Professional Documents
Culture Documents
Hien Quoc Ngo, Erik G. Larsson and Thomas L. Marzetta, Energy and Spectral Efficiency of
Very Large Multiuser MIMO Systems, 2013, IEEE Transactions on Communications, (61), 4,
1436-1449.
http://dx.doi.org/10.1109/TCOMM.2013.020413.110848
Abstract—A multiplicity of autonomous terminals simultane- Recently, there has been a great deal of interest in MU-
ously transmits data streams to a compact array of antennas. MIMO with very large antenna arrays at the BS. Very large ar-
The array uses imperfect channel-state information derived rays can substantially reduce intracell interference with simple
from transmitted pilots to extract the individual data streams.
The power radiated by the terminals can be made inversely signal processing [8]. We refer to such systems as “very large
proportional to the square-root of the number of base station MU-MIMO systems” here, and with very large we mean arrays
antennas with no reduction in performance. In contrast if perfect comprising say a hundred, or a few hundreds, of antennas,
channel-state information were available the power could be simultaneously serving tens of users. The design and analysis
made inversely proportional to the number of antennas. Lower of very large MU-MIMO systems is a fairly new subject that is
capacity bounds for maximum-ratio combining (MRC), zero-
forcing (ZF) and minimum mean-square error (MMSE) detection attracting substantial interest [8]–[11]. The vision is that each
are derived. An MRC receiver normally performs worse than ZF individual antenna can have a small physical size, and be built
and MMSE. However as power levels are reduced, the cross-talk from inexpensive hardware. With a very large antenna array,
introduced by the inferior maximum-ratio receiver eventually things that were random before start to look deterministic. As
falls below the noise level and this simple receiver becomes a a result, the effect of small-scale fading can be averaged out.
viable option. The tradeoff between the energy efficiency (as
measured in bits/J) and spectral efficiency (as measured in Furthermore, when the number of BS antennas grows large,
bits/channel use/terminal) is quantified for a channel model that the random channel vectors between the users and the BS
includes small-scale fading but not large-scale fading. It is shown become pairwisely orthogonal [10]. In the limit of an infinite
that the use of moderately large antenna arrays can improve number of antennas, with simple matched filter processing at
the spectral and energy efficiency with orders of magnitude the BS, uncorrelated noise and intracell interference disappear
compared to a single-antenna system.
completely [8]. Another important advantage of large MIMO
Index Terms—Energy efficiency, spectral efficiency, multiuser systems is that they enable us to reduce the transmitted power.
MIMO, very large MIMO systems
On the uplink, reducing the transmit power of the terminals
will drain their batteries slower. On the downlink, much of
I. I NTRODUCTION the electrical power consumed by a BS is spent by power
In multiuser multiple-input multiple-output (MU-MIMO) amplifiers and associated circuits and cooling systems [12].
systems, a base station (BS) equipped with multiple antennas Hence reducing the emitted RF power would help in cutting
serves a number of users. Such systems have attracted much the electricity consumption of the BS.
attention for some time now [2]. Conventionally, the commu- This paper analyzes the potential for power savings on
nication between the BS and the users is performed by orthog- the uplink of very large MU-MIMO systems. We derive
onalizing the channel so that the BS communicates with each new capacity bounds of the uplink for finite number of BS
user in separate time-frequency resources. This is not optimal antennas. While it is well known that MIMO technology can
from an information-theoretic point of view, and higher rates offer improved power efficiency, owing to both array gains
can be achieved if the BS communicates with several users in and diversity effects [13], we are not aware of any work that
the same time-frequency resource [3], [4]. However, complex analyzes power efficiency of MU-MIMO systems with receiver
techniques to mitigate interuser interference must then be used, structures that are realistic for very large MIMO.1 We consider
such as maximum-likelihood multiuser detection on the uplink both single-cell and multicell systems, but focus on the anal-
[5], or “dirty-paper coding” on the downlink [6], [7]. ysis of single-cell MU-MIMO systems since: i) the results
are easily comprehensible; ii) it bounds the performance of
Manuscript received Dec. 15, 2011; revised May 2, 2012 and Aug. 20,
2012; accepted Nov. 1, 2012. The associate editor coordinating the review a multicell system; and iii) the single-cell performance can
of this paper and approving it for publication was B. Clerckx. This work be actually attained if one uses successively less-aggressive
was supported in part by the Swedish Research Council (VR), the Swedish frequency-reuse (e.g., with reuse factor 3, or 7). Our results are
Foundation for Strategic Research (SSF), and ELLIIT. E. Larsson was a Royal
Swedish Academy of Sciences (KVA) Research Fellow supported by a grant different from recent results in [14] and [15]. In [14] and [15],
from the Knut and Alice Wallenberg Foundation. Parts of this work were the authors derived a deterministic equivalent of the SINR
presented at the 2011 Allerton Conf. Commun., Control and Comput. [1]. assuming that the number of transmit antennas and the number
H. Q. Ngo and E. G. Larsson are with the Department of Electrical
Engineering (ISY), Linköping University, 581 83 Linköping, Sweden (Email:
nqhien@isy.liu.se; egl@isy.liu.se). 1 After submitting this work, other papers have also addressed the tradeoff
T. L. Marzetta is with Bell Laboratories, Alcatel-Lucent, 600 Moutain between spectral and energy efficiency in MU-MIMO. An analysis related to
Avenue, Murray Hill, NJ 07974, USA (Email: tom.marzetta@alcatel- the one presented here but for the downlink was given in [16]. However, the
lucent.com). analysis of the downlink is quantitatively and qualitatively different both in
Digital Object Identifier xxx/xxx what concerns systems aspects and the corresponding the capacity bounds.
2 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XXX 2012
of users go to infinity but their ratio remains bounded for where G represents the M ×K channel matrix between the BS
the downlink of network MIMO systems using a sophisticated and the K users, i.e., gmk , [G
G]mk is the channel coefficient
√
scheduling scheme and MISO broadcast channels using zero- between the mth antenna of the BS and the kth user; pux
forcing (ZF) precoding, respectively. The paper makes the is the vector of symbols simultaneously transmitted by the K
following specific contributions: users (the average transmitted power of each user is pu ); and
• We show that, when the number of BS antennas M grows n is a vector of additive white, zero-mean Gaussian noise.
without bound, we can reduce the transmitted power of We take the noise variance to be 1, to minimize notation, but
each user proportionally to 1/M if the BS has perfect without loss of generality. With this convention, pu has the
channel
√ state information (CSI), and proportionally to interpretation of normalized “transmit” SNR and is therefore
1/ M if CSI is estimated from uplink pilots. This holds dimensionless. The model (1) also applies to wideband chan-
true even when using simple, linear receivers. We also nels handled by OFDM over restricted intervals of frequency.
