On the Generalized Lognormal Distribution

T. L. Toulias1 and C. P. Kitsos1

1
Technological Educational Institute of Athens, Department of Mathematics, Greece
{t.toulias, xkitsos}@teiath.gr

Abstract

This paper introduces, investigates and discusses the order generalized Lognormal dis-
tribution (GLD). Under certain values of the extra shape parameter the usual Lognormal,
LogLaplace and LogUniform distribution are obtained, as well as the degenerate Dirac dis-
tribution. The shape of all the members of the GLD family is extensively discussed. The
cumulative distribution function is evaluated through the generalized error function, while
series expansion forms are derived. Moreover, the moments for the GLD are also studied.

Keywords: generalized error function, Lognormal distribution, LogUniform distribution,

LogLaplace distribution, moments.

1 Introduction
Lognormal distribution has been widely applied in many dierent aspects of life sciences, including
Biology, Ecology, Geology and Meteorology as well as in Economics, Finance and Risk Analysis,
see [3]. Also, it plays an important role in Astrophysics and Cosmology, see [16, 1, 2] among others,
while for Lognormal expansions see [6].
In principle, the Lognormal distribution is dened as the distribution of a random variable whose
logarithm is normally distributed, and usually is formulated with two parameters. Furthermore,
LogUniform and LogLaplace distributions can be similarly dened with applications in Finance,
see [18] and [14]. Especially, the powertail phenomenon of the LogLaplace distributions [13]
attracts attention quite often in Environmental Sciences, Physics, Economics, Finance as well as
in longitudinal studies [4]. Recently, LogLaplace distributions have been proposed for modeling
growth rates as stock prices, [17], and currency exchange rates, [14].
In this paper a generalized form of Lognormal distribution is introduced, involving a third,
shape, parameter. With this generalization, a family of distributions is emerged, that combines
theoretically all the properties of Lognormal, LogUniform and LogLaplace distribution, depend-
ing on the value of this third parameter.

1
 The generalized order Lognormal distribution (GLD) is the distribution of a random vec-
tor whose logarithm follows the order Normal distribution, an exponential power generalization
of the usual Normal distribution, introduced by [7] and [8]. This family of pdimensional gener-
alized Normal distributions, denoted by Np (, ), is equipped with an extra shape parameter ,
and constructed to play the role of Normal distribution for the generalized Fishers entropy type
information, see also [11] and [12].

The density function fX of a pvariate, order normally distributed random variable Y

Np (, ), with location vector Rp , positive denite scale matrix Rpp and shape parameter
R \ [0, 1] is given by [7],
{ }
1/2
fY (y) = fY (y; , , ) = Cp |det | exp 1
Q (y) 2(1) , y Rp , (1)

where Q is the quadratic form Q (y) = (y )T 1 (y ), = (, ) while Cp being the

normalizing factor
( p2 + 1)
1 p 1
Cp = p/2 1 .
1 ( ) (2)
(p )

From (1), notice that the secondordered Normal is the known multivariate normal distribution,
i.e. N2p (, ) = N p (, ), see also [11] and [9].

In Section 2, a generalized form of the Lognormal distribution is introduced, which is derived

from the univariate family of N (, 2 ) = N1 (, 2 ) distributions, denoted by LN (, ), in-
cludes the LogLaplace distribution as well as the LogUniform distribution. The shape of the
LN (, ) members is extensively discussed while it is connected to the tailing behavior of LN
through the study of the c.d.f. In Section 3, an investigation of the moments of the generalized
Lognormal distribution, as well as the special cases of LogUniform and LogLaplace distributions,
is presented.

The generalized error function that briey provided here, plays an important role to the de-
velopment of LN (, ), see Section 2. The generalized error function denoted by Erf a and the
generalized complementary error function Erfca = 1 Erf a , a 0, [5], are dened respectively as

x
et dt,
(a+1) a
Erf a (x) :=

x R. (3)
0

The generalized error function, can be expressed (changing to ta variable), through the lower
incomplete gamma function (a, x) or the upper (complementary) incomplete gamma function

2
(a, x) = (a) (a, x), in the form

(a) (1 ) (a) [ (1) ( )]

Erf a (x) =
a, x
a
=

a a1 , xa , x R, (4)

see [5]. Moreover, adopting the series expansion form of the lower incomplete gamma function,

x
(1)k a+k
(a, x) := ta1 et dt = x , x, a R+ , (5)
k!(a + k)
0 k=0

a series expansion form of the generalized error function is extracted,

(a + 1) (1)k
Erf a (x) = xka+1 , x, a R+ . (6)
k!(ka + 1)
k=0

Notice that, Erf 2 is the known error function erf, i.e. Erf 2 (x) = erf(x), while Erf 0 is the function

of a straight line through the origin with slope (e )1 . Applying a = 2 the known incomplete

gamma function identities such as (1/2, x) = erf x, and (1/2, x) = (1 erf x) =

erfc x, x 0 is obtained. Moreover, while Erf a 0 = 0 for all a R+ . and

(1)
lim Erf a x = 1 (a) a , a R+ ,
x

as (a, x) (a) when x +.

2 The order Lognormal Distribution

The generalized univariate Lognormal distribution is dened, through the univariate generalized
order Normal distribution, as follows.

Definition 2.1. When the logarithm of a random variable X follows the univariate order Nor-
mal distribution, i.e. log X N (, 2 ), then X is said to follow the generalized Lognormal
distribution, denoted by LN (, ), i.e. X LN (, ).

The LN (, ) it is referred as the (generalized) order Lognormal distribution (GLD).

Like the usual Lognormal distribution, the parameter R it is considered logscaled, while the
non logscaled (i.e. e when is assumed logscaled) is referred as the location parameter of
LN (, ). Hence, if X LN (, ) then log X is a order normally distributed variable, i.e.
log X N (, 2 ). Therefore, the location parameter R of X is in fact the mean of Xs

3
natural logarithm, i.e. E[log X] = , while
( )
( )2 1 3 1

( ) 2 , and

Var[log X] = 1 (7)
1

( ) ( )
1
5 1
Kurt[log X] = ( ) , (8)
2 3 1

see [9] for details on N .

