2012 IEEE International Symposium on Information Theory Proceedings

Pilot Contamination Precoding in Multi-Cell Large

Scale Antenna Systems
Alexei Ashikhmin Thomas Marzetta
Bell Laboratories Bell Laboratories
Alcatel-Lucent Inc. Alcatel-Lucent Inc.
Murray Hill, NJ 07974 Murray Hill, NJ 07974

Abstract—An LSAS entails a large number (tens or hundreds) base stations were proposed recently in [3]. Unfortunately, in
of base station antennas serving a much smaller number of all these techniques SINR values remain finite and do not grow
terminals, with large gains in spectral-efficiency and energy- indefinitely with the number of base station antennas M .
efficiency compared with conventional MIMO technology. Until
recently it was believed that in multi-cellular LSAS, even in In this paper we allow a limited collaboration between base
the asymptotic regime, as the number of service antennas tends stations and propose an outer multi-cellular pre-coding, which
to infinity, the performance is limited by directed inter-cellular we call pilot contamination pre-coding. In the asymptotic
interference. The interference results from unavoidable re-use of regime, as M tends to infinity, this precoding allows us to
reverse-link pilot sequences (pilot contamination) by terminals in construct interference free multi-cell wireless systems. We
different cells.
We devise a new concept that leads to the effective elimi- further proceed with analysis of LSASs with finite M .
nation of inter-cell interference in TDD LSAS systems. This The paper is organized as follows. First, in Section II, we
is achieved by outer multi-cellular pre-coding, which we call describe our system model, assumptions, and TDD protocol.
pilot contamination pre-coding (PCP). The main idea of PCP Then, in Section III, we explain the pilot contamination
is that each base station linearly combines messages aimed to
terminals from different cells that re-use the same pilot sequence.
problem and show that the pilot contamination results in inter-
Crucially, the combining coefficients depend only on the slow- cell interference that does not vanish as the number of base
fading coefficients between the terminals and the base stations. station antennas M is growing. In Section IV we propose
Each base station independently transmits its PCP-combined the pilot contamination precoding and show that it allows
symbols using conventional linear pre-coding that is based on achieving interference free TDD LSASs. Finally, in Section
estimated fast-fading coefficients. Further we derive estimates for
SINRs and a capacity lower bound for the case of LSASs with
V we derive estimates for SINRs and a capacity lower bound
PCP and finite number of antennas M . for the case of LSASs with PCP and finite number of antennas
M.
I. I NTRODUCTION
The exponential increase in demand for high data rates, as II. S YSTEM M ODEL
well as the higher user density in cellular networks require
We consider a cellular network composed of L hexagonal
new ways of mitigating interference allowing a larger number
cells, each consisting of a central M -antenna base station and
of users to share bandwidth. This, along with the Green Touch
K single-antenna terminals that share the same bandwidth.
initiative to decrease the power consumption in communica-
We assume that Orthogonal Frequency-Division Multiplexing
tions networks, motivates the analysis of cellular systems with
(OFDM) is used. Consequently, we consider a flat-fading
a very large number of base station (BTS) antennas. Such
channel model for each OFDM subcarrier. For a given sub-
systems have been studied extensively.
carrier we denote by
In [5], [1] the authors derived estimates for SINR values

in a non-cooperative cellular network in which the number of gmikl = hmikl βikl (1)
BTS antennas tends to infinity. It is shown in [5], [1], that not
all interference vanishes, and therefore, SINR does not grow the channel (propagation) coefficient between the m-th an-
indefinitely. The reason is that the channel estimates made tenna of the i-th BTS and the k-th terminal of the l-th cell,
at the BTS contain not only the desired channel vector and Fig. 1. The fist factor in (1) is the fast fading coefficient
additive white noise, but also components directed towards hmikl ∼ CN (0, 1). The second factor in (1) is slow fading
users from other cells who are assigned non-orthogonal pilot coefficient (log normal and geometric decay). Coefficients βikl
sequences. are constant with respect to frequency and BTS antenna index.
A number of techniques were proposed for achieving higher The channel coefficients of the i-th base station form the
SINRs. The numerical results obtained in [5], [1], [2] show channel vector gikl = (g1ikl , g2ikl , . . . , gM ikl ) ∈ C1×M , and
that these techniques provide breakthrough data transmission the corresponding fast fading coefficients form fast fading vec-
rates for noncooperative cellular networks. Advanced network- tor hikl = (h1ikl , h2ikl , . . . , hM ikl ) ∈ C1×M . We assume that
MIMO TDD systems that allow some collaboration between fading coefficients are i.i.d. and therefore hikl ∼ CN (0, IM ).

