4.4 Exploiting Statistial CSIT in Spatially Correlated Channels
4.4.3 ML oarse Channel Estimation with CQI Feedbak
islarge. Asuboptimal,yeteient,greedyuserseletionshemeanbeusedinstead,similar
to the approahin [11℄. Here weextendthis sheme for MMSE linearbeamformingwith
long-term spatial information and instantaneous salar CQI feedbak
γ k
. The proposedgreedyuserseletionalgorithmisgiveninTable4.2. Inthisalgorithm,usersareaddedone
by one to the set. The omplexity anbe redued by onsidering only usersexeeding a
threshold
γ th
. Theuser with the highest CQI is examined at eah time, and it is addedto thesetof sheduledusers
S
onlyifitresultsinsum-rateinrease. Weshould notethattheoverallperformaneofgreedyuserseletiondepends heavilyonwhetherthepreoding
matrixanbereproessedeahtimeauserisadded,whihinturndependsontheformof
thehannelfeedbakavailableatthetransmitter.
4.4.3 ML oarse Channel Estimation with CQI Feedbak
Asstatedbefore,theorrelationmatrixprovidesusefulinformationaboutthespatial
han-nel harateristis, espeially if it is ill-onditioned, however it does notreveal any
infor-mation aboutthe qualityof the urrenthannel realization. Inorder to exploit multiuser
diversity,theshedulerrequiresproperlydesignedinstantaneouslow-ratefeedbak
γ k
,whihTable4.2: GreedyUserSeletionwithStatistialCSIT
GreedyUserSeletionwithstatistialCSIT
Ateahtimeslot
t
1. Initialize
S = ∅
andG = ∅
.2. Selettheusersthat exeedthethreshold
γ th
G = {∀ k ∈ { 1, . . . , K }| γ k ≥ γ th }
3. Selettheuserwith thehighestCQIvalue
k max =
argmax
1 ≤ k ≤ K γ k
S ← S ∪ { k max } , G ← G \ S
4. Repeat
k ∗ =
argmin
k ∈G Tr (
Ψ( ˆ S ) + βI − 1 2 )
S ← S ∪ { k ∗ } , G ← G \ S
until
|S| = M
5. Returnuserset
S
anbeameasureofthequalityoftheurrenthannel. Inthissetion,werestritourselves
toRayleighfadingorrelatedhannels,i.e.
h ¯ k = 0
,andweproposeasimpleframeworkinwhihlong-termstatistialhannelknowledgeisombinedwithshort-termpartialCSITas
ameanstoprovideaoarsehannelestimateateahslot.
MLEstimationwith BeamGain Information
Weadoptherethefeedbakstrategy1andonsiderthateahuser
k
feedsbakthesquaredmagnitudeofthe hannelwith abeamforming vetor
z k ∈ C M × 1
, i.e.γ k = h H k z k
2
. The
beamforming vetors an be interpreted as pilot signals during the training phase or as
the preferredbeamformer in a two-stage preoding and sheduling approah (see Setion
3.4). Thisbeamformeranbehosenrandomlyoritanbeoptimizedbasedonlong-term
statistialinformation.
Optimized training vetors As the orrelation matrix of eah user is known at the
transmitter side, the training vetors
z k
an be optimized. Briey speaking, an eienttrainingodebookanontain
N = N p + N l + N r
vetors,where theindiesp,l,rindiateprinipal,loal, and random,respetivelyasexplained below. Theodebook onstrution
1)Basedoneahuser'sstatistialCSIT,theodebookwillontain
N p
prinipaleigenvetors of theovarianematrixR k
(N p = 1
forMISOhannels).2)Inthisstep,weselet
N l
vetorsintheloalareaofeahprinipalstatistialdiretionv k
asameanstoaountforthosehannelrealizationsthatsteertheprinipalsingularvetor
in aloalityoftheprinipalstatistial diretion. Theloal areaoftheprinipalstatistial
diretion is dened by a one around
v k
and is haraterized by the angle between the trainingandtheprinipalstatistialvetors.3)Duringthethirdstep,wegenerate
N r
vetorsthatareoutsidetheonedened instep2andaountforthehannelrealizationsinwhihthediretionoftheprinipalrightsingular
vetor(orvetorhannel)isfarfromthestatistial(mean)hanneldiretion. Thesevetors
anbehosenrandomly orastheones that oversoptimallytheremainingspae, outside
theone,whihisrelatedtotheGrassmannianlinepakingproblem. Intheidealase,the
size of
N r
shouldbeadapted basedonthestrengthoforrelation,asitgivesameasure on thefrequenythatthesedeviationsour.Random training vetors Forsimpliity, werather adopt alow-omplexity approah
and onsider a randomopportunisti beamformingsetting [9℄. In this setting, we assume
that thevetors
z k
are isotropiallydistributed and hosenrandomly, i.e.z k = q m
where{ q m } M m=1
are the olumns of the unitary matrixQ
. We ombine the information ex-trated from the orrelation matrix with a salar instantaneous feedbak in the form ofγ k = | h H k ˜ q k | 2
,wherethevetorz k = ˜ q k
ishosenbyuserk
as˜
q k =
argmax
m=1,...,M | h H k q m | 2
(4.12)Clearly,thistypeofsalarCQIprovidesajointinstantaneous measureofthequalityofthe
urrenthannel realizationand its diretionof thehannelinstantaneously. Although the
amount of spatial information enapsulated into this metri annot be deomposed from
the hannel gain information, it is partiularly useful for userswith strong hannels, i.e.