Let Y := log X N (, 2 ) with density function as in (1) and X = g(Y ) = eY . Then, the
density function fX of X LN (, ) can be written, through (1), as
{   }
  1  log x  1
exp  
( 1 )  d 1 
  1
fX (x) = fX (x; , , ) = fY g (x)  g (x) = fY (log x) x = .
dx 2( 1
)
1/ ( 1 )x

(9)
The probability density function fX , as in (9), is dened on R+ = R+ \ 0, i.e. LN (, ) has zero
threshold. Therefore, the following Denition extends Denition 1.

Definition 2.2. When the logarithm of a random variable X + follows the univariate order
Normal distribution, i.e. log(X+) N (, 2 ), then X is said to follow the generalized Lognormal
distribution with threshold R, i.e. X LN (, ; ).

It is clear that when X LN (, ; , ) then log(X ) is a order normally distributed

variable, i.e. log(X ) N (, 2 ) and thus, is the mean of (X )s natural logarithm while
Var[log X] is the same as in (7).
Let Y = log(X + ) N (, 2 ). The density function of X = eY LN (, ; ) is given
by fX (x) = fX (x ), x > 0.
log(x )
Let z = . Then, the limiting threshold density value of fX (x) with x + ,
implies that
{ ( )}

lim+ fX (x) = 1 C1 lim exp z + z = 1 C1 e(sgn )() ,
1

|z|
1 1 (10)
x z

and therefore
0, (1, +),
lim+ fX (x) = (11)
x + , (, 0),

4
i.e. the fX s dening domain, for the positiveordered Lognormal random variable X, can be
extended to include threshold point , by letting fX () = 0.
The generalized Lognormal family of distributions LN is a wide range family bridging the Log
Uniform LU, Lognormal LN and LogLaplace LL, as well as the degenerate Dirac D distributions.
Indeed:

Theorem 2.1. The generalized Lognormal distribution LN (, ), for order values of = 0, 1, 2, ,

is reduced to

D(e ), = 0,




LU(e , e+ ), = 1,
LN (, ) = (12)

LN (, ), = 2,




LL(e , 1 , 1 ), = .

Proof. From the denition (1) of N the order values is a real number outside the closed interval
[0, 1]. Let X LN (, ) with density function fX as in (9). We consider the following cases:

(i) The limiting case = 1. Let x R+ such that | log x| 1. Using the gamma function ad-
ditive identity (z +1) = z (z), z R+ in (9) we have that LN 1 (, ) = lim1+ LN (, )
with
1
x [e , e+ ],
2x ,
fX1 (x) := lim+ fX (x) = (13)
1 0, x (, e ) (e+ , +),

which is the density function of the LogUniform distribution LU(a, b), 0 < a < b, with
1 1
a = e and b = e+ , i.e. = 2 log(ab) and = 2 log ab . Therefore, rstordered
Lognormal distribution is in fact the LogUniform distribution, with vanishing threshold
density, fX1 (0) = limx0+ fX1 (x) = 0. For the purposes of statistical application the Log
Uniform moments are not the same as the model parameters, i.e., although X = , X =

/ 3.

(ii) The normal case = 2. It is clear that LN 2 (, ) = LN (, ), as fX2 coincides with the
Lognormal density function, and therefore the second-ordered Lognormal distribution is in
fact the usual Lognormal distribution.

(iii) The limiting case = . We have LN (, ) := lim LN (, ) with

/
e
}
1
{  x , x (0, e ],
1  log x  2
fX (x) := lim fX (x) = exp   = (14)
2x
/
e x , x > e ,
+1

5
 which coincides with the density function of the known LogLaplace distribution (symmetric
LogExponential distribution) LL( , , ) with = e and = = 1/, see [13]. There-
fore, the inniteordered LogNormal distribution is in fact the LogLaplace distribution,
with threshold density


0, < 1,

fX (0) = lim+ fX (x) = 1, = 1, (15)
x0


+ , > 1.

For the purposes of statistical application the LogLaplace moments are not the same as the

model parameters, i.e., although X = , X = 2.

(iv) The limiting case = 0. It is

( )
1 kk
lim fX (e ) = lim e0 = +, (16)
0 2e k k!

through the Stirlings asymptotic formula. Assuming now x = e we have, through (16) and
the Stirlings asymptotic formula, that

| log x |
fX0 (x) := lim fX (x) = 0 1
e = 0, (17)
0 2 2 2 x

i.e. LN 0 (, ) := lim0 LN (, ) = D(e ) as fX0 coincides with the Dirac density

function, with the (non logscaled) location parameter e of LN 0 (, ) being the singular
(innity) point. Therefore, the zeroordered Lognormal distribution LN 0 is in fact the
degenerate Dirac distribution with pole at the location parameter of LN 0 (with vanishing
threshold density fX0 (0) = limx0+ fX0 (x) = 0.

From the above limiting cases (i), (iii) and (iv) the dening domain R \ [0, 1] of the order values
, used in (1), it is safely extended, to include the values = 0, 1, , i.e. can now be dened
outside the open interval (0, 1). Eventually, the family of the order Normals can include the
LogUniform, Lognormal, LogLaplace and the degenerate Dirac distributions, as (12) holds.

From the above Theorem 2, (11) and (14) the domain of the density functions fX (x), x > 0 can
also be extended, to include the threshold point x = 0, by setting fX (0) := 0 for all nonnegative
ordered Lognormals, i.e. for all 0 [1, +), while for the LogLaplace case of = + with
= 1 by setting fX+ (0) := 12 e .

6
 From the fact that N0 (, ) = D(), see [9], one can say that the degenerate LogDirac distri-
bution, say LD(), equals LN 0 (, ) and hence, through Theorem 1, we can write LD() = D(e ).

