Physical Communication 48 (2021) 101432

Contents lists available at ScienceDirect

Physical Communication
journal homepage: www.elsevier.com/locate/phycom

Full length article

Pilot and data power allocation for spatially correlated massive MIMO
systems with imperfect CSI
∗ ∗
Yixin Xu a , , Fulai Liu b , , Ruiyan Du b , Zhenxing Sun a
a
School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China
b
Institute of Engineering Optimization and Smart Antenna, Northeastern University at Qinhuangdao, Qinhuangdao 066004, China

article info a b s t r a c t

Article history: This paper investigates pilot and data power allocation for a multi-cell spatially correlated massive
Received 4 July 2020 multiple-input multiple-output (MIMO) system with imperfect channel state information (CSI). By
Received in revised form 20 May 2021 exploiting the availability of statistical CSI, the uplink closed-form spectral efficiency (SE) expressions
Accepted 15 July 2021
are derived for both full covariance knowledge case and no prior statistical information case. In order to
Available online 17 July 2021
enhance the accuracy of channel estimation, novel pilot power allocation schemes are firstly designed
Keywords: by minimizing the sum normalized mean square error (NMSE) of all users. Then, based on the optimal
Massive MIMO pilot power, data power allocation schemes are proposed to maximize the sum SE of system. For
Power allocation solving these non-convex problems, an iterative optimization algorithm is developed to obtain the local
Correlated Rayleigh fading optimal solution by exploiting second-order channel statistics of users. Compared with the traditional
method, numerical results show that the proposed pilot and data power allocation schemes can achieve
higher SE for spatially correlated massive MIMO system.
© 2021 Elsevier B.V. All rights reserved.

1. Introduction scheme can achieve higher SE than the jointly power allocation
mentioned in [10].
Massive multiple-input multiple-output (MIMO) system has All aforementioned power allocation schemes assume that
been investigated for beyond fifth-generation (5G) wireless net- the channel is spatially uncorrelated Rayleigh fading. However,
works, since it can provide much higher spectral efficiency (SE) the practical channels of massive MIMO system are spatially
than traditional MIMO [1–3]. In such system, tens of users are correlated [12]. To this end, the ergodic SE expression with per-
simultaneously served via base station (BS) which is equipped fect channel state information (CSI) was derived for single-cell
with hundreds of antennas [4,5]. However, the performance of spatially correlated massive MIMO in [13] and [14]. Then, for
system is seriously degraded by pilot contamination [6]. multi-cell spatially correlated massive MIMO system with imper-
Power allocation plays an important role to alleviate the ef- fect CSI, the effects of pilot contamination can be decreased by
fects of pilot contamination and enhance the SE in massive MIMO adopting the user grouping and pilot assignment [15]. Moreover,
system [7–11]. For single-cell massive MIMO system, a scheme by using different channel estimators, the closed-form uplink and
of pilot power allocation was developed to improve the per- downlink SE expressions were derived and analyzed for max-
imum ratio combining in [16]. However, the power allocation
formance of system by enhancing channel estimation quality
scheme is not considered in these works, which is very important
in [7]. Then, a joint pilot and data power control problem was
in massive MIMO system as mentioned before.
formulated to improve the SE of system, which can be solved
In addition, spatially correlated massive MIMO has recently
by utilizing the general inner approximation algorithm [8]. For
attracted a significant research interest in several scenarios [17–
multi-cell massive MIMO system, a scheme of pilot power alloca-
20]. In [17], a beam-domain full-duplex massive MIMO was pro-
tion based on the relative channel estimation error was designed
posed to realize co-time co-frequency uplink and downlink trans-
to improve the performance of channel estimation in [9]. After
mission. After that, a beam-domain hybrid time switching and
that, a joint pilot and data power scheme was proposed to max-
power splitting protocol was investigated to intelligently sched-
imize the sum SE of system [10]. Later on, a disjoint pilot and
ule users and sensors [18]. Then, a beam-domain secure trans-
data power control was shown in [11], and it was proved that this
mission scheme was designed to against a full-duplex moving
eavesdropper [19]. By exploiting the time and spatial properties
∗ Corresponding authors. of channels, a hybrid beamforming scheme was proposed for
E-mail addresses: [email protected] (Y. Xu), [email protected] (F. Liu). high-speed railways communication in [20]. Moreover, a set of

https://doi.org/10.1016/j.phycom.2021.101432
1874-4907/© 2021 Elsevier B.V. All rights reserved.
Y. Xu, F. Liu, R. Du et al. Physical Communication 48 (2021) 101432

potential scenarios under which massive MIMO can fail in Rician 2.2. Channel estimation
fading channels is identified in [21], which indicates that mas-
sive MIMO can be intrinsically effective at combating inter-user Without loss of generality, assume that the orthogonal pilot
interference. sequences are assigned to users in one cell and this pilot set is
Motivated by the above discussion, this paper aims to design reused for other cells [1]. In uplink channel estimation phase, all
pilot and data power allocation schemes for spatially correlation users simultaneously transmit their orthogonal pilot signals to
massive MIMO system with imperfect CSI. By considering the the corresponding BS in each  cell
2 for performing channel estima-
availability of statistical CSI, the uplink closed-form SE expres- tion. Let φk ∈ Cτ p ×1 , where φk  = 1, denote the pilot sequence
sions are derived for both full covariance knowledge case and no used by user k. Thus, the pilot signal Y j ∈ CM ×τp received at BS j
prior statistical information case. In these expressions, the impact can be given by
of spatially correlated channel, pilot power, data power and pilot
L K
contamination are all taken into consideration. Note that pilot ∑ ∑ √
power exerts considerable influence on the accuracy of channel Yj = pln hjln φH
n + N p, (1)
estimation since the uplink pilots are only utilized to perform l=1 n=1

