3.7 Beam Power Control with Beam Gain Information
3.7.2 Beam Power Alloation for more than two beams
Forthegeneralaseof
B > 2
beams,ananalytialtreatmentof(3.20)doesnotunfortunately seemtratable, beauseofthe lakofonvexity. Therefore, weproposehereasuboptimal- yet eient- iterativealgorithm that aims to inrease systemthroughput by alloating
poweroverthebeams. Thealgorithmtriestoidentifytheextremepointsofthesumrateand
ndthepowervetor
P
thatmaximizes(3.20). Theextremumofthesumratefuntionanbe foundanalytially using Lagrangianduality theory and onsidering the
Karush-Kuhn-Tuker(KKT)onditions. LetWloG
B = M
anddenetheobjetivefuntionG (P) =
Inordertosolvetheoptimizationproblem
max
wemayformulatetheLagrangianfuntion as
L ( p , µ, ν) = G ( p ) +
where
ν ≥ 0
andµ ≥ 0
aredualvariables. Theostfuntion isneitheronvexnotonavewithrespetto
{ P m } M m=1
,thereforeaglobaloptimalsolutionforanyhannelmodelishardtoobtain. However,KKTonditionsareneessaryforextremum,whether loal orglobal,
of
G (P)
. BydierentiatingwithrespettoP m
,wend∂ G (P)
∂P m
+ ν m − µ = 0, 1 ≤ m ≤ M
(3.33)P m ≥ 0, 1 ≤ m ≤ M P − X
m
P m ≥ 0
TheKKT onditionsare neessaryand suient ifand onlyif the Hessian of (3.32) is a
negativedenitematrix. Forsuhlassofhannels,aglobalmaximumisidentiedthrough
theKKTonditionsabove. Forgeneralhannels,theKKTpointsanbeaglobalorloal
maximum,asaddle-point,orevenaglobalorloalminimum.
Iterative BeamPower Control Algorithm
Performingtransformationoftheprimalproblem(3.20)intoitsdualandsolvingthelatter
byKKTonditionsdoesnotguaranteeglobaloptimalprimalsolution. Astheprimalisnot
aonvexoptimizationproblem, thereouldbeaduality gap. Nevertheless,weproposean
iterativealgorithm,inspiredbytheiterativewater-lling(IWF)algorithm[73℄andtheKKT
solutionof (3.31), as ameans to identify the extremepointson the boundary of
P M
. InthisIterativeBeamPowerControlAlgorithm,eahuseriterativelymaximizesitsownrate
byperforming single-userwater-llingand treatingthemultiuserinterferenefromall the
otherusers(beams)asnoise. Clearly,ouralgorithmdoesnotseektondaglobaloptimum,
howeveritanprovidesigniantsum-rateimprovement.
AlgorithmI Let
P (0)
betheinitialpointandI (P (i) ) = σ 2 + P
j 6 =m P j (i) η k j j
betheinter-ferenefuntionat
i
-thiteration. Thestepsof thealgorithmaresummarizedin Table3.1.IterativeBeamPowerControlAlgorithm
Step1(Initialization)Set
P (0) = 0
Step2Foriteration
i = 1, 2, . . .
,ompute∀ k m ∈ S
:λ (i) k m = I (P η km m (i−1) ) = η kmm
σ 2 + P
j6=m P j (i−1) η kj j
Step3(Water-lling): let
π (i)
bethesolutionof:π (i) =
argmax
π ≥ 0, P
m π m ≤ P
X
k m ∈S
log 2
1 + π m λ (i) k m
Step4(Update): let
P (i) = π (i)
Table3.1: IterativeBeam PowerControlAlgorithmforSum-RateMaximization
Someobservationsareinorder:
At eah iteration
i
, oneλ (i) k m = η km m
(σ 2 + P
j 6 =m P j (i − 1) η km j )
is alulatedfor eah user
k m
usingP j (i − 1) , j 6 = m
,itiskeptxedandtreatedasnoise. GiventhetotalpoweronstraintP
,the`water-llingstep'isaonvexoptimization problem similarto multiuserwater-llingwith
ommonwater-llinglevel. Thus,alltransmitpowersin
P
assignedtobeamsarealulatedsimultaneouslyinordertomaintainaonstantwater-llinglevel. Thealgorithmomputes
iteratively thebeam power alloation that leadsto sum rate inreaseand onverges to a
limitvaluegreaterorequaltothesumrateofequalpoweralloation. Formally,thepower
assignedtobeam
m
atiterationi
yieldsP m (i) = [µ − 1/λ (i) k m ] +
,withX
k m ∈S
[µ − 1/λ (i) k m ] + = P
,where
µ
is theommonwater-llinglevel. Thebeampowerontrol for strategy3assigns transmitpowersoverthebeamsaordingto theiterativesolutionwhen theahievedsumrateishigherthanthatoftheboundarypoints.
