The Pareto-Lévy Law and the Distribution of Income

Author(s): Benoit Mandelbrot

Source: International Economic Review , May, 1960, Vol. 1, No. 2 (May, 1960), pp. 79-106
Published by: Wiley for the Economics Department of the University of Pennsylvania
and Institute of Social and Economic Research, Osaka University

Stable URL: https://www.jstor.org/stable/2525289

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

and Wiley are collaborating with JSTOR to digitize, preserve and extend access to International
Economic Review

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 INTERNATIONAL
ECONOMIC
REVIEWV
REVIEWMay, 1
~~~~~~~~~VO
THE PARETO-LEVY LAW AND THE DISTRIBUTION OF INCOME*

BY BENOIT MANDELBROT'

SUMMARY

THE PURPOSE of this paper is twofold. On the one hand, we wish

to give an account of a set of new models for certain distribution
properties of an important class of economic quantities, which includes
"income" (see [15], [17]). On the other hand, we think that the tools
which we shall use are as important as the results which we hope to
achieve: that is, we intend to draw the economist's attention to the
great potentital importance of "stable non-Gaussian" probability dis-
tributions. To give a sharper focus to our paper, we shall develop
our main points within the frame of a theory of income distribution;
but our approach will be immediately translatable in terms of similar
quantities, and our theory may well be more reasonable, or the em-
pirical fit better, in the case of some other quantities. We may thus
paraphrase what a famous author said of Brownian motion: it is possible
that the properties which we shall study are identical to those of income;
however, the information available to us regarding incomes is so lacking
in precision that we cannot really form a judgment on the matter.
This paper will be almost exclusively theoretical. Our point of de-
parture will be an interesting empirical observation, namely, that over
a certain range of values of income U, its distribution is not markedly
influenced either by the socio-economic structure of the community
under study, or by the definition chosen for "income." That is, these
two elements may at most influence the values taken by certain param-
eters of an apparently universal distribution law. This law was dis-
covered by V. Pareto in 1897 [24]; actually, several different statements
have at times been called "Pareto's law." In the introduction we shall
very carefully distinguish between these statements, and we shall
comment upon some existing theories of income distribution.
* Manuscript received June 19, 1959.
' The work in this paper was started while the author was associated with the Uni-
versity of Geneva (under partial support of the Rockefeller Foundation), then with the
University of Lille and the Ecole Polytechnique in Paris.

79

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
80 BENOIT MANDELBROT

We shall then introduce a new version of Pareto's law which be-

longs to the class of "stable" distributions, and which we shall designate
as "the Pareto-Levy lawv" (P-L)c' This law also explains partially the
data relative to the middle-income range.
The initial motivation for this law will be essentially twofold. First,
among all the possible interpolations of the weak (asymptotic) Pareto
law, the P-L law is the only one which strictly satisfies one (strong)
form of the condition of invariance oi the distribution, relative to the
definition of income. Second, a P-L distribution is a possible linmit
distribution of sums of random variables. That is, as we shall demon-
strate, the Gaussian distribution is not the only possible such limit (as
is too generally assumed), and it is unnecessary (as well as insufficient)
to try to save the limit argument by applying it to log U, instead of
U, as is done in some theories leading to the "log-normal law" for U.
The P-L law turns out to have many other very desirable proper-
ties. it has also the drawback that it is limited to the interpolation
of Pareto's law, when its index a is such that 1 < a < 2, (see 1.1 for
a definition of a).
In a series of other articles, which has started with [18], [19], [20],
we also study the IP-L stochastic processes, a family of models of
income variation. These processes agairn preserve certain desirable
properties of the classical Gaussian processes without such drastic de-
fects as the fact that income variation cannot possibly be Gaussian.
It turns out that in the high income range, a P-L Markovian pro-
cess can be approximated by a randonm walk of log U: that is, we
are able to derive the fundarnental initial assumptions of a model
due to Champernowne [2]. Moreover, a P-L Markovian process be-
haves quite reasonably in the intermediate range of incomes.

1. INTRODUCTION

1.1. The strong Pareto law. Let us begin with a basic variant,
which will be referred to as the strong Pareto law. Let P(u) be
the percentage of individuals with an income U (over some fixed period
of reference) exceeding some number u (u is assumed to be a conr-
tinuous variable). The strong Pareto law asserts that:
((u/?u)-, when u > u0

1P when u < u.
n This associates Pareto's name with that of Paul Levy, wh
tributions intensively. (See in particular Levy [10], [11], [12].) To the best of our knowl-
edge the application of these laws to economics is entirely new as well as are the definition
and properties of the "Pareto-Levy" processes of [18), [19), [20).

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LEVY LAW 81

Then, the "density" p(u) = - dP(u)/du satisfies

I a(u 0)u-(a+l when u > u?

O(U) = 0 when u < ul.
This distribution is fully specified by two "state variables": u0, a scale
factor, and a, which will turn out to be some kind of index of ine-
quality of distribution. The strong law leaves the value of a undeter-
mined (although in some cases, one immediately states that a > 1).
We may also find the stronger statement that a -_ 3/2: the corresponding
variant of the strong law will be referred to as the strongest Pareto
law. Graphically, the strong law implies that the ("double logarithmic")
graph of y log P, as a function of v -log t, is a straight line.
The empirically observed P(u) are of course percentages relative to
finite populations. They will further be considered as frequencies rel-
ative to a random sample, drawn from an infinite population. [Em-
pirically, the law holds as well for samples of a few hundred (the
"burghers" of a city-state of the Renaissance) or of close to a hundred
million (the taxpayers in the USA).] That is, U will be treated as
a random variable with values u, and the curve describing the varia-
tion of U in timne will be treated as a random function.
1.2. The weak Pareto law. The strongest Pareto law is now
acknowledged to be empirically unjustified (and it should be noted that
many purported "disproofs" of "the" Pareto law apply to this variant
only). On the contrary, there is little question of the validity of the
Pareto law ifi "sufficiently" large values of u are concerned. The best
way of taking care of the limitations of the strong law is simply to
state that:

P(u) "behaves like" (u/u0)? as U -- oo

This is the weak Pareto law. Such a statement is useful only if

the exact definition of "behaves like" conforms to the empirical evi-
dence and (taking advantage of the indeterminacy of such evidence)
lends itself to easy mathematical manipulation. The following defini-
tion of "behaves like" is usually found to be adequate:

P(lu), (u/u? Mo
That is:

P() -1, as u oo
(uIuo)}0

or,

P(u) = {1 + e(u)}(u/u')0, where e(u) - 0, as u oo

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
82 BENOIT MANDELBROT

Graphically, this means that the curve (log P, log u) is asymptotic to

the straight line which represents the strong Pareto law.
1.3. Interpolation of the weak Pareto law. The range of values
of u which is accounted for by Pareto's analytic expression cove:s
only a part of the total population. Elsewhere the "density" -dP(u)/du
is represented by a quite irregular curve, the shape of which depends
in particular upon the breadth of coverage of the data considered.
This has been particularly clearly emphasized by H. P. Miller [22].
He has shown that the income distribution is skewed and presents
several maxima, if one includes all individuals, even those with no
income and part-time workers, and if one combines the incomes of
men and women. However, the separate income distributions oI most
of the different occupational categories, as distinguished by the census,
are both regular and fairly symmetric; the main source of skewness
in the overall distribution and hence in Pareto's law can be traced to
the inclusion of self-employed persons and managers together with all
wage earners. One may also note that the method of reporting in-
come differs by occupational categories, and that, as a result, the cor-
responding data are not equally reliable.