978-1-4673-2579-0/12/$31.00 ©2012 IEEE 1137

 where zi ∈ CM ×K is the additive noise. Without loss of
generality, we assume that the entries of zi are i.i.d. CN (0, 1)
random variables and that all gains are scaled accordingly.
The i-th base station estimates the vectors giki for terminals
√
located in the same cell as ĝiki 
= Yi · ψk† / τ ρr , k =
1, . . . , K, which results in
L

ĝiki = giki + gikl + zi , (3)
l=1,l=i

Fig. 1. The channel coefficient gmikl between the m-th antenna of the i-th z ψ†
where zi = √iτ ρkr ∼ CN (0, τ 1ρr IM ). The estimate (3) is
cell and the k-th terminal in the l-th cell
contaminated, as it involves the channel vectors of users from
different cells that use pilot ψk .
We further assume that the fast fading channels of different The base station uses the estimates ĝiki for computing
terminals are statistically independent. precoding beamforming vectors. For M → ∞ it is convenient
We assume a time block fading model. Thus fast fading to use conjugate beam-forming defined by
vectors hikl stay constant during coherence intervals of T †
wki = ĝiki . (4)
OFDM symbols. These vectors are assumed to be independent
in different coherence blocks. With this precoding the base station transmits with the power
We assume reciprocity between uplink and downlink chan- ρf ||ĝiki ||2 , which varies from one coherent interval to another.
nels, i.e. that βikl and hmikl are the same for these channels. For transmitting with a constant power the base station should
Let us consider a Time-Division Duplexing (TDD) scheme, use normalized beamforming vector:
† †
Fig. 2. Each coherence interval is organized in four phases: ĝiki ĝiki
wki = = √ . (5)
• first, each terminal sends uplink data to its BTS; ĝiki  αki M
• second, all terminal synchronously send pilot sequences The scalar
ĝiki 
of length τ ; αki = √
M
• base stations use these pilots to estimate the correspond-
ing channel vectors and use the obtained estimates for is a normalization factor.
maximum ratio combining processing of the uplink data After computing vectors wki BTS transmits from its M
antennas the vector
• base stations transmit downlink data to their terminals
using the channel estimates as beamforming vectors; vi = w1i q1i + . . . + wKi qKi ,
where qki is the signal intended to the k-th terminal in the i-th
cell.
The k-th terminal of the i-th cell receives the signal
L 
 K
√
xki = ρf glki wk l qk l + zki , zki ∈ CN (0, 1). (6)
l=1 k =1
Fig. 2. Coherence interval of T = 11 OFDM symbols
III. P ILOT C ONTAMINATION P ROBLEM IN LSAS
Below we describe details of this protocol for downlink The analysis of the asymptotic behavior of the SINR was
transmission. We assume that K terminals located in the same conducted in [5] for the precoding defined in (4) and in [1]
sell use orthogonal pilots. The maximum number of orthogonal for precoding (5). Below we summarize the obtained results.
τ -tuples is τ , therefore K ≤ τ . The coherence intervals We assume precoding (5). The following well known
of fast moving terminals are short, typically T ∈ [4, 30]. Lemma will be useful for us.
So τ should be small. Therefore users in the neighboring
Lemma 1. Let x, y ∈ CM ×1 be two independent vectors with
cells unavoidably use nonorthogonal pilots, or reuse the same
distribution CN (0, c I). Then
pilots. We assume that in all cells the same set ψ1 , . . . , ψτ of
orthogonal pilots (ψi ψj† = δij ) is used. While not optimal, x† y a.s. x† x a.s.
lim = 0 and lim = c. (7)
this choice completely eliminates intra-cell interference as M →∞ M M →∞ M
M → ∞, as we detail below. Using the fact that the channel vectors of different terminals
The k-th terminals in all cells synchronously send ψk . The are independent, and applying the above lemma, we can derive
i-th base station receives the M × τ matrix 2
the asymptotic behavior of αki :
 K
L  L

√  2 a.s. 1
Yi = ρr τ gikl ψ k + zi , (2) lim αki = βikl + . (8)
M →∞ τ ρr
l=1 k=1 l=1

