Hindawi Publishing Corporation Journal of Probability and Statistics Volume 2013, Article ID 432642, 15 pages http://dx.doi.org/10.1155/2013/432642
Research Article On the Generalized Lognormal Distribution Thomas L. Toulias and Christos P. Kitsos Technological Educational Institute of Athens, Department of Mathematics, Ag. Spyridonos & Palikaridi Street, 12210 Egaleo, Athens, Greece Correspondence should be addressed to Thomas L. Toulias;
[email protected] Received 11 April 2013; Revised 4 June 2013; Accepted 5 June 2013 Academic Editor: Mohammad Fraiwan Al-Saleh Copyright Β© 2013 T. L. Toulias and C. P. Kitsos. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. This paper introduces, investigates, and discusses the πΎ-order generalized lognormal distribution (πΎ-GLD). Under certain values of the extra shape parameter πΎ, the usual lognormal, log-Laplace, and log-uniform distribution, are obtained, as well as the degenerate Dirac distribution. The shape of all the members of the πΎ-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 πΎGLD are also studied.
1. Introduction Lognormal distribution has been widely applied in many different aspects of life sciences, including biology, ecology, geology, and meteorology as well as in economics, finance, and risk analysis, see [1]. Also, it plays an important role in Astrophysics and Cosmology; see [2β4] among others, while for Lognormal expansions see [5]. In principle, the lognormal distribution is defined as the distribution of a random variable whose logarithm is normally distributed, and usually it is formulated with two parameters. Furthermore, log-uniform and log-laplace distributions can be similarly defined with applications in finance; see [6, 7]. Specifically, the power-tail phenomenon of the Log-Laplace distributions [8] attracts attention quite often in environmental sciences, physics, economics, and finance as well as in longitudinal studies [9]. Recently, Log-Laplace distributions have been proposed for modeling growth rates as stock prices [10] and currency exchange rates [7]. 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, which combines theoretically all the properties of Lognormal, Log-Uniform, and Log-Laplace distributions, depending on the value of this third parameter.
The generalized πΎ-order Lognormal distribution (πΎ-GLD) is the distribution of a random vector whose logarithm follows the πΎ-order normal distribution, an exponential power generalization of the usual normal distribution, introduced by [11, 12]. This family of π-dimensional generalized normal distributions, denoted by NππΎ (π, Ξ£), is equipped with an extra shape parameter πΎ and constructed to play the role of normal distribution for the generalized Fisherβs entropy type of information; see also [13, 14]. The density function ππ of a π-variate, πΎ-order, normally distributed random variable π βΌ NππΎ (π, Ξ£), with location vector π β Rπ , positive definite scale matrix Ξ£ β RπΓπ , and shape parameter πΎ β R \ [0, 1], is given by [11]. ππ (π¦) = ππ (π¦; π, Ξ£, πΎ) = πΆπΎπ |det Ξ£|β1/2 exp {β
πΎβ1 πΎ/2(πΎβ1) }, ππ (π¦) πΎ
(1)
π¦ β Rπ , where ππ is the quadratic form ππ (π¦) = (π¦ β π)T Ξ£β1 (π¦ β π), π = (π, Ξ£), while πΆπΎπ being the normalizing factor πΆπΎπ = πβπ/2
Ξ (π/2 + 1) πΎ β 1 π((πΎβ1)/πΎ)β1 . ( ) πΎ Ξ (π ((πΎ β 1) /πΎ))
(2)
2
Journal of Probability and Statistics
From (1), notice that the second-ordered normal is the π known multivariate normal distribution; that is, N2 (π, Ξ£) = π N (π, Ξ£); see also [13, 15]. In Section 2, a generalized form of the Lognormal distribution is introduced, which is derived from the univariate family of NπΎ (π, π2 ) = N1πΎ (π, π2 ) distributions, denoted by LNπΎ (π, π), and includes the Log-Laplace distribution as well as the Log-Uniform distribution. The shape of the LNπΎ (π, π) members is extensively discussed while it is connected to the tailing behavior of LNπΎ 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 Log-Laplace distributions, is presented. The generalized error function, that is briefly provided here, plays an important role in the development of LNπΎ (π, π); see Section 2. The generalized error function denoted by Erfπ and the generalized complementary error function Erfcπ = 1 β Erfπ , π β₯ 0 [16], are defined, respectively, as Erfπ (π₯) :=
Ξ (π + 1) π₯ βπ‘π β« π ππ‘, βπ 0
π₯ β R.
(3)
The generalized error function can be expressed (changing to π‘π variable), through the lower incomplete gamma function πΎ(π, π₯) or the upper (complementary) incomplete gamma function Ξ(π, π₯) = Ξ(π) β πΎ(π, π₯), in the form Erfπ (π₯) =
1 Ξ (π) Ξ (π) 1 π 1 πΎ( ,π₯ ) = [Ξ ( ) β Ξ ( , π₯π )] , βπ βπ π π π π₯ β R; (4)
see [16]. Moreover, adopting the series expansion form of the lower incomplete gamma function, π₯
β
(β1) π₯π+π , π! + π) (π π=0
πΎ (π, π₯) := β« π‘πβ1 πβπ‘ ππ‘ = β 0
π
π₯, π β R+ , (5)
a series expansion form of the generalized error function is extracted: Erfπ (π₯) =
Ξ (π + 1) β (β1)π π₯ππ+1 , β βπ π=0 π! (ππ + 1)
π₯, π β R+ . (6)
Notice that, Erf2 is the known error function erf, that is, Erf2 (π₯) = erf(π₯), while Erf0 is the function of a straight line through the origin with slope (πβπ)β1 . Applying π = 2, the known incomplete gamma function identities such as πΎ(1/2, π₯) = βπ erf βπ₯ and Ξ(1/2, π₯) = βπ(1 β erf βπ₯) = βπ erfc βπ₯, π₯ β₯ 0 are obtained. Moreover, Erfπ 0 = 0 for all π β R+ and lim Erfπ π₯ = Β±
π₯ β Β±β
1 1 Ξ (π) Ξ ( ) , βπ π
as πΎ(π, π₯) β Ξ(π) when π₯ β +β.
