HAL Id: inria-00288866
https://hal.inria.fr/inria-00288866v2
Submitted on 2 Apr 2009
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Directionally Convex Ordering of Random Measures, Shot Noise Fields and Some Applications to Wireless
Communications
Bartlomiej Blaszczyszyn, D. Yogeshwaran
To cite this version:
Bartlomiej Blaszczyszyn, D. Yogeshwaran. Directionally Convex Ordering of Random Measures, Shot Noise Fields and Some Applications to Wireless Communications. Advances in Applied Probability, Applied Probability Trust, 2009, 41, pp.623-646. �10.1239/aap/1253281057�. �inria-00288866v2�
SURES, SHOT NOISE FIELDS AND SOME APPLICATIONS TO
WIRELESS COMMUNICATIONS
BART LOMIEJB LASZCZYSZYN,
INRIA/ENSandMath. Inst. UniversityofWro law
D.YOGESHWARAN,
INRIA/ENS
Abstrat
Diretionally onvex (dx) ordering is a tool for omparison of dependene
strutureofrandomvetorsthatalsotakesintoaountthevariabilityofthe
marginaldistributions. Whenextendedtorandomeldsitonernsomparison
of all nite dimensional distributions. Viewing loally nite measures as
non-negative elds of measure-values indexed by the bounded Borel subsets
of the spae, in this paper we formulate and study the dx ordering of
random measures on loally ompat spaes. We show that the dx order
is preserved under some of the natural operations onsidered on random
measures and point proesses, suh as deterministi displaement of points,
independent superposition and thinning as well as independent, identially
distributed marking. Furtheroperationssuhas positiondependent marking
and displaement of points are shown to preserve the order on Cox point
proesses. We also examinethe impat of dx orderon the seondmoment
properties,inpartiularonlusteringandonPalmdistributions. Comparisons
of Ripley's funtions, pairorrelation funtions as wellas examples seem to
indiatethatpointproesseshigherindxorderlustermore.
As the main result, we show that non-negative integral shot-noise elds
with respetto dx ordered random measures inheritthis ordering from the
measures. Numerous appliations of this result are shown, in partiular to
omparisonofvariousCoxproessesandsomeperformanemeasuresofwireless
networks,inbothofwhihshot-noiseeldsappearaskeyingredients. Wealso
mentionafewpertinentopenquestions.
Keywords: stohasti ordering, diretional onvexity, random measures, ran-
domelds,pointproesses,wirelessnetworks
2000MathematisSubjetClassiation: Primary60E15
Seondary60G60,60G57,60G55
Postaladdress: ENSDITREC,45rued'ulm,75230Paris,FRANCE.
Postaladdress:ENSDITREC,45rued'ulm,75230Paris,FRANCE.
Point proesses (p.p.) have been at the entre of various studies in stohasti
geometry, both theoretial and applied. Most of the work involving quantitative
analysis of p.p. have dealt with Poisson p.p.. One of the main reasons being that
harateristis of Poisson p.p. are amenable to omputations and yield nie losed
formexpressionsinmanyases. ComputationshavebeendiÆultingreatmanyases,
evenforCox(doublystohastiPoisson)p.p..
Comparisonofpointproesses Toimproveuponthissituation,qualitative,om-
parativestudiesofp.p. haveemergedasusefultools. Therstmethodofomparison
ofp.p. hasbeenouplingorstohastidomination(see[18,20,32℄). Inourterminology,
theseareknownasstrongorderingof p.p.. Whentwop.p. anbeoupled,oneturns
outtobeasubsetoftheother. Thisorderingisveryusefulforobtainingvariousbounds
andprovinglimittheorems. However,usingitoneannotomparetwodierentp.p.
withsamemeanmeasures. An obviousexampleis anhomogeneousPoisson p.p. and
astationary Coxp.p. with thesameintensity. The questionarises ofwhat ordering
is suitable for suh p.p.? This is an important questionsine it is expeted that by
omparing p.p. of the same intensity one should ahieve a tighter bound than by
oupling. Forsomemoredetailsonstrongorderingofp.p. andneedforotherorders,
seeremarksin [29,Setion5.4andSetion 7.4.2℄.