The above reasons make it unlikely that a single theory could ever
explain all the features of the income distribution or that a single
empirical formula could ever represent all the data. As a result, it
has been frequently suggested that several different models may be
required to explain the empirical P(u); unfortunately, we know of no
empirical check of this conjecture. In any case, the present paper
will be devoted almost exclusively to a theory of high income data
and of the weak Pareto law. It is unlikely that the interpolation of
the results of our model will be able to explain finally all the middle
income data. However, we shall not examine this point in great
detail.
1.4. Two laws of income distribution which are in contradiction
with the weak Pareto law.
A. Pareto himself has suggested the following possible improvement
in his law:

p(u) = - dP/du - ku-0'4+1exp(- bu) (where b > 0)e

However, the weak Pareto law must be at least approximately cor-

rect for large u. Hence, the parameter b must be very small, and
the evidence against the hypothesis that b 0 is not strong. The
choice, therefore, between the hypotheses b = 0 and b # 0 may be
legitimately influenced by ease of mathematical manipulation and ex-

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LtVY LAW 83

planation. We shall see that the weak Pareto law (b = 0) has some
crucial properties, which wsTill be used as a basis for a theory, and
which disappear if b / 0. We shall therefore disregard the case b t 0
(the more so since the formula 'above provides no improvement for
middle values of ut). (See 2.4 for a further discussion of b = 0.)
B. The best known statement about income distribution, apart from
the weak Pareto law, is the log-normal law [8], [1], which claims that
the variable log U [or perhaps better log (U-u'), where u'>0] is well
represented by the Gaussian distribution. The empirical evidence for
this is that the graph of (P, u) on log-normal paper seems to be straight.
Such a graph emphasizes a different range of values of u from that
of the Pareto graph, so that the graphical evidence for the two laws
is not contradictory. The motivation for the log-normal law is, how-
ever, largely theoretical (see 1.5). Roughly speaking, a log-normal U
can be explained by assuming that log U is the sum of many additive
components.
1.5. Some existing models of income distribution considered as
"thermodynamic" theories. There is a great temptation to consider
the exchanges of money which occur in economic interaction as analo-
gous to the exchanges of energy which occur in physical shocks between
gas molecules. In the loosest possible terms, both kinds of interactions
6"should" lead to "similar" states of equilibrium. That is, one "should"
be able to explain the law of income distribution by a model similar
to that used in statistical thermodynamics: many authors have done
so explicitly, and all the others of whom we know have done so im-
plicitly.
Unfortunately, the Pareto P(u) decreases much more slowly than
any of the usual laws of physics, so that if one wants to apply the
physical theory mechanically, one must somehow argud that U is a less
intrinsic variable than some slowly increasing function V(U). For
that, one must renounce the additive properties of U. The seemingly
universal choice for V is V = log U [or V' r log (U - u')]. This choice
is suggested by the fact that one plots log<u empirically. But it can
also be traced back to the "moral wealth" of Berpouilli, and it ap-
parently can be justified by some law of proportionate effect, or by
some kind of Weber-Fechner law. If this choice of V is granted, one
has to explain the normal law for the middle zone of v's, and the
exponential law P(v) = exp {- a(v - v0)} for the upper zone of v's.
Indeed, many existing models of the Pareto distribution are reduci-
ble to the observation that (for any a) exp (- av) is a possible bar-
ometric density distribution in the atmosphere. Alternatively, con-

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
84 BENOIT MANDELBROT

sider a set of Brownian particles, floating in a gas at a uniform

temperature and density, the whole being enclosed in a semi-infinite
tube with a closed bottom and an open top; assume further that the
field of gravity is uniform. Then the density distribution which the
Brownian particles attain as a state of final equilibrium is the ex-
ponential. This is the result of a compromise between two forces, both
uniform along the tube, i.e., the force of gravity, which alone would
tend to pull all particles to the bottom, and the influence of heat
motion, which alone would tend to diffuse them all to infinity. Clearly,
the models of income distribution we are now considering involve an
interpretation of the forces of diffusion and of gravity.
Unfortunately, such interpretations are never sufficiently intuitive
to exclude, for example, the removal of the counterpart of the bot-
tom of the tube. The trouble is that such a removal changes every-
thing. In fact, there is no steady limit state any longer, because all
the Brownian particles diffuse to ininus infinity. It is true that the
Gaussian distribution does appear as a conditional distribution for V
(this distribution is valid if one assumes that all the particles start
from the same point), but this cannot really be a basis for a model
of the log-normal law.
This difficulty is already very clear in the simplest model of a dif-
fusion, the so-called random walk, in which both time and V are
quantified. Time is integral and V is a multiple of a unit v". This
model (apparently introduced into economics by Champernowne [2]) fur-
ther assumes (1) that the variation of V is Markovian, that is, V(t+1)
is a function only of V(t) and of chance, and (2) that the probability
that V(t + 1) - v(t) = kv'" becomes independent of v(t) as v increases.
Additional assumptions are required to get either Pareto's law or the
log-normal law, and the difference between these additional assump-
tions has no intuitive meaning, which makes both conclusions uncon-
vincing. However, the models which lead to the exponential are still
slightly better: in fact, we can, perhaps, argue that the apparent
normality of the "density" p(v) in the central zone of v's simply
means that - logp(v) may be represented by a parabola in that zone,
whereas for large v's, it is represented by a straight line. Such a
parabolic interpolation needs no limit theorems of probability for its
justification; it applies to any regularly concave curve, at least in the
first approximation.
In the case of models of the weak Pareto law, further non-intuitive
assumptions are necessary to rationalize the empirical fact that a > 1
in all cases.

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LEVY LAW 86

We shall not analyze any of the other existing translations into

economic terms of various models of the normal or exponential distri-
butions, which occur in statistical thermodynamics, such as gas equilib-
rium or Brownian motion. Because of their common "thermodynamic"
character, these theories are essentially equivalent. In particular,
they do not assume any institutional income structure; all income re-
cipients are treated as entrepreneurs. A rewording of the classical
theories, however, can make them applicable to many possible institu-
tional structures of wage and salary earners; see Lydall [13]3.
In this article (see also [18]) we shall attempt to show that the
analogy of statistical physics need not be abandoned, if we want to
avoid the weaknesses which mar existing theories. We shall not re-
quire the transformnation V log U, and we shall not try to justify
the introduction of the economic counterparts of such conditions as
the presence or the absence of a bottom to the tube in which Brownian
motion is studied. That is, we shall not try to force income into the
structure of statistical thermodynamics (even implicitly); rather, we
shall attempt explicitly to generalize the statistical methods of thermo-
dynamics, in order to cover the economic concept of income.
We shall be successful only in the case where 1 < a < 2, a restriction
which we shall discuss repeatedly.

2. PARETO-LEVY RANDOM VARIABLES

2.1. An analysis of the definition of income. One of the principal

claims of this paper is that it is impossible to "explain" why Pareto's
law, and not some other law, is satisfied by income distributions, with-
out also studying the inseparable problem raised by the fact that
essentially the same law continues to be Jollowed by the distribution
Of "income," despite changes in the def,inition of this term. Such an
invariance is of course very important to census analyzers, because
of the small effect of large changes in methods of estimating income.
To study this problem one needs a practical way of expressing the
different definitions of U. We shall argue that there are several
ways of distinguishing different sources of U. Therefore, U may be
written in different ways as a sum of elements, such as: A, agricult-
3 We had indenpendently rediscovered this model and had discussed it in the original
text of this paper, as submitted on June 19, 1959. The coincidence of the predictions
of the P-L model and of the Lydall model for 1<a <2, shows that the distribution
which corresponds to the least amount of organization could be "frozen" without modi-
fying it, reinterpreted as a wage distribution and then let to evolve along conceptually
quite different lines. This fact has great relevance to the problem of the value of a
and to the fact that a has recently tended to increase beyond 2, in highly organized
industrial societies.

