2006, Vol. 40, No. 2, pp. 21–27 Bangladesh
ON GENERALIZED ORDER STATISTICS FROM EXPONENTIAL DISTRIBUTION
M. Ahsanullah
Department of Management Sciences
Rider University, Lawrenceville, NJ 08648-3099, USA
summary
Let X be a non-negative random variable (r.v.) with cumulative distribution function (cdf) F. SupposeX(r, n, m, k), (r= 1,2, . . . , n), m is a real number and k≥1, is the rth smallest generalized order statistic. Several distributional prop- erties of the generalized order statistics from exponential distribution are given.
Some characterizations of the exponential distribution based on the generalized order statistics are presented.
Keywords and phrases: Generalized Order Statistics, Exponential Distribution, Characterization.
1 Introduction
Let X be a random variable (r.v.) whose probability density function (pdf) f is given by
f(x) = 8
<
:
θe−θx, x >0, θ >0, 0, otherwise.
(1.1) We denote Xi∈E(θ) if the pdf of Xiis as given in (1.1). SupposeX(1, n, m, k), . . . , X(n, n, m, k), (k≥1,m >−1 is a real number), are n generalized order statistics. Their joint pdff1,2,...,n(x1, x2, ..., xn) can be written as (see Kamps 1995, pp 50 51)
f1,2,...,n(x1, x2, ..., xn) = 8
>>
><
>>
>:
kQn−1
j=1γjQn−1
i=1(1−F(xi))mf(xi)(1−F(xn))k−1f(xn) F−1(0)< x1< ... < xn< F−1(1),
0, otherwise
(1.2)
whereγj=k+ (n−j)(m+ 1) andf(x) = dxdF(x). If m = 0 and k = 1, then X(r,n,m,k) reduces to the ordinary rth order statistic and (1.2) is the joint pdf of the n order statistics,X1,n< .. < Xn,n.If k = 1 and m = -1, then (1.2) is the joint pdf of the first n upper record values of the i.i.d. r.v.’s with distribution function F(x) and pdf f(x). For various distributional properties of order statistics and
c
Institute of Statistical Research and Training (ISRT), University of Dhaka, Dhaka 1000, Bangladesh.
upper record values, see David ( 1981) and Ahsanullah (2004) respectively. Galambos and Kotz (1978) discussed extensively the characterization of exponential distribution by order statistics.
Rossberg ( 1972) characterized the exponential distribution by considering the independence of Xr,n−Xr−1,n and Xr−1,n for 1 < r < n. Ahsanullah ( 1978 a) gave a characterization of the exponential distribution by the independence of XU(n)- XU(n−1)and XU(n−1),n >1, where XU(n) is the nth upper record value. In this paper some characterizations of the exponential distribution based on the generalized order statistics are presented.
2 Main Results
The marginal pdffr,n,m,k(x) of X(r,n,m,k) can be written as fr,n,m,k(x) = cr−1
(r−1)1(1−F(x))γr−1f(x)gr−1m (F(x)), (2.1) where
gm(x)) = 1
m+ 1(1−(1−x)m+1, m >−1 =−ln(1−x), m=−1, x∈(0,1) andcr−1=γ1γ2...γr.
Since limm→−1 1
m+1(1−(1−x)m+1=−ln(1−x), we will takegm(x)) = m+11 (1−(1−x)m+1) for all m with limm→−1gm(x)) =−ln(1−x).
The cdf ofFr,n,m,k(x) of X(r,nm,k) is Fr,n,m,k(x) =
Z x
−∞
cr−1
(r−1)1(1−F(u))γr−1f(u)gmr−1(F(u))du
= It(r, γr
m+ 1)
= Γr(α(x)), m=−1, wheret= 1−(1−F(x))m+1,It(u, v) = B(u,v)1 Rt
owu−1(1−w)v−1dw,is the incomplete Beta function, α(x) =−ln(1−F(x)) and Γr(y) = Γr1 Ry
0 ur−1e−udu.
