Lecture Notes in Statistics 160

Edited by [Link], P. Diggle, S. Feinberg, K. Krickeberg,

I. Olkin, N. Wermuth, and S. Zerger
Springer
New York
Berlin
Heidelberg
Barcelona
Hong Kong
London
Milan
Paris
Singapore
Tokyo
Tomasz Rychlik

Projecting Statistical Functionals

, Springer
Tomasz Rychlik
Institute of Mathematics
Polish Academy of Sciences
Chopina 12
87100 Toruil
Poland
trychlik@[Link]

Library of Congress Cataloging-in-Publication Data

Rychlik, Tornasz.
Projecting statistical functionals / Tornasz Rychlik
p. cm. - (Lecture notes in statistics; 160)
Includes bibliographical references and index.
ISBN-13:978-0-387-95239-0 e-ISBN-13:978-1-4612-2094-7
DOl: 10.10071978-1-4612-2094-7

I. Statistical functionals. 2. Hilbert space. I. Title. II. Lecture notes in

statistics (Springer-Verlag); v. 160.
QA273.6 .R93 2001
519.2'4-dc21 00-069240

Printed on acid-free paper.

C 2001 Springer-Verlag New York, Inc.

All rights reserved. This work may not be translated or copied in whole or in part without the
written permission of the publisher (Springer-Verlag New York, Inc., 175 Fifth Avenue, New York,
NY 10010, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer
software, or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use of general descriptive names, trade names, trademarks, etc., in this publication, even if the
former are not epecia\ly identified, is not to be taken as a sign that such names, as understood by
the Trade Marks and Merchandise Marks Act, may be accordingly used freely by anyone.

Camera-ready copy provided by the author.

9 8 7 654 3 2 1

ISBN-13:978-0-387-95239-0 SPIN 10793590

Springer-Verlag New York Berlin Heidelberg

A member ofBerteismannSpringer Science+Business Media GmbH
Preface

About 10 years ago I began studying evaluations of distributions of or-

der statistics from samples with general dependence structure. Analyzing
in [78] deterministic inequalities for arbitrary linear combinations of order
statistics expressed in terms of sample moments, I observed that we obtain
the optimal bounds once we replace the vectors of original coefficients of
the linear combinations by the respective Euclidean norm projections onto
the convex cone of vectors with nondecreasing coordinates. I further veri-
fied that various optimal evaluations of order and record statistics, derived
earlier by use of diverse techniques, may be expressed by means of projec-
tions. In Gajek and Rychlik [32], we formulated for the first time an idea of
applying projections onto convex cones for determining accurate moment
bounds on the expectations of order statistics. Also for the first time, we
presented such evaluations for non parametric families of distributions dif-
ferent from families of arbitrary, symmetric, and nonnegative distributions.
We realized that this approach makes it possible to evaluate various func-
tionals of great importance in applied probability and statistics in different
restricted families of distributions.
The purpose of this monograph is to present the method of using pro-
jections of elements of functional Hilbert spaces onto convex cones for es-
tablishing optimal mean-variance bounds of statistical functionals, and its
wide range of applications. This is intended for students, researchers, and
practitioners in probability, statistics, and reliability. Intended as a refer-
ence book it could also be used as a textbook for a specialized course in the
subject area. Numerous open problems are formulated in the text. I hope
that they stimulate some readers to undertake research in this direction.
vi Preface

The prerequisities are upper-level undergraduate courses in probability

theory, mathematical statistics, and functional analysis, including an ele-
mentary theory of Hilbert spaces. Certainly, some knowledge of nonpara-
metric statistical inference and reliability theory would be beneficial. It is
important to become acquainted with the content of Chapter 2 before read-
ing the following ones. Chapters 3 through 6 can be studied independently,
with some exceptions. Some results of Sections 5.3,5.4, and 5.6 are deduced
from lemmas contained in Sections 3.2, 3.3, and 4.5, respectively. The re-
sults of Chapter 7 are based on applications of second moment bounds
presented in Sections 4.1, 4.2, 6.2, and 6.3.
This work was done with the support of the Polish State Committee for
Scientific Research (KBN) Grant 2 P03A 014 13. The idea of writing the
book was suggested by Leslaw Gajek. Besides working together on spe-
cific problems, we carried on comprehensive discussions that enabled us to
crystallize original vague ideas into a definite plan. As he undertook other
important obligations, he could not participate in completing the project.
Some results were obtained in cooperation with Andrzej Okolewski whose
contribution is gratefully acknowledged. Significant parts of the content
of this book were presented and discussed during regular Thursday semi-
nars on applied probability held in the Institute of Mathematics of Polish
Academy of Sciences in Warsaw. Critical remarks and comments of the par-
ticipants had a strong positive impact on the final outcome. The assistance
of several people made it possible to improve the presentation. Many help-
ful suggestions were provided by John Kimmel, the Executive Editor of the
Statistics Division at Springer-Verlag, and one of the reviewers. The copy-
editor corrected a number of linguistic and style mistakes. Jan K. Kowalski
helped to prepare the final Jh.1EXversion of the manuscript in camera-ready
form. It is a pleasure to offer my sincere thanks to them.

Tomasz Rychlik
January, 2001
Contents

Preface v

List of Tables ix

1 Introduction and Notation 1

1.1 Introduction. 1
1.2 Notation.......... 5

2 Basic Notions 11
2.1 Elements of Hilbert Space Theory 11
2.2 Statistical Linear Functionals . . . 16
2.3 Restricted Families of Distributions. 25

3 Quantiles 33
3.1 General and Symmetric Distributions . . . . . . . . . . . . 33
3.2 Distributions with Monotone Density and Failure Rate . .. 36
3.3 Distributions with Monotone Density and Failure Rate on
the Average . . . . . . . . . . . . . 44
3.4 Symmetric Unimodal Distributions 50
3.5 Open Problems . . . . . . . . . . . 54

4 Order Statistics of Independent Samples 55

4.1 General and Symmetric Distributions . . . . . . . . . . . . 56
4.2 Life Distributions with Decreasing Density and Failure Rate 60
viii Contents

4.3 Distributions with Monotone Density and Failure Rate on

the Average . . . . . . . . . . . . . 69
4.4 Symmetric Unimodal Distributions 79
4.5 Bias of Quantile Estimates. 82
4.6 Open Problems. . . . . . . . . . . 91

5 Order Statistics of Dependent Observations 95

5.1 Dependent Observations with Given Marginal Distribution 96
5.2 General and Symmetric Distributions . . . . . . . . . . . . 100
5.3 Distributions with Monotone Density and Failure Rate . . . 105
5.4 Distributions with Monotone Density and Failure Rate on
the Average . . . . . . . . . . . . . . . . . . . . . . 114
5.5 Symmetric Unimodal and U-Shaped Distributions 115
5.6 Bias of Quantile Estimates ... 118
5.7 Extreme Effect of Dependence. 121
5.8 Open Problems . . . . . 127

6 Records and kth Records 131

6.1 Dependent Identically Distributed Observations . . . . . . . 131
6·.2 General and Symmetric Distributions . . . . . . . . . . . . 133
6.3 Life Distributions with Decreasing Density and Failure Rate 136
6.4 Increments of Records 140
6.5 Open Problems . . . . . . . . . . . . . . . . . . . . . . . . . 143

7 Predictions of Order and Record Statistics 145

7.1 General Distributions . . . . . . . . . . . . 146
7.2 Distributions with Decreasing Density and Failure Rate 149
7.3 Open Problems . . . . . . . . . . . . . . . . . . . . . . . 154

8 Further Research Directions 157

References 163
Author Index 171
Subject Index 173
List of Tables

3.1 Sharp uniform mean-variance bounds on quantiles for vari-

ous families of distributions. . . . . . . . . . . . . . . . . .. 53

4.1 Sharp uniform mean-variance bounds on expectations of or-

der statistics from independent samples of size 20 for various
families of distributions. . . . . . . . . . . . . . . . . . . .. 81
4.2 Sharp uniform variance bounds on upper bias deviations of
estimators Xj:20 of pth quantiles, p = j/20, for various fam-
ilies of distributions (independent case). . . . . . . . . . .. 92

5.1 Sharp uniform mean-variance bounds on expectations of or-

der statistics from dependent samples of size 20 for various
families of distributions. . . . . . . . . . . . . . . . . . . . . 117
5.2 Sharp uniform variance bounds on upper bias deviations of
estimators lj:20 of pth quantiles, p = j /20, for various fam-
ilies of distributions (dependent case). . . . . . . . . . . .. 122
1
Introduction and Notation

1.1 Introduction
This work presents a method of using projections of functions onto convex
cones in Hilbert spaces for determining sharp bounds on values of statistical
functionals over general and restricted families of distributions, expressed in
terms of moment parameters of the distributions. The method is based on
representing the statistical functionals and families of distributions as fixed
elements and convex cones, respectively, in a common real Hilbert space.
Then the norm of the projection of the element onto the cone provides the
optimal bound. The distribution for which the bound is attained is derived
by a simple transformation of the projection.
The advantage of the projection method lies in providing definite an-
swers for numerous simply stated but nontrivial problems of theoretical
and practical importance. The method enables us to optimally evaluate
various random objects in terms of the first two moments which are the
most classical parameters describing the population. Natural restrictions on
the random structure of observations are allowed. Although some results
presented here were proven earlier by means of specific tools, our unified
approach provides simpler proofs and indicates mutual relations among
various evaluations. On the other hand, for numerous problems presented
in the book, we do not see alternative ways of solving them without refer-
ence to projections. The idea of projections has been effectively exploited
in various aspects of statistical inference, for instance in the least squares
and minimum distance estimation. The majority of applications are based
2 1. Introduction and Notation

on projections onto finite-dimensional linear subspaces. The novelty of our

approach consists in using solutions of projection problems onto convex
cones in function spaces for precise evaluations of statistical functionals.
We illustrate the general idea by an example.
EXAMPLE 1. Suppose that random variables XI, ... ,Xn are independent
identically distributed (LLd.) with a common distribution F. We are inter-
ested in evaluating the upper deviation of the expectation of the jth order
statistic EFXj :n from the population mean P.F = EFX1 in the standard
deviation units UF = (VarFX1)1/2. We can Write

EFXj:n - P.F = 11 [F-l(X) - P.F][/;:n(X) - 1] dx, (1.1)

where
/;:n(X) = n(; =:)xi- 1 (1- x)n- j

is the density function of the jth order statistic from the LLd. standard uni-
form sample. This may be interpreted as the inner product of the centered
quantile function F-l_P.F with /;:n -1 in the Hilbert space £2([0, 1), dx) of
the square integrable functions on the unit interval. Applying the Schwarz
inequality to (1.1), and noting that IIF-l - [Link] = UF, we obtain
(1.2)
The equality is attained here if the arguments of the inner product are
proportional; that is,
F- 1 (x) - P.F = a[/;:n(x) -1] (1.3)
for some a ~ O. All the possible centered quantile functions of distributions
with a finite variance form the convex cone of nondecreasing functions
integrating to 0 in £2([0, 1),dx). The cone is further denoted by Co.
rr j = n, then the right-hand side of (1.3) actually increases, and inte-
grates to O. Therefore bound (1.2) is then tight, and (1.3) enables us to
determine a (power) distribution that attains the bound (see Gumbel [36],
and Hartley and David [38]). Otherwise /;:n -1 is not nondecreasing, and
(1.2) cannot be sharp.
The most natural idea is to replace /;:n - 1 by the closest element of
the family of centered quantile functions Co. We call it the projection of
/;:n - 1 onto CO and denote it by pO(/j:n - 1). The idea does work here
and in many other problems considered in the monograph. Namely,
[EFXj:n - P.F]/UF $ B°(j,n) = IlpO(fj:n -1)11 (1.4)
for all F with u} < 00, and we get the equality in (1.4) if
F-l(X) - P.F = P°(fj:n -1)(x) E CO,
(1.5)
UF IIP°(fj:n - 1)11
 1.1 Introduction 3

which defines distribution function F with the extreme normalized expec-

tation of X j : n . If j = 1, then /;:n - 1 is decreasing, and the projection is
constant
p°(fj:n - 1) = o.
Hence we have EFXl :n ~ /JF which is clear by relation X l :n ~ Xl. If
1 < j < n, then the increasing-decreasing function /;:n - 1 is replaced by
nondecreasing
p°(fj:n -1)(x) = /;:n(min{x,a.}) -1, (1.6)
where a. is a point of increase of /;:n determined by an equation (see
Moriguti [58]). A detailed construction of (1.6) and verification of (1.4)
and (1.5) is presented in Section 4.1 (see also Example 2 of Section 2.1).
This is based on an old general method due to Moriguti based on con-
vex minorants, and (1.6) is actually the derivative of the greatest convex
minorant of the antiderivative of /;:n - 1.
For the first time in the context of evaluating statistical functionals,
however, the notion of projection appeared in Rychlik [78], where some
deterministic bounds on linear combinations of order statistics in terms
of sample mean and variance were established. The general idea of using
projections for sharp evaluations of statistical functionals was formulated in
Gajek and Rychlik [32]. The most recent review of results is presented in the
expository paper by Rychlik [86]. In principle, we provide here an exposition
of recent research: a significant number of papers we refer to have not been
published yet, and some results have not been presented elsewhere. Since
this work is entirely devoted to the projection method, we consistently omit
discussing other approaches and many related results. Therefore references
to results obtained by different methods are rare and laconic here. It is
worth pointing out that considering Hilbert spaces and convex cones is
essential for the method. For instance, the projections of linear statistical
functionals in V-spaces with p '" 2 do not provide analogous evaluations.
The structure of the book reHects the purpose of presenting bounds
on various statistical functionals over different classes of distributions by
means of the common method based on projections. Fundamental notions
are introduced in Chapter 2. Some facts of general Hilbert space theory
taht are applied in our method are collected in Section 2.1. Inner product
formulae for some functionals with statistic81 interpretations are worked
out in Section 2.2. In Section 2.3 special classes of distributions of practical
importance in probability, mathematical statistics, and reliability are char-
acterized by convex cones of respective quantile functions. Some partial
orders of distributions that are useful in defining the classes are introduced
there. Consecutive chapters are devoted to specific functionals, and bounds
on the functionals over various classes are discussed in respective sections.
Titles of sections refer to families of distributions with intuitive interpreta-
tions. In fact, more general results are presented. We often consider families
4 1. Introduction and Notation

of distributions related to a fixed general one with respect to partial or-

ders defined in Section 2.3. Some results are specified for specially chosen
extreme elements of the families, for instance for uniform and exponential
distributions that generate some classes with natural intuitive properties.
The main results are presented in Chapters 3 to 7 in a unified way:
each sharp bound on a fixed functional over a fixed class of distributions
is explicitly written, together with a formula for the distribution function
for which the bound is attained. Different bounds for various classes of
distributions are also compared numerically. Special emphasis is laid on
construction of projections, the essence of our method. Since in the prob-
lems we study, general methods for constructing projection functions are
not known, various tools are needed for solving specific problems. Usually
we first describe the shape of the projections up to several real parameters
by means of geometric arguments, and then determine the optimal parame-
ters analytically. Some bounds are expressed by complicated formulae that
should be evaluated by use of subtle tools of numerical analysis. There
is an apparent intentional contrast between precise formulation of results
and informal justification. A detailed verification is provided for selected
problems of bounds on quantiles and order statistics in Chapters 3 and 4,
respectively. Presented here are the most typical geometric and analytic
arguments used for determining projections of step and smooth functions.
The other proofs are merely' sketched and we refer the reader to the orig-
inal papers for complete details. In particular, we omitted formal proofs
of the results of Chapters 5 and 6, which are also substantially based on
projecting step and smooth functions, respectively.
The results presented here are far from being complete. There are still
many interesting open problems that can be formulated for statistical func-
tionals and families of distributions described here. In the concluding sec-
tions of the chapters some open problems are stated. All of them can be
reformulated as projection problems onto convex cones. They are more diffi-
cult than standard exercises and problems usually contained in monographs
with the purpose of enabling the reader to gain a deeper understanding of
the text. We hope that our problems stimulate readers to undertake re-
search and deliver original and interesting evaluations of statistical func-
tionals by means of projections. There are also other functionals and fami-
lies of distributions that are not studied here and for which the projection
method would work. Some of them are presented in Chapter 8, where pos-
sible further research directions are indicated. Extending the evaluation
problems to these functionals and families generate new interesting ques-
tions. In order to solve them, one should possibly develop techniques and
methods different from ones presented here.
 1.2 Notation 5

1.2 Notation
We tried to set a unified notation for the whole work. A list of symbols is
presented below. Only notation used locally is not included there. We gen-
erally preserved standard symbols. However, some other notation may look
strange at the first glance. For instance, numerous convex cones of nmc-
tions were considered in the text. They were denoted by C with various
superscripts and subscripts whose meaning is explained in the list below
(see also Section 2.3). The projection operator onto a given cone is written
as P with the same upper and lower indices. Moreover, the indices appear
in the notation of sharp bounds on functionals that are obtained by means
of projections on specified convex cones. For instance, B°(j, n) calculated
in Example 1, denotes the sharp mean-variance bound on the expectation
of the jth order statistic of an independent identically distributed sample
of size n with arbitrary marginal distribution that has a finite variance. The
bound is obtained by means of projection pO of a properly chosen functional
(see Section 2.2) onto the convex cone Co of quantile functions of all distri-
butions with finite variance, centered about the respective mean. Convex
cones of quantile functions for families of distributions satisfying various
restrictions and respective projection operators need more sophisticated
notation. Nevertheless, we tried to introduce the symbols in a coherent
manner and believe that the reader shall get used to them.

List of symbols
(1l, (. , .)) (real) Hilbert space with inner product (.,.)
11·11 = (., .)1/2 - norm in (1l, (. ,.))
L2([a, d), w(x) dx) Hilbert space of functions 9 : [a, d) I-t !R
t
satisfying g2(X)W(X) dx < 00 with inner
a d
product (g, h) = fa g(x)h(x)w(x) dx
for a positive weight nmction w(x)
C¥.,(3, ... scalars, parameters of functions
C¥.*,(3., ••. optimal scalars, parameters of projections
g,h, .. . functions, elements of Hilbert spaces
G,H, .. . antiderivatives of g, h, .. . , respectively
gH(x) = g(H(x)) -composition of functions H andg
H(x) greatest convex minorant of H(x)
h(x) (right) derivative of H(x)
l(x) constant function equal to 1
lA(X) indicator of A (= 1 on A and 0 elsewhere)
x+ = max{x,O} - positive part of a number
lxJ = max{k :5 x: kinteger}-floorofa number
6 l. Introduction and Notation

Th(·) = (h, .) - continuous linear functional on

a Hilbert space
ThO/II·11 normalized linear functional
F(x) marginal distribution function of population
f(x) density function of F
.\F(X) = f(x)/[I - F(x)] - failure rate of F
F-1(x) = sup{y: F(y) ~ x}, 0 ~ x < I, - quantile
function of F
F-l(P) quantile of order p, 0 < p < I
Fh,(x) = [F(x)-F(y)]/[I-F(y)],x ~ y, -distribu-
tion function of X under condition X > y
I1:..(x) = Fjfl(X + y) = [F(x + y) - F(y)]![I- F(y)],
x ~ 0, - distribution function of X - Y
under condition X > y
Flz(x) = F(x)/F(z), x ~ z, - distribution
function of X under condition X ~ z
FI~(x) = [F(x)-F(y)]/[F(z)-F(y)], y~x~z,
- distribution function of X under condi-
tiony<X~z
J.I.=J.l.F = f~ F-l(X) dx - mean of F
m2 =m} = fo 2[F-l(X)]2 dx - second raw moment of F
q2 =q} = mF - 2 - variance
J.l.F . 0f F
<; = <;F = f; W-1(x) - [Link] dx - mean absolute
deviation of F
[Link]
2111
mean of Ffu
mpo second raw moment of FI~
III
M2_M2
- pO = [q} + (J.l.F - y)2]/[I_ F(y)]
III
'Pn(F) family of all distributions on ~n with common
marginal F
U(x) = x, 0 ~ x ~ I, -standard uniform distribution
function
V(x) = 1- exp( -x), x ~ 0, - standard exponential
distribution function
W(x) a (distinguished) distribution function
a=aw (finite) left endpoint of (interval) support of W
d=dw right endpoint of (interval) support of W
w(x) density function of W,
weight function in L2([aw,dw),w(x)dx)
J.l.w(f3) = Ew(XIX > (3)
q~(f3) = Varw(XIX > (3)
itw(f3) = Ew(XIX ~ (3)
q~(f3) = Varw(XIX ~ (3)
 1.2 Notation 7

fi.W{P) = Ewmin{X,p}
v?v{P) = Ew[min{X,p}]2
u?v{P) = Varwmin{X,p}
11w{a,p) = Ew{X - a) 1 (ti,d) (X)
1?~{a,p) = Varw{X - a) 1 (ti,d) (X)
f/w{P) = 11w{P, P) = Ew{X - P)+
v?v{P) = Ew[{X _ P)+]2
J~{P) = 1?~{P,P) = Varw{X - P)+
'iJw{P) = Ew{P-X)+
t?~{P) = Varw{P - X)+
XI, ... ,Xn. ... independent identically distributed (LLd.)
random variables with distribution function F
YI, ... ,Yn, ... possibly dependent identically distributed
random variables with common marginal F
X j :n jth order statistic of Xl, .. . , X n, 1 :5 j :5 n
Y;:n jth order statistic of YI , .. . , Yn, 1 :5 j :5 n
Ej=l CjXj :n L-statistic of independent sample
Ej=l Cj Y;:n L-statistic of dependent sample
C = (CI, ... ,en) vector of coefficients of L-statistic
Ln nth occurrence time of (first) record (increase
in sequence of sample maxima Xj:j, j ~ 1)
Rn = XLn - nth value of (first) record,
L~) nth occurrence time of (kth) record (increase
in sequence of kth greatest order statistics
X j +1-k:j, j ~ k), k ~ 1
~) = XL~.)+1-k:L~.) - nth value of kth record
/;:n{x) = n(j=Dxj - I {I- x)n- j l[O,lj (x) - density
function of jth order statistic of LLd. stan-
dard uniform sample of size n, 1 :5 j :5 n,
expectation functional for Xj:n
Fj:n{x) = E:=j (:)xk{1 - x)n-k, 0 :5 x :5 1,
- distribution function of /;:n{x)
Gj:n{X) distribution function of jth order statistic
of dependent identically distributed sample
of size n
Gj:n{x) = (nx + 1- j)l[(j-I)/n,lj{x)/{n + 1- j) - sto-
chastically largest distribution function
of jth order statistic of dependent sample
of size n with standard uniform marginal
gj:n{X) = nl[(j-l)/n,l) {x)/{n + 1 - j) - density func-
tion of Gj:n, expectation functional for Y;:n
8 1. Introduction and Notation

Gc(x) the greatest convex function satisfying

G(jjn) ~ ~{=l Ck, 0 ~ j ~ n, 0 ~ x ~ 1
9c(x) = ~;=l dj l[U-l)/n,j/n) (x) - (right) derivative of
Gc(x), expectation functional for ~;=l CjYj:n
fn(x) = [-In(l- x)]nl[O,l)(x)jn! - density function
of nth value of (first) record of LLd. standard
uniform sequence, expectation functional of Rn
Fn(x) = (1 - x) ~;=o[-ln(l - x)J3 jj!, 0 ~ x ~ 1,
- distribution function of fn(x)
f~k)(x) = knH[-ln(l - x)]n(l- x)k- 1l[o,1)(x)jn!
- density function of nth value of kth
record of LLd. standard uniform sequence,
expectation functional for R~k)
F~k) (x) (1- x)k ~;=o kj[-ln(l- x)J3 Ii!, 0 ~ x ~ 1,
- distribution function of f~k) (x)
~c convex order of distribution functions: F ~c W
if F-1W(x) is convex on [aw,dw)
~* star order of distribution functions: F ~* W if
F- 1 W(x) is starshapedj that is,
F-1W(x)j(x-aw) is nondecreasing on [aw,dw)
~s s-order of symmetric distribution functions:
F ~s W if F-1W(x) is convex on [J1.w,dw)
C/' {g E L2([O, 1), dx): 9 is nondecreasing} -fami-
ly of quantile functions F-1(x) of arbitrary
distribution functions F(x) with finite variance
CO = {g E C/' : fol g(x) dx = O} - family of respect-
ive centered quantile functions F- 1 (x) - J1.F
C+ = {g E C/' : g(O) = O} - family of quantile
functions of life distributions with aF = 0
cs {g E C/' : g(x) = -g(l - x-)} - family of
centered quantile functions of symmetric
distributions
C- any ofC/',Co,C+,C s
Civ = {gW: 9 E C·} C L2([aw,dw),w(x)dx)
- family of compositions of W(x) with (center-
ed) quantile functions of C· (with aw replaced
by J1. w in the last case)
C·~c(:~c)w = {g E Civ : 9 is convex (concave) on [aw,dw)}
- family of compositions of W(x) with (center-
ed) quantile functions of C· for F ~c (~c)W
C·~.(~.)w = {g E Civ : 9 is (anti)starshaped on [aw, dw)}
- family of compositions of W(x) with (center-
ed) quantile functions of C- for F ~* (~*)W
 1.2 Notation 9

c·b(:::S.)W = {g E Cw
' : 9 is convex (concave) on r"w,dw)}
lfO
- family of compositions of symmetric W (x)
with centered quantile functions of C' for
symmetric P t. (js)W

In particular:

(centered) quantile functions of distributions

with decreasing (increasing) density
(centered) quantile functions of distributions
with decreasing (increasing) density on the
average
centered quantile functions of symmetric
unimodal (U-shaped) distributions
compositions of V(x) = 1 - exp( -x) with
(centered) quantile functions of distributions
with decreasing (increasing) failure rate
compositions of V(x) = 1- exp( -x) with
(centered) quantile functions of distributions
with decreasing (increasing) failure rate on
the average
any of above defined convex cones
projection ontoC:
sharp bound on p-l(P) determined by projec-
tion ontoC:
B = B:(j,n) sharp bound on expectation of Xj:n (indepen-
dent case) determined by projection ontoC:
C = C:(j,n) sharp bound on expectation of }j:n (dependent
case) determined by projection ontoC:
D = D:(k,n) sharp bound on expectation of ~k) determined
by projection onto C:
2
Basic Notions

2.1 Elements of Hilbert Space Theory

We recall here some basic facts about the Hilbert spaces that are used in
the sequel. They can be found in textbooks on functional analysis (see, e.g.,
Balakrishnan [9]). A pair (1i, (., .)) is called a real inner product space if
1i is a real linear space and the function (.,.) : 1i x 1i I---t ~, referred to
further as the inner product, is linear in each argument when the other is
fixed, symmetric under rearrangement of arguments, and positive if both
arguments are identical and nonzero. These properties imply the Schwarz
inequality
'Vg, hE 1i (g, h) ~ [(g, g)(h, h)]1/2. (2.1)
This is trivial when either of the arguments is zero. Otherwise we con-
clude (2.1) from the relations

o~ (g - ~~:~jh'9 - ~~:~jh) (h,h) = (g,g)(h,h) - (g,h)2.

This also shows that (2.1) becomes the equality iff 9 = 0, h = 0, or 9 = a:h
for some a: > o. We use (2.1) for verifying that the function
Ilhll = (h,h)1/2
h I---t
defines a norm in 1i. H (1i, II ·ID is complete, then
(1i, (. ,.)) is called the
Hilbert space. The Riesz representation theorem asserts that every linear
continuous functional defined on a Hilbert space can be written as
Th(g) = (g, h), 9 E 1i,
12 2. Basic Notions

for some h E 1l. By (2.1) again, 11Th II = Ilhll. One can see that the normal-
ized nonzero functional Th(g)/lIgll, 0 :F 9 E 1l, attains its maximum IIhll
at 9 = ah with a > O.
In numerous statistical problems, it is important to maximize a linear
normalized functional over a convex cone in a Hilbert space. We say that
C C 1l is a convex cone if f, 9 E C implies that af + f3g E C for arbitrary
a, f3 ~ o. If h E C, then the solution of our restricted maximization problem
coincides with that of the general one. Otherwise we show that h should
be replaced by its projection Ph onto C, that is, the element of C that is
least distant from h. This can be deduced from the following theorem (cf.
Balakrishnan [9, Section 1.4]).
Theorem 1 If h is an arbitrary element of a real Hilbert space 1l and C is
a closed convex cone in 1l, then there exists a unique projection Ph of h
onto C that is characterized by two relations

VgEC (g, h) < (g, Ph), (2.2)

(Ph,h) = (Ph,Ph). (2.3)

This is a refinement of the statement that there is a uniquely defined pro-

jection Ph of arbitrary h E 1l onto a closed convex set C C 1l, and Ph
satisfies
v gEe (g, h - Ph) ~ (Ph, h - Ph) (2.4)
(see Balakrishnan [9, Section 1.4]). The projection is the only point of C
that satisfies (2.4).
Observe that for Ph :F 0 relation (2.2) combined with (2.1) gives

v O:F 9 E C Th(g)/llgll ~ IIPhl1 > O. (2.5)

Setting 9 = aPh for some positive a and using (2.3), the equality holds
in (2.5). If Ph = 0 then, due to (2.2), Th is nonpositive on C, and clearly
Th(Ph) = O. Due to the fundamental significance of Theorem 1 for our
further considerations, we recall its proof here.

PROOF OF THEOREM 1. We first show that for every h E 1l there exists

a unique Ph E C that minimizes distance Ilg - hll over all 9 E C . The claim
is trivial if h E C. Otherwise there exists a sequence gn E C, n ~ 1, such
that
lim IIgn - hll = inf Ilg - hll = D > 0,
n~oo gEC

say. For arbitrary two elements of the sequence we have

Ilgn; h _ gm 2- hl1
2
+ Ilgn; h + gm 2- hl1 2

= ~(lIgn - hW + Ilgm - hW)· (2.6)

 2.1 Elements of Hilbert Space Theory 13

For arbitrary f > 0, the right-hand side of (2.6) is not greater than D2 + f
if nand m are large enough. Since (gn + gm)/2 E C, the latter term of the
left-hand side is not less than D2. Therefore we have

which, by completeness of 1£ and closedness of C, implies that gn, n ~ 1,

has a unique limit Ph E C, say. Relations

Ilign - hll - IIPh - hili::; IIgn - Phil -+ 0, as n -+ 00,

the former being concluded from the triangle inequality, imply that actually
IIPh-hll =D.
Now we verify that (2.4) completely characterizes Ph. For arbitrary gEe
we define a function Dg on the unit interval [0,1] as
Dg(O:) = IIh - (1 - o:)Ph - o:gll2
= IIh - Phll 2 + 20:(h - Ph, Ph - g) + 0:2 II Ph _ g1l2.
We have

D~(O) = 2(h - Ph,Ph - g), (2.7)

D~(o:) 211 Ph - gll2 ~ O. (2.8)
By (2.8), relation (2.4), identical with nonnegativity of (2.7), is the nec-
essary and sufficient condition for the nondecreasing of both D~ and D g •
Therefore

Now we use the fact that C is a convex cone. Plugging 9 = o:Ph into (2.4),
we obtain
(0: - 1)(Ph, h - Ph) ::; 0, 0: ~ 0,
which yields (Ph, h - Ph) = O. This proves (2.3), and in combination
with (2.4) gives (2.2) •

Here we concentrate on functionals on Hilbert space £2([a, d), w(x) dx),

for some -00 < a < d::; +00, w : [a, d) I-t !R+. The space consists of square
integrable functions on interval [a, d) with a positive weight function w, and
the inner product is defined by

(g, h) = ld g(x)h(x)w(x) dx. (2.9)

Here the Schwarz inequality takes on the form

]1/2
fa d
g(x)h(x)w(x) dx::;
[
fa d
g2(x)w(x) dx fa d
h 2(x)w(x) dx (2.10)
14 2. Basic Notions

Below we present two examples of projections onto convex cones contained

in L2([a, d), w(x) dx). The former has a simple solution. In the latter, the
form of projection is more complicated and depends on the weight func-
tion w.

EXAMPLE 2. Let

C+ = {h E L2([a,d),w(x)dx): h ~ a}.
Verifying (2.2) and (2.3) we deduce that

P+h = h+ = max{h,O}
is the projection of h onto C+ for every h E 1-£. Also, one can check di-
rectly that h+ is actually the nonnegative function closest to h for arbitrary
weight w . •

EXAMPLE 3. Consider the set C~ of all nondecreasing functions in the

Hilbert space L2([a, d), w(x) dx). We assume that the constant and linear
functions belong to the space. Then

W(y) = 111 w(x) dx (2.11)

is a finite, strictly increasing, and absolutely continuous function. The same

holds for its well-defined inverse W-l : [0, W(d)) I-t lR+. By finiteness of

(d (wed)
(h,l) = 1a h(x)w(x) dx = 10 hW-1(x) dx, (2.12)

we can define an absolutely continuous function

Hw(y) = foil hW-1(x) dx, °~ y < W(d), (2.13)

and its greatest convex minorant Hw, with a nondecreasing derivative hw,
say.
We prove that h"- defined as h"- = hw W E C~ is the projection of h
onto C~ by checking (2.2) and (2.3). For the former one, we need the
following lemma (cf. Marshall and Olkin [55, Proposition A.2.(iii) , p. 444]).
Lemma 1 If G ~ H are functions of bounded variation on an interval
[A, D), which are equal at the endpoints, then

LD g(x) G(dx) ~ LD g(x) H(dx) (2.14)

holds for every non decreasing function 9 for which both the integrals exist.
 2.1 Elements of Hilbert Space Theory 15

PROOF (cf. Marshall and Proschan [56]). The statement is easily verified
for all indicator functions l[y,D)(X), A < y < D. Therefore this is true for
all positive combinations of the indicator functions, and, by the Lebesgue
monotone convergence theorem, for arbitrary nonnegative nondecreasing
functions as well. Therefore the reversed inequality holds for the nonnega-
tive nonincreasing functions. Finally, we represent an arbitrary nondecreas-
ing function as the difference of two nonnegative terms, the nonincreasing
and nondecreasing ones, and apply the above statements to both parts . •

Since Hw ~ Hw satisfy the assumptions, for arbitrary 9 E C~ we can

write

(g, h) =
l o
W (d)
gW-1(y)hW-1(y) dy

=
l o
W (d)
gW-1(y) Hw(dy)

<
l o
W (d)
gW-1(y) Hw(dy)

= lad g(x)h'(x)w(x) dx

= (g,h,). (2.15)

In order to derive (2.3), we first thoroughly analyze relations in pairs

Hw,Hw, and hW-l,hw, and h,h,. Observe that the open set {Hw <
Hw} is a (possibly empty) union of countably many (at most) disjoint
open intervals, Ui(W(b i ), W(Ci)), say. The function Hw is linear on each
interval, and coincides with Hw at the endpoints. Therefore

and we have
16 2. Basic Notions

Thus the equality holds for the integrals over the whole of the open set
W-l({Hw < Hw}). For the remaining part W- 1 ({Hw = Hw}) the con-
clusion holds, since Hw = Hw implies that h" = h there. Summing up,
we have
(h,h,,) = (h",h,,),
which is the desired conclusion . •
The construction of the L2-projection onto the family of monotone func-
tions under uniform weighting was presented in Moriguti [58]. For the gen-
eral case we refer to Rychlik [85]. For simplicity, we treated elements of
L2-spaces as functions rather than equivalence classes up to almost sure
equality, and we follow this convention later. For example, h E C+ gener-
ally means almost sure nonnegativity, and monotonicity can be precisely
defined by comparisons of integrals over the intervals of the same weight.
In Examples 2 and 3 we were able to determine projections for arbitrary h,
but there are no general rules for constructing projections onto other con-
vex cones. Then we apply arguments suitable for specific functions h and
cones. Usually, we first try to describe the shape of the projection in a para-
metric way, and then precisely determine the parameters. We finally note
that more often the characterization (2.2) and (2.3) is used for determining
projections of principal interest. In the problems under study we first try to
find the projection by minimizing the distance of the Hilbert space point,
representing a functional, to a cone. The solution of the auxiliary problem
satisfies (2.2) and (2.3) in particular which are needed for estimating values
of the functional over the cone.

2.2 Statistical Linear Functionals

Investigations of statistical procedures treated as functionals on distribu-
tion functions were initialized by von Mises [101]. The general theory is
presented in Serfling [95] and Prakasa Rao [73]. Here we confine ourselves
to some statistical functions that can be represented as linear functionals on
Hilbert spaces. Assume that a random variable X has a distribution func-
tion F, finite mean J.I. = J.l.F, and second raw moment m 2 = m}. Changing
variables, we write

(2.16)

(2.17)

where
F- 1 (x) = sup{y: F(y):::; x}, 0:::; x < 1, (2.18)
 2.2 Statistical Linear Functionals 17

is the right-continuous quantile function of X. We can say that mF is

the norm of F-l in L2([0, 1), dx), and J.'F = (F-l, I). The family of
all possible quantile functions is identical to the convex cone of (right-
continuous versions of) nondecreasing functions in L2 ([0, 1), dx). For the
variance u 2 = u~, we have

Observe that the functions F-l- J.'F form the convex cone of nondecreasing
functions integrating to O. Our purpose is to determine sharp bounds for
normalized statistical functionals represented as T(F-l ) /mF and T(F- l -
J.'F)/UF for general and restricted classes of quantile functions. Now we
present exemplary linear functionals of statistical importance acting on
quantile functions in L2([0, 1), dx). Considering bounds on narrower classes
of distributions, it is convenient to transform the quantile functions so that
other L2-spaces are studied.
Quantiles F-l(P) of order 0 < p < 1. They characterize distributions by
describing levels that divide respective populations into subsets contain-
ing desired proportions of elements. Quantiles are often used for defining
critical levels of tests and interval estimates. In order to evaluate F-l(P)
in terms of moments, we represent it as a limit of continuous linear L2_
functionals

F-l(P) = q'"p
lim -I-lq F-l(x) dx = lim (F-l, _I_I(p q»),
q- p p q'"p q- p ,
(2.20)

where IA denotes the indicator function of set A. Precisely, a distribu-

tion function may have nonunique quantiles, and (2.20) defines the upper
quantile of order p.

Expectations of order statistics of independent samples. Let Xl' ... ' Xn

be independent identically distributed random variables with a common
distribution function F. Then the jth order statistic Xj:n, 1 $ j $ n, is the
jth smallest value in the sequence Xl' ... ' X n . The respective distribution
function is
P(Xj:n $ x) = P(at least j among Xl, ... ,Xn are $ x)

= ~ (~) Fk(x)[1 - F(x)]n-k

= Fj:nF(x), (2.21)
say. Since

f 1m ()
X = Fj:n
' (X ) = n (n -1)
j _ 1 X j-l( 1 - X )n- ,
j (2.22)
18 2. Basic Notions

we have

EFXj:n 1 +00

-00 x Fj:nF{dx)

= 11 F- 1 {x)/i:n{X) dx

= (F- 1 ,/i:n). (2.23)

Note that Fj:n and /i:n are the distribution and density functions, re-
spectively, of the jth order statistic of the standard uniform Li.d. sam-
ple of size n. Order statistics are directly used for estimating quantiles
of order j /n and describing the lifetime of the j -out-of-n reliability system
which contains n independent identical elements and operates until at least
n + 1 - j of its elements do. We can also study the expectations of linear
combinations of order statistics (so called L-statistics)

(2.24)

which have numerous applications in statistical inference. For example,

those with Ej=1 Cj = 1 (inchlding the sample mean l/n Ej=1 Xj:n, sample
median X(n+1)/2:n for odd n, trimmed means l/{n - 2k) Ej~:+1 Xj:n for
j ~ n/2-1) and ones satisfying E~=1 Cj ~ 0, 1 ~ k ~ n-1, and Ej=1 Cj =
o (including sample range Xn:n - Xl:n and sample interquartile distance
X n+1- Ln/ 4 J:n - X Ln/ 4 Jon) are used for estimating the location and dispersion
ofthe population, respectively. Moreover, the projection method enables us
to evaluate precisely uniform convergence rates of estimates for particular
families of distributions. In the exemplary case of quantile estimation, this
is possible by analyzing EFXj:n - F-l{j/n) and EF{Xj:n - X kj:kn ) for
k > 1. For a comprehensive treatment of the theory of order statistics
we refer the reader to David [22], Arnold et al. [7], and Balakrishnan and
Rao [12]. The best references for their applications are Balakrishnan and
Cohen [11], and Balakrishnan and Rao [13].

