See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/227183242

Fractals and Economics

Chapter · November 2010

DOI: 10.1007/978-1-4419-7701-4_25

CITATIONS READS

2 4,410

2 authors:

Misako Takayasu Hideki Takayasu

Tokyo Institute of Technology Sony Computer Science Laboratories, Inc.
236 PUBLICATIONS   3,632 CITATIONS    325 PUBLICATIONS   6,249 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Appearance of Unstable Monopoly State Caused by Selective and Concentrative Mergers in Business Networks View project

Econo-Socio-Physics View project

All content following this page was uploaded by Misako Takayasu on 29 January 2014.

The user has requested enhancement of the downloaded file.

444 Fractals and Economics

Fractals and Economics

MISAKO TAKAYASU1 , HIDEKI TAKAYASU2
1 Tokyo Institute of Technology, Tokyo, Japan
2 Sony Computer Science Laboratories Inc, Tokyo, Japan

Article Outline
Glossary
Definition of the Subject
Introduction
Examples in Economics
Basic Models of Power Laws
Market Models
Income Distribution Models Fractals and Economics, Figure 1
Future Directions Fractured pieces of plaster fallen on a hard floor (provided by H.
Bibliography Inaoka)

Glossary fractured pieces of plaster fallen on a hard floor. There are

Fractal An adjective or a noun representing complex con- several large pieces, many middle size pieces and countless
figurations having scale-free characteristics or self- fine pieces. If you have a microscope and observe a part
similar properties. Mathematically, any fractal can be of floor carefully then you will find in your vision several
characterized by a power law distribution. large pieces, many small pieces and countless fine pieces,
Power law distribution For this distribution the proba- again in the microscopic world. Such scale-invariant na-
bility density is given by a power law, p(r) D c  r˛1 , ture is the heart of the fractal. There is no explicit defini-
where c and ˛ are positive constants. tion on the word fractal, it generally means a complicated
Foreign exchange market A free market of currencies, scale-invariant configuration.
exchanging money in one currency for other, such as Scale-invariance can be defined mathematically [42].
purchasing a United States dollar (USD) with Japanese Let P( r) denote the probability that the diameter of
yen (JPY). The major banks of the world are trading a randomly chosen fractured piece is larger than r, then
24 hours and it is the largest market in the world. this distribution is called scale-invariant if this function
satisfies the following proportional relation for any posi-
tive scale factor  in a considering scale range:
Definition of the Subject
Market price fluctuation was the very first example of frac- P( r) / P( r) : (1)
tals, and since then many examples of fractals have been The proportional factor should be a function of , so we
found in the field of Economics. Fractals are everywhere can re-write Eq. (1) as
in economics. In this article the main attention is focused
on real world examples of fractals in the field of economics, P( r) D C()P( r) : (2)
especially market properties, income distributions, money
Assuming that P( r) is a differentiable function, and dif-
flow, sales data and network structures. Basic mathemat-
ferentiate Eq. (2) by , and then let  D 1.
ics and physics models of power law distributions are re-
viewed so that readers can start reading without any spe- rP0 ( r) D C 0 (1)P( r) (3)
cial knowledge.
As C 0 (1) is a constant this differential equation is readily
integrated as
Introduction 0 (1)
P( r) D c0 r C : (4)
Fractal is the scientific word coined by B.B. Mandel-
brot in 1975 from the Latin word fractus, meaning “frac- P( r) is a cumulative distribution and it is a non-in-
tured” [25]. However, fractal does not directly mean frac- creasing function in general, the exponent C 0 (1) can be re-
ture itself. As an image of a fractal Fig. 1 shows a photo of placed by ˛ where ˛ is a positive constant. Namely, from
 Fractals and Economics 445

the scale-invariance with the assumption of differentiabil-

ity we have the following power law:

P( r) D c0 r˛ : (5)

The reversed logic also holds, namely for any power law
distribution there is a fractal configuration or a scale-in-
variant state.
In the case of real impact fracture, the size distribution
of pieces is experimentally obtained by repeating sieves of Fractals and Economics, Figure 2
various sizes, and it is empirically well-known that a frac- Sierpinski gasket
tured piece’s diameter follows a power law with the expo-
nent about ˛ D 2 independent of the details about the
material or the way of impact [14]. This law is one of the for any fractal phenomena including shapeless quantities.
most stubborn physical laws in nature as it is known to In such cases the power law exponent is the most impor-
hold from 106 m to 105 m, from glass pieces around us to tant quantity for quantitative characterization of fractals.
asteroids. From theoretical viewpoint this phenomenon is According to Mandelbrot’s own review on his life the
known to be described by a scale-free dynamics of crack concept of fractal was inspired when he was studying eco-
propagation and the universal properties of the exponent nomics data [26]. At that time he found two basic prop-
value are well understood [19]. erties in the time series data of daily prices of New York
Usually fractal is considered geometric concept intro- cotton market [24]:
ducing the quantity fractal dimension or the concept of (A) Geometrical similarity between large scale chart and
self-similarity. However, in economics there are very few an expanded chart.
geometric objects, so, the concept of fractals in economics (B) Power law distribution of price changes in a unit time
are mostly used in the sense of power law distributions. interval, which is independent of the time scale of the
It should be noted that any geometrical fractal object unit.
accompanies a power law distribution even a determinis-
tic fractal such as Sierpinski gasket. Figure 2 shows Sier- He thought such scale invariance in both shape and
pinski gasket which is usually characterized by the fractal distribution is a quite general property, not only in price
dimension D given by charts but also in nature at large. His inspiration was cor-
rect and the concept of fractals spread over physics first
log 3 and then over almost all fields of science. In the history of
DD : (6)
log 2 science it is a rare event that a concept originally born in
economics has been spread widely to all area of sciences.
Paying attention to the distribution of length r of white tri- Basic mathematical properties of cumulative distribu-
angles in this figure, it is easy to show that the probability tion can be summarized as follows (here we consider dis-
that a randomly chosen white triangle’s side is larger than tribution of non-negative quantity for simplicity):
r, P( r), follows the power law,
1. P( 0) D 1 , P( 1) D 0.
˛ log 3 2. P( r) is a non-increasing function of r.
P( r) / r ; ˛DDD : (7)
log 2 3. The probability density is given as p(r)   drd P( r).
As for power law distributions there are three peculiar
Here, the power law exponent of distribution equals to the
characteristics:
fractal dimension; however, such coincidence occurs only
4. Difficulty in normalization. Assuming that P( r) D
when the considering distribution is for a length distribu-
c0 r˛ for all in the range 0  r < 1, then the nor-
tion. For example, in Sierpinski gasket the area s of white
malization factor c0 must be 0 considering the limit
triangles follow the power law,
of r ! 0. To avoid this difficulty it is generally as-
log 3 sumed that the power law does not hold in the vicinity
P( s) / s ˛ ; ˛D : (8) of r D 0. In the case of observing distribution from real
log 4
data there are naturally lower and upper bounds, so this
The fractal dimension is applicable only for geometric difficulty should be necessary only for theoretical treat-
fractals, however, power law distributions are applicable ment.
446 Fractals and Economics

