This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LWC.2018.2841375, IEEE Wireless
Communications Letters
1

NOMA-Aided Cell-Free Massive MIMO Systems

Yikai Li, Student Member, IEEE and Gayan Amarasuriya Aruma Baduge, Member, IEEE

Abstract—A cell-free massive multiple-input multiple-output sum rate is derived for a design of practically viable cell-
system with non-orthogonal multiple-access (NOMA) is investi- free massive MIMO-NOMA system. In contrast to [3]–[8], the
gated. An achievable sum rate is derived and compared against joint detrimental effects of intra-cluster pilot contamination,
the orthogonal multiple access (OMA) counterpart. Thereby, the
detrimental effects of intra-cluster pilot contamination, inter- inter-cluster interference, and imperfect SIC are investigated.
cluster interference and imperfect successive interference can- This achievable sum rate is compared against the cell-free
cellation (SIC) are investigated. The number of users served massive MIMO-OMA counterpart. Our results reveal that
simultaneously by NOMA can be significantly higher than that NOMA outperforms OMA in terms of the sum rate in the
of OMA. Nevertheless, the achievable sum rate of NOMA can be regime of large number of users. However, the intra-cluster
lower than that of OMA even in the regime of low number
of users due to intra-cluster pilot contamination and error pilot contamination and residual-interference of imperfect SIC
propagation of imperfect SIC. degrade the sum rate gains of NOMA in the regime of low
number of users.
Index Terms—NOMA, massive MIMO, performance analysis
Notation: zH , z∗ , and [z]k denote the conjugate-transpose,
I. I NTRODUCTION conjugate, and the kth element of a vector z, respectively.
In cell-free massive multiple-input multiple-output (MIMO) E[z] and Var[z] are the expectation and variance operators of
systems, multiple users are served simultaneously in the a random variable z, respectively.
same time-frequency resource block via spatial multiplexing II. S YSTEM , CHANNEL AND SIGNAL MODELS
by spatially-distributed access points (APs) [1], [2]. Cell- A. System and Channel Model
free massive MIMO can provide significant spectral/energy We consider the downlink transmission of a NOMA-aided
efficiency gains by reaping benefits of both network MIMO cell-free massive MIMO system in which M single-antenna
and (co-located) massive MIMO [1], [2]. APs serve KN single-antenna users spatially distributed in
The orthogonal multiple-access (OMA) techniques are ap- N clusters in the same time-frequency resource block. Each
proaching their fundamental spectral efficiency limits. To cluster consists of K users. APs are connected to a central
this end, non-orthogonal multiple-access (NOMA) can be processing unit (CPU) via a perfect backhaul network [1].
exploited to meet the conflicting demands on high spec- The downlink channel between the mth AP and the kth user
tral/energy efficiency, massive connectivity with low latency, in the nth cluster, where m ∈ {1, · · · , M }, k ∈ {1, · · · , K}
and high reliability with user-fairness of next-generation wire- and n ∈ {1, · · · , N }, can be modeled as
less standards [3]. In NOMA, multiple users can be served in
1/2
the same time/frequency/spatial resource block [4]. hmnk = ζmnk h̃mnk , (1)
A combination of MIMO and NOMA has been shown to
where ζmnk captures the large-scale fading, which is assumed
boost the sum rates [5], [6]. The complexity of such systems
to be known a priori as it changes very slowly and hence
can be significantly reduced by decomposing the MIMO-
only needs to be estimated once about every 40 coherence
NOMA into multiple single-input single-output NOMA sub-
time intervals [9]. Moreover, h̃mnk ∼ CN (0, 1) is circularly
channels [6], [7]. In [8], the design of massive MIMO NOMA
symmetric Gaussian distributed with zero mean unit variance
systems is investigated, and the outage probability is derived
and accounts for quasi-static Rayleigh fading.
for perfect user-ordering and limited feedback. In [8], the
users are grouped into spatial-clusters, and a null-space based B. Channel State Information Acquisition
precoder is used to cancel the inter-cluster interference [8]. The APs estimate the uplink channels via the pilots trans-
However, such null-space based precoding [6]–[8] can be mitted by the users, and thereby, the downlink channels are
prohibitively complicated for massive MIMO. The detrimental obtained by exploiting the time-division duplexing (TDD)
effects of estimated/imperfect channel state information (CSI) channel reciprocity. To minimize the channel estimation over-
at APs, imperfect successive interference cancellation (SIC), head, the users within the same cluster are allocated the
and statistical CSI at users have not yet been studied for same pilot sequence with τ symbols in length. The N pilot
NOMA-aided massive MIMO. This gap in NOMA-aided sequences allocated for N clusters are mutually orthogonal,
massive MIMO literature motivates this work. and hence τ ≥ N . The pilot sequence allocated for K users in
In this letter, the design of cell-free massive MIMO- the nth cluster is denoted as φn ∈ Cτ ×1 satisfying kφn k2 = 1
NOMA systems is investigated. Superposition-coded signals and φHn φl = 0 for n 6= l. The pilot signal received at the mth
are spatially-multiplexed via conjugate beamforming into mul- AP can be written as
tiple user-clusters in the same time-frequency resource block N X K
p √ X
by the APs. The users within each cluster are served via power- ym = τ p p hmnk φn + nm , (2)
domain NOMA with user-ordering and SIC. An achievable n=1 k=1