Order statistics of dependent samples. Suppose that Y 1 , • •. ,Yn are pos-

sibly dependent identically distributed random variables with a common
marginal F, and }jon, 1 ~ j ~ n, are the respective order statistics. As in
the previous case, }j:n may represent the lifetime of an {n + 1 - j)-out-
of-n system of elements with identical failure probability, but here each
element affects the other ones somehow. It may happen in particular that
Y1 = ... = Yn , if the damage of a single element causes the immedi-
ate damage of all remaining ones. It was shown in Rychlik [79] that for
 2.2 Statistical Linear Functionals 19

c = (Cl, ... , cn) E ~n, we have

(2.25)

where 'Pn(F) denotes the family of all joint distributions P on ~n with

identical marginals F, and ge is (the right) derivative of function Ge , being
the greatest convex one satisfying
j
Ge(O) =0, Ge(jln)~LCi, j=I, ... ,n. (2.26)
i=l

Formula (2.25) is valid for arbitrary coefficients C1, • •• , Cn, and distribution
function F with a finite expectation. The supremum is attained for some
distributions in 'Pn(F). A detailed characterization of the distributions as
well as arguments leading to (2.25) are presented in Section 5.1. Maximizing
(2.25) over a family of marginals F, we first determine the F providing the
extreme value of (F-1, ge), and then take the joint distributions in 'Pn(F) ,
for which the expectation of the L-statistic is actually equal to (F-1, ge) for
the specified F. By definition, ge is a nondecreasing step function with n-l
jumps at most, located at some points of the form j In, j = 1, ... ,n - 1. In
particular, for arbitrary 1 ~ j ~ n,

sup
PeP.. (F)
Eplj:n =
n
n
+1-
.1 1

J . (j-1)/n
F-1(x) dx

= (F- 1, n +~ _ j l[(j-1)/n,l) )

= (F- 1 , gj:n), (2.27)

say (cf. also Caraux and Gascuel [19], Rychlik [76]). One can more easily de-
termine projections of simple step functions presented in (2.25) and (2.27)
than of polynomials appearing in respective formulae (2.24) and (2.23) for
the LLd. case. This explains the surprising fact that we have more results
and of simpler forms for arbitrarily dependent samples than for standard
independent observations. Note finally that the projection method allows
us to measure sensitivity of L-statistics upon dependence by evaluating

(2.28)

in chosen classes of marginal distributions. These evaluations allow statis-

ticians to choose L-statistics that are robust against violations of indepen-
dence assumptions.
20 2. Basic Notions

Record values. Record values in numerical sequences are ones that exceed
all the preceding ones. For a random sequence Xi, j ~ 1, record values,
and respective record occurrence times L n , n ~ 1, are random increasing
sequences. By convention, we assume

(2.29)

and further put

Ln = min{j > L n- l : Xi > XL,._J, (2.30)

Rn = XL,., n ~ 1. (2.31)

Due to another convention, the first value of record occurs at time 1 and
equals Xl. Like extreme order statistics, records are applied in estimating
strength of materials, predicting natural disasters, sport achievements, and
the like. They were first studied by Chandler [20]. Comprehensive studies of
records can be found in Ahsanullah [1] and Arnold et al. [8]. IT observations
are independent identically distributed and the distribution does not have
an atom at its right support endpoint, then the sequence of records is
infinite almost surely. Formulae (2.29), (2.30), and (2.31) are well defined
for arbitrary original sequences, but we assume further that Xi, j ~ 1,
are independent and have an identical continuous distribution function F,
say. It is obvious that if a current value of a record is given, then the
conditional distribution of the next one is identical with the distribution of
the parent variable under the condition that it exceeds the actual record
value. In particular, for records R~ of an i.i.d. sequence xy, j ~ 1, with
the standard exponential distribution function

V(x) = 1 - exp( -x), x ~ 0,

we have

P(R~+l - R~ > YIR~ = x) P(X[ > Y + xlX[ > x)

= exp(-y)
= P(X[ > y) (2.32)
for arbitrary x, y ~ O. It follows that the record increments R%, Rf -
Rr; , . .. ,R~V - R~ , ... are independent standard exponential random vari-
ables, and Rn has the gamma distribution r(n + 1, 1) with shape parameter
n + 1 and scale 1. The transformation

produces an Li.d. sequence with common distribution function F. By strict

monotonicity, it preserves the record occurrence times. Therefore

(2.33)
 2.2 Statistical Linear Functionals 21

and
EFRn = EvF-1V{R~)
= (>0 F- 1 {1 _ e- xn e- z dx
Z)

10 n!

= 11
F-1{x)fn{x) dx

= (F-1,fn) (2.34)
with
fn{x) = [-In{l- x)]n In!,
which is a desired inner product representation.
kth record values. An increasing sequence of record values arises from
a nondecreasing sequence of sample maxima (Xn :n ), n ~ I, by crossing
out all repetitions. For arbitrary fixed k, the sequence of kth greatest order
statistics (Xn+1-k:n), n ~ k, is nondecreasing as well. By analogy, we can
define occurrence times and values of kth records in the following way
L(k)
0 = k, Itok) = X 1:k, (2.35)
L(k)
n = min{j > L~k21 : Xj > X L(Io)
n-l
H-k:L(Io) },
n-l
(2.36)
R(k)
n = X L~Io)+1_k:L~Io), n~1. (2.37)

The kth records were introduced by Dziubdziela and Kopocinski [26]. There
is also another convention of defining kth record occurrence times that
consists in subtracting k - 1 from L~k) defined in (2.35) and (2.36). This
implies that we start counting records only when the first k observations
are carried out. In particular, the first value of the kth record occurs at
moment 1 then. However, we are concerned with record values here, which
are not affected by the particular definitions of occurrence times. In contrast
with standard records, a random variable X L(Io) observed at the kth record
time L~k) for k ~ 2 does not necessarily become the kth record value
immediately. Generally, we have

XL~Io) = XL~Io)+1_i:L~Io) ~ XL~Io)+1_k:L~Io)

for some 1 ~ i ~ k. Thus it becomes the ith record first and will become the
kth record when k - i greater values occur in the sequence. A key relation
that allows us to establish the distribution function of kth records is

p{R(k)
nH
> yIR(k)
n
= x) = [1- F{Y)] k
1 _ F{x) , Y
> x, (2.38)

{cf. (2.32)), which means that the distribution of the next value of the kth
record, when the current one is known, is the same as the distribution of
22 2. Basic Notions

the minimum of k original variables Xj under the condition that they are
greater than the current record. Relation (2.38) shows that distributions of
nth values of kth records from the i.i.d. sample with distribution function F
and standard first records of the i.i.d. sample with distribution function

F1:kF(x) = 1- [1- F(x)]k

(Le., that of X 1:k) coincide. Therefore, by (2.34),

=
1
EFR~k) EV(F1:kF)-lV(R:;)
= 00
F- 1(1 _ e-z/k) xn e-Z dx
o n!
= 11 F-1(X)f~k)(x) dx

= (F-1 ,f~k»), (2.39)

where

k,~ 1, n ~ 0, (2.40)

is the density function of the nth value of the kth record of the LLd. stan-
dard uniform sequence, and fA 1) = fn defined in (2.34). Formula (2.38) is
true for arbitrary F. However, for the discontinuous F function F-1 F1-;-iV
is not strictly increasing and may transform a record in the exponential se-
quence into a number equalizing a previous score. Therefore (2.39) is true
only for continuous F.

kth lower record values. The consecutive decreasing elements S~k), n ~ 1,

of the nonincreasing sequence of kth order statistics Xk:i, i ~ k, are called
kth lower record values. H Xi, i ~ 1, are LLd. standard uniform, so are
1 - Xi, i ~ 1. If the upper records R~k) occur in the original sequence,
we observe the lower records S~k) in the latter. H F is continuous, then
transformation F-1 provides observations with distribution function F,
and preserves their order. Therefore we have
S(k)
n = F- 1(1 _ R(k»)
n ,

and, by (2.39) and (2.40),

EFS~k) = 11 F- 1(x)fAk)(1 - x) dx

= 1
o
1 kn+1
F-1(x) _ _ (_lnx)nxk-1 dx.
n!
 2.2 Statistical Linear Functionals 23

Conditional expectations of order statistics. Suppose that the common

distribution function F of Xl, ... ,Xn is absolutely continuous. Then the
conditional distribution of Xj:n given that Xi:n = y for some 1 :$ i < j :$ n
is the same as the distribution of the (j - i)th order statistic obtained from
a sample of size n - i from a population with distribution function

D ( ) _ F(x) - F(y)
.c'11I x - I _ F(y) , x ~ y,

which is actually the distribution function of Xl provided that Xl > y.

In fact, order statistics X j : n , 1 :$ j :$ n, of continuous populations form
Markov chains with respect to j. This means that conditional distribution
of X j :n when X i1 :n = Yl < ... < Xik:n = Yk, 1 :$ i l < ... < ik < j :$ n,
depends only on the value of Xik:n. The relations are useful in predicting
further failures of elements of reliability systems, and in statistical inference
based on partial (truncated) observations. The proofs, based on the use of
joint distributions of several order statistics, can be found in Arnold et
al. [7, Section 2.4]. In particular, we have

for 1 :$ i l < ... < ik < j :$ n, and

EF(Xj:nIXi:n = y)
=
11
EFIIIXj-i:n-i

= F-l(F(y) + [1 - F(Y)]X)/;-i:n-i(X) dx

= r
10
l
F-l( )f... . (x -
F(Y)) l[F(II),l)(X) dx
X }-I.n-I 1 - F(y) 1 - F(y)
(2.41)

for 1 :$ i < j :$ n. Similar results are obtained for the reversed conditioning.
The conditional distribution of the X j :n when Xk:n = z is given for k > j is
identical with the distribution of the jth order statistic from the sample of
size k -1 with the common distribution function of Xl under the condition
that Xl :$ z which equals

F lz( x ) = F(x)
F(z) , x:$ z.

H both Xi:n = Y and Xk:n = z for i < j < k are known, then the condi-
tional distribution of X j :n coincides with that of Xj-i:k-l-i of the doubly
truncated population with distribution function

p,IZ( ) = F(x) - F(y)

III X F(z) _ F(y) , y < x :$ z
24 2. Basic Notions

(see also Arnold et al. [7, Section 2.4]). Respective analogues of (2.41) are

EF(Xj:n!Xk:n = z)

11°
= EFlzXj:k-l

= F- 1 ( )f·. (~) l[o,F(z»(X) d (2.42)

x J.k-l F(z) F(z) x

EF(Xj:n!Xi:n = y, Xk:n = z)
= EFI,Xj-i:k-l-i

1°
III

1 -1 ( X - F(y) ) l[F(y),F(z»(X)
= F (X)!i-i:k-l-i F(z) _ F(y) F(z) _ F(y) dx (2.43)

for i < j < k and y < z. Order statistics of discontinuous populations

do not have the above-mentioned properties in general. Distribution the-
ory and dependence structure of order statistics from discrete populations
were thoroughly discussed in Nagaraja [62). Conditional expectations of
functions of adjacent order statistics for general populations were exam-
ined by Franco and Ruiz [30).
Conditional expectations of record statistics. We already referred to the
fact that the distribution of the (m + 1)st record value in an atomless
Li.d. sequence Xi, i ~ 1, under the condition that Rm = Y is that of the
original observation Xi under condition Xi > y. This is actually the record
value number 0 from the sequence with distribution function Fly. Looking
for further record values, we can merely confine ourselves to the X s that
exceed level y. The (m+2)nd record value is the first one among Xs greater
than y that is also greater than the first Xi exceeding y. In other words, this
is the first record value from the population with distribution Fly' Arguing
inductively, we conclude that the distribution of Rn under the condition
that Rm = y for some m < n is identical with the distribution of the
(n - m - l)st record value in an LLd. sequence with common distribution
F truncated on the left at y. Moreover, sequences of record values have
the Markov structure (see, e.g., Nevzorov and Balakrishnan [64, p. 527)).
Analogous properties have kth record values for general k. Justifying the
claim, it suffices to recall the fact that distributions of the kth and first
record values from distributions F and FuF = 1- (1- F)k, respectively,
coincide. One can easily check that

(FUF)IY(x) = [1 - (1- F)k)IY(x)

= 1- [l-F(X)f
1- F(y)
= 1 - [1 - Fly(x))k
= Fu(Fly)(x).
 2.3 Restricted Families of Distributions 25

Accordingly, for continuous parent distributions F, 1::; m < n, and k ~ 1,

we have

= EFl:kFlyRn-m-1
= EF R(k)
III n-m-l

r
10
1 F-1( )ik) (x - F(Y)) l[F(y),I)(X) d
x n-m-l 1 _ F(y) 1 _ F(y) x
r 1 k n- m (1 - x)k-l
10 F- 1(x) (n - m - 1)![1 - F(y)]k
1 - F(Y)] n-m-l
x [ In 1- x l[F(y),I) (x)dx. (2.44)

Predictions of further record values are of vital interest in constructing

protection devices. Reconstruction of previous record values is also possible.
Nagaraja [61] proved that in the LLd. samples with continuous distribution
function F, the conditional distribution of Rm under the condition Rn = z
is identical with the distribution of the mth order statistic from the Li.d.
sample of size n - 1 ~ m with the distribution function

p1Z()= V- 1F(min{x,z})
x V-IF(z)'
Accordingly, elementary calculations lead us to the linear functional repre-
sentation

EF(RmIRn = z)
[Link] :n -

11
1

= (pIZ)-1 (x)fm:n-l (x) dx

=
rF
10
1 -1
(x)fm:n-l
( -In(I-X)) l[o,F(z)) (x)
-In(1 _ F(z)) -(1 _ x) In(1 _ F(z)) dx.

Analysis of the functional is certainly a challenging task. Franco and Ruiz

[29] studied conditional expectations of functions of the nearest previous
record.

2.3 Restricted Families of Distributions

We formally define the convex cone of quantile functions considered in the
previous section as

C/' = {g E £2([0, 1),dx) : 9 - nondecreasing, right continuous}. (2.45)

26 2. Basic Notions

Right continuity assumption is inessential for problems of L2- projections.

It is introduced here in order to preserve the consistency with definitions
of quantile functions, and One can simply replace an arbitrary nondecreas-
ing function by the right continuous version. We regard elements of COnvex
COnes defined here and statistical functionals as being right continuous.
Observe that projecting functions onto (2.45) we maximize normalized sta-
tistical functionals in 11F- 1 11 = mF units. In order to get more subtle
evaluations in terms of mean and variance, we should consider the class of
p-l - [Link] functions defined as

CO = {g E c/': 11 g(x)dx = O}. (2.46)

It is also of interest to study life distributions, which, by definition, have

the left support endpoint at zero, and generate the following family of
nonnegative quantile functions

C+ = {g E C/' : g(O) = O}. (2.47)

In this case we have another pair of location and scale parameters: the left
endpoint of support a F = 0 and the square root of the second moment m F,
respectively.

The family of symmetric distributions (about the respective expectation

J.L = [Link]) is characterized by equivalent relations
p(x - J.L) = 1- P(J.L - x-), (2.48)
p-1(x)_J.L = -[F-1(I-x-)-J.L]. (2.49)

It is convenient to study the upper halves of p-1 - [Link], which form the
cone

CS = {g E L2([~, 1), 2dx) : 9 - nondecreasing, 9 (~) = O} , (2.50)

and extend the functions to the whole unit interval using (2.49). This im-
plies the modification of functionals which consists in symmetrizing them
about 1/2. Indeed, by (2.49), this yields

T h (F-1 - J.L) = 11 [F-1(x) - J.L]h(x) dx

= r1
J1/2
2[p-1(X) _ J.L]hS(x) dx (2.51)

for
h 8 (x) = [h(x) - h(l- x-)]/2. (2.52)
Observe that the norm of p-1 - [Link] in L 2 ([1/2, 1), 2dx) is aF.
 2.3 Restricted Families of Distributions 27

In various models, existence and monotonicity of the density function

of the parent distribution F is assumed. H the density is nonincreasing
(nondecreasing) then F is increasing concave (convex) on its support, and
F-1 is increasing convex (concave, respectively). We define
cC u =
_c
{g E C/ : 9 - convex}, (2.53)
C~u = {g E C/ : 9 - concave}, (2.54)

°
with C~ u (C~ u) and C; u (C~ u) denoting intersections of (2.53) «2.54),
respecti~ely) "Vrith linear-~ubsp~es of functions integrating to (vanishing
at 0, respectively). The apparently awkward notation is justified below. We
say that F(x) succeeds the standard uniform distribution function
U(x) = x, ° ~ x ~ 1,
in the convex order (written as F tc U) if F- 1U = F-1 is convex on [0,1).
The reversed relation F :jc U is defined by the convexity of U- 1 F = F on
the support of F, being equivalent with the concavity of F-1. By conven-
tion, we call distributions satisfying F tc U and F :jc U the decreasing
and increasing density distributions, respectively. The convex order, de-
fined in van Zwet [100], is a partial order of (absolutely) continuous distri-
bution functions, invariant under location and scale transformations (see
Dharmadhikari and Joag-dev [25, Theorem 9.1, p. 217]). Therefore we can
generalize (2.53) and (2.54) as
C{.,w = {g E L 2([aw,dw),w(x)dx) : 9 - nondecreasing
and convex}, (2.55)
C~w = 2
{g E L ([aw,dw),w(x)dx) : 9 - nondecreasing
and concave}, (2.56)
by taking compositions F-1 W for an arbitrarily fixed distribution func-
tion W with support [a,d) = [aw,d w ), and density w. We also introduce
convex cones C~cw (C~cw) and C~cw (C~cw) by adding conditions

ld g(x)w(x) dx = 0,
g(a) = 0,
respectively, to definition (2.55) «2.56), respectively). These definitions are
justified by properties of compositions of F- 1 with W. In particular, we
have

ld[F-1W(y)]2W(y) dy = 11 [F-1(XW dx < 00, (2.57)

ld[F-1W(y) - IJ]w(y)dy = 11 [F-1(X) - IJ] dx = 0, (2.58)

F- 1W(a) = F- 1(0). (2.59)

28 2. Basic Notions

The families of life distributions being in convex order with exponential

distribution are studied in reliability theory. Observe that F ~c V implies
convexity of the hazard junction

V-l F(x) = -In[l- F(x)],

and nondecrease of

Function AF(X), called the failure (hazard) rate, describes the infinitesi-
mal probability of failure at a short period of time following x under the
condition of surviving until x. Condition F ~c V defines the family of dis-
tributions with increasing failure rate (IFR, for short). Likewise, F tc V
coincides with nonincrease of AF, and defines the family of decreasing failure
rate (DFR) distributions. We specify here bounds on statistical functionals
restricted to IFR and DFR distributions. Since U -<c V, every DFR distri-
bution has a decreasing density, and the increase of density implies that of
the failure rate.
van Zwet [100] (see also Lawrence [49]) defined a counterpart ~. of
convex order for symmetric distributions, and called it s-order. We have
F ~8 W for F and W symmetric about [Link] and [Link], respectively, if
F-lW(x) is concave for x ~ [Link]. This is equivalent with convexity of
F-l W on the lower half of the support of W, and concavity-convexity
of W- l F on the support of F. Combining constructions leading to (2.50),
(2.55), and (2.56), we define the convex cones

CtcW = {g E L 2 ([[Link],dw),2w(x)dx) : g([Link]) = 0, g - nondecreasing,

convex}, (2.60)

and C~ w replacing the convexity assumption by concavity in (2.60). It is

worth-pointing out here that F t. (~.)U define the classes of symmetric
unimodal (U -shaped) distributions. Multiplying the weight function by 2 (cf.
also (2.50)) allows us to confine ourselves to L2-spaces defined by means
of probabilistic measures. Obviously, the constant multiplicator does not
affect projections.
The star order ~. of continuous life distributions, introduced in Barlow
and Proschan [15], is more general partial order than the convex one (see,
e.g., Dharmadhikari and Joag-dev [25, Theorem 9.1, p. 217]). By definition,
F ~. W iff W- l F is starshapedj that is, W-l F(x)Jx is nondecreasing on
the support of F. If F ~* V in particular, then

AF(X) = -In[l -
x
F(x)] =.!.
x
r AF(Y) dy
10 (2.61)
 2.3 Restricted Families of Distributions 29

is nondecreasing, and we can say that P has an increasing failure rate on

the average (IFRA, for brevity). DFRA life distributions P are defined by
relation P !:::* V. Also, P ~* (!:::*)U means that

P(x)
x
=.! r f(y) dy
x Jo
is nondecreasing (nonincreasing, respectively). Accordingly, the relation de-
fines the family of life distributions with increasing (decreasing) density on
the average: although f may be multimodal, the larger values in [aF, dF)
are more (less) probable than the smaller ones. The star order is scale in-
variant. In order to make it invariant with respect to translations as well,
we generalize the definition as follows: P !:::* (~*)W iff aF,aW are finite
and [F-IW(X) - p-lW(aw)]/(x - aw) is nondecreasing (nonincreasing)
on [aw, dw). The definition enables to establish mean-variance bounds on
statistical functionals by projecting them onto convex cones of p- 1 W - [Link]
described by the generalized star relations:

_.
c~ w = {g E L2([aw,dw),w(x)dx) : g(x), g(x) - g(aw) are
x-aw
nondecreasing and

[d W g(x)w(x) dx = O}, (2.62)

Jaw
_.
c~ w {g E L2([aw,dw),w(x)dx) : g(x) isnondecreasing,
g(x) - g(aw) is nonincreasing,
x-aw
[d W g(x)w(x) dx = O}. (2.63)
Jaw
Special emphasis is laid on cases W = U, V. For a detailed treatment of
stochastic orders and their applications, we refer the reader to monographs
of Dharmadhikari and Joag-dev [25], and Shaked and Shantikumar [96].

We adopt the convention of denoting the projection onto a given convex

cone by writing P with the same upper and lower indices that appear in
the notation of the cone. For example, P~ wh denotes the projection of h
onto C~. w. Although the original quantile functions of restricted families of
distributions determined by the orders discussed above form convex cones,
we prefer considering compositions p-l W. The reason is that the latter
have natural analytic and geometric properties. Our process of determining
the projection of a given functional consists of choosing an arbitrary start-
ing point in the cone and constructing consecutive approximations improv-
ing the previous ones. We first try to describe the shape of the projection
function and then calculate optimal parameters. It is therefore essential for
the first step that we are able to check immediately if a function proposed
30 2. Basic Notions

for an approximation actually belongs to the cone. The only assumption

appearing in definitions of our cones that cannot be verified at first glance
is weighted integrability to 0 (see, e.g., (2.46), (2.60), (2.62), and (2.63)).
One can overcome the problem using the following lemma (cf. Rychlik [87,
Lemma 1]).
Lemma 2 Suppose that C is a subset of a real Hilbert space 11. such that
gEe implies 9 + cgo E C for some go E 11. and all real c. If ho E 11. has
a projection Pho onto C, then
(Ph o, go) = (ho, go).
PROOF. For arbitrary fixed gEe and c E lR, IIg+cgo-hoI1 2 is minimized
by
eo = eo(g) = (h o - g,go).
(go,go)
The projection Pho necessarily has the form 9 + eogo for some gEe, and
satisfies
(g + eogo, go) = (ho,go) .•
Let C~ be any of the above-considered convex cones such that

ld g(x)w(x) dx = (g,l) = 0
for all 9 E C~ is assumed. Let C( denote the extension of C~ by dropping the
integral condition. Note that each C( is translation invariant (i.e., fulfills
the assumption of Lemma 2 with go = 1). Therefore (h, 1) = 0 implies that
PI' h E C~ and coincides with p2 h. Otherwise we replace h by
(h,l)
ho =h- (1,1/ =h - (h,I), (2.64)

noting that we confined ourselves to weights generated by probability mea-

sures. Then, by (2.2) and (2.3),

V 9 E C~ Th(g) = (g,h) = (g,h- (h,I)) = Tho (g) ::; 11P2hollllgll,

Th(p2ho) = Tho(P2ho) = 11P2ho1l 2 •
Since (ho,l) = 0, we have ~ho = P!'ho. It follows that it suffices to
replace functional h by (2.64) and project it onto C(, without bothering
about the integral condition. In fact, translation invariance of C( yields

P!'(ho) = P!,(h) - (h,l)

so that we are reduced to projecting the original h onto C(, and subtracting
the constant from the result.
 2.3 Restricted Families of Distributions 31

Another problem that one should be aware of is the existence of pro-

jections and probability distributions attaining bounds on functionals in
restricted families. Functional representations of expectations of records,
and conditional expectations of order and record statistics are valid for
absolutely continuous distributions. Moreover, the families of absolutely
continuous distributions determined by the convex and star orders form
convex cones that are not closed. In fact, the cones of absolutely continu-
ous distributions are dense in the respective convex cones defined in this
section. We project the functionals onto the closed convex cones, and the
projections are border points of them. In many cases they correspond to
distributions that are not absolutely continuous, and so do not belong to
the families under study. Nevertheless, we write down formulae describ-
ing these distributions, because they enable the reader to guess easily the
forms of absolutely continuous elements of restricted families with values of
functionals arbitrarily close to optimal bounds. Strictly speaking, the weak
convergence in the topological sense is sufficient here. However, we can eas-
ily find sequences F;l, k ~ 1, of quantile functions of absolutely continuous
distributions with assumed properties that tend to the desired limit F- 1 ,
say, in the norm of L2([0, 1), dx). Note that, by change of variables,

Fk-1W ~ F-1W in L 2 ([aw,dw),w(x)dx)

is equivalent to
F k- 1 ~ F- 1 in L2([0, l),dx).
More intuitively, we can express the relation in terms of random variables.
If Fk, k ~ 1, are distribution functions of a given family, and F generated
by the projection method does not belong to the family, then Fk, k ~ 1,
attains in the limit the bound attained by F if

that is, F;l(X) converges to F-1(X) in the mean square for a standard
uniform random variable X.
For instance, (2.53) contains the quantile functions of distributions with
decreasing density, and, possibly, an atom at the left endpoint ap of sup-
port. Suppose that the solution of the projection problem, dual to a problem
of evaluating a linear statistical functional over decreasing density distribu-
tions, has an atom. This can be replaced by absolutely continuous (uniform,
say) components with increasing concentration about ap so that we obtain
decreasing density distributions with quantile functions tending to that of F
in L2([0, 1), dx). Moreover, these absolutely continuous approximations can
be modified so as to preserve desired moments of limiting F.
3
Quantiles

The main results of this chapter come from Rychlik [90]. The bounds for
quantiles of general distributions were obtained by Moriguti [58]. Those for
symmetric and symmetric unimodal distributions may also be concluded
from the Chebyshev and Gauss inequalities, respectively. Vysochanskii and
Petunin [102] presented a refinement of the Gauss inequality for unimodal
distributions. Further generalizations can be found in Dharmadhikari and
Joag-dev [25, Section 1.5]. We also notice that the Markov inequality yields

F- 1 (P) ~ fJ.F
I-p

for quantiles of nonnegative random variables. Another implication of the

Markov inequality is the second moment bound

-1(p) mF
F ~ (I _ p)1/2 .

3.1 General and Symmetric Distributions

For these two cases, there is no need to think of a quantile value as
a limit of L2-functionals (cf. (2.20», because we can directly use results
of Moriguti [58], summarized below. These enable us to analyze integrals
of monotone functions with respect to arbitrary, not necessarily absolutely
continuous, distribution functions.
34 3. Quantiles

Lemma 3 If H : [a, d) t-+ ~ is right continuous and of bounded variation,

H- denotes its left continuous version, and li is the right derivative of the
greatest convex minorant fl of H, then

1 d
g(x) H(dx) ~ 1 d
g(x)li(x) dx ~
[
1d
g2(x) dx l d
li2 (x) dx
]1/2
(3.1)

for every nondecreasing function g.

The former relation in (3.1) becomes an equality iff 9 is constant in every
interval contained in {fl < min{H,H-n, and right (left) continuous at
every discontinuity point of H (if any) such that H > H- (H < H-,
respectively) there. The latter relation in (3.1) becomes an equality iff either
li = 0 or 9 = ali for a ~ O.
Also, under assumption H(a) = H-(d), function 9 can be replaced by
arbitrary translation 9 + c in the middle and last expressions of (3.1), and
the conditions for equality.
We tacitly assume that all integrals in (3.1) are well defined and finite.
Observe that the conditions for the latter equality imply those for the
former, because fl is linear and li is constant on every interval of the set
{fl < min{ H, H-}}. The arguments in the proof of Lemma 3 are similar to
those used in Example 3 of Section 2.2. Since we have a uniform weighting
here, introducing compositions with the anti derivative is redundant. More-
over, the first step of constructing the derivative li of the greatest convex
minorant fl of the antiderivative H of the projected function h also may
be omitted, because in Lemma 3 we start from a distribution function H.
On the other hand, this is not necessarily continuous here which results in
a more sophisticated condition for equality in (3.1).
Theorem 2 (general distributions) For all 0 < p < 1,
F-1(p) - [Link] ~ (_p_)1/2 (3.2)
(TF 1- p

The equality in (3.2) holds iff F is the two-point distribution valued at

+ [Pj(I - p)]1/2(T with probabilities p and 1- p,
J.L - [(1 - p)jp]1/2(T and J.L
respectively.
The theorem immediately follows from Lemma 3 by putting

H(x) = 1[p,1)(x) - U(x)

with

{ -1 if o ~ x <p,
li(x) = ~, if p ~ x < 1,
P
Illil1 2 ,
I-p
 3.1 General and Symmetric Distributions 35

and
F-l(X) - J1. h(x)
a = Ilhll·
Moriguti [58] also showed that
F- 1 (q) _ F-l(P) (1 1)1/2
--='------"'--'-
aF
< -- + -
- 1- q p , O<p<q<l. (3.3)

Special cases of (3.2) and (3.3) for the median and interquartile distance
yield

F- 1 (~) - J1.F < aF,

F- 1 (~) _ F- 1 (~) < 23 / 2 aF.

Likewise for symmetric distributions and 1/2 ~ p < 1, we can write

F- 1 (P) _ J1. < /,11/2 (F-l(x) - J1.] 1[p,1) (dx)
< /,11/2 (F-l(X) _ J1.]l[p,l)'(x)dx
< 2- 1 / 2 11 1 [p,l)'lla
[2(1 - p)r 1 / 2 a, (3.4)
because
-- x-p
l[p,I)(X) = 1- P l[p,l) (x)
is the greatest convex minorant of l[p,I). For p < 1/2 we trivially obtain

F- 1 (P) - J1.F ~ F- 1 (P) - F- 1 (~) ~ O.

Theorem 3 (symmetric distributions) We have F-l(P) ~ J1.F for ev-
ery quantile of order p < 1/2.
For p ~ 1/2, we have
F-l(P) - J1.F 1
aF ~ [2(1 _ p)]1/2 ' (3.5)
with the equality attained in (3.5) for the three-point distribution

p (X = J1. ± [2(1 :P)]1/2) = 1 - p,

P(X = J1.) = 2p-l.
Observe that for p = 1/2 bounds (3.2) and (3.5) coincide, and (3.5) is
sharper for p > 1/2.
36 3. Quantiles

3.2 Distributions with Monotone Density and

Failure Rate
We start with the problem of evaluating the upper quantiles F- 1 (P) for
all F !:c W. Our auxiliary projection problem is to find a function in
C( w minimizing the distance to the indicator function of the interval
[W- 1 (P), W- 1 (q)) for some q > p. The construction is based on the follow-
ing two lemmas.
Lemma 4 Let

h = M1[b,c) E L 2 ([aw,d w ),w(x)dx)

for a < b < c ::; d and M > O. Then for every 9 E C( w there exists
ga.{3 E C(
_c
w defined as -

(3.6)

for some a :2: 0, 13 ::; b, a(b - 13) < M, such that

Ilga.{3 - hll ::; IIg - hll·

The two-parametric class (3.6) consists of nonnegative constants, increasing
linear functions, and broken lines with two pieces: horizontal and increasing
ones.

PROOF OF LEMMA 4. Note first that if g(a) < 0, then g+ = max{g, O} E

C(w lies closer to h than the original g. If g(a) :2: 0 and g(b) :2: h(b),
toen h ::; 9 :j:. h and (g,l) > (h, 1). Since IIg + C - hll is minimized in the
class of translations 9 + C, of a given function 9 by the real C satisfying
(g + C,l) = (h,l) (cf. Lemma 2), we improve approximation of h by
adding a negative constant and truncating at level O.
These modifications lead us to functions satisfying 0 ::; g(a) ::; g(b) <
h(b). Function 9 = 0 can be replaced by a positive constant. A correction
of a nonnegative nonzero function 9 E C( w that vanishes at c consists in
setting 0 in [c, d), and adding a positive cC:nstant. Therefore we can further
assume that g(c) > 0, and define

"y = inf{x E (b,c] : g(x) :2: h(x)}.

We easily check that the function

g-y(x) = max {gb~ =~(b) (x - b) + 9(b),0}

lies between hand g, and can be alternatively parameterized as (3.6) with
respective restrictions on parameters .•
 3.2 Distributions with Monotone Density and Failure Rate 37

Lemma 5 Let the assumptions of Lemma 4 hold and

ld w(x) dx = ld h(x)w(x) dx = 1. (3.7)

If
JbC(X - a)w(x) dx
fC ( )
Jb W x dx
~
a
d( ) ( )l
x - a w x dx, (3.8)

then p( wh = 1.
Othe1wise there exists a unique fJ* < b that solves

JbC(X - fJ)w(x) dx J~ax{{3,a}(x - fJ)2w(X) dx

C = d ~.~
Jb w(x) dx Jmax{{3,a}(x - fJ)w(x) dx

and

a* = a*(fJ*) = h.
[ d(x - fJ*)w(x) dx
]-1
>0 (3.10)

such that
/' () _ ( _ (x - fJ*)+ (3.11)
Pt,c wh x -a. x-fJ*)+ - d' •
Jmax{{3.,a}(y - fJ*)w(y) dy
Precisely, if relation "~" holds in (3.9) with fJ replaced by a, then fJ* ~ a,
and (3.11) is linear on [a, d). In the opposite case, fJ* > a and the projection
is actually a two-piece broken line.
PROOF. Our purpose is to minimize

D(a,fJ) = Ilga{3-hW

= a2 1d (x - fJ)2w(x) dx

r
max{{3,a}
2a 1
- Jbc w(x) dx Jb (x - fJ)w(x) dx + Jbc w(x) dx (3.12)

with respect to a and fJ. For fixed fJ < b, we easily find optimal

JbC(x - fJ)w(x) dx
a* = a* (fJ) = d > 0, (3.13)
J: w(x) dx Jmax{{3,a}(x - fJ)2w(x) dx
which plugged into (3.12) gives
38 3. Quantiles

By differentiation,

dD(a.(j3),j3) 2 fbC(x - j3)w(x) dx

=
dj3 [fbc w(x) dxJ2 f~ax{/3,a} (x - j3)2w(X) dx

x [l b
C
w(x) dx
d
r
lmax{/3,a}
(x - j3)2w(X) dx

- 1r d

max{/3,a}
(x - j3)w(x) dx r (x -
1b
j3)w(x) dxl.

Since the factor in the first line is positive, the sign of the derivative is
identical with that of the last two which is denoted by K(j3). Observe that

- Ew(X - blX > b)Ew(X - bib < X < c)

> Varw(XIX > b) > 0 (3.15)

for a random variable X with distribution function W. Furthermore,

K'(j3) = r w(x) dx lbr (x - j3)w(x) dx

d

lmax{/3,a}

r (x - j3)w(x) dx
- lbr w(x) dx lmax{/3,a} d

evaluated at j3 = b satisfies
K'(b)
d = Ew(Xlb < X < c) - Ew(XIX > b) < O. (3.16)
fb w(x) dx fbc w(x) dx

l
Finally,
K"(j3) = -w(j3)
c
(x - j3)w(x) dx ~0 (3.17)

under the convention that w(j3) = 0 for j3 < aw. By (3.17) and (3.16),
K'(j3) is constant for j3 < a and nonincreasing to K'(b) < 0 for a ~ j3 < b.
If K'(a) ~ 0, which means that (3.8) holds, then K'(j3) ~ 0 and K(j3)
is nonincreasing and, by (3.15), positive for all j3 ~ b. Therefore (3.14) is
minimized at j3 = -00, which implies that the projection is a constant.
This amounts to (h,l) = 1, which proves the first assertion.
If K'(a) > 0, then K(j3) changes its sign once from minus to plus at
some j3 < b, at which (3.14) is minimized. Equations (3.9) and K(j3) = 0
are equivalent, and allow us to rewrite (3.13) as (3.10). Note that K(a) 2:: 0
 3.2 Distributions with Monotone Density and Failure Rate 39

implies that the solution to (3.9) satisfies (3 $ a, and that the projection is
linear. In the opposite case a broken line with break at (3 > a is obtained.
This proves the final claim of Lemma 5. •
Using (3.11) and writing (3
'-*
= max{(3.,a}, we obtain
IIPgcwh -IW = IIPSw h 11 2 - 1

= -.
I; (x - (3.)2W(X) dx - [J;-. (x - (3.)w(x) dX]2
[J:. (x - (3.)w(x) dx]2 (3.18)

P~cwh(x) -1
IlPgcw h -III
(x - (3.)+ - I; (y - (3.)w(y) dy
= {d -. d }1/2' (3.19)
It!.. (y - (3.)2 W(Y) dy - [It!.. (y - (3.)w(y) dy]2
Letting c ~ b, we reduce the right-hand sides of (3.8) and (3.9) to b - a
and b - (3, respectively. Analyzing the resulting formulae, we derive bounds
for quantiles. We introduce some notation before presenting them. For
a random variable X with distribution function W, and arbitrary (3 E (a, d),
denote the expectation and variance of X under the condition of exceeding
level (3 by
I;xw(x)dx
J.'w ((3) = Ew(XIX > (3) = d ' (3.20)
1/3 w(x) dx
= Varw(XIX > (3)
I; x 2w(x)dx - I-''tv((3)
=
I; w(x) dx (3.21)

We also define a distance between the pth quantile and the conditional
mean in respective standard deviation units
8 ((3) = W-1(p) - I-'w((3) (3.22)
w ow ((3) .
Theorem 4 (P tc W) IfW- 1 (p) $ I-'w, then p-l(p) $ I-'F.
If J.'w < W-l(p) $ J.'w + u~/(I-'w - aw), then
p-l(p) - I-'F < r ( ) _ W- 1 (p) - I-'w
_ uw aw - , (3.23)
UF UW
and the equality holds for the location-scale transformations of W, that is,
for
P(x) = W ( I-'w + uw- X-I-')
u- . (3.24)
40 3. Quantiles

Finally, if W-l(P) > [Link] + a'tv/([Link] - aw), then there exists a unique
solution (3* = (3*(P) E (aw, W-l(P)) to equation
[W- 1 (P) - [Link] «(3)][[Link] «(3) - (3] = a'tv«(3), (3.25)
and

F-l(P) - [Link] < A = AO (P) = [8'tv«(3*) + W«(3*)] 1/2 (3.26)

aF - tcW 1- W«(3*)

The bound in (3.26) is attained by

F(x) =W ((3* + Utw«(3*) - (3*][1- W«(3*)] (1 + A7)) l[/J_q/A,oo) (x).
(3.27)
Distribution function (3.27) has a jump of height W«(3*) < p at the left end-
point, and shares the shape of W on its support. This is a life distribution
if

Mean-variance bounds for F !:c U, V are specified in Propositions 1 and 2,

respectively.
Proposition 1 (decreasing density) If 0 < p::; 1/2, then F-l(P) ::; [Link]
holds.
If 1/2 < p ::; 2/3, then

F-l~~ - [Link] ::; 2\1'3 (p _ ~) , (3.28)

which becomes an equality for the uniform distribution on [I' - .;aa, I' +
.;aa].
If 2/3 < p < 1, then

F-l(P) - [Link] < [ 9p - 5 ] 1/2 (3.29)

aF - 9(1- p)
This is an equality for the mixture of the Dirac distribution concentrated at
I' - 3a[(1 - p)/(9p - 5)]1/2 and the uniform one on [I' - 3a[(l - p)/(9p-
5)]1/2, I' + 3a(3p-1)/[(l- p)(9p- 5)]1/2] with weights 3p - 2 and 3(1- p),
respectively.
Proposition 2 (decreasing failure rate) If 0 < p::; 1-e- 1 Rl 0.63212,
then F-l(P) ::; [Link].
If 1 - e- 1 < p ::; 1 - e- 2 Rl 0.86466, then bound
[F-l(P) - [Link]]/aF ::; -In(l - p) - 1, (3.30)
is attained by the exponential distribution function with location I' and
scale a.
 3.2 Distributions with Monotone Density and Failure Rate 41

If 1 - e- 2 < p < 1, then for

'Y = 'Yv(P) = (1 - p)e 2 E (0,1),

we get
F- 1(P) - f,LF < (_2__ 1)1/2 (3.31)
O"F - 'Yv(P)
This is an equality if F is the combination of an atom at f,L-0"['Y/(2-'Y)]1/2,
and the exponential distribution with location h(2-'Y)j1/2f,L-'Y0" and scale 0",
with respective probabilities 1 - 'Y and 'Y.
We see that bounds (3.29) and (3.31) tend to infinity if p /' 1, and the
same holds generally for (3.26). It follows from the fact that for b = W- 1(P)
large enough, f3 = f3(b) satisfies

I;
x(x - f3)w(x) dx
b = -'--d.,------- (3.32)
1/3 (x - f3)w(x) dx
(cf. (3.9) for f3 > a). Therefore b(f3) /' d, as f3 /' d, and the same holds for
the inverse. Furthermore, 1 - W(f3) ~ 0, whereas the nominator of A~ w
remains bounded below from zero, as b /' d. _0

Now we proceed to the class of distributions determined by relation F :-;c

W. We look for the projection of an indicator function onto cone C( w. _0

Lemma 6 Under the hypotheses of Lemma 4, for every g E C( w there

exists gOti3'Y E C{o w defined as _0

(3.33)

for some a > 0, and a ::; [3 ::; b ::; 'Y ::; c with a(-y - [3) ::; M such that

PROOF. We start with eliminating all 9 E C~w satisfying g(c) ::; 0.