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
86 BENOIT MANDELBROT

ural, commercial, or industrial incomes; B, incomes in cash or in kind;

C, ordinary taxable income or capital gains; D, incomes of different
members of a single taxDaying unit, etc... Let the -income categories
in a certain decomposition of U be labeled as Us (1 ? i <1 N). We shall
assume that every method of reporting or estimating income corre-
sponds to the observation of the sum of the U, corresponding to so
subset of indices i. This quite reasonable consideration imposes the
restriction that the scale of incomes themselves is more intrinsic than
any transformed variables, such as log U.'
Let the incomes in the categories U, be considered as statistically
independent. This assumption idealizes the actual situation. A priori,
this abstraction may seem a bad first approximation, since when in-
come is divided into two categories U' and U", such as agricultural
and industrial incomes, the observed values u' and u" are usually very
different. We shall find, however, that in the weak Pareto case, when
the parts are independent, one would actually expect them to be very
unequal; bence, although such an inequality cannot be a confirmation
of independence, at least it does not contradict it in this case.
Under these assumptions, the only probability law for income which
could possibly be observed must be such that if U' and U" follow
this law (up to a scale transformation, and up to the choice of origin),
then U' E U" must also follow the same law (the sign 0D designates
the addition of random variables). That is, whatever a'>O, b', a">O, b",
there exist two constants a > 0 and b, such that

(a'U + b') 0D (a"U + b") = aU + b.

Such a probability law is said to be "stable" (under addition), and
its density as well as the variable U itself are "stable." The family
of all laws which satisfy this requirement has been constructed by
Paul Levy [10]: it includes, naturally, the Gaussian law, but it also
includes certain non-Gaussian laws, each of which turns out to satisfy
the weak Pareto law, with some 0 < a < 2. In other words, the addi-
tivity property of U, and the behavior of P(u), under the weak Pareto
law (both of which disappear if the scale of U is changed) turn out
to be precisely sufficient and necessary for the application of Le'vy's
theory of stable laws.
Further, the parameters of a non-Gaussian stable law can be ar-
ranged so that E(U) < co, which requires 1 < a < 2, and so that
p(- u) decreases very much faster than p(u), when u o oo: in the
I The same kind of division may also be performed in the direction of time. The prob-
lem arises that it is unlikely that the year be an intrinsic unit of time. If it is not, how
can a single law hold irrespective of the time unit chosen? See [181, [191, [291.

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LtVY LAW 87

latter case, the stable law may be called positive, an abbreviation for
"gmaximally skewed in the positive direction."
Since the usual terminology of probability (stable distributions) does
not serve our purpose, we propose to refer to "positive" stable dis-
tributions with 1 < a < 2, as "Pareto-Levy distributions."
It is not strictly true that the distribution of income is invariant
over the whole range of u, with respect to a change in definition of
income. We argue, however, that if the variable U is not Gaussian
and is very skewed, the most reasonable "first order" assumption
about income is that it is a P-L variable.
The density p(u) of the P-L law unfortunately cannot be expressed
in a closed analytic formi, but is determined by its bilateral Laplace
transform (valid for b > 0):

G(b) - exp bu) I dP(u) 5 exp ( bu)p(u)du

= exp {(bu*)&' + Mb},

which depends upon three parameters: a (1 < a < 2), u* (which is a

positive scale parameter), a-nd 111 (which is identical to E(U)).
G(b) yields the result:

P(u) - uO[u'F(1 - a)]O for u cX

where F(1 - a) is the Euler gamma-function.

For other values of u, the Pareto-Levy distribution had to be com-
puted. A few sample graphs of densities appear in Figure 1. More
detailed tables will be published by Mandelbrot and Zarufaller [21].
We see that as long as a is not close to 2, the P-L density curve
very rapidly becomes indistinguishable from a strong Pareto curve of
the same a. [For this, the origin of the strong Pareto curve must
be chosen properly and one must have u uX*F(1 - a).] The two
curves converge n-ear u = E(U), when a is in the neighborhood of 3/2,
and at even smaller values of u, when a is less than 3/2: that is, the
asymptotic behavior of P(u), derived from G(b), is very rapidly im-
plemented. As for large negative values of u, by the P-L law they
have a probability which is not zero, but decreases so rapidly as ui co,
that they may be safely disregarded; one finds that

log log p(u)]>- a log I uI e

a- 1

This is a faster decrease than in the case of the Gaussian density-

see Appendix I.

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
88 BENOIT MANDELBROT

02

I L (D~~~~~~~

O D 49 CV
0O00 0 0 0 00 0o Il
0 qD t N ? @ eP ? C
N CJ C4 C4 clJ
(n)d ?

N~~~~~

_+-
+

5 2 0 0 O 0

+
_ n

0Q (D C\ N N Q O, 0.O

+
(n) d

+0 4 z
NN

CD C D CNJ 0O 0 e C\J O CD eD v N O

(n) d

CQ NQ Q o

I I l I I I I I I I 0 0 0 ( I

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LEVY LAW 89

Finally, near E(U) and the most likely value of U, the P--L curve
has the kind of skewness which one finds in the empirical data and
which one hopes to derive in a theoretical curve.
Let a now approach 2. At the limit, the P-L density will tend
toward a Gaussian density. Close to the limit, the P-L density already
resembles a Gaussian one. Only for large values of u, (which have
a very small probability) is the Gaussian decrease of P(u) replaced by
a Paretian decrease.
In other terms, the P-L density is worth considering only if a is
not "too close" to 2. Even in this range, however, the prediction for
intermediate values of u is quite inadequate to cover the complete
curve of incomes, as described for example by Miller [22]. But as
part-time and unskilled workers are eliminated we more nearly approx-
imate such a curve. On the other hand, the income distribution of
unskilled workers is fairly symmetric and has a fairly small dispersion.'
Since the distribution of unskilled wages differs from that of high
incomes, the P-L law would a priori explain only the latter incomes.
If we examine any income category as defined by the Bureau of Census,
we can seldom tell in advance whether its income mechanism is closer
to that of the unskilled workers or to that of the largest income re-
cipients. Therefore, if we wish to determine how wide are the cate-
gories to which the P-L law applies, the only solution is to start from
the P-L interpolation of high incomes, and then to see how many
other incomes (and which) must be added to obtain a good fit. This
establishes a distinction between two kinds of income categories. The
reasonableness of this distinction should be checked on independent
grounds. For the time being we shall be content to observe that for
a far enough from 2 the prediction made by the P-L law is "reasona-
ble." (cf. 2.6.)
We shall conclude this section with a few statements concerning
non-P-L stable distributions. The Gaussian distribution is of course
stable, and it is the only stable distribution with a finite variance.
Further, all the stable variables with a finite mean are differences of
P-L variables scaled by arbitrary positive coefficients. The bilateral
generating function is no longer defined, and we must consider the
usual characteristic function (cof.). For a P-L variable, the c.f. is im-
mediately obtained as

9D(?)=3 _ e1"p(u)du = exp[-iM - _'I (ut)*{1 i' tg.53r}cos aj.

6 We may even think that their incomes would become non-random if their occupations
were entirely interchangeable and if their changes in type of employment were without
obstruction and costless.

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 90 BENOIT MANDELBROT

Hence, denoting by (1 + e)/(1 - )

components, the general stable variable, with E(U) < co, has a c.f. of
the form:

= exp[- iMC - Kl(u*)W{1 - tg t7r }cos t]

where u> ?0, 1,8I < I and 1 < a < 2.

