compound-Poisson distributions
Citation for published version (APA):
Harn, van, K., & Steutel, F. W. (1975). Interpolating between the compound-geometric and compound-Poisson distributions. (Memorandum COSOR; Vol. 7510). Technische Hogeschool Eindhoven.
Document status and date: Published: 01/01/1975
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.
STATISTICS AND OPERATIONS RESEARCH GROUP
Memorandum COSOR 75-10
Interpolating between the compound-geometric and compound-Poisson distributions
by
K. van Harn and F.W. Steutel
by
K. van Harn and F.W. Steutel
O. Summary
The starting point of this paper is the characterization of the compound-Poisson, i.e. the infinitely divisible lattice distributions (class C
1) and the subset of compound-geometric lattice distributions (class CO) by the non-negativity of recursively defined quantities (section 1). In section 2 we generalize these recursion relations by introducing sequences c (a) for n a £ [0,1] and thus we obtain classes of distributions C t which we wish to
a
be increasing with a. This can be achieved by an appropriate choice of cn(a) (section 3). For the classes C we obtain other properties, which generalize
a
known. properties of the class C
I • Another subdivision of this class is given
in section 4, where we consider the compound-negative-binomial lattice dis-tributions.
1. Preliminaries
We only consider lattice distributions (p) 0' i.e. distributions on the n n2
nonnegative integers, with PO > O. The corresponding probability generating
function (pgf) "" P(z) :=
L
n=O n p z , nis then unequal to zero on
Izl
< p for some p > O. If (p ) >0 (or P) isinfi-n infi-
n-nitely divisible (inf div), then P is unequal to zero on the closed unit disk
Izl
S I (cf. [4]).For an inf div pgf P we have the following representation (see [1]).
Theorem 1.1. A pgf P is inf div if and only if there is a A > 0 and a pgf Q), such that
(I. 1 ) P(z)
=
exp[A(QI(z) - l)J •Thus, the class of inf div distributions on the nonnegative integers coin-cides with the class of compound-Poisson distributions on the nonnegative integers. We recall the definition of such distributions.
Definition 1.2. A compound (lattice) distribution is a distribution with pgf P(z)
=
G(Q(z». where G and Q are pgf's. The corresponding random va-riable x can then be written as+ ••• + x , -n
where ~'~1'~2"" are independent, while n has pgf G and each of the ~i has pgf Q.
Example 1.3.
- p
i) _n has a geometric (p) distribution: G(z)
=
~--.-- t so aaompound-geome-- pz
tria distribution (p ) >0 has pgf
n n-(1.2) P ( z) = -:---~-i..,-~ 1 - P
- pQO(z) •
ii) n has a Poisson (A) distribution: G(z)
=
exp[A(z - J)J, so a aompound-Poisson distribution (p ) >0 has pgfn
n-(1. 3) P(z)
=
exp[A(Ql(Z) - I)J •Definition 1.4.
Co
is the set of all compound-geometric distributions.C
1 is the set of all compound-Poisson distributions, which, according to theorem 1.1, is equal to the set of all inf div distributions. If P is a pgf corres-ponding to a distribution in a class C, we shall also say, that P E C.
Remark. To some of the generating functions, corresponding to distributions in
Co
or CI, we add an index 0 or I, to be able to fit them in the more ge-neral notation of the following section. Further we denote the coefficient of zn of a generating function Ra(z) by rn(a), or rn, if no confusion is possible, while the sequence (rn(a»n~O is simply denoted by rn(a).We now formulate some more or less known theorems, which will be the start-ing point for our investigations in the next sections. For the sake of com-pleteness we write down the proofs in detail and in a few lemma's we give some properties of the appearing quantities. We start with the definition of absolute monotoniclty, and then consider the distributions (pn)n~O in
C) •
Definition 1.5. A function R on the complex numbers is called
absoZuteZy
mo-notone
(abs mon), if there are rn ~ 0 (n=
0,1,2, ••• ) and p > 0, such that00
R(z)
=
I
for Izi < P •n=O
Theorem 1.6. A pgf P is inf div (P €
C
I) if and only if p' (z)
RI(z) := P(z) is absolutely monotone. Proof. Let P €
C
I• Then there are A >
0
and pgf QI' such that log P(z)=
A(QI (z) - 1) •Hence:
R, (21) = p' (z)/P(z) = [log P(z)]' = ,AQi (z) , which is a abs mon function (izi < 1).
Conversely, let P'(z)/P(z)
=
R1(Z) be abs mon, i.e.00
P'(z)/P(z)
=
R1(z)=
for Iz 1 < p ,with r (1) ~ 0 for all nand p > O. Integrating from 0 to z (1211 < p), we
n obtain ~ rn (1) zn+1 log P(z)
=
log PO + L n=O n + 1 121 1 < p , 00 r (I) or, if we define A := -logPo
and Qt(z) :=*
L
nn+]n=O P(z) = exp[A(Ql(Z) - 1)], 1 Z I < p , with A > 0 and Q 1 abs mono n+I 21
Since P'(z) = P(z).RI(z), Izi < p, we have the following relations
(I.4) n = 0,1,2, ••• t
where r (1) ~ 0 for all n. It follows (see lemma ].7), that n
00 r (1)
\' n
L _ _ '!"<oo_
n=O n + 1
New Ql(z) is convergent for Izl S 1, so that exp[A(QI(z) - I)J is an analytic continuation of P(z) to the whole disk Izl S 1_ As P itself is analytic on
I z
I
S I, we haveP(z)
=
exp[A(Ql(Z) - )J, forI
z I S I •From P (1)
=
I we get now Qt(l)=
1 , so Q) is a pgf and P is a compound-Poisson pgf, i.e. PEe 1 •From Ql(l) = 1 we then obtain:
00 r (1) 00 r (1)
I
n 1 ,L
n It -logPo
0r
=
or:=
= •n=O n + 1 n=O n + 1
Remark. For the quantities
R),
It and Q} from the preceding proof we have the relations( I .5)
=
AQj (z), A=
-logPo =
z
J
R) (u)du •o
Some properties of the recursions (1.4) are now summarized in the following lemma.
Lemma 1.7. The sequence r (I), defined by (1.4), and its generating function
n
RI have the following properties
i) R) has a radius of convergence p > 0, and p' (z)
RI (z)
=
P(z) for Izi < p • ii) If rn(l) ~ 0 for all n, then(1.,6)
00 r (I)
\' n
L -~'r<oo, n=O n + )
so that RI(z) is convergent for Izi < I.