channel estimation. Meanwhile, the SE of system can be improved where( N p) refers to the independent additive noise with i.i.d
by enhancing the accuracy of channel estimation [11]. Based on C N 0, σ 2 elements and pli denotes the pilot transmit power for
these key principles, the schemes of pilot and data power alloca- user i in cell l.
tion are designed to increase the accuracy of channel estimation In order to estimate the channel hjjk , the received pilot signal
and the sum SE for multi-cell spatially correlated massive MIMO in (1) is correlated with the pilot φk as follow
systems, respectively. The main contribution of this paper can be
L
summarized as follows: √ ∑ √
y jjk = Y j φk = pjk hjjk + pik hjik + N p φk . (2)
• To measure error of channel estimation, the normalized i̸ =j
mean square error (NMSE) metric is adopted, for which
the closed-form expressions are derived under full covari- If the BS has no prior information, the least-square (LS)
ance knowledge case and no prior statistical information method can be used for channel estimation. Then, the estimated
LS
case. For these cases, pilot power allocation schemes are channel ĥjjk can be expressed as [16]
formulated by minimizing the sum NMSE to improve the
LS 1
accuracy of channel estimation, which can be transformed ĥjjk = √ y jjk . (3)
into geometric program (GP). pjk
• Based on the optimal pilot power, novel data power allo- If the BS has full knowledge of covariance matrix, the
cation schemes are designed for full covariance knowledge minimum-mean-squared-error (MMSE) method can be used for
case and no prior statistical information case by maximizing MMSE
channel estimation. Then, the estimated channel ĥjjk can be
the sum SE of system. To solve this non-convex problem,
calculated as [22]
an iterative optimization algorithm is developed to obtain a
local optimum. MMSE √
jjk y jjk ,
1
ĥjjk = pjk R jjk Q − (4)
The rest of this paper is organized as follows. The system ∑L
where Q jjk = l=1 plk R jlk + σ I M .2
model of spatially correlated massive MIMO system is introduced
in Section 2. Pilot power allocation schemes are discussed in
2.3. Uplink data transmission
Section 3. In Section 4, data power allocation schemes are for-
mulated to maximize the sum SE of system. Furthermore, an
During uplink data transmission phase, the BS detects data
iterative optimization algorithm is proposed to solve this non-
signals with the maximum ratio combining (MRC) detector. Then,
convex problem. Simulation results are provided in Section 5.
the received data signal r jk at BSj can be modeled as
Finally, conclusions are drawn in Section 6.
Notation: (·)H and (·)−1 denote the conjugate transpose, matrix K
√ H ∑ √ H
inversion, respectively. E {·} denotes the expectation. tr(·) denotes rjk = qjk ĥjjk hjjk xjk + qjn ĥjjk hjjn xjn
the trace of matrix. Finally, the notation C N (0, R ) denotes the n=1
n̸ =k
complex Gaussian distribution with mean 0 and covariance R.
L K
∑ ∑ √ H H
2. System model + qln ĥjjk hj ln xln + ĥjjk nd , (5)
l=1 n=1
l̸ =j
2.1. Channel model
where xjk ∼ C N (0, 1) refers to the data signal intended for user
Consider an uplink transmission of L-cell spatially correlated k in cell j and qjk denotes the transmit
( data) power allocated for
massive MIMO system, where each cell contains one M-antennas user k in cell j. Moreover, nd ∼ C N 0, σ 2 represents the vector
BS and K single-antenna users. The whole system operates in time of additive Gaussian noise and ĥjjk stands for the MRC detection
division duplex (TDD) model. Let τc samples define the length vector.
of each coherence interval, where τp samples are used for pilot Based on (5), a lower-bound expression on uplink SE of user k
transmission and the rest of τc − τp samples are utilized for data in cell j can be written as [22, Th. 4.4]
transmission.
τc − τp
SEUL jk ),
log2 (1 + SINRUL
As for channel model, this paper considers the spatially corre- jk = (6)
lated Rayleigh fading channels. The channel vector between τc
( user
k in cell l and BS in cell j, is modeled as hjlk ∼ C N 0, R jlk ,
)
where the effective signal-to-interference-plus-noise ratio (SINR)
where R jlk ∈ CM ×M stands for the positive semi-definite spa- ⏐ { H }⏐2
qjk ⏐E ĥjjk hjjk ⏐
⏐ ⏐
tial covariance matrix describing the spatial correlation of the
non-line-of-sight (NLoS) components and large-scale fading vari-
UL
SINRjk = {⏐ ⏐2 } }⏐2 {⏐ ⏐ } .
⏐ H ⏐2
⏐ { H
∑L ∑K ⏐ H
n=1 qln E ⏐ĥjjk hj ln ⏐ − qjk ⏐E ĥjjk hjjk ⏐ + σ 2 E ⏐ĥjjk ⏐
⏐ ⏐ ⏐
ations [22]. l=1

2
Y. Xu, F. Liu, R. Du et al. Physical Communication 48 (2021) 101432

⏐ ⏐2
qjk pjk ⏐tr(R jjk (Q jjk )−1 R jjk )⏐
SINRMMSE
jk = ∑L ∑K ∑L ⏐2 (7)
qlk plk ⏐tr(R jlk (Q jjk )−1 R jjk )⏐ + tr(R jjk (Q jjk )−1 R jjk )σ 2
⏐
l=1 n=1 qln tr(R jln R jjk (Q jjk )−1 R jjk ) + l =1
l̸ =j

Box I.

