Download as pdf or txt
Download as pdf or txt
You are on page 1of 28

Ph.D.

Thesis
Anders Johansen
Discrete scale invariance and other cooperative
phenomena in spatially extended systems with threshold
dynamics

CATS
Niels Bohr Institute
University of Copenhagen
0
Discrete scale invariance and other cooperative
phenomena in spatially extended systems with
threshold dynamics
Anders Johansen
December 31, 1997
Contents
1 Introduction 5
1.1 Imitation and Synchronisation . . . . . . . . . . . . . . . . . . . . 6
1.2 Fractals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Hierarchical systems . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.4 1=f -noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Scale invariance and discrete scale invariance . . . . . . . . . . . . 11
1.6 Exactly solvable models with discrete scale invariance . . . . . . 14
1.6.1 Potts model on a hierarchical lattice . . . . . . . . . . . . 14
1.6.2 Di usion on a random lattice . . . . . . . . . . . . . . . . 19
1.6.3 An experimental hierarchical model . . . . . . . . . . . . . 21
1.7 Possible log-periodic signatures in cracking . . . . . 1
. . . . . . . 22
2 Hierarchical structures in growth models 27
2.1 Laplacian growth models . . . . . . . . . . . . . . . . . . . . . . . 27
2.1.1 Numerical results . . . . . . . . . . . . . . . . . . . . . . . 29
2.1.2 Possible origin of discrete scale invariance in DLA . . . . . 34
2.2 The thermal fuse model . . . . . . . . . . . . . . . . . . . . . . . 37
2.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4 Paper I: D. Sornette, A. Johansen A. Arneodo, J.-F. Muzy and H.
Saleur, Phys. Rev. Lett. 76, 251-254 (1996) . . . . . . . . . . . . 43
3 The stock market and nancial crashes 44
3.1 How to model the stock market? . . . . . . . . . . . . . . . . . . . 45
3.2 Discrete scale invariance in nance . . . . . . . . . . . . . . . . . 52
3.3 Fitting the stock market index . . . . . . . . . . . . . . . . . . . . 57
3.4 \Last minute" results . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.6 Paper II: Eur.Phys.J. B 1, pp. 141-143 (1998) . . . . . . . . . . . 68
3.7 Paper III: D. Sornette and A. Johansen, Physica A 245, 411-422
(1997) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.8 Paper IV: D. Sornette, A. Johansen and J.P. Bouchaud, J. Phys.
I. France 6, 167-175 (1996) . . . . . . . . . . . . . . . . . . . . . 70
4 Model epidemics 71
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.3 A mean- eld model . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
1 This section is based on notes belonging to a Ph.D.-course at Denmark's Technical Univer-
sity and is written by Henrik Myhre Jensen.

1
4.6 Paper V: A. Johansen, Physica D, vol. 78, 186-193, (1994). . . . . 96
4.7 Paper VI: A. Johansen, J. Theo. Biol. 178, 45-51 (1996) . . . . . 97
5 Earthquakes and stick-slip 98
5.1 Earthquakes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.2 Fault models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.3 A stick-slip experiment . . . . . . . . . . . . . . . . . . . . . . . . 100
5.3.1 Dynamic phases . . . . . . . . . . . . . . . . . . . . . . . . 102
5.3.2 Mean- eld Burridge-Knopo model and branching . . . . . 107
5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.5 Paper VII: A. Johansen et al. , Phys. Rev E, Vol. 48, 4779-90
(1993). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.6 Paper VIII: A. Johansen, P. Dimon and C. Ellegaard Wear, Vol.
172, 93-97 (1994). . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6 Earthquake prediction 115
6.1 Paper IX: A. Johansen et al, J. Phys. I. 6 1391-1402 (1996) . . . 118
7 Conclusion and perspectives 119
A SOC or POC? 121
A.1 Paper X: D. Sornette, A. Johansen and I. Dornic, J. Phys. I.
France 5, 325-35 (1995) . . . . . . . . . . . . . . . . . . . . . . . . 121
B 101 years of The Dow Jones Average 122
C The method of surrogate data 128
D GARCH(1,1) model of the stock market 130
D.1 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
D.2 Initial values of the parameters . . . . . . . . . . . . . . . . . . . 131
D.3 Objective function . . . . . . . . . . . . . . . . . . . . . . . . . . 131
E Con rmation of authorship 133

2
Preface
The systems considered in this thesis are what one may call spatially extended
systems with threshold dynamics. Systems with threshold dynamics, or relax-
ation systems, abound in nature and examples can be found in very di erent
areas such as biology ( re- ies, epidemics), geophysics (earthquakes, volcanoes),
economy (the nancial markets), engineering (stick-slip in friction, fracturing) to
mention a few. All of these complex systems consist of a large number of units,
be it individuals, traders, asperities etc. Due to some kind of coupling between
the units, a collective behaviour is observed, which due to an external forcing of
the system may vary with time. The systems presented in the following chapters
have the common feature that they, despite a large number of internal degrees of
freedom, can \phase-lock" in a collective behaviour, which can be described by a
few global quantities.
Besides a more general introduction to some concepts from statistical physics,
my thesis is divided into 5 chapters corresponding to the 5 di erent areas that
have been investigated: Large stock market crashes, growth models, stick-slip
in friction, earthquakes and epidemics. The purpose of the ve chapters is to
present the papers published by myself and co-authors, as well as giving a more
personal view on the di erent subjects, emphasised by the use of \I" as opposed
to \we", including some additional results which have emerged after the articles
have been published. While the connection between the rst 4 areas is rather
obvious (at least to me!), the last area is primarily connected via the concept of
threshold dynamics and complex systems. However, my own personal interest
in the subject of disease spreading and epidemics forbid me to exclude this very
important subject from my thesis.
That the title of my thesis is what it is and not \An Experimental Investi-
gation of Acoustic Emission from Failure Processes" is literally pure accident(s).
I've always believed that misfortunes, such as delays, sloppy work, plain old bad
luck etc. had a Gaussian distribution. What I'm painfully aware of now is that
this is only true in the context of \a grand-canonical averaging" and do not apply
to individuals. It has become quite clear during my Ph.D.-studies that I live in
the tail . First, I spent months in France on a purchased measuring device that
2

did not meet the promised standards; later, weeks were spent on a CAMAC-
crate for the same reason; then it took the workshop 1 year and not the promised
4 month to construct the pulling device (I do not blame the workmen of the
workshop!) and even after the work was nished it had to be dismantled twice
due to construction errors; the electronic workshop used almost 5 month on the
some rather straightforward pre-amps; after weeks of frustrating work with DMA
data transfer between the scope and PC, I nd that my new GPIB-card does not
2The only other possibility is that some supernatural power is strongly suggesting that I
quit experimental physics.

