# On upper bounds on expectations of gOSs based on DFR and DFRA distributions

We focus on the problem of establishing the optimal upper bounds on generalized order statistics which are based on the underlying cdf belonging to the family of distributions with decreasing failure rate and decreasing failure rate on the average. This issue has been previously considered by Bieniek [Projection bounds on expectations of generalized order statistics from DFR and DFRA families, Statistics, 2006; 40: 339–351], who established upper nonnegative mean-variance bounds with use of the projections of the compositions of density functions of the uniform generalized order statistic and the exponential distribution function onto the properly chosen convex cones. In this paper we obtain possibly negative upper bounds, by improving the zero bounds obtained by Bieniek for some particular cases of gOSs. We express the bounds in the scale units generated by the central absolute moments of arbitrary orders. We also describe the attainability conditions.

There are no comments yet.

## Authors

• 2 publications
• ### Optimal upper bounds on expected kth record values from IGFR distributions

The paper concerns the optimal upper bounds on the expectations of the k...
02/04/2019 ∙ by Agnieszka Goroncy, et al. ∙ 0

• ### Optimal moment inequalities for order statistics from nonnegative random variables

We obtain the best possible upper bounds for the moments of a single ord...
06/13/2018 ∙ by Nickos Papadatos, et al. ∙ 0

• ### Bounded Statistics

If two probability density functions (PDFs) have values for their first ...
08/04/2018 ∙ by Pranava Chaitanya Jayanti, et al. ∙ 0

• ### On infinite covariance expansions

In this paper we provide a probabilistic representation of Lagrange's id...
06/19/2019 ∙ by Marie Ernst, et al. ∙ 0

• ### Buffered environmental contours

The main idea of this paper is to use the notion of buffered failure pro...
02/28/2019 ∙ by Kristina Rognlien Dahl, et al. ∙ 0

• ### Distorted stochastic dominance: a generalized family of stochastic orders

We study a generalized family of stochastic orders, semiparametrized by ...
09/10/2019 ∙ by Tommaso Lando, et al. ∙ 0

• ### Chernoff information of exponential families

Chernoff information upper bounds the probability of error of the optima...
02/14/2011 ∙ by Frank Nielsen, et al. ∙ 0

##### This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

## 1 Introduction

Consider a random sample

of i.i.d. random variables with common cdf

and finite absolute moment of order ,

 σpp=E|X1−μ|p=1∫0|F−1(x)−μ|pdx, (1.1)

with

 μ=EX1=1∫0F−1(x)dx. (1.2)

Let

denote the standard uniform distribution function and

the standard exponential distribution function, i.e. with the density , for .

We focus on the generalized order statistics

, introduced by Kamps (1995a, 1995b). They can be defined for an arbitrary vector of positive coefficients

by the following quantile transformation

 X(r)γ=F−1(1−r∏i=1Bi),1≤r≤n, (1.3)

where

denote independent beta distributed random variables with distributions

, respectively (see Cramer and Kamps, (2003)). The model of generalized order statistics (gOSs, for short) constitutes a unified approach containing many other popular models of ordered random variables, e.g. ordinary order statistics, record values, sequential order statistics, progressively censored type II order statistics and many more.
Let

denote uniform gOSs, i.e. those arising from the standard uniform probability distribution function

. By

we denote the probability density function of

, , which, in the general case is given by

, ,

where and stands for the particular Meijer’s G-function (see Mathai, Chapters 2 and 3, (1993)). We may further assume, that , since the marginal distribution of the single generalized order statistic does not depend on the ordering of the parameters .

This paper is devoted to the bounds on expectations of properly standardized gOSs, and its aim is to establish upper nonpositive (possibly negative) bounds on

 EX(r)γ−μσp,p≥1, (1.4)

for the restricted families of the underlying distributions , i.e. ones with decreasing failure rate or decreasing failure rate on the average. Such families of distributions can be defined in terms of the convex and star orders. First, we say that the distribution function succeeds some other fixed distribution function in the convex transform order, and write , if the composition is convex on the support of . In particular, if , then we say that belongs to the family of distributions with decreasing density (DD) and if , then is an element of decreasing failure rate class of distributions (DFR). Indeed, since is convex on , then the cumulative hazard function

 Λ(x)=V−1(F(x))=−ln[¯F(x)],

is concave on the support of . Hence its derivative

 λF(x)=Λ′(x)=f(x)1−F(x),

called the failure rate function is nonincreasing. On the other hand, it is said that the distribution function belongs to the family with decreasing failure rate on the average distributions (DFRA) if , i.e. is nondecreasing on .

