IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO.
5, MAY 2019 1
Massive MIMO with Spatially Correlated Rician
Fading Channels
Özgecan Özdogan, Student Member, IEEE, Emil Björnson, Senior Member, IEEE,
and Erik G. Larsson, Fellow, IEEE
Email: {ozgecan.ozdogan, emil.bjornson, erik.g.larsson}@liu.se
Abstract—This paper considers multi-cell Massive MIMO
(multiple-input multiple-output) systems where the channels
are spatially correlated Rician fading. The channel model
arXiv:1805.07972v2 [cs.IT] 13 Jun 2019
is composed of a deterministic line-of-sight (LoS) path and
a stochastic non-line-of-sight (NLoS) component describing a
practical spatially correlated multipath environment. We derive
the statistical properties of the minimum mean squared error
(MMSE), element-wise MMSE (EW-MMSE), and least-square
(LS) channel estimates for this model. Using these estimates
for maximum ratio (MR) combining and precoding, rigorous
closed-form uplink (UL) and downlink (DL) achievable spectral
efficiency (SE) expressions are derived and analyzed. The asymp-
totic SE behavior when using the different channel estimators are
also analyzed. Numerical results show that the SE is higher when
using the MMSE estimator than the other estimators, and the
performance gap increases with the number of antennas.
Index Terms—Massive MIMO, spatially correlated Rician
fading, channel estimation, spectral efficiency.
I. I NTRODUCTION
Massive MIMO (multiple-input multiple-output) is the key
technology for increasing the spectral efficiency (SE) in future
cellular networks, by virtue of beamforming and spatial mul-
tiplexing [1]. A Massive MIMO base station (BS) is equipped
with a massive number (e.g., a hundred) of individually
steerable antennas, which can be effectively used to serve tens
of user equipments (UEs) simultaneously on the same time-
frequency resource. The canonical form of Massive MIMO
operates in time-division duplex (TDD) mode and acquires
channel state information (CSI) by using uplink (UL) pilot
signaling and channel reciprocity [2]. The CSI is used for
coherent UL receive combining and downlink (DL) transmit
precoding.
The achievable SEs of Massive MIMO systems with imper-
fect CSI have been rigorously characterized and optimized for
fading channels modeled by either spatially uncorrelated [3] or
spatially correlated [2], [4] Rayleigh fading. Communication
with fading-free (line-of-sight) LoS propagation has also be
Manuscript received May 17, 2018; revised September 3, 2018 and Novem-
ber 14, 2018; accepted January 5, 2019. Date of publication January 21,
2019; date of current version May 15, 2019. This work was supported in
part by ELLIIT and in part by the Swedish Research Council. This paper
was presented at the IEEE SPAWC 2018. The associate editor coordinating
the review of this paper and approving it for publication was X. Chen.
(Corresponding author: Özgecan Özdogan)
The authors are with the Department of Electrical Engineering
(ISY), Linköping University, 581 83 Linköping, Sweden (e-mail: ozge-
can.ozdogan@liu.se; emil.bjornson@liu.se, erik.g.larsson@liu.se).
Color version of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
treated [2], [5]. However, practical channels can consist of
a combination of a deterministic LoS path and small-scale
fading caused by multipath propagation, which can be modeled
by the Rician fading model [6].
The performance of Massive MIMO with Rician fading
channels is much less analyzed than with Rayleigh fading. The
single-cell case was studied in [7]–[9] under the assumption
of spatially uncorrelated Rician fading channels and zero-
forcing (ZF) processing. Approximate SE expressions for the
UL and DL were provided in [7] and [8], [9], respectively. The
multi-cell case was studied in [10]–[12], assuming spatially
uncorrelated Rician fading within each cell and spatially
uncorrelated Rayleigh fading across cells. Approximate SE
expressions were derived in the UL with ZF combining [11]
and in the DL with ZF [10] or regularized ZF precoding [12].
Note that these are the prior works that consider imperfect CSI,
which is the practically relevant scenario, while prior works
assuming perfect CSI can be found in the reference lists of
[7]–[12].
A. Main Contributions
There are three major limitations of the prior works. First,
the fading was modeled as spatially uncorrelated, although
practical channels are correlated, due the finite number of
scattering clusters [2]. Second, the inter-cell channels were
modeled by Rayleigh fading, although it may happen that a
UE has LoS paths to multiple BSs (e.g., in parks, dense small-
cell deployments, or when serving unmanned aerial vehicles
(UAVs)). The existence of an LoS path depends on environ-
mental factors. In the simulation part, we use a probabilistic
approach based on the 3GPP model [13] to achieve a realistic
scenario. Third, only approximate SE expressions were derived
in closed form in prior works, which only provide insights
into special operational regimes, such as having asymptotically
many antennas. In this paper, we address these shortcomings:
• We consider a multi-cell scenario with spatially correlated
Rician fading channels between the pairs of BSs and UEs
that are determined statistically where other pairs have
spatially correlated Rayleigh fading channels. Previously,
this channel model has only been used for single-cell
scenarios with perfect CSI [14], [15].
• We derive the minimum mean squared error (MMSE),
element-wise MMSE (EW-MMSE) and least-square (LS)
channel estimators and characterize their statistics. Us-
ing these estimates for MR combining and precoding,
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 2

we compute rigorous closed-form UL and DL SEs and of the antennas at BS j to UE k in cell l is determined by the
discuss their structure. normalized trace as
• We analyze asymptotic behavior of UL and DL SEs under 1
j j
spatially correlated Rician fading when using the different βlk = tr(Rlk ). (2)
Mj
estimators.
j
• We compare the UL and DL SEs with MMSE, EW- where βlk is called the large-scale fading coefficient.
MMSE and LS estimation numerically, considering both
correlated and uncorrelated Rician and Rayleigh fading. III. C HANNEL E STIMATION
The conference version of this paper [16] only considered
Each BS requires CSI for receive processing. Therefore, τp
the UL and only used the MMSE and LS estimators.
samples are reserved for performing UL pilot-based channel
Reproducible research: All the simulation results can be
estimation in each coherence block, giving room for τp mu-
reproduced using the Matlab code and data files available at:
tually orthogonal pilot sequences. These pilot sequences are
https://github.com/emilbjornson/rician-massive-mimo
allocated to different UEs and the same sequences are reused
by UEs in multiple cells. The deterministic pilot sequence of
II. C HANNEL AND S YSTEM M ODEL UE k in cell j is denoted by φ jk ∈ Cτ p and kφ jk k 2 = τp . We
We consider a Massive MIMO system with L cells where define the set
each cell consists of one base station (BS) with M j antennas
P jk = (l, i) : φ li = φ jk , l = 1, . . . , L, i = 1, . . . , Kl ,

that serves K single-antenna user equipments (UEs). The (3)
system operates in TDD mode where the channel responses with indices of all UEs in the system that utilize the same
remain constant over a coherence block of τc samples. Also, pilot sequence as UE k in cell j (including the UE itself). The
we assume that the channel realizations are independent p
received pilot signal Y j ∈ C M j ×τ p at BS j is
between any pair of coherence blocks. The size of τc is
determined by the carrier frequency and external factors such Kj Kl
L Õ
p
Õ √ j
Õ √ j p
as the propagation environment and UE mobility [3]. The Yj = p jk h jk φTjk + pli hli φTli + N j , (4)
samples are used for three different tasks: τp samples for k=1 l=1 i=1
uplink (UL) pilot signals, τu samples for UL data transmission l,j

and τd samples for downlink (DL) data transmission where p

where N j ∈ C M j ×τ p has independent and identically dis-
τc = τp + τu + τd . Both UL and DL channels are estimated tributed NC (0, σul2 )-elements, with variance σul2 . To estimate the
by uplink pilot signals by exploiting channel reciprocity in the j p
channel hli , BS j multiplies Y j with the UE’s pilot sequence
TDD protocol.1
φ li to obtain
∗
The channel response between UE k in cell l and the BS
j
in cell j is denoted by hlk ∈ C M j . Each element of hlk
j p p √ j
Õ √
j p
y jli = Y j φ li∗ = pli τp hli + pl0 i0 τp hl0 i0 +N j φ li∗ . (5)
corresponds to the propagation channel from the UE to one (l0,i0 )∈ Pl i \(l,i)
j
of the BS’s M j antennas. The superscript of hlk indicates the p
BS index and the subscript identifies the index of the cell and The processed received pilot signal y jli ∈ C M j is a sufficient
j
the UE. While the channel responses are the same in the UL statistics for estimating hli [2]. We will now consider three dif-
and DL of a coherence block, for notational convenience, we ferent channel estimators, which rely on different amounts of
j j
use hlk for the UL channel and (hlk ) H for DL channel. statistical channel knowledge. The statistical distributions (the
In this paper, we consider spatially correlated Rician fading mean vector and covariance matrices) can be estimated using
j
channels. Each channel vector hlk , ∀ j, l ∈ 1, . . . , L and the sample mean and sample covariance matrices in practice
∀k ∈ 1, . . . , K, is modeled as a realization of the circularly [2], [17]–[20]. Note that a small change in the UE location
symmetric complex Gaussian distribution2 may result in a significant phase-shift of the LoS component.
j

j j
 More specifically, if the UE moves half a wavelength away
hlk ∼ NC h̄lk , Rlk , (1) from the BS, the phase of the channel response changes by
j 180◦ . This phase shift, however, will be identical for all BS
where the mean h̄lk ∈ C M j corresponds to the LoS component
j antennas, and may therefore be accurately tracked in practice.
and Rlk ∈ C M j ×M j is the positive semi-definite covariance
matrix describing the spatial correlation of the non-line-of-
sight (NLoS) components. The small-scale fading is described A. MMSE Channel Estimator
j j
by the Gaussian distribution whereas Rlk and h̄lk model the Based on the processed received pilot signal in (5), the
j
macroscopic propagation effects, including pathloss, shadow- BS can apply MMSE estimation to obtain an estimate of hli
fading, and the antenna gains and radiation patterns at the as shown in the following lemma. Notice that the Bayesian
transmitter and receiver. The average channel gain from one MMSE estimator requires that the statistical distributions are
1 We assume that the hardware is fully synchronized for reciprocity; see [2,
fully known.
Sec. 6] for a review of calibration techniques.
Lemma 1: The MMSE estimate of channel from BS j to
2 Strictly speaking, circularly symmetric random variable must have zero UE i in cell l is
mean, but we consider the common extension of this terminology to the case j j √ j j

p p

when it is sufficient that the non-zero-mean part is circularly symmetric. ĥli = h̄li + pli Rli Ψli y jli − ȳ jli , (6)
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 3

p √ j j

j

where ȳ jli = pl0 i0 τp h̄l0 i0 and h̃li ∼ NC 0 M , Σ̃li ,
Í
(l0,i0 )∈ Pli (14)
!−1   −1
n o −1 j j j j j j j j
where Σli = pli τp Dli Λli Ψli Λli Dli and Σ̃li = Rli −
Õ
j p j
Ψli = τp Cov y jli = pl0 i0 τp Rl0 i0 + σ 2 I M j . (7)
j j j j j j j
(l0,i0 )∈ Pli pli τp Rli Λli Dli − pli τp Dli Λli Rli + Σli .
j j
The estimation error h̃li = hli − ĥli has the covariance matrix
j Proof: We can easily get the desired result using the same
approach as was employed to derive the MMSE estimator,
j j j j j
Cli = Rli − pli τp Rli Ψli Rli (8) but estimating each element separately using only the signal
j j j
obtained at that antenna and then computing the resulting
and the mean-squared error is MSE = E{khli −ĥli k 2 } = tr(Cli ). statistics.
j j j j
The MMSE estimate ĥli and the estimation error h̃li are In contrast to MMSE estimation, ĥli and h̃li are correlated
independent random variables and distributed as
n o
j j j j j j
  with E ĥli (h̃li ) H = pli τp Dli Λli Rli − Σli except in the special
j j j j
ĥli ∼ NC h̄li, Rli − Cli , (9) case when all the covariance matrices are diagonal.
 