π β R+ ,
(7)
2. The πΎ-Order Lognormal Distribution The generalized univariate Lognormal distribution is defined, through the univariate generalized πΎ-order normal distribution, as follows. Definition 1. When the logarithm of a random variable π follows the univariate πΎ-order normal distribution, that is, log π βΌ NπΎ (π, π2 ), then π is said to follow the generalized Lognormal distribution, denoted by LNπΎ (π, π); that is, π βΌ LNπΎ (π, π). The LNπΎ (π, π) is referred to as the (generalized) πΎ-order Lognormal distribution (πΎ-GLD). Like the usual Lognormal distribution, the parameter π β R is considered to be log-scaled, while the non log-scaled π (i.e. ππ when π is assumed log-scaled) is referred to as the location parameter of LNπΎ (π, π). Hence, if π βΌ LNπΎ (π, π), then log π is a πΎ-order normally distributed variable; that is, log π βΌ NπΎ (π, π2 ). Therefore, the location parameter π β R of π is in fact the mean of πβs natural logarithm, that is, E[log π] = π, while Var [log π] = (
πΎ 2((πΎβ1)/πΎ) Ξ (3 ((πΎ β 1) /πΎ)) 2 ) π, πΎβ1 Ξ (((πΎ β 1) /πΎ))
Kurt [log π] =
Ξ (πΎ β 1/πΎ) Ξ (5 (πΎ β 1/πΎ)) ; Ξ2 (3 ((πΎ β 1)/πΎ))
(8)
(9)
see [15] for details on NπΎ . Let π := log π βΌ NπΎ (π, π2 ) with density function as in (1) and π = π(π) = ππ . Then, the density function ππ of π βΌ LNπΎ (π, π) can be written, through (1), as σ΅¨σ΅¨ π σ΅¨σ΅¨ σ΅¨ σ΅¨ ππ (π₯) = ππ (π₯; π, π, πΎ) = ππ (πβ1 (π₯)) σ΅¨σ΅¨σ΅¨ πβ1 (π₯)σ΅¨σ΅¨σ΅¨ σ΅¨σ΅¨ ππ₯ σ΅¨σ΅¨ = ππ (log π₯) =
1 π₯
(10)
σ΅¨πΎ/(πΎβ1) σ΅¨ exp {β ((πΎ β 1) /πΎ) σ΅¨σ΅¨σ΅¨(log π₯ β π) /πσ΅¨σ΅¨σ΅¨ } 2π ((πΎ β 1)/πΎ)
1/πΎ
Ξ ((πΎ β 1) /πΎ) π₯
.
The probability density function ππ , as in (10), is defined in Rβ+ = R+ \0; that is, LNπΎ (π, π) has zero threshold. Therefore, the following definition extends Definition 1. Definition 2. When the logarithm of a random variable π + π follows the univariate πΎ-order normal distribution, that is, log(π + π) βΌ NπΎ (π, π2 ), then π is said to follow the generalized Lognormal distribution with threshold π β R; that is, π βΌ LNπΎ (π, π; π). It is clear that when π βΌ LNπΎ (π, π; π, ), log(π β π) is a πΎ-order normally distributed variable, that is, log(π β π) βΌ NπΎ (π, π2 ), and thus, π is the mean of (π β π)βs natural logarithm while Var[log π] is the same as in (8). Let π = log(π + π) βΌ NπΎ (π, π2 ). The density function of π = ππ β π βΌ LNπΎ (π, π; π) is given by ππ (π₯) = ππ(π₯ β π), π₯ > 0.
Journal of Probability and Statistics
3
Let π§ = (log(π₯ β π) β π)/π. Then, the limiting threshold density value of ππ (π₯) with π₯ β π+ implies that lim ππ (π₯)
π₯ β π+
= πβ1 πΆπΎ1 lim exp {βπ§ (π + π§ β ββ
π πΎ β 1 1/(πΎβ1) )} (11) β |π§| π§ πΎ
= πβ1 πΆπΎ1 π(sgn πΎ)(ββ) ,
(ii) The βnormalβ case πΎ = 2: it is clear that LN2 (π, π) = LN(π, π), as ππ2 coincides with the Lognormal density function, and therefore the second-ordered Lognormal distribution is in fact the usual Lognormal distribution. (iii) The limiting case πΎ = Β±β: we have LNΒ±β (π, π) := limπΎ β Β±β LNπΎ (π, π) with ππΒ±β (π₯) := lim πππΎ (π₯) = πΎ β Β±β
and therefore
σ΅¨σ΅¨ log π₯ β π σ΅¨σ΅¨ 1 σ΅¨ σ΅¨σ΅¨ exp {β σ΅¨σ΅¨σ΅¨ σ΅¨σ΅¨} σ΅¨ σ΅¨σ΅¨ 2ππ₯ π σ΅¨
βπ/π
0, πΎ β (1, +β) , lim+ ππ (π₯) = { π₯βπ +β, πΎ β (ββ, 0) ;
π { (1βπ)/π { , π₯ β (0, ππ ] , { { 2π π₯ = { π/π { {π { π₯β(π+1)/π , π₯ > ππ , { 2π
(12)
(15)
that is, the ππ βs defining domain, for the positive-ordered Lognormal random variable π, can be extended to include threshold point π by letting ππ (π) = 0. The generalized Lognormal family of distributions LNπΎ is a wide range family bridging the Log-Uniform LU, Lognormal LN, and Log-Laplace LL distributions, as well as the degenerate Dirac D distributions. We have the following.
which coincides with the density function of the known Log-Laplace distribution (symmetric logexponential distribution) LL(πσΈ , πΌ, π½) with πσΈ = ππ and πΌ = π½ = 1/π; see [8]. Therefore, the infiniteordered log-normal distribution is in fact the LogLaplace distribution, with threshold density
Theorem 3. The generalized Lognormal distribution LNπΎ (π, π), for order values of πΎ = 0, 1, 2, Β±β, is reduced to
(16)
D (ππ ) , { { πβπ π+π { { {LU (π , π ) , LNπΎ (π, π) = {LN (π, π) , { { { 1 1 { LL (ππ , , ) , { π π
πΎ = 0, πΎ = 1, πΎ = 2,
(13)
(i) The limiting case πΎ = 1: let π₯ β Rβ+ such that | log π₯ β π| β€ 1. Using the gamma function additive identity Ξ(π§ + 1) = π§Ξ(π§), π§ β R+ , in (10), we have LN1 (π, π) = limπΎ β 1+ LNπΎ (π, π) with ππ1 (π₯) := lim+ πππΎ (π₯) πΎβ1
π₯ β [ππβπ , ππ+π ] ,
For the purposes of statistical application, the LogLaplace moments are not the same as the model parameters; that is, although ππ = π, ππ = β2π. (iv) The limiting case πΎ = 0: we have
πΎ = Β±β.
