Performance Analysis ofCell-free Massive MIMO

Systems under Limited Feedback

Jia nfcng Shi
Xudong Yin Jianxin Dai
Nationa lMobile Communications
College ofTetecomnmnications School a/Science
ResearchLaboratory. Southeast
& Information Engineering Nanjing University oj Postsand
University
,\'wy"ingUniversity ofPostsand Telecommunications
Na njing 2 10096. Chin a
Telecomm unications Nanjing 2 10023. Chi na
shij ianfeng-1!:sc u,edu,cn
Nanjing 2 10003. China .\ ·ationa l.\fobife Communications
26977 18806 1!:qq.com ResearchLaboratory: Southeast
University. Nanjing 2100%. Ch ina
daijx'1!;njupl.cdu.cn

Ab.ftruct- I n this paper , we consider ace ll-freem assive C urrently, some studiesh ave been performedo n the cell-
mu ltiple-inpu t mu ltiple- outpu t (:\ 11:\ 10 ) sys temw hi chco ntai n s free massive MIMO systems. [12}a nalyzedt he minimu m
a large num ber ofa ccess point s (APs)a nd a much smaller rate and5% -outage rate performance under thecondi tion of
numbero f users. Eac h APob tain s th e chann el param eters imperfectch annel informatio n and assumedt hatt he
estimatedfr om t he uplink quantization bits and applies complete ly orthogonalp ilots area llocated to the users. In
maximum ratio tr ansmis sion (:\IRT) inthe downlink [ 13]. thea uthors deriv ed the closed-form express ion for
t ransmission.Durin g the tr ansmi ssion of feedback, we ccnstder downli nk achievable rateu nder the Ray leigh fading channel.
a qu antization techniquea ndlimit edca pac ity, We redesignth e
[ 14] proposeda random vectorq uantization code book design
eodebookwhic h co ntai ns channe l direction inform atio n (C[)J )
method. In[15 ]. the effect of large-scale fadin g was added to
a nd channel q ualityinf ormation (CQI) . We derivet he dosed-
the originalG rassma nn codebookw hich minimizes chord
form expre ssion for t he down link ac hievab le rate a nd a pp ly
max-min power controlal gorithm to guaranteeth e users ' data d istance to design ac odebookb ased onfu ll channel
ra tes. In order toin crease the perform ance • .... ea lso propose a n information. Althoug h this codebooki ss uboptimal,i t is
optimal bita llocation a lgorithm. Simulationres ults show that mores uitable for distributed antenna systems (DAS). [16]
th e pr oposeda lgorithms ca na chieve good system performance. analyzed the statistics o ft he quantized channela ndt he real
channe l. [17] proposed that thech annel quantizat ion needs
Keywords-s-Cett-free m assive MUdD, ch(lnlleJ (Ju(lnti:atitm, feedbac k channel directio n information (COl) and channel
codeb(wk d e ~ · i1( l . tlchievable rate. maximize radio transmission quality information (CQI), while it did not ind icate the
feedback ratio of CQ I and COl. Tot he best of our
I. INTRODUCTION knowled ge. inc ell-free massive MIMO systems,th ere has
beenn ore searc hes on limitedfee dback.
Due to the advanta ges ofhi gh energy efficiency and
spectru m utilization, large-sca lem ultiple-input multiple- Inthi s paper . weco nsider ace ll-free massive MIMO
output (MIMO) has been considered as ap romising 5G syste mu nderthe ass umptionsth at its channel obeys
wireless access technolo gy [1-5]. In gene ral, there are mainly Rayleigh fading and APs haveno channel statei nformatio n
twoca tegories of massive MIMO : the co-loc atedma ssive (CS I). First,w ere design the codebooka nd carryo n channel
MIMO andd istributed massive MIMO . where antenna arrays estimation by channel quantiza tion [18·20 ). Then, we derive
arelocate da t the bases tation and spreado ut over aw ide the closed-form express ion ford ow nlink achievable rate and
area respectively [6- 10]. Since the anten nas of the d istributed tackle with the power andbi t allocat ion problem. Lastly. we
mass ive MIMO are spread outov er a largea rea. it can propo se the optimal bita llocation algorit hm to improve the
reducet he impacto f shadow fadin g, and its coveragei s rate performance.
wider than co-locat ed massiveM IMO at the cost o f higher The rema inder o ft his pape r is organized as follows.
backhaul [11].H owever. both of thet wo categories inherit Sectio n II descri bes thece ll-free massive MIMO down link
the shortco ming oft he traditiona l massive MIMO system. system. In Section III. we redesign the codebooka nd carry
i.e., thep erformance is typica lly limited by intra-ce ll onc hannel estimation . In Section IV, we derivet he closed-
interference. T hus in this pape r, wedi scuss a newma ssive form expression for datar ate and solvet hepo wer andbi t
MIMOmo del, called cell-freem assive MIMO. whichi s free a llocationprob lem. In Section V • We provide numerica l
from ce lls. results and somedi scuss ions. Finally, Section VI conc ludes
In [ I]. the authors considered the distribu ted mass ive this paper.
MIMO systemsw here plentyo fa ccess points( APs) are
Thes ymbols usedin this article are as follows: ( A l' ,
d istributed overawi de areaa nd serve a muchsm aller
num ber of users. AllA Ps arecon nectedt o thece ntral ( A)H repr esentt he matr ix transpose and conjugate transpose .
processing unit( C PU) via a backhaul network. All usersg et
service froma ll APss imultaneously. Sinceno cell is
E {o} • IHI denote Eucl idean norm and expectation
invo lved, whic h is the origin of its name. Ob viously. iti s a respect ively.
combinat ion oft he co-locatedm ass ive MIMO and
distributed massive MIMO. thereforeita lsoi nherits the II. C ELL-F REE M ASS[VE M IMO S VSTHI MODEL
advantages of them. Furthermore, sincet he users are clo ser
We consider a ce ll-free massive MIMO systemwi th
tot he APs,i tca np rovide higher coverage performance. And
alsod ue to the largerco verage. antenna selection for each M APs and K (M » K) users, both of whicha re randoml y
userb ecomes a very important issue. locatedi n a large area. Eachu ser ise quipped with a single
antenna and APs equipped with N , antennas , respec tively.

