IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS
Performance Analysis of Massive MIMO for
Cell-Boundary Users
Yeon-Geun Lim, Student Member, IEEE, Chan-Byoung Chae, Senior Member, IEEE, and
Giuseppe Caire, Fellow, IEEE
Abstract—In this paper, we consider massive multiple-input
multiple-output (MIMO) systems for both downlink and uplink
scenarios, where three radio units (RUs) connected via one
digital unit (DU) support multiple user equipments (UEs) at the
cell-boundary through the same radio resource, i.e., the same
arXiv:1309.7817v2 [cs.IT] 10 Jul 2015
time-frequency slot. For downlink transmitter options, the study
considers zero-forcing (ZF) and maximum ratio transmission
(MRT), while for uplink receiver options it considers ZF and
maximum ratio combining (MRC). For the sum rate of each
of these, we derive simple closed-form formulas. In the simple
but practically relevant case where uniform power is allocated
to all downlink data streams, we observe that, for the downlink,
vector normalization is better for ZF while matrix normalization
is better for MRT. For a given antenna and user configuration, we
also derive analytically the signal-to-noise-ratio (SNR) level below
which MRC should be used instead of ZF. Numerical simulations
confirm our analytical results.
Index Terms—Massive MIMO, cell-boundary users, ergodic
achievable rate, matched filter, zero-forcing, normalization, pre-
coding, and combining filter.
I. I NTRODUCTION
Multiple-input multiple-output (MIMO) wireless communi-
cation techniques have evolved from single-user MIMO (SU-
MIMO) to multi-user MIMO (MU-MIMO) systems [1]. To
approach the capacity of the MIMO broadcast channel, [2],
[3] proposed simple zero-forcing (ZF)-based linear algorithms,
where the transmitter and the receivers are equipped with
multiple antennas. The authors in [4] intensively investigated
the optimality of the linear matched type-combining filter and
they assumed an infinite number of antennas at the receiver.
[4] proved that a simple linear beamforming (coordinated
beamforming in the paper) asymptotically approaches the sum
capacity achieved by dirty paper coding (DPC).
Recently, massive MIMO (a.k.a. large-scale MIMO) has
been proposed to further maximize network capacity and to
conserve energy [5], [6], [7]. The authors in [6] proposed
massive MIMO systems that use simple linear algorithms
such as maximum ratio combining (MRC) for the uplink
and maximum ratio transmission (MRT) for the downlink.
To further increase sum rate performances, several network
MIMO algorithms with multiple receive antennas have been
proposed [8], [9]. These systems, however, assume that the
Y.-G. Lim and C.-B. Chae are with the School of Integrated Technology,
Yonsei University, Korea. Email: {yglim, cbchae}@yonsei.ac.kr.
G. Caire is with the School of Electrical Engineering and Computer Science,
Technical University of Berlin, Germany. Email: caire@tu-berlin.de.
1
network supports a maximum of three users through a rela-
tively small number of transmit antennas.1
Massive MIMO systems in multi-cell environments have
also been studied in [5], [7], [10], [11]. Multi-cell massive
MIMO is prone to some critical issues. These include pi-
lot contamination, which becomes, in time division duplex
(TDD) systems, the main capacity-limiting factor, especially
when MRT is used. Joint spatial division and multiplexing
is proposed in [12] to employ frequency division duplex
(FDD) systems. The authors in [13] proposed a pilot alignment
algorithm for a cognitive massive MIMO system. The authors
in [7] investigated downlink performance with MRT and ZF
precoder for a massive MIMO system.2 In [5], the authors
studied uplink performance with MRC, ZF, and a minimum
mean square error (MMSE) filters for massive MIMO. It was
shown that transmit energy can be conserved by the power-
law 1/M with perfect channel state information (CSI)
scaling √
and 1/ M with imperfect CSI at the base station (BS),
where M represents the number of BS antennas. In [11],
the authors showed theoretically and numerically the impact
of pilot contamination and proposed a multi-cell MMSE-
based precoding algorithm to reduce both intra- and inter-cell
interference. In [11], MRT precoding was used; the inter-user
interference is eventually eliminated once the transmitter has
a large enough number of antennas.
The assumption of an infinite number of antennas at the
BS for a finite number of users somehow trivializes many
problems (e.g., under this limit MRC/MRT have the same
performance as ZF). A more meaningful system scaling is
considered in [10], [15], [16], where the number of antennas
per BS and the number of users both go to infinity with a
fixed ratio. In this case, the results of the infinite number
of BS antennas per user can be recovered by letting this
ratio become large. This more refined analysis, however,
illuminates all the system performance regimes. For example,
in [16] the “massive MIMO” regime is defined as the regime
where pilot contamination dominates with respect to multi-
user interference; it is observed that this regime occurs only
when the number of BS antennas per user is impractically
large. These conclusions are also reached, independently and
in parallel, in [10]. In particular, [10] considers a multi-cell
architecture formed by small clusters of cooperating BSs. The
authors proposed a system where the users are partitioned
1 Note that more than three users can be supported if there is a common
message, i.e., for a clustered broadcast channel.
2 In [14], the authors investigated the performance of MRT and ZF in large-
scale antenna systems, but paid little attention to normalization techniques.
2 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS
into homogeneous classes and the downlink MIMO precod-
ing scheme is optimized for each class. Then, a scheduler
optimally allocates the time-frequency transmission resource
across the different user classes, yielding an inherent multi-
mode multi-cell massive MIMO system. One of the main
findings of [10] is that it is convenient to serve users at the
cell center in a single-cell mode, while users at the cell edge
should be served by small cooperative clusters formed by the
closest three neighboring cell/sectors.
This paper, motivated by the results in [10], focuses on the
cell-edge user performance-the system bottleneck for both the
uplink and the downlink. The massive MIMO system under
consideration consists of multiple radio units (RUs) connected
to one another by optical fibers, and further connected to
a centralized digital unit (DU), as illustrated in Figs. 1(a)
and 2. Through the optical fibers, each RU can share data mes-
sages and channel state information. With respect to all three
neighboring BSs forming a cluster, the cell-edge users have
symmetric and spatially isotropic channel statistics. Hence,
the system is equivalent to, as shown in Fig. 1(b), a single-cell
massive MIMO system with cell-edge users that are located in
the low signal-to-noise-ratio (SNR) regime (for the high SNR
regime, we provide some results to provide a detailed analysis,
such as for determining which normalization method is better
for MRT). As in the prior work referenced above, we consider
the performance of linear precoding/filtering schemes such as
ZF and MRT for the downlink and ZF and MRC for the uplink.
In particular, we consider two possible normalizations of the
precoding filters for the downlink, referred to as vector and,
respectively matrix normalization. The main contributions of
this paper are as follows:
• Tighter ergodic achievable sum rate of ZF and
MRT/MRC: Tighter ergodic achievable sum rate of
ZF and MRT/MRC: We provide a methodology of
simple but quite accurate approximations while con-
sidering the property of near deterministc, which is
defined in Section III-E. This approximation method
is valid for massive MIMO systems at low/high
SNR. The effective channel matrix tends to an
identity matrix when MRT/MRC are assumed with
perfect CSI in massive MIMO systems. This means
that the random channels become near deterministic
due to the property of the law of large numbers. If
the number of user is large, the sum of the inter-
ference powers cannot become near deterministic,
which means the sum of those still have randomness
and do not converge to zero. The researchers in [5]
concluded that with perfect CSI at the BS and a
large M , the performance of a MU-MIMO system
with a transmit power per user scaling with M is
equal to the performance of a single-input-single-
output system. Assuming a large number of users,
the conclusion in [5] does not hold because the sum
of the interference powers does not converge to zero.
From this aspect, considering the condition of near
deterministic and the large number of users is mathe-
matically significant in analyzing the performance of
massive MIMO systems. We investigate whether the
signal-to-interference-plus-noise-ratio (SINR) term
in the ergodic sum rate is able to approximate near
deterministic such as the form of the expectation
when M goes to infinity in the low or high SNR
regime. From this approximation of the SINR term,
we derive simple approximations for the ergodic
achievable sum rate of ZF and MRT/MRC at the
low and high SNR regimes considering two simple
power normalization methods at the downlink.
The derived approximations are accurate and far
simpler to evaluate than the asymptotic expressions
given in [10], [15], [16], which were obtained
through asymptotic random matrix theory [17] and
usually given in terms of the solution of multiple
coupled fixed-point equations. Thanks to their sim-
plicity, the proposed approximations are suitable to
analyze the low and high SNR regimes. From the
derived approximations, we investigate a suitable
normalization method of MRT and transceiver mode
selection algorithms. In their asymptotic expressions,
the authors in [7], [14], [16] did not consider normal-
ization methods; thus the expressions were unable
to classify which normalization method was better.
Numerical results demonstrate the tightness of our
analysis.3
• Downlink precoding normalization methods: We
compare matrix and vector normalization for down-
link precoding, in the simple and practical case of
uniform power allocation over all downlink streams.
It is well known that these normalizations, with opti-
mal (waterfilling) power allocation and ZF precoding
yield identical results [18]. However, with practical
suboptimal power allocation and in the finite antenna
regime these normalizations are generally not equiv-
alent. Most prior work on multi-user MIMO paid
scant attention to this issue. For example, the authors
in [19], considered matrix normalization and those
in [20] considered vector normalization in [20] with
neither making any mention of why different nor-
malizations were used; [19] is part I of their papers
dealing with linear precoding (ZF/MMSE), and [20]
is part II focusing on non-linear precoding (vector
perturbation). Thus, the comparison between vector
and matrix normalization remains an open problem.
To solve this issue, we find a suitable normalization
method for MRT precoding by using the proposed
approximations and a suitable normalization method
for ZF precoding by using the arithmetic-geometric
inequality.
• Transceiver mode selection algorithms: We pro-
pose two transceiver mode selection algorithms from
transmit power and the number of active users per-
spectives. In [21], it is concluded that ZF is better
3 In fact, the achievable sum rate can be slightly enhanced in a low SNR
regime by using regularized ZF (a.k.a. MMSE). This strategy, however,
requires knowing all users noise variances at the transmitter, which requires
additional feedback. Thus, in this paper we focus on ZF and/or MRT/MRC.
MINOR REVISION 3

