Knowledgeof the hannel stateby the transmitter of aommuniationssystem hasbeen
demonstratedto bebeneialtowirelessommuniations,partiularlyinmultiuserMIMO
systems. The majorimportane of theavailabilityofhannel knowledge hasbeenalready
reognizedin [7℄, bypointingout that lakofperfet CSITresultsin total lossofdegrees
offreedom, in ontrast to what happensin single-userormultiple aess MIMO shemes.
The MIMO BC with no CSIT is degraded no matter whether the reeivers have CSIR
ornot. Hene, when the users havethe same number of antennas, it anbe shown that
superpositionodingisthesameastime-sharing. Therefore,thesumapaityisthesame
asifthereisonlyoneuserinthesystemandnomultiuser diversitygains anbeexpeted.
Thesigniantdiereneonthesumratebehaviorbetweenmultiuserandsingle-userMIMO
revealstheardinalroleofCSITinmultiuserMIMOdownlinksystems,presentedindetail
inthefollowingsetions.
2.6.1 Channel Knowledge at the Transmitter
Inmultiuser MIMO literatureit is often assumedthat the reeiverenjoys lose-to-perfet
hannel knowledge, whereasthe transmitter hasdierentlevelsof CSIT, ranging from no
CSITatalltofullCSIT.Theassumptionthatthereeiverhasauratehannelinformation
isoftenreasonableespeiallyin thedownlink,wherepilotsymbol-basedhannelestimation
ismoreeientsinetheterminalsanshareaommonpilothannel. Channelaquisition
at the transmitter relies on hannel measurements at a reeiver, sine the transmitter is
informedbythereeiveron thehannel statein animpliitorexpliitway. Themethods
available to gatherCSI at thetransmitter mainly relyonhannel reiproityorfeedbak.
Insystemsforwhih hannel reiproity annotbeexploited,the needfor CSIT feedbak
plaes a signiant burden on uplink apaity. The feedbak load is further exaerbated
inhigh-mobilitysystems(suhas3GPP-LTE,WiMAX,et.) wherethehannelonditions
hangerapidly and in widebandsystems, wheremorefeedbaktrainingisrequireddue to
2.6.2 Capaity saling laws in MIMO BC systems
ThedominantroleofCSITin multiusermulti-antennasystemsanbeidentiedby
study-ing theasymptotiapaitygrowthunder dierentassumptionson CSIT.Speially, the
fundamentalroleplayedbythemultipleantennasinexpandingthehannelapaityisbest
apprehendedbyexamininghowthesumratesaleswiththetransmitpowerandthenumber
ofativeusers.
Full CSI atthe Transmitter
High Power Regime: Thesalinglawofthesum-rateapaityofMIMO BC forxed
N k = N
,M
,andK
andlargeP
isgivenby[44,46℄P lim →∞
C DP C
log P = min(M, max(N, K))
(2.42)The above result implies that at high SNR,the apaity exhibits linear growth with the
numberoftransmitantennas. Furthermore,thenumberofreeiveantennas peruserplays
verylittlerolein theapaityofMIMObroadasthannelsomparedto
M
(providedthatK > M
).Large
K
Regime: The saling law of the sum-rate apaity of MIMO BC for xedN k = N
,M
,andP
andlargeK
isgivenby[44℄K lim →∞
C DP C
log log KN = M
(2.43)The result in (2.43) indiates that, with full CSIT, the system an enjoy a multiplexing
gainof
M
,obtainedbytheBSseletingandsendingdatatoM
arefullyseletedusersoutof
K
(multiuserdiversity). Sine eah userexhibitsN
independentfadingoeients, the totalnumberofdegreesoffreedomformultiuserdiversityisKN
,thusgivingtheextragainlog log KN
.In ontrast, if the BS seletsand transmitsonly to the userwith maximumrate, the
apaityoftime-sharing,
C T S
,isgivenby[44℄K lim →∞
C T S
min(M, N ) log log K = 1
(2.44)Fromtheaboveresults,itisevidentwhytheapaitysalinglawsprovidetheneessary
justiationforthegreatappealofmultiuserMIMOsystems. Thespatialmultiplexinggain
of
M
,whihisthepre-logfatorofthesumrate,impliesalinear(inthenumberoftransmitantennas)inreasein apaityforno additionalpower. Theorrespondinggainis realized
by simultaneouslytransmittingindependent data streams in the same frequeny band to
spatially separableusers.