978-1-7281-2373-8/19/$31.00 ©2019 IEEE

All A Ps areco nnected toth eC PUv ia ab ackhaul network . [ I I]p ropo sed cod ebook desig n schemewit h co nsidering
Thel inks from the A Ps tot heu sers andtheu sers to the A Ps LSF, shown as follo ws:
share thes ametime -freq uency resource under T OO mode .
I). According to thed istribution charac teristics of fast
Thesys tem model is shown inFi g. I.
fad ing cha nnel, it des igns a root codebook {F" ..., F, } based
onG rassman n schemeor Rand omVe ctor Quantizat ion
(RVQ )s cheme,whi ch is used to quantifyth e colu mn space
of the fast-fadin g directio nm atrix .
2). Accordingtot he movi ng speed,p eriodica lly replace
the large-scale fading matrix , andth e root codebook
{FI' ..., FI } is rotat ed by A pi yie ld ing Ui = A pi F; .

3).The glo bal codebook is {WI" '" WI J where

,
W,oU,(U; U,p .
B.Ana lysis a/bitd ivision /actor
During the limited feed back transmission , the tota l
AP7 number ofv alid feedbackbit s fora ce rtainu ser is defined
as BkHal , includ ing the number of C Ol feed back bits B("/>/
and CQ I feedback bits B ("(! / . A. is used to represe nt the bit
d ivision factor, then.