derive closed-form lower bounds on the uplink achievable The channel matrix G models independent fast fading,
rates for finite M , for the cases of perfect and imperfect geometric attenuation, and log-normal shadow fading. The
CSI, assuming MRC, ZF, and minimum mean-squared coefficient gmk can be written as
error (MMSE) receivers, respectively. See Section III. p
gmk = hmk βk , m = 1, 2, ..., M (2)
• We study the tradeoff between spectral efficiency and
energy efficiency. For imperfect CSI, in the low trans- where hmk is the fast fading√ coefficient from the kth user to
mit power regime, we can simultaneously increase the the mth antenna of the BS. βk models the geometric attenu-
spectral-efficiency and energy-efficiency. We further show ation and shadow fading which is assumed to be independent
that in large-scale MIMO, very high spectral efficiency over m and to be constant over many coherence time intervals
can be obtained even with simple MRC processing at the and known a priori. This assumption is reasonable since the
same time as the transmit power can be cut back by orders distances between the users and the BS are much larger than
of magnitude and that this holds true even when taking the distance between the antennas, and the value of βk changes
into account the losses associated with acquiring CSI very slowly with time. Then, we have
from uplink pilots. MRC also has the advantage that it can
G = H D 1/2 (3)
be implemented in a distributed manner, i.e., each antenna
performs multiplication of the received signals with the where H is the M × K matrix of fast fading coefficients
conjugate of the channel, without sending the entire base- between the K users and the BS, i.e., [H
H ]mk = hmk , and
band signal to the BS for processing. Quantitatively, our D is a K × K diagonal matrix, where [D
D ]kk = βk .
energy-spectral efficiency tradeoff analysis incorporates
the effects of small-scale fading but neglects those of B. Review of Some Results on Very Long Random Vectors
large-scale fading, leaving an analysis of the effect of We review some limit results for random vectors [17]
large-scale fading for future work. See Section IV. T
that will be useful later on. Let p , [p1 ... pn ] and q ,
T
[q1 ... qn ] be mutually independent n × 1 vectors whose
II. S YSTEM M ODEL AND P RELIMINARIES elements o are i.i.d. zero-mean o random variables (RVs) with
A. MU-MIMO System Model
n n
2 2
E |pi | = σp2 , and E |qi | = σq2 , i = 1, ..., n. Then from
We consider the uplink of a MU-MIMO system. The system the law of large numbers, we have
includes one BS equipped with an array of M antennas that 1 H a.s. 2 1 a.s.
receive data from K single-antenna users. The nice thing p p → σp , and p H q → 0, as n → ∞. (4)
n n
about single-antenna users is that they are inexpensive, simple, a.s.
and power-efficient, and each user still gets typically high where → denotes the almost sure convergence. Also, from
throughput. Furthermore, the assumption that users have single the Lindeberg-Lévy central limit theorem, we have
antennas can be considered as a special case of users having 1 d
√ p H q → CN 0, σp2 σq2 , as n → ∞ (5)
multiple antennas when we treat the extra antennas as if they n
were additional autonomous users.2 The users transmit their
d
data in the same time-frequency resource. The M × 1 received where → denotes convergence in distribution.
vector at the BS is
√ C. Favorable Propagation
y = pu Gx + n (1)
Throughout the rest of the paper, we assume that the fast
2 Note
that under the assumptions on favorable propagation (see Section fading coefficients, i.e., the elements of H are i.i.d. RVs with
II-C), having n autonomous single-antenna users or having one n-antenna zero mean and unit variance. Then the conditions in (4)–(5)
user (where the antennas cooperate in the encoding), represent two cases
with equal energy and spectral efficiency. To see why, consider two cases: are satisfied with p and q being any two distinct columns of
the case of 2 autonomous single-antenna users of which each spends power G . In this case we have
P , and the case of one dual-antenna user with a total power constraint
of 2P . Then, the sum rates for the two cases are the same and equal to GH G H H H 1/2
P 0 hH
= D 1/2 D ≈ D, M K
h k2 h k2 M M
P kh P kh
=log2I+ N1 [h 1 ,
log2 1+ N1 +log2 1+ N2 h1 h 2 ]
0 0 0 0 P hH
and we say that we have favorable propagation. Clearly, if all
2
where h i is the channel vector between the ith user (or ith antenna) to the
BS, and N0 is the variance of noise. fading coefficients are i.i.d. and zero mean, we have favorable
NGO et al.: ENERGY AND SPECTRAL EFFICIENCY OF VERY LARGE MULTIUSER MIMO SYSTEMS 3
propagation. Recent channel measurements campaigns have From (1) and (8), the received vector after using the linear
shown that multiuser MIMO systems with large antenna detector is given by
arrays have characteristics that approximate the favorable- √
r = puA H Gx + A H n . (10)
propagation assumption fairly well [10], and therefore provide
experimental justification for this assumption. Let rk and xk be the kth elements of the K × 1 vectors r and
To understand why favorable propagation is desirable, con- x , respectively. Then,
sider an M × K uplink (multiple-access) MIMO channel H , K
√ √
where M ≥ K, neglecting for now path loss and shadowing
X
rk = pu a H
k g k xk + pu aH H
k g i xi + a k n (11)
factors in D . This channel can offer a sum-rate of i=1,i6=k
K
X where ak and g k are the kth columns of the matrices A and G,
log2 1 + pu λ2k (6)
R= respectively. For a fixed channel realization G, the noise-plus-
k=1
interference Pterm is a random variable with zero mean and
where pu is the power spent per terminal and {λk }K k=1 are
K
variance pu i=1,i6=k |a aH 2 ak k2 . By modeling this term
k g i | + ka
the singular values of H , see [13]. If the channel matrix is as additive Gaussian noise independent of xk we can obtain a
normalized such that |Hij | ∼ 1 (where ∼ means equality of lower bound on the achievable rate. Assuming further that the
PK
the order of magnitude), then 2
k=1 λk = kH H k2 ≈ M K. channel is ergodic so that each codeword spans over a large
Under this constraint the rate R is bounded as (infinite) number of realizations of the fast-fading factor of G ,
log2 (1 + M Kpu ) ≤ R ≤ K log2 (1 + M pu ) . (7) the ergodic achievable uplink rate of the kth user is
The lower bound (left inequality) is satisfied with equality
aH 2
if λ21 = M K and λ22 = · · · = λ2K = 0 and corresponds
pu |a g
k k |
RP,k = E log2 1 + . (12)
K
to a rank-one (line-of-sight) channel. The upper bound (right
H g |2 + ka
P
2
p u a
|a k i a k k
inequality) is achieved if λ21 = · · · = λ2K = M . This occurs if
i=1,i6=k
the columns of H are mutually orthogonal and have the same To approach this capacity lower bound, the message has to be
norm, which is the case when we have favorable propagation. encoded over many realizations of all sources of randomness
III. ACHIEVABLE R ATE AND A SYMPTOTIC (M → ∞) that enter the model (noise and channel). In practice, assuming
P OWER E FFICIENCY wideband operation, this can be achieved by coding over the
frequency domain, using, for example coded OFDM.