Obviously, 9 ::; g+ = max{g,O} ::; h i g+. The first two relations show

°
that g+ provides a better approximation. The latter two imply that this
can be further improved by adding a constant C > such that

(g+ + C, 1) = (h,I).
We can repeat the truncation and translation operations several times until
we eventually obtain a function positive at c.
42 3. Quantiles

Then we assume that g( e) > 0, and separately study three cases: g(b) :5 0,
g(a) :5 0 < g(b), and g(a) > O. In the first one, we construct the line
passing through (b,O) and tangent to the graph of gin (b, e], and truncate
it at level gee). The resulting function has the desired form, runs above 9
and below h in [a, e), and the reversed relations hold in [e, d). In the second
case, we take the linear function secant to 9 at (3 E [a, b) satisfying g((3) = 0
and b. The truncation defined above provides the conclusions of the former
case. Finally, if g(a) > 0, then it suffices to replace point ((3,0) by (a, 0) in
the construction of the second case .•

The function (3.33) is a linear increasing function on the left which

changes into a constant at point , E [b, e]. It is convenient to reparam-
eterize (3.33) as
gO:/3'Y(x) = a[min{x,,} - (3] + 1.
Then for fixed, E [b, e] we minimize the squared distance

D(a, (3,,) = l d
[a(min{x,,} - (3) + 1- h(x)]2w(x)dx (3.34)

with respect to a and (3. However, this is a simple matter, because (3.34)
is a convex quadratic function in both arguments. Therefore we have
Lemma 7 Under the hypotheses of Lemma 5, we have

p{
-<
wh(x) = a("{Hmin{x,,} - (3("{)] + 1 (3.35)

for some, E [b, e] with

(3("{) = ld min{x, ,}w(x) dx, (3.36)

L<min{:Il,'Y}w(:Il)d:ll _ (3( )
1< w(:Il)d:ll '
a(,) = (3.37)

In the description of quantile bounds for F ~c W, we use moments of

right truncated random variables with distribution function W (cf. (3.20)
to (3.22)):

= Ew min{X,,} = ld min{x,,}w(x) dx, (3.38)

= Varwmin{X,,}

= l d
[min{x,,}]2 W(X) dx - ti~("{), (3.39)

W-1(P) - tiw(,)
= (3.40)
OW("{)
 3.2 Distributions with Monotone Density and Failure Rate 43

Notice that (3.36) and (3.37) can be expressed as

fJ("{) = PW('Y) E [aw,'Y),

( ) Ew(min{X,'Y}IX> b) - PW('Y) 0
0: 'Y = -2 ( )
uw'Y
> ,
respectively. Therefore (3.35) is actually a nondecreasing concave function.
In order to deduce the optimal bounds on quantiles, we do not need to
calculate precisely parameter 'Y in Lemma 7, because we merely look for

g. = limP{ wh E C~ w.
c'-"b -" -"

Note that (3.36) and (3.37) tend to fJ(b) = Pw(b), and

o:(b) = d b - fJ(b) = W-l~~ - Pw(b) (3.41)

fa min2{x, b}w(x) dx - fJ2(b) uw(b)

respectively, uniformly in all 'Y E [b, e], and therefore p{ wh tends to

-"
g.(x) = o:(b)[min{x, b} - fJ(b)] +1 (3.42)
in the L2-norm as e ~ b. The norm

A~ w(p) = Ilg. - 111 = o:.(b)ow(b) = c5w(b) (3.43)

-"
(cf. (3.38) to (3.40), and (3.41), and (3.42» provides the optimal bound.
Writing
F-1W(X) - J-tF _ g.(x) -1 _ min{x,b} - Pw(b)
UF - A~ w(p) - uw(b)
-"
(cf. (3.41) to (3.43)), we easily determine the distribution function that has
a pth quantile at A~ w(p). In conclusion, we get the following reSult.
-"
Theorem 5 (F ~c W) Then for arbitrary 0 < p < 1 and b = W-1(p), we
have
F-l(p) - J-tF <
_ c5w (b) , (3.44)
UF
and the equality in (3.44) holds for

F(x) = { W (Pw(b) + uw(b)7) ' if a:~e < c5w(b), (3.45)

1, if a:~" ~ c5w(b).

Formula (3.45) defines the distribution function of an affinely transformed

random variable with distribution function W right truncated at its pth
quantile. This becomes a life distribution if

J-tF = uF[jtw(b) - aw]/uw(b).

44 3. Quantiles

Bounds (3.44) continuously increase from 0 at p = 0 to (dw - I'w)/o'W

at p = 1, which is finite if W has a finite support. For large p, these are
close to respective quantiles of parent W. However, they are positive for all
quantiles in contrast to negative values of standardized quantiles of small
orders for all F tc W (including W itself).
Proposition 3 (increasing density) For arbitrary 0 < p < 1, we have
F-1(P) _ I'F < (~)1/2
(3.46)
OF - 4 - 3p
Inequality (3.46) becomes an equality for a mixture of the uniform distribu-
tion on the interval [I' - 0.;3(2 - p)/[P(4 - 3p)]1/2, I' + 0[3P/(4 - 3p)j1/2]
and the degenerate measure concentrated at 1'+0[3P/(4-3P)j1/2 with prob-
abilities p and 1 - p, respectively.
Proposition 4 (increasing failure rate) For arbitrary 0 < p < 1, we
have
F-l (P) - I'F < -In(1 - p) - P (3.47)
OF - [P(2 - p) + 2(1 - p) In(1 - p)]1/2'
The equality in (3.47) is attained by the distribution of the random variable
min{X, -In(1 - p) - p}
Y - I' + O-::--:-~"---,---'--:--,-,:-;;....:;.,.=
- [P(2 - p) + 2(1 - p) In(1 - p)]1/2
for a standard exponentially distributed X.

3.3 Distributions with Monotone Density and

Failure Rate on the Average
First we consider quantile bounds for F t. W for which the projection
of h = Ml[b,c) onto the family of nondecreasing starshaped functions is
needed.
Lemma 8 For h defined in Lemma 4, and for every g E C( w there exists
ga/3 E C{. w defined as _.

ga/3(X) = a(x - a)l[b,d) (x) + (3 (3.48)

for some a, (3 ~ 0, a(b - a) + (3 < M, such that

Ilga/3 - hll ~ Ilg - hll·
3.3 Distributions with Monotone Density and Failure Rate on the Average 45

Candidates for projection Pt wh have a constant value (3 E [0, h(b)) in

[a, b), a jump at b to ga/3(b) <.. h(b), and a linear part in [b, d) that can be
extended to the left so as to pass through (a, (3) = (a, ga/3 (a)). A constant
projection (3 is also possible.

PROOF OF LEMMA 8. The first steps of reasoning are similar to those of

Lemma 4. If 9 E C{.w, so does g+. Moreover, the latter lies closer to h. If
g(c) ~ 0, then a zero function better approximates h, and can be further
improved by constant 1 = (h, 1). If
g(b) ~ h(b) = a::;x<d
max h(x),

then a downward translation of 9 gives a better approximation. Therefore

we can confine ourselves to functions 9 E C?w such that 0 ~ g(a) ~
g(b) < h(b) and g(c) > O. Let "( E (b, c] be tne smallest point at which 9
exceeds h. Put
g("() - g(a)
g')'(x) = g(a) + "( _ a (x - a)1[b,d)(X). (3.49)

This is the closest function to h in [a, b) among all nondecreasing starshaped

functions starting from level g(a) ~ O. Moreover, 9 E C{.w implies that
g(x) ~ g')'(x) ~ h(x), if x E [b,,,(),
g(x) ~ g')'(x) ~ h(x), if x E ["(,d).
This gives the desired conclusion, because (3.49) can be easily reparame-
terized as (3.48) . •

Lemma 9 Let assumptions of Lemma 4 and (3.7) hold. If

fbC(x - a)w(x) dx ld( ) ( )d (3.50)

rc ( ) ~ x - a w x x,
Jb W x dx b

_.
then Pt wh = 1. Otherwise

PSwh(x) = a*(x - a)1[b,d) (x) + (3* (3.51)

for a* > 0, 0 ~ (3. < 1 defined by
fbe(X-a)w(x) dx r d( ) ( )
J'eb W ()x dx - Jb X - a w x dx
(3.52)

(3. = (3.53)
46 3. Quantiles

PROOF. By Lemma 8, it suffices to minimize the function

D(a,f3) = Ilga.8 - hl1 2

= la
d[ l[b C) (X)
f3 + a(x - a)l[b,d)(X) - Ibc ~(x) dx
]2 w(x) dx
= f32 - 2f3 - It ;)
b W x dx
r
+ a 2 d(x - a)2w(x) dx
Jb

2a [IbC(jb;w~~~~ dx - f31d (x - a)w(x) dX] (3.54)

with respect to parameters a ~ 0 and 0 ~ f3 < 1/[fbc w(x) dxl. Fixing a,

we minimize (3.54) at

f3.(a) =1 - a ld (x - a)w{x) dx. (3.55)

Plugging it into (3.54), we obtain

D(a, P. (an ~ ,,' { 1.' (z - a)'w(x) dx - [1.' r}

(x - a)w(x) dx

- 2a [K(j:-;~~~; dx - ld(x - a)w(x) dx]

+ 1 -1 (3.56)
Ibc w(x) dx .
This is a quadratic function of a, with a positive coefficient of the quadratic
term. Under the nonnegativity condition, (3.56) is minimized at the max-
imum of 0 and (3.52). If (3.50) holds, then the optimal value is a = 0,
and, by (3.55), f3 = 1. This proves the first statement. Otherwise (3.56) is
minimized by (3.52), for which (3.55) coincides with (3.53) . •

Notice that the denominator of (3.52) is positive by the Schwarz inequal-

ity with Ibd w(x) dx < 1, and so is the numerator if (3.50) does not hold.
Therefore (3.51) is actually a nondecreasing starshaped function. Observe
finally that (3.53) is positive, because its numerator, divided by 1 - W(b),
can be interpreted as

Ew«X - a)21X > b) - Ew{X - alb < X < c)Ew{X - alX > b)
> Varw{XIX > b) > O.
Moreover, by (3.55), f3 < 1 < 1/[f: w{x) dxl, and so (3.53) satisfies both
constraints deduced from the geometric arguments presented in Lemma 8.
3.3 Distributions with Monotone Density and Failure Rate on the Average 47

By the arguments of Section 2.3,

Pg.w(h -1) = P{.wh-1.

This is 0 if (3.50) holds. Otherwise, we have

JbC(x-a)w(x) dx J,d( ) ( )
I.c w(x) dx - b X - a w x dx
[Link]-III b 1/2' (3.57)
{Jbd(X-a)2 w(x)dX - [Jbd(x-a)w(x)dxf}

p~.wIt(x)-l _ (x-a)I[b,d)(x)-Jbd(y-a)w(y)dy· ( )
IIP~. wit-III - { d d 2}1/2' 3.58
- Jb (y-a)2 w(y)dy - [Jb (y-a)w(y)dy]

Letting c '\t b, and using the notation of Theorem 4, we state its analogue
for F ~* W.
Theorem 6 (F ~* W) Using (3.20) and (3.21) with f3 = b = W- 1(P),
write
17=17w(a,b) = Ew[(X - a)I[b,d)(X))

= ld (x - a)w(x) dx = (1 - p)[Link](b), (3.59)

1J2 =1J~(a, b) = Varw[(X - a)I[b,d) (X))

1.' (x-a)'w(x)dx - [I.' (x-a)w(x)dx r (3.60)

If 17w(a, b) ~b-a, thenF-1(p) ~[Link].

Otherwise
F- 1(P)-[Link] AO ()_b-a- 17w (a,b)
(3.61)
O"F ~ t.w P - 1Jw(a, b) ,

and bound (3.61) is attained by

0, if x-IJ < _!l.

F(x) = { p, if
iT
_!l.< X-IJfJ' < b-a-f/ (3.62)
fJ- fJ '
X-IJ > b-a-f/
iT -
W (b - 7] + 1Jx~IJ), if iT - fJ •

Obviously, A~.w(P) ~ A~cw(P), because F ~c W implies F b W. It

follows that (3.61) tends to infinity as p /" 1. The main difference between
(3.27) and (3.62) is that the atom of the latter is distant from the smooth
part. Special cases of distributions dominating the uniform and exponential
ones in the star order are presented in Propositions 5 and 6.
48 3. Quantiles

Proposition 5 (decreasing density on the average ) If we have p ~

v'2 - 1 ~ 0.41421, then F-l(P) ~ J1.F.
Ifp> v'2 -1, then

F- 1 (P) - J1.F < V3(rr + 2p - 1)

(3.63)
O'F - 8u(P)

with
8~(P) = 12t?~(0,p) = 1 + 6p2 - 4p 3 - 3p4. (3.64)
Relation (3.63) becomes an equality for the mixture of an atom at J1. -
V30'(1- p2)/8u(P) with probability p, and the uniform distribution on [J1. +
V30'(p2 + 2p - 1)/8u(P), J1. + V30'(1 + p2)/8u(P)] with probability 1 - p.

Proposition 6 (decreasing failure rate on the average) Letusdeno-

te by Po ~ 0.553567 the unique zero of the strictly increasing function

vv(P) = p[1 -In(1 - p)]- 1, 0 < p < 1. (3.65)

If p ~ pO, then F-l(P) ~ J1.F.

Otherwise
F- 1 (P) - J1.F < vv(P)
(3.66)
O'F - 8v (P)
for

8~(P) = t?~(0, -In(1 - p» = (1 - p)p[1 -In(1 - p)]2 +1- p. (3.67)

Bound (3.66) is sharp. This is attained by F being a combination of the

jump distribution at J1. - 0'(1 - p)[1 -In(1 - p)]/8v(P) and the exponential
distribution with location J1. + O'Vv (P)/8 v (P) and scale 0'(1 - p)/8 v (P).

Theorem 7 asserts that general bound (3.2) cannot be improved in classes

of distributions defined by F ~* W. Moreover, we conclude analogous
results for all F preceding a fixed W in partial order more general than the
star one, for example the superadditive and Laplace transform orders (see,
e.g., Shaked and Shantikumar [96]).
Theorem 7 (F ~* W) If density function w(x) is bounded on a neighbor-
hood of aw, then for arbitrary 0 < p < 1 there exist sequences of absolutely
continuous distribution functions Fk ~* W with arbitrary common mean J1.
and variance 0'2 such that

lim
p,-1(p)_
k
(
J1.= ~
)1/2
k-too 0' 1- p

In particular, the statement holds for distributions with increasing density

and failure rate on the average.
3.3 Distributions with Monotone Density and Failure Rate on the Average 49

The claim can be justified by some informal arguments. To fix the ideas,
we consider the approximation of two-point distribution F defined in The-
orem 2 by Fk ~* U, that is, ones with nondecreasing Fk{x)/{x - ak) on
respective supports [ak, dk). Note that F is constant beside of its two jumps,
and the starshaped Fk have to increase at rate 1/{x - ak) at least on their
supports. However, letting ak -t -00, we relax conditions on the increase
rate of Fk itself. Accordingly, such Fk can be starshaped and approximate
well any two-point distribution at and beside jump points. The formal proof
of Theorem 7 is constructive. It is also possible to provide special construc-
tions of sequences satisfying the statement for specific W with unbounded
densities about respective left support ends.

PROOF OF THEOREM 7. Due to Theorem 2 the general bound (3.31) is

attained by the quantile functions

F-1(x) = { JLF - (jFJ¥- if 0 ~ x < p,

JLF + (jFJt!:p if p~x<l.

Under a change of variables, we have

1
F- W{x) -
(jF
JLF -gp
_ () _
x - {i!.l--
~
I-p'
p' if a ~ x < b = W-1(P),
if b ~ x < d.
(3.68)
Our proof consists in constructing a sequence gk E C~ w, k -t 00, that
converges to (3.68) in L2([aw, dw), w(x)dx). For sufficiently large k, we
define piecewise linear continuous nondecreasing starshaped functions

k4 (x - a) - k - J¥-, if a ~ x ~ a + k- 3 ,

-J!=1!. p , if a + k- 3 ~ X ~ b,

k~
b-a - k- J!=1!.
p ,
if b< x
-
<
-
b + b-a
kv'p(l-p) ,

Jt!:p, if b+ b-a < X < d.

kv'p(l-p) -

Observe that

= 1 a+k- 3
[k 4 (x - a) - k]2W(X) dx

+ rb+ Io",:71""-P) [k x - a _ k _ 1 ]2 w(x) dx

lb b- a Vp(1 - p)

W (b+ b-a ) - W(b)

w(x) kv'p(l-p)
< sup -3k + p(l-p) -t 0
a$x$a+k- 3
50 3. Quantiles

as k --+ 00. The same holds for

because

and

Relations
[Fk"lW(x) - /l-F]/UF = 91e(X)
define a sequence of Fie -<* W with common mean /l- and variance u 2 for
which Fk"l(p) /" Jp/(l- p) .•

3.4 Symmetric Unimodal Distributions

Let W be an absolutely continuous distribution function of a symmetric
random variable. We are interested in evaluating [F-l(P) - /l-F]/UF for all
symmetric distributions satisfying F tB W. A dual problem is to project

_ l!W-l(l_q~.W-l(l_P»(Z) if < 1/2

hB(x) _ { 2 q-p) P, (3.69)
- +l!W-l(~'W-l(q»(Z)
2 q p)
'f
1 P>
_ 1/2 ,

for q '\t ponto e:,.c w as defined in (2.60). If p < 1/2, then hB $ O. Its
projection onto the family of nonnegative functions e+ ::> e:,.cw is P+h B =
o which actually belongs to e:,. w, and so P:" whB = O. This sequence of
elaborated arguments leads us-to the trivial c~nclusion

Otherwise we have
Lemma 10 Let

hB = Ml[b,c) E L 2 ([[Link],dw),2w(x)dx)
for /l-w < b < c $ dw and M > O. Then for every 9 E e:,.c w there exists
9a{3 E etc w defined as -

9a{3(X) = a(x - (3)+

for some a ~ 0, /l-w $ (3 $ b, a(b - (3) < M such that

 3.4 Symmetric Unimodal Distributions 51

The only difference between the statements of Lemmas 4 and 10 is that

there is a lower constraint on parameter 13 in the latter. IT b is close to
[Link], the constraint plays a significant role leading to conclusions essentially
different from those of Lemma 5.
Lemma 11 If for f.L = [Link], we have
IbC(x - f.L)w(x) dx I:(x - f.L)2w(X) dx
rc < d ' (3.70)
Jb w(x) dx - I" (x - f.L)w(x) dx
then
P:- w h8 (x) = IbC(y];~ f.L)w(y) dy d X - f.L . (3.71)
_c 2 b w(y) dy I" (y - f.L)2 w(y) dy
Otherwise there exists a unique f.L < 13. < b that solves
IbC(x - f3)w(x) dx I;(x - f3)2w(X) dx
C = d (3.72)
Ib w(x) dx I{3 (x - f3)w(x) dx
(ef, (3.9)) such that

P:- w h8 (x) = d (x - 13.)+ (3.73)

_c 2I{3.(y - f3.)w(y) dy
Under (3.70), we have

(3.74)

= (3.75)

Otherwise
I;. (x - f3.)2 w(x) dx
(3.76)
d ] 2'
2 [ I{3.(x - f3.)w(x) dx

(x - 13.)+
= 1/2 . (3.77)
[2 I;. (y - f3.)2 w(y) dy]
Letting e ":>I b, recalling (3.20) through (3.22), and introducing

~w = EwlX - [Link] =2 i d
(x - f.L)w(x) dx (3.78)

(the latter under the symmetry assumption), we obtain conclusions similar

to ones obtained in Theorem 4.
52 3. Quantiles

Theorem 8 (F ts W) If p $ 1/2 (i.e., W-l(P) $ I'w), then we have

F-l(P) $ I'F.
If I'w < W- 1 (P) $ I'w + u'f..,/t;w, then (3.23) holds, with the equality
attained by (3.24).
Finally, if W- 1 (P) > I'w + u'f..,/t;w, then there exists a unique {3* =
{3.(P) E (I'w, W- 1 (P» solving Equation (3.25) such that

F- 1 (P) -I'F < A

UF -
= AS
t.w
(P) ={ 8'f..,{{3*) + 1
2[1- W({3.)]
}1/2 (3.79)

The equality in (3.79) holds for

F(x) ={ 1- W ({3* - 2[p,w{{3.) - {3:]A 7 ), if 7 < 0, (3.80)

W ({3* + 2[l'w({3*) - {3*]A 7 ), if 7 ~ O.

Symmetric distribution function (3.80) has a jump of height 2W{{3.) -1 <

2p-l at its center of symmetry. If W- 1 (P) = I'w+u'f..,/t;w, then I'w solves
(3.25), bounds (3.23) and (3.79) are equal, and (3.80) coincides with (3.24).
Putting W = U, we get
Proposition 7 (symmetric unimodal distributions) Ifp $ 1/2, then
F-l(P) $I'F.
If 1/2 $ p $ 5/6, then

F-l~~ -I'F $ 2V3 (p _ ~) , (3.81)

which becomes the equality for the uniform distribution on interval [I' -
vau, I' + vau].
If 5/6 $ p < 1, then

F- 1 (P) -I'F < ! (_2_)1/2 (3.82)

UF - 3 1- p
The bound becomes the equality for a mixture of the atom at I' with proba-
bility 6p - 5, and uniform distribution on interval [I' - uh/2(1- p), I' +
u/J2{1- p)] with probability 6(1- p).
Observe that (3.82) is 1.5 times less than respective bound (3.5) for sym-
metric distributions.

In Table 3.1, we numerically compare bounds on standardized quantiles

[F-l(p) - I'Fl/uF of orders p = 0.05(0.05)0.95 for nine families of distri-
butions: general (G), symmetric (S), and symmetric unimodal (SUN) ones,
distributions with decreasing density {~O) and failure rate (OFR) , and
those on the average (OOA and OFRA, respectively), and with increasing
density (10) and failure rate (IFR). Bounds for distributions with increas-
ing density and failure rate on the average are identical with the general
 TABLE 3.1. Sharp uniform mean-variance bounds on quantiles for various families of distributions.
p G S SUN DD DFR DDA DFRA ID IFR
0.05 0.22942 0 0 0 0 0 0 0.19739 0.19782
0.10 0.33333 0 0 0 0 0 0 0.28475 0.28601
0.15 0.42008 0 0 0 0 0 0 0.35603 0.35874
0.20 0.5 0 0 0 0 0 0 0.42008 0.42465
0.25 0.57735 0 0 0 0 0 0 0.48038 0.48741
0.30 0.65465 0 0 0 0 0 0 0.53882 0.54905
0.35 0.73380 0 0 0 0 0 0 0.59660 0.61097 tIo)

0.40 0.81650 0 0 0 0 0 0 0.65465 0.67435 ~

0.45 0.90453 0 0 0 0 0.13511 0 0.71375 0.74030

0.50 1 1 0 0 0 0.32163 0 0.77460 0.81005
0.55 1.10554 1.05409 0.17321 0.17321 0 0.50913 0 0.83793 0.88503
0.60 1.22474 1.11803 0.34641 0.34641 n
0 0.70235 0.15541 0.90453 0.96708
0.65 1.36277 1.19523 0.51962 0.51962 0.04982 0.90763 0.33725 0.97531 1.05869
i
~
0.70 1.52753 1.29099 0.69282 0.69389 0.20397 1.13406 0.53901 1.05131 1.16351 ~
0.75 1.73205 1.41421 0.86603 0.88192 0.38629 1.39582 0.77040 1.13389 1.28721 ~
0.80 2 1.58114 1.03923 1.10554 0.60944 1.71781 1.05060 1.22474 1.43941 t;j
0;'
0.85 2.38048 1.82574 1.21716 1.40106 0.89712 2.15063 1.42081 1.32613 1.63860 S".
0.90 3 2.23607 1.49071 1.85592 1.30641 2.82311 1.98883 1.44115 1.92754 [
0.95 4.35890 3.16228 2.10819 2.80872 2.10081 4.24076
~ -
3.18532 1.57425 2.44875 ~.

C11
tIo)
54 3. Quantiles

ones. Certainly, the general bounds are merely valid for wider classes of
life distributions (e.g., NBU, HNBUE and C-class defined in Klefsjo [46])
which are generated by more general orders than the star order. On the
other hand, there are still interesting open problems of accurate bounds on
quantiles of distributions succeeding a given one in these general orders.

3.5 Open Problems

1. What are the bounds on standardized quantiles in families of distribu-
tions defined by relation F ~B W for a fixed symmetric distribution
function W? Of special interest are the symmetric U-shaped distri-
butions, alternatively defined by F ~B U.
2. Mean-variance bounds on quantiles of distributions determined by
F ~* W coincide with general bounds under very mild conditions
on fixed W. The reason is that without disturbing values of mean,
variance, and quantile of a given order we are able to choose elements
of the class with an arbitrarily remote left endpoint of support which
approximate two-point distributions attaining the general bounds.
This is impossible if we fix the left endpoint (aF = 0, say) and
the second raw moment m~. Then we have the problem of evalu-
ating F-l(p)/mF for all life distributions determined by F ~* W
with a nontrivial solution. Also, still unknown are the bounds on
F-l(p)/mF for F coming from the classes of
(a) general distributions,
(b) distributions succeeding (preceding) fixed W in the convex or-
der, and
(c) distributions succeeding fixed W in the star order.
3. Quantile differences F-l(q) - F-l(P), 0 < p < q < 1, are measures
of population dispersion less sensitive to errors in statistical sam-
pling than the standard deviation UFo Formula (3.3) evaluates the
ratio [F-l(q) - F-1(P)]/UF for arbitrary F. What are the respective
bounds if F belongs to the restricted families of
(a) symmetric distributions,
(b) distributions determined by relations in the convex order,
(c) distributions determined by relations in the star order, and
(d) distributions determined by relations in the s-order?
Of special interest, especially in symmetric populations, are quantile
differences for symmetric pairs p and q = 1 - p.
4
Order Statistics of Independent
Samples

Bounds for expectations of order statistics from general i.i.d. samples, due
to Moriguti [58], are presented in Section 4.1. Bounds of Sections 4.2 and 4.4
for populations with decreasing density and failure rate, and symmetric uni-
modal ones, were obtained by Gajek and Rychlik [33]. Results of Section 4.3
for restricted families determined by the star order come from Rychlik [89].
Section 4.5, partially based on Okolewski and Rychlik [65], is devoted to
the study of quantile estimation bias in various nonparametric families of
distributions.
We do not discuss here bounds for the sample maximum and range from
discrete populations obtained by LOpez-Blazquez [50, 51]. We also merely
mention sharp inequalities due to Papadatos [69] for the expectations of
order statistics and their differences of nonnegative samples expressed in
terms of the population mean. Blom [17] and van Zwet [100] developed
another method of evaluating the expectations of order statistics from re-
stricted families of distributions defined by means of stochastic orders using
quantiles of the parent distribution. The method is based on the Jensen in-
equality. Papadatos [67, 68] derived some attainable bounds on the variance
of order statistics from general and symmetric distributions, respectively,
measured in the standard deviation units of the parent distribution. Anal-
ogous evaluations for covariances of order statistics are known, but only
small samples of size n = 2 and 3 were treated (see, e.g., Ma [52] and
Papadatos [70]). Bounds and approximations for moments of order statis-
tics were reviewed in David [22, Chapter 4], Arnold and Balakrishnan [5,
Chapters 4,5], and Rychlik [84].
56 4. Order Statistics of Independent Samples

4.1 General and Symmetric Distributions

The problem of calculating mean-variance bounds on expectations of order
statistics lies in determining projections of density functions (2.22) onto
the convex cone (2.46). Observe that

fn:n(x) = nxn- 1, 0 < x < 1, (4.1)

is actually nondecreasing, and so p.l' fn:n = fn:n, and P°(fn:n - 1) =

fn:n - 1. Therefore

EFXn:n - JLF < Ilf.

UF - n.n
_ 111 = n- 1
(2n _ 1)1/2
,. ., (n /2)1/2 (4.2)

with
1 1/(n-1}
n- 1 x-JL )]
F(x) = [
~ 1+
(
(2n _ 1)1/2 U
_ (2n - 1)1/2 < X - JL < (2n _ 1)1/2 (4.3)
n-1 - U - ,

attaining the bound. Distribution (4.3) is a location-scale transformation

of a power distribution, and this is uniform for n = 2. The result was
published independently by Gumbel [36] and Hartley and David [38]. More
generally, the respective bound for L-statistics

n n
EF LCj(Xj:n - JLF)/UF ~ LCj(fj:n -1) (4.4)
j=l j=l

is sharp if the argument of the norm is nondecreasing. By differentiation,

n n-1
LCjfj:n(x) = n L(Cj+1 - Cj)/;:n-1(X), (4.5)
j=l j=l
we conclude that nondecrease of the sequence of coefficients is a sufficient
condition. This was implicitly exploited by Plackett [72] and Nagaraja [60]
in calculating optimal bounds for the sample range Xn:n - X1:n and selec-
tion differentials l/k Ej=n+1-k X j:n, respectively.
In the case of general bounds, the Moriguti [58] projection method should
be used (cf. Rychlik [84, Theorem 7, p. 121]). Here we confine ourselves
to single order statistics. Observe first that EFX1:n ~ JLF is implied by
relation X 1:n ~ Xl. Calculating bounds for nonextreme order statistics
X j :n , 2 ~ j ~ n - 1, we refer to Moriguti [58, Example 2, p. 111]. We
first aim at determining p.l' /;:n. Note that /;:n(x) > 0 for 0 < x < 1, and
increasing-decreasing with the maximum at (j -l)/(n -1). Therefore Fj:n
 4.1 General and Symmetric Distributions 57

is strictly increasing on [0,1], convex on [0, (j - 1)/(n - 1)], and concave

on [(j - 1)/(n - 1),1]. The slopes of tangent lines
la(x) = /i:n(a)(x - a) + Fj:n(a)

°
continuously increase for a E [0, (j - 1)/(n - 1)] and so do , 0 (1), ranging
from to 'U-1)/(n-1)(I) > Fj:n (l) = 1. The line 10 • for some 0.* E (0, (j-
1)/(n-l» such that 10 • (1) = 1 becomes a part of the lower convex envelope
of Fj:n on [0.*,1], and the remaining part coincides with Fj:n . Therefore,
the greatest convex minorant Fj:n of Fj:n has form

p. ( ) _ { Fj:n(x),
3: nX -
if °
~ x ~ 0.*,
/i:n(a*)(x - 1) + 1, if 0.* ~ x < 1,
(4.6)

for a unique 0.* = a*(j, n) E (0, (j - 1)/(n - 1)) satisfying

(1 - a*)/i:n(a*) =1- Fj:n(a*). (4.7)
The derivative of (4.6) is
p/' /i:n(x) = hn(x) = /i:n(min{x,a*}). (4.8)
Using
(i~j-2) (m+n-~-j)
Am(x)/i:n(x) =n '-\m+,;:-~)' fi+j-1:m+n-1(X), (4.9)

we calculate

IIhnl1 2 = lao f;'n(x) dx + (1- a*)f;'n(a*)

(~-ne:=n
n enn- i ) F2j-1:2n-1 (0.*)
+ (1 - a*)f;'n(a*). ([Link])
We have thus proven
Theorem 9 (general distributions) For arbitrary F and j = 1, we
have EFX1:n ~ [Link], which is attained by the atom measure at [Link].
For 2 ~ j ~ n - 1, inequality
[EFXj : n - [Link]]/UF ~ B = B°(j,n) = (1lhnW _1)1/2 (4.11)
is sharp (see (4.7) and (4.10»), and this is attained by the distributionfunc-
tion

F(x) = {° ,
f:-1 (1 +
):n
BZ-I:')
".'
'J71
il
~
~". -<_.!.
".
B'
_.!.B -< Z-I:' < f;,n(a.)-1
B' (4.12)
1, 71 Z-I:' > Ji:n(a.)-1
'J ". - B .

For j = n, (4.2) is the best bound, attainable by (4.3).

58 4. Order Statistics of Independent Samples

Distribution function (4.12) has an absolutely continuous component (in-

verse of a polynomial of degree n - 1), and a jump of height 1 - (}* at
the right end. With few exceptions, the distribution does not have an ex-
plicit formula. Balakrishnan [10] used a relationship between binomial and
negative binomial distributions for reducing polynomial equation (4. 7) of
degree n to one of degree j - 1. This makes possible writing explicit values
of solutions (}* and bounds B°(j, n) for j = 2,3, n - 2, n - 1. For j = 2 in
particular (}* = (n _1)-2, and (4.10) amounts to

_ 2 n 2{n - 1) [ n 2n - 3 {n - 2)2n-1]
Ilh:nll = (2n _ 3){2n _ 1) 1 - (n _ 1)4(n-1) . (4.13)

In Chapter 7 we need more rough approximations of EFXj:n in mF units.

=
Observe that functions p/' /i:n !j:n for 2 :::; j :::; n -1 and p/' fn:n fn:n =
vanish at 0, and so coincide with the projections onto C+ C C/'. Therefore,
we have

EFXj:n/mF < Ilhnll, 2:::; j:::; n -1, (4.14)

EFXn:n/mF < Ilfn:nll = n/(2n - 1)1/2, (4.15)

(cf. (4.10) and (4.2), respectively). The bounds are attained by

0, if ~:::; 0,
F{x) = { fj-;~ (lIhnll~), if 0:::; ~ < f(I'ii~~il)' (4.16)
1 if .L > hn(a.)
, m - IIhnll'
for 2 :::; j :::; n - 1, and

0, if ~:::; 0,
F{x) = { [(2n _ 1)-1/2~]1/(n-1) , if 0:::; ~ < {2n - 1)1/2, (4.17)
1, if ~ ~ (2n - 1)1/2

for the sample maximum. Both (4.16) and (4.17) are life distributions.
Density function hn is decreasing and so projection p/' hn = 1 f/. C+,
but the constant can be approximated in L2{[0, 1), dx) by functions

This implies that the trivial bound for the sample minimum

(4.18)

is approximated by the sequence of two-point life distributions

 4.1 General and Symmetric Distributions 59

The problem of determining bounds analogous to (4.11) for symmetric

parent distributions can be solved by use of the trick of folding the func-
tional about 1/2. Case j = n is the only one for which
Sj:n(x) = fj:n(x) - /;:n(1 - x) = /;:n(x) - fn+l-j:n(x) (4.20)
is nondecreasing. Explicit bound

EFXn:n - J.I. = n r
11/2
[F-1(X) - J.I.][x n- 1 - (1- X)'~-1] dx

[I.;, dxt'
r
< n [ZR-1 - (1- z)R- 11' u/v'2

= {2(;:~ 1) [1- c:~n _1] u

(4.21)
(cf. (4.2» was obtained by Moriguti [57] through use ofthe Schwarz inequal-
ity. The extreme distributions for which the equality holds have quantile
functions equal to x n- 1 - (1- x)n-1 up to affine transformations.
Bounds for general order statistics of symmetric populations are derived
by projecting differences Sj:n onto the family of nondecreasing functions
in L2([1/2, 1), 2dx). For this purpose we first analyze variability of the
differences, using the auxiliary results of Gajek and Rychlik [33, Lemma 3,
p.167].
Lemma 12 Consider the function defined in (4.20) for x E [1/2,1) and
(n + 1)/2 < j < n. This is nonnegative, and equal to 0 at 1/2 and 1, and
increasing-decreasing. For j ::; n - 2 function (4.20) is concave-convex if
j ::; [n + (3n - 5)1/2]/2, and convex-concave-convex othennise. Also, Sn-1:n
is concave for n ::; 7, and convex-concave for n ~ 8.
Now we merely apply the first statement of the lemma. Further properties
are needed in Section 4.4. Observe first that, by symmetry, Sj:n ::; 0 for
j ::; (n + 1)/2, and so

EFXj:n - J.I. = f1 [F-1(x) - J.I.]Sj:n(x)dx::; 0, (4.22)

11/2
since F-1 - J.l.F ~ 0 on [1/2,1). IT j = (n + 1)/2, then Sj:n = 0, and (4.22)
becomes the equality for any symmetric F. For (n + 1)/2 < j ::; n - 1, we
repeat arguments leading to (4.8), and obtain
p. /;:n(x) = Sj:n(min{X, a.}), 1/2::; x < 1, (4.23)
for a. E [1/2,1) defined by equation
(1 - a)[/;:n(a) - fn+l-j:n(a)] = Fn+l-j:n(a) - Fj:n(a) (4.24)
60 4. Order Statistics of Independent Samples

(cf. (4.7)). Using (4.9), we calculate

'-(2nnif'
(2!-2W n - 2i )
liPs /i:nW = n [F2j-1:2n-1 (a*) + F2n-2jH:2n-1 (a*)

F2j-1:2n-1 (~) - F2n-2jH:2n-1 (~)]

(n-1) 2
2n (~:n~l) [Fn:2n - 1(a*) - ~]
+ (1- a*)s~:n(a*). (4.25)

Theorem 10 (symmetric distributions) For 1 ~ j ~ (n + 1)/2, we

have EFX j:n ~ Jl.F for all symmetric parent distributions of the sample.
This becomes the equality if either F is the Dirac measure at Jl.F or j =
(n + 1)/2.
For (n + 1) /2 < j ~ n - 1, we have

(4.26)

defined in (4.24) and (4.25), with the equality for

0, zf x-y < _ Si:n(O.)

2B'
if _ Sj:n (0.) < x-y <
17
F(x) = { S:-l (2B
3:n
x -y)
17' 2B -
Sj:n (0.)
2B'
(4.27)
if x-y > Sj:n (0.)
17

1, 17 - 2B '

(see (4.23) and (4.25)).

For j = n we have (4.21) which becomes the equality for (4.27) with
a* = 1 and sn:n(l) = n.
Distribution (4.27) has a smooth component with two atoms of measure
1 - a* (= 0 for j = n) at the endpoints of support.

4.2 Life Distributions with Decreasing Density and

Failure Rate
Unlike the problems studied above, we present here bounds in terms of the
square root of the second raw moment. In fact, except for the scale unit
we also have a location parameter in the model. This is the population
minimal value which for the life distributions amounts to o. The problem
is to evaluate the expected lifetime of (n + 1 - j)-out-of-n systems of inde-
pendent elements whose identical distribution functions satisfy F ~c W for
some fixed W. The solution is based on determining the projection P:j: wh _0
 4.2 Life Distributions with Decreasing Density and Failure Rate 61

of h = h:nW onto the convex cone

c~
_c
w = {g E L 2 ([aw,dw),w(x)dx) : 9 - nondecreasing, convex,
g(aw) = O}. (4.28)

In the sequel we frequently abstract from the specific form of h and assume
the following

h(x), w(x) >0 for a < x < d, h(a) = 0, (4.29)

ld h(x)w(x) dx = ld w(x) dx = 1, (4.30)

ld x 2 w(x) dx < 00, (4.31)

h(x) - bounded, h" exists, (4.32)

'Va < x <b h'(x),h"(x) > 0,

'V b < x <c h"(x) < 0 < h'(x), (4.33)
'V c < x <d h'(x) < o.
However, it is worth pointing out that for the representation h = hn W,
(4.29) through (4.33) are actually conditions on W, and some of them are
naturally satisfied. For j =f 1, we have (4.29) and (4.30), and boundedness
and monotonicity assumptions (with c d for j = =
n). An equivalent formu-
lation of (4.31) is finiteness of m~, and existence of h" requires differentia-
bility of density w on (aw,dw). Crucial assumptions are ones describing
regions of convexity and concavity of h. These were chosen to cover the
important cases of uniform and exponential weights, without pretending to
develop a general theory. One can easily check that the assumptions are
=
satisfied for 2 ~ j ~ n - 1 and both W U, V (with a b in case j = 2), =
and for j = n and W = =
V with c d. We point out that for j =
1 the
trivial bound
EFX1:n ~ [Link] ~ mF

is valid for arbitrary F, being attainable for degenerate distributions. In

the case j = =
nand W U, we have

P:cufn:n
_ = fn:n E C~ u,_c

and so the bound for F tc U is identical with that for general F.

We now present the solution to the projection problem. Lemma 13 de-
scribes geometrical properties of the projection. This allows us to restrict
ourselves to continuous functions that are identical with h on the left and
have linear extensions on the right.
62 4. Order Statistics of Independent Samples

Lemma 13 For fixed h satisfying (4.29) through (4.33), set

h(x), if a ~ x ~ p,
haP(x) ={ h(P) + a(x - P), if P ~ x < d. (4.34)

For every 9 E ct w there exists haP E ct w such that

-< -<

IlhaP - hll ~ IIg --: hll·

PROOF. Take an arbitrary nonzero 9 E ct w' Since g(a) = h(a) = 0, it
makes sense to define -<

'Y = sup{x E [a, b] : g(x) = h(x)}.

We now analyze mutual relations between 9 and h in [b, d), where h is first
concave increasing and then possibly decreasing.
Suppose first that 9 > h in a right neighborhood of b. Then these cases
are possible:
(i) g> h on the whole (b,d),
(ii) 9 and h are merely tangent in a single possibly degenerate subinterval
of (b, c), or
(iii) 9 crosses h at a point 8 E (b, c), say, and either 9 < h holds on (8, d)
or 9 and h cross each other at a single point A E (8, d).
In each case we construct modifications gi, i = 1,2,3, of the original 9 that
have form (4.34) and are less distant from h than g.
(i) We have g(x) > h(x) for all x E (-y, d). Define

_ { h(b) + h'(b)(x - b), if a ~ x ~ b,

h(x) = h(x), if b ~ x ~ c,
h(c), if c~ x < d.

We easily see that it is nondecreasing, concave, and it < g. Therefore the

regions above the graph of 9 and beneath the graph of it are disjoint planar
sets, and there exists a straight line h(x) separating the sets from each
other. This line must have a common point, say P E ['Y, b], with h. Then
g1 = max{h,h} has the desired form. Moreover, IIg1 - hll ~ IIg - hll,
because h(x) = g1(X) for x E [a,p] and h(x) ~ g1(X) ~ g(x) for x E [P,d).
(ii) Functions h and 9 ~ h are concave and convex, respectively, in (b, c),
and have a tangent point there. Accordingly, there exists a line 12 separating
the curves in the interval. Note that 12 crosses h at some P E h, b), because
it runs beneath 9 and over the concave part of h, and h('Y) = g(-y). For
g2 = max{h,12} we easily verify the conclusions of the previous case.
 4.2 Life Distributions with Decreasing Density and Failure Rate 63

(iii) Extend the definition of 6. as follows

6. = sup{x ~ 0: g(x) :$ h(x)}.
In the former subcase 6. = d which is possible only if d < 00. Note that
g(6.) = lim g(x),
x/11
h(6.) = lim h(x),
x/11
are well defined and finite. Define

1011(x) = g(o) + g(6.1 =:(0) (x - 0). (4.35)

Then we have
g(x) :$ 1000(x) :$ h(x), if 0:$ x :$ 6.,
h(x) :$ 1011(x) :$ g(x), if 6.:$ x < d,
and the latter holds in a left neighborhood of 0, as well. By convexity of g,
we have
1011(x) :$ g(x), a:$ x :$ o.
This relation, combined with g( 'Y) = h('Y), implies the existence of (3 E b, oj
such that 1011 ((3) = h((3) and 1611 ~ h' ((3). Therefore

( x) - {h(X), if a:$ x :$ (3,

g3 - 1011(x), if (3:$ x < d,
is the desired convex nondecreasing modification of g.