j j
h̃li ∼ NC 0 M , Cli . (10)
C. LS Channel Estimator
Proof: The proof follows from the standard MMSE j j
If the BS has no prior information regarding Rli and h̄li , the
estimation of Gaussian random variables that are observed in
non-Bayesian LS estimator can be utilized to get an estimate
Gaussian noise [2], [21]. j
j of the propagation channel hli . The LS estimate is defined as
Note that the estimation error covariance matrix Cli does j p √ j
not depend on the mean values. In other words, the estimation the value of ĥli that minimizes ky jli − pli τp ĥli k 2 , which in
error is not affected by the LoS components since these this case is
j 1 p
are known and can be subtracted from the received signals. ĥli = √ y . (15)
Moreover, the channel estimates of UEs in the set Pli are pli τp jli
not independent, despite the assumption that the channels Lemma 3: The LS estimator and estimation error are corre-
are independent. This is known as pilot contamination and lated random variables and distributed as
happens since the UEs use the same pilot sequence. UE  
( j, k) ∈ Pli has the channel estimate j 1 p 1 j
ĥli ∼ NC √ ȳ jli, (Ψli )−1 , (16)
√   pli τp pli τp
j j j j p p
ĥ jk = h̄ jk + p jk R jk Ψli y jli − ȳ jli (11) 
1 1

j j p j j
h̃li ∼ NC h̄li − √ ȳ jli, (Ψli )−1 − Rli . (17)
j p pli τp pli τp
and it is correlated with ĥli in (6) since y jli appears in both
j j
expressions and Ψli = Ψ jk . We will utilize the distributions Proof: The proof is given in Appendix C.
of the channel estimates and estimation errors in Sections IV This lemma shows that the statistics are more complicated
and V when analyzing the UL and DL SE. than when using the MMSE and EW-MMSE estimators.
For example, the estimation error has non-zero mean, which
B. Element-wise MMSE Channel Estimator needs to be accounted for when analyzing the communication
performance.
If the BS does not have knowledge of the entire covariance
matrices, the EW-MMSE estimator can be implemented as an
alternative [2], [22]. In this method, only the diagonals of the IV. U PLINK S PECTRAL EFFICIENCY WITH MR COMBINING
covariance matrices are needed and the correlation between the
During data transmission, the received signal y j ∈ C M j at
elements are ignored by the estimator. As a result, there are
BS j is
no matrix inversions and thus the computational complexity
is greatly reduced as compared to the MMSE estimator in Kj
Õ Kl
L Õ
Õ
j j
Lemma 1. yj = h jk s jk + hli sli + n j , (18)
Lemma 2: The EW-MMSE estimate of channel from BS j k=1 l=1 i=1
l,j
to UE i in cell l is
j j √ j j

p p
  
ĥli = h̄li + pli Dli Λli y jli − ȳ jli , (12) where n j ∼ NC 0 M j , σul2 I M j is additive noise. The UL signal
j j
from UE k in cell l is denoted by slk ∈ C and has power
where Dli ∈ C M j ×M j and
h Λ i li ∈ C
M j ×M j are diagonal matrices
 plk = E |slk | 2 . The first term in (18) is the desired signal
j j j
with Dli = diag Rli : m = 1, . . . , M j and Λli = and the latter terms denote interference and noise, respectively.
mm
 hÍ
j
i  −1 BS j selects the receive combining vector v jk ∈ C M j based
(l0,i0 )∈ Pl i pl0 i0 τp Rl0 i0 + σ I M j :m= 1, . . . , M j .
diag 2
mm on its CSI and multiplies it with y j to separate the desired
j
The distributions of EW-MMSE estimate ĥli and the signal from its UE k from interference. As in [2, Th. 4.4], the
j ergodic UL capacity of UE k in cell j is lower bounded by
estimation error h̃li are correlated and distributed as
  τu  
SEuljk = log2 1 + γ ul
j j j
ĥli ∼ NC h̄li, Σli , (13) τc jk [bit/s/Hz], (19)
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 4

with the effective SINR fading Massive MIMO systems. The signal terms in the
p jk |E{v H
j 2 numerator depend on the estimation quality and the LoS
jk h jk }|
γ ul
jk = , component. The  former is  reduced by pilot contamination,
L Õ Kl 2 τ tr R j Ψ j R j j j
Õ
j j since p p = p jk tr R jk − C jk , which is the
jk hli | } − p jk |E{v jk h jk }| + σul E{kv jk k }
2 2 2 2
pli E{|v H H jk jk jk jk
l=1 i=1
transmit power multiplied with the trace of the covariance
(20) matrix of the channel estimate in (9).
where the expectations are with respect to all sources of In the denominator, the relation between the covariance ma-
j j
randomness. Since SEul jk is below the capacity, it is an ergodic trices R jk and Rli , and the inner product of LoS components
achievable SE. The effective SINR γ ul jk
can be computed h̄ jjk and h̄lij determine how large the interference terms are. If
numerically for any combining scheme and channel estimator. the covariance matrices span different subspaces, or one has
We will show that it can be computed in closed form when very small eigenvalues (e.g., due to weak large-scale fading),
using MR combining, based on each of the three channel there will be little interference from the NLoS propagation.
estimators derived in Section III. Similarly, there is little interference from the LoS propagation
j j
when h̄ jk and h̄li are nearly orthogonal.
A. Uplink Spectral Efficiency with MMSE estimator The non-coherent interference term ξliul does not increase
j j
If the MMSE estimator in (6) is used, we obtain a closed- with M j , unless h̄ jk and h̄li are nearly parallel vectors. The
form expression for the SE in (19) as in the next theorem. coherent interference term Γliul involves the pilot-contaminating
j
Theorem 1: If MR combining with v jk = ĥ jk is used based UEs, which are (l, i) ∈ P jk \( j, k), and it grows linearly with
on the MMSE estimator, then M j . The term ν uljk
grows with M j and depends on the norm of
n
j
o 
j j j

j 2 desired UE’s LoS component.
E vH jk h jk = p jk τp tr R jk Ψ jk R jk + k h̄ jk k , (21)
 
(22) B. Uplink Spectral Efficiency with EW-MMSE Estimator
j j j j
E kv jk k 2 = p jk τp tr R jk Ψ jk R jk + k h̄ jk k 2,

If the EW-MMSE estimator in (12) is used, we obtain
j 2
and E{|v H h
jk li | } is given in (23), at the bottom of this page. a closed-form expression for the SE in (19) as in the next
Plugging these expressions into the SINR in (20) yields theorem.
j

j j j

j 2 Theorem 2: If MR combining with v jk = ĥ jk is used based
p jk τp tr R jk Ψ jk R jk + p jk k h̄ jk k
2
on the EW-MMSE estimator, then
γ ul,mmse
jk
= ,
L Õ Kl
n o  
j j j j j
Õ Õ E vH h jk = p jk τp tr D jk Λ jk D jk + k h̄ jk k 2, (27)
pli ξli +
ul
pli Γli − p jk ν jk + σul
ul ul 2 jk
l=1 i=1 (l,i)∈ P j k \(j,k)

j

j
E kv jk k 2 = tr Σ jk + k h̄ jk k 2,

(24) (28)
j
k h̄ j k k 4
where ν ul = , ξliul and Γliul correspond and E{|v H
j 2
jk hli | } is given in (29), at the top of next page.
 
jk p j k τ p tr
j j j j
R j k Ψ j k R j k + k h̄ j k k 2
to LoS-related interference, non-coherent interference, and Plugging these expressions into the SINR in (20) gives γ ul,ew
jk
coherent interference, respectively. The latter two are given in (30).
at the bottom of this page. Proof: The proof is given in Appendix E.
Proof: The proof is given in Appendix D. This SINR expression in Theorem 2 is more complicated
The rigorous closed-from SINR expression in (24) provides than when using the MMSE estimator, but can be interpreted
important and exact insights into the behaviors of Rician in an analogous way.

  
 H j 2
   2
j j j j j j j j j j j j  j j
E v jk hli  = p jk τp tr Rli R jk Ψ jk R jk + p jk τp (h̄li ) H R jk Ψ jk R jk h̄li + (h̄ jk ) H Rli h̄ jk + (h̄ jk ) H h̄li 
 2
 p jk pli τ 2 tr R j Ψ j R j  + 2√ p jk pli τp Re tr R j Ψ j R j (h̄ j ) H h̄ j
   n   o
(l, i) ∈ P jk

+

p li jk jk li jk jk li jk (23)
0
 (l, i) < P jk .


   H  H  2
j j j j j j j j j j j j  j j
p jk τp tr Rli R jk Ψ jk R jk + p jk τp h̄li R jk Ψ jk R jk h̄li + h̄ jk Rli h̄ jk + (h̄ jk ) H h̄li 
ξliul =   , (25)
j j j j
p jk τp tr R jk Ψ jk R jk + k h̄ jk k 2
 2     H 
√
 
j j j  j j j j j
p jk pli τp2 tr Rli Ψ jk R jk  + 2 p jk pli τp Re tr Rli Ψ jk R jk h̄li h̄ jk

Γliul =   . (26)
j j j j
p jk τp tr R jk Ψ jk R jk + k h̄ jk k 2
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 5

  
 H j 2
 
j j j j j j j j j j
E v jk hli  = χliul = tr Rli Σ jk + (h̄ jk ) H Rli h̄ jk + (h̄li ) H Σ jk h̄li + |(h̄ jk ) H h̄li | 2
2
√
    n o
 p jk pli τ 2 tr D j Λ j D j + τ
j j j j H j

2 p p tr D Λ D Re (h̄ ) h̄ (l, i) ∈ P jk

jk li p
+

p li jk jk li jk jk jk li (29)
0
 (l, i) < P jk .