for cell center users, i.e., high SNR, and MRT is

better for cell-boundary users, i.e., low SNR, in a
downlink system. However, [21] did not consider
transceiver mode selection as a function of SNR
(i.e., of the transmit power, for a given pathloss law
and cell geometry) for the same class of edge users.
In this paper, we explain how much transmit power
and/or number of active users are needed for ZF to
provide a better sum rate than MRT (for downlink)
or MRC (for uplink). In particular, we find the
optimal MIMO mode selection scheme in terms of
closed-form thresholds of the transmit power, where
the thresholds depend on the number of edge users.
Note that since our system model is simplified by making Fig. 1: (a) System model of multi-RU massive MIMO scenario
assumptions, our analytical results are more accurate than the with cell-boundary users. (b) System model of single-RU
work given demonstrated in [10], [15], [16]. Furthermore, our- massive MIMO scenario with cell-boundary users.
closed form expressions are based on the assumption of infinite
M , and when M goes to infinity the approximation is quite
accurate. Even when M is finite, though still a large number, and assume perfect CSI at the RU. We also assume that the
the approximations are quite accurate. In prior work on mas- channel is flat fading and the elements of a channel matrix are
sive MIMO with linear receivers/precoders, reserchers have modeled as independent complex Gaussian random variables
provided the simple lower bounds of the sum rate performance with zero mean and unit variance. The channel between the
[5], [7], [14], or the complex closed-form expressions of that cloud BS (one DU and three RUs) and the k-th user is denoted
[16], which are less accurate than our analysis. We anticipate by an 1 × M row vector h Tk (k = 1, 2, · · · , K). A K × M
our contributions to yield a wide range of insights for related channel matrix H between the cloud BS and all UEs consists
studies, such as those on performance analysis on MIMO, of channel vectors h Tk . Let g k denote the column vector of
power normalization methods, and the trade-off between ZF transmit precoding and sk represent the transmit symbol for
and MRT/MRC. the k-th UE at downlink. Similarly, let w k denote the column
This paper is organized as follows. In Section II, we intro- vector of receive combining filter for the k-th UE at uplink.
duce the considered system model and problem statement with Also, let nk be the additive white Gaussian noise vector. Then,
respect to precoding normalization methods and beamforming the received signal at the k-th UE is expressed by
techniques. In Section III, we introduce some mathematical
K
motivations and preliminaries useful for analysis. In Sec- p X p
tion IV, we analyze i) the ergodic performance of ZF and yk = Pth Tk g k sk + Pth Tk g ` s` +nk (1)
| {z }
MRT precoding, ii) which precoding normalization method `=1,`6=k
desired signal | {z }
is better for each precoder, and iii) the ergodic performance interference
for cell-boundary users with the best normalization method. where, Pt denotes the total network transmit power across
In Section V, we provide an approximation of the achievable three RUs. Also, the received signal for the k-th UE at the
ergodic uplink sum rate. In Section VI, we propose transceiver cloud BS is expressed by
mode selection algorithms with i) a power threshold and ii)
K
the number-of-users cross point of ZF- and MRT-precoding p X p
techniques. Numerical results are shown in Section VII. Sec- rk = Puw Tk h k xk + w Tk nk
Puw Tk h ` x` +w (2)
| {z }
tion VIII presents our conclusions and future work. `=1,`6=k
desired signal | {z }
interference

II. S YSTEM M ODEL : M ASSIVE MIMO
Consider a massive MIMO system as shown in Figs. 1
and 2. One DU (cloud BS) controls three RUs and K users.
Each RU is connected with one another by optical fibers.
Fig. 1(a) shows that the cloud BS provides a massive MIMO
environment to cell-edge users under the assumption that RU
has a relatively small number of antennas (more practical in the
recent antenna configuration; we will use this system model
in Section VII). Fig. 1(b) illustrates the equivalent model of
Fig. 1(a) considered as single-cell massive MIMO systems
with cell edge users. We assume that the cloud BS has M
antennas and each user equipment (UE) is equipped with one
antenna. In this paper, we do not consider pilot contamination
where, Pu and xk denote the transmit power per each user and
the transmit symbol of the k-th user at uplink, respectively.
A. Downlink
Eq. (1) contains the desired signal, interference, and noise
terms. To eliminate the interference term, we use the following
precoding:
ZF : F = H ∗ (H
H H ∗ )−1 = [ff 1 f 2 · · · f k · · · f K ],
MRT : F = H ∗ = [ff 1 f 2 · · · f k · · · f K ]
where F is a precoding matrix consisting of each column
vector f k .
4 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS

1) Asymptotic downlink sum rate: In Section IV, we derive

the simple and tight approximations of the ergodic achievable
sum rate of ZF and MRT. We also find that ZF with vector
normalization is better than ZF with matrix normalization
while MRT with matrix normalization is better than MRT
with vector normalization. The proposed ergodic achievable
sum rates of ZF with vector normalization at low SNR and
MRT with matrix normalization at low SNR are given by
 
Pt (M − K + 1)
RZFvec , DLL = K log2 1 + ,
K
 
Pt (M + 1)
RMRTmat , DLL ≈ K log2 1 + .
Pt (K − 1) + K
Fig. 2: Block diagram of multi-RU massive MIMO system. 2) Asymptotic uplink sum rate: We also evaluate the ap-
One DU consists of three RUs connected by optical fibers. proximations of the ergodic achievable sum rate of ZF and
This model is regarded as a single-RU massive MIMO system. MRC in the uplink case. The proposed ergodic achievable sum
rates of ZF at low SNR and MRC at low SNR are given by
 
Pu M
To satisfy the power constraint, we consider two meth- RMRC, ULL ≈ K log2 1 + ,
Pu (K − 1) + 1
ods, i.e., vector/matrix normalizations. The normalized trans-
RZF, ULL ≈ K log2 {1 + Pu (M − K + 1)} .
mit beamforming vectors (columns of a precoding ma-
with vector/matrix normalizations are given as g k =
trix) √ 3) Transceiver mode selection algorithm: In Section VI, we
f k /( K||ff k ||) and g k = f k /||FF ||F , respectively. Note propose two transceiver mode selection algorithms from i) a
that vector normalization imposes equal power per downlink transmit power and ii) the number of active users perspectives.
stream, while matrix normalization yields streams with differ- We explain how much transmit power and/or the number of
ent power. In this paper, to simplify, we do not consider a active users are needed for ZF to provide a better sum rate
power optimization that could yield a complexity problem in than MRT/MRC. The thresholds are given in Lemmas 6-10.
very large array antenna systems.
1) ZF/MRT with vector normalization: The received signal III. M ATHEMATICAL M OTIVATIONS AND P RELIMINARIES
at the k-th UE can be expressed as follows: In this section, we introduce some mathematical motiva-
K tions and preliminaries to evaluate an asymptotic analysis for
p fk p X f`
yk = Pth Tk √ sk + Pth Tk √ s` + nk . network massive MIMO, which will be used in Sections IV
K||ff k || `=1,`6=k
K||ff ` || and V.
(3)
2) ZF/MRT with matrix normalization: Similarly, we can A. Achievable Rate Bound
rewrite the received signal with matrix normalization as such: In this paper, to maximize the achievable sum rate of
K downlink/uplink systems, we evaluate the closed forms of each
p fk X p f` system’s performance. The achievable rates are bounded as
yk = Pth Tk sk + Pth Tk s` + nk . (4)
||F
F ||F ||F
F ||F follows:
`=1,`6=k
( )   
1 S
log2 1 +
≤ E log2 1 +
B. Uplink E I+N I +N
S
Similar to the downlink system, to eliminate the interference  
S