Form >−1, we have
Fr,n,m,k(x)−Fr+1,n,m,k(x) = It(r, γr
m+ 1)−It(r+ 1, γr+1
m+ 1)
= It(r, γr
m+ 1)−It(r+ 1, γr
m+ 1−1)
= Γ(r+m+1γr )
Γ(r+ 1)Γ(m+1γr )tr(1−t)m+1γr −1
= γ1γ2...γr
r!(m+!)r[1−(1−F(x))m+1]r(1−F(x))m+1]
γr+1 m+1
= cr−1
r! (1−F(x))γr+1[1−(1−F(x))m+1
m+! ]r
= fr+1,n,m,k(x)(1−F(x))
γr+1f(x) , if f(x)6= 0.
Form=−1,
Fr,n,m,k(x)−Fr+1,n,m,k(x) = fr+1,n,m,k(x)(1−F(x))
f(x) , f(x)6= 0
The joint pdffr,s,n,m,k(x, y) of X(s,n,m,k) and X(r,n,m,k),r < ss is given by fs,r,n,m,k(x, y) = cs−1
(r−1)1(s−r−1)!(1−F(x))mf(x)gr−1m (F(x))
× [h(y)−h(x)]s−r−1(1−F(y))γs−1f(y), x < y, whereh(x) = m+11 (1−x)m+1form6=−1 andh(x) =−ln(1−x) form=−1, x∈(0,1).
The conditional pdf of X(s,n,m,k) given X(r,n,m,k)=x, 1≤r < s≤n,has the following form fs|r,n,m,k(y|x) = cs−1
cr−1(s−r−1)!(hm(F(y))−hm(F(y)))s−r−1
× (1−F(y))γs−1
(1−F(x))γr+1f(y), y≥x.
The conditional pdf of X(r+1,n,m,k) given X(r,n,m,k) is fr+1|r,n,m,k(y|x) =γr+1
„1−F(y) 1−F(x)
«γr+1
f(y)
1−F(y), y≥x. (2.2)
Theorem 2.1: If Xk∈E(θ), k = 1,2,... and X’s are independent, thenγ1(X(1, n, m, k)∈E(θ) and X(r,n,m,k)=d θ
r
P
j=1 Wj
γj, where W′jsare i.i.d with cdfF(x) = 1−e−x, x≥0.
Proof. See Ahsanullah (2000).
It follows from Theorem 2.1. that Dr+1 = γr+1(X(r+ 1, n, m, k)−X(r, n, m, k)) and Dr = γr(X(r, n, m, k)−X(r−1, n, m, k)), r≥ 1, with X(0,n,m,k) =0, are identically distributed.The following Theorem gives a characterization of the exponential distribution using this property.
An absolutely continuous ( with respect to Lebesgue measure) cumulative distribution function F with probability density function f, the ratio 1−Ff(x)(x) for 1−F(x) >0 is called hazard rate. We will say F blongs to class C if r(x) is either monotone increasing or decreasing.We write Dr = γr{X(r, n, m, k)−X(r−1, n, m, k)}, r >1.
Theorem 2.2: Let X be a non-negative r.v. having an absolutely continuous (with respect to Lebesgue measure) strictly increasing distribution function F(x) with F(0) = 0 and F(x)<1 for allx >0. Then the following properties are equivalent
(a) X has an exponential distribution with density as given in (1.1);
(b) For some fixed r, (1≤r < n), the statistics Dr+1and Dr, r≥1, are identically distributed and F belongs to class C.
Proof: From Theorem 2.1, it is easy to show that (a) =⇒(b). We will prove here that (b) =⇒(a).
P(Dr+1≥x) = 1−FDr+1(x)
= cr−1
(r−1)!
Z ∞ x
Z∞ 0
(1−F(y))mf(y)gmr−1(F(y))
× (1−F(y+ z γr+1
))γr+1−1f(y+ z γr+1
))dzdy
= cr−1
(r−1)!
