You are on page 1of 31

Physica A 391 (2012) 1323–1353

Contents lists available at SciVerse ScienceDirect

Physica A
journal homepage: www.elsevier.com/locate/physa

Minireview

Measuring statistical evenness: A panoramic overview


Iddo I. Eliazar a,∗ , Igor M. Sokolov b
a
Department of Technology Management, Holon Institute of Technology, P.O. Box 305, Holon 58102, Israel
b
Institut für Physik, Humboldt-Universität zu Berlin, Newtonstr. 15, D-12489 Berlin, Germany

article info abstract


Article history: Motivated by the question ‘‘how equal is the distribution of wealth within a given human
Received 7 July 2011 population?’’ economics devised an impressive toolbox of quantitative measures of societal
Received in revised form 5 September 2011 egalitarianism including the Lorenz curve and the following indices: Gini, Pietra, Hoover,
Available online 18 September 2011
Amato, Hirschman, Theil and Atkinson. These quantitative measures – considered in the
broader context of general data-sets with positive values – are, in effect, general gauges
Keywords:
of statistical evenness. While the application of Gini’s index grew beyond economics
Statistical heterogeneity
Evenness gauges
and reached diverse fields of science, the aforementioned ‘‘evenness toolbox’’ has largely
Lorenz curve remained within the confines of the social sciences. The aim of this Paper is to expose this
Gini’s index ‘‘evenness toolbox’’ to the physics community by presenting a comprehensive evenness-
Pietra’s index based approach to a fundamental problem in science—the measurement of statistical
Hoover’s ‘‘Robin Hood’’ index heterogeneity.
Amato’s index © 2011 Elsevier B.V. All rights reserved.
The curvature index
Hirschman’s index
Theil’s index
Atkinson’s indices
Rényi’s entropies
Rényi’s indices
Pareto’s 20–80 rule
Pareto’s probability law
Lorenzian fractality
Power-laws
Rank distributions

Contents

1. Introduction.......................................................................................................................................................................................... 1324
2. The Lorenz curve .................................................................................................................................................................................. 1327
2.1. The Lorenz curve: a census approach..................................................................................................................................... 1327
2.2. The Lorenz curve: a probabilistic approach ........................................................................................................................... 1328
2.3. Discussion................................................................................................................................................................................. 1329
2.3.1. The census approach vs. the probabilistic approach .............................................................................................. 1329
2.3.2. Normalization ........................................................................................................................................................... 1330
2.4. Two oligarchy models ............................................................................................................................................................. 1331
2.4.1. Oligarchy model I...................................................................................................................................................... 1331
2.4.2. Oligarchy model II .................................................................................................................................................... 1331
3. The Gini and Pietra indices .................................................................................................................................................................. 1331

∗ Corresponding author. Tel.: +972 507 290 650.


E-mail addresses: eliazar@post.tau.ac.il (I.I. Eliazar), igor.sokolov@physik.hu-berlin.de (I.M. Sokolov).

0378-4371/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.physa.2011.09.007
1324 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

3.1. The distance indices................................................................................................................................................................. 1332


3.2. The Gini index .......................................................................................................................................................................... 1332
3.2.1. Gini’s index: probabilistic representation............................................................................................................... 1333
3.2.2. Gini’s index: stochastic representation ................................................................................................................... 1333
3.3. The Pietra index ....................................................................................................................................................................... 1334
3.3.1. Pietra’s index: probabilistic representation............................................................................................................ 1334
3.3.2. Pietra’s index: stochastic representation ................................................................................................................ 1335
3.4. Discussion................................................................................................................................................................................. 1335
3.4.1. Gini’s index vs. Pietra’s index................................................................................................................................... 1335
3.4.2. Shifts and their effect on the distance indices ........................................................................................................ 1335
4. Pietra’s index revisited ........................................................................................................................................................................ 1336
4.1. Pietra’s index and Hoover’s ‘‘Robin Hood’’ index .................................................................................................................. 1336
4.2. Pietra’s index and renewal processes ..................................................................................................................................... 1336
4.3. Pietra’s index and M /G/∞ systems ....................................................................................................................................... 1337
4.4. Pietra’s index and financial call options ................................................................................................................................. 1337
5. The Amato and curvature indices ....................................................................................................................................................... 1338
5.1. The Amato index ...................................................................................................................................................................... 1338
5.2. The curvature index................................................................................................................................................................. 1339
6. Evenness and randomness .................................................................................................................................................................. 1340
6.1. The measurement of randomness .......................................................................................................................................... 1340
6.2. The Rényi indices ..................................................................................................................................................................... 1341
6.3. The Hirschman and Theil indices............................................................................................................................................ 1342
7. Pareto’s 20–80 rule and probability law............................................................................................................................................. 1343
7.1. Pareto’s 20–80 rule .................................................................................................................................................................. 1344
7.1.1. Lorenz-curve reconstruction.................................................................................................................................... 1344
7.1.2. Evenness bounds ...................................................................................................................................................... 1345
7.2. Pareto’s probability law........................................................................................................................................................... 1346
7.2.1. Pareto’s probability law and power-law Lorenz curves ......................................................................................... 1346
7.2.2. Lorenzian fractality and power-law Lorenz curves ................................................................................................ 1346
8. A Lorenzian limit law for rank distributions ...................................................................................................................................... 1347
8.1. Rank distributions and their Lorenz curves ........................................................................................................................... 1347
8.2. The Lorenzian limit law ........................................................................................................................................................... 1347
8.3. Network topologies ................................................................................................................................................................. 1348
9. Poissonian population models ............................................................................................................................................................ 1348
9.1. The topology of Poissonian populations ................................................................................................................................ 1349
9.2. The evenness of Poissonian populations ................................................................................................................................ 1350
9.3. Discussion: evenness vs. randomness .................................................................................................................................... 1350
10. Conclusions........................................................................................................................................................................................... 1350
Acknowledgment ................................................................................................................................................................................. 1352
Appendix............................................................................................................................................................................................... 1352
References............................................................................................................................................................................................. 1352

1. Introduction

How to measure the statistical heterogeneity of a given data-set? This question stands at the very foundation of almost
all fields of science, and has different context-dependent answers. In the physical sciences there are two main classes of
gauges of statistical heterogeneity. The first class is commonly applied in the context of transport processes, and quantifies
the dispersion of the processes under consideration via gauges such as Mean Square Displacement (MSD) [1,2]. The MSD
is essentially equivalent to the standard deviation — which is perhaps the most common gauge of statistical heterogeneity
applied in science at large. The second class is commonly applied in the context of the statistical physics of large systems, and
quantifies the randomness of the systems under consideration via gauges such as entropy [3,4]. Entropy is also the elemental
gauge of statistical heterogeneity applied in information theory [5,6].
In the social sciences the question of statistical heterogeneity is often addressed in the context of societal egalitarianism:
How equal is the distribution of wealth in a given human population? Societal egalitarianism is measured via the class of
‘‘equality indices’’ [7,8], of which the most commonly applied gauge is Gini’s index [9]. Gini’s index takes values in the unit
interval; the closer the index is to zero – the more equal the distribution of wealth; the closer the index is to one – the more
unequal the distribution of wealth. Entropy and Gini’s index thus tackle the measurement of statistical heterogeneity via
diametrically different perspectives: entropy quantifies the randomness of the data-sets under consideration, whereas Gini’s
index quantifies the evenness of the data-sets under consideration.
Although stemming from an economic origin, Gini’s index is far from being confined to the measurement of societal
egalitarianism alone. Rather, Gini’s index is a quantitative gauge of statistical evenness in the context of general data-sets
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1325

with positive data values. And indeed, in recent years the application of Gini’s index has spread beyond economics to diverse
fields of science. Examples include: anthropology — demographic changes and kinship in anthropological populations [10];
astrophysics — the analysis of galaxy morphology [11]; developmental biology — gene expression in embryonic stem
cells [12]; ecology — the effect of biodiversity on ecosystem functioning [13]; ecology — the extinction risk of mammal
populations [14]; econophysics — wealth inequality in the minority game [15]; environmental sciences — the effect of
climate change on marine biodiversity [16]; finance — the analysis of inter-trade time intervals [17]; medical chemistry
— the analysis of kinase inhibitors [18]; population biology — the transmission of infectious agents [19]; sustainability
science — land-change [20]; sustainability science — vulnerability of underdeveloped countries to climate-related
extreme events [21].
Gini’s index is only one possible gauge of statistical evenness. Indeed, economics has established a whole toolbox of
gauges devised to measure societal egalitarianism: the Lorenz curve [22]; Pietra’s index [23]; Hoover’s ‘‘Robin Hood’’ index [24];
Amato’s index [25]; Hirschman’s index [26]; Theil’s index [27]; Atkinson’s indices [28]. To date, none of the aforementioned
gauges has attained – beyond economics and the social sciences – the popularity of Gini’s index. Yet so, the recent
growth in the application-popularity of Gini’s index was accompanied by recent theoretical advances – published in
statistical physics journals – in ‘‘evenness-based’’ measurement of statistical heterogeneity: Lorenzian analysis of infinite
Poissonian populations [29]; the measurement of statistical heterogeneity via Pietra’s index [30]; the diversity of Poissonian
populations [31]; the maximization of statistical heterogeneity [32]; Gini characterization of extreme-value statistics [33];
the Pietra term structure of financial assets [34]; the connection between randomness and evenness, and Rényi’s indices [35];
the emergence of Pareto’s law from Zipf’s law [36]; the Heapsian structure of growth processes whose evolution is governed
by Zipf’s law [37]; Lorenzian limit law for rank distributions [38].
The goal of this Paper is to provide a panoramic overview of the measurement of statistical evenness, and to expose the
physics community to the rich world stretching out beyond Gini’s index. We believe that as physics marches on towards
‘‘new frontiers’’ such as Econophysics [39–41], Complex Networks [42,43] and Social Dynamics [44,45], the assimilation of the
economical toolbox of evenness gauges into the machineries of statistical physics is of significant importance. The paper is
self-contained, and its remainder is organized as follows:

• Section 2 reviews the Lorenz curve — a universally calibrated method to measure the distribution of wealth within human
populations.
• Section 3 reviews the Gini and Pietra indices — two gauges of societal egalitarianism based on curve-distances in the
space of Lorenz curves.
• Section 4 further investigates Pietra’s index, its special properties, and its coincidence with Hoover’s ‘‘Robin Hood’’ index.
• Section 5 reviews the Amato index and establishes the curvature index — two gauges of societal egalitarianism based,
respectively, on the length and curvature of Lorenz curves.
• Section 6 reviews Rényi’s indices — a family of entropy-based gauges of societal egalitarianism which generalize and
connect together the Hirschman and Theil indices, and which are analogous to Atkinson’s indices.
• Section 7 explores Pareto’s 20–80 rule and Pareto’s probability law from a ‘‘Lorenzian perspective’’.
• Section 8 establishes a ‘‘Lorenzian limit law’’ which provides a universal macroscopic statistical classification of rank
distributions and network topologies.
• Section 9 addresses the measurement of statistical evenness in the context of populations modeled by Poisson
processes.

As noted already, the economic indices of societal egalitarianism are – in the border context of general data-sets with
positive data values – gauges of statistical evenness. Thus, the following terms shall henceforth be used interchangeably:
‘‘gauges of societal egalitarianism’’, ‘‘gauges of statistical evenness’’, and the shortcut ‘‘evenness gauges’’. Readers interested
in economic-oriented overviews of the measurement of statistical evenness may further consult [7,8].
A note about notation. Throughout the Paper: (i) i.i.d. is the acronym for ‘‘independent and identically distributed’’; (ii) ⟨X ⟩
denotes the mean of a random variable X ; (iii) φ −1 (·) denotes the inverse function of a monotone increasing (decreasing)
function φ(·); (iv) (x)+ = max(0, x) denotes the positive part of a real number x; (v) in Tables 1–4 we use the following
special functions:

1. The tail probability of the standard Normal law:


∫ ∞  
1 −1
Φ (z ) = √ exp x2 dx (1)
z 2π 2
(−∞ < z < ∞).
2. The Gamma function:
∫ ∞
Γ (p) = exp(−x)xp−1 dx (2)
0

(p > 0).
1326 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

3. The tail probability of the Gamma law:


∫ ∞
1
Γ (p; y) = exp(−x)xp−1 dx (3)
y Γ (p)
(y ≥ 0; p > 0).

Table 1
Examples of probability laws. This table presents seven classes of probability laws whose
statistical evenness shall be hereafter analyzed. All the probability-law classes depend on
a positive-valued parameter p. The probability laws’ characterizing details – range, tail
probability, and mean – are specified in the respective columns. At the parameter value p = 1
the Beta I and Beta II probability laws yield the Uniform probability law. At the parameter
value p = 1 the Weibull probability law yields the exponential probability law.
Law Range Tail Ψ (v) = Mean ⟨W ⟩ =
−(1+p)
Pareto I (1, ∞) v 1+p
p
Pareto II (0, ∞) (1 + v)−(1+p) 1
p

1 − vp
p
Beta I (0, 1) 1+p

Beta II (0, 1) (1 − v)p 1


1+p
exp(p)−1
Zipf (1, exp(p)) 1− 1
p
ln(v) p

Log-normal (0, ∞) Φ ( 1p ln(v)) exp( 12 p2 )

Weibull (0, ∞) exp(−v p ) Γ ( 1p )

Table 2
Examples of Lorenz curves. This table presents the Lorenz curves corresponding to the seven probability-law
classes of Table 1. In the case of the exponential probability law – i.e., the Weibull probability law with
exponent p = 1 – the Lorenz curve admits the explicit form Lexp (x) = x − x ln(x). The limiting Lorenz curves
obtained as the parameter p tends to zero (third column) and to infinity (fourth column) include: (i) the ‘‘line
of perfect equality’’ Lpe (x) = x which represents the purely communist societal scenario; (ii) the ‘‘line of
perfect inequality’’ Lpi (x) = 1 which represents the purely monarchic societal scenario; (iii) the Lorenz
curve Lexp (x) which corresponds to the exponential probability law.