   2
j j j j
p jk p jk τp tr D jk Λ jk D jk + k h̄ jk k 2
γ ul,ew
jk
= . (30)
Õ Kl
L Õ    2    
j j j j j j
pli χliul − p jk p jk τp tr D jk Λ jk D jk + k h̄ jk k 2 + σul2 tr Σ jk + k h̄ jk k 2
l=1 i=1

j
C. Uplink Spectral Efficiency with LS Estimator ignores the NLoS paths. If MR combining with v jk = h̄ jk is
used based on such a Mean Only (MO) estimator, then
If the LS estimator in (15) is used, we obtain a closed-form n o n o
j j
jk h jk = E kv jk k = k h̄ jk k 2,
H 2
expression for the SE in (19) as in the next theorem. E vH (35)
p
Theorem 3: If MR combining with v jk = √ p j1k τ p y j jk is used n o
j 2 j j j j j
based on the LS estimator, then E |v H
jk hli | = (h̄ jk ) H Rli h̄ jk + |(h̄ jk ) H h̄li | 2 . (36)
Õ √ pli
Plugging these expressions into the SINR in (20) gives γ ul,mo
n o
j j j H j
E vH h
jk jk = tr(R jk
) + √ (h̄li ) h̄ jk , (31) jk
p jk as
(l,i)∈ P j k
j
p jk k h̄ jk k 2
n o 1 
j 1 
p .
E kv H
jk k
2
= tr (Ψ jk )−1 + k ȳ k 2, (32) Õ Kl
L Õ
pli
p jk τp
 
p jk τp2 j jk j j j j j j
(h̄ jk ) H Rli h̄ jk + |(h̄ jk ) H h̄li | 2 − p jk k h̄ jk k 2 + σul2
j
l=1 i=1 k h̄ jk k 2
j 2
and E{|v H jk hli | } is given in (33), at the top of next page where
(37)
p √ j j j j Note that if UE ( j, k) does not have an LoS component then
x̄ jk = ȳ j jk − pli τp h̄li and (Ω jk )−1 = (Ψ jk )−1 − pli τp Rli .
we have no information regarding its channel and the SINR
Plugging these into (20) gives γ ul,ls in (34) where χliul,ls is
jk in (37) becomes zero. Thus, the MO estimator is only useful
defined in (33).
for UEs that have an LoS path.
Proof: The proof is given in Appendix F.
Note that the Rayleigh fading counterpart of (34) can be easily V. D OWNLINK S PECTRAL E FFICIENCY WITH MR
obtained by setting all the mean vectors to zero. In this case, P RECODING
the difference in SE between the MMSE and LS/EW-MMSE
Each coherence block contains τd DL data transmissions,
estimators can be rather small [2]. However, the loss in SE
where the transmitted signal from BS l is
incurred by using the LS estimator under Rician fading can
Kl
be quite large depending on the dominance of LoS paths. Õ
Since the mean values are not utilized as prior information, xl = wlk ςlk , (38)
k=1
the interference terms are larger than when using the MMSE
estimator. The LS estimates of the pilot-contaminating UEs are where ςlk ∼ NC (0, ρlk ) is the DL data signal intended for
equal up to a scaling factor. Compared to the SE with MMSE UE k in the cell and ρlk is the signal power. The transmit
p j precoding vector wlk determines the spatial directivity
estimator, the inner product of ȳ j jk and h̄li determines how  of the
2 = 1,
large the corresponding interference terms are instead of the transmission. The precoding vector satisfies E kwlk k
such that E kwlk ςlk k 2 = ρlk is the transmit power allocated
j j

inner product of h̄ jk and h̄li .
to this UE. The received signal y jk ∈ C at UE k in cell j is
Kj
D. Uplink Spectral Efficiency with Mean Only Estimator
Õ
j
y jk = (hljk ) H w jk ς jk + (h jk ) H w ji ς ji
If the BS only knows the mean values of the UEs’ channels i=1
i,k
but not the covariance matrices, we can use this information as Kl
L Õ
a channel estimate, without the need for sending pilots. The
Õ
+ (hljk ) H wli ςli + n jk , (39)
mean values vary much slower compared to the small-scale l=1 i=1
fading coefficients. Thus, using the MO estimator may greatly l,j
reduce the computational complexity since the channels are  
not estimated for each coherence block. However, it is only where n jk ∼ NC 0, σdl2 is i.i.d. additive receiver noise with
meaningful when the UEs have dominant LoS paths since it variance σdl2 . The first term in the above equation denotes the
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 6

n o  
j 2 j j p j
p jk τp2 E |v H
jk hli | = p jk τp2 χliul,ls = τp tr Rli (Ψ jk )−1 + |(ȳ j jk ) H h̄li | 2
p j p j j j



 (ȳ j jk ) H Rli ȳ j jk + τp (h̄li ) H (Ψ jk )−1 h̄li (l, i) < P jk
τ 2 |tr(R j )| 2 + x̄ H R j x̄ + τ (h̄ j ) H (Ω j )−1 h̄ j
+ li p


p li n jk li jk  p li jk li (33)
 √ p H j j

p H j j j j
o
 +2 pli τp Re (ȳ j jk ) h̄li tr Rli + (ȳ j jk ) Rli h̄li + x̄ H h̄li k h̄li k 2 (l, i) ∈ P jk .


 jk

 √ 2
j pl i j j 
p jk tr(R jk ) + (l,i)∈ P j k (h̄li ) H h̄ jk 
 Í
√
pjk
γ ul,ls
jk
=  2   , (34)
Kl
L Õ  √  tr (Ψ
j −1
) k ȳ
p
k 2
Õ
ul,ls j
Õ pli j H j jk j jk
pli χli − p jk tr(R jk ) + √ (h̄li ) h̄ jk  + σul ­­ +
  2
© ª
®
l=1 i=1
p jk p τ
jk p p jk τp2 ®
 (l,i)∈ P j k 
« ¬

    H
 H l 2
   2
j
E wli h jk  = pli τp tr Rljk Rlli Ψlli Rli + pli τp h̄ljk Rlli Ψlli Rlli h̄ljk + (h̄ljk ) H h̄lli 
k ĥlli k 2

(43)
 
E
 2     H 
l Ψl Rl  + 2√ p p τ Re tr Rl Ψl Rl
 
τ 2 l l

 H  p p
jk li p tr R jk li p h̄ h̄ (l, i) ∈ P jk
+ h̄lli Rljk h̄lli +


jk li li  jk li li li jk
0 (l, i) < P jk ,



desired signal, the second term is the intra-cell interference, the DL SINR in (41) gives
and the third term is the inter-cell interference. The ergodic
DL capacity of UE k in cell j is lower bounded by [2, Th. 4.6]
 
j j j j
ρ jk p jk τp tr R jk Ψ jk R jk + ρ jk k h̄ jk k 2
τd   γ dl,mmse
jk
= ,
Kl
L Õ
SEdl = log2 1 + γ dl [bit/s/Hz] (40) Õ Õ
jk
τc jk ρli ξlidl + ρli Γlidl − ρ jk ν dl
jk + σdl
2

l=1 i=1 (l,i)∈ P j k \(j,k)

with  (44)
  
 n o 2 where ν dl
jk
= k h̄ljk k 4 p jk τp tr Rljk Ψljk Rljk + k h̄ljk k 2 , ξlidl
j 
ρ jk E w H h and Γlidl correspond to non-coherent interference in (45), co-

jk jk 
γ dl
jk = , herent interference in (46) respectively.
Kl
L Õ    o 2
 H l 2
Õ  n
H j 
ρli E wli h jk  − ρ jk E w jk h jk  + σdl
2 Proof: The proof is similar to the uplink case in Appendix

l=1 i=1 D and is omitted.
(41) This SINR expression resembles the UL counterpart in
where the expectations are with respect to all sources of Theorem 1 to a large extent. The transmit power and noise
randomness. In the following subsections, the effective SINR variance are denoted differently and some of the indices are
γ dl
jk
is computed for MR precoding when using the different switched in the interference terms, as expected from the UL-
channel estimators. DL duality [2, Sec. 4.3]. Hence, the interpretation of the SINR
is qualitatively the same as in the uplink.

A. Downlink Spectral Efficiency with MMSE Estimator

If the MMSE estimator in (6) is used, we obtain a closed- B. Downlink Spectral Efficiency with EW-MMSE Estimator
form expression for the DL SE in (40) as in the next theorem.
ĥ
j If the EW-MMSE estimator in (12) is used, we obtain a
Theorem 4: If MR precoding with w jk = r n jk o is used closed-form expression for the DL SE in (40) as in the next
j
E k ĥ j k k 2
theorem.
based on the MMSE estimator, then ĥ
j

Theorem 5: If MR precoding with w jk = r n jk o is used

j
n o r   E k ĥ j k k 2
j j j j j
jk h jk =
E wH p jk τp tr R jk Ψ jk R jk + k h̄ jk k 2, (42) based on the EW-MMSE estimator, then
 
j j j j
p jk τp tr D jk Λ jk D jk + k h̄ jk k 2
  
 H l 2 n o
and E wli h jk  is given in (43), at the top of this page where E wH
j
= , (47)
jk h jk r  
j j 2
tr Σ jk + k h̄ jk k
 
E k ĥlli k 2 = pli τp tr Rlli Ψlli Rlli + k h̄lli k 2 . Inserting these into

 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 7

   H  H  2
pli τp tr Rljk Rlli Ψlli Rlli + pli τp h̄ljk Rlli Ψlli Rlli h̄ljk + h̄lli Rljk h̄lli + (h̄ljk ) H h̄lli 
 
ξli =
dl
  , (45)
pli τp tr Rlli Ψlli Rlli + || h̄lli || 2
 2     H 
√
 
p jk pli τp tr R jk Ψli Rli  + 2 p jk pli τp Re tr R jk Ψli Rli h̄li h̄ jk
2 l l l  l l l l l

Γlidl =   . (46)
pli τp tr Rlli Ψlli Rlli + || h̄lli || 2

  
 H l 2
 
k ĥlli k 2 E wli h jk  = χlidl = tr Rljk Σlli + (h̄lli ) H Rljk h̄lli + (h̄ljk ) H Σlli h̄ljk + |(h̄lli ) H h̄ljk | 2

E
2
√
    n o
 p jk pli τ 2 tr Dl Λl Dl + 2 p jk pli τp tr Dljk Λlli Dlli Re (h̄lli ) H h̄ljk

(l, i) ∈ P jk

+

p jk li li (48)
0
 (l, i) < P jk .