There is a voluminous literature concerning the bounds on varieties of gOSs, in particular order statistics and their linear combinations (so called -statistics), records and

th record values. The classical results on the bounds on the expected sample range expressed in terms of standard deviation units were determined by Plackett (1947). Later, at the beginning of the 1950s, Moriguti (1953) introduced the greatest convex minorant method which allowed to establish bounds for the arbitrary order statistics. A year later Gumbel (1954) and Hartley and David (1954) independently derived the generalized results on the expectation of the sample maximum and of the range, which were published in the same journal issue. Undoubtedly a turning point in the research was the introduction of the method of projection, proposed by Gajek and Rychlik (1996). It was immediately applied by Gajek and Rychlik (1998) in order to establish optimal bounds for the expectations of order statistics based on the life distributions with decreasing density or failure rate. Many applications of this method are comprehensively described in Rychlik (2001). Further, Rychlik (2002) considered optimal mean-variance bounds on order statistics from families of distributions determined by the star ordering, in particular distributions with decreasing density and failure rate on the average. Danielak (2003) obtained the bounds for expectations of the trimmed means from distributions with decreasing density and decreasing failure rate.

The general bounds on the expectations of gOSs can be found in Cramer et al. (2002) and in Goroncy (2014). Projection mean-variance bounds for gOSs from restricted families were established by Bieniek (2006, 2008) who considered distributions with decreasing failure rate and decreasing failure rate on the average, as well as distributions with decreasing density and decreasing density on the average. Recently Goroncy (2017) completed his results of the latter paper with the possibly negative upper bounds for some particular cases of gOSs based on the DD and DDA distributions.

## 2 Main results

We first focus on the case of distribution functions with decreasing failure rate, i.e., ones for which are convex on . Note that we can express the expectations of the gOSs in terms of the following integral

 EX(r)γ−μσp=∞∫0F−1(V(x))−μσp(^fγ,r(x)−1)v(x)dx, (2.1)

for , and

 ^fγ,r(x)=fγ,r(V(x)),0

Below we recall the results of Bieniek (see Theorem 4.2, (2006)), who established upper nonnegative bounds on the expectations of generalized order statistics based on DFR distributions. He applied the projection of the composition of the uniform gOS density with the exponential distribution function onto a properly chosen convex cone. For this purpose we introduce

 ϱj,r=r∑i=j1γi,1≤j≤r,

and the following function

 α∗(y)=12(r−1∑j=1ϱj,rγj^fγ,j(y)+(1γ2r−1)^fγ,r(y)),y≥0,

which corresponds to the slope of the best linear approximation of the form of the projected function on .

###### Theorem 1.

If and then .
Fix now and parameters , , where the second smallest of them is strictly greater than 1. Let be the th gOS based on the distribution function from DFR family of distributions with finite mean and positive variance for .
If , then .
If , then

 EX(r)γ−μσ2≤ϱ1,r−1,

and the bound is attained for the following exponential distribution

If , then

 EX(r)γ−μσ2≤C=C1,r(γ), (2.3)

where

 C2=y∗∫0(^fγ,r(x))2e−xdx+e−y∗{(^fγ,r(y∗))2+2α∗(y∗)^fγ,r(y∗)+2(α∗(y∗))2}−1,

where is the smallest positive solution to the following equation

 r−1∑j=11γj(1−ϱj,r2)^fγ,j(y)−(γr−1)22γ2r^fγ,r(y)=0.