Law Lorenz L(x) = limp→0 L(x) = limp→∞ L(x) =


p
Pareto I x 1 +p Lpi (x) Lpe (x)
p
Pareto II (1 + p)x 1+p − px Lpi (x) Lexp (x)
1+p
Beta I 1 − (1 − x ) p Lpi (x) Lpe (x)
1+p
Beta II (1 + p)x − px p Lpe (x) Lexp (x)
1−exp(−px)
Zipf 1−exp(−p)
Lpe (x) Lpi (x)
Log-normal Φ (Φ −1 (x) − p) Lpe (x) Lpi (x)
Weibull Γ ( 1+p p ; − ln(x)) Lpi (x) Lpe (x)

Table 3
Probability laws and their Gini and Pietra indices. This table presents the Gini
and Pietra indices corresponding to the seven probability-law classes of
Table 1.
Law Gini G = Pietra P =
1 pp
Pareto I 1+2p (1+p)1+p

Pareto II
1+p
1+2p
( 1+p p )p
1 pp
Beta I 1+2p (1+p)1+p

Beta II
p
1+2p
( 1+p p )1+p
1+exp(−p)
ln( 1−exp(−p) )
2 1 1 pe
Zipf 1−exp(−p)
− p 1−exp(−p)
− p

1 − 2Φ ( √ ) 1 − 2Φ ( )
p p
Log-normal 2
2
−1/p
Weibull 1−2 Γ ( 1p ; Γ ( 1+p p )p )
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1327

Table 4
Probability laws and their Rényi indices. This table presents the Rényi indices corresponding to the seven
probability-law classes of Table 1. The Rényi indices with exponents α ̸= 1 are given in the second column;
the Rényi indices with exponents α = 1 (the Theil-index counterpart) and α = 2 (the Hirschman-index
counterpart) are given, respectively, in the third and fourth columns. In the pair of Paretian examples the
Rényi indices (α ̸= 1) are restricted – due to moment-convergence – to the exponent range 0 < α < 1 + p.
Law 1 − Rα = 1 − R1 = 1 − R2 =
1
Pareto I
1+p
p
( 1−α+p )
p 1−α
1+p
p
exp(− ) 1
p
1− 1
p2
1
Pareto II 1
p
( Γ (1+α)ΓΓ(p(1)−α+p) ) 1−α 1
p
Γ ′ ( p)
exp( Γ (p) − Γ ′ (2)) 1
2
(1 − 1p )
1
Beta I
p
1+p
( α+ p
)
1+p 1−α p
1+p
exp( 1+
1
p
) 1 − (1+1p)2
1
Beta II 1
1+p
( Γ (Γ1+α)Γ (2+p) 1−α
(1+α+p)
) 1
1+p
Γ ′ (2+p)
exp( Γ (2+p) − Γ ′ (2)) 1
2
(1 + 1
1+p
)
1
1−exp(−p) 1 1−exp(−α p) 1−exp(−p) 2 1−exp(−p)
Zipf p
( α 1−exp(−p)
) 1−α p
exp(1 −
p
exp(p)−1
) p 1+exp(−p)

Log-normal exp(− α2 p2 ) exp(− 12 p2 ) exp(−p2 )


α+p
Γ( ) 1 Γ ′ ( 1 +p ) Γ ( 1p )2
Weibull Γ ( 1+p p )( p
) 1−α Γ ( 1+p p ) exp(− p
)
Γ ( 1+ p
p )
Γ ( 1p ) 2pΓ ( 2p )

2. The Lorenz curve

In 1905 the American statistician Max Lorenz devised a remarkable quantitative method for measuring the distribution
of wealth within human populations — a method referred to nowadays as the Lorenz curve [22]. The Lorenz curve y = L(x)
(0 ≤ x, y ≤ 1) of a given human population reads out as follows: the top 100x% of the population are in possession of
100y% of the total population’s wealth. The Lorenz curve measures the distribution of wealth within human populations in
a universally calibrated way, using an intrinsic scaling mechanism. No matter what the population considered is, no matter
in what currency wealth is measured, and no matter the range of the population’s wealth values—the Lorenz curve will
always represent the population’s distribution of wealth in a universally-calibrated functional form.
In this section we review the Lorenz curve using two different approaches: a census approach (Section 2.1), and a
probabilistic approach (Section 2.2). The exposition of the two approaches is followed by a short discussion (Section 2.3),
and is exemplified by two models of oligarchic societal scenarios (Section 2.4).

2.1. The Lorenz curve: a census approach

Consider a population consisting of n members labeled by the index i = 1, . . . , n. Let wi denote the wealth of the ith
population member, and let w = (w1 + · · · + wn )/n denote the average population wealth. With no loss of generality,
assume that the labeling of the population-members is non-increasing in their wealth: w1 ≥ w2 ≥ · · · ≥ wn ≥ 0.
The aggregate wealth of the k richest population members is w1 + · · · + wk and hence – following the definition of the
Lorenz curve – the value of the population’s Lorenz curve at the point k/n is given by

w1 + · · · + wk
 
k
L = (4)
n w1 + · · · + wn
(k = 1, . . . , n). In between the rational points k/n (k = 1, . . . , n) the Lorenz curve is linearly interpolated. The Lorenz curve
obtained is thus a piecewise linear function, whose slopes are given by

wk
  k  k−1 
k L −L
1L = n n
= (5)
n 1
n
w
(k = 1, . . . , n). The properties of the Lorenz curve follow straightforwardly from Eqs. (4) and (5):
1. Scale invariance. The Lorenz curve is invariant with respect to the currency by which wealth is measured: if the
measurement of wealth is changed by a positive scale factor s (namely: wi → swi ) then the Lorenz curve is left
unchanged.
2. Boundary conditions. The Lorenz curve initiates at the origin L(0) = 0 (by definition), and terminates at L(1) = 1.
3. Monotonicity. The Lorenz curve is monotone non-decreasing. Namely, the slopes of the Lorenz curve are non-negative
valued:
 
k
1L ≥0 (6)
n
(k = 1, . . . , n).
1328 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

4. Concavity. The Lorenz curve is concave. Namely, the slopes of the Lorenz curve are monotone non-increasing:
   
k k+1
1L ≥ 1L (7)
n n
(k = 1, . . . , n − 1).
5. Bounds. The Lorenz curve is bounded from below by the ‘‘line of perfect equality’’ Lpe (x) = x, and is bounded from above
by the ‘‘line of perfect inequality’’ Lpi (x) = 1 (see explanation below):

x ≤ L(x) ≤ 1 (8)
(0 ≤ x ≤ 1).
6. Inversion. The wealth values of the population members can be reconstructed from the Lorenz curve via the inversion
formula
 
k
wk = w · 1L (9)
n
(k = 1, . . . , n).
The lower and upper bounds stated in Property #5 stem from two extreme societal scenarios—communism and
monarchy. Communism yields the ‘‘line of perfect equality’’ Lpe (x) = x and monarchy yields the ‘‘line of perfect
inequality’’ Lpi (x) = 1, as we now explain.
Communism. Consider a purely communist society in which the population wealth is equally distributed amongst its
members. Namely, if the total population wealth is w then each population member is in possession of an equal part of
it: w1 = w2 = · · · = wn = w/n. In this scenario L(k/n) = k/n (k = 1, . . . , n) and hence in the limit n → ∞ the Lorenz
curve converges to the, so called, ‘‘line of perfect equality’’ : Lpe (x) = x (x ≥ 0). The line of perfect equality Lpe (x) is the lower
bound of all Lorenz curves since the minimum of the sum w1 + · · · + wk – the aggregate wealth of the k richest population
members – is indeed attained by the ‘‘communist scenario’’.
Monarchy. Consider a purely monarchic society in which the population wealth is held exclusively by a single monarch.
Namely, if the total population wealth is w then the monarch is in sole possession of all the wealth — leaving all other
population members completely impoverished: w1 = w and w2 = · · · = wn = 0. In this scenario L(k/n) = 1
(k = 1, . . . , n) and hence in the limit n → ∞ the Lorenz curve converges to the, so called, ‘‘line of perfect inequality’’ :
Lpi (x) = 1 (x > 0). The line of perfect inequality Lpi (x) is the upper bound of all Lorenz curves since the maximum of the
sum w1 + · · · + wk – the aggregate wealth of the k richest population members – is indeed attained by the ‘‘monarchic
scenario’’.

2.2. The Lorenz curve: a probabilistic approach

Consider a population in which the distribution of wealth is governed, statistically, by the probability density function
ψ(w) (w ≥ 0). Namely, if a member of the population is sampled at random then its wealth – henceforth denoted W – is a
random variable whose probability law is governed by the probability density function ψ(w). Let the function
∫ ∞
Ψ (v) = ψ(w)dw = Pr(W > v) (10)
v
(v ≥ 0) denote the corresponding tail probability. Namely, Ψ (v) is the probability that a randomly-sampled member of the
population is richer than the value v . The mean of the random variable W – i.e., the mean wealth of a randomly-sampled
member of the population – is given by
∫ ∞ ∫ ∞
⟨W ⟩ = wψ(w)dw = Ψ (v)dv. (11)
0 0
∞
The population members who are richer than the value v constitute a fraction of size v ψ(w)dw of the entire
∞
population, and are in aggregate possession of v wψ(w)dw wealth. In particular, the total population wealth is given
∞
by 0
wψ(w)dw . Hence – due to the definition of the Lorenz curve – the population’s Lorenz curve is implicitly given by
 ∞
wψ(w)dw
∫ ∞
L ψ(w)dw = v∞ (12)
v 0
wψ(w)dw
(v ≥ 0). Differentiating both sides of Eq. (12) with respect to the variable v , and using the tail probability Ψ (v) and the
mean wealth ⟨W ⟩, we obtain that
1
L′ (Ψ (v)) = v (13)
⟨W ⟩
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1329

(v ≥ 0). Using the change of variables u = Ψ (v) we further obtain that


1
L ′ ( u) = Ψ − 1 ( u) (14)
⟨W ⟩
(0 ≤ u ≤ 1). Finally, integrating Eq. (14) and using the boundary condition L(0) = 0, we conclude that
∫ x
1
L(x) = Ψ −1 (u)du (15)
⟨W ⟩ 0
(0 ≤ x ≤ 1). The properties of the Lorenz curve given by Eq. (15) are identical to the Lorenz-curve properties stated in
Section 2.1 (see explanation below):
1. Scale invariance. The Lorenz curve is invariant with respect to the currency by which wealth is measured: if the
measurement of wealth is changed by a positive scale factor s (namely: W → sW ) then the Lorenz curve is left
unchanged.
2. Boundary conditions. The Lorenz curve initiates at the origin L(0) = 0 (by definition), and terminates at L(1) = 1.
3. Monotonicity. The Lorenz curve is monotone non-decreasing:
L′ (x) ≥ 0 (16)
(0 ≤ x ≤ 1).
4. Concavity. The Lorenz curve is concave:
L′′ (x) ≤ 0 (17)
(0 ≤ x ≤ 1).
5. Bounds. The Lorenz curve is bounded from below by the line of perfect equality Lpe (x) = x, and is bounded from above
by the line of perfect inequality Lpi (x) = 1. Namely:
x ≤ L(x) ≤ 1 (18)
(0 ≤ x ≤ 1).
6. Inversion. The reconstruction of the tail probability Ψ (v) from the Lorenz curve L(x) is given by the inversion
formula
v
 
Ψ (v) = (L ) ′ −1
(19)
⟨W ⟩
(v ≥ 0).
Explanation. Scale invariance follows from Eq. (15) by a straightforward change of variables. Setting v = 0 in Eq. (12)
yields the boundary condition L(1) = 1. Monotonicity follows from Eq. (14). Differentiating Eq. (14) yields L′′ (x) =
−1/[⟨W ⟩ψ(Ψ −1 (u))] which, in turn, implies concavity. The upper bound L(x) ≤ 1 follows from Eq. (12). Eq. (14) implies
that L′ (0) = wmax /⟨W ⟩ ≥ 1, where 0 < wmax ≤ ∞ is the upper bound of the random variable W . The lower bound
L(x) ≥ x follows from the combination of the facts that the Lorenz curve is monotone, L′ (0) ≥ 1 and L(1) = 1. The
inversion formula of Eq. (19) follows straightforwardly from Eq. (14).
Examples. Examples of various classes of probability laws and their corresponding Lorenz curves are given in Tables 1 and 2.
Table 1 specifies the characterizing details of the probability-law classes, and Table 2 presents both the Lorenz curves and
the limiting Lorenz curves of the probability-law classes. Graphical illustrations of Lorenz curves are given in Figs. 1 and 2.

2.3. Discussion

2.3.1. The census approach vs. the probabilistic approach


We described the Lorenz curve via two different approaches: census and probabilistic. The census approach (presented
in Section 2.1) was based on ‘‘census data’’ — the sequence {wi }ni=1 representing the precise wealth values of all members of
the population under consideration. The probabilistic approach (presented in Section 2.2) was based on ‘‘sample data’’—the
random variable W representing the wealth of a randomly-sampled member of the population under consideration.
The probabilistic approach generalizes the census approach. Indeed, sampling a random population-member from a
population with wealth values {wi }ni=1 yields a random variable W which is uniformly distributed over the set of outcomes
{wi }ni=1 . On the other hand, the census approach can be regarded as a discretized approximation of the probabilistic approach.
Indeed, consider a population in which the distribution of wealth is governed by the probability density function ψ(w)
(w ≥ 0), and let the random variable W denote the wealth of a randomly sampled member of the population. Now, discretize
the probability density function ψ(w) into n equiprobable bins, represent bin i by a single wealth level wi (i = 1, . . . , n),
and set Wn to be a random variable which is uniformly distributed over the set {wi }ni=1 . Then, it is possible to show that the
discretized random variable Wn converges in law, as n → ∞, to the random variable W .
Thus, in the case of large populations (n ≫ 1), the census approach and the probabilistic approach are essentially
equivalent. Due to this observation we shall henceforth – in the context of large populations (n ≫ 1) – apply the census and
probabilistic approaches interchangeably.
1330 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

Fig. 1. Lorenz curves. The graphs of five Lorenz curves, corresponding to the Pareto I class of probability laws (see Table 2), are depicted. The Lorenz curves
are characterized – in an increasing order – by the parameter values p = ∞ (the line y = x), p = 3/2, p = 1/2, p = 1/6, and p = 0 (the line y = 1).

Fig. 2. Lorenz curves. The graphs of five Lorenz curves, corresponding to the Beta I class of probability laws (see Table 2), are depicted. The Lorenz curves
are characterized – in an increasing order – by the parameter values p = ∞ (the line y = x), p = 3/2, p = 1/2, p = 1/6, and p = 0 (the line y = 1).

2.3.2. Normalization
Scaling the random variable W – the wealth of a randomly-sampled member of the population – by its mean ⟨W ⟩ yields
the normalized random variable
W
Ŵ = . (20)
⟨W ⟩
The normalized random variable Ŵ shall turn out to play a key role in the sequel.
The tail probability of the normalized random variable Ŵ is given by

Ψ̂ (v) = Pr(Ŵ > v) = Ψ (⟨W ⟩v) (21)


(v ≥ 0). Substituting Eq. (21) into Eq. (15) yields the following ‘‘normalized representation’’ of the Lorenz curve:
∫ x
L(x) = Ψ̂ −1 (u)du (22)
0

(0 ≤ x ≤ 1). Consequently, the reconstruction of the normalized tail probability Ψ̂ (v) from the Lorenz curve L(x) is given
by the inversion formula

Ψ̂ (v) = (L′ )−1 (v) (23)


(v ≥ 0).
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1331

2.4. Two oligarchy models

We conclude this section with two models of oligarchic societal scenarios. These models exemplify the application of
both the census and probabilistic approaches described above, and will serve us in the sequel.

2.4.1. Oligarchy model I


Consider a purely oligarchic society in which the total population wealth is held exclusively by an oligarchy which
constitutes a fraction f (0 < f < 1) of the entire population. Assume that amongst the oligarchy wealth is equally
distributed. Namely, if m = fn denotes the oligarchy size, and if w denotes the total population wealth, then: (i) each
oligarchy member is in possession of an equal part of the total wealth w1 = · · · = wm = w/m; (ii) all the ‘‘non-
oligarchy’’ population members are completely impoverished wm+1 = · · · = wn = 0.
For 1 ≤ k ≤ m we have L(k/n) = k/m = (k/n)/f , and hence in the limit n → ∞ the following piecewise-linear Lorenz
curve is obtained:

1x 0 ≤ x ≤ f,
L(x; f ) = f (24)
1 f ≤ x ≤ 1.

This oligarchy model spans a full spectrum of societal scenarios ranging from the communist extreme to the monarchic
extreme. Indeed, in the limit f → 1 – i.e., as the oligarchy tends to communism – the Lorenz curve L(x; f ) converges to the
line of perfect equality Lpe (x) = x. On the other hand, in the limit f → 0 – i.e., as the oligarchy tends to a monarchy – the
Lorenz curve L(x; f ) converges to the line of perfect inequality Lpi (x) = 1.
In terms of the probabilistic approach this oligarchy model is represented by a binomial law: the random variable W –
which represents the wealth of a randomly-sampled member of the population – attains either the value w/m = (w/n)/f
with probability f , or the value 0 with probability 1 − f . The corresponding mean is ⟨W ⟩ = w/n, and the corresponding
normalized random variable Ŵ = W /⟨W ⟩ is given by

1 w.p. f ,
Ŵ = f (25)
0 w.p. 1 − f .

