Asymptotics in Poisson order statistics
Citation for published version (APA):
Brands, J. J. A. M. (1988). Asymptotics in Poisson order statistics. (EUT report. WSK, Dept. of Mathematics and Computing Science; Vol. 88-WSK-08). Technische Universiteit Eindhoven.
Document status and date: Published: 01/01/1988
Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)
Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page numbers.
Link to publication
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement:
www.tue.nl/taverne Take down policy
If you believe that this document breaches copyright please contact us at: openaccess@tue.nl
providing details and we will investigate your claim.
Eindhoven University of Technology
Department of Mathematics and Computing Science
ASYMPTOTICS IN POISSON ORDER STATISTICS
by I.I.A.M.Brands
EDT-report 88-08
July1988
AMS subject classification: 41A60, 62E20
ASYMPTOTICS
IN
POISSON
ORDER
STATISTICS
by
JJA.M. Brands
Department of Mathematics, Eindhoven University of Technology,
The Netherlands ABSTRACf
In order statistics sums involving incomplete gamma functions are met. The asymptotic behaviour of such sums is studied, going beyond the results obtain-able by the central limit theorem.
1. Introduction
Some colleagues*) of the author have posed the following problem: Determine the asymptotic behaviour forJ.l.-+ooof the sums
(1) S (J.1,m, n):= ~ /m(J.1, k)(l-/(J.1,k»/1 ,
k=O 00
(2) T(J.1,m,n):= ~ (J.1-k)r(J.1, k)(1-/ijI.,k»/1 ,
k=O
where m and n are positive integers and Ii
(3) /(J.1,k):=(k!rl
I
e-'tkdt (J.1>O,kel'fo)o
This problem arose in the study oftheexpectation and variance of the order statistics in a random sample from the Poisson distribution with large meanJ.l. In section 2 we present the results. A brief description of the derivation is given in section 3. The details of the derivation are given in sections 4 to 8.
- 2-2. Results
The sums S and T have the following asymptotic behaviour:
(4) SijJ.,m,n)=A(m,n)J,1112+B(m,n)+OijJ.-112) ijJ.~-) , (5) TijJ.,m,n)=C(m,n)IJ.+D(m,n)J,1112+0(1) (IJ.~-) , where GIG (6) A (m,n)
=..J2
J
r(x)r(-x)dx , (7) B (m,n)=
-2/3J
xr(x)r(-x)dx , (8) C(m,n)=-3B(m,n) , 00 (9) D(m,n)=..J2
J
r(x)l'(-x)( 213-x2)dx , -eo and x (to)f
(x):= n-l12J
e-s2ds (XEJR.) .Clearly f can be expressed in the errorfunction but fonnulas (6) to (10) do not become simpler in doing so. Some coefficients are
(11) A (1,1)
=
2A(1,2)=
2A(2, 1)=
n-l12 B (n,n)=
0 (nEIN){3
B(1,2)=-B(2,1)=B(I,3)=-B(3,1)= 12 n-I D(1,1)=2D(1,2)= 2D{2,1)=
t
n-1I2 . 3. Sketch of the derivationThe results are obtained by taking the following steps.
(i) The sums are approximated by sums over Iu-kISIJ.213 with an error of 0 (e-Cflll3)
ijJ.~oo)
where c is some positive constant
(ii) ForkE [IJ.-C IJ.213, IJ.+c IJ.213]the asymptotic behaviour of1ijJ., k)forIJ.~oois detennined. Letx E JR. be defmed by
- 3-whereh(s):=-s+log(l+s) (s>-l) .
Roughlyx =UL-k) (2J1)-112. ThenIUL,k) has a complete asymptotic expansion in powers ofll-1I2wichis unifonn withrespecttox e JR,X=OULI/6) UL~oo).
I
UL,
k) :::f
(x)-21/2T 11t-I12e-x1J1-112+ ....
(J1~00),where fis defined by (10), i.e.
Ik -J11
~
c J12J3~
I
I(J1, k) - {f(x)+
e-x1f
ql(x)J1-l/2}I
~
B 1l-(N+l)12.1=1
(iii) The sums ~ are approximated by integrals
J ...
