Upload
pierre
View
216
Download
3
Embed Size (px)
Citation preview
84
10 - POISSON STREAMS
10.1 Privileged transitions
The situation is that of sections 6,7,8 and 9. The main features of interest in the lJ"esent section are the existence of a stationary poi nt process
N = I cST and of a sequence of marks (U , n e ZI) where U ~(X l' nf'V n n nn-
~N,et'p) ~here 0n-1' Xn,Yn ) e
GxSxGx( n M). Recall that X • X(T ) is ses s n n the state at time Tn' a~d an-1 is the
o 0 site of A(X 1) whi ch triggers the transi tion X 1 + X , i.e. X 1 --!!:l.)X • If PN n- . n- n n- n is the Palm probability associated with (N,at,p), we have (Eq. (8.1.2» :
Let
H C GxScG
s be a subset of lJ"i vileged transi tions g + ~.
Let now NH be the point p-ocess associated with H, Le.
N ( .) • I 1 {( X ~ X)eM I cST (.) --H nf'V n-1' n-1' n n
Fran (10.1.1)
Calling"'H the right-hand side of (10.1.5) and applying (10.1.4), we have
P~(X_l·g, ~-ls, XO"'8', YO € ii> •
1 H(g,S,g')
and therefore
ACs ,g) A
----"- peg ,s ,g') PA(g)-{ s}( B)I "'H A(g)
(10.1 .2)
(10.1.3)
( 1 O. 1 .4)
( 10.1 .5)
(10.1.6)
pON (X o"'8' ,ioeS) • I 1H(g,s,g') ACs,g) p(g,s,g') (A ( PA(g)-{s} B)/L. 10.1.7) H (g ,s )eGxS A(g) °H
F. Baccelli et al., Palm Probabilities and Stationary Queues© Springer-Verlag Berlin Heidelberg 1987
Let l.5 recall that
= x g'
85
P A( g' ) (8) •
10.2 Sufficient conditions for Poissonian streCIIS
( 1 O. 1. 8)
Suppose that for all ~ e G, 8 € ( 0 (H n M »), the right-hand sides of aes s s
(10.1.7) and (10.1.8) are equal. It is then clear that the restrictions of P and
P~ to a( ~ ; n <: 1) are equal, where (~, n e Z/) is the sequence of p:>ints of NW H
From the characterization of Poisson processes of 87, Part 1, it follows that NH is a Poisson process.
The following condition is sufficient for this to happen
(g ,s ,g' ) € H ~ A( g') n s' = (A( g) - {s} n S'),
x = g'
L 1 H( g ,s ,g') x ).( s ,g) p (g ,s ,g' ) (s,g)eSxG g
L x ).(s,g) p(g,s,g') (g,s,g' )eH g
Indeed (10.2.1a) implies that for all B € ~ (Hs n Ms)
PA(g)_{s}(fl) & PA(g' )(8) , if (g,s,g') e H
and (10.2. lb) concl udes the proof. a
( 1 O. 2. 1 a)
(10.2.1b)
(10.2.2)
APPmroIX
Appendix 1. Change of scale
With each macrostate g e G, we associate a measurable flow (T:, tell) on
(n M, ® Ms) by s€S s f£S
s Note that if s r;. A(g), c(s,g) - 0, so that Ttc(s,g) ms ~ ms.