3
meet the rates promised by the manual and that it can only transfer alphanumeric
strings and not arrays of bytes; at some point, it appeared as if I had a NIM-rack,
which de ed causality (it turned out to be a glitch in the power-supply). The
list goes on and I owe my sanity to the wise words of Predrag Cvitanovic \Don't
be angry about all the things that do not work. Be happy about all the things
that actually do work". With scarcely three months left of my Ph.D.-studies, I
nally got a set of pre-amps that worked. I then realized that it would be an
impossible task to nish the experiment in time as well as writing a dissertation
and I gave up trying to include the experiment in my thesis. Hence, what was
originally thought to be the backbone of this dissertation has been reduced to a
single section (section 1.7) and the thesis consists primarily of the work I've been
doing while waiting for things to arrive, people xing their mistakes and so forth.
First, I would like to thank my son Thorlek for his patience with his mother
and father, who have arranged things so inconveniently as to nish their studies
at the same time. Besides my di erent co-authors, I would like to thank Clive
Ellegaard, Tomas Bohr and especially Didier Sornette to whom I owe most of my
scienti c training after nishing my M.Sc. and whose scienti c enthusiasm and
energy never ceases to astonish me. I thank Marc Oxborrow for supplying me
with the diagram of the pre-amps and bringing the use of piezo-electric lms to my
attention. I also thank Mogens Levinsen, Klaus Lindeman, Sune Hrluck, Jacob
Sparre Andersen and Morten Stocklund Johansen for help concerning the inter-
facing between measuring device and P.C. and Freddy Christiansen for checking
my (ab)use of the English language. Furthermore, I owe them gratitude together
with the rest of the students of CATS for patiently having listen to my complaints
and out-bursts of frustration. I also thank Mads Svendsen from SE-banken for
providing the Hang Seng index and Ole Winter Christiansen and Morten Sparre
Andersen from Denmark and Greenlands Geological Survey for commenting on
the work done in relation to the earthquake in Kobe. I thank Mogens Hgh
Jensen, Henrik Jensen and Professor Dietrich Stau er for accepting to sit in the
committee. Last, I would like to thank the di erent organisations which have
contributed to what has been my salary during my studies. They include the
Novo Foundation, the EEC Human Capital and Mobility Program, the Danish
Research Academy, the French Embassy in Copenhagen, the SARC Foundation,
the Ib Henriksens Foundation and the Danish Science Foundation, which funds
CATS.
The Danish Ph.D.-reform of 1993 deserves a special comment. It's a historical
example of the utmost stupidity, which has cost me and other Ph.D.-students both
time and money. How can you f.ex. force Ph.D.-students to spend six months
abroad and at the same time remove their possibility of deducting their expenses?
The list of irrational and illogical features of the law is almost endless and clearly
illustrates that a law does not have to make any sense - it only has to be passed
by parliament. The fact that the law is presently being revised shows some sign
of intelligent behaviour and provides hope for the future of the nation.
4
1 Introduction
Many dynamical systems in nature are spatially extended systems, which do not
only interact with their environment but are driven by it. In fact, more or less
all systems on a so-called macroscopic scale are examples of such systems. Fur-
thermore, a large number of such systems are what one usually refers to as \out-
of-equilibrium systems" and what keeps these systems away from equilibrium is
exactly the coupling to an external source of \energy" of some kind. Spatially
extended systems have until recently generally been modelled by partial di er-
ential equations, in short P.D.E.'s. Unfortunately, contrary to equations which
have no explicit space-dependence, known as ordinary di erential equations or
O.D.E.'s, analytic solutions to P.D.E.'s are in general dicult to obtain. One is
then often left with trying to get numerical solutions, replacing scienti c genius
with brute (computer-) force. Using a computer to solve a P.D.E. means replac-
ing the continuous equation with a di erence equation which is discrete in both
space and time, and hoping that the discretisation has not left out any signi cant
physics. The basic assumption is still that of a continuum in space and time and
the discretisation is only a consequence of the binary nature of todays computers.
However, as we shall see in the following chapters, there are many systems which
are indeed discrete spatially or temporally (or both) and where a continuum de-
scription represented by a set of O.D.E.'s or P.D.E.'s may not be appropriate.
The tool that one may use instead and which has become increasingly popular
with the increasing computing-power of computers is cellular automata. Cellu-
lar automata are the simplest possible spatially extended dynamical systems in
which a variable with a numerable (often nite) number of states is de ned on a
discrete space referred to as the \lattice". Speci c examples are the Potts model
to be studied below and the model of epidemics studied in chapter 4. Let us
for completeness mention that the question between discreteness and continuum
can appear in yet another fashion. If the variable instead of having a numerable
number of allowed values has a continuum of states one generally talks about
coupled-map-lattices.
The discussion of how to solve P.D.E.'s of course assumes that it's in the rst
place possible to write down some sort of eld theory for the problem in question.
For many seemingly very simple problems this has proven to be very dicult.
Examples are geometrical models such as di usion-limited-aggregation models,
percolation models besides various spin models. And when the question concerns
the brain, how AIDS spreads in society or the development of earthquakes, it
seems quite impossible to derive a full P.D.E. for the problem.
As we shall repeat several times in this thesis, the power of using cellular
automata to model spatially extended systems is that one can get estimates
of global quantities from very simple assumptions of how the local dynamics
works. In order to illustrate these concepts further, let us consider some of the
systems under investigation here. Two examples come from society: disease
5
spreading and the stock market. Society consists of individuals and is hence
discrete with a nite population. Furthermore, only a nite number of \states"
exist for each individual with respect to the two subjects given above. When it
comes to disease spreading only a few basic states exists: susceptible, infectious,
latent, immune, dead or some combination of these. The stock market is similar.
Here the actors on the stock market only have three possible states: buying,
selling or waiting. Furthermore, one may argue that these two systems have
a discrete time-evolution due to the fact that people sleep and stock markets
close . Also, when it comes to fracturing and earthquakes one may argue that
3