   2 .    
j j j j j j
ρ jk p jk τp tr D jk Λ jk D jk + k h̄ jk tr Σ jk + k h̄ jk k 2
γ dl,ew
jk
=    2 . (49)
j j j j 2
L
ÕÕ K l
χ dl p jk τp tr D jk Λ jk D jk + k h̄ jk k
ρli   li − ρ jk   + σdl2
j j 2
l=1 i=1 tr Σli + k h̄li k
l l 2 tr Σ jk + k h̄ jk k

    !
 H l 2  l H l 2
n o  H  
and E wli h jk  is given in (48), at the top of next page E |wliH hljk | 2 = h̄lli Rljk h̄lli +  h̄li h̄ jk  k h̄lli k 2 .
 
where E k ĥlli k 2 = tr Σlli + k h̄lli k 2 . Inserting these expres-

(54)
sions into the DL SINR in (41) gives in (49). γ dl,ew Inserting these expressions into the DL SINR in (41) gives
jk
Proof: The proof is similar to the uplink case in Appendix γ dl,mo
jk
as
E and is omitted. j
ρ jk k h̄ jk k 2
.
Kl
L Õ
C. Downlink Spectral Efficiency with LS Estimator Õ ρli  l H l l 
j
( h̄li ) R h̄
jk li + |(h̄l H l 2
li ) h̄ jk | − p jk k h̄ jk k 2 + σdl2
If the LS estimator in (15) is used, we obtain a closed-form
l=1 i=1
k h̄lli k 2
expression for the DL SE in (40) as in the next theorem. (55)
j
ĥ j k
Theorem 6: If MR precoding with w jk = r n o is used
j
E k ĥ j k k 2
VI. A SYMPTOTIC A NALYSIS
based on the LS estimator, then
n o √ In this section, we will analyze the asymptotic behavior
j j p j j
E vH tr(R jk ) + (l,i)∈ P j k √ p jlki (h̄li ) H h̄ jk of Rician fading channels when using MR based on the
Í
n o jk h jk
j
E wH jk h jk = q  =r     , different channel estimators. We make the following technical
E kv jk k 2 1 j −1
τp tr (Ψ jk )
p
+ k ȳ j jk k 2 assumptions:
p j k τ p2
(50) Assumption 1: For l, j = 1, . . . , L and i = 1, . . . , Kl , the
j j
spatial covariance matrix Rli satisfies lim sup kRli k2 < ∞ and
 2 
and E wliH hljk  is given in (51), at the top of next page
 
Mj
 
j
  
p 2
 lim inf M1 j tr Rli > 0.
where E k ĥli k = p τ 2 τp tr (Ψli )
1 l −1
+ k ȳlli k . Inserting
 l 2
Mj
li p
Assumption 2: For l, j = 1, . . . , L and i = 1, . . . , Kl , the LoS
these expressions into the DL SINR in (41) gives γ dl,ls in (52) j j
jk component h̄li satisfies lim sup M1 j k h̄li k 2 < ∞.
where the interference term χlidl,ls is defined in (51). Mj
Proof: The proof is similar to the uplink case in Appendix Assumption 3: For l, j = 1, . . . , L, i = 1, . . . , Kl and
j j
F and is omitted. k = 1, . . . , K j , the LoS components h̄ jk and h̄li satisfy
 
 j j
lim M1 j (h̄ jk ) H h̄li  → 0 if ( j, k) , (l, i).
Mj
D. Downlink Spectral Efficiency with Mean Only Estimator The first assumption is standard in the asymptotic analysis
j
h̄ for Massive MIMO [4] and implies that array gathers an
If MR precoding with w jk = jk
j is used based on the MO
h̄ j k amount of signal energy that is proportional to the number
estimator, then n o of antennas and this energy originates from many spatial di-
j j
jk jk = k h̄ jk k,
E wH h (53) rections. The other assumptions are discussed in Section VI-E.
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 8

n o  
p
E k ĥlli k 2 E |wliH hljk | 2 = pli τp2 χlidl,ls = τp tr Rljk (Ψlli )−1 + |(ȳ jli ) H h̄ljk | 2


p p



 (ȳ jli ) H Rljk ȳ jli + τp (h̄ljk ) H (Ψlli )−1 h̄ljk (l, i) < P jk
  2
+ p jk τp2 tr(Rljk ) + x̄liH Rljk x̄li + τp (h̄ljk ) H (Ωlli )−1 h̄ljk


(51)
 
√
 n   o
p p
 +2 p jk τp Re (ȳ jli ) H h̄ljk tr Rljk + (ȳ jli ) H Rljk h̄ljk + x̄liH h̄ljk k h̄ljk k 2 (l, i) ∈ P jk .





 √ 2     
j pl i j j  j p
ρ jk p jk τp2 tr(R jk ) + (l,i)∈ P j k (h̄li ) H h̄ jk  τp tr (Ψ jk )−1 + k ȳ j jk k 2
 Í
√
pjk
γ dl,ls
jk
=  √ 2 , (52)
j p j j 
p jk τp2 tr(R jk ) + (l,i)∈ P j k √ p jlki (h̄li ) H h̄ jk 
Í
Kl
L Õ
pli τp2 χlidl,ls

Õ
ρli   − ρ jk   + σdl2
p 2 j −1 p
l=1 i=1 τp tr (Ψli )
l −1
+ k ȳlli k τp tr (Ψ jk ) + k ȳ j jk k 2

We recall the definition of spatially orthogonal matrices from as M1 = · · · = ML → ∞, it follows that

[2].
Definition 1: For l = 1, . . . , L and i = 1, . . . , Kl , two spatial
 
j j j j
j j ρ jk p jk τp tr R jk Ψ jk R jk + ρ jk k h̄ jk k 2
covariance matrices Rli and R jk dl,mmse
γ jk −    2 → 0.
are asymptotically spatially orthogonal if p jk pli τp2 tr Rljk Ψlli Rlli 
 
Õ
ρli  
1  j j  (l,i)∈ P j k \(j,k) pli τp tr Rlli Ψlli Rlli + k h̄lli k 2
tr Rli R jk → 0 as M j → ∞. (56) (58)
Mj
Proof: The proof is given in Appendix G.
Definition 1 implies that both covariance matrices are These theorems show that the SINRs are generally upper
strongly rank-deficient. In spatially correlated fading, it holds bounded by simplified asymptotic SINR expressions, which
if the BS is equipped with a ULA and the channels from depend on the covariance matrices and mean values. It is
two UEs have non-overlapping supports of their multipath only in the special case of asymptotically spatially orthogonal
angular distributions [23, Theorem 1]. However, such angular matrices that the SE grows without limit with rate log2 (M),
separations are unlikely to occur in practice, at least in sub-6 which is consistent with the results for correlated Rayleigh
GHz cellular networks. fading in [2], [23].

A. Asymptotic Analysis of Spectral Efficiency with MMSE B. Asymptotic Analysis of Spectral Efficiency with EW-MMSE
Estimator Estimator

Theorem 7: Under Assumptions 1–3, it follows that γ ul,mmse Theorem 9: Under Assumptions 1–3, it follows that γ ul,ew
jk
jk
j j
grows without bound as M j → ∞ if R jk is asymptotically grows without bound as M j → ∞ if D jk is asymptotically
j j
spatially orthogonal to Rli for all (l, i) ∈ P jk \( j, k). If this spatially orthogonal to Dli for all (l, i) ∈ P jk \( j, k). If this
is not the case, then under Assumption 1–3, as M j → ∞, it is not the case, then under Assumption 1–3, as M j → ∞, it
follows that follows that
 
j j j j j j j j
p2jk τp tr(R jk Ψ jk R jk ) + p jk k h̄ jk k 2 p2jk τp tr D jk Λ jk D jk + p jk k h̄ jk k 2
γ ul,mmse
jk
−    2 → 0. ul,ew
γ jk −     2 → 0.
j j j 
p jk τp2 tr Rli Ψ jk R jk  τ 2 tr D j Λ j D j

Õ Õ p p
jk li p li jk jk
pli2   pli  
j j j j
(l,i)∈ P j k \(j,k) p jk τp tr R jk Ψ jk R jk + k h̄ jk k 2 (l,i)∈ P j k \(j,k)
j j j j
p jk τp tr D jk Λ jk D jk + k h̄ jk k 2
(57) (59)
Proof: The proof is given in Appendix G. Proof: The proof is given in Appendix H.
Theorem 8: Under Assumptions 1–3, it follows that γ dl,mmse jk Theorem 10: Under Assumptions 1–3, it follows that γ dl,ew jk
grows without bound as M1 = · · · = ML → ∞ if Rljk grows without bound as M1 = · · · = ML → ∞ if Dlli
is asymptotically spatially orthogonal to Rlli for all (l, i) ∈ is asymptotically spatially orthogonal to Dljk for all (l, i) ∈
P jk \( j, k). If this is not the case, then under Assumption 1–3, P jk \( j, k). If this is not the case, then under Assumption 1–3,
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 9

as M1 = · · · = ML → ∞, it follows that E. Discussion on the Asymptotic Behaviors of Rician Fading


ρ j k p j k τ p tr

j j j
Djk Λjk Djk

j
+ k h̄ j k k 2
2 Channels
 
j j
tr Σ j k + k h̄ j k k 2 To describe the intuition behind Assumptions 2-3 and Defi-
γ dl,ew
jk
−     2 → 0. (60) nition 1, we consider a uniform linear array (ULA) with omni-
Õ p jk pli τp2 tr Dljk Λlli Dlli directional antennas, where the LoS component is modeled as
ρli   [2, Sec. 1.3]
(l,i)∈ P j k \(j,k) tr Σlli + k h̄lli k 2
q h j j
iT
Proof: The proof is given in Appendix H. j j,LoS
h̄li = βli 1 e j2πd H sin(ϕl i ) . . . e j2πd H (M−1) sin(ϕl i ),
The implications from these theorems are similar to the (63)
MMSE estimation case, except that it is the diagonals of the where βli
j,LoS
is the large-scale fading coefficient, dH ≤ 0.5 is
covariance matrices that need to be asymptotically spatially the antenna spacing parameter (in fractions of the wavelength),
orthogonal to achieve an asymptotically unbounded SE. This j
and ϕli is the angle of arrival (AoA) to the UE seen from the
is a more restrictive condition. j j,LoS
BS. Utilizing this model, we have M1 j k h̄li k 2 = βli , which
C. Asymptotic Analysis of Spectral Efficiency with LS Estima- is a finite value for any M j . Hence, Assumption 2 holds.
tor The magnitude of the inner product of the LoS components
of two different UEs is
Theorem 11: Under Assumptions 1–3, as M j → ∞, it
follows that 
j −1 
 MÕ

  q j j  m 
 j H j j j 2πd H (sin(ϕ j k )−sin(ϕl i ))
(h̄ jk ) h̄li  = β jk βli 

  e 
j j
p2jk τp2 tr(R jk ) + k h̄ jk k 2

 m=0 
γ ul,ls
jk
− 2 → 0. q
j j j j
= β jk βli g1 (ϕ jk , ϕli ),

2 τ 2 tr(R j )+ k h̄ j k 2
pli (64)
Í p li li
(l,i)∈ P j k \(j,k) √ p j k  j 
j j 
tr(R j k )+ (l, i)∈P j k √ p (h̄li ) H h̄ j k 
Í
li
(61) where
Proof: The proof is given in Appendix I. 
j j

πd H M j (sin(ϕ j k )−sin(ϕl i ))
Theorem 12: Under Assumptions 1–3, as M1 = · · · = ML → sin

 j j
j j
 if sin(ϕ jk ) , sin(ϕli ),
g1 (ϕ jk , ϕli ) =

 
j j

∞, it follows that sin πd H (sin(ϕ j k )−sin(ϕl i ))
j j
if sin(ϕ jk ) = sin(ϕli ).
2 

j j
ρ j k p j k τ p2 tr(R j k )+ k h̄ j k k 2
 Mj



j

p (65)
τ p tr (Ψ j k )−1 + k ȳ j j k k 2 j j
γ dl,ls
jk
−  2 → 0. (62) Hence, if the LoS components h̄ jk and h̄li do not have
ρli p j k τ p2 tr(Rlj k )+ k h̄lj k k 2 j j
Í same AoA (more precisely, if sin(ϕ jk ) , sin(ϕli )) then,
(l,i)∈ P j k \(j,k) τ tr((Ψ l )−1 )+ k ȳ k 2
p  
p  j j
li lli
lim M1 j (h̄ jk ) H h̄li  → 0 holds [2, Sec. 1.3]. This implies that
Proof: The proof is given in Appendix I. Mj
In contrast to the case with the MMSE and EW-MMSE Assumption 3 holds, except in the extreme case when AoAs
estimators, we notice that the SE does not grow without bound are exactly the same.
when using the LS estimator.