If 9 < h on some right neighborhood of b (and so of 'Y), then there is

a single point 6. > b at most at which 9 and h cross each other. If there is
no such 6., we set 6. = d, being finite (cf. Case (iii)). For each y E [a, b),
define the linear function tangent to h at y as
Iv(x) = h(y) + h'(y)(x - y). (4.36)
Notice that the slopes of (4.36) increase, and Yl < Y2 implies IVl < IV2
on [Y2, d). Function y t-+ Iv(o), y E [a, bJ, is strictly increasing and continu-
ous, and satisfies
Ib(6.) ~ h(6.) ~ g(6.).
Moreover, relation 1')'(6.) > h(6.) implies

l~ = h'(-y) =
> h(6.1 ~(-y) ~ g(6.1 =~('Y) ~ g'('Y+). (4.37)

If 'Y > a, then, by (4.37), we have h'('Y) < g'(-y-), and h < 9 on a left
neighborhood of 'Y. Summing up, under condition g(-y+) < h(-y+), three
cases are possible:
64 4. Order Statistics of Independent Samples

(iv) 1'Y(.~) :$ h(~),

(v) 1'Y(~) > h(~) with "( = a, and
(vi) ''Y(~) > h(~) with g("(-) > h(,,(-),
which are considered consecutively in the remainder of the proof.
(iv) Relations 1'Y(~) :$ h(~) :$ lb(~) imply existence of a a E [,,(, b] for
which 16(~) = h(~) holds. Then

g(x) :$ 16(x) :$ h(x), x E [a, ~].

If (~, d) "# 0, then the reversed inequalities hold there. Therefore the convex
nondecreasing function

g4(X) = max{h(x),16(x)} = h(x)1[a,6) (x) + 16(X)1(6,d) (x)

is a better approximation of h than g.
(v) The linear function g5(X) = l'Ya(x) secant to 9 at "( = a and ~
(cf. (4.35)) runs between 9 and h in the whole domain [a,d).
(vi) In this case there exists a E [a,,,() such that g(a) = h(a), and g(x) >
h(x) for x E (a,,,(). Apply (4.35) again for defining the straight line [Link]. We
have
g(x) :$16a(x) :$ h(x), x E ["(,d],
and
g(x) :2: 16a(x) :2: h(x), xE [~,d),

under condition ~ < d (cf. Case (iv)). For x < ,,(, we proceed as in
Case (iii). We have

h("(-) < 16a("(-) :$ g("(-),

16a(a) :$ g(a) = h(a),
which imply that 16a(ao) = h(ao) for a unique ao E [a,,,(), and
h(x) < 16a(x) :$ g(x), x E [ao,"().

It follows that h'(ao) < l~a' and

g6(X) = h(x)1[a,6o) (x) + 16a(x)1(60,d)(X)
is an element of ct w that lies closer to h than the original g. This ends
the proof. • _c

Observe that hQfj E ctc

w requires a :$ f3 :$ b, and either a :2: h'(f3) > 0
rr
for f3 > a or a :2: 0 for = a. If a = b in particular, then the projection is
a nondecreasing linear function

P:
_c
wh(x) = a.(x - a)
 4.2 Life Distributions with Decreasing Density and Failure Rate 65

whose slope 0:. can be easily determined. Note that for fixed {3 E [a, b],

D(o:,{3) = Ilha,8 - hW = [Link][o:(X - {3) + h({3) - h(X)]2w(X) dx (4.38)

is a convex quadratic function of 0:, and under restriction ha,8 E ct w is

minimized at _c

&(a) = max{O, 0:. (a)}, (4.39)

&({3) = max{h'({3),o:.({3)}, {3 > a, (4.40)
with
J;(x - {3)[h(x) - h(,8)]w(x) dx
0:.({3) = d ,(4.41)
J,8 (x - {3)2w(x) dx
being the optimal slope without restrictions. Consecutive reasoning steps
consist in eliminating {3 > a for which h'({3) > 0:. ({3), and analyzing
D(o:. ({3), {3) with continuous derivative

dD (0:. ({3) , {3)

d{3
= 2[0:.({3) - h'({3)] [Link][h(X) - ha.(,8),8(x)]w(x) dx

= 2K({3)L(,8) , (4.42)
say. We minimize D (0:. ({3),,8) by determining the set
/C = {,8 E (a, b) : K({3) ~ O},
and analyzing sign changes of L in /C. If /C = (a,l\;] for some I\; E (a, b]
(which actually holds in special cases W = U, V), then L({3) is positive,
negative, or negative-positive, with zero at some .oX E /C, and so optimal
{3. = a, 1\;, and .oX in the respective cases. Specific forms of

h(x) = !;:n(x) = nBj-l,n-l(X),

where
Bk,m(X) = (;)xk(1- x)m-k, 0 ~ k ~ m < 00,
are the Bernstein polynomials of degree m, enable us to determine final
forms of P:cu!;:n. Differentiating and integrating the Bernstein polyno-
mials we ontain linear combinations of these polynomials of smaller and
greater degree, respectively. These operations were applied in defining fac-
tors of (4.42) which coincide with combinations of Bernstein polynomials
of degree n + 1 up to positive functional multiplicators. The coefficients of
the combinations depend on j and n. The basic tool in analyzing the sign
changes of the combinations is the following well-known variation dimin-
ishing property of the Bernstein polynomials proved in Schoenberg [93, p.
252]).
66 4. Order Statistics of Independent Samples

Lemma 14 The number of zeros of a given nonzero combination of Bern-

stein polynomials
m
B(x) =L akBk,m(X), x E (0,1), (4.43)
k=O

does not exceed the number of sign changes of the sequence ao, ... , am. The
first and last signs of (4.43) are identical with the signs of the first and last
nonzero elements of ao, ... , am, respectively.
This is equivalent to saying that Bk,m, °
$ k $ m, for fixed m form
a Chebyshev system on [0,1] (see Karlin and Studden [45, Theorem 1.4.1]).
In addition, differentiation and integration operators defined in (4.41) and
(4.42), acting on compositions

h(x) = /i:n V(x) = nBj-l,n-l V(x),

do not lead beyond the polynomials of degree n in V(x) up to positive
factors and can be further rewritten as combinations of Bk,n V(x). Therefore
Lemma 14 is useful in analyzing variability of (4.42) for h(x) = /i:n V(x)
as well.
Theorem 11 (decreasing density) If 2 $ i $ min{2(n + 1)/3,n - I},
then
EFXj:n < v'3_
i_ (4.44)
mF - n+l'
which is the equality if F is the uniform distribution on [0, v'3mF].
If 2(n + 1)/3 < i $ n - 1, then

EFXj:n $ B = B~cu(j,n) = U(/i:n)a•.[Link], (4.45)

mF -

(cf. (4.34)) for

(2;-2) 2n-2;
= n '-1 n_n-' F 2j-l:2n-l(.8.) + (1- .8.)fI:n(f3.)
n

+ (1 - .8.)2a./i:n(.8.) + (1 - .8.)3a~/3, (4.46)

and .8. = .8.(j,n) being the smaller of the smallest positive zeros of poly-
nomials
 4.2 Life Distributions with Decreasing Density and Failure Rate 67

j (n - j + 3)'
Ku(x) = 3 L(j + 1 - k)!k:n+2(X) - (n _ j)! . !;-l:n+2(X)
k=1
. 3 (n - j + 2)!
+ (n - J -"2) (n _ j)! hn+2(X), (4.48)

j-l 3j - k
Lu(x) = L(n+ 1- -2-)!k:n+2(X)
k=1
(n - j)(n - j -I)! () (4.49)
- 4 j:n+2 X .

Bound (4.45) is attained by

0, if m -<0 ,
£

!j-;~ (B!n) , if 0< £ < f;,,,({3.)

F(x) ={ R + Bii-f;,,,({3.) - m - B '
zf /;,,,({3.) < £ < /;',,({3.)+a.(I-{3.)
fJ* a. ' B -m- B '
1, if £ > /;',,({3.)+a.(I-{3.)
m - B .
(4.50)
Finally, for j =n we have (4.15) which becomes the equality for (4.17).
Theorem 12 (decreasing failure rate) Set
j 1
[Link];,,, = EVXj:n = L n + 1- k' 1~j ~ n. (4.51)
k=1
If [Link];,,, ~ 2, then
EFXi:n < [Link];,,, (4.52)
mF -.../2'
which becomes the equality for the exponential distribution with the scale
parameter mF /.../2.
Otherwise

(4.53)

for
(21- 2)(2"-2;)
= n '-(2"" if' F2j-l:2n-l(')'.. )
+ (1-1'.. )[2a~+2a .. hnh'.. )+ !J,n(')''')] , (4.54)

(4.55)
68 4. Order Statistics of Independent Samples

and 'Y. being the minimum of the smallest positive zeros of

Kv(x) = LI'Vi+ 1 -.,n+l-lofk:n+1(X)

k=l
(n - j +2)!
2 (n _ j)! !i-l:n+1(X)

+ 2(n - j - ~)(n - j + l)/j,n+1(x), (4.56)

Lv(x) = L[2 -1'V;+l-lo,n+l_.l!k:n+1(X)

k=l
(n + 1- j)/j,n+1(x). (4.57)

Bound (4.53) is attained by

0, i'Il £
m - < 0,
F(x) = { fJ-;~ (Brii) , il 0< £ <
!;,n(-r.)
'I - m - B '
1 - (1 - 'Y.) exp ( il
'I
£
m -
>
!;,n("Y.)
B .
(4.58)

Formulae (4.45) through (4.47), and (4.53) through (4.55) are concluded
from (4.34) and (4.41). Moreover, (4.48) and (4.49) as well as (4.56) and
(4.57) differ from K and L defined in (4.42) by positive functional factors
common for all j, and original variables are replaced by V (x) in the lat-
ter case. All bounds are achieved by absolutely continuous distributions.
The left-hand parts of (4.50) and (4.58) are the inverses of polynomials
equal to /j,n up to scale factors, and cannot be written explicitly. They
have uniform and exponential right tails, respectively. Bound (4.15) was
derived directly from the Schwarz inequality. Using the following integral
approximations of harmonic series

n+ 1 n + 1/2
In n+ 1 -J. < I'V;on < In n +1/2- J. ,

we deduce that condition 1'V;,n ::; 2 providing bound (4.52) is true for
j ::; (1 - e- 2 )(n + 1/2) and false for j ~ (1 - e- 2 )(n + 1). The difference
between both estimates is (1 - e- 2 )/2 ~ 0.43233 which implies that for
given n the condition has to be directly checked for one j at most. Bounds
of Theorem 12 are tighter than those of Theorem 11, because a narrower
class of distributions was treated there. This is confirmed by numerical
comparisons presented in Gajek and Rychlik [33, Table I, p. 162].
4.3 Distributions with Monotone Density and Failure Rate on the Average 69

4.3 Distributions with Monotone Density and

Failure Rate on the Average
The crucial points in calculating mean-variance bounds for EpXj : n for F
with decreasing density and failure rate on the average lies in deriving pro-
jections p~.u(fj:n -1) and p~. v (fj:n V-I), respectively, onto convex cones
of nondecreasing starshaped functions that are orthogonal to constants. In
fact, we consider a more general problem of projecting h = hn W satisfying
(4.29) to (4.33) onto C( w which gives
-*

Note first that

p(
-*
wh(a) E [h(a), h(c».
Functions starting from a higher level uniformly majorize h, and contradict
the statement of Lemma 2. Ones with g(a) < h(a) are eliminated by

gh(a) = max{g, h(a)} E C{.w'

Therefore in Lemma 15, describing the shape of projection functions, we
confine ourselves to functions starting from a range point of h.
Lemma 15 Define functions

h((3), if a ~ x ~ (3,
hf3a(x) = { h(x), if (3 ~ x ~ a, (4.59)
h(al:::~(f3) (x - a) + h((3) , if a ~ x < d,

for some a =f. a ~ (3 ~ a < c. For every a ~ (3 < c and 9 E C( w satisfying

g(a) = h((3) there exists a =f. a E [(3, c) such that hf3a E C( ;- and
-*

Il h f3a - hll ~ Ilg - hll· (4.60)

Functions (4.59) are obviously nondecreasing, and constant functions

hf3f3(x) == h((3), a < (3 < c,

are also possible here. However, starshapedness is not apparent here. For
fixed (3 E [a, c), let

sf3( "')
.... -_ h(a) - h((3) , a =f. a E [(3, c), (4.61)
a-a
denote the slopes of lines

lBA(X) = sf3(a)(x - a) + h((3) (4.62)

70 4. Order Statistics of Independent Samples

passing through B = (a, h(fJ)) and points A = (a, h(a)) , fJ $ a < c, of the
graph of h (cf. the last line of (4.59)). Write
8a (a) = h'(a) ~ 0
for completeness. We first prove the following.
Lemma 16 For every fJ E [a,c) there exists a(fJ) E B(fJ) = [max{fJ,b},c)
such that (4.61) increases on (fJ,a(fJ)) and decreases on (a(fJ),c).
In consequence, (4.59) is actually starshaped if a E [fJ,a(fJ)]. Observe that
a(fJ) is the point where lBA.{,8) is tangent to h and therefore can be deter-
mined by
h(a) - h(fJ) = h'(a)(a - a). (4.63)

PROOF OF LEMMA 16. The statement is trivial for fJ = a = b with

a(fJ) = a. Indeed, by concavity of h on (a, c), for any a < al < a2 < c
point Al = (al,h(al)) lies above lBA2 (cf. (4.62» and so 8a (al) > 8a (a2)'
We now claim that 8,8 is increasing-decreasing on B(fJ) if either fJ > a = b
or fJ ~ a :F b. The proof consists in checking the falsity of the contradiction:
for every fixed a E B(fJ) there exist al < a < a2, ai E B(fJ) , such that
8,8(ai) ~ 8,8(a), i = 1,2. IT lBA is tangent to h at a, it runs above all graph
points A', a :F a' E B(fJ). Accordingly, 8,8(a') < 8,8(a). IT lBA is a secant
line to the curve, then these cross each other once more in B(fJ) at most.
IT h - IBA changes its sign from - to + at a, then for all B(fJ) 3 al < a
we have
h(al) < lBA(al),
8,8(al) < 8,8 (a).
IT the sign changes from + to -, then the analogous relations hold for
B(fJ) 3 a2 > a. Therefore our claim is actually true.
By continuity of 8,8, it remains to show that (4.61) is increasing in (fJ, b)
when a $ fJ < b. Note that
hh{,8)(X) = max{h(fJ),h(xH
is convex in (a, b) and strictly convex in (fJ, b). For every fJ $ al < a2 $ b
point Al lies below lBA2' and hence 8,8(al) < 8,8(a2). This completes the
proof. •

PROOF OF LEMMA 15. IT g(a(fJ)) ~ h(a(fJ)), then we can take h,8li{,8).

Since h(x) < h(fJ) on [a, fJ), we have h < h,8li{,8) $ 9 there, because con-
stant h(fJ) is the smallest nondecreasing function starting from level h(fJ).
Also, h,8li{,8) = h is the optimal approximation of h on [fJ, a(fJ)]. Finally,
for x > a(fJ), we have

h(x) < h(a(fJ)) - h(fJ) (x - a) + h(fJ) < g(x).

- a(fJ) - a -
4.3 Distributions with Monotone Density and Failure Rate on the Average 71

The former inequality holds, because the middle term defines the line that
is tangent to h at a«(3) and majorizes h. The latter is a consequence of
the fact that the line is the smallest starshaped function in [a«(3) , d) that
passes through Band ..4«(3) = (a(.B) , h(a«(3))). The above arguments show
that hf3ii:(f3) actually satisfies (4.60).
However, h -:f:. hf3ii:(f3) ~ h and, due to Lemma 2, the approximation can
be further improved by a downward translation of hf3ii:(f3). We consider the
case g(a«(3)) < h(a«(3)) now. Define

0= inf{x > a«(3) : g(x) ~ h(x)},

putting 0 = d if 9 < h in (a«(3), d). Take

lBD(X) = h(01 =~«(3) (x - a) + h«(3).

We first prove that lBD < h in (a«(3),o). If (c,o) -:f:. 0, then lBD - h is
strictly increasing there, and lBD(O) - h(o) ::; O. For x E (a«(3) , min{ c, o})
the inequality follows from the strict convexity of IBD - h in the interval,
and its nonpositivity at the endpoints. Now we check that lBD - h changes
the sign in (a, a«(3)) once at most. Assume that a is the largest point of
sign change in (a, a«(3)). Evidently lBD - h is positive and negative on the
left and right to a, respectively. If a E [b, a «(3)) , then convexity of lBD - h
implies its positivity in (b, a). This is also positive in (a, b) by its concavity
there, and nonnegativity at the endpoints. If a E (a, b), it suffices to repeat
the above argument with b replaced by a. Assume that unique a exists.
We can write
lBD(X) = h(a) - h«(3) (x - a) + h«(3).
a-a
Since lBD ~ h«(3) , we get a E [(3, a«(3)). Note that in the class of non-
decreasing starshaped functions passing through B = (a, h«(3)) and D =
(0, g(o)), function lBD is maximal in [a, oj and minimal in [0, d). Therefore

g(x) ::; lBD(X) ::; h(x), if x E [a, oj,

h(x) ::; lBD(X) ::; g(x), if x E [0, d).

Also, we have
g(x) ::; h«(3) ::; g(x), x E [a, (3].
Consequently, hf3a defined as h«(3), h, and IBD in [a, (3j, [(3, aj, and [a, d),
respectively, lies closer to h than g. Since (3 ::; a < a«(3) , we see that
hf3a E cC w·
If (3 ~·a, it may happen that lBD has no sign changes in [a,o). Then
IBD(a) = h(a) = 0 and lBD < h in (a,o), and lBD > h in (0, d). The
last relation is a consequence of concavity and ultimate decrease of h. If
lBD = 0::; h, referring to Lemma 2 we decrease the distance to h by adding
72 4. Order Statistics of Independent Samples

a positive constant I~D. H IBD has a positive slope, we can take a line IBD
running through (8,g(8)) with a slightly smaller slope. Observe that IBD
lies closer to h than the original one in [a, 8] and [8, d). Both modifications
lead to linear functions that cross h once at some a < ii«(3). Therefore we
are in position to apply the construction of the previous paragraph, which
ends the proof. •
Below we determine the optimal parameters. For a ~ (3 ~ c and (3 ~ a ~
c, a '" a, set

K(a, (3) = fad [ha.a(x) - h(x)] (x - a)w(x)dx, (4.64)

L(a, (3) = 1d [ha.a(x) - h(x)] w(x) dx. (4.65)

Then for a < (3 ~ C write

k«(3) = K«(3, (3) = !.ad [h«(3) - h(x)] (x - a)w(x) dx, (4.66)

£«(3) = L«(3, (3) = 1d [h«(3) - h(x)] w(x) dx = h«(3) - 1. (4.67)

Lemma 17 Let Pbe the unique zero of (4.67) in (a, c). If

k(p) = i d
[1 - h(x)] (x - a)w(x) dx ~ 0, (4.68)

then P{. wh = h/J/J = 1.

Otherwise there exists a unique pair «(3.. , a .. ), a ~ (3.. < p, max{(3.. , b} <
a .. < c, determined by equations
K(a,(3) = 0, (4.69)
L(a,(3) = 0, (4.70)

_.
such that Pt' wh = ha•.a., defined as in (4.59).
Since h is strictly increasing from h(a) = 0 to h(c) = suph > 1 (cf. (4.29)
and (4.30)), p is actually well defined.
PROOF OF LEMMA 17. By Lemma 15, we should minimize

D«(3, a) = 1d [[Link](x) - h(xWw(x) dx

= 1.a [h«(3) - h(xWw(x) dx

+ fad[IBA(X) - h(xWw(x) dx (4.71)
4.3 Distributions with Monotone Density and Failure Rate on the Average 73

(cf. (4.61) and (4.62)) with respect to two parameters (3 E [a, c) and a :f;
0: E [max{(3,b},o((3)) c (a,c). Fixing (3 and differentiating (4.71) with
respect to 0:, we obtain

8D((3,0:)
80:
= 0:: a [h'(o:) -l~A] id[lBA(X) - h(x)]w(x) dx

= _2_[h'(0:) -l~A]K((3,o:) (4.72)

o:-a
(cf (4.64)). If (3 < 0: < 0((3), then h-lBA changes the sign from - to + at 0:
and hence the expression in brackets is positive. This vanishes at 0((3), but
0((3) cannot be optimal, because hf3(i(f3) ;::: h. Analyzing the sign of (4.72),
it suffices to concentrate on (4.64). We have

8K((3,0:)
J::l
vO:
_ h'(o:) -l~A
-
0: - a
I d( _ )2 ()d
0<
X a wx x>,0

which implies that (4.72) is the product of a positive function and increasing
K((3, .). Since the integrand is positive for 0: = 0((3), we have K((3, 0((3)) >
O. If 0: = (3 > a, then lBA = h((3). It follows that K((3) < 0 as (3 '\.- a and
K((3) > 0 for (3 /" c.
We can summarize the behavior of (4.71) as follows. If (3 is small enough
then (4.72) is negative for 0: close to (3, and changes its sign at an 0:.((3) E
((3,0((3)) where (4.64) vanishes and the unique minimum of D((3,·) is at-
tained. If (3 ;::: S satisfying K(S) = 0 then (4.72) is positive for all 0: > (3.
Then D((3,·) is minimized at 0:.((3) = (3 which gives a constant approxi-
mation h{3f3 = h((3). It remains to choose (3 E [a,e) such that ((3,0:*((3))
minimizes (4.71), where 0:*((3) > (3 satisfies K((3,o:*((3)) = 0 for (3 < Sand
0:*((3) = (3 for (3 ;::: S.
By Lemma 2, a necessary condition for that is

L((3,o:*((3)) = l d
[h{3o<.(f3)(x) - h(x)]w(x) dx = 0 (4.73)

(cf (4.65)). It is clear that (4.67) strictly increases from negative L(a) to
positive L(c) = L(c,o:*(c)). We have L($) = 0 if

hSS = h($) = ld h(x)w(x) dx = 1.

We show that L((3, 0:* ((3)) is also increasing, when 0:*((3) > (3 is determined
by (4.69). Consider
74 4. Order Statistics of Independent Samples

dL((3, a*((3»
d(3

(4.74)

Plugging
8K(p,a.(p»
00. ((3) = 813
8K(p,a. (13»
d(3
8a.
= h'((3) f:.(p) [X - a*((3)](x - a)w(x) dx >0
[h'(a*((3» -lkA.(p)l f:.(p) (X - a)2w(x) dx

into (4.74), we obtain

dL((3, a.((3» h'((3)

=
d(3 f:.(p) (X - a)2w(x) dx
x {[p W(X) dx [d (x _ a)2w(x) dx
ia ia.(p)

+ ['I. W(X) dx [d (x _ a)2w(x) dx

ia.(p) ia.(p)

[ [d (x_a)W(X)dx]2}. (4.75)
ia.(p)

The last two lines are positive by the Schwarz inequality and so is the whole
expression in the curly brackets. Because h' ((3) and the denominator are
positive as well, the same holds for (4.75).
/J
We are thus led to the following conclusions. H ~ /J then ((3*, a. ((3.» =
/J)
(/J, is the unique pair satisfying necessary condition (4.73) for minimiz-
ing(4.71). This gives the first statement of Lemma 17. To see that ~ /J /J
coincides with (4.68), we note that (4.66) satisfies

lim /«((3) <0< lim /«((3)

p'\"a p/,c
4.3 Distributions with Monotone Density and Failure Rate on the Average 75

and
[('(f3) = h'(f3) i d
(x - a)w(x) dx > 0.

Therefore [( is incr~asing and has a sin&.le ze!o at p. The same holds for L
which vanishes at f3. Hence conditions f3 ~ f3 and (4.68) are equivalent. If
p < p, then
L(P) = lim_L(f3,a.(f3» > 0,
f3/'f3
and we can make L(f3, a. (f3» smaller by decreasing f3. There is some f3. E
[a, P) that satisfies (4.73), because the opposite contradicts the existence
of the solution. Taking a. = a.(f3.), we see that (4.73) and (4.70) are
identical, and (4.69) holds by the definition of a.(f3.) . •

Theorem 13 (F t. W) Suppose that the density w of Wand h = hn W

for some 2 ~ j ~ n satisfy (4.29) through (4.33).
P
If for unique = P(j, n) E (0, (j - 1)/(n - 1» satisfying

(4.76)

we have
fd . [1- hnW(x)](x - a)w(x)dx ~ 0, (4.77)
lW-l(f3)
then EFXj:n ~ J1.F for all F b W.
Otherwise there exists a pair (a., f3.), aw ~ f3. < W- 1 «j -1)/(n - 1»,
aw < a. E [f3., W-l«j -1)/(n -1))), determined by equations

hnW(a) - h nW(f3)
a-a
Ia
d(x _ a)2w(x)dx

+ id[jj:nW(f3) - hnW(x)](x - a)w(x) dx = 0, (4.78)

fJ,nW(JJ) [I." w(z)dx+ {W(Z)dx]

a-a a
I
+hnW(a) - hn W (f3) d(x _ a)w(x)dx

-Fj:nW(f3) + Fj:nW(a) -1 = 0, (4.79)

such that
EFXj:n - J1.F
0'
<
-
B = B O~.w (.}, n) (4.80)
F -
76 4. Order Statistics of Independent Samples

for

B2 = (fj:n W(!3.W [1 13
• w(x) dx + i~ w(x) dx 1
3-(2nnif
(2!-2)(2n-~i)
+ n [F2j-1:2n-1 W(a.) - F2j-1:2n-1 W(!3.)]

+ 2hn W (!3./j:n W (a.) - hn W (!3.) {d (x _ a)w(x) dx

a. - a 10:.
+ [hnW(a.) __ h nW (!3.)] 2 {d (x _ a)2w(x) dx -1. (4.81)
a. a 10:.

!
The equality in (4.80) is attained by

0, it X-IJ<_l-hnW(.B.)
~ q B'
r-1(Bx-1J + 1) it _l-li: nW (i3.) < X-IJ
F(x)= 3:n q ,
~ B - q
< _l-lj:n W(o:.)
W (a + (0:. hn
-a)[B9+l- hn W(i3.)])
W(o:.)-hn W(i3.) ,
it
~
X-IJ
q -
B
> _ 1-
B
'
Ij:n W(o:.)
.
(4.82)
Thivial bounds identical with general ones for the sample minimum are
consequences of
P~.w(h -1) = P: wh -1 = 0
- -'
under (4.68), rewritten as (4.77). These apparently hold for small order
statistics. Equations (4.78) and (4.79) follow from (4.69) and (4.70). Plug-
ging in W = U, V we specify (4.77) to (4.82).
Proposition 8 (decreasing density on the average) If for given 2 ~
j ~ n - 1, and ~ defined in (4.76)

(4.83)

holds, then EFXj : n ~ /LF.

Otherwise there are unique 0 < !3. < ~, !3. < a. < (j - 1) / (n - 1) that
solve equations

= 0, (4.84)

= 0, (4.85)

and then
(4.86)
4.3 Distributions with Monotone Density and Failure Rate on the Average 77

for

The equality in (4.86) holds for the location-scale family of distributions

0,
:-l(BX-/f + 1)
f 1:n u '

F(x) =
a.[B T H-/j,n(.8.)]
Ij,n(a.)-lj,n({3.) ,

1,

Proposition 9 (decreasing failure rate on the average) If for 2 ~

j ~ nand [3 defined in (4.76) we have
j

L
A

(1-[3)[1-1n(I-[3)]- F~+~_t:~) +In(I-[3)Fn+1-j:n([3) ~ 0, (4.89)

k=l

then EFXj :n ~ /-LF.

Otherwise there exist 0< (3. < -In(l- [3), (3. < a. < In(n - 1)/(n - j)
solving equations

(a +2+~) e-afi,nV(a) - (1 +~) e-afi,nV((3)

1
-L
j
n +1_ k Fk:n V(a) - aFj:n V(a) 0, (4.90)
k=l

a
(1- e-{3 - e: ) fi,nV((3) +(1 +~) e- a fi,nV(a)
-1 + Fj:n V((3) - Fj:n V(a) = 0, (4.91)

such that
EFXj:n - /-LF <
_
B _ BO
-
(. )
)-- V ),n , (4.92)
UF -'
where
78 4. Order Statistics of Independent Samples

B2 = (1 - e- f3 • + 2a;2e -O:·) [/i:n V(.8.W

2(2 + a*)a;2 e -o:· /i:n V (.8.)/i:n V(a.)
+ (2 + 2a. + a~)a;2e-a. [/i:n V(a.W - 1

'-(2"" iT
(2!-2W"-~;)
+ n [F2j-l:2n-l V(.8*) - F2j-l:2n-l V(a*)]. (4.93)

I
Bound (4.92) becomes the equality for

0, If X-J.! < _1-1;,,, V(f3.)

B'
.f
q
:-l (B X-J.! + 1)
f J:n _1-/;:" V(f3.)
B
<
x-J.!
-
F(x) = q ,
< _1-/;:n V(a.)
q

B '
1 a.[B9+1-I;,,, V(f3.)] zf x-J.! > _1-/;:n V(a.)
- exp (- 1;,,, V(a.)-I;,,, V(f3.) ), q - B •
(4.94)
Distribution functions (4.50) and (4.58) are similar to (4.88) and (4.94),
respectively. The essential difference is that the latter ones have jumps of
height .8., V (.8.), respectively. If

[1 - /i:n V(.8*)]aF
[Link] = BO (.)
tov),n
, (4.95)

then (4.94) is actually a DFRA life distribution starting at o.

Under assumptions of Proposition 8, the sample maximum attains gen-
eral bound (4.2), because the density of (4.3) is decreasing. Surprisingly,
general bounds (4.11) and (4.4) are also attained for arbitrary order statis-
tics from the samples with increasing density on the average. The same
holds for the distribution of the narrower class of IFRA distributions, and,
more generally, for families of distributions satisfying F j* W, when con-
ditions (4.29) to (4.33) hold. This is a consequence of the fact that non-
decreasing functions p/' /i:nW, 2 ~ j ~ n (see (4.1) and (4.8)) may be
approximated in £2([aw,dw),w(x)dx) with any desired accuracy by se-
quences hk' k ~ 1, of antistarshaped nondecreasing functions starting from
sufficiently low level hk(a). For example, we can take

where .8k \.t a, and ak are sufficiently large. Alternatively, we can also an-
alyze the norm convergence of hk W- 1 in £2([0,1), dx). For more reasoning
details we refer the reader to Rychlik [89]. Summarizing, we have
 4.4 Symmetric Unimodal Distributions 79

Theorem 14 (F:j. W) For 2 $ j $ n -1, h = /;:nW and w = dW/dx

satisfying (4.29) to (4.33), general bounds (4.11) are attained in the limit by
sequences of absolutely continuous distribution functions FIc :j. W whose
quantile functions tend in L2([0,1),dx) to that of (4.12). In particular,
the general bounds cannot be improved in the classes of distributions with
increasing density and failure rate on the average. Analogous conclusions
hold for the sample extremes.

4.4 Symmetric Unimodal Distributions

For j $ (n+1)/2, we have EFXj:n $ [Link] for all F !:s U which is guaranteed
by the symmetry assumption only. Otherwise the bounds for standardized
expectations of order statistics are nontrivial, because

EUXj:n - [Link] = V3 (~ - 1) > O.

au n+1
In the case j = n, the equality in (4.21) holds for a symmetric unimodal
distribution. In the remaining cases (n + 1)/2 < j < n, we obtain the
optimal bounds by projecting

Sj:n(x) = /;:n(x) - fn+1-j:n(x), 1/2 $ x < 1,

onto Ct.2U-l' Applying Lemma 12 for verification of (4.29) through (4.33),
we are- in a position to describe the shape of projections by means of
Lemma 13. It immediately follows that for (n + 1)/2 < j $ min{n-
2, [n + (3n - 5)1/2]/2} and j = n - 1 $ 6, the projection is linear, and the
respective bounds are attained by uniform samples. In fact, the uniform dis-
tributions provide the optimal bounds for a wider range of order statistics.
The necessary and sufficient conditions are presented in Theorem 15.
Theorem 15 (symmetric unimodal distributions) For (n + 1)/2 <
j $ n -1, put

Kb(x) = Ku(x) - Kii(x), (4.96)

LU(x) = Lu(x) - Lu(x), (4.97)
for 1/2 $ x < 1, where Ku and Lu are defined in (4.48) and (4.49),
respectively, and Kii and Lu are respective modifications of (4.48) and
(4.49) that consist in replacing j by n + 1 - j.
If LlI is positive on a right neighborhood of 1/2, then

EFXj:n - [Link]
O'F
< V3 (~ _
- n +1
1) '
(4.98)

where the equality holds for F uniformly distributed on [J.t - V30', J.t + ..;30'].
80 4. Order Statistics of Independent Samples

Otherwise, under notation (4.34) we obtain

(4.99)

where

3j[Fn+2-j:n+1 ((3.) - Fj+1:n+1 ((3.)]

=
(1- (3.)3(n + 1)
3(3.[Fn+l- j:n((3.) - Fj:n((3.)] 3sj:n((3.) (4.101)
(1 - (3.)3 2(1 - (3.) ,

(3. = min {(3 u

> ~ : Kt;((3)L ((3) = 0 } . (4.102)

The equality in (4.99) is achieved by

0, zf X-/I: < _ Sj,n (13.)+<>. (1-,8.)

U hB '
if _ S;'n(,8.)+<>.(l-,8.)
-
1 - (3 + Sj,n(,8.) + hB X-/I:
* 0... q' £r", hB
< < _ Sj,n(t3.)
x-J!.
- U hB'
-

F(x) = 8-:- 1 ( '2B x -J!.) zf - 8;'n(,8.) < x-J!. < Sj,n(,8.)

3:n V ~ u ' hB - u - hB'
(3 - Sj,n (13.) + VZB x-J!. zf Sj,n (13.) < x-J!.
* a... a. IT' hB - u
< Sj,n (,8.)+<>. (1-,8.)
- hB '
1, zf x-J!. > 8j,n (,8.)+<>. (1-,8.)
u - hB
(4.103)
Analysis similar to that in the proof of Theorem 11 is applied here. Re-
placing h = /i:n by h = 8j:n = /i:n - fn+1-j:n in (4.38) and (4.42) we write
D S (a, (3), K 8 ((3), £S ((3), choosing optimal slopes (4.41) for various (3, and
try to determine (3 E [1/2,b) that minimizes DS(a.((3),(3) under condition
K8((3) 2:: O. Since K((3) and L((3) of (4.42) are linear operators acting on h,
we have
 4.4 Symmetric Unimodal Distributions 81

TABLE 4.1. Sharp uniform mean-variance bounds on expectations of order statis-

tics from independent samples of size 20 for various families of distributions.
j G S SUN DDA DFRA
10 0.56881 0 0 0.02526 0
11 0.64211 0.17773 0.08256 0.19442 0
12 0.72127 0.50155 0.24742 0.36024 0
13 0.80855 0.74687 0.41240 0.52453 0.08848
14 0.90714 0.90447 0.57730 0.69045 0.24189
15 1.02182 1.00151 0.74232 0.86310 0.41773
16 1.16054 1.08224 0.90735 1.05096 0.62534
17 1.33774 1.18095 1.07220 1.26912 0.88206
18 1.58450 1.32488 1.24648 1.54902 1.22562
19 1.98814 1.57364 1.54236 1.97731 1.76097
20 3.04243 2.26455 2.26463 3.04423 3.03006

with a positive factor M((3) independent of j (see (4.42) and comments

following Theorem 12). A thorough analysis leads us to the conclusions
that Ku((3) is +- on (1/2, b), and LfA(3) is either + or -+ on (1/2, c).
If LU(I/2+) > 0, then (3* = 1/2 is the solution that implies linearity of
projection and the resulting quantile function. Otherwise optimal (3. > 1/2
is the smaller of zeros of Ku and Lb. Then

and (4.100), (4.101), and (4.103) are derived from general formulae by
elementary calculations.
Note that (4.103) has a density symmetric about [Link], a finite support
with uniform ends, and a (principally infinite) peak at the center. The
statement of Theorem 15 is weaker than those of Theorems 11 and 12 in
that we were not able to determine explicitly the pairs (j, n) for which
L(a) > 0, and the resulting optimal bounds are determined by the minimal
distributions W in the class with respect to the order.
In Table 4.1 numerical evaluations of mean-variance bounds are pre-
sented for the jth order statistics, 10 ~ j ~ 20, of i.i.d. samples of size
n = 20, coming from general (G), symmetric (S), symmetric unimodal
(SUN) populations, and those with decreasing density and failure rate on
the average (DDA and DFRA, respectively). For j ~ 9, all bounds are triv-
ial except for the first case. The values of the third column were presented
in Gajek and Rychlik [33, Table III]. Numerical bounds for the DDA and
DFRA samples of size 15 can be found in Rychlik [89].
82 4. Order Statistics of Independent Samples

4.5 Bias of Quantile Estimates

Order statistic Xj:n is the most natural nonparametric estimate of quan-
tile p - l (P) if j I n is close to p, although some more sophisticated smooth
estimates based on kernels and Bernstein polynomial estimates were also
constructed (see, e.g., Sheather and Marron [97], Huang and Brill [40], and
Cheng [21]). If j(n)ln -+ p and p-l(P) is unique, then Xj(n):n -+ p-l(P)
almost surely. Optimality of order statistics in quantile estimation for finite
samples under various criteria was proved by Zielinski [103, 105, 106]. Here
we employ the projection method for gauging the bias EFXj :n - p - l (P) of
quantile estimation by sample quantiles. For convenience we assume that
j In = p precisely. We first consider general distribution functions. Since
both the upper and lower bias deviations are significant here, we present
the lower bounds for the difference as well.
Theorem 16 (general distributions) For arbitrary 1 ~ j < n < 00,

(4.104)

For j = 2 and n = 3,4,5,6, and 3 ~ j < n < 00,

EFXj :n - p-l(P) > -B/'(· n ) _ _ Pj:n(P) (4.105)
(TF - - ), ,p - (P(1- p)]1/2·

Both (4.104) and (4.105) are attained by the (limiting) two-point distribu-
tion

For j = 1 < n < 00,

EFX1:n - p-l(P) > -B = -B/'(1 n ) (4.107)
(TF - - - "p,
where
2
B 2 = 2nn_ 1 Pl:2n-l (1) + n n_
;;;
2
1 Pl:n (1);;; . (4.108)

The equality in (4.107) holds for

P(x) =

(4.109)
 4.5 Bias of Quantile Estimates 83

For j = 2 and n ~ 7, we have

EFX2 :n - F-l(P)
-----..;;;....;.. ~
-B
_-- -B/'(2
_ )
,n,p, (4.110)
O'F

where

B2 = (2n ~ ~t1;~ 1) [F 2n - (~)3: 1 - F3 :2n - 1 (a*)]

+ a*#n(a*) + n~ 2Fi,n (~) , (4.111)

and a. E (l/(n - 1), 2/n) is uniquely defined by

[(n - 1)2x 2 + (n - l)x + 1](1 - x)n-2 = 1. (4.112)

Bound (4.110) is attained by the distribution function

<_

1
0, it X-J.! hn(a.)
'J u B'
f -l(-B~) if - hn(a.) < z-J.! < _ hn(2/n)
F( ) 2:n - u ' B - u - B'
(4.113)
X = ~, it
'J
_ hnJ2/n)
l!.
<
-
< nF2:n]2/n)
X-J.!
u (n-2)l!. '
1, it X-J.! > nF2:n (2/n)
'J u - (n 2)l!.

Note that
B(j,n,p) - B(j,n,p) ~ [P(1- p)tl/2. (4.114)
Precisely, we have equality in (4.114) under conditions of (4.105), and strict
inequality holds only in the exceptional cases treated in the last two state-
ments. This implies that the bias oscillation presumably does not depend
on the sample size, is smaller for the central quantiles, and increases to
infinity on the tails. The only positive effect of the sample increase is that
the absolute deviation of the bias decreases to 1/[4p(1 - p)j1/2, because
Fj:n(P) -+ 1/2 by the de Moivre-Laplace theorem. However, for no quan-
tile the bias oscillation tends to zero. This is obvious, because we take into
account the distributions with nonunique pth quantiles (estimation prob-
lems in such cases are discussed in Feldman and Tucker [28]). Removing
such distributions would not help here either, because the rate of conver-
gence of Xj:n to F- 1 (P) depends on the slope increase of Fat F-l(P), and
it is impossible to determine uniform rates without imposing conditions on
the slope (cf., e.g., Zielmski [104]). In fact, all the bounds of Theorem 16
are optimal for the class of strictly increasing F. However, the conclusions
on attainability should be formulated more carefully then.
The proof of Theorem 16 is based on the representation
84 4. Order Statistics of Independent Samples

and Lemma 3. The upper bound (4.104) is the L2- norm of the derivative
of the greatest convex minorant of Fj :n - 1[p,1), and the centered quantile

from 0, increases on [O,p), jumps down to Fj:n(P) - 1 < 0, increases to

on [p, 1), and is concave on [p, 1), its greatest convex minorant has two linear
°
function of (4.106) is proportional to the derivative. Since F j : n -1[p,1) starts

pieces on [O,p] and [p, 1] with slopes -[l-Fj:n (P)]/p and [l-Fj:n(p)]/(l-p),
respectively. An easy computation leads us to the final claim.
In order to get the lower bounds it suffices to find the greatest convex mi-
norant for 1[p,1) - Fj:n. Distribution function Fl:n is concave, and therefore
the greatest convex minorant of 1[p,1) (x) - Fj:n(x) coincides with -Fl:n(X)
and the straight line F 1:n (P)(x -1)/(1- p) on [O,p) and [p, 1), respectively.
For 2 :5 j < n, the problem is that function 1[p,1) (x) - Fj:n(x) = -Fj:n(x)
°
for :5 x :5 p = j / n is decreasing concave-convex with the inflection point
(j - l)/(n - 1) < p. Two cases are possible: either the straight line join-
ing (0, -Fj:n(O)) = (0,0) with (p, -Fj:n(P)) lies entirely below the graph
of -Fj:n, or the line crosses the graph there. In the former case, the line
becomes a part of the greatest convex minorant. In the latter one, the mino-
rant consists of the line passing through (0,0), and tangent to the graph at
a point 0:* E ((j -1)/(n-1),p), and -Fj:n(x) itself on [o:*,p]. In both cases,
the line passing through (p, -Fj:n(P)) and (1, 1[p,1)(1) - Fj:n(l)) = (1,0)
is the remaining part of the minorant. The problem of settling which case
actually holds is equivalent to checking the sign of the expression

b(j,n) = Fj :n (~) - ~hn (~) . (4.115)

°
We have b(j,n) :5 in the first case, which leads to (4.105), and (4.108)
holds otherwise. A thorough study carried out in Okolewski and Rych-
lik [65] shows that (4.115) is positive for j = 2 with n ;::: 7 only. This
is based on analysis of sign changes of sequences defined as integrals of
a fixed function with sequences of totally positive functions (see Karlin [43]
and [44, Chapter 1] for details). We preserved the dual notation j/n and p
in Theorem 16 on purpose. The reason is that the formulae remain true if

°
p ~ jfn as well. For instance, both (4.104) and (4.105) hold for all j, n with
< p = (j - l)/(n - 1) < 1. In fact, the former is true for all 1 < j < n
with p ;::: (j - l)/(n - 1) and so is the latter for all 1 < j < n with
p:5 (j - l)/(n - 1).
Now we restrict ourselves to the distributions following a given W in the
convex and star orders. We present the upper bounds only. As above, all
are true for arbitrary j,n with p ;::: (j - l)/(n - 1). The auxiliary dual
problem to solve is to maximize the functional
EFXj:n - F-1(P)

= lim fl F-1(x) [hn(X) - _1_ 1[p,q)(X)] dx

q'"pjo q-p
 4.5 Bias of Quantile Estimates 85

over C~cw and c~. w' For convenience, we replace the constant subtrahend
in the orackets by a variable one to make the difference constant. This does
not affect the limit

EFXj:n - F-1(p)

= ¥~ld F-1W(x) [hnW(X)I[B,d)\[b,C) (x)

1-Fj:nW(c) + Fj:nW(b) 1 ()] ()d (4.117)
W(c) _ W(b) [b,c) x W X x.