these systems have a discrete nature both spatially and temporally. Spatially,
discreteness comes from the fact that we are dealing with connected systems
of faults/ bres/grains, i.e., heterogeneous systems. The loading is in general
expected to be continuous, but it's the re-distribution of the stress- eld due to
local failure which essentially determines the dynamic of the system and infer
a discrete time evolution. This becomes quite clear in the context of cellular
automata. If we f.ex. take the Burridge-Knopo model or the bre model to be
discussed in chapter 5, then the proper way of simulating the model is of course
not to apply an incremental increase of the stress eld, but instead to increase it
to a value corresponding to the \breaking" value of the weakest element, hence
making time discrete in some sense.

1.1 Imitation and Synchronisation


In reading the history of crowds, we nd that, like individuals, they have their
whims and their peculiarities; their seasons of excitement and recklessness, when
they do not care what they do. We nd that whole communities suddenly x their
minds upon an object and go mad in it's pursuit; that millions of people become
simultaneously impressed with one delusion, and run after it, til their attention is
caught by some new folly more captive than the rst. We see one nation suddenly
seized, from its highest to its lowest members, with a erce desire of military
glory; another suddenly become crazed upon a religious scruple; and neither of
them recovering its senses until it has shed rivers of blood and sowed a harvest of
groans and tears, to be reaped by its posterity. At an early age in the annals of
Europe its populations lost their wits about the sepulchre of Jesus, and crowded
in frenzied multitudes to the Holy Land; another age went mad for the fear of the
devil, .... Men, it has well been said think in herds; it will be seen that they go
mad in herds, while they recover their sense slowly, one by one.
Charles Mackay 1814-1889, Memoirs of Extraordinary Popular Delusions and
the Madness of Crowds. Boston L.C., Page & company, 1932
3 The notion of a discrete time evolution in models of nancial markets has been applied
by f.ex. B. B. Mandelbrot by the use of \trading times" opposed to continuous time, see f.ex.
reference [20] of chapter 3.

6
A special feature of discrete spatially extended systems is the possibility of
imitation and synchronisation. There are numerous examples of collective, or
crowd, behaviour especially in sociology and biology, but also in chemistry and
physics such behaviour can be found. Some examples from society are given
in the quotation above and let us just mention two additional with a direct re-
lationship to the topics considered here. Stock market crashes are examples of
cooperative behaviour, where traders more or less simultaneously lose faith in the
present value of the market and only slowly realize the over-reaction that a crash
represent . Also, out-breaks of epidemics are signs of cooperative behaviour \in
4

which the well imitate the infected, but the infected do not imitate the well -
rather they spontaneously transform back to good health after a recovery time"
[1]. Other examples in biology are the synchronised ashing of re- ies or beating
of pacemaker cells in the heart, the collective movement of sh in a school and
women whose menstrual cycles \phase up" [2]. When we move on to chemistry
and physics we meet the famous Belousov-Zhabotinsky reaction, where a stag-
gering number of molecules synchronise, and the Ising model, which models the
phase-transition seen in ferro-magnets tuned to a critical temperature. More or
less all of these system are governed by threshold dynamics and may be referred
to as relaxation systems in some speci c sense.
On a macroscopic level, relaxation systems can be described as 3-state dy-
namical systems, which are driven out of equilibrium towards a state of collapse
after which they relax and again reach a state of equilibrium. Depending on the
process of collapse, i.e., if it's reversible or if it leads to the destruction of the sys-
tem, this process may repeat itself in the same xed order: build-up, collapse and
relaxation. Furthermore, the process of relaxation is in general much faster than
the time of build-up. This is usually referred to as separation of time scales and
is one of the main di erences between relaxation systems and f.ex. a harmonic
oscillator.
Non-trivial cooperative e ects are often seen when relaxation oscillators are
coupled together in an ensemble and the aim here is not to give a complete re-
view of all the work that has been done in the eld. Instead, we will focus on two
speci c features, namely imitation and synchronisation. A lot of the work on syn-
chronisation has been done in the context of pulse-coupled oscillators. Whereas
the theorem of Mirollo and Strogatz [3] gives sucient conditions for synchro-
nisation to occur in the case of a fully interconnected population of oscillators,
very little is known analytically in the case of short-range interactions between
a few units. The capital idea of Mirollo and Strogatz is that of \absorption",
which means that once two oscillators has become synchronised they will remain
so. Even though the mathematical proof of Mirollo and Strogatz cannot be ex-
4One may of course adapt a converse point of view and regard the surge preceding a stock
market crash as an e ect of traders slowly losing their senses and suddenly recover them with
a crash as consequence, as argued in reference [3] of chapter 3.

7
tended to local interactions, the results obtained in the epidemic model presented
in chapter 4 is an example of a model with only nearest-neighbour interactions,
where synchronisation occur due to such an absorption mechanism.

1.2 Fractals
Many new concepts have been invented with the understanding of complex sys-
tems in statistical physics. Examples are scale invariance, fractals, 1=f -noise etc.
In order to understand the reason for the physics community's great interest in
these phenomena, we should rst consider the statistical properties of some so-
called \ordinary systems". Most spatially extended systems are spatially uniform
on suciently large length scales. Examples are the distribution of molecules in
a gas, defects in a solid, grains in rock-materials, trees in a forest, clouds in the
sky and so forth. In all of these examples the positions of the objects are more or
less random and we refer to them as heterogeneous (or inhomogeneous) systems
in contrast to f.ex. a perfect crystal with ions on a regular lattice. However,
both crystal and forest can be labelled as being spatially uniform in the following
context. If we pick a box of size jxj and measure some quantity of the system,
which could be the total number of objects or the cumulative mass, then this
observable will be proportional to the volume jxjd of the box provided that the
box is suciently large , but do not exceed the size of the system. Dividing with
5

the volume gives us the density, which then will be constant for an uniform distri-
bution of objects, be it random or on a lattice. We can then say that the objects
ll the d-dimensional space they inhabit, they are space- lling. Note, however,
that we have introduced two length scales here: A lower cut-o corresponding
to some \building block", such as the unit cell in a crystal, and an upper cut-o
corresponding to the system size.
Systems, which are not space- lling are also quite common in nature. Exam-
ples are particle trajectories (ballistic), lines and surfaces of natural objects etc.
Here, one or more direction in space is excluded and the distribution of points is
hence restricted in that direction. Using the box-counting method above again
gives that the \mass" no longer grows with the volume of the box jxjd, but in-
stead with jxjD , where D < d is some integer. Take f.ex. a tree. The mass of
the tree grows as Ah, where A is the cross-section of the tree and h is the height
above the ground. This means that the mass of the tree is proportional to the
box-size h and not the volume of the box h . 3