1138
 Denote the terms in the double sum of (6) by Qk l and let • First, we assume that all signals qki are accessible to all
Sk l = |Qk l |2 . Again using Lemma 1, we obtain base stations across the entire network.
 2 • Second, we assume that the slow fading coefficients βikl
ρf βlki
Sk l a.s. , k  = k, can be accurately estimated and made available to all base
lim = α 2
kl (9)
M →∞ M 0, k  = k. stations or alternatively to a network hub.
Remark We assume that qki and βikl are accessible across
In (6) Ski is the signal power and all other terms Sk l ,
the entire network only to obtain a simple theoretical model.
contribute to interference. From (9) we see that only the
In a real system it will be enough if qki and βikl are available
terminals that use the same pilot ψk create interference that
only to a limited number of cell neighboring the i-th cell.
does not vanish as M → ∞. The reason for this is that the
We would like to point out that the slow fading coefficients
beamforming vectors of these terminals contain components
βikl are easy to estimate since they are constant over the M
directed towards the k-th terminal in the i-th cell generating
antennas, frequency, and over many time slots. In a typical
directed interference that does not vanish as M → ∞. The
system with 20MHz bandwidth and the useful symbol duration
variance of the additive noise, however, is unitary regardless
of 2/3 × 100 microseconds there are about 1400 tones per
of the number of BTS antennas, thus rendering the effect of
OFDM symbol. We assign to each terminal a dedicated OFDM
the noise null in the asymptotic region.
tone. The terminals transmit pilots of duration 1 (for instance
Equations (8) and (9) define the asymptotic behavior of the
ψkl = 1 for all k and l), simultaneously in the assigned tones.
numerator an denominator of the SINR, which are formed
The i-th base station receives in the appropriate tone: yi =
√
of independent variables. Therefore, we have the following 1 †
βikl h
ikl ψkl , and computes the estimate as β̂ikl = M yi yi .
theorem (analysis for uplink is similar).
It is easy to see that
Theorem 1. [5] [1] The downlink SINR of the k-th terminal 1 † a.s.
in the i-th cell for precoding (4) is lim y yi = βikl .
M →∞ M i
2
D βiki So we need only one OFDM symbol for getting these esti-
ςik = L , (10)
2
βlki mates. Terminals that are located in remote cells can reuse
l=1,l=i
the same OFDM tone. If this separation is not enough the
and for precoding (5) the SINR is terminals may use orthogonal pilots ψkl of length n. In this
2
βiki 2
/αki case the same estimation procedure can be applied and only
D
ςik = L , (11) every (1400 × n)-th terminal will be reusing the same pilot.
2 /α2
βlki
l=1,l=i kl Now we re-examine equation (6). Instead of treating the
L terms Qkl as interference and estimating their powers, we look
2
with αkl = j=1 βlkj + τ 1ρr .
For uplink transmission the SINR is at the terms themselves:
√
1 ρf
U
2
βiki √ Qkl = √ ĝlki wkl qkl
ςik = L , (12) M M
2
βlki   L 
l=1,l=i ρf βlki  † †
= hlki βlkl1 hlkl1 + zl qkl (13)
IV. P ILOT C ONTAMINATION P RECODING AND M αkl
l1 =1
I NTERFERENCE F REE LSAS
Applying Lemma 1 we obtain
The SINR values from Theorem 1 can be improved by √
different techniques (for instance, we can allow different 1 a.s. ρf βlki
lim √ Qkl = qkl .
transmit powers ρf,l ; their optimization gives a significant gain M →∞ M αkl
in SINRs [1]). In all these techniques, however, the SINRs From the above expression and (9 we get
do not grow with M . One may try to use network MIMO 
1 √ β1ki β2ki βLki
approach for canceling the inter-cell interference. Network √ xki = ρf qk1 + qk2 + . . . + qkL .
MIMO, however, is based on the assumption that the base M αk1 αk2 αkL
(14)
stations send to each other estimates ĝikl , which is not feasible
Let us denote
in LSAS systems. ⎛ ⎞ ⎛ ⎞
One possible conclusion of these arguments can be that xk1 qk1
1 ⎜ ⎟ ⎜ . ⎟
in both noncooperative LSASs and LSASs with cooperation xk = √ ⎝ ... ⎠ and qk = ⎝ .. ⎠
SINRs do not grow with M beyond certain limits. Below M
xkL qkL
we show that this conclusion is wrong. We demonstrate
that a limited collaboration between base stations allows us and
to radically resolve the pilot contamination problem and to ⎛ ⎞
β1k1 /αk1 ... βLk1 /αkL
construct interference free LSASs. ⎜ .. .. ⎟
We make the following assumption on collaboration be- Bk = ⎝ . . ⎠.
tween base stations. β1kL /αk1 ... βLkL /αkL