Conversely, if a sequence r (1) with r (I) ~ 0 for all n and satisfying (1.6)
n n
Proof. From (1.4) we obtain n por = {n + l)p I - \ ' p r n n+ L k n-k ' k=1 so that n Po
I
r nI
~ (n + 1 )p n+ 1 + \' L P k n-k 'I
r I k=l or Hence, if we define N 1~I(x) :=I
(x > 0, N E: E) : n=O N N n 2pol~1 (x) :s;l
(n+ I)p IX + nl l
Pk 1r _k1xn ~n=O n+ n=O k=O n
00 N N
I
nl
l
Irn_klxn~
~ (n+l)p IX + Pk n=O n+ k=O n=k N N k+ :s; p'(x) +l
P kI
Irn1x n ~ p'(x) + p(x)I~1 (x) • k=O n=ONow, if we choose a Xo > 0, such that P(x
O) < 2PO' then we have p' (x
O)
1~I(xo) :s; 2p _ P(x ) < 00 for all N ,
o
000
from which we can conclude that
L
Irnlx~ <00,
n=Oradius of convergence p ~ xo. Taking generating mediately have
P'(z)
=
P(Z)Rt{Z) •Let r ~ 0 for all n. Then we can write n
00
and R1(z)
=
L
r zn has an=O n
im-eo eo n ""
n~o ~
k~O
Pkrn- k=
~ ~ r n-k l. Pk l.-n ""
k=O n=k n 1 - Po "" l. \' P n+1 n=O eo co r eo r=
l
Pkl
n + ]n + k <!: POl n:-1 'k""O n=O n=O
so that 00 r - P
l
-.!!.. s
0 < co , n=O n+J Po eo d eo r I ( ) \' n zn+ t l.'t < I. As R z = -d l.~) z n=O n rand
I
n+i
n z n+l convergent for I zl n=Ofollows that R
1(z)
is
convergent for Izl < 1. Conversely, let rn <!: 0 for alleo r
n and a:=
L
n:-I < 00. Choose Po E (O,IJ and define Pn' n <!: 1 byn=O n (n + l)p 1 == \' P r n+ l. k n-k' k=O n=O,I,2, ••••
It follows that Pn <!: 0 for all n. Now define qn := Pn/PO' n ~ Ot then also
n=0,1,2, ••• ,
with qo
=
1. If ~ is the counting measure on {0,1,2, ••• } and fn(k) k <!: 0, n <!: 0, then fn(k) S fO(k) for all nandf
fO d~ = a <00,
sothe dominated-convergence theorem
r k : == '!'"k-+--1-+-n- , that by eo rk
I
limI
k + 1 + = lim f (k)d~(k) 1J:'7<X> k=O n ~ n lim f (k)d~(k) = 0 • nSo we can choose nO and aO < 1, such that
00 r k V
l
S a O • n<!:nO k=O k + 1 + n eoWe now estimate
L
qn as follows. For N > nO we have n=OAs a O < N+I
I
n=l q = n 1, we have N+lI
qn S n=t NI
qn+l=
n=O NI
n=O n + N "" r kI
qnI
k + 1 n=O k=O S a + n for N > nO n -1 0 00 aI
and soI
no
-1L
n=O 1 - aO n=O qn' n=O qn""
=: c < 00.
q S nSince qn
=
Pn/PO' it follows that c ~ I and n~o Pn=
POc, so that (pn)n~Obecomes a probability distribution, if we take PO == l/c.
0
Remark. The r n (1) may be expressed explicitly in the P : from the first n n equations of (1.4) by Cramer's rule we obtain
Po
0
PI PI Po ZPZ (I.7) r n-l (l) -n det 3P3 = Po P 2 p] PO. Po.
np Pn- l.
(I ~ ...
C PI nFrom theorem 1.6 and lemma 1.7 one easily derives the following characteriza-tion of inf div lattice distribucharacteriza-tions, first given by Katti [3J.
Theorem 1.8. A distribution (p ) >0 is inf div if and only if the quantities n n-r n (I), n == 0,1, ••• defined by n (n + 1)p I"
I
Pkrn_k(I) n+ k=O are nonnegative.Proof. Let (p ) >0 be inf dive According to lemma 1.7i) the generating func-n func-
n-tion Rt of the r
n(!) exists and is equal to P'(z)/P(z) for Izi < p. As this function is abs mon, by theorem 1.6 we have: r (1) ~ 0 for all n. Conversely,
n
let rn(l) ~ 0 for all n. Then (lemma 1.7ii» the generating function R) of the r n (I) exists and is equal to P'(z)/P(z) for Izl < 1, so that P'(z)/P(z) is abs mono The inf div of P follows now from theorem 1.6.
0
We now turn to the distributions in CO' and derive some analogous results.
Theorem 1.9. A pgf P is compound-geometric (P E CO) if and only if
P(z) - PO
RO (z):= zp (z)
is absolutely monotone.
Proof. Let P E CO' Then there are p E (0,1) and pgf QO' such that
I-p
P(z) - 1 - PQo(z) • Hence:
which is an abs mon function (Izl ~ 1).
Conversely, let RO be abs mon, i.e.
n r (O)z ,
n for I zl < p,
with rn(O) ~ 0 for all nand p > O. As P(z) - Po
=
zP(z)RO(Z)' we havePo
P (z)
=
-:----=--~ - zRO(z) ,I zl
< p 1or, if we define p
:=
I - Po and QO(z) :=p
zRO(z):_ I-p
with p € (0,1) and Q
O abs mono Since
P(Z) - Po
z
=
P(z).RO(z),I zl
< p ,we have the following relations
(I .8) n = 0,1,2, ••• ,
where r (0) 2 0 for all n, from which we can derive (see lemma 1.10), that
n
I
n=O
r (0) < 1 •
n
Now RO(z), and hence QO(z), is convergent for Izi ~ 1, and, as
I
n=O
r (0) < 1,
n I ,
1 -
t .
~--~~ __ ~s an analytic continuation of P(z) to the closed unit disk
1 - pQo z)
Izi ~ 1. As P itself is analytic for Izi ~ 1, we have
From P (1) I-p P (z ) = -:---"!!"-~""-': - pQO(z) , = 1 we now get Q O (1) metric pgf: P € CO' From Q
O (I) = I we then obtain:
for
I
zI
~ 1 • = 1 , so Q O is a pgf 00~
RO(!)=
I, or:I
n=O and P is a compound-geo-r (0) = p = I - po· nRemark. The following relations now hold for the quantities R
O' p and QO
from the preceding proof
or if we take qo =
o ,
(1.9)
00
D
RO(z)
=
~
PQO(z), p = 1 - Po=
I
rn(O) =RO(!)' Qo(z)=t
zRa(z) • n=OLemma 1.10. The sequence r (0), defined by (1.8), and its generating
func-n
tion RO have the following properties
i) RO has a radius of convergence p > 0, and
(1.10)
p(z) - PO RO(z) = zP(z)
ii) If r (0) ~ 0 for all n, then
n (l.lI)
I
n=O r (0) < 1 , n for 1 z 1 < p •so that RO(z) is convergent for
Izl
~ I.Conversely, if a sequence r n (0) with r n (0) ~ 0 for all n and satisfying
(1.11)
is
given, then (1.8) defines a unique probability distribution (pn)n~O.Proof. Exactly as in the proof of lemma 1.7 we can show, that RO has a radius of convergence p > 0 (replace the factor n+ t by 1). Then, taking generating functions, from (1.8) we immediately obtain
Let r (0) ~
o
for all n. Then we can write nco co n co 00 co
1 - Po =
I
Pn+l =I
I
Pkr -k (0) =I
PkI
r (0) =L
r (0)n=O n=O k=O n k=O n=O n n=O n
so that 00
I
n=O 00 r (0)=
n 1 - Po < I ,and RO(z)
=
I
r (O)zn is convergent for n=O nConversely, let r ~ 0 for all n and a
:=
n1 z 1 ~ J. 00
I
rn < 1. Choose Po € (O,lJ,de-n=O
,
fine P
n' n ~ I by (1.8) (then Pn ~ 0 then also
for all n) and define qn := PnlpO' n ~ 0,
We can write now 00
I
qn+1=
n=O Hence, as a < I, 00I
qn+1 = n=O 00 00I
qkL
k=O n=k 00 r n-k=
a(1 + I~
a ,orL
qn=
~---a-
=: c < 00 • n=O 00 00Since qn
=
Pn/PO' it follows that c ~ 1 andL
n=O a probability distribution, if we take PO=
~•
P n
=
POc, so (p ) >0 becomes nn-o
Remark. The determinant representation (cf. (I. 7) for the reO) now becomes
n Po
0
PI PI Po P2 (1.12) r n- 1 (0) = Po -n det P2 PI Po P3.
• Po Pn- 1 PI PnAs an analogue to theorem 1.8 we now have from theorem 1.9 and lemma 1.10 (see also [4J):
Theorem 1.11. A distribution (p ) >0 ~s compound-geometric if and only if n
n-the quantities r (0), n n
=
0,1, ••• defined byare nonnegative.
