2 o
. Thehannel gainof eahpath
φ p
isassumedtobezero-meanomplexGaussiandistributedandallpathshaveunit variane.
2.4 Capaity of MIMO Broadast Channels
Theompleteharaterizationoftheapaityregionofmulti-antennabroadasthannelwas
the foremosttheoretial hallenge in multiuserinformation theoryoverthelast veyears.
TheanalysisofbroadasthannelswasinitiatedbyCover[14℄andtheirapaityisgenerally
knownonlyinspeialases,wherethesignalssenttotheusersanbeorderedaordingto
their `strength'. In ontrastto single-usersystems wherethe apaityis asingle number,
the apaityof amultiuser systemwith
K
users isharaterizedbyaapaity region, i.e.a
K
-dimensional rateregion,where eah pointis avetorofratesahievablebyalltheK
userssimultaneously. Arate vetorisahievableifthere exists aoding shemefor whih
theerrorprobabilityforallusersisarbitrarysmallastheodebloklengthinreases. The
maximumofthesumoftheommuniationratesistheso-alledsum-ratepointandlieson
theboundaryoftheapaityregion. Clearly,sinethe
K
userssharethesamebandwidth,atradeoarisesbetweenthereliableommuniationuserrates: ifonewantstoommuniate
at ahigherrate,theotherusersmayneedto lowertheirrates.
A large lass of broadast hannels, known as `more apable' hannels [15℄, ontains
two important ategories as speial ases: `degraded' and `less noisy' hannels. Roughly
speaking,abroadasthannelisdegradedwhentheusersanbeorderedfromthestrongest
totheweakestinanaturalorder.Forinstane,aSISObroadasthannelisdegraded,sine
theusersanbeorderedaordingtotheir
| H k | 2
,andtheapaityregionanbeahievedbysuperposition oding [14℄. However,MIMObroadasthannels aregenerallynon-degraded
asthereisnotanaturalwaytoorderhannelmatries.
2.4.1 Capaity with perfet CSI at the transmitter
Although the haraterization of the general (fading) broadast apaity region is a long
standing problemin multiuserinformation theory,substantialprogresshasbeenmade for
Gaussian MIMO hannels. Despite not being degraded, the Gaussian MIMO BC oers
signiant struture that an be exploited to haraterize its apaity region. The key
theoretial toolfor haraterizing the MIMO BC apaity region with full CSI,the Dirty
Paper Coding (DPC), wasrevealed by the seminalwork of Caire and Shamai (Shitz)[7℄.
doesindeedahievetheapaityofa2-userMISObroadasthannel. Theresultsof[7℄were
extendedandgeneralizedby[1618℄,untilthefullharaterizationofMIMOGaussianBC
apaityregion(foranyompatsetofinputovarianesandnotonlyunderatotalpower
onstraint)byWeingartenetal.[8℄,establishingtheoptimalityofDPCasapaity-ahieving
strategy.
Assuming noise with unit variane and given a set of positive semi-denite matries
P k ≥ 0, ∀ k
thatsatisfythepoweronstraintTr n
TheDPC regionis givenbytheonvexhullofalltheahievableratesas
C DP C =
onvandisshowntobeequivalenttotheapaityregionofMIMO broadasthannel[8℄.
Theapaityexpression(2.29)anbesimplied asfollows:
C DP C = E H
TheoneptofdirtypaperodingwasintroduedbyCosta[6℄,whoshowedthatforasalar
GaussianhannelwithAWGNandaninterferingGaussiansignalknownnon-ausallyatthe
transmitter(butnotatthereeiver),theapaityisthesameasiftherewasnoadditive
in-terferene,orequivalentlyasifthereeiveralsohadknowledgeoftheinterferene. Inother
words,dirtypaperoding allowsnon-ausally knowninterferenetobe`pre-subtrated'at
thetransmitterwithnoinreaseinthetransmitpower. Assume,withoutlossofgenerality,
thattheenodingproessisperformedin asendingorder. Theenoderrstpiksa
ode-wordfor
i
-threeiver,andthenhoosesaodewordforreeiver(i + 1)
-threeiverwithfull(non-ausal)knowledgeoftheodewordintendedforreeiver
i
. Thus,theenoderonsiderstheinterferenesignalaused byusers
j < i
asknownnon-ausally and subsequently, thei
-thdeodertreatstheinterferenesignalausedbyusersj > i
asadditionalnoise.Uplink-Downlinkduality
Themain tool that failitatedthe extensionof theworkin [7℄ andsimplied the problem
of nding the apaity region of MIMO BC was the uplink-downlink duality, introdued
in[1719℄. Theoneptofuplink-downlinkdualityanbeseen,ingeneral,astheequivalene
between the performane of a lass of reeive and transmit strategies when the role of
transmittersand reeiversare reversed. This equivalene has beenobserved in seemingly
dierent ontexts in the literature. For instane, in point-to-point links, the duality is
that the apaityregion oftheMIMO BC,
C DP C
with poweronstraintP
is equalto theapaityregionoftheso-alleddualMIMOMAC,
C MAC
withsumpoweronstraintP
.C DP C (P, H 1...K ) = [
T r { P K k=1 P k ≤ P }
C MAC ( P 1...K , H T 1...K )
(2.31)where theunionistakenoverallmatries
P k ≥ 0 ∀ k
suhthatTr n P K
k=1 P k ≤ P o
.