Then, by using a technique proposed in [22], for the case of Proposition 2. For the no prior statistical information case, the
full covariance knowledge, the close-form expression of SINR can closed-form expression of NMSE can be given by
be written as in Box I. ∑L
Moreover, for the case of no prior statistical information, the tr( i̸ =j pik R jik + σ 2 I M )
ηjkLS = . (10)
closed-form expression of SINR can be obtained in the following pjk tr(R jjk )
proposition.
Proof. See Appendix.
Proposition 1. If the LS estimator in (3) is used, the SINR expression
in (6) can be written as Moreover, the closed-form NMSE expression for full covari-
ance knowledge case can be expressed as [22]
SINRLS
jk
⏐2 tr(R jjk − pjk R jjk (Q jjk )−1 R jjk )
ηjkMMSE = .
⏐
qjk pjk ⏐tr(R jjk )⏐ (11)
= ∑L ∑K ∑L ⏐2 tr(R jjk )
qlk ⏐tr(R jlk )⏐ + σ 2 tr(Q jjk )
⏐
l=1 n= 1 qln tr(R j ln Q jjk ) + pjk l=1
l̸ =j
Note that a matrix inversion is involved in (11), which is very
(8) difficult to be solved in the pilot power optimization problem. To
address this issue, the trace lemma is introduced as follows.
Proof. By following the same methodology as in [16], and after
some simple algebra, the (8) can be obtained. Lemma 1 ([23]). For any positive semi-definite matrices {A, B} ∈
tr(B)
It can be observed from (7) and (8) that the pilot power and CM ×M , it holds that tr(A−1 B) ≥ tr(A) .
data power are both handed in the expressions of SINRUL jk for
different cases. Thus, the SE of system can be heavily affected by By applying Lemma 1, the upper bound of (11) can be written
the pilot and data power allocation. A common way to enhance as
the SE is maximizing sum SE of system by considering both pilot pjk tr(R jjk R jjk )
and data power as variables. However, for spatially correlated η̃jkMMSE = 1 − ∑L . (12)
massive MIMO system, it is difficult to deal with such problem tr(R jjk )tr( l=1 plk R jlk + σ 2 I M )
since the (7) and (8) contain the inverse matrix computation
and matrix computation that involve the pilot power variable.
Particularly, by invoking the fact that pilots are only used for 3.2. Pilot power allocation with no prior statistical information
channel estimation, the pilot power should be allocated based
on channel estimation quality instead of sum SE, which can also For the case of no prior statistical information, a pilot power
help to increase the SE of system [11]. To this end, a disjoint pilot optimization problem can be formulated as
and data power allocation is designed to enhance the accuracy of L
∑ K
∑
channel estimation and the sum SE of system in this paper. minimize ηjkLS
{pjk } j=1 k=1 (13)
subject to 0 < pjk ≤ pmax ∀j, k.
3. Proposed pilot power allocation
where pmax denotes the maximum transmit pilot power at the
user.
In this section, the pilot power allocation schemes are de- By invoking the fact that R jjk is the positive semi-definite
veloped for both no prior statistical information case and full spatial covariance matrix and the pilot power pjk is non-negative.
covariance knowledge case to enhance the accuracy of channel Then, the non-convex optimization problem (13) can be regarded
estimation, which can also increase the SE of system. as a GP due to the fact that the objective function in (13) is a
posynomial [24], which can be reformulated to a convex prob-
lem and solved efficiently via employing the standard convex
3.1. NMSE optimization solvers such as CVX [25].

In order to measure the accuracy of channel estimation, the 3.3. Pilot power allocation with full covariance knowledge
NMSE metric is adopted, which can be used to perform pilot
power allocation [9]. To this end, the NMSE of user k in cell j is
For the case of full covariance knowledge, a pilot power opti-
defined as
mization problem can be written as
{ 2 } L K
∑ ∑
E hjjk − ĥjjk  η̃jkMMSE
 
minimize
ηjk = {  }
2
. (9) {pjk } j=1 k=1 (14)
E hjjk 
subject to 0 < pjk ≤ pmax ∀j, k.
3
Y. Xu, F. Liu, R. Du et al. Physical Communication 48 (2021) 101432

√
Recall that R jjk is the positive semi-definite matrix and pjk is By using Lemma 2 and assuming ρjk = qjk , the optimization
non-negative. Then, the problem (14) can be rewritten as problem (16) can be rewritten as [27]
L K L K
∑ ∑ pjk tr(R jjk R jjk ) ∑ ∑ wjk ejk 1
maximize ∑L maximize (− + log2 (wjk ) + )
l=1 plk R jlk + σ I M )
{pjk } tr(R jjk )tr( 2
j=1 k=1 (15) {wjk ≥0},{ρjk ≥0}, ln 2 ln 2
j=1 k=1
{cjk } (20)
subject to 0 < pjk ≤ pmax ∀j, k.
subject to ρ ≤ qmax 2
jk ∀j, k.
Note that the denominator of (15) is a posynomial function and
numerator function is monomial. This problem is also a GP op- To solve (20), an iterative algorithm is introduced to optimize
timization problem, which can be solved efficiently by CVX as {ρjk }, {cjk }, and {wjk }. Specifically, by using the optimal power
mentioned before. (t −1) (t )
solutions {ρjk }, the {cjk } in the tth iteration can be obtained
according to (18), i.e.,
4. Proposed data power allocation
(t )
ρjk(t −1) Re(tr(R jjk ))
In this section, the data power allocation schemes are designed cjk = ∑
L ∑K (t −1) 2 1 ∑L (t −1)
⏐2 ,
ρ tr(R jln Q jjk ) + l=1 (ρlk )2 ⏐tr(R jlk )⏐ +
⏐ 1
l=1 n=1 ( ln ) p pjk
tr(Q jjk )
with the optimal pilot power to maximize the sum SE of system. jk

