4.2.1 User Seletion in time-orrelated hannels
Considerthat the hannel exhibits orrelationfrom oneshedulingtime slot to theother.
Evidently, in suh ongurations, the sheduling deisions exhibit in turn some form of
orrelationoversuessiveintervals. Inotherwords,hannelorrelationinthetimedomain
reates temporal redundany, whih an be exploited asmeans to either redue feedbak
rate or inreasethesystemthroughput. If thehannel varies slowly,then learlythebest
user in termsof hannel qualityat urrent time slot
τ
ishighly likelyto be thebest userat the subsequent time slots
τ + T c
. Therefore, the fat that previously seleted usersarehighly likelyto remaingoodanbefurtherexploitedduring theuserseletionproess.
Temporalorrelationhasbeenexploitedinpaketswithdesign,eitherbyusingamaximum
weightmathingalgorithm[81℄orbyarandomizedalgorithmexploitingtemporalorrelation
of queue states [82℄. In [83℄, the authors proposed arandomized sheduler that exploits
temporalorrelationsinslowfadinghannels.
4.2.2 Beamformingand Sheduling exploitingtemporalorrelation
Sine shedulingand linearpreoding is theleitmotiv of this dissertation,weaddress the
problemhowtoexploittemporalorrelationandenhanethesum-rateperformaneof
low-novelSDMA sheduling/preoding sheme, oined asMemory-based Opportunisti
Beam-forming (MOBF). Theshemebuildsonmulti-beamrandombeamforming[9℄presentedin
Setion 2.9.3, and exploits memory in thehannel asameans to ll the gap to sum-rate
optimality.
Inanutshell, MOBFreplaestherandomseletionofpreodingmatrieswitha
ombi-nationofrandomandpastfeedbak-aidedbeamformingmatriesthat arekeptinmemory.
Theshemean beseenassuessiverenementof thepreodingmatrixinside the
oher-enetimeofthehannel. Whentheoherenetimeofthehannelishigh(e.g. largeDoppler
spread),MOBFapproahesthesumapaityofoptimalunitarypreoderwithperfetCSIT.
Forunorrelatedi.i.d. hannels, theperformaneof theproposedshemeremainssuperior
to that of [9℄ at the expense of moderate additional feedbak (two SINR valuesper user
insteadofone).
Interestingly,theshemeanbeseentoalsorelatetoreentusefulresults[84℄,presented
toimprovethedelayperformaneofthesingle-beamopportunistibeamforming[53℄. In[84℄
shedulingislimitedtooneuserandtemporalhannelorrelationisexploitedthroughthe
useof a xed set of beamsdetermined in advaned. This sheme does notautomatially
reah theperformaneof afull CSIT senario,sinethe temporal orrelationwasusedto
restore fairness and users with long waiting times are prioritized. Another sheme that
exploits temporal orrelation in orthogonal frequeny division multiple aess (OFDMA)
systemshavebeenproposedin [85℄.
4.2.3 Memory-based Opportunisti Beamforming
Asstated before, MOBFbuilds onrandom beamforming (f. Setion 2.9.3),in whih the
transmittergeneratesat eahtime slot
t
aB × B ( B ≤ M )
unitarypreoding matrixQ (t)
randomly, as a means to redue the feedbak burden and omplexity requirements, i.e.
W (t) = Q (t) = [ q 1 (t) . . . q B (t)]
. InonventionalRBF[9℄,anewrandomunitarypreoding isgenerated and used forserving theseleted usersat eah time slot. Hene,anykind ofstrutureinthephysialhannelisnotexploited. Memory-basedOpportunisti
Beamform-ing attempts to exploit memory in the hannel by making at eah time slot animproved
seletionof theunitarypreodingmatrixbasedonpastCQIinformation. Temporal
orre-lation is exploited by memorizingthe previous best shedulingdeision(s), i.e. thegroup
of seleted users
S
for arandom preoderQ (t)
, and omparing it with the next randommathings
Q(t + i)
fori = 1, . . . , T c
.Speially,weonsiderthattheBShasaodebook(set)of`preferred'unitarymatriesof
size
U
:Q = { Q 1 , Q 2 , . . . , Q U }
(4.1)with
Q ⊆ U (M, M )
,whereU (M, M )
denotestheunitarygroupofdegreeM
,i.e. thegroupof
M × M
unitarymatriesdeningtheomplexStiefelmanifold. Thenotionof`preferred' is used in the sense of (relative) maximizationof the sum rate among pastused randombeamformingmatries.
Ateahtimeinstant
t
,theunitarymatrixofthepreferredset,denotedQ ˜
anddenedasthepreoderthathasprovidedthehighestsumrateinprevioustimeslots,isappliedanditssum
rateismeasured(updated)underurrenthannelonditions. Theahievablesumrateof
Q ˜
attimeslot
t + 1
isomparedwiththat ofanew,randomlygeneratedunitarymatrixQ r
,andthebeamformingmatrixthatoersthehighestsumrateisseletedfortransmission. In
thephaseofupdatingtheodebook
Q
,thesumratevalueofQ ˜
intheodebookisupdated,and thenewlygeneratedrandompreoder
Q r
isaddedinto theodebookifandonlyifitssumrateishigherthanthesumrateof theodebook matrixwiththeminimumsumrate.
Let
S t
denote the set of seleted users at eah sheduling windowt
andH ( S t )
be theorrespondingsubmatrixof
H = [h T 1 . . . h T K ] T
. WithR (Q, S t )
wedenotethesumratewhenunitary beamformingmatrix
Q
isused forservingthe usersbelongingtoS t
. The stepsoftheproposedalgorithmareoutlinedinTable4.1.