term, and to maximize the SNR in (2), we use the following ≤ log2 1 + E
I +N
combining filter at the RUs:
by using Jensen’s Inequality of convex and concave functions
H ∗H )−1H ∗ = [w
ZF : W = (H w 1 w 2 · · · w k · · · w K ], where S, I, and N represent signal power, interference power,
MRC : W = H ∗ = [ww1 w 2 · · · w k · · · w K ] and noise power, respectively. Note that we use these bounds
only for ZF cases and show that our results, based on the
where W is a combining filter matrix consisting of each bounds, are more accurate than prior work [5], [7], [16].
column vector w k . Here, we do not consider normalization
since it does not change SNR values in the uplink scenario.
B. Expectation and Variance of Random Vectors
Lemma 1: Let h k and h ` (k 6= `) be M × 1 vectors
C. Summary of the Main Results whose elements are independent identically distributed (i.i.d.)
We first summarize the main results based on our analyses. complex Gaussian random variables with zero mean and unit
These results are mathematically motivated from some random variance.
matrix theorems, shown in Section III. 1) E[||hhk ||2 ] = M ,
MINOR REVISION 5

hk ||2 ] = M .
Var[||h F. Ergodic Achievable Sum Rate of Massive MIMO Systems
∗
hkh ` ] = 0, at the low/high SNR regime
2) E[h
h∗kh ` ] = M .
Var[h
Lemma 4: Let Xv and vi be a norm of a random vec-
hk ||4 ] = M 2 + M ,
3) E[||h
tor v (M × 1) and the i-th entry of v , respectively,
hk ||4 ] = 4M 3 + 10M 2 + 6M .
Var[||h n o
∗
h k h ` |2 ] = M ,
4) E[|h i.e., Xv = v12 + v22 + · · · + vM 2
. Since E X1v =
n o n o
h∗kh ` |2 ] = M 2 + 2M .
Var[|h 1
E v2 +v2 +···+v2 1
= E M (v2 +v2 +···+v2 )/M ≈ M E v2 1

Proof: See Appendix A. 1 2 M 1 2 M n o { i}

1 1 1 1
and E{X = E{v12 +v22 +···+vM
≈ , E con-
v} 2
} M E{vi2 } Xv
1
C. Effective Channel verges to E{Xv } as M goes to infinity.
In this Lemma, we assume that the desired signal and the
Lemma 2: In massive MIMO systems, the transmit energy interference plus noise terms are norms of a random vector
can be conserved by power-scaling law 1/M with perfect CSI. (M × 1), and some of those terms in the low/high SNR
1
1) lim E[ M hk ||2 ] = 1,
||h regime are near deterministic as satisfying the condition of
M →∞
1
lim Var[ M hk ||2 ] = 0.
||h (5); thus we assume at least one of the those terms has the
M →∞
1 ∗
2) lim E[ M h kh ` ] = 0, same property of Xv . From the assumption, if Sn≈ E o {S}
M →∞ S
1 ∗ from (5) with a low/high SNR assumption, then E I+N ≈
lim Var[ M h kh ` ] = 0. n o n o
M →∞ 1 1 E{S}
Generally, a transceiver uses MRT or MRC in massive E {S} E I+N . Also, E {S} E I+N ≈ E{I+N } as M
n o
MIMO, which means that the effective channel of the desired 1 1
goes to infinity using E Xv ≈ E{Xv } . Similarly, for I + N
signal becomes one and the interference signal becomes zero n o
S
as the number of antennas (M ) goes to infinity, as illustrated: converges to E {I + N } cases, we also obtain E I+N ≈
E{S}
E{I+N } .
So we could obtain the following approximation of
1 a.s. SINR when M goes to infinity in the low or high SNR regime:
H H ∗ −−→ I K , as M → ∞.
M  
S E(S)
E ≈ . (6)
I +N E(I + N )

D. Signal and Interference Power From (6), the lower bound of the ergodic sum rate is the
Lemma 3: In a similar way, the expectation and the vari- same as the upper bound of the ergodic sum rate. Thus we
ance of the signal and the interference power are given by could also get the following approximation of the ergodic sum
1) lim E[ M12 ||hhk ||4 ] = 1, rate when M goes to infinity in the low or high SNR regime:
M →∞
lim Var[ M12 ||h
hk ||4 ] = 0.     
M →∞
S E(S)
E log2 1 + ≈ log2 1 + .
h∗kh ` |2 ] = 0,
2) lim E[ M12 |h I +N E(I + N )
M →∞
h∗kh ` |2 ] = 0.
lim Var[ M12 |h
M →∞
Note that the terms M1 1
hk ||4 and M
||h h∗kh ` |2 do not converge
|h
to M and zero, respectively, as M goes to infinity since their
variance does not go to zero, which means that they still have G. Arithmetic-Geometric Inequality
randomness.
Lemma 5: Let b1 , b2 , ..., bK be random variables. We can
obtain the following inequality through Arithmetic-geometric
E. Chebyshev’s Inequality Inequality defined in [22]:
2
Let X be a random variable with variance σX , c and K   !
 be scalars, and Y = 1c X be a random variable with
X 1 1
log2 1 + ≥ K log2 1 + PK .
variance σY2 = c12 σX2
, respectively. If c is not very large in k=1
Kbk k=1 bk
comparison with E{X}, but c2 >> σX 2
, then Y = 1c X is
σ2 σ2 Proof:
near deterministic, i.e., P [|Y − E{Y }| > ] ≤ Y2 = 2Xc2 . For
fixed  > 0, !
K  
2 1 X 1 1
σX log2 1 + ≥ log2 1 + 1 PK
if → 0, then Y ≈ E{Y } (5) K Kbk
c2 k=1 K k=1 Kbk
K   !
with high probability. We use (5) to approximate a SINR term X 1 1
⇔ log2 1 + ≥ K log2 1 + PK .
in ergodic sum rate expressions. Kbk
k=1 k=1 bk
As an example, to analyze the rate in the high SNR regime,
if Var[ P1t ||h
hk ||2 ] = P12 M converges to zero, then P1t ||h
hk ||2
t
1 2 1
converges to Pt E{|h hk | } = Pt M . This is a simple application of Jensen’s Inequality.
6 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS

IV. A SYMPTOTIC D OWNLINK S UM R ATE FOR By using Lemma 5, (9) can further be expressed as
C ELL - BOUNDARY U SERS (  !)
 1 2

In this section, we derive the achievable rate bounds, and K log2 1 + E Pt  
||F
F ||F 
show which normalization method is suitable for ZF- and   
MRT-type precoding at the downlink. We assume that the cell- 1
= K log2 1 + E Pt
boundary users are in the low SNR regime (Pt Var {Xv } → 0) H H ∗ )−1 )
tr((H
( !)
and that the cell center users are in the high SNR regime 1
( P1t Var {Xv } → 0); we also assume that the desired power = K log2 1 + E Pt PK
k=1 ||ff k ||2
term or the interference power term becomes near determin-   
istic in the low/high SNR regime. Based on our analytical 1
≤ K log2 1 + E Pt .
results, we will also show which precoding technique is K||ff k ||2
desired for cell-boundary users. So the second upper bound (without an expectation form) of
matrix normalization in the ZF case can be given by
 
A. Ergodic Performance Pt (M − K + 1)
RU 2
ZFmat , DL = K log 2 1 + . (10)
1) Achievable rate bounds for ZF precoding: K
i) Lower bounds: The lower bound of the ergodic sum rate 2) Achievable rate for MRT precoding:
for the ZF precoding is well known, as follows [7]: i) Vector normalization-low SNR regime: From (3), we can
  derive the ergodic achievable sum rate of vector normalization
Pt (M − K)
RL ZFvec , DL = R L
ZFmat , DL = K log 2 1 + in low SNR as follows:
K  
Pt M
RMRTvec , DLL ≈ K log2 1 + . (11)
using the property of Wishart matrices [17]. Pt (K − 1) + K
ii) Vector normalization-upper bound: From (3), we can Proof: See Appendix B1.
derive the SINR of the upper bound of vector normalization ii) Vector normalization-high SNR regime: Similarly, we
in the ZF case, as given by can get the ergodic achievable sum rate of vector normalization
 T √ f k 2 in high SNR as follows:
   

S
 
 P h
t h k K||ff k ||
 
  
E =E Pt (M + 1)
I +N  Pt P K  T √ f ` 2
  RMRTvec , DLH ≈ K log2 1 + . (12)
hk K||ff ||  + 1  Pt (K − 1) + K
 
`=1,`6=k h `
Proof: See Appendix B2.
 