The major benet of the uplink-downlink duality is that the apaity region of the
downlinkanbealulatedthroughtheunionofregionsofthedualuplink,whihisonvex
and whose boundary an be alulated using interior-point methods [20℄. An additional
benet is from an optimization theory point of view, sine by exploiting the duality the
dimensionalityoftheoptimizationproblemissigniantlyredued. Inmanypratialases,
the number of transmit antennas in the broadasthannel is greater than the numberof
reeiveantennas of anyof thereeivers. Therefore, insteadofoptimizing over
K
matriesofsize
M × M
,weneedtooptimizeoverK
matriesofsizesN × N
. Notethattheuplink-downlink duality only holds under a total power onstraint, and extensions of the DPC
optimalityto generalonstraintsettings(e.g. per-antennapoweronstraint)arebased on
themoregeneraloneptofmin-maxduality[8,21℄.
On theoptimal numberof users with non-zeroalloated power
Multiuser information theory advoatesfor transmittingto multiple users simultaneously
by properlydistributingthespatial dimensionsamong thebest groupof usersasameans
to boost the system throughput. A natural question that arises is how many users an
besimultaneously ative, and how the spatial dimensions are distributed among them. Yu
and Rhee[22℄obtainedatheoretialupperbound onthenumberofsimultaneouslyative
usersbyountingthenumberofvariablesandunknownsin thesetofKarush-Kuhn-Tuker
(KKT)optimalityonditionsforthesum-ratemaximizationproblem. Thisboundindiates
that inthedownlinkhannelmaximizingthesumrateentailsshedulingatmost
M 2
userssimultaneously. In pratie, simulations show that typially the number of ative users
is four times the number of transmit antennas in the high SNR regime using optimum
ovarianematries,and thatshedulingupto
M
users,although suboptimal, resultsto a small apaityloss. In[23℄,it wasindependentlyshownthat under ertainonditionsin avetordownlink with
K
usersanda BSwith twotransmitantennas, thenumberof usersthatanbesimultaneouslyservedanbehigherthantwo. Thepoweralloatedtothe
k
-thuser is no longer a water-lling proedure, but it is found by the KKT onditions. Note
that whenrestritingto linearpreodingtehniques,aswedointhis thesis,thenumberof
servedusersisdiretlylimitedbythenumberofdegreesoffreedomattheBS,i.e.
M
.2.4.2 Capaity with no CSI at the transmitter
The Gaussian MIMO BC with no CSIT is still degraded nomatter whether the reeivers
haveCSIRornot,assumingthatthetransmitterorthereeiversareequippedwithmultiple
antennas [7℄. In that ase, the apaity region is ahieved by superposition oding [24℄.
When the users have the same number of antennas, it an be shown that superposition
oding isthe sameastime sharing. In thisase, the sumapaity isthe sameasifthere
simultaneously. TheapaityregionoffadingMIMOBC isanopenproblemoftheoretial
interest. Theapaityregionisnotexpliitlyharaterized,andonly asymptotiallytight
boundsurrentlyexist. ThefadingMISOBCisonsideredin[25℄assumingthedistribution
ofthefadingoeientsisisotropi. Itwasshownthattheapaityregionisequivalentto
thatofthefadingsalarBC,resultinginamultiplexing gainofone. Whenthetransmitter
hasinompleteCSIonthefadingrealization,thepre-logfator(multiplexinggain