Define the following r.v. on (n , ® M) s€S f£S s
x (m ) inf { 1 s ; s e A( g) /,
c (s ,g)
where x (m ) is the first lXlint of ms strictly to the right of O. 1 s
Lanma
Proof
PAC ) is Tg-invariant. - g --,
Def ine the mappi ng k s,g
k (m) ( .) - L Ii ( ) (. ) • s,g s i£l/ Xi ms
C (s ,g)
(A.1.1)
(A.1.2)
( A. 1 • 3)
-1 This mapping performs a change of scale, so that if we denote n ~ P ok ,
s,g s s ,g (Ms,HS,nS,g) is a stationary point !l"ocess with the intensity >,(s,g) - >'sc(s,g). If
s e A(g), >.(s,g) > 0, and the Palm !l"obability nO associated with t ,n -1
ok. s ,g
s ,g s s,g is
Define the !l"obability Q on ( n Ms ' ® Hs) by seA ( g) f£A( g)
-Q = ® f£A( g)
n s,g
The Palm probability QO of Q with respect to
N = L seA(g)
is gi yen by
m s
(A. 1.11)
(A.1.5)
13 0 -
I A( g) I L
1-1
87
(A.l.6)
where {u 1' ••• ,ul A(g) I } is an enuneration of A(g). Clearly QO is T T' -invariant,
where'f' is the r.v.definedon ( II M, ® H)by seA(g) s £A(g) s
( A. 1. 7)
Denote by k the mappi ng from II Minto itself def ined by k «m , s eAt g») g seA(g) s g s
• (k (m), s e A(g». If T is the restriction of T to ( II Ms' ® H), and s,g s seA(g) £A(g) s
if T~ is the restriction of T: to this space, then T~ is the image of T t , by kg'
Therefore the probability
o IA(g)1 Q - L
1-1
is 18- invar i ant. 'f
The lanma is now established since QO is the restriction of P A(g) to
( II M, ® H), and since T:m & ms when s fI A(g). 0 seA(g) S £A(g) S s
(A. 1. 8)
Rancrk 1 : The point p"ocess N defined by (A.l.5) admits the intensity A(g) under = --Q, and the inversion formula giving Q in terms of QO is
'f' Q( A) • A (g) E':Qo [f 1( (T m , seA ( g » e A) d u ] , o u s
-1 - -1 Therefore, since Q • QOkg and QO • QO 0 kg ,
T Q(A) • A(g) EQo [f l((18m ,s e A(g» e A) du), o u s
R ancrk 2 : We have
EA [t] = E- [t']. P A(g) %
(A € ® £A(g)
H ). s
(A € ® H). £A(g) s
( A. 1. 9)
( A. 1 .10)
(A.l.l1)
88
Therefore in view of the construction of (Z(t), t G- R ) in !!6, and of the
definition ofT,
(A.1.'3)
where the 1 ast i nequal1ty f oll ows from ass un pt ion A.2 of !!7. U si ng (7.'.4) we obtain
E O[ T, ] c K LX- K P g£G g
(A.,.,4)
Appendix 2. Proof' of Insensl t1Ylty (Proof' of' (8.1.1»
We aoopt the notation A(g) - (u" ••• u/ A(g) / }. Let g e G and B € ~ 'DIs be
fixed. Ck is the event
C~=(Tgr' (B n (xO(m ) -O}, t uk
where xO(m s ) is the_ first point of ms less or equal to zero. Considering all the
possible values for °0 ' we get
pO [ X (0) • g, X (T,) • g', Y (T,) e B J
/A(g)/O = L p [X (0) - g, X (T,) • gf, Y( T,) e B n {x (m ). 0 J]
k _, 0 uk
/ A(g) / 0 g • L p [X(O) -g, Y(O) e Ck, X(T,) - gf J.
k -,
Using the conditional independence of X( T,) and Y(O) knowing (X (0) ,~O) (see ~6. 4), we can rewrite the last expression as
where the R.H.S. is obtained fran the very definition of pO (see (6.3.6».