From onvex to dx order Tworandomvariables X andY with thesamemean
E(X) = E(Y) an be ompared by how "spread out" their distributions are. This
statistial variability (in statistial ensemble) is aptured to a limited extent by the
variane, but more fully by onvex ordering, under whih X is less than Y if and
only if for all onvex f, E(f(A)) E(f(B)). In multi-dimensions, besides dierent
statistial variability of marginal distributions, two random vetorsan exhibit dif-
ferentdependene properties of theiroordinates. Themost evidentexample hereis
omparisonofthevetoromposedofseveralopiesofonerandomvariabletoavetor
omposed of independent opies sampled from the same distribution. A useful tool
foromparisonofthedependene strutureofrandomvetorswithxedmarginalsis
thesupermodular order. Thedxorderisanotherintegralorder(generatedbyalass
of dxfuntions in thesame manneras onvexfuntions generatethe onvexorder)
thatanbeseenasageneralizationofthesupermodularone,whih inadditiontakes
intoaountthevariabilityofthemarginals(f[29,Setion3.12℄). Itanbenaturally
extendedtorandomeldsbyomparisonofallnite dimensionaldistributions.
extension that onsists in dx ordering of loally nite measures (to whih belong
p.p.) viewed as non-negativeelds of measure-valueson all bounded subsets of the
spae. Weshowthat thedxorderispreservedundersomeofthenaturaloperations
onsideredonrandommeasuresandpointproesses,suhasindependentsuperposition
and thinning. Also, we examine the impat of dx order on the seond moment
properties, inpartiularonlustering,andPalmdistributions.
Integral shot-noise elds Many interesting harateristis of random measures,
bothinthetheoryandinappliationshavetheformofintegralsofsomenon-negative
kernels. We all them integral shot-noise elds. For example, many lasses of Cox
p.p., with the most general being Levy based Cox p.p. (f. [14℄), have stohasti
intensityelds,whihareshot-noiseelds. Theyarealsokeyingredientsofthereently
proposed,so-alled\physial"modelsforwirelessnetworks,aswewillexplaininwhat
follows(see also [1,8,11℄). It is thus partiularly appealing to study the shot-noise
eldsgeneratedbydxorderedrandommeasures.
Sine integralsare linearoperators on thespae of measures,and knowing that a
linearfuntion of avetoris trivially dx, it is naturallyto expet that the integral
shot-noiseeldswithrespettodxorderedrandommeasureswillinheritthisordering
fromthemeasures. However,thispropertyannotbeonludedimmediatelyfromthe
nite dimensional dx ordering of measures. The formal proof of this fat that is
themain resultofthis paperinvolvessomeargumentsfrom thetheoryofintegration
ombined with the losure property of dx order under joint weak onvergene and
onvergeneinmean.
Ordering in queueing theory and wireless ommuniations The theory of
stohastiorderingprovideselegantandeÆienttoolsforomparisonofrandomobjets
andisnowbeingusedinmanyelds. Inpartiularinqueueingtheoryontext,in[33℄,
Rossmadeaonjeturethatreplaing astationaryPoissonarrivalproess inasingle
server queue by a stationary Cox p.p. with the same intensity should inrease the
averageustomer delay. Therehavebeenmanyvariationsof theseonjetures whih
are now known as Ross-type onjetures. They triggered the interest in omparison
ofqueueswithsimilar inputs([6,25,31℄). Thenotionofadxfuntion waspartially
developedandusedinonjuntionwiththeprovingofRoss-typeonjetures([21,22,
34℄). Muhearliertotheseworks,aomparativestudyofqueuesmotivatedbyneuron-
ringmodelsanbefoundin[16℄. Alsoomparisonofvarianesofpointproessesand
breproesseswasstudiedin[36℄andheneitanbeonsideredasaforerunnertoour
artile. TheappliabilityoftheseresultshasgeneratedsuÆientinterestinthetheory
andStoyan([29℄). Asmostworksonorderingofp.p. weremotivatedbyappliations
toqueueingtheory,resultswereprimarilyfousedonone-dimensionalpointproesses.
An attempt to retifythe lakofworkin higher dimensionswasmadein [24℄,where
omparisonresultsforshot-noiseeldsofspatialstationaryCoxp.p. weregiven. The
resultsof[24℄ arethestartingpointofourinvestigation.
Our interest in ordering of point proesses, and in partiular in the shot-noise
eldstheygenerate,hasrootsin theanalysisofwirelessommuniations,wherethese
objetsareprimarily usedto modelthesoalled interferenethat isthetotal power
reeivedfrom manyemitterssatteredin theplaneorspaeandsharingtheommon
Hertzianmedium. Aordingto anewemergingmethodology,theinterferene-aware
stohasti geometry modeling of wireless ommuniationsprovidesa way of dening
and omputing marosopi properties of large wireless networks by someaveraging
overall potential random patterns for node loations in an innite plane and radio
hannelharateristis,inthesamewayasqueuingtheoryprovidesaveragedresponse
timesorongestionoverallpotentialarrivalpatternswithinagivenparametrilass.
These marosopi properties will allow one to haraterize the key dependeniesof
the network performane harateristis in funtion of a relatively small number of
parameters.
In the above ontext, Poisson distribution of emitters/reeiver/users is often too
simplisti. Statistis show that the real patterns of users exhibits more lustering
eets (\hotsspots")thanobserved in anhomogeneousPoisson pointproesses. On
theotherhand,goodpaket-ollision-avoidanemehanismsshemeshouldreatesome
\repulsion" in the pattern of nodes allowedto aess simultaneouslyto the hannel.