5. Divergence R 1of moments. As for moments defined by telling non-stationary situation is to check the power law
hr n i  0 r n p(r)dr, hr n i D 1 for n  ˛. In the exponent of S( f ) / f 1Cˇ in the vicinity of f D 0, for
special case of 2  ˛ > 0 the basic statistical quantity, 0 > ˇ the time series is non-stationary.
the variance, diverges,  2  hr2 i  hri2 D 1. In the Three basic examples of fractal time series are the fol-
case of 1  ˛ > 0 even the average can not be defined lowings:
as hri D 1.
6. Stationary or non-stationary? In view of the data anal- 1. White noise. In the case that fx(t)g is a stationary inde-
ysis, the above characteristics of diverging moments pendent noise, the autocorrelation is given by the Kro-
is likely to cause a wrong conclusion that the phe- necker’s function, C(T) D ıT , where
nomenon is non-stationary by observing its averaged (
1; T D 0
value. For example, assume that we observe k samples ıT D
fr1 ; r2 ; : : : ; r k g independently from the power law dis- 0; T ¤ 0:
tribution with the exponent, 1  ˛ > 0. Then, the sam-
ple average, hri k  1k fr1 C r2 C    C r k g, is shown to The corresponding power spectrum is S( f ) / f 0 . This
diverge as, hri k / k 1/˛ . Such tendency of monotonic case is called white noise from an analogy that super-
increase of averaged quantity might be regarded as a re- position of all frequency lights with the same amplitude
sult of non-stationarity, however, this is simply a gen- make a colorless white light. White noise is a plausible
eral property of a power law distribution. The best way model of random phenomena in general including eco-
to avoid such confusion is to observe the distribution nomic activities.
2. Random walk. This is defined by summation of a white
directly from the data. Pt
noise, X(t) D X(0) C sD0 x(s), and the power spec-
Other than the power law distribution there is another im- trum is given by S( f ) / f 2 . In this case the autocor-
portant statistical quantity in the study of fractals, that is, relation function can not be defined because the data is
the autocorrelation. For given time series, fx(t)g, the auto- non-stationary. Random walks are quite generic mod-
correlation is defined as, els widely used from Brownian motions of colloid to
market prices. The graph of a random walk has a frac-
hx(t C T)x(t)i  hx(t)i2 tal property such that an expansion of any part of the
C(T)  ; (9)
hx(t)2 i  hx(t)i2 graph looks similar to the whole graph.
3. The 1/f noise. The boundary of stationary and non-
where h   i denotes an average over realizations. The au- stationary states is given by the so-called 1/f noise,
tocorrelation can be defined only for stationary time series S( f ) / f 1 . This type of power spectrum is also widely
with finite variance, in which any statistical quantities do observed in various fields of sciences from electrical
not depend on the location of the origin of time axis. circuit noise [16] to information traffics in the Inter-
For any case, the autocorrelation satisfies the following net [53]. The graph of this 1/f noise also has the fractal
basic properties, property.
1. C(0) D 1 and C(1) D 0
2. jC(T)j  1 for any T  0.
Examples in Economics
3. R 1 Wiener–Khinchin theorem holds, C(T) D
The
In this chapter fractals observed in real economic ac-
0 S( f ) cos 2
f d f , where S(f ) is the power spectrum
defined by S( f R)  hbx( f )b
x( f )i, with the Fourier trans- tivities are reviewed. Mathematical models derived from
form, b
x( f )  x(t)e2 i f t dt. these empirical findings will be summarized in the next
chapter.
In the case that the autocorrelation function is charac- As mentioned in the previous chapter the very first ex-
terized by a power law, C(T) / T ˇ , ˇ > 0, then the ample of a fractal was the price fluctuation of the New
time series fx(t)g is said to have a fractal property, in the York cotton market analyzed by Mandelbrot with the daily
sense that the autocorrelation function is scale-indepen- data for a period of more than a hundred years [24]. This
dent for any scale-factor,  > 0, C(T) / C(T). In the research attracted much attention at that time, however,
case 1 > ˇ > 0 the corresponding power spectrum is there was no other good market data available for scientific
given as S( f ) / f 1Cˇ . analysis, and no intensive follow-up research was done un-
The power spectrum can be applied to any time se- til the 1990s. Instead of earnest scientific data analysis arti-
ries including non-stationary situations. A simple way of ficial mathematical models of market prices based on ran-
 Fractals and Economics 447

dom walk theory became popular by the name of Financial

Technology during the years 1960–1980.
Fractal properties of market prices are confirmed with
huge amount of high resolution market data since the
1990s [26,43,44]. This is due to informationization of fi-
nancial markets in which transaction orders are processed
by computers and detail information is recorded auto-
matically, while until the 1980s many people gathered
at a market and prices are determined by shouting and
screaming which could not be recorded. Now there are
more than 100 financial market providers in the world
and the number of transacted items exceeds one million.
Namely, millions of prices in financial markets are chang-
ing with time scale in seconds, and you can access any mar-
ket price at real time if you have a financial provider’s ter-
minal on your desk via the Internet.
Among these millions of items one of the most repre-
sentative financial markets is the US Dollar-Japanese Yen
(USD-JPY) market. In this market Dollar and Yen are ex-
changed among dealers of major international banks. Un-
like the case of stock markets there is no physical trad-
ing place, but major international banks are linked by
computer networks and orders are emitted from each
dealer’s terminal and transactions are done at an electronic
broking system. Such a broking system and the computer
networks are provided by financial provider companies
like Reuters.
The foreign exchange markets are open 24 hours and
deals are done whenever buy- and sell-orders meet. The Fractals and Economics, Figure 3
minimum unit of a deal is one million USD (called a bar), Dollar-Yen rate for 13 years (Top). Dark areas are enlarged in the
and about three million bars are traded everyday in the following figure [30]
whole foreign exchange markets in which more than 100
kinds of currencies are exchanged continuously. The to- previous chapter. This geometrical fractal property can be
tal amount of money flow is about 100 times bigger than found in any market, so that this is a very universal market
the total amount of daily world trade, so it is believed that property.
most of deals are done not for the real world’s needs, but However, it should be noted that this geometrical frac-
they are based on speculative strategy or risk hedge, that tal property breaks down for very short time scale as typi-
is, to get profit by buying at a low price and selling at cally shown in Fig. 4. In this figure the abscissa is 10 min-
a high price, or to avoid financial loss by selling decreas- utes range and we can observe each transaction separately.
ing currency. Obviously the price up down is more zigzag and more dis-
In Fig. 3 the price of one US Dollar paid by Japanese crete than the large scale continuous market fluctuations
Yen in the foreign exchange markets is shown for 13 shown in Fig. 3. In the case of USD-JPY market the time
years [30]. The total number of data points is about 20 scale that this breakdown of scale invariance occurs typi-
million, that is, about 10 thousand per day or the aver- cally at time scale of several hours.
aged transaction interval is seven seconds. A magnified The distribution of rate change in a unit time (one
part of the top figure for one year is shown in the second minute) is shown in Fig. 5. Here, there are two plots of cu-
figure. The third figure is the enlargement of one month mulative distributions, P(> x) for positive rate changes
in the second figure. The bottom figure is again a part of and P(> jxj) for negative rate changes, which are al-
the third figure, here the width is one day. It seems that at most identical meaning that the up-down symmetry of
least the top three figures look quite similar. This is one of rate changes is nearly perfect. In this log–log plot the es-
the fractal properties of market price (A) introduced in the timated power law distribution’s exponent is 2.5. In the
448 Fractals and Economics