We are interested in projecting hj:n,b,c(x) which denotes the expression in

the square brackets.
Lemma 18 Ifh E L2([aw,dw),w(x)dx) is nonnegative on [a,b), negative
constant on [b, c), and nonnegative nonincreasing on [c, d) for a < b < c <
d, and
ld h(x)w(x) dx = 0, (4.118)

then for every g E c~_c w there exists ga/3"Y E c~_c w defined as

(4.119)

with a, fJ > 0 and b :::; 'Y :::; c such that

Function hj:n,b,c is orthogonal to constants and satisfies the other assump-

tions of Lemma 18 if p 2: (j -1)/(n-1). The orthogonality property (4.118)
is important here: it allows us to restrict assumptions on hj:n,b,c to sign and
monotonicity conditions. In other lemmas describing the shape of projec-
tions, the concavity, and convexity in some regions play a significant role
as well.

PROOF OF LEMMA 18. Since 9 is nondecreasing and integrates to 0, the

only possible constant function is equal to O. Otherwise we have g(a) <
0< g(d-). H g(b) 2: 0, then

g+ = max{g,O} E Cf':w
lies closer to h, and, by Lemma 2, so does
86 4. Order Statistics of Independent Samples

Repetitive application of the procedure allows us to restrict ourselves to

functions satisfying 0 > g(b) = -{3, say. Similar arguments, with g+ re-
placed by gh(b) = max{g, h(b)} excludes the functions for which g(a) <
h(b) = infa<z<d h(x). We are therefore in a position to assume

h(b) ~ g(a) ~ g(b) ~ 0,

and define
6 = inf{x~: g(x) ~ h(x)}, (4.120)
setting 6 = d if d < 00 and g(d-) :5 h(d-).
If 6 > c, the straight line leo secant to 9 at e and 6 has a positive slope

g(6) - g(e)
°= 6 -e) ,

and runs above 9 in [e,6] and below 9 elsewhere. In particular, we have

g(b) ~ le6(b). It follows that Ic6 crosses the constant function

at a '"'I E [b, e). We easily see that the broken line

satisfies

g(x) ~ gQ{j'Y(a;) ~ hex) if x E (a, b) U (e,6), (4.121)

h(x) ~ gcr.8'Y(x) ~ g(x) if x E (b,e) U (6, d), (4.122)

which is the desired conclusion.

If 6 = e we replace le6 by the line Ie tangent to 9 at e. It runs beneath g,
and crosses lb at some '"'I E [b, e). Its truncation at the levellbh) = -{3 has
form (4.119), and satisfies (4.121) and (4.122) with (e,6) = 0.•

Lemma 19 Under the hypotheses and notation of Lemma 18, we have

P~_c wh(x) = gcr •.8.'Y. (x)

for some '"'I. E [b, c),

Jd (x-'"'I.)h(x)w(x)dx
0.=0.('"'1.) = 'Y. 2,(4.123)
J~ (x-'"'I.)2w(x)dx- [J~ (x-'"'I.)w(x)dx]
d
{3.={3.h.) = o.h.)l (x-'"'I.)w(x)dx. (4.124)
'Y.
 4.5 Bias of Quantile Estimates 87

PROOF. For fixed a", the function

is minimized with respect to translation parameter (3 at

(3.(a,,) = l d
[a(x-,)l["!,d)(x)-h(X)]W(X)dX

= a i d
(x - ,)w(x) dx

(cf. (4.124)). Furthermore, for fixed, E [b,c],

=
D(a, (3. (a, ,), ,)

f.' {h(X) - a [(X - 7)1[,.,) (x) - .t 1

(y -7 )w(y) dy

is a convex quadratic function in a, with the global minimum at a.(,),

rw(x) <Ix

defined as in (4.123). It is obvious that

is a continuous function that attains its minimum at some , •. Observe

that the numerator of the right-hand side of (4.123) is positive. By the
Schwarz inequality with condition f~ w(x) dx < 1, the denominator is pos-
itive as well. Therefore a.b.) > 0, which together with (4.124) yields
(3. = (3. (a. b.), ,.) > O. The solution to our minimization problem is
actually a convex nondecreasing function . •

If c '\, b, so does , •. For h = hj:n,b,c we have

o < -I "Y.
c
(x - ,.)h(x)w(x) dx

= 1-Fj:n W (c)+Fj :nW(b)l c( _ ) ()d

W(c) _ W(b) "Y. x ,. W x x
< (c - ,.)[1- Fj:nW(c) + Fj:nW(b)]-+ 0 (4.125)
and so

fbd(X - b)2W(x) dx - [fbd(X - b)w(x) dX] 2'

= a.(b) ld(X-b)W(X)dX,
88 4. Order Statistics of Independent Samples

as c \.t b. We easily see that

EFXj :n - F- 1(P)

< fad [F-1W(X) - JLFjga:.(b).8.(b)b(X) dx

Jbd(x - b)hnW(x)w(x) dx (4126)

< 1/2 uF ,
2}
.
{ Jbd(x - b)2W(x) dx - [Jbd(X _ b)w(x) dX]

which becomes the equality if

F- 1W(x) - JLF _ (x - b)I[b,d) (x) - Jbd(X - b)w(x) dx

- 1/2·
UF
{Jbd(X - b)2W(X) dx - [Jbd(X - b)w(x) dxf}

Theorem 17 (F tc W) Set

fj = fjw(b) = Ew(X - b)+ = ld (x - b)w(x) dx, (4.127)

fjw;,n (b) = Ew(Xj :n - b)+

= ld (x - b)hnW(x)w(x) dx, (4.128)

J2 = J~(b) =

=
Varw(X - b)+

I.' (x -b)w(x)dx - [I.' (x- b)w(x)dx r· (4.129)

If (j - 1)/(n - 1) ~ P E (0,1), then for b = W-1(P)

EFXj :n - F- 1(P) < fjw;,n (b)

(4.130)
UF - Jw(b) ,

and we get the equality here if

-1 AX-JL
F(x) = W(W (P) + 1] + t9-U -)I[tL- i)/J,+oo) (x). (4.131)
A

Note that

fjw(b) = 1]w(b, b),

Jw(b) = t9w(b,b),

where the right-hand sides were defined in (3.59) and (3.60).

 4.5 Bias of Quantile Estimates 89

Proposition 10 (decreasing density) For (j - l)/(n - 1) ~P E (0,1),

we have
EFXj:n - F-l(p) < 2V3 nh [1- Fj+l:n+l(P)] - p[l- Fj:n(P)]. (4.132)
O'F - (1 - p)3/2(1 + 3p)l/2
Bound (4.132) is attained by the mixture of the atom at J.' - 0'[3(1- p)/(l +
3p)]1/2 with weight p, and the uniform distribution on [J.t - 0'[3(1- p)/(l +
3p)Jl/2, J.' + 0'(1 + p){3/[(1 - p)(l + 3p)]P/2] with weight 1 - p.
Proposition 11 (decreasing failure rate) For (j - l)/(n - 1) ~ P E
(0,1), the inequality

EFXj:n - F-1(p) < 1 ~ 1- Fk:n(P) (4.133)

O'F - (1 - p2)1/2 L..J n + 1 - k
k=l

is tight and attained by the mixture of the atom at J.' - 0'[(1- p)/(l +p)Jl/2,
and the exponential distribution with scale parameter 0'/ (1- p2)1/2 starting
from the atom, with respective weights p and 1 - p.

For F t. W, the solution has a similar and even simpler form.

Lemma 20 Under the assumptions of Lemma 18 for every 9 E C~ w there
exists go:{3 E C~ w defined as
_0
_0

go:{3(x) = a(x - a)1[c.d) (x) - (3 (4.134)

with a, (3 > 0 such that
Ilgo:{3 - hll ~ IIg - hll·
PROOF. By arguments analogous to those of Lemma 18, it suffices to
confine ourselves to the family of starshaped nondecreasing functions 9
such that g(b) = -(3 for some (3 E [h(b), 0]. For all x ~ b we have
g(x) ~ -(3 ~ 0 ~ h(x).
This means that the approximation of h is improved by
g-{3 = max{g, -(3} E C{.w·
Recalling (4.120), we define (4.134) with the slope parameter

a = g(6) + (3 > O.
6-a
Then the function (4.134) is the smallest possible one for arguments x E
(b,6), and the greatest one for x E (6, d) among all starshaped functions
starting from (a, -(3) and passing through (6,g(6)). We have thus proved
(4.121) and (4.122), with (4.119) replaced by (4.134), which is precisely the
assertion of the lemma. •
90 4. Order Statistics of Independent Samples

Lemma 21 Under the assumptions of Lemma 18 and notation of Lemma

20, we have

for

t(x - a)h(x)w(x) dx
(4.135)
Jcd(x - a)2w(x) dx - [t(x - a)w(x) dX] 2'

13* = Ct* ld (x - a)w(x) dx. (4.136)

The proofis similar to that of Lemma 19. We first find the solution (4.136)
to the minimization problem of 119at3 - hW with respect to 13. Then we
look for Ct* minimizing 119at3.(a)-y - hll 2 and find (4.135), which is positive
by arguments analogous to those presented in the last part of the proof of
Lemma 19.
Taking h(x) = hn W(x) for x ~ c, and passing to the limit, we get

II 9a.t3. 11 -- Jbd(x - a)hnW(x)w(x)dx

1/2 ' (4.137)
{Jbd(X - a)2w(x) dx - [Jbd(X - a)w(x) dxf}

and the normalized function 9a.t3. /119a •.8.11 identical to (3.58). This implies
that the distribution functions which attain the bounds in Theorem 18,
and Propositions 12 and 13 coincide with those attaining the bounds of
Theorem 6, and Propositions 5 and 6, respectively.
Theorem 18 (P ~* W) Under the notation (3.60), b = W-1(P), and

TlWj:n (a, b) = Ew(Xj:n - a)I(b,d)(Xj:n)

= ld (x - a)hnW(x)w(x) dx, (4.138)

and for (j - l)/(n - 1) ::; p E (0,1) the following inequality is sharp

EFX j :n - p-1 (P) ::; B~. w(j, n,p) = Tl;i:( (a~~) . (4.139)
(J'F - wa,
Proposition 12 (decreasing density on the average) If (j -l)/(n-
1) ::; p E (0,1), then the following inequality is sharp
EFX j :n - p-1(P) -0
< Bt.u(j, n,p)
2v'3j 1 - Pj+1:n+1 (P)
= (4.140)
n +1 9u(P)
 4.6 Open Problems 91

Proposition 13 (decreasing failure rate on the average) If we have

(j - l)/(n - 1) $ p E (0,1), then the following inequality is sharp
EpXj :n - F-l(p) -0 ( )
$ Bt.v j,n,p
Up
",j l-F.:n(P) I (1 )[1 F (p)]
= [Link]=l n+l-k 8;(p)-P - j:n • (4.141)

Notations 8u(p) and 8v (p) , that appear in (4.140) and (4.141), were defined
in (3.64) and (3.67), respectively. It is worth pointing out that the upper
bias deviations in large samples are significantly different for the classes
of distributions determined by the convex and star orders. To see this,
suppose that p is fixed and take a sequence of order statistics X j : n , j =
j(n),n -+ 00, such that p ~ (j -l)/(n -1) -+ p. If F?::c W, then (4.130)
holds and, since X j :n -+ b = W-l(p) almost surely, (4.128) tends to O.
On the other hand, (4.129) does not depend on the sample size and is
positive. In consequence, the bound in (4.130) tends to 0, and so do those
of (4.132) and (4.133). If F ?::. W, then (4.138) tends to (b- a)/2 > 0, and
(3.60) remains fixed positive. Therefore the right-hand sides of all (4.139)
to (4.141) have positive limits.
Table 4.2 contains values of extreme upper deviations of order statistics
estimates Xj:n. 1 $ j < n = 20, of quantiles of order p = j In for the fam-
ilies of general (G), decreasing density (~O) and failure rate (OFR), and
decreasing density, failure rate on the average (OOA and OFRA, respec-
tively). In general populations, the estimates are more stable for the central
quantiles than for the extreme ones which confirms the theoretical analysis
of bias oscillation (cf. (4.114». Otherwise the deviations are increasing in j.
It follows from the fact that all the families are more concentrated on the
left, and the upper quantiles are more dispersed and thus more difficult to
be estimated.

4.6 Open Problems

1. Natural questions are the mean-variance bounds on expectations of
single order statistics for populations determined by the convex or-
der. In particular, what are the evaluations for distributions with
monotone density and failure rate?
2. General second moment bounds are presented in (4.18), (4.14), and
(4.15). Analogous results for distributions with decreasing density and
failure rate can be found in Section 4.2, but the problem is unsolved
when either density of failure rate is increasing. More generally, we
ask for the second moment bounds for families of life distributions
determined by the convex order. Likewise, what are the respective
bounds for distributions determined by the star order?
92 4. Order Statistics of Independent Samples

TABLE 4.2. Sharp uniform variance bounds on upper bias deviations of esti-
mators Xj:20 of pth quantiles, p = j/20, for various families of distributions
(independent case).

j G DD DFR DDA DFRA

1 1.64485 0.05658 0.01795 0.11742 0.03627
2 1.30582 0.08034 0.02683 0.20820 0.06764
3 1.13394 0.09763 0.03406 0.28585 0.09838
4 1.02862 0.11226 0.04069 0.35251 0.12916
5 0.95804 0.12575 0.04718 0.40974 0.16022
6 0.90860 0.13894 0.05379 0.45924 0.19170
7 0.87348 0.15240 0.06075 0.50272 0.22379
8 0.84894 0.16660 0.06828 0.54180 0.25671
9 0.83279 0.18203 0.07660 0.57798 0.29076
10 0.82380 0.19919 0.08603 0.61264 0.32638
11 0.82139 0.21876 0.09694 0.64709 0.36418
12 0.82548 0.24158 0.10989 0.68269 0.40502
13 0.83648 0.26888 0.12569 0.72092 0.45011
14 0.85539 0.30247 0.14557 0.76365 0.50125
15 0.88410 0.34523 0.17163 0.81339 0.56122
16 0.92588 0.40204 0.20756 0.87393 0.63458
17 0.98657 0.48192 0.26077 0.95145 0.72938
18 1.07691 0.60355 0.34835 1.05688 0.86128
19 1.21205 0.80859 0.51815 1.20493 1.06011
 4.6 Open Problems 93

3. Differences of order statistics, especially the sample range Xn:n -Xl:n

and quasiranges Xj:n - XnH-j:n, (n + 1)/2 < j < n, are used as
simple and robust estimates of dispersion of the parent population.
Spacings Xj:n - Xj-l:n, 2 $ j $ n, represent times between con-
secutive failure times of elements of composite systems. In general
populations, expectations of sample range and quasiranges were pre-
cisely evaluated by Plackett [72] and Moriguti [58], respectively. The
former was derived by direct application of the Schwarz inequality.
For the latter, the greatest convex minorant construction was used.
In both cases, the bounds are twice as great as those on the jth order
statistic in the symmetric populations (see (4.21) and (4.24) through
(4.26)). It is important to find evaluations for differences of order
statistics, especially the quasiranges and spacings, in the restricted
families of distributions considered here.
4. Completely untouched domains of potential interest are bounds on
most popular L-statistics: trimmed means, some best linear unbiased
estimates (so called BLUEs) in parametric models, the Gini mean
difference 2/[n(n -1)] E7=1 (2j - n -l)Xj:n, and many others, both
in general and restricted families of distributions.
5. The results of Section 4.5 should be completed by analyzing the bias
of quantile estimates for dual families of distributions defined by rela-
tions F ~c W and F ~* W. Evaluations of the bias in the symmetric
populations, with possible additional restrictions, are not known, ei-
ther. Also, some L-statistics can be studied in a further perspective.
We point out here that in the problems of bias evaluation, the lower
bounds are of key interest as well. On the other hand, the upper and
lower bounds on EFXj:n - EFXkj:kn for various k > 1 would provide
precise evaluations of rates of convergence of quantile estimates over
different classes of distributions.
6. The order statistic X j:n is used for estimating F-1(p) if jln is close
to p. This is justified by strong consistency of the estimates of the
unique quantiles for which j In -+ p is only needed. H n is fixed and
jln < p < (j + l)ln, then the (j + l)st order statistic is a possibly
better candidate. Evaluating EF X j : n - F-l (p) for various j, and given
n and p, we can find the order statistic whose bias is minimal in
a given class of distributions, and recommend it as the most bias-
robust estimate of the pth quantile. Therefore it is worth extending
results of Section 4.5 to the cases p =j:. j In.
5
Order Statistics of Dependent
Observations

Assume that Y1 , .•. , Y n are possibly dependent and identically distributed.

Recalling arguments of Rychlik [79), in Section 5.1 we conclude sharp
bounds (2.25) and (2.27) on expectations of general L-statistics and single
order statistics, respectively, depending on the common marginal distri-
bution of the observations. Next we apply the projections of functionals
defined in (2.25) and (2.27) for establishing respective moment bounds
over general and restricted families of marginals. In Section 5.2 general,
symmetric, and nonnegative observations are treated, and respective de-
terministic bounds for arbitrary samples are concluded. The results cited
are from Rychlik [81), but some earlier partial solutions are also mentioned.
In the remainder of Chapter 5 we confine ourselves to single order statis-
tics. In Section 5.3 we present mean-variance and second moment bounds
on the expectations of order statistics for families of parent distributions
related to a given one in the convex order. The mean-variance bounds for
F ~c W and F jc W are not published elsewhere, except for DFR and
IFR distributions, given in Rychlik [87). The results for general W and
W = U (i.e., for the decreasing density distributions) are presented here
as well. Rychlik [84] established second moment bounds for F ~c (jc)W
with general W. The decreasing density and failure rate distributions were
studied in Gajek and Rychlik [32). Analogous results for increasing ones
come from Rychlik [84]. Section 5.4 deals with mean-variance bounds for
order statistics based on samples with common marginal distributions be-
ing in a star relation with a fixed one W. For ones determined by the
exponential W = V, which includes important classes of life distributions
with monotone failure rate, respective bounds were established in Rych-
96 5. Order Statistics of Dependent Observations

lik [87]. Especially, it was shown that general bounds (5.13) are attained
by the IFRA distributions. In fact, the claim can be extended to families
of distributions defined by F j. W for general W. We also write explicitly
the bounds for F ~. U that have decreasing densities on the average. The
mean-variance bounds for order statistics with symmetric unimodal and
U-shaped distributions of parent variables, described in Section 5.5, were
obtained in Gajek and Rychlik [32] and Rychlik [84], respectively. We call
a symmetric distribution U-shaped if it has nonincreasing and nondecreas-
ing density on the lower and upper halves of its support. Bias of quantile
estimation in dependent samples is studied in Section 5.6. Section 5.7 deals
with the extreme deviation of expected order statistics under violating the
independence assumption.
It is worth pointing out that Papadatos [71] established sharp uniform
bounds on distribution functions of order statistics under the assumption
that the maxima of the subsamples of a given size have a specified common
distribution. The special cases were the independent and arbitrarily depen-
dent identically distributed samples discussed in this book. Balakrishnan
et al. [14] presented the best possible mean-variance bounds on the ex-
pectations of order statistics from the samples taken without replacement
from finite populations. As the size of a population increases to infinity, the
bounds reduce to those of the Li.d. samples from arbitrary populations (cf.
Section 4.1). On the other hand, in the case of exhaustive drawing with-
out replacement, these results coincide with ones for arbitrarily dependent
samples with arbitrary common marginal distributions, presented in Sec-
tion 5.2, which are actually attained by exhaustive sampling models. For
some quantile bounds on order statistics of dependent observations and
respective L-statistics, we refer the reader to Rychlik [84, Section 5].

5.1 Dependent Observations with Given Marginal

Distribution
Suppose that arbitrarily dependent Y1 , ... , Yn have a fixed common distri-
bution function F with a finite mean JLF. We aim to justify bound (2.25)
for the expectation of an arbitrary combination of order statistics under all
possible interdependencies of observations and to specify the conditions of
its attainability. For the formal proof we refer the reader to Rychlik [79]
(see also Rychlik [84]).
First we characterize vectors (G 1: n , ... , Gn :n ) of all possible distribution
functions of order statistics Yi.:n,' .. , Yn :n by relations
n
:2: G j :n = nF, (5.1)
j=l

G1:n ~ G2:n ~ ... ~ Gn :n . (5.2)

 5.1 Dependent Observations with Given Marginal Distribution 97

The former immediately follows from

n n

L l(-oo,z) (Yj:n) =L l(-oo,z) (Yj), x E In, (5.3)

j=l j=l

by taking expectations of both sides of (5.3). The latter is a consequence

of relations Y1:n $ Y 2 :n $ ... $ Y n:n . There are many ways of construct-
ing ordered variables Yj:n, 1 $ j $ n, with distributions satisfying (5.1)
and (5.2) (the simplest one consists in taking Gi~(X), 1 $ j $ n, for X
being a standard uniform variable). There are also many ways of construct-
ing random variables Yj, 1 $ j $ n, with identical marginal distribution
function F, whose order statistics have given distribution functions satis-
fying (5.1) and (5.2) (the simplest one consists in random rearranging Yj:n,
1 $ j $ n).
Now we solve the problem of minimizing E~=l CjGj:n(X) for fixed co-
efficients c = (C1, ... , en) of the L-statistic under study and distribution
functions satisfying (5.1) and (5.2) valued at arbitrary point x. This is
a linear programming problem that has the solution
n
min L cjGj:n(x) = GeF(x) (5.4)
j=l

for Ge being the greatest convex function on the unit interval satisfying
(2.26). Recalling Lemma 1, we obtain

Ep LCjYj:n
j=l
= (00 y (t. C;G',.(dy»)
< 1+
-00
00
y GeF(dy)

= 11 F- 1(x) Ge(dx)

= 11 F-1(x)ge(X) dx

= (F-1,ge), (5.5)

which is the desired conclusion.

Analyzing values Gj:n(X), 1 $ j $ n, x E In, for the solutions to (5.4),
we are able to determine supports and some mutual relations of respective
order statistics. They depend on properties of function Ge , whose graph
is a (broken) line with breaks (if any) at some multiplicities of lIn. Let
o = ko < k1 < ... < k" = n, 1 $ K, $ n, be the sequence of integers such
that each kiln, 1 $ i $ K, - 1, is a breakpoint of Ge . Let 0 = 10 < It <
98 5. Order Statistics of Dependent Observations

... < I).. = n, K, ~ >. ~ n, be the sequence of integers for which

n
GcU In) = 2: Ck
k=1

holds. We have

{O,n} C K={k i : I~i~K,}

C C = {li: 1 ~ i ~ >.}
c {O, ... ,n}.
Then equality in (5.4) for all x E ~ and some Gj :n satisfying (5.1) and
(5.2) implies

P(F- 1(k i _dn) ~ Yl j_1 +1:n = Ylj:n ~ F-l(kdn)) = 1 (5.6)

for all 1 ~ i ~ K" and 1 ~ j ~ >. such that ki - 1 ~ lj-l < lj ~ ki. Relation
C = K uniquely determines the distributions of all order statistics that
solve (5.4) and attain equality in (5.5). Then

G-.3: n (X ) -_ nF(x) - k i - 1
,
ki - ki- 1
for ki - 1 < j ~ ki . If C "I K, there are also other solutions. For instance,
for the sample mean we have

Gc(x) = x,
K = {O,n},
C = {O, ... ,n},
and (5.6) simply means that each }j:n should belong to the domain of Y1 .
Indeed,
1 n
Ep- 2:}j:n =
11
F-1(x)dx = [Link]
n j=l 0

for any type of dependence in variables.

In the special case of single order statistics }j:n, functions Gc and gc
have the forms

n+~-j (x_ j : 1 )+,

n
n + 1 -). l[(i-l)/n,l) (x),
respectively, which implies (2.27). Since here

K = {O,j -I,n} C C = {O,I, ... ,j -I,n},

 5.1 Dependent Observations with Given Marginal Distribution 99

relation
P{Yj-l:n ~ F-1{{j - l)jn) ~ Yj:n = Yn :n ) = 1 (5.7)
combined with (5.1) and (5.2) are respective conditions for equality. The
stochastically largest (Le., uniformly smallest) distribution function

G .. F{ ) = [nF{x) + 1 - j]+
,.n X n + 1 -3.
of the jth order statistic is uniquely determined. There are various ways of
constructing dependent samples with the same marginal distribution and
the extreme distribution of Yj:n. The simplest one is a random rearrange-

e: X)
ment of

Y F-
e: 1)
1:n = ... = Yj-l:n = 1 1

< F-
e:
1

< F- 1 1 + [1 - j : 1] X)
= Yj:n = ... = Yn :n
for some X uniformly distributed on [0,1]. For the sample minimum, (2.27)
and (5.7) imply the trivial claims EFY1:n ~ J.l.F, becoming the equality
for identical observations Y1 :n = Yn :n = Y1 • Condition (5.7) excludes the
possibility of constructing an absolutely continuous joint distribution of the
sample with the stochastically maximal jth order statistic except for the
sample maximum.
In the first paper in this field of research Mallows [53] constructed a den-
sity function of the sample with identical uniform marginals that provide
the maximal expectation of the sample maximum. Lai and Robbins [47]
extended the construction to arbitrary possibly nonidentical marginal dis-
tributions. Lai and Robbins [48] and Tchen [98] constructed infinite se-
quences of variables with identical and arbitrary distributions, respectively,
such that all sample maxima are stochastically maximal. Bounds (2.27) for
general order statistics of identically distributed samples were proved in-
dependently in Caraux and Gascuel [19] and Rychlik [76]. In the former,
some inequalities for nonidentically distributed observations were presented
(with conditions for equality established in Rychlik [82]). In the latter, the
problem of constructing sequences with stochastically extreme order statis-
tics was also discussed. Asymptotic properties of sequences of stochastically
extreme maxima and other order statistics were studied in Lai and Rob-
bins [48] and Rychlik [77], respectively.
100 5. Order Statistics of Dependent Observations

In the rest of this chapter we present tight bounds for expected order and
L-statistics of dependent samples for various families of marginal distribu-
tions. We also indicate the marginals for which the bounds are attained.
Formally, in each case one should say that these are attained by the joint
distributions specified for arbitrary marginals and fixed L-statistics by (5.1)
and (5.2) with (5.6) (replaced by (5.7) for single order statistics in partic-
ular), and the given marginal. However, we drop repeating the reference to
the construction of the joint probability and confine ourselves to describing
the optimal marginal.

5.2 General and Symmetric Distributions

We first study the optimal bounds for general L-statistics based on de-
pendent samples with arbitrary common marginal distribution F. For an
arbitrarily fixed sequence of coefficients c = (C1' ... , cn ), we construct func-
tion gc that defines the functional of the expectation of the respective L-
statistic (see (2.25) and (5.5)) by differentiating function Gc determined
by (2.26). Using the notation of Section 5.1, we have
n n

gc(x) = Ldjl[(j-1)/n,j/n)(x) = Ldki1[ki_l/n,k;/n)(X) (5.8)

j=1 i=1
for a nondecreasing sequence dj , 1 $ j $ n, with the specified increasing
subsequence dki' 1 $ i $ /'i, $ n, of distinct values. Precisely, we have

(5.9)

_ 1 n n r1
d=-Ldj =LCj=Gc(l)= 10 gc(x)dx. (5.10)
n j=1 j=1 0

We claim that the optimal mean-variance bound for the expectation of

a given L-statistic amounts to the Euclidean norm of the respective vector
(dj - d:)/n, 1 $ j $ n.
Theorem 19 (general distributions) Under notation (5.8) to (5.10),
with G c defined in (2.26), we have
1/2
[
~t~-iP
3=1
]
(5.11)
 5.2 General and Symmetric Distributions 101

If C = 0, then (5.11) becomes the equality for degenerate distributions.

Otherwise the equality holds for the ,,-point marginal distribution

P dk. - ii)
(y, =J.L+U-- ki - ki-1
1 = , (5.12)
C n

PROOF. Verification of (5.11)

n

Ep L Cj(}j:n - 1') = 101F- 1(X)[ge (X) - d] dx

j=1

= 101[F-1(X) - J.L][ge(X) - d] dx

< {10 1[F-1(X) _ J.L]2dx 101[ge(X) _ d] 2dX} 1/2

= Cu

ii
is based on (2.25), (5.10), (2.10), and (5.8). The Schwarz inequality provides
the sharp bound here, because Be - is nondecreasing, and so

with a = u / C > 0 actually defines a quantile function with the desired

mean and variance. An elementary algebra enables us to derive the respec-
tive distribution (5.12) . •
Especially, for the single order statistics we have
Corollary 1 (general distributions) Inequality

Ep}j:[Link] <C=CO(' n) =( j - l .)1/2 (5.13)

Up - J, n + 1- J
is sharp and becomes the equality for the two-point marginal

( U) j-l
P Y1=J.L- C =-;;:-=I-P(Yi=J.L+uC). (5.14)

Bounds (5.13) for the sample maximum and general j were presented in
Arnold [2] and Gascuel and Caraux [34], respectively. We proceed now to
present analogous results for symmetrically distributed random variables.
Below in Theorem 20 we omit the formal presentation of the distribution
that attains the bound, because this needs introducing a rather complicated
notation.
102 5. Order Statistics of Dependent Observations

Theorem 20 (symmetric distributions) Inequality

n
EF "lj·[Link]
L.J Cj· ~ C 8( c ) =
[
1"
-2
nL.J (dj-dnH- j )2
]1/2
, (5.15)
.
3=1 (TF n j= L( n+3)/2 J
where L·J denotes the floor of a number, is tight. Equality is attained by
a unique (up to location-scale tmnsformations) marginal symmetric distri-
bution supported on n points at most whose probabilities are multiplicities
of l/n.
PROOF. Inequality (5.15) follows from

EF
j=l
n
2: Cj(lj;n - J.L) = I 1/2
1
[F-1(X) - [Link]~(x) dx (5.16)

with
gc(x) - gc(1- x-)
g~(x) = 2
n
= "L.J dj - d2nH - j 1[(j-1)/n,j/n) ( x ) , (5.17)
j=L(n+3)/2J

1/2 ~ x < 1, and the Schwarz inequality. Note that (5.17) is a nonnegative
non decreasing piecewise constant function with L(n + 1)/2J values at most.
Thus its antisymmetric extension onto [0,1/2) is also a nondecreasing step
function with n values at most and jumps at some points j In, 1 ~ j ~
n - 1. The extension coincides with the quantile function of the extreme
distribution providing equality in (5.15) up to an affine transformation.
This justifies the latter claim of Theorem 20.•

Although the algorithm of determining gc and g~ is simple, we are not able

to write explicitly respective formulae for general linear combinations of
order statistics. For the single order statistics, however, we obtain
gj,n(x) = n +n1 -J.1[max{(j-1)/n,1-(j-1)/n},1) (x), 1/2 ~ x < 1. (5.18)
By an easy computation, we conclude
Corollary 2 (symmetric distributions) Inequality
_EF_Y-=.j_:n_-...:..J.L_F
(TF
<
-
C= [2(n + n
1 - J)
. { j -
. mm n +1- J
., II}] 1/2
(5.19)

is sharp and becomes the equality for the three-point marginal distribution

P(Y1=J.L) = 2Ii-1_!1
n 2'
(5.20)

(5.21)
 5.2 General and Symmetric Distributions 103

Inequality (5.19) and its special case for j = n can be found in Gascuel
and Caraux [34] and Arnold [2], respectively. Here we merely mention the
second moment bounds for general L-statistics

EF [Link]:n < 11 p-1(X)[gc(x)]+ dx

< II(gc)+llmF

[~ t,(dj)~] 'l'mF (5.22)

and for single order statistics

EFlj:n ~ ( ~
n+ -)
.)1/2 mF (5.23)

of nonnegative samples. Bounds (5.22) and (5.23) are attained by

P(Y1 = 0) = kio (5.24)

n
ki - ki-1
P (Y1 = mF 11(::)+11) =
n
io <i ~ K, (5.25)

with io = max{O ~ i ~ K: dki ~ O}, and

P(Y1 = 0) = j : 1 = 1- P (Y1 = [n + ~ _ j f/2 mF) , (5.26)

respectively. If dn ~ 0 and so II (gc)+ II = 0, then (5.24) and (5.25) are clearly

replaced by the Dirac measure at 0. In fact, (5.24) with (5.25), and (5.26)
are not unique solutions. The second relation in (5.22) shows that p-1 may
have various forms on [0, kio/n) provided that nonnegativity, nondecrease,
and moment conditions are not violated (see Rychlik [81] for more details).
Rychlik [81] (and Arnold [3] for the cases of sample maximum Yn :n and
range Yn:n - Y1 :n ) presented more general sharp bounds in terms of cen-
tral absolute moments of various orders based on the HOlder inequality
instead of the Schwarz one. These are also attainable by discrete marginal
distributions with probabilities kdn for some integer ki. This form of
solution allows us to deduce analogous inequalities for deterministic se-
quences. Randomly rearranging a sequence of (not necessarily distinct)
numbers Y1, ... ,Yn, we obtain a random sequence of dependent identically
distributed random variables with expectation, second raw moment, vari-
ance
1 n
fi = - LYj,
n j=1
104 5. Order Statistics of Dependent Observations

respectively, and deterministic order statistics Y1:n $ ... $ Yn:n' Using

(5.11), (5.15), and (5.22), we conclude optimal bounds

t
j=1
Cj (Yj:n

S
- y) < [
~tcP;-cP
3=1
]
1/2

, (5.27)

t Cj{Yj:n - y) < [
2~ t (dj - dn+1_j)2
]
1/2

, (5.28)
j=1 S j=L(n+3)/2J
1/2
Ln
CjYj:n < [
~ t,{dj)~ ]
, (5.29)
j=1 m

for general, symmetric, and nonnegative sequences of numbers, respectively.

Sharpness of (5.27) is verified by putting Yj = dj , 1 $ j $ n, and with
reference to (5.12).
The problem of best deterministic bounds for L-statistics in terms of var-
ious sample parameters has a long history. We confine ourselves to those de-
rived by means of the Schwarz inequality. Samuelson [92] raised and solved
the problem of how much a single observation can deviate from the sam-
ple mean in the standard deviation units. Samuelson's paper stimulated
intensive investigations of the problem and its modifications: alternative
proofs, rediscoveries of earlier results, and extensions. Six different proofs
were reviewed by Arnold and Balakrishnan [5]. The earliest proofs found
in the literature were due to Thompson [99] and Scott [94]. We do not at-
tempt to present a complete record of consecutive contributions, referring
the reader to Arnold [4] for a comprehensive bibliography, and Olkin [66] for
a recent review, with yet another proof. Scott [94] established the bound
for deviations of Yn-1:n from the mean in the standard deviation units.
The bounds for arbitrary order statistics follow directly from Mallows and
Richter [54], and were explicitly stated by Boyd [18] and Hawkins [39]. Mal-
lows and Richter [54] established the inequalities for selection differentials
2:1=1 Yi:n/j, and 2::=n+1-k Yi:n/ k, and their differences. The respective
results for Yn:n - Y1:n, Yn-1:n - Y1:n, and the differences of arbitrary or-
der statistics were derived by Nair [63], David et al. [24], and Fahmy and
Proschan [27] (implicitly in Arnold and Groeneveld [6]), respectively, and
for the L-statistics with nondecreasing coefficients by David [23]. Bounds
(5.27) through (5.29) for general L-statistics come from Rychlik [81].
 5.3 Distributions with Monotone Density and Failure Rate 105

5.3 Distributions with Monotone Density and

Failure Rate
Determining mean-variance bounds for F tc W by the projection method,
we look for the element of C(
_c
w least distant from the function
n
h(x) = n +1_ j I[U-l)/n,l) W(x)
n
= +
n 1 -J.1[W- ((j-l)/n),dw)(X).
1

Applying Lemmas 4 and 5 with b = W-1(U -1)/n) and c = dw, we obtain

the solution. Note that there are no trivial zero projections and bounds for
b > aw, because
EwX < [Link](b) = Ew(XIX > b)
contradicts (3.8). Hence the projection is linear increasing with a possible
constant left part, and the condition for distinguishing the cases is given in
the last part of Lemma 5.
Theorem 21 (F tc W) If for b = W-1«j - 1)/n) we have
(J'2
[Link](b) :::; [Link] + w , (5.30)
[Link]-aw
then inequality
EF}j:n - [Link] < [Link](b) - [Link]
(5.31)
(J'F - (J'w
holds true and becomes the equality for

F(x) =W ( [Link] X - f.L)

+ (J'w-(J'- . (5.32)

Otherwise there exists a unique (3* = (3*(j,n) E (aw,b) satisfying equa-

tion
(J'~({3) = [[Link](b) - [Link] ({3)][[Link] ({3) - (31 (5.33)
such that
(5.34)

for

iJw({3) = Ew(X - (3)+ = i d

(x - (3)w(x) dx, (5.35)

J~({3) =

=
Varw(X - (3)+

I.' (x - P)'w(x) dx - [I.' (x - P)w(x) dx r (5.36)

106 5. Order Statistics of Dependent Observations

(cf. (4.127) and (4.129)). Bound (5.34) is attained by

X - JL)
F(x) =W (
[3* + T}W([3*) + t1w([3*)-u-
A A

1 [/J-"'1iw (,B.)/t1w(,B.),oo) (x).

(5.37)
Note that sup EwYj:n = JLw(b) which confirms the first statement of The-
orem 21. Condition (5.30) is certainly satisfied by small order statistics.
Definitions (5.35) and (5.36) are analogues of (3.59) and (3.60), respec-
tively, valued at ([3, [3). Distribution function (5.37) has a jump of height
W([3) < (j - 1)/n at JL - ufJ/-D and a density right to the point.
Proposition 14 (decreasing density) If (j - 1)/n ~ 1/3, then

(5.38)

which becomes the equality for F being the uniform distribution function
on fJ.t - V3u, JL + V3u].

n)
Otherwise
EFYj:n - JLF < ~ (9 j - 9 - 1/2 (5.39)
UF - 3 n+ 1- j ,
and the equality holds for the mixture of the Dirac distribution at JL -
3u [en + 1 - j)/(9j - 9 - n)]1/2 and the uniform distribution on

[JL -
3 (n + 1 - j )
u 9j _ 9 _ n
1/2 3j - 3 + n
,JL + U [en + 1 - j)(9j - 9 - n)]1/2
1
with respective probabilities 3(j - 1)/(2n) - 1/2 and 3(n + 1 - j)/(2n).
Proposition 15 (decreasing failure rate) If (j - 1)/n ~ 1 - e- 1 ~
0.63212, then
EFYj:n - JLF < I n
(5.40)
UF -
n
n+ 1 -J.,
and the equality holds in (5.40) for the exponential distribution with location
JL - U and scale u.
Otherwise for
'Y ='Yv(j, n) = ( 1 - j -1)
-n- e,
we have

EFYj:n - JLF
UF
< C =d
- toV
(. ) = [
J,n
2 -1]
'Yv(j,n)
1/2
, (5.41)

which becomes the equality for

(5.42)
 5.3 Distributions with Monotone Density and Failure Rate 107

Note that (5.42) is the mixture of the exponential distribution with location
J.t - ulO and scale ul(-YO), with probability;, and the Dirac measure at
J.t - u I 0 with probability 1 - ;. This is a D FR life distribution if J.t = u I 0 .
For small order statistics (5.40) is attained by a life exponential distribution
with J.t = u.
JOw we consider analogous bounds for F ~c W. First we try to describe
P-jcwh for
n
h(x) = n+1_ j l[W-1((j-l)/n),dw) (x)

1
= 1 - W(b) l[b,d)(X).