2.4.2. Oligarchy model II


Consider an oligarchic society in which a fraction f2 of the total population wealth is held by an oligarchy which
constitutes a fraction f1 of the entire population (0 < f1 < f2 < 1). Assume that both amongst the oligarchy and amongst
the ‘‘non-oligarchy’’ – henceforth termed the ‘‘proletariat’’ – wealth is equally distributed. Namely, if m = f1 n denotes the
oligarchy size, and if w denotes the total population wealth, then: (i) each oligarchy member is in possession of an equal
part of the total oligarchy wealth w1 = · · · = wm = w f2 /m; (ii) each proletariat member is in possession of an equal part
of the total proletariat wealth wm+1 = · · · = wn = w(1 − f2 )/(n − m).
For 1 ≤ k ≤ m we have L(k/n) = kf2 /m = (f2 /f1 )(k/n). On the other hand, for m ≤ k ≤ n we have L(k/n) =
f2 + (k − m)(1 − f2 )/(n − m) = f2 + ((1 − f2 )/(1 − f1 ))(k/n). Hence, in the limit n → ∞ the following piecewise-linear
Lorenz curve is obtained:

f2
 x 0 ≤ x ≤ f1 ,


f1
L(x; f1 , f2 ) = (26)
1 − f2
f 2 + (x − f1 ) f1 ≤ x ≤ 1.


1 − f1
Namely, the Lorenz curve L(x; f1 , f2 ) is the linear interpolation passing through the points L(0) = 0, L(1) = 1, and
L(f1 ) = f2 (the first two points are the universal boundary conditions of all Lorenz curves).
In terms of the probabilistic approach this oligarchy model is represented by a binomial law: the random variable W
– which represents the wealth of a randomly-sampled member of the population – attains either the value w f1 /m =
(w/n)(f2 /f1 ) with probability f1 , or the value (w/n)((1 − f2 )/(1 − f1 )) with probability 1 − f2 . The corresponding mean
is ⟨W ⟩ = w/n, and the corresponding normalized random variable Ŵ = W /⟨W ⟩ is given by

f2
w.p. f1 ,


f1

Ŵ = (27)
 1 − f2

 w.p. 1 − f1 .
1 − f1

3. The Gini and Pietra indices

How equal – or unequal – is the distribution of wealth within a given human population? The Lorenz curve facilitates a
natural and straightforward quantitative approach to tackle the measurement of societal egalitarianism. Let y = L(x) be the
1332 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

Lorenz curve of the human population under consideration. Since the line of perfect equality represents a purely communist
society, the closeness of the Lorenz curve y = L(x) to the line of perfect equality Lpe (x) = x quantifies the societal
egalitarianism of the population considered: the closer the curves – the more equal the distribution of wealth; the further
away the curves – the more unequal the distribution of wealth.
In this section we follow the aforementioned ‘‘distance approach’’ to quantitatively measure societal egalitarianism. We
begin with the introduction of a family of ‘‘distance indices’’ (Section 3.1), and then focus on two specific distance indices:
Gini’s index (Section 3.2), and Pietra’s index (Section 3.3). We conclude the section with a discussion (Section 3.4).

3.1. The distance indices

The ‘‘Lp distance’’ between two general curves y = K1 (x) and y = K2 (x) (0 ≤ x ≤ 1) is given by

∫ 1
 1p
‖K1 − K2 ‖p = |K1 (x) − K2 (x)| dx
p
, (28)
0

where the parameter p takes values in the range 1 ≤ p ≤ ∞ [46].1 The Lp distance between the Lorenz curve y = L(x)
and the line of perfect equality Lpe (x) = x is thus given by ‖L − Lpe ‖p . The maximal Lp distance from the line of perfect
equality Lpe (x) = x is attained by the line of perfect inequality Lpi (x) = 1, and is given by ‖Lpi − Lpe ‖p = (1 + p)−1/p .
We define the ‘‘distance index’’ Dp of the population under consideration as the normalized Lp distance between the
population’s Lorenz curve y = L(x) and the line of perfect equality Lpe (x) = x. Namely:

 ∫ 1  1p
‖L − Lpe ‖p
Dp = = ( 1 + p) (L(x) − x) dx .
p
(29)
‖Lpi − Lpe ‖p 0

Eq. (29) defines an entire family {Dp }1≤p≤∞ of distance indices—all taking values in the unit interval: 0 ≤ Dp ≤ 1. The
lower bound Dp = 0 indicates coincidence of the Lorenz curve y = L(x) with the line of perfect equality Lpe (x) = x, and
the upper bound Dp = 1 indicates coincidence of the Lorenz curve y = L(x) with the line of perfect inequality Lpi (x) = 1.
Every distance index Dp is thus a gauge of societal egalitarianism: the smaller the index – the more equal the distribution
of wealth; the greater the index – the more unequal the distribution of wealth.
As an illustrative example consider the oligarchic Lorenz curve L(x; f ) given by Eq. (24). A simple calculation implies
that the corresponding distance index is Dp (f ) = 1 − f , where f is the ‘‘oligarchy fraction’’. The distance index Dp (f ) is
monotone decreasing with respect to the oligarchy fraction f . As the oligarchy tends to monarchy (f → 0) the distance
index Dp (f ) converges to its upper bound (Dp (f ) → 1), and as the oligarchy tends to communism (f → 1) the distance
index Dp (f ) converges to its lower bound (Dp (f ) → 0).
Eq. (29) defines the distance indices {Dp }1≤p≤∞ in terms of the population’s Lorenz curve y = L(x). Yet, we would like
to obtain more elementary representations of the distance indices: (i) a ‘‘probabilistic representation’’ of distance indices
in terms of the function Ψ (v) – the tail probability governing the population’s distribution-of-wealth; (ii) a ‘‘stochastic
representation’’ of distance indices in terms of the random variable W — the wealth of a randomly-sampled member of
the population. In general, it is not possible to attain simple closed-form probabilistic and stochastic representations of
the distance indices {Dp }1≤p≤∞ . However, for the two following ‘‘boundary cases’’ closed-form probabilistic and stochastic
representations are attainable: (i) the distance index D1 – called Gini’s index – to be addressed in Section 3.2; (ii) the distance
index D∞ – called Pietra’s index – to be addressed in Section 3.3.

3.2. The Gini index

The Gini index G of a given human population is defined as its D1 distance index:
∫ 1
G = D1 = 2 (L(x) − x)dx. (30)
0

Gini’s index was devised by the Italian statistician Corrado Gini in 1912 [9], and it equals twice the area captured between the
population’s Lorenz curve y = L(x) and the line of perfect equality Lpe (x) = x. Examples of various classes of probability
laws and their corresponding Gini indices are given in Table 3, and graphical illustrations of Gini’s index are given in Fig. 3.

1 At the parameter value p = ∞ the L distance is given by ‖K − K ‖ = max


∞ 1 2 ∞ 0≤x≤1 |K1 (x) − K2 (x)|.
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1333

Fig. 3. Gini and Pietra indices. The Gini and Pietra indices, corresponding to the Pareto I and Beta I classes of probability laws (see Table 3), are depicted as
functions of the parameter p (represented by the x-axis). The solid line is the graph of the Gini indices, and the dashed line is the graph of the Pietra indices.

3.2.1. Gini’s index: probabilistic representation


Let us now follow the probabilistic approach described in Section 2.2. Our aim is to establish a formula for Gini’s index
in terms of the tail probability Ψ (v) which governs the population’s distribution of wealth.
Using Eq. (15), and thereafter changing the order of integration, implies that
1 1 x
∫ ∫  ∫ 
1
L(x)dx = Ψ (u)du dx −1
0 ⟨W ⟩ 0
0
∫ 1
1
= (1 − u)Ψ −1 (u)du. (31)
⟨W ⟩ 0
Applying the change of variables w = Ψ −1 (u) to the right-hand-side of Eq. (31), and using Eq. (11), further implies that
∫ 1 ∫ ∞
1
L(x)dx = (1 − Ψ (w))wψ(w)dw
0 ⟨W ⟩ 0
∫ ∞
1
= 1+ (Ψ (w)2 )′ w dw. (32)
2⟨W ⟩ 0
Applying integration-by-parts to the integral appearing on the right-hand-side of Eq. (32), we obtain that the area captured
beneath the Lorenz curve y = L(x) is given by
∫ 1 ∫ ∞
1
L(x)dx = 1 − Ψ (v)2 dv. (33)
0 2⟨W ⟩ 0
Finally, substituting Eq. (33) into Eq. (30), and using Eq. (11), we conclude that
∞ ∞
Ψ (v)2 dv Ψ (v)(1 − Ψ (v))dv
G = 1− ∞ 0
= 0
∞ . (34)
0
Ψ (v)dv 0
Ψ (v)dv
Eq. (34) is an explicit probabilistic formula for Gini’s index G — representing it in terms of the tail probability Ψ (v) governing
the population’s distribution-of-wealth.

3.2.2. Gini’s index: stochastic representation


We now turn to establish a formula for Gini’s index in terms of the normalized random variable Ŵ = W /⟨W ⟩ — which
represents the normalized wealth of a randomly-sampled member of the population. Using the probability density function
ψ(w) and and thereafter changing the order of integration implies that
∫ ∞ ∫ ∞ ∫ ∞  ∫ v 
Ψ (v)(1 − Ψ (v))dv = ψ(w1 )dw1 ψ(w2 )dw2 dv
0 0 v 0
∫ ∞ ∫ ∞
= (w1 − w2 )+ ψ(w1 )ψ(w2 )dw1 dw2
0 0
∫ ∞ ∫ ∞
1
= |w1 − w2 |ψ(w1 )ψ(w2 )dw1 dw2 . (35)
2 0 0
1334 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

Hence, setting W1 and W2 to be two i.i.d. copies of the random variable W , Eq. (35) implies that
∫ ∞
1
Ψ (v)(1 − Ψ (v))dv = ⟨(W1 − W2 )+ ⟩ = ⟨|W1 − W2 |⟩. (36)
0 2
Substituting Eq. (36) into Eq. (34), and using Eq. (11), we obtain that
⟨(W1 − W2 )+ ⟩ ⟨|W1 − W2 |⟩
G= = . (37)
⟨W ⟩ 2⟨W ⟩
Finally, using normalization we conclude that:
1
G = ⟨(Ŵ1 − Ŵ2 )+ ⟩ = ⟨|Ŵ1 − Ŵ2 |⟩, (38)
2
where Ŵ1 and Ŵ2 are two i.i.d. copies of the normalized random variable Ŵ . Eq. (38) is an explicit stochastic formula for
Gini’s index G — representing it in terms of the normalized random variable Ŵ . Moreover, Eq. (38) provides a clear-cut
understanding of the meaning of Gini’s index: Gini’s index equals half the mean absolute difference between two normalized
and independent population-samples Ŵ1 and Ŵ2 .

3.3. The Pietra index

The Pietra index P of a given human population is defined as its D∞ distance index:
P = D∞ = max (L(x) − x). (39)
0≤x≤1

Pietra’s index was devised by the Italian statistician Gaetano Pietra in 1915 [23], and it equals the maximal vertical distance
between the population’s Lorenz curve y = L(x) and the line of perfect equality Lpe (x) = x. Examples of various classes of
probability laws and their corresponding Pietra indices are given in Table 3, and graphical illustrations of Pietra’s index are
given in Fig. 3.

3.3.1. Pietra’s index: probabilistic representation


Let us now follow the probabilistic approach described in Section 2.2. Our aim is to establish a formula for Pietra’s index
in terms of the tail probability Ψ (v) which governs the population’s distribution of wealth. To that end we focus on the
difference function Λ(x) = L(x) − x (0 ≤ x ≤ 1).
Since the Lorenz curve y = L(x) is concave, and since the line of perfect equality Lpe (x) = x is linear —the difference
function Λ(x) is concave. Moreover, using Eq. (14), the derivative of the difference function Λ(x) is given by
1
Λ′ (x) = Ψ −1 (x) − 1 (40)
⟨W ⟩
(0 ≤ x ≤ 1). Due to its concavity, the difference function Λ(x) attains a unique global maximum on the unit interval. The
point x∗ at which the global maximum is attained is the unique solution of the equation Λ′ (x∗ ) = 0. Hence, using Eq. (40)
we obtain that x∗ = Ψ (⟨W ⟩) and, consequently, we further obtain that the value of the global maximum is given by
P = Λ(Ψ (⟨W ⟩)) = L(Ψ (⟨W ⟩)) − Ψ (⟨W ⟩). (41)
Using Eq. (15), and thereafter applying the change of variables w = Ψ −1
(u), implies that
∫ Ψ (⟨W ⟩)
1
L(Ψ (⟨W ⟩)) = Ψ −1 (u)du
⟨W ⟩ 0
∫ ∞
1
= wψ(w)dw. (42)
⟨W ⟩ ⟨W ⟩
Substituting Eq. (42) into Eq. (41) further implies that
∫ ∞ ∫ ∞
1
P = wψ(w)dw − ψ(w)dw
⟨W ⟩ ⟨W ⟩ ⟨W ⟩
∫ ∞
1
= (w − ⟨W ⟩)ψ(w)dw. (43)
⟨W ⟩ ⟨W ⟩
Finally, applying integration-by-parts to the right-hand-side of Eq. (43) we conclude that
∫ ∞
1
P = Ψ (v)dv. (44)
⟨W ⟩ ⟨W ⟩
Eq. (34) is an explicit probabilistic formula for Pietra’s index P — representing it in terms of the tail probability Ψ (v)
governing the population’s distribution-of-wealth (the population’s mean wealth ⟨W ⟩, in turn, is given by Eq. (11)).
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1335

3.3.2. Pietra’s index: stochastic representation


We now turn to establish a formula for Pietra’s index in terms of the normalized random variable Ŵ = W /⟨W ⟩—which
represents the normalized wealth of a randomly-sampled member of the population.
The right hand side of Eq. (43) implies that
∫ ∞
1
P = (w − ⟨W ⟩)+ ψ(w)dw. (45)
⟨W ⟩ 0

Eq. (45), in turn, implies that

⟨(W − ⟨W ⟩)+ ⟩
P = . (46)
⟨W ⟩
Finally, using normalization we conclude that:

P = ⟨(Ŵ − 1)+ ⟩. (47)

Eq. (47) is an explicit stochastic formula for Pietra’s index P — representing it in terms of the normalized random variable Ŵ .
Moreover, Eq. (47) provides a clear-cut understanding of the meaning of Pietra’s index: Pietra’s index is the mean exceedance
of a normalized population-sample Ŵ over its mean ⟨Ŵ ⟩ = 1.