Pt
=E iii) Matrix normalization-low/high SNR regime: From (4),
K||ff k ||2
(a) Pt (M − K + 1)
we can evaluate the ergodic achievable sum rate of matrix nor-
= (7) malization in low/high SNR by using the following formation:
K  
Pt (M + 1)
n (a)o results from  the diversity order of ZF
where RMRTmat , DLL/H ≈ K log2 1 + . (13)
Pt (K − 1) + K

E ||ff 1k ||2 = M − K + 1 , as shown in [23]. From
(7), the upper bound of vector normalization in the ZF case Proof: See Appendix B3 and B4.
can be represented as
  B. Comparison between Vector and Matrix Normalizations
Pt (M − K + 1)
RU ZFvec , DL = K log 2 1 + . 1) Performance comparison of ZF: To find which normal-
K H ∗ −1
ization technique is better in ZF, we let bk = [(H H P)t ]kk in
iii) Matrix normalization-upper bound: From (4), the SINR Lemma 5 directly.
of the upper bound of matrix normalization in the ZF case can K  
be expressed as
X Pt
log2 1 +
K[(H H H ∗ )−1 ]kk
 T f k 2
 
k=1
 

S
 
 Pt hhk ||FF ||F  
 !
E =E Pt
I +N  T f ` 2
  ≥ K log2 1 + PK
 Pt PK H H ∗ )−1 ]kk
hk ||FF ||F  + 1  k=1 [(H
 
`=1,`6=k h
( 
 1 2
 ) ⇔ RZFvec , DL ≥ RZFmat , DL (14)
= E Pt    . (8) K
||F
F ||F  H H ∗ )−1 ]kk = ||ff k ||2 and H H ∗ )−1 ]kk =
P
where [(H k=1 [(H
PK 2 2
k=1 k|f
f k || = ||F
F || . From (14), since the desired power
From (8), the first upper bound (with an expectation form) of and the interference power are one and zero respectively, the
matrix normalization in the ZF case can be represented as power normalization per user only affects the performance of
ZF. From the perspective of Jesen’s Inequality, the sum rate
(  !)
 1 2

U1
RZFmat , DL = K log2 1 + E Pt    . (9) with the different power allocation per user is the upper bound
||F
F ||F 
of the sum rate with the same power allocation per user at the
MINOR REVISION 7

instant channel and arbitrary SNR. We can conclude that, in V. A SYMPTOTIC U PLINK S UM R ATE FOR
the ZF case, vector normalization is always better than matrix C ELL -B OUNDARY U SERS
normalization.
We have focused on a downlink scenario with a sum power
2) Performance comparison of MRT: From (11)-(13), a constraint. In this section, we investigate an uplink case, where
comparison of the ergodic achievable sum rate is given by each user has its own power constraint. From (2), the ergodic
achievable sum rate for the uplink, RUL , is
RMRTmat , DLL ' RMRTvec , DLL (15)
"K ( )#
RMRTmat , DLH ≈ RMRTvec , DLH (16) X w Tk h k |2
Pu |w
RUL = E log2 1 + PK .
at low and high SNR, respectively. It is well known that k=1 Pu `=1,`6=k |w w Tk h ` |2 + ||w
w k ||2
approach methods that maximize the desired signal are better (18)
than those that mitigate the interference power at low SNR. From (18), we can derive the ergodic achievable uplink sum
The effective desired channel gain can be maximized with rate of MRC, RMRC, UL , as follows:
MRT. The desired power per user scales down in proportion to
its power in the vector normalization while the desired power RMRC, UL
per user scales down in the same proportion over all users. This "K ( )#
X hk ||4
Pu ||h
means that the gains of better channels with vector normaliza- =E log2 1 + PK .
tion tend to scale down more than the gains of better channels k=1 Pu `=1,`6=k h∗kh ` |2 + ||h
|h hk ||2
with matrix normalization. This yields the difference of the (19)
desired power term of (15) and (16) at low SNR. Therefore,
We approximate the ergodic achievable sum rate of MRC as
we confirm that, for MRT precoding, matrix normalization
follows:
is always better than vector normalization at low SNR. We
conclude, however, that there is marginal performance gap i) High SNR regime:
between vector normalization and matrix normalization at high
 
Pu (M + 1)
SNR. RMRC, ULH ≈ K log2 1 + . (20)
Pu (K − 1) + 1
ii) Low SNR regime:
C. Ergodic Achievable Sum Rate for Cell-boundary Users  
with the Best Normalization Method Pu M
RMRC, ULL ≈ K log2 1+ . (21)
Pu (K − 1) + 1
As explained in Section IV-B, we conclude that the suitable
normalization methods are vector normalization for ZF and Proof: See Appendix C
matrix normalization for MRT. We assume that the transmit Similar to ZF precoding, the ergodic sum rate for ZF for
power (Pt ) is small for cell-boundary users (low SNR regime). uplink at low SNR, RZF, ULL , is
Using the property of ZF precoding, the ergodic achievable   
U Pu
sum rate of ZF is represented as RZF, ULL ≈ RZF, UL = K log2 1 + E
H ∗H )−1 ]k,k
[(H
  
S = K log2 {1 + Pu (M − K + 1)} .
E log2 1 +
I +N
( !)
Pt |hhTk g k |2
= E log2 1 + PK
Pt `=1,`6=k |h hTk g ` |2 + 1 VI. T RANSCEIVER M ODE S ELECTION
hTk g k |2


= E log2 1 + Pt |h A. Algorithm
(b)
hTk g k |2 ) In this section, we propose two transceiver mode selection

≈ log2 1 + Pt E(|h (17)
algorithms from i) transmit power and ii) the number of
 T 2
where (b) results from (5) with Pt Var |h hk g k | → 0. Eq. (17) active users perspectives. To provide a mathematically simple
indicates that the achievable sum rate of ZF precoding can solution, we first introduce Lemma 6 and Lemma 7 that use a
approach its upper bounds at low SNR by (5). Thus, the power threshold as follows:
ergodic achievable sum rate of ZF with vector normalization Lemma 6: The power threshold to select a better precoder
at low SNR is given by for downlink is given by
K2
 
Pt (M − K + 1)
RZFvec , DLL ≈ K log2 1 + . Pth, DL = . (22)
K (K − 1)(M − K + 1)

From (13), we find the ergodic achievable sum rate of matrix
normalization in low SNR:
 
Pt (M + 1)
RMRTmat , DLL ≈ K log2 1 + .
Pt (K − 1) + K
If the RUs have more transmit power than the power
threshold Pth, DL , the ZF precoder provides a better sum rate
performance.
Proof: To derive (22) for cell-boundary users, we use
the low SNR approximation for ZF and MRT. By letting
8 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS

RZFvec , DLL ≥ RMRTmat , DLL , we can get (22) as follows:

RZFvec , DLL − RMRTmat , DLL ≥ 0
Pt (M − K + 1) Pt (M + 1)
⇔ − ≥0
K Pt (K − 1) + K
K2
⇔ Pt ≥ Pth, DL = .
(K − 1)(M − K + 1)