Thisrisesquestionsabouttheanalysisofnon-Poissonmodels,whihouldbetosome
extent takled on the ground of the theory of stohasti ordering. Interestingly, we
shallshowthat thereareertainperformaneharateristisinwirelessnetworksthat
improvewithmorevariabilityintheinputproess.
Theremainingpartoftheartileisorganizedasfollows. Inthenextsetion,
wewillpresentthemaindenitionsandstatethemainresultsonerningdxordering
of the integral shot-noise elds. Setion 3 will explore the various onsequenes of
ordering ofrandom measures. Theproofs of themain resultsare given in Setion 4.
Examples illustrating the use and appliation of the theorems shall be presented in
Setion 5. Setion 6 will sketh some of the possible appliations of results in the
ontext of wireless ommuniations. Finally, we onlude with some remarks and
questions in Setion 7. Thereis anAppendix (Setion 8) ontainingsomeproperties
ofstohastiordersandtheirextensionsthatareusedin thepaper.
TheorderonR n
shalldenotetheomponent-wisepartialorder,i.e.,(x
1
;:::;x
n )
(y
1
;:::;y
n )ifx
i y
i
foreveryi.
Denition2.1. We say that a funtion f : R d
! R is diretionally on-
vex(dx)ifforeveryx;y;p;q2R d
suhthat px;yqandx+y=p+q,
f(x)+f(y)f(p)+f(q):
Funtionf issaidtobediretionallyonave(dv)iftheinequalityintheabove
equationisreversed.
Funtionf issaiddiretionally linear(dl)ifitisdxanddv.
Funtionf =(f
1
;:::;f
n ):R
d
!R n
issaidtobedx(dv)ifeahofitsomponent
f
i
is dx(dv). Also, weshall abbreviateinreasing and dx by idx and dereasing
anddxbyddx. Similarabbreviationsshallbeusedfordv funtions. Moreover,we
abbreviatenon-negativeandidxbyidx +
.
Inthefollowing,letFdenotesomelassoffuntionsfrom R d
toR. Thedimension
disassumedtobelearfrom theontext. Unlessmentioned,whenwestateE(f(X))
forf 2FandX arandomvetor,weassumethattheexpetationexists,i.e.,foreah
randomvetorX weonsider thesub-lassofFforwhih theexpetationsexistwith
respetto(w.r.t) X.
Denition2.2. SupposeX andY arereal-valuedrandomvetorsofthesame
dimension. ThenXissaidtobelessthanY inForderifE(f(X))E(f(Y))for
allf 2F(forwhihbothexpetationsarenite). WeshalldenoteitasX
F Y.
Suppose fX(s)g
s2S
and fY(s)g
s2S
are real-valued random elds, where S is
an arbitraryindex set. We say that fX(s)g
F
fY(s)g iffor everyn 1and
s
1
;:::;s
n
2S,(X(s
1
);:::;X(s
n ))
F (Y(s
1
);:::;Y(s
n )):
Inthe remainingpartofthe paper,wewill mainly onsider Fto bethelass ofdx,
idxandidvfuntions;thenegationofthesefuntionsgiverisetodv;ddvandddx
orders respetively. If F is thelass of inreasing funtions,weshall replaeF by st
(strong)in theabovedenitions. Thesearestandardnotationsusedin literature.
As onerns random measures, we shall work in the set-up of [17℄. Let E be a
loally ompat, seond ountable Hausdor (LCSC) spae. Suh spaes are polish,
i.e., ompleteandseparablemetrispae. LetB(E) betheBorel-algebraandB
b (E)
bethe -ring of bounded, Borel subsets (bBs). Let M =M( E) be thespae of non-
negativeRadonmeasuresonE. TheBorel-algebraMisgeneratedbythemappings
7!(B)forallB bBs. Arandommeasure isamappingfromaprobabilityspae
(;F;P)to(M;M). Weshallallarandommeasureap.p. if2N,thesubsetof
ountingmeasuresinM. Further,weshallsayap.p. issimpleifa.s. (fxg)1for
allx 2E. Throughout, weshalluse for anarbitraryrandom measureand fora
p.p.. Arandommeasure anbeviewedasarandomeldf(B)g
B2B
b (E)
:Withthis
viewpointandthepreviouslyintroduednotionoforderingofrandomelds,wedene
orderingofrandommeasures.
Denition2.3. Suppose
1
() and
2
() are random measures on E. We say that
1 ()
dx
2
()ifforanyI
1
;:::;I
n
bBsin E,
(
1 (I
1 );:::;
1 (I
n ))
dx (
2 (I
1 );:::;
2 (I
n
)): (1)
Thedenitionissimilarforotherorders,i.e.,whenFisthelassofidx=idv=ddx=ddv=st
funtions.