3.4. Discussion

3.4.1. Gini’s index vs. Pietra’s index


Comparing the stochastic formulae appearing in Eqs. (38) and (47) pinpoints the difference between Gini’s index G and
Pietra’s index P . The term (Ŵ1 − Ŵ2 )+ appearing in Eq. (38) is the exceedance of the normalized population-sample Ŵ1 over
the normalized population-sample Ŵ2 . Namely, in the computation of Gini’s index a pair of normalized population-samples
are chosen independently and at random, and the exceedance of the first normalized population-sample over the second
normalized population-sample is considered: (Ŵ1 − Ŵ2 )+ .
On the other hand, the term (Ŵ − 1)+ appearing in Eq. (47) is the exceedance of the normalized population-sample Ŵ
over its mean ⟨Ŵ ⟩ = 1. Namely, in the computation of Pietra’s index a single normalized population-sample is chosen at
random, and the exceedance of the normalized population-sample over its unit mean is considered: (Ŵ − 1)+ .
The aforementioned distinction between the Gini and Pietra indices tells us when to favor one over the other. In cases
where the difference between two independent population-samples is of interest — Gini’s index G should be applied to use.
However, in cases where the exceedance of a population-sample over the population-mean is of interest — Pietra’s index
P should be applied to use. Hence, from a ‘‘propagation perspective’’, Gini’s index G measures the mean absolute deviation
between two i.i.d. random trajectories, whereas the Pietra’s index P measures the exceedance of a random trajectory over
its mean.

3.4.2. Shifts and their effect on the distance indices


We conclude this section with an analysis of the effect of shifts on the distance indices {Dp }1≤p≤∞ . Consider a population
whose distribution-of-wealth is characterized by the Lorenz curve y = L(x), and assume that the population’s government
has decided to endow each and every population-member a fixed sum of size s. How does this governmental endowment
change the population’s distribution of wealth?
Let W and Ws denote, respectively, the wealth of a randomly-sampled member of the population before and after the
endowment. Clearly, the endowment is a shift operation translating the pre-endowment random variable W to the post-
endowment random variable Ws = W + s. Hence, the tail probability and the mean of the random variable Ws transform,
respectively, to Ψs (v) = Ψ (v−s) (v ≥ s) and ⟨Ws ⟩ = ⟨W ⟩+s. Eq. (15), in turn, implies that the population’s post-endowment
Lorenz curve is given by
   
s ⟨W ⟩
Ls (x) = x+ L(x) (48)
⟨W ⟩ + s ⟨W ⟩ + s
(0 ≤ x ≤ 1). Note that in the limit s → 0 the post-endowment Lorenz curve Ls (x) converges back to the pre-endowment
Lorenz curve L(x), and in the limit s → ∞ the post-endowment Lorenz curve Ls (x) converges to the line of perfect equality
Lpe (x) = x.
Substituting Eq. (48) into Eq. (29) implies the following connection between the population’s post-endowment and pre-
endowment distance index Dp :

⟨W ⟩
Dp (W + s) = Dp (W ) (49)
⟨W ⟩ + s
1336 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

(in Eq. (49) we slightly abused notation — denoting by Dp (W +s) and Dp (W ), respectively, the population’s post-endowment
and pre-endowment distance index Dp ). Eq. (49) emphasizes the nonlinear effect of the shift operation W → W + s on
distance index Dp : an additive shift of size s applied to the random variable W results in a multiplicative evenness reduction of
factor ⟨W ⟩/(⟨W ⟩ + s). An alternative representation of this nonlinear effect is given by the following connection:
 
1
Dp (W + s) = qDp (W ) ⇔ s = − 1 ⟨W ⟩ (50)
q
(s > 0; 0 < q < 1). For example, if we want to ‘‘double the evenness’’ – i.e., attain the multiplicative factor q = 1/2 – then
a shift of size s = ⟨W ⟩ needs to be applied.

4. Pietra’s index revisited

In this section we revisit Pietra’s index and interpret it in the context of: the Bolshevik revolution (Section 4.1),
renewal processes (Section 4.2), M /G/∞ systems (Section 4.3), and financial call options (Section 4.4). These interpretations
naturally connect Pietra’s index – and rather surprisingly so – to remote and seemingly unrelated fields of science.

4.1. Pietra’s index and Hoover’s ‘‘Robin Hood’’ index

Consider the census setting of Section 2.1: A population consisting of n members labeled by the index i = 1, . . . , n, where
wi denotes the wealth of member i, and where w = (w1 + · · · + wn )/n denote the average population wealth.
Now, assume that the population undergoes a Bolshevik revolution which turns the population into a purely communist
society. The Bolshevik revolution redistributes the population’s wealth applying the ‘‘wealth transformation’’ wi → w . The
‘‘Bolshevik tax’’ set on member i of the population is thus given by ti = wi − w ; this tax is positive for the members of the
population who are richer than average, and is negative for the members of the population who are poorer than average.
Hence, the proportion of the total population wealth which was transferred from rich to poor – in order to attain an equal
distribution of wealth – is given by the ratio
(w1 − w)+ + · · · + (wn − w)+
R= . (51)
w1 + · · · + wn
The ratio R was devised by the American economist Edgar Malone Hoover in 1936 [24] as a gauge of societal egalitarianism,
and is often termed the ‘‘Robin Hood’’ index — due to the act of taking from the rich and giving to the poor.
Combining together Eqs. (5) and (51) implies that
n    
− k 1
R= 1L −1 . (52)
k=1
n + n

In turn, for large populations (n ≫ 1) Eq. (52) implies that


∫ 1
R≈ (L′ (x) − 1)+ dx. (53)
0

The derivative L′ (x) = Ψ −1 (x)/⟨W ⟩ is monotone decreasing from the level L′ (0) ≥ 1 to the level L′ (1) = 0 (recall
Eq. (14)). On the other hand, L′ (x) = 1 if and only if x = Ψ (⟨W ⟩). This observation, combined with Eq. (41), implies that
∫ 1 ∫ Ψ (⟨W ⟩)
(L′ (x) − 1)+ dx = (L′ (x) − 1)dx
0 0

= L(Ψ (⟨W ⟩)) − Ψ (⟨W ⟩) = P , (54)


where P is Pietra’s index.
Hence, we obtained that – for large populations (n ≫ 1) – Hoover’s ‘‘Robin Hood’’ index coincides with Pietra’s index.
Namely, in the context of the Bolshevik revolution the Pietra index P is: the proportion of the total population wealth which
needs to be redistributed in order to attain a purely communist society.

4.2. Pietra’s index and renewal processes

Consider a stochastic sequence of time epochs T0 < T1 < T2 < · · ·, whose differences {Tn − Tn−1 }∞ n=1 are i.i.d. random
variables. In probability theory such stochastic sequences are termed ‘‘renewal processes’’ [47]. The best-known example of
renewal processes is the Poisson process — in which the differences are exponentially distributed [48]. Renewal processes are
widely applied in science and engineering to model the occurrences of random events. Applications of renewal processes
range from Continuous Time Random Walks (CTRWs) – in which an ‘‘event’’ is a jump of the CTRW [2], to service and queuing
systems – in which an ‘‘event’’ is the arrival of a new customer [49].
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1337

Given a renewal process, let W denote its inter-event waiting time. Namely, W is the random time elapsing between the
occurrence of two consecutive events the renewal process tracks. Consider an external observer initiating its observation
(of the renewal process) at an arbitrary time point τobs ≫ T0 , and let Wobs denote the observer’s waiting time till observing
an event. Namely, Wobs is the random time elapsing between time τobs and the first event occurring after time τobs .
The connection between the tail probabilities of the inter-event waiting time W and the observer’s waiting time Wobs is
given by the formula
∫ ∞
1
Pr(Wobs > t ) = Pr(W > t ′ )dt ′ (55)
⟨W ⟩ t

(t ≥ 0) [50]. The waiting times W and Wobs coincide, in law, if and only if they are exponentially distributed — that is, if and
only if the renewal process is a Poisson process. In general however, the statistics of the observer’s waiting time Wobs can
differ rather markedly from those of the inter-event waiting time W . For example, if the inter-event waiting time W has an
infinite variance then the observer’s waiting time Wobs will have an infinite mean [50].
Combining together Eq. (55) and Eq. (44) we obtain the following ‘‘renewal representation ’’ of Pietra’s index:
P = Pr(Wobs > ⟨W ⟩). (56)
Namely, in the context of renewal processes, the Pietra index P of the inter-event waiting time W is: the probability that the
observer’s waiting time Wobs will exceed the duration t = ⟨W ⟩ — the mean inter-event waiting time, and the renewal process’s
intrinsic time-scale. Eq. (56) implies that Pietra’s index is, in effect, a natural quantitative measure for the eccentricity of
renewal processes.

4.3. Pietra’s index and M /G/∞ systems

Consider a particle-system described as follows: independent particles arrive at the system stochastically in time,
following a Poisson process [48]; each particle, after having arrived at the system, spends in the system a random sojourn
time; the particles’ sojourn times are i.i.d. random variables. Such particle-systems stem from the modeling of Geiger–Müller
counters (‘‘type-II counters’’ [51]), and constitute the basic model of infinite-server queuing systems and of infinite-
broadband transmission channels. In queuing theory such particle-systems are termed ‘‘M /G/∞ systems ’’ [52].2
Given an M /G/∞ system, let W denote the random duration of the particles’ sojourn times. Assume that the M /G/∞
system initiates at time τ0 , and consider an observation of the system initiating at time τobs ≫ τ0 . The correlation between
the random number of particles present in the system at time τobs , and the random number of particles present in the system
at time τ obs + t, is given by the M /G/∞ auto-correlation function
∫ ∞
1
ρM /G/∞ (t ) = Pr(W > t ′ )dt ′ (57)
⟨W ⟩ t

(t ≥ 0) [53].
Based on their underlying sojourn times, M /G/∞ systems are capable of generating a wide spectrum of correlation
 ∞ M /G/∞ systems are capable of generating Long Range Dependence (LRD): non-integrable auto-
structures. In particular,
correlation functions 0 ρM /G/∞ (t )dt = ∞ [54,55] (the phenomena of LRD is often referred to as the ‘‘Joseph Effect’’ [56]).
Specifically, an M /G/∞ system generates LRD if and only if the sojourn time W has an infinite variance [53].
Combining together Eq. (57) and Eq. (44) we obtain the following ‘‘M /G/∞ representation’’ of Pietra’s index:

P = ρM /G/∞ (⟨W ⟩). (58)


Namely, in the context of M /G/∞ systems, the Pietra index P of the sojourn time W is: the magnitude of the M /G/∞ auto-
correlation function at the time-lag t = ⟨W ⟩—the particles’ mean sojourn time, and the M /G/∞ system’s intrinsic time-scale.
Eq. (58) implies that Pietra’s index is, in effect, a natural quantitative measure for the temporal correlations of M /G/∞
systems.

4.4. Pietra’s index and financial call options

Consider a financial investor contemplating between two investment alternatives: a riskless investment – say a bank
account, and a risky investment – say a stock. The investor initiates its investment at time τ0 , and terminates its investment
at time τ0 + t. One dollar, invested at time τ0 in the riskless bank account, will yield at time τ0 + t the deterministic and
predictable value VBank . On the other hand, investing the same dollar at time τ0 in the risky stock will yield at time τ0 + t the
stochastic and unpredictable outcome which we denote W . The exceedance of the risky stock investment over the riskless
bank investment is given by (W − V Bank )+ .

2 In the ‘‘M /G/∞ notation’’ : the letter M represents a Markovian (Poisson) process governing the particles’ arrivals; the letter G represents a General
distribution of the particles’ sojourn times; the symbol ∞ represents an infinite particle-processing capacity.
1338 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

A ‘‘call option’’ – contingent on the risky stock investment – is a financial contract assuring its holder the right, but not the
obligation, to purchase one stock unit at a specified time (called ‘‘maturity’’) and at a specified price (called ‘‘strike’’) [57].
Assume, with no loss of generality, that the value of one stock unit at time τ0 is one dollar. Then, a call option maturing at
time τ0 + t with strike s will yield, at maturity, the stochastic and unpredictable payoff (W − s)+ . The theory of financial
derivatives asserts that the value, at time τ0 , of a call option maturing at time τ0 + t with strike s is given by the following
‘‘risk-neutral’’ option-pricing formula:
⟨(W − s)+ ⟩∗
VCall (s) = (59)
VBank
(s > 0) [58]. The mean ⟨·⟩∗ appearing in Eq. (59) is with respect to the, so called, ‘‘risk-neutral probability’’—a probability
under which the mean stock value coincides with the value of the riskless bank investment: ⟨W ⟩∗ = VBank [58].
Combining together Eqs. (59) and (46) we obtain the following ‘‘call-option representation ’’ of Pietra’s index:
P = VCall (⟨W ⟩∗ ). (60)
Namely, in the context of financial call options, the Pietra index P of the stock-investment outcome W is: the value of a
call option with payoff (W − VBank )+ — the exceedance of the risky stock investment over the riskless bank investment. Eq. (60)
implies that Pietra’s index is, in effect, an intrinsic risk-pricing mechanism applied by financial markets to price call options.
The connection between Pietra’s index and financial call options further leads to the notion of the ‘‘Pietra term structure of
financial assets’’ [34].

5. The Amato and curvature indices

In Section 3 we focused on ‘‘Lp gauges’’ of societal egalitarianism based on the distance between the Lorenz curve
y = L(x) and the line of perfect equality Lpe (x) = x. In this section we shift our focus to gauges of societal egalitarianism
which are based on intrinsic characteristics of Lorenz curves: length and curvature.
Intuitively it is rather evident that the longer the Lorenz curve y = L(x), or the more curved the Lorenz curve — the
more distant it is from the line of perfect equality Lpe (x) = x. Hence, both the length and the curvature of the Lorenz curve
y = L(x) can serve as quantitative measures of societal egalitarianism. In this section we: (i) review Amato’s index – which is
a length-based gauge of societal egalitarianism (Section 5.1); (ii) establish the curvature index – which is a curvature-based
gauge of societal egalitarianism (Section 5.2).

5.1. The Amato index

The length of a general curve y = K (x) (0 ≤ x ≤ 1) is given by the formula:


∫ 1 
|K | = 1 + (K ′ (x))2 dx. (61)
0

Let y = L(x) be the Lorenz curve of a human population under consideration. Since the Lorenz curve y = L(x) is
encompassed by the triangle whose vertices are the points√ {(0, 0), (0, 1), (1, 1)}, and since it is monotone increasing and
concave, its length is: (i) bounded from below by the length 2 of the triangle edge connecting the vertices (0, 0) and (1, 1);
(ii) bounded from above by the aggregate length 1 + 1 = 2 of the two other triangle edges—the vertical edge connecting
the vertices (0, 0) and (0, 1), and the horizontal edge connecting the vertices (0, 1) and (1, 1).
The Amato index A [25] of the population considered is defined as the normalized length of its Lorenz curve y = L(x).
Namely:

|L| − 2
A= √ . (62)
2− 2
The index A was devised by the Italian statistician Vittorio Amato in 1968 [25] as a gauge of societal egalitarianism, and was
reinvented by Nanak Kakwani in 1980 [59]. Amato’s index A takes values in the unit interval: 0 ≤ A ≤ 1. The lower bound
A = 0 indicates coincidence of the Lorenz curve y = L(x) with the line of perfect equality Lpe (x) = x, and the upper bound
A = 1 indicates coincidence of the Lorenz curve y = L(x) with the line of perfect inequality Lpi (x) = 1.
As an illustrative example consider the oligarchic Lorenz curve L(x; f ) given by Eq. (24). A simple calculation implies
that the corresponding Amato index is given by

( 1 + f 2 + (1 − f )) − 2

A(f ) = √ , (63)
2− 2
where f is the ‘‘oligarchy fraction’’. The Amato index A(f ) is monotone decreasing with respect to the oligarchy fraction f .
As the oligarchy tends to monarchy (f → 0) the Amato index A(f ) converges to its upper bound (A(f ) → 1), and as the
oligarchy tends to communism (f → 1) the Amato index A(f ) converges to its lower bound (A(f ) → 0).
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1339

Let us now follow the probabilistic approach described in Section 2.2. Our aim is to obtain a formula for Amato’s index in
terms of the normalized random variable Ŵ = W /⟨W ⟩ — which represents the normalized wealth of a randomly-sampled
member of the population. Using Eq. (14), and thereafter the change of variables w = Ψ −1 (x), implies that the length of the
Lorenz curve y = L(x) is given by
∫ 1 
|L| = 1 + (L′ (x))2 du
0
 
Ψ −1 (x)
2
w
∫ 1
 ∫ ∞  2
= 1+ dx = 1+ ψ(w)dw
0 ⟨W ⟩ 0 ⟨W ⟩
  
 2
W
= 1+ . (64)
⟨W ⟩
Finally, using normalization and substituting Eq. (64) into Eq. (62) we conclude that:
 √
⟨ 1 + Ŵ 2 ⟩ − 2
A= √ . (65)
2− 2
Eq. (65) is an explicit stochastic formula for Amato’s index A — representing it in terms of the normalized random variable
Ŵ . Amato’s index – from an analytic perspective – is considerably less tractable than the Gini and Pietra indices described
in Section 3, as well as from the Rényi indices to be established in Section 6.