(We may note that the same formula, with 0 < a < 1, gives another
family of stable variables, with E( U) = oo. The stable variables which
do not fall into either of the above families are those of Gauss and
of Cauchy and finally a small family of variables connected with those
of Cauchy.)
2.2. Another (equivalent) property of the stable laws. The stable
laws have another (equivalent) property: they are the only possible
limit laws of weighted sums of identical and independent random
variables.8 That is, if U is decomposed into a sum of a large number
of components Us, we need not resort to the above argument of ob-
servational invariance; we know that (1) if the sum is not Gaussian
(which is a conspicuous fact), (2) if the expected value of the sum is
finite (which is also a fact), and (3) if the probability of -u is much
less than that of u, for u > 0 and large, then no hypothesis about the
distribution of the parts is necessary to conclude that the sum can
only be a Pareto-Levy variable. We shall develop these points in
more detail, and in 2.6 we shall show that the above restrictions may
be reduced to broad "qualitative" properties, such as 'the likelihood
that two independent variables U' and U" contribute unequally to the
sum U' 6f U".

B Every stable law is such a limit. To prove this, assume that the variables Us them-
selves are stable; by induction of stability, the e-sum v 1 U8 will be a variable
of the form a(N)U + b(N) and hence {a(N)I-I{(ZfN1Ui) - b(N)} will have the same
distribution function as U. Conversely, assume that a certain normed sum of the variables
Ui has a limit. Write the N-th normed sum, taken in the sense of E, as:
N

A(N) Us - B(N)
S-1

;A() .A(n) U- B(n)- A( A(m) - +C(N(m) Q )

where N = n + m. By letting both n and m tend to infinity, we find that the definition
of a stable variable must be satisfied by the limit of A(N)Zt..lUL - B(N) + C(N).
We should also mention the following theorem of Gnedenko. For the convergence of
N' Y i -1Ui - B(N) to a stable limit, it is necessary and sufficient that 0 < a < 2,
and that for u > 0, F(u) - 1 - P(u) 1 - {1 T e'(u)j(u/ul)-, where e'(u) -O 0, as u e,
and for u < 0, F(u) {1 + eff(u)1(u/u2)w, where e"(u) -O 0, as u - oo.

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LE VY LAW 91

It should be noted that an explanation of the distribution of income

in terms of the addition of many contributions does not involve the
log U transformation. This avoids the usual limit argument, which
leads to the log-normal law and contradicts Pareto's law.
These results may be better than expected because the limit distribu-
tion might not hold for U, which is a sum of only a few random com-
ponents. That is, if one wants to increase the number of components
of U, one must abandon the hypothesis of independence at some point:
the more U is subdivided the less independent the components become.
This is a difficulty which is not proper to the problem of income, but
is acutely present in all social science applications of probability theory.7
Of course, the principle of the difficulty appears in physics for example
in explaining why some kind of noise is Gaussian. But in most physi-
cal problems there exists a sufficiently large zone between the systems
which are so small that they are impossible to subdivide and those which
are so large that they can no longer be considered as homogeneous.
No such zone exists in most problems of economics, so that a seemingly
successful application of a limit theorem may seem too good to be true.
To sum up, it cannot be strictly true that the additive components
are indepenident and hiave the same distribution (up to scale). However,
the "likelihood" of any distribution is greatly increased if this is the
only one reducible to limit arguments. Hence, if a sum of many
co,mponents is not Gaussian, is skewed, and such that E(U) < o,
then the most reasonable first assumiption concerning the sum is that
it follows the Pareto-Levy law.
More precisely, two extreme interpretations are possible. The "mini-
mal" interpretation regards the Pareto-Levy distribution as just an-
other law which is correct asymptotically, is sutfficiently easy to handle,
and is useful in the first approximation (after all, the Gaussian it-
self is frequently a good first approximation to distributions which,
actually, are certainly not Gaussian).
At the other extreme we may take the Pareto-Le6vy law entirely
seriously, and try to check its ability to predict some properties of
income distribution which otherwise would seem independent of the
weak Pareto law. We think that sucih predictions were in fact achieved.
This provides some claim for a "maximal" interpretation of the P-L
law, which regards its invariance and limit properties as being "ex-
plicative. "
2.3. Stable laws are very well known in abstract probability theory
In estimating a mean by an empirical average, often the sample is too small or the
many available data are strongly correlated.

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
92 BENOIT MANDELBROT

(cf. Le6vy [10], [111, [12], Gnedenko and Kolmogoroff [8]). We might,
therefore, have introduced the P-L law without special motivatior
other than the fact that the known asymptotic behavior of P(u) and
its recently computed behavior for intermediate u make it an attrac-
tive interpolation for the income data (Mandelbrot [18]). Unfortunately,
the behavior of tlle P-L law is in many other ways quite different
from that to which most statisticians are accustomed from their con-
stant handling of the Gaussian law. The only way of really showing
how far this law is suited to the present problem is to sketch its
theory, starting with a heuristic introduction to its properties (from
the viewpoint of addition and of the study of extremal values), and
ending with some rigorous results.
We see how the three main drawbacks of the classical theories
vanish. Many variants of the same argument ali lead to the same
result; no change of scale of U is necessary, the behavior of P(u)
being exactly what is needed, and a is "near" 3/2. In fact, a should
not be too close to 2, and cannot be greater than 2, a point which
requires some elaboration.
2.4. Concerning the sign of a -2 and the variance of U. A few
empirical data claim to lead to a > 2. To explain this by a limit argu-
ment requires most specific and unlikely assumptions. Similarly, the
observational invariance of the stable distribution cannot hold if a > 2,
unless it is weakened (2.5) to require that the sum of a few weak
Pareto variables with a > 2 is a weak Pareto variable with a > 2; as
the number of addends increase (2.8), the weak Pareto law holds for
a decreasing zone of values of u and finally the sum tends towards
a Gaussian variable.
We find that the sign of a - 2 distinguishes two sets of different
random variables. The occurrence of a > 2, or even a < 2 but close
to 2, may mean a predominance of salaries (Lydall's income model [13]
is valid for all values of a). (Another explanation may be found in
Mandelbrot [19].) On the other hand, several of the examples where
a > 2 are encountered occur in communities of Oceania which are
very much less self-contained than, for example, Great Britain or the
U.S.A. Their distributions of U may perhaps be truncated. In any
case, the sign of a - 2 raises an important empirical problem: Is the
fit of the weak law equally good regardless of this sign, and does this
law represent a higher percentage of the data where a < 2 ?
A second consequence of the limitation 1 < a < 2 is that the second
moment of U is then infinite, the first moment being finite. This is
easily seen, because by integration by parts,

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LEVY LAW 93

E(U) udF(u) - y tdPut) | P(u)du < X

_ 0 , 0 0 0o

and D(U) = E(U2) - [E(U)]2 where

E(U2) u2d.F(u) = u'dP(u) = 2 uP(u)du = co

(We assume that the behavior of P(u) for negative u does not lead
to a convergence problem for E(U).)
The last result holds for both strong and weak Pareto variables if
1 < a < 2, but it would, for example, cease to hold for the density
p(u) = ku-10`'1 exp(- bu) of 1.4. This provides a new and important
test of our conjecture that b =_ 0. The finite E(U) means that if u,
are samples from a Pareto distribution, the empirical mean E,
Et- Us/N tends to E( U) with probability 1 (Kolmogoroff's strong law
of large numbers); in addition, EN is a good estimate for E(U), if N
is large. Now consider DN = E'- (u, - EN)21N. If a> 2, DN tends to a
finite limit D( U), of which it is a good estimate, and E '-, (u, -E)
tends to a Gaussian variable. This result changes little if one adds an
exponential factor with small b. However, if a < 2, and b 0 the
lirnit of DN is infinity (one may show that D, grows without limit
like N2'-1) but if b becomes > 0, D(U) is finite. Therefore, the use-
fulness of the exponential factor exp(- bu) may be tested by check-
ing whether DN keeps increasing with N in the case of the largest
of the samples available. We could not make the direct test, but an
indirect test results from the following observation: the ordering of
the different populations by "increasing inequality" should presumably
be identical with their ordering- by decreasing a; on the other hand,
more usual measures of inequality are given by Dy, DN/EN or VDK14EN.