diewith an error which, forIIl-kISIl213 IIl-kISIl213
every positive number r, is OUL-r) UL~oo). Then these integrals are transfonned into
00
integrals over x and then approximated by integrals
J
dx with errors of the kind4. The truncation of the sum
The functionIUL, k)interpreted as a function of the real variable ke [0,00) is decreasing on [0,00). This statement follows from
11 00
~
IUL,
k)=(r(k+1»-2J
dtJ
e-t--ttk~
log(tt-1)dt<O o 11Let~1l-J.L2J3:=a. Then substitutingt=k(1+s) we have
l-JUL,
kJ:;;1-JUL,a)=(r(a+l)r1e-aaiol+1J
eah(s)ds ,a-I11-1
where h(s):=-s+log(l+s) . The function h is concave and negative on (-1,00) , whence
h (sJ:;;h(J1a-l_1)
+
(s-J1a-1+
1)h'(J1a-l-l) (s;::J,J.a-l-1).It follows that
00
0< e-a aa+l (r(a
+
1))-1J
eah(s)~e-aaa+l(r(a+1))-1eah (a-lll-l) (a h' (J1a-l-1)r1a-I11-1
-4-:lit 1/3 II]
0<
I:
<~. ~312e-TJ.l.=
O(e""1l ) ~-+oo)Q!;kS;a
k-lll-l
Let ~~+J1213
.
Then I(~,k)=e-kkk+l(r(k+I)rlJ
ekh(s)tis. Since h(s~-1I2s2 on (-1, 0] -1we have
k-lll-l
I~, k)~e-kkk+l(r(k+l))-1
J
e-1/2ks2 tis -100
~e-kkk+l(r(k+I»-1
J
e-1J2ks2tis
l-k-Ill
-~ekkk+l(r(k+1»-1
J
e-1/2k(l_k-1Il)stis l-k-IllHence, for both of the sums
Hence,inboth cases
where
c
is a positive number.S. The incomplete gamma function As we have seen already the substitution
(12) t=k(I+s)
in the integral representation ofI(~,k)gives k-'1l-1
(13) I~,k)=(k
!r
1e-kkk+1J
ekh(s)tis , -1where
(14) h(s):= -s+log(I+s) (s>-I). We introduce a new integration variable
y
by5 -(15) y:=k1l2Ih(s)1 112 sgn(S) (s>-I).
Then we get (16)
where
(17) X:=kll2lh(k-l~-I)lll2sgn(JJ.-k).
To study the transformation (15) we introduce first in (13) (18) t=lh(s)l l12sgn(s).
Then
(19)
..J2
t=S(1-2/3s+2/4s2-2/5s3+....)112 (IsI <1) ,=s-1I3s2+7/36s3+ ... (lsl<l) ,
the radius of convergence being one since h(s)/s2 has no zeros inside the unit circle. By the Biirmann-Lagrange theorem we can expand s as a powerseries in 't with a positive radius of con-vergence, say p.
We calculate (20)
(21)
The transformation (18) changes the integral (13) into
k-l12x
(22) I (JJ.,k)=(k!)-1
e-
kkk+1I
e-
k't2 : d't-00
We shall study the asymptotic behaviour of this integral fork~ooand fixed xeR. Therefore we
shall denote the right hand of (22)byi(x,k).
Inorder to use (21) we must truncate the interval of integration. Now it is easily shown that
where
c
is a positive number.Inside the circle I'tI:~1I2p the powerseries in (21) is also an asymptotic series, i.e. for every Ne lNwehave
(23)
Hence
-
6-ds _C"
..fi'
d't ="'42 +4/3't
+
6~-81135~+'.. +CN-l1'-1+0(1') (l'tIS1I2p)(24) i(x,k)
=
(k!r
le-kkk+lJ
e-k"'(..fi'+~
t+....+cN_ll'-l+O(-eN))d'tl'tISll2p, 't<k-lI2x
Now we change the lower bound into - 0 0and, evenually, the upper bound into k-l12x, thereby
.1:1/3
making an error of the kind0(e-C
) (k~-)uniformly in XEJR.
Then we substitutet=k-l12yand we get
(25) i(x,k)=(k
!r
le-.l:kk+1I2 (e-l (..fi'+4/3k-112y+~
k-l y 2+ . ..--kl/3
+0(e-e ) (k~oo) uniformly inXE JR.
x
Since
J
e-l O(k-Nl2y N) dy=
o
(k-Nl2) (k~-)unifonnly inxER, we have --(26)i
(x,k)=.8J!Q&
xJ
e-y2(..fi'+4/3k-l/2y+' .. +cN_ 1k-{N-l)l2yN-l) dy uniformly in XEm.,
where Hence N-l(28) i(x,k)=g(k) ,,£ctfi(x)k-l12+O(k-N/2) (k~oo)uniformlyin XE JR, 1=0
where the
c/s
are given by (21) and x7
-Integration by parts gives (30) Ji(X)=e-X
2
PI(X)+
l~li
r(1;1 )(21Cr1flf(x) (IeIN) . wherefisdefined by (10) andl
(/-l~
(31) PI(x):=-2-3fl1C-1flr(1+1)
L
(r(1+1 _s»-lxl-I-2s (IeIN) .. 2 s=O 2
Letting
x~oo
wefind thatJi(oo)=O if1is odd and Ji(oo)=r{1;1)(21Cr1fl ifIis even.Since limIUJ.,k)=l we get the asymptotic series for (g(k»-1 by lettingx~ooin (28). We have
Il--t""
(32) 1::::g(k) iClJi(oo)k-l'2
(k~oo)
.1=0
Comparing (28) (30) and (32) we see that the factor withwhichf (x) occurs in (28) has the same asymptotic expansion as the factorg(k) in (32). Hence
(33) I(x,k»::::f(x)+g(k)- LCle-- x2PI(x)k-lfl (k~oo)uniform1yinxeJR.