Table4.1: Memory-basedOpportunistiBeamformingAlgorithm
Memory-basedOpportunisti Beamforming(MOBF)Algorithm
First phase: (`best'unitarymatrixseletion)
Step 0Initializeodebook
Q = { Q 1 , Q 2 , . . . , Q U }
,eahwithsumrate
R (Q i ), i = 1, . . . , U
Ateahtimeslot
t
,Step 1Generateanewrandompreoder
Q r
Step 2Selet
Q ˜ ∈ Q
:Q ˜ =
argmax
Q i ∈Q R (Q i )
Step 3Apply
Q ˜
,olletupdatedfeedbakfromtheusersandalulate
R ( Q, ˜ S t+1 )
Step 4If
R ( Q, ˜ S t+1 ) > R (Q r , S t )
,Q ∗ → Q ˜
,elseQ ∗ → Q r
Seond phase: (Updateofodebook
Q
)Step 5Updatethevalue
R ( Q) ˜
inthesetQ
Step 6If
[ R (Q r ) > R (Q min )]
,Q min → Q r
,whereQ min =
argmin
Q i ∈Q R (Q i )
)Some ommentsare in order: The algorithm outlined in Table 4.1 presents a general
frameworkformemory-based,randomizedshedulingin slowtime-varyinghannels. First,
in pratie, at eah time slot
t
thesetQ
ontains only onepreoder matrix(U = 1
), i.e.the onethat has provided the highestsystem throughputup to the urrent time instant.
Seondly,althoughMOBFisbasedonRBFforpreodinganduserseletion,ourproposed
shemeisnotonlyrestritedtosuhsystems. Theideaofmemory-basedpreodinganbe
also applied to systemswhere the usersutilizeaodebook toquantize theirhannelsand
feed bak quantized CDI. If the hannel is strongly orrelated, the above oneptan be
used to redue the feedbakloadby dereasingthefeedbak reporting rate. At eah slot,
additional CDI is then fed bak only if it is suiently dierent than the one previously
reported. Alternatively,ifweenforeCDIreportingateahtimeinstant,usersmayhavethe
PerformaneAnalysis
Theunderlying ideabehindthe sum-rateanalysis of MOBFis the following: theproess
of memorizing at eah sheduling slot the sum-ratemaximizing preoding matrixan be
seenasarandomsearh ofbeamformingongurationsin thespaeof orthogonalunitary
preoders. Evidently, the performane ofsuh sheme depends on thedistribution of the
sumrateonditionedto aertainhannel realization
H ( S )
forthe seletedgroupofusersS
. Tosimplify ouranalysis,wexthehanneloftheseleteduserstoaertainrealizationH
andweanalyzethepropertiesofX i = R ( Q i , H )
,whihrepresentsthesumrateprovidedby random unitary matries
Q i
for a given hannelH
. Therefore,{ X i } ∞ i=1
is a randomproesswhosedistributiondependsontheunderlyingrandomvariable
Q i
. Forxedhannelrealization,
X i
is i.i.d. fori
withassoiated PDFf X ( · )
andCDFF X ( · )
. Ifthe hannel isquasi-stati,memory-basedbeamforming aimsat ndingthe unitarybeamformingmatrix
Q ∗
from thefeasible setof unitarymatriesU
that maximizes thesumrate. This anbemathematiallywritten as:
Q ∗ =
argmax
Q i ∈U R (Q i ) =
argmax
1 ≤ i ≤|U| X i
(4.2)Note that this optimizationreturns oneoutof possibly many globalmaximizers
Q ∗
sinetheglobal maximizer isnot unique, i.e.
R (Q ∗ ) = R (Q ∗ Q ′ H )
, for anyQ ′ ∈ U
. However,themaximumvalueofthesumrate,
X ∗ = R ( Q ∗ , H )
, isuniqueoverthesetU
.Assumingthat the set ofunitary matries
U
isnite with ardinality|U|
, then for|U|
i.i.d. randomunitarymatries
{ Q i } |U| i=1
,theahievablesumrateX ∗
isgivenbyX ∗ = max
1 ≤ i ≤|U| X i = Z ∞
0
xdF X |U| (x)
(4.3)Forasymptotiallylarge
|U|
, thedistribution ofmax
1 ≤ i ≤|U| X i
onverges-afterpropershiftingandsaling-toalimitingdistribution(l.d.) ofGumbel,FréhetorWeibulltype. However,
astheexatform of the CDF
F X (x)
is diultto obtain, theexatl.d. is diultto beinferred. Hene,weresortto thefollowingresultin ordertoderivetheasymptoti(in
|U|
)onvergeneofouralgorithm.
Proposition 4.1: Consider ahannel with memory
L = T T c s
, whereT c
is the hanneloherene time, and
T s
is the slot duration. ForL → ∞
, the sum rate of memory-based beamformingR MOBF
onvergesto the apaity of optimum unitary beamformingR ∗
for agivenhannel
H
:R MOBF → R ∗ = max
Q ∈U R ( Q i , H )
(4.4)Proof. Theproofisgivenin Appendix4.A.
The above result implies that the maximum of the sum rate oered by using various
preoders
Q i
onvergesasymptotiallytotheoptimumapaityofunitarybeamformingR ∗
.As aresult, the orrespondingunitary preoding matrix, denoted
Q ∗
, whih orrespondstothematrixthat maximizesthesumrateonvergestooneofthepossiblymanyoptimum
unitarypreoders. Therefore,ifthehannel isquasi-stati(verylarge
L
),theodebook ofMOBFwillontainanoptimalbeamformingmatrix,i.e. aunitarymatrixthat maximizes