Fractals and Economics, Figure 4

Market price changes in 10 minutes

Fractals and Economics, Figure 6

USD-JPY exchange rate for a week (top) Rate changes smaller
than 2 are neglected (middle) Rate changes larger than 2 are
neglected (bottom)

any market. Up-down symmetry also holds universally in

short time scale in general, however, for larger unit time
Fractals and Economics, Figure 5 the distribution of price changes gradually deforms and
Log–log plot of cumulative distribution of rate change [30] for very large unit time the distribution becomes closer to
a Gaussian distribution. It should be noted that in special
cases of market crashes or bubbles or hyper-inflations the
original finding of Mandelbrot, (B) in the previous chap- up-down symmetry breaks down and the power law dis-
ter, the reported exponent value is about 1.7 for cotton tribution is also likely to be deformed.
prices. In the case of stock markets power laws are con- The autocorrelation of the time sequence of price
firmed universally for all items, however, the power expo- changes generally decays quickly to zero, sometimes ac-
nents are not universal, taking value from near one to near companied by a negative correlation in a very short time.
five, typically around three [15]. Also the exponent values This result implies that the market price changes are ap-
change in time year by year. parently approximated by white noise, and market prices
In order to demonstrate the importance of large fluc- are known to follow nearly a random walk as a result.
tuations, Fig. 6 shows a comparison of three market prices. However, market price is not a simple random walk. In
The top figure is the original rate changes for a week. The Fig. 7 the autocorrelation of volatility, which is defined
middle figure is produced from the same data, but it is con- by the square of price change, is shown in log–log scale.
sisted of rate changes of which absolute values are larger In the case of a simple random walk this autocorrela-
than 2, that is, about 5 % of all the data. In the bottom tion should also decay quickly. The actual volatility au-
curve such large rate changes are omitted and the residue tocorrelation nearly satisfies a power law implying that
of 95 % of small changes makes the fluctuations. As known the volatility time series has a fractal clustering property.
from these figures the middle figure is much closer to the (See also Fig. 31 representing an example of price change
original market price changes. Namely, the contribution clustering.)
from the power law tails of price change distribution is Another fractal nature of markets can be found in the
very large for macro-scale market prices. intervals of transactions. As shown in Fig. 8 the transac-
Power law distribution of market price changes is tion intervals fluctuate a lot in very short time scale. It is
also a quite general property which can be confirmed for known that the intervals make clusters, namely, shorter in-
 Fractals and Economics 449

Fractals and Economics, Figure 7

Autocorrelation of volatility [30]

Fractals and Economics, Figure 10

Income distribution of companies in Japan

also a power law [35]. A company’s income is roughly

given by subtraction of incoming money flow minus out-
going money flow, which can take both positive and nega-
tive values. There are about six million companies in Japan
and Fig. 10 shows the cumulative distribution of annual
Fractals and Economics, Figure 8 income of these companies. Clearly we have a power law
Clustering of transaction intervals
distribution of income I with the exponent very close to
1 in the middle size range, so-called the Zipf’s law,

P(> I) / I ˇ ; ˇ D 1: (10)

Although in each year every company’s income fluctuates,

and some percentage of companies disappear or are newly
born, this power law is known to hold for more than 30
years. Similar power laws are confirmed in various coun-
tries, the case of France is plotted in Fig. 11 [13].
Observing more details by categorizing the companies,
it is found that the income distribution in each job cate-
gory follows nearly a power law with the exponent depend-
ing on the job category as shown in Fig. 12 [29]. The impli-
Fractals and Economics, Figure 9 cation of this phenomenon will be discussed in Sect. “In-
Power spectrum of transaction intervals [50] come Distribution Models”.
A company’s size can also be viewed by the amount of
whole sale or the number of employee. In Figs. 13 and 14
tervals tend to gather. To characterize such clustering ef- distributions of these quantities are plotted [34]. In both
fect we can make a time sequence consisted of 0 and 1, cases clear power laws are confirmed. The size distribution
where 0 denotes no deal was done at that time, and 1 de- of debts of bankrupted companies is also known to follow
notes a deal was done. The corresponding power spectrum a power law as shown Fig. 15 [12].
follows a 1/f power spectrum as shown in Fig. 9 [50]. A power law distribution can also be found in per-
Fractal properties are found not only in financial mar- sonal income. Figure 16 shows the personal income dis-
kets. Company’s income distribution is known to follow tribution in Japan in a log–log plot [1]. The distribution
450 Fractals and Economics

Fractals and Economics, Figure 11

Income distribution of companies in France [13]

Fractals and Economics, Figure 14

The distribution of employee numbers [34]

Fractals and Economics, Figure 12

Income distribution of companies in each category [29]

Fractals and Economics, Figure 15

The size distribution of debts of bankrupted companies [12]

is clearly separated into two parts. The majority of peo-

ple’s incomes are well approximated by a log-normal dis-
tribution (the left top part of the graph), and the top few
percent of people’s income distribution is nicely charac-
terized by a power law (the linear line in the left part of the
graph). The majority of people are getting salaries from
companies. This type of composite of two distributions
is well-known from the pioneering study by Pareto about
100 years ago and it holds in various countries [8,22].
Fractals and Economics, Figure 13 A typical value of the power exponent is about two,
The distribution of whole sales [34] significantly larger than the income distribution of com-
 Fractals and Economics 451

Fractals and Economics, Figure 16 Fractals and Economics, Figure 17

Personal income distribution in Japan [1] The distribution of the amount of transferred money [21]

panies. However, the exponent of the power law seems to

be not universal and the value changes county by country
or year by year. There is a tendency that the exponent is
smaller, meaning more rich people, when the economy is
improving [40].
Another fractal in economics can be found in a net-
work of economic agents such as banks’ money transfer
network. As a daily activity banks transfer money to other
banks for various reasons. In Japan all of these interbank
money transfers are done via a special computer network
provided by the Bank of Japan. Detailed data of actual
money transfer among banks are recorded and analyzed
for the basic study. Fractals and Economics, Figure 18
The total amount of money flow among banks in The number distribution of active links per site [20]
a day is about 30  1012 yen with the number of trans-
actions about 10 000. Figure 17 shows the distribution of
the amount of money at a transaction. The range is not link number distribution follows a power law. In the ter-
wide enough but we can find a power law with an expo- minology of recent complex network study, this property
nent about 1.3 [20]. is called the scale-free network [5]. The scale-free net-
The number of banks is about 600, so the daily trans- work structure among these intermediate banks is shown
action number is only a few percent of the theoretically in Fig. 19.
possible combinations. It is confirmed that there are many There are about 10 banks with large link numbers
pairs of banks which never transact directly. We can de- which deviate from the power law, also small link num-
fine active links between banks for pairs with the averaged ber banks with link number less than four are out of the
number of transaction larger than one per day. By this cri- power law. Such small banks are known to make a satellite
terion the number of links becomes about 2000, that is, structure that many banks linked to one large link num-
about 0.5 percent of all possible link numbers. Compared ber banks. It is yet to clarify why intermediate banks make
with the complete network, the actual network topologies fractal network, and also to clarify the role of large banks
are much more sparse. and small banks which are out of the fractal configuration.
In Fig. 18 the number distribution of active links per In relation with the banks, there are fractal properties
site are plotted in log–log plot [21]. As is known from other than cash flow and the transaction network. The dis-
this graph, there is an intermediate range in which the tribution of the whole amount of deposit of Japanese bank
452 Fractals and Economics

Fractals and Economics, Figure 19

Scale-free network of intermediate banks [20]

Fractals and Economics, Figure 21

Distribution of bank numbers historically behind a present
bank [57]

Fractals and Economics, Figure 20

Distribution of total deposit for Japanese banks [57] Power law
breaks down from 1999

is approximated by a power law as shown in Fig. 20 [57].