But in view of the lemma of Append1x "
89
Since
we finally get
( A. 2. 1)
For any u 6- A(g) such that g ~ g' is allaied, we have
PA(g)_{U}(B) - PA(g')_{v}(B) if (A(g')-A(g» n S'· {v}. (A. 2. 2)
To show this, use the birth restriction property of h.5, which implies that if u -
u is immediately reactivated after the transition g + g', then necessarily u e- s' since v -£ S' i also use the JToperty that for Poisson JTocesses stationary
probab1lity and+ Palm JTobab1lity coincide on all events implying only the
restriction toB - {O} of the point JTocess (see h of Part 1). We have similarly
PA(g)-{u}( B) K P A(g') (B) if A(g') n S' c A(g) and u fI A(g') ( A. 2. 3)
and
PA(g)_{U}(B) KPA(g')_{u}(B) ifA(g') n S'C A(g) and ueA(g'). ( A. 2. 4)
Let now u 6- A(g') n S'. Write the local balance equation for (g' ,u),
multiply both sides by PA(g')_{u}(B). Using (A.2.2), (A.2.3), (A.2.4) and
o relation (7.1.4) linking Xg and xg ' we obtain
and
the
o A ( u ,g') P ( B) L ° L Xg' A(g') A(g')-{u} - _ Xg
gE:Gu sfA ( g )
peg ,s ,g' ) Hs ,g) ( A(g) PA(g)-{ s} B)
t X 0 P ( ') AC u ,g) P ( B) + L g g,u,g A(g) A(g)-{u} •
gE:Gu
(A. 2.5)
Similarly, multiplying both sides of (7.5.2) by "A(g,)(B), and using (A.2.2),
( A. 2. 3), (A. 2. 4), we get
o t ACs ,g' ) xg' L - --- PA(g' )_( s}( B) •
sfA(g')nS' A(g')
~ xO ~ A(S,g) g I A(g)::>A(g')n S' g sfA(g)nS' A(g) p(g,s,g') PA(g)_{s}(B). (A.2.6)
90
The aumnation of thia relation and of relation (A.2.5) when u apans A(g')nS'
yields (uae the birth restriction p"0perty)
o xg' ~
BfA(g' )
A (a ,g' ) t t 0 A( a ,g) PA(g' )_{ a}( B) - L L Xgp(g ,a ,g') PA(g)-{ a}( B)
A(g') g£G BfA(g) A(g)
that ia to Bay in view of (A.2.1)
o A
Xg' P A( g' ) (B) - P( X (T,) = g', Y (T,) e B), Q.e.d.
Appendix 3. The trans1tion II .. Jcs (Proof' c:L (8.1.2»
Reproducing canputations already done at the begining of the preaent appendix,
we get
Appendix II. Proof c:L (8. 3. 5)
and
By construction, ~
o A
xeo) -Xo' YeO) -(T_T yo). o
Denote by Zo the R. H. S. of (A. 4.1). The Palm inveraion formula yields
p[X(O) - g, yeO) e B] = ptxo=g, ~ € B]
T,
, ~[f , (XOOTu·g, ZOOTueB) du).
E~T1J 0
But for u e- [0, T, ) ,
&. ZOOT - (T OY ) u u 0 •
(A. 4.1)
eA.4.2)
91
Hence
[A ...D [JT1 ~'::--.. A ]
P X(O)a g, Y(O) e B] - ---E""N 1(X 0"'8, TU Yo € B)du.
t'tJ: T1 ] 0
(A. 4. 3)
In view of the definition of rP, the R.H.S. of (A.4.3) reduces to
K.A(g)
where we have used (A. 1 .1 0) and (A. 1 .14) to get the last expr essi ons. Equation (8.3.5) follows fran equation (7.'.4).