These two methods of ordering populations have been compared and
found entirely contradictory. This result ceases to be absurd if one
takes account of the fact that the values of N in the different sam-
ples which were compared range from 102 to 108; and it seems that
even if D(U) were finite, it would not be approached, even with the
largest samples. In that case, irrespective of any theory, it is pref-
erable to take b = 0, and D(U) = co. Further, Ino function of DN
is adequate to compare degrees of inequality, except perhaps between
samples of identical size.
Another test of the usefulness of the approximation D = co is pro-
vided by the relative contribution to DN of the largest of the u,: this
is very large (close to 1/2 for the Wisconsin incomes [23]), as might be
predicted from the theory of the Pareto-Levy law. The usual procedure

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
94 BENOIT MANDELBROT

of truncating U in order to avoid D = co distorts the whole problem.

2.5. Heuristic study of the addition of tvo independent random
variables, particularly in the Gaussian and weak Pareto cases. Let
U' and U" be two independent random variables, having the same
probability density p(u). We shall compare the behavior of p(u) for

large u, and that of p2(u) | p(x)p(u - x)dx, which is the density

of the sum U = U'QD U". It will be assumed that u may vary to
+ co.

One basic case is

p(u') = C exp(- bu') if u' > 0,

0 if u' <0.
Then,

p2(u) 0
= C2exp(-b
= C exp(- bu)dx = C2u exp(- bu).

Note that in this case log[p(u')] = logC-btb' is represented by a

straight line. The two next simplest cases are those where -log [p(u')]
has the same convexity, up or down, over the whole range of varia-
tion of u', so that the expression - log p(x) - log p(u - x) has an ex-
tremum for x = u/2. These two cases are illustrated in Figure 2.
One kind of convexity arises when d2 logp(u)/du2 < 0, for all u.
Then p(u) decreases rapidly as u - co, and the integrand p(x)p(u - x)
has a maximum for x = u/2. If that maximum is strong enough, the
integral } p(x)p(u - x)dx is likely to be made up to a great extent
of the contributions of the x's in a small interval near u/2. Hence,
a large value of u is likely to come from two contributions u' and
u" which are almost equal. (This is not obvious, and in particular it
is not true if the concavity of -log p is turned the other way.)
Consider for example the Gaussian density

P(X)= e exp 2x

It is known that

P2(U) =exp

so that the Gaussian character of density is preserved in addition,

except for the value of a. The same result can also be obtained heu-

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LVYY LAW 95

ristically, by arguing that p(x)p(u - x) stays near its maximum value

(u/2W}2 over some interval D/2 on each side of u/2, and is negli-
gible elsewhere. In that case, one gets the approximate estimate:

p2(A) - p(uf2)p(u/2)D - exp (-. 4)>

In other words, the approximate estimate is correct, if one takes D

a/i/ independent of u.
Suppose now that p(u) decreases slowly, so that d' log p(u)/du2 > 0,
for all u. Then p(x)p(u - x) has a minimum for x = u/2. It may

Ul~~~~~~~~~

U U

U U
'a '~~~~~~~2

0 0
_~~~~~~~4

p (U 1) p (U/

FIGURE 2: DISTRIBUTION OF U' WHEN U'G U" IS KNOWN

be noted that p(x) must - 0 as x -- co and must either = 0 or 0

as x -. - co. This is compatible with a - log p(x) concave downwards,
but only if the range of variation of x is truncated froin below. As-

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
96 BENOIT MANDELBROT

sume that u' can only be 2 u?, so that its most probable value is u0.
In that case, p(x)p(u - x) will have two maxima, for x = u? and for
x = u - u?. If they are strong enough, p2(zt) will be composed mostly
of the contributiorns of the neighborhoods of these maxirmla. Compare
then the following two expressions:

p2(u) = 2 p(x)p(u - x)dx, and 2 p(x)p(u)dxo

o o

These two expressions differ only in the small contributions of x's very
different from x - 0; that is,

p2(u) A 2p(u) p(x)dx

0

Finally, if u is large, p(x)dx p(x)dx, so that

o O

p,(u) 2 p(u)e

Hence,

p2(x)dx = P,(u) 2P(u).

Note that a large valze of u is now likely to be the sum of a relatively

very small value of either u' (or u") and of a value of u" (or u')
very close to u - u'. The two addends are likely to be of very un-
equal size; but the problem is entirely symmetric, so that the mean
value E(u'/u) is still u/2, by compensation. (See also Appendix II.)
We have in addition proved that the distribution of the largest of
two variables U' and U" of the slowly decreasing type has the same
asymptotic behavior as the distribution of their sum. Another deriva-
tion of this result starts from the derivation of the distribution of
max((U', U"). Clearly, Pm(u), the probability that u be larger than
max(U', U") is the probability that ut be larger than both u' and u".
Hence,

I - Pm(u) = [1 - P(u)]2.

For large u and small P(u), this becomes

Pm(u) ~ 2P(u) .

That is, for slowly decreasing densities, Pm(u) P,(u).

A prototype of the slowly decreasing variable is the strong Pareto
variable. In that case, we can write:

P2(u) ~ 2P(u) = P(2'0 u).

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LEVY LAW 97

That is, the sum of two independent and idential strong Pareto
variables is a weak Pareto variable, with the same a and a scale
factor u? multiplied by 21/.
Likewise, any weak Pareto distribution will be invariant in addition
up to the value of u?. The proof requires an easy refinement of the
previous argument, to cover the case where -logp(x) is not concave
or convex all through the range of x, but d2logp(x)/dx2 becomes and
stays negative for large values of x (which implies a quite regular
behavior for -log p(x) in that region). One can show in this way that
the weak Pareto law is preserved in the addition of two (or of a few)
independent random variables: there is no self-contradiction in the
observed fact that this law holds for parts of income as well as for
the whole. That is, the exact definition of the term "income" may
not be a matter of great concern. But, conversely, it is unlikely that
the observed data on P(u) for large u will be useful in discriminat-
ing among several different definitions of "income."
The weak Pareto and the Gaussian are the only laws strictly having
the above invariance ("stability") property. They will be distinguished
by a criterion of "equality" versus "inequality" between u' and in",
when u = u' + u" is known and large. (2.6.) In 2.8 we shall cite a
further known result concerning stable probability distributions.
We may also need to know the behavior of p,(u), when the density
p'(u') of U' decreases slowly and p"(u") decreases rapidly. In that case,
a large u is likely to be equal to u', plus some "small fluctuation."
In particular, a Gaussian error of observation concerning a weak Pareto
variable is quite negligible for large u.
2.6. The problem of addition and of division into two, in the case
of stable variables. We have shown that the behavior of the sum of
two variables is determined mainly by the convexity of -log p(u): we
shall later show that this criterion is in general insufficient to study
the addition of many variables. However, if we limit ourselves to
stable random variables, the convexity of -log p(u') is sufficient to
distinguish between the case of the Gaussian and that of all other stable
distributions. That is, these two cases may be distinguished by the
criterion of approximate equality of the parts of a Gaussian sum con-
trasted with the great inequality between these parts in the case of
all other distributions, in particular the Pareto-Levy one.