Fig. l . A cell-free massive MIMO system.\\hcrc thc dashed and solid Bn J/ =).B."""
( I)
lines represent backhaul and access links. respectively. B l "Q1 =(1- ).) B m"" ,

Thed ifferen cebet ween ce ll-free massiveM IMO and Basedon the conclusion of [Ill, the realc hannel vector
traditionalm assiveM IMOi s thatallA Psse rvethetar get h ml canb e decomposed as
user simultaneous ly intheform er. Ino ne coh erence interva l,
itn eeds tocom pleteth ree transmi ssio ns: the uplink
quanti zation bits transmi ssion. the downlink signal
transmission andt he uplin ks ignaltran sm iss ion.Us ers send
where s..... representsth e direction infonnation of theC Ol
qua ntizationb its tot he AP s first and then A Psc hoose
channe ls betw ee nu sers and A Psfro mt hec odebook . Int he quan tizat ion error, sca lar z.... represe nts the magnitude of the
seco ndt ransmission , A Psp rec ede the transmitteds ignals C O l quant izatio n errorand
with the obtained channel statei nformation in thefi rst
transmis sion andu sers sends ignals toth e APs. Toa void (3)
excha nge of CSI between AP s andu sers. we adopt the
maximumra tiot ransmissi on (MRT) in the dow nlink and Noteth at sm.! is distributedi n theze ro spaceo f h....
maximum rat io combining(M RC) in the uplink . In the
follo wing, we mai nlys tudyt he firstt wop hase s. iso topicall y, thus E [s... J = 0 However, z... mea ns the
acc uracy of C Ol and
III. CIIANNELQUANTlZATION
E {=... }= N~- I r b",·,. / ( N,- l ). (4)
Suppo se thatt he user'sc han nelv ector h... is ,
decomposed into gml and timl , where gml = llhmlll IS CO l
Ass ume thatt he CQ I Ilh",. II/." is quantized byq mi' g iven
and 11 m• = h..l /llhmlll isC Q I. In adownl ink mu lti-user MIMO
by
system,CO li sg ene rally used tod esig n a precodin g mat rix
toeli minate inter-user interference and obtain multi-user
diversity ga in.C Q I isg enerallyu sed for ada ptive resource
allocatio nincludin g powera llocation , subcarrierallo catio n wher et heva riable e..l isth e quan tizat ion error. whic h
andan tenna selectio n, etc. belo ngs tot he rounding error and e m.! is uniforml y
I, 1
A. Codebook Design d istributed in [_2X,VJ - 2I1, VJ - ] . Thus,
In general, the codebookd esign is basedo n COL
Howe ver, in this paper weneedto take intoacco unt both E{ , . ,} o O. (6)
COl and CQ I. Theco debook designba sed on all infonnatio n
indica tes thatt he large-scalefa ding (LSF)i s alsoc onsidere d
in the cod ebook desig n scheme. Due to thes low ly chan ging
E (( 'N - E {eN lJ} 2'''' /3 . (7)
natu re of LSF, iti s reasonable to assu me thatthet rans mitter T hen theC SI obta ined by the feedback isw ritte n as
and receiver have perfect knowledge of LSF. The literature
 (8)
(16 )
C. Channel estimation
Boththe transmitter andth e receiver keepth e same
channel codebook which is mentioned abovean d we
assume that thetra nsmitter know s perfect CS l.B y
introduce chordal distance as a metric . the math ematica l .If I
re prese ntat ion of the quantization processis VI .II' .!. 14 "L..'7..i"I · h .. . h.- .' I '
- VPj' ( 18)
.. =1
h... = W,= arg min d 2 ( h... .w) (9) Wetreatt he second, thirdandfo urtht erms in ( [5 ) asth e
W, E!W,.. _." ,! ' effectivenoise ', andthe firstit emi s uncorre latedwi th other
where d ( h..l ' Wi) represents the distance between h... and terms in the worst case. Soweca n obtain theachievablerate
Wi ' The transmitterse nds the position of quantized channel of thek -th userf or cell-free massive MIMO under limited
to there ce iver and the receive r searches the channelfrom feedback
thecodeboo k accord ing top osition.