Proposition 2.1. The mode of the positiveordered Lognormal random variable X LN (, ),

X < e (1, +), is given by

Mode X = e , (18)

with corresponding maximum density value,

exp{( )/ }
max fX = fX (Mode X ) = . (19)
2( 1
)
1/ ( 1 )

Proof. Recall the density function of X LN (, ) as in (9), and let m = Mode X > 0. Then
d
it holds dx fX (m; , , ) = 0, i.e.

(1 ) 1
|log m | 1 d
dx |log x |x=m = m

. (20)

dx |x| = sgn x, x R, it is
d
From

m = e , (21)

provided that x < e . Otherwise (21) holds trivially, as (20) implies = 0, i.e. m = e . Moreover,
d
dx fX (x) > 0 when

1
1 + sgn(log x ) 1 | log x | 1 > 0, x > 0, (22)

and thus, fX is a strictly ascending density function on (0, e ) when > 1, and also on

d
(e , e ) when < 0. Similarly, with dx fX (x) < 0, fX is a strictly descending density function


on (e
, +) when > 1, and also on (0, e ) (e , +) when < 0. Especially, for < 0,
the point e is a nonsmooth point of fX , as
[  1 ]
 1
C1  
= + lim sgn(log x ) 
log x 
lim d fX (x) = +. (23)
xe dx (e )2 xe

Therefore, the positiveordered Lognormals formed by a unimodal density function with mode as
in (18), and corresponding maximum density as in (19), see Figures 1(a1), 1(a2) and 1(a3).

Proposition 2.2. The global mode point of the negativeordered Lognormal random variable X
LN (, ), (, 0), is (in limit) the threshold 0 which is an innite (probability) density point.
Moreover, the location parameter (i.e. e ) of LN (, ) is a nonsmooth (local) mode point for

7
all X<0 that corresponds to locally maximum density

1
fX (e ) = , (24)
2( 1
)
1/ ( 1
)e

while e is a local minimum (probability) density point with corresponding locally minimum

density fX (e ).

Proof. The negativeordered Lognormals formed by density functions admitting threshold 0 (in
limit) for their global mode point (of innite density), shown in (11). Moreover, from the above
discussed monotonicity of fX in Proposition 2.1 all the negativeordered Lognormals admit also
e as a local nonsmooth mode point, and exp{ } as a local minimum density point, with

densities as in (24) and fX (e ) respectively, see Figures 1(b1), 1(b2) and 1(b3).

Furthermore, Proposition 2.2 holds (in limit) for random variables X1 and X which provide
LogUniform and LogLaplace distributions respectively. Indeed, for given a, b R+ , X1
LU(a, b) = LN 1 (, ) with = 1
2 log(ab) and = 1
2 log ab , we get, through (18) and (19), that

Mode X1 := Mode X1+ = e = a with corresponding maximum density (i.e. the maximum
value of the density function),
1
e ( 1
)

1
max fX1 = fX1 (a) = lim 1 = . (25)
2 1 ( + 1)
+
a log ab

Moreover, the nonzero minimum density (i.e. the minimum, but not zero, value of the density
1
( )
function) is obtained at x = b with min fX1 = fX1 (b) = 2b = 1/ b log ab . These results are in
accordance with the LogUniform density function in (13).
For X LN (log , 1/) = LL(, , ) we evaluate, through (18) and (19), that


0, < 1,

Mode X+ = /e = 1, (26)



, > 1,

with the corresponding maximum density value being innite, i.e. max fX = fX (0) = +,
provided < 1, and max fX+ = /(2), provided 1. The same result can also be derived
through (24) as . These results are in accordance with the LogLaplace density function
in (14), although for = 1, Mode X can be dened, through (14), for any value inside the
interval (0, ].

8
 The above discussion on behavior of the modes with respect to shape parameter is formed in
the following Propositions.

Proposition 2.3. Consider the positiveordered Lognormal family of distributions LN (, ) with

xed parameters , and 1. When rises, i.e. when we move from LogUniform to Log
Laplace distribution inside the LN family, the mode points of LN are:

strictly increasing from e (LogUniform case) to e (LogLaplace case) provided that

1
< 1 (with their corresponding maximum density values moving smoothly from 2 e to
+).

xed at e1 for all LN 1 (, = 1) (with the corresponding maximum density values

moving smoothly from 12 e1 to 12 e ).

strictly decreasing from e (LogUniform case) to threshold 0 (LogLaplace case) provided

1
that > 1 (with their corresponding maximum density values moving smoothly from 2 e
1
to 2 e ).

Proof. Let X LN (, ). Mode X is a smooth monotonous function of (, 0) (1, +)

for positive and negativeordered X , as

d
d Mode X = log()e . (27)

For X LN (, ) we evaluate, through (18) and (19), that

e , < 1,

1
Mode X+ = e = 1, (28)



0, > 1,

with the corresponding maximum density value being innite, i.e. max fX = fX (0) = +,
provided > 1, and max fX+ = 1/(2e ), provided 1.
Assume that 1. Considering (25), (28) with (27) and Proposition 1 2.1 the results for the
positiveordered Lognormals hold.

Proposition 2.4. For the negativeordered Lognormal family of distributions LN (, ) with

< 0, when rises, i.e. when we move from LogLaplace to degenerate Dirac distribution inside
the LN family, the local minimum (probability) density points of LN are:

9
 strictly increasing from threshold 0 (LogLaplace case) to e1 (Dirac case) provided that
< 1.

xed at e1 for all LN (, = 1).

strictly decreasing from e (LogLaplace case) to e1 (Dirac case) provided that > 1.

d d
Proof. Assume now that < 0. From (27) it is d (e )<0 when > 1 and d (e ) >0

when < 1. Therefore, the local minimum density point e (see Proposition 2.2) for > 1 is

decreasing from e | = e to Mode X0 = e1 through (18). When = 1, e = e1

for all < 0, while for < 1, e increases from e | = 0 to Mode X0 = e1 through
(18).