where pp is the pilot transmit power, and nm ∼

The authors are with the Department of Electrical and Computer Engi-
neering, Southern Illinois University, Carbondale, IL, USA 62901, Email: CN τ ×1 (0τ ×1 , Iτ ) represents the additive white Gaussian
{yikai.li,gayan.baduge}@siu.edu. noise (AWGN) vector at the mth AP.

2162-2337 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LWC.2018.2841375, IEEE Wireless
Communications Letters
2

k−1 N K
√ X p X X p
ỹnk = Pnk cnk xnk + cnk Pnk0 xnk0 + cn0 k xn0 + Pnk00 [cnk xnk00 − E[cnk ] x̂nk00 ] + nnk (12)
k0 =1 n0 =1,n0 6=n k00 =k+1
| {z } | {z }
inherent intra-cluster interference after SIC error propagation or residual interference due to imperfect SIC

K N
In order to estimate the channel hmnk , the received pilot = √P c x + c XpP 0 x 0 + X c 0 x 0 + n ,(10)
signal at the mth AP (ym p
) is projected onto φn as [1] | nk{znk nk} nk
0 =1,k0 6=k
nk nk
0 =1,n0 6=n
n k n nk
|{z}
k n
K [desired signal]
} [AWGN]
√ X | {z } | {z
p H p H [intra-cluster interference before SIC] [inter-cluster interference]
ỹmn = φn ym = τ pp hmnk + φn nm . (3)
PM ∗
vmn PM v∗
k=1 where cnk = m=1 hmnk |vmn | , cn0 k = m=1 hmnk |vmn00 |
In the case that any two pilot sequences are either identical mn
p and nnk ∼ CN (0, 1). In order to apply the power-domain
or orthogonal, ỹmn becomes a sufficient statistic. Thus, the
p NOMA, the users in the nth cluster are assumed to be ordered
MMSE estimate of hmnk given ỹmn can be derived as [10]
√ based on the effective channel strengths as follows [7], [8]:
p∗ τ pp ζmnk
E[ỹmn hmnk ] p p M ∗
2 M ∗
2 M ∗
2
ĥmnk = p ỹ = ỹmn . (4) X vmn X vmn X vmn
|2 ] mn
PK
E[|ỹmn 1 + τ pp k=1 ζmnk ĥmn1 ≥ ĥ mn2 ≥ · · · ≥ ĥ mnK .(11)
m=1
|vmn | m=1
|vmn | m=1
|vmn |
p
By using the fact that ỹmn is Gaussian distributed, ĥmnk can
be rewritten as In power-domain NOMA, higher powers are allocated for
√ the users with lower channel strengths yielding Pn1 ≤ · · · ≤
ĥmnk = ηmnk vmn , (5)
Pnk ≤ · · · ≤ PnK [7], [8]. Hence, within the nth cluster, the
where vmn ∼ CN (0, 1), and ηmnk is defined as kth user is always able to decode the signal intended for the
2
h i τ pp ζmnk lth user for ∀l ≥ k provided that the kth user can decode
ηmnk = E |ĥmnk |2 = PK . (6)
1 + τ pp k=1 ζmnk its own signal. Thus, the kth user can successively cancel the
intra-cluster interference from the lth user before decoding its
The channel estimation error is defined as εmnk = hmnk −
own signal for ∀l ≥ k. The kth user treats the signals from
ĥmnk , where εmnk
 and ĥmnk are statistically independent the users for ∀l < k as interference [7], [8]. The optimal
[10]. Moreover, E |εmnk |2 = ζmnk − ηmnk .
power allocation and optimal user-clustering are left as open
C. Signal Model problems for the sake of brevity.
In the downlink data transmission, the APs employ a Remark 1: In TDD cell-free massive MIMO, the instanta-
conjugate beamformer designed based on the locally estimated neous CSI is not available at the user nodes [1]. However,