Two other classes of lattice distributions, which are known to be inf div, are the log-convex and the completely monotone distributions. For the latter we give a useful characterization in theorem 1.14 (see [4J).
Definition 1.12. A distribution (p ) >0 1S called log-convex if
n
n-n ~ I •
Definition 1.13. A distribution (p ) >0 is called completely monotone if
n
n-n ~ 0 and k ~ 0 •
A
is the set of all completely monotone lattice distributions.Theorem 1.14. (p ) >0 is completely monotone if and only if there is a finite n
n-measure ~ on [0,1) such that
or, equivalently, (pn)n~O is a mixture of geometric distributions: 1
Pn
=
I
(I - p)p dF(p) , no
with F is a distribution function on [O,IJ.
We conclude this section with a theorem concerning the relation between the classes
A,
B, Co
andC}
of (inf div) distributions (see [2J, [4J and [5J).Theorem 1.15.
A
cB
cCo
cC
I, or: the following implications hold:
(pn)n~O completely monotone ~
- (p) n n~O compoundgeometric
-Proof.
A
cB:
Let (pn)~O be completely monotone. Then (theorem 1.14) 1I
(1 - p)p dF(p), n n ~a ,
a
with F a distribution function on [O,IJ. If ~ is a random variable with dis-tribution function F, x := (I
-~) ~~Hn+l)
and1..
:= (I-~) ~~Hn-1),
then,us~ng Schwarz's inequality,we have
so that (p ) >0 is log-convex. n
n-B c CO: Let (p ) >0 be log-convex. If PI = 0, then it follows by induction
2 n n_
from Pn ~ Pn+)Pn-I' that Pn =
a
for all n ~ I, which is a trivial case. So PI > 0, but then it follows by induction from P~
p2 IIp 2' that p >a
n n- n- n
for all n. Now we can write P > k >
---
Pk-
t k=l, •••,n,
and so
(1.13) k= I, •••
,n.
We want to use theorem 1.11 and prove the nonnegativity of the r (0) by in-n duction from (1.13): rO(O) = P1/pO>0. Let rk(O) ~O for k~n-l, then it follows
P n par n (0)
so r (0) ~ 0, and (theorem 1.11): (p ) >0 € CO.
n n
n-Co c
C
I: We can simply prove this by showing that a compound-geometric distri-bution is also compound-Poisson.
I-p
If P(z) = I _ PQO(z) €
CO'
then(1.14) A := -loge) -p) > 0, Q) (z) :=
log(I - pQo (z» 1 ""
logO - p) =
!
n~1
n
(pQO(Z» nis abs mon, with QI(I) = I, while
An other method uses the theorems 1.6 and 1.9: If P € CO' then
P(z)
with RO abs mon, from which we derive
Rl(Z)
=
P'(z)/P(z)=
[log P(z)]'=
[-log(1 - zRO(z»]'=
[zRO(z)] , I=
1 - zRO(z)=
p;
P(z)[zRO(z)]' ,which is an abs mon function, so: P € C I' 0
Remark. The relations between R
t and RO' derived in the preceding proof, (1.15) [ZRO(z)J'
=
1 - zRO(Z) t and (1.16) RI(Z)=
P(z)[ZR (z)]t POa
'
give by equating the coefficients of sequences r (I) and r (0) ([zRO(z)]'
n n
zn the following relations between the
00
=
I
(n + l)r (O)zn) n=O n (1.17) and (1.18) r (1) = n (n + l)r n (0) + 1 n r (I) -I
(k + l)rk(O)P n_k • nPo
k=OFrom both relations we immediately obtain: If r (0) ~ 0 for all n, then n
(1.19) r (I) ~ (n + I)r (0),
n n n = 0,1,2, ••••
In view of the theorems 1.8 and 1.11 we can formulate the following
Corollary 1.16. If, for a given distribution (p ) >0' the r (0), n ~ 0,
de-n de-n- n
fined by
n
Pn+1
=
I
Pkr _k(O),are all nonnegative, then the r (I), n ~ 0, defined by
n
(n + I)Pn+1 = n
=
0,1,2, •••are all nonnegative, too.
2. Interpolating between 1 and n + I
In the preceding section we have given characterizations of the distributions
(pn)n~O in
Co
andC]
by the nonnegativity of the recursively defined quanti-ties r (0) and r (1) (n n n ~ 0) (see theorem 1.8 and 1.11). This suggests the possibility of dividing the class of distributionsC)\C
O (cf. corollary 1.16) in a number of increasing subclasses of distributions, characterized in a similar way.
Let us introduce therefore sequences (cn(a»n~O ' defined for a € [O,IJ, with
the following properties
(2.1)
c n (0)
=
1 and c n (1)=
n + 1 for all n ~ 0 ,c (a) is nondecreasing 1n n for each a, and in a for each n • n
For such a sequence cn(a), and a lattice distribution (pn)n~O' define the sequence (r (a» n n_ >0 by
(2.2) n
=
0,1,2, ••• ,with generating function
R (z) :
=
a 00r
n=O n r n (a)z •For a € [O,)J define the following class of distributions
v
>0 r (a) ~ O} ,n- n
or, in terms of generating functions,
C
a:=
{pgf PI
R is abs mon} • aOf course, for a
=
0 we get the recursion(1.8)
with the rn(O), RO andCo
from section 1, and for a
= ]
the recursion (1.4) with the rn(l), R) and C1 from that section.Examples of c (a) we can obtain by an obvious interpolation between c (0) :::: 1
n n
and c (1) :::: n+ I, or, if C (z) is the generating function of the sequence
n a -I -2
cn(a), between CO(z) = (I - z) and C
I (z) :::: (I - z)
Example 2.1. (the properties (2.1) are easily seen to hold)
i) C (z) a -I -I 00 kn H) C (z) a (l - z) (I - az)
:::: I
a z , so m=O 2 n n+l c (a) = 1 +a +a + ••• +a :::: - a n -a 00 00 00 niii) C (z) :::: (1 - z) -) (1 - az) -a =
I
z mI
( )(-az) = -a kI
I
( k ) -a (-a) k n z ,a
m=O k=O k ncO k=O
n
Ca)(_a)k n (a+~-I )ak c (a) =
I
::::I
n
k=O k k=O
iv)
c
(z) = (l-z) [J-(I-a)zJ::::(l-z) -2 - 1 - 2 +az(l-z)=
a 00 coI
nI
(n + I) z n so c (a) 1 + an •=
z + az,
:::: n n=O ncO v) cn(a) = (n + 1 )a.
vi) c (a) = ( 1 + an) a
.
n n k vii) c (a) =I
a k a n k=O ,n n with a k ,n ~ 0, aO ,n :::: 1 andI
k=1 a=
n. k,nIn the following lemma we summarize the main properties of the sequence rn(a) and its generating function R (cf. lemma 1.7 and 1.10).
a
Lemma 2.2. For every a € [O,lJ and a fixed sequence c (a) we have
n
i) Ra has a radius of convergence p > 0, and
(2.3)
(2.4)
0>
R (z)
=
p(z)-lI
c (a)p Izn,a n=O n n+ for
I
zI
< P •so that R a (z) is convergent for Izi < 1. If, furthermore, lim c (a) =: c(a) < 0>, then
n n-+<>o (2.5) iii)
L
n=O r (a) < c(a) , nso that Ra(z) is convergent for
Izi
~ I.Conversely, every sequence r (a) with r (a) ~ 0 for all n, and
satisfy-n n
ing (2.4), or, if lim c (a) =: c(a) < 0>, (2.5), by (2.2) defines a
pro-n
n-+<>o
bability distribution (pn)n~O € Ca'
Proof. We can prove i) and the first part of ii) in the same way as in lemma 1.7 (with the factor n+ 1 replaced by c (a». For the second part of ii) we
n
can write, since c (a) is nondecreasing in n,
n
so
1 - PO
=
L
Pn+1=
n=O0> n 0> 0>
I
c (a)L
Pkrn_k(a)~ ~
L
PkL
rn(a) ,n=O n k=O c\aJ k=O n=O
L
rn (a) ~ (I - Po)c(a) < c(a) • n=OFollowing again the proof of lemma 1.7, we see (using again the monotonicity of c (a», that for iii) it is sufficient to prove
n
(2.6)
0> rk(a)
lim
L
< I •n-+<>o k=O ck+n(a)
"" rk(lX) 00 rk(lX) lim
I
C k (lX) .,.I
lim Ck+n(lX) =o ,
n-roo k=O +n k=O yt+<X>
so that (2.6) holds. I f lim C (lX)
n
=
C(lX) < 00,
then yt+<X> 00 rk(lX) 00 rk(lX) limI
Ck+n(lX)=
I
c(lX),
n-+<» k=O k=Oand this is less than I, i f (2.5) holds.