In recent years large banks merged making a few mega
banks and the distribution is a little deformed. Histori-
cally there were more than 6000 banks in Japan, however,
now we have about 600 as mentioned. It is very rare that
a bank disappears, instead banks are merged or absorbed. Fractals and Economics, Figure 22
The number distribution of banks which are historically Distribution of in-degrees and out-degrees in Japanese com-
behind a present bank is plotted in Fig. 21, again a power pany network [34]
law can be confirmed.
Other than the example of the bank network, network
structures are very important generally in economics. In Still more power laws in economics can be found in
production process from materials, through various parts sales data. A recent study on the distribution of expen-
to final products the network structure is recently studied diture at convenience stores in one shopping trip shows
in view of complex network analysis [18]. Trade networks a clear power law distribution with the exponent close to
among companies can also be described by network termi- two as shown in Fig. 23 [33]. Also, book sales, movie hits,
nology. Recently, network characterization quantities such news paper sales are known to be approximated by power
as link numbers (Fig. 22), degrees of authority, and Page- laws [39].
ranks are found to follow power laws from real trade data Viewing all these data in economics, we may say that
for nearly a million of companies in Japan [34]. fractals are everywhere in economics. In order to under-
 Fractals and Economics 453

Superposition of Basic Distributions

A power law distribution can also be easily produced by
superposition of basic distributions.
Let x be a Gaussian distribution with the probability
density given by
p
R R 2
p R (x) D p e 2 x ; (11)
2

and R be a 2 distribution with degrees of freedom ˛,

1 ˛/2
˛ R
w(R) D 2

˛
 R 2 1 e 2 : (12)
2

Then, the superposition of Gaussian distribution, Eq. (11),

Fractals and Economics, Figure 23 with the weight given by Eq. (12) becomes the T-distribu-
Distribution of expenditure in one shopping trip [33] tion having power law tails:

Z1
stand why fractals appear so frequently, we firstly need to p(x) D W(R)p R (x)dR
make simple toy models of fractals which can be analyzed 0
completely, and then, based on such basic models we can

˛C1 1
make more realistic models which can be directly compa- Dp 2

˛ / jxj˛1 ; (13)
rable with real data. At that level of study we will be able to
2 (1 C x 2 ) ˛C1
2

predict or control the complex real world economy.

which is P(> jxj) / jxj˛ in cumulative distribution. In
the special case that R, the inverse of variance of the nor-
Basic Models of Power Laws mal distribution, distributes exponentially, the value of ˛
In this chapter we introduce general mathematical and is 2. Similar super-position can be considered for any basic
physical models which produce power law distributions. distributions and power law distributions can be produced
By solving these simple and basic cases we can deepen our by such superposition.
understanding of the underlying mechanism of fractals or
power law distributions in economics. Stable Distributions
Assume that stochastic variables, x1 ; x2 ; : : : ; x n , are inde-
Transformation of Basic Distributions pendent and follow the same distribution, p(x), then con-
sider the following normalized summation;
A power law distribution can be easily produced by vari-
able transformation from basic distributions. x1 C x2 C    C x n  n
Xn  : (14)
n1/˛
1. Let x be a stochastic variable following a uniform dis-
tribution in the range (0, 1], then, y  x 1/˛ satisfies If there exists ˛ > 0 and n , such that the distribution
a power law, P(> y) D y ˛ for y  1. This is a useful of X n is identical to p(x), then, the distribution belongs to
transformation in case of numerical simulation using one of the Levy stable distributions [10]. The parameter ˛
random variable following power laws. is called the characteristic exponent which takes a value
2. Let x be a stochastic variable following an exponen- in the range (0, 2]. The stable distribution is character-
tial distribution, P(> x) D ex , for positive x, then, ized by four continuous parameters, the characteristic ex-
y  ex/˛ satisfies a power law, P(> y) / y ˛ . As ponent, an asymmetry parameter which takes a value in
exponential distributions occur frequently in random [1, 1], the scale factor which takes a positive value and
process such as the Poisson process, or energy distri- the location parameter which takes any real number. Here,
bution in thermal equilibrium, this simple exponential we introduce just a simple case of symmetric distribution
variable transformation can make it a power law. around the origin with the unit scale factor. The probabil-
454 Fractals and Economics

ity density is then given as Another entropy approach to the power laws is to gen-
Z1 eralize the entropy by the following form [56],
1 ˛
p(x; ˛) D eix ejj d : (15) R
1
2
1 p(x)q dx
1
x0
For large jxj the cumulative distribution follows the power Sq  ; (19)
q1
law, P(> x; ˛) / jxj˛ except the case of ˛ D 2. The
stable distribution with ˛ D 2 is the Gaussian distribution. where q is a real number. This function is called the q-en-
The most important property of the stable distribu- tropy and the ordinary entropy, Eq. (15), recovers in the
tion is the generalized central limit theorem: If the distri- limit of q ! 1. Maximizing the q-entropy keeping the
bution of sum of any independent identically distributed variance constant, so-called a q-Gaussian distribution is
random variables like X n in Eq. (14) converges in the limit obtained, which has the same functional form with the
of n ! 1 for some value of ˛, then the limit distribution T-distribution, Eq. (12), with the exponent ˛ given by
is a stable distribution with the characteristic exponent ˛.
q3
For any distribution with finite variance, the ordinary cen- ˛D : (20)
tral limit theory holds, that is, the special case of ˛ D 2. For 1q
any infinite variance distribution the limit distribution is
This generalized entropy formulation is often applied
˛ ¤ 2 with a power law tail. Namely, a power law realizes
to nonlinear systems having long correlations, in which
simply by summing up infinitely many stochastic variables
power law distributions play the central role.
with diverging variance.

Random Multiplicative Process

Entropy Approaches
Stochastic time evolution described by the following for-
Let x0 be a positive constant and consider a probability
mulation is called the multiplicative process,
density p(x) defined in the interval [x0 ; 1), the entropy
of this distribution is given by x(t C 1) D b(t)x(t) C f (t) ; (21)
Z1
S p(x) log p(x)dx : (16) where b(t) and f (t) are both independent random vari-
x0
ables [17]. In the case that b(t) is a constant, the distribu-
tion of x(t) depends on the distribution of f (t), for exam-
Here, we find a distribution that maximizes the entropy ple, if f (t) follows a Gaussian distribution, then the distri-
with a constraint such that the expectation of logarithm of bution of x(t) is also a Gaussian. However, in the case that
x is a constant, hlog xi D M. Then, applying the varia- b(t) fluctuates randomly, the resulting distribution of x(t)
tional principle to the following function, is known to follows a power law independent of f (t),
01 1
Z1 Z
P(> x) / jxj˛ ; (22)
L   p(x) log p(x)dx  1 @ p(x)dx  1A
x0 x0 where the exponent ˛ is determined by solving the follow-
0 1
Z1 ing equation [48],
C 2 @ p(x) log xdx  M A (17)
x0
hjb(t)j˛ i D 1 : (23)

the power law is obtained, This steady distribution exists when hlog jb(t)ji < 0
  Mlog 1 and f (t) is not identically 0. As a special case that b(t) D 0
x x0
P( x) D : (18) with a finite probability, then a steady state exists. It is
x0 proved rigorously that there exists only one steady state,
In other words, a power law distribution maximizes the and starting from any initial distribution the system con-
entropy in the situation where products are conserved. verges to the power law steady state.
To be more precise, consider two time dependent ran- In the case hlog jb(t)ji  0 there is no statistically
dom variables interacting each other satisfying the rela- steady state, intuitively the value of jb(t)j is so large that
tion, x1 (t)  x2 (t) D x1 (t 0 )  x2 (t 0 ), then the equilibrium x(t) is likely to diverge. Also in the case f (t) is identically
distribution follows a power law. 0 there is no steady state as known from Eq. (21) that
 Fractals and Economics 455

log jx(t)j follows a simple random walk with random noise where N is the total number of sites and f j (t) is the injected
term, log jb(t)j. mass to the site j.
The reason why this random multiplicative process The characteristic function, Z(; t)  hex j (t) i,
produces a power law can be understood easily by consid- which is the Laplace transform of the probability density,
ering a special case that b(t) D b > 1 with probability 0.5 satisfies the following equation by assuming uniformity,
and b(t) D 0 otherwise, with a constant value of f (t) D 1.
In such a situation the value of x(t) is 1 C b C b2 C   C b K Z(; t C 1)

with probability (0:5)K . From this we can directly evaluate N 2 1 1
the distribution of x(t), D Z(; t)2 C Z(; t) C he f j (t) i :
N N N
  (26)
b KC1  1
P  D 2KC1 i: e: The steady state solution in the vicinity of  D 0 is ob-
b1
(24) tained as
log 2
P( x) D 4(1 C (b  1)x)˛ ; ˛D : p
log b Z() D 1  h f i1/2 C    : (27)