No CSI at the Transmitter
IntheabseneofCSIT,usermultiplexingisgenerallynotpossible,astheBSdoesnotknow
HighPower Regime: Thesaling lawofthe sum-rateapaityofMIMO BC forxed
N k = N
,M
,andK
,satisesP lim →∞
C DP C
log P = min(M, N )
(2.45)whih impliesthat at high SNR theapaity isessentiallythe sameasthat ofa
point-to-pointMIMOsystem. Inotherwords,TDMAisoptimalin thisregime.
Large
K
Regime: The saling law of the sum-rate apaity of MIMO BC for xedN k = N
,M
,andP
andlargeK
isK lim →∞
C DP C
log log KN = 0
(2.46)Inontrastto (2.43),there is nomultiusergainsinethetransmitter hasnoknowledge of
theusershannelsin ordertoexploitthem.
NotethattheaboveresultsholdundertheassumptionofperfetCSIR.Theimpatoflak
ofCSI at bothends of theMIMO network and in theasymptotially high SNR regime is
studiedin [25,47℄,where itis shownthat boththemultiuserdownlinkand thesingleuser
apaitysaledouble logarithmiallywiththeSNR.
Information theoretidesign guidelines
TheaboveapaitygrowthresultshighlightseveralfundamentalaspetsofmultiuserMIMO
systems,whihomeinmuhontrastwiththeonventionalsingle-userMIMOsetting. The
designguidelinesthat anbeextratedaresummarizedasfollows:
•
CapaitysalinglawsadvoateforservingmultipleuserssimultaneouslyinanSDMA fashion, with asuitably hosenpreoding sheme at the transmitter. Although themultiplexing gain is limited by thenumber of transmit antennas, the numberof
si-multaneouslyservedusersisinpriniplearbitrary. Howmanyandwhihusersshould
eetively beservedwithnon-zeropowerat anygiveninstantoftimeisthe problem
addressedbytheresourealloationstrategy.
•
Unlikein thepoint-to-pointMIMO setting,thespatial multiplexingofdierentdata streamsan bedonewhileusersareequipped withsingle-antennareeivers,thusen-abling the apaity gains of MIMO while maintaining low ost for user terminals.
Having multiple antennas at theterminalanthus beviewedasoptional equipment
allowingextradiversitygainforertainusersorgivingtheexibilitytoward
interfer-eneanelingandmultiplexingofseveraldatastreamstosuhusers(reduingthough
thenumberofotherusersservedsimultaneously).
•
ThemultiplexinggainofM
inthedownlinkomesattheonditionoflosetoperfetCSIT. IntheabseneofCSIT, usermultiplexing isgenerallynotpossible,astheBS
justdoesnotknowinwhih`diretion'toformspatialbeams. Thus,theompletelak
ofCSI knowledgeredues themultiplexing gainto one. Thisis akeydierene with
point-to-pointMIMO,inwhihtheasymptotiapaityisnotsensitivetoCSIT,and
exeption lies in senarioswith terminal devies having enoughantennas to remove
o-streaminterfereneatthereeiver(
N k ≥ M
). Inthelatterase,theBSmaydeidetoeithermultiplexseveralstreamstoasingleuserorspreadthestreamsovermultiple
users, ahieving an equivalent multiplexing gain in bothases. This is onditioned
howeverontheindividualuserhannelstobefullrank.