CSI, which is acquired via channel reciprocity and the uplink as the number of APs grows sufficiently large, the underlying
MMSE channel estimation. The data signal intended for the channels harden. Hence, E[cnk ] can be used as the effective
K users in the nth cluster is superposition coded as channel gain for decoding xn at the kth user in the nth
X K p cluster [1]. The acquisition of E[cnk ] is not practically difficult
xn = Pnk xnk , (7) because it depends only on the statistical properties of the
k=1 channels and remains fixed for several coherence intervals.
where xnk and Pnk are the data signal and the transmit power Remark 2: Due to intra-cluster pilot contamination, channel
allocated for the kth user in the nth cluster, respectively, for estimation errors and statistical CSI knowledge at the users,
n ∈ {1, · · · , N } and k ∈ {1, · · · , K}. Furthermore, xnk and perfect SIC is not practically viable. Hence, the post-processed
xml for m, n ∈ {1, · · · , N } and k, l ∈ {1, · · · , K} satisfy signal at the kth user node in the nth cluster after the imperfect
( SIC can be written as (12) at the top of this page.
1, if n = m and k = l In (12), x̂nk is an estimate of xnk for ∀n, k. Without loss
E[xnk x∗ml ] = (8)
0, otherwise. of generality, xnk is assumed to be drawn from a Gaussian
h i P
2 K
Therefore E |xn | = k=1 Pnk = Pn , where Pn accounts distribution with zero-mean and unit variance. Hence, x̂nk and
xnk are assumed to be jointly Gaussian distributed with a
for the total signal power allocated for the nth cluster.
normalized correlation coefficient ρnk as
The transmitted signal at the mth AP can be written as
N N
xnk = ρnk x̂nk + enk , (13)
X ĥ∗mnk X ∗
vmn
tm = xn = xn , (9) 2 2
|v mn | nk ∼ CN (0, 1), enk ∼ CN (0, σenk /[1 + σenk ]) and
where x̂p
n=1 ĥmnk n=1
2
ρnk = 1/ 1 + σenk . x̂nk and enk are statistically independent.
where a short-term power constraint (STPC) is used in the
conjugate beamformer design. The total average transmit III. ACHIEVABLE S UM R ATE A NALYSIS
power at the mth P AP for all N clusters can be written as With the statistical CSI knowledge at the users, the signal
2 N
Ptm = E[|tm | ] = n=1 Pn . at the kth user in the nth cluster can be written as
The KN users in N clusters are served simultaneously by p
M APs. Hence, the signal received at the kth user in the nth ỹnk = Pnk E[cnk ] xnk + ñnk , (14)
cluster can be written as where the first term captures desired signal component and the
M
X second term represents the effective noise, which is defined in
ynk = hmnk tm + nnk
m=1
(15) at the top of next page. By using (14), an achievable rate

2162-2337 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LWC.2018.2841375, IEEE Wireless
Communications Letters
3

k−1 N K
√ X p X X p
ñnk = Pnk (cnk − E[cnk ]) xnk + cnk Pnk0 xnk0 + cn0 k xn0 + Pnk00 [cnk xnk00 − E[cnk ] x̂nk00 ] + nnk (15)
k0 =1 n0 =1,n0 6=n k00 =k+1