By using a large antenna array, we can reduce the transmit-
Proposition 1: Assume that the BS has perfect CSI and that
ted power of the users as M grows large, while maintaining a
the transmit power of each user is scaled with M according
given, desired quality-of-service. In this section, we quantify
to pu = EMu , where Eu is fixed. Then,3
this potential for power decrease, and derive achievable rates
of the uplink. Theoretically, the BS can use the maximum- RP,k → log2 (1 + βk Eu ) , M → ∞. (13)
likelihood detector to obtain optimal performance. However, Proof: We give the proof for the case of an MRC receiver.
the complexity of this detector grows exponentially with K. With MRC, A = G so a k = g k . From (12), the achievable
The interesting operating regime is when both M and K are uplink rate of the kth user is
large, but M is still (much) larger than K, i.e., 1 K M . ( !)
g 4
It is known that in this case, linear detectors (MRC, ZF mrc p u kg k k
RP,k = E log2 1 + .
and MMSE) perform fairly well [8] and therefore we will
PK
pu i=1,i6=k |gg H 2 g k k2
k g i | + kg
restrict consideration to those detectors in this paper. We treat (14)
the cases of perfect CSI (Section III-A) and estimated CSI
(Section III-B) separately. Substituting pu = EMu into (14), and using (4), we obtain (13).
By using the law of large numbers, we can arrive at the same
A. Perfect Channel State Information result for the ZF and MMSE receivers. Note from (3) and (4)
that when M grows large, M 1
G H G tends to D , and hence the
We first consider the case when the BS has perfect CSI,
ZF and MMSE filters tend to that of the MRC.
i.e. it knows G . Let A be an M × K linear detector matrix
Proposition 1 shows that with perfect CSI at the BS and a
which depends on the channel G . By using the linear detector,
large M , the performance of a MU-MIMO system with M
the received signal is separated into streams by multiplying it
antennas at the BS and a transmit power per user of Eu /M
with A H as follows
is equal to the performance of a SISO system with transmit
r = AHy . (8) power Eu , without any intra-cell interference and without any
We consider three conventional linear detectors MRC, ZF, and fast fading. In other words, by using a large number of BS
MMSE, i.e., antennas, we can scale down the transmit power proportionally
to 1/M . At the same time we increase the spectral efficiency
G for MRC
H −1
K times by simultaneously serving K users in the same time-
A= G G G for ZF (9) frequency resource.
G GH G + 1 I −1
3 As mentioned after (1), p has the interpretation of normalized transmit
for MMSE
u
pu K SNR, and it is dimensionless. Therefore Eu is dimensionless too.
4 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XXX 2012
1) Maximum-Ratio Combining: For MRC, from (14), by Therefore, the kth column of A is given by [18]
the convexity of log2 1 + x1 and using Jensen’s inequality,
−1
Λ −1
H 1 k gk
we obtain the following lower bound on the achievable rate: a k = GG + I M g k = H −1 (23)
pu g k Λk g k + 1
mrc mrc
RP,k ≥ R̃P,k PK
( PK )!−1 where Λ k , i=1,i6=k g ig H i + pu I M . Substituting (23) into
1
pu i=1,i6=k |gg H
k g i |2 +kgg k k2 (12), we obtain the uplink rate for user k:
, log2 1+ E . (15)
pu kgg k k4 = E log2 1 + g H −1
mmse
RP,k k Λk g k
Proposition 2: With perfect CSI, Rayleigh fading, and
(a)
1
M ≥ 2, the uplink achievable rate from the kth user for MRC = E log2 −1
1 H
1 − gH pu I M + G G gk
can be lower bounded as follows:
k
!
p u (M − 1) β k
mrc
(16)
R̃P,k = log2 1 + PK . 1
pu i=1,i6=k βi + 1 = E log2 −1
1 − G H p1u I M + GG H G
Proof: See Appendix A. kk
If pu = Eu /M , and M grows without bound, then
(b)
1
(24)
!
Eu
(M − 1) β k
= E log2 −1
mrc
R̃P,k = log2 1+ E MPK → log2 (1+βk Eu) . (17)
I K + pu G H G
i=1,i6=k βi +1
u
M kk
Equation (17) shows that the lower bound in (16) becomes where (a) is obtained directly from (23), and (b) is obtained
equal to the exact limit in Proposition 1 as M → ∞. by using the identity
2) Zero-Forcing Receiver: With ZF, A H =
1
−1
1
−1
−1 H H H
G I M + GG G= IK + G G GH G
GH G G H , or A H G = I K . Therefore, a H
k g i = δki , pu pu
where δki = 1 when k = i and 0 otherwise. From (12), the −1
uplink rate for the kth user is = I K − I K + pu G H G .
By using Jensen’s inequality, we obtain the following lower
bound on the achievable uplink rate:
p u
(18)
zf
RP,k = E log2 1 + .
1
−1
(25)
GH G mmse
RP,k mmse
≥ R̃P,k = log2 1 +
E {1/γk }
kk
Using the approximate PDF of γk given by (26), we have analysis takes into account the fact that pilot signals cannot
the following proposition. take advantage of the large number of receive antennas since
Proposition 4: With perfect CSI, Rayleigh fading, and channel estimation has to be done on a per-receive antenna
MMSE, the lower bound on the achievable rate for the kth basis. All results that we present take this fact into account.
user can be approximated as Denote by E , Ĝ G −G G. Then, from (33), the elements of the ith
column of E are RVs with zero means and variances pp ββii+1 .
mmse
R̃P,k = log2 (1 + (αk − 1) θk ) . (29)
Furthermore, owing to the properties of MMSE estimation, E
Proof: Substituting (26) into (25), and using the identity is independent of Ĝ G. The received vector at the BS can be
[21, eq. (3.326.2)], we obtain rewritten as
H √ √
(34)
Γ (αk ) r̂r = Â
A Gx − puE x + n .
puĜ
mmse
R̃P,k = log2 1 + θk (30)
Γ (αk − 1)
Therefore, after using the linear detector, the received signal
where Γ (·) is the Gamma function. Then, using Γ (x + 1) = associated with the kth user is
xΓ (x), we obtain the desired result (29). K
√ √
Remark 1: From (12), the achievable rate RP,k can be aH
X
r̂k = puâ g
ĝ
k k x k + p u aH
â g i xi
k ĝ
rewritten as i=1,i6=k
a H g |2 K
|a √
RP,k = E log2 1 + Hk k
X
a k Λ ka k − pu aH
â aH
k ε i xi + â k n (35)
( !) i=1
1/2 2 −1/2
kaaH Λ k kΛΛ g k k 2
≤ E log2 1 + k k H k
where â ak , ĝg i , and ε i are the ith columns of Â
A, Ĝ
G, and E ,
a k Λ ka k
respectively.