Proof. From definition (1) of NπΎ the order πΎ value is a real number outside the closed interval [0, 1]. Let ππΎ βΌ LNπΎ (π, π) with density function πππΎ as in (10). We consider the following cases.
{ 1 , = { 2ππ₯ {0,
0, π < 1, { { ππΒ±β (0) = lim+ ππΒ±β (π₯) = {1, π = 1, { π₯β0 {+β, π > 1.
(14)
π₯ β (ββ, ππβπ ) βͺ (ππ+π , +β) ,
which is the density function of the Log-Uniform distribution LU(π, π), 0 < π < π, with π = ππβπ and π = ππ+π ; that is, π = (1/2) log(ππ) and π = (1/2) log(π/π). Therefore, first-ordered Lognormal distribution is in fact the Log-Uniform distribution, with vanishing threshold density, ππ1 (0) = limπ₯ β 0+ ππ1 (π₯) = 0. For the purposes of statistical application the LogUniform moments are not the same as the model parameters; that is, although ππ = π, ππ = π/β3.
lim πππΎ (π₯) =
πΎ β 0β
lim
πππΎ (π₯) ,
π:=[(πΎβ1)/πΎ] β β
π₯ β Rβ+ ,
(17)
where [π] is the integer value of π β R. For the value π₯ = ππ , the p.d.f. as in (10) implies limβ πππΎ (ππ ) =
πΎβ0
ππ 1 ( lim ) β
π0 = +β, 2πππ π β β π!
(18)
through Stirlingβs asymptotic formula π! β β2ππ(π/ π)π as π β β. Assuming now π₯ =ΜΈ ππ , (10), through (18), implies σ΅¨ σ΅¨σ΅¨ σ΅¨log π₯ β πσ΅¨σ΅¨σ΅¨ 1 β
0 β
= 0; ππ0 (π₯) := limβ πππΎ (π₯) = σ΅¨ 2 πΎβ0 π 2β2ππ π₯
(19)
that is, LN0 (π, π) := limπΎ β 0β LNπΎ (π, π) = D(ππ ) as ππ0 coincides with the Dirac density function, with the (non-log-scaled) location parameter ππ of LN0 (π, π) being the singular (infinity) point. Therefore, the zero-ordered Lognormal distribution LN0 is in fact the degenerate Dirac distribution with pole at the location parameter of LNπΎ β 0β (with vanishing threshold density ππ0 (0) = limπ₯ β 0+ ππ0 (π₯) = 0). From the above limiting cases (i), (iii), and (iv), the defining domain R \ [0, 1] of the order values πΎ, used in (1), is safely
4
Journal of Probability and Statistics
extended to include the values πΎ = 0, 1, Β±β; that is, πΎ can now be defined outside the open interval (0, 1). Eventually, the family of the πΎ-order normals can include the LogUniform, Lognormal, Log-Laplace, and the degenerate Dirac distributions as (13) holds. From Theorem 3, (12), and (15), the domain of the density functions πππΎ (π₯), π₯ > 0, can also be extended to include the threshold point π₯ = 0 by setting πππΎ (0) := 0 for all nonnegative-ordered Lognormals, that is, for all πΎ β 0 βͺ [1, +β), while for the Log-Laplace case of πΎ = +β with π = 1 by setting ππ+β (0) := (1/2)πβπ . From the fact that N0 (π, β
) = D(π), see [15], one can say that the degenerate log-Dirac distribution, say LD(π), equals LN0 (π, β
), and hence, through Theorem 3, we can write LD(π) = D(ππ ). Proposition 4. The mode of the positive-ordered Lognormal random variable ππΎ βΌ LNπΎ (π, π), π < ππ πΎ β (1, +β), is given by πΎ
Mode ππΎ = ππβπ ,
(20)
with corresponding maximum density value, max πππΎ = πππΎ (Mode ππΎ ) =
exp {(ππΎ ) /πΎ β π} 1/πΎ
2π((πΎ β 1)/πΎ)
Ξ ((πΎ β 1) /πΎ)
.