One would like to have, for all lX and 8 € [0,1]
(2.7)
We now investigate for sequences C (lX), satisfying (2.1), to what extend n
this implication holds.
Lemma 2.3. For all n ~ 0 and all lX,S € [O,lJ
(2.8) Proof. Define A (z) := lX 00 (convergent for JzJ < I) ,
then it follows from lemma 2.2i): AlX(z)
=
P(z)RlX(z) and AS(z)=
P(Z)RS(Z), from whichEquating the coefficients of zn, we obtain (2.8).
Theorem 2.4. For all lX € [O,lJ
This is an immediate consequence of the following: If r (0) ~ 0 for all n, then
n
i) For all a € [O,IJ and all n
(2.9)
o
ii)If
(2.10)
then
and
cn+1(a)
c (a) is nondecreasing in a for all n ,
n
r (a) n
c (a) is nondecreasing in a for all n ,
n
r (a) is nondecreasing in a for all n • n
Proof. Let r (0) ~
a
for all n (or: (p ) >0 € CO), According to (2.8) (withn n
n-S
=
0, and hence ckCS)=
I) we can writen-l
=
cn (a)Pn+1 -L
Pk+l rn_1_k(a)=
k=Oso that
po[r (a) - c (a)r n n n (0)]
=
=
which is nonnegative, as rn(O) ~ 0 for all nand cn(a) is nondecreasing in n. We conclude
r (a) ~ c (a)r (0),
n n n n
=
0,1,2,.,. ,so that all r (a) are nonnegative, and (p ) >0 €
C
N ,n ( ) n n- u
C i a n+
Furthermore, let c (a) be nondecreasing in a for all n. Then we have for
n
S > a and k ::;; n
cnca) n-l ct+1(S) n-I c t +1 (a) ck (a) =
R.~k
cg, (13) ;:::.t~k
cg, (a) orc (a) n
(2. II ) for a < S, k
=
O,I, ••• ,n • From (2.2) and (2.8) we obtain for a < Spoe c (a)r n n (S) - c n (S)r n (a)]
=
which is nonnegative on account of (2.11). It follows that c (a)r (S) ~ c (S)r (a),
n n n n for S > a and all n ,
r (a)
n
which implies that c (a) is nondecreasing in a for all n. Consequently
n c (8) r (8) ~ n() r (a) ~ r (a), n c a n n n for S > a ,
because c (a) is nondecreasing in a. So we finally have, that r (a) is
non-n n
decreasing in a for all n.
0
Remark. One easily verifies, that condition (2.10) is satisfied in the follow-ing cases n+J c (a)
=
(a+n) I - a n n ' I-a a , I + an and (n + I) •We now consider distributions (p ) >0 in
C
with a > O. It turns out that forn n- a
many choices of c (a) we do not have the desired property (2.7). We shall give n some simple, necessary conditions for (2.7), from which we obtain counter-examples for almost all our choices of the sequence cn(a). To this end in the following lemma we introduce a special distribution.
Lemma 2.5. For a
O E (0,1), the for all n ~ I by (2.2) defines which
(2. 12)
sequence rn(a
O) with rO(aO)
= )
and rn(a O) = 0 a probability distribution(p )
'0 EC
t forn n", a
O
Proof. The sequence rn(a
o)
satisfies the conditions of lemma 2.2iii) and hence defines a distribution (~p) € C by the recursion, from which wen n~O 0.
0 obtain: cn(aO)P
n+1
=
Pn ' and from this (2.12) immediately follows.0
Lemma 2.6.i) If there is a 0.
0 € (0,1) such that
(2.13)
then for r 2(a) corresponding to (pn)n~O we have: r 2(1) < O. ii) If there is a 0.
0 € (0,1) such that
(2.14)
then for r 2(a) corresponding to (pn)n~O we have: r 2(0.0) = 0 and ri(a
o)
< 0 (here c1 and c2 are supposed to be differentiable).So, in both cases we do not have (2.7).
Proof. For the rn(a) corresponding to
(p )
>0 we have on account of (2.12)n
n-from which we successively obtain
I
= '( )
c 1(0.) - I , c] 0.It follows, that, if (2.13) holds, then
and consequently, (pn)~O in C
ao' but not in Ct' If (2.14) holds, then
whence (pn)n~O in C
ao' but not in
e
a for a E (aO,aO + e), with £ > 0suffi-ciently small.
0
In the following lemma, for the different choices of cn(a) (cf. lemma 2.1) we list values of aO' for which (2.13) and (2.14) hold.
Lemma 2.7.
i) c (a)
n
=
(a:n): satisfies (2.13) for aO >13 -
1 (~0.732), and (2.14) for aO >H15-
1) (~0.618).n+l - a c (a) =
-"":""---n I-a satisfies neither (2.13) nor (2.14).
ii)
iii) c (a)
=
nn
I
(a+~-l)ak:
satisfies (2.13) and (2.t4) for aO > 1 - £, withk=O
£ positive and sufficiently small.
iv) c (a)
=
) + an:n satisfies (2.13) and (2.14) for aO > ~
.
cn(a)
=
(u'+ 1) : a satisfies (2.13) for a 1 - £ (£ positive andsuffi-O >
2 2 ciently small) and (2.14) for a
O > log log 3.
v)
c n (a)
=
( 1 + an)a: satisfies (2.14) for aO > 1 - £ (e: positive andsuf-vi)
ficiently small) •
We further have (cf. example 2.1ii) and vii»:
Lemma 2.8. If c
J is linear and c2 quadratic in at then in order to exclude (2.13) and (2.14) for all a
O E (0,1) it is necessary and sufficient that for all a E (0,1)
Proof. As c (0) = I and C (I)
n n n + I, cl and c2 already have the following form
with b
1
o.
Further, from the requirement cI(a) ~ c2(a) for all a E (0,1) weI
obtainib ~ I _ a for all a € (0,), so: b ~ I.
Now suppose, that (2.13) does not hold for all a
O E (0,1). Then it follows
> 2a - 1
that b - a(2 _ a) for all a E (0,1), which implies: b ~ I. So we obtain (2.15)
as the only possibility for c
1 and cZf which gives no counter-examples
accord-ing to lemma 2.7ii).