As is known from this discussion, the mechanism of From this behavior the following power law steady distri-
this power law is deeply related to the above mentioned bution is obtained.
transformation of exponential distribution in Sect. “Trans- 1
formation of Basic Distributions”. P( x) / x ˛ ; ˛D : (28)
2
The power law distribution of a random multiplicative
By introducing a collision coefficient depending on the
process can also be confirmed experimentally by an elec-
size of particles power laws with various values of expo-
trical circuit in which resistivity fluctuates randomly [38].
nents realized in the steady state of such aggregation with
In an ordinary electrical circuit the voltage fluctuations
injection system [46].
in thermal equilibrium is nearly Gaussian, however, for
a circuit with random resistivity a power law distribution
holds. Critical Point of a Branching Process
Consider the situation that a branch grows and splits with
probability q or stops growing with probability 1  q as
Aggregation with Injection
shown in Fig. 24. What is the size distribution of the
Assume the situation that many particles are moving ran- branch? This problem can be solved in the following way.
domly and when two particles collide they coalesce mak- Let p(r) be the probability of finding a branch of size r,
ing a particle with mass conserved. Without any injection then the next relation holds.
of particles the system converges to the trivial state that
X
r1
only one particle remains. In the presence of continuous p(r C 1) D q p(s)p(r  s) : (29)
injection of small mass particles there exists a non-trivial sD1
statistically steady state in which mass distribution follows
a power law [41]. Actually, the mass distribution of aerosol
in the atmosphere is known to follow a power law in gen-
eral [11].
The above system of aggregation with injection can
be described by the following model. Let j be the discrete
space, and x j (t) be the mass on site j at time t, then choose
one site and let the particle move to another site and parti-
cles on the visited site merge, then add small mass particles
to all sites, this process can be mathematically given as,

8
<x j (t) C x k (t) C f j (t) ;
ˆ prob D 1/N
x j (tC1) D x j (t) C f j (t) ; prob D (N  2)/N
:̂
f j (t) ; prob D 1/N
Fractals and Economics, Figure 24
(25) Branching process (from left to right)
456 Fractals and Economics

Multiplying y rC1 and summing up by r from 0 to 1, It is known that for small positive  the statistically steady
a closed equation of the generating function, M(y), is ob- distribution x is well approximated by a Gaussian like the
tained, case of thermal fluctuations. For  close to 1 the fluctua-
tion of x is very large and its distribution is close to a power
1
X law. In the limit  goes to 1 and the distribution converges
M(y)1Cq D q y  M(y)2 ; M(y)  y r p(r) : (30)
rD0
to Eq. (28), the aggregation with injection case. For inter-
mediate values of  the distribution accompanies a fat tail
Solving this quadratic equation and expanding in terms of between Gaussian and a power law [49].
y, we have the probability density,

p(r) / r3/2 eQ(q)r ; Q(q)  log 4q(1  q) : (31) Fractal Tiling

A fractal tiling is introduced as the final basic model. Fig-
For q < 0:5 the probability decays exponentially for large ure 25 shows an example of fractal tiling of a plane by
r, in this case all branches has a finite size. At q D 0:5 the squares. Like this case Euclidean space is covered by var-
branch size follows the power law, P( r) / r1/2 , and the ious sizes of simple shapes like squares, triangles, circles
average size of branch becomes infinity. For q > 0:5 there etc. The area size distribution of squares in Fig. 25 follows
is a finite probability that a branch grows infinitely. The the power law,
probability of having an infinite branch, p(1) D 1M(1),
is given as, P( x) / x ˛ ; ˛D 1/2 : (34)
p
2q  1 C 1  4q(1  q)
p(1) D ; (32) Generalizing this model in d-dimensional space, the dis-
2q tribution of d-dimensional volume x is characterized by
which grows monotonically from zero to one in the range a power law distribution with an exponent, ˛ D (d  1)/d,
q D [0:5; 1]. It should be noted that the power law distri- therefore, the Zipf’s law which is the case of ˛ D 1 realizes
bution realizes at the critical point between the finite-size in the limit of d D 1. The fracture size distribution mea-
phase and the infinite-size phase [42]. sured in mass introduced in the beginning of this article
Compared with the preceding model of aggregation corresponds to the case of d D 3.
with injection, Eq. (28), the mass distribution is the same A classical example of fractal tiling is the Apollonian
as the branch size distribution at the critical point in gasket, that is, a plane is covered totally by infinite num-
Eq. (31). This coincidence is not an accident, but it ber of circles which are tangent each other. For a given
is known that aggregation with injection automatically river pattern like Fig. 26 the basin area distribution follows
chooses the critical point parameter. Aggregation and a power law with exponent about ˛ D 0:4 [45]. Although
branching are reversed process and the steady occur- these are very simple geometric models, simple models
rence of aggregation implies that branching numbers keep may sometimes help our intuitive understanding of frac-
a constant value on average and this requires the critical tal phenomena in economics.
point condition. This type of critical behaviors is called the
self-organized criticality and examples are found in vari-
ous fields [4].

Finite Portion Transport

Here, a kind of mixture of aggregation and branching
is considered. Assume that conserved quantities are dis-
tributed in N-sites. At each time step choose one site ran-
domly, and transport a finite portion,  x j (t), to another
randomly chosen site, where  is a parameter in the range
[0, 1].

x j (t C 1) D (1   )x j (t) ;
Fractals and Economics, Figure 25
x k (t C 1) D x k (t) C  x j (t) : (33) An example of fractal tiling
 Fractals and Economics 457