Pnk |cnk |2
γnk = ! (21)
k−1 N K
P 2 P 2 P 2 2
Pnk0 |cnk | + Pn0 |cn0 k | + Pnk00 |cnk xnk00 − E[cnk ] x̂nk00 | + σnk
k0 =1 n0 =1,n0 6=n k00 =k+1

for the kth user in the nth cluster can be derived by using the number of users that can be served simultaneously in the
worst-case Gaussian technique as follows [1]: same time-frequency resource block by the proposed NOMA-
aided massive MIMO does not depend on the number of users
4
!!
. X
2
R̃nk = φlog 1 + Pd PIi + σnk , (16) within a cluster. Although, the number of clusters depend
i=1
on the maximum orthogonal pilot sequence length that can
where φ = (τC −τ )/τC is a pre-log factor, τC is the coherence be accommodated within a coherence interval, the maximum
interval, τ is the pilot sequence length, Pd is the desired signal number of users served can grow without bound theoretically.
power, and PIi ’s for i ∈ {1, · · · , 4} account for the powers of However, when the number of users of each cluster grows
the effective interference terms, which can be defined as large, the sum rate gain may become infinitesimal and the
Pd = Pnk |E[cnk ]|2 and PI1 = Pnk Var[cnk ] , (17a) complexity of SIC becomes prohibitively complicated.
Remark 5: In order to guarantee user-fairness, the transmit
k−1
!
X
Pnk0 E |cnk |2 ,
 
PI2 = (17b) power can be allocated for each AP based on a max-min
k0 =1
fairness optimal power control algorithm to yield a common
N
PI3 =
X
Pn0 E |cn0 k |2 ,
 
(17c)
achievable rate for all users across all clusters. Nevertheless,
n0 =1,n0 6=n
it is left as an open research problem.
K
X IV. N UMERICAL R ESULTS
Pnk00 E |cnk xnk00 − E[cnk ] x̂nk00 |2 .
 
PI4 = (17d)
k00 =k+1
In the NOMA case, the same pilot sequence is shared among
By evaluating the expectation and variance operators in users within a cluster, and the pilots assigned for multiple
(17a)-(17d), the achievable sum rate of the proposed NOMA- clusters are mutually orthogonal. In the OMA case, each
aided cell-free massive MIMO system can be written as user is assigned with an orthogonal pilot. The minimum pilot
N X
X K sequence lengths for NOMA and OMA are τNOMA = N and
R̃ = R̃nk , (18) τOMA = KN , respectively. The pre-log factors for NOMA
n=1 k=1
and OMA cases are defined as φNOMA = (τc − N )/τc and
where R̃nk is defined in (16) and PIj for j ∈ {1, · · · , 4} can
φOMA = (τc − N K)/τc , respectively.
be derived in closed-form as (see Appendix for the derivation)
!2 The users are clustered based on their spatial locations,
M M 
π X √ X πηmnk  and all clusters of users are uniformly distributed over a
Pd = Pnk ηmnk ζmnk −
, PI1 = Pnk , (19a)
4 m=1 m=1
4 1000 m × 1000 m square area. The users located within each
k−1
! M
!2 M 
cluster are uniformly scattered around the center point of a
X π X √ X πη mnk

PI2= Pnk0  ηmnk + ζmnk − , (19b) cluster. Moreover, M = 100 APs are uniformly distributed in
4 m=1 4
k0 =1 m=1 the given area. The path-loss is modeled as ζmnk = d−β mnk ,
where β is the path-loss exponent and dmnk is the distance
 
X N XM
PI3=  Pn0  ζmnk , (19c) between the mth AP and the kth user node in the nth cluster.
n0 =1,n0 6=n m=1 In Fig. 1, the achievable sum rates of the cell-free massive
K

M  M
!2  MIMO systems with NOMA and OMA are plotted as a
X X πηmnk
 π(1−ρ 00) X
nk √
PI4 = Pnk00 ζmnk− + ηmnk . (19d) function of the number of users. The analytical sum rate
00 m=1
4 2 m=1 curves, which are plotted by using (18), are compared against
k =k+1