0
a
It is a pity, that cn(a) = (n + I)a and cn(a)
=
(a:n)(~
r{an+ 1) , n~
00) do not have the required properties, because they seem to give a "better" sub-division ofC]\C
O' than e.g.c (a) = I + a + ••• + an
=
n
n+I - a
I - a. ('" I _ a. ' n ~ 00) •
However, since the latter yields no counter-examples and is the most obvious in view of lemma 2.8, we shall consider it in more detail in the next section. If we drop the monotonicity of cOCa), then we can consider sequences cn(a)
a 1 n+l-a. -1
like (n + a) , and r(I _ a) ( n+l) • The latter we obtain by means of
"frac-tional differentiation": a-I
=
z from which 00 \' r(n + I) l. r(n - a + ]) n=I n-a z = 00 \' (n + I)! n l. f(n + 2 - a) Z , n=O (n + I)! cnCa.)=
f(n + 2 - a) = n+l-a -Ir
(I - a) ( n+ I )We shall not persue this in this report.
We conclude this section with a property of
C ,
which holds for every choice aTheorem 2.9. If (p) €
C
t then also (q(Y» €C
for all Y € [O.IJ,n n~O a n n~O a •
with
n
qn (y) .-.-
PTYT
Y P n' n=
0,1,2, ••••Proof. Let (p ) >0 € C t so rn(a)
~
0 for all n. Define the sequence r(Y)(a)n n- a n by n c (a)q(Y)
=
n n+lL
q (Y)rCY)(a) k n-k t n = 0,1,2, •••• k=O On the other side we can writen+1 n+l n
c (a)q(Y) - Y c (a)p Y
~
r (a)=
n n+1 -
PTYJ
n n+l=
PTYJ
L Pk n-k so that r(Y)(a)=
n k=O=
n \' (y) n-k+l ( ) L qk Y r n-k a k=O n+l ( ) 0 Y r n a ~ for all n •It follows by definition, that (qCY» €
C •
n n~O a
3. Case c (a) n
=
I + a + a 2 + ••• + a nIn this section we investigate the recursion relations (2.2) with n+l
- a
1 - a n
=
0,1,2, ••••o
So, the quantities r (a), R (z) and the class C , defined in section 2, now n a a correspond to this choice of c (a).
n
First, we formulate lemma 2.2 more precisely.
Lemma 3.1. For all a € [0,1) we have
i) Ra has a radius of convergence p > 0, and
(3.1) R (z) = 1 I
a - a z { 1
P (az) }
00
(3.2)
I
n=O
r (0.) <
~--n - 0. •
so that Ro.(Z) is convergent for Izi S I.
iii) Conversely, every sequence r (a) with r (0.) <:: 0 for all n, and
satisfy-n n
ing (3.2). by (2.2) defines a unique probability distribution. (p ) "'0 E C • n n", 0.
Proof. For Iz\ < 1 we can write
00 00 \' n 1 \' n t.. c (o.)p IZ
=
1 {t.. Pn+lz n=O n n+ - 0. n=O 00 \' n+l n} t.. 0. Pn+)z == n=O P(Z) - Po p(o.z) - P = "':"'-_ { O}=
- 0. z Z= "':"'--
- 0. Z {P(z) - P(o.z)}Now (3.1) immediately follows from (2.3). As lim c (0.)
=
~---,
the rest isn 1 - 0.
a reformulation of lemma 2.2. n-+<x>
o
Remark. In lemma 3.1 we only consider 0. < t. The case 0.
=
I, which isessen-tially different, has been treated in lemma 1.7. It is a limiting case in the following sense: lim R (z) == o.tl 0. lim ":'1---0.- z o.t) p' (z)
=
~P~("'zp){I _ P(az)}
= )
lim p(z) - P(o.z)=
P(z)
PTZr
at) z - o.zSince (3.1) gives an explicit expression of R in P, we can also prove
theo-0.
rem 2.4 using generating functions.
Lemma 3.2. For all 0. € [0,1) we have
(3.3) and
(3.4)
RO(Z) - o.RO(o.z) Ro. (z)
=
"'!"')---a 1 - o.zRO (o.z)R (z) 0.
p(o.z) RO(z) - o.RO(o.z)
= ----
or, in terms of the coefficients, (3.5) and (3.6) r (a) n n-I k+l = c (a)r (0) +
I
a rk(0)rn_1_k(a) n n k=O I n n-k r (a)= --
l
ck(a)rk(O)a P-k n Po k=O n 1 Po PoProof. As RO(z)
=
z
{I -PTZJ}'
we have P(z) = I _ ZRO(Z) • Substituting thisin (3.1) we obtain (3.3):
I 1 - ZRO(Z) RO(z) - aRO(az)
R (z) =~--{I - }
=~-a - a z 1 - azRO (az) - (l 1 - azRO (az)
from which, using (3.1) once more, (3.4) follows. The relations (3.5) and
(3.6) now follow from (3.3) and (3.4) by equating the coefficients of zn.
0
From this lemma we immediately obtain the abs mon of Ra from that of RO' and so: Co c C
a, while the inequality (2.9) follows from (3.5) or (3.6). Further-more, letting a
+
1 in lemma 3.2, we get the relations (1.15), ••• ,(1.18).The following lemma expresses P in terms of R • a
Lemma 3.3. For all a € [0,1) we have (with the p from lemma 3.1)
(3.7) I f P E: (3.8) 00 p(z) = Po
n
k=OC , then (3.7) holds for Izi ::;; 1 and we a 00 1 - (I -
(l)a~
(ak) P(z) =n
a k k,
k=O 1 - (I - a)a zR (a z) a for Izi can write for Izi ::;; 1 •Proof. According to (3.1) for Izl < p we can write
P(z)
=
[I - (1 - a)zR (z)]-IP(az) , afrom which we obtain for every n € E
P(z) or
n-l IT k=O k k [J - (J - a)a zR (a z)J a
Po
This tends to PTZ) , for n +
00,
from which (3.7) follows. If P E Ca' thenon account of lemma 3.1 the derivation above is valid for
Izi
$ I. Takingz == I in (3.7) we get
k k
p == IT [1- (1 - a)a R (a )J ,
°
k=O aand (3.8) follows from (3.7).
o
Lemma 3.3 gives us a characterization of C • For a =
°
and 1 we already have aone: Co is the set of compound-geometric distributions and C) the set of com-pound-Poisson, or inf div distributions. C now appears here as the rather
a
special set of infinite products of compound-geometric pgf's, given by (3.8).
Condensing the notation a little, we have:
Lemma 3.4. For a E (0,1): P E C if and only if there is p E (0,1) and pgf
a
Q with Q(O)
=
0, such that(3.9) P(z)
=
00 k
IT 1 - pg(a ) k •
k=O 1 - pQ(a z)
Proof. Let P E C • Then P has the
a zR (z)
representation (3.8), which becomes (3.9), a
if we define Q(z) := R (1) and pgf, while p € (0,1) onaaccount
p := (1 - a)R (I). As R is abs mon, Q is a
a a
of (3.2).
Conversely, if P has the representation (3.9), then we have for the corres-ponding R a R (z) == ~~-a - a z = " ' ! " ' - -
1..
{I - (l - pQ (z»} = - a z ()O k IT 1 - pQ(a z) }=
k+l k=O 1 - pQ(a z) p Q(z) - a z which, as Q(O)=
0, is an abs mon function. So: P € C •a
o
Theorem 3.5. For all a € [O,lJ we have: C
a C C10 Or: every pgf P with an ab-solutely monotone R is inf dive
Proof. Let P E Ca with a € (0,1). Then P has the representation (3.9), or (3.10) co P(z)
=
IT k=O I - 'lJ' k - 'lJ' 8 (z) , k k k k kwith 'lJ'k
:=
pQ(a ) E (0,1) and 8k(z) := Q(a z)!Q(a ) a pgf. 80, P is an
infi-nite product of compound-geometric pgf's, which is inf dive
o
Remark. We can also prove theorem 3.5 using the following relation between RI and Ra: from (3.7) we obtain
so (3.11)
00
P'(z) d d k k
R1 (z) = P{z)
= -
log P(z) =- - log IT [I - (l -a.)a. zR (a z)J =dz dz k~O a
co
= -
I
.Ld loge I - (1 - a.)a.kzR (a.kz) J =I
k=O z a. k=O co [zR (akz)J' R1 (z) = (l -a)
I
ak----a.-.,k~-"';"k-
• k=O 1 - (l - a)a zR (a z) a (l-a.)a.k[zR (a.kz)J' a. k k ' I - (I - a)a zR (a z) aNow, if Ra is abs mon, then it follows from (3.11), that R] is abs mon, too.