Results of Lemmas 6 and 7 may be applied here. However, setting c = dw

also results in reduction of the number of parameters of functions that
are the candidates for being projections. In contrast to the conclusion of
Lemma 6, the level at which the linear functions are possibly broken is
specified and coincides with the level of the original h.
Lemma 22 If h(x) = Ml[b,d)(X), then for every 9 E C~w there exists
ga/3 E C~
_c
w defined as -

ga/3(x) = M + amin{x -,8, o}

for some a ~ ° and ,8 ~ b such that

Ilga/3 - hll ~ Ilg - hll·

Lemma 23 Under the assumptions of Lemma 22 and (3.7), we have

p{wh(x)= d 1 +[Link]{x-,8.,O}, (5.43)

_c fb w(x) dx
where ,8. > b is the unique solution to the equation

,8 [l b w(x) dx 111 xw(x) dx - 1b xw(x) dx 111 w(x) dx1

= 1b w(x)dx 111 x w(x)dx 2 -l b
xw(x)dx 111 xw(x)dx, (5.44)

with i3 = min{,8, d}, and

_ _
a. - a.(,8.) - d d}
f: {/3
.
w(x) dx
• (5.45)
fb w(x) dx famln ., (,8. - x)w(x) dx
108 5. Order Statistics of Dependent Observations

Precisely, if

d~
,Q
= a
t X2W(X) dx - taxw(x)bdx I: xw(x) dx '
t w(x) dxbad (5.46)
Ia w(x) dx Ia xw(x) dx - Ia xw(x) dx
fJ

(Le., the left-hand side of (5.44) is not greater than the right-hand side for
(3 = d), then (3. = /J, and (5.43) is linear. This may happen if d is finite
and b is close to d. A probabilistic interpretation of Lemma 23 is given in
Theorem 22. For a < (3 < d, we define

fJ.w«(3) = Ew(XIX < (3) = It xw(x) dx , (5.47)

It w(x) dx
q~«(3) = Varw(XIX < (3)
It x 2w(x)dx - fJ.~«(3)
= (5.48)
It w(x) dx
= i/w«(3) =
i/ Ew«(3 - X)+ = 1 13
«(3 - x)w(x) dx, (5.49)

· [· r
192 = 19~ «(3) = Varw«(3 - X)+

= fa «(3-x)2w(x)dx - fa «(3-x)w(x)dx (5.50)

(cf. (3.20) and (3.21) as well as (5.35) and (5.36».

Theorem 22 (F ~c W) If for b = W- 1 (U - l)jn)

m~ - fJ.w(b)[Link] <d (5.51)

[Link] - fJ.w(b)
holds, then there exists a unique (3. E (b, d) satisfying the equation

q~«(3) = [(J.w«(3) - fJ.w(b)][{3 - fJ.w«(3)] (5.52)

such that

EFYj:n - [Link] < C = ~-<.w},

(. n) = +j -1 1 . 19w«(3.) (5.53)
O'F - - n - } 71w (R.) v
fJ.

(cf. (5.33) and (5.34»). The equality in (5.53) holds if

F(x) = { W ((3. - i/«(3.) + 19«(3.) ~) if ~ < ;f~:~' (5.54)

1 ;1 :r:-e > 13.
, IJ IT - "(13.)·

OthenlJise we have (5.31) which becomes the equality for (5.32).

 5.3 Distributions with Monotone Density and Failure Rate 109

Distribution function (5.54) jumps from W(.8*) > (j -1)/n to 1 at its right
endpoint. H W = U, both cases are possible. Since dv = +00, the first one
is only applicable for the IFR distributions.
Proposition 16 (increasing density) If (j - 1)/n < 2/3, then
EFlj:n - JLF < [(j - 1)(8n - 9j + 9)]1/2
(5.55)
UF - 3(n + 1 - j) ,

which becomes the equality for the combination of the uniform distribution
on
4n - 3j + 3 [ 9j - 9 ] 1/2]
[ JL - U [(j _ 1)(8n _ 9j + 9)]1/2' JL + U 8n - 9j + 9 '

and an atom at the right end of the interval, with respective coefficients
3(j -1)/(2n) and 1- 3(j -1)/(2n).
If (j - 1)/n ~ 2/3, then (5.38) holds with the equality for the uniform
distribution on lJL - J3u, JL + J3u].
Proposition 17 (increasing failure rate) For {3* > In[n/(n + 1 - j)]
defined by the equation

1 - (1 + {3)e-.8 n +1- j n
ln (5.56)
e- - +
.81{3= ·
3- l 1·
n+-3
and
(5.57)
we have
EFlj:n - JLF < j - 1 (Jv({3*) (5.58)
UF - n + 1- j e-.8· -1 + {3*.
The equality in (5.58) holds if

0,
F(x)= { l-exp(-I+e-.8·-(Jz~,,),
1,
(5.59)
This is the exponential distribution with location JL - u(1 - e- fJ .) / (J and
scale U /(J, right truncated at JL + u(e-.8· -1 + {3*)/8. The jump probability
is e- fJ ••

Likewise, for determining respective second moment bounds for life dis-
tributions F tc W and F ~c W, we need to find best approximations
of
n
h(x) = n + 1 _ j 1[(i-1)/n,1) W(x)
110 5. Order Statistics of Dependent Observations

by nondecreasing convex and concave functions that additionally obey the

condition of vanishing at aw. We assume here that W is a life distribution
and aw = O. The respective projections have similar parametric forms

P:;-" wh(x) = o:(x - (3)+,

~ .min{~,1}.
n+ -) 0:

In the first case, under the conditions of the latter statement of Theorem 21,
the optimal parameter 13 defined in (5.33) is positive which implies that
PSwh E C~"w provides the solution to our problem. Then we have

(cf. (5.34) to (5.36)). Otherwise P:;

wh is the linear function crossing the
origin, with the optimal slope and-~orm

Ibd xw(x) dx JLw(b)

0: .. (0) - = --2-' (5.61)
Ib w(x) dx fo x 2 w(x) dx mw
d d

fbdxw(x) dx JLw(b)
(5.62)
= mw '
d
fb w(x) dx
[
fod
x 2w(x) dx
] 1/2

respectively.
Theorem 23 (F !:c W) If for b = W-1«j -1)jn)

(5.63)

holds true (ef. (5.30)), then

EFYj:n JLw(b)
--<
- - -, (5.64)
mF mw

(ef. (5.62)), whieh is attained for

F(x) =W (m :"F)' w

Otherwise for 13.. E (0, b) determined by (5.33) with

y2 = y~(13.. ) = Ew[(X - 13.. )+]2 = ld

/3.
(x - 13.. )2 w (x) dx (5.65)
 5.3 Distributions with Monotone Density and Failure Rate 111

we have
EF}j:n < VW({3.)
(5.66)
mF - J.,tW({3.) - (3. '
which becomes the equality for

F(x) = W ({3. + v~) 1[0,00) (x). (5.67)

Again, we point out analogies between Theorems 21 and 23. Under equiv-
alent conditions (5.30) and (5.63), we obtain analogous bounds (5.31) and
(5.64), respectively. In the opposite case, the bounds are determined by
the same parameter {3. and related by (5.60). Moreover, both the extreme
distributions (5.37) and (5.67) are location-scale modifications of W with
identical jump W ({3.) at the left support end.
Proposition 18 (decreasing density) If (j - l)/n ~ 1/3, then

EF}j:n
mF
~ J3
2
(1 + 1) ,
j -
n
(5.68)

where the equality holds if F is the uniform distribution function on the

interval [0, J3mF].
Otherwise
1/2
EF}j:n < ~ ( 2n )
(5.69)
mF - 3 n+ 1-j ,
which becomes the equality for the mixture of the atom at 0 with prob-
ability 3(j - 1)/(2n) - 1/2 and the uniform distribution on the interval
[0, mF [2n/(n + 1 - j)]1/2] with probability 3(n + 1 - j)/(2n).

Proposition 19 (decreasing failure rate) If (j - l)/n < 1 - e- 1 ~

0.63212, then
EF}j:n < ...!... (In n . + 1) , (5.70)
mF - J3 n+ 1-)
which is the equality for the exponential life distribution with scale mF/V2.
Otherwise
1/2
EF}j:n < [ 2n ]
(5.71)
mF - (n + 1 - j)e
The bound becomes the equality for the mixture of the exponential distribu-
tion with scale mF {n/[2e(n + 1 - j)]}1/2 and zero, with weights [1 - (j -
l)/n]e and 1- [1- (j -l)/n]e, respectively.

Projection of h = Ml[b,d) for M = n/(n + 1 - j), 0 = aw < b =

W- 1 ((j - l)/n) < dw onto ctw, desired for calculating second moment
bounds for samples with life distributions satisfying F ~c W, is described
in Lemma 24 (cf. Rychlik [84, Theorem U(b), p. 131]).
112 5. Order Statistics of Dependent Observations

Lemma 24 Projection of h = Ml[b,d) onto ctw in L2([0, dw), w(x) dx)

has the form -
ptwh(x) = Mmin{xja*, I}. (5.72)

ld ld
If
d xw(x) dx > x 2w(x) dx, (5.73)

then a* E (b, d) is the unique solution to

a lOt xw(x) dx = lOt x 2w(x) dx. (5.74)

Otherwise
(5.75)

Under (5.73) the projection is actually a broken line with a break at a*

that belongs to the domain of h. Then

IIP~cwh112 = M2 [~~ lOt> x 2w(x)dx+ i~ W(X)dX]

= Ot
M2 [ ~* l xw(x)dx + i~ W(X)dX]
< (5.76)

(5.77)
say, and
ptwh(x) 1 . {x }
(5.78)
IlPtwhll = :y mm a*' 1 .
If (5.73) is not true (which is possible for dw < 00 only), then the projection
is actually linear and satisfies

(5.79)

ptwh(x) x
(5.80)
IIP~whll
_c

Theorem 24 (F ~c W) If
2
n mw d (5.81)
n + 1 - j JLw(b) < w
for b = W- 1 «j - l)jn), then there exists a unique b < a* < dw solving
W(a)tLw(a)[a - tLw(a)J = W(a)a~(a) + [(j - l)jnJatLw(b) (5.82)
 5.3 Distributions with Monotone Density and Failure Rate 113

(cf. (5.47) and (5.48») such that for

= ;;lv(a.) =
ld
;;2 Ew(min{X,a.})2

= lo
a•
x 2w(x) dx + a~
a.
w(x) dx (5.83)

we have
EFYj:n < n ;;w(a.)
(5.84)
mF - n+ 1-j a.
The equality in (5.84) holds for

....L<~
mF D, (5.85)
....L>~
mF - D'

If (5.81) does not hold, then bound

EFYj:n < JLw(b)

(5.86)
mF - mw
is attained by F(x) = W(mwx/mF)'
Relation (5.81) is always true for W supported on the whole positive half-
axis. Otherwise this is satisfied for small order statistics. Life distribution
(5.85) has a finite support with probability mass 1 - W(a.) at the right
endpoint. Below we specify results for distributions with increasing density
and failure rate.
Proposition 20 (increasing density) If (j - 1)/n < 1/../3 R;j 0.57735,
then
EFYj:n < n (1_~j-1)1/2 (5.87)
mF - n + 1 - j ../3 n
The equality holds for a mixture of the uniform distribution on the interval
[0,mF/[1- 2(j -1)/(../3n)]1/2] and the Dirac distribution concentrated at
mF/[1-2(j-1)/( ../3nW/2]1/2 with probabilities ../3(j-1)/n and 1-../3(j-
1)/n, respectively.
If (j - 1)/n > 1/../3, then sharp bound

EFYj:n ::; ../3 (1 + j - 1) (5.88)

mF 2 n

is attained by the uniform distribution on the interval [0, ../3mF]'

Proposition 21 (increasing failure rate) For a. > In[n/(n + 1 - j)]
uniquely defined by equation

!(1 - e- a )
a
- e- a = (1 _ j - 1) (1 -In
n
n
n+1-J
.) (5.89)
114 5. Order Statistics of Dependent Observations

and
(5.90)
we ha1Je
EFY;:n < n iiv(a.}
(5.91)
mF - n+1-j a.
The equality holds iff

0,
F(z} = { 1 - exp (-ii ':F) (5.92)
1,

which is a combination of a right truncated exponential distribution with

a pole at the truncation point [Link].

5.4 Distributions with Monotone Density and

Failure Rate on the Average
The results of this section are based on assertions of Lemmas 8 and 9 for
b = W-l«j - 1}jn} and c = dw. Therefore one can expect some analo-
gies with inequalities for quantiles of order p derived for b = W- 1 (P} and
c \, b. Observe that norm (3.57) depends on c, and the normalized projec-
tion (3.58) does not. It follows that bounds for quantiles and expected order
statistics substantially differ, but they are attained by the same distribu-
tions when p = (j - 1}jn. Note that condition (3.50) is false for c = dw,
and there are no trivial bounds EFY;:n $ IJF for j ~ 2. In Theorem 25 and
Propositions 21 and 22 we omit explicit descriptions of distributions for
which bounds are attained, and refer the reader to respective assertions in
Theorem 6 and Propositions 5 and 6 in which p and b should be replaced
by (j - 1}jn and W- 1 «j - 1)jn), respectively.
Theorem 25 (F!::. W) For 2 $ j $ n < 00, with b = W-l«j - 1)jn},
and the notation of (3.20) and (3.60), the following inequality is tight

EFY;:n -IJF <j - 1 IJw(b) .

(5.93)
UF - n Dw(a, b}

Proposition 22 (decreasing density on the average) For 2 $ j $

n < 00, bound

EFY;:n -IJF < v'3(j - 1)(n: + j - 1}

(5.94)
UF - n28u(~)

(cf. (3.64)) is the best possible.

 5.5 Symmetric Unimodal and U-Shaped Distributions 115

Proposition 23 (decreasing failure rate on the average) For 2 :$ j

:$ n < 00, bound
EF}j:n-J-tF < (j-1)(1+.1nn:IT=:1)
(5.95)
UF - nOv(~)

(cf. (3.67)) is the best possible.

Theorem 26 (F ~* W) For arbitrary continuous distribution function W

with a finite second moment there exists a sequence Fk ~ .. W, k = 1,2, ... ,
of distribution functions that have densities and finite second moments, and
satisfy
.
11m EFk }j:n - J-tFk ( j - 1 ) 1/2
sup =. (5.96)
k-too PE'P.. (Fk) UFk n +1- J
In particular, bound (5.96) is sharp for distributions with increasing density
and failure rate on the average.

Theorem 26 asserts that general bound (5.13) is attained among F ~ .. W

under mild assumptions on the maximal element W. The proof consists in
constructing a sequence of antistarshaped superpositions F k- 1W - J-tFIo (with
F,;IW(aw) \.t -00) that integrate to 0, and tend to 1[W-1((j-l)/n),dw) in
L2([aw,dw),w(x)dx). Actually, we can prove an analogous claim for arbi-
trary limiting function 2:7=1 djl[w-1((j-l)/n),W-1(jjn» with nondecreas-
ing coefficients dj , 1 :$ j :$ n, which means that general bound (5.11) for
an arbitrary L-statistic cannot be improved once we restrict the family of
distributions to any of the form {F: F ~* W}.

5.5 Symmetric Unimodal and U-Shaped

Distributions
Theorem 27 (symmetric unimodal distributions) If 0 < (j-1)/n :$
1/3, then
EF}j:n - J-tF < 2[n(j -1)]1/2
(5.97)
UF - 3(n + 1- j) .
This becomes the equality if Yl = J-t with probability 1 - 3(j - l)/n and
is uniformly distributed on [J-t - u[n/(j -1)]1/2,J-t + u[n/(j -1)]1/2] with
probability 3(j - l)/n.
If 1/3:$ (j - l)/n :$ 2/3, then

EF}j:n - J-tF <

_ raj - 1 ,
v~ (5.98)
UF n

which is the equality iff Y 1 is uniformly distributed on [J-t - V3u, J-t + v'3u].
116 5. Order Statistics of Dependent Observations

If (j - l)/n > 2/3, then

EF1j:n - I'F <~( n )1/2 (5.99)

UF - 3 n + 1- j
Here the equality is attained by the miXture of the atom at I' and the uni-
form distribution on [JL - u[n/(n + 1 - jW/ 2, I' + u[n/(n + 1 - j)]1/2] with
probabilities 3(j - l)/n - 2 and 3[1 - (j - l)/n], respectively.

Theorem 28 (symmetric U-shaped distributions) If either (j -l)/n

< 1/2 - 1/(2v'3) ~ 0.21132 or (j - l)/n > 1/2 + 1/(2v'3) ~ 0.78868 then
inequality (5.98) holds true with the respective condition for equality.
For 1/2 - 1/(2v'3) < (j - l)/n < 1/2 + 1/(2v'3) bound

(5.100)

is sharp, and this becomes the equality for the combination of uniform dis-
tribution on

[I' - (1- 4v'31~ - ~D1/2 ,I' + (1- 4v'31~ - ~D1/21

with probability 2v'31(j - l)/n - 1/21 and two atoms at the ends of the
interval, with identical probabilities 1/2 - v'31(j - l)/n - 1/21.

Proofs of Theorems 27 and 28 are based on projecting folded function-

als (5.18) onto convex cones C~c2U-1 and C~c2U-1. Observe that

B() B ()
gj:n X = n +j-1
1- jgn+1-j:n x ,

and the same holds for the respective projections and their norms. This
explains the relation between the bounds in both theorems and the fact
that the distributions with extreme expectations of jth smallest and largest
order statistics are identical. We can immediately conclude similar bounds
for some L-statistics.
In contrast with the independent case, all the bounds are nontrivial for
any order statistic except that of the sample minimum. Table 5.1 con-
tains numerical values of mean-variance bounds for order statistics from
dependent samples of size n = 20 coming from the following families of
distributions: general (G), symmetric (S), symmetric unimodal (SUN) and
U-shaped (SUS), distributions with decreasing density and failure rate (DD
and DFR, respectively), and with decreasing density and failure rate on the
average (DDA and DFRA, respectively). In fact, all the bounds depend on
TABLE 5.1. Sharp uniform mean-variance bounds on expectations of order statistics from dependent samples
of size 20 for various families of distributions.
j G S SUN SUS DD DFR DDA DFRA ID IFR
2 0.22942 0.16644 0.15692 0.08660 0.08660 0.05129 0.09028 0.05250 0.21558 0.21583 Q1
3 0.33333 0.24845 0.23424 0.17321 0.17321 0.10536 0.18543 0.10999 0.31208 0.31285 g,
4 0.42008 0.32219 0.30376 0.25981 0.25981 0.16252 0.28232 0.17247 0.39167 0.39325
5 0.5 0.39529 0.37269 0.34641 0.34641 0.22314 0.37897 0.23995 0.46398 0.46673
6 0.57735 0.47141 0.44444 0.43341 0.43301 0.28768 0.47458 0.31254 0.53287 0.53725
7 0.65465 0.55328 0.52164 0.52398 (")

8 0.73380 0.64359 0.60622 0.62188

0.51962
0.60624
0.35668
0.43078
0.56934
0.66414
0.39046 0.60045 0.60705
f.
0.47409 0.66814 0.67779 e.c:::
9 0.81650 0.74536 0.69282 0.73080 0.69389 0.51083 0.76038 0.56408 0.73703 0.75084 ~
10 0.90453 0.86244 0.77942 0.85499 0.78496 0.59784 0.85983 0.66139 0.80802 0.82760 ~
11 1 1 0.86603 1 0.88192 0.69318 0.96490 0.76749 0.88192 0.90962
12 1.10554 1.05409 0.95263 1.04499 0.98758 0.79851
rJ
Q..
1.07834 0.88455 0.95939 0.99878
13 1.22474 1.11803 1.03923 1.09620 1.10554 0.91629 1.20402 1.01574 1.04083 1.09750 ~
14 1.36277 1.19523 1.12583 1.15493 1.24084 1.04984 1.34731 1.16593 1.12587 1.20916
15 1.52753 1.29099 1.21716 1.22261 1.40106 1.20521 1.51632 1.34278 1.21244 1.33867 Ia.
16 1.73205 1.41421 1.33333 1.30023 1.59861 1.39393 1.72425 1.55911 1.29904 1.49376 t:I
fij.
17 2 1.58114 1.49071 1.38564 1.85592 1.63670 1.99448 1.83837 1.38564 1.68751
18 2.38048 1.82574 1.72133 1.47224 2.21944 1.97612 2.37741 2.22936 1.47224 1.94488 8:
0"
19 3 2.23607 2.10819 1.55885 2.80872 2.52143 2.99846 2.85796 1.55885 2.32232 a
20 4.35890 ~.
3.16228 2.98142 1.64545 4.09607 3.70340 4.35840 4.22194 1.64545 2.99772

.....
.....
~
118 5. Order Statistics of Dependent Observations

value (j-1)/n only, because they were derived by means of projecting func-
tions of (j-1)/n. Therefore all estimates for Y;:n hold also true for any Yk:m
provided that (k - l)/m =
(j - l)/n. In particular, the moment bounds
for the sample medians Yn :2n+1 do not change under increase of the sample
size. Moreover, one can check that all the bounds for restricted families of
distributions presented in this chapter increase continuously in (j - l)/n.
Consequently, Table 5.1 may provide fair approximations of bounds for Yk:m
from various families of parent distributions with (k-1)/m ~ (j -1)/20 for
some 2 ::; j ::; 20. Some tables for second moment bounds were presented in
Gajek and Rychlik [33]. Also, the third column (SUN) of Table 5.1 comes
from that paper. The numerical results for distributions with monotone fail-
ure rate and monotone failure rate on the average were earlier presented in
Rychlik [87].

5.6 Bias of Quantile Estimates

As in Section 4.5, we consider the problem of evaluating the bias for esti-
mation of population quantiles p-l(P) by sample quantiles X j:n, with the
only difference being that dependent observations are possible now. The
results stated here are valid for (j - l)/n ::; p < 1. The problem of es-
tablishing the upper bounds on the bias in general populations consists in
finding the derivative of the greatest convex minorant of the nonmonotone
distribution function
0, if o<x<
- -
i=l
n '
na:+1-j . i=l
Gj:n(x) -l(p,I)(X) = { ,(+1-{' If n::; X < p,
n a:-l
n+l-j'
if p <
_ x < 1,
calculating its £2- norm, and the form of the normalized derivative. The
lower bound amounts to the negative of the upper one for

p-l(P) + sup
PE'P" (F)
Ep(-Xj:n) = r P-l(x)(l(p,I) - G_e)(dx),
10
1

where ej is the jth unit vector, and

G- ei (x) = - min {jx, 1}

(cf. (2.25) and (2.26)). Again, we use the Moriguti [58] algorithm based on
the greatest convex minorant. The final results are shown below.
Theorem 29 (general distributions) For all (j - l)/n ::; p < 1, we
have
(5.101)
 5.6 Bias of Quantile Estimates 119

with

Q/'(j,n,p) = (_1_ + ~)1/2,

1-p j-1

(j/'(j,n,p) = n (1_p)I/2
n+1-j p
The upper bound in (5.101) is attained by the two-point marginal distribu-
tion (4.106). The lower one is attained by the three-point distribution

P (X=~-U [1_P~0P_1)/nr/2) =
j-1
n
j-1
P(X =~) = p-~,

(j _ l)/n ] 1/2)
P (X
[
= ~+U1_p+ (j -l)/n = 1-p.

Unlike in the independent case, there is no meaningful evaluation of the

bias oscillation Q/'(j,n,p) + (j/'(j,n,p). However,

I!/'(j,n,p) j - 1 ] 1/2 1
C/'(j,n,p) = Fj:n(P) [ p(n + j _ 1 _ np) ~ 2' (5.102)
B-/' (j,n,p) n+1-j 1
(j/'(j,n,p) = [1 - Fj:n(P)] n(l _ p) ~ 2' (5.103)

as j /n ~ p, which means that asymptotic maximal bias is twice as great

when the independence assumption is violated.

In order to determine the upper bias deviations for populations with

Ftc Wand F t. W, we need to project the step functions
0, if 0 <
-
x < W- 1 (i=.!.)
n '
hj:n,p,q(x) = { n+~-j - q':p' if W-l(P) $ x < W-1(q), (5.104)
n+l-j' elsewhere,
onto the convex cones of nondecreasing convex and nondecreasing star-
shaped functions, respectively, in L2([aw, dw), w(x)dx). Fortunately, func-
tions (5.104) obey the assumptions of Lemmas 18 through 21 which provide
exact formulae for the projections. As in (4.125),

o < - r (x -
J-y. -y.)hj:n,p,q(x)w(x) dx

=
n(l - q + p) + 1 - j
(n + 1 _ j)(q _ p)
r
J-y. (x - -y.)w(x) dx

< (c- -y. )n(l- q +1

p )+l- j
. ~,
0
n+ -3
120 5. Order Statistics of Dependent Observations

as q '\t p and c = W- 1 (q),'Y* '\t b = W- 1 (P). It follows that the bound for
F tc W has the form of (4.126) with hnW(x) replaced by gj:nW(x) =
n1[b,d)(x)j(n + 1- j). For F t* w, we have (4.137) with hnW replaced
by gj:n W. In both cases the normalized projection does not depend on the
specific form of h. Therefore the upper bounds in the dependent case are
attained by the same marginal distributions that provide the analogous
bounds for independent observations. Note that the same holds when F is
general.
Theorem 30 (F tc W) If (j -l)jn ~ p < 1 and b = W- 1 (P), then, with
notation (4.127) and (4.129), we have
EpYj:n - F- 1 (P)
< n 1/w(b)
--=........::..;.;.;.....--=-'- (5.105)
- n + 1 - j 'i?w (b)
A '

Inequality (5.105) is sharp.

Proposition 24 (decreasing density) If (j - l)jn ~ p < 1, then in-
equality
EpYj:n - F- 1(P) < n 3P ) 1/2 (3 - (5.106)
Up - n +1- j 1 + 3p
is sharp.
Proposition 25 (decreasing failure rate) If (j - l)jn ~ p < 1, then
inequality
EpYj:n- F - 1(p) < n (1_p)1/2
(5.107)
Up - n +1- j 1 +P
is sharp.

Theorem 31 (F t* W) If (j -l)jn ~ p < 1 and b = W- 1 (p), then, with

notation (3.59) and (3.60), we have
EpYj:n - F- 1 (P) -0 ( )
Up
< C t • w j,n,p
n 7]w(a, b)
= n+1-j'i?w(a,b)"
(5.108)

Inequality (5.108) is sharp.

Proposition 26 (decreasing density on the average) If (j - l)jn ~
p < 1, then the following inequality is sharp

EpYj:n - F- 1 (P) -0 ( )
< C t • v j,n,p
Up

= n(l + p) [ 3(1 - p) ] 1/2 5 109

n + 1 - j 1 + P + 7p2 + 3p 3 . (. )
 5.7 Extreme Effect of Dependence 121

Proposition 27 (decreasing failure rate on the average) If we have

(j - 1) / n ::::; p < 1, then the following inequality is sharp

-0
Ct.v(j,n,p)
1 -p
}
1~
n {
= n+1-j 1 + p[l-1n(l-p)J2
(5.110)

IT j /n -+ p, then bounds (5.105) through (5.107) do not tend to 0, unlike

the independent case. However, comparing asymptotic values of the upper
bias deviations for F toO W, U, and V we obtain nontrivial results:

B~.w(j,n,p) (b - a) Jbd w(x) dx

-0 -+
Ct. w(j, n,p) 2 Jbd(x - a)w(x) dx
W- 1 (P) -aw 1
= 2[[Link](W-l(P» - awl
<-
2'
(5.111)

B~.u(j, n,p) p
-+ 2(1 + p)'
(5.112)
C~.u(j, n,p)
B~. v(j, n,p) -In(l-p)
-0 ) -+ 2[1 -In(l - p)]'
(5.113)
Ct.v(j,n,p

IT p '\, 0, all the limits in (5.111) to (5.113) converge to O. Differentiating the

first one with respect to b, we check that this is increasing. The limiting
values of the right-hand sides of (5.112) and (5.113) for p /'" 1 are 1/4
and 1/2, respectively. We recall (5.103) to point out that the respective
limits in the general case amount to 1/2 for all p.
Table 5.2 is the counterpart of Table 4.2 for the dependent case. This
enables us to compare respective maximal upper bias deviations for finite
samples of size n = 20. The effect of dependence is actually notable in all
the families. Just as for the independent samples of general populations,
the quantile estimation is less precise on the tails. The same concerns the
dependent samples with decreasing density functions. For the DFR and
DFRA distributions, the upper bias deviations increase in j and p in com-
mon with the LLd. case.

5.7 Extreme Effect of Dependence

The problem we consider here stems from the robust statistics whose do-
main is studying sensitivity of statistical procedures against violations of
standard assumptions about statistical models. Robustness of L-statistics
which are popular tools in robust and nonparametric inference was studied
122 5. Order Statistics of Dependent Observations

TABLE 5.2. Sharp uniform variance bounds on upper bias deviations of estima-
tors lj:20 of pth quantiles, p = i/20, for various families of distributions (depen-
dent case).
j G DD DFR DDA DFRA
1 4.35890 1.57425 0.95119 1.71534 0.99748
2 3.15789 1.51700 0.95214 1.75672 1.04201
3 2.64497 1.47348 0.95925 1.77758 1.08589
4 2.35294 1.44088 0.96058 1.78338 1.12918
5 2.16506 1.41737 0.96825 1.77967 1.17204
6 2.03670 1.40175 0.97840 1.77127 1.21477
7 1.94681 1.39329 0.99127 1.76200 1.25780
8 1.88422 1.39159 1.00716 1.75471 1.30171
9 1.84257 1.39655 1.02647 1.75159 1.34727
10 1.81818 1.40836 1.04973 1.75437 1.39544
11 1.80907 1.42749 1.07763 1.76456 1.44744
12 1.81444 1.45479 1.11111 1.78373 1.50480
13 1.83450 1.49150 1.15142 1.81369 1.56953
14 1.87044 1.53947 1.20024 1.85672 1.64422
15 1.92450 1.60128 1.25988 1.91583 1.73235
16 2 1.68034 1.33333 1.99488 1.83837
17 2.10042 1.78017 1.42374 2.09772 1.96708
18 2.22222 1.89832 1.52944 2.22108 2.11701
19 2.29416 1.97386 1.60128 2.29389 2.22208
 5.7 Extreme Effect of Dependence 123

by many authors (we refer merely to monographs of Huber [41] and Hampel
et al. [37] and references given there). For standard i.i.d. parametric mod-
els, many violations of marginals were studied. Dependence-robustness for
location models was analyzed in Rychlik [80]. The projection method pro-
vides tools for determining expectation sensitivity of arbitrary L-statistics
against dependence of observations
n n

sup EF LCjlj:n - EF LCjXj:n

t
PE'P.. (F) j=1 j=1

= 11o
p-l(X) [Gc(X) -
3=1
cj/;:n(x)] dx, (5.114)

for various families of parent marginals P.

We present some preliminary results of Rychlik [88], where robustness
of single order statistics coming from general populations was studied. To
evaluate robustness of the jth order statistic in standard deviation units, we
first determine the projection of hj:n = gj:n - /;:n onto Co in L2([O, 1), dx).
Since hj:n is the difference of density functions, pOhj:n = p? hj:n. Thus we
are reduced to establishing projections onto the cone of nondecreasing func-
tions that can be determined by means of the greatest convex minorants.
The solution is trivial for the sample minimum, since

(5.115)
is actually increasing and thus p? h1: n = h1:n . Another simple solution
emerges for the sample maximum. Then
_nxn-l, if 0<x<1_1
hn:n(x) = { n (1- x n - 1) , i f 1_1<x<1
- n' (5.116)
n - - ,

has antiderivative
if 0 < x < 1 - 1
H (x) - { _xn
' - - n' (5.117)
n:n - n(x -1) + 1- x n , if 1- ~ :::; x:::; 1,

which is decreasing on [0, I-lin], increasing on [1- lin, 1], and concave
on both intervals. Therefore

- (1 --1 )n-l x if 0 <x<l--

H
=
1
n:n(x) ={ n (1 - nf)n (x -' 1) , if 1 1n :(
-
x <
- ,
i' (5.118)

and

P?hn:n(x) = { -( - r-)n
(I 1 t- 1
'
if 0<x<1_1
- n' (5.119)
n1- n , if 1- 1n <x<1.
- -
124 5. Order Statistics of Dependent Observations

A deeper analysis is needed for 2 ::; j ::; n - 1. Then hj : n decreases from 0

at 0 to - /i:n((j -1)/n), jumps up by n/(n+ 1- j), runs down to hj:n((j-
1)/(n - 1)) where /i:n is maximized and eventually increases to hj:n(l) =
n/(n+ 1- j). It is important to know the sign of hj:n{(j -1)/(n-l)). It can
be shown that this is negative for small and moderate j, and the proportion
of js for which this is true increases to 1, as n becomes large. In this case,
a thorough analysis shows that either Hj:n((j -1)/n) lies above the greatest
convex minorant which consists the line tangent to H j : n at 0 and some
s > (j - 1)/(n - 1), and Hj:n itself right to s, or Hj:n((j - 1)/n) spans
the minorant which has two linear pieces joining at (j -1)/n and coincides
with Hj:n in the right part. If hj:n((j - 1)/(n - 1)) ~ 0, then the convex
minor ant has the form analogous to the latter one, with the only difference
being that the slope of the second linear piece is positive. Analytic forms of
assumptions and resulting projection are precisely described in Lemma 25.

Lemma 25 Set

0< x ::; 1, (5.120)

. 1
~
n
<x <
-
1. (5.121)

If hj:n«j -1)/(n -1)) < 0, then there exist unique r E «j -1)/(n -1), 1)
such that hj:n(r) = 0 and s E [(j - 1)/(n - 1),r) such that either s =
(j - 1)/(n -1) if Sj:n((j -1)/(n -1)) ::; hj:n((j -1)/(n -1)) or s is the
solution to Sj:n(x) = hj:n(x) otherwise.
If

hj:n ( nj-l)
_ 1 < 0 and Sj:n(s) ::; Sj:n (j-l)
-n- , (5.122)

then
p/' hj:n{x) = {Sj:n(s), II:ff 0::; x < s, (5.123)
hj:n(x), s ::; x ::; 1.

If

then there exists a unique t E «j -1)/(n-l), 1) such that Tj:n(t) = hj:n(t),

and
if O<x<.i.=l
- n '
if .i.=l<x<t
n - -,
(5.125)
if t::; x ::; 1.
 5.7 Extreme Effect of Dependence 125

For the prevailing number of cases, we have (5.122) with s > (j -1)/(n-l).
Then (5.123) is continuous, and can be written as

This form is easier to handle in numerical calculations. Function (5.125)

has the only jump at (j -1)/n. Using projections (5.115), (5.119), (5.123),
and (5.125), we are in a position to write optimal bounds for the extreme
dependence effect.

Theorem 32 (general distributions) Bounds

EP Y 1:n - EpX1:n
sup
PEPn(P) ap
= A = A(I)
L..l. L..l.
n - 1
,n = ----:-
(2n - 1)1/2
(5.126)

are attained by the marginal distribution functions

F(x) = 1- -;
[1 ( n- x-p,
1 - (2n _ 1)1/2 -a-
1 )] 1/(n-1)
,

-(2n _ 1)1/2 ~ X - P, ~ (2n - 1)1/2. (5.127)

a n-l

If (5.122) holds for 2 ~ j ~ n - 1, then

(5.128)

for

(5.129)

Bound (5.128) is attained by

0, it X-f,L < Sj:n{S)

~ u A'
{ it Sj:n(S) < X-f,L < ~
F(x) = ~'~1 (~~) ~
it
A-U-A'
hj:n{s) < X-f,L < n/(nH-j)
(5.130)
3:n u' ~ A-u- A'
it X-f,L > n/{nH-j)
1, ~ u - A
126 5. Order Statistics of Dependent Observations

Under conditions (5.124) we have (5.130) with

t1 2 =
n
j _ 1 F j :n
2 (j~
- 1) + ( t-
j -
~ 1) [n
+ n 1 _ j - /i:n{t)
]
2

n 2{1- t) 2n
+ (n + 1 _ j)2 n +1_ j [1 - Fj:n{t)]
2i-2 2.. -2i)
+ n ·-1 .. _ .. - · [1 - F2j-l:2n-l (t)]. (5.131)

The supremum is attained by (5.130) with s, Sj:n(s), and hj:n{s) replaced

by (j - l)ln, Sj:n«j - l)ln) and hj:n{t), respectively.
Finally; the bound

EFYn:n - EFXn:n _ (n - 1)n-l/2

sup (5.132)
PE1'.. (F) UF - nn-l

is attained by the two-point marginal distribution concentrated on points

J.t - ul(n _1)1/2 and J.t + u(n - 1)1/2 with respective probabilities 1- lin
and lin.
In the problem of measuring the effect of dependence, it is of vital interest
to evaluate the lower deviations

EXj :n - EYj:n
A -(.
u. J,n
)
= sup
PE1'.. (F) UF

Fortunately, once we consider the class of arbitrary marginals, we can get

an immediate answer. Indeed, introducing

1 $ j $ n, with common distribution function

F-(x) = 1 - F([Link] - x-)

and order statistics

Xj,n = [Link] - Xn+l- j:n
(with the same notation for Ys), we have

-In
. f EYj:n - EXj :n
FE:Fp UF
EF-Y';+l-j:n - EF-X;+l_j:n
= sup
F-E:F UF
= t1(n + 1- j,n). (5.133)
 5.8 Open Problems 127

Accordingly, we conclude that the lower bound on the deviation of the ex-
pectation of the jth smallest order statistic under dependence is identical
with the upper bound on the deviation of the jth largest one. Similar ar-
guments explain the coincidence of (5.126) with the mean-variance bound
for the maximum in the i.i.d. sample {see (4.2)), and the fact that these
bounds are attained by distribution functions of variables mutually sym-
metric about J.t (see (4.3) and (5.127)). This is a consequence of relations
sup EFY1:n - EFX1:n = [Link] - EFX1:n = EF-X:L;'n - [Link]-' (5.134)
PE'P.. (F)

In fact, introducing dependence of Xi on [Link], being necessary in (5.134),

is redundant in arguing that leads to (5.133). Comparing (5.13) and (5.14)
for j = n with (5.132) we observe that

sup EFYn:n - EFXn:n

supEFYn:n - [Link]
= (1 _ .!.)
n
n-l ":It e- 1 Rj 0.36788, (5.135)

and both the suprema in (5.135), taken over all P E 'Pn{F) are achieved
by the same distribution function.
Numerical calculations based on Theorem 32 show that for general margi-
nal distributions central order statistics are more robust under dependence
than the extreme ones. We can also see that each jth smallest order statistic
is mOre sensitive than the respective jth greatest one. In fact, Rychlik [88]
measured dependence-robustness of order statistics in terms of scale pa-
rameters generated by central absolute moments of order 1 ~ p ~ 00. Nu-
merical analysis shows similarity of conclusion for cases p = 1 and p = 2,
which were presented above. For p = +00 (i.e., for bounded observations),
the conclusions are diametrically opposite. An interesting fact to note here
is that (5.135) holds true for the classes of marginal distributions with finite
pth moment for arbitrary 1 ~ p ~ 00.

5.8 Open Problems

1. What are the second moment bounds on order statistics of depen-

dent samples from populations determined by the star order? It is of
particular interest for F~* W for which the mean-variance bounds
coincide with the general ones. In the case of life distributions, con-
ditions of attainability require unbounded increase of the second raw
moments, which seems to be unrealistic in many practical problems.
128 5. Order Statistics of Dependent Observations

2. From the results for single order statistics we can immediately con-
clude bounds for some L-statistics. For instance, we have

1
L ¥i:n
k
= sup
PE'P,.(F)
EF k
+ -
1
J
.
l=j
(5.136)

j 1
L
k-1
= sup EF [ -Y}:n +- ¥i:n
PE'P,.(F) n n 1=j+1
n+l-k ]
+ n Yk:n, (5.137)

1 ~ j ~ k ~ n, (cf. (2.25) and (2.26)), and therefore all the bounds

derived in this chapter for a given jth order statistic coincide with
the bounds for trimmed and Winsorized means ((5.136) and (5.137),
respectively) that reject j - 1 smallest observations. Moreover, for
2 ~ j ~ n, we have

=
Therefore we obtain bounds for spacings Y}:n - y}-l:n in standard de-
viation units by multiplying all respective mean-variance bounds for
single jth order statistics of various families of sample distributions
by factor n/U - 1). In all these cases, the bounds are attained by
the same joint distributions. This is not so for the differences of other
pairs of order statistics. Then, due to (2.25) and (2.26), we have

sup EF(Y}:n - ¥ion)

PE'Pn(F)

= 11 p-1(x) [n + ~ _ j lW-1)/n,1)(X) - T1[O,i/n)(X)] dx

when 2 ~ i+ 1 < j ~ n. Therefore we should cope with the projections

of three-valued step functions onto various convex cones. Gajek and
Rychlik [32] proved that the projection of an n-valued step function
(which corresponds with general L-statistics of dependent samples of
size n at least) onto C~cw is a broken line with n pieces at most. We
suspect a similar conClusion for projections onto other cones deter-
mined by the convex order. However, it is a difficult analytic problem
to find the best approximations in such large parametric classes.
 5.8 Open Problems 129

3. Studies of bias in quantile estimation in the independent and depen-

dent samples were carried out for the same families of marginals.
Formulating unsolved problems in the latter case, we can directly re-
fer to the last two open problems of the previous chapter, where other
classes of interest were indicated. It seems that projection problems
may be solved more easily here, because the functional has a simpler
form.
4. We have no evaluations of the dependence effect on order statistics
from restricted families of marginal distributions. For some families
(e.g., those of symmetric and symmetric unimodal distributions) the
lower bounds could be deduced from the upper ones. We also deduce
the lower bounds for distributions with increasing density once we
find the upper bounds for distributions with decreasing density, and
vice versa. However, this is not so if we take into account distribu-
tions with monotone failure rate, and many other pairs of mutually
dual families. We also mention the importance of analogous evalua-
tions ofrobustness against dependence for many popular L-statistics:
spacings, ranges, trimmed means, and the like.
6
Records and kth Records

In comparison with evaluations of other statistical functionals discussed

here, investigations for record values are still at a preliminary stage, and
only a few results are presented now. Examples of Section 6.1 show that
the range of record values can be arbitrarily large when all types of inter-
dependence among the original variables are admitted. For the case of LLd.
sequences with general and symmetric distributions, mean-variance bounds
on standard and kth records, due to Nagaraja [59] and Raqab [74], respec-
tively, are presented in Section 6.2. In Section 6.3 second moment bounds
for distributions with decreasing density and failure rate are cited from
Gajek and Okolewski [31]. Finally, we discuss evaluations of Rychlik [83]
for increments of first records coming from various families of parent dis-
tributions. Note that Raqab [75] derived pth absolute moment bounds on
expectations of first records in general and symmetric populations based
on the HOlder inequality. Nagaraja [59] used the Jensen inequality for de-
riving some quantile bounds on expectations of records (see also Arnold
and Balakrishnan [5, Section 6.2]). Some bounds and approximations can
be found in Arnold et al. [8, Sections 3.8 and 3.9].