with adjustable positive parameters, fc0 ; c1 ; : : : ; c k g. By

the effect of this modulation on variance, the distribution
of price difference becomes superposition of Gaussian dis-
tribution with various values of variance, and the distribu-
tion becomes closer to a power law. Also, volatility cluster-
ing occurs automatically so that the volatility autocorrela-
tion becomes longer.
There are many variants of ARCH models, such as
GARCH and IGARCH, but all of them are based on purely
probabilistic modeling, and the probability of prices going
up and that of going down are identical.
Another type of market price model has been pro-
posed from physics view point [53]. The model is called
the PUCK model, an abbreviation of potentials of unbal-
anced complex kinetics, which assumes the existence of
market’s time-dependent potential force, UM (x; t), and the
Fractals and Economics, Figure 26 time evolution of market price is given by the following set
Fractal tiling by river patterns [45]
of equations;
ˇ
d ˇ
x(tC1) x(t) D  UM (x; t)ˇˇ C f (t) ; (37)
Market Models dx xDx(t)x M (t)
In this chapter market price models are reviewed in view b(t) x 2
UM (x; t)  ; (38)
of fractals. There are two approaches for construction of M1 2
market models. One is modeling the time sequences di-
where M is the number of moving average needed to de-
rectly by some stochastic model, and the other is model-
fine the center of potential force,
ing markets by agent models which are artificial markets
in computer consisted of programmed dealers. 1 X
M1
The first market price model was proposed by Bache- xM (t)  x(t  k) : (39)
M
lier in 1900 written as his Ph.D thesis [3], that is, five years kD0
before the model of Einstein’s random walk model of col- In this model f (t) is the external noise and b(t) is the curva-
loid particles. His idea was forgotten for nearly 50 years. ture of quadratic potential which changes with time. When
In 1950’s Markowitz developed the portfolio theory based b(t) D 0 the model is identical to the simple random walk
on a random walk model of market prices [28]. The the- model. When b(t) > 0 the market prices are attracted to
ory of option prices by Black and Scholes was introduced the moving averaged price, xM (t), the market is stable, and
in the 1970s, which is also based on random walk model of when b(t) < 0 prices are repelled from xM (t) so that the
market prices, or to be more precise a logarithm of market price fluctuation is large and the market is unstable. For
prices in continuum description [7]. b(t) < 2 the price motion becomes an exponential func-
In 1982 Engle introduced a modification of the sim- tion of time, which can describe singular behavior such as
ple random walk model, the ARCH model, which is the bubbles and crashes very nicely.
abbreviation of auto-regressive conditional heteroscedas- In the simplest case of M D 2 the time evolution equa-
ticity [9]. This model is formulated for market price differ- tion becomes,
ence as,
b(t)
x(t C 1) D  x(t) C f (t) : (40)
x(t) D (t) f (t) ; (35) 2
As is known from this functional form in the case b(t) fluc-
where f (t) is a random variable following a Gaussian dis- tuates randomly, the distribution of price difference fol-
tribution with 0 mean and variance unity, the local vari- lows a power law as mentioned in the previous Sect. “Ran-
ance (t) is given as dom Multiplicative Process”, Random multiplicative pro-
cess. Especially the PUCK model derives the ARCH
X
k
 (t)2 D c0 C c k (x(t  k))2 ; (36) model by introducing a random nonlinear potential func-
jD1 tion [54]. The value of b(t) can be estimated from the
458 Fractals and Economics

Fractals and Economics, Figure 27

Tick intervals of Poisson process (top) and the self-modulation
process (bottom) [52]

data and most of known empirical statistical laws includ-

ing fractal properties are fulfilled as a result [55].
The peculiar difference of this model compared with fi-
nancial technology models is that directional prediction is
possible in some sense. Actually, from the data it is known
that b(t) changes slowly in time, and for non-zero b(t) Fractals and Economics, Figure 28
the autocorrelation is not zero implying that the up-down The power spectrum of the self-modulation process [52]
statistics in the near future is not symmetric. Moreover in
the case of b(t) < 2 the price motion show an exponen-
tial dynamical growth hence predictable.
As introduced in Sect. “Examples in Economics” the
tick interval fluctuations can be characterized by the 1/f
power spectrum. This power law can be explained by
a model called the self-modulation model [52]. Let t j be
the jth tick interval, and we assume that the tick interval
can be approximated by the following random process,

1 X
K1
t jC1 D  j t jk C g j ; (41)
K
kD0

where j is a positive random number following an ex-

ponential distribution with the mean value 1, and K is
an integer which means the number of moving average, Fractals and Economics, Figure 29
The distribution of normalized time interval [50]
g j is a positive random variable. Due to the moving aver-
age term in Eq. (41) the tick interval automatically make
clusters as shown in Fig. 27, and the corresponding power
spectrum is proved to be proportional to 1/ f as typically minds move quicker if the market becomes busier. By this
represented in Fig. 28. self-modulation effect transactions in markets automati-
The market data of tick intervals are tested whether cally make a fractal configuration.
Eq. (41) really works or not. In Fig. 29 the cumulative Next, we introduce a dealer model approach to the
probability of estimated value of j from market data is market [47]. In any financial market dealers’ final goal is to
plotted where the moving average size is determined by the gain profit from the market. To this end dealers try to buy
physical time of 150 seconds and 400 seconds. As known at the lowest price and to sell at the highest price. Assume
from this figure, the distribution fits very nicely with the that there are N dealers at a market, and let the jth dealer’s
exponential distribution when the moving average size is buying and selling prices in their mind B j (t) and S j (t). For
150 seconds. This result implies that dealers in the mar- each dealer the inequality, B j (t) < S j (t), always holds. We
ket are mostly paying attention to the latest transaction for pay attention to the maximum price of fB j (t)g called the
about a few minutes only. And the dealers’ clocks in their best bid, and to the minimum price of fS j (t)g called the
 Fractals and Economics 459

best ask. Transactions occur in the market if there exists

a pair of dealers, j and k, who give the best bid and best ask
respectively, and they fulfill the following condition,

B j (t)  S k (t) : (42)

In the model the market price is given by the mean value

of these two prices.
As a simple situation we consider a deterministic time
evolution rule for these dealers. For all dealers the spread,
S j (t) B j (t), is set to be a constant L. Each dealer has a po-
sition, either a seller or a buyer. When the jth dealer’s po-
sition is a seller the selling price in mind, S j (t), decreases
every time step until he can actually sell. Similar dynamics
is applied to a buyer with the opposite direction of motion.
Fractals and Economics, Figure 30
In addition we assume that all dealers shift their prices in
Price evolution of a market with deterministic three dealers
mind proportional to a market price change. When this
proportional coefficient is positive, the dealer is catego-
rized as a trend-follower. If this coefficient is negative, the
dealer is called a contrarian. These rules are summarized
by the following time evolution equations.

B j (t C 1) D B j (t) C a j S j C b j x(t) ; (43)

where Sj takes either +1 or 1 meaning the buyer position

or seller position, respectively, x(t) gives the latest mar-
ket price change, fa j g are positive numbers given initially,
fb j g are also parameters given initially.
Figure 30 shows an example of market price evolu-
tion in the case of three dealers. It should be noted that
although the system is deterministic, namely, the future
price is determined uniquely by the set of initial values,
resulting market price fluctuates almost randomly even in
the minimum case of three dealers. The case of N D 2 Fractals and Economics, Figure 31
gives only periodic time evolution as expected, while for Cumulative distribution of a dealer model for different values
of b. For weaker trend-follow the slope is steeper [38]
N  3 the system can produce market price fluctuations
similar to the real market price fluctuations, for example,
the fractal properties of price chart and the power law dis-
on the dealer model. Experiments on the market is either
tribution of price difference are realized.
impossible or very difficult for a real market, however, in
In the case that the value of fb j g are identical for all
an artificial market we can repeat occurrence of bubbles
dealers, b, then the distribution of market price difference
and crashes any times, so that we might be able to find
follows a power law where the exponent is controllable by
a way to avoid catastrophic market behaviors by numeri-
this trend-follow parameter, b as shown in Fig. 31 [37].
cal simulation.
The volatility clustering is also observed automatically for
large dealer number case as shown in Fig. 32 (bottom)
which looks quite similar to a real price difference time Income Distribution Models
series Fig. 32 (top). Let us start with a famous historical problem, the St. Pe-
By adding a few features to this basic dealer model it is tersburg Paradox, as a model of income. This paradox was
now possible to reproduce almost all statistical characteris- named after Daniel Bernoulli’s paper written when he was
tics of market, such as tick-interval fluctuations, abnormal staying in the Russian city, Saint Petersburg, in 1738 [6].
diffusions etc. [58]. In this sense the study of market be- This paradox treats a simple lottery as described in the
haviors are now available by computer simulations based following, which is deeply related to the infinite expected
460 Fractals and Economics

Bernoulli’s answer to this paradox is to introduce the

human feeling of value, or utility, which is proportional to
the logarithm of price, for example. Based on this expected
utility hypothesis the fair value of X is given as follows,