!ppendix 5. Proof of (9.1.3)
Let
d(g)- r xg' r p(g',s.g)A(s,g') g ilG seA(g' )
50
and for g,g e G , let 50
A(g' ,g) • 6(g' ,g) O(so,g) - A(g» + ~ p(g' ,s,g) A(s,g'), sf'A( g' )
s.'o50
where 6 denotes the Kronecker sl'llbol. With these notations, for g e global balance equation of (r,A) reads
+
( A. 5.1)
(A. 5.2)
the
If p(.,s ,.) : G x G +lR is an arbitrary substochastic matrlx, the state o So So
space G can be partitioned into one inessential class KO and a fanily of So
essential cl~ses of irreduclbility K"K 2, ... Clearly, a(g) ·0 (see (9.3.5» for geK.LetK -
o be the subset of Gs where a(g) > 0 and K be the subset where a(g) < O. N oU ce 0 + •
that transi tions between K and K are Impossi bl e. In vi ew of equation (9.3.4) the substraction 0,.(9.2.8) - A (A.5.3» yields the relation
So
92
L b (g') A( g' ,g) g' £G
So
where b (g) C A, x ( )-A x , g e G • Let a, g So g So
{g e G I b(g) > a I and So
{g e G I b(g) < a I. So
+ + Let g6-J. Since transitions from K toJ are imjXlssible, (A.6.4) reads
L + b (g') A( g' ,g) g'£J
a (g) • L _ b(g') A( g' ,g) g'£J
+ (A II - A ) {d(g) + A, L fa (g') peg' ,So,g) A(So,g') , So g' £KO ,
+ A, L + fa (g') (p(g' ,So,g) - 6(g,g'» A(So,g') I. g' £K ,
(A.5.4)
(A.5.5)
Assune that A,II - A < 0, which is always jXlssible. &unmation of the ri@l1t So
hand side of equation (A.5.5) for all g e J+ yields a non-jXlsitive result (use the relation 1: + peg' ,So,g) .,). Hence the sunmation of the left hand sides has to
g€K
yield a norrpositive result too, that is
A, A2 __ L + a(g) + L + beg') (- L + A(g',g»:S; o. (A.5.6) AO g€K g' £J gEa
Fran the very definition, A(g,g') i: 0, g f, g' and A(g' ,g) - 0, so that
1: + A( g' ,g) i: o. g€J
+ Hence, the left hand side of (A.5.6) is zero which implies that K..(1.
Let us show that K is empty too. The suretraction «9.2.8)-II.(A.5.3) yields
A2 _ a(g) ." (ft
\ l. x -IIx) A(' ) A g' £G a, (g' ) g' g ,g • So So
( A. 5.7)
93
Summation of (A. 5.7) entails
L a (g) - 0 g£G
So
(use the convergence of the series l: g€G
So which conci udes the proof of (9.3.1).
x I A( g ,g) I ). Hen ce K g
APPENDIX 6. Proof of the converse theorEIII in the general case
( A. 5. 8)
+ .q entails K a(J,
Let C = {g e G / c(so,g) ~ OJ. For g e c, the immediate activity assunption So
entails that roth d(g) and the second term in the right hand side of (A.5.3)
vanish. Hence the i nsensi ti vity bal ance equation is tri vi all Y sati sf! ed for all g in C. Furthermore, if g e c, equation (A.5.4) together with the relation a(g) = 0
entail that
L b(g') A( g' , g) • 0, g e C.
~-£C
We can r ewr i te (A. 6. 1) as
b(g) M (g) = L b(g') u (g' , g ), g e C, ~-£C
(A. 6.1)
(A.6.2)
where (M(g), u(g,.), g e C) is the substochastic infinitesimal generator on C
defined by
and
u(~ ,g) - I peg' ,s,g) a(s,g') sf.A( g' )
s.*so
M(g) = A(g) - u(g,g).
(A.6.3)
( A. 6. 4)
Hence, beg) - 0, at least for all g such that this infinitesimal generator on C
is strictly substochastic or equi valently such that
L A ( g ,g') < O. ~-£C
Let g & C. Using this fact in equations (9.2.8) and (9.2.9) together with the
propertya(g) = 0, g e G , we get respectively (A. 6.5) and (A. 6.6) belCAol. So
and
- (1
94
Xg, A(g' ,g) + A x C(So,g) - A (l-A IA) Xg c(so,g) So g 2 So 1
A
-~) L Al t -£G
So
a (l-Il) (d(g) + L Xg' p(g',so,g) A C(So,g'», t -£G So
So
(A.6.5)
where we have used the property "A(g' ,g) > 0, g i C implies L A(g' ,g) < 0".
Elimination of the term L t -£G
So
gee X, A(g',g) yields the desired result. The Il"oof of g
the converse theorem is thus concluded.