5.2. The curvature index

The curvature of a general curve y = K (x) (0 ≤ x ≤ 1) is given by the formula:

|K ′′ (x)|
κ(x) = (66)
(1 + (K ′ (x))2 )3/2
(0 ≤ x ≤ 1) [60]. Let y = L(x) be the Lorenz curve of a human population under consideration. Applying Eq. (66) to the
Lorenz curve y = L(x) implies that the curve’s average curvature is given by the following ‘‘curvature index’’ :
1
|L′′ (x)|

C= dx. (67)
0 (1 + (L′ (x))2 )3/2
Let us now follow the probabilistic approach described in Section 2.2. Our aim is to obtain a formula for the curvature
index in terms of the normalized random variable Ŵ = W /⟨W ⟩ — which represents the normalized wealth of a randomly-
sampled member of the population. Applying the change of variables u = L′ (x) to the right-hand-side of Eq. (67), and using
the concavity of the Lorenz curve (L′′ (x) ≤ 0), implies that
∫ L′ (0)
1
C= du. (68)
L′ (1) (1 + u2 )3/2
On the other hand Eq. (14) implies that: (i) L′ (1) = Ψ −1 (1)/⟨W ⟩ = wmin /⟨W ⟩, where wmin is the lower bound of the
random variable W ; (ii) L′ (0) = Ψ −1 (0)/⟨W ⟩ = w max /⟨W ⟩, where wmax is the upper bound of the random variable W .
The bounds satisfy 0 ≤ wmin ≤ wmax ≤ ∞. Substituting the limits wmin /⟨W ⟩ and wmax /⟨W ⟩ into the integral appearing on
the right-hand-side of Eq. (68), and thereafter calculating this integral, yields

1 1
C=   2 −   2 . (69)
⟨W ⟩ ⟨W ⟩
1+ wmax
1+ wmin

Finally, using normalization we conclude that:

1 1
C=   2 −   2 , (70)
1 1
1+ ŵmax
1+ ŵmin

where ŵmin = wmin /⟨W ⟩ and ŵmax = wmax /⟨W ⟩ are, respectively, the lower and upper bounds of the normalized random
variable Ŵ = W /⟨W ⟩. Eq. (70) is an explicit stochastic formula for the curvature index C —representing it in terms of the
normalized random variable Ŵ .
1340 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

2 −3/2
Noting that the function ∞φ(u) = (1 + u ) (u ≥ 0) appearing on the right-hand-side of Eq. (68) is a probability
density function—i.e., that 0 φ(u)du = 1—implies that the curvature index C takes values in the unit interval: 0 ≤ C ≤ 1.
The lower bound C = 0 is attained if and only if ŵmin = ŵmax —i.e., if and only if the normalized random variable Ŵ is
degenerate: Ŵ = ⟨Ŵ ⟩ = 1. A degenerate normalized random variable Ŵ corresponds to the societal scenario of pure
communism, and hence the lower bound C = 0 indicates coincidence of the Lorenz curve y = L(x) with the line of perfect
equality Lpe (x) = x. On the other hand, the upper bound C = 1 is attained if and only if ŵmin = 0 and ŵmax = ∞—i.e., if and
only if the normalized random variable Ŵ ranges over the entire positive half-line (0, ∞). Hence, the upper bound C = 1
does not indicate coincidence of the Lorenz curve y = L(x) with the line of perfect inequality Lpi (x) = 1. Rather, the upper
bound C = 1 merely indicates that the wealth W of a randomly-sampled member of the population can admit all positive
values: 0 < W < ∞.
The curvature index C displays a markedly different behavior than all other evenness gauges discussed in this Paper—
the Gini, Pietra and Amato indices discussed so far, as well as the Rényi indices to be established in Section 6. On the one
hand, the curvature index C is easily computable: its calculation requires only the lower and upper bounds of the normalized
random variable Ŵ . On the other hand, the curvature index C is useless in cases where wealth values span the entire positive
half-line: 0 < W < ∞.

6. Evenness and randomness

So far we discussed evenness gauges based on the Lorenz curve y = L(x). In this section we tackle the measurement of
evenness via a fundamentally different approach: the measurement of randomness. The section is organized as follows. We
begin with a short exposition regarding the measurement of randomness (Section 6.1), in which we introduce a family of
randomness gauges that underlie Rényi’s entropies. We then establish Rényi’s indices – a family of entropy-based evenness
gauges – and show how this family is related to the Lorenz curve y = L(x) (Section 6.2). Thereafter we explain how the
family of Rényi indices generalizes and connects together two economic gauges of societal egalitarianism: Hirschman’s index
and Theil’s index (Section 6.3).

6.1. The measurement of randomness

How random is a given probability law? In this subsection we give a brief exposition to the measurement of the randomness
of probability laws — establishing a family of randomness gauges that underlie Rényi’s entropies [61].
Consider a general probability law Π = {πi }ni=1 . Namely, the probability law Π assigns the probabilities {πi }ni=1 to
n different outcomes labeled i = 1, . . . , n. Let Sα = {X1 , . . . , Xα } be a random sample consisting of α (α = 2, 3, . . .)
independent random variables drawn from the probability law Π . The random sample Sα appears deterministic if and only
if all random variables {X1 , . . . , Xα } happen to yield the same outcome — an event occurring with probability
n
(πi )α .

Pr(X1 = · · · = Xα ) = (71)
i =1

The probability Pr(X1 = · · · = Xα ) is a quantitative measure of the ‘‘randomness’’ of the probability law Π — the larger it is
the more ‘‘deterministic’’ the law, and the smaller it is the more ‘‘random’’ the law.
Let us turn now to the case where Π is a probability law on the real line, governed by an arbitrary probability density
function φ(x) (−∞ < x < ∞). In this continuous case the probability Pr(X1 = · · · = Xα ) trivially equals zero, and the
non-trivial counterpart of the Eq. (71) is given by the limiting probability
∫ ∞
1
lim Pr(max(Sα ) − min(Sα ) ≤ δ) = φ(x)α dx (72)
δ→0 δ α−1 −∞

(the term max(Sα ) − min(Sα ) appearing on the left-hand-side of Eq. (72) is the range of the random sample Sα ). It is
straightforward to note that if the scale of the random variables {X1 , . . . , Xα } is changed by the positive factor s (namely:
Xj → sXj , j = 1, . . . , α ) then the right-hand-side of Eq. (72) grows non-linearly by the positive factor s1−α . Hence, if we
wish to turn the right-hand-side of Eq. (72) into a gauge of randomness which responds linearly to changes of scale then we
need to raise it by the power 1/(1 − α).
Returning back to the probability law Π = {πi }ni=1 , and raising the probability Pr(X1 = · · · = Xα ) by the power 1/(1 −α),
we arrive at the following gauge of randomness:
 1
 1−α
n
(πi )α

Rα (Π ) = . (73)
i =1

The special case α = 2 – corresponding to a sample of size 2 – yields Simpson’s index [62] which is a measure of biodiversity
commonly applied in population biology and ecology. Namely, Simpson’s index is the reciprocal of the probability that
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1341

two independent random draws from a given probability distribution Π coincide: R2 (Π ) = 1/ Pr(X1 = X2 ). In sociology
Simpson’s index is often referred to as ‘‘Blau’s index’’ [63]. In physics Simpson’s index is often referred to as the ‘‘inverse
participation ratio’’, and examples of its application include local vibrations in disordered systems [64,65] and localization of
quantum states [66]. The gauge of randomness Rα (Π ) is termed in physics the ‘‘generalized inverse participation ratio ’’ [67].
The right-hand-side of Eq. (73) is well-defined for all positive exponents α — albeit for α = 1. Nonetheless, taking the
limit α → 1 in Eq. (73) yields the limit
 
n

R1 (Π ) = exp − πi ln(πi ) . (74)
i=1

Thus, Eq. (73)–(74) establish a whole family of randomness gauges {Rα (Π )}α>0 .
The randomness gauges {Rα (Π )}α>0 have a special multiplicative feature: If Π1 , . . . , Πm are independent probability
laws, then the randomness level of the product probability law Π1 × · · · × Πm equals the product of the randomness levels
of the independent laws:
Rα (Π1 × · · · × Πm ) = Rα (Π1 ) · · · Rα (Πm ). (75)
In other words, if m independent random experiments are conducted simultaneously then the randomness level of the
‘‘meta experiment’’ (Π1 × · · · × Πm ) equals the product of the randomness levels of the single experiments (Π1 , . . . , Πm ).
The Rényi entropies are defined as the logarithms of the randomness gauges {Rα (Π )}α>0 [61]. Namely, ln(Rα (Π ))
is the Rényi entropy corresponding to the randomness gauge Rα (Π ), and the Rényi entropy ln(R1 (Π )) coincides with
the Boltzmann–Gibbs entropy [3]. In information theory the Boltzmann–Gibbs entropyis often referred to as ‘‘Shannon’s
entropy’’ [68]. The connection between Simpson’s index and the Boltzmann–Gibbs entropy is discussed in Ref. [69]. The
multiplicative feature of the randomness gauges {Rα (Π )}α>0 (characterized by Eq. (75)) translates into the following additive
feature of Rényi’s entropies: If m independent random experiments are conducted simultaneously then the entropy of the
‘‘meta experiment’’ (Π1 × · · · × Πm ) equals the sum of the entropies of the single experiments (Π1 , . . . , Πm ).

6.2. The Rényi indices

Consider the census setting of Section 2.1: A population consisting of n members labeled by the index i = 1, . . . , n, where
wi denotes the wealth of member i, and where w = (w1 + · · · + wn )/n denotes the average population wealth.
The population’s wealth values {wi }ni=1 induce a probability law Π = {πi }ni=1 given by

wi 1 wi
πi = = . (76)
w1 + · · · + wn n w
Namely, the probability πi associated with member i is its ‘‘share’’ of the total population’s wealth. A natural way to measure
the evenness of the population’s wealth values {wi }ni=1 is via the randomness of the probability law Π : the more ‘‘random’’ the
law – the more unequal the population’s distribution of wealth; the more ‘‘deterministic’’ the law – the more equal the
population’s distribution of wealth. We note that the distribution of the probability π1 – in the case where the population’s
wealth values {wi }ni=1 are i.i.d. random variables – is studied in detail in Ref. [70]; see also Ref. [71] for the special case n = 2
– which turns out to yield surprising results.
Following this randomness-based approach, we now turn to compute the randomness gauge Rα (Π ) (given by Eqs. (73)
and (74)) of the wealth-induced probability law Π (given by Eq. (76)). Assume, with no loss of generality, that the labeling of
the population members is non-increasing in their wealth: w1 ≥ w2 ≥ · · · ≥ wn ≥ 0. Then, for large populations (n ≫ 1),
we obtain that:
The case α ̸= 1:
1
 α  1−α
1 wi
n 

Rα ( Π ) =
i =1
n w
1 1

n   α  1−α ∫ 1
 1−α
− i 1 α
=n 1L ≈n (L (u)) du

i =1
n n 0

α 1 α 1
1
Ψ − 1 ( u) ∞
w
∫   1−α ∫   1−α
=n du =n ψ(w)dw
0 ⟨W ⟩ 0 ⟨W ⟩
 α  1−α1
W
=n (77)
⟨W ⟩
(in the transition from the first line to the second line we used Eq. (5); in the transition from the second line to the third line
we used Eq. (14); in the third line we applied the change of variables w = Ψ −1 (u)).
1342 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

The case α = 1:
 
1 wi 1 wi
n 
−  
R1 (Π ) = exp − ln
i=1
n w n w
   
n   
− i i 1
= n exp − 1L ln 1L
i=1
n n n
 ∫ 1 
≈ n exp − (L′ (u)) ln(L′ (u))du
0

Ψ (u) Ψ (u)
 ∫ 1  −1   −1  
= n exp − ln du
0 ⟨W ⟩ ⟨W ⟩
w w
 ∫ ∞    
= n exp − ln ψ(w)dw
0 ⟨W ⟩ ⟨W ⟩
    
W W
= n exp − ln (78)
⟨W ⟩ ⟨W ⟩
(in the transition from the first line to the second line we used Eq. (5); in the transition from the third line to the fourth line
we used Eq. (14); in the transition from the fourth line to the fifth line we applied the change of variables w = Ψ −1 (u)).
Based on Eqs. (77) and (78) we introduce a family of entropy-based evenness gauges {Rα }α>0 which we term ‘‘Rényi
indices ’’ [35]. In terms of the Lorenz curve y = L(x) of the population under consideration, the Rényi indices {Rα }α>0 are
given by:
1
∫
1
 1−α
α

(L (u)) du

α ̸= 1,



1 − Rα = 0
 ∫ 1  (79)

(L (u)) ln(L (u))du
′ ′
α = 1.

exp −

0

On the other hand, in terms of the normalized random variable Ŵ = W /⟨W ⟩ – which represents the normalized wealth of
a randomly-sampled member of the population – the Rényi indices {Rα }α>0 are given by:
 1
⟨Ŵ α ⟩ 1−α α ̸= 1,
1 − Rα = (80)
exp(−⟨Ŵ ln(Ŵ )⟩) α = 1.
The Rényi indices {Rα }α>0 take values in the unit interval: 0 ≤ Rα ≤ 1 (α > 0). This assertion follows straightforwardly
from Jensen’s inequality which asserts that if φ(·) is a convex function then: φ(⟨Ŵ ⟩) ≤ ⟨φ(Ŵ )⟩ [72]. Indeed: (i) for 0 < α < 1
1
set φ(x) = −xα to obtain that ⟨Ŵ α ⟩ 1−α ≤ 1; (ii) for α = 1 set φ(x) = x ln(x) to obtain that exp(−⟨Ŵ ln(Ŵ )⟩) ≤ 1; (iii) for
1
α > 1 set φ(x) = xα to obtain that ⟨Ŵ α ⟩ 1−α ≤ 1.
As an example consider the oligarchic Lorenz curve L(x; f ) given by Eq. (24). A simple calculation implies that the
corresponding Rényi indices are given by Rα (f ) = 1 − f , where f is the ‘‘oligarchy fraction’’. The Rényi indices Rα (f )
are monotone decreasing with respect to the oligarchy fraction f . As the oligarchy turns into a monarchy (f → 0) the R ényi
indices Rα (f ) converge to their upper bound (Rα (f ) → 1), and as the oligarchy turns into communism (f → 1) the Rényi
indices Rα (f ) converge to their lower bound (Rα (f ) → 0).
Examples of various classes of probability laws and their corresponding R ényi indices are given in Table 4. We note that
the family of Rényi indices {Rα }α>0 is somewhat similar to the family of Atkinson indices {Aε }0≤ε≤1 [28] whose stochastic
representation – in the context of large populations (n ≫ 1), and in terms of the normalized random variable Ŵ = W /⟨W ⟩
– is given by:
 1
⟨Ŵ 1−ε ⟩ 1−ε ε ̸= 1,
1 − Aε = (81)
exp(⟨ln(Ŵ )⟩) ε = 1.