−1
= E log2 1 + g H (31) Since ĜG and E are independent, Â A and E are independent
k Λk g k .
too. The BS treats the channel estimate as the true channel,
The inequality is obtained by using Cauchy-Schwarz’ inequal-
and the part including the last three terms of (35) is considered
ity, which holds with equality when a k = cΛ Λ−1
k g k , for any as interference and noise. Therefore, an achievable rate of the
c ∈ C. This corresponds to the MMSE detector (see (23)).
uplink transmission from the kth user is given by (36) shown
This implies that the MMSE detector is optimal in the sense
at the bottom of the page.
that it maximizes the achievable rate given by (12). Intuitively, if we cut the transmitted power of each user, both
the data signal and the pilot signal suffer from the reduction
B. Imperfect Channel State Information in power. Since these signals are multiplied together at the
In practice, the channel matrix G has to be estimated at the receiver, we expect that there will be a “squaring effect”. As a
BS. The standard way of doing this is to use uplink pilots. consequence, we cannot reduce power proportionally to 1/M
A part of the coherence interval of the channel is then used as in the case of perfect CSI. The following proposition shows
for the uplink training. Let T be the length (time-bandwidth that√it is possible to reduce the power (only) proportionally to
product) of the coherence interval and let τ be the number 1/ M .
of symbols used for pilots. During the training part of the Proposition 5: Assume that the BS has imperfect CSI,
coherence interval, all users simultaneously transmit mutually obtained by MMSE estimation from uplink pilots, and that
orthogonal pilot sequences of length τ symbols. The pilot the transmit power of each user is pu = √EM u
, where Eu is
sequences used by the K users can be represented by a τ × K fixed. Then,
√
matrix ppΦ (τ ≥ K), which satisfies Φ H Φ = I K , where RIP,k → log2 1 + τ βk2 Eu2 , M → ∞.
(37)
pp , τ pu . Then, the M × τ received pilot matrix at the BS is
given by Proof: For MRC, substituting â ak = ĝg k into (36), we
√ obtain the achievable uplink rate as
Y p = ppGΦ T + N (32) (
mrc
where N is an M × τ matrix with i.i.d. CN (0, 1) elements. RIP,k =E log2 1+
The MMSE estimate of G given Y is !)
pu kĝg k k4
1 1 .
G = √ Y pΦ ∗D̃
Ĝ D = G + √ W D̃ D (33) pu
PK
|ĝg H g i |2 +pu kĝg k k2
PK βi
g k k2
pp pp i=1,i6=k k ĝ i=1 τ pu βi +1 +kĝ
−1 (38)
√
where W , N Φ ∗ , and D̃
D , p1p D −1 + I K . Since Substituting pu = Eu / M into (38), and again using (4)
Φ H Φ = I K , W has i.i.d. CN (0, 1) elements. Note that our along with the fact that each element of ĝg k is a RV with zero
( !)
aH
pu |â g k |2
k ĝ
RIP,k = E log2 1+ PK (36)
aH g i |2 + pu kâa k k2 K βi
a k k2
P
pu i=1,i6=k |â k ĝ i=1 τ pu βi +1 + kâ
6 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XXX 2012
p β2
mean and variance pp βpk +1
k
, we obtain (37). We can obtain the Proposition 7: With ZF processing using imperfect CSI,
limit in (37) for ZF and MMSE in a similar way. Rayleigh fading, and for M ≥ K + 1, the achievable uplink
Proposition 5 implies that with imperfect CSI and a large rate for the kth user is bounded as
M , the performance of a MU-MIMO system with an M -
antenna array
√ at the BS and with the transmit power per user τ p2u (M − K) βk2
set to Eu / M is equal to the performance of an interference- zf
R̃IP,k = log2
1 + K
.
free SISO link with transmit power τ βk Eu2 , without fast P pu β i
(τ pu βk + 1) τ pu βi +1 +τ pu βk +1
fading. i=1
Remark 2: From the proof of Proposition 5, we see that (42)
if we cut the transmit power proportionally to 1/M α , where √
α > 1/2, then the SINR of the uplink transmission from√the Similarly, with pu = Eu / M , when M → ∞, the
kth user will go to zero as M → ∞. This means that 1/ M achievable uplink rate and its lower bound tend to the ones
is the fastest rate at which we can cut the transmit power of for MRC (see (40)), i.e.,
each user and still maintain a fixed rate.
→ log2 1 + τ βk2 Eu2 , M → ∞ (43)
zf
R̃IP,k
Remark 3: In general, each user can use different transmit
powers which depend on the geometric attenuation and the which equals the rate value obtained from Proposition 5.
shadow fading. This can be done by assuming that the kth
3) MMSE Receiver: With imperfect CSI, the received vec-
user knows βk and performs power control. In this case, the
tor at the BS can be rewritten as
reasoning leading to Proposition 5 can be extended to show
that to achieve the same rate as in a SISO system using √ √
y= Gx −
puĜ pu E x + n . (44)
transmit power Euq , we must choose the transmit power of
the kth user to be MEτuβk . Therefore, for the MMSE receiver, the kth column of Â
A is
Remark 4: It can be seen directly from (14) and (38) that given by
the power-scaling laws still hold even for the most unfavorable −1
propagation case (where H has rank one). However, for this H 1 √
ak = Ĝ
â GĜ
G + Cov (− puE x + n ) ĝg k
case, the multiplexing gains do not materialize since the pu
intracell interference cannot be cancelled when M grows −1
Λk ĝg k
Λ̂
without bound. = −1 (45)
1) Maximum-Ratio Combining: By following a similar line ĝg H Λk ĝg k + 1
k Λ̂
of reasoning as in the case of perfect CSI, we can obtain lower
where Cov (aa) denotes the covariance matrix of a random
bounds on the achievable rate.
vector a, and
Proposition 6: With imperfect CSI, Rayleigh fading, MRC
processing, and for M ≥ 2, the achievable uplink rate for the K K
!
