InSetion2.5,wepresentedshemesthatdealwiththeoptimizationoftheinputovariane
matriesorthepreodingdesign. Inthissetion,adierentapproahisfollowedandwetry
toidentifytheoptimalseletionofuserstobeserved. FollowingtheseminalworkofKnopp
andHumblet[1℄, multiuserdiversity reeivedan inreaseattention in theeld ofresoure
alloation for wireless networks, shattering the traditional viewof fading asdetrimental.
Inthis work,the authorsprovided novelinsights to thequestionof `whih user should be
servedinordertomaximizethe sumrate'andgaveriseto anovelsetoftehniques,oined
as opportunisti ommuniation. Simply speaking, opportunism reommends sheduling
thebest user(i.e. theuser withthe mostfavorablehannelonditions) in eahoherene
intervalinordertomaximizethesystemthroughput.
ConsideraMISO
K
-userbroadasthannel, forwhih thesum rateapaityis upperbounded by
Clearly,thesumrateismaximizedwhenonlythestrongestuserisassignednon-zeropower
P k = P
,i.e.C BC = E
Traditionally, hannel fading was viewed as a soure of unreliability that has to be
mitigated. Animportantmeanstoopewithfadingisdiversity,whihanbeobtainedover
time (interleaving of oded bits), frequeny (ombining of multipaths in spread-spetrum
orfrequeny-hoppingsystems)andspae(multipleantennas). Thebasiideaistoimprove
performanebyreatingseveralindependentsignalpathsbetweenthetransmitterandthe
reeiver. Theseminal work of [1℄gavetheidea that in theontextof multiuser diversity,
fadingan be onsideredasasoure ofrandomizationthat anbeexploited. This isdone
bydynamially shedulingtransmissions (resoures)among the usersasa funtion of the
theirpeaks. Inonephrase,underopportunistitransmission,wetransmitwhenandwhere
thehannel isgood.
2.7.1 Asymptoti Sum-rate Analysis with Opportunisti
Shedul-ing
Multiuser diversity is a form of diversity inherent in wireless networks, provided by the
independent time-varyinghannels arossthedierentusers. Themultiuser diversitygain
omes from thefatthat theeetivehannelgain,denoted as
g k
, isimprovedfromg k
tomax 1 ≤ k ≤ K g k
. Theamountofmultiuserdiversitygaindepends ruiallyonthetailof thedistribution of
g k
, implying that the heavier thetail, the morelikelythere isa userwitha verystronghannel, andthe largerthemultiuser diversitygain. Therefore, thehannel
statistis has an impat on system throughput. In the following, we derive the sum-rate
growthfordierenthannelgaindistribution.
Fading
Weonsider thatthehannelsofallusersarei.i.d. Rayleighfading,thus
g k
ishi-squareddistributed with
2M
degrees of freedom, i.e.g k ∼ χ 2 (2M)
ifg k = k h k k 2
org k ∼ χ 2 (2)
ifg k = | h k | 2
. The limiting distribution(l.d.) of ahi-squarerandomvariable is of Gumbel typeanditanbeshownthatthemaximumvalueofK
i.i.d.g k ∼ χ 2 (2M )
randomvariablessatises[9℄
Pr { log K + (M − 2) log log K + O(log log log K)
≤ max
1 ≤ k ≤ K g k ≤ log K + M log log K + O(log log log K) }
≥ 1 − O 1
log K
(2.48)
Therefore,forlarge
K
,max
1 ≤ k ≤ K g k
behavesaslog K
withhighprobability,thusR ≈ log log K+
log P + o(1)
. Thelargerthenumberofusers,thestrongertendstobethestrongesthannelandthelargerthemultiuserdiversitygain.
Itan beeasilyshownthat thelimitingdistribution of Rieanand Nakagami fadingis
ofGumbeltype. However,themultiuserdiversitygainissigniantlysmallerintheRiean
ase omparedto theRayleighase. Exponentialandgammadistributions also belong to
themaximumdomainofattrationofaGumbeldistribution.