(21)
Proof. Recall the density function of ππΎ βΌ LNπΎ (π, π) as in (10), and let π = Mode ππΎ > 0. Then it holds that (π/ππ₯) πππΎ (π; π, π, πΎ) = 0; that is, 1/(πΎβ1) π σ΅¨σ΅¨ π 1σ΅¨ σ΅¨ σ΅¨ ( σ΅¨σ΅¨σ΅¨log π β πσ΅¨σ΅¨σ΅¨) σ΅¨σ΅¨log π₯ β πσ΅¨σ΅¨σ΅¨π₯=π = β . π ππ₯ π From (π/ππ₯)|π₯| = sgn π₯, π₯ β R, we have πΎ
π = ππβπ ,
(22)
provided that π₯ < π . Otherwise, (23) holds trivially, as (22) implies π = 0; that is, π = ππ . Moreover, (π/ππ₯)πππΎ (π₯) > 0 when σ΅¨1/(πΎβ1) σ΅¨ 1 + sgn (log π₯ β π) ππΎ/(1βπΎ) σ΅¨σ΅¨σ΅¨log π₯ β πσ΅¨σ΅¨σ΅¨ > 0, π₯ > 0, (24) and thus, πππΎ is a strictly ascending density function on πΎ
(0, ππβπ ) when πΎ > 1 and also on (ππβπ , ππ ) when πΎ < 0. Similarly, with (π/ππ₯)πππΎ (π₯) < 0, ππ is a strictly descending πΎ
density function on (ππβπ , +β) when πΎ > 1 and also on πΎ (0, ππβπ ) βͺ (ππ , +β) when πΎ < 0. Specifically, for πΎ < 0, the π point π is a nonsmooth point of ππ, as lim
π
π₯ β ππ ππ₯
πππΎ (π₯) = β
πΆπΎ1 (πππ )2
Proposition 5. The global mode point of the negative-ordered Lognormal random variable ππΎ βΌ LNπΎ (π, π), πΎ β (ββ, 0), is (in limit) the threshold 0 which is an infinite (probability) density point. Moreover, the location parameter (i.e., ππ ) of LNπΎ (π, π) is a nonsmooth (local) mode point for all ππΎ 1,
with the corresponding maximum density value being infinite; that is, max πππΎ = πππΎ (0) = +β, provided π < 1, and
Journal of Probability and Statistics
eβ2/3
e2/3
1.5 fXπΎ βΌ π©πΎ 1 is decreasing from ππβπ |πΎ β ββ = ππ
Proposition 6. Consider the positive-ordered Lognormal family of distributions LNπΎ (π, π) with fixed parameters π, π and πΎ β₯ 1. When πΎ rises, that is, when one moves from Log-Uniform to Log-Laplace distribution inside the LNπΎ family, the mode points of LNπΎ are
It is easy to see that for the Log-Laplace case LL(π, π, π), πΎ the local minimum density point ππβπ of ππΎ 1 coincides (in limit) with the local nonsmooth mode point ππ of ππΎ ; see Figure 1(b3). Also, notice that the local minimum
(i) strictly increasing from ππβπ (Log-Uniform case) to ππ (Log-Laplace case) provided that π < 1 (with their corresponding maximum density values moving smoothly from (1/2π)ππβπ to +β), (ii) fixed at ππβ1 for all LNπΎβ₯1 (π, π = 1) (with the corresponding maximum density values moving smoothly from (1/2)π1βπ to (1/2)πβπ ), (iii) strictly decreasing from ππβπ (Log-Uniform case) to threshold 0 (Log-Laplace case) provided that π > 1 (with their corresponding maximum density values moving smoothly from (1/2π)ππβπ to (1/2π)πβπ ). Proof. Let ππΎ βΌ LNπΎ (π, π), Mode ππΎ is a smooth monotonous function of πΎ β (ββ, 0) βͺ (1, +β) for positive-and negative-ordered ππΎ , as πΎ π Mode ππΎ = βππΎ log (π) ππβπ . ππΎ
(29)
For πΒ±β βΌ LNΒ±β (π, π) we evaluate, through (20) and (21), that Mode π+β
ππ , { { πβ1 = {π , { {0,
π < 1, π = 1, π > 1,
(30)
with the corresponding maximum density value being infinite; that is, max πππΎ = πππΎ (0) = +β, provided π > 1, and max ππ+β = 1/(2πππ ), provided π β€ 1. Assume that πΎ β₯ 1. Considering (27), (30) with (29) and Proposition 4, the results for the positive-ordered Lognormals hold. Proposition 7. For the negative-ordered Lognormal family of distributions LNπΎ (π, π) with πΎ < 0, when πΎ rises, that is, when one moves from Log-Laplace to degenerate Dirac distribution inside the LNπΎ family, the local minimum (probability) density points of LNπΎ are (i) strictly increasing from threshold 0 (Log-Laplace case) to ππβ1 (Dirac case) provided that π < 1, (ii) fixed at ππβ1 for all LNπΎ (π, π = 1), (iii) strictly decreasing from ππ (Log-Laplace case) to ππβ1 (Dirac case) provided that π > 1.
πΎ
to Mode π0 = ππβ1 through (20). When π = 1, ππβπ = πΎ ππβ1 for all πΎ < 0, while for π < 1, ππβπ increases from πΎ ππβπ |πΎ β ββ = 0 to Mode π0 = ππβ1 through (20).
πΎ
density point ππβπ , πΎ < 0, for the Dirac case D(ππ ), is the limiting point ππβ1 although the (probability) density in D(ππ ) case vanishes everywhere except at the infinite pole ππ . Figure 1 illustrates the probability density functions πππΎ curves for scale parameters π = 2/3, 1, 3/2 of the positiveordered lognormally distributed ππΎ βΌ LNπΎβ₯1 (0, π) in Figures 1(a1)β1(a3), respectively, while the p.d.f. of negativeordered lognormally distributed ππΎ βΌ LNπΎ βππ , into the other two integrals, we obtain 0
πΉ|π| (π₯) = πΆπΎ1 β«
ββ
Γβ«
exp {β
(1/π) log 2
0
πΎ β 1 πΎ/(πΎβ1) } ππ€ + πΆπΎ1 |π€| πΎ exp {β
π(π₯)
+ πΆπΎ1 β«
(1/π) log 2
π (π₯) :=
π
log (π₯ + π ) β π , π
πΎ β 1 πΎ/(πΎβ1) } ππ§ |π§| πΎ
exp {β
πΎ β 1 πΎ/(πΎβ1) } ππ§, |π§| πΎ
π₯ β R+ , (56)
(53) and hence
where ππ is the p.d.f. of π. For example, see [17] on the folded normal distribution. However, the density function ππ is defined in (βππ , +β) due to threshold βππ , while it vanishes elsewhere; that is, π (βπ₯) + ππ (π₯) , 0 β€ π₯ β€ ππ , π|π| (π₯) = { π ππ (π₯) , π₯ > ππ .