1139
Then from (28) we have Let ⎛ ⎞
β1k1 ... β1kL
1 √ ⎜ .. ⎟
√ xk = ρf Bk qk . (15) U
Bk = ⎝ . ..
⎠.
M .
βLk1 ... βLkL
Now we can describe the pilot contamination precoding.
Downlink PCP Then
The i-th base station estimates the slow fading coefficients √
βikl , k = 1, . . . , K; l = 1, . . . L, and transmits them to the (ŝk1 , . . . , ŝkL ) = ρr B U 
k (sk1 , . . . , skL ) .
network hub (alternatively to all other base stations). The Uplink PCP
network hub computes the PCP precoding matrices The base stations estimate their slow fading coefficients βikl
Ak = B−1
k , k = 1, . . . K (16) and transmit them to the network hub. The hub computes
√
and transmits the l-th row akl of Ak to the l-th base station. A k = ( ρr B Uk)
−1
, k = 1, . . . , K,
The l-th base station computes the signals and passes the l-th row akl of Ak to the l-th base station.

qkl = akl qk . (17) The i-th base station computes the estimates ŝki , k =
1, . . . , K, according to (19) and passes them to all other base
and transmits from its M antennas the vector stations. The i-th base stations reconstruct the symbol ski as
  
vl = w1l q1l + w2l q2l + . . . + wKl qKl . ski = aki · (ŝk1 , . . . , ŝkL ) .
The end of Downlink PCP The end of uplink PCP
From (15), (17), and (16) we obtain U
Similar to the downlink case we have limM →∞ ςkl = ∞.
⎛  ⎞ ⎛ ⎞ ⎛ ⎞
qk1 qk1 qk1
1 ⎜ ⎟ ⎜ . ⎟ ⎜ ⎟ V. F INITE M A NALYSIS
√ xk = Bk ⎝ ... ⎠ = Bk Ak ⎝ .. ⎠=⎝
..
. ⎠.
M 
For a finite M we make small changes in the TDD protocol.
qKL qKL qKL
A. TDD Protocol and Downlink PCP
From this equation it follows that both the additive noise and
interference vanishes and we have SINR: The i-th base station receives at the m-th antenna the m-th
D
row ymi of the matrix Yi defined in (2). It further multiples
lim ςkl → ∞. (18) ymi by ψk† and obtains the processed pilot signal
M →∞

Thus, under our network assumptions, we constructed a noise L


√
free and interference free multi-cell LSAS. ymik = ymi ψk† = ρr τ gmikl + zmik , zmik ∈ CN (0, 1).
l=1
Remark In this algorithm vectors wkl are computed solely on
the basis of ĝlkl . An exchange of M dimensional vectors is From ymik we get MMSE linear estimates:
not required (opposite to the network MIMO). √
ρr τ βikl ymik
Let us consider now the uplink transmission. Let the k-th ĝmikl = L . (20)
1 + ρr τ s=1 βiks
user in the l-th cell transmit skl . The i-th base station receives
K
L 
The estimation error g̃mikl = ĝmikl − gmikl is uncorrelated
 √ with the estimate ĝmikl . It is easy to see that

yi = ρr gikl skl + zi .
l=1 k=1   2
ρr τ βikl
E |g̃mikl |2 = βikl − L , (21)
To decode the signal ski the base station applies Maximum 1 + ρr τ s=1 βiks
Ratio Combining by multiplying the received signal yi by the
and
scaled channel estimate from (3) and obtains ∗ ρr τ βikl βiki
E {ĝmikl ĝmiki }= L . (22)
1 ∗ 1 + ρr τ s=1 βiks
ŝki = √ ĝiki yi (19)
M Under the conjugate beam-forming precoding (4) the m-th
L L K
1  ∗ √  antenna of the i-th base station transmits
= √ gikl1 ρr gikl skl
M l1 =1 l=1 k=1
K
1  ∗
L vmi =√ ymik qki , (23)
 L 
 K γ
∗ k=1
+ gikl z
1 i
+ zi 
gikl skl + zi zi .
l1 =1 l=1 k=1 where γ is a normalization coefficient chosen so that to satisfy
the average power constraint per cell:
Here ∗ denotes complex conjugation of all entries of the  
L M
corresponding vector. Applying Lemma 1, we obtain 1 2
a.s. √ E |vml | = 1. (24)
ŝki = ρr (βik1 sk1 + βik2 sk2 + . . . + βikL skL ) . L m=1
l=1

1140
The k-th terminal in the l-th cell receives the signal Theorem 2.
 2
L 
 M  L 
√ M ρf ρr τ  j=1 βjkl akjl 
xkl = ρf gmjkl vmj + zkl , (25) D
ςkl = , (34)
j=1 m=1 T1 + M · T2
If we use downlink PCP we have where
K K L  K 
L   L