Lemma 7: The power threshold to select a better receive

combining filter at uplink is given by
1
Pth, UL = . (23)
M −K +1
If each UE has larger transmit power per user than Pth, UL , Fig. 3: The difference of the gradient between ZF and MRT
the solution employing ZF at the RUs provides a better sum at Kcross when P is very small (almost zero) and M is much
rate performance. larger than P . The difference is always positive (> 0).
Proof: To evaluate (23), we use the low SNR approxima-
tion of MRC, i.e., (21). From RL ZF, UL ≥ RMRC, ULL , we can
obtain (23) for uplink as follows: Proof: Both RZFvec , DLL and RMRTmat , DLL are concave
functions. Also, unlike RZFvec , DLL , RMRTmat , DLL is a monotonic
RZF, ULL − RMRC, ULL ≥ 0 increasing function; thus, two cross points exist: one is when
Pu M the number of users K is one; the other is when the number
⇔ Pu (M − K + 1) − ≥0
Pu (K − 1) + 1 of users K has (24) with a large M approximation (this
1 approximation results from satisfying Kcross, DL = 1 condition)
⇔ Pu ≥ Pth ,UL = .
M −K +1 as follows:
RMRTmat , DLL − RZFvec , DLL ≥ 0
Lemma 6 helps the RUs select one of the precoders, i.e., ZF
or MRT, with respect to the transmit power of the cloud BS. Pt (M + 1) Pt (M − K + 1)
⇔ −
Also, the power policy of the cloud BS could be adjusted by Pt (K − 1) + K K
the power threshold that is a function of M and K. Therefore, Pt M Pt (M − K + 1)
≈ − ≥0
the RUs could find a suitable precoding mode according to Pt (K − 1) + K K
the user’s location. Similarly, Lemma 7 could be applied to Pt (M + 1)
the uplink case. ⇔ K ≥ Kcross, DL = .
1 + Pt
The proposed power threshold would be affected by a
specific number of users, so a power cross point, that refers Lemma 10: The user cross point, Kcross, UL , to select a
to Pcross , exists where MRT or MRC is always better for better receive combining filter at uplink when the RUs have
any number of active users. Since Lemmas 6 and 7 are larger transmit power than Pcross, UL , is given by
monotonic increasing functions of K (Pth, DL/UL (K = k+1) >
Pth, DL/UL (K = k)), Pth, DL and Pth, UL have minimum values 1
Kcross, UL = M + 1 − . (25)
at K = 2. These points become Pcross, DL and Pcross, UL . Pu
Lemma 8: If the transmit power of the RUs/UEs is lower If the number of users, K, is larger than Kcross, UL , MRC
than Pcross , MRT or MRC is always better than ZF in terms provides a better sum rate performance.
of sum rate. The power cross point, Pcross , at downlink and Proof: Similar to Lemma 9, we can obtain (25) as follows:
uplink are given by
RMRC, ULL − RZF, ULL ≥ 0
4 1
Pcross, DL = , Pcross, UL = . Pu M
M −1 M −1 ⇔ − Pu (M − K + 1)
Pu (K − 1) + 1
1
Now we investigate Kcross , which is a transceiver mode ⇔ K ≥ Kcross, UL = M + 1 − .
selection threshold when the transmit power at the transceiver Pu
is larger than Pcross and when the number of active users varies.
Lemma 9: If the RUs have more transmit power than Pcross , Lemmas 8-10 provide a proper solution for the low SNR
the user cross point at downlink, Kcross, DL , for selecting a regime like a cell-boundary. For example, if the users have
better precoder is given by very low SNR, which means Pt or Pu is always lower than
Pcross, DL or Pcross, UL , the cloud BS should use MRT or MRC
Pt (M + 1)
Kcross, DL = . (24) to increase a sum rate. Also, the cloud BS should use MRT
1 + Pt
or MRC for users having transmit power larger than Pcross
If the number of users K is larger than Kcross, DL , MRT (especially in the low SNR regime) when the number of active
precoder provides a better sum rate performance. users is larger than Kcross .
MINOR REVISION 9

(a) Achievable rate of ZF at low SNR. (a) Achievable rate of MRC at high SNR.

(b) Achievable rate of MRT at low SNR. (b) Achievable rate of MRC at low SNR.

Fig. 4: Achievable rate vs. the number of cell-boundary users, Fig. 5: Achievable rate vs. the number of cell-boundary users,
where M = 24, K = [1, 24], and total SNR = −13.8 dB. where M = 24, K = [1, 24], and total SNR = (a) 13.8 dB,
(b) −13.8 dB.

B. Performance Comparison for a Large Number of Users

Case ergodic sum rate of MRC when M = K
 
Pt
We derive the ergodic sum rate of ZF with vector nor- lim RMRT, DL (M = K) = M log2 1 +
malization at low SNR (or upper bound of ZF with vector M →∞ Pt + 1
normalization) when M = K lim RMRC, UL (M = K) = M log2 (1 + 1) = M. (27)
M →∞

Pt
 In the special case of MRC, if the transmit power per user is
lim RZFvec , DLL (M = K) = lim K log2 1 + scaling down with M from the sum power constraint Pu, sum ,
M →∞ K→∞ K
 K/Pt i.e., Pu = Pu, sum /M , then
Pt
= lim Pt log2 e ln 1 + 
Pu, sum

K→∞ K lim RMRC, UL (M = K) = M log2 1 + .
= Pt log2 e. (26) M →∞ Pu, sum + 1
(28)
This is equal to the result in [19]. From this result, we are From (27) and (28), we conclude that the performance of MRT
able to gather insights into user scheduling. We also derive and MRC is bounded by M when M = K, even SNR goes
the ergodic sum rate of MRT with any normalization and the to infinity. This differs from the result in [5] (RMRC, UL =
10 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS

M log2 (1+Pu, sum )) because the authors in [5] did not consider
the case of a large number of users.
Next, at Kcross, DL in downlink, we check the difference of
the gradient between the rates of ZF and MRT. If the gradient
of the rate of ZF is larger than that of MRT, the rate of ZF
with vector normalization is larger than that of MRT when
K < Kcross, DL . In the other case, the rate of MRT with matrix
normalization is larger than that of ZF when K ≥ Kcross, DL .
The difference of the gradient between the rates of ZF and
MRT is expressed as

GMRTmat , DLL − GZFvec , DLL

(Pt + 1)2 (M + 1)(Pt + 1)
= − (29)
(M + 1)P ln 4 M P ln 2(2M +1)
where GMRTmat , DLL denotes the gradient of the RMRTmat , DLL
curve at Kcross, DL . Similarly, GZFvec , DLL is the gradient of the
RLZFvec , DL curve at Kcross, DL . In general, cell-boundary users
have relatively low SNR and, as we assumed, the cloud BS (a) Achievable rate at downlink.
has large-scale antennas, meaning M is much larger than
Pt . Therefore, if Kcross, DL exists, (29) is always positive. We
also confirm this through numerical comparisons as shown
in Fig. 3. From this observation, we realize that MRT precod-
ing is suitable for cell-boundary users if the number of active
users is larger than Kcross, DL .