(54)
πΉ|π| (π₯) = Ξ¦π (0) + [Ξ¦π (
log 2 ) β Ξ¦π (0)] π
+ [Ξ¦π (π (π₯)) β Ξ¦π ( = Ξ¦π (
log 2 )] π
log (π₯ + ππ ) β π ), π
(57)
10
Journal of Probability and Statistics eβ2/3
e2/3
0.9
0.9
0.8
0.8
0.7
0.7
0.6 0.5 0.4 0.3
0.5 0.4 0.3 0.2
0.1
0.1 1
2
x
X1 βΌ βπ°(eβ2/3 , e2/3 ) X1.1,1.2,...,1.9 X2 βΌβπ©(0, 2/3) X3,4,...,10 X+β βΌ ββ(1, 3/2, 3/2)
0
3
e1
0.6
0.2 0 0
eβ1
1
FXπΎ βΌ π©πΎ (0, 1)
FXπΎ βΌ π©πΎ (0, 2/3)
1
1
0
x
X1 βΌ βπ°(eβ1 , e1 ) X1.1,1.2,...,1.9 X2 βΌβπ©(0, 1) X3,4,...,10 X+β βΌ ββ(1, 1, 1)
Xββ βΌ ββ(1, 3/2, 3/2) Xβ10,β9,...,β1 Xβ0.9,β0.8,...,β0.1
2
3
Xββ βΌ ββ(1, 1, 1) Xβ10,β9,...,β1 Xβ0.9,β0.8,...,β0.1
(b)
(a) eβ3/2
1 0.9
FXπΎ βΌ π©πΎ (0, 3/2)
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
1
2
3
x X1 βΌ βπ°(eβ3/2 , e3/2 ) X1.1,1.2,...,1.9 X2 βΌβπ©(0, 3/2) X3,4,...,10 X+β βΌ ββ(1, 2/3, 2/3)
Xββ βΌ ββ(1, 2/3, 2/3) Xβ10,β9,...,β1 Xβ0.9,β0.8,...,β0.1
(c)
Figure 2: Graphs of the c.d.f. πΉππΎ , ππΎ βΌ LNπΎ (0, π), for π = 2/3, 1, 3/2, and various πΎ values.
with Ξ¦π being the c.d.f. of the standardized r.v. π βΌ NπΎ (0, 1). From (38) and the fact that Erfπ 0 = 0, π β Rβ+ , it is clear that β1 (57) implies MADππΎ = Med|ππΎ β ππ | = πΉ|π π (1/2) = 0, for πΎ βπ | every ππΎ βΌ LNπΎ (π, π), and the theorem has been proved.
3. Moments of the πΎ-Order Lognormal Distribution For the evaluation of the moments of the generalized Lognormal distribution, the following holds.
Journal of Probability and Statistics
11
(π‘) Proposition 16. The π‘th raw moment πΜπ of a generalized lognormally distributed random variable π βΌ LNπΎ (π, π) is given by
Γ Ξ ((2π + 1)
π
2 2 ππ‘π β (2π‘ π ) 1 = Ξ (π + ) , β βπ π=0 (2π)! 2
(π‘) πΜπ
β πΎ 2π((πΎβ1)/πΎ) ππ‘π (π‘π)2π = ) ( β Ξ ((πΎ β 1) /πΎ) π=0 (2π)! πΎ β 1
(π‘) πΜπ
Example 17. For the second-ordered lognormally distributed π βΌ LN2 (π, π), (58) implies
(58)
πΎβ1 ) πΎ
and coincides with the moment generating function of the πΎ(π‘) order normally distributed log π; that is, πlog π (π‘) = πΜπ .
R+
=
πΎ β 1 σ΅¨σ΅¨σ΅¨σ΅¨ log π₯ β π σ΅¨σ΅¨σ΅¨σ΅¨πΎ/(πΎβ1) 1 1 } ππ₯, πΆπΎ β« π₯π‘β1 exp {β σ΅¨ σ΅¨σ΅¨ σ΅¨σ΅¨ π πΎ σ΅¨σ΅¨σ΅¨ π R+ (59) (πΎβ1)/πΎ
and applying the transformation π§ = ((πΎ β 1)/πΎ) (log π₯ β π), π₯ > 0, we get (π‘) πΜπ
=
πΆπΎ1 (
(1/π)
πΎ (πΎβ1)/πΎ πΎ (πΎβ1)/πΎ ππ§} ) ) β« exp {π‘π + π( πΎβ1 πΎβ1 R
1 (2π)! Ξ (π + ) = 2π βπ, 2 2 π!
(60) Through the exponential series expansion exp {π‘(
πΎ ) πΎβ1
β
π((πΎβ1)/πΎ)
π
πΎ (π‘π) ( ) π! πΎ β 1 π=0
ππ§} = β
π§π , (61)
we have β
β 1 1 2 2 π (π‘π)2π π‘π = π ( π‘π) β π π=0 2 π! π=0 π! 2
(π‘) πΜπ = ππ‘π β
π‘π+(1/2)(π‘π)2
=π
π‘ β R+ ,
which is the π‘th raw moment of the usual lognormally (1) distributed π βΌ LN(π, π), with mean ππ := πΜπ = E[π] = (π‘) 2 exp{π + (1/2)π }. This is true as πlog π (π‘) = πΜπ = exp{π‘π + (1/2)(π‘π)2 } is the known moment-generating function of the normally distributed log π βΌ N(π, π2 ). (π‘) Theorem 18. The πth central moment (about the mean) ππ of a generalized lognormally distributed random variable π βΌ LNπΎ (π, π) is given by (π) ππ =
π π π πππ π ) ππβπ , β ( ) (β π ππ Ξ ((πΎ β 1) /πΎ) π=0 π
π β N, (67)
πΎ 2π((πΎβ1)/πΎ) πΎβ1 (ππ)2 π ( Ξ ((2 π + 1) ) ), πΎβ1 πΎ π=0 (2 π)! β
ππ = β
π β N. (68) (π) Proof. From the definition of the πth central moment ππ we have
β
2π
π
πΎ (π‘π) ( ) πΎ β 1 (2π)! π=0
ππ‘π β
(69)
2π((πΎβ1)/πΎ)
while using the binomial identity we get π
π π (π) ππ = β ( ) (βππ ) β« π₯πβπ ππ (π₯) ππ₯ π R+ π=0
Γ β« π§2π exp {βπ§πΎ/(πΎβ1) } ππ§. R+
(62) Finally, substituting the normalizing factor πΆπΎ1 as in (2) into (62) and utilizing the known integral [16],
R+
(66)
R+
πΎ ) πΎβ1
β« π₯π πβππ₯ ππ₯ =
,
π
(πΎβ1)/πΎ
π
(65)
(π) := E [(π β ππ ) ] = β« (π₯ β ππ ) ππ (π₯; π, π, πΎ) ππ₯, ππ
it is obtained that (π‘) = 2πΆπΎ1 ( πΜπ
π β N,
where
Γ exp {β|π§|πΎ/(πΎβ1) } ππ§.