To this end, the closed-form expression in (7) and (8) are used to (21)
formulate the proposed data power allocation for both no prior
∗
statistical information case and full covariance knowledge case. and for the tth iteration, the ejk defined in (19) can be expressed
as in Box II.
(t )
4.1. Data power allocation with no prior statistical information Then, the optimal solution of {wjk } in the tth iteration can
(t ) (t )
be written as wjk = 1
∗(t) . After obtaining the optimal {cjk } and
ejk
By neglecting the constant (τc − τp )/τc in (6), the data power
allocation for no prior statistical information case can be formu- {wjk(t ) } in the tth iteration, to obtain the optimal {ρjk(t ) } in the tth
lated as iteration, the optimization problem (20) can be simplified as
L K L K

wjk(t ) e(jkt )
∑ ∑ ∑ ∑
maximize log2 (1 + SINRLS
jk ) minimize
{qjk } (t )
{ }
j=1 k=1 (16) ρjk j=1 k=1 (23)
subject to 0 < qjk ≤ qmax ∀j, k. (t ) 2
subject to 0 < (ρjk ) ≤ qmax ∀j, k.
where qmax stands for the maximum data transmit power at a
where
user.
(t ) (t ) (t )
To solve this non-convex problem, an iterative algorithm is ejk = 1 − 2Re(ρjk cjk tr(R jjk ))
proposed to efficiently obtain the sub-optimal solutions. Specifi- L K
1
⏐ ⏐2 ∑
cally, by exploiting the MMSE detection to solve xjk from rjk in (5),
+ ⏐cjk(t ) ⏐ ( (t )
∑
(ρln )2 tr(R jln Q jjk )
⏐ ⏐
the detection problem can be written as cjk∗ = arg mincjk ejk , where pjk
{⏐ ⏐2 } l=1 n=1
ejk = E ⏐cjk rjk − xjk ⏐ is the mean square error (MSE) [26]. For L
∑ (t )
⏐2 1
(ρlk )2 ⏐tr(R jlk )⏐ + tr(Q jjk )).
⏐
the case of no prior statistical information, the closed-form of ejk +
can be rewritten as pjk
l=1
( L K L
(t )
⏐ ⏐2 ∑ 1 ⏐2 To find the optimal ρjk , the Lagrange function of optimization
∑ ∑ ⏐
ejk = ⏐cjk ⏐ qln tr(R jln Q jjk ) + qlk ⏐tr(R jlk )⏐
pjk (23) can be written as
l=1 n=1 l=1
)
L K
1 √ (t )
∑ ∑ (t ) (t ) (t )
+ tr(Q jjk ) + 1 − 2Re( qjk cjk tr(R jjk )) (17) L({ρjk }) = (wjk ejk ) + λjk ((ρjk )2 − qmax ) (24)
pjk
j=1 k=1

The optimal detection coefficient cjk∗ can be calculated by solving (t )

the first derivative ejk in respect of cjk to zero, then where λjk ≥ 0 is the Lagrange multiplier and note that (ρjk )2 ≤
qmax .
√ (t )
Re( qjk tr(R jjk )) Taking the partial derivatives of L(ρjk ) in respect of ρjk to zero,
cjk∗ = ∑ (t )
the optimal solution of ρjk can be obtained as
L ∑K ∑L ⏐ ⏐2
l=1 n=1 qln p1 tr(R jln Q jjk ) + l=1 qlk ⏐tr(R jlk )⏐ + 1
pjk
tr(Q jjk )
jk

(18) wjk(t ) Re(cjk(t ) tr(R jjk ))

∗
ρjk(t ) = ⏐ ⏐2 ⏐ ⏐2 ⏐
Substituting (18) into (17), the MMSE ejk can be given by ∑L ∑K (t ) ⏐ (t ) ⏐ ∑L (t ) ⏐ (t ) ⏐
⏐2
n=1 wln ⏐cln ⏐ tr(R jln Q j ln ) + l=1 wln ⏐cln ⏐ ⏐tr(R jlk )⏐ + λjk
1
l=1 pln
⏐2
(25)
⏐
qjk ⏐tr(R jjk )⏐
e∗jk = 1 − ∑ ⏐2
L ∑K ∑L
1
⏐ 1
l =1 n=1 qln pjk tr(R jln Q jjk ) + l=1 qlk ⏐tr(R jlk )⏐ + pjk
tr(Q jjk ) Moreover, the Lagrange multiplier λjk must satisfy the com-
(t )
(19) plementary slackness condition λjk ((ρjk )2 − qmax ) = 0. Since
(t ) (t ) (t )
the obtained {cjk }, {wjk } and {ρjk } are all optimal solution in
which is equal to (1 + SINRLS
jk )
−1
. tth iteration, the proposed iterative algorithm will converge to a
stationary solution [28]. Moreover, the procedure of the proposed
wb
Lemma 2 ([27]). Let f (w ) = − ln 2
+ log2 w+ ln12 and w be a positive solution is⏐ summarized in Algorithm 1. Particularly, only when
real number, then maximizef (w ) = − log2 b, where the optimal
⏐
⏐∑L ∑K (t ) ∑L ∑K (t −1)
satisfying ⏐ SEjk − SEjk ⏐ ≤ ε, where ε ≥
⏐
w>0 j=1 k=1 j=1 k=1
value w = ∗ 1
b
. 0 is a given error tolerance, Algorithm 1 will be stopped.
4
Y. Xu, F. Liu, R. Du et al. Physical Communication 48 (2021) 101432