This distribution has been used so far only to derive the distribution
of U' @ U". Suppose now that the value u of U is given and that
we wish to study the distribution of u' or of ui" = u - u'. If the
a priori distribution of U' is Gaussian, with mean M and variance a',

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
98 BENOIT MANDELBROT

then the a priori distribution of U is Gaussian with mean 2M and

variance 2a2, and the conditional distribution of u', given u, is:

(2w)-"2r-1'a-1[ 2 exp-(u 2 j
p(uIu) = 2-'12(2,7)-112r-1 expL- (u -2M)2]

= 2'12(27')-11or1-1 exp[- (u'-u/2)2]

This is Gaussian, with mean value u/2 and variance a2/2. A striking
feature of the result is that the law of u' depends on u only through
the mean value of U' (see the right-hand side of Figure 2).
Let us now consider the non-Gaussian case. Insofar as our theory
is adequate in income studies, the problem of division will appear in
such questions as the following: if we know the sum of the agricul-
tural and industrial incomes of an individual and if the a priori dis-
tributions of both these quantities are Pareto-Levy with the same a,
then what is the distribution of the agricultural income? This case
is more involved than the Gaussian one, because we do not know any
explicit analytic form for the distribution of u' or u"; we can, how-
ever, do some numerical plotting (see the left-hand side of Figure 2).
If the sum u is very large, we find that the distribution of u' has
two very sharp maxima, near u'max and u - uma. As u decreases,
the shape of this distribution of u' will change, instead of being simply
translated, as in the Gaussian case. When u becomes small, more
maxima may appear. They will then all merge, and the distribution
of u' will differ little from that which is valid in the Gaussian case.
Finally, as u becomes negative and very large, the distribution of u'
will remain one with a single maximum.
Hence, bisection provides a very sharp distinction in this respect
between the Gaussian and all other stable laws.
Now, consider a fairly small number N; what is the distribution of
(1/N)-th of a stable variable? In the Gaussian case this (1/N)-th re-
mains Gaussian, whatever N may be; its mean value is u/N, and its
variance (N - 1)u2/N. In the non-Gaussian case, each part of a large
u may be small or it may be close to u; most of the N parts will be
small, but the largest of them will be likely to be close to the whole.
The situation is less intuitive when N becomes very large. How-
ever, Levy has proved that the necessary and sufficient condition for
the limit of the e-sum E U1 to be Gaussian is that the value us

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LEVY LAW 99

of the largest of the smmnands be negligible compared to the whole.

On the contrary, if the limit is stable non-Gaussian and such that
E( U) 0 O, both the sum and the largest of the summands will increase
roughly as N,11, and one can even show (Darling [4]) that their ratio
tends towards a limit, which is a random number having a distribu-
tion dependent upon a.
Let us now return to the discussion of 2.2. We argued there that
U is the sum of N components, without knowing N. A posteriori
this turns out to be quite acceptable, because the largest (o; the few
largest) of the components will cover a substantial part of the whole,
essentially independent of the number N of components. This emi-
nently desirable feature of the P-L theory is an important confirma-
tion of its usefulness.
If, however, a P-L income is small, its components are likely to be
of the same order of magnitude as in the Gaussian case. This has
an important bearing upon the problem touched near the end of 2.1.
Assume that one has a Census category in which income is usually
rather small and which is such that when U is decomposed into parts,
the sizes of the parts tend to be proportional to their a priori sizes.
One may assimilate this behavior to that of a Gaussian unskilled work-
er's income (the decomposition. may now refer to such things as the
lengths of time during which parts of income were earned). But such
behavior may also be that of a P-L variable, considered for small
values of u. As a result, the seemingly fundamental problem of split-
ting income into two parts so that only one follows the P-L law is
bound to have some solutions which are unassailable, but impossible
to justify positively. Hence, it is questionable whether this problem
is really fundamental.
2.7. Addition of many weak Pareto variables; possible asymptotic
invariance of the weak Pareto density if 0 < a < 2; non-invariance
in the case a>>2. Applying the reasoning of 2.5 to the eD-sum WN
Ejf1U, of N weak Pareto variables, we obtairn the following value
for the function PN of the variable Wy,

PN(u) ~ NP(u) - P(uN/110)e

This relationship may be expressed alternatively as follows: let a be
any small probability, let t(a) be the value of U such that P[u(a)] - a,
and let w,(a) be the value of WNV such that PN[wN(a)] = a. The above
approximation then becomes
WN(a) o NNlwu(a) w
Either way, the distribution of WN`N11X would be independent of

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
100 BENOIT MANDELBROT

N for large values of w; in other

Actually, there are two obvious limitations to the validity of the ap-
proximation PN . NP.
The first limitation is that when a > 1 and hence E(U) < oo this
approximation can, at best, be valid if E(U) = 0. This gives the only
possible choice of origin of Us such that when N -o co, the distribution
of WyN-10 can tend to a limit, over a range of wg such that PN(WN)
does not decrease to zero.8
The second limitation is that when a> 2, the relationship Pg NP
can at best hold in a zone of values wN such that the total probability
P-(WN) -0 as N-f oo. This is due to the fact that U now has a finite
variance D( U), and it is possible to apply the classical central limit
theorem to U. That is, we can now assert that, when N-. oo,

Pr { NE(U) < x tends to | exp(-y'/2)dy.

[ND(U)]12 V/2i'r

This holds over a range of values of x,

as N - Xo. That is, an increasingly larg
eventually enter into the Gaussian zone, in which WN - NE( U) diffuses
like N"'l (that is, more rapidly than Nl""). As a counterpart, the total
probability of the values of x, such that P--NP, must tend to zero as
N-00o.
Note that the N1'2 diffusion is not radically changed if U is tru
cated to less than some fixed bound: that is, the N"'l diffusion re
sents the bahavior of those values of WN which are sums of comp
small contributions, whereas the N""o diffusion represents the be
of sums of a single large term and of many very small ones.
We can now draw some conclusions concerning the relationship be-
tween the behavior of the largest of N terms U, and the bel a 7ior of
their sum. In the weak Pareto case, the two problems are identical
if N 2, whatever a; but as N increases they become distinct prob-
lems. The problem of the maximum is the only one which remains
simple and continues to lead to a weak Pareto variable. On the con-
trary, the concavity of -log p(u') is not a sufficiently stringLent criterion
8 To show this, write V= U + c; then VY = U- + Nc and VNN-'la = UNN-li0 +
cNI-1"x. When N-- co, the last term will increase without limit, so that VNN-11/ can-
not have a non-trivial limit distribution, if UvN-1l has one. Further, if UNN-NI0 has
a limit distribution. UY!N will have the degenerate limit 0, so that we must assume
that E(U) =0.

If, on the contrary, a < 1, the above argument fails, because Nl-l ten
therefore WfN-1/0 could have a limit distribution on a non-decreasing r
of VU. whatever the origin of U (anyway E(U) -oo).

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LEVY LAW 101

to discriminate between those cases where P- NP does or does not

apply, over a range of values of u having a fixed probability.
2.8. Infinite divisibility of the stable distributions. If U is stable,
we can write, designating by E the sums in the sense of D3,

U = A(N) U, - B(N) = 1 {A(N)Ut - B(N)IN} = V

where the Ug are independent values of U and where VN, = A(N) U, -
B(N)/N. Hence, for every N, U can be considered as a sum of N
independent and identically distributed variables, V, This is the
definition of infinite divisibility for a random variable. Suppose that
U is a P-L variable with E(U) 0. Then the infinite divisibility of
U is made obvious by writing the c.f. g)(-) in the following fashion:

log (?) = C (ecx -- ix) ! d(x-0) I

To divide a P-L variable by N we need only replace C by C/N. This
preserves the form of the function log q(s), as it should, becaulse
(1/N)-th of a stable variable is itself stable. If s is chosen small
enough, we may approximate logcp(C) by the integral log9($, e), re-
stricted to the range (a < x < co). Further, the term itC x d(x-3) I
of log p(?, c) amounts to a non-random translation of U. The essential
part of log 9(?), is, therefore, the part C (e0x - 1) Id(x-0) , which is
a limit of approximations of the form (eiX - 1) I A(x-) 1.