(πΎβ1)/πΎ
(64)
and through the gamma function identity
(π‘) , we Proof. From the definition of the π‘th raw moment πΜπ have (π‘) = E [ππ‘ ] = β« π₯π‘ ππ (π₯) ππ₯ πΜπ
π‘ β R+ ,
Ξ ((π + 1) /π) , ππ(π+1)/π
π, π, π β Rβ+ ,
(63)
we obtain (58). Moreover, for π := log π βΌ NπΎ (π, π2 ) we have ππ (π‘) = (π‘) E[ππ‘π ] = E[ππ‘ ] = πΜπ , and the proposition has been proved.
π
(70)
π π (πβπ) . = β ( ) (βππ ) πΜπ π π=0 Applying Proposition 16, (70) implies that (π) ππ =
π π π β [(π β π) π]2 π πππ π ) β β ( ) (β π ππ π=0 (2 π)! Ξ ((πΎ β 1) /πΎ) π=0 π
Γ(
πΎ 2 π((πΎβ1)/πΎ) Ξ ((2 π + 1) ((πΎ β 1) /πΎ)) , ) πΎβ1 (71)
12
Journal of Probability and Statistics
while taking the summation index π until π β 1, we finally obtain (67), and the theorem has been proved. Example 19. Recall Example 17. Substituting (66) and the 2 mean ππ = ππ+(1/2)π into (70), the second-ordered lognormally distributed π βΌ LN2 (π, π) provides (π) ππ
π
2 2 π = β ( ) (β1)π πππ+(1/2)[π+(πβπ) ]π , π π=0
π β N,
(72)
2
2
2 (2) := Var [π] = ππ = π2π+π (ππ β 1) , ππ
(73)
which are the πth central moment and the variance, respectively, of the usual lognormally distributed π βΌ LN(π, π). The same result can be derived directly through (67) for πΎ = 2 and the use of the known gamma function identity, as in (65). 2 := Var[π], Theorem 20. The mean ππ := E [π], variance ππ coefficient of variation πΆππ , skewness π π and kurtosis π
π of the generalized lognormally distributed π βΌ LNπΎ (π, π) are, respectively, given by
ππ =
ππ π, Ξ ((πΎ β 1) /πΎ) 1
2 2 ππ = β ππ +
πΆππ2 = Ξ (
4π
πΎ β 1 β1 )] πΎ
3π
Γ (π π4 β 4π ππ π3 +
(81)
2 6π2π ππ π2
β
3 4ππ ππ π1 ) .
π π, Ξ ((πΎ β 1) /πΎ) 2
(75)
πΎ β 1 π2 ) 2 β 1, πΎ π1
(76) π3π π, 3 ππ Ξ ((πΎ β 1) /πΎ) 3
π
π = β πΆππβ4 β 6πΆππβ2 β 4
(77)
ππ π4π + 4 π, πΆππ ππ Ξ ((πΎ β 1) /πΎ) 4 (78)
where the sums ππ , π = 1, . . . , 4, are given by (68). Proof. From Proposition 16 we easily obtain (74), as ππ := (1) . From Theorem 18 we have πΜπ (2) ππ
=
2 ππ
πΎ β 1 β1 2π + [Ξ ( )] (π π2 β 2ππ ππ π1 ) . (79) πΎ
Hence, substituting π1 from (74), (75) holds. Moreover, the squared coefficient of variation is readily obtained via (75) and (74). By definition, skewness π π is the standardized (3) 3 third (central) moment; that is, π π := Skew[π] = ππ /ππ . Theorem 18 provides that ππ =
βπΆππβ3 3π
+
Example 21. For the second-ordered lognormally distributed π βΌ LN2 (π, π), utilizing (65) into (68) we get ππ = 2 2 βππ(π π )/2 , π β Nβ . Applying this to Theorem 20 we derive (after some algebra) 2
ππ = ππ+(1/2)π ,
2
πΎ 3 [ππ Ξ( 2π
β 1 β1 )] πΎ π
Γ (π π3 β 3π ππ π2 + 3π ππ π1 ) .
2
2 ππ = π2π+π (ππ β 1) ,
πΆππ = βππ β 1, 2
2
2
π π = (ππ + 2) βππ β 1,
2
which are the mean, variance, coefficient of variation, skewness, and kurtosis, respectively, of usual lognormally distributed π βΌ LN(π, π). For the usual lognormally distributed random variable π βΌ LN, it is known that Mode π < Med π < ππ . The following corollary examines this inequality for the LNπΎ family of distributions. Corollary 22. For the πΎ-ordered lognormally distributed ππΎ βΌ LNπΎ (π, π), it is true that Mode ππΎ β€ Med ππΎ = (ππ )ππΎ β€ πππΎ . The first equality holds for the Log-Laplace distributed π+β with π < 1 as well as for all the negative-ordered ππΎ 1: from (20) we have πΎ
Mode ππΎ = ππβπ < ππ = Med ππΎ . (80)
2
π
π = π4π + 2π3π + 3π2π β 3, (82)
2
(74)
2π
π π = β πΆππβ3 β πΆππβ1 +
:=
4 Ξ( π
π = πΆππβ4 + [ππ
Substituting ππ , π = 1, 2, 3, from (74), (75), and (77), we obtain (78).
while
2 ππ
Substituting π1 and π2 from (74) and (75), we obtain (77). Finally, kurtosis π
π is (by definition) the standardized fourth (4) 4 (central) moment; that is, π
π := Kurt[π] = ππ /ππ , which provides, through Theorem 18, that
(84)
For the Log-Laplace case of π+β , it holds Mode ππΎ = ππ = Med ππΎ ,
(85)
Journal of Probability and Statistics
13
provided that π < 1, while for π β₯ 1 we have π
Mode ππΎ = 0 < π = Med ππΎ .