I V. D OWN LINK SIGNAL TRANSMISS ION R" = log1 + IDS• • I'] ( 19)
Weap ply the MRT preceding for thedow nlink signal ,. a [ E1IBV.I'l+t, E(IUiul'l +1
transm ission. With imperfect CSI. APs can only regard the
channel esti mates asthetrueva lues. The receiveds ignal of Thus, the closed -form expressionisded uced to
thek ·th user is givenb y

rd.• =
"
Lh.. 1 x .. + W d. l • ( 10)
.. ~ I

where w d .• - eN (0, I) is add itive white Gaussian noisea nd (2 0)

x.. IS the signaltra nsmitted bythem-th AP.
Mathematically ,

( 11) where
2
where P. with £ {Ipln= I is thes ignal intended to the A.."= '1... ( «: + q ~ . 2- ; "", ) X...' (2 1)
k-th user, '7mt is the powercon trol coefficient which should
satisfythefo llowing inequality X
... N,,
:..!!:Jl.
= 1_ N, -1 2 /.', - l +_1_
N 2
:.!!:J!!..
" ,- 1
' (22)

Ell, .I')sp j. (12)

Basedo n ( 11), ( 12) can be rewritten as
B
...
= n
' f",.
q'
.. J
[
1_ N , -I 2
N.
, ""'- J
/.',-1 (2 3)
<
L 'l... P ~ . =::;I . m =I. 2.···.M. (1 3)
Proof: See Appendix A.
. =1
Weca n seefro m (19) thatth eperform ance isn ot only
Substituting (I I) into(1 0). the rece ived signa l can be related tot he power allocation factor '7... . buta lso the bit
further expresseda s
number ofC Ol and CQI. Also we can see thatw hen 8cm
rd.• = "
Lh... x .. + W d .1
and BelJl becomevery large, all items containing A orB will
.. - I become O. It meansthat nothing todowit hth emand it also
( 14)
is consistent with thequantitat ivem odel.

V . PO \\.'E R CO NTROLAN D BIT [)lVISION

Based on (14), thek ·thu sercandet ect the
data PI from rd .l • Throught hema x-min powerco ntrol, cell-free massive
MIMO can service everyu ser fairly regardless of their
Next, wew illd erive a closed-formexp ression fort he locat ion. Mathematica lly, the max-minpo wer allocatio n
downlinkac hievable rate.Weass ume that each useronly problemi s formulated as
knows estimated CSI rather than perfectC SI.Th en, the
max min R ef
rece ived signa l rd.• in( 14) canb erewr itten as [21·22] Iq.. 1 ' = I..... K d ,J

, ,
rd.• = DSlq. + BUlq. + L U I.II,q« + W d .l , ( 15)
subjectto L fJ...q;" =::; 1 m = I,2, ··, M (24)

".
hi

'1... ;:: 0, k =I,..., K. m= I. ..., A-l.

where
 Theprob lem in (24) is quasico ncave (see [8, Proposition low-ra teare as. Mo reo ver, the max-m in powerco ntrol
ID, which ca n beso lved by the bisectionme thod (see [23, algorithm ca n significant ly improvet hen etwor k th rough put
Ch.4.2.5)). sincet hat max-min power co ntrol algor ithm can make the
performance o feach user moreave rage andim prove the
Reca llt hatthetotal number of bits is B""" I ' where so me to tal performance of the sy ste m. T he results we ll reflect the
of the m are used to quantify CQI, and othe rs are used to advantages of max-mi n power co ntrol algorithm .

