Statistics & Probability Letters 6 (1988) 159-162 February 1988
North-Holland
WHEN ARE INTERMEDIATE PROCESSES OF THE SAME STOCHASTIC ORDER?
Bruce COOIL
Owen Graduate School of Management, Vanderbilt University, Nashville, TN, USA
Received April 1987
Revised June 1987
Abstract: Let Z(~ n) represent the ruth largest order statistic in a random sample of size n. Here we study the process Z[~,~I,
t > 0, where re(n) is an intermediate sequence such that m ---,oo, m/n ~ 0 as n ~ oo.
Keywords: intermediate order statistics, differentiable domains of attraction, extremal distribution.
Let Z~~) >/Z: t") >I --- >/Z~ ~) represent the de-
scending order statistics formed from a random
sample of size n taken from a common distribu-
tion function F. When re(n) is an intermediate
sequence such that
m~ and m/n~O asn~oo, (1)
intermediate order statistics of the form 7(~)
L'[mt],
t > 0, have been used in a variety of ways to make
statistical inferences about the upper tail of F
(here [x] refers to the integer part of x, and t is
fixed with respect to n). For examples of such
estimators see Pickands (1975), Davis and Resnick
(1984), Hall and Welsh (1984, 1985), Csbrg~5 et al.
(1985), Smith and Weissman (1985), and Smith
(1986). One can study the asymptotic joint distri-
bution of such estimators by considering the sto-
chastic process of the form
to~n)(t ) =- k [ L'[ mt]y(n) -- Q( mt/n ) )/a~ ~) (2)
where a~ ") > 0 does not depend on t and where
Q(.) is the quantile function of the upper tail of
F,
Q(p)=inf{y:l-F(y)<~p}, 0<p~<l. (3)
Research supported by the 1987 Dean's Fund of the Owen
Graduate School of Management and the Vanderbih Univer-
sity Research Council.
0167-7152/88/$3.50 © 1988, Elsevier Science Publishers B.V. (North-Holland)
Of particular interest are conditions under which
there exists a scaling sequence a~ ") > 0 such that
to~")(t) has a limiting stochastic process to(t) as
n --* oo. In this paper we find a necessary condi-
tion for the existence of sequences m(n) and a~ n)
such that to~")(t), t > 0, has a nondegenerate limit-
ing process to(t), which applies whenever F is
continuously differentiable in the upper tail. (This
condition has already been shown to be sufficient.)
We will say to~")(t) has a limiting stochastic
process to(t) and write to~")(t) ~D to(t), t > 0,
whenever all finite dimensional distributions of
the process to~n)(t) converge to those of to(t):
(to n) ( tl ) ..... to(ran) ( tk )) d( to(tl) ..... to( tk ))
(4)
as n---,oo for any values 0<t 1< ... <t k<oo.
The weak convergence of the stochastic process
to~n)(t) to a nondegenerate limit to(t) is of particu-
lar interest because it implies that the centered
intermediate order statistics Z (n) -
I-,,1 Q(mt/n) are
asymptotically of the same stochastic order for all
t. When this is true, it becomes possible to esti-
mate, among other things, functions of extreme
quantiles of the form Q(mt/n), t > O, or the limits
of such functions as n-~ oo, by using the same
functions of the corresponding intermediate order
statistics even when these functions do not rescale
159