VII. N UMERICAL RESULTS

For numerical comparisons, we assume that each RU has
eight transmit antennas; thus the cloud BS has a total of
24 antennas. Note that any number of antennas can be
used and this constraint is not really related to our system.
This assumption is based on the parameters of 3GPP LTE-
advanced; Release 10 supports eight Node B antennas [24].
Fig. 4(a) shows the achievable sum rate of ZF for downlink
at low SNR. We compare the simulation results with their
theoretical upper bound. As mentioned in Section IV, the
ergodic achievable sum rate of ZF with vector normalization
approaches its upper bound at low SNR while that of ZF
with matrix normalization approaches its first upper bound.
Note that the first upper
n bound o is obtained by Monte Carlo’s
simulation because E ||F1||2 is unknown. Fig. 4(b) also de-
F
scribes the achievable sum rate of MRT with 1/M (−13.8 dB)
total SNR. Our achievable sum rate result is almost the
same as the numerical results for vector normalization. There
is, however, a gap between the simulation results and the
proposed achievable sum rate for matrix normalization. This
is because the proposed achievable sum rate form is accurate
when SNR is lower than 1/M . Our analysis is more accurate
than the closed forms in [10], [15], [16], which are the same as
the lower bound of ZF. In addition, our expression of ergodic
sum rate is more accurate nthan the closedo forms in [10], [16]
that are given by K log2 1 + PtPK+K tM
and these could be
the lower bounds of our expression. From this comparison,
we could confirm (as was also shown in Section IV) that
ZF with vector normalization is better. In contrast, MRT with
matrix normalization is better at getting an improved sum rate
performance at low SNR.
(b) Achievable rate at uplink.
Fig. 6: Achievable rate vs total SNR, where M = 24, K = 20,
and Pth = (a) 6.2 dB, (b) −7 dB.
Figs. 5(a) and 5(b) show that the results from (20) and (21)
are approximately the same as the ergodic achievable uplink
sum rate of MRC where Pu ≥ M and Pu ≤ 1/M , respectively.
Note the large gap between the ergodic achievable uplink sum
rate and the lower bound of MRC with finite M shown in
[5]. The legend, Simulation, indicates the ergodic achievable
uplink sum rate of MRC (19) while the proposed analysis in
Figs. 5(a) and 5(b) indicates the approximation shown in (20)
and (21), respectively.
n Note that the lower
o bound in [5] is given
L Pu (M −1)
by RMRC = K log2 1 + Pu (K−1)+1 and the closed form in
n o
[16] is given by RMRC = K log2 1 + PPu K+1 uM
. In addition,
the sum rate approaches to M when Mn = K at high o SNR,
Pu, sum
and the sum rate approaches to M log2 1 + Pu, sum +1 when
M = K at low SNR. This shows that the results of (27) and
(28) could hold even M is finite.
MINOR REVISION
(a) Achievable rate at downlink.
(b) Achievable rate at uplink.
Fig. 7: Achievable rate vs total SNR, where M = 24, K =
[1, 24], (a) Pt = Pcross,DL = −7.6 dB, (b) Pu = Pcross,UL =
−13.6 dB.
In Fig. 6(a), we illustrate the achievable sum rate of MRT
precoding and ZF precoding at downlink. Fig. 6(b) illustrates
the achievable sum rate of MRC and ZF at uplink (Pth = 6.2 dB
and −7 dB are calculated by Lemmas 6 and 7 with M = 24
and K = 20 at downlink and uplink, respectively). Note that
a cross point of MRT (or MRC) curve and ZF curve is the
power threshold. It is well known that MRT/MRC performs
better than ZF at low SNR. This is the same in the massive
MIMO systems, and we can easily find a borderline between
MRT/MRC region and ZF region with the proposed simple
equation. Figs. 7(a) and 7(b) show that MRT/MRC is always
better than ZF at very low SNR regardless of K. Used for
simulations were Pt = Pcross,DL = −7.6 dB at downlink and
Pu = Pcross,UL = −13.6 dB at uplink, and M = 24. This
result verifies Lemma 8. We summarize also our conclusions
11
(a) Achievable rate at downlink.
(b) Achievable rate at downlink.
Fig. 8: Achievable rate vs. the number of cell-boundary users,
where M = 24, K = [1, 24], and total SNR = (a) 0 dB, (b)
5 dB.
in Tables I, II, and III.
In Fig. 8(a), we also compare the achievable sum rates of ZF
precoding with MRT precoding when the total transmit SNR
is 0 dB. It shows that ZF with vector normalization is better
while MRT with matrix normalization is better for achieving
higher sum rates. Fig. 8(b) illustrates the achievable sum rates
of ZF- and MRT-precoding with −5 dB transmit SNR. The
result is similar to that found in Fig. 8(a). As mentioned in
Section VI, in the low SNR regime, using MRT precoding is
generally better when the number of active users is larger than
Kcross, DL . Also, we realize through Figs. 8(a) and 8(b) that as
SNR decreases, Kcross, DL shifts to the left. This means that
the cloud BS could determine a precoding by Kcross, DL at low
SNR.
The performance of MRT/MRC increases as K increases
12
(a) Achievable rate of MRT at low SNR.
(b) Achievable rate of MRC at low SNR.
Fig. 9: Achievable rate vs. the number of antennas at the BS,
where K = 10, M = [10, 100], and total SNR = -20 dB.
while the performance of ZF decreases as K increases. This
is because the ergodic sum rate of ZF at M = K goes to a
very small constant at downlink, as shown in (26). Similarly,
at uplink, the ergodic sum rate of ZF at M = K could also
close to a very small value since its lower bound is zero.
Figs. 9(a) and 9(b) show that the achievable sum rate
of MRT/MRC at low SNR. To predict the tightness of our
analysis for infinite M , we compare the simulation result of
our analysis with that of the closed form in [16] and the lower
bound in [7] and [14] in the downlink scenario with a large
number of antennas at the BS (up to M = 100). Similarly,
we compare the simulation result of our analysis with that
of the closed form in [16] and the lower bound in [5] in the
uplink scenario. Both numerical results of MRT/MRC show
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS
TABLE I: Precoding normalization techniques in network
massive MIMO systems.
Precoding normalization technique
ZF Vector normalization ≥ Matrix normalization
MRT Matrix normalization ≥ Vector normalization
TABLE II: Optimal switching point in network massive MIMO
systems.
Pth Kcross
K2 Pt (M +1)
Downlink (K−1)(M −K+1) 1+Pt
1 1
Uplink M −K+1
M +1− Pu
TABLE III: Desired technique in network massive MIMO
systems. ZF (if K ≤ Kcross ) and MRT/MRC (if K ≥ Kcross ).
Kcross Precoding technique
Cell-center Large Zero-forcing
Cell-boundary Small MRT/MRC
that, when M goes to infinity, our analyses are quite tighter
than the work given in [5], [7], [14], [16]. Note that the closed
form in [16] is very tight when M = 40 but as M increases not
as tight as our analysis in the low SNR regime and the uplink
scenario. The computation complexity is the same as the lower
bounds in [5], [7], [14] but the accuracy is better than the
closed form in [16]. Therefore, we conclude that the proposed
ergodic sum rate is tighter and simpler than the previous work.
VIII. C ONCLUSIONS
In this paper, we proposed massive MIMO systems support-
ing multiple cell-boundary UEs. For precoding designs, we
first derived the achievable sum rate bounds of zero-forcing
(ZF) and the approximation of the ergodic achievable sum
rate of maximum ratio transmission (MRT) with vector/matrix
normalization. Through analytical and numerical results, we
confirmed that vector normalization is better for ZF and that
matrix normalization is better for MRT. We also investigated
the optimal mode-switching point as functions of the power
threshold and the number of active users in a network. Ac-
cording to the mathematical and numerical results the BS can
select a transceiver mode to increase the sum rate for both
downlink and uplink scenarios. We anticipate our analysis
providing insights for related studies, such as those on per-
formance analysis of MIMO, power normalization methods,
and the trade-off between ZF and MRT/MRC. In future work,
we will consider more practical scenarios including limited
cooperation among RUs and cooperation delay.
A PPENDIX
A. Proof of Lemma 1
Let hk,m be the m-th element of h k . Since hk,m is an i.i.d.
complex Gaussian random variable with zero mean and unit
variance, i.e., hk,m ∼ CN (0, 1), |hk,m |2 is a Gamma random
MINOR REVISION 13

variable with unit shape parameter and unit scale parameter, the ergodic achievable sum rate of vector normalization. We
i.e., |hk,m |2 ∼ Γ(1, 1) from the relationship between Rayleigh can then obtain the following equations:
distribution and Gamma distribution. Therefore, we can say
hk ||2 ≈ Pt M
Pt ||h (32)
that |hk,m |2 is an exponential random variable with unit
parameter (λ = 1), i.e., |hk,m |2 ∼ Exp(1) by the property at low SNR, and
of Gamma distribution and exponential distribution. Since the 1 1
n-th moment of the exponential random variable is n! hk ||2 ≈ M
||h (33)
λ , we can Pt Pt
obtain E(|hk,m |4 ) = 2, E(|hk,m |6 ) = 6, and E(|hk,m |8 ) = 24.
at high SNR from Lemmas 1 and (5).
1) Proof of Lemma 1. 3): The expectation of ||h hk ||4 is given
Next, we assume that Var[Pt ||h h∗kh ` |2 ] con-
hk ||4 ] and Var[|h
by
verge to zero at high SNR to evaluate the ergodic achievable
hk ||4 ] = E{(|hk,1 |2 + |hk,2 |2 + · · · + |hk,M |2 )2 }
E[||h sum rate of matrix normalization. Also, Var[ P1t ||H H ||2F ] con-
verges to zero at high SNR. Similar to the vector normalization
= M E(|hk,m |4 ) + M P2 E(|hk,m |2 |hk,i |2 )
case, we can obtain the following equations:
= 2M + M (M − 1) = M 2 + M,
Pt ||h h∗k h` |2 ≈ Pt M
hk ||4 ≈ Pt (M 2 + M ), Pt |h (34)
where i 6= m. The notation nPr denotes a permutation
hk ||4 is also derived by
operator. The variance of ||h at low SNR, and
K
1 1 X 1
hk ||4 ]
Var[||h H ||2F =
||H hk |2 ≈ KM
|h (35)
4 2 2 Pt Pt Pt
= M Var(|hk,m | ) + M P2 Var(|hk,m | |hk,i | ) k=1

+ M P2 Cov(|hk,m |2 |hk,i |2 , |hk,i |2 |hk,m |2 ) at high SNR.