Convergene Issues As stated before, no global maximum is guaranteed due to the
lakofonvexityofsum-ratemaximizationproblem. Therefore,wedonotexpetthat the
onvergene pointof theiterativealgorithm begenerallyaglobal optimal powersolution.
Interestingly, it an be shown that the onvergene leads to a Nash equilibrium, when
onsidering that eah userpartiipatesin anon-ooperativegame. Theonvergeneto an
equilibrium pointanbeguaranteedsine
I ( P )
isastandardinterferenefuntion [73,74℄.The proof of existene of Nash equilibrium follows from an easy adaptation of the proof
in[75℄. However,theuniquenessoftheseequilibriumpointsannotbeeasilyderivedforthe
aseofarbitraryhannels.
Letusnowderiveanalytiallytheonvergenepointofthe2-beamaseusingthe
itera-tivealgorithmandompareitwiththeoptimalbeampowersolutiongivenbyTheorem3.2.
Atthesteadystate,sayiteration
s
,wehavethat( P 1 (s) = µ − 1/λ (s) k 1
fromthesumpoweronstraint. Upononvergeneofthealgorithm,
wehave that
P i (s) = P i (s − 1) , i = 1, 2
, whih results into a systemof equationsAP T = b
It an be observed that (3.35) is dierent from (3.25a). Fortunately, it still provides a
heuristipoweralloationalgorithmandasshownthroughsimulationsinSetion3.9,there
is not a signiant redutionin sum rate by alloating the powerover beams using this
algorithm.
Reinterpretationin termsofSuessiveConvex Approximation
Inthis setion, weresortto GeometriProgramming (GP) [72℄ whih representsthestate
beome averypopular and powerfultehnique asit provides eient solutions in power
ontrol problems with non-linear objetive funtions and spei SINR onstraint, by
re-vealing thehidden onvexitystruture. Furthermore,the proposed solutionsareveryfast
and numerially eient, often exhibiting polynomial time omplexity. In partiular, we
apitalizeontheso-alledsuessive onvexapproximation (SCA)tehnique[72,76℄,whih
isshowntobeonvergentandturnsoutthat itoftenomputesthegloballyoptimalpower
alloation. Interestingly, the heuristi iterative algorithm proposed in Table 3.1 nds an
equivalentinterpretation,sineapplyingSCAtoourbeampowerontrolproblemresultsin
thesameiterativealgorithm. Werstlowerbound
log(1 +
SINR)
intheobjetivefuntionforsome
a
andb
[76℄:log(1 +
SINR) ≥ a log(
SINR) + b
(3.36)Applying(3.36)into theoptimizationproblem(3.20)resultsintherelaxation
max P
whih stillremainsanon-onvexproblemsinetheobjetivefuntion isnotonavein
P
.However,usingthetransformation
P ˜ m = log(P m )
wehavethefollowingonavemaximiza-tionproblem:
DeningtheLagrangianfuntion as
D ( ˜ P , λ) =
weonsiderthedualproblem(3.38)thatis
min
λ max
P ˜ D ( ˜ P, λ)
. Thedualsolutionoftheinnermaximizationproblem is given by the stationary point of the Lagrangian funtion (3.38)
with
λ
xed. DierentiatingwrtP ˜ m
andapplyingtheinversetransformationP m = e P ˜ m
weformthefollowingxed-pointequation
∂ D
Remarkably, this xed point-equationprovidesthe samepoweralloation algorithm asin
Table3.1for
a m = 1, ∀ m
(wlog)where thepowersanbeupdatediterativelyusing(3.39).However,wenotethatazerodualitygapannotbeguaranteedformallyduetolakof
on-vexity,implyingthat notheoretialargumentanshowonvergenetotheglobaloptimum
forgenerallassofhannels.