We may thus represent U as a limit of sums of Poisson variables.
To each increment dx of the variable x there is a corresponding con-
tribution to U equal to x multiplied by a Poisson variable of expected
value C d(x-c) . This means that a Pareto-Levy variable may be
considered as a sum of variables, each of wvhich is closely relatQd
to the strong Pareto law. (The strong Pareto law may be truncated
at any e > 0, because the term i<r of log p(O) takes care of the
divergence near 0 of the integral 5(et;2 - 1) Id(x-y) .)
This relationship between the strong Pareto and the P-L laws and
the role of the convergence factor iXx I d(x-of) I are made quite ciear
in a very intuitive physical problem of Holtsmark [9].
The problem is that of the attraction exerted on a star by an in-
finite uniform cloud of identical stars. Let us postpone convergence
problems and consider first a very large sphere of radius R, with:n
which N stars of unit mass are distributed at random, uniformly
and independently. A final star is located at the center 0 of the sphere,
and we wish to compute the resultant of thle Newtonian attractiots

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
102 BENOIT MANDELBROT

exerted on 0 by the N other stars. Units will be chosen in such a way

that two stars of unit mras6 attract each other with the force r-2 = u.
Let D = N(4R32r/3)-' be the average density of stars, and let u* = R-2.
Consider first the attraction of the stars located within a thin pencil
(or infinitesimal cone) covering dS spherical radians, having its apex
at 0, and extending in one direction only from 0. This pencil is a sum
of cells, each of which is contained between some radius r and some
radius r+dr. The volume of such a cell is d V = dSd(r3) = dS Jd(u-812)J.
Therefore, if there were a single star in the pencil dS, the probability
of its being in the volume d V would be given by the strong Pareto
distribution:

| d Sd(u8) = d(u/u*) |
dS.RB

The characteristic function of this distribution is a fairly involved func-

tion p(.u*). If there were N stars in the pencil, the probability that
the attraction on 0 is u would have the c.f. qN(Cu*), which becomes
increasingly more involved as N - co.
The problem is simplified if (D remaining constant) R - co, and
N -> co. It then becomes possible to assume that the number of stars
in the cell of volume d V is not fixed but is given by a Poisson ran-
dom variable, with the same D as the expected density. It is clear
that one can easily go from, one case to the other by slightly chang-
ing the distribution of stars for large R, an area of small values of u.
In this Poisson approximation, the stars located in the volume d V
will exert a total force which is a multiple of u = r-2, the multiplier
being a Poisson variable with expected value D-dS. I d(u-312) 1. That
is, the total force exerted on 0 will be the sum of a number of in-
dependent discrete jumps. (The mean relative number of jumps, with a
value between some a and some b, will be D.dS.(a-312 -b- '2)/D.dS.u*-'2,
i.e., it will follow the strong Pareto law.) The c.f. of the total con-
tribution of the pencil dS will then be approximated by the integral

log p() = (D * dS) 2 (e';" - 1) I d(u-812)

The fact that the integral is extended to u = o raises no convergence

difficulty, but the careless extension of the integration down to u = 0
(R = o) would lead to a divergent sum. The physical reason for
this is that, although each of the distant stars contributes little at-
traction, their number is such that their total mean attraction is in-
finite. However, one may disregard this infinite mean value, because
the difference between the attraction and its mean is finite; indeed,

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LEVY LAW 103

the fluctuation of the contribution of far away stars has the charac-
teristic function

R2
cp(I) = exp (D * dS) (eu" 1 -iu)} I d(zu3/2)

which converges and tends to zero as R - co. For the sake of con-
venience, the same correction itu may also be used for all other values
of u, since its effect for large u is only to add a finite translation to
U. Hence, the difference between the attraction of the stars in the
pencil dS, and the mean value of this attraction, is a "positive"
stable variable, or "Pareto-Le'vy" variable, with a 3/2.
The meaning of the stability of this attraction is that if two clouds
of stars are distinguished by their colors (say red and blue) but have
the same density and fill the samrre pencil dS, then the forces exerted
on 0 by red or blue stars alone, or bv both together, differ only by
a scale factor and not in the analytic form of their d stributions.
It is also evident why large negative values of u are unlikely com-
pared to large positive values. A large negative u can occur only if
there is an abnormally small number of stars. Moreover, the absence
of any stars nea-r 0 is a quite likely event, but alone it can at best
give a bounded negative u. Therefore, a large negative u must also
contain the negative contributions of stars missing far away from 0;
each of these stars contributes little to U, so that the nuinber of miss-
ing stars must be very large, and this is very unlikely.
On the contrary, an unboundedly large positive u may be obtained
from the presence oI a single star near 0, irrespective of the density
of stars elsewhere; such an event is far more likely than the combi-
nation of events required for a negative u. It is easy to check the
fact that the distribution oi U has the same asymptotic behavior for
large u as the distribution of the attraction of the nearest star.

3. CONCLUSION

It is to be hoped that this first application of the stable laws will

stimulate interest in a more detailed stuidy of their properties. Most
of the usual procedures of statistics must be revised because of the
infinite variance of U; such new problenms arise, as that of the best
choice of the origin u", in such fashion that the P-L density p(u) and
the strong Pareto density [(u - ut)/u']-t` coincide over as large a rang
of values of u as possible.
Another series of problems concerns the comparison of the P-L curve
with the empirical data in the region of intermediate values of u.

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
104 BENOIT MANDELBROT

A final problem concerns the sign of a - 2. We have referred to

it several times, but a further discussion can be pursued only within
the framework of a theory of P-L processes. (One indication may
already be found in Section 2 of [19J.) To settle this problem, it will
probably be necessary to introduce some dependence between the addi-
tive components of U; this must be done carefully, however, if one
wants to avoid obtaining a wholly indeterminate answer. See also
footnote 3.

International Business Machines Corporation, U. S. A.

APPENDIX I

The behavior of the P-L density for large negative arguments is not classical. To derive
it, we note first that the form of the bilateral generating function G(b) (which is not com-
monly used) may be deduced from the commonly used characteristic function, with the
help of some standard theorems on Fourier transforms in the complex plane. From the
existence of G(b) it follows that p(u) must decrease faster than any form exp(-lbul),
when u - .
To show exactly how fast p(u) decreases, write for convenience v -u and fv)
- log p(v); Alv) will increase faster than any linear form of v. Write further:

G(b) J exp(bv)p(v)dv = exp [by - fAv)]dv c exp [h(v)]dv.

If b is large, the integrand exp(h) must have a maximum for the value of v such that
b f'(w). Near that point, we can write:

h(v) [bw - f '(w)] - (1/2)(v - w)9f"(w) + (1/6Xv - w)88f()(w) + ' - '

If we could limit oneself to terms of order 0 and 2, we would find for G(b) the approxi-
mation exp [bw - f'(w)][2it/f"(w)]-112. Let us investigate the possibility of equating this
approximation to exp(b'), with a flv) of the form Kvy; if we take account of the term
exp [bw - f'(w)] alone, the approximation may indeed be made equal to exp (bx) by tak-
ing c a(a - 1)-1. The term [21/f "(w)]-1"2 weakens the result somewhat, since, instead
of having log [-log p(v)] a log K + c log v, we can only assert that log ( - log p)/log v -c
as v - oo. Finally, consider the terms other than those of orders 0 and 2: we note that
the term in (v - w)2 gives a non-negligible contribution to G(b), only as long as (v - w)
is of the order of magnitude of [ff"(W)]-1/2 ~- W-1[2(*-1)]-1. Then the term in (v - W)3 is
of the order of magnitude of W-([2(W.1)-l and is negligible; similarly the terms of
higher order do not modify the behavior of p(v).