The bound is attained for the distribution function

Note that for cases with , Bieniek derived zero bounds without describing the attainability conditions. In general, the positivity and negativity of the upper bounds on the expected gOSs depends on the parameters , and the restrictions imposed on the parent distribution function . The projection method is appropriate in cases when it results with nonconstant projections (and positive upper bounds) of functions , which depend basically on their shapes. Otherwise this procedure returns zero bounds and another method should be employed in order to obtain possibly negative ones. Our objective is to improve Bieniek’s bounds in some particular cases.

In the first proposition we present two most general cases of optimal nonpositive bounds, which can either be zero or strictly negative.

###### Proposition 1.

Let be the th generalized order statistic based on parameter vector , and a DFR parent distribution function with expectation .

• If , then the bound is attained by the exponential baseline distribution functions.

• If , then for every DFR baseline distribution function we have .

Proof. Formula can be rewritten as

 X(r)γ=F−1(1−r∏i=1(1−Ui)1/γi),1≤r≤n,

where , , are i.i.d. standard uniform random variables. Further on we can represent them as , , with being i.i.d. standard exponential. Therefore

 X(r)γ=F−1(1−r∏i=1exp(Viγi))=F−1(V(r∑i=1Viγi)).

If is the exponential distribution function with location parameter and scale parameter , and , then

 EX(r)γ=r∑i=1E(λVi+θ)γi=λ+θ=μ.

Suppose now that . Then obviously and . Composition for being DFR is possibly first constant, and then strictly increasing. Consequently,

 F−1(V(r∑i=1Viγi))≤F−1(V(r∑i=1Vicγi)),

and the inequality is strict for sufficiently large arguments. Since any positive combination of has a positive probability measure of these arguments, it follows that

 EX(r)γ=EF−1(V(r∑i=1Viγi))

where the last inequality follows from Theorem 1.

In the next proposition we present the most general results for the upper negative bounds on with .

###### Proposition 2.

Fix and , , are such that . Let be the th gOS based on the distribution function from DFR family of distributions with finite mean and absolute central moment with . We then have the following bound

 EX(r)γ−μσp≤−inf0≤α<∞Bp(α), (2.4)

where

 Bp(α)=b(α)−1[e−αp(1−e−α)+α+e−α∫α(α+e−α−x)pe−xdx+∞∫α+e−α(x−e−α−α)pe−xdx]1/p, (2.5)

with

 b(α)=eα(1−r∑j=1ϱj,rγj^fγ,j(α))−1. (2.6)

The equality holds (possibly is approached in the limit) for the mixtures of atoms at and shifted exponential distributions with the following cdf

 (2.7)

with probabilites and , respectively, for

 c1 = b(α0)Bp(α0)σp, c2 = μ−e−α0c1,

where is the parameter in which the infimum of is attained (possibly approached in the limit).

Proof. The proof is analogous to the proofs presented in papers of Goroncy (2017, Proposition 1) and Rychlik (2009, Theorem 2), therefore we present only the main steps of the reasoning. This particular case requires consideration of the composition of the density function of the uniform th gOS and the exponential distribution function , which is given by , and .
Let us rewrite as the following functional

 Th(g)=∞∫0h(x)g(x)v(x)dx,

represented by

 h(x)=^fγ,r(x)−1,

with

 g(x)=F−1(V(x))−μσp. (2.8)

Further, consider a convex and bounded subset of , which consists of nondecreasing and convex functions that integrate to zero and have the unit th norm. Obviously, since belongs to DFR family of distributions, functions are elements of . According to Bieniek (2006, Theorem 4.2) and our setting of parameters, we assume that . Transforming into we also transform set into , which is a set that consists of all the nondecreasing and convex functions of , which satisfy

 T1(~g) = ∞∫0~g(x)v(x)dx=0, (2.9) Th(~g) = ∞∫0h(x)~g(x)v(x)dx=−1. (2.10)

Noticing that , for and , which increases as increases, we conclude that the problem of finding the supremum of over and the problem of determining

 sup{||~g||p,~g∈~Gp}, (2.11)

are equivalent. Since the norm is a convex functional, we may simplify the calculations by confining ourselves only to the extreme elements of , i.e. maximizing the norm on the subset of functions of the form