--,,----
quantify C Ol. Now , weneedtocon side r howtoa llocate bits
eac h part to improves yste m performance. Ass ume that all
0.9 _ _ r.we......, __
usershavethe same b it allocation factor. Toimp rove the
performance of the entire system, we ai matmax imizingt he
overall rate tofi nd o ut theo ptima l bit allocationf actor .
o.
, ~ 07

L R':f
•
~ 06
max
n,'II' H,1)/ h i d .• ~ 0.5

subject to 0 < Bn!1 < 8 """1 (25) ia H

0< s.; < B.,." 03

o.a
Be/){ + B(VI s: BMal'
0'
It is clear that prob lem in (25) isal inea r op timization coL ~"-- ---" ~_- ~," ,--- - -~,, ,--- -- -~~-_ -.J
proble m and the optima l value is achieved when the third
P... ·user dow"'in k lhroogf>put( MbiUslHz)
constraint iseq ual. Additio nally, theop tima l solut ion canbe
obtained by di ffere ntiating .
I'ig. 2. Cumulative distribution ofper-user downlink throughput.
When weapp ly power allocation and bita llocation , we
first considerp owe r allocatio n anda ssume thattheb its are
allocat ed even ly. Thenco nsider the b it a llocat ion. Here we
ca n prov ide a newa lgorithm thatweset Bn.! 1= 0 : B""" I . c.a
When B(VI changes, wes o lve (24)and thenc hoose th e o.
bestso lution. j 0.1
E
:Ii 0.6
Co mparing thesetwoa lgorithm s, we findth at the •
~ 0.5
com plexityo f the newa lgorithm isal most B"ool times that
of the origi nal algor ithm for thatwe dono t co nsider the ~ o.~
com plex ity of so lving (25), butt he newa lgorithm ca n get 8 0.3
the betters o lution. Co nsidering that theco mp lexit y of 0' _ _ QWmOl b;j_..__ago_connd
__ QWmol b;jallocabon& opbaml _oonIt<Il
so lving (24) is verylow ,s otheco mp lex ity of thenew 0.' __ ,,_b;j_...-....geo_connd
algor ithm is sti ll acceptable. --QWmolb;j-...-....geo--

•
0
• e 1012 14
P ... ·use< dow nlink lhroog~
16 16
MbiUl>!Hz)
20 22 24

V I. SIMU1.ATlON RESUU
In thiss ection, weprov idet hen umerical results tover ify Fig. 3. Comparison of different hit allocat ionsc hemesand power control.
the performance of our proposed algor ithms. In a
Ikrn x I kmsq uare area, all A Ps andu sersa re di stribu ted Fig. 3 shows thec umu lative d istr ibution of the per-user
random ly. T he LS F coe fficie nt p,... is given by the COST downlinkth roughp utunde r different power con trol andb it
a llocat ion schemes.Weaddt he influence oft hebit
lI ata model:
allocation factor in this figurec ompared w ith thefigure
P"", = I O - IJ 6-3 IIOS,.(d.. j, X.. 1IO• (26) above . We can see from itt hat the performance of max- min
power control scheme is better under the sameb it allocation
where d",. is th edista nceb etween th em oth A P and th e schemea nd the performance of optimalbita llocatio n
k- th user in ki lometers . lOu .. I IO) repres ents the shadow scheme is better under the sam e power control scheme. The
effect with X mJ - N (O,a ,z'-I ) ' Un less othe rwise specified, cond ition for achieving thebes t performa nce isa pply ing
max-minp owerco ntrol and optima l bita llocation scheme.
thevar iance a ~ isset to8 dB. T he carrier frequency T he reaso n is thatt he max -min powe r co ntro l sc heme
is 1,. = 1.9 G Hza nd the bandwid th is B = 20M Hz. The ens ures that each user' s perfor mance tends tobe consisten t
ma ximumt ransm it powerforeac h A Pi s 200 mW.Th e and opti mal b it allocation sc heme ai msa ti mprovi ng the
numbe rs of APs and usersare 100 and 40 , respectively . overall performance of all users. T hese two algorithm s have
d ifferent purposes. Sowecanach ieve best res ults by
Fig. 2 show s the cumulative distribu tion of theper- user co mbining the m.
down link th ro ughput under the max- min power con trol and
average power a llocation when the quantization b its are Fig. 4 showsh ow the per-use r throu ghput cha nges with
allocat ed uniform ly. Iti sse en that the throu gh put of the then umbero fq uantizatio n bits. Wec an seefro m itcle arly
averagepowe r a llocat ion a lgorit hmh as awide range that theth rough put of opti malb ita llocationi s alway s better
distribu tion andt he d ifference betweenth e maximum than that of averageb it allocatio n. As then um bero f
throughputand the m ini mumt hroughput is very large. In quan tiza tionb its increases, the per -user's throughput also
ad ditio n, a largep ercentage of users are co ncentrat ed in increases. In addition , whe nth en umber of quantizati on b its
reac hes ace rtain value, thepe r-user throughput of optimal complex ity of the channele stimation and power allocat ion
bit allocation scheme becomess table, while that of average a lgorithms willal so decrease.