In the past two decades it has become clear that some systems lie in between
the case of an uniform distribution with constant density or systems with an
integer co-dimension d ? D. These spatially extended systems are referred to as
\fractals". If we again use the box-counting method, fractals are characterised
5For a perfect crystal this correspond to the size of the unit cell, which is in the order of a
few 
A, whereas for clouds we talk about 100-1000 km.

8
by the fact that f.ex. the cumulative mass in a box of size jxj will obey a re-
lation M (jxj) / jxjD , where D < d is not an integer. Numerous examples of
fractal structures have been found in nature (geological fault structures and frac-
tures), in the laboratory (viscous ngering and fungus growth) and in simulations
(percolation and di usion-limited-aggregation) [4, 5, 6, 7] to mention a few.
Another way of characterising fractals is by the observation that fractals are
self-similar. Qualitatively, self-similarity means that an object or structure looks
the same regardless whether we look at it through a microscope, a magnifying
glass or just with our eyes. Uniform systems, on the contrary, are not self-
similar and we have no problem of telling if the object has been magni ed or not.
However, it is important to remember that fractals, man-made or in nature, also
have a lower cut-o (as well as an upper cut-o due to the nite size of any real
system) due to some inherent length scale or building block and if we increase
the magni cation of the microscope too much, the self-similarity disappears.

1.3 Hierarchical systems


A special class of self-similar systems are so-called hierarchical systems. The
di erence between a hierarchical system and a fractal is that the latter is con-
tinuous, whereas hierarchical systems are discrete in the sense that a preferred
ratio between structures at di erent scales exists. Whereas hierarchical systems
in a strict mathematical sense are more or less restricted to man-made struc-
tures, such as the Sierpinski gasket shown in gure 1 or the Egyptian pyramids,
many systems in nature and society are hierarchical in a more relaxed sense. A
well-known example from society is the way individuals are organised in various
bodies, such as the military, large companies, the authorities etc. Also other
animals tend to group together in a hierarchical fashion, the young around the
mother, mothers around a dominant male and so forth. In physics, we nd hi-
erarchical structures in relation to branching processes, the cascade of eddies in
turbulence, Titius-Bode law etc. In the following, we will see some examples
of systems which have a hierarchical structure also in a quantitative sense. Let
me mention that I'm well aware of the fact that the basis of the renormalisation
group theory by K.G. Wilson [8] is exactly the existence of a hierarchy of scales.
However, this does not mean that the systems which can be described within the
framework of the renormalisation group theory are hierarchical systems. This,
since no preferred scaling ratio exist. We will return to this question in section
1.5.

1.4 1=f -noise


Also temporal fractals exist in terms of the temporal behaviour of some uctuat-
ing time-signal. Temporal fractals has been coined 1=f -noise (\one-over-f-noise"),

9
which refers to the f ? -behaviour,  1, of the power spectrum for small fre-
quencies.
The most common tool for analysing temporal signals is the power spectrum
or, similarly, the spectral density S (!), de ned as the absolute square of the
Fourier transform of the signal. The reason is that with the invention of the
Fast Fourier transform (FFT) the computer-time used in order to calculate the
power spectrum grows as N ln N whereas the the time used in calculating the
more intuitive (discrete) auto-correlation function
N ?
c ( ) = N ? 1
X 
1
2
x (i) x (i +  ) ? x ; 2
(1)
x i

grows as N , where N is the number of points in the time series. Since the Wiener-
2

Khintchine theorem states that the auto-correlation function is the Fourier trans-
form of the power spectrum the two are equivalent.
Let us consider the power spectrum of a few simple dynamical processes. If
a temporal process is quasi-periodic (truly periodic systems are rarely seen in
nature) with some frequency ! this means that the coecients of the Fourier
series of the time signal
1
X
x (t) = ai cos (!it + i) (2)
i=0
are all approximately 0 except for a few that belong to some narrow interval of
frequencies [! ? ! : ! + !]. This means that the power spectrum will have a
well-de ned peak around !.
For many random processes, the auto-correlation function (as de ned above)
will be 1 for  = 0 and decay very rapidly to 0. This corresponds to a power
spectrum which is constant except for the highest frequencies. (Such a power
spectrum is called \ white", since all frequencies carry equal weight.) F.ex. ther-
mal uctuations have a white power spectrum. Another process with a white
power spectrum is the Brownian motion of a particle. The equation of motion
for a particle of mass m subjected to a uctuating force  (t) as well as viscous
friction is
mx ? x_ =  (t) (3)
h (t)  (t )i =  (t; t ) :
0 0 (4)
Making a variable change y = x_ then give the power spectrum
S (!) / +1 ! : ) c ( ) / e?=c ; c = ?
2 2
1
(5)
of the velocity y. That an exponentially decaying auto-correlation function cor-
responds to a 1=f -behaviour of the power spectrum (other reasons for a 1=f -
2 2

behaviour exists, see f.ex. paper VIII in section 5.6) means that one generally
10
will see such behaviour for small enough frequencies. This since temporal cor-
relations in real systems in general must vanish for suciently long times. This
also explains the reason for the large interest in 1=f -noise, since such behaviour
implies that the auto-correlation function decays algebraically. This means that
long time temporal correlations exist and that systems exhibiting 1=f -noise have
very long \memory" of their past history.
Let us nally mention that natural systems claimed to exhibit 1=f -noise range
from noise in resistors (perhaps the only really convincing natural example) over
river-level uctuations (the Nile) to quasars.