(86)
For π = 1, the inequality (84) clearly holds. (ii) The negative-ordered Lognormal case πΎ < 0: from Proposition 4 the inequality as in (86) holds. Moreover, if Mode ππΎ is considered as the nonsmooth local mode point of the negative-ordered ππΎ then the equality as in (85) holds. From the above cases and (83), the corollary holds true. Corollary 23. The raw and central moments of a Log-Uniformly distributed random variable π βΌ LU(π, π), 0 < π < π, are given by (π‘) = πΜπ
ππ‘ β ππ‘ , π‘ log (π/π)
(π) ππ =
(π β π)π logπ (π/π)
π‘ β R, +
Thus, letting π := π1 βΌ LN1 (π, π) = LU(π, π) with π = (1/2) log(ππ) and π = (1/2) log(π/π), it holds (recall the exponential odd series expansion) that (π‘) (π‘) = lim+ ππ = πΜπ πΎ πΎβ1
π‘ β R, (95) (1) and hence (87) holds. Moreover, ππ := ππ = E[π] = (1/2π) π+π πβπ (π β π ), and therefore (89) holds. Working similarly, (67) implies π ππ π β [(π β π) π]2π π (π) (π) ππ ππ , = lim+ ππ = π ( ) β ) (β β πΎ π πΎβ1 ππ π=0 (2π + 1)! π=0
π β N. (96)
(87)
1 log (π/π)
Using the exponential odd series expansion, the above expansion becomes
π πβπ πβπ πβ1 π (π β π) (π β π ) , Γβ( ) π π (π β π) log (π/π) π=0
π β N,
(π) ππ =
(88) respectively, while the mean, variance, coefficient of variation, skewness, and kurtosis of π are given, respectively, by πβπ , log (π/π)
(89)
(π β π)2 (π β π) (π + π) , + 2 2 log (π/π) log (π/π)
(90)
ππ = 2 ππ =
πΆππ = β 1 + ππ =
π+π π log , 2 (π β π) π
(π β π) (π3 β π3 ) 1 π4 β π4 [ β 4 4 4 log (π/π) ππ 3 log2 (π/π) (π β π)3 (π + π) (π β π)4 +3 β 3 ]. log3 (π/π) log4 (π/π)
(93)
Corollary 24. The raw and central moments of a Log-Laplace distributed random variable π βΌ LL(π, π, π) are given by (π‘) = πΜπ
πΎβ1 + 1) . πΎ (94)
ππ‘ π2 > ππ‘ , π2 β π‘2
π > π‘, π‘ β R,
π π2(π+1) π (π) , = ππ β ( ) ππ π (1 β π2 )π [π2 β (π β π)2 ] π=0
(98)
π > π, π β N. (99)
The mean, variance, coefficient of variation, skewness, and kurtosis of π are given, respectively, by ππ = 2 = ππ
β πΎ 2π((πΎβ1)/πΎ) ππ‘π (π‘π)2π+1 ) ( β π‘πΞ ((πΎ β 1) /πΎ + 1) π=0 (2π + 1)! πΎ β 1
Γ Ξ ((2π + 1)
π β N,
2 := and, through (89), we obtain (88). Moreover, for π = 2, ππ (2) 2 Var[π] = ππ β ππ implies (90), and hence (91) also holds. (3) (4) For π = 3 and π = 4, through ππ and ππ , we obtain (92) and (93), respectively.
Proof. Recall Proposition 16 with ππΎ βΌ LNπΎ (π, π). Through the gamma function additive identity (58) can be written as (π‘) = πΜπ πΎ
π π π(πβπ)π β πβ(πβπ)π πππ π π , ) β ( ) (β π 2π π=0 π ππ (π β π)
(97)
(91)
π3 β π3 1 (π β π)2 (π + π) (π β π)3 [ ], β3 +2 3 3 2 ππ 3 log (π/π) 2 log (π/π) log (π/π) (92)
π
π =
ππ‘π β (π‘π)2π+1 ππ‘(π+π) β ππ‘(πβπ) , = β π‘π π=0 (2π + 1)! 2π‘π
π2 π2 (2π2 + 1) (π2 β 4) (π2 β 1)
πΆππ = ππ =
ππ2 > π, π2 β 1
1 β 2π2 + 1 , π π2 β 4
2 (15π4 + 7π2 + 2) π (π2 β 9)
β
2
π > 1,
(100)
,
(101)
π > 2,
(102)
π > 2,
π2 β 4 3
(2π2 + 1)
,
π > 3,
(103)
14
Journal of Probability and Statistics
π
π =
3 (8π8 + 212π6 + 95π4 + 33π2 + 12) (π2 β 4) (π2
β
16) (π2
β
9) (2π2
2
+ 1)
,
(104)
π > 4. Proof. Let ππΎ βΌ LLπΎ (π, π, π) = LNπΎ (log π, 1/π, 1/π). For πΎ = Β±β, that is, πΎ/(πΎ β 1) = 1, the raw moments as in (58) provide β π‘ 2π (π‘) (π‘) π‘ = πΜπ = π ( ) , πΜπ β Β±β π π=0
Acknowledgment π‘ β R,
(105)
as π = πΒ±β , while through the even geometric series expansion, it is 1 π‘ β π‘ π β π‘ π (π‘) = [ ( + (β π ) ) ] πΜπ β β Β±β 2 π π π=0 π=0
(106)
π 1 π = ππ‘ ( + ), 2 πβπ‘ π+π‘ provided that π > π‘, and hence (98) holds. Moreover, ππ := (1) = E[π], and hence (100) holds. πΜπ Working similarly, (67) implies π
π (βππ /π) π (π) = ππ π2 β ( ) , ππ 2 π π β (π β π)2 π=0
out, in which nonclosed forms as well as approximations were obtained and investigated in various examples. This generalized family of distributions derived through the family of the πΎ-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.