6.3. The Hirschman and Theil indices

Consider the census setting of Section 2.1: A population consisting of n members labeled by the index i = 1, . . . , n,
where wi denotes the wealth of member i, and where w = (w1 + · · · + wn )/n denotes the average population wealth. The
Hirschman index [26] and the Theil index [27] of the population are defined, respectively, by

1 −  wk 2
n
H= (82)
n k=1 w
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1343

Fig. 4. Rényi indices. The Rényi indices, corresponding to the Pareto I class of probability laws (see Table 4), are depicted as functions of the parameter
p (represented by the x-axis). The solid line is the graph of the Rényi indices with exponent α = 1 (the Theil-index counterpart), and the dashed line is
the graph of the Rényi indices with exponent α = 2 (the Hirschman-index counterpart). The dashed line should be considered only for parameter values
p > 1.

and

1 −  wk  w 
n
k
T = ln . (83)
n k=1 w w
Hirschman’s index H was devised by the German-born economist Albert Otto Hirschman in 1945 [26]. Hirschman’s index
was reinvented in 1950 by the economist Orris Herfindahl, and is commonly – yet mistakenly – referred to as the ‘‘Herfindahl
index’’ or the ‘‘Herfindahl–Hirschman index’’ [73]. Hirschman’s index is widely applied in economics and in competition law
as a quantitative measure of concentration. Theil’s index T was devised by the Dutch economist Henri Theil in 1967 [27].
Theil’s index is a quantitative measure of economic inequality which is based on an information-theory approach — as is
evident from the resemblance of Theil’s index to the Boltzmann–Gibbs entropy.
For a large population (n ≫ 1) it is straightforward to note that Eqs. (82) and (83) yield, respectively, the stochastic
approximations

H ≈ ⟨Ŵ 2 ⟩ (84)

and

T ≈ ⟨Ŵ ln(Ŵ )⟩, (85)

where the random variable Ŵ = W /⟨W ⟩ represents the normalized wealth of a randomly-sampled member of the
population.
Comparing Eq. (84)–(85) to Eq. (80) reveals the connection of the Hirschman and Theil indices to the Rényi indices
{Rα }α>0 established in Section 6.2. Indeed, considering the Hirschman and Theil indices to be given, respectively, by the
right-hand-sides of Eq. (84)–(85), we obtain that: H = 1/(1 − R2 ) and T = − ln(1 − R1 ). Examples of various classes
of probability laws and their corresponding R1 Rényi indices (the Theil-index counterpart) and R2 Rényi indices (the
Hirschman-index counterpart) are given in Table 4; graphical illustrations of these Rényi indices are given in Figs. 4 and 5.
The family of Rényi indices {Rα }α>0 provides a unified theoretical framework which both generalizes and connects
together Hirschman’s index H and Theil’s index T . The advantage of the Rényi indices {Rα }α>0 – over Hirschman’s index
H and Theil’s index T – is that they establish a family of continuous and consistent gauges of statistical evenness, which
are all calibrated to the unit interval (0, 1).

7. Pareto’s 20–80 rule and probability law

The Italian economist Vilfredo Pareto revolutionized the social sciences with his studies regarding the distribution of
wealth in human populations. In this section we show how two of Parteo’s discoveries are intimately related to the Lorenz
curve and to the measurement of societal egalitarianism: Pareto’s 20–80 rule (Section 7.1), and Pareto’s probability law
(Section 7.2).
1344 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

Fig. 5. Rényi indices. The Rényi indices, corresponding to the Beta I class of probability laws (see Table 4), are depicted as functions of the parameter p
(represented by the x-axis). The solid line is the graph of the Rényi indices with exponent α = 1 (the Theil-index counterpart), and the dashed line is the
graph of the Rényi indices with exponent α = 2 (the Hirschman-index counterpart).

7.1. Pareto’s 20–80 rule

In their best-seller Linked, Albert–Laszlo Barabási write ([74], page 66): ‘‘Outside academia Pareto is best known for one
of his empirical observations. . . A careful observer of economic inequalities, he (Pareto) saw that 80% of Italy’s land was owned
by only 20% of the population. More recently, Pareto’s Law or Principle, known also as the 20–80 rule, has been turned into a
Murphy’s Law of management: 80% of customer service problems are created by only 20% of consumers, 80% of decisions are made
during 20% of meeting time, and so on’’. In terms of the Lorenz curve Pareto’s rule is given by 0.8 = L(0.2).
In this subsection we consider as given a specific ‘‘Pareto-type’’ information
f2 = L(f1 ) (86)
(0 < f1 < f2 < 1). From this given information we shall show how to derive either the entire Lorenz curve y = L(x), or
bounds on the corresponding evenness gauges.

7.1.1. Lorenz-curve reconstruction


In general it is impossible to reconstruct the entire Lorenz curve y = L(x) from the information that the Lorenz curve
passes through the specific point (f1 , f2 ). Yet so, in some particular cases such a reconstruction is possible. For example, if
the probability law which governs the population’s distribution-of-wealth is characterized by a single parameter p, and if
this parameter can be explicitly extracted from the information f2 = L(f1 ) — then a Lorenz-curve reconstruction is indeed
possible. We present three examples of such reconstruction:
Paretian distribution of wealth. Assume that the population’s distribution of wealth is governed by the Pareto I probability
law of Table 1. Then, the information f2 = L(f1 ) implies that
ln(f2 )
p= . (87)
ln(f1 ) − ln(f2 )
In particular, in the case of Pareto’s 20–80 rule we obtain that p ≃ 0.16 — yielding, in turn, the Gini index G ≃ 0.76 and the
Pietra index G ≃ 0.63.
Beta distribution of wealth. Assume that the population’s distribution of wealth is governed by the Beta I probability law of
Table 1. Then, the information f2 = L(f1 ) implies that
ln(1 − f1 )
p= . (88)
ln(1 − f2 ) − ln(1 − f1 )
In particular, in the case of Pareto’s 20–80 rule we obtain that p ≃ 0.16 — yielding, in turn, the Gini index G ≃ 0.76 and the
Pietra index G ≃ 0.63.
Log-normal distribution of wealth. Assume that the population’s distribution of wealth is governed by the Log-
Normal probability law of Table 1. Then, the information f2 = L(f1 ) implies that3

p = Φ −1 (f1 ) − Φ −1 (f2 ). (89)

3 In Eq. (89) the function Φ (·) denotes the tail probability of the standard Normal law—see Eq. (1).
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1345

Fig. 6. Pareto’s 20–80 rule. The Lorenz curves, corresponding to the Pareto I and Beta I probability laws (see Table 2) which satisfy Pareto’s 20–80 rule, are
depicted. The solid line is the graph of the Pareto I Lorenz curve which passes through the point (0.2, 0.8), and the dashed line is the graph of the Beta I
Lorenz curve which passes through the point (0.2, 0.8).

In particular, in the case of Pareto’s 20–80 rule we obtain that p ≃ 1.68 — yielding, in turn, the Gini index G ≃ 0.77 and the
Pietra index G ≃ 0.59.
Note that in the case of Pareto’s 20–80 rule both the Pareto I and Beta I probability laws yield the same parameter
p ≃ 0.16. The coincidence of the inferred parameter p of the Pareto I and Beta I probability laws holds, in general, whenever
the specific point (f1 , f2 ) satisfies the condition f1 + f2 = 1. Graphical illustrations of the Pareto I and Beta I Lorenz curves
satisfying Pareto’s 20–80 rule are given in Fig. 6.
We emphasize that while the three aforementioned probability laws can all be set to follow Pareto’s 20–80 rule, the
resulting population statistics can be markedly different. Indeed, calculating the variance Var (Ŵ ) of the corresponding
normalized random variable Ŵ we obtain that: (i) Var (Ŵ ) = ∞ — in the case of the Pareto I probability law with the
‘‘20–80 parameter’’ p ≃ 0.16; (ii) Var (Ŵ ) ≃ 2.89 — in the case of the Beta I probability law with the ‘‘20–80 parameter’’ p ≃
0.16; (iii) Var (Ŵ ) ≃ 15.84 — in the case of the Log-Normal probability law with the ‘‘20–80 parameter’’ p ≃ 1.68.

7.1.2. Evenness bounds


As discussed above, the information that the Lorenz curve y = L(x) passes through the specific point (f1 , f2 ) is, in general,
insufficient in order to reconstruct the entire Lorenz curve. Yet, since the Lorenz curve y = L(x) is monotone non-decreasing
and concave, the following pair of chords lie beneath the Lorenz curve: (i) the chord connecting the points (0, 0) and (f1 , f2 );
(ii) the chord connecting the points (f1 , f2 ) and (1, 1). However, the union of these two chords constitutes the Lorenz curve
L(x; f1 , f2 ) of the Oligarchy Model II (Eq. (26)). Consequently, we obtain that the Lorenz curve y = L(x) is bounded from
below by the Lorenz curve y = L(x; f1 , f2 ). Namely:
L(x) ≥ L(x; f1 , f2 ). (90)
Eq. (90), in turn, implies the following evenness bounds.
Gini’s index. The area captured between the Lorenz curve y = L(x) and the line of perfect equality Lpe (x) is bounded from
below by the area captured between the Lorenz curve y = L(x; f1 , f2 ) and the line of perfect equality Lpe (x). Consequently,
the Gini index G is bounded from below by the Gini index G(f1 , f2 ). Using Table 6 we conclude that:
G ≥ f2 − f1 . (91)

Pietra’s index. The maximal vertical gap between the Lorenz curve y = L(x) and the line of perfect equality Lpe (x) is
bounded from below by the vertical gap at the point x = f1 — which is the maximal vertical gap between the Lorenz
curve y = L(x; f1 , f2 ) and the line of perfect equality Lpe (x). Consequently, the Pietra index P is bounded from below by
the Pietra index P (f1 , f2 ). Using Table 6 we conclude that:
P ≥ f2 − f1 . (92)
Amato’s index. Due to the concavity of the Lorenz curves, the length of the Lorenz curve y = L(x) is bounded from below
by the length of the Lorenz curve y = L(x; f1 , f2 ). Consequently, the Amato index A is bounded from below by the Amato
index A(f1 , f2 ). Using Table 6 we conclude that:

(f1 )2 + (f2 )2 + (1 − f1 )2 + (1 − f2 )2 − 2
 
A≥ √ . (93)
2− 2
1346 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

7.2. Pareto’s probability law

In his studies Pareto came up with a dramatic empirical discovery regarding the distribution of wealth in human
populations: the frequency of individuals with wealth greater than the level l follows, asymptotically, a decreasing power-
law in the variable l [75]. The striking feature of Pareto’s discovery was its empirical universality — all human populations
investigated by Pareto appeared to be asymptotically governed by power-law wealth statistics (albeit with different
exponents). Named after Pareto, probability laws governed by power-law tails are nowadays referred to as ‘‘Paretian’’. In
this subsection we connect together Pareto’s probability law, power-law Lorenz curves, and fractality.

7.2.1. Pareto’s probability law and power-law Lorenz curves


Consider a human population whose distribution of wealth is governed by the Paretian tail probability
v 1+p
min
Ψ (v) = (94)
v
(v ≥ vmin ), where p is a positive parameter,4 and where vmin is a positive lower-bound value. The ‘‘Pareto I’’ example of
Table 2 (supplemented by the scale-invariance of Lorenz curves) implies that the corresponding Lorenz curve is given by
the power-law
p
L(x) = x 1+p (95)
(0 ≤ x ≤ 1). Conversely, substituting the power-law Lorenz curve of Eq. (95) into the inversion formula of Eq. (23) yields
the normalized Paretian tail probability

p/(1 + p)
 1+p
Ψ̂ (v) = (96)
v
(v ≥ p/(1 + p)).
Hence, the class of Paretian probability laws is equivalent to the class of power-law Lorenz curves. We now turn to show
that the class of power-law Lorenz curves is characterized by a unique ‘‘fractal’’ feature.

7.2.2. Lorenzian fractality and power-law Lorenz curves


Assume that we are constructing the Lorenz curve of a given human population — yet that we do so using a lower cutoff
level l (l > 0). Namely, rather than considering the entire population, we consider only the sub-population of people who
are richer than the level l. Thus, each cutoff level l induces a Lorenz curve Ll (x) that quantifies the distribution of wealth
within the ‘‘richer-than-l ’’ sub-population. Our goal in this section is to characterize the class of ‘‘Lorenz fractal ’’ populations:
populations whose Lorenz curves Ll (x) are invariant with respect to the cutoff level l applied.
Let Wl denote the wealth of a randomly-sampled member of the ‘‘ richer-than-l’’ sub-population. The probability tail of
the random variable Wl is given by
Ψ (v)
Ψl (v) = Pr (Wl > v) = Pr (W > v | W > l) = (97)
Ψ (l)
(v > l). Eq. (97) implies that the mean ⟨Wl ⟩ of the random variable Wl is given by

Ψ (v) Ψ (ls)
∫ l ∫  ∫ ∞ 
⟨Wl ⟩ = dv + dv = l 1 + ds (98)
0 l Ψ (l) 1 Ψ (l)
(the right-hand-side of Eq. (98) follows from the middle part of Eq. (98) via the change of variables v = ls). On the other
hand, Eq. (97) implies that the tail probability of the normalized random variable Ŵl = Wl /⟨Wl ⟩ is given by
Ψ (⟨Wl ⟩v)
Ψ̂l (v) = Pr(Ŵl > v) = (99)
Ψ (l)
(v ≥ l/⟨Wl ⟩). Setting

Ψ (ls)

ψ(l) = ds (100)
1 Ψ (l)
(l > 0), and substituting Eq. (98) into Eq. (99) we conclude that:
Ψ (l(1 + ψ(l))v)
Ψ̂l (v) = (101)
Ψ (l)
(v ≥ 1/(1 + ψ(l))).

4 The positivity of the parameter p is required in order to ensure that the probability law governed by the tail probability Ψ (v) possesses a finite mean.
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1347

Due to Eq. (22) the Lorenz curve Ll (x) is l-invariant if and only if the tail probability Ψ̂l (v) is l-invariant. A necessary
condition for the tail probability Ψ̂l (v) to be l-invariant is that its range be l-invariant – which, in turn, holds if and only if
the function ψ(l) is l-invariant. From Eq. (100) it is evident that the function ψ(l) is l-invariant if and only if the function
Ψ (v) is a power-law. Since the function Ψ (v) is a tail probability then its admissible power-law form is the Paretian form of
Eq. (94) — implying, in turn, that the corresponding Lorenz curve admits the power-law form of Eq. (95). Hence, we conclude
that: The class of ‘‘Lorenz fractal’’ populations coincides with the class of populations characterized by power-law Lorenz curves.
This result provides a Lorenzian perspective to the ‘‘fractality’’ of the class of Paretian probability laws. For a detailed study
of ‘‘fractality’’ in the context of probability laws defined on the positive half-line the readers are referred to Refs. [76,77].

8. A Lorenzian limit law for rank distributions

In Section 7 we explored the close connection between Pareto’s probability law and power-law Lorenz curves. In this
section we show – using a ‘‘Lorenzian analysis’’ – that Pareto’s probability law is the universal stochastic limit law emerging
from rank distributions. We begin with a short description of rank distributions (Section 8.1), establish a ‘‘Lorenzian limit
law’’ for rank distributions (Section 8.2), and obtain from this limit law a universal macroscopic statistical classification of
network topologies (Section 8.3).