D. Asymptotic Analysis of Spectral Efficiency with Mean Only VII. N UMERICAL R ESULTS
Estimator
Under Assumptions 1-3, if UE ( j, k) has an LoS path, it In this section, the closed-form SE expressions derived in
follows that γ ul,mo
jk
grows without bound as M j → ∞. In the the previous sections are validated and evaluated by simulat-
DL, under Assumptions 1-3, if UE ( j, k) has an LoS path, it ing a Massive MIMO cellular network. We have a 16-cell
follows that γ ul,mo
jk
grows without bound as M1, . . . , ML → ∞. setup where each cell covers a square of 250 × 250 m. The
 j  H j 2
 
network has a wrap-around topology. This layout is selected
pli  h̄ j k h̄l i 
Proof: In the UL, the LoS related term in (37) to guarantee that all BSs receive equally much interference
from all directions. There are K = 10 UEs per cell and these
j
k h̄ j k k 2
goes to zero as M j → ∞ for (l, i) , ( j, k) due to Assumption are uniformly and independently distributed in each cell, at
j H j j
h̄ j k Rli h̄ j k j distances larger than 35 m from the BS. Each UE is assigned
2 and 3. Noting the j ≤ kRli k2 is finite as M j →
k h̄ j k k 2 to the BS that provides largest channel gain considering all the
∞ due to Assumption 1 completes the proof. In the DL, the combinations of the UE and BSs in the system. The location
expressions for γ dl,mo
jk
contains the same matrix expressions as of each UE is used when computing the large-scale fading and
ul,mo
γ jk , except that indices (l, i) and ( j, k) are swapped in the nominal angle between the UE and BSs.
interference terms. Each BS is equipped with a ULA with half-wavelength
Note that we assume that the mean vectors are perfectly antenna spacing [2, Sec. 1]. Thus, the LoS component from UE
known, however small phase-shifts may effect the performance i in cell l to BS j is given by (63). For the covariance matrices,
severely as discussed in Section III. we consider N = 6 scattering clusters and the covariance
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 10

matrix of each cluster is modeled by the (approximate3 ) 22

Gaussian local scattering model [2, Sec. 2.6], such that 20

j,NLoS Õ
βli N 2 18
σϕ
 2
j
j π(s−m) cos(ϕl i, n )
e π(s−m) sin(ϕli, n ) e
h i
j −
= 2
,

Average sum SE [bit/s/Hz/cell]

Rli 16
s,m N n=1 14
(66)
j,NLoS j
where βli is the large-scale fading coefficient and ∼ ϕli,n 12

j j
U[ϕli − 40◦, ϕli + 40 ] is the nominal AoA for the n
◦ 10

cluster. The multipath components of a cluster have Gaussian 8

distributed AoAs, distributed around the nominal AoA with 6

the angular standard deviation (ASD) σϕ = 5◦ . We consider 4

MMSE with Rician fading
EW-MMSE with Rician fading
communication over a 20 MHz channel and the total receiver LS with Rician fading
MMSE with Rayleigh fading
2
noise power is −94 dBm. Each coherence block consists of EW-MMSE/LS with Rayleigh fading

τc = 200 samples and the same τp = 10 pilots are allocated 0

10 20 30 40 50 60 70 80 90 100

in each cell. The pilots are randomly assigned to the UEs in Number of antennas (M)

every cell in the sense that the kth UE in two cells, that belong
to the same pilot group, uses the same pilot. Fig. 1: Average UL sum SE for K = 10 as a function of the
Based on the 3GPP model in [13], the existence of an LoS number of BS antennas for different channel estimators.
path depends on the distance. The probability of LoS for the
channel between UE (l, i) and BS j is
( j
300−dli j Fig. 1 shows the sum UL SE averaged over different UE
, 0 < dli < 300 m,
Pr(LoS) = 300
j
(67) locations and shadow fading realizations, when using MR
0, dli > 300 m.
combining based on either the MMSE, LS or EW-MMSE
If the LoS path exists then the corresponding large-scale fading estimators. As a reference, we also provide curves for Rayleigh
coefficient is modeled (in dB) as fading with the same covariance matrices, representing the
j
 
j j
case when all the LoS components are blocked but the small-
βli = −30.18 − 26 log10 dli + Fli, (68) j
scale fading remains (i.e., the average channel gain E{khli k 2 }
j is smaller, as it would be the case in practice). The curves
where Fli ∼ N (0, σsf2 ) is the shadow fading with σsf = 4. The are generated using the closed-form expressions from Sec-
j j
Rician factor is calculated as κli = 13 − 0.03dli [dB] and is tion IV and the “” markers are generated by Monte Carlo
used to compute the large-scale fading parametersr for the LoS simulations. The fact that the markers overlap with the curves
j
j,LoS κ j confirms the validity of our analytical results. As expected,
and NLoS paths in (63) and (66) as βli = j
li
β and
κl i +1 li the highest UL SE is obtained when the MMSE estimator is
r
j,NLoS j
βli = j
1
βli (in linear scale). If the LoS path does not employed, since the LoS component and spatial correlation
κl i +1
are known and utilized. For Rician fading, the performance
exist then the large-scale fading parameter is modeled (in dB) of MMSE and EW-MMSE are very close and EW-MMSE
as   performs better than LS since it utilizes knowledge of the
j j j
βli = −34.53 − 38 log10 dli + Fli, (69) channels’ mean values. In the case of Rayleigh fading, MMSE
j estimation is still the better choice, while the sum SE with LS
where Fli ∼ N (0, σsf2 ) is the shadow fading with σsf = 10. In and EW-MMSE are identical since the estimates are the same
j,LoS j,NLoS j
this case, βli = 0 and βli = βli (in linear scale). up to a scaling factor.
We apply the heuristic UL power control policy from [2,
Fig. 2 shows cumulative distribution function (CDF) curves
Sec. 7.3], where the transmit power UE k in the cell j is
for the SE per UE. The randomness is due to random UE
j
βjk locations and shadow fading realizations. For UEs with good
, > ,

 ul
 p
 max ∆ j
β j,min channels, the MMSE and EW-MMSE estimators give the same

p jk =

j
βjk
j (70)

 p ul ∆ β j,min , ∆ ≤ ,
SE since the estimation errors are anyway small. On the other
 max β j
 j
β j,min hand, there is a noticeable difference between MMSE and EW-
 jk

j MMSE for the UEs with the weakest channel conditions.

where the maximum power is pul max = 10 dBm and β j,min = Fig. 3 shows the average sum DL SE over different UE
j j j
min(β j1, . . . , β jk , . . . , β jK ). This policy allows the UE with locations and shadow fading realizations, with MR precoding
the weakest channel to transmit at full power and require the based on the MMSE, LS or EW-MMSE estimators. The same
remaining UEs to reduce their power so their uplink SNRs are behaviors are observed as in the UL. Fig. 4 shows the average
at most ∆ = 10 dB higher. For simplicity, the same transmit UL SE with uncorrelated fading as a function the number of
power is used in the DL and as in the UL: ρ jk = p jk for each antennas. In this case, EW-MMSE and MMSE coincide since
UE. the spatial covariance matrices are diagonal. These estimators
3 This expression gives accurate results when each cluster has a small ASD, are better than LS when having Rician fading since the mean
as is the case here. vectors are utilized to improve the estimates. In contrast, all
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 11

1 20

0.9 18

0.8
Cumulative Distribution Function (CDF)

16

Average sum SE [bit/s/Hz/cell]

0.7
14

0.6
12

0.5
10
0.4
8
0.3

6
0.2 MMSE
EW-MMSE
LS 4 MMSE/EW-MMSE with Rician fading
0.1 LS with Rician fading
MMSE/EW-MMSE/LS with Rayleigh fading
2
0 10 20 30 40 50 60 70 80 90 100
0 1 2 3 4 5 6 7 8
Number of antennas (M)
SE per UE [bit/s/Hz]

Fig. 2: CDF of the UL SE per UE with M = 100 for different Fig. 4: Average UL SE with spatially uncorrelated Rician and
j j,NLoS
channel Rayleigh fading with Rli = βli IM .

30
22

20
25

18
Average sum SE [bit/s/Hz/cell]
Average sum SE [bit/s/Hz/cell]

16 20

14

15
12

10
10

MMSE
6 MMSE with Rician fading 5 MO
EW-MMSE with Rician fading EW-MMSE
LS with Rician fading LS
4 MMSE with Rayleigh fading
EW-MMSE/LS with Rayleigh fading 0
10 20 30 40 50 60 70 80 90 100
2
10 20 30 40 50 60 70 80 90 100 Number of antennas (M)
Number of antennas (M)
Fig. 5: Average UL sum SE for K=10 as a function of BS
Fig. 3: Average DL sum SE for K = 10 as a function of the antennas for different channel estimators where all UE-BS
number of BS antennas for different channel estimators. pairs have an LoS path.

the vectors give the same performance in the case of Rayleigh in a fraction 1/ f of the cells. The increased number of pilots
fading, since the estimates are equal up to a deterministic reduces the pre-log factor in (19) since τu = τc −τp , but it also
scaling factor, which cancel out in the SINR expressions. increases the instantaneous SINR in (20). There is a trade-off,
In Fig. 5, we consider a scenario with spatially correlated in that a larger reuse factor implies less pilot contamination but
Rician fading channels between every pair of BSs and UEs larger pre-log penalty. The LS estimator is more sensitive to
which may not be possible in practice. Compared the case pilot contamination since it cannot suppress pilot interference
where only some UEs have LoS (e.g. Fig. 1), the average UL by using spatial processing, while the MMSE estimators can
sum SEs are higher since the existence of an LoS component suppress this interference and therefore function even with a
improves the SE. Note that if the MO estimator is used and tighter reuse, and that this explains why the optimal reuse
some UEs in the cell have Rayleigh fading channels then their factor is different for the different estimators.
SEs would become zero with the MO estimator, which is why
we do not consider this case in the other simulations. VIII. C ONCLUSION
Fig. 6 shows the effect of three different ways to reuse This paper studied the UL and DL SE of a multi-cell
pilot sequences across the cells on the average sum UL SEs Massive MIMO system with spatially correlated Rician fading
for different estimators. The integer f denotes the pilot reuse channels. We derived rigorous closed-form achievable SE
factor where τp = f K. This means that there are f times more expressions when using either MMSE, EW-MMSE, or LS
pilots than UEs per cell and the same subset of pilots is reused estimation. The expressions provide exact insights into the
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 12

25
that
f=1
f=2   2
1   
f=4
H 12
E |x By| = tr (Rx ) BRy  + tr BRy B H Rx + | x̄ H Bȳ| 2
2
 H

 2 

20
   
1 1
+ 2Re tr (RxH ) 2 BRy2 ȳ H B H x̄ + x̄ H BRy B H x̄ + ȳ H B H Rx Bȳ.
Average sum SE [bit/s/Hz/cell]