 L

1  ∗  1  ∗ T1 =
1
+ ρf βjkl 1 + ρr τ βjns |anju |2
vmj =√ ymjk qkj =√ ymjk akjl qkl , (26) L
γ γ j=1 n=1 s=1 u=1
k=1 k=1 l=1

where akjl , l = 1, . . . , L, form the j-th row of Ak . To satisfy and  2

 L 
the power-constraint (24) we choose:   
  T 2 = ρ f ρr τ  βjkl akju  .
K L L L 
M   u=l  j=1 
γ= 1 + ρr τ βiks |akil |2 . (27)
L i=1 s=1 It is proven in [4] that assuming the worst case Gaussian
k=1 l=1

B. SINR estimates and Capacity Lower-Bound for PCP noise in the effective channel we obtain a capacity bound:
 D

Substituting (26) into (25), we obtained Ckl > log2 1 + ςkl . (35)
 M L K L
ρf     ∗
The expression (34) allows one to simulate SINRs in a fast
xkl = gmjkl ymjn anju qnu + zkl . way without generating M -dimensional vectors.
γ m=1 j=1 n=1 u=1
Theorem 2 shows that reduction ρf by the factor of 2 can be
After some manipulations we can write this expression in the compensated by doubling M . Thus LSASs can be very useful
following form for building energy efficient wireless systems.
(1) (2) (3) (4) (5) If PCP is not used, that is akjl = δjl , we have:
xkl = skl qkl + wkl + wkl + wkl + wkl + wkl , (28)
Theorem 3. If PCP is not used then
where 2
 D M ρf ρr τ βlkl
M L
ρf     ςkl = ,
skl = ∗
E ĝmjkl ymjk akjl , (29) R1 + M · R 2
γ m=1 j=1 where
and  K 
L   L


1
(1)
wkl = zkl , R1 = + ρf βjkl 1 + ρr τ βjns
j=1 n=1
L s=1
 M L K
(2) ρf    ∗  and R2 = ρf ρr τ
 2
βukl .
wkl = − g̃mjkl ymjn qnj , (30) u=l
γ m=1 j=1 n=1 2
D βlkl
From this we have limM →∞ ςkl =  which coin-
 M L
2
u=l βukl
(3) ρf    ∗  cides with (12). If we use PCP with Ak = B−1 then T2 = 0
wkl = ĝmjkl ymjn qnj , (31) k
γ m=1 j=1 and
n=k
D
lim ςkl =∞
M →∞
 M L
(4) ρf   ∗ as it is supposed to be according to (18).
wkl = (ĝmjkl ymjk
γ m=1 j=1
R EFERENCES
∗ 
− E{ĝmjkl ymjk })qkj , (32) [1] F. Fernandes, A. Ashikhmin, T. Marzetta, “Interfernce Reduction on
 M
 L
 L

Cellular Networks with Large Antenna Arrays,” IEEE International
(5) ρf  ∗
 Conference on Communicaions (ICC), Ottawa, Canada, 2012.
wkl = E ĝmjkl ymjk akju qku . (33) [2] J. Hoydis, S. ten Brink, M. Debbah, “Massive-MIMO: How Many
γ m=1 j=1 u=1,u=l Antennas do We Need,” arXiv:1107.1709v2.
[3] H. Huh, G. Caire, H.C. Papadopoulos, S.A. Ramprashad, “Achieving
The terminal can compute skl in advance and use it for “Massive-MIMO” Spectral Efficiency with a Not-so-Large Number of
estimation its SINR value. Thus s2kl is the signal power. All Antennas,” arXiv:1107.3862v2.
[4] T. Matzetta, ”How Much Training is Required for Multiterminal
other terms in (28) form the effective noise. The quantity MIMO?” in Proc. Asilomar Conference on Signals, Systems and Com-
(30) corresponds to channel estimation error, (31) to non- puters, Pacific Grove, CA, USA, Oct./Nov. 2006, pp.359–363.
orthogonal channels, (32) to beam-forming gain uncertainty, [5] T. Marzetta, “Noncooperative Cellular Wireless with Unlimited Numbers
of Base Station Antennas,” , IEEE Trans. on Wireless Communications,
and (33) to pilot-contamination. All these noise components 9, pp. 3590 –3600, 2010.
are mutually uncorrelated.
Estimating (29) and and combining (30)-(33), after calcu-
lations, we obtain for given Ak and βjkl :

1141