(t −1) 2 2
(ρjk
⏐ ⏐
∗(t) ) ⏐tr(R jjk )⏐
ejk = 1 − ∑L (22)
∑K
ρ(t −1) 2 1 ∑L
ρ (t −1) 2 ⏐tr(R jlk )⏐2 +
⏐ ⏐ 1
l=1 n=1 ( ln ) p tr(R jln Q jjk ) + l=1 ( lk ) pjk
tr(Q jjk )
jk

Box II.

The iterative optimization method in Algorithm 1 can also be

used for (30) to obtain an optimal solution. At iteration t, the
optimization variables are updated as follows:
·cjk is updated as cjk(t ) = Re(ρjk(t −1) tr(R jjk (Q jjk )−1 R jjk ))×
( L K
∑ ∑ (t −1) 2
(ρln ) tr(R jln R jjk (Q jjk )−1 R jjk ) + tr(R jjk (Q jjk )−1 R jjk )
l=1 n=1
L
)−1
⏐tr(R jlk (Q jjk )−1 R jjk )⏐2
∑ (t −1) 2
(ρ
⏐ ⏐
+ lk ) pjk (31)
4.2. Data power allocation with full covariance knowledge l=1

·wjk is updated as wjk(t ) = 1

∗(t) where ejk
∗(t)
= 1 −
For the case of full covariance knowledge, the data power ejk
(t −1)
allocation optimization problem can be formulated as (ρ jk tr(R jjk (Q jjk )−1 R jjk ))2 ×
L K
∑ ∑
SINRMMSE
( L K
maximize log2 (1 + jk ) ∑∑ (t −1) 2
{qjk } j=1 k=1 (26) (ρln ) tr(R jln R jjk (Q jjk )−1 R jjk ) + tr(R jjk (Q jjk )−1 R jjk )
l=1 n=1
subject to 0 < qjk ≤ qmax ∀j, k. )−1
L
⏐tr(R jlk (Q jjk )−1 R jjk )⏐2
∑ (t −1) 2
(ρ
⏐ ⏐
Similar to no prior statistical information case, the ejk can be + lk ) pjk (32)
written as l=1
( L K
⏐ ⏐2 ∑ ∑ ·ρjk is updated as in Box III.
ejk = cjk
⏐ ⏐ qln pjk tr(R jln R jjk (Q jjk )−1 R jjk )
l=1 n=1 5. Numerical results
L
)
∑ ⏐ ⏐2
+ pjk tr(R jjk (Q jjk )−1 R jjk ) + qlk (pjk )2 ⏐tr(R jlk (Q jjk )−1 R jjk )⏐ In order to evaluate the performance of the proposed pilot
l=1 and data power allocation, the results of simulation compared
√ with the traditional method are presented. Specifically, a typical
+ 1 − 2Re( qjk pjk cjk tr(R jjk (Q jjk )−1 R jjk )). (27)
uplink massive MIMO system with 4 hexagonal cells is consid-
Then, taking the partial derivatives of ejk in respect of cjk to ered, where the BS is placed at center of each cell and the users
zero, the cjk∗ can be presented as are randomly distributed in one cell. Assuming that the system
√ uses 20MHz of bandwidth and the noise variance is -96 dBm.
cjk∗ = Re( qjk pjk tr(R jjk (Q jjk )−1 R jjk )) The large-scale fading coefficient βjlk can be calculated as βjlk =
( L
∑ K
∑ −148.1 − 37.6 log10 (djlk /1km) + zjlk [22], where djlk stands for the
× qln pjk tr(R jln R jjk (Q jjk )−1 R jjk ) + pjk tr(R jjk (Q jjk )−1 R jjk ) distance from user k in cell l to BSj, and zjlk refers to the shadow
l=1 n=1
fading with the standard derivation of 7. Moreover, the spatial
)−1 correlation matrix R jlk can be defined as [29]
L
∑ ⏐2
. ∗ ∗ M −1
⏐
qlk (pjk )2 ⏐tr(R jlk (Q jjk )−1 R jjk )⏐
⎡ ⎤
+ (28) 1 rjlk ··· (rjlk )
l=1 ⎢ ∗ M −2 ⎥
⎥
⎢rjlk 1 ··· (rjlk )
∗
Substituting (28) into (27), the ejk can be obtained as R jlk = βjlk ⎢. .. .. .. (34)
⎢ ⎥
⎢.. .
⎥
√ . . ⎥
e∗jk = 1 − ( qjk pjk tr(R jjk (Q jjk )−1 R jjk ))2 ×
⎣ ⎦
( L K (rjlk )M −1 (rjlk )M −2 ··· 1