VII. CONC LUSION

Int his paper , we analyze the performance of cell-free
mass ive MIMO under limited feed backtaki ng into accoun t
the effects of the number of bit, bit allo cationf actor, channe l
estimation , power control and complexitya nalysis.
Wer edesig ned thec odebook andd erived closed-for m
rate expression. First we appliedt he max-minp owerc ontro l
method in each AP to guarantee user's performance. Then
wes tud iedt he impacto fb ita llocation factor andpropo sed
the opti ma l bit allocation scheme toi mprov e the system
a performance. The results demonstrated the performance
,,;--~--:;:---:;---;~-:::--~---:;---: impro veme nt canb e achieved by our proposed algorithms.
o s rn 15 20 25
Howev er,w e consider that thet ransmitter knows theCS I
Numberol b~
perfectlya nd we canf urtherd iscuss the imper fect situatio n.
Fig.4. Per-user throughputun der differentnumber o fb it.
ACKNO\V(.EIX.iMENT
bita llocation schemei ss till increas ing slowly. Besides, the T he resea rch wass upportedinpa rtb yN ationalNatura l
d iffere nce of per-user throughpu t betwee n differe nt sche mes Scie nce Foundation of China (61701254),Na tural Scie nce
increases with the number o f bits first and then decreases. Foundation of Jian gsu Province (BK20 170901), the open
The reason is thatw henth en umber of bits iss mall,t here research fund of Natio nalMobil eCo mmunicationsR esearc h
arefew allocated plans andthepe rfomance is sirnilia r. Laborato ry, SoutheastU niversity(20 (901 5), andNU PTSF
When the number of bits increases,wec an getma ny (NY218038).
allocation plans. Thus, thep erformanc e o f optimal
allocationpla n willbem uch bette r that of average bit Append ix A
allocation.Whenthen umbe r ofb itreac hes a large value, the DS. isca lculated by
number of bits allocat ed toea ch phase iss ufficie nt, sothe
gap becom e small.

" ,-- - - --:;---- - ------- ---,

"
0.75

8~ 0.7

0.65
(27)
ce
E(IBu,l'l
.. w ---:;,----:;- --::-- ---:;---;o-- --.J.
" '" 20 JO
Nu_ oIb f
40
<E{M(t."i.h.,h:'- E{t."ih..h:,}]'j (28)
Fig. 5. Bilqua nufkation factor versus thenumberofquantizenon bus.

Fig.5s hows howt heb it allocationfa ctor changes with

the number of bits. When the numbero f bits iss mall, bit
=t."m(E (Ih.,h:.I')-IE{hmh:.l')
p1
allocation facto r changes quickly. And thenit floats up and
do wn inave rys mall rangea s the numbe r of bits is bigger ll :. 1' l
E h• • •