X H
X βi 1
kth user is lower bounded by Λk ,
Λ̂ ĝg iĝg i + + I M . (46)
i=1
τ p u β i + 1 pu
i=1,i6=k
mrc
τ pu (M − 1) βk2 Similarly to in Remark 1, by using Cauchy-Schwarz’ inequal-
R̃IP,k = log2 1+ .
K ity, we can show that the MMSE receiver given by (45) is the
βi + (τ +1) βk + p1u
P
(τ pu βk + 1) optimal detector in the sense that it maximizes the rate given
i=1,i6=k
by (36).
(39)
√ Substituting (45) into (36), we get the achievable uplink rate
By choosing pu = Eu / M , we obtain for the kth user with MMSE receivers as
→ log2 1 + τ βk2 Eu2 , M → ∞. (40) −1
mrc
n o
R̃IP,k mmse
RP,k = E log2 1 + ĝg H Λ
Λ̂ g
ĝ
k k k
Again, when M → ∞, the asymptotic bound on the rate equals
K
!−1 −1
the exact limit obtained from Proposition 5.
X βi 1 H
=−E log2I K + + G Ĝ
Ĝ G .
2) ZF Receiver: For the ZF receiver, we have âaH g i = δki .
k ĝ τ pu βi +1 pu
i=1
From (36), we obtain the achievable uplink rate for the kth
kk
user as (47)
Again, using an approximate distribution for the SINR, we
p can obtain a lower bound on the achievable uplink rate in
zf u
RIP,k = E log2 1+ K .
−1 closed form.
P pu β i H
+1 G
Ĝ G
Ĝ
Proposition 8: With imperfect CSI and Rayleigh fading, the
τ pu βi+1
i=1 kk
(41) achievable rate for the kth user with MMSE processing is
approximately lower bounded as follows:
Following the same derivations as in Section III-A2 for the
case of perfect CSI, we obtain the following lower bound on
mmse
R̃IP,k = log2 1 + (α̂k − 1) θ̂k (48)
the achievable uplink rate.
NGO et al.: ENERGY AND SPECTRAL EFFICIENCY OF VERY LARGE MULTIUSER MIMO SYSTEMS 7
√
where where pux i is the K × 1 transmitted vector of K users in
(M − K + 1 + (K − 1) µ̂)
2 the ith cell; n l is an AWGN vector, n l ∼ CN (0, I M ); and
α̂k = , G li is the M × K channel matrix between the lth BS and
M − K + 1 + (K − 1) κ̂
the K users in the ith cell. The channel matrix G li can be
M − K + 1 + (K − 1) κ̂ represented as
θ̂k = ω β̂k (49)
M − K + 1 + (K − 1) µ̂ 1/2
P −1 G li = H liD li (52)
K βi τ p βk2
where ω , 1
i=1 τ pu βi +1 + pu , β̂k , τ pu βuk +1 , µ̂ and where H li is the fast fading matrix between the lth BS and the
κ̂ are obtained by using following equations: K users in the ith cell whose elements have zero mean and unit
K variance; and D li is a K×K diagonal matrix, where [D D li ]kk =
1 X 1
µ̂ = βlik , with βlik represents the large-scale fading between the
K −1 K−1 K−1
i=1,i6=k
M ω β̂i 1 − M + M µ̂ +1 kth user in the i cell and the lth BS.
1) Perfect CSI: With perfect CSI, the received signal at the
K
X ω β̂i lth BS after using MRC is given by
κ̂ 1 +
2
K−1 K−1
L
i=1,i6=k M ω β̂i 1 − + M µ̂ +1 √ √ X
M
rl = pu G H
ll G ll x l + p u GH H
ll G lix i + G ll n l . (53)
K
X ω β̂i µ̂ + 1/(K − 1) i=1,i6=l
= 2 . (50) Eu
With pu = M , (53) can be rewritten as
K−1 K−1
i=1,i6=k M ω β̂i 1 − M + M µ̂ +1
p GH G L
Table I summarizes the lower bounds on the achievable 1 √ X GH ll G li 1
√ r l = Eu ll ll x l + pu xi + √ GH
ll n l .
rates for linear receivers derived in this section, distinguishing M M
i=1,i6=l
M M
between the cases of perfect and imperfect CSI, respectively. (54)
Here C (x) , log2 (1 + x).
From (4)–(5), when M grows large, the interference from
We have considered a single-cell MU-MIMO system. This
other cells disappears. More precisely,
simplifies the analysis, and it gives us important insights into
how power can be scaled with the number of antennas in very 1
√ r l → EuD llx l + D 1/2
p
nl
ll ñ (55)
large MIMO systems. A natural question is to what extent this M
power-scaling law still holds for multicell MU-MIMO systems. where ñ
nl ∼ CN (0, I ). Therefore, the SINR of the uplink
Intuitively, when we reduce the transmit power of each user, transmission from the kth user in the lth cell converges to a
the effect of interference from other cells also reduces and constant value when M grows large, more precisely
hence, the SINR will stay unchanged. Therefore we will have
the same power-scaling law as in the single-cell scenario. The SINRP
l,k → βllk Eu , as M → ∞. (56)
next section explains this argument in more detail. This means that the power scaling law derived for single-cell
systems is valid in multicell systems too.
C. Power-Scaling Law for Multicell MU-MIMO Systems 2) Imperfect CSI: In this case, the channel estimate from
the uplink pilots is contaminated by interference from other
We will use the MRC for our analysis. A similar analysis
cells. The MMSE channel estimate of the channel matrix G ll
can be performed for the ZF and MMSE detectors. Consider
is given by [11]
the uplink of a multicell MU-MIMO system with L cells
L
!
sharing the same frequency band. Each cell includes one BS X 1
equipped with M antennas and K single-antenna users. The Gll =
Ĝ G li + √ W l D̃ D ll (57)
i=1
pp
M × 1 received vector at the lth BS is given by
√
L hwhere D ll is a diagonal
i D̃ P matrix where −1the kth diagonal element
L
X
yl = pu G lix i + n l (51) D ll
D̃ = βllk β
i=1 lik + 1
pp . The received signal at
kk
i=1 the lth BS after using MRC is given by
H
TABLE I
r̂r l = Ĝ
Gll y l
L OWER BOUNDS ON THE ACHIEVABLE RATES OF THE UPLINK L
!H L
!
TRANSMISSION FOR THE k TH USER .
