patterns, 22 - recognition, 333
- power law, 98 - theorem, 95
- process, 89, 178, 180, 320 signal, 38, 39, 50, 53, 198, 211,
- scaling, 104 218, 227, 283, 287, 292, 364
- structure, 25, 27, 28, 186 signal-to-noise ratio, 50
- structures, 74 size-shape problem, 331
- superposition, 29 sleep stages, 283, 286-288
- similarity, 23, 25, 27, 29, 68, space
69, 75, 77, 86-88, 90, 107, - and time, 89
116, 126, 135, 137, 142, 145, - branching process, 90
154, 168, 182, 190, 191, 198, - division, 117, 126
223,239,266,301,307,348 - Euclidean, 88, 192, 221
- similarity in phyllotaxis, 156 - filling, 104,227,318
- similarity of airways, 128 - filling tree, 79, 80
- similarity of the tree, 156 - partition, 202
- sustaining epidemics, 327 - representation, 41, 42
- written programs, 356 spectral
serial slices, 117, 119, 122, 126, - analysis, 60, 283, 284
127 - density, 98
shape - features, 59
396 Index
spiral, 3, 69, 83, 141, 144, 146-150, - phenomena, 88

268-271, 273 - resistance, 251, 256, 257, 260
Steiner problem, 104, 108, III - roughness, 241
stochastic - texture, 68, 75
- differential equation, 97 - wrinkles, 74
- growth model, 91 synthesis, 165, 166, 168, 170, 172,
- iteration, 90 335
- mechanisms, 63
- model, 324 terminal
- process, 23, 60, 89, 91, 98,
- airways, 76
223
- branches, 117, 126, 128
- simulation, 106
- branching, 80
strategy, 83, 161-163, 274, 307
- bronchioles, 116
surface
- cap of epithelium, 76
- antigens, 192
- transferase, 195
- area, 24, 68, 70, 71, 73, 74,
- twig, 68
82, 107,360
texture, 9
- complexity, 74, 198
- density, 72, 74 texture
- density estimates, 72 - analysis, 308, 313
- dynamics, 371 - irregular, 292
- elements, 75 - of bone, 302
- fractal, 75, 87,253 - of pure liver tissue, 313
- fractional dimension, 190 - of the fractal, 6
- function, 333 time
- immunophenotype, 191, 196, - axis, 89
198 - constant, 30, 174, 210
- impedance, 258 - course of development, 210
- irregularity, 97, 256 - delay, 285
- layer, 363 - distribution, 239
- localization, 251 - fractal, 22, 31
- markers, 192-194 - interval, 34
- membrane antigens, 191 - of observation, 31
- membrane immunoglobulin, - period, 281
195 - process, 32, 177
- molecules, 72 - process fractal, 29
- morphogenesis, 88 - scale, 22, 29, 180
- of a linear membrane, 244 - series, 31, 38-40,43,47,53,
- of the air-blood barrier, 73 59, 60, 100, 132
- of the electrode, 258 - step, 40
- of the kidney, 237 - structure in dynamic, 5
- of the lung, 4,68, 69, 73,81 topographies, 201
- of the pulmonary alveoli, 251 trabecular bone, 292, 294, 298, 305
- patterns, 190 turbulence, 4, 81, 132, 238

Fractals in Biology and Medicine PDF

Uploaded by

Copyright:

Available Formats

Fractals in Biology and Medicine PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fractals in Biology and Medicine PDF

Uploaded by

Copyright:

Available Formats

R C A Sil

Fractals in biology and medicine / edited by T. F. Nonnenmacher, G. A. Losa, E. R. Weibel

Deutsche Bibliothek Cataloging-in-Publication Data

Fractals in biology and medicine / ed. by T. F. Nonnenmacher

Preface: Summary of the Symposium . ...... . .. . . ..... ...... . . ... . ... 1

Fractals in Biological Design and Morphogenesis . .... . . . . . ...... ..... 67

Fractals in Molecular and Cell Biology 159

Fractal Structure and Metabolic Functions 231

Estimation of the Correlation Dimension of All-Night

Fractals in Pathology ......... . ...................................... 291

Modelling ....... . ........................................ .. ......... 323

List of Speakers .............................................. . . . .... 373

List of Participants .......... . .. . . . . . . . . . . . . . . . . . . . . . . .... . 376

Index.......... . .................................................... 382

Under the Auspices of

- INTERNATIONAL SOCIETY FOR STEREOLOGY

Financial Support was Received from

- DIPARTIMENTO DELL'ISTRUZIONE E DELLA CULTURA DEL

the perspective of living systems must be an interdisciplinary endeavour. One of

1 The fractal nature of morphogenesis in biological systems

2 Importance of fractal analysis for understanding

chaotic complexity of heart frequency, appears to be a safety measure, since the

3 Assessing complexity of biostructures in health and disease

4 The role of the eye in morphology

2 Examples that show that, for Cantor dusts, a given fractal

I IIIIIIIII I IIII I I III II I I II

3 Implications of the examples in Section 2; the loose

Such a source of possible controversy is by no means hypothetical. As a

4 A possible measure of lacunarity: via the filling

lim [log L (E) j log E] = 1 - D. (3)

On the line, 1 - D, is a set's codimension C. This result of Bouligand holds

'Circumstantial evidence suggests that in the case of a fractal it is impossible for

One can imagine using max{L(E)(2E)0-1} to define IjAF, and calling AF

10gL(E) - (1 - D) log(2E) (6)

The values of both expressions for 1/ max{L(E)(2E)0 - 1} are plotted on Fig-

5 Generalization, random dusts, and a connection between

6 The tedious issue of how to define fractals, fractal

7 Bibliographic guidance concerning self-affinity

8 Bibliographic guidance concerning multifractals

I (P) = I, or I (P) = «true,» if the point P belongs to the set 5;

a fractal interpretation of certain stochastic, chaotic-like processes. More recently,

2 Quantitative Structural Analysis of Self-Similar Cell Profiles

on the experimental device involved and on magnification as we will show in this

2.1 Measurements of the Fractal Dimension

2.2 Estimation of Scaling Limits

2.3 Estimation of the Fractal Dimension by Asymptotic Fractal Formulas

Formula (3) has been suggested by Barth (private communication). Taking g = -1

the individual D-values. Therefore, we were able to calculate a mean value D =

3 Scale-Invariant Protein Dynamics

duration between events which is longer than an order-of-magnitude occurs with

where aN [20] is a normalization constant and ~ = f / T.

GN(~) = T~~ A CVm IN(vm, ~/)..) + ~~ (A -voe-~/).. + A-Nvoe-~/)..N) (8)

where Vm = 1 - (In p + 27rim)/ In A is a complex-valued exponent and IN denotes

IN(vm, x) = dwwvm-Ie- w = 'Y(vm, x) - 'Y(vm, XA -N+I). (9)

A re-interpretation of the parameters aoo I 7 In >. = (3 L0, J1, + 1 = (3, u>. =

dN(t)/dt = -(l/r)N(t) (12)

with the exponential solution N(t) = N(O)exp(-t/T) to a fractional differential

Liouville-Riemann integral operator oD-;-{3, defined by [21] [27]

with 0 < ,8 < 1 and replacing (I/T)oDt1N(t) by (I/T{3)oD-;-{3N(t), the inte-