 ~gα(x)=a(α)+b(α)(x−α)1[α,∞)(x),α∈(0,∞), (2.12)

with coefficients and complying with conditions and . The substantiation of the fact that extreme elements of are broken lines defined in , is exactly the same as in Rychlik (2009, pp.59–60). Therefore we claim that

 supEX(r)γ−μσp=sup{Th(g),g∈Gp}=−inf0≤α<∞1||~gα||p. (2.13)

We first calculate the coefficients of . Using the following integral calculations obtained by Bieniek (see Lemma 3.1, (2006))

 ∞∫α^fγ,r(x)e−xdx = e−αr∑j=11γj^fγ,j(α), ∞∫α(x−α)^fγ,r(x)e−xdx = e−αr∑j=1ϱj,rγj^fγ,j(α),

we obtain that

 T1(~g(α)) = ∞∫0~gα(x)e−xdx=a(α)+b(α)e−α, (2.14) Th(~g(α)) = ∞∫0(^fγ,r(x)−1)~gα(x)e−xdx=b(α)e−α(r∑j=1ϱj,rγj^fγ,j(α)−1), (2.15)

and together with conditions and , conclude that

 a(α)=−b(α)e−α,

where is given by . We are now in the position of determining the norm of , which is required for establishing ,

 ||~gα||pp = ∞∫0∣∣a(α)+b(α)(x−α)1[α,∞)(x)∣∣pe−xdx=bp(α)[e−pα(1−e−α) +α+e−α∫α(α+e−α−x)pe−xdx+∞∫α+e−α(x−α−e−α)pe−xdx],

and eventually results in the bound .

The equality in is attained for the distribution functions , which satisfy the following condition

 F−1α(V(x))−μσp=~gα(x)||~gα||p.

The above proportion can be rewritten as

 F−1α(1−e−x)=⎧⎪⎨⎪⎩μ−σpb(α)e−αBp(α),0≤x<α,μ+σpb(α)Bp(α)(x−α−e−α),α≤x<∞,

which is consistent with the equality conditions described in the proposition.

Note that it is not easy to determine the integrals in of Proposition 2 in general, especially for noninteger . However, we managed to confirm the zero bound and specify the equality conditions for the special case of the first generalized order statistic, i.e. and arbitrarily chosen . This is possible due to the fact that the density of gOS in this case can be easily determined and that simplifies calculations. These results are presented in the proposition below, while the special case with is considered further in a separate corollary.

###### Proposition 3.

Let be the first generalized order statistic with parameter , based on the distribution function from DFR family of distributions with mean and finite th central absolute moment , for . If , then , which holds for any DFR distribution function . If , then we have the following upper bound

 EX(1)γ−μσp≤0.

The equality above is attained in limit by sequences of distributions , , which are the mixtures of atoms in and exponential distributions with the location and scale parameters with probabilities and respectively, for , and

 Np(α)=⎡⎢⎣αp−αp+1+αe−α⎛⎜⎝α∫0ypeydy+Γ(p+1)⎞⎟⎠⎤⎥⎦1/p.

Proof. The case and , is obvious, since for any parent distribution .

Suppose . Rychlik (see proof of Theorem 1, (2009)), who obtained analogous results for the common order statistics, has shown that distribution functions , , which are obviously DFR, have mean equal to and th absolute central moment equal to . Moreover, the quantile function of such distributions is given by

 F−1α(x)=⎧⎪ ⎪ ⎪ ⎪⎨⎪ ⎪ ⎪ ⎪⎩μ−σpNp(α)α,0

It is sufficient to check that

 EX(1)γ−μσp⟶0,asα⟶0.

First note that for the first uniform gOS we have the following density and the cumulative probability functions

 fγ,1(x)=γ1(1−x)γ1−1, Fγ,1(x)=1−(1−x)γ1,

respectively, where . Therefore using the formula

 ∞∫α(x−α)^fγ,1(x)e−xdx=1γ1e−αγ1,

we obtain

 0 > EFαX(1)γ−μσp=1∫0F−1α(x)−μσp(fγ,1(x)−1)dx =1Np(α)1∫1−α(−ln(1−x)+lnα)(fγ,1(x)−1)dx =1Np(α)∞∫−lnα(y+lnα)(fγ,1(1−e−y)−1)e−ydy =1Np(α)⎛⎜⎝∞∫−lnα(y+lnα)fγ,1(1−e−y)e−ydy−α∫∞0ye−ydy⎞⎟⎠ =α1−1p1γ1αγ1−1−1{αp−1−αp+e−α[α∫0ypeydy+Γ(p+1)]}1/p,

which tends to 0 if , and completes the proof.