where the correlation coefficient is defined as rjlk = µejθjlk , µ ∈

∑∑
qln pjk tr(R jln R jjk (Q jjk )−1 R jjk ) + pjk tr(R jjk (Q jjk )−1 R jjk )
l=1 n=1 [0, 1] denotes the correlation magnitude and θ jlk represents the
L
)−1 array boresight of the user incidence angle. The maximum pilot
∑ ⏐2 and data powers are assumed to be fixed to pmax = qmax =
.
⏐
+ qlk (pjk ) ⏐tr(R jlk (Q jjk )−1 R jjk )⏐
2
(29)
200mW for every user.
l=1
The numerical results will be compared as follows: (1) Equal
which is equal to (1 + SINRMMSE
jk )−1 . pilot and data power allocation (EPPA+EDPA); (2) Proposed data
Then, the optimization problem (26) can be reformulated as power allocation (DPA) with equal pilot power; (3) Proposed pilot
L K
∑ ∑ wjk ejk 1 and data power allocation (PPA+ DPA).
minimize ( − log2 (wjk ) + ) Fig. 1 shows the NMSE per user against the number of BS
{wjk ≥0},{ρjk ≥0}, ln 2 ln 2
j=1 k=1
(30) antennas when K = 6. It is clear that the proposed pilot power
{cjk } allocation can achieve higher performance of channel estimation
subject to ρ ≤ qmax 2
jk ∀j, k. than that of equal pilot power allocation for both LS and MMSE
5
Y. Xu, F. Liu, R. Du et al. Physical Communication 48 (2021) 101432

wjk(t ) cjk(t ) Retr(R jjk (Q jjk )−1 R jjk )

ρjk(t ) = ⏐ ⏐2 ⏐ ⏐2 ⏐ (33)
∑L ∑K (t ) ⏐ (t ) ⏐ ∑L (t ) ⏐ (t ) ⏐ ⏐ −1 R )⏐2
w w
⏐
c tr(R R (Q ) −1 R ) +
l=1 n=1 ln ⏐ ln ⏐ jln jjk jjk jjk l=1 lk ⏐clk ⏐ tr(R jlk (Q jjk ) jjk

Box III.

Fig. 3. Sum spectrum efficiency per cell against the number of BS antennas
when K = 6.
Fig. 1. The NMSE against the number of BS antennas.

Fig. 3 shows the sum spectrum efficiency per cell against

the number of BS antennas by using the LS and MMSE channel
estimator, where the number of users is K = 6. For the MMSE
channel estimator, the proposed pilot and data power allocation
can obtained much higher sum spectrum efficiency per cell than
that of only data power control as well as equal pilot and data
power allocation. This demonstrates that the proposed pilot and
data power allocation can achieve the best sum spectrum effi-
ciency, since it does not only considers the pilot power allocation
to enhance the accuracy of channel estimation, which is an ef-
fective way to indirectly increase the sum spectrum efficiency of
system, but also utilizes the data power allocation to directly en-
hance the sum spectrum efficiency of system. Similar to the case
of MMSE channel estimation, the proposed pilot and data power
allocation can achieve the best sum spectrum efficiency per cell
Fig. 2. Sum spectrum efficiency per cell against the number of iterations.
with the LS channel estimation. However, the performance gap
between the proposed pilot and data power allocation and only
data power allocation is small for the LS channel estimation.
channel estimator. Particularly, the performance gap between the Fig. 4 shows the sum spectrum efficiency per cell against
proposed pilot power allocation and equal pilot power allocation the number of users in each cell for the LS and MMSE channel
becomes larger for LS channel estimator. This means that the pro- estimator cases when M = 200. For the MMSE channel estimator
posed pilot allocation can significantly enhance the performance case, it is clear that the proposed pilot and data power allocation
of channel estimation by using LS channel estimator compare to can achieve higher sum spectrum efficiency per cell than that of
the MMSE channel estimation. In addition, the proposed pilot other schemes. On the other hand, the proposed pilot and data
power allocation can not only enhance the accuracy of channel power allocation can achieve the best sum spectrum efficiency
estimation, but also improve the performance of system. This per cell with the LS channel estimation. Similar to Fig. 3, the
will be demonstrated in the following subsection through the performance gap between the proposed pilot and data power
simulation results. allocation and only data power allocation is very small for the
In Fig. 2, the convergence of the proposed iterative data power LS channel estimation.
allocation algorithm in Section 4 is investigated for both the LS To further illustrate the effects of the proposed pilot and data
and MMSE channel estimator case when M = 100 and K = 6. power allocation, the cumulative distribution function (CDF) of
It is clear that the sum spectrum efficiency per cell tends to be sum spectrum efficiency per cell for both LS and MMSE channel
stable after 40 iterations. By using the MMSE channel estimation, estimator is shown in Fig. 5 when M = 400 and K = 8. In this fig-
the sum spectrum efficiency per cell at the stationary point is ure, the proposed power allocation is compare against the equal
increased by about 19% compared with the initial point. In con- power control scheme to demonstrate the importance of power
trast, for the LS channel estimation, the sum spectrum efficiency allocation. It is intuitive that the proposed pilot and data power
per cell at the stationary point is about 20% higher than that at allocation can obtain much better sum spectral efficiency for both
the initial point. Besides, the MMSE channel estimation gives the LS and MMSE channel estimator than that of equal power control
spectrum efficiency that is up to 11% higher than the LS at the scheme and only proposed data power allocation. Specifically, the
stationary point. performance of MMSE channel estimator is much better than LS
6
Y. Xu, F. Liu, R. Du et al. Physical Communication 48 (2021) 101432

allocation can achieve higher SE compared with traditional power

allocation for spatially correlated massive MIMO system.