Before we investigate, whether the implication (3.12)
holds generally, we prove, that the special distribution (pn)n~O €
C
ao (see lemma 2.5), from which we obtained counter-examples for many choices of cn(a), belongs to every
C
e
withe
~ a.O'
Lemma 3.6. For the r (a) corresponding to the distribution
(p )
>0 € C fromn n n- a.O
lemma 2.5 we have
(3.13) n
=
O,t, ••• , and a. € [O,t) •From this it follows that (3.14)
and (3.15)
00
'"
Proof. On account of (3.7) for p(z)
:=
I
p zn we have, as R(z)-n=O n aa
,
from which for a E [a,l) we obtain
' " 00 1 - (I -a )akz
R", (z) == ~_1.. {I _ P (a z)}
=
-:-~1..
{I _ II 0 0 }"'" - a z 'i?(z) - a z k=a 1 - (1 -
aa)a~a1!
Hence, for m ~ J
1..u
m-l k R m(z) ==-
II [I - (I-
tla)aOz] },
m z k==O a O - a 0which is a polynomial in z of degree (m - l) t so that rn(aa) m
=
o
for n ~ m, or(3. t 6)
However, with (2.12),
+. •• ,
which is a polynomial in a of degree n. Hence, (3.13) immediately follows from (3.16). From (3.13) we easily obtain (3.14) and prove (3.15) by proving the equivalent inequalities
(3.17) c (a) n II (8 - aa) k ~
n k=1
using mathematical induction. For n == I:
n k
c (8) II (a - a O)' n k==l
Suppose that (3.17) holds for a fixed n, then
n+l k n+1 k n+l k c+l(a) IT (13 - 0. 0)
=
C (a) IT (13 - 0.0) + IT a(S - 0.0) ~ n k=l n k=1 k=1 n k n+1 k ~ (S - an+l)c (13) IT (a - 0. 0) + IT 13 (a - 0.0) ~a
n k=J k=t n+1 k n+J k ~ c (13) IT (a - aO) + Sn+l IT (a - 0. 0) = n k=1 k=lThe result of the preceding lemma gives uS hope, that (3.12) does indeed hold. Therefore, we look for a convenient expression for RI3 in terms of Ra.
Lemma 3.7. For a,S € [0,1) the following relations between Ra and RS hold
(3.18) and (3.19) I - (1 - S)zRS(z) 1 - (I - a)zR (z) a 1 - (1 - S)azRS(az)
=
1 - (I - a)l3zR (Sz) a RQ (z) ==""1-~
1..
{I - IT k k 1 - (1 - a)a zR (a z) k a k } 00 p S z k=O 1 - (1 _ 0.)0. SzR (a Sz) a P(az) Proof. According to (3.1) we have: J - (I - a)zR (z) = ~~- , soa P(z)
[I - (I - a)zRa(z)][1 - (I - S)azRS(az)]
=
P(az) P(z) • P(az) = P(Saz)_ P(Sz)
- P(z) • P(Sz) -P(aSz) - [I - (I - S)zRQ(z)][1 - (I - a)l3 zRro(Sz)]
p '"
which gives (3.18). By iteration of (3.18) we obtain for every n ~ 1
o
n-l 1 - (I -a)akzR (akz)
a n n 1 - (I -13)zR
S(Z) == k=O IT - - - -... - (I - a)a SzR (a I3z) k--=.:..-k-.[ I - (I - S)a ZRS(a z)] • a
Since the last factor tends to 1, if n +
00,
(3.19) follows.o
Though (3.19) expresses RS proof of the abs mon of RS' fore, we return to relation
explicitly in R , it does not provide an easy a
if we know that R is abs man and i f S ~ a. There-a
Lemma 3.8. For a,B ~ [0,1) we have (3.20)
or, in terms of the coefficients, (3.21)
with
(3.22) o (a,B)
:=
~ l (B k+1 - a n-k+1 )rk(a)rn_k(S).n k=O
Proof. Writing out (3.18) we obtain (3.20), from which, by equating the ff ' . f n+1 (3 21)
coe ~Clents 0 Z t • follows.
Lemma 3.9. If P £
C ,
then for all n ~ 1 a(3.23) C (a)r (B) ~ C (S)r (a),
n n n n
lIn
for S £ [a , I J •
Proof. Let P £
C •
According to (3.21), (3.23) is equivalent toa
(3.24) o n-l(a,S) ~ 0 for
B
~ [a IIn
,IJ.o
I ( ) ( ) () () Of S S [ ",lIn+I,IJ
For n
= :
00 a,a=
a - a rO a rO 8 ~ or B ~ a. uppose £ ~and 0k(a,B) ~
a
for k = 1,2, ••• ,n-l. Then it follows from (3.21), that rk(B) ~ 0 for k=
1,2, ••• ,n, so that with (3.22)n+1 n
o (a,a) ~ (a - a)
I
rk(a)r _k(8) ~ 0 •n k=O n
Thus, (3.24) has been proved by induction.
Remark. Using (3.21) and (3.22), we can prove something more than (3.23), namely
c (a)r (8) ~ c (S)r (a), for 8 ~ a, if n
=
0,1, ••• ,5,n n n n a 3/n-2
for ~ ~ a , if n ~ 6 •
IJ
In the following lemma we formulate some relations, from which another par-tial result follows.
Lemma 3.10.
i) If a - Sy with a,y € [O,IJ, then
(3.25)
ii) For all 13 € [O,IJ
(3.26) iii) For (3.27) iv) IfP (3.28) (1 + S)R 2(z)
=
Ra(Z) + SRa(Sz) - 13(1 - 8)zR S(Z)RS(Sz) • 13all 13 € [0, I ] and all n ~ 0
2 S n k
oneS ,13)
=
+ 13 (I - an+2)2
S rk(S)rn_k(S) •k=O
€ C 2 with 8 € [O,IJ, then
S
2 2
c (13 )r (a) ~ c (8)r (13), n == 0,1,2, ••••
n n n n
v) For aIlS € [O,lJ C 2 c Cst
a
Proof. If a - Sy, then, on account of (3.19), we can write
== k+1 k+1 00 1 - (1 - a) a zR ( a z) == [1 - (I -a)zR (z)J II a == a k-O 1 - (1 - a)akSzR (ak8z) a
=
[ l -k k 00 I - (1 - a)a 8zR (a 8z) I (1 - a) zR (z) J { II k a k } - ==a k-O 1- (1 -a)a y8zR (a ySz)
a
- [I - {I -a)zR {z)J[l -a (1 -y)azR {Sz)J-1 ,
y
from which (3.25) follows. Taking y
=
a, and mUltiplying out, we obtain (3.26),Hence. with (3.21), I - Sn+2 2 n k
=
{(1 +8)r 1(8) +8(1-8)I
8 rk(S)rn_kCS)} -1 - -13 2 n+ k=O 1 - on+2 2 0 2 n k .., (0 ) .., (1 _ on+ ) \' 0 (0) (0) r n+J IJ= T"'+i3
..,
l.. I-' r k .., rn- k .., 1 - 13 k=Owhich is (3.27). I f furthermore. P €: C 2' then all r (8 2) are nonnegative
a
nand from (3.21) and (3.27) by induction we obtain for n
=
0,1.2 •••• ,which gives (3.28). From this we see, that all r (8) are nonnegative, too,
n
Corollary
3.11.