Z ∞ 0
(1−F(y))mf(y)gr−1m (F(y))
× int∞x (1−F(y+ z γr+1
))γr+1−1.f(y+ z γr+1
))dzdy
= cr−1
r!
Z ∞ 0
(1−F(y))mf(y)gr−1m (F(y))
× (1−F(z+ x γr+1
))γr+1dy (2.3)
Similarly, for r≥2,
P(Dr≥x) = 1−FDr(z)
= cr−2
(r−2)!
Z ∞ 0
(1−F(y))mf(y)gr−2m (F(y))
× (1−F(y+ z γr+1
))γr+1−1.f(y+ z γr
))dzdy
= cr−2
(r−1)!
Z ∞ 0
(1−F(y))mf(y)gr−2m (F(y))
× Z ∞
x
(1−F(y+ z γr+1
))γr+1−1.f(y+ z γr
))dzdy
= cr−2
(r−2)!
Z ∞ 0
(1−F(y))mf(y)gr−2m (F(y))(1−F(y+ x γr
))γrdy
= cr−1
(r−1)!
Z ∞ 0
gr−1m (F(y))(1−F(y+ x γr
))γr−1f(y+ x γr
))dy
= cr−1
(r−1)!
Z ∞ 0
gr−1m (F(y))(1−F(y+ x γr
))γrr(y+ x γr
))dy (2.4)
The relation (2.4) is also true for r=1. Since Dr+1and Drare identically distributed, we must have from (2.3) and (2.4).
0 = cr−1
(r−1)!
Z ∞ 0
gmr−1(F(y))[(1−F(y))mf(y) (1−F(z+ x γr+1
))γr+1
− (1−F(y+ x γr
))γrr(y+ x γr
)]dy (2.5)
If F has increasing hazard rate, then ln (1-F) is concave and we have for m≥1, ln(1−F
„ z+ x
γr
«
) = ln{1−F
„(m+ 1)z γr
+γr+1
γr
(z+ x γr+1
)
« }
≥ m+ 1 γr
ln(1−F(z)) +γr+1
γr
ln(1−F
„ z+ x
γr+1
« ) i.e. (1−F
„ z+ x
γr
«
)γr ≥(1−F(z))m+1(1−F
„ z+ x
γr+1
« )γr+1
(2.6) Substituting (2,6) in (2.5), we obtain
0 = cr−1
(r−1)!
Z ∞ 0
gr−1m (F(y))[(1−F(y))mf(y)
× (1−F(z+ x γr+1
))γr+1−(1−F(y+ x γr
))γrr(y+ x γr
)]dy
≤ cr−1
(r−1)!
Z ∞ 0
gr−1m (F(z))(1−F(z))m+1(1−F(z+ x γr+1
))γr+1[r(z)−r(z+ x γr
)]dx
≤ 0.
Thus (2.5) to be true we must haver(z) =r(z+γx
r) for all z and almost all x.Hence the hazard rate is constant and F(x) = 1-e−θx,for all x≥0 and any realθ >0. If F has decreasing hazard rate, then we obtain similarly the equalityr(z) =r(z+γx
r) for all z and almost all x.
If k = 1 and m = -1, then from the Theorem 2.2, we obtain the result based on r upper records ( see Ahsanullah (1978 a)).. If k = 1 and m = 0, then we obtain the characterization result of the exponential distribution given by Ahsanullah ( 1978 b) based on the identical distribution of XU(n)
- XU(n−1)and XU(n+1)−XU(n).
It can be that under the condition of the existence of the first moment, the condition (b) of Theorem 2.2 can be replaced by the equality of E(Dr+1) and E(Dr). However we need the condition that F belongs to class C.For an alternative proof of the characterization of the exponential distribution using the equality property of E(Dr+1) and E(Dr),see Cramer, Kamps and Raqab(2000).
Let F be the distribution function of a nonnegative random variables. We will call F ”new better than used” (NBU) if 1−F(x+y)≤(1−F(x))(1−F(y)), x, y≥0,and F is ”new worse than used”
(NWU) if 1−F(x+y) ≥(1−F(x))(1−F(y)), x, y≥0.We will say that F belongs to the class C1 if F is either NBU or NWU. The following theorem gives a characterization of the exponential distribution if F belongs to class C1.