CRediT authorship contribution statement

Yixin Xu: Conceptualization, Methodology, Software,

Resources, Writing – original draft, Writing – review & editing.
Fulai Liu: Conceptualization, Validation, Resources, Writing – re-
view & editing, Supervision. Ruiyan Du: Validation, Supervision.
Zhenxing Sun: Software, Resources, Writing – review & editing.

Declaration of competing interest

The authors declare that they have no known competing finan-

cial interests or personal relationships that could have appeared
Fig. 4. Sum spectrum efficiency against the number of users when M = 200. to influence the work reported in this paper.

Funding

This work was supported by the National Natural Science

Foundation of China under Grant No. 61971117, by the Natural
Science Foundation of Hebei Province, China under Grant No.
2020501007. All authors read and approved the final manuscript.

Availability of data and materials

Data sharing not applicable to this article as no datasets were

generated for analyzed during the current study.

Appendix. Proof of Proposition 2

Fig. 5. CDF of sum spectrum efficiency when M = 400 and K = 8. Proof. The numerator of (11) can be calculated as
{ 2 }
E hjjk − ĥjjk 
 
for all power allocation schemes since the channel estimation ⎧ 2 ⎫
errors for MMSE are always small. Moreover, the proposed data ⎪
⎨ L √  ⎪
∑ pik 1  ⎬
power allocation with equal pilot power still performs better =E  √ hjik + √ n p

than the equal pilot and data power allocation, which means that ⎩ i̸=j pjk pjk  ⎪
⎪
⎭
even without pilot power allocation, the proposed data power ⎧ ⎫
allocation can also enhance the spectrum efficiency of system. L √ L √
pik 1 p 1
⎨∑ ⎬
ik
∑
H
Particularly, the proposed power allocation has about 19% sum =E ( √ hjik + √ np ) ( √ hjik + √ np )
spectrum efficiency per cell improvement compared to equal ⎩ pjk pjk pjk pjk ⎭
i̸ =j i̸ =j
power control scheme with the MMSE channel estimation, which ⎧ ⎫
L
demonstrates that the proposed pilot and data power allocation pik 1
⎨∑ ⎬
can effectively improve the performance of system. In addition, = tr(E hjik hHjik + n p nH
p )
⎩ pjk pjk ⎭
for the case of LS channel estimation, the improvement of sum i̸ =j

spectrum efficiency per cell is only about 14% compared with L

∑ pik σ2
equal power control scheme. = tr( R jik + I M ). (35)
pjk pjk
i̸ =j

6. Conclusion And the denominator of (9) is given by

{  }
2
E hjjk  = E hHjjk hjjk = tr(R jjk ).
{ }
In this paper, the pilot and data power allocation schemes are (36)
investigated in multi-cell massive MIMO system over spatially
Then, substituting (35) and (36) into (9), the (10) can be obtained.
correlated Rayleigh fading channel. By considering the availabil-
ity of statistical CSI, the uplink closed-form SE expressions are
References
derived for both full covariance knowledge case and no prior
statistical information case. After that, the pilot power allocation
[1] T.L. Marzetta, Noncooperative cellular wireless with unlimited numbers
is first performed to enhance the accuracy of the channel esti- of base station antennas, IEEE Trans. Wirel. Commun. 9 (11) (2018)
mation. Then, the scheme of data power allocation is developed 3590–3600.
to maximize the sum spectrum efficiency of system by using [2] E.G. Larsson, F. Tufvesson, O. Edfors, T.L. Marzetta, Massive MIMO for next
the optimal pilot power obtained in the pilot power allocation generation wireless systems, IEEE Commun. Mag. 52 (2) (2014) 186–195.
[3] E. Björnson, E.G. Larsson, T.L. Marzetta, Massive MIMO: Ten myths and one
scheme. Finally, in order to solve this non-convex problem, an critical question, IEEE Commun. Mag. 54 (2) (2016) 114–123.
iterative optimization algorithm is designed for both two cases. [4] A. Gupta, R.K. Jha, A survey of 5G network: Architecture and emerging
Simulation results show that the proposed pilot and data power technologies, IEEE Access. 3 (2015) 1206–1232.