For all a €:[O,IJ
we haveC
cC
1 cC
1 c ••• cC
c ••• cC
1 •
a a2 a4 2-m
a
D
We have now obtained in a rather laborious way a few results (theorem 3.5, lemma 3.9 and corollary 3.11), which indicate, that (3.12) is true. Indeed, we can prove (3.12) by a method, inspired by still another proof of
IIC
acCI",
which follows from the next theorem and lemma.
Theorem 3.12.
1.' ) If P €
CIt
then for all a €[O,IJ
P(z) P(az) is abs mon and P(az) P(a)P(z) €Cl
o ii) For all a € [0,1) we have the following characterization ofC :
a (3.29) P €
C
a ~ P(a)P(z) P(az) €C
O'Proof. If P €
CI'
then P has the representation (1.1), from which it followsthat
(3.30)
p~(!;fz) =
exp[A(Q(Z) - Q(az) + Q(a) - I)] •Now let P € C • First we see from a
(3.31) P(a)P(z) P(az) :: pea)
1 - (1 - a)zR (z) , a
that P(a)P(z) . P(az) 1.S a pg. f If we e 1ne p:= d f' 1 - P() a an d Q( ) z:= j _ 1 -pea) a Z R ( ) a Z , then it follows that p € [0,1) and Q is a pgf (as R (1) == 1 -
pea»~.
Hencea I-a
P(a)P(z) I - ~
P(az)
=
1 - pQ{z) €Co •
However, this and its converse, i.e. (3.29), we can immediately prove using h f P(a)P(z) the following relation between the RO corresponding to t e pg P(az) (de-noted by R6a
»
and Ra:(3.32) RO (a) (z) ==
z
I {I - P(a)P(z)JP(az)} pea)=
z
1 {I - P(az) P(z) }=
( I - a Ra ) () Z •Thus R(a) is abs mon if and only if R is abs mon, which is, by definition,
, a
a equivalent to (3.29).Lemma 3.13. The following relation between R) and Ra'holds
(3.33) or (3.34) [zR (z)]' a
=
I - (1 - a)zR (z) • a In terms of the coefficients(3.35) c (a)r (1)
n n
Proof. On account of (3.1) we can write
[ P(az)], P(az)P'(z) P'(az)
(l -a)[zRa(Z)]' == - P(z) == p(z)2 - a P(z) ==
== P(az) {R (z) - aRI(az)} ,
pez) I
o
from which (3.33) follows. Using (3.1) once more we obtain (3.34), from which, equating the coefficients of zn, we get (3.35) by using
(3.36) and (3.36)'
~
a {R1(Z) - aR 1(az)}=
L
n-O n c (a)r (1) z , n n [zR (Z)JI ... aL
n=O (n+ l)r (a)zn • nCorollary 3.14. For all a € [O,IJ we have C
a c CI. Furthermore, if P € Ca, then
(3.37)
o
Proof. Let P € Ca. Then Ra is abs mon, so that the abs mon of Rl follows from
(3.33) and (3.29), or immediately from (3.34). By induction we obtain the
in-equality (3.37) from (3.35).
o
We now try to find a relation between Ra and RS similar to the relation (3.33) between Ra and R). On the one hand we can write for (J - a)[zR
a (z) ] t the limit
zR(zl - SzR.(Sz)
(I -a)lim a a ... lim
~
:~
{Ra(z) - SRa(SZ)} ,st) z - SZ Stl
and on the other hand, according to (3.33), P(az) {R (Z) - aR
1(az)} ... lim P(az) {RQ(z) - aRQ(az)} •
P(z) 1 Stt P(Bz) ~ ~
Equating the two expressions, we may hope, that the following relation holds
1 -
a
1 8 {R (z) - SR
(Sz)}
- a a
or, symmetrizing in a and S,
(3.38) PC I -az a ) {R (z) - SR (Sz)} a a
The truth of (3.38) is expressed in the following lemma.
Lemma 3.15. The following relation between Ra and RS holds (3.39)
Proof. According to (3.1) the left-hand side of (3.38) can be written as
~(;z~
{RS(z) - aRS(az)}=
zP(Bz) {(I -:t:») -
(1 -:t:~»)}
=
1 {p(aSz) 1 }
=
Z
P(az)P(6z) -PTZJ '
which is symmetric in a and
S,
and hence equal to the right-hand side of(3.38). Rewriting (3.38) gives (3.39).
o
Now, using relation (3.39), we can easily prove the desired property (3.12).
Theorem 3.16. For all a,S € [O,IJ the following implication holds
Proof. The case a ~ 1,
a
=
1 has been treated in corollary 3.14. Now let a < 8 <follows lation
(3.40)
1 and P € Ca' Then also P € CI, and on account of lemma 3.12i) it
P(z) P(8z)
that , and therefore P( ) , is abs mono Considering the
re-p(i
z) az (3.39) and noting that1 -
B
{Ra(Z) - BRa(Bz)}=
L
n=On c n (B)r n (a)z ,
p(az)
we obtain the abs mon of RB from that of Ra and P(az) • Hence P €
CS'
o
Using lemma 3.15 we prove inequalities, stronger than r n (B) ~ 0 (n ~ 0), and of which (1.19), (2.9), (3.15), (3.23), (3.28) and (3.37) are special cases.
Corollary 3.17. If P € C
a, then we have for all a € [a,l]
(3.41) n
=
0,1,2, ••• ,and so: every r is nondecreasing on [a,I].
n
Proof. As we saw in the proof of theorem 3.16, if P € C and a a < at we can ~ite
ro
p(az) t n
P(az)
=
n~O sn(a,a)z ,coeffi-cients of zn, it follows from (3.39) and (3.40), that
c (a)r (13)
=
n n
from which (using
So ...
I and sn ~ 0) we obtain (3.41). The last assertion of this corollary is obtained from (3.41) by the fact, that cn(a) isnondecreas-ing in a.
o
From theorem 3.16 we obtain an extension of corollary 1.16, as follows.
Corollary 3.18. If the r (a), n ~ 0, defined by n n+J n - a
~
Pkrn_k(a) , n=
0,1,2, ••• Pn+l=
I-a k=Owith a E: [O,lJ, are all nonnegative, then the r (13), n ~ 0, defined by n
_ I3n+1 n
Pn+1
=
I
Pkrn_k(13), n=
0,1,2, ••• I - 13k==O
with 13 € [a,IJ, are all nonnegative, too. We then, in fact, have
n+J _ I3n+1
- a r (13) ~ r (a), 0,1,2, •••
1 - 13 n
=
.
I-a n n
Theorem 3.16 says, that for a ~ 13 the abs mon of RI3 follows from that of Ra' As on account of lemma 3.1iii) an abs mon R may be any abs mon function f
a
with f(l) < 1 ,with lemma 3.7 we can formulate the following assertion - a
about abs mon functions.
Corollary 3.19. If ~ is an abs mon function with ~(O)
=
a
and ~(l) < I, then the function~, fora
~ a ~B
~ 1 defined byor by
I - ~(z) I - ~(z) J-~(az)'" J-cp(Sz)'
~(z) ... 1
In theorem 2.9 we proved the following implication (now formulated in terms of generating functions)
(3.42) P C _ P(yz) C
€ a P(y) € at for y € [O,IJ •
If we denote the generating function of the r (a) corresponding to the pgf
(y) P(yz) (y). n
P (z):= P{y) by Ra (1n the sequel we shall use such a notation without
further explication). then with (3.1) we can write
Hence (3.43)
(I - a) zR (Y)(z)
a
p(y) (az) P(yaz)
... 1 - ... 1 - ... (1 - a)yzR (yz) •
p(y)(z) P(yz) a
R(y)(z) ... yR (yz) ,
a a
from which (3.42) immediately follows.