1.5 Scale invariance and discrete scale invariance


The spatial and temporal fractals considered above are consequences of what one
refers to as scale invariance. The concept of scale invariance is quite general.
A system is called scale invariant if some observable of the system o (x) is, in
some speci c sense, left unchanged by a change of scale in the control parameter
x ! x. Speci cally, we demand that there exist a number  (), such that
o (x) =  () o (x) (6)
in order to call a system scale invariant. A perhaps more intuitive de nition of
scale invariance can be reached by the following transformation. If we instead of
x use ln x as control parameter, then a change of scale in x ! x correspond to
nothing but a translation of the new control parameter ln x ! ln  + ln x. The
consequence on the observable is
ln (o (ln x)) ! ln (o (ln x + ln )) + ln ; (7)
which is also nothing but a translation of o. Hence, scale invariance of an observ-
able o (x) is nothing but translational invariance of the logarithm of the observable
ln (o (ln x)) .
6

Why is this then so interesting? A solution to equation (6) is a simple power


law
o (x) = cx , = ? ln  (8)
ln 
and power laws are generally referred to as scale invariant functions. This, since
a change of scale x ! x give x !  x which means that the ratio
o (x) = ? (9)
o (x)
6 Since the logarithm can be thought of as an di eomorphism between additive and multi-
plicative groups, this is not that surprising.

11
Figure 1: The iteration process giving the Sierpinski gasket

is independent of the original value (or scale) of x. If we do the same thing for
f.ex. an exponential e?cx we instead get
x ! x => oo((x x) = exp (x ? x) ;
) (10)
which very much depends on x.
Until now, we have not put any restrictions on , but assumed equation (6)
to be valid for arbitrary choices of . This need not always be the case. F.ex.
on a crystal lattice only translations which are commensurate with the lattice
spacing make any sense. However, since scale invariance of some observable o (x)
can be expressed as translational invariance of the logarithm of the observable,
we are not talking about the usual kind of lattice, as we know it from f.ex. solid
state physics, but that of a hierarchical lattice. An example of such a lattice is
the Sierpinski gasket shown in gure 1.
Here, the hierarchical lattice is generated by a recursive process, where each step
consists of removing the middle triangle starting with a single triangle of unit
length, see gure 1. Going from the i-th iteration to the i + 1-th, we increase the
number of \cells" N by a factor of 3 and the length l by a factor of 2. After n
such iterations, we have l = 2n and N = 3n giving
N (l) = lD ; ln 3  1:585:
D = ln (11)
2
Note, however, that relation (11) only holds for re-scaling of l which are powers
of two and hence the Sierpinski gasket is only invariant or self-similar under such
12
magni cations. What about arbitrary re-scaling factors  then? Well, increasing
l continuously from say 2n to 2n leaves N (l) unchanged until the value l = 2n
+1 +1

is reached and at which it jumps by a factor 3. If we de ne a continuous version


of relation (11)

N (x) = xD x ) D (x) = ln ln
( ) N (x) ; (12)
x
then D (x) will decrease from the value 1:585 for 2n < x < 2n and rst reach
+1

it again as x reaches the value 2n . This means that


+1

 
ln
N (x) = x p ln 2 ;
D x (13)
where p is a periodic function in ln x, i.e., log-periodic in x, with a period of 1.
Writing P in terms of its Fourier series, we get
  1   1
ln x
p ln 2 =
X ln x
ck exp i2k ln 2 =
X
ck xi k= :
2 ln 2
(14)
k ?
= 1
k ? = 1

If we compare this with equation (13), we see that D is replaced by


i2 ;
Dk = D + k ln (15)
2
where k in principle can run between 1. However, since N (x) is real, the
Fourier series is symmetric and ck = c?k and keeping only the rst term in the
expansion, we get
  
c ln x
N (x) = x 1 + 2 c cos 2 ln 2 :
D 1
(16)
0

An alternative way of arriving at equation (15) is to use the condition for scale
invariance equation (6) and its power law solution directly, giving
o (x) =  () o (x) ) cxD = c (x)D ; (17)
which is nothing but
1 = ei k = D ; k 2 N:
2
(18)
If we allow complex solutions to this equation , i.e., k 6= 0, we get exactly equation
(15) with
Dk = ln  + i2k ; (19)
ln  ln 
13
with
D = ln  and  = 2: (20)
ln 
Before we move on to some more physical examples of discrete scale invariance
than that of the Sierpinski gasket, let us recapitulate. From the discussion above
it should be clear that discrete scale invariance of some observable is equivalent
to considering complex solutions to an equation like equation (6). Speci cally,
this corresponds to power laws with complex exponents and, since the observable
o (x) is real, log-periodic oscillations decorating the over-all power law behaviour.
That o (x) is a log-periodic function of some variable x means that we can write
o (x) = f (x) p (ln x) ; (21)
where p is some periodic function of period ln . This means that log-periodicity
is directly related to the existence of a preferred scaling ratio  and that observed
log-periodic structures in data indicate that the responsible processes have a pre-
ferred scale. This is a very important observation, since it provides for additional
constraints on the underlying physics. The explosion of the last decade in the
number of systems which are claimed to exhibit scale invariance means that ap-
parently many di erent mechanism can generate scale invariance . This means
7

that the usefulness of scale invariance as a modelling constraint has become rather
limited. Discrete scale invariance on the other hand provides for very speci c in-
formation on the underlying physical structures.