Remark 3: The sum rate in (18) is practically achievable with Monte-Carlo simulations of (18) with the same worst-case
the availability of statistical CSI knowledge at the users. In Gaussian assumption, and the ergodic sum rate upper bound
Section IV, the achievable rate in (18) is compared with the in (20). The coherence interval is set to τc = 196, and hence,
ergodic sum rate, which is only achievable at the genie-aided the maximum number of users, which can be served in the
users with instantaneous CSI via a stronger capacity bound. same time-frequency resource block by OMA is limited to
This ergodic sum rate is an upper bound for the achievable 196. However, NOMA can serve twice the users that of OMA
sum rate in (18) and can be written as when two users are grouped into each cluster. Nevertheless,
N X
K Fig. 1 clearly reveals that OMA significantly outperforms
X
R̃ < R̄ = φ E[log (1 + γnk )] , (20) NOMA in terms of the achievable sum rate in the regime of
n=1 k=1 low number of users. This observation is due to the fact that
where γnk is defined as (21) at the top of this page. the performance of NOMA is hindered by the intra-cluster
Remark 4: The proposed NOMA-aided massive MIMO is a pilot contamination due to the shared pilots within clusters
solution to user-overloading, where the number of data streams and error propagation due to imperfect SIC. For instance, the
is larger than the spatial degrees-of-freedom. The maximum residual interference resulted from imperfect SIC (ρnk = 0.1)

2162-2337 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LWC.2018.2841375, IEEE Wireless
Communications Letters
4

20 WCG-SR (Sim)
WCG-SR (Ana) - OMA
E-SR (Sim) - OMA
Var[cnk ] = E |cnk − E[cnk ]|2
 
WCG-SR (Ana) - NOMA - Perfect SIC
E-SR (Sim) - NOMA - Perfect SIC  " M # 2
WCG-SR (Ana) - NOMA - Imperfect SIC (ρ = 0.1) M ∗ ∗
Achievable Sum Rate [bits/s/Hz]

15
X v X v
E-SR (Sim) - NOMA - Imperfect SIC (ρ = 0.1)
= E hmnk mn − E hmnk mn 
m=1
|v mn |
m=1
|v mn |

M
" # M
∗ 2 2
v∗

X vmn X
= E hmnk − E hmnk mn . (23)
|vmn | |vmn |
10 m=1 m=1

By substituting (5) and (22) into (23), we have

NOMA M M
X X π
Var[cnk ] = ζmnk − ηmnk . (24)
4
5 OMA m=1 m=1
2
From (22) and (24), E[|cnk | ] can be computed as
E |cnk |2 = Var[cnk ] + |E[cnk ]|2
 

M √
M M
!2
X X π X πηmnk
0 = ζmnk − ηmnk + . (25)
0 100 200 300 400 m=1 m=1
4 m=1
2
Number of User Nodes 2
Fig. 1. The achievable sum rate versus the number of users. For NOMA, Next, E[|cn0 k | ] for n0 6= n can be computed

as
2 = 0 dBm, and 2
each cluster has two users. Here, β = 2, M = 100, σnk M
X v∗ 0 
Pn = 38 dBm, which is allocated to two users within a cluster based on a E |cn0 k |2 = E  hmnk mn
 
3:7 ratio for ∀n. In the legend, the acronyms WCG-SR and E-SR stand for |vmn0 |
the worst-case Gaussian sum rate and ergodic sum rate, respectively. m=1
" M ! M
!#
∗
degrades the achievable sum rate by 0.52 bits/s/Hz over that
X vmn0
X ∗ vm0 n0
=E hmnk hm0 nk
of the perfect SIC for 100 simultaneously served users. This |vmn0 | |vm0 n0 |
m=1 m0 =1
observation reveals that the sum rate loss of imperfect SIC
" M # M
X X
becomes significant when the estimate x̂nk is less correlated =E |hmnk |2 = ζmnk . (26)
m=1 m=1
with xnk for ∀n, k due to error propagation.
By using E[x∗nk00 x̂nk00 ]=E[xnk00 x̂∗nk00 ] = ρnk00 , PI4 is given by
K
V. C ONCLUSION X
Pnk00 E |cnk xnk00 − E[cnk ] x̂nk00 |2
 