X 1 √ X
D ll
= D̃ G li + √ W l pu nl . (58)
G lix i +n
i=1
pp i=1
Perfect CSI Imperfect CSI √
p (M −1)βk τ pu (M −1)βk 2 With pu = Eu / M , we have
MRC C u K C
K
L X L L
βi+(τ +1)βk+p1 1 GH
li G lj GH
li n l
P P
pu βi +1 (τ pu βk +1) −1 p X X
u
i6=k i6=k
3/4
D
D̃ ll r
r̂ l = E u x j +
M 3/4
2
M i=1 j=1
M i=1
τ pu (M −K)βk
ZF C (pu (M −K) βk) C
K
L
1 X WH 1 WH
βi
+τ βk+p1
l G li l nl
P
(τ pu βk+1) τ p β +1
i=1 u i . (59)
u
+√ 3/4
x i + √
MMSE C ((αk − 1) θk ) C (α̂k − 1) θ̂k τ i=1 M τ Eu M 1/2
8 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XXX 2012
By using (4) and (5), as M grows large, we obtain by the transmit power expended (in Joules/channel use).
L Typically, increasing the spectral efficiency is associated with
1 −1 p X 1 increasing the power and hence, with decreasing the energy-
D
D̃ ll r
r̂ l → Eu D lix i + √ wl
w̃ (60)
M 3/4 i=1
τ Eu efficiency. Therefore, there is a fundamental tradeoff between
the energy efficiency and the spectral efficiency. However, in
where w̃
w l ∼ CN (0, I M ). Therefore, the asymptotic SINR of
one operating regime it is possible to jointly increase the
the uplink from the kth user in the lth cell is
energy and spectral efficiencies, and in this regime there is
2
τ βllk Eu2 no tradeoff. This may appear a bit counterintuitive at first, but
SINRIP
l,k → , as M → ∞. (61)
PL 2 2
τ i6=l βlik Eu + 1 it falls out from the analysis in Section IV-A. Note, however,
√ that this effect occurs in an operating regime that is probably
We can see that the 1/ M power-scaling law still holds. of less interest in practice.
Furthermore, transmission from users in other cells consti- In this section, we study the energy-spectral efficiency
tutes residual interference. The reason is that the pilot reuse tradeoff for the uplink of MU-MIMO systems using linear
gives pilot-contamination-induced inter-cell interference which receivers at the BS. Certain activities (multiplexing to many
grows with M at the same rate as the desired signal. users rather than beamforming to a single user and increasing
Remark 5: The MMSE channel estimate (57) is obtained the number of service antennas) can simultaneously benefit
by the assumption that, for uplink training, all cells simul- both the spectral-efficiency and the radiated energy-efficiency.
taneously transmit pilot sequences, and that the same set of Once the number of service antennas is set, one can adjust
pilot sequences is used in all cells. This assumption makes other system parameters (radiated power, numbers of users,
no fundamental difference compared with using different pilot duration of pilot sequences) to obtain increased spectral-
sequences in different cells, as explained [8, Section VII-F]. efficiency at the cost of reduced energy-efficiency, and vice-
Nor does this assumption make any fundamental difference versa. This should be a desirable feature for service providers:
to the case when users in other cells transmit data when the they can set the operating point according to the current traffic
users in the cell of interest send their pilots. The reason is demand (high energy-efficiency and low spectral-efficiency,
that whatever data is transmitted in other cells, it can always for example, during periods of low demand).
be expanded in terms of the orthogonal pilot sequences that
are transmitted in the cell of interest, so pilot contamination A. Single-Cell MU-MIMO Systems
ensues. For example, consider the uplink training in cell 1 of
a MU-MIMO system with L = 2 cells. Assume that, during We define the spectral efficiency for perfect and imperfect
an interval of length τ symbols (τ ≥ K), K users in cell 1 CSI, respectively, as follows
are transmitting uplink pilots Φ T at the same time as K users K
X K
T −τ X A
A A A
in cell 2 are transmitting uplink data X 2 . Here Φ is a τ × K RP = R̃P,k , and RIP = R̃IP,k (62)
T
matrix which satisfies Φ H Φ = I K . The received signal at base k=1 k=1
station 1 is where A ∈ {mrc, zf, mmse} corresponds to MRC, ZF and
√ √ MMSE, and T is the coherence interval in symbols. The
Y 1 = ppG 11Φ T + puG 12X 2 + N 1
energy-efficiency for perfect and imperfect CSI is defined as
where N 1 ∈ CM ×τ is AWGN at base station 1. By projecting 1 A 1 A
the received signal Y 1 onto Φ ∗ , we obtain A
ηP = A
R , and ηIP = R . (63)
pu P pu IP
√ √
Y 1 , Y 1Φ ∗ = ppG 11 + puG 12X̃
Ỹ X 2 + Ñ
N1 The large-scale fading can be incorporated by substituting (39)
where X̃ ∗
X 2 , X 2Φ , and Ñ ∗
N 1 , N 1Φ . The kth column of Ỹ
Y1 and (42) into (62). However, this yields energy and spectral
is given by efficiency formulas of an intractable form and which are
√ √ very difficult (if not impossible) to use for obtaining further
ỹy 1k = ppg 11k + puG 12x̃
x2k + ñ
n1k insights. Note that the large number of antennas effectively
removes the small-scale fading, but the effects of path loss
where g 11k , x̃
x2k , and ñ
n1k are the kth columns of G 11 , X̃
X 2 , and
and large-scale fading will remain. This may give different
N 1 , respectively. By using the Lindeberg-Lévy central limit
Ñ
√ users vastly different SNRs. As a result, power control may
theorem, we find that each element of the vector puG 12x̃ x2,k
be desired. In principle, a power control factor could be
(ignoring the large-scale fading in this argument) is approxi-
included by letting pu in (39) and (42) depend on k. The
mately Gaussian distributed with zero mean and variance Kpu .
optimal transmit power for each user would depend only on the
If K = τ , then Kpu = pp and this result means that the
large-scale fading, not on the small-scale fading and effective
effect of payload interference is just as bad as if users in cell
power-control rules could be developed straightforwardly from
2 transmitted pilot sequences.