It is easy to see that for the LogLaplace case LL(, , ), the local minimum density point

e , of X<0 with > 1 coincides (in limit) with the local nonsmooth mode point e of X ,

see Figure 1(b3). Also, notice that the local minimum density point e , < 0, for the Dirac
case D(e ), is the limiting point e1 although the (probability) density in D(e ) case vanishes
everywhere except the innite pole e .
Figure 1 illustrates the probability density functions fX curves for scale parameters =
2/3, 1, 3/2 of the positiveordered lognormally distributed X LN 1 (0, ) in the left sub
gures (a1)(a3) respectively, while the p.d.f. of negativeordered lognormally distributed X
LN <0 (0, ) aredepicted in the right subgures (b1)(b3) respectively. Moreover, the density

points e on fX are also depicted (small circles over p.d.f. curves with their corresponding
ticks on xaxis). According to the above Proposition 2.4, in the left subgures (a1)(a3), i.e. for
positiveordered X1 , these density points represent the mode points on fX while in the right
subgures (b1)(b3), i.e. for negativeordered X<0 , represent the local minimum density points
on fX curves.
For the evaluation of the cumulative distribution function (c.d.f.) of the generalized Lognormal
distribution, the following Theorem stated and proved.

Theorem 2.2. The c.d.f. FX of a order Lognormal random variable X LN (, ) is given

by
{ }
log x 1
FX (x) = 1
+ Erf 1
( 1
)

(29)
2
2 ( 1
) ( 1 )
( ( ) )
1 1 log x 1
= 1 1 , , x R+ . (30)
2 ( 1
)

10
Figure 1: Graphs of the density functions fX , X LN (0, ), for = 2/3, 1, 3/2, and various
positive (left subgures) and negative (right subgures) values.

11
Proof. From density function fX , as in (9), we have

x x {   }
1  log t  1
FX (x) = FX (x; , , ) = fX (t)dt = C1 t1 exp 1
  dt.
0 0

log t
Applying the transformation w = , t > 0, the above c.d.f. is reduced to

log x

{ } ( )

log x
FX (x) = C1 |w|
exp 1 1 dw = Z , (31)

where Z is the c.d.f. of the standardized order Normal distribution Z = 1

(log X )
N (0, 1). Moreover, Z can be expressed in terms of the generalized error function. In particular

z { } z { }

Z (z) = C1 exp |w| 1 dw = Z (0) + C exp 1
1 1
|w|
1 dw,
0

and as fZ is a symmetric density function around zero, we have

z { } z {  }
 1
 1 1 
Z (z) = 1
2 + C1 exp 1
|w| 1 dw = 1
2 + C1 exp ( ) w dw,
0 0

and thus
1
( 1
)
z
1
{
}
Z (z) = 1
2

+ C1 ( 1 ) exp u 1 du. (32)
0

Substituting the normalizing factor, as in (2), and using (3) we obtain

{ }
1 1
Z (z) = 1
+ 21 Erf 1 ( )
z , z R, (33)
2
2 ( 1
+ 1) ( 1 )

and nally, through (31), we derive (29), which forms (30) through (4).

It is essential for numeric calculations, to express (29) considering positive arguments for Erf.
Indeed, through (32), it is
{  }
sgn(log x ) 1 1  log x 
FX (x) = 1
+ Erf ( )
  , (34)
2
2 ( 1
) ( 1 )
1

12
while applying (4) into (34) it is obtained
(   )
1 + sgn(log x ) sgn(log x )  log x  1
FX (x) = 1 1
,   . (35)
2 2 ( 1
)

As the generalized error function Erf a is dened in (4), through the upper incomplete gamma
function (a1 , ), series expansions can be used for a more numericaloriented form of (4). Here
some expansions of the c.d.f. of the generalized Lognormal distribution are presented.

Corollary 2.1. The c.d.f. FX can be expressed in the series expansion form

(   )k
( ) 1  log x  1
( 1
1
 
)

log x
FX (x) = 1
+ , x R+ . (36)
2 2 1
( )
k![(k + 1) 1]
k=0

Proof. Substituting the series expansion form of (6) into (34) and expressing the innite series
using the integer powers k, the series expansion as in (36) is derived.

Corollary 2.2. For the negativeordered lognormallydistributed random variable X with =

1
1n R , n N, n 2, the nite expansion is obtained

 
sgn(log x ) nk  log x k/n
n1
FX (x) = 1
+ 1
sgn(log x ) {   }   . (37)
2 2
 log x 1/n k=0 k!  
2 exp n  

Proof. Applying the following nite expansion form of the upper incomplete gamma function,

n1
xk
(n, x) = (n 1)!ex , x R, n N = N \ 0,
k
k=0

into (35) we readily get (37).

Example 2.1. For the (1)ordered lognormally distributed X1 (i.e. for n = 2) we have
 
 log x 
1+2  
FX1 (x) = 1
2 + 12 sgn(log x ) sgn(log x ) {  } ,
 log x 
2 exp 2  

13
while for the (1/2)ordered lognormally distributed X1/2 (i.e. for n = 3), it is
  ( )
3  log x  log x 2
1+3  
3
+ 9
FX1/2 (x) = 1
2 + 12 sgn(log x ) sgn(log x ) {  } .
 log x 
2 exp 3 3  

Example 2.2. For the secondordered Lognormal random variable X2 LN 2 (, ), we immediate

derive, from (29), that
( ) ( ) ( )
log x 1 1 logx 1 1 logx
FX2 (x) = X2 = 2 + 2 Erf 2 = 2 + 2 erf , (38)
2 2

i.e. the c.d.f. of the usual Lognormal is derived, as it is expected due to LN 2 = LN , see Theo-
rem 2.1.