15 (72)
Proof: The proof is given in Appendix B.
10
Lemma 6: [2, Lemmas B.7-8] Consider the positive semi-
definite matrices A ∈ C N ×N and B ∈ C N ×N . It holds that
tr (AB) ≤ kAk2 tr (B) , (73)
5
  1
tr A−1 B ≥ tr (B) , (74)
kAk2
0
MMSE EW-MMSE LS where k.k2 denotes the spectral norm which gives the largest
eigenvalue of A.
Fig. 6: Average UL sum SE for different pilot reuse factors
and different estimators.
A PPENDIX B
P ROOF OF L EMMA 4 AND L EMMA 5
1
operation and interference behavior when having Rician fading For the proof of Lemma 4, note that x = Rx2 wx + x̄ and
1
channels. These expressions can be utilized for more efficient y = Ry2 wy + ȳ where wx ∼ NC (0, I N ) and wy ∼ NC (0, I N ) are
user scheduling, pilot allocation, power control, and other independent vectors. Hence,
resource allocation in practical systems. We observed that the
E |x H By| 2 =

spatial correlation plays an important role and the existence
2
of an LoS component improves the achievable SE in Massive


   
 
MIMO. In addition, the MMSE estimator performs better than
  
1 1
 1 1  
E x̄ BRy2 wy + wx (Rx ) 2 BRy2 wy + x̄ Bȳ + wx (Rx ) 2 Bȳ .
 H
 H H H H H


the other estimators for both spatially correlated Rayleigh and  } |{z} | {z } 
Rician fading, while the LS estimator gives the lowest SE. In

 | {z } | {z 
 c f
 
d b
practice, the covariance matrices and the mean vectors might
 
(75)
not be known perfectly. Hence, the practical performance lies
between the MMSE/EW-MMSE and LS estimators since it is We compute each term as cc∗ = | x̄ H Bȳ| 2 and
highly probable that the mean is known up to a random phase- 
1

H 12 H
shift and covariance matrices are known with some error. E {dd } = E x̄ BRy wy wy (Ry ) B x̄ = x̄ H BRy B H x̄ (76)
∗ H 2 H

 
1
H 12
E { f f } = E ȳ B Rx wx wx (Rx ) Bȳ = ȳ H B H Rx Bȳ (77)
∗ H H 2 H

A PPENDIX A ( 2 )
 H H 1 1
U SEFUL RESULTS E {bb } = E wx (Rx ) 2 BRy wy 
∗

2

   
Lemma 4: Consider the vectors x ∼ NC (x̄, Rx ), with mean H 12
1 1 1
= E E wx (Rx ) BRy wy wy (Ry ) B (Rx )wx wx
H H

2 2 H H 2
vector x̄ ∈
C N and covariance matrix Rx ∈ C N ×N , and y ∼
NC ȳ, Ry with mean vector ȳ ∈ C N and covariance matrix 
1 1
  
Ry ∈ C N ×N . Also, B ∈ C N ×N is a deterministic matrix and x = E wxH (RxH ) 2 BRy B H (Rx2 )wx = tr BRy B H Rx (78)
and y are independent vectors. It holds that
The remaining terms are zero due to the circular symmetry
  properties and independence of wx and wy . This completes
E |x H By| 2
= tr BRy B H Rx + x̄ H BRy B H x̄

the proof of Lemma 4.
1
+ȳ H B H Rx Bȳ + | x̄ H Bȳ| 2 . (71) For the proof of Lemma 5, note that x = Rx2 w + x̄ and
1
y = Ry2 w + ȳ, where w ∼ NC (0, I N ) same. Hence,
Proof: The proof is given in Appendix B.
E |x H By| 2 =

Lemma 5: Consider the vectors x ∼ NC (x̄, Rx ), with mean  2
vector x̄ ∈
C N and covariance matrix Rx ∈ C N ×N , and y ∼ 




 
 
NC ȳ, Ry with mean vector ȳ ∈ C N and covariance matrix 1 1
 1 1  
E x̄ BRy w + w (Rx ) BRy w + x̄ Bȳ + w (Rx ) Bȳ .
 H
 H H 2 H H H 2


 2 2 
Ry ∈ C N ×N . Also, B ∈ C N ×N is a deterministic matrix. The  } |{z} | {z } 
 | {z } |
 {z 

vectors x and y are correlated and they are constructed as 
 d b c f
 

1 1
x = Rx2 w + x̄ and y = Ry2 w + ȳ where w ∼ NC (0, I N ). It holds (79)
 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 13

 The first term is identical with E kv jk k =

2

Similar to the proof of Lemma 4, we compute MMSE
  estimator.
1 1
∗ = k x̄ H BR 2 k 2 , E f f ∗ = k ȳ H B H R 2 k 2 , E H
 } =
E {dd ∗
 }  y { }  x {bc j j
E ĥ jk ĥ jk . Then,
1 1 1 1
tr (RxH ) 2 BRy2 ȳ H B H x̄, E {cb∗ } = tr (RyH ) 2 B H Rx2 x̄ H Bȳ,  
n o   H 
j j j
cc∗ = | x̄ H Bȳ| 2 and E vH = 2
=

h
jk jk E kv jk k tr E ĥ jk
ĥ jk
( 2 )  
j j j j
= p jk τp tr R jk Ψ jk R jk + k h̄ jk k 2 .
 H H 1 1
E {bb } = E w (Rx ) 2 BRy w
∗ (85)

2

  2
1    The last term in (20) can be written as
1
= tr (RxH ) 2 BRy2  + tr BRy B H Rx .

(80) (  )
  j  H j 2
  
 H j 2
E v jk hli  = E  ĥ jk hli 
The other terms are zero due to the circular symmetry property
of w. This finishes the proof of Lemma 5. (
  j  H j 2
 ) (
  j  H j 2
 )
= E  ĥ jk ĥli  + E  ĥ jk h̃li  , (86)
A PPENDIX C
P ROOF OF L EMMA 3 j j j
since h̃li and the pair (ĥ jk , ĥli ) of estimates are independent
The LS estimate is clearly Gaussian distributed. Its and estimation error has zero mean.
j j
mean
n ovalue and n covariance o matrix can be calculated n o as For (l, i) ∈ P jk , ĥ jk and ĥli are not independent. The
j p p
E ĥli = E √ p 1 τ p y jli = √ p 1 τ p ȳ jli and Cov ĥli =
j j j √ j j p p
li li
channel estimates are ĥ jk = h̄ jk + p jk R jk Ψ jk (y j jk − ȳ j jk )
  j j √ j j p p p
and ĥli = h̄li + pli Rli Ψ jk (y j jk − ȳ j jk ) where y j jk ∼
 H
p p p p j
1
p τ 2 E y jli
− ȳ jli
y jli
− ȳ jli
= pli1τ p (Ψli )−1 . Similarly,   −1 
li p 
p j
the mean and covariance of the estimation error can be NC ȳ j jk , τp Ψ jk . We reformulate these estimates using
computed as   1/2
j p p
a matrix Q−1/2 = √1τ p Ψ jk that gives Q−1/2 (y j jk − ȳ j jk ) =
n o
j
n
j j
o
j 1 p
E h̃li = E hli − ĥli = h̄li − √ ȳ , (81) w ∼ NC (0, I N ). Then, the channel estimates are written as
pli τp jli
j √ j j p p j
ĥ jk = p jk R jk Ψ jk Q1/2 Q−1/2 (y j jk − ȳ j jk ) + h̄ jk
n o n o n o n o
j j j j j j j
Cov h̃li = E hli (hli ) H − E hli (ĥli ) H − E ĥli (hli ) H √   1/2
j j j
n o n o n o = p jk τp R jk Ψ jk w + h̄ jk , (87)
j j j j
+E ĥli (ĥli ) H − E h̃li E (h̃li ) H (82)
√   1/2
j j j j
ĥli = pli τp Rli Ψ jk w + h̄li . (88)
where
n
j j
o 1 n
p j
o 1
√
 1/2
E ĥli (hli ) H = √

E y jli (hli ) H Using Lemma 5 for Rx2 = p jk τp R jk Ψ jk
j j j
, x̄ = h̄ jk and
pli τp  1/2
1
√

j j j
1

 Õ √
 
 Ry2 = pli τp Rli Ψ jk , ȳ = h̄li gives (89) at the top of this
j j H p ∗ j H
=√ pl0 i0 τp hl0 i0 (hli ) + N j φli (hli )
 

E j j j j
pli τp   (l0,i0 )∈ Pli 
 page by using the fact that pli τp Rli Ψ jk Rli = Rli − Cli . To
(  )
 √    j  H j 2
j
Õ pl0 i0 j j H calculate E  ĥ jk h̃li  in (86), we use Lemma 4 since
= Rli +

√ h̄l0 i0 (h̄li ) (83)
(l0,i0 )∈ P
pli j j
li
ĥ jk and h̃li are independent. Noting that the estimation error
n o j
j j j j j
and E hli (hli ) H = Rli + h̄li (h̄li ) H . Inserting all terms into h̃li has distribution NC (0, Cli ), we obtain
(82) gives the covariance matrix of the estimation error as (  )
  j  H j 2 
j j j
  H
j j
shown in Lemma 3. E  ĥ jk h̃li  = p jk τp tr Cli R jk Ψ jk R jk + h̄ jk Cli h̄ jk .


(90)
A PPENDIX D
Substituting (89) and (90) back into (86) gives the final result
P ROOF OF T HEOREM 1
n o for (l, i) ∈ P jk as in (23).
j 2 j j
We need to characterize the terms E v H

  jk jk , E kv jk k
h For the case of (l, i) < P jk , ĥ jk and hli are independent.
j
2 Using Lemma 4, we get
and E v H h in (20). We begin with calculating

jk li  (  )
  j  H j 2  H
j j j j j
n o n o n o n o E  ĥ jk hli  = p jk τp h̄li R jk Ψ jk R jk h̄li (91)
j j H j j H j j H j
E vH h
jk jk = E (ĥ jk
) h jk
= E ( ĥ jk
) ĥ jk
+ E (ĥ jk
) h̃ jk
.
   H  2
n o (84) j j j j j j j  j j
+ p jk τp tr Rli R jk Ψ jk R jk + h̄ jk Rli h̄ jk + (h̄ jk ) H h̄li  .
j j
The second term E (ĥ jk ) H h̃ jk is zero due to the inde-
pendence of estimate and estimation error when using the This finishes the proof of Theorem 1.
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 14

(  )
  j  H j 2  
j j
 2
j 

j

j j j
 
 j j
2
E  ĥ jk ĥli  = p jk pli τp2 tr Rli Ψ jk R jk  + p jk τp tr Rli − Cli R jk Ψ jk R jk + (h̄ jk ) H h̄li 
 

    H   H  H 
√ j j j j j j j j j j j j

j
+ 2 p jk pli τp Re tr Rli Ψ jk R jk h̄li h̄ jk + p jk τp h̄li R jk Ψ jk R jk h̄li + h̄ jk Rli − Cli h̄ jk , (89)

 2  n o √ √ n o
j j H j j H j j H j j j H j j j
E v H h = (h ) h̄ ( h̄ ) h + p p τ (h ) h̄ (h − h̄ ) Λ D h

jk li  E li jk jk li jk li p E li jk li li jk jk li
n o √ √ n o
j j j H j j j j j j j j j j
+ p jk E (hli ) H D jk Λ jk x jk − x̄ jk x jk − x̄ jk Λ jk D jk hli + p jk pli τp E (hli ) H D jk Λ jk (hli − h̄li )(h̄ jk ) H hli



  H 
j j j j j j j j j j
+ p jk pli τp2 E (hli ) H D jk Λ jk (hli − h̄li ) hli − h̄li Λ jk D jk hli . (93)