2.6.3 Partial Channel State Information
The oftenunrealistiassumption oflose toperfet CSIT, aswell asthe onsiderablegap
betweentheahievablesumrateoffullCSITomparedtothenoCSITase,havemotivated
researh workon shemesemployingpartial CSIT. PartialCSIT orlimitedfeedbak refers
to anypossibleform of inompleteinformationonthehannel. Thisterm inludes,butis
notlimitedto,salarCQIfeedbak(e.g. estimateofreeivedSINR),quantizedCSIT
(quan-tization ofhannel vetor),hanneldiretion information, statistialCSIT, et. Multiuser
MIMO shemesrelyingonpartialCSITlieat theheartofthis dissertation.
Thepratial,thoughsuboptimal,approahesdesribedinSetion2.5.1areshowntobe
highly sensitiveto hannel estimationerrors,thus diulttobeimplementedwithpartial
CSIT. The low-omplexity alternative of downlink beamforming and sheduling, despite
being less sensitive to CSIT imperfetions, requires full CSI as a means to minimize the
multiuserinterferene[12℄. Fortunately,worklike[9℄demonstratesthattheoptimalapaity
salingofMIMOBC(i.e.
M log log K
)assumingK
single-antennausers,anbeahievedforK → ∞
eventhoughthetransmitterreliesonsalarCQI.Severalshemesbasedonpartial CSITareshowntoahievelosetoDPCsum-rateperformaneinsomeasymptotiregimes.However,themajorityoftheseapproahesbeomeinevitablyinterferenedominatedathigh
SNR sine theerrorintrodued(and theinreasein inter-user interferene) dueto partial
CSIT saleswithSNR.Hene,inthelargepowerregime,suhshemesexhibitasumrate
eilingbehaviorandfailto ahievefull multiplexinggain.
Itwouldhavebeenawedtoonludethatpartial CSITleadsneessarilytoaollapse
ofmultiplexing gain. Thismultiplexinggainlossanbemitigatedbyusingavariable-yet
nite -ratefeedbak hannel. In[10℄,Jindal showedthat thefeedbakloadperusermust
inrease approximatelylinearlywith thenumberof transmitantennas aswell aswith the
transmitpower(indB)in ordertoahievethefullmultiplexinggain. Inthisthesis,wetry
toshedsomelightontheseissues,byproposingseveralrobustlinearbeamformingshemes
with limitedfeedbak. Theinterferene dominatedbehaviorofsuh shemesis studied in
detailandseveralofourproposalsprovidemeanstoirumventthesum-rateeilingeet.
2.6.4 Statistial Channel Knowledge at the Transmitter
AnotherkindofpartialhannelstateknowledgethatanbeobtainedattheBSwithlittleor
nofeedbakoverheadisthestatistial CSIT.As seond-orderhannel statistisvarymuh
slowerin time ompared to the hannel realization itself, expliit statistial CSIT anbe
onveyedperiodiallyto theBS resultingin littleuplink overhead. Impliit knowledge on
thehannelstatistisanbeobtainedwithoutanyadditionalfeedbakbyaveraginguplink
measurements(statistialreiproity).
•
Channel Mean Information (CMI), whih refers to the ase where the mean of the hannel distribution is available while the ovariane matrix is unknown and oftenassumedaswhite.
•
Channel Covariane Information (CCI), whih refersto the asewhere themean is assumed zero (as it is assumed to vary rapidly) and the information regarding therelativegeometry of the propagation paths is available through anon-white spatial
ovarianematrix.
Channelknowledgeaquisition using ovariane feedbakan be applied to bothtime
di-visionduplex(TDD) and frequenydivision duplex(FDD) systems. Inontrast to
deter-ministireiproityinTDD systems,thehannel statistisoftheuplinkand thedownlink
remainrelatedinFDDandthedierenebetweenthefrequenybandsanbeoveromeby
usingfrequenyalibrationmatrix. Long-termstatistial hannelknowledgeisassumedin
Chapter4, whereweshow howstatistial CSIT anbe ombinedwith instantaneous
low-rate CQI feedbak to inrease system throughput by seleting spatially ompatible users
withlargehannelsgains.