APPENDIX II

We can exhibit the invariance of the weak Pareto variables in a specific example. Let
O < a < 1, and consider the discrete variable having the following Z(b) as discrete (one-
sided) generating function:

Z(b) = - exp( -bn)p(n) = 1 - C(1 - e-b)0 (where 0 < C < 1) .

n-0

All p(n) are positive and less than 1, and n I p(n) = 1. For large tn,

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 THE PARETO-LEVY LAW 105

p()-Cn-(a,+')
PC) r(- a)

The sum of two variables of this type has the generating function:

ZA(b) Z2(b) - 1 - 2C(1 - eb)* + C2(1 - e-1)2j

so that

P2(n) - 2Cn(+) - C2n-(2%+l) = 2p(n) + correction.

[r(-a) r(- 2a)

For large n, the second part of the right-hand term becomes negligible compared to the
first. If a = 1/2, it is zero, so that the range of values of n in which it may be neglected
increases as a tends to 1/2.
If 1 < a < 2, we must add a factor in (1 - e-b) to obtain an acceptable generating func-
tion; consider for example:

Z(b) = 1 - C(1 - e-b) + C'(1 - eb)

where 0 < aC' ? C ? C' + 1, so that C' ? (a - 1)-1.

Now

2Cn-(04+1) 2CC'(a + 1)n-(,x1) C'2n-(2x+l)

r(- a) r(- a) '( - 2a)

For large n, the second and third terms become negligible for all a. The ratio of the
coefficients of the first and of the second term depends little upon a; but the ratio of
the coefficients of the third and first terms is ruled by r(- - a)Ir(- -2 a), which is zero for
a 3/2, but may become large elsewhere. As a result, the range of values of n over
which the third term is important may be large.
Each time a increases past an integral value, the sign of C(1 - e-b) must be changed,
and another polynomial term must be added to Z(b), to keep it a generating function.
The number of corrective terms of p2(n) - 2p(n) increases, as well as the range of values
of n in which the corrective terms are appreciable.
Similarly, as more than two terms are added, pN(n) - Np(n) is vitiated over an in-
creasing range of values of n. Let N -e oo, and observe the weighted sums of the varia-
bles Us, whose values are the integers n. (Z remains a summation in the sense of (B.)
For 0 < a < 1, it is sufficient to consider the expression WN =N-" , since its
g.f. is ZN(N-1/ab), which tends to exp(-Cbm) when N --* co, as it should.
If 1 < a < 2, one must consider the expression WN = N-Il z ( Ut - M), where M, the
mean value of Ut, is easily found to be C. The g.f. of WN is clearly exp(NCb)ZN(N-l Ib)
and when N -*0o, it tends to exp (Cbm), as it should. It is easily seen that if we choose
for M a value different from C, the g.f. of WN will not tend to a non-degenerate ex-
pression.
If the value of a is higher, no linear renorming of Us can eliminate from log[Z(b)]
the square term Kb2, and hence, the best normed sum of the Us is the classically normed
N-112Z(Ui - M), which tends to a Gaussian, whatever the value of a.

APPENDIX III

This appendix concerns the transformation V=log U, which is used in most classical
theories of income distribution.
The strong objections against this transformation do not apply at all in the case of a law
due to Estoup and Zipf, which is formally similar to that of Pareto, but relates to word

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms
 106 BENOIT MANDELBROT

frequencies, and which we have studied since 1951 (cf. [14], [16]). In the formal expression
of that law U is replaced by the- "rank" of a word, where words are crdered by decreasing
frequencies. Hence log U has an intrinsic meaning as "cost of coding"; in particular,
the addition of "costs" is quite meaningful.
In the case of income, the transformation V=log U is justified in our [18] and [191,
and in a more detailed article which ought to appear soon.

REFERENCES

[ 1 ] AITCHISON, J. AND J. A. C. BROWN, The Lognormal Distribution, Cambridge: Cam-

bridge University Press, 1957.
[ 2 1 CHAMPERNOWNE, G., Economic Journal, Vol. -63 (1953), pp 318 ff.
[ 3 ] CHANDRASEKHAR, S., Reviews of Modern Physics, Vol. 15 (1943), pp 1 ff, (reprinted
in N. Wax, Noise and Stochastic Processes, New York: Dover, 1954).
[4] DARLING, D., Transactions of the American MUathematical Society, Vol. 73 (1952),
pP 95 ff.
15 1 EINSTEIN, A., The Theory of Brownian Motion, New York: Dover, 1956 (reprint).
[6] FRICHET, M., Comptes Rendus de l'Acadenie des Sciences, Paris, Vol. 249 (1959),
p. 1837.
[7 1 GIBRAT, R., Les inigalitUs economiques, Paris: Sirey, 1932.
[ 81 GNEDENKO, B. V. AND A. N. KOLMOGOROFF, Limit Distributions for Sums of In-
dependent Random Variables (English translation by K. L. Chung), Reading (Mass.):
Addison Wesley Press, 1954.

[91 HOLTSMARK, J., Annalen der Physik, Vol. 58 (1919), pp 577 ff.
[10] IwVY, PAUL, Calcul des ProbabilitUs, Paris: Gauthier-Villars, 1925.
[11] ----, Theorie de laddition des variables aUatoires, Paris: Gauthier-Villars, 1937
(second edition: 1954).
[12] , Processus stochastiques et mouvement Brownien, Paris: Gauthier-Villars,
1948.
[13] LYDALL, H. F., Econometrica, Vol. 27 (1959), pp 110 if.
[141 MANDELBROT, B., in Communication Theory, ed. Willis Jackson, London: Butter.
worth, 1953.
[151 , Memoranduim of the University of Geneva Mathematical Institute, January
16, 1956.
[16] , in Logique, Langage et Thiorie de l'Information, Apostel, L., B. Mandel-
brot and R. Morf, Paris: Presses Universitaires de France, 1957.
[17] , Memorandum of the University of Lille Mathematical Institute, Novem-
ber, 1957.
[181 - , Comptes Ren-dus de 1'Acad(mie des Sciences, Paris, Vol. 249 (1959), pp.
613-615.
[19] -, Comptes Rendus de l'Acadimie des Sciences, Paris, Vol. 249 (1959), pp.
2153-2155.
[20] , Comptes Rendus de l'Acadgmie des Sciences, Paris, Vol. 250 (1960), pp.
451-453.
[211 MANDELBROT, B. AND F. ZARNFALLER, to be published.
[22] MILLER, H. P., The Income of the American People, New York: J. Wiley, 1955.
[23] HANNA, F. A., J. A. PECKHAM, AND S. M. LERNER, Analysis of Wisconsin Income,
(Vol. 9, Studies in Income and Wealth) New York: National Bureau of Economic
Research, 1948.
[241 PARETO, V., Cours d'Economie Politique, Lausanne and Paris: 1897.
[25] WOLD, H. A. 0. and P. WHITTLE, Econometrica, Vol. 25 (1957), pp. 591 ff.

This content downloaded from

146.199.60.115 on Mon, 19 Dec 2022 20:16:43 UTC
All use subject to https://about.jstor.org/terms