Log-normal Shadowing
Weonsidernowthattheeetivehannelgain
g k
isdominatedbylog-normalshadowing.It an be shown that the maximum value of
K
i.i.d. log-normal distributed r.v. with logarithmimeanµ s
andvarianeσ 2 s
,satises[48℄Pr { b K − a K log log K ≤ max
1 ≤ k ≤ K g k ≤ b K + a K log log K } ≥ 1 − O 1
log K
where
b K =
exp{ √
2 log Kσ s + µ s }
anda K = b K σ s / √ 2 log K
.Hene,thethroughputsaleslike
R ≈ √
2 log Kσ s + µ s
Pathloss
Consider now a more realisti senario, in whih the users are loated randomly over a
ellgivenby a disk of radius
R
around the serving BS. The hannel gainonsists in theprodutbetweenavariable representingthepath lossand avariable representingthefast
fadingoeient,i.e.
g k = L k γ k
,whereL k
isthepathlossbetweenuserk
andtheBS(f.Set.2.1),and
γ k
istheorrespondingnormalizedomplexfadingoeient.We onsider auniform distributionof the population in eah ell. Thus the distane
betweenuser
k
andtheBS,d k
,isar.v. withnon-uniform distributionf D (d)
givenbyf D (d) = 2d/R 2 , d ∈ [0, R]
(2.49)Further,therandomproess
d k
anbeonsideredi.i.d. arossusersandells,ifusersineahellaredroppedrandomlyineahdisk. Theonsideredoverageregionanbeassimilated
withtheinside areaofeah disk, inadisk-pakingregionofthe2Dplane. Usersdropped
outside the disks andropped from the analysis, as these will not aet the saling law.
Assuming
R = 1
fornormalization,thedistributionofL k = βd − k ǫ
isgivenbyf L (x) =
( 2
ǫ ( x β ) − 2 ǫ 1 x
withx ∈ [β, ∞ ) 0
withx / ∈ [β, ∞ )
(2.50)
The distributionof
L k
is remarkable in that it diers strongly from fast fadingdistribu-tions,due toitsheavy tailbehavior. Formally,
L k
followsaPareto-typedistribution andis aregularlyvaryingrandom variablewith exponent− 2/ǫ
, i.e.lim
t →∞
1 − F α (x)
1 − F α (tx) → t 2/ǫ
. Aninterestingaspet of regularly varying r.v. is that they are stable with respet to
multi-pliationwithotherindependent r.v. withnite momentsaspointedoutbythefollowing
theorem:
Theorem2.1[49℄: Let
X
andY
betwoindependentr.v. suhthatX
isregularlyvaryingwith exponent
− η
. AssumingY
has nite momentE { Y η }
, then the tail behavior of theprodut
Z = XY
isgovernedby:1 − F Z (z) → E { Y η } (1 − F X (z))
whenz → ∞
(2.51)Theideabehindthistheoremisthatwhenmultiplyingaregularlyvaryingr.v. withanother
onewith nite moment, oneobtainsa heavy tailedr.v. whose tail behavioris similar to
therst one,upto a saling. Sine
γ k
hasnite moments, thetailbehaviorofg k
anbeharaterizedby:
1 − F g (x) → E { γ k 2 ǫ } β
x 2 ǫ
when
x → ∞
(2.52)Therefore,
g k
isalsoregularlyvaryingwithexponent− 2 ǫ
,whihimpliesthat[50℄K lim →∞
Pr
{ max
1 ≤ k ≤ K g k ≤ β E { γ k 2/ǫ } ǫ/2 K 2 ǫ x } = e − x −2/ǫ ∀ x > 0,
(2.53)Usingtheaboveresult,wean showthatthethroughputsalesforasymptotiallylarge
K
as
R ≈ ǫ
2 log K
(2.54)Observe that a muh greater throughput growth than in the ase of fading is obtained.
loss aross the user loations in the ell. As the distribution of pathloss belongs to the
maximum domain of attration of Fréhet type, alogarithmi apaitygrowth with
K
isahieved. However,theshedulingdeisionsaretakeninaquiteunfairfashion admittedly,
sinethesheduler tendstoseletusersloserto theaess pointasmoreusersareadded
to thenetwork.