the resulting expressions. However, the introduction of such
power control may bring new trade-offs, for example that
IV. E NERGY-E FFICIENCY VERSUS S PECTRAL -E FFICIENCY of fairness between users near and far from the BS. In
T RADEOFF addition, the spectral versus energy efficiency tradeoff relies
The energy-efficiency (in bits/Joule) of a system is defined on optimization of the number of active users. If the users
as the spectral-efficiency (sum-rate in bits/channel use) divided have grossly different large-scale fading coefficients, then the
NGO et al.: ENERGY AND SPECTRAL EFFICIENCY OF VERY LARGE MULTIUSER MIMO SYSTEMS 9
50.0
Bounds
given by
40.0 Simulation τ (M − 1) p2u
mrc T −τ
RIP = K log2 1 + ,
30.0 T τ (K − 1) p2u + (K + τ ) pu + 1
1 mrc
20.0 mrc
ηIP = R . (64)
MRC, ZF, MMSE pu IP
Spectral-Efficiency (bits/s/Hz)
10.0
Perfect CSI
We have
0.0 mrc 1 mrc
50 100 150 200 250 300 350 400 450 500 lim ηIP = lim R
Number of Base Station Antennas (M)
pu →0 pu IP
pu →0
30.0 T −τ (log2 e) τ (M − 1) pu
Bounds = lim K = 0 (65)
Simulation
pu →0 T τ (K − 1) p2u + (K + τ ) pu + 1
20.0 and
1 mrc
mrc
lim ηIP = lim RIP = 0. (66)
pu →∞ pu →∞ pu
10.0 MRC, ZF, MMSE
Equations (65) and (66) imply that for low pu , the energy
Imperfect CSI efficiency increases when pu increases, and for high pu mrcthe en-
0.0 ∂R
50 100 150 200 250 300 350 400 450 500 ergy efficiency decreases when pu increases. Since ∂pIP > 0,
Number of Base Station Antennas (M)
u
∀pu > 0, RIP mrc
is a monotonically increasing function of pu .
Therefore, at low pu (and hence at low spectral efficiency), the
Fig. 1. Lower bounds and numerically evaluated values of the spectral energy efficiency increases as the spectral efficiency increases
efficiency for different numbers of BS antennas for MRC, ZF, and MMSE
with perfect and imperfect CSI. In this example there are K = 10 users, the and vice versa at high pu . The reason is that, the spectral
coherence interval T = 196, the transmit power per terminal is pu = 10 dB, efficiency suffers from a “squaring effect” when the received
and the propagation channel parameters were σshadow = 8 dB, and ν = 3.8. data signal is multiplied with the received pilots. Hence, at
pu 1, the spectral-efficiency behaves as ∼ p2u . As a
consequence, the energy efficiency (which is defined as the
issue will arise as to whether these coefficients should be fixed spectral efficiency divided by pu ) increases linearly with pu . In
before the optimization or whether for a given number of more detail, expanding the rate in a Taylor series for pu 1,
users K, these coefficients should be drawn randomly. Both we obtain
ways can be justified, but have different operational meaning
mrc
∂RIP 1 ∂ 2 RIP
mrc
in terms of scheduling. This leads, among others, to issues mrc mrc
RIP ≈ RIP |pu =0 + pu + p2u
∂pu 2 ∂p2
with fairness versus total throughput, which we would like to pu =0 u pu =0
avoid here as this matter could easily obscure the main points T −τ
= K log2 (e) τ (M − 1) p2u . (67)
of our analysis. Therefore, for analytical tractability, we ignore T
the effect of the large-scale fading here, i.e., we set D = I K . This gives the following relation between the spectral effi-
Also, we only consider MRC and ZF receivers.4 ciency and energy efficiency at pu 1:
For perfect CSI, it is straightforward to show from (16), r
T −τ
(20), and (63) that when the spectral efficiency increases, the mrc
ηIP = mrc .
K log2 (e) τ (M − 1) RIP (68)
energy efficiency decreases. For imperfect CSI, this is not T
always so, as we shall see next. In what follows, we focus We can see that when pu 1, by doubling the spectral
on the case of imperfect CSI since this is the case of interest efficiency, or by doubling M , we can increase the energy
in practice. efficiency by 1.5 dB.
1) Maximum-Ratio Combining: From (39), the spectral 2) Zero-Forcing Receiver: From (42), the spectral effi-
efficiency and energy efficiency with MRC processing are ciency and energy efficiency for ZF are given by
τ (M − K) p2u
T −τ
4 When M is large, the performance of the MMSE receiver is very close
zf
RIP = K log2 1 + , and
T (K + τ ) pu + 1
to that of the ZF receiver (see Section V). Therefore, the insights on energy 1 zf
versus spectral efficiency obtained from studying the performance of ZF can zf
ηIP = R . (69)
be used to draw conclusions about MMSE as well. pu IP
!
T −τ τ (M − 1) p2u 1 mrc
mrc
Rmul = K log2 1+ , and ηmul
mrc
= R (73)
pu IP
T 2 2
τ K L̄ − 1 + β L̄ − 1 (M − 2) pu + L̄ (K + τ ) pu + 1
!
T −τ τ (M − K) p2u 1 zf
zf
Rmul = K log2 1+ , and ηIP
zf
= R (74)
pu ml
T τ K L̄2 − L̄β + β − 1 p2u + L̄ (K + τ ) pu + 1
10 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XXX 2012
40.0
Perfect CSI, MRC
are used for the simulation in the Section V. Here, we consider
Imperfect CSI, MRC p = E M a simplified channel model, i.e., D ll = I K , and D li = βII K ,
Perfect CSI, ZF where β ∈ [0, 1] is an intercell interference factor. Note that
Imperfect CSI, ZF
Spectral-Efficiency (bits/s/Hz)
30.0 from (57), the estimate of the channel between the kth user
Perfect CSI, MMSE
Imperfect CSI, MMSE
in the lth cell and the lth BS is given by
−1 L p
1 X 1
20.0 ĝg llk = L̄ + h llk + hlik + √ w lk . (72)
βh
pp pp
i6=k
PL √
where L̄ , (L − 1) β + 1. The term i6=k βh hlik represents
10.0 the pilot contamination, therefore
p = E
M PL √
hlik k2
i6=k E k βh
= β (L − 1)
Eu = 20 dB hllk k2 }
E {kh
0.0
50 100 150 200 250 300 350 400 450 500 can be considered as the effect of pilot contamination.
Number of Base Station Antennas (M) Following a similar derivation as in the case of single-
Fig. 2. Spectral efficiency versus the number of BS antennas M for MRC,
cell MU-MIMO systems, we obtain the spectral efficiency
ZF, and MMSE processing at the receiver, with perfect CSI and with imperfect and energy efficiency for imperfect CSI with MRC and ZF
CSI (obtained from uplink pilots). In this example K = 10 users are receivers, respectively, as (73) and (74) shown at the bottom of
served simultaneously, the reference transmit power is Eu = 20 dB, and
the propagation parameters were σshadow = 8 dB and ν = 3.8.
the previous page. The principal complexity in the derivation is
the correlation between pilot-contaminated channel estimates.