Note that the first generalized order statistic has the distribution identical with the minimum of i.i.d. random variables with the distribution function , which shares the DFR property. Therefore, knowing the bounds for the ordinary order statistics, we also obtain the bounds for the first generalized order statistics, with use of this transformation. This is why results of Proposition 3 above coincide with the result of Rychlik (see Theorem 1, (2009)), who considered upper nonpositive bounds on low rank common order statistics in the same model of DFR family of distributions.

The case of bounds given in scale units generated by the mean absolute deviation , is treated separately in the corollary below, which follows from Proposition 2. Note that in this case

 ||~gα||1=2e−α−e−αb(α), (2.16)

since the denominator of reduces to .

###### Corollary 1.

Fix and , , are such that . Let be the th gOS based on the distribution function from DFR family of distributions with finite absolute mean deviation defined in for . We have the following bound

 EX(r)γ−μσ1≤−inf0≤α<∞B1(α),

where

 B1(α)=12ee−α[1−r∑j=1ϱj,rγj^fγ,j(α)],0≤α<∞. (2.17)

The equality holds (possibly is approached in the limit) for mixtures of atoms in and shifted exponential distributions with and with probabilites and , respectively, where is the parameter in which the infimum of is attained (possibly approached in the limit).

In particular, bounds for the first generalized order statistics, given in the absolute mean deviation units are strictly negative, presented in the corollary below.

###### Corollary 2.

Let be the first generalized order statistic with parameter , based on the distribution function from DFR family of distributions with finite mean absolute deviation . We have the following upper bound

 EX(1)γ−μσ1≤−inf0<β≤112eβ(1−1γβγ−1). (2.18)

The equality conditions in correspond to those described in Corollary 1, with , where is the point in which the infimum of the rhs of is attained.

Note that here we have

 b(α)=eα(1−1γ1e−α(γ1−1))−1,

which for , together with results in the desired bound .

The numerical values of bounds for the first gOSs expressed in units for some particular cases of the parameter are presented in Table 1. Parameter stands for the argument of the rhs of , for which the infimum is attained.

The numerical simulations show that for increasing values of , the value of decreases from 1 to 0 while the bound gradually decreases to . It eventually stabilizes its value in beginning with , when the infimum of the rhs of is attained in . This coincides with results of Goroncy (2014, Theorem 4) in the general case of the arbitrary underlying distribution function for .

Let us now consider gOSs which are based on DFRA distribution functions. The case of upper nonnegative bounds on for was solved by Bieniek (2006, Theorem 4.3). However, for some particular cases of gOSs, the projection method, which was used there, resulted in zero bounds. Recently, Goroncy (2017) presented bounds on expectations of the standardized gOS arising from decreasing density on the average distributions. This method of obtaining possibly negative bounds will be applied here in order to obtain respective results for decreasing failure rate on the average distributions.

The results presented below are just the immediate application of the Proposition 3 of Goroncy (2017) in case when a particular condition presented in Bieniek (2006, Theorem 4.3) is satisfied.