4 2 2 1) Proof of (11):
+ 4 M P2 Cov(|hk,m | , |hk,i | |hk,m | )
hk ||4
  
+ 4 M P3 Cov(|hk,m |2 |hk,i |2 , |hk,m |2 |hk,j |2 ) XK  Pt ||√||hKhhk || 2

RMRTvec , DLL = E  log2 1 + PK ∗
hk h ` |2
|h

= 20M + 3M (M − 1) + 3M (M − 1) k=1
 Pt `=1,`6=k ||√Kh 2
+ 1 
h` ||
+ 16M (M − 1) + 4M (M − 1)(M − 2)  
hk ||2
||h

 Pt K 
= 4M 3 + 10M 2 + 6M. (30) = KE log2 1 + PK ∗ 2

|h
hkh ` |
 Pt `=1,`6=k K||h h` ||2 + 1 
The variance terms and the covariance terms in (30) are   
calculated by (a)
 Pt MK

≈ KE log2 1 + PK |h ∗
hkh ` | 2

Var(|hk,m |4 ) = 20,  Pt `=1,`6=k K||h h` || 2 + 1 
Var(|hk,m |2 |hk,i |2 ) = 3,
 
M
(b)
 Pt K 
Cov(|hk,m |2 |hk,i |2 , |hk,i |2 |hk,m |2 ) = 3, ≈ K log2 1 + PK h∗ 2
E(|h kh ` | )
 Pt `=1,`6=k E(K||h h` ||2 ) + 1

Cov(|hk,m |4 , |hk,i |2 |hk,m |2 ) = 4,  
(c) Pt M
Cov(|hk,m |2 |hk,i |2 , |hk,m |2 |hk,j |2 ) = 1 = K log2 1 +
Pt (K − 1) + K
where i 6= j 6= m.
where (a) results from (32).4 Approximation (b) and equality
h∗kh ` |2 is
2) Proof of Lemma 1. 4): The expectation of |h
(c) can be obtained by Lemmas 4 and 1, respectively.
h∗kh ` |2 ] = E(|h∗k,1 h`,1 + h∗k,2 h`,2 + · · · + h∗k,M h`,M |2 )
E[|h 2) Proof of (12):
= M E(|h∗k,m h`,m |2 ) + M P2 E(h∗k,m h`,m h∗k,i h`,i ) hk ||4
  
XK  Pt ||√||hKh hk ||2

=M RMRTvec , DLH = E  log2 1 + PK |h ∗
hkh ` | 2

k=1
 P t `=1,`6=k || Kh
√
h` ||2
+ 1 
h∗kh ` |2 is also expressed as
where i 6= m. The variance of |h  
||h
h k || 4

 Pt K||hhk ||2 
h∗kh ` |2 ] = Var(|h∗k,1 h`,1 + h∗k,2 h`,2 + · · · + h∗k,M h`,M |2 )
Var[|h = KE log2 1 + PK |h ∗
hkh ` | 2

 Pt `=1,`6=k K||h h` || 2 + 1 
= M Var(|h∗k,m h`,m |2 )
hk ||4
" ( ||h
)#
+ M P2 Var(h∗k,m h`,m h∗k,i h`,i ) (d) Pt KM
≈ KE log2 1 + PK |hh∗ 2
kh ` |
= 3M + M (M − 1) = M 2 + 2M (31) Pt `=1,`6=k KM +1
 
(e) Pt (M + 1)
where Var(|h∗k,m h`,m |2 ) is 3, and Var(h∗k,m h`,m h∗k,i h`,i ) is 1 ≈ K log2 1 +
Pt (K − 1) + K
(i 6= m). Note that the covariance terms in (31) are all zeros.
where (d) results from (33) and (e) can be obtained by
Lemmas 1 and 4.
B. Proof of Ergodic Sum Rate of MRT
4 Since Var[||h hk ||2 ](= M ) < Var|hh∗k h` |2 ](= M 2 + 2M ) from Lemma 1,
First, we suppose that Var[Pt ||hhk ||2 ] and Var[ P1t ||h
hk ||2 ] h∗k h ` |2 ] at low SNR. We
hk ||2 ] fastly converges to zero than Var[Pt |h
Var[Pt ||h
converge to zero at low and high SNR, respectively, to evaluate use this property for the rest of the proofs.
14 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS

3) Proof of (13): in the low SNR regime 2) Proof of (21):

RMRC, ULL
RMRTmat , DLL
" ( )#
hk ||4
Pu ||h
 
hk ||4
 = E K log2 1+
XK  Pt ||h 2
||H ||F
H
 Pu
PK
h∗kh ` |2 + ||h
|h hk ||2
`=1,`6=k
= E log2 1 + PK h∗
|h 2

kh ` |
  
k=1
 Pt `=1,`6=k ||H H ||2F
+ 1  Pu ||h
hk ||2 
" ( )# = E K log2 1 + PK h∗ h |2
|h

hk ||4
Pt ||h  Pu `=1,`6=k ||hhkk ||` 2 + 1 
= KE log2 1 + PK
Pt `=1,`6=k |h h∗kh ` |2 + ||H H ||2F
  
 P u M 
≈ E K log2 1 +
" ( )#
Pt (M 2 + M )
(f ) ∗

PK h h |2
|h
≈ KE log2 1 + PK  Pu `=1,`6=k ||hhkk ||` 2 + 1 
Pt `=1,`6=k M + ||H H ||2F  
Pu M
≈ K log2 1 +
 
(g) Pt (M + 1)
≈ K log2 1 + Pu (K − 1) + 1
Pt (K − 1) + K
using the same methods as in (36).
where (f ) can be obtained by using (34) directly. (g) can also
be derived by Lemma 4 and E(||H H ||2F ) = M K. R EFERENCES
[1] D. Gesbert, M. Kountouris, R. W. Heath, Jr., C.-B. Chae, and T. Salzer,
4) Proof of (13): in the high SNR regime “Shifting the MIMO paradigm: From single user to multiuser commu-
nications,” IEEE Sig. Proc. Mag., vol. 24, no. 5, pp. 36–46, Oct. 2007.
[2] Q. Spencer, A. L. Swindlehurst, and M. Haardt, “Zero-forcing methods
RMRTmat , DLH for downlink spatial multiplexing in multiuser MIMO channels,” IEEE
 
hk ||4
 Trans. Sig. Proc., vol. 52, pp. 462–471, Feb. 2004.
X K  Pt ||h H ||2F
||H
 [3] C.-B. Chae, D. Mazzarese, N. Jindal, and R. W. Heath, Jr., “Coordinated
= E log2 1 + PK |h ∗
hkh ` | 2
 beamforming with limited feedback in the MIMO broadcast channel,”
k=1
 P t `=1,`6=k ||H H ||2F
+ 1  IEEE Jour. Select. Areas in Comm., vol. 26, no. 8, pp. 1505–1515, Oct.
" ( )# 2008.
||hhk ||4 [4] C.-B. Chae and R. W. Heath, Jr., “On the optimality of linear multiuser
= KE log2 1 + PK MIMO beamforming for a two-user two-input multiple-output broadcast
h∗kh ` |2 + P1t ||H
`=1,`6=k |h H ||2F system,” IEEE Sig. Proc. Lett., vol. 16, no. 2, pp. 117–120, Feb. 2009.
" ( )# [5] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral
(h) ||hhk ||4 efficiency of very large multiuser MIMO systems,” IEEE Trans. Comm.,
≈ KE log2 1 + PK vol. 61, no. 4, pp. 1436–1449, April 2012.
h∗k h` |2 + P1t M K
`=1,`6=k |h [6] T. L. Marzetta, “Noncooperative cellular wireless with unlimited num-
bers of base station antennas,” IEEE Trans. Wireless Comm., vol. 9,
 