10.0 We can see that the spectral efficiency is a decreasing
Perfect CSI, MRC Eu = 5 dB
function of β and L. Furthermore, when L = 1, or β = 0, the
Imperfect CSI, MRC
Perfect CSI, ZF
results (73) and (74) coincide with (64) and (69) for single-cell
8.0 MU-MIMO systems.
Spectral-Efficiency (bits/s/Hz)
Imperfect CSI, ZF
Perfect CSI, MMSE
Imperfect CSI, MMSE
V. N UMERICAL R ESULTS
6.0 p = E
M
A. Single-Cell MU-MIMO Systems
We consider a hexagonal cell with a radius (from center
4.0
to vertex) of 1000 meters. The users are located uniformly
at random in the cell and we assume that no user is closer
to the BS than rh = 100 meters. The large-scale fading is
2.0
modelled via βk = zk /(rk /rh )ν , where zk is a log-normal
p = E
30.0
18.0 1 bit/s/Hz MRC 2 bits/s/Hz MRC
27.0
ZF ZF
15.0 MMSE
Required Power, Normalized (dB)
MMSE
12.0 21.0
Fig. 4. Transmit power required to achieve 1 bit/channel use per user Fig. 5. Same as Figure 4 but for a target spectral efficiency of 2 bits/channel
for MRC, ZF, and MMSE processing, with perfect and imperfect CSI, as use per user.
a function of the number M of BS antennas. The number of users is fixed to
K = 10, and the propagation parameters are σshadow = 8 dB and ν = 3.8.
as Fig. 2. The gap between the performance of MRC and Fig. 6 shows the relative energy efficiency versus the the
that of ZF (or MMSE) is reduced compared with Fig. 2. This spectral efficiency for MRC and ZF. The relative energy
is so because the relative effect of crosstalk interference (the efficiency is obtained by normalizing the energy efficiency by
12 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XXX 2012
10
4
-20 dB by jointly choosing K, τ , and pu to maximize the energy-
M=100
efficiency subject a fixed spectral-efficiency, i.e.,
Relative Energy-Efficiency (bits/J)/(bits/J)
-10 dB
10
3
A
arg max ηIP , s.t. RIP
A
= const., K ≤ τ ≤ T
ZF pu ,K,τ
10
4
-20 dB
magnitude. This is possible with simple linear processing such
MRC
as MRC or ZF at the BS, and using channel estimates obtained
Relative Energy-Efficiency (bits/J)/(bits/J)
ZF
-10 dB from uplink pilots even in a high mobility environment where
10
3
half of the channel coherence interval is used for training.
β =0.04
Generally, ZF outperforms MRC owing to its ability to cancel
0 dB intracell interference. However, in multicell environments with
β =0.32
10
2
strong pilot contamination, this advantage tends to diminish.
β =0.11
MRC has the additional benefit of facilitating a distributed per-
10 dB antenna implementation of the detector. Quantitatively, with
10
1
MRC, 100 antennas can serve about 50 terminals in the same
K =1,
time-frequency resource, each terminal having a fading-free
20 dB throughput of about 1 bpcu, and hence the system offering a
M =1,
10
0
β =0 sum-throughput of about 50 bpcu. These conclusions are valid
under a channel model that includes the effects of small-scale
Reference Mode M = 100, L = 7 Rayleigh fading, but neglects the effects of large-scale fading
10
-1
(see the discussion after (63)).
0 1010 20 30 40 50 60 70 80 90
Spectral-Efficiency (bits/s/Hz)
A PPENDIX
Fig. 8. Same as Figure 6, but for a multicell scenario, with L = 7 cells, A. Proof of Proposition 2
and coherence interval T = 196.
From (15), we have
( PK )!−1
2
p u i=1,i6=k i|g̃ | + 1
uniformly at random in the first cell, and the 2nd user is mrc
R̃P,k = log2 1+ E (75)
located uniformly at random in one of the 6 nearest-neighbor pu kgg k k2
cells of the 1st cell. Let β̄1 and β̄2 be the large scale fading gH g
g g
from the 1st user and the 2nd user to the 1st BS, respectively. where g̃i , kgkg ki . Conditioned on g k , g̃i is a Gaussian RV
k
(The large scale fading is modelled as in Section V-A1.) with zero mean and variance βi which does not depend on
Then we compute β as E β̄2 /β̄1 . By simulation, we obtain g k . Therefore, g̃i is Gaussian distributed and independent of
β = 0.32, 0.11, and 0.04 for the cases of (σshadow = 8 dB, g k , g̃i ∼ CN (0, βi ). Then,
ν = 3.8, freuse = 1), (σshadow = 8 dB, ν = 3, freuse = 1), ( PK
pu i=1,i6=k |g̃i |2 + 1
)
and (σshadow = 8 dB, ν = 3.8, freuse = 3), respectively, where E
freuse is the frequency reuse factor. pu kgg k k2
Fig. 8 shows the relative energy efficiency versus the spec-
K
X 1
tral efficiency for MRC and ZF of the multicell system. The
2
= pu E |g̃i | +1 E
reference mode is the same as the one in Fig. 6 for a single-cell pu kgg k k2
i=1,i6=k
system. The dotted line shows the performance for the case of
K
M = 1, K = 1, and β = 0. The solid and dashed lines show
X 1
= pu βi + 1 E
. (76)
the performance for the cases of M = 100, and L = 7, with pu kgg k k2
i=1,i6=k
different intercell interference factors β of 0.32, 0.11, and 0.04.
Each point on these curves is computed by jointly choosing Using the identity [22]
τ , K, and pu to maximize the energy efficiency for a given E tr W −1 = m/(n − m) (77)
spectral efficiency. We can see that the pilot contamination
where W ∼ Wm (n, I n ) is an m×m central complex Wishart
significantly degrades the system performance. For example,
matrix with n (n > m) degrees of freedom, we obtain
when β increases from 0.11 to 0.32 (and hence, the pilot
contamination increases), with the same power, pu = 10 1 1
E = , for M ≥ 2. (78)
dB, the spectral efficiency and the energy efficiency reduce g
pu kg k k 2 pu (M − 1) βk
by factors of 3 and 2.7, respectively. However, with low Substituting (78) into (76), we arrive at the desired result (16).
transmit power where the spectral efficiency is smaller than
10 bits/s/Hz, the system performance is not affected much B. Proof of Proposition 3
by the pilot contamination. Furthermore, we can see that From (3), we have
in a multicell scenario with high pilot contamination, MRC −1 1 −1
achieves a better performance than ZF. E GHG = E H HH
kk βk kk
−1
VI. C ONCLUSION 1 H
= E tr H H
Kβk
Very large MIMO systems offer the opportunity of increas-
(a) 1
ing the spectral efficiency (in terms of bits/s/Hz sum-rate) by = , for M ≥ K + 1 (79)
one or two orders of magnitude, and simultaneously improving (M −K) βk
the energy efficiency (in terms of bits/J) by three orders of where (a) is obtained by using (77). Using (79), we get (20).
14 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XXX 2012