Theorem 2.3.
Let X be a non-negative r.v. having an absolutely continuous (with respect to Lebesgue mea- sure) strictly increasing distribution function F(x) with F(0) = 0 andF(x)<1 for allx >0. Then the following properties are equivalent
(a) X has an exponential distribution with density as given in (1.1);
(b) For some fixed r, (1≤r < n), the statisticsDr+1/γr+1and X(1,n-r,m,k ) are identically distributed and F belongs to class C1.
Proof.
From (2.3) , we have P(Dr+1≥xγr+1) = cr−1
r!
Z ∞ 0
(1−F(y))mf(y)gmr−1(F(y))(1−F(x+x))γr+1dy (2.7) Substituting F(x) =1- e−θxin (2.7), we obtain
P(Dr+1≥xγr+1) = cr−1
r!
Z ∞ 0
e−mθyθe−θy [ 1
m+ 1(1−e−θ(m+1)y]r−1e−θγr+1(x+y)dy
= cr−1
r! e−θγr+1x Z∞
0
θe−yθγr[ 1
m+ 1(1−e−θ(m+1)y]r−1dy
= e−θγr+1x. We now prove that (b) =⇒(a).
From (2.3) , we have
P(Dr+1≥xγr+1) = cr−1
r!
Z∞ 0
(1−F(y))mf(y)gmr−1(F(y))(1−F(y+x))γr+1dy
= Z ∞
0
fr,n,m,k(y)(1−F(y+x))γr+1/(1−F(y))γr+1dy (2.8) The pdf of X(1,n-r,m,k) is
f1,n−r,m,k(x) =γr+1(1−F(x))γr+1f(x) (2.9) Thus
P(X(1, n−r, m, k)≥x) = (1−F(x))γr+1 (2.10) Since P(Dr+1≥xγr+1) =P(X(1,n-r,m,k)≥x),we obtain from (2.6) and (2.10)
0 = Z ∞
0
fr,n,m,k(y)[(1−F(y+x))γr+1/(1−F(y))γr+1−(1−F(x))γr+1]dy (2.11) If F belongs to class C1,then for (2.11) to be true , we must have
(1−F(y+x))γr+1/(1−F(y))γr+1 = (1−F(x))γr+1∀x≥0 and almost all y≥0. (2.12) The solution of (2.12) ( see Acz´el, J. (1966)) is
F(x) = 1−e−θx, x≥0 and any arbitrary, θ >0.
References
[1] Acz´el, J. (1966). Lectures on Functional Equations and Their Applications. Academic Press, New York.
[2] Ahsanullah, M. (1978a ). Record values and the exponential distribution. Ann. Inst. Statist.
Math. 30, A, 429-433.
[3] Ahsanullah, M. (1978b ).On a characterization of the exponential distribution by spacings.
Ann. Inst. Statist. Math. 30, A, 163-166.
[4] Ahsanullah,M. (2000). Generalized order statistics from exponential distribution. JSPI 85,85- 91.
[5] Ahsanullah,M. (2004).Record Values - Theory and Applications. University Press of America, Lanham, Maryland, USA.
[6] Cramer, E, Kamps,U. and Raqab,M.Z. (2003).Characterizations of exponential by spacings of generalized order statistics. Applications Mathematicae, 30,3, 257-265.
[7] David, H.A. (1981). Order Statistics (second edition). Wiley, New York.
[8] Galambox, J. and Kotz, S. (1978). Characterizations of Probability Distributions. Lecture Notes in Mathematics, No. 675, Springer Verlag, 1978, New York.
[9] Kamps, U. (1995). A concept of generalized order statistics. B. G. Teubner Stuttgart, Germany.
[10] Rossberg, H.J. (1977). Characterization of distribution functions by the independence of order statistics, , A, 34, 111- 120.