###### Proposition 4.

Let , , be the th gOS based on the distribution function which belongs to DFRA family of distributions and fix . Let be the only zero of the equation in the interval , for being the smallest inflection point of . If the following condition is satisfied

 r∑j=11γj(ϱj,r+^β)^fγ,r(^β)≤1+^β,

then the following bound holds

 EX(r)γ−μσp≤−inf0<α<∞B∗p(α), (2.19)

where

 B∗p(α) = ⎧⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪⎨⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎪⎩1bα[((α+1)e−α)p(1−e−α)+(α+1)e−α∫α(−x+(α+1)e−α)pe−xdx+∞∫(α+1)e−α(x−(α+1)e−α)pe−xdx⎤⎦−1/p,0<α<α0,1bα[((α+1)e−α)p(1−e−α)+∞∫α(x−(α+1)e−α)pe−xdx]−1/p,α≥α0, bα = −eα[r∑j=11γj^fγ,j(α)(α+ϱj,r)−α−1]−1,

with being the only solution of the equation in . The equality in is attained for distribution functions such that

 F−1α∗(V(x))=⎧⎪⎨⎪⎩−bα∗(α∗+1)e−α∗B∗p(α∗)σp+μ,0

where is the argument for which the infimum of is attained. If the infimum of is attained in the limit, then the equality is also attained in the limit by sequences of distributions described in .

Numerical calculations show, that in some cases (e.g. for particular settings of progressively type II censored order statistics), the infimum of is approached in limit for and equals to zero. This is not surprising, but it is not easy to show analytically that in general the zero infimum of is attained in the limit, at the right exponential support interval.

## Acknowledgements

The research was supported by the Polish National Science Center Grantno. 2011/01/D/ST1/04172. The author is very grateful to anonymous referees, whose helpful comments and remarks helped to significantly improve the manuscript, especially to the one who suggested addition of Proposition 1.

## References

• [1] Bieniek, M., (2006), Projection bounds on expectations of generalized order statistics from DFR and DFRA families, Statistics, 40: 339–351.
• [2] Bieniek, M., (2008), Projection bounds on expectations of generalized order statistics from DD and DDA families, J. Statist. Plann. Inference, 138: 971–981.
• [3] Cramer, E., Kamps, U., (2003), Marginal distributions of sequential and generalized order statistics, Metrika, 58: 293–310.
• [4] Cramer, E., Kamps, U., Rychlik, T., (2002), Evaluations of expected generalized order statistics in various scale units, Appl. Math. (Warsaw) 29: 285–295.
• [5] Danielak, K., (2003), Sharp upper mean-variance bounds for trimmed means from restricted families, Statistics, 37: 305–324.
• [6] Gajek, L., Rychlik, T., (1996), Projection method for moment bounds on order statistics from restricted families. I. Dependent case. J. Multivar. Anal., 57: 156–174.
• [7] Gajek, L., Rychlik, T., (1998), Projection method for moment bounds on order statistics from restricted families. II. Independent case, J. Multivar. Anal., 64: 156–182.
• [8] Goroncy, A., (2014), Bounds on expected generalized order statistics, Statistics, 48: 593–608.
• [9] Goroncy, A., (2017), Upper non-positive bounds on expectations of generalized order statistics from DD and DDA populations, Comm. Stat. Theor. Meth., 46:24: 11972–11987.
• [10] Gumbel, E.J., (1954), The maxima of the mean largest value and of the range, Ann. Math. Stat., 25: 76–84.
• [11] Hartley, H.O., David, H.A., (1954), Universal bounds for mean range and extreme observation, Ann. Math. Stat., 25: 85–99.
• [12] Kamps, U., (1995a), A Concept of Generalized Order Statistics, Teubner, Stuttgart.
• [13] Kamps, U., (1995b), A concept of generalized order statistics, J. Statist. Plann. Inference, 48: 1–23.
• [14] Mathai, A. M., (1993), A Handbook of Generalized Special Functions for Statistical and Physical Sciences, Clarendon Press, New York.
• [15] Moriguti, S., (1953), A modification of Schwarz’s inequality with applications to distributions. Ann. Math. Statist. 24: 107–113.
• [16] Plackett, R.L., (1947), Limits of the ratio of mean range to standard deviation. Biometrika 34: 120–122.
• [17] Rychlik, T., (2001), Projecting statistical functionals, Springer-Verlag, New York.
• [18] Rychlik, T., (2002), Optimal mean-variance bounds on order statistics from families determined by star ordering, Appl. Math., Warsaw, 29: 15–32.
• [19] Rychlik, T., (2009), Non-positive upper bounds on expectations of low rank order statistics from DFR populations, Statistics, 43: 53–63.