7
Y. Xu, F. Liu, R. Du et al. Physical Communication 48 (2021) 101432

[5] J.G. Andrews, et al., What will 5G be? IEEE J. Sel. Areas Commun. 32 (6) [26] Q. Shi, M. Razaviyayn, Z.Q. Luo, H. Chen, An iteratively weighted MMSE
(2014) 065–1082. approach to distributed sum-utility maximization for a MIMO interfering
[6] J. Jose, A. Ashikhmin, T.L. Marzetta, S. Vishwanath, Pilot contamination and broadcast channel, IEEE Trans. Signal Process. 59 (9) (2011) 4331–4340.
precoding in multi-cell TDD systems, IEEE Trans. Wirel. Commun. 10 (8) [27] B. Wang, L. Dai, Z. Wang, N. Ge, S. Zhou, Spectrum and energy efficient
(2011) 2640–2651. beam space MIMO-NOMA for millimeter-wave communications using lens
[7] T.C. Mai, H.Q. Ngo, M. Egan, T.Q. Duong, Pilot power control for cell-free antenna array, IEEE J. Sel. Areas Commun. 35 (10) (2017) 2370–2382.
massive MIMO, IEEE Trans. Veh. Technol. 67 (11) (2018) 11264–11268. [28] Q. Zhang, Q. Li, J. Qin, Robust beamforming for nonorthogonal multiple-
[8] H.V. Cheng, E. Björnson, E.G. Larsson, Optimal pilot and payload power access systems in MISO channels, IEEE Trans. Veh. Technol. 65 (12) (2016)
control in single-cell massive MIMO systems, IEEE Trans. Signal Process. 10231–10236.
65 (9) (2017) 2363–2378. [29] S. Loyka, Channel capacity of MIMO architecture using the exponential
[9] P. Liu, S. Jin, T. Jiang, Q. Zhang, M. Matthaiou, Pilot power allocation correlation matrix, IEEE Commun. Lett. 5 (9) (2001) 369–371.
through user grouping in multi-cell massive MIMO systems, IEEE Trans.
Commun. 65 (4) (2017) 1561–1574.
[10] O. Saatlou, M.O. Ahmad, M.N.S. Swamy, Joint data and pilot power allo-
cation for massive MU-MIMO downlink TDD systems, IEEE Trans. Circuits Yixin Xu was born in Hebei, china, in 1981. He received
Syst. II 66 (3) (2019) 512–516. the M.S. degree from Yanshan University, Qinhuang-
[11] Hieu Trong Dao, S. Kim, Disjoint pilot power and data power allocation dao, china, in 2011. He is currently pursuing the
in multi-cell multi-user massive MIMO systems, IEEE Access 6 (2018) Ph.D. degree in signal and information processing with
66513–66521. Northeastern University. His research interest includes
[12] X. Gao, O. Edfors, F. Rusek, F. Tufvesson, Massive MIMO performance eval- signal processing for communications and Massive
uation based on measured propagation data, IEEE Trans. Wirel. Commun. MIMO.
14 (7) (2015) 3899–3911.
[13] J. Zhang, C.K. Wen, S. Jin, X. Gao, K.K. Wong, On capacity of large-
scale MIMO multiple access channels with distributed sets of correlated
antennas, IEEE J. Sel. Areas Commun. 31 (2) (2013) 133–148.
Fulai Liu was born in Hebei, China, in 1975. He
[14] H. Tataria, P.J. Smith, L.J. Greenstein, P.A. Dmochowski, M. Matthaiou,
received the M.S. degree and the Ph.D. degree from
Impact of line-of-sight and unequal spatial correlation on uplink MU-MIMO
Northeastern University, Shenyang, China, in 2002 and
systems, IEEE Wirel. Commun. Lett. 6 (5) (2017) 634–637.
2005, respectively. Since 2010, he is a professor at
[15] P. Li, Y. Gao, Z. Li, D. Yang, User grouping and pilot allocation for spatially
Northeastern University, Qinhuangdao, China. His re-
correlated massive MIMO systems, IEEE Access 6 (2018) 47959–47968.
search interests include array signal processing and its
[16] Ö. Özdogan, E. Björnson, E.G. Larsson, Massive MIMO with spatially
applications, cognitive radio, etc.
correlated rician fading channels, IEEE Trans. Commun. 67 (5) (2019)
3234–3250.
[17] X. Xia, K. Xu, D. Zhang, Y. Xu, Y. Wang, Beam-domain full-duplex massive
MIMO: Realizing co-time co-frequency uplink and downlink transmission
in the cellular system, IEEE Trans. Veh. Technol 66 (10) (2017) 8845–8862. Ruiyan Du was born in 1978. She received her B.S.
[18] K. Xu, Z. Shen, Y. Wang, X. Xia, D. Zhang, Hybrid time-switching and power degree from Hebei Normal University, Shijiazhuang,
splitting SWIPT for full-duplex massive MIMO systems: A beam-domain China, in 1999, the M.S. degree from Yanshan Uni-
approach, IEEE Trans. Veh. Technol. 67 (8) (2018) 7257–7274. versity, Qinhuangdao, China, in 2006, and the Ph.D.
[19] Z. Shen, K. Xu, X. Xia, W. Xie, D. Zhang, Spatial sparsity based secure degree from Northeastern University, Shenyang, China,
transmission strategy for massive MIMO systems against simultaneous in 2012. She is an associate professor at Northeastern
jamming and eavesdropping, IEEE Trans. Inf. Forensics Secur. 15 (2020) University, Qinhuangdao, China. Her research interests
3760–3774. include wireless communications, signal processing for
[20] K. Xu, Z. Shen, Y. Wang, X. Xia, Location-aided mMIMO channel tracking communications.
and hybrid beamforming for high-speed railway communications: An
angle-domain approach, IEEE Syst. J. 14 (1) (2020) 93–104.
[21] M. Matthaiou, P.J. Smith, H.Q. Ngo, H. Tataria, Does massive MIMO fail in Zhenxing Sun received B.S. degree and M.S. degree
Ricean channels? IEEE Wirel. Commun. Lett. 8 (1) (2019) 61–64. from Northeast Petroleum University, Daqing, China,in
[22] E. Björnson, J. Hoydis, L. Sanguinetti, Massive MIMO networks: Spectral, 2003 and in 2011, respectively. He has been a doctoral
energy, and hardware efficiency. IFound, Trends Signal Process. 11 (2017) student in Northeastern University, Shenyang, China,
154–655. since 2015. Currently, he is an associate professor at
[23] S. Wang, M. Wu, Z. Jia, Matrix Inequality, Science Press, 2006, pp. 130–131. Department of Electronics and Information Engineering
[24] M. Chiang, Geometric Programming for Communication Systems. Hanover, Northeast Petroleum University at Qinhuangdao, China.
MA, USA, 2005, Now. His research interesting includes interference man-
[25] S. Boyd, L. Vandenberghe, Convex Optimization, Cambridge University agement technique in Ultra-Dense Networks, network
Press, 2004. coding in full duplex relay networks.