(i) Pt (M + 1)
≈ K log2 1 + no. 11, pp. 3590–3600, Nov. 2010.
Pt (K − 1) + K [7] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Mazetta, O. Edfors,
and F. Tufvesson, “Scaling up MIMO: Opportunities and challenges with
large arrays,” IEEE Sig. Proc. Mag., vol. 30, no. 1, pp. 40–60, Jan. 2013.
where (h) results from (35) while (i) can be obtained by [8] C.-B. Chae, S. Kim, and R. W. Heath, Jr., “Network coordinated beam-
Lemmas 1 and 4 as well. forming for cell-boundary users: Linear and non-linear approaches,”
IEEE Jour. Select. Topics in Sig. Proc., vol. 3, no. 6, pp. 1094–1105,
Dec. 2009.
[9] C.-B. Chae, I. Hwang, R. W. Heath, Jr., and V. Tarokh, “Interference
aware-coordinated beamforming in a multi-cell system,” IEEE Trans.
Wireless Comm., vol. 11, no. 10, pp. 3692–3703, Oct. 2012.
C. Proof of Ergodic Sum Rate of MRC [10] H. Huh, G. Caire, H. C. Papadopoulos, and S. A. Ramprashad, “Achiev-
ing massive MIMO spectral efficiency with a not-so-large number of
1) Proof of (20): antennas,” IEEE Trans. Wireless Comm., vol. 11, no. 9, pp. 3226–3239,
Sep. 2012.
[11] J. Jose, A. Ashikhmin, T. L. Marzetta, and S. Vishwanath, “Pilot
contamination and precoding in multi-cell TDD systems,” IEEE Trans.
RMRC, ULH Wireless Comm., vol. 10, no. 8, pp. 2640–2651, Aug. 2011.
" ( )#
[12] J. Nam, J.-Y. Ahn, A. Adhikary, and G. Caire, “Joint spatial division
hk ||4
Pu ||h and multiplexing: Realizing massive MIMO gains with limited channel
= E K log2 1+ PK
Pu `=1,`6=k h∗kh ` |2 + ||h
|h hk ||2 state information,” in Proc. of Conf. on Info. Scien. and Systems, March
" ( )# 2012, pp. 1–6.
hk ||4
||h [13] M. Filippou, D. Gesbert., and H. Yin, “Decontaminating pilots in
= E K log2 1 + PK 1
cognitive massive MIMO networks,” in Proc. of Int. Symp. on Wireless
`=1,`6=k h∗kh ` |2 +
|h h 2
Pu ||h k ||
Comm. Systems, Aug. 2012, pp. 816–820.
" ( )# [14] H. Yang and T. L. Marzetta, “Performance of conjugate and zero-forcing
hk ||4
||h beamforming in large-scale antenna systems,” IEEE Jour. Select. Areas
≈ E K log2 1 + PK 1 in Comm., vol. 31, no. 2, pp. 172–179, 2013.
h∗kh ` |2 +
|h
`=1,`6=k Pu M [15] H. Huh, A. M. Tulino, and G. Caire, “Network MIMO with linear
  zero-forcing beamforming: Large system analysis, impact of channel
Pu (M + 1)
≈ K log2 1+ . (36) estimation, and reduced-complexity scheduling,” IEEE Trans. Info. Th.,
Pu (K − 1) + 1 vol. 58, no. 5, pp. 2911–2934, May 2012.
[16] J. Hoydis, S. ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need?” IEEE Jour.
using (32), (33), Lemmas 1 and 4. Select. Areas in Comm., vol. 31, no. 2, pp. 160–171, Feb. 2013.
MINOR REVISION 15

[17] A. M. Tulino and S. Verdu, “Random matrix theory and wireless
communications,” Foundations and Trends in Comm. and Info. Th.,
vol. 1, no. 1, 2004.
[18] G. Caire, N. Jindal, M. Kobayashi, and N. Ravindran, “Multiuser MIMO
achievable rates with downlink training and channel state feedback,”
IEEE Trans. Info. Th., vol. 56, no. 6, pp. 2845–2866, June 2010.
[19] C. Peel, B. Hochwald, and A. Swindlehurst, “A vector-perturbation
technique for near-capacity multiantenna multiuser communication-part
I: channel inversion and regularization,” IEEE Trans. Comm., vol. 53,
no. 1, pp. 195–202, Jan. 2005.
[20] B. Hochwald, C. Peel, and A. Swindlehurst, “A vector-perturbation
technique for near-capacity multiantenna multiuser communication-part
II: perturbation,” IEEE Trans. Comm., vol. 53, no. 3, pp. 537–544, March
2005.
[21] R. H. Y. Louie, M. R. McKay, and I. B. Collings, “Maximum sum-
rate of MIMO multiuser scheduling with linear receivers,” IEEE Trans.
Comm., vol. 57, no. 11, pp. 3500–3510, Nov. 2009.
[22] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
Products. Academic Press, 2007.
[23] K. K. Wong and Z. Pan, “Array gain and diversity order of multiuser
MISO antenna systems,” Int. J. Wireless Inf. Networks, vol. 2008, no. 15,
pp. 82–89, May 2008.
[24] S. Sesia, I. Toufik, and M. Baker, LTE, The UMTS Long Term Evolution:
From Theory to Practice. Wiley, 2012.
Yeon-Geun Lim (S’12) received his B.S. degree in
Information and Communications Engineering from
Sungkyunkwan University, Korea in 2011. He is now
with the School of Integrated Technology, Yonsei
PLACE University, Korea and is working toward the Ph.D.
PHOTO degree.
HERE His research interest includes massive MIMO and
interference management techniques for smart small
cell networks.
Guiseppe Caire (S’92 - M’94 - SM’03 - F’05) was
born in Torino, Italy, in 1965. He received the B.Sc.
in Electrical Engineering from Politecnico di Torino
(Italy), in 1990, the M.Sc. in Electrical Engineering
PLACE from Princeton University in 1992 and the Ph.D.
PHOTO from Politecnico di Torino in 1994. He was a re-
HERE cipient of the AEI G.Someda Scholarship in 1991,
has been with the European Space Agency (ESTEC,
Noordwijk, The Netherlands) from May 1994 to
February 1995, was a recipient of the COTRAO
Scholarship in 1996 and of a CNR Scholarship
in 1997. He has been visiting Princeton University in Summer 1997 and
Sydney University in Summer 2000. He has been Assistant Professor in
Telecommunications at the Politecnico di Torino, Associate Professor at
the University of Parma, Italy, Professor with the Department of Mobile
Communications at the Eurecom Institute, Sophia-Antipolis, France, and a
professor of Electrical Engineering with the Viterbi School of Engineering,
University of Southern California, Los Angeles, CA. He is now a professor
with the School of Electrical Engineering and Computer Science, Technical
University of Berlin, Germany. He served as Associate Editor for the IEEE
TRANSACTIONS ON COMMUNICATIONS in 1998-2001 and as Associate
Editor for the IEEE TRANSACTIONS ON INFORMATION THEORY in
2001-2003. He received the Jack Neubauer Best System Paper Award from the
IEEE Vehicular Technology Society in 2003, and the IEEE Communications
Society & Information Theory Society Joint Paper Award in 2004 and in 2011.
Giuseppe Caire is Fellow of IEEE since 2005. He has served in the Board of
Governors of the IEEE Information Theory Society from 2004 to 2007, and as
President of the IEEE Information Theory Society in 2011. His main research
interests are in the field of communications theory, information theory, channel
and source coding with particular focus on wireless communications.

Chan-Byoung Chae (S’06 - M’09 - SM’12) is

an Assistant Professor in the School of Integrated
Technology, College of Engineering, Yonsei Univer-
sity, Korea. He was a Member of Technical Staff
PLACE (Research Scientist) at Bell Laboratories, Alcatel-
PHOTO Lucent, Murray Hill, NJ, USA from 2009 to 2011.
HERE Before joining Bell Laboratories, he was with the
School of Engineering and Applied Sciences at
Harvard University, Cambridge, MA, USA as a Post-
Doctoral Research Fellow. He received the Ph.D.
degree in Electrical and Computer Engineering from
The University of Texas (UT), Austin, TX, USA in 2008, where he was a
member of the Wireless Networking and Communications Group (WNCG).
Prior to joining UT, he was a Research Engineer at the Telecommunications
R&D Center, Samsung Electronics, Suwon, Korea, from 2001 to 2005.
While having worked at Samsung, he participated in the IEEE 802.16e
standardization, where he made several contributions and filed a number of
related patents from 2004 to 2005. His current research interests include
capacity analysis and interference management in energy-efficient wireless
mobile networks and nano (molecular) communications. He has served/serves
as an Editor for the IEEE T RANS . ON W IRELESS C OMM ., IEEE T RANS . ON
S MART G RID and IEEE/KICS J OUR . C OMM . N ETS. He is a Guest Editor
for the IEEE J OUR . S EL . A REAS IN C OMM . (special issue on molecular,
biological, and multi-scale comm.). He is an IEEE Senior Member.
Dr. Chae was the recipient/co-recipient of the IEEE INFOCOM Best
Demo Award (2015), the IEIE/IEEE Joint Award for Young IT Engineer
of the Year (2014), the Haedong Young Scholar Award (2013), the IEEE
Signal Processing Magazine Best Paper Award (2013), the IEEE ComSoc AP
Outstanding Young Researcher Award (2012), the IEEE VTS Dan. E. Noble
Fellowship Award (2008), the Gold Prize (1st) in the 14th/19th Humantech
Paper Contests, and the KSEA-KUSCO scholarship (2007). He also received
the Korea Government Fellowship (KOSEF) during his Ph.D. studies.