6.1 Dependent Identically Distributed

Observations
The results of the previous chapter prove that numerous nontrivial in-
equalities can be established for expectations of order statistics under the
132 6. Records and kth Records

assumption that all observations have an identical distribution, but they

are arbitrarily dependent. However, the assumption admits probabilistic
models with peculiar properties of records. The examples presented be-
low show that stating the problem of bounds for the expectation of record
values of arbitrarily dependent observations makes no sense.
In the sequence of identical variables Y1 = Y2 = ... , the primary record
value flo = Yi can never be improved, and kth records cannot be defined
for k ~ 2. On the other hand, we can construct a sequence of identically
distributed variables for which the first record value may be arbitrarily
large. Precisely, the notion of being arbitrarily large depends here on prop-
erties of distribution F of single observations. If F has an atom at the right
endpoint of its support dF, then R1 = dF. If dF is finite and F is continu-
ous at dF, then R1 can be arbitrarily close to dF. If dF is infinite, then R1
may have an arbitrarily large value. The conclusions follow from the next
theorem.
Theorem 33 For an arbitrary positive integer n, there exists a sequence
of standard uniform random variables lj, j ~ 1, for which R1 ~ 1 - lIn.
The proof is constructive. Combining the construction with transformation
F-1(lj), j ~ 1, we obtain a sequence of dependent random variables with
common distribution function F whose first record value satisfies R1 ~
F-1(1 - lIn). This is arbitrarily large in the sense described above.

PROOF. Fix n, and take n random variables Ui, which are uniformly dis-
tributed on intervals [(i-l)/n, i/n], 1 ~ i ~ n. Here Ui may be independent
or arbitrarily dependent, for example generated for a single uniform random
variable Uo by means of transformations

.
U . -_ i - I
n
+ Uo
, 1~ i ~ n.

Now create n random vectors

1 ~ i ~ n. (6.1)

Let (Y1 , •.. , Yn ) coincide with a randomly chosen vector Ui. We easily
check that lj, 1 ~ j ~ n, are uniformly distributed on [0,1]. Now extend
the sequence by adding i.i.d. random variables lj, j > n, with the same
uniform distribution, and independent of (Y1 , .•. , Yn). If (Yi, ... , Yn ) = Ui
for 1 ~ i < n, then L1 = i + 1, and

R1 = Un E[1 - ~,1] .
In order to complete the proof we notice that otherwise L1 > nand

R1 E (Un, 1] C [1 - ~, 1] .•
 6.2 General and Symmetric Distributions 133

Analogous assertions can be proved for general kth records. It suffices

to replace single Ui by k-tuples of (independent, say) variables uniformly
distributed on [(i - l)/n, i/n], 1 :5 i :5 n, generating vectors Ui (see (6.1»
of length kn. There were some attempts at imposing stronger conditions on
the dependence structure of observations (e.g., exchangeability in Arnold
and Balakrishnan [5, Section 6.3], stationary Markov property in Biondini
and Siddiqui [16]), but they do not lead to representations for which our
projection method works. In the remainder of this chapter we therefore
confine ourselves to record values of independent sequences.

6.2 General and Symmetric Distributions

Bounds on expectations of kth records for general and symmetric distribu-
tions were analyzed in Raqab [74] by means of Moriguti [58] projection of
functions (2.40) based on greatest convex minorants. Special cases of first
records for which (2.40) are convex functions varying from 0 to 00, were
solved in Nagaraja [59] by use of the Schwarz inequality. For general F, we
therefore have

EFRn-ILF
< -1
{I I
[-In(1-x)] 2n dx-1
f/2

: 10 hn(x)dx-1f/2
rTF (n!)2 0

= [C) I

= [(~) - l f / 2 (6.2)

n ~ 1, which becomes the equality for an affine transformation of the

Weibull distribution

F(x) = 1- exp ( - [n! (1 + nX ~ IL)] lIn) l[p-uID,oo)(X) (6.3)

with shape parameter l/n and

Observe that (6.3) has decreasing density and failure rate.

IT k ~ 2, then (2.40) vanishes at 0 and 1, and increases on [0, V(n/(k -
1»], and ultimately decreases. Analysis similar to that in Section 4.1 yields
p0(f~k) -1)(x) = p/' f~k)(x) -1 = f~k)(x) -1
= f~k)(min{x,a*}) -1,
where a. is uniquely defined in the following theorem.
134 6. Records and kth Records

Theorem 34 (general distributions) For given k ~ 2 and n ~ 1, de-

fine u* = u*(k,n) E (0, V(nJ(k -1))) as the solution to

(1 - x)f~k)(x) = 1- F~k)(x)
k
Ln -:-d-In(1
j
= (1 - x)k - x)]j, (6.4)
j=O J.

where f~k)(x) and F~k)(x) are the density function defined in (2.40) and
the respective distribution function. Then

(6.5)

for

The equality in (6.5) holds for

(6.7)

For n = 2,3, Equation (6.4) is solved by

u*(k,2) = 1 - exp ( - k(k ~ 1)) ,

1 + (2k - 1)1/2)
u*(k,3) = 1 - exp ( - k(k _ 1) ,

respectively, and then (6.6) has complicated explicit representations. Dis-

tribution function (6.7) has a finite support with a smooth density function,
and a pole with probability 1 - u* at the right end.
Analogous second moment bounds have slightly simpler forms

EFRn
mF
< c:) 1/2, n ~ 1, (6.8)
EFR~k)
< 11!~k)lI, n ~ 1, k ~ 2 (6.9)
mF
 6.2 General and Symmetric Distributions 135

(cf. (6.6)), and are attained by

F(x) = 1 - exp ( - [V(2n)! :f/n) l[o,oo)(x), (6.10)

0, if ~ $ 0,
F(x) = { (f~k»)-1(IIfi,k)II~), if 0 $ ~ < f,1~~)i,)' (6.11)
1 if a: > f~k)().)
, m - IIl~k II '
respectively. Both (6.10) and (6.11) define life distributions. Formally, in-
equalities (6.8) and (6.9) hold true for n = 0, and they amount to 1 then.
They are attained by degenerate random variables concentrated at mF > O.
Obviously, this can be approximated by life random variables in the mean
square with an arbitrarily desired accuracy.
Since
Sn(x) = fn(x) - fn(l- x), x E [~, 1) ,
is the difference of increasing and decreasing functions with identical values
at 1/2, this is strictly increasing from 0 to 00. Evaluating expectations of
the first records in symmetrically distributed sequences we simply use the
Schwarz inequality

EFRn - J1.F < /1

1/2
[F-1(x) - J1.]Sn(x) dx

< D 8 (I,n)uF = DUF, (6.12)

where

2D2 = /1 s;(x) dx

C:) -(n~)2 10
1/2

1
= Inn x Inn(1 - x) dx. (6.13)

The bound in (6.12) is attained by

F(x) = S;;1 ( V2D x : J1.) , (6.14)

which has a smooth symmetric density with the infinite support and sym-
metry center J1..
Raqab [74] also considered kth records in symmetrically distributed pop-
ulations using the greatest convex minorant approach. He proposed numeri-
cal evaluations that substantially improved bounds for general populations
(6.5) and (6.6), and nonsharp ones derived in Grudzien and Szynal [35]
based on direct application of the Schwarz inequality. However, Raqab's
paper lacks a theoretical justification for use of specific constructions of
the greatest convex minorant and precise description of the scope of appli-
cability. Therefore we do not present details here.
136 6. Records and kth Records

6.3 Life Distributions with Decreasing Density and

Failure Rate
We now aim at describing the bounds on EFR~k) /mF for the i.i.d. sequences
with the life distributions with decreasing density and failure rate. There-
fore we need to project (2.40) and their compositions with the exponential
distribution function
k +1
n
f~k)V(x) = __ xne-(k-1)z, k ~ 1, n ~ 0, (6.15)
n!
onto convex cones cit u and cit v' respectively. We first exclude trivial
=
cases. IT n 0, then _c _c

EFmk) = EFX1:k :5 /l-F :5 mF

for arbitrary F. Referring to the projection method, we obtain the same
conclusion. For n = 0 both (2.40) and (6.15) are nonincreasing concave
functions. It easily follows that their projections are constants amounting
to 1 by Lemma 1, with norm 1. We refer to (4.19) for construction of se-
quences that attain the bound in the limit. IT k = 1 and n ~ 1, then both
(2.40) and (6.15) are convex increasing, and so belong to cit u and cit v'
respectively. This implies that evaluations of respective exp~ctation ruiIc-
tionals by the Schwarz inequality provide the sharp bounds in both cases
coinciding with (6.8) valid for arbitrary F. Another way of verifying the
sharpness of (6.8) in the classes under study consists in checking that dis-
tribution function (6.10) providing equality in (6.8) has decreasing density
and failure rate.
In the remaining cases k ~ 2 with n ~ 1 we make use of auxiliary results
of Section 4.2. First we check that the pairs

(h(x), w(x)) = (J~k)(x), 1[0,1) (x)),

(h(x),w(x)) = (J~k)V(x),e-z1[o,oo)(x))

obey conditions (4.29) to (4.33). In fact, only verification of (4.33) is non-

trivial here. An easy computation shows that (6.15) is convex increasing on
=
(a, b) (0, (n- Vn)/(k-1)), concave increasing on (b, c) «n- Vn)/(k- =
=
1),n/(k-1)), and decreasing on (c,d) (n/(k-1), 00). Note that a b for =
the first record values. More effort is needed for establishing similar con-
clusions for (2.40). Namely, this is concave increasing on (0, V(l/(k - 1)))
and decreasing on (V(l/(k - 1)),1) for n = 1 and all k ~ 2. IT n ~ 2 and
k ~ 2, then f~k) is convex increasing on (0, b), concave increasing on (b, c),
and decreasing on (c,l) for c = V (n/(k - 1)) and

V(n - 1), if k = 2,
b = { V (n(2k-s)+[n 2+4 n(k-1)(k-2)]1/2) 'f k> 3 (6.16)
2(k-1)(k-2) ,1 _,
 6.3 Life Distributions with Decreasing Density and Failure Rate 137

This enables us to apply Lemma 13 to establish the parametric forms of

projections of f~k) and f~k)V. Since we have no convex increasing pieces
for n = 1, the projections are then linear. Consecutive reasoning steps
are similar to those in the proof of bounds on order statistics. The main
difference is that calculating (4.41) and (4.42) now we do not deal with
combinations of Bernstein polynomials, but obtain combinations of density
functions fM.) with the same j and various m. Without going into details,
we try to convince the reader by presenting formulae for derivatives and
indefinite integrals of f~k) and f~k)V:

f~1c]l (x) - (k - l)fAk)(x)

(fAk»), (x) = I-x
11 fAk)(y)dy = 1 ~x t
m=O
f!::)(x),

(fAk)V)'(x) = f~1c]l V(x) - (k - l)fAk)V(x),

1 f~)V(y)
00
dy = T L f!::)V(x).
-z

m=O
n

It is possible but unnecessary to replace positive functional coefficients

in the right-hand sides by more cOniplicated constants with simultaneous
replacement of k by neighboring integers. Gajek and Okolewski [31] estab-
lished the following variation diminishing property for the combinations of
density functions of kth record values in the LLd. uniform sequences.
Lemma 26 For every positive integer j and m, the number of zeros in
(0,1) of a given nonzero linear combination

L amfM.)(x),
n
fa(x) = x E (0,1),
m=O

does not exceed the number of sign changes of the sequence a = (ao, ... , an).
The first and the last signs of the sum are identical to the signs of the first
and last nonzero elements of a, respectively.
Note that it suffices to prove the claim for combinations of
·m+1
g(xjj,m + 1) = fM.)V(x)e- = J m!Z x me- jZ l[o,oo)(x) (6.17)

which are the density functions of r(j, m + 1) gamma distributions. This

follows from the fact that for fixed j, the one-parametric family of gamma

°
every positive integer i, and arbitrary nondecreasing sequences < Xl <
... < Xi, < VI < ... < Vi, we have
°
densities g(Xjj, v), v > 0, is strictly totally positive of order ooj that is, for
138 6. Records and kth Records

(see, e.g., Karlin [44, Chapter 1]). The variation diminishing property for
linear combinations of density functions of totally positive families is stated
in Karlin and Studden [45, Corollary 4.1].
A thorough analysis of (4.41) and{4.42) enables us to establish bounds
described in the following two theorems. We point out apparent analogies
with Theorems 11 and 12. Here we take into account k ~ 2 and n ~ 1, as
the remaining cases k = 1 and n = 0 were discussed above.
Theorem 35 (decreasing density) If (1 l/kt+l : ; 3, then
+
E F R(k)
mF
n <v'3 [
-
1-( _k_ )n+l]
k+1'
(6.18)

which is the equality if F is the uniform distribution on [0, v'3mF].

Otherwise, with notations (4.34) and (4.41), we have

E R(k)
F
mF
n ::; D = D~cu(k,n)
-
= 11(J~k»o< •.[Link], (6.19)

lr
for

11(J~k»o<•.8.W = k(2k ~ n
+l C:)Fi~k-l)(,8*)
+ (1- ,8.)[f~k)(,8.W + (1- ,8.)2a.f~k){,8.)
+ {I - ,8.)3a ;/3, (6.20)

(6.21)

where ,8. < b (the first inflection point of f~k) defined in (6.16») is the
smaller of the smallest positive zeros of combinations

Ku{x) = ~~ [1- (k~l)n-m] f::){x)

- kfn(k)
- 2 (x) +(k- 2 5) f n(k)
- 1 (x ) , (6.22)
 6.3 Life Distributions with Decreasing Density and Failure Rate 139

Lu(x) = ~~
2k m=O
[3 (_k_)n-m -1]
k+ 1
,(Ie)(x)
m

_
4k(k
(k - l)(k - 2) lie) ( )
+ 1) n-l X • (6.23)

Inequality (6.19) becomes the equality if

(6.24)
Theorem 36 (decreasing failure rate) If n $ 2k - 1, then

EFmale ) n+ 1
--"--- < - - (6.25)
mF - V2k'
which becomes the equality for the exponential distribution with the scale
parameter mF/V2.
Otherwise, under notations (4.34) and (4.41), we have

for

+
k(_k_) (2n)
n
2k - 1
2n+l F.(21e-l) ( )
2n 'Y. , (6.27)

n +1 (Ie)
= 2k(1 - 'Y.) [1 - Fn+l h'.)]

V-l h'.) [1 _ p(le) ('Y )] - !,(Ie) ('Y) (6.28)

2(1 - 'Y.) n· 2 n .,

where 'Y. < V«n - .fii)/(k - 1)) is the minimum of the smallest positive
zeros of

Kv(x) = ~ n - m,(Ie)( ) _
~ 2k2 m x
1- k 3 ,(Ie) ( )
k2 n-2 X
m=O
+ (k - 1)2(2k - 1) ,(Ie) ( ) (6.29)
2k2 n-l X ,
140 6. Records and kth Records

L ()
VX
= ~
L...J
2k-n+m,(k)( )_ (k-1)2,(k) ( )
2k2 m X 2k2 n-l X . (6.30)
m=O
The equality holds in (6.26) if
0, if .at. < 0 (10)
m -,
F(x) = (f~k»-l
{ (D;i) , if O::;;i::; in A-r·) ,
1 - (1 - , .. ) exp ( - D:;_!:Ie)(-y.»), if ;i ~ i~Ie~-r.).
(6.31)
We see that the extreme expectations of kth record values are attained by
the uniform and exponential distributions in Theorems 35 and 36, respec-
tively, if n is small. This is true for n = 1 and all k in particular, which
is an immediate consequence of the linearity of respective projections im-
plied by lacking a convex piece in the left parts of i1k) (x) and i1 k)V(x).
Distribution functions F$;!) (x) of the jth record values appearing in (6.20),
(6.21), (6.27), and (6.28) have alternative representations
F$;!){x) = G(-ln(l- x)jj,m + 1),
where the right-hand side denotes the gamma distribution functions with
densities defined in (6.17), composed with the standard exponential quan-
tile function. Absolutely continuous distribution functions (6.24) and (6.31)
have the same form up to a multiplicative factor on the left, and uniform

amount to 1 - 13. and 1 - '*'

and exponential right tails, respectively. The contributions of the latter
respectively.

6.4 Increments of Records

Here we present some standard deviation bounds on expectations of record
increments

EF(Rn - Rn-d = 11 [F-l(X) - #tHin{x) - in-l (x)] dx, n ~ 2. (6.32)

We omit case n = 1, because estimates for

EF{RI - Ro) = EFRI - #tF
were already presented. Properties of functions
CPn(x) = in(x) - in-l (x)
[-In(l - x)]n [-In{l - x)]n-l
=
n! (n -I)!
= in-leX) [-In{! -x) -1] n ~2, (6.33)
 6.4 Increments of Records 141

are crucial for determining its projection onto various convex cones. We
easily check that each CPn integrates to 0, starts from the origin, decreases
to
cp(1 _ e-n+l) = _ (n - l)n-1 ,
n!
and increases to +00 at 1 passing through the horizontal axis at 1 - e- n .
Its antiderivative «Pn needed for the Moriguti [58] projection, is therefore
concave decreasing, convex decreasing, and convex increasing in [0,1 -
e-n+l], [1 - e-n+l, 1 - e-n+l], and [1 - e- n , 1], respectively. This vanishes
°
at and 1, and is negative in the middle. Thus we deduce that its greatest
convex minorant <I>n is linear in [0, a*] for some a* E [1_e- n+l, 1- e-n+l],
that is determined by equation
XCPn(X) = «pn(x) = Fn(x) - Fn-1(X) = fn(x) (6.34)
(cf. (6.4)), and coincides with «Pn elsewhere. Finally,
o -,
P CPn(x) = «pn(x) = cpn(max{a*,x}).
Theorem 37 (general distributions) For n 2:: 2 we have

EF(Rn - R n- 1) ~~ = ~(n), (6.35)

O"F

where

~2(n) = (2n - 2) (1- a*) [2~1 (n~*)j + (1 _!) (na*)2n-1] (6.36)

n-l J!~ n (2n-l)!
3=0

for unique a* E (1 - e-n+l, 1 - e-n+l) satisfying equation

-In(1 - x) = nx. (6.37)
Equality in (6.35) holds for

F(x) = CPn-1 ( x
~-O"- -11-) l[I'-O"rpn(a.)/~,oo)(x). (6.38)

Formula (6.37) is a reduction of (6.34), and 11<I>~112 is rewritten as (6.36).

We can also show that
2n - 2) 1/2 2 1 n-

~(n) '" ( n _ 1 '" (mr)1/4'

which is the rate of increase of the extreme expectation of the (n - l)st

record value (cf. (6.2)). Distribution (6.38) has jump a* and a density with
infinite support right to the jump point. Observe that the contribution of
the smooth component 1- a* < e-n+l is very small.
142 6. Records and kth Records

We are able to establish similar bounds for populations with monotone

density and failure rate functions, once we apply the following observation.
Lemma 27 Decreasing-increasing functions IPn(x) and 1/Jn(x) = IPn(1 -
e- Z ), n ~ 2, are convex on their intervals of increase.
PROOF. For
xn x n- l
1/Jn(x) = n! - (n - 1)!
that increases for x ~ n -1, we simply verify the claim by repeated differ-
entiation. Since IPn is the superposition 1/Jn(x) with the increasing convex
function V-I (x) = -In(1 - x), we have

IP~(x) = 1/J~V-I(x)[V-I(x)12 + 1/J~V-I (x)(V-I)"(x) > 0

for x > 1 - e- n+1, which is the minimum point of IPn. The relation is
implied by the positivity of all the above components of IP~(x) .•

Proposition 28 (decreasing density and failure rate) For F tc U

bound (6.35) is the best possible.
The conclusion for F tc V follows from V h U. The statement of the
proposition can be proved by checking that (6.38) has a decreasing density,
which is equivalent to the convexity of IPn(max{a., x}). Since a. > 1 -
e- n+l , this immediately follows from Lemma 27.
Proposition 29 (increasing density and failure rate) For F ~c U

EF(Rn - Rn-l) < V3 (6.39)

O'F - 2n '

which is the equality for the uniform distribution on an arbitrary interval

of length 2V30'.
For every F :::;c V
EF(Rn - Rn-l) ~ 1, (6.40)
O'F
which becomes the equality for the exponential distribution with scale 0'.
Both assertions are deduced by means of analogous arguments. The cru-
cial steps of the proofs consist in showing that projections of IPn and 1/Jn
onto the cones of nondecreasing concave functions are linear. By Lemma 2,
neither IPn (1/Jn) nor the projection can majorize the other. Since both IPn
and 1/Jn are first decreasing and then strictly convex increasing, every nonde-
creasing concave function may cross either of them at two points at most.
Taking linear functions that pass through the crossing points improves
the approximations of IPn and 1/Jn. Standard arguments of the projection
method lead us to the final conclusions.
 6.5 Open Problems 143

6.5 Open Problems

1. An annoying still unsolved problem is to find the bounds on the
nth values of the kth records in symmetric populations for general n
and k. The trouble here lies in describing variability of the function

for 1/2 < x < 1. This is necessary for application of the greatest
convex minorant construction. Furthermore, we could ask for respec-
tive stricter bounds for symmetric unimodal distributions. Note that
these are unknown in the case of first records as well.
2. What are the mean-variance bounds for kth records of Li.d. sam-
ples with decreasing density and failure rate? For distributions with
increasing density and failure rate neither second moment nor mean-
variance bounds are known. A more general problem consists in estab-
lishing analogous results for distributions preceding and succeeding
a given one in the convex order.
3. What are the respective evaluations for the record values coming from
populations with distributions determined by the star order relations
with a fixed one?
4. Results of Section 6.4 should be completed by considering increments
of kth records R~k) - R~k21 in various classes of parent distributions.
It is of interest to compare kth record increments for different k.
5. We have practically no evaluations for the lower records. Generally,
they cannot be concluded from analogous results determined for the
upper records.
6. Establishing bounds on EF(Rn - Xn:n) in various families of distri-
butions we evaluate the rate of increase of the gap between the ele-
ments of nondecreasing sequences of sample maxima and their strictly
increasing subsequence. Analyzing EF(~k) - X n+1-k:n), we derive
similar evaluations for the kth largest order statistics.
7
Predictions of Order and Record
Statistics

In this chapter we evaluate expected increments of future order and record

statistics in the i.i.d. samples under conditions that some previous values
are known. These are important for predicting prospective failures in relia-
bility systems and shock models on the grounds of former data. The results
are based on representations of conditional expectations of order statistics
and records in terms of unconditional expectations of other ones presented
in Section 2.2, and evaluations of the latter derived in Sections 4.1, 4.2, 6.2,
and 6.3. These provide bounds in terms of conditional second moments

)_ 1
1 [F-l(X) - y]2 dx
2 _ E ((X _ )21X _F_--,l(~y)~~.,.....,...._ _ (7.1)
mFI~ - F Y >Y - 1 _ F(y) ,
where y is the value of a previous observation. We have
Jo1[F-l(x) - y]2 dx
2
mpo
III
< 1- F(y)
EF(X _y)2
= 1- F(y)
u} + (IJ.F - y)2
= 1- F(y)
= M;o =M,
III
(7.2)

say. The equality in (7.2) holds if

F- 1 (x) = y, x E [O,F- 1 (y)];
146 7. Predictions of Order and Record Statistics

that is, F has a jump at the left endpoint y of its support. It occurs that the
optimal constants for bounds expressed in terms of m ~o are identical with
. III
those measured in larger units M ~o • These have a more intuitive meaning
III
depending on moments and value of the parent distribution at a fixed ob-
servation point, and allow us to determine distributions attaining bounds
uniquely up to the three mentioned parameters. The former are indepen-
dent of the distribution at points preceding y, and so admit an ambiguity
in the description of extremal distributions. Therefore we choose represen-
tations of bounds in terms of M ~o -units. Results for general distributions,
III
and ones following the uniform and exponential distributions in the convex
order, are stated in Sections 7.1 and 7.2, respectively. Formal proofs will
be published in Rychlik [91].

7.1 General Distributions

We first consider predictions of order· statistics. Our aim is to estimate
expected increments of order statistics Xj:n - Xi:n for some 1 :::; i < j :::; n
when X i : n = Y is known. By (2.41),

EF(Xj:n - Xi:nIXi:n = y) = EFIIIXj-i:n-i - Y = EFI~ Xj-i:n-i, (7.3)

where F is a continuous distribution function with finite mean !JF and

variance u}, and

o F(x + y) - F(y)
1'111(x) = F11I(x + y) = 1 _ F(y) , x ~ 0, (7.4)

is the distribution function of X - y under the condition that X > y. Set

F(y) = p. Below we describe the maximal values of (7.3) for all distributions
with given !JF, UF, and F(y) = p measured in scale units

Note first that quantile functions of (7.4)

(7.5)

form the convex cone C+ of nondecreasing elements of L2 ([0, 1), dx) starting
from 0. Combining (7.3) and (7.4) with (2.23), and further applying the
projection method, the Schwarz inequality, and (7.2), we obtain
 7.1 General Distributions 147

EF(Xj:n - xi:nlxi:n = y) = 11 [F-l(p + (1- p)x) - y]!J-i:n-i(X) dx

< 11 [F-l (p + (1 - p)x) - Y]!j-i:n-i(X) dx

< 111t-i:n-illmFoIy
< 111t-i:n-i 11MFI~ , (7.6)

where !j-i:n-i is the projection of !J-i:n-i onto the family of nondecreas-

ing functions derived by means of the greatest convex minorant method.
The equality in (7.6) holds when p-l(p + (1 - p)x) - Y is proportional to
It-i:n-i, and equal to zero for arguments z = p + (1 - p)x E [P,1), and
z E [O,p), respectively. Note that for j = n constructing the projection is
redundant, because fn-i:n-i is actually increasing. If j = i + 1, the pro-
jection is constant amounting to 1. Explicit representations of bounds are
described in Theorem 38.
Theorem 38 (order statistics, general distributions) Sharp bound

EF(Xi+l:n - Xi:nIXi:n = y) < 1 (7.7)

MFo -
Iy

is attained by the two-point distribution

P(X = y) = p = 1 - P(X = Y + MFoIy ). (7.8)

For i +2 ~ j ~ n - 1, we have

EF(Xj :n - Xi:nIXi:n = y) < 111-··

-
.11 (7.9)
M FO 3-t:n-"
Iy

where the optimal constant 111t-i:n-ill is defined in (4.10) with parameter

0:.= o:.(j - i,n - i) determined by Equation (4.7). The equality in (7.9)
holds for

if ~ < 0,
°
- M
·· .(~)
if < ~ < f J-"n-. 1=1>
Ilfi-i:n-ill'
.. . ( .2i.::.2. )
if ~ > f J-"n-. 1-p
M - IIfi-i:n-ili
(7.10)
Finally,

EF(Xn:n - Xi:nIXi:n = y) < n- i (7.11)

MFo - (2n - 2i - 1)1/2
III
148 7. Predictions of Order and Record Statistics

holds and becomes the equality for the combination of degenerate and power
distributions
0, if z'M" < 0,
F(x)= { p+(I-p)[ 1. Z-"]l/(n-i-l) i~ 0< Z-" <v'2n-2i-l
v'2n-21-1 M ' :/ - M =-_:::-;----::-'
1, if z'M" ~ v'2n - 2i - 1.
(7.12)
All the distribution functions (7.8), (7.10), and (7.12) have jumps at y.
Also, (7.10) has an atom at the right end of its support, and (7.12) does not.
We emphasize the fact that basic relation (7.3) and, accordingly, issuing
inequalities (7.7), (7.9), and (7.11) hold true for continuous distributions F.
They are the best possible, but attainable merely in the limit by continuous
distributions tending to (7.8), (7.10), and (7.12), respectively, in the sense
described earlier (see, e.g., the last two paragraphs of Section 2.3). The
same reservations concern the other results of this chapter.

We now proceed to records and make use of identity

EF(R(k)
n
- R(k)IR(k)
m m
= y) = EFoIv R(k)
n-m-l> n > m, (7.13)

concluded from (2.44). Since for n = m + 1 we have

EF(R~~l - R~)IR~) = y) = EFI~ X1:k,
the evaluations of the first statement of Theorem 38 (see (7.7) and (7.8»
concerning the conditional expectations of spacings apply to the differences
of consecutive record values as well. We omit the case in our further con-
siderations. Mimicking arguments applied in (7.6), we obtain

EF (R~k) - R~) IR~) = y) = 11 [F-l (p + (1 - p)x) - y]J~kJm-l (x) dx

~ 11 [F-l(p + (1- p)x) - y]~kJm-l (x) dx
~ lI~kJm-lI1MFI~' (7.14)

For k = 1, we have ~kJm-l = f~kJm-l which are actually increasing.

Otherwise the Moriguti [58] construction of nondecreasing approximations
is necessary.
Theorem 39 (kth record values, general distributions) For k =1
(standard record values), we have

EF(Rn - RmlRm = y) < (2n - 2m - 2) 1/2

(7.15)
MFo - n-m-l '
Iv
 7.2 Distributions with Decreasing Density and Failure Rate 149

which becomes the equality for the mixture of Dirac and Weibull distribu-
tions

F(X)={P+(l- p )[l-ex p (- [v'(2n-2m-2)!7] l/(n-m-l)]} 1[0,00) (7)'

(7.16)
If k ~ 2, then

(k) (k)1 (k) )

EF ( Rn - Rm m:;, = y < Ilf-(k) II (7.17)
M FO - n-m-l'
Iy

where the right-hand side is defined in (6.6) with Q* = Q*(k,n - m - 1)

coming from (6.4). The equality in (7.17) holds for (7.10) with !i-i:n-i
(k)
replaced by fn-m-l'

7.2 Distributions with Decreasing Density and

Failure Rate
Here we confine ourselves to restricted families of parent distributions de-
fined by relation F tc W with a given W. We first notice that F tc W
does not entail the same for F1~ which appears in unconditional represen-
tations (7.3) and (7.13) of conditional expectations of the increments. This
is shown by the following counterexample.

EXAMPLE 4. Set F(x) = x l / 3 and W(x) = x l / 2 , 0 ~ x ~ 1. Obviously

F-lW(x) = X 3 / 2 is convex, and so F >-c W. Take now

FI~/S(x) = 2(x + 1/8)1/3 - 1, 0 ~ x ~ 7/8,

with the quantile function

(F1l/S) 1 + 1) 3 - 11,
o -1 (x) = g[(x O~x~1.

The composition

has the second derivative

negative for x E [0,1), which proves F1~/s -<c W .•

150 7. Predictions of Order and Record Statistics

However, relation P tc W is preserved if we compare F1~ with similar

= =
modification WI~ for properly chosen z such that W(z) P(y) p. Indeed,
we have

(7.18)

whose convexity is implied by that of p-l W(x). Note that the reversed im-
plication is false, because we can extend FI~ and WI~ on the left to y and z,
respectively, in an arbitrary way. Compositions (7.18) form the convex cone

+ 2 ( w(z +x) )
CtcWI~ C L [0, dw - z), 1 _ W(z) dx .

In particular, for uniform W = U, compositions (7.18) have the form

and

In the exponential case W = V, we have

and the identity

holds for all positive z.

We are now in a position to formulate the main results of this section.

We first focus on differences of consecutive order statistics and records.
Projecting the respective functionals we obtain

= y) ~ [Link]~ ~ mFI~ ~ MFI~'

EF(Xi+l:n - Xi:nIXi:n (7.19)

EF(R~~l - R~)IR~) = y) ~ [Link]~ ~ mFI~ ~ MFI~ (7.20)

(cf. (7.1)). Equalities in (7.19) and (7.20) hold if P-l(X) is constant for
x E [p, 1). Convexity and nondecrease of P-1W(x) impose the single value
of the quantile function on the whole domain. The conditional expectations
have a practical sense if y is the only support point of P, and they are equal
to 0, but all the evaluations are meaningless then.
Some results of this section can be immediately concluded from the pre-
vious one. For instance, (7.12) is a decreasing density distribution function
 7.2 Distributions with Decreasing Density and Failure Rate 151

which implies that (7.11) is the best possible bound for the restricted class.
In the remaining cases 3 ~ i + 2 ~ j < n we have
EF(Xj:n - Xi:nIXi:n = y)
= 1 1 P
- [F-l(x + p) - y]!i-i:n-i (1 ~ p) /'~ p
~ 10f1-P[F-l(X + p) - y]P: u!i-i:n-i
_c
(-1
x )
p -
dx
I-p
< lIP: u!i-i:n-illm~oIII
_c .

< lIP:
_c
u!i-i:n-iIlM~o.
III
(7.21)

The equality is a consequence of (7.3), (7.5), and a change of variables. The

first inequality follows from (2.2) and relation

P: U O !i-i:n-i(X) = P:
_c Ip _c
u!i-i:n-i (-1
x ),
_ P

which is implied by the equivalence of claims

9 P Ect UoIp <==>

(-1-·_ -) _c
g(.) Ec~cu'

r~
and the identity

1 1 P
- ~j-i:n-i (1 ~ p) -g (1 ~ p) 1 p= 1 1
[!i-i:n-i(x)-g(X Wdx.
The second inequality in (7.21) follows from the Schwarz inequality com-
bined with changes of variables, and (7.2) implies the last one. Equality in
(7.21) is obtained if

p,+ f..
tcU 3-I:n-t I-p
.( z )
F-1(x+p)-y = + M~o, o ~ X < 1, (7.22)
II P>-c u !i-i:n-ill III
F-1(x) = y, o ~ x <po (7.23)
Recalling Theorem 11, we can now state the following assertions.
Theorem 40 (order statistics, decreasing density) If i +2 ~ j <
min{(2n + i + 2)/3, n - I}, then

EF(Xj:n - Xi:nIXi:n = y) < va j - i (7.24)

M~o - n+l-i'
III

which becomes the equality for F being the convex combination of the Dirac
measure at y and the uniform distribution on [y, y + yaM] with respective
coefficients p and 1 - p.
152 7. Predictions of Order and Record Statistics

If (2n + i + 2)/3 < j ~ n - 1, then

EF(Xj : n - Xi:nIXi:n = y) < II(f····)

M -
II
J-,.n-I 0I.{3. , (7.25)
FI~

where the right-hand side is defined by (4.46) with 0. = 0.(13.) and 13. =
13.(j - i, n - i) uniquely determined by (4.47) through (4.49). Inequality
(7.25) becomes the equality for

< 0,
1
??::::1l.
M
o < ??::::1l. < !i-im-.({3.)
71 - M IIUj-':n-.)".[Link]'
-
~J 1;-':n-.({3.) < x-y < 1;-':n-.({3.)+OI.(I-{3.) (7.27)
IIUj-':n-.)".p.11 - M - IIUj-':n-.)".p. II '
x-y > !j-i:n-i ({3. )+01. (1-{3.)
M - IIUj-i:n-;}".II.11 '
respectively.
In the case F ~c V, the projection method is essentially exploited for pre-
dicting sample maxima as well. By arguments similar to ones used in (7.21),
for i + 2 ~ j ~ n we deduce that

EF(Xj :n - Xi:nIXi:n = y)
= 1 00
[F-l(p + (1 - p)V(x)) - yjli-i:n-i V(x)v(x) dx

< l°O[F-l(p+ (l-p)V(x)) -yjP~cvli-i:n-N(x)v(x)dx

< lIP:
_c
v/i-i:n-iVIIMFo,
Iy (7.28)

with the equality valid under conditions

F- 1 (p+(1-p)V(x)) -y =
P: +v Ii-i:n-i V(x) Mpo,
c
x~O,
IIPtc vli-i:n-NII Iy

F- 1 (x) = y, O~x<p

(cf. (7.22) and (7.23)). Finally, we refer to Theorem 12.

Theorem 41 (order statistics, decreasing failure rate) Under nota-
tion (4.51), if {LVj-i:n-i ~ 2, then

(7.29)
 7.2 Distributions with Decreasing Density and Failure Rate 153

This becomes the equality if F is the mixture of an atom at y and the

exponential distribution with location y and scale M/../2, with respective
coefficients p and 1 - p.
Otherwise

(7.30)

with the nonn and its parameters a. = a.h.) and ,. =

defined in (4.54) through (4.57). Equality holds in (7.30) for
,.(j - i, n - i)

(7.31)

(7.32)

respectively.

Note that (7.16) has decreasing density and failure rate. This means that
general bounds (7.15) are the optimal ones for the predictions of the values
of the first records from populations with decreasing density and failure
rate. Therefore it remains to study differences of nonsuccessive values of
kth records for k ~ 2. In order to obtain desired conclusions, we are reduced
to replacing !;-i:n-i(X) by f~".lm-l(x) in (7.21) and (7.28), and references
to Theorems 35 and 36.
Theorem 42 (kth records, decreasing density) If (l+l/kt- m ::; 3,
we have

(7.33)

which is the equality if F is the mixture of the unifonn distribution on [y, y+

JaM] and an atom at y.
In the opposite case, we have

defined in (6.20) with parameters a. = a.(/3.) and /3. = /3.(k,n - m -1)

defined in (6.21) to (6.23). Bound (7.34) is attained by the distribution
function given in (7.26) and (7.27) with !;-i:n-i replaced by f~".lm-l.
154 7. Predictions of Order and Record Statistics

Theorem 43 (kth records, decreasing failure rate) If m +2 ~ n ~

m+ 2k, then
(7.35)

This is the equality for the combination of the exponential distribution with
location parameter y and scale M/V2, and an atom at y.
Otherwise
(k) (k)1 (k) )
(
EF Rn - Rm Rm = y <
-
II(f(k) V) -1 II (7.36)
M FO n-m-l ct. V (-y.) ,
III

where the bound is defined by (6.27), and the parameters 0. = 0.(1'.) and
1'. = 'Y.(k,n-m-l) are determined from (6.28) to (6.30). Inequality (7.36)
becomes the equality for F defined in (7.31) and (7.32) with /i-i:n-i replaced
by f~k}.m-l'
Distribution functions (7.26), (7.31), and their modifications described in
Theorems 42 and 43 have unbounded derivatives in the right neighborhoods
of the left endpoints of their supports. The shape conditions entail unique
extensions of them to the left: only a jump at y and no mass of the left are
admitted. This means that the analogous bounds expressed in mFO -units
III
are attained by the same distributions only. This is not so in the remaining
cases. It is worth noticing that bounds (7.24) and (7.33) as well as (7.29)
and (7.35) are attained by the same elements of families with decreasing
density and failure rate, respectively.

7.3 Open Problems

1. What are the sharp bounds on conditional expectations of increments

of order and record statistics from populations with increasing density
and failure rate? More generally, we ask about respective evaluations
for the families of distributions determined by the convex order rela-
tions with a fixed abstract distribution function W.

2. Solve the analogous problem for the distributions defined by the star
order relations.

3. Applying (2.42) and the projection method makes it possible to re-

cover the conditional of past failures. Similar results for previous
record values can be established by exploiting the dependence struc-
ture of records.
 7.3 Open Problems 155

4. One can try to retrieve missing values of order statistics using (2.43).
However, the problem is that the quantile functions of doubly trun-
cated distributions Ell:do not form a convex cone, and one cannot
simply use our projection method here.

5. In contrast to the previous results, all the bounds presented in Chap-

ter 7 are expressed in terms of sophisticated scale units

MFo = [u} + ([Link] - y)2] 1/2

Iy I-F(y)

Is it possible to replace them by simpler and more intuitive ones?

Certainly, the problem also surpasses the range of direct applications
of projections onto convex cones.
8
Further Research Directions

In the monograph we mostly focused on the optimal upper bounds, but

certainly the lower ones are needed for evaluating the actual ranges of the
functionals over given classes of distributions. Generally, the best upper
and lower bounds are not symmetric about zero, but the latter can also be
derived by means of our projection method. To this end we should analyze
the negatives of the functionals under study. Only the functional corre-
sponding to the order statistics and L-statistics (see (2.27) and (2.25)) in
the dependent case needs a more subtle transformation. Changing the signs
of coefficients Cj, 1 ~ i :5 n, in (2.26) results in constructing a functional
g-e different from -ge. Also, one should realize that generally projecting
a functional and its negative are different problems that should be solved
separately by use of specific arguments. With few exceptions they cannot
be derived one from the other in a simple way.
Here we confined ourselves to statistical functionals defined for finite
samples, but various notions of asymptotic statistics are represented in
that form. As an example we mention the limits

of sequences of L-statistics

~
n
Cj,nXj:n = ~
n
h ~ - h (i- -1)]
[(i) ;- Xj:n
158 8. Further Research Directions

with coefficients determined by a smooth weight function h. From a prac-

tical point of view, we are interested in uniform bias evaluation for finite
sample estimates of asymptotic quantities in various classes of distributions.
In our example, it is to analyze the differences

An interesting topic of investigation is analyzing functionals of general-

ized order statistics extensively presented in Kamps [42]. The generalized
order statistics X (j, n, m, k), j = 1, ... n, based on distribution function F
with real parameters m, k satisfying

1}j = k + (n - j)(m + 1) ~ 1, 1 $ j $ n,

have expectations

EFX(j n m k)=
, "10
r F- 1(x) (j-1)!(m+1) (1_x)l1i- 1[1_(1_x)m+1]j- 1dx
1 j.
TIi=l1},

for m i: 1, and

EFX(j, n, -1, k) = 11 F- 1(x) g{:\1~ (1 - x)l1i-1[-ln(1 - X)]j-1 dx.