1=1
(Le. after truncation the (absolute) error is smaller thanCk-(N+l)12 withCindependent ofx).
According to the results of section 4 we restrict ourselves to values of ke[1J.;1213,IJ.+I1213].Then x=0UJ.l/6) UJ.~oo).
Inorder to get an asymptotic series forIUJ.,k) forlJ.~oowe have to express k as a function ofIJ. andx. From(17) we havekh{lJ.k-l_1)=-x2,sgnx=sgnUJ.-k). Puttingk=IJ.(1-'I') and 1J.-1flx=z we
get (34) 'I'+(1-'I')10g(1-'I')=z2 whence (35) Then (36)
i
~
=z2 (I'l'l <1), sgnz=sgn'l'. ;=2 (i-1)i- 8-00 (37) 'F 1:dizi (Iz I<r) . i=1 Hence 00 (38) k9J.- 1:difJ.l-iI2Xi (Ix I<r~). i=1
Clearly, forfJ. sufficiently large,xis within the range of convergence sincex=O(fJ.1/6) (fJ.~oo).
A few coefficientsdiare calculated. (39)
(42)
We need also expansions fork-1J2 , k-1andk-312 :
Dearly, there is a positive number'0 such that for all
ae
JR the function (1-'I')Q has a power-series expansion in z which converges for Iz I<r0; if00
(40) (l-'I')Q= 1:a/a)zj (I z I<ro)
j:4J
then
00
(41) kQ9J.Q(1-'l')Q= 1:aj(a)fJ.a-il2xj (Ix I<rofJ.1/2) .
j:4J We calculate k-l12
=
fJ.-ll2+rl12fJ.-lx+~fJ.-3I2x2+13
-fi
...,-2x 3+ ... 12 36 (43) (44) k-I9J.-1+21J2fJ.-3/2x+S/3fJ.-2x2+ ... k-312= fJ.-3/2+3.2-1J2fJ.-2x+ ...Anasymptotic expansion forg(k)can be determined from (32).
00
(45) g(k) :::: "Lglk-l12 (k~oo)
1=0
Some coefficientsglare
(46) go=I, g2=-1I12andg/=Oif 1
=
odd.-
9-(47) I(Il,k)
-=f
(x)+e-:iL
QI(X)Il-112 (IJ.-+oo) 1=1Theql'sare polynominals. We calculate
(48) (49) q1(x)=_21I2Tlx-112 ()= 5 -112 q2 X
- - x
x
126. The replacement of the sum by an integral We have already
(50)
By induction one can prove easily that
(51) I II. 00
~J(J.l,k)=(r(k+l»-l-l
I
dtI ...
· 0 0J
e-t--'tl- ' " -'t,(ttl· .. td'L(t,tit ...,tl)dtl ... dtl ,o
where the functionsL(to,tl' ...,t/)are defined by
(52) L(to)=1
With methods simular to those used in the treatment ofI (IJ.,k) in section 5 we can prove easily that
(54)
IJ3
with an error of the kind
o
(e-CII. ) (IJ.-+oo). Furthennore, if we restrict ourselves to values ofksuch that IJ.l.-kISJ.l.213, then the integrals inII. 1I.+2l3 1I.+211.2IJ
(51) canbereplaced by
J
I· ..
I
1I.-211.2IJ 11.+211.:113 1I.-211.2IJ
Then it follows from (51) that for IIl-k ISJ.l.213
(56)
10
-From (52) and (53) it follows that
[ 11+211213J 1+1 M ' +l~ M/log 213 (IeIN0), 11-21ol
whence, by (54) and (55), we have for all ke IV
d '
I - , I (IJ.,k)I
=
0 (IJ.-II3) (IJ.-.+oo) (IeIN0). dkNow we apply the Euler-Maclaurin sumformula: For every fixed reINwe have
q q (57)
'LI
(k)=I
1
(x)dx+
1/21
(q)+
1/21
(P) k=p p+
i:.