1.6 Exactly solvable models with discrete scale invariance


1.6.1 Potts model on a hierarchical lattice
Scale invariance have a more general setting than the geometrical interpretation
given above. In critical phenomena, scale invariance is often expressed in terms of
the partition function or similarly the free energy. In f.ex. phase transitions, the
free energy is invariant under a scale change close to the critical point, de ned as
the point where the free energy or one of its derivatives has a singularity. We will
here consider a speci c model, the Potts model on a hierarchical lattice, where
the free energy exhibits log-period oscillations as the critical point is approached.
The Potts model is a q-state spin model on a lattice, where the interaction energy
E between two spins i ; j is de ned as
E = ?J (i ; j ) ; (22)
7A speci c example is the power law distribution of avalanches in the various earthquake
models mentioned in chapter 5

14
Figure 2: The iteration process giving the hierarchical diamond lattice

giving a Hamiltonian
X
H = ?J  (i ; j ) ; (23)
<ij>
where the sum runs over all nearest-neighbour spins. In the hierarchical version
of the model considered here, the spins are positioned on a hierarchical diamond
lattice constructed by the iteration process shown in gure 2. On the n-th level
corresponding to a magni cation of 2n we have (2 + 4n) spins connected by 4n
2
3
bonds. The partition function of that level is de ned as
X
Zn (J ) = e? J=kT  ; ; ( ) ( 0
)
(24)
fg
where the sum runs over all spin con gurations fg of that level. We now want
to apply the real-space renormalisation scheme commonly used on spin-models,
where we decimate the number of spins by replacing the original coupling constant
J between neighbouring spins by a renormalised coupling constant J~ between
next-nearest neighbouring spins. We do this by summing out the contribution of
the two middle spins in units consisting of a single diamond, see g 3.
The contribution from such a unit with 4 spins to the partition function is
e? J=kT  1 ;2  2 ;3  3 ;4  4 ;1 :
( )( ( )+ ( )+ ( )+ (
(25) ))

Summing out the two middle spins  and  gives


2 4
? ?J=kT 
2e +q?2 for  6= 
2
1 (26)
3
? ? J=kT 
e 2
+q?1 for  =  ;
2
1 (27)
3

15
Figure 3: Spin decimation by real-space renormalisation

since we have q possible spins. We can rewrite this as


 ?2J=kT 2 ! !
?
2e?J=kT +q?2
2
1+ e +q?1 ? 1  ( ;  ) (28)
2e ?J=kT +q?2 1 3

This we have to do for all 4 sub-units that a \super-unit" on the next level consists
of, all giving a contribution as the one above. In order to simplify notation we
set X = eJ=kT . We can now write the contribution of the super-unit to the
renormalised partition function as
(2X + q ? 2) X~  1 ;3 ;
2 ( )
(29)
where
 
X~   1 ;3
( ) ~
= 1 + X ? 1  ( ;  ) 1 (30)
3

 ? J=kT 
X~ = 2ee?J=kT ++ qq ?? 12 :
2 2

(31)

Here X~ de nes the renormalised coupling constant J~. We now have the two maps
de ning our renormalisation ow as we move through the di erent levels of the
lattice:
 
X + q ?
Xn? = X + q ? 2   (Xn)
n 1
2 2

(32)
1
n
Zn? (Xn? ) = Zn (Xn) (2Xn + q ? 2)?  ; (33)
n 24
1 1

16
where 4n is the number of bonds we have removed in order to get to the n ? 1-th
level. It's because of the hierarchical structure of the lattice that we can obtain
an exact relation for the renormalisation ow and calculating the initial condition
simply gives
Z (X ) = q (X + q ? 1) :
1 1 1 (34)
Taking the logarithm on both sides of equation (32) for the partition function
and dividing with the number of bonds (the free energy is an extensive quantity)
gives the corresponding recursion relation for the free energy
fn? (Xn? ) = 4fn (Xn) ? 2 ln (2Xn + q ? 2)
1 1 (35)
or in the general form in which it is usually written
 
f X~ = 1 f (X ) + g (X ) ; (36)
where g is a non-singular function.
Returning to the RG-map (31) for the coupling, we see that it has two trivial
xed points X , which are independent of q,
1
X = 1 (37)
corresponding to the two trivial situations with either totally de-coupled spins
or a situation where all spins have the same value. Besides the two trivial xed
points, we have a critical xed point Xc which depends on q
3:38 ; q=2
4 ; q=3
Xc  4:54 ; q=4 : (38)
5:03 ; q=5
:::
Here we have only kept the real-valued solutions. The crucial di erence between
the two trivial xed points and the critical xed point can be found in the eigen-
values of the RG-map linearised close to the xed points. As usual, the two trivial
xed points are attractive meaning that if we start above or below Xc then we
will end up in the corresponding xed point after in principle in nitely many iter-
ation of the RG-map. We can regard the two regimes belonging to X  = 1 and
X  = 1 as a high-temperature regime, where the spins de-couple due to thermal
uctuations, and a low-temperature where the interaction energy between spins
dominates and where a majority of spins will have the same value. On the other
hand, the xed point Xc is repulsive.

17
It can be shown [9] that equations (32) and (35) may be solved for the free
energy by
1
f (X ) =
X 1 g  n (X )
( )
(39)
n=0
n
and that this sum is singular for X = Xc. This, since the repulsive nature of the
xed point Xc means that the derivative of the map  is larger than 1. Quite
generally, this implies that close to Xc we have
f (K ) / (Xc ? X )z : (40)
Using this, equation (35) and the fact that we can linearise equation(32) close to
Xc, i.e.,
 (X )  Xc +  (Xc ? X ) ; (41)
gives
(Xc ? X )z  1 ( (Xc ? X ))z ) z = : (42)

The general solution to this equation is complex, which means that the free energy
close to the critical point can be written in the general form
1
X
f (X )  (Xc ? X ) z
cn cos (!n ln (Xc ? X ) + n) ; (43)
n=0

where
z = ln  and ! = 2n : (44)
ln  ln 
Hence, the free energy of the model exhibits log-periodic oscillations as the critical
value for the coupling constant is approached. Unfortunately, the amplitude of
the oscillations found in this model are only on the order of 10? compared to4