users that are verylikely to be sheduled. It an bealso shown that the hoie of
q ˜ k
isequivalentto seletingthebeamoverwhih user
k
experiene thehighestreeivedSINRk
in [9℄. Assume that user
k
hasits maximumSINR on beami
outof thej ∈ { 1, . . . , M }
,Sine
f (x)
isalwaysmonotonouspositiveforx ∈ (0, c)
, wehavethati =
argmax
We propose aML estimation framework that ombines long-term statistial knowledge
andinstantaneousCSITprovidedbythefeedbakmetri
γ k
. Thisfeedbakallowsustopikuserswhose hannels span spatially separated ones of multipath and havegood hannel
gains. This so-alled Constrained Maximum Likelihood (CML) hannel estimate is the
onethat maximizesthelog-likelihoodfuntion of thePDF (4.5) onditionedto thesalar
feedbakonstraint
γ k = | h H k q ˜ k | 2
:h ˆ k =
argmax f (h | γ k )
(4.15)Thisresultstothefollowingoptimization problem:
max
h
k
h H k R k h k
s.t. | h H k q ˜ k | 2 = γ k
(4.16)
Itanbeeasilyshownthat(4.16)isequivalenttosolvingthefollowinggeneralizedeigenvalue
problem(GEV):
R k h k = λ Φ k h k
,whereΦ k = ˜ q k q ˜ H k
. Themaximumgeneralizedeigenvalue oftheHermitianmatrixpair(R k , Φ k
),withΦ k > 0
isdened asλ max (R k , Φ k ) = sup { λ |
det(λΦ k − R k ) = 0 } = sup
q 6 =0
h H k R k h k h H k Φ k h k
(4.17)
Thesolutionof(4.16),in theviewofthegeneralizedRayleigh-Ritzquotient,isgivenby
h ˆ k =
argmax
h k
h H k R k h k h H k Φ k h k
(4.18)
whihorrespondstothedominantgeneralizedeigenvetor,denotedas
u k
,assoiatedwiththelargestpositivegeneralizedeigenvalueoftheHermitianmatrixpair(
R k , Φ k
). Therefore,theMLhannelestimateisgivenby
h ˆ k =
√ γ k
| q ˜ H k u k | u k
(4.19)OrthogonalBasis expansion
ThesolutionoftheCMLestimateasageneralizedeigenvalueproblemrequiresthe
om-putation of the prinipal generalized eigenvetor at eah time slot, thus it may exhibit
remarkableomputational omplexity in pratie. In order to failitate thealulation of
the oarse estimate, we derive an equivalent hannel estimation framework in whih the
hannel of the
k
-th user is expressed as alinear ombination of orthogonal vetors.Al-thoughanyorthogonalbasisanbeused, intheaseofrandomtrainingvetorsitismore
naturaltohoosethebeamformingvetors
{ q m } M i=1
asourorthonormalbasis. Inthatase, thehannelvetoranbeexpressedash H k =
M
X
m=1
α m q H m
(4.20)where
α m
arethe(omplex) weightsoftheorthogonalexpansion.Consider, without loss of generality, that
q 1
orresponds to the best beam hosen byuser
k
. Substituting (4.20) into (4.16), and solvingtheoptimization problem (4.16)using Lagrangemultipliers,weobtainthattheoptimalweightsb opt = [α 2 , · · · , α M ] T
equaltob opt = − α 1 A − 1 c
(4.21)where
and
α 1 = √ γ k
sothat theinstantaneousCQIfeedbakonstraintissatised.Observingthesimilarityinthestrutureofmatrix
A
withthatofQ T R − k 1 Q ∗
,theomputa-tional omplexityof thematrixinversionof
A
anbefurtherreduedthroughuseofblokmatrixdeomposition. Denote
F = Q T R − k 1 Q ∗
,thenTheinverse
A − 1
anbeeasily obtainedusingtheequation:S A − 1
ML ChannelEstimation with Channel NormFeedbak
ConsidernowthattheinstantaneousCQImetritakesontheformofthehannelnorm,i.e.
γ k = k
hk k 2
. Clearly,theabovediretion independentCQI feedbakprovide less instanta-neous spatialinformation thanγ k =
h H k z k
2
. However,foruserswith largehannel gain,
thusforusersthataremorelikelytobeseleted,hannelnormfeedbakprovidesomeform
of additionalspatial information(espeially in Riean hannels). Moreover,the largerthe
hannel gain,the moreaurate this hannel diretional information. There is howevera
dierenebetweenRayleighandRieanhannels. IntheRayleighase,thesignambiguity
onthediretionannotbeeliminated,whereasin Rieanhannels(non-zeromean)thereis
additionalCDIonthesignoflargehannelrealizations.
SimilarlytoSetion4.4.3,weformulateaoarseMLhannelestimateassumingthatthe
hannel norm of user
k
is known, whih results in the followingonstrained optimization problem:max h k
h H k R k h k
s.t. k h k k 2 = γ k
(4.22)
Thesolutionof (4.22)isgivenby
h ˆ k = √ γ k u k
(4.23)where