8.1. Rank distributions and their Lorenz curves

Rank distributions are objects we encounter ubiquitously and on a daily basis, and are often visualized in the form of ‘‘pie
charts’’. The abstract description of rank distributions is as follows. Consider an ensemble consisting of n different elements.
Each element has a size, and the elements are ranked in a non-increasing order of their sizes r = 1, 2, . . . , n. Namely, the
element ranked r = 1 is the largest one, the element ranked r = 2 is the second largest, and so forth. Let S (r ) denote
the size of the element with rank r, and let A(r ) = S (1) + · · · + S (r ) denote the aggregate size of the r largest elements
(r = 1, . . . , n). The ensemble’s rank distribution is given by

S (r )
F (r ) = (102)
A(n)
(r = 1, . . . , n). Namely, F (r ) is the relative size of the rth largest element—measured with respect to the aggregate size of
the entire ensemble.
The following five examples illustrate the ubiquity of rank distributions: (i) Demography: the elements are cities
composing a given state, and F (r ) is the fraction of the state’s population domiciled in the rth largest city. (ii) Linguistics:
the elements are different words appearing in a given text, and F (r ) is the occurrence-frequency of the rth most common
word in the text. (iii) Economics: the elements are firms competing in a given market, and F (r ) is the market-share of the
rth largest firm. (iv) Politics: the elements are political parties in a given democracy, and F (r ) is the fraction of parliament-
seats held by the rth largest party. (v) Sports: the elements are nations taking part in a given Olympiad, and F (r ) is the
fraction of Olympic medals won by the rth champion nation.
The Lorenz curve corresponding to the rank distribution of Eq. (102) is constructed as follows. The aggregate size of the
r largest ensemble elements is A(r ), and the aggregate size of the entire ensemble is A(n). Hence, if we deem the size of the
elements to represent their ‘‘wealth’’ then the ensemble’s Lorenz curve is given by
r  A(r )
L = (103)
n A(n)

at the rational points { 0n , 1n , . . . , nn } (here and hereinafter we set A(0) = 0). In between the rational points the Lorenz curve
is interpolated—the common interpolation being linear.

8.2. The Lorenzian limit law

Consider now vast ensembles which we model via the limit n → ∞. In what follows we assume that the aggregate
function A(r ) (r = 0, 1, . . . , n) emanates from a monotone non-decreasing function defined on the non-negative half-line:
A(x) (x ≥ 0). The natural interpolation of the Lorenz curve of Eq. (103) is thus given by L(x) = A(nx)/A(n). Consequently,
in the limit n → ∞ we obtain the limiting Lorenz curve

A(nx)
L(x) = lim (104)
n→∞ A(n)
(0 < x ≤ 1). Eq. (104) leads to the notion of regular variation [78].
A non-negative valued function φ(x) (x ≥ 0) is termed regularly varying if the limit ψ(x) = limn→∞ φ(nx)/φ(n) exists
for all x > 0. It is straightforward to note that if φ(x) is regularly varying then ψ(xy) = ψ(x)ψ(y) holds for all x, y > 0. This
observation, in turn, implies that the limit ψ(x) is a power-law: ψ(x) = xϵ , where ϵ is the corresponding real-valued
1348 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

‘‘exponent of regular variation’’. Regularly varying functions with exponent ϵ = 0 are termed slowly varying. The class
of slowly varying functions includes asymptotically constant functions, logarithms, powers of slowly varying functions,
and logarithms of slowly varying functions. A regularly varying function φ(x) with exponent ϵ admits the representation
φ(x) = φ0 (x)xϵ where φ0 (x) is a slowly varying function. The notion of regular variation plays a key role in Probability
Theory, as it arises naturally in the context of stochastic limit laws [78].
Eq. (104) thus implies the following Lorenzian Limit Law (LLL) for rank distributions [38]: A vast ensemble admits a Lorenz
curve if and only if the function A(x) is regularly varying — in which the limiting Lorenz curve is given by the power-law

L(x) = xϵ (105)

(0 < x ≤ 1). Moreover, since Lorenz curves are monotone non-decreasing and concave, the regular-variation exponent ϵ is
confined to the range 0 ≤ ϵ ≤ 1. Hence, the LLL establishes that in the case of vast ensembles there are only three possible
rank-distribution scenarios:
• Monarchic Scenario — characterized by the exponent value ϵ = 0. This scenario yields the line of perfect inequality
L(x) = 1 = Lpi (x) which, in turn, represents a ‘‘monarchic’’ vast ensemble.
• Communist Scenario — characterized by the exponent value ϵ = 1. This scenario yields the line of perfect equality
L(x) = x = Lpe (x) which, in turn, represents a ‘‘communist’’ vast ensemble.
• Paretian Scenario— characterized by the intermediate exponent values 0 < ϵ < 1. This scenario yields the class of Lorenz
curves which represent – following the first part of Section 7.2 – Paretian vast ensembles governed by the Paretian tail
probability
v ϵ
 1−ϵ
min
Ψ (v) = (106)
v
(v ≥ vmin ).
Note that the Monarchic Scenario and the Communist Scenario are deterministic limit laws, whereas the Paretian Scenario
is a stochastic limit law.

8.3. Network topologies

Networks are fundamental in an overwhelmingly wide span of scientific fields, and the single most important element of
a network is its topology [42,43]. The LLL established in Section 8.2 provides a universal macroscopic statistical classification
of network topologies. Given a network set its connected-components to be the ensemble elements, and set the size of each
element to be its number of nodes. In this network setting the LLL is interpreted as follows:

• The Monarchic Scenario represents a network which, on the macroscopic scale, is totally connected: one giant connected-
component dominates the entire network. This scenario is characterized by the Lorenzian exponent ϵ = 0.
• The Communist Scenario represents a network which, on the macroscopic scale, is totally disconnected: the network is
fragmented into a ‘‘uniform dust’’. This scenario is characterized by the Lorenzian exponent ϵ = 1.
• The Paretian Scenario represents a ‘‘fractal’’ phase-transition between connectedness and disconnectedness: connected-
components of all scales emerge, and their sizes are governed by power-law statistics. This scenario is parametrized by
the Lorenzian exponent 0 < ϵ < 1.
In this network-conceptualization the Lorenzian exponent ϵ quantifies the macroscopic degree-of-connectedness of large
networks: ϵ = 0 representing connectedness, ϵ = 1 representing disconnectedness, and 0 < ϵ < 1 representing
intermediate levels of macroscopic connectivity. In the network-conceptualization we set the size of an element to be its
number of nodes. However, other size-measures can be considered as well. For example, the size of an element can be
its number of links, or – in the context of physical networks – a physical measure such as area, volume, mass, etc. Note
that whatever the size-measure considered — the LLL yields the same macroscopic statistical result. Thus, the emergence
of power-law statistics at the transition between connectedness and disconnectedness is not only independent of the
underlying network topology — it is also independent of the specific size-measure applied. The application of the LLL to
network topologies is in agreement with the physical theories of phase transitions — which predict the emergence of power-
law statistics at the critical phase-transition thresholds [79].

9. Poissonian population models

In Section 2 we followed two different approaches to model the wealth of human populations: census and probabilistic.
The census approach was based on ‘‘census data’’ — the sequence {wi }ni=1 representing the precise wealth values of all
members of the population under consideration. On the other hand, the probabilistic approach was based on ‘‘sample data’’ —
the random variable W representing the wealth of a randomly-sampled member of the population under consideration. In
this section we explore yet another modeling approach: Poisson processes [48].
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1349

Poisson processes are the most commonly applied statistical model for the random scattering of points in general
domains, with applications ranging from queuing systems [49] to insurance and finance [80]. In the context of wealth-
modeling, a human population is a collection of points scattered along the positive half-line — each point representing the
wealth of a single population member.
A Poisson process X on the positive half-line, with intensity λ(x) (x > 0), is governed bythe following statistics [48]:
(i) the number NI of points residing in the interval I is Poisson-distributed with mean ⟨NI ⟩ = I λ(x)dx, i.e.,

⟨NI ⟩k
Pr(NI = k) = exp(−⟨NI ⟩) (107)
k!
(k = 0, 1, 2, . . .); (ii) the numbers of points residing in disjoint intervals are independent random variables.
In particular, the mean number of points of the Poisson process X which exceed the level l is given by the following ‘‘tail
intensity’’ function:
∫ ∞
Λ(l) = ⟨N(l,∞) ⟩ = λ(x)dx (108)
l

(l > 0). Henceforth, we term a Poisson process X on the positive half-line a ‘‘Poissonian population ’’ if its tail intensity
function Λ(l) is monotone decreasing from infinity (i.e., liml→0 Λ(l) = ∞) to zero (i.e., liml→∞ Λ(l) = 0). As we shall
momentarily explain, Poissonian populations are the relevant Poisson processes in the context of wealth-modeling.
The remainder of this section is organized as follows. We begin with describing the topology of Poissonian populations
(Section 9.1), thereafter we turn to explore the measurement of evenness of Poissonian populations (Section 9.2), and
conclude with a discussion (Section 9.3).

9.1. The topology of Poissonian populations

Human populations – albeit in the case of purely communist societies – always posses a ‘‘richest population member’’.
In the case of a population modeled by a Poisson process X, the wealth of the richest population member is represented
by the maximal point M = maxx∈X (x) of the Poisson process X. The maximal point M is no greater than the level l if and
only if the Poisson process X has no points residing above the level l. Hence, Eqs. (107)–(108) imply that the cumulative
distribution function of the maximal point M is given by

Pr(M ≤ l) = Pr(N(l,∞) = 0) = exp(−Λ(l)) (109)

(l > 0).
The cumulative distribution function Pr(M ≤ l) is proper if and only if it is monotone increasing from the level
liml→0 Pr(M ≤ l) = 0 to the level liml→∞ Pr(M ≤ l) = 1. Eq. (109) implies that the cumulative distribution function
Pr(M ≤ l) is proper if and only if the tail intensity function Λ(l) is monotone decreasing from the level liml→0 Λ(l) = ∞ to
the level liml→∞ Λ(l) = 0. Hence, we assert that: A Poisson process X on the positive half-line has a proper maximal point if
and only if it is a Poissonian population. This is the reason why Poissonian populations are the relevant Poisson processes in
the context of wealth-modeling.
Poissonian populations exhibit a rather surprising phenomena: they consist of infinitely many points [31]. Specifically,
a Poissonian population X can be represented by an infinite monotone decreasing sequence of random points: X1 > X2 >
X3 > · · · (the point X1 being the maximal point M). This topological structure implies that a Poissonian population X has
infinitely many points residing below any given level l, and has finitely many points residing above any given level l (l > 0).
The ‘‘existence theorem’’ of the theory of Poisson processes ([48], Section 2.5), further implies that the points residing above
the level l are i.i.d. random variables governed by the tail probability

Λ(v)
Ψl (v) = (110)
Λ(l)
(v ≥ l). Moreover, setting

Λ(ls)

ψ(l) = ds (111)
1 Λ(l)
(l > 0), and repeating the analysis conducted in Eqs. (97)–(101) – albeit replacing the function Ψ (·) by the function Λ(·) –
we obtain that the normalized tail probability (corresponding to the tail probability of Eq. (110)) is given by

Λ(l(1 + ψ(l))v)
Ψ̂l (v) = (112)
Λ(l)
(v ≥ 1/(1 + ψ(l))). With Eq. (112) at hand, we are now in position to explore the evenness of Poissonian populations.
1350 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

9.2. The evenness of Poissonian populations

Since Poissonian populations are infinite objects, their evenness cannot be measured via any of the evenness gauges
discussed above. Indeed, all the evenness gauges are based on the random variable W — which represents the wealth
of a randomly-sampled member of the population. However, in an infinite population the draw of a ‘‘random sample’’ is
impossible — for there is no uniform distribution on an infinite set. Thus, the random variable W fails to exist in the context
of Poissonian populations. On the other hand, for each cutoff level l (l > 0) there exists a random variable Wl — which
represents the wealth of a randomly-sampled member of the finite sub-population X ∩ (l, ∞). Hence, for each cutoff level
l (l > 0) all evenness gauges discussed above can be applied.
The situation for Poissonian populations is therefor as follows. Let the index I denote the evenness gauge to be applied,
and let X be the Poissonian population under consideration. Rather than having one single measure of evenness of the
population, we have an infinite family of indices {Il }l — the index Il measuring the evenness of the population members with
wealth exceeding the level l. The measurement of evenness is thus dependent on the cutoff level l applied. Consequently,
the Poissonian population X can be assigned one single measure of evenness if and only if the indices {Il }l are invariant with
respect to the cutoff level l applied: Il = const (l > 0).
Now, the indices {Il }l are l-invariant if and only if the normalized tail probability Ψ̂l (v) of Eq. (112) is l-invariant. The
analysis following Eq. (101) – albeit replacing the function Ψ (·) by the function Λ(·) – asserts that the normalized tail
probability Ψ̂l (v) of Eq. (112) is l-invariant if and only if the function Λ(·) is a power-law:
 1+p
1
Λ(l) = c (113)
l
(l > 0), where the coefficient c and the exponent p are positive parameters.5 Substituting Eq. (113) into Eq. (110) yields the
Paretian tail probability of Eq. (94) with lower bound vmin = l. Namely:
 1+p
l
Ψl (v) = (114)
v
(v ≥ l). Thus, using the notion of ‘‘Lorenzian fractality’’ introduced in Section 7.2 we conclude that: A Poissonian
population X admits one single measure of evenness if and only if it is Lorenz fractal.

9.3. Discussion: evenness vs. randomness

It is illuminating to compare – in the context of Poissonian populations – the measurement of evenness to the
measurement of randomness. To that end, let E denote a Rényi entropy to be applied, and let X be the Poissonian population
under consideration. As in the case of the measurement of evenness, the entropy cannot be calculated for the entire
Poissonian population. Rather than having one single entropy measure of the population, we have an infinite family of
entropies {El }l — the entropy El measuring the randomness of the finite sub-population X ∩ (l, ∞). Consequently, the
Poissonian population X can be assigned one single measure of randomness if and only if the entropies {El }l are invariant
with respect to the cutoff level l applied: El = const (l > 0).
An analysis given in Appendix asserts that the entropies {El }l are l-invariant if and only if the function Λ(·) is an
exponential:
Λ(l) = c exp(−pl), (115)
where the coefficient c and the exponent p are positive parameters. In turn, substituting Eq. (115) into Eq. (110) yields the
exponential tail probability
Ψl (v) = exp(−p(v − l)) (116)
(v ≥ l).
Poissonian populations thus pinpoint a striking difference between the measurement of evenness and the measurement
of randomness. In the context of Poissonian populations the measurement of evenness leads to Paretian probability laws,
whereas the measurement of randomness leads to exponential probability laws. Thus, considering ‘‘fractality’’ as invariance
with respect to the cutoff level l we can conclude that: power-law structures are fractal in the context of evenness-measurement,
and exponential structures are fractal in the context of randomness-measurement.