Denition2.4. LetSbeanysetandE aLCSCspae. Givenarandommeasureon
E andameasurable(intherstvariablealone)responsefuntionh(x;y):ES !
R +
where
R +
denotestheompletionofpositivereal-linewithinnity,the(integral) shot-
noise eldisdenedas
V
(y)=
Z
E
h(x;y)(dx): (2)
Withthisbriefintrodution,wearereadytostateourkeyresultthatwillbeproved
inSetion 4.1.
Theorem2.1. 1. If
1
idx(resp.idv)
2
, then fV
1 (y)g
y2S
idx(resp.idv)
fV
2 (y)g
y2S .
2. LetE(V
i
(y))<1,forally2S,i=1;2:If
1
dx
2
,thenfV
1 (y)g
y2S
dx
fV
2 (y)g
y2S .
Therstpartof theabovetheoremfor theone-dimensionalmarginalsofbounded
shot-noiseeldsgeneratedbylowersemi-ontinuousresponsefuntionsisprovedin[24℄
for the speial ase of spatial stationary Cox p.p.. It is onspiuous that we have
generalizedtheearlierresulttoagreatextent. Thismoregeneralresultwillbeusedin
manyplaesinthispaper,inpartiulartoproveorderingofindependently,identially
marked p.p. (Proposition 3.2), Ripley's funtions (Proposition 3.4), Palm measures
(Proposition 3.5), independently marked Cox proesses (Proposition 3.7), extremal
shot-noise elds (Proposition 4.1). Apart form these results, Setions 5 and 6shall
amplydemonstrateexamplesandappliationsthat shallneedTheorem2.1.
WeshallnowgiveasuÆientonditionforrandommeasurestobeordered,namely
thattheondition(1)inDenition 2.3needsto beveriedonlyfordisjointbBs. The
neessityistrivial. Thisisamuheasieronditionandwillbeusedmanytimesinthe
remainingpartofthepaper.
Proposition 3.1. Suppose
1
() and
2
() are two random measures on E. Then
1 ()
dx
2
()if and only if ondition (1)holds for all mutually disjoint bBs. The
sameresultsholdstrue for idxandidvorder.
Proof. We need to prove the 'if' part alone. We shall prove for dx order and
the same argument is valid for f being idx oridv. Let ondition (1) be satised
for all mutually disjoint bBs. Let f : R n
+
! R be dx funtion and B
1
;:::;B
n be
bBs. We an hoose mutually disjoint bBs A
1
;:::;A
m
suh that B
i
= [
j2Ji A
j for
all i. Hene (B
i ) =
P
j2Ji (A
j
): Now dene g : R m
+
! R n
+
as g(x
1
;:::;x
m ) =
( P
j2J1 x
j
;:::; P
j2Jn x
j
): Theng isidl andsof Æg is dx. Moreover,f((B
1 );:::;
(B
n
))=fÆg((A
1
);:::;(A
m
))andthustheresultfordxfollows.
3.1. Simple OperationsPreserving Order
Point proesses are speial asesof random measures and as suh will besubjet
to theonsidered ordering. It is known thateahp.p. onaLCSC spae E anbe
representedasaountablesum= P
i
"
X
i
ofDirameasures("
x
(A)=1ifx2Aand
0otherwise)insuhawaythatX
i
arerandomelementsinE. Weshallnowshowthat
allthethreeordersdx;idx;idvpreservesomesimpleoperationsonrandommeasures
andp.p.,asdeterministimapping,independentidentiallydistributed(i.i.d.) thinning
andindependentsuperposition.
Let:E !E 0
beameasurable mappingtosomeLCSCspaeE 0
. Bytheimageof
a(random)measurebyweunderstand 0
()=( 1
()). Notethat theimageof
ap.p. byonsistsin deterministidisplaementofallitspointsby.
Let = P
i
"
x
i
. By i.i.d. marking of , with marks in some LCSC spae E 0
,
we understand a p.p. on the produt spae E E 0
, with the usual produt Borel
-algebra, dened by
~
=
P
i
"
(xi;Zi)
, where fZ
i
gare i.i.d. random variables (r.v.),
soalled marks,on E 0
. By i.i.d. thinningof ,weunderstand = P
i Z
i
"
xi , where
Z
i
arei.i.d. 0-1Bernoulli randomvariables r.v.. The probability PfZ =1gis alled
theretentionprobability. Superpositionofp.p. isunderstoodasadditionof(ounting)
measures. MeasuresonCartesianprodutsofLCSCspaesarealwaysonsideredwith
theirorrespondingprodutBorel-algebras.
Proposition 3.2. Suppose
i
;i=1;2are random measuresand
i
;i=1;2arep.p..