   2
j j j j j j j j j j  j j
ξliulp jk τp kRli k2 tr R jk Ψ jk R jk + p jk τp k h̄li k 2 kR jk Ψ jk R jk k2 + kRli k2 k h̄ jk k 2 + (h̄ jk ) H h̄li 
≤   . (101)
Mj j j j j
p jk τp M j tr R jk Ψ jk R jk + M j k h̄ jk k 2

p j k pli τ p2
 2
√
  n   o
j j j  j j j j j
tr Rli Ψ jk R jk  2 pli p jk Re tr Rli Ψ jk R jk M1 j (h̄ jk ) H h̄li

Γliul Mj
=   +   . (102)
Mj j j j j
p jk τp tr R jk Ψ jk R jk + k h̄ jk k 2
j j j j
p jk τp tr R jk Ψ jk R jk + k h̄ jk k 2

A PPENDIX E Appendix C. The last term that is needed to calculate in (20)

P ROOF OF T HEOREM 2 is
     
 H j 2 1  p H j 2
E v jk hli  =
n o
We need to characterize the terms E v H
j
h jk , E kv jk k 2 E (y j jk ) hli  (94)

  
jk p jk τp2
 H j 2
n o
j 2
in (20). Calculations of E v H

and E v jk hli  jk h jk and  
 Õ √
 
1
 
j H j p ∗ H j 
= 0 i 0 τp (h 0 0 ) h + (N φ ) h ,
 

 2
E  pl j jk
E  kv jk k  are rather straightforward. For the third term p jk τp  (l0,i0 )∈ P
2   li li li  
 
 2 j k
j j j
E v H
 
jk li  , if (l, i) < P jk then ĥ jk and hli are independent
h

p
where the noise term has distribution N j φ ∗jk ∼
vectors. Directly applying Lemma 4 gives  
NC 0, σul τp I M and kφ jk k = τp .
2 2
     −1 
 j H j 2
  
j j j j j p p j
E (ĥ jk ) hli  = tr Rli Σ jk + (h̄ jk ) H Rli h̄ jk If (l, i) < P jk , note that y j jk ∼ NC ȳ j jk , τp Ψ jk and
p j
j j j

 j j
2 y j jk and hli are independent random vectors. Directly utilizing
+ (h̄li ) H Σ jk h̄li + (h̄ jk ) H h̄li  . (92) Lemma 4 gives
 2     −1 
 p j j j p j p
j j
If (l, i) ∈ P jk then ĥ jk and hli are not independent vectors. E (y j jk ) H hli  = τp tr Rli Ψ jk + (ȳ j jk ) H Rli ȳ j jk
√ j p
Decomposing y j jk using x jk = y j jk − pli τp hli as (y j jk −   −1  2
j j j  p j
p H √ j j
ȳ j jk ) = x jk − x̄ jk + pli τp (hli − h̄li ) gives (93) at the top

+ τp (h̄li ) H Ψ jk h̄li + (ȳ j jk ) H h̄li  . (95)
of previous page. Computing each term
 in (93) and noticing
   p j
j j j j j j
that tr R jk Λ jk D jk = tr D jk Λ jk D jk leads to (29). If (l, i) ∈ P jk , then y j jk and hli are not independent random
p j
vectors since contains
y j jk hli .
We decompose the terms in
√ j
(94) by using x jk = y j jk − pli τp hli . Then we obtain
A PPENDIX F 
√ 2 
j j H j
E x H h + p τ (h ) h = (96)

P ROOF OF T HEOREM 3 jk li li p li li 
 2    
j  j H j 2
+ τ
n o 2
We need to characterize the terms E v H
j
, E kv jk k 2 E x H h p ) h
 
h jk li  li p E (hli li 
 jk jk
2 
j √ n
j j H j
o √ n
j H j j H
o
and E v H h in (20). The first two terms are given in + pli τp E x H jk hli (hli ) hli + pli τp E (hli ) hli (hli ) x jk .

jk li 
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 15

Further, we compute each term beginning with expressions for γ dl,mmse

jk
contains the same matrix expressions
as γ ul,mmse , except that indices (l, i) and ( j, k) are swapped in
  
 H j 2
 
j j j jk
E x jk hli  = τp tr Rli (Ω jk )−1 + x̄ H jk Rli x̄ jk the interference terms. The proofs follows analogously.
 2
j j j j
+ τp (h̄li ) H (Ω jk )−1 h̄li + x̄ H h̄ (97) A PPENDIX H

jk li 
P ROOF OF T HEOREM 9 AND T HEOREM 10
j j j
where (Ω jk )−1 = (Ψ jk )−1 − pli τp Rli . The other terms are The proof begins with dividing the numerator  and denom- 
j j j

 j j
2     2
j 
  
j
2 inator of γ ul,ew
jk
in (30) by M j and p jk τp tr D Λ D
jk jk jk
+
E (hli ) H hli  = tr Rli  + tr Rli

j
k h̄ jk k 2 . Then, the numerator is strictly positive and finite as
j
 
j
 H
j j j j M j → ∞ due to Assumptions 1 and 2. The interference
+ 2k h̄li k 2 tr Rli + 2 h̄li Rli h̄li + k h̄li k 4, (98) j j j j
term becomes χliul M j (p jk τp tr D jk Λ jk D jk + k h̄ jk k 2 ) and


n o n o n o the same procedure is applied as in the proof of Theo-

j j H j j j H j
E xH h
jk li li(h ) hli
= E x H
jk E h (h
li li
) hli rem 7. The non-coherent interference goes asymptotically
H j
 
j H j j H j j H j zero due  to
 Assumptions 1-3. For the noise term, note that
= x̄ jk h̄li tr Rli + x̄ jk Rli h̄li + x̄ jk h̄li (h̄li ) h̄li, (99) j j
tr Σ j k + k h̄ j k k 2

j j j

j
is bounded since the trace expres-
n o p j k τ p tr D j k Λ j k D j k + k h̄ j k k 2
j j j j j sions can not grow faster than
E (hli ) H hli (hli ) H x jk = (h̄li ) H Rli x̄ jk    j due  to Assumption 1.
M
j j
  σul2 tr Σ
jk
+ k h̄ j k k 2
j j j j j
+ tr Rli (h̄li ) H x̄ jk + (h̄li ) H h̄li (h̄li ) H x̄ jk . (100) Then the noise term 
j j j

j
goes to zero
M j (p j k τ p tr D j k Λ j k D j k + k h̄ j k k 2 )
as M j → ∞ due to Assumptions 1 and 2. The coherent
Combining the results for (l, i) ∈ P jk and (l, i) < P jk give
   interference term
 H j 2
the final result of E v jk hli  in (33). p j k pl i τ p2
 
j j j
2
Mj tr Dli Λ jk D jk
  (104)
j j j j
A PPENDIX G p jk τp tr D jk Λ jk D jk + k h̄ jk k 2
P ROOF OF T HEOREM 7 AND T HEOREM 8
is  bounded  since the expression
The proof begins with dividing the numerator and denom- j j j j
p jk τp tr D jk Λ jk D jk + k h̄ jk k 2 in the denominator scales
inator of γ ul,mmse in (24) by M j . The numerator becomes
p 2j k τ p
 jk  with M j and the traces in the numerator can not grow faster
j j j p j
Mj tr R Ψ R
jk jk jk
+ Mj kj k h̄ jk k 2 . This term is strictly positive than M j due to Assumption 1. For this term, we note that
and finite as M j → ∞ due to the Assumptions 1 and 2. The 1  j j j  1 j  j j 
non-coherent interference in (25) satisfies (101) at the top of tr Dli Λ jk D jk ≤ Λ jk tr Dli D jk . (105)
Mj Mj 2
next page.  
j j
If (l, i) , ( j, k) then the upper bound in (101) goes to It goes to zero if M1 j tr Dli D jk → 0 for all (l, i) ∈
zero as M j → ∞ due to Assumptions 1–3. The LoS-related P jk \( j, k), which happens under asymptotic spatial orthogo-
j
ν ul k h̄ j k k 4 nality. In this case, the whole denominator vanishes and the
term Mj
jk
= 
j j j

j
is strictly finite as
p j k τ p M j tr R j k Ψ j k R j k +M j k h̄ j k k 2 SINR goes asymptotically to infinity. Otherwise, it is only
M j → ∞ due to Assumption 2. This term cancels out the non- these terms that remain asymptotically in the denominator.
coherent interference term ξliul /M j for (l, i) = ( j, k) in (101). This completes the proof of the UL part. In the DL, we
σ2
The noise term Mulj also goes to zero as M j → ∞. Considering divide the numerator and denominator of γ dl,ew jk
in (49) by
 
j j j j 2
the coherent interference in (26), we obtain (102) at the top p jk τp tr D jk Λ jk D jk +k h̄ jk k . The numerator becomes strictly
of this page. positive and finite due to Assumptions 1 and 2. The inter-
The second term in (102) goes to zero due to Assump- χldli
ference term is     . By
tions 1–3.  The first term is bounded, since the expression (tr(Σ ll i )+ k h̄ll i k 2 ) p j k τ p tr D jj k Λ jj k D jj k + k h̄ jj k k 2
j j j j
p jk τp tr R jk Ψ jk R jk + k h̄ jk k 2 in the denominator scales with applying similar process as in the proof of Theorem 8, we
M j and the traces in the numerator can not grow faster than can complete the proof of Theorem 10.
M j due to Assumption 1. We note that
A PPENDIX I
1   j j j   1 j  j j  P ROOF OF T HEOREM 11
tr Rli Ψ jk R jk  ≤ Ψ jk tr Rli R jk . (103)
Mj Mj 2
The proof begins with dividing the numerator
and denominator of γ ul,ls
 
j j
It goes to zero if M1 j tr Rli R jk → 0 for all (l, i) ∈ jk
in (34) by M j and
 √ 
j pjk j H j 
tr(R jk ) + (l,i)∈ P j k √ pl i (h̄li ) h̄ jk .
Í
The numerator

P jk \( j, k), which happens under asymptotic spatial orthogo-  √ 
nality. In this case, the SINR grows without bound as specified  pjk j pjk j H j 
becomes  M j tr(R jk ) + (l,i)∈ P j k M j √ p (h̄li ) h̄ jk . Note
Í
in the theorem. Otherwise, the SINR will asymptotically only pjk j
li

contain these terms in the denominator. This finishes the that Mj tr(R jk ) > 0 due to Assumption 1 and the only non-
1 j
proof for the UL SE with MMSE estimator. In the DL, the zero term in the second part is Mj k h̄ jk k 2 due to Assumption
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 16

  Í 
j j j j
kRli k2 tr (Ψ jk )−1 kRli k2 (l,i)∈ P j k pli τp βli + σul
 √
pjk
 = √
pjk
, (106)
j j j j j j
M j tr(R jk ) + (l,i)∈ P j k (h̄li ) H h̄ jk tr(R jk ) + (h̄li ) H h̄ jk
Í Í
√
pl i (l,i)∈ P j k √ pl i