Example 2.3. For the inniteordered Lognormal X LN (, ), setting 1

= 1, we
obtain through (36) and the exponential series expansion, that

( )
1 1 log x
FX (x) = 2 + 2 Erf 1

(  )
 log x  k+1
= 1
2 2 sgn(log x )
1
(k+1)! 
1

k=0
{  }
 log x 
= 1
2 + 1
2 sgn(log x ) 1
2 sgn(log x ) exp   , (39)

and hence
1 / 1/
2e x , x (0, e ],
FX (x) = / (40)

1 e , x (e , +),
2x1/
which is the c.d.f. of the LogLaplace distribution as in (14). This is expected as LN (, ) =
LL(e , 1/, 1/), see Theorem 2.1.

It is interesting to mention here that the same result can also derived through (37), as this nite
expansion can be extended for n = 1, which provides (in limit) the c.d.f. of the inniteordered
Lognormal distribution.

Example 2.4. Similarly, for the rstordered random variable X1 LN 1 (, ), the expansion

14
(36) can be written as
(   )k
1 log x  1
( 1
1
( )
 
) log x

FX (x) = 1
+ 1 + ( 1) ,
2
2 ( 1
+ 1)
k![(k + 1) 1]
k=1

log x
and provided that 1, we obtain

1 log x log x +
FX1 (x) = lim+ FX (x) = 2 + 2 (1 + 0) = 2 ,
1

with FX1 (e ) = 0 and FX1 (e+ ) = 1. Therefore,

0, x (0, e ),

FX1 (x) = 1
(log x + ), x [e , e+ ], (41)


2

1, x (e+ , +),

coincides with the c.d.f. of the LogUniform distribution LU(a = e , b = e+ ) as in (14). This
is expected as LN 1 (, ) = LU(e , e+ ), see Theorem 2.1.

Table 1 provides the probability values P;1 = Pr{X i}, i = 1

2 , 1, 2, . . . , 5 for various
X LN (0, 1). Notice that P;1 = 1/2 for all values due to the fact that 1 = e |=0 = Med X
(see Theorem 2), i.e. the point 1 coincides with the invariant median of the LN (0, 1) family
discussed above. Moreover, the last two columns provide also the 1st and 3rd quartile points q1;
and q3; of X , i.e. Pr{X qk; } = k/4, k = 1, 3, for various values. These quartiles evaluated
using the quantile function QX of r.v. X , i.e.

{ }
QX (P ) := inf x R+ | FX (x) P
{ [ ( )] 1 }
1 1

= exp sgn(2P 1) 1
, |2P 1| , P (0, 1), (42)

for P = 1/4, 3/4, that derived through (35). The values of the inverse upper incomplete gamma
function 1 ( 1
, ) were numerically calculated.
Figure 2 illustrates the c.d.f. FX curves, as in (34), for certain r.v. X LN (0, ), and
for scale parameters = 2/3, 1, 3/2 in the 3 subgures respectively. Moreover, the 1st and 3rd
quartile points QX (1/4) and QX (3/4) are also depicted (small circles over c.d.f. curves with their
corresponding ticks on xaxis).

15
Figure 2: Graphs of the c.d.f. FX , X LN (0, ), for = 2/3, 1, 3/2 and various values.

16
Table 1: Probability mass values P;i = Pr{X i}, i = 12 , 1, 2, . . . , 5, and the 1st and 3rd quartiles
q1; , q3; for various generalized lognormally distributed X LN (0, 1).

P;1/2 P;1 P;2 P;3 P;4 P;5 q1; q3;

50 0.2501 0.5000 0.7499 0.8326 0.8739 0.8987 0.4998 2.0008
10 0.2505 0.5000 0.7495 0.8297 0.8698 0.8940 0.4990 2.0038
5 0.2508 0.5000 0.7492 0.8264 0.8652 0.8887 0.4982 2.0071
2 0.2515 0.5000 0.7485 0.8187 0.8539 0.8756 0.4964 2.0145
1 0.2521 0.5000 0.7479 0.8097 0.8408 0.8601 0.4945 2.0223
1/2 0.2524 0.5000 0.7476 0.7989 0.8248 0.8410 0.4925 2.0303
1/10 0.2528 0.5000 0.7482 0.7757 0.7895 0.7984 0.4986 2.0426
1 0.1534 0.5000 0.8466 1.0000 1.0000 1.0000 0.6065 1.6487
3/2 0.2381 0.5000 0.7619 0.8848 0.9437 0.9721 0.5172 1.9334
2 0.2441 0.5000 0.7559 0.8640 0.9172 0.9462 0.5094 1.9630
3 0.2472 0.5000 0.7528 0.8505 0.8989 0.9267 0.5049 1.9804
4 0.2481 0.5000 0.7519 0.8452 0.8917 0.9188 0.5034 1.9867
5 0.2486 0.5000 0.7514 0.8425 0.8878 0.9145 0.5025 1.9899
10 0.2494 0.5000 0.7506 0.8375 0.8810 0.9068 0.5011 1.9954
50 0.2499 0.5000 0.7501 0.8341 0.8761 0.9013 0.5002 1.9992
0.2500 0.5000 0.7500 0.8333 0.8750 0.9000 0.5000 2.0000

Theorem 2.3. The (non logscaled) location parameter e is in fact the geometric mean as well
as the median for all generalized lognormally distributed X LN (, ). Moreover, this median
is also characterized by vanishing median absolute deviation.