PI4 =
A power-domain NOMA technique has been investigated k00 =k+1
for cell-free massive MIMO systems, and the achievable sum K
X
Pnk00 E |cnk |2 + (1 − 2ρnk00 ) (E[cnk ])2 . (27)
   
rate has been derived. The OMA counterpart outperforms =
NOMA in terms of the achievable sum rate in the regime k00 =k+1

of the low number of users because the latter suffers from R EFERENCES
intra-cluster pilot contamination and imperfect SIC. However, [1] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta,
“Cell-Free Massive MIMO Versus Small Cells,” IEEE Trans. Wireless
NOMA can serve twice the users than that of OMA when Commun., vol. 16, no. 3, pp. 1834–1850, Mar. 2017.
two users are grouped into each cluster, while achieving a [2] L. D. Nguyen, T. Q. Duong, H. Q. Ngo, and K. Tourki, “Energy
sum rate gain over OMA when the number of simultaneously Efficiency in Cell-Free Massive MIMO with Zero-Forcing Precoding
Design,” IEEE Commun. Lett., vol. 21, no. 8, pp. 1871–1874, Aug.
served users grows large due to the proposed pilot assignment. 2017.
Hence, the proposed NOMA-aided massive MIMO can sup- [3] Z. Ding, X. Lei, G. K. Karagiannidis, R. Schober, J. Yuan, and V. K.
port massive connectivity in next-generation wireless systems. Bhargava, “A Survey on Non-Orthogonal Multiple Access for 5G
Networks: Research Challenges and Future Trends,” IEEE J. Sel. Areas
A PPENDIX Commun., vol. 35, no. 10, pp. 2181–2195, Oct 2017.
[4] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and
K. Higuchi, “Non-Orthogonal Multiple Access (NOMA) for Cellular
In this appendix, derivations of E[cnk ], Var[cnk ] and Future Radio Access,” in In Proc. IEEE 77th Veh. Techno. Conf. (VTC
E |cnk |2 are outlined. By using the fact that εmnk = hmnk − Spring), Jun. 2013, pp. 1–5.
ĥmnk , E[cnk ] can be computed as [5] Q. Sun, S. Han, C. L. I, and Z. Pan, “On the Ergodic Capacity of
#" MIMO NOMA Systems,” IEEE Wireless Commun. Lett., vol. 4, no. 4,
 v∗ M 
mn
X pp. 405–408, Aug. 2015.
E[cnk ] = E ĥmnk + εmnk [6] Z. Ding, F. Adachi, and H. V. Poor, “The Application of MIMO to Non-
|vmn |
m=1 Orthogonal Multiple Access,” IEEE Trans. Wireless Commun., vol. 15,
M √ no. 1, pp. 537–552, Jan. 2016.
" M #
X  1/2  v∗ πηmnk
(a) (b)
mn
X
=E ηmnk vmn + εmnk = , (22) [7] M. Zeng, A. Yadav, O. A. Dobre, G. I. Tsiropoulos, and H. V. Poor,
m=1
|v mn |
m=1
2 “Capacity Comparison Between MIMO-NOMA and MIMO-OMA With
Multiple Users in a Cluster,” IEEE J. Sel. Areas Commun., vol. 35,
where the step (a) is based on (5). Moreover, vmn ∼ no. 10, pp. 2413–2424, Oct. 2017.
2 [8] Z. Ding and H. V. Poor, “Design of Massive-MIMO-NOMA With
RCN √ |vmn | ∼ exp(1). Thereby, E[|vmn |] =
(0, 1), and hence,
∞ √ −x Limited Feedback,” IEEE Signal Process. Lett., vol. 23, no. 5, pp. 629–
0
xe dx = π/2. By exploiting the fact that ĥmnk and 633, May 2016.
εmnk are independent, the step (b) is obtained. [9] T. L. M. A. Ashikhmin and L. Li, “Interference Reduction in Multi-
cell Massive MIMO Systems I: Large-Scale Fading Precoding and
By using the fact that hmnk and hm0 nk are independent Decoding,” Nov. 2014, arXiv:1411.4182[cs.IT].
when m 6= m0 for m, m0 ∈ {1, · · · , M }, n ∈ {1, · · · , N } and [10] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
k ∈ {1, · · · , K}, Var[cnk ] can be computed as follows: Theory. Englewood Cliffs, NJ: Prentice Hall, 1993.

2162-2337 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.