the leading behaviour [9] and, as we shall see later, much smaller than the results
obtained from the analysis of numerical and real data, to be presented in the
following. The small amplitude of the oscillations is not the only problem with
the hierarchical Potts model, when we wish to discuss model-mechanism(s) for
log-periodicity. The model above, as well as most other spin-models, are quasi-
static equilibrium models where the state of the model is completely determined
by the value of the control parameter. Furthermore, the singular behaviour of
the free energy stems from the fact that the RG-map for  in equation (32)
close to the xed point Xc has derivatives larger than 1, i.e., Xc is repulsive.
This is the general case in critical phenomena and means that it is necessary
18
to force or tune the coupling constant (or whatever parameter that controls the
critical behaviour) to its critical value. Hence, we need an additional ingredient
in order to have a dynamical model, which reaches the critical point due to some
dynamical process. The concept of self-organized criticality has been suggested
as a possible generic mechanism for the existence of scale invariance in nature
[10]. We will not discuss the relevance of self-organized criticality to the existence
of scale invariance and discrete scale invariance in nature here. Instead, we note
that a rm theoretical framework for the concept of self-organized criticality for
non-conservative systems is still missing and that the progress in the eld to
a large part consists of analysis of various numerical models all characterised
by a very slow driving. We refer the interested reader to paper X reproduced
in appendix A for a discussion of whether the models claimed to exhibit self-
organized criticality are truly self-organising or not. See also [11]. For a more
main stream treatment of the concept of self-organized criticality the reader may
consult [2] and the references therein.
Another problem with the Potts model is that the hierarchical structure of the
lattice responsible for the log-periodic correction to the usual power law behaviour
close to the critical point is build into the de nition of the model. What we
need is a simple physical model, where the hierarchical structure evolves as a
consequence of the dynamics of the model and is not put in by hand. Whereas,
we make no claim to have answered the rst objection concerning the amplitude
of the oscillations, we will in chapter 2 see examples of dynamically generated
hierarchical structures.
In spite of the short-comings of the model described above, we believe it
provides valuable information as an example of an exactly solvable model with
complex exponents and log-periodic oscillations due to discrete scale invariance.
1.6.2 Di usion on a random lattice
Another exactly solvable dynamical model, which exhibit log-periodic oscillations
in a physical observable is the random walk model presented in [12]. Consider a
random walker on a 1-d lattice and assume that the bonds between two adjacent
sites acts like \diodes" such that they only allow movements in one direction. The
dynamics of this model is of course trivial, since a walker can't jump the barriers
and hence will get trapped more or less instantly. Let us therefore assume that a
small fraction p of the diodes which allow jumps to, say, the left, has been slightly
damaged and will also allow jumps to the right with a small probability q. This
means that the two parameters of the model are assumed to full- ll the relation
q  p  1: (45)
For parameter values chosen according to inequality (45), very clear log-periodic
oscillations are found in the mean displacement hx (t)i as a function of time t,
see gure 4.
19
Figure 4: hx (t)i t? = as a function of time t. Reproduced from [12].
1 2

The exact solution of this problem can be found for a more general choice
of parameter values than those given by inequality (45), see [12]. Since the
derivation is rather lengthy, we will not repeat it here. Instead we will present a
less rigorous derivation based on inequality (45).
As long as the random walker do not meet one of the damaged diodes he
will proceed with a constant speed to the right. After some time, he will meet a
cluster of damaged diodes of size k that he has to pass. The number of steps tk
needed to do so can be estimated as
 k
tk  q 1 ? q  k ; (46)
where the ratio is the average number of jumps to the left per jump to the right
and hence the average transmission rate to the right. The approximation made in
saying that the transmission rate of k \ lters" is the transmission rate of a single
\ lter" to the power k assuming a small transmission rate is well-known from f.ex.
optics. The density of clusters of k damaged diodes is pk (1 ? p) which means
2

that the average distance between two such clusters of size k can be approximated
by p?k . Since we have assumed q  p, this means that the time-of-passage of the
rst clusters of size k is completely dominated by tk . Within this picture, this
means that multiplying (renormalising) tk with  gives the mean passage time of
the next cluster of size k + 1, whose average distance is a factor p? further away.
1

Hence, renormalising time by  correspond to renormalising space by p. We can


thus write down an approximate RG-equation for the displacement
x (t)  px (t) + g (t) : (47)
where g (t) is some regular function taking into account everything the rst term
of equation (47) has missed. This equation is of the same kind considered in
20
the previous section and as previously only discrete re-scalings in factors of 
are de ned. To proceed from here, we assume as usual that we can neglect the
regular term. This leads to a power law solution of equation (47),
x (t)  t ) (48)
t  (t) p;
 
(49)
and using the previous trick 1 = ei n nally gives
2

x (t)  t P (! ln t) ; (50)
where
ln p and ! = 2 :
v = ln (51)
 log
In fact, this recovers the exact solution of [12]. Furthermore, if we return to gure
4, we see a good agreement between numerical simulations and the prediction for
 and ! above. Inserting the parameter values used in the simulations shown
gives  = 0:5 and ! = 2:6, which is really remarkably close to what can be
extracted from the data. This especially considering that the parameter values
used in the numerical simulations,  = 0:09 and p = 0:3, are quite far from
full- lling inequality (45).
That time is discrete in the way we have de ned the model above is not
important. What is essential in generating the log-periodic oscillations is the dis-
crete nature of the lattice. This means that the above model gives an interesting
example of how a hierarchical space (in terms of the distance between clusters)
result in hierarchical temporal scales due to a multiplicative equivalence between
space and time.
Also, random walks with a xed bias direction on randomly diluted lattices in
3 dimensions with densities far above the percolation threshold show log-periodic
oscillations in the e ective exponent k,
2
r / tk ; (52)
versus time [13]. Here hr i is the mean square displacement of the walk. The
2

physical mechanism responsible for the observed log-periodic signatures is similar


to that of the one-dimensional situation discussed above. As the walker moves
through the lattice he will get trapped in dead ends of depths n with a trapping
time that increases exponentially with n [13] similarly to the Arrhenius factor in
statistical physics.
1.6.3 An experimental hierarchical model
Log-periodic oscillation have been observed experimentally as well [14]. Speci -
cally, the magneto-resistance of Al-wires organised on a Sierpinski gasket exhibits
log-periodic oscillations as a function of the magnetic eld, see gure 5.
21
Figure 5: Logarithmic derivative of magneto-resistance. The left gure is experi-
mental data and the right the corresponding RG calculation. Reproduced from
[14].