X1 X1
U(2n ) log(2n )
XD nC1
D D1Clog 2  1:69 ; (45)
nD0
2 nD0
2nC1

where the utility function, U(x) D 1Clog x, is normalized

to satisfy U(1) D 1. This result implies that the appropri-
ate entry fee X should be about two dollars.
The idea of utility was highly developed in economics
for description of human behavior, in the way that human
preference is determined by maximal point of utility func-
tion, the physics concept of the variational principle ap-
plied to human action. Recently, in the field of behavioral
finance which emerged from psychology the actual obser-
vation of human behaviors about money is the main task
and the St. Petersburg paradox is attracting attention [36].
Although Bernoulli’s solution may explain the human
behavior, the fee X D 2 is obviously so small that the
bookmaker of this lottery will bankrupt immediately if the
entrance fee is actually fixed as two dollars and if a lot of
people actually buy it. The paradox is still a paradox.
To clarify what is the problem we calculate the distri-
bution of income of a gambler. As an income is 2n with
probability 2n1 , the cumulative distribution of income
is readily obtained as,
Fractals and Economics, Figure 32
Price difference time series for a real market (top) and a dealer
model (bottom) P( x) / 1/x : (46)

This is the power law which we observed for income dis-

value problem in probability theory and also it has been tribution of companies in Sect. “Examples in Economics”.
attracting a lot of economists’ interest in relation with the The key of this lottery is the mechanism that the prize
essential concept in economics, the utility [2]. money doubles at each time a head appears and the coin
Assume that you enjoy a game of chance, you pay toss stops when a tail appears. By denoting the number
a fixed fee, X dollars, to enter, and then you toss a fair coin of coin toss by t, we can introduce a stochastic process or
repeatedly until a tail firstly appears. You win 2n dollars a new lottery which is very much related to the St. Peters-
where n is the number of heads. What is the fair price of burg lottery.
the entrance fee, X?
Mathematically a fair price should be equal to the ex- x(t C 1) D b(t)x(t) C 1 ; (47)
pectation value, therefore, it should be given as,
where b(t) is 2 with probability 0.5 and is 0 otherwise. As
1
X 1 introduced in Sect. “Random Multiplicative Process”, this
XD 2n  D 1: (44)
2nC1 problem is solved easily and it is confirmed that the steady
nD0
state cumulative distribution of x(t) also follows Eq. (46).
This mathematical answer implies that even X is one mil- The difference between the St. Petersburg lottery and the
lion dollars this lottery is generous enough and you should new lottery Eq. (47) is the way of payment of entrance
buy because expectation is infinity. But, would you dare to fee. In the case of St. Petersburg lottery the entrance fee
buy this lottery, in which you will win only one dollar with X is paid in advance, while in the case of new lottery you
probability 0.5, and two dollars with probability 0.25, . . . ? have to add one dollar each time you toss a coin. This new
 Fractals and Economics 461

Fractals and Economics, Figure 33

Theoretical predicted exponent value vs. observed value [29]
Fractals and Economics, Figure 34
Numerical simulation of income distribution evolution of Japa-
nese companies [32]
lottery is fair from both the gambler side and the book-
maker side because the expectation of income is given by
hx(t)i D t and the amount of paid fee is also t.
This type of company’s income model can be gener-
Now we introduce a company’s income model by gen-
alized to take into account the effect of company’s size
eralizing this new fair lottery in the following way,
dependence on the distribution of growth rate. Also, the
magnitude of the random force term can be estimated
I(t C 1) D b(t)I(t) C f (t) ; (48)
from the probability of occurrence of negative income.
Then, assuming that the present growth rate distribution
where I(t) denotes the annual income of a company, b(t)
continues we can perform a numerical simulation of com-
represents the growth rate which is given randomly from
pany’s income distribution starting from a uniform distri-
a growth rate distribution g(b), and f (t) is a random noise.
bution as shown in Fig. 34 for Japan and in Fig. 35 for USA.
Readily from the results of Sect. “Random Multiplicative
It is shown that in the case of Japan, the company size dis-
Process”, we have a condition to satisfy the empirical rela-
tribution converges to the power law with exponent 1 in
tion, Eq. (10),
20 years, while in the case of USA the steady power law’s
Z slope is about 0.7 and it takes about 100 years to con-
hb(t)i D bg(b) D 1 : (49) verge [31]. According to this result extremely large com-
panies with size about 10 times bigger than the present
This relation is confirmed to hold approximately in actual biggest company will appear in USA in this century. Of
company data [32]. course the growth rate distribution will change faster than
In order to explain the job category dependence of the this prediction, however, this model can tell the qualita-
company’s income distribution already shown in Fig. 12, tive direction and the speed of change in very macroscopic
we plot the comparison of exponents in Fig. 33. Empiri- economical conditions.
cally estimated exponents are plotted in the ordinate and Other than this simple random multiplicative model
the solutions of the following equation calculated in each approach there are various approaches to explain empiri-
job category are plotted in the abscissa, cal facts about company’s statistics assuming a hierarchical
structure of organization, for example [23].
hb(t)ˇ i D 1 : (50)
Future Directions
The data points are roughly on a straight line demonstrat-
ing that the simple growth model of Eq. (48) seems to be Fractal properties generally appear in almost any huge
meaningful. data in economics. As for financial market models, em-
An implication of this result is that if a job category pirical fractal laws are reproduced and the frontier of
is expanding, namely, hb(t)i > 1, then the power law ex- study is now at the level of practical applications. How-
ponent determined by Eq. (50) is smaller than 1. On the ever, there are more than a million markets in the world
other hand if a job category is shrinking, we have an expo- and little is known about their interaction. More research
nent that is larger than 1. on market interaction will be promising. Company data
462 Fractals and Economics