1
Proof. Considering (34) and the fact that Erf a 0 = 0, a R+ , it holds that Med X = FX
(1/2) =
e . For the geometric mean (g )X = eE[log X ] , we readily obtain (g )X = e as log X
N (, 2 ) with E[X ] = . A dispersion measure for the median is the socalled median absolute
deviation or MAD, dened by MAD(X ) = Med |X Med X |. For X LN (, ), it is
X Med X = X e LN (, ; e ), i.e. X e follows the generalized Lognormal
distribution with threshold e . Furthermore, |Y | is the folded distribution case of Y := X e
which is distributed through p.d.f. of the form

f|Y | (x) = fY (x) + fY (x), x R+ , (43)

where fY is the p.d.f. of Y . For example, see [15] on the folded Normal distribution. However, the

17
density function fY is dened on (e , +) due to threshold e , while vanishes elsewhere, i.e.

fY (x) + fY (x), 0 x e ,
f|Y | (x) = (44)
fY (x), x > e ,

Therefore, the c.d.f. of |Y | is given by

x e e x

F|Y | (x) = f|Y | (t)dt = fY (t)dt + fY (t)dt + fY (t)dt, x R+ . (45)

0 0 0 e

Applying the transformation w = 1

[log(e

t) ], t < e into the rst integral of (45) and
z= 1
[log(t + e ) ], t > e into the other two integrals, we obtain that

1
0 { } log 2
{ }

F|Y | (x) = C1 exp |w|
1 1 1
dw + C exp 1
|z| 1 dz+
0

g(x)
{ }

log(x+e )
C1 exp 1
|z| 1 dz, g(x) := , x R+ ,
1
log 2

and hence
[ ( ) ] [ ( )] (
)
F|Y | (x) = Z (0) + Z log 2 Z (0) + Z (g(x)) Z log 2 = Z log(x+e

)
, (46)

with Z being the c.d.f. of the standardized r.v. Z N (0, 1). From (33) and the fact that
1
Erf a 0 = 0, a R+ , it is clear that (46) implies MAD X = Med |X e | = F|X (1/2) = 0,
e |

for every X LN (, ), and Theorem has been proved.

3 Moments of the order Lognormal Distribution

For the evaluation of the moments of the generalized Lognormal distribution the following holds.
(t)
Proposition 3.1. The tth raw moment X of a generalized lognormally distributed random
variable X LN (, ), is given by

et (t)2n 2n 1 ( )
(t) 1
X = ( ) (2n + 1) , (47)
( 1
) n=0
(2n)! 1

18
and coincides with the moment generating function of the order normally distributed log X, i.e.
(t)
Mlog X (t) = X .

(t)
Proof. From the denition of the tth raw moment X we have
{   }
(t)  log x  1
X = E[X ] =t t
x fX (x)dx = 1 1
C x t1
exp 1
  dx,
R+ R+

1
and applying the transformation z = ( 1
)
1
(log x ), x > 0, we get
{ } { }
(t) 1 1

X = C1 ( 1 )
exp t + n( 1 ) z exp |z| 1 dz.
R

Through the exponential series expansion

{ } n
1 (t) 1
exp t( 1 ) z = ( )n z n ,
n! 1
n=0

it is obtained that

{ }
(t) 1 (t)2n 2n 1
X =
2C1 ( 1 ) et ( 1 ) z 2n exp z 1 dz. (48)
(2n)!
n=0 R+

Finally, substituting the normalizing factor C1 as in (2) into (48) and utilizing the known integral,
[5],

( m+1
n )
xm ebx dx = n, m, b R+ ,
n
m+1 , (49)
nb n
R+

we obtain (47).
(t)
Moreover, for Y := log X N (, 2 ) it is MY (t) = E[etY ] = E[X t ] = X , and Proposition
has been proved.

Example 3.1. For the secondordered lognormally distributed X LN 2 (, ), (47) implies

(t) et (2t2 2 )n ( )
X = n + 12 , t R+ ,
n=0 (2n)!

19
and through the gamma function identity

( ) (2n)!
n + 12 = 2n , n N, (50)
2 n!

it is



(t) (t)2n (1 )
2 2 n 1 2
X = et n = et 1
2t = et+ 2 (t) , t R+ , (51)
2 n! n!
n=0 n=0

which is the tth raw moment of the usual lognormally distributed X LN (, ), with mean
(1) (t)
X := X = E[X] = exp{ + 12 2 }. This is true as Mlog X (t) = X = exp{t + 12 (t)2 } is the
known moment generating function of the normally distributed log X N (, 2 ).
(t)
Theorem 3.1. The kth central moment (about the mean) X of a generalized lognormally dis-
tributed random variable X LN (, ), is given by

k ( )
(k) ek k ( X )n
X = Skn , k N, (52)
( 1 e
) n=0 n

where

( )
(k)2m 2m 1
Sk = ( 1 ) (2m + 1) 1
, k N. (53)
(2m)!
m=0

(k)
Proof. From the denition of the kth central moment X it is

(k)
X := E[(X X )k ] = (x X )k fX (x; , , )dx,
R+

while using the binomial identity we get

k ( ) k ( )
(k) k k (kn)
X = (X )n xkn fX (x)dx = (X )n X . (54)
n=0 n n=0 n
R+

Applying Proposition 3.1, (54) implies that

k ( ) ( )
(k) ek k X n [(k n)]2m 2m 1 1
X = ( ) ( ) (2m + 1) , (55)
( 1 e (2m)! 1
) n=0 n m=0

while taking the summation index n until k 1, we nally obtain (52), and Theorem has been
proved.

20
 1 2
Example 3.2. Recall Example 3.1. Substituting (51) and the mean X = e+ 2 into (54), the
secondordered lognormally distributed X LN 2 (, ), provides

k ( )
(k) k 1 2 2
X = (1)n ek+ 2 [n+(kn) ] , k N, (56)
n=0 n

while ( )
(2) 2 2
2
X := Var[X] = X = e2+ e 1 , (57)

which are the kth central moment and the variance, respectively, of the usual lognormally dis-
tributed X LN (, ). The same result can derived directly through (52), for = 2 and the use
of the known gamma function identity, as in (50).

2
Theorem 3.2. The mean X := E[X], variance X := Var[X], coecient of variation CVX ,
skewness X and kurtosis X of the generalized lognormally distributed X LN (, ) are, re-
spectively, given by
e
X = S1 , (58)
( 1
)

e2
2
X = 2X + S2 , (59)
( 1
)

S2
CVX2 = ( 1
) 1, (60)
S12
e3
X = CVX3 CVX1 + S ,
3 ( 1 ) 3
(61)
X

X e4
X = CVX4 6CVX2 4 + 4 S4 , (62)
CVX X ( 1
)

where the sums Si , i = 1, . . . , 4 are given by (53).