√ n
p j
 
j p j j
o √ j
 
j
2 pli τp Re (ȳ j jk ) H h̄li tr Rli + (ȳ j jk ) H Rli h̄li 2 pli τp k h̄li k 2 tr Rli
 √
p
 −  √
p
 → 0, (107)
j j j j j j
M j tr(R jk ) + (l,i)∈ P j k √ pj k (h̄li ) H h̄ jk M j tr(R jk ) + (l,i)∈ P j k √ pj k (h̄li ) H h̄ jk
Í Í
li li

 
j p
σul 2 τ p tr (Ψ j k )−1 + k ȳ j j k k 2 [7] Q. Zhang, S. Jin, K.-K. Wong, H. Zhu, and M. Matthaiou, “Power
3. The noise term p j k τ p2 M j 
 √
pl i

scaling of uplink massive MIMO systems with arbitrary-rank channel
j Í j H j 
tr(R j k )+ (l, i)∈P j k √ p j k (h̄l i ) h̄ j k  means,” IEEE Journal of Selected Topics in Signal Processing, vol. 8,
goes asymptotically to zero since the first factor goes no. 5, pp. 966–981, Oct. 2014.
to zero and the second factor is bounded due to [8] C. Kong, C. Zhong, M. Matthaiou, and Z. Zhang, “Performance of
downlink massive MIMO in Ricean fading channels with ZF precoder,”
Assumptions
 1–3. For the interference  term, by using in IEEE International Conference on Communications (ICC), Jun. 2015.
j j j j
tr Rli (Ψ jk )−1 ≤ kRli k2 tr (Ψ jk )−1 , the upper bound of [9] Y. Hu, Y. Hong, and J. Evans, “Angle-of-arrival-dependent interference
modeling in Rician massive MIMO,” IEEE Transactions on Vehicular
the interference term is (106) at the top of previous page Technology, vol. 66, no. 7, pp. 6171–6183, Jul. 2017.
and it goes asymptotically to zero due to Assumptions 1 [10] L. Zhao, T. Yang, G. Geraci, and J. Yuan, “Downlink multiuser massive
and 2. Another term is (107) at the top of previous page MIMO in Rician channels under pilot contamination,” in IEEE Interna-
tional Conference on Communications (ICC), May 2016.
asymptotically due to Assumption 1–3. The inequalities [11] L. Wu, Z. Zhang, J. Dang, J. Wang, H. Liu, and Y. Wu, “Channel
 H estimation for multicell multiuser massive MIMO uplink over Rician
p j p j p
ȳ j jk Rli ȳ j jk ≤ ||Rli ||2 k ȳ j jk k 2 (108) fading channels,” IEEE Transactions on Vehicular Technology, vol. 66,
no. 10, pp. 8872–8882, Oct. 2017.
j j j j j [12] L. Sanguinetti, A. Kammoun, and M. Debbah, “Asymptotic analysis
(h̄li ) H (Ω jk )−1 h̄li ≤ k(Ω jk )−1 k2 k h̄li k 2 (109)
of multicell massive MIMO over Rician fading channels,” in IEEE
j j j j j
(h̄li ) H (Ψ jk )−1 h̄li ≤ k(Ψ jk )−1 k2 k h̄li k 2 (110) International Conference on Acoustics, Speech and Signal Processing
(ICASSP), Mar. 2017.
j j H j j H j 2 [13] 3rd Generation Partnership Project, Technical Specification Group Radio
x̄ H
jk h̄li (h̄li ) h̄li ≤ k h̄li x̄ jk k2 k h̄li k (111)
Access Network; Spatial channel model for Multiple Input Multiple
Output (MIMO) simulations, Mar. 2017, 3GPP TR 25.996 V14.0.0.
are also useful. Using these inequalities, Assumptions 1 and [14] J. Zhang, C.-K. Wen, S. Jin, X. Gao, and K.-K. Wong, “On capacity of
2, and noting that large-scale MIMO multiple access channels with distributed sets of cor-
 2 related antennas,” IEEE Journal on Selected Areas in Communications,
j j vol. 31, no. 2, pp. 133–148, Feb. 2013.
pli τp2 tr(Rli ) + pli τp2 k h̄li k 4 [15] H. Tataria, P. J. Smith, L. J. Greenstein, P. A. Dmochowski, and
 √
p
 (112) M. Matthaiou, “Impact of line-of-sight and unequal spatial correlation
j j j
M j p jk τp2 tr(R jk ) + (l,i)∈ P j k √ pj k (h̄li ) H h̄ jk
Í
on uplink MU-MIMO systems,” IEEE Wireless Communications Letters,
li
vol. 6, no. 5, pp. 634–637, Oct. 2017.
has a positive finite limit as M j → ∞, we obtain (61). This [16] O. Özdogan, E. Björnson, and E. G. Larsson, “Uplink spectral efficiency
of massive MIMO with spatially correlated rician fading,” in 2018 IEEE
finishes the proof for the UL. 19th International Workshop on Signal Processing Advances in Wireless
In the DL, we divide the numerator and denominator of Communications (SPAWC), June 2018, pp. 1–5.
γ dl,ls
jk
in (52) by M j . The numerator is positive and finite as [17] E. Björnson, L. Sanguinetti, and M. Debbah, “Massive MIMO with
imperfect channel covariance information,” in Proc. ASILOMAR, 2016.
σ2 [18] D. Neumann, M. Joham, and W. Utschick, “Covariance matrix estima-
M j → ∞ due to Assumptions 1 and 2. The noise term Mdlj
tion in massive MIMO,” IEEE Signal Processing Letters, vol. 25, no. 6,
goes to zero as M j goes to infinity. For interference term, we pp. 863–867, June 2018.
follow the approach from the UL part to obtain (62). [19] K. Upadhya and S. A. Vorobyov, “Covariance matrix estimation for
massive MIMO,” IEEE Signal Processing Letters, vol. 25, no. 4, pp.
546–550, April 2018.
R EFERENCES [20] S. Haghighatshoar and G. Caire, “Massive mimo pilot decontamination
and channel interpolation via wideband sparse channel estimation,” IEEE
[1] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, “Massive MIMO Transactions on Wireless Communications, vol. 16, no. 12, pp. 8316–
for next generation wireless systems,” IEEE Communications Magazine, 8332, Dec 2017.
vol. 52, no. 2, pp. 186–195, Feb. 2014. [21] S. M. Kay, Fundamentals of statistical signal processing: Estimation
[2] E. Björnson, J. Hoydis, and L. Sanguinetti, “Massive MIMO networks: theory. Prentice Hall, 1993.
Spectral, energy, and hardware efficiency,” Foundations and Trends
R [22] N. Shariati, E. Björnson, M. Bengtsson, and M. Debbah, “Low-
in Signal Processing, vol. 11, no. 3-4, pp. 154–655, 2017. complexity polynomial channel estimation in large-scale MIMO with
[3] T. L. Marzetta, E. G. Larsson, H. Yang, and H. Q. Ngo, Fundamentals arbitrary statistics,” IEEE Journal of Selected Topics in Signal Process-
of Massive MIMO. Cambridge University Press, 2016. ing, vol. 8, no. 5, pp. 815–830, Oct 2014.
[4] J. Hoydis, S. Ten Brink, and M. Debbah, “Massive MIMO in the UL/DL [23] H. Yin, D. Gesbert, M. Filippou, and Y. Liu, “A coordinated approach
of cellular networks: How many antennas do we need?” IEEE J. Sel. to channel estimation in large-scale multiple-antenna systems,” IEEE J.
Areas Commun., vol. 31, no. 2, pp. 160–171, Feb. 2013. Sel. Areas Commun., vol. 31, no. 2, pp. 264–273, 2013.
[5] H. Yang and T. L. Marzetta, “Massive MIMO with max-min power
control in line-of-sight propagation environment,” IEEE Transactions
on Communications, vol. 65, no. 11, pp. 4685–4693, Jul. 2017.
[6] D. Tse and P. Viswanath, Fundamentals of wireless communications.
Cambridge University Press, 2005.
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 67, NO. 5, MAY 2019 17

Özgecan Özdogan (S’18) received her B.Sc and

M.Sc. degrees in Electronics and Communication
Engineering from Äřzmir Institute of Technology,
Turkey in 2015 and 2017 respectively. She is cur-
rently pursuing the Ph.D. degree in communication
systems at Linköping University, Sweden.

Emil Björnson (S’07-M’12-SM’17) received the

M.S. degree in Engineering Mathematics from Lund
University, Sweden, in 2007. He received the Ph.D.
degree in Telecommunications from KTH Royal In-
stitute of Technology, Sweden, in 2011. From 2012
to mid 2014, he was a joint postdoc at the Alcatel-
Lucent Chair on Flexible Radio, SUPELEC, France,
and at KTH. He joined Linköping University, Swe-
den, in 2014 and is currently Associate Professor and
Docent at the Division of Communication Systems.
He performs research on multi-antenna commu-
nications, Massive MIMO, radio resource allocation, energy-efficient com-
munications, and network design. He is on the editorial board of the IEEE
Transactions on Communications (since 2017) and the IEEE Transactions on
Green Communications and Networking (since 2016). He is the first author
of the textbooks “Massive MIMO Networks: Spectral, Energy, and Hardware
Efficiency” (2017) and “Optimal Resource Allocation in Coordinated Multi-
Cell Systems” from 2013. He is dedicated to reproducible research and has
made a large amount of simulation code publicly available.
Dr. Björnson has performed MIMO research for more than ten years and has
filed more than ten related patent applications. He received the 2018 Marconi
Prize Paper Award in Wireless Communications, the 2016 Best PhD Award
from EURASIP, the 2015 Ingvar Carlsson Award, and the 2014 Outstanding
Young Researcher Award from IEEE ComSoc EMEA. He also co-authored
papers that received best paper awards at the conferences WCSP 2017, IEEE
ICC 2015, IEEE WCNC 2014, IEEE SAM 2014, IEEE CAMSAP 2011, and
WCSP 2009.

Erik G. Larsson (S’99-M’03-SM’10-F’16) received

the Ph.D. degree from Uppsala University, Uppsala,
Sweden, in 2002. He is currently Professor of Com-
munication Systems at Linköping University (LiU)
in Linköping, Sweden. He was with the KTH Royal
Institute of Technology in Stockholm, Sweden, the
George Washington University, USA, the University
of Florida, USA, and Ericsson Research, Sweden.
His main professional interests are within the areas
of wireless communications and signal processing.
He has co-authored some 170 journal papers on
these topics, and the two Cambridge University Press textbooks Space-Time
Block Coding for Wireless Communications (2003) and Fundamentals of
Massive MIMO (2016). He is co-inventor on 19 issued U.S. patents.
He is a member of the IEEE Signal Processing Society Awards Board
(2017–2019), an editorial board member of the IEEE Signal Processing
Magazine (2018–2020), and a member of the steering committee for the
IEEE Transactions on Wireless Communications (2019–2022). From 2015 to
2016 he was chair of the IEEE Signal Processing Society SPCOM technical
committee. From 2014 to 2015 he was chair of the steering committee for
the IEEE Wireless Communications Letters. He was General Chair of the
Asilomar Conference on Signals, Systems and Computers in 2015, and its
Technical Chair in 2012. He was Associate Editor for, among others, the IEEE
Transactions on Communications (2010-2014) and the IEEE Transactions on
Signal Processing (2006-2010).
He received the IEEE Signal Processing Magazine Best Column Award
twice, in 2012 and 2014, the IEEE ComSoc Stephen O. Rice Prize in
Communications Theory in 2015, the IEEE ComSoc Leonard G. Abraham
Prize in 2017, and the IEEE ComSoc Best Tutorial Paper Award in 2018.