10. Feller W (1971) An Introduction to Probability Theory and Its

Applications, 2nd edn, vol 2. Wiley, New York
11. Friedlander SK (1977) Smoke, dust and haze: Fundamentals of
aerosol behavior. Wiley-Interscience, New York
12. Fujiwara Y (2004) Zipf Law in Firms Bankruptcy. Phys A
337:219–230
13. Fujiwara Y, Guilmi CD, Aoyama H, Gallegati M, Souma W (2004)
Do Pareto-Zipf and Gibrat laws hold true? An analysis with Eu-
ropean firms. Phys A 335:197–216
14. Gilvarry JJ, Bergstrom BH (1961) Fracture of Brittle Solids. II Dis-
tribution Function for Fragment Size in Single Fracture (Exper-
imental). J Appl Phys 32:400–410
15. Gopikrishnan P, Meyer M, Amaral LAN, Stanley HE (1998) In-
verse Cubic Law for the Distribution of Stock Price Variations.
Fractals and Economics, Figure 35 Eur Phys J B 3:139–143
Numerical simulation of income distribution evolution of USA 16. Handel PH (1975) 1/f Noise-An “Infrared” Phenomenon. Phys
companies [32] Rev Lett 34:1492–1495
17. Havlin S, Selinger RB, Schwartz M, Stanley HE, Bunde A
(1988) Random Multiplicative Processes and Transport in
so far analyzed show various fractal properties as intro- Structures with Correlated Spatial Disorder. Phys Rev Lett 61:
1438–1441
duced in Sect. “Examples in Economics”, however, they 18. Hidalgo CA, Klinger RB, Barabasi AL, Hausmann R (2007) The
are just a few cross-sections of global economics. Espe- Product Space Conditions the Development of Nations. Sci-
cially, companies’ interaction data are inevitable to ana- ence 317:482–487
lyze the underlying network structures. Not only money 19. Inaoka H, Toyosawa E, Takayasu H (1997) Aspect Ratio Depen-
flow data it will be very important to observe material flow dence of Impact Fragmentation. Phys Rev Lett 78:3455–3458
20. Inaoka H, Ninomiya T, Taniguchi K, Shimizu T, Takayasu H
data in manufacturing and consumption processes. From (2004) Fractal Network derived from banking transaction –
the viewpoint of environmental study, such material flow An analysis of network structures formed by financial institu-
network will be of special importance in the near future. tions. Bank of Japan Working Paper. http://www.boj.or.jp/en/
Detail sales data analysis is a new topic and progress is ronbun/04/data/wp04e04.pdf
expected. 21. Inaoka H, Takayasu H, Shimizu T, Ninomiya T, Taniguchi K
(2004) Self-similarity of bank banking network. Phys A 339:
621–634
22. Klass OS, Biham O, Levy M, Malcai O, Solomon S (2006) The
Bibliography
Forbes 400 and the Pareto wealth distribution. Econ Lett
90:290–295
Primary Literature
23. Lee Y, Amaral LAN, Canning D, Meyer M, Stanley HE (1998) Uni-
1. Aoyama H, Nagahara Y, Okazaki MP, Souma W, Takayasu H, versal Features in the Growth Dynamics of Complex Organiza-
Takayasu M (2000) Pareto’s law for income of individuals and tions. Phys Rev Lett 81:3275–3278
debt of bankrupt companies. Fractals 8:293–300 24. Mandelbrot BB (1963) The variation of certain speculative
2. Aumann RJ (1977) The St. Petersburg paradox: A discussion of prices. J Bus 36:394–419
some recent comments. J Econ Theory 14:443–445 http://en. 25. Mandelbrot BB (1982) The Fractal Geometry of Nature. W.H.
wikipedia.org/wiki/Robert_Aumann Freeman, New York
3. Bachelier L (1900) Theory of Speculation. In: Cootner PH (ed) 26. Mandelbrot BB (2004) The (mis)behavior of markets. Basic
The Random Character of Stock Market Prices. MIT Press, Cam- Books, New York
bridge (translated in English) 27. Mantegna RN, Stanley HE (2000) An Introduction to Eco-
4. Bak P (1996) How Nature Works. In: The Science of Self-Orga- nomics: Correlations and Complexity in Finance. Cambridge
nized Criticality. Springer, New York Univ Press, Cambridge
5. Barabási AL, Réka A (1999) Emergence of scaling in ran- 28. Markowitz HM (1952) Portfolio Selection. J Finance 7:77–91
dom networks. Science 286:509–512; http://arxiv.org/abs/ 29. Mizuno T, Katori M, Takayasu H, Takayasu M (2001) Statisti-
cond-mat/9910332 cal laws in the income of Japanese companies. In: Empirical
6. Bernoulli D (1738) Exposition of a New Theory on the Measure- Science of Financial Fluctuations. Springer, Tokyo, pp 321–
ment of Risk; Translation in: (1954) Econometrica 22:22–36 330
7. Black F, Scholes M (1973) The Pricing of Options and Corporate 30. Mizuno T, Kurihara S, Takayasu M, Takayasu H (2003) Analy-
Liabilities. J Political Econ 81:637–654 sis of high-resolution foreign exchange data of USD-JPY for 13
8. Brenner YS, Kaelble H, Thomas M (1991) Income Distribution in years. Phys A 324:296–302
Historical Perspective. Cambridge University Press, Cambridge 31. Mizuno T, Kurihara S, Takayasu M, Takayasu H (2003) In-
9. Engle RF (1982) Autoregressive Conditional Heteroskedasticity vestment strategy based on a company growth model. In:
With Estimates of the Variance of UK Inflation. Econometrica Takayasu H (ed) Application of Econophysics. Springer, Tokyo,
50:987–1008 pp 256–261
 Fractals and Economics 463

32. Mizuno T, Takayasu M, Takayasu H (2004) The mean-field 47. Takayasu H, Miura H, Hirabayashi T, Hamada K (1992) Statisti-
approximation model of company’s income growth. Phys A cal properties of deterministic threshold elements–The case of
332:403–411 market price. Phys A 184:127–134
33. Mizuno T, Toriyama M, Terano T, Takayasu M (2008) Pareto law 48. Takayasu H, Sato AH, Takayasu M (1997) Stable infinite vari-
of the expenditure of a person in convenience stores. Phys A ance fluctuations in randomly amplified Langevin systems.
387:3931–3935 Phys Rev Lett 79:966–969
34. Ohnishi T, Takayasu H, Takayasu M (in preparation) 49. Takayasu M, Taguchi Y, Takayasu H (1994) Non-Gaussian dis-
35. Okuyama K, Takayasu M, Takayasu H (1999) Zipf’s law tribution in random transport dynamics. Inter J Mod Phys B
in income distribution of companies. Phys A 269:125– 8:3887–3961
131. http://www.ingentaconnect.com/content/els/03784371; 50. Takayasu M (2003) Self-modulation processes in financial mar-
jsessionid=5e5wq937wfsqu.victoria ket. In: Takayasu H (ed) Application of Econophysics. Springer,
36. Rieger MO, Wang M (2006) Cumulative prospect theory and Tokyo, pp 155–160
the St. Petersburg paradox. Econ Theory 28:665–679 51. Takayasu M (2005) Dynamics Complexity in Internet Traffic. In:
37. Sato AH, Takayasu H (1998) Dynamic numerical models of Kocarev K, Vatty G (eds) Complex Dynamics in Communication
stock market price: from microscopic determinism to macro- Networks. Springer, New York, pp 329–359
scopic randomness. Phys A 250:231–252 52. Takayasu M, Takayasu H (2003) Self-modulation processes and
38. Sato AH, Takayasu H, Sawada Y (2000) Power law fluctua- resulting generic 1/f fluctuations. Phys A 324:101–107
tion generator based on analog electrical circuit. Fractals 8: 53. Takayasu M, Mizuno T, Takayasu H (2006) Potentials force ob-
219–225 served in market dynamics. Phys A 370:91–97
39. Sinha S, Pan RK (2008) How a “Hit” is Born: The Emergence 54. Takayasu M, Mizuno T, Takayasu H (2007) Theoretical analysis
of Popularity from the Dynamics of Collective Choice. http:// of potential forces in markets. Phys A 383:115–119
arxiv.org/PS_cache/arxiv/pdf/0704/0704.2955v1.pdf 55. Takayasu M, Mizuno T, Watanabe K, Takayasu H (preprint)
40. Souma W (2001) Universal structure of the personal income 56. Tsallis C (1988) Possible generalization of Boltzmann–Gibbs
distribution. Fractals 9:463–470; http://www.nslij-genetics. statistics. J Stat Phys 52:479–487; http://en.wikipedia.org/wiki/
org/j/fractals.html Boltzmann_entropy
41. Takayasu H (1989) Steady-state distribution of generalized ag- 57. Ueno H, Mizuno T, Takayasu M (2007) Analysis of Japanese
gregation system with injection. Phys Rev Lett 63:2563–2566 bank’s historical network. Phys A 383:164–168
42. Takayasu H (1990) Fractals in the physical sciences. Manchester 58. Yamada K, Takayasu H, Takayasu M (2007) Characterization of
University Press, Manchester foreign exchange market using the threshold-dealer-model.
43. Takayasu H (ed) (2002) Empirical Science of Financial Fluctua- Phys A 382:340–346
tions–The Advent of Econophysics. Springer, Tokyo
44. Takayasu H (ed) (2003) Application of Econophysics. Springer,
Tokyo Books and Reviews
45. Takayasu H, Inaoka H (1992) New type of self-organized criti- Takayasu H (2006) Practical Fruits of Econophysics. Springer,
cality in a model of erosion. Phys Rev Lett 68:966–969 Tokyo
46. Takayasu H, Takayasu M, Provata A, Huber G (1991) Statisti- Chatterjee A, Chakrabarti BK (2007) Econophysics of Markets and
cal properties of aggregation with injection. J Stat Phys 65: Business Networks (New Economic Windows). Springer, New
725–745 York

View publication stats