(1)
Proof. From Proposition 3.1 we easily obtain (58), as X := X . From Theorem 3.1 it is
[ ( )]1 ( )
(2)
2
X := X = 2X + 1
e2 S2 2e X S1 .

Hence, substituting S1 from (58), (59) holds. Moreover, the squared coecient of variation is
readily obtained via (59) and (58). By denition skewness X the standardized third (central)

21
 3 (3)
moment, i.e. X := Skew[X] = X /X . Theorem 3.1 provides that
[ ]1 ( )
X = CVX3 + X
3
( 1
) e3 S3 3e2 X S2 + 3e X S1 .

Substituting S1 and S2 from (58) and (59), we obtain (61). Finally, kurtosis X is (by denition)
4 (4)
the standardized fourth (central) moment, i.e. X := Kurt[X] = X /X , provides, through
Theorem 3.1, that
[ ]1 ( )
X = CVX4 + X
4
( 1
) e4 S4 4e3 X S3 + 6e2 2X S2 4e 3X S1 .

Substituting Si , i = 1, 2, 3, from (58), (59) and (61), we obtain (62)

Example 3.3. For the secondordered lognormally distributed X LN 2 (, ), utilizing (50) into

(53) we get Sn = e(n )/2 , n N . Applying this to Theorem 3.2 we derive (after some algebra)
2 2

1 2 2
( 2
)
X = e+ 2 , 2
X = e2+ e 1 , CVX = e2 1, (63)

( 2 ) 2 2 2
X = e + 2 e2 1, X = e4 + 2e3 + 3e2 3, (64)

which are the mean, variance, coecient of variation, skewness and kurtosis respectively, of usual
lognormally distributed X LN (, ).

For the usual lognormally distributed random variable X LN , it is known that Mode X <
Med X < X . The following Corollary examines this inequality for the LN family of distributions.

Corollary 3.1. For the ordered lognormally distributed X LN (, ), it is true that Mode X
Med X = (g )X X . The rst equality holds for the LogLaplace distributed X+ with < 1
as well as for all the negativeordered X<0 where Mode X is considered to be the local (non
smooth) mode point of X . The second equality holds for the degenerate Dirac case of X0 .

Proof. From (58) and Theorem 2.3 it is

Med X = (g )X = e < X , (65)

for every X LN (, ). The above inequality becomes equality for the limiting Dirac case of
X0 . For the relation between the mode and the median of X the following cases are considered:

22
 (i) The positiveordered Lognormal case > 1. From (18) it is

Mode X = e < e = Med X . (66)

For the LogLaplace case of X+ , it holds

Mode X = e = Med X , (67)

provided that < 1, while for 1 we have

Mode X = 0 < e = Med X . (68)

For = 1, the inequality (66) clearly holds.

(ii) The negativeordered Lognormal case < 0. From Proposition 4 the inequality as in
(68) holds. Moreover, if Mode X is considered as the nonsmooth local mode point of
the negativeordered X then the equality as in (67) holds.

From the above cases and (65), Corollary holds true.

Corollary 3.2. The raw and central moments of a LogUniformly distributed random variable
X LU(a, b), 0 < a < b, are given by

(t) bt at
X = , t R, (69)
t log ab

k1 ( ) ( )
(k) (a b)k 1 k (a b)n bkn akn
X = + , k N, (70)
logk ab log ab n=0 n (k n) logn ab

respectively, while the mean, variance, coecient of variation, skewness and kurtosis of X are
given, respectively, by
ba
X = , (71)
log ab

2 (b a)2 (b a)(b + a)
X = 2 b + , (72)
log a 2 log ab

b+a
CVX = 1 + 2(ba) log ab , (73)
[ ]
1 b3 a3 (b a)2 (b + a) (b a)3
X = 3 3 +2 , (74)
X 3 log ab 2 log2 ab log3 ab

23
 [ ]
1 b4 a4 (b a)(b3 a3 ) (b a)3 (b + a) (b a)4
X = 4 4 +3 3 . (75)
X 4 log ab 3 log2 ab log3 ab log4 ab

Proof. Recall Proposition 3.1 with X LN (, ). Through the gamma function additive iden-
tity (47) can be written as

et ( )
(t) (t)2m+1 2m 1
X = ( 1 ) (2m + 1) 1 +1 .
t ( 1
+ 1) m=0 (2m + 1)!

Thus, letting X := X1 LN 1 (, ) = LU(a, b) with = 1

2 log(ab) and = 1
2 log ab , it holds (recall
the exponential oddseries expansion) that

(t) (t) et (t)2m+1 et(+) et()
X = lim+ X = = , t R.
1 t m=0 (2m + 1)! 2t

(1)
and hence (69) holds. Moreover, X := X = E[X] = 1
2 (e
+
e ) and therefore (71) holds.
Working similarly, (52) implies

k ( )
(k) (k) k ( X )n [(k n)]2m
X = lim X =e k
e , k N.
(2m + 1)!
n=0 n
1+
m=0

Using the exponential odd series expansion, the above expansion become

k ( )

k k ( )n e(kn) e(kn)
= e2
(k)
X eX , k N,
(k n)

n=0 n

(2)
2
and, through (71), we obtain (70). Moreover, for k = 2, then X := Var[X] = X 2X implies
(3) (4)
(72) and hence (73) also holds. For k = 3 and k = 4, through X and X , we obtain (74) and
(75) respectively.