If m = 0 and k = 1, then X(j, n, m, k) reduce to the standard order statis-

tics Xj:n, 1 $ j $ n, of the independent sample with the common distri-
bution function F. If F is absolutely continuous, k is a positive integer,
and m = -1, then X(j,n,m,k) do not depend on parameter n, and coin-
cide with values of kth records R~k). For other choices of parameters, the
generalized order statistics represent observations of some censoring and
truncation schemes and complex reliability and shock models, for example
sequential order statistics and Pfeifer's record model in which the failure
probabilities of surviving elements change at the failure moments of the
other ones.
As an example of other classes of distributions that admit our projection
approach are the distributions determined by the relations of superadditive
order with a given W. We say that F succeeds W in the superadditive order
and write F t+ W if F-1 W is superadditivej that is,

(8.1)
for positive x and y. The reversed inequality defines F ~+ W. The su-
peradditive order implies the star one and so describes larger classes of
 8. Further Research Directions 159

distributions. In fact, relation (8.1) is applied for the life distributions and
should be satisfied for all 0 = aw ~ x + y < dw, but it can be modi-
fied by subtracting F-1W(aw) = F-I(O) from all the terms in (8.1) when
aw "# O. Especially for W = V, composition

V-I F(x) = -In[1 - F(x)]

defines the hazard function of F, and (8.1) describes the new worse than
used (NWU) distributions. The interpretation of the subadditivity of the
hazard function is that surviving time x + y by a single device is more
likely than under replacement by a new one in meantime x. The reversed
relation defines the new better than used (NBU) distributions. Since mean-
variance bounds presented here for F :5. W cannot be improved in general
populations, the same holds for the larger classes F:5+ W, and the NBU
distributions in particular. However, a difficult problem is to determine
bounds for F !:+ W, because the superadditivity does not allow a natural
graphical interpretation. Verifying the property of a function at fixed x + y,
we should study its values at all pairs x and y. Since
n
9j:n(X) =
n + 1 -). l[(j-I)/(n,l) (x)
is superadditive for (j - lIn ~ 1/2, we have the conclusion of Corollary 1
for the subadditive distribution functions F !:+ U then. We conjecture
that the bound on quantiles and order statistics of dependent samples of
the NWU populations are attained by Poisson distributions.
The method based on the greatest convex minor ants provides the best
L2-approximations of general functions by monotone ones. This was fruit-
fully exploited in determining sharp mean-variance and second moment
bounds on various statistical functionals in general and symmetric popu-
lations. For other families of distributions, the projection heavily depends
on properties of functions being projected. The reason is that there are no
known procedures which allow us to determine projections of general func-
tions onto corresponding convex cones of quantile functions and their mod-
ifications. What we strongly need here is, for example a general method of
projecting onto the family of convex functions. It is not clear here whether
one should try to generalize the Moriguti [58] method of convex minor ants
for this problem, or develop quite a different approach.
We obtain usually nons harp bounds on the statistical functionals defined
in Section 2.2 in terms of

and IIF-I - JtFllp, p > 1, by using the Holder inequality instead of the
Schwarz one. These are scale parameters expressed in terms of pth roots
160 8. Further Research Directions

of the raw and central absolute moments, respectively, of order p. As we

mentioned in the Introduction, general LP-projections onto convex cones
do not allow characterizations similar to (2.2) and (2.3) and, accordingly,
do not give us optimal bounds in the scale units generated by pth moments.
However, the L2 -projections onto the monotone functions can be modified
so that we derive sharp evaluations in terms of the above scale parameters.
Indeed, evaluating a nonzero statistical functional over the class of general
quantile functions by means of (2.2) and the Holder inequality, we have

Th(F- 1) = 11 F-l(X)h(x) dx

< 11 F-l(x)h(x) dx

< IIhllqllF-lllp (8.2)

with q = pi (p - 1) and h = p.l' h defined by means of the greatest convex

minorant construction (see Example 3, Section 2.1). The Holder inequality
becomes the equality if either F-l is zero or

(8.3)

for some positive a. In particular, the equality holds for

(8.4)

which is actually a quantile function with a desired pth moment, because

each XP' p > 1, is strictly increasing. Since the intervals on which (8.4) and
the original h are constant coincide, the former satisfies the conditions of
Lemma 3 for attaining equality in the second line of (8.2). Summing up,
we have
IIF-1I1p (I _ _
= Ilxphllp 10 Xph(x)h(x) dx
= IIhllqllF-lllp
= _ qIIIIF-l ilp
IIhll IIXphllpXph-II p'
which proves that (8.4) attains the equality in (8.2) that is actually sharp.
Obvious modifications lead to sharp bounds for general and symmetric dis-
tributions in terms of IIF-l- JLFllp. An important question is now whether
L 2-projections onto convex cones of functions that obey conditions more
stringent than monotonicity only admit modifications that provide sharp
bounds in terms of general pth norms.
 8. Further Research Directions 161

Finally, it is of interest if projections onto convex sets that are not neces-
sarily convex cones provide meaningful evaluations of functionals. IT so, we
could significantly extend the class of functionals for which respective eval-
uations hold by adding ones that act directly on the distribution functions,
densities, and other characteristics of distributions.
References

[1] Ahsanullah, M. (1995), Record Statistics, Nova Sci., Commack, NY.

[2] Arnold, B.C. (1980), Distribution-free bounds on the mean of the

maximum of a dependent sample, SIAM J. Appl. Math. 38, 163-167.

[3] Arnold, B.C. (1985), p-Norm bounds on the expectation of the max-
imum of possibly dependent sample, J. Multivar. Anal. 17, 316-332.

[4] Arnold, B.C. (1988), Bounds on the expected maximum, Commun.

Statist. Theor. Meth. 17, 2135-2150.

[5] Arnold, B. C. and N. Balakrishnan (1989), Relations, Bounds and

Approximations for Order Statistics, Lecture Notes in Statistics, Vol.
53, Springer-Verlag, New York.

[6] Arnold, B.C. and R.A. Groeneveld (1974), Bounds for deviations be-
tween sample population statistics, Biometrika 61, 387-389.

[7] Arnold, B.C., N. Balakrishnan, and H.N. Nagaraja (1992), A First

Course in Order Statistics, Wiley, New York.

[8] Arnold, B.C., N. Balakrishnan, and H.N. Nagaraja (1998), Records,

Wiley, New York.

[9] Balakrishnan, A. V. (1981), Applied Functional Analysis, 2nd ed.,

Springer-Verlag, New York.
164 References

[10] Balakrishnan, N. (1993), A simple application of binomial-negati-

ve binomial relationship in the derivation of sharp bounds for mo-
ments of order statistics based on greatest convex minorants, Statist.
Probab. Lett. 18, 301-305.
[11] Balakrishnan, N. and A.C. Cohen (1991), Order Statistics and Infer-
ence: Estimation Methods, Academic, Boston.
[12] Balakrishnan, N. and C.R. Rao (eds.) (1998), Order Statistics: Theory
& Methods, Handbook of Statistics, Vol. 16, North Holland, Amster-
dam.
[13] Balakrishnan, N. and C.R. Rao (eds.) (1998), Order Statistics: Appli-
cations, Handbook of Statistics, Vol. 17, North Holland, Amsterdam.
[14] Balakrishnan, N., C. Charalambides, and N. Papadatos (2001),
Bounds on expectation of order statistics from a finite popula-
tion, with an insight into Hartley-David-Gumbel, Samuelson-Scott,
Arnold-Groeneveld and some other bounds, submitted for publica-
tion.
[15] Barlow, R.E. and F. Proschan (1966), Inequalities for linear com-
binations of order statistics from restricted families, Ann. Math.
Statist. 37, 1574-159l.
[16] Biondini, R. and M.M. Siddiqui (1975), Record values in Markov
sequences, in: Statistical Inference and Related Topics, Vol. 2 (M.L.
Puri, ed.), Academic, New York, 291-352.
[17] Blom, G. (1958), Statistical Estimates and Transformed Beta- Varia-
bles, Almqvist and Wiksells, Uppsala.
[18] Boyd, A.V. (1971), Bound for order statistics, Publ. of the Elec-
trotechnical Faculty of Belgrade Univ., Math. Phys. Series 365,31-32.
[19] Caraux, G. and O. Gascuel (1992), Bounds on distribution functions
of order statistics for dependent variates, Statist. Probab. Lett. 14,
103-105.
[20] Chandler, K.N. (1952), The distribution and frequency of record val-
ues, J. Roy. Statist. Soc., Ser. B 14, 220-228.
[21] Cheng, C. (1995), The Bernstein polynomial estimator of a smooth
quantile function, Statist. Probab. Lett. 24, 321-330.
[22] David, H.A. (1981), Order Statistics, 2nd ed., Wiley, New York.
[23] David, H.A. (1988), General bounds and inequalities in order statis-
tics, Commun. Statist. Theor. Meth. 17, 2119-2134.
 References 165

[24] David, H.A., H.O. Hartley, and E.S. Pearson (1954), The distribution
of the ratio, in a single normal sample, of range to standard deviation,
Biometrika 41, 482-493.
[25] Dharmadhikari, S. and K. Joag-dev (1988), Unimodality, Convexity,
and Applications, Academic, New York.
[26] Dziubdziela, W. and B. Kopocinski (1976), Limiting properties of the
kth record values, Zastos. Mat. 15, 187-190.
[27] Fahmy, S. and F. Proschan (1981), Bounds on differences of order
statistics, Amer. Statist. 35,46-47.
[28] Feldman, D. and H.G. Thcker (1966), Estimation of non-unique quan-
tiles, Ann. Math. Statist. 37, 451-457.
[29] Franco, M. and J.M. Ruiz (1996), On characterization of continuous
distributions by conditional expectation of record values, Sankhya
A 58, 135-14l.
[30] Franco, M. and J.M. Ruiz (1999), Characterization based on condi-
tional expectations of adjacent order statistics: A unified approach,
Proc. Amer. Math. Soc. 127, 861-874.
[31] Gajek, L. and A. Okolewski (2001), Projection method for moment
bounds on record statistics from restricted families, submitted for
publication.
[32] Gajek, L. and T. Rychlik (1996), Projection method for moment
bounds on order statistics from restricted families. I. Dependent case,
J. Multivar. Anal. 57, 156-174.
[33] Gajek, L. and T. Rychlik (1998), Projection method for moment
bounds on order statistics from restricted families. II. Independent
case, J. Multivar. Anal. 64, 156-182.
[34] Gascuel, O. and G. Caraux (1992), Bounds on expectations of order
statistics via extremal dependences, Statist. Probab. Lett. 15, 143-148.
[35] Grudzien, Z. and D. Szynal (1985), On the expected values of kth
records and associated characterizations of distributions, in: Proba-
bility and Statistical Decision Theory, Vol. A (F. Konecny, J. Mo-
gyor6dy, and W. Wertz,eds.) Reidel, Dordrecht, 119-127.
[36] Gumbel, E.J. (1954), The maxima of the mean largest value and of
the range, Ann. Math. Statist. 25, 76-84.
[37] Hampel, F.R., E.M. Ronchetti, P.J. Rousseeuw, and W.A. Stahel
(1986), Robust Statistics. The Approach Based on Influence Func-
tions, Wiley, New York.
166 References

[38] Hartley, H.O. and H.A. David (1954), Universal bounds for mean
range and extreme observation, Ann. Math. Statist. 25,85-99.
[39] Hawkins, D.M. (1971), On the bounds of the range of order statistics,
J. Amer. Statist. Assoc. 66, 644-645.

(40) Huang, M.L. and P. Brill (1999), A level crossing quantile estimation
method, Statist. Probab. Lett. 45, 111-119.
[41] Huber, P.J. (1981), Robust Statistics, Wiley, New York.

[42] Kamps, U. (1995), A Concept of Generalized Order Statistics, Teub-

ner, Stuttgart.
[43] Karlin, S. (1957), P olya type distributions, II, Ann. Math. Statist. 28,
281-308.

[44) Karlin, S. (1968), Total Positivity, Vol. I, Stanford Univ. Press, Stan-
ford, CA.
[45) Karlin, S. and Studden, W.J. (1966), Tchebyshev Systems: With Ap-
plications in Analysis and Statistics, Interscience, New York.
[46] Klefsjo, B. (1983), A useful ageing property based on the Laplace
transform, J. Appl. Prob. 20, 615-626.
[47] Lai, T.L. and H. Robbins (1976), Maximally dependent random vari-
ables, Proc. Nat. Acad. Sci. U.S.A. 73,286-288.
[48] Lai, T.L. and H. Robbins (1978), A class of dependent random vari-
ables and their maxima, Z. Wahrsch. Verw. Gebiete 42,89-111.
[49] Lawrence, M.J. (1975), Inequalities for s-ordered distributions, Ann.
Statist. 3, 413-428.
[50] Lopez-Blazquez, F. (1998), Discrete distributions with maximum
value of the maximum, J. Statist. Plann. Inference 70, 201-207.
[51) Lopez-Blazquez, F. (2000), Bounds for the expected value of spacings
from discrete distributions, J. Statist. Plann. Inference 84, 1-9.
[52) Ma, C. (1992), Variance bound of function of order statistic, Statist.
Probab. Lett. 13, 25-27.
[53] Mallows, C.L. (1969), Extrema of expectations of uniform order
statistics, SIAM Rev. 11,410-411.
[54] Mallows, C.L. and D. Richter (1969), Inequalities of Chebyshev type
involving conditional expectations, Ann. Math. Statist. 40, 1922-
1932.
 References 167

[55] Marshall, A.W. and I. Olkin (1979), Inequalities: Theory of Majoriza-

tion and Its Applications, Academic, New York.
[56] Marshall, A.W. and F. Proschan (1970), Mean life of series and par-
allel systems, J. Appl. Probab. 7, 165-174.

[57] Moriguti, S. (1951), Extremal properties of extreme value distribu-

tions, Ann. Math. Statist. 22, 523-536.

[58] Moriguti, S. (1953), A modification of Schwarz's inequality with ap-

plications to distributions, Ann. Math. Statist. 24, 107-113.

[59] Nagaraja, H.N. (1978), On the expected values of records, Austral.

J. Statist., 20, 176-182.

[60] Nagaraja, H.N. (1981), Some finite sample results for the selection
differential, Ann. Inst. Statist. Math. 33, 437-448.

[61] Nagaraja, H.N. (1988), Some characterizations of continuous distri-

butions based on regressions of adjacent order statistics and record
values, Sankhya A 50, 70-73.

[62] Nagaraj a, H.N. (1992), Order statistics from discrete distributions

(with discussion), Statistics 23, 189-216.

[63] Nair, K.R. (1948), The distribution of the extreme deviate from the
sample mean and its studentized form, Biometrika 35, 118-144.

[64] Nevzorov, V.B. and N. Balakrishnan (1998), A record of records, in:

Order Statistics: Theory (3 Methods (N. Balakrishnan and C.R. Rao,
eds.), Handbook of Statistics, Vol. 16, North Holland, Amsterdam,
515-570.

[65] Okolewski, A. and T. Rychlik (2001), Sharp distribution-free bounds

on the bias in estimating quantiles via order statistics, Statist. Probab.
Lett., to appear.
[66] Olkin, I. (1992), A matrix formulation on how deviant can an obser-
vation be, Amer. Statist. 46,205-209.

[67] Papadatos, N. (1995), Maximum variance of order statistics, Ann.

Inst. Statist. Math. 47, 185-193.
[68] Papadatos, N. (1997), A note on maximum variance of order statistics
from symmetric populations, Ann. Inst. Statist. Math. 49, 117-121.

[69] Papadatos, N. (1997), Exact bounds for the expectations of order

statistics from non-negative populations, Ann. Inst. Statist. Math. 49,
727-736.
168 References

[70] Papadatos, N. (1999), Upper bound for the covariance of extreme

order statistics from a sample of size three, Sankhyii A 61, 229-240.

[71] Papadatos, N. (2001), Distribution and expectation bounds on order

statistics from possibly dependent variates, Statist. Probab. Lett., to
appear.

[72] Plackett, R.L. (1947), Limits of the ratio of mean range to standard
deviation, Biometrika 34, 120-122.

[73] Prakasa Rao, B.L.S. (1983), Nonparametric Functional Estimation,

Academic, Orlando, FL.

[74] Raqab, M.Z. (1997), Bounds based on greatest convex minorants for
moments of record values, Statist. Probab. Lett. 36, 35-41.

[75] Raqab, M.Z. (2000), On the moments of record values, Commun.

Statist. Theory Meth. 29, 1631-1647.

[76] Rychlik, T. (1992), Stochastically extremal distributions of order

statistics for dependent samples, Statist. Probab. Lett. 13, 337-341.

[77] Rychlik, T. (1992), Weak limit theorems for stochastically largest or-
der statistics, in: Order Statistics and Nonparametrics. Theory and
Applications (I.A. Salama and P.K. Sen, eds.), North-Holland, Ams-
terdam, 141-154.

[78] Rychlik, T. (1992), Sharp inequalities for linear combinations of ele-

ments of monotone sequences, Bull. Polish Acad. Sci. Math. 40, 247-
254.

[79] Rychlik, T. (1993), Bounds for expectation of L-estimates for depen-

dent samples, Statistics 24, 1-7.

[80] Rychlik, T. (1993), Bias-robustness of L-estimates of location against

dependence, Statistics 24, 9-15.

[81] Rychlik, T. (1993), Sharp bounds on L-estimates and their expec-

tations for dependent samples, Commun. Statist. Theory Meth. 22,
1053-1068. Erratum in Commun. Statist. Theory Meth. 23,305-306.

[82] Rychlik, T. (1995), Bounds for order statistics based on depen-

dent variables with given nonidentical distributions, Statist. Probab.
Lett. 23, 351-358.

[83] Rychlik T. (1997), Evaluating improvements of records, Appl. Math.

(Warsaw) 24, 315-324.
 References 169

[84] Rychlik T. (1998), Bounds on expectations of L-estimates, in: Order

Statistics: Theory & Methods (N. Balakrishnan and C.R. Rao, eds.),
Handbook of Statistics, Vol. 16, North-Holland, Amsterdam, 105-145.
[85] Rychlik, T. (1999), Error reduction in density estimation under shape
restrictions, Ganad. J. Statist. 27, 607-622.
[86] Rychlik, T. (2000), Evaluating statistical functionals by means of
projections onto convex cones in Hilbert spaces: Part I and II, in:
Applied Mathematics Reviews Vol. 1 (G.A. Anastassiou, ed.) World
Sci., Singapore, 407-489.
[87] Rychlik, T. (2001), Mean-variance bounds for order statistics from
dependent DFR, IFR, DFRA and IFRA samples, J. Statist. Plann.
Inference, to appear.
[88] Rychlik, T. (2001), Stability of order statistics under dependence,
Ann. Inst. Statist. Math., to appear.
[89] Rychlik, T. (2001), Optimal mean-variance bounds on order statistics
from families determined by star order, submitted for publication.
[90] Rychlik, T. (2001), Sharp mean-variance inequalities for quantiles of
distributions determined by convex and star orders, submitted for
publication.
[91] Rychlik, T. (2001), Predictions of increments of order and record
statistics in nonparametric families of distributions, in preparation.
[92] Samuelson, P.A. (1968), How deviant can you be? J. Amer. Statist.
Assoc. 63, 1522-1525.
[93] Schoenberg, I.J. (1959), On variation diminishing approximation
methods, in: On Numerical Approximation: Proc. of Symp., Madi-
son, 1958 (R.E. Langer, ed.), Univ. Wisconsin Press, Madison.
[94] Scott, J.M.C. (1936), Appendix to paper by Pearson and Chandra
Sekar, Biometrika 28, 319-320.
[95] Serfling, R.J. (1980), Approximation Theorems of Mathematical
Statistics, Wiley, New York.
[96] Shaked, M. and J.G. Shantikumar (1994), Stochastic Orders and
Their Applications, Academic, Boston.
[97] Sheather, S.J. and J.S. Marron (1990), Kernel quantile estimators, J.
Amer. Statist. Assoc. 80, 410-416.
[98] Tchen, A. (1980), Inequalities for distributions with given marginals,
Ann. Probab. 8, 814-827.
170 References

[99] Thompson, W.R. (1935), On a criterion for the rejection of observa-

tions and the distribution of the ratio of deviation to sample standard
deviation, Ann. Math. Statist. 6, 214-219.
[100] van Zwet, W.R. (1964), Convex Transformations of Random Vari-
ables, Math. Centre Tracts, Vol. 7, Mathematisch Centrum, Amster-
dam.
[101] von Mises, M. (1947), On the asymptotic distribution of differentiable
statistical functions, Ann. Math. Statist 18, 309-348.
[102] Vysochanskii, D.F. and Y. Petunin (1979), Justification of the three-
sigma rule for unimodal distributions, Theor. Probab. Math. Statist.
21,25-36.

[103] Zielinski, R. (1988), A distribution-free median-unbiased quantile es-

timator, Statistics 19, 223-227.
[104] Zielinski, R. (1998), Uniform strong consistency of sample quantiles,
Statis. Probab. Lett. 37, 115-119.
[105] Zielinski, R. (1999), Best equivariant nonparametric estimator of
a quantile, Statist. Probab. Lett. 45, 79-84.
[106] Zielinski, R. (2001), PMC-optimal nonparametric quantile estimator,
Statistics, to appear.
Author Index

Ahsanullah, M., 20, 163 Dziubdziela, W., 21, 165

Anastassiou, G.A., 169
Arnold, B.C., 18, 20, 24, 55, 101, Fahmy, S., 104, 165
103, 104, 131, 133, 163 Feldman, D., 83, 165
Franco, M., 24, 25, 165
Balakrishnan, A. V., 11, 12, 163
Balakrishnan, N., 18, 24, 55, 58, Gajek, L., v, 3, 55, 59, 68, 81, 96,
96, 104, 131, 133, 163, 118, 128, 131, 137, 165
164, 167, 169 Gascuel, 0., 19,99, 101, 103, 164,
Barlow, R.E., 28,164 165
Biondini, R., 133, 164 Groeneveld, R.A., 104, 163
Grudzien, Z., 135, 165
Blom, G., 55, 164
Gumbel, E.J., 2, 56, 165
Boyd, A.V., 104, 164
Brill, P., 82, 166 Hampel, F.R., 123, 165
Hartley, H.O., 2, 56, 165, 166
Caraux, G., 19, 99, 101, 103, 164, Hawkins, D.M., 104, 166
165 Huang, M.L., 82, 166
Chandler, K.N., 20, 164 Huber, P.J., 123, 166
Charalambides, C., 164
Cheng, C., 82, 164 Joag-dev, K., 27-29, 33, 165
Cohen, A.C., 18, 164
Kamps, U., 158, 166
David, H.A., 2, 18, 55, 56, 104, Karlin, S., 66, 84, 138, 166
164-166 Klefsjo, B., 54, 166
Dharmadhikari, S., 27-29,33, 165 Konecny, F., 165
172 Author Index

Kopocinski, B., 21, 165 111, 118, 123, 127, 128,

131, 146, 165, 167-169
Lai, T.L., 99, 166
Langer, R.E., 169 Salama, LA., 168
Lawrence, M.J., 28, 166 Samuelson, P.A., 104, 169
Lopez-Bhizquez, F., 55, 166 Schoenberg, LJ., 65, 169
Scott, J.M.C., 104, 169
Ma, C., 55, 166 Sen, P.K., 168
Mallows, C.L., 99, 104, 166 Serfiing, R.J., 16, 169
Marron, J.S., 82,169 Shaked, M., 29, 48, 169
Marshall, A.W., 14, 15, 167 Shantikumar, J.G., 29, 48, 169
Mogyorody, J., 165 Sheather, S.J., 82, 169
Moriguti, S., 3, 16, 33, 35, 55, 56, Siddiqui, M.M., 133, 164
59, 93, 118, 133, 141, Stahel, W.A., 165
148, 159, 167 Studden, W.J., 66, 138, 166
Szynal, D., 135, 165
Nagaraja, H.N., 24, 25, 56, 131,
133, 163, 167 Tchen, A., 99, 169
Nair, K.R., 104, 167 Thompson, W.R., 104, 170
Nevzorov, V.B., 24, 167 Tucker, H.G., 83, 165

van Zwet, W.R., 27, 28, 55,170

Okolewski, A., 55, 84, 131, 137,
von Mises, M., 16, 170
165, 167
Vysochanskii, D.F., 33, 170
Olkin, 1., 14, 104, 167
Wertz, W., 165
Papadatos, N., 55, 96, 164, 167,
168 Zielinski, R., 82, 83, 170
Pearson, E.8., 165
Petunin, Y., 33, 170
Plackett, R.L., 56, 93, 168
Prakasa Rao, B.L.S., 16, 168
Proschan, F., 15,28,104,164,165,
167
Puri, M.L., 164

Rao, C.R., 18, 164, 167, 169

Raqab, M.Z., 131, 133, 135, 168
Richter, D., 104, 166
Robbins, H., 99, 166
Ronchetti, E.M., 165
Rousseeuw, P.J., 165
Ruiz, J.M., 24, 25, 165
Rychlik, T., v, 3, 16, 18, 19, 30,
33,55,56,59,68,78,81,
84, 95, 96, 99, 103, 104,
Subject Index

atom, see degenerate distribution 89,92,106,111,117,120,

122, 138, 139, 142, 151-
Bernstein polynomial, 65, 66, 82, 153
137
bias, 82-92, 118-122 decreasing density on the average
binomial distribution, 58 (DDA) distribution, 29,
48,53,76,77,81,90,92,
Chebyshev inequality, 33 115, 117, 120, 122
Chebyshev system, 66 decreasing failure rate (DFR) dis-
conditional expectation tribution, 28, 40, 41,53,
oforder statistics, 22-24,146- 67,68,89,92,106,111,
148, 150-153 117, 120, 122, 139, 140,
of record values, 24, 25, 148, 142, 152-154
149, 153, 154
convergence decreasing failure rate on the av-
almost sure, 82, 91, 93 erage (DFRA) distribu-
in mean square, 31 tion, 28, 29, 48, 53, 77,
weak, 31 78,81,91,92, 115, 117,
convex cone, 1,2, 12-16,25-31 121, 122
convex order, 27, 28, 36-44, 60- degenerate distribution, 40, 41,
68,84-89,105-114,119, 44, 48, 58, 60, 89, 106,
120, 136-140, 149-154 107, 109, 111, 113, 116,
148, 149, 151, 153, 154
de Moivre-Laplace theorem, 83
decreasing density (DD) distribu- Dirac distribution, see degenerate
tion, 27, 40, 53, 66, 67, distribution
174 Subject Index

exponential distribution, 20, 40, kth lower record value, 22, 143
41, 44, 48, 67, 89, 106, kth (upper) record occurrence ti-
109, 111, 114, 139, 142, me, 21
153 kth (upper) record value, 21, 22,
131-143, 148, 149, 153,
failure rate, 28 154

gamma distribution, 20, 137, 140 L-statistic, 18, 19,56,93-97,100-

Gauss inequality, 33 104, 115, 116, 121, 123,
general distribution, 25, 26, 34, 128, 129, 157, 158
53, 57, 81, 82, 92, 100, Laplace transform order, 48
101, 117-119, 122, 125, Lebesgue monotone convergence
126, 134, 141, 147-149 theorem, 15
generalized order statistic, 158 life distribution, 26, 58, 60-68, 78,
Gini mean difference, 93 107, 136-140
greatest convex minorant, 3, 14, linear functional, 1, 11, 16-25
34, 35, 57, 84, 93, 118,
123, 124, 133, 135, 141, Markov inequality, 33
143, 147, 159, 160 Markov property, 24, 133

negative binomial distribution, 58

hazard function, 28, 159
new better than used (NBU) dis-
hazard rate, see failure rate
tribution, 54, 159
Hilbert space, 1, 2, 11-16
new worse than used (NWU) dis-
HOlder inequality, 103, 131, 159,
tribution, 159
160
norm, 1, 11
increasing density (ID) distribu- order statistic
tion, 27, 44, 53,109, 113, of dependent sample, 18, 19,
117, 142 95-129
increasing density on the aver- of independent sample, 2, 3,
age (IDA) distribution, 17, 18, 55-93, 123-127,
29, 48, 79, 115 146-148, 150-153
increasing failure rate (IFR) dis-
tribution, 28, 44, 53, 109, Poisson distribution, 159
113,114,117, 142 power distribution, 2, 56, 148, 149
increasing failure rate on the av- prediction
erage (IFRA) distribu- of order statistics, 146-148,
tion, 28, 29, 48, 79, 115 150-153
inner product, 2, 11, 13 of records, 25, 148, 149, 153,
inner product space, 11 154
projection, 1-3, 12-16,29-31
j-out-of-n reliability system, 18,
60 quantile, 17, 33-54, 82-92, 118-
Jensen inequality, 55, 131 122
jump, see degenerate distribution quantile function, 2, 16, 17,25-31
 Subject Index 175

record increment, 140-142, 148, uniform distribution, 27, 40, 44,

149, 153, 154 48, 66, 79, 89, 106, 109,
record occurence time, 20 111, 113, 115, 116, 138,
record value, 20,21,131-143,148, 142, 151, 153
149
Riesz representation theorem, 11 variation diminishing property,
robust estimate, 19, 93, 121, 123, 65, 137, 138
127
Weibull distribution, 133, 149
s-order,28, 50-52, 79-81,115,116 Winsorized mean, 128
sample interquartile distance, 18
sample maximum, 55, 56, 58, 59,
99, 101, 103, 104, 126,
127
sample median, 18
sample minimum, 57,58, 125, 127
sample quantile, 82, 118
sample quasirange, 93
sample range, 18, 55, 56, 93, 103,
104, 129
Schwarz inequality, 2, 11, 13, 46,
59, 68, 74, 87, 93, 101-
104, 133, 135, 136, 146,
151, 159
selection differential, 56, 104
spacing, 93, 128, 147, 150
star order, 28, 29, 44-50, 69-79,
89-91,114,115,120,121
starshaped function, 28
superadditive function, 158
super additive order, 48, 158
symmetric distribution, 26, 35, 53,
60,81,102,117
symmetric U-shaped distribution,
28,116, 117
symmetric unimodal distribution,
28, 52, 53, 79-81, 115-
117

three-point distribution, 35, 102,

119
totally positive function, 84, 137
trimmed mean, 18, 93, 128
two-point distribution, 34, 58, 82,
101, 116, 119, 126, 147
Lecture Notes in Statistics Vol. 105: Constantine Gatsonis, James S. Hodges,
Robert E. Kass, and Nozer D. Singpurwalla (Editors),
For information about Volumes I to 86 Case Studies in Bayesian Statistics, Volume n. x, 354
please contact Springer-Verlag , pages, 1995.

Vol. 87: J. Milller, Lectures on Random Voronoi V~l.106: ~d Niederreiter and Peter Jau-Shyong
Tessellations. vii, 134 pages, 1994. ShIUe (Editors), Monte Carlo and Quasi-Monte Carlo
Methods in Scientific Computing. xiv, 372 pages, 1995.
Vol. 88: J. E. Kolassa, Series Approximation Methods
Vol. 107: Masafumi Akahira and Kei Takeuchi Non-
in Statistics. Second Edition. ix, 183 pages, 1997.
i
Regular Statistical Estimation. vii, 183 pages, 995.
Vol. 8?: P. Cheeseman, and R W. Oldford (Editors),
Vo~. 108: 'Yesley L. Schaible (Editor), Indirect
Selectmg Models from Data: AI and Statistics lV. xii,
487 pages, 1994. EstImators m US Federal Programs. viii 195 pages
1995. "
Vol.. ~O: A. Csenki, Dependability for Systems with a
PartitIOned State Space: Markov and Semi-Markov Vol. 109: Helmut Rieder (Editor), Robust Statistics
Theory and Computationallmplementation. x, 241 Data Analysis, and Computer Intensive Methods. xiv,
pages, 1994. 427 pages, 1996.

Vol. 91: J.D. Malley, Statistical Applications of Vol. 110: D. Bosq, Nonparametric Statistics for
Jordan Algebras. viii, 10 I pages, 1994. Stochastic Processes. xii, 169 pages, 1996.

Vol. 92: M. Eerola, Probabilistic Causality in Vol. III: Leon Willenborg and Ton de Waal
Longitudinal Studies. vii, 133 pages, 1994. Statistical Disclosure Control in Practice. xiv', 152
pages, 1996.
Vol. 93: Bernard Van Cutsem (Editor), Classification
and Dissimilarity Analysis. xiv, 238 pages, 1994. Vol. 112: Doug Fischer and Hans-J. Lenz (Editors),
Learning from Data. xii, 450 pages, 1996.
Vol. 94: Jane F. Gentleman and G.A. Whitmore
(Editors), Case Studies in Data Analysis. viii, 262 Vol. 113: Rainer Schwabe, Optimum Designs for
pages, 1994. Multi-Factor Models. viii, 124 pages, 1996.

Vol. 95: Shelemyahu Zacks, Stochastic Visibility in Vol. 114: C.C. Heyde, Yu. V. Prohorov, R. Pyke, and
Random Fields. x, 175 pages, 1994. S.T. Ra.c~ev (Edit~rs), A~ens Conference on Applied
Probablhty and TIme Series Analysis Volume I:
Vol. 96: Ibrahim Rahimov, Random Sums and Applied Probability in Honor of J.M. Gani. viii, 424
Branching Stochastic Processes. viii, 195 pages, 1995. pages, 1996.

Vol. 97: R. Szekli, Stochastic Ordering and Vol. 115: P.M. Robinson and M. Rosenblatt (Editors),
Dependence in Applied Probability. viii, 194 pages Athens Conference on Applied Probability and Time
lW~ , Series Analysis Volume n: Time Series Analysis in
Memory ofE.J. Hannan. viii, 448 pages, 1996.
Vol. 98: Philippe Barbe and Patrice Bertail The
Weighted Bootstrap. viii, 230 pages, 1995: Vol. 116: Genshiro Kitagawa and Will Gersch,
Smoothness Priors Analysis of Time Series. x, 261
Vol. 99: C.C. Heyde (Editor), Branching Processes: pages, 1996.
Proceedings of the First World Congress. viii, 185
pages, 1995. Vol. 117: P~ul Glasserman, Karl Sigman, and David
D. Yao (EdItOrs), Stochastic Networks. xii, 298,1996.
Vol. 100: Wlodzimierz Bryc, The Normal
Distribution: Characterizations with Applications. viii, Vol. 118: Radford M. Neal, Bayesian Learning for
139 pages, 1995. Neural Networks. xv, 183, 1996.

Vol. 101: H.H. Andersen, [Link], D. Sl!Irensen, Vol. 119: Masanao Aoki and Arthur M. Havenner
and P.S. Eriksen, Linear and Graphical Models for the Applications of Computer Aided Time Series '
Multivariate Complex Normal Distribution. x, 184 Modeling. ix, 329 pages, 1997.
pages, 1995.
Vol. 120: Maia Berkane, Latent Variable Modeling
Vol. 102: A.M. Mathai, Serge B. Provost, and Takesi and Applications to Causality. vi, 288 pages, 1997.
Hayakawa, Bilinear Forms and Zonal Polynomials. x,
378 pages, 1995. Vol. 121: Constantine Gatsonis, James S. Hodges,
Robert E. Kass, Robert McCulloch, Peter Rossi and
Vol; 103: Anestis Antoniadis and Georges Oppenheim Nozer D. Singpurwalla (Editors), Case Studies in
(EdItors), Wavelets and Statistics. vi, 411 pages, 1995. Bayesian Statistics, Volume III. xvi, 487 pages, 1997.

Vol. 104: Gilg U.H. Seeber, Brian J. Francis, Reinhold Vol. 122: .Timothy G. Gregoire, David R Brillinger,
Ha~[Link], and G~riele Steckel-Berger (Editors),
Peter J. Dlggle, Estelle Russek-Cohen, William G.
Stattsttcal Modellmg: 10th International Workshop, Warren, and Russell D. Wolfinger (Editors), Modeling
Innsbruck, July 10-14th, 1995. x, 327 pages, 1995. Longitudinal and Spatially Correlated Data. x, 402
pages, 1997.
Vol. 123: D.Y. Lin and T.R. Fleming (Editors), Vol. 142: Gyilrgy Terdik, Bilinear Stochastic Models
Proceedings of the First Seattle Symposium in and Related Problems of Nonlinear Time Series
Biostatistics: Survival Analysis. xiii, 308 pages, 1997. Analysis: A Frequency Domain Approach. xi, 258
pages, 1999.
Vol. 124: Christine H. Muller, Robust Planning and
Analysis of Experiments. x, 234 pages, 1997. Vol. 143: Russell Barton, Graphical Methods for the
Design of Experiments. x, 208 pages, 1999.
Vol. 125: Valerii V. Fedorovand Peter Hackl, Model-
Oriented Design of Experiments. viii, 117 pages, 1997. Vol. 144: L. Mark Berliner, Douglas Nychka, and
Timothy Hoar (Editors), Case Studies in Statistics and
Vol. 126: Geert Verbeke and Geert Molenberghs, the Atmospheric Sciences. x, 208 pages, 2000.
Linear Mixed Models in Practice: A SAS-Oriented
Approach. xiii, 306 pages, 1997. Vol. 145: James H. Matis and Thomas R. Kiffe,
Stochastic Population Models. viii, 220 pages, 2000.
Vol. 127: Harald Niederreiter, Peter Hellekalek,
Gerhard Larcher, and Peter Zinterhof (Editors), Monte Vol. 146: Wim Schoutens, Stochastic Processes and
Carlo and Quasi-Monte Carlo Methods. xii, 448 Orthogonal Polynomials. xiv, 163 pages, 2000.
pages, 1997.
Vol. 147: Jiirgen Franke, Wolfgang HlIrdle, and
Vol. 128: L. Accardi and C.C. Heyde (Editors), Gerhard Stahl, Measuring Risk in Complex Stochastic
Probability Towards 2000, x, 356 pages, 1998. Systems. xvi, 272 pages, 2000.

Vol. 129: Wolfgang Hlirdle, Gerard Kerkyacharian, Vol. 148: S.E. Ahmed and Nancy Reid, Empirical
Dominique Picard, and Alexander Tsybakov, Bayes and Likelihood Inference. x, 200 pages, 2000.
Wavelets; Approximation, and Statistical
Applications. xvi, 265 pages, 1998. Vol. 149: D. Bosq, Linear Processes in Function
Spaces: Theory and Applications. xv, 296 pages,
Vol. 130: Bo-Cheng Wei, Exponential Family 2000.
Nonlinear Models. ix, 240 pages, 1998.
Vol. 150: Tadeusz Calinski and Sanpei Kageyarna,
Vol. 131: Joel L. Horowitz, Semiparametric Methods Block Designs: A Randomization Approach, Volume
in Econometrics. ix, 204 pages, 1998. I: Analysis. ix, 313 pages, 2000.

Vol. 132: Douglas Nychka, Walter W. Piegorsch, and Vol. 151: HAkan Andersson and Tom Britton,
Lawrence H. Cox (Editors), Case Studies in Stochastic Epidemic Models and Their Statistical
Environmental Statistics. viii, 200 pages, 1998. Analysis. ix, 152 pages, 2000.

Vol. 133: Dipak Dey, Peter Miiller, and Debajyoti Sinha Vol. 152: David Rlos Insua and Fabrizio Ruggeri,
(Editors), Practical Nonparametric and Semiparametric Robust Bayesian Analysis. xiii, 435 pages, 2000.
Bayesian Statistics. xv, 408 pages, 1998.
Vol. 153: Parimal Mukhopadhyay, Topics in Survey
Vol. 134: Yu. A. Kutoyants, Statistical Inference for Sampling. x, 303 pages, 2000.
Spatial Poisson Processes. vii, 284 pages, 1998.
Vol. 154: Regina Kaiser and Agustin Maravall,
Vol. 135: Christian P. Robert, Discretization and Measuring Business Cycles in Economic Time Series.
MCMC Convergence Assessment. x, 192 pages, 1998. vi, 190 pages, 2000.

Vol. 136: Gregory C. Reinsel and Raja P. Velu, Vol. 155: Leon Willenborg and Ton de Waal,
Multivariate Reduced-Rank Regression. xiii, 272 Elements of Statistical Disclosure Control. xvii, 289
pages, 1998. pages, 2000.

Vol. 137: V. Seshadri, The Inverse Gaussian Vol. 156: Gordon Willmot and X. Sheldon Lin,
Distribution: Statistical Theory and Applications. xi, Lundberg Approximations for Compound
360 pages, 1998. Distributions with Insurance Applications. xi, 272
pages, 2000.
Vol. 138: Peter Hellekalek and Gerhard Larcher
(Editors), Random and Quasi-Random Point Sets. xi, Vol. 157: Anne Boomsma, Marijtje A.J. van Duijn,
352 pages, 1998. and Tom A.B. Snijders (Editors), Essays on Item
Response Theory. xv, 448 pages, 2000.
Vol. 139: Roger B. Nelsen, An Introduction to
Copulas. xi, 232 pages, 1999. Vol. 158: Dominique Ladiray and Benoit Quenneville,
Seasonal Adjustment with the X-II Method. xxii, 225
Vol. 140: Constantine Gatsonis, Robert E. Kass, pages, 200 1.
Bradley Carlin, Alicia Carriquiry, Andrew Gelman,
Isabella Verdinelli, and Mike West (Editors), Case Vol. 159: Marc Moore (Editor), Spatial Statistics:
Studies in Bayesian Statistics, Volume IV. xvi, 456 Methodological Aspects and Some Applications. xvi,
pages, 1999. 282 pages, 2001.
Vol. 141: Peter Muller and Brani Vidakovic (Editors),
Bayesian Inference in Wavelet Based Models. xi, 394 Vol. 160: Tomasz Rychlik, Projecting Statistical
pages, 1999. Functionals. ix, 169 pages, 2001.