By means of similar methods we can obtain other properties of Ca' We formulate them in the following theorems.
Theorem 3.20. A pgf p. which is the limit of a sequence pg£ts P £ C ,
be-n a
longs to C (i.e. C is closed under weak convergence).
a a
Proof. Let P £ C for n
~
1 so R(n)(z) ...n a ' a
But then we have, if P(z) ... lim P (z).
n n-+<= R (z) =
-:--_1.
{l _ P(az)} ... a - a z P(z) ... lim R(n)(Z) ; n-+<= awhich is abs mon, too, so: P f C •
a
Theorem 3.21. For all a £ [O,IJ we have:
1 Pn (az) _ a
z
{I - ~ (z) } is abs mono n P (az) _ a~
{I - lim pn(z) } = n-+<= n (3.44) P € C a - P(y)P(z) P(yz) £C
a' f or y € [ 0 ) ,a€ CO' for y € [atlJ •
Proof. Define p(y)(z)
:=
P~1~!5z)
t for y E [O,lJ. If P € CI, then on account
of theorem 3.12i) we know that p(y) E
C
I, too, for all y. Now let a € [0,1)
and P E
C •
Then we can writea
(J-a)zR(y)(Z)
=
I _p(y)(az) a I _ P(az) P(yz) =a p(y)(z) p(yaz) P(z)
P(yz) {P(ayz) p(az)} P(yz) ( ) { () ( ) }
= P(ayz) P(yz) - P(z) = P(ayz) ] - a z Ra Z -YRa yz ,
so that
(3.45) R(y)(z) - P(yz) fR (z) - yR (yz)} •
a P(ayz) a a
Using (3.39) we can rewrite this
(3.46) R(y)(z) a = I -1 - y P(az) {R ( ) a P(yaz) y z - a R (a )} y.z •
As according to theorem 3.12i), ptyz» is abs mon, using (3.45) we obtain for )' ayz
all y E [O,lJ the abs mon of R(y from that of R • Hence p(Y)
a a
y € [0,1]. However, P €
C
for y ~ a, so on account of (3.29)that p(Y) €
Co
for all y!
[a,IJ.Theorem 3.22. For all a € [O,IJ we have:
€
C
for all a'W'e can conclude
o
(3.47) P E
C
~ n-l IT P( k )Yk
Z €C ,
a k-O P(y ) Y
lIn for n ~ 1, y E [a ,1].
Proof. Take a fixed n ~ 1 and define
n-I k
p(y)(z) := IT P(y z) k=O p(yk)
for y E [0,1] •
If P €
C
I' then p(y) as product of pgf's in
C)
belongs toC
I ' too, for all y. Now let a € [0,1) and P € C • Fora
€ [0,1) we can writea or, with
e
=
y, (1 - y)zR(y)(z) y p(ynz) "" (1 n) R ()=
1 - P(z) - y z n z yHence
R(Y)(z) = c l(Y)R (z) •
Y n- n
Y
(3.48)
As R n is abs mon for Y
~
a 1/n, it follows from (3.48), that~Y)
is abs monfor
~ ~
alIn.
Hence P (y) € C for Y~
a IIn.
0
Y
Remark. Inview of theorem 3.16 it is evident, that theorem 3.22 is equivalent
to the assertion that for all Y € [O,IJ:
(3.49)
We can easily prove (3.49) using lemma 3.4: If P € C , then there are
yn
P € (0,1) and pgf Q with Q(O)
=
0, such thatBut then co nR. P(z)
=
IT 1 - pQ(y )nl
•
R.=O I - pQ(y z) n~1 PCykz) k k=O p(y ) Still another proof ofn
(3.49) we obtain from the characterization (3.29): If
P € C n' then P(y )P(z)
Y p(ynz)
n-l k
€ CO. But for p(y)(z)
:=
IT P(YkZ) we havek=O P(y )
Corollary 3.23. For all n ~ I and all a € [O,IJ we have
(3.50) n-I k IT 1 - pQ(a ) c
C
k
'"
a ' k"O 1 - pQ (a z) with p € [0,1) and Q a pgf. I-p Proof. Take P(z)=
I _ pQ(z) € Co in (3.47).o
Theorem 3.24. For all ~ € [O,IJ we have:
(3.51) for Y E [0, 1] •
Proof. If P €
C
I, then, according to theorem 1.1, pY €
C
1 for all Y ~ O. Now let ~ € [0,1) and P € C • Then we can write~
(y) pY (~z)
(I - ~)zR (z)::: 1 - :::
~ pY(z)
so that
!..
[(I - ~)ZR(Y) (z)J __ (P(~z»Y-l d [P(~z)]=
dZ ~ - Y P(z)
rz
P(z)Hence (3.52)
As on account of theorem 3.12i)
p~(~;jz)
€ CI, and so
(p~(~;jz»]-Y
€ C1 for Y € [0,1], it follows that (P(z) )I-Y is abs mon for Y € [O,IJ. Now, usingP(~z) ( )
(3.52), we obtain the abs mon of R~Y from that of R~, and we conclude: pY €
C
~.
o
use characterization (3.29) for the proof of theorem Remark. If we wish to
3.24, then we have to p € [0. 1) and A a pgf,
1 -
!
prove (3.51) for ~
=
O. If P(z) ::: 1 _ pA z) € CO' with then we can write [p (z) JY=
I~ ~Btz)
, withq
:=
1 - (] - p)Y , B(z) :=1.
{I - (t -pA(z»Y} =1.
{t -1:
q q n=O <X>::: 1.
~r
(n-l-Y)( A(z»n q l.. n n-l p , n=tNow let P E: C a.. Then, according to (3.29), P(a.)P(z) E: Cat so that
P(a.z) P(a.)Yp(z)Y
= (P(a. )P(Z»Y Eo Co '
P(a.z)Y P (a. z)
too, for Y E: [O,lJ. Using (3.29) once more, we conclude that pY E: Ca. for
Y E: [0,1].
We conclude the properties of the classes C with the assertion that u
a. a<1
is dense in C
1 in the following sense: Theorem 3.25. If P E:
C
I , then there is an increasing sequence an' with lim a. '"' 1, and there are pgf's P E C ,such that
n+<'O n n a. n
P(z)
=
lim P
(~),n+<'O n for
I zl
~ I •Proof. Let P E: CI' Then there is a X > 0 and a pgf Q, such that
Take a. :=
n
(3.53)
P(z) = exp[X(Q(z) - I)J •
I
- ~ • If for n > A we define the pgf's Pn by
P (z) n n n-I
=
IT k=O k 1 - A!n.Q(a ) nk
'
I - A!n.Q(a. z) n C qthen, on account of corollary 3.23, we know that P nEe a. • As P (1) = P n (I) = I ,
n
it is sufficient to prove the convergence of P (z) to P(z) for z E: [0.1). So,
n
take a fixed z E: [0,1) and write
with n-I = IT k=O {J + A[Q(z) - IJ
=
nHence (3.54) For k < n we estimate k ~ zQ' (z) • (t - (l ) t n
because (t - (lk) z < 1;; k < z and Q is convex. As for k < n
n n.
-
(l~
= I - (1 - ; , ) k~
1 - (I - k ; ' )~
*
n n
we finally have for some C(z) > 0 and all k < n (3.55)
n I n
As lim Q(z)
=
Q(l) = I and lim (l=
lim (I -~)=
I, we further have forztl n~ n n~ n
all k < n
Using this and (3.55) we conclude from (3.54) that
(3.56)
Now it follows that
and
n-I
p (z)
~
IT {I +A[Q(Z~
- IJ + Ifk(n)\}~
n k=O
~ {I + A[Q(Z) - IJ + e(n)}n + exp[A(Q(Z) - I)J,