=E11(qm+em)( ~ k m hm+,[z:;"'m )q•• q')

than 18.As the number of bits increase, the range become
small andfi nallyi t will approach a stable value. The reason
is thatwh en the numbero f bitsiss mall, itwillallocat e all
the added bits toon esi de which can contr ibute more to the
performance, and whenthe numbe r of bits beco mes larger,
= E(q:. +q ~ . e~ . HI ~ l- z••h;.+,[z:;"'mh:.I')
q:. +q ~. 2';")(E 11~
itwillalso allocate addedb its totheoth ers ide.
Comparing Fig. 4a nd Fig. 5, weca n findth at we needto = ( h :'I ') +E 11,[z:;",.,h:.I'j)
use almost 30bit s toquan tize thechannelwhe n we allocate
theb itse venly, but we only need almost halfwh en weapp ly
the proposed algorithm and thee ntirebit overhead has
=(9'mk""
+q2 2 - 2R, "'
3
)( I _N NN
, - 1 2 ~:" ~ + _ 1_2 - N~ !i )
dropped by half. Ina dd ition tot her educedov erhead, the
, ,
(29)
where J ..k isin theNU LL space of h... and it is a unit na nn [6] J . Wang , II. Zhu , andN . Gome s, " Distributed antenna systems for
mobile communications inhighspeedt rains." IEEE Journal on
vecto r. Selected Areasin Communications.v ol. 30. PI'.67 5-683 , May 2012

[
17]H,Zhu , "Performance comparison between distributed anten na and
1£{ h mJ i : kW ~q ~ . I _ N~ ~1 2 ~!i , - 1 ) (30) microccllular svstems," IEEE Journal on Selec ted Areas in
Communications , vol. 29, pp. 1151· 1163, no. 6, 2011.
[8]II,Zhu , S, Karachonrzuis and D.T oum pakans. "Lo w-complexity
Then Resource A llocationand us Application to Distri buted Anten na
E(IBU,I' l Systems :' IEEE Wireless Communications. \ '01. 17. no.3, pp. 44-50,
June, 20 10.

-_PJ'f '"" (
L.. l7... q...
.. ; 1
2 -~" ,.- )
• + q..2k -
J
X (3 1)
(91 C. Pan. II. Zhu, N. J Gomes , and J wang, "Joint l'recodingand RRII
selection for User-centric Green MIMOC-R AN," IEEE Transactio ns
on wireles s Communications, vel. 16. pp. 2891·2906 , no. 5, May
20 17.

1- N, - 12 -N~ :; +_1_ 2 -:"31 ) - q ~ k [ 1- N, - 12 ~:"! ; ) (lO ] 11 . Wei. D. Wang, II.Z hu, J. Wang, S Sun. andX. 11 . You. "Mutual
coupling calibrat ion for multiuser massiveMIMO systems. " IEEE
[ N, N, N, Transactionson Wireless Communications. vnl.15 , pp. 606-619,
January 20 16.

+'. 1') [IIJ S,Zhou , M. znao. X. Xu, J Wang, andY . vee . " D i s t r i bu t ~'d wireless
communication system: A new architect ure forfu ture public wireless
acce ss," IEEECommlln. Mag., vol. 41 , no, 3, pp. 108- 113, Mar,2003 .

p1 l l ~ " i h ~ h ~ n
= E (12] E. Navcbi. A. Ashikhmi n. T, L Marzella. and 11 . Yang." Cell-free
massive MIMO systems :' in Proc. 4 7111 Asilomar Co nference all
S ignufs. Syslems and Computers (ACSSC), Pacific Grove. CA USA,
= p ~f E fi iJl7"",l7...h....h.. h:k.h : ,) 1'1'.695---699, Nov. 20 15,.
1"'1 ", I (13] H, Q. Ngo. A. As hikhmin, 1-1, Yao g. E. G,L a rsson, andT . L.
Marzella," Cell-freemassiveM IMO versu s smallcells : ' JEtE T ran.~
= pJ E { ~~ ,[ri: ( q ;,. .q ~ . + q ;' , e ~ , ) Wireless Commun., Jan . 2017,
[14] C. K, Au- yeung, S. Member, andD,J . Love. "Onthe Performance of
Random vector Quantization Limited Feedback Bcemtorrmng in a
:0:( JI-zork, hork, ii::,+ F:;s ork ,ii::.)(JI - :""ii.. ii:, + .,J;;s .. , i ~ .) } M ISO System " [EEE Transactions 00 Wireless