Bu[fU-l)(q)_fU-l)(P~
/=1 (21)!1
+o[
/1J"')(%111
(tf'p1 .Taking/(k)= Im(IJ.,k)(I-1{IJ.,k»/I,
p=f
11;12131,
q
=ll1+
I1213J '
we find, for every reIN,using (54), (56) and (57)1l+tL213
(58)
'L
rQ.t,k) (1-/(1J., k»/I=
J
r(IJ.,k)(I-1(IJ.,k»/IdieIk-jL151l213 1l-jL213
Simularly
1l+tL:n
(59)
'L
(u-k)/m(IJ.,k)(l-/{IJ.,k»/I=
J
(IJ.-k)r(IJ.,k) (l-/(IJ., k»/I dk Ik-jLISIl:n 1l-jL:n7. The asymptotic behaviour of Sand T
According to section 6 the sums
'L
canbe replaced by integrals with small errors whichIIl-k 1511213
are, for every r>O , 0(IJ.-r) (IJ.-.+oo). Then these integrals are transformed into integrals
B{jl.) dk
I ...
dx dxwhereA(11) andB(IJ.)are asymptotically equivalent with 1/2+2 11116 (IJ.-.+oo)11
-By means of (38) and (47) we get complete asymptotic expansions for the two integrands.
(60)
(61)
[m{JJ.,k)(1-[(JJ.,k»": :::
:i:
SI(X)~-l/2 (JJ.~oo)
,'~l
bothuniformly inXElR ,x
=
O{JJ.l/6)(JJ.~00). The functionsSI(X) andtl(X)are absolutely integr-able over(-00,00). Now we proceed as follows. LetNEIN.ThenJI.+J1213
t
[m{JJ.,k)(l-/{JJ.,k»" : dx JI.--JI. N B{JI.) [ N1 lJ
=L
~-l12J
sl(x)dx+O ~--21"7 '~l -A{JI.)It is easily seen that the functions SI(X) are of the form
2 2 2
p(x,e-X
,1
(x» e-x +q(x,e-X ,/(x» I (x)(l-I (x»,
where p and q are polynomials. Hence
J
Is,(x)ldxandJ
Is,(x)ldxareo
(e-eJl.II3)(JJ.~00).
X<-A(JI.) x>B{JI.)It follows that S (J..l,m,n)has a complete asymptotic expansion
00
(62) S{JJ.,m,n):::
i
~-l12
J
sl(x)dx(JJ.~00)
/=-1 . - 0
A simular argument holds forT(JJ.,m,n).
00
(63) T{JJ.,m,n):::
:i:
J..l-I12J
tl(x)dx(JJ.~oo).
1~2 We calculate (64) S-l (x)=.J2!"'(x)!"(-x) (65) SO(X)=2/31t-I12 [nr(X)f"-I(-x)-mr-l(X)f"(-X~
e-x2-2/3xf"'(x)f"(-x) (66) L2(X) =2x !"'(x)!"(-x) (67) Ll(x)=
23123-11t-112[n!"'(x)f"-I(-x)-mr-l(X)f"(-X~
xex2-21I2X2r(x)!"(-x) 2The term with factor
e-
x in the right hand of (65) gives 0 upon integration.2
Integration by parts of the term with factor xe-x in the right hand of (67) gives
00
12 -8. Computation of some coefficients
~ ~
A(1, 1)="2
J
f (x)(I-f (x»dx=-"21t- 1I2J
x (l-2f (x»e-X2dx --~ ~ =2"21t-1/2J
f (x)xe-X2dx="21t-1J
e-2x2dx=1t-l12 --~ A (l,2)=A(2, 1)= 112"2J
(f(x)f2(-x)+f(-x)f2(x»dx --~ =112"2J
f(x)f(-x)dx=lI2A(l,I) --B(1,1)=-2/3J
x! (x)f (-x)dx=O . B (1,2)=-B(2, 1)=-2/3J
x!(x)f2(-x)(f(x)+f(-x»dx --~ =-2/3J
x!(x)f3(-x)dx=B(1,3)=-B(3,1) B(1,2)=-2/3J
x(f (x}-2f2(x)+f3 (x»dx ~= 1I31t-1I2
J
x 2e-x\l-4f (x)+3f2(x»dx~ = 1I61t-lJ2
J
e-x2(1-4f (x)+3f2(x)+1t-1I2xe-x2 (-4+6f (x»)dx ~ ~=
1/6J
(1-4f (x)+3f2(x»!, (x)dx+1t-1J
xe-2x2f (x)dx --00 ..J?: 1/4 -3/2J
_3x2dx 3-1 = 1te
=
12 1t . --~ D(1,1)=2/3A(I,I}-"2J
x 2f(x)f(-x)dxfi
~ =2/3A(1,I)-2~
1t-1/2J
x 3e-X2f(x)dx--- -- 13-=2/3A(1 1)-