π β N,
(107)
provided π > π, and hence, through (100), the central moments (99) are obtained. (2) 2 2 := Var[π] = πΜπ β ππ , Moreover, for π = 2 and due to ππ (101) holds true, while for π = 3 and π = 4 we derive, through (3) (4) and ππ , (103) and (104), respectively. ππ Example 25. For a uniformly distributed r.v. π βΌ U(π, π) = N1 (π, π) with π = π β π and π = π + π, it holds that LU := ππ βΌ LU(ππβπ , ππ+π ) due to Theorem 3, and therefore LU is a Log-Uniform distributed r.v. as LU βΌ LU(ππ , ππ ). Applying (87), the known moment-generating function of the uniformly distributed π βΌ U(π, π) is derived; that is, (π‘) = (ππ‘π β ππ‘π )(1/π‘(π β π)). ππ(π‘) := E[ππ‘π] = πΜLU Similarly, for a Laplace distributed r.v. πΏ βΌ L(π, π) = NΒ±β (π, π), it holds that LL := ππΏ βΌ LL(ππ , 1/π, 1/π) due to Theorem 3, and therefore LL is a Log-Laplace distributed random variable. Applying (98), we derive the known momentgenerating function of the Laplace distributed πΏ βΌ L(π, π); (π‘) = ππ‘π (1 β π‘2 π2 )β1 . that is, ππΏ (π‘) := E[ππ‘πΏ ] = πΜLL
4. Conclusion The family of the πΎ-order Lognormal distributions was introduced, which under certain values of πΎ includes the Log-Uniform, Lognormal, and Log-Laplace distributions as well as the degenerate Dirac distribution. The shape of these distributions for positive and negative shape parameters πΎ as well as the cumulative distribution functions, was extensively discussed and evaluated through corresponding tables and figures. Moreover, a thorough study of moments was carried
The authors would like to thank the referee for his valuable comments that helped improve the quality of this paper.
References [1] E. L. Crow and K. Shimizu, Lognormal Distributions, Marcel Dekker, New York, NY, USA, 1988. [2] A. Parravano, N. Sanchez, and E. J. Alfaro, βThe Dependence of prestellar core mass distributions on the structure of the parental cloud,β The Astrophysical Journal, vol. 754, no. 2, article 150, 2012. [3] F. Bernardeau and L. Kofman, βProperties of the cosmological density distribution function,β Monthly Notices of the Royal Astronomical Society, vol. 443, pp. 479β498, 1995. [4] P. Blasi, S. Burles, and A. V. Olinto, βCosmological magnetic field limits in an inhomogeneous Universe,β The Astrophysical Journal Letters, vol. 514, no. 2, pp. L79βL82, 1999. [5] F. S. Kitaura, βNon-Gaussian gravitational clustering field statistics,β Monthly Notices of the Royal Astronomical Society, vol. 420, no. 4, pp. 2737β2755, 2012. [6] G. Yan and F. B. Hanson, βOption pricing for a stochasticvolatility jump-diffusion model with log-uniform jumpamplitudes,β in Proceedings of the American Control Conference, 2006. [7] T. J. Kozubowski and K. PodgΒ΄orski, βAsymmetric Laplace distributions,β The Mathematical Scientist, vol. 25, no. 1, pp. 37β 46, 2000. [8] T. J. Kozubowski and K. PodgΒ΄orski, βAsymmetric Laplace laws and modeling financial data,β Mathematical and Computer Modelling, vol. 34, no. 9β11, pp. 1003β1021, 2001. [9] M. Geraci and M. Bottai, βQuantile regression for longitudinal data using the asymmetric Laplace distribution,β Biostatistics, vol. 8, no. 1, pp. 140β154, 2007. [10] D. B. Madan, βThe variance gamma process and option pricing,β The European Financial Review, vol. 2, pp. 79β105, 1998. [11] C. P. Kitsos and N. K. Tavoularis, βLogarithmic Sobolev inequalities for information measures,β IEEE Transactions on Information Theory, vol. 55, no. 6, pp. 2554β2561, 2009. [12] C. P. Kitsos and N. K. Tavoularis, βNew entropy type information measures,β in Proceedings of the Information Technology Interfaces (ITI β09), Cavtat, Croatia, June 2009. [13] C. P. Kitsos and T. L. Toulias, βNew information measures for the generalized normal distribution,β Information, vol. 1, no. 1, pp. 13β27, 2010. [14] C. P. Kitsos and T. L. Toulias, βEvaluating information measures for the -order Multivariate Gaussian,β in Proceedings by IEEE of the 14th Panhellenic Conference on Informatics (PCI β10), pp. 153β 157, Tripoli, Greece, September 2010. [15] C. P. Kitsos, T. L. Toulias, and P. C. Trandafir, βOn the multivariate πΎ-ordered normal distribution,β Far East Journal of Theoretical Statistics, vol. 38, no. 1, pp. 49β73, 2012.
Journal of Probability and Statistics [16] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Products, Elsevier, 2007. [17] F. C. Leone, L. S. Nelson, and R. B. Nottingham, βThe folded normal distribution,β Technometrics, vol. 3, pp. 543β550, 1961.
15
Advances in
Operations Research Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Advances in
Decision Sciences Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Journal of
Applied Mathematics
Algebra
Hindawi Publishing Corporation http://www.hindawi.com
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Journal of
Probability and Statistics Volume 2014
The Scientific World Journal Hindawi Publishing Corporation http://www.hindawi.com
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
International Journal of
Differential Equations Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Volume 2014
Submit your manuscripts at http://www.hindawi.com International Journal of
Advances in
Combinatorics Hindawi Publishing Corporation http://www.hindawi.com
Mathematical Physics Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Journal of
Complex Analysis Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
International Journal of Mathematics and Mathematical Sciences
Mathematical Problems in Engineering
Journal of
Mathematics Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Volume 2014
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Discrete Mathematics
Journal of
Volume 2014
Hindawi Publishing Corporation http://www.hindawi.com
Discrete Dynamics in Nature and Society
Journal of
Function Spaces Hindawi Publishing Corporation http://www.hindawi.com
Abstract and Applied Analysis
Volume 2014
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
International Journal of
Journal of
Stochastic Analysis
Optimization
Hindawi Publishing Corporation http://www.hindawi.com
Hindawi Publishing Corporation http://www.hindawi.com
Volume 2014
Volume 2014