2 )(1-,v ~ ~ 1 2 ~
Communications , \01. 6, no.2, 1'1'. 458-462, 2007.
= PJif}....( q~ k' +q;". ,2B,,,, + ~2.\, ~ ] (I S] X. Chen , Z. Zhang, and I I. Che ri, "Ondistr ibuted ante nna systems
., , 1 3 .\, with limited feedback preceding opportunities and
challenges , "wireless C ommunications, vo1.l 7, no. 2, PI'. 80-88,
(3 2) April20 lO.
Substit uting (27), (3 1), and (32) into (19),w e obta in (20). (16 ] N,Jindal. ""MIMO Broadcast Channels With Finite-Rate Feedback."
IEEE Transaction s onInfor mation Theo ry, \'01. 52, no. I I, ['I' 5045 -
506 0,2006,
R EFERE NCES
(1 71 D l.i. etc., "Multi-cell hybrid channel infromationf eedback for
downlink mult i-cel1 preceding, " Wireless Communications and
[ I]H . Q. Ngo . A. Ashikbmin, II. Yang. E. G. Larsson, and T. L Networking Conference(WCN C ), 1'1'. 1340·1345,20 11.
Marze lla. "Cell-f ree massive MIMO: Uniformly great servic e for [18] II,Zhuand J. Wang,"C hunk-base d resourceallo cat ion in OFDMA
everyone." in Proc.I EEE 1m. Workshop Signal Process. Adv. sy-stems • Part I: chun k allocation: ' IEEE Transactio ns on
Wirt'!ess CO/ll/llun, (SPAWC), Stockholm, Sweden, pp. 201-205, June Communications . \'01. 57. no.9 . Pl'. 2734-274 4. Sept. 2009.
201 5. (19] 11, ZhuandJ . Wang, "Chun k-based resourc e allocat ion in O H )M A
[2] V. Jungnickcl eta ll." Therole of small cells .co ordinated multipoint. systems - Part[I : joint chunk , power andbita llocation : ' IEEE
and massiveMIMO in 5G." in IEEE Communications Magazine. \'01 Transactionson Communicat ions,v ol. 60, no.2 , pp. 499-509, Feb.
52, no. 5, pp. 44·5 1, May 20 14. 20 12.
[3]A L. Swindlehurst, E. Ayenoglu.P . Hcydan andF .C apolinu [20J II, Zhu. "Radio resourc e allocation for OFDM A systems inh igh
"Millime ter-wave massive MIMO : the next wireless revolution "," in speed environments," IEEE Journal on Selected Areas in
IEE E C ommunicationsMagazin e, vot. 52, no. 9. pp. 56-62 . Communications, \ '01. 30, pp. 748-759 , May 2012 .
Septe mber 20 14. [21] B. lIassibi and n. M.Ho chwald. "Hnw much training is neededi n
[4] C. XWang et all. "Cellular architecture and key technologies for 50 multiple-antenna wireless link s?" IEEE Trans. In! Theory, \' 01. 49, no.
wireless communication networks ." inI EEE Communications 4, Pl'. 951-963, Apr. 2003.
Magazine , \'01. 52, no,2 . pp. 122-130. February 20 14.
(22J 11, Yang andT . L Marzcua, "Capacity performance of multiccll
[5] F. Boccardi, R,W . l leath ,A Lozano. T.I Marzcua and P. Popovski, large-seale antenna systems :' in Proc. 5 1s1 AllertonC onf Co r omun.,
"Five disruptiv e technology directions for5G ," in IFE E Control. Comput., Monticello, IL. USA, pp. 668--675 , oe. 20 13.
Co mm unications Magazine. \'01. 52, no.2 , pp. 74-80. February 201 4. [23] S. Boyd and L. Vandenberghe. Convex Optimization Ca mbridge.
U,K,: CambridgeUniv.Pre ss, 2004,