10. Conclusions

Physics and economics devised diametrically different approaches to measure the statistical heterogeneity of general
data-sets. Physics was interested in quantifying the randomness of large systems, and established the powerful concept

5 Indeed, substituting Eq. (112) into Eq. (111) implies that the exponent p must be positive in order to ensure integrability.
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1351

Table 5
Gauges of statistical evenness. The table summarizes all gauges of statistical evenness discussed: the Gini,
Pietra, Amato, curvature, and Rényi indices. In the second column the indices are represented in terms of
the population’s Lorenz curve y = L(x). In the third column the indices are represented in terms of the
normalized random variable Ŵ — the normalized wealth of a randomly-sampled member of the population.
In the stochastic representation of Gini’s index Ŵ1 and Ŵ2 are two i.i.d. copies of the normalized random
variable Ŵ . In the stochastic representation of the curvature index ŵmin and ŵmax are, respectively, the lower
and upper bounds of the normalized random variable Ŵ . The Rényi indices with exponents α ̸= 1 are given
in the sixth row; the Rényi indices with exponents α = 1 (the Theil-index counterpart) and α = 2 (the
Hirschman-index counterpart) are given in the last two rows.
Index Lorenzian representation Stochastic representation
1
Gini G = 2 0
(L(x) − x)dx 1
2
⟨|Ŵ1 − Ŵ2 |⟩
Pietra P = max0≤x≤1 (L(x) − x) ⟨(Ŵ − 1)+ ⟩
1 √ √ √ √
0 1+(L′ (x))2 dx− 2 ⟨ 1+Ŵ 2 ⟩− 2
Amato A = √ √
2− 2 2− 2
|L′′ (x)|
1 1 1
Curvature C = 0 (1+(L′ (x))2 )3/2
dx  − 
1
1+( 1
)2 1+(
ŵmin
)2
ŵmax
1 1
α α
1
Rényi 1 − Rα = ( (L (x)) dx) 1−α
0

⟨Ŵ ⟩ 1−α
1 ′
Rényi 1 − R1 = exp(− (L (x)) ln(L′ (x))dx) exp(−⟨Ŵ ln(Ŵ )⟩)
 1 0′
Rényi 1 − R2 = 1/( 0 (L (x))2 dx) 1/⟨Ŵ 2 ⟩

Table 6
Oligarchy models. The table summarizes – in the context of the two Oligarchy Models of Section 2.3 –
all gauges of statistical evenness discussed: the Gini, Pietra, Amato, curvature, and R ényi indices. In the
Oligarchy Model I the parameter f denotes the ‘‘oligarchy fraction’’ (0 < f < 1). In the Oligarchy Model II the
parameter f1 denotes the ‘‘oligarchy fraction’’, and the parameter f2 denotes the fraction of the population’s
total wealth held by the oligarchy (0 < f1 < f2 < 1). The Rényi indices with exponents α ̸= 1 are given
in the sixth row; the Rényi indices with exponents α = 1 (the Theil-index counterpart) and α = 2 (the
Hirschman-index counterpart) are given in the last two rows.
Index Oligarchy I Oligarchy II

Gini G = 1−f f2 − f1
Pietra P = 1−f f2 − f1
√ √ √ √ √
1+f 2 +(1−f )− 2 (f1 )2 +(f2 )2 + (1−f1 )2 +(1−f2 )2 − 2
Amato A = √ √
2− 2 2− 2
Curvature C = √1  1
−  1
1+f 2 f
1+( f1 )2
1−f
1+( 1−f1 )2
2 2
α (1−f2 )α
( (f(f)2α−
) 1
Rényi 1 − Rα = 1−f 1 + (1−f1 )α−1
) 1−α
1

Rényi 1 − R1 = 1−f ( ff12 )f2 ( 11− )


f1 1−f2
−f2
Rényi 1 − R2 = 1−f ( ff12 + 1−f1 )
1−f2 −1

of entropy. Economics, on the other hand, was interested in quantifying wealth-equality within human populations,
and established an impressive toolbox of quantitative measures of societal egalitarianism. These measures of societal
egalitarianism – in the context of general data-sets with positive values – are, in effect, general gauges of statistical evenness.
This paper presented a panoramic overview of the measurement of statistical evenness: the quantitative measurement of
statistical heterogeneity based on the notion of evenness — rather than on the notion of randomness. Three categories of
evenness gauges were presented and discussed in detail:

• Evenness gauges based on distances in the space of Lorenz curves. These gauges include the Gini and Pietra indices — the
latter index coinciding with Hoover’s ‘‘Robin Hood’’ index.
• Evenness gauges based on the length and curvature of Lorenz curves — Amato’s index and the curvature index.
• Evenness gauges based on Rényi’s entropies. These gauges generalize and connect together the Hirschman and Theil
indices, and are analogous to Atkinson’s indices.

All the aforementioned evenness gauges are indices taking values in the unit interval (0, 1). A small index value indicates
a high degree of statistical evenness, whereas a large index value indicates a high degree of statistical heterogeneity. In
the context of human populations the lower bound 0 represents pure communism — the most even societal scenario in
which wealth is equally distributed amongst all the population members. On the other hand (albeit in the case of the
curvature index), the upper bound 1 represents pure monarchy — the most uneven societal scenario in which one single
monarch possesses all wealth, and all other population members are completely impoverished. Lorenzian and stochastic
representations of all the aforementioned evenness gauges are summarized in Table 5. The aforementioned evenness gauges,
in the case of the Oligarchy Models of Section 2.3, are summarized in Table 6.
1352 I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353

Acknowledgment

The Authors gratefully acknowledges Shlomi Reuveni for having created the figures.

Appendix

In the appendix we describe the measurement of randomness in the context of Poissonian populations. Since the
measurement of randomness does not require positive-valued data sets, we henceforth consider Poissonian populations
on the entire real line. Namely: A Poisson
∞ process X on the real line, with intensity λ(x) (x real), is a Poissonian population
if its tail intensity function Λ(l) = l λ(x)dx (l real) is monotone decreasing from infinity (i.e., liml→−∞ Λ(l) = ∞) to zero
(i.e., liml→∞ Λ(l) = 0).
The ‘‘existence theorem’’ of the theory of Poisson processes ([48] Section 2.5) implies that the points of a Poissonian
population X residing above the level l are i.i.d. random variables governed by the probability density function φl (x) =
λ(x)/Λ(l) (x ≥ l). Hence, the Rényi entropy corresponding to the finite sub-population X ∩ (l, ∞) is given by
 ∫ ∞ 
1
 ln φl (x)α dx α ̸= 1,
1−α


l
El = ∫ ∞ (117)
φl (x) ln(φl (x))dx α = 1.

−

l

Applying the change of variables y = x − l we obtain that


 ∫ ∞ 
1
 ln φl (l + y)α dy α ̸= 1,
1−α


0
El = ∫ ∞ (118)
φl (l + y) ln(φl (l + y))dy α = 1.

−

0

Eq. (118) implies that the entropies {El }l are l-invariant if and only if the function φl (l + y) (l real, y ≥ 0) is l-invariant. In
particular, for y = 0 the l-invariance implies that:

λ(l)
= const (119)
Λ(l)
(l real). Since Λ′ (l) = −λ(l) Eq. (119) implies that the tail intensity Λ(·) is an exponential:
Λ(l) = c exp(−pl) (120)
(l real), where the coefficient c and the exponent p are positive parameters.

References

[1] J.P. Bouchaud, A. Georges, Phys. Rep. 195 (1990) 12.


[2] R. Metzler, J. Klafter, Phys. Rep. 339 (2000) 1.
[3] E.T. Jaynes, Phys. Rev. 106 (1957) 620.
[4] M. Kardar, Statistical Physics of Particles, Cambridge University Press, Cambridge, 2007.
[5] T.M. Cover, Elements of Information Theory, 2nd ed., Wiley, New York, 2006.
[6] R.M. Gray, Entropy and Information Theory, 2nd ed., Springer, New York, 2011.
[7] P.B. Coulter, Measuring Inequality: A Methodological Handbook, Westview Press, Boulder, 1989.
[8] F.A. Cowell, Measuring Inequality, 3rd ed., 2000, http://sticerd.lse.ac.uk/research/frankweb/MeasuringInequality/index.html.
[9] C. Gini, 1912, Reprinted in :Variabilita e Mutabilita E. Pizetti and T. Salvemini, Memorie di Metodologica Statistica, Libreria Eredi Virgilio Veschi, Rome,
1955.
[10] E.A. Hammel, PNAS 102 (2005) 2248.
[11] R.G. Abraham, S. van den Bergh, P. Nair, Astrophys. J. 558 (2003) 218.
[12] Z. Ouyang, Q. Zhou, W.H. Wong, PNAS 106 (2009) 21521.
[13] L. Wittebolle, et al., Nature 458 (2009) 623;
S. Naeem, Nature 458 (2009) 579.
[14] A.D. Davidson, et al., PNAS 26 (2009) 10702.
[15] K.H. Ho, F.K. Chow, H.F. Chau, Phys. Rev. E 70 (2004) 066110.
[16] G. Beaugrand, M. Edwards, L. Legendre, PNAS 107 (2010) 10120.
[17] N. Sazuka, J. Inoue, Physica A 383 (2007) 49;
N. Sazuka, J. Inoue, E. Scalas, Physica A 388 (2009) 2839.
[18] P.P. Graczyk, J. Med. Chem. 50 (2007) 5773.
[19] M.E.J. Woolhouse, et al., PNAS 94 (1997) 338.
[20] R.R. Rindfuss, et al., PNAS 101 (2004) 13976.
[21] A.G. Patt, et al., PNAS 107 (2010) 1333.
[22] M.O. Lorenz, Pub. Amer. Stat. Assoc. 9 (1905) 209.
[23] G. Pietra, Atti del Reale Istituto Veneto di Scienze, Lettere ed Arti, tomo LXXIV, parte II, 1914–15, p. 775.
[24] E.M. Hoover, Rev. Econ. Stat. 18 (1936) 162.
[25] V. Amato, Metodologia Statistica Strutturale, vol. 1, Cacucci, Bari, 1968.
[26] A.O. Hirschman, National Power and the Structure of Foreign Trade, University of Califorina Press, Berkeley, 1945.
I.I. Eliazar, I.M. Sokolov / Physica A 391 (2012) 1323–1353 1353

[27] H. Theil, Economics and Information Theory, Rand McNally and Co., Chicago, 1967.
[28] A.B. Atkinson, J. Econ. Theoret. 2 (1970) 244.
[29] I. Eliazar, Physica A 386 (2007) 318.
[30] I. Eliazar, I.M. Sokolov, Physica A 389 (2010) 117.
[31] I. Eliazar, I.M. Sokolov, Phys. Rev. E 81 (2010) 011122.
[32] I. Eliazar, I.M. Sokolov, Physica A 389 (2010) 3023.
[33] I. Eliazar, I.M. Sokolov, Physica A 389 (2010) 4462.
[34] I. Eliazar, Physica A 390 (2011) 699.
[35] I. Eliazar, Physica A 390 (2011) 1982.
[36] I. Eliazar, J. Phys. A: Math. Theoret. 44 (2011) 022001.
[37] I. Eliazar, Physica A 390 (2011) 3189–3203.
[38] I. Eliazar, M.H. Cohen, Physica A 390 (2011) 4293.
[39] B.B. Mandelbrot, Fractals and Scaling in Finance: Discontinuity, Concentration, Risk, Springer, New York, 1997.
[40] J. Voit, The Statistical Mechanics of Financial Markets, Springer, New York, 2005.
[41] R.N. Mantegna, H.E. Stanley, An Introduction to Econophysics: Correlations and Complexity in Finance, Cambridge University Press, Cambridge, 2007.
[42] R. Albert, A.L. Barabási, Rev. Modern Phys. 74 (2002) 47.
[43] R. Cohen, S. Havlin, Complex Networks: Structure, Robustness and Function, Cambridge University Press, Cambridge, 2010.
[44] S.N. Durlauf, H.P. Young (Eds.), Social Dynamics, MIT Press, Boston, 2004.
[45] C. Castellano, S. Fortunato, V. Loreto, Rev. Modern Phys 81 (2009) 591–646.
[46] H.L. Royden, Real Analysis, 3rd ed., Macmillan, New York, 1988.
[47] D.R. Cox, Renewal Theory, Longman, London, 1962.
[48] J.F.C. Kingman, Poisson Processes, Oxford University Press, Oxford, 1993.
[49] R.W. Wolff, Stochastic Modeling and the Theory of Queues, Prentice-Hall, London, 1989.
[50] S.M. Ross, Applied Probability Models with Optimization Applications, Holden-Day, San Francisco, 1970.
[51] L. Takacs, Introduction to the Theory of Queues, Oxford University Press, Oxford, 1962.
[52] D. Gross, C.M. Harris, Fundamentals of Queueing Theory, Wiley, New York, 1974.
[53] I. Eliazar, Queueing Syst. 55 (2007) 71.
[54] D.R. Cox, Long-range dependence: a review, in: H.A. David, H.T. David (Eds.), Statistics: An Appraisal, Iowa State University Press, Ames, 1984,
pp. 55–74.
[55] G. Rangarajan, M. Ding (Eds.), Processes With Long-Range Correlations: Theory and Applications, Springer, New York, 2003.
[56] B.B. Mandelbrot, J.R. Wallis, Water Resour. Res. 4 (1968) 909.
[57] J.C. Hull, Options, Futures, and Other Derivatives, 3rd ed., Prentice-Hall, London, 1989.
[58] N. Bingham, R. Kiesel, Risk-Neutral Valuation: Pricing and Hedging of Financial Derivatives, 2nd ed., Springer, New York, 2004.
[59] N.C. Kakwani, Income Inequality and Poverty, Methods of Estimation and Policy Applications, Oxford University Press, Oxford, 1980.
[60] H.S.M. Coxeter, Introduction to Geometry, Wiley, New York, 1969.
[61] A. Rényi, Proc. 4th Berkeley Symp. Math. Stat. Prob. 1960, 1961, p. 547.
[62] E.H. Simpson, Nature 163 (1949) 688.
[63] P.M. Blau, Inequality and Heterogeneity: A Primitive Theory of Social Structure, Free Press, New York, 1977.
[64] R.J. Bell, P. Dean, D.C. Hibbins-Butler, J. Phys. C 3 (1970) 2111.
[65] R.J. Bell, P. Dean, Phil. Mag. 25 (1972) 1381.
[66] D.J. Thouless, Phys. Rep. 13 (1974) 93.
[67] C. Castellani, L. Peliti, J. Phys. A: Math. Gen. 19 (1986) L429.
[68] C.E. Shannon, W. Weaver, The Mathematical Theory of Communication, University of Illinois press, Urbana, 1949.
[69] M.O. Hill, Ecology 54 (1973) 427.
[70] C. Mejia-Monasterio, G. Oshanin, G. Schehr, J. Stat. Mech. (2011) P06022.
[71] G. Oshanin, Yu. Holovatch, G. Schehr, Proportionate vs disproportionate distribution of wealth of two individuals in a tempered Paretian ensemble,
Physica A 390 (2011) 4340. doi:10.1016/j.physa.2011.06.067.
[72] W. Feller, An Introduction to Probability Theory and its Applications, vol. 2, 2nd ed., Wiley, New York, 1971.
[73] A.O. Hirschman, Amer. Econ. Rev. 54 (1964) 761.
[74] A.L. Barabási, Linked, Perseus Publishing, Cambridge, MA, 2002.
[75] V. Pareto, Cours d’économie Politique, Droz, Geneva, 1896.
[76] I. Eliazar, J. Klafter, Phys. Rev. E 77 (2008) 061125.
[77] I. Eliazar, J. Klafter, Risk Decisions Anal. 1 (2009) 155.
[78] N.H. Bingham, C.M. Goldie, J.L. Teugels, Regular Variation, Cambridge University Press, Cambridge, 1987.
[79] H.E. Stanley, Introduction to Phase Transitions and Critical Phenomena, Oxford University Press, Oxford, 1987.
[80] P. Embrechts, C. Kluppelberg, T. Mikosch, Modelling Extremal Events for Insurance and Finance, Springer, New York, 1997.

You might also like