Corollary 3.3. The raw and central moment of a LogLaplace distributed random variable X
LL(, , ), are given by
(t) t 2
X = > t , > t, t R, (76)
2 t2
k ( )

(k) k 2(n+1)
X = k
, > k, k N. (77)
n=0 n (1 2 )n [ 2 (k n)2 ]

24
The mean, variance, coecient of variation, skewness and kurtosis of X are given, respectively, by

2
X = > , > 1, (78)
2 1

2 2 2 (2 2 + 1)
X = , > 2, (79)
( 2 4)( 2 1)2

1 2 2 + 1
CVX = , > 2, (80)
2 4

2(15 4 + 7 2 + 2) 2 4
X = , > 3, (81)
( 9)
2 (2 2 + 1)3

3(8 8 + 212 6 + 95 4 + 33 2 + 12)( 2 4)

X = , > 4. (82)
( 2 16)( 2 9)(2 2 + 1)2
Proof. Let X LL (, , ) = LN (log , 1/, 1/). For = , i.e.
1 = 1, the raw
moments as in (47), provide

(t) (t)
X = X = t (t/)2k , t R,
k=0

as X = X , while through the even geometric series expansion, it is

[ ]


( )
(t/)k = 12 t
(t)
X = 12 t (t/)k + t +t ,
+
k=0 k=0

(1)
provided that > t and hence (76) holds. Moreover, X := X = E[X] and hence (78) holds.
Working similarly, (52) implies
( )n
k ( )
X
(k) k
X = k 2 , k N,
n=0 n 2 (k n)2

provided > k and hence, through (78), the central moments (77) are obtained.
(2)
2
Moreover, for k = 2 and due to X := Var[X] = X 2X , (79) holds true, while for k = 3 and
(3) (4)
k = 4 we derive, through X and X , (81) and (82) respectively.

Example 3.4. For a uniformly distributed r.v. U U(a, b) = N1 (, ) with a = and

b = +, it holds that LU := eU LU(e , e+ ) due to Theorem 2.1, and therefore LU is a Log
Uniform distributed r.v. as LU LU(ea , eb ). Applying (69) the known moment generating function

25
 (t)
of the uniformly distributed U U(a, b) are derived, i.e. MU (t) := E[etU ] = LU = (etb eta ) t(ba)
1
.
Similarly, for a Laplace distributed r.v. L L(, ) = N (, ), it holds that LL := eL
LL(e , 1/, 1/) due to Theorem 2.1, and therefore LL is a LogLaplace distributed random vari-
able. Applying (76) we derive the known moment generating function of the Laplace distributed
L L(, ), i.e. ML (t) := E[etL ] = LL = et (1 t2 2 )1 .
(t)

4 Conclusion
The family of the order Lognormal distributions was introduced, which under certain values of
includes the LogUniform, Lognormal, LogLaplace as well as the degenerate Dirac distributions.
The shape of these distributions for positive and negative shape parameters as well as the
cumulative distribution functions, were extensively discussed and evaluated through corresponding
Tables and Figures. Moreover, a thorough study of moments was carried out, in which non-
closed forms as well as approximations were obtained and investigated in various examples. This
generalized family of distributions derived through the family of the order Normal distribution, is
based on a strong theoretical background as the Logarithmic Sobolev Inequalities provide. Further
examinations and calculations can be produced while an application to real data is upcoming.

Acknowledgments
The authors would like to thank the referee for his valuable comments that helped improve the
quality of this paper.

References
[1] Bernardeau, F. & Kofman, L. (1995). Properties of the cosmological density distribution
function. Monthly Notices of the Royal Astrophys. J., 443, 479498.

[2] Blasi, P., Burles, S. & Olinto, A. V. (1999). Cosmological magnetic eld limits in an inhomo-
geneous Universe. The Astrophysical Journal Letters 514(2), L79L82.

[3] Crow, E. L., & Shimizu, K. (1998). Lognormal Distributions - Theory and Applications, M.
Dekker, New York, NY & Basel.

[4] Geraci, M., & Bottai, M. (2007). Quantile regression for longitudinal data using the asym-
metric Laplace distribution. Biostatistics 8(1), 140154.

26
 [5] Gradshteyn, I. S. & Ryzhik, I. M. (2007). Table of Integrals, Series, and Products. Elsevier.

[6] Kitaura, F. S. (2012). Non-Gaussian gravitational clustering eld statistics. Monthly Notices
of the Royal Astronomical Society (to appear).

[7] Kitsos, C. P. & Tavoularis, N. K. (2009a). Logarithmic Sobolev inequalities for information
measures. IEEE Trans. Inform. Theory, 55(6), 25542561.

[8] Kitsos, C. P. & Tavoularis, N. K. (2009b). New entropy type information measures. Proceed-
ings of the Information Technology Interfaces (ITI 2009), Cavtat, Croatia, June 2225.

[9] Kitsos, C. P., Toulias, T. L. & Trandar, C. P. (2012). On the multivariate ordered normal
distribution. Far East J. Theor. Stat. 38(1), 4973.

[10] Kitsos, C. P. & Toulias, T. L. (2011). On the multivariate ordered normal distribution. Far
East J. Theor. Stat. 35(2), 95114.

[11] Kitsos, C. P. & Toulias, T. L. (2010a). New information measures for the generalized normal
distribution. Information 1, 1327.

[12] Kitsos, C. P. & Toulias, T. L. (2010b). Evaluating information measures for the order Mul-
tivariate Gaussian. Proceedings by IEEE of the 14th Panhellenic Conference on Informatics
(PCI 2010), 153157, Tripoli, Greece, Sept. 1012.

[13] Kozubowski, T. J. & Podgorski, K. (2001). Asymmetric Laplace laws and modeling nancial
data. Math. Comput. Modelling 34, 10031021.

[14] Kozubowski, T. J., & Podgorski, K. (2000). Asymmetric Laplace distributions. Math. Sci. 25,
3746.

[15] Leone, F. C., Nelson, L. S., & Nottingham, R. B. (1961). The folded normal distribution.
Technometrics 3(4), 543550.

[16] Parravano, A., Sanchez, N. & Alfaro, E. J. (2012). The Dependence of prestellar core mass
distributions on the structure of the parental cloud. The Astrophysical Journal (to appear).

[17] Madan, D. B. (1998). The variance gamma process and option pricing. European Finance
Rev. 2, 79105.

[18] Yan, G. & Hanson, F. B. (2006). Option pricing for a stochasticvolatility jump-diusion
model with loguniform jumpamplitudes. Proceedings of the American Control Conference.

99

27