1.7 Possible log-periodic signatures in cracking8


The stress eld locally at the tip of a crack of a linear elastic solid can be written
as
ij (r; ) = p 1 KI fijI () + KII fijII () + KIII fijIII () :
? 
(53)
2r
Here the Roman indices refers to the symmetric (with respect to the crack plane)
in-plane stress eld (mode I ), the anti-symmetric in-plane stress eld (mode II )
and the anti-symmetric out-of-plane stress eld (mode III ), respectively, see
gure 6.
The functions fijI;II;III are normalised
f I = f II = f III = 1 for  = 0;
22 12 23 (54)
so that we have the usual interpretation of the the stress intensity factors in front
of the crack tip on the plane of the crack
p
(KI ; KII ; KIII ) = 2r ( ;  ;  ) : 22 12 23 (55)
The relative crack face displacements behind the crack tip can be written as
r  
32r  ? 
E
( ;  ;  ) =
 E KII ; KI ; 2 KIII (56)
1
1 2 3

E = E= (1E? plane stress


) plane strain :
2 (57)

22
Figure 6: Mode I, II and III loading

Here E is Young's modulus and  Poisson's ratio. Satisfying the compati-


bility (or continuity) conditions is equivalent to requiring that the Airy stress
9

function  satis es the biharmonic equation


10

r  = 0; 4
(58)
where
rr = 1r r + r1 
2

2 2
(59)
 
r = ? r 1r  (60)

 = r :
2

2
(61)
 has separable solutions of the form
 = Kr  [C cos (( ? 1) ) + C sin (( ? 1) ) +
1+
1 2

C cos (( + 1) ) + C sin (( + 1) )] ;


3 4 (62)
8 This section is based on notes belonging to a Ph.D.-course at Denmark's Technical Univer-
sity and is written by Henrik Myhre Jensen.
9 It simply states that knowing the displacement at some point ( ) and calculating the
x; y; z

displacement at some point ( 0


x ;y ;z
0
) by integrating the strain eld over some path between
0

the two points, then this integral should be independent of the chosen path.
10 Deriving stresses from Airy's stress function ensures that the equilibrium equations in
absence of volume forces are automatically satis ed.

23
where K depends on load and geometry. The Ci's and  are all real and can
be determined by the boundary conditions for the displacement r ;  and the
stress eld rr ; r ;  speci c to the problem in question. This gives a set linear
equation s
A  C~ = 0; (63)
with a characteristic equation det(A) = 0 determining . In the case of a crack
lying in a bi-material interface complex values for the Ci's and  are found [15].
For the stress eld the characteristic equation becomes
? 
(( )  + 1) +  ( + 1) ? 4 (  +  ) (  +  ) sin () sin () (64)
1 2 2 1
2
1 2 2 2 1 2
2 2

= 0;
giving
 = n +n ++1 i ; n = 0; 1; 2; :::;
1
2 (65)
where subscript 1 and 2 refers to the two materials, respectively, and  is the
bi-material constant
 = 2r 1 ln 1 ? ; (66)
1+
=  (( ?
1 1) ?  ( ? 1) :
2

+ 1) +  ( + 1)
2 1
(67)
1 2 2 1

The dominant singularity is given by  = 1=2 + i. The stress at the tip of the
crack lying at the bi-material interface is then given by
ij (r; ) = p 1 Re Kri fijI (; ) + Im Kri fijII (; ) + fijIII () (68)
? ?  ?  
2r
K = KI + iKII :
This means that the stress eld exhibits log-periodic oscillations as a function of
the distance r to the crack tip in a bi-material interface when subjected to mode
I and II loading. The stress eld at an interface crack between two anisotropic
materials or at an interface ending in a free surface is usually oscillatory as well.
Since the energy release is determined by the work done by the stresses opening
the crack one may speculate on whether log-periodic signatures can be present in
acoustic emissions from cracking. An experimental investigation of this possibility
was, as mentioned in the preface, one of the original purposes of my Ph.D.-studies
at CATS.
Yet an example possible log-periodic signatures in cracking has been provided
by Ball and Blumenfeld [16]. In their theoretical analysis, logarithmic oscillations
24
in quasi-static crack growth was predicted and it constitutes perhaps one of the
rst examples of log-periodic oscillations in non-tree structures. From linear
stability analysis of the stress eld around a two-dimensional wedge subjected to
combined mode I and II loading, they found an instability where the local stress
eld is large. This is conceptually similar to the instability in DLA of locally
high gradients of the eld near the interface, see also chapter 2. They argued
that because of this instability the dynamics may couple to sub-dominant terms
in the stress eld and generate branches. Their argument was that, since these
sub-dominant terms exhibit logarithmic oscillations, so will the branching in the
cracking pattern resulting in a growing dendrite with logarithmically periodic
branches.

25
References
[1] E. Callen & D. Shapero, 1974, Physics Today, July, p. 23 (1974).
[2] A. Corral, Complex Behaviour in Slowly Driven Dynamical System: Sand-
piles, Earthquakes, Biological Oscillators ..., Ph.D.-thesis, Institute of
Physics, University of Barcelona (1997).
[3] R.E. Mirollo and S.H. Strogatz, SIAM J. Appl. Math, vol. 50, p. 1645 (1990)
[4] B.B. Mandelbrot, The fractal geometry of nature, San Francisco: W.H. Free-
man (1982).
[5] J. Feder Fractals, New York: Plenum Press (1988).
[6] D. Stau er & A. Aharony, Introduction to Percolation Theory, London,
Washington DC: Taylor & Francis (1992).
[7] P. Meakin, Prog. Solid St. Chem. 20, p. 135 (1990).
[8] K.G. Wilson and J. Kogut, Physics Reports 12, p. 75 (1974).
[9] B. Derrida, L. De Seze and C. Itzykson, J. Stat. Phys. 33, 559 (1983); B.
Derrida, C. Itzykson and J.M. Luck, Commun.Math.Phys. 94, p. 115 (1984)
[10] P. Bak, How nature works: the science of self-organised criticality New York:
Copernicus (1996).
[11] Didier Sornette and Ivan Dornic, Phys. Rev. E, 54, 3334-3338 (1996).
[12] Bernasconi J. and W.R. Schneider, J.Phys.A15, p. L729 (1982)
[13] Dietrich Stau er and Didier Sornette, http://xyz.lanl.gov/ps/cond-
mat/9712085
[14] B. Doucot et al. , Phys. Rev. Lett., vol. 57, p. 1235 (1986)
[15] M. L. Williams, Bull. Seismo. Soc. 49, p. 199 (1959).
[16] R.C. Ball and R. Blumenfeld, Phys. Rev. Lett. 65, p. 1784 (1990); Reply,
Phys. Rev. Lett. 68, p. 2254 (1992).

26

You might also like