Received March 6, 2018, accepted April 10, 2018, date of publication April 19, 2018, date of current version

May 9, 2018.
Digital Object Identifier 10.1109/ACCESS.2018.2828403

Cooperative User Scheduling in

Massive MIMO Systems
XIANG CHEN1,2 , FENG-KUI GONG 1 , (Member, IEEE), HANG ZHANG2 ,
AND GUO LI1 , (Student Member, IEEE)
1 State Key Laboratory of ISN, Xidian University, Xi’an 710071, China
2 Science and Technology on Information Transmission and Dissemination in Communication Networks Laboratory, Shijiazhuang 050081, China
Corresponding author: Feng-Kui Gong ([email protected])
This work was supported in part by the National High-tech Research and Development Program of China under Grant 2014AA01A704,
in part by the Joint Fund of the Ministry of Education of China under Grant 6141A02022338, and in part by the Opening Project of Science
and Technology on Communication Networks Laboratory under Grant KX162600027.

ABSTRACT Taking advantage of distributed computation and the characteristic of device-to-device (D2D)
communication among users, a cooperative user scheduling (CUS) scheme is first proposed. Through
reducing the number of users which feedback their channel state information (CSI) to the base station,
the problem of huge feedback in massive multiple-input multiple-output (MIMO) systems can be, to a
large extent, solved. The property of D2D allows users to exchange information, so the users can calculate
their correlation coefficients with the selected best user locally. Then, these coefficients are used to filter
out these strongly correlated users without greatly affecting the system sum rate. Specially, to decrease
the computational complexity and reduce the number of feedback users, the cell users are divided into
several groups randomly. The users in the same group can determine by themselves, which need to feedback
their CSIs, so the global optimal selection problem is decomposed into several local optimization problems.
Through theoretical analysis on the CUS scheme, we obtain the lower bound expression of system sum rate.
Simulation results indicate that the uplink feedback resources can be greatly saved with the CUS scheme
while the proposed scheme’s influence on the system sum rate is negligible.

INDEX TERMS Massive MIMO, cooperative user scheduling, D2D, user filtering, huge feedback.

I. INTRODUCTION cell to select the optimal user set to maximize the system
Recently, the next-generation wireless communication sum rate. For example, in the traditional MIMO systems,
becomes the focus of attention, worldwide researchers Yoo and Goldsmith [5] propose a low-complexity semi-
have achieved numerous achievements on the fifth- orthogonal user selection (SUS) algorithm based on the
generation (5G) wireless communication. Since proposed, exact CSI. The SUS algorithm iteratively selects the user with
massive antenna array is regraded as one of the key tech- the greater channel norm and the correlation coefficient. The
niques in the 5G communication. In massive multiple-input SUS algorithm can effectively decrease the computational
multiple-output (MIMO) systems, the base station (BS) is complexity, while it needs the exact CSIs of all the users.
equipped with dozens or even hundreds of antennas to The limitation makes the algorithm difficult to be applied
serve multiple users. Through deeply developing the wireless to massive MIMO systems. Since, in the frequency divi-
resources in the spatial dimension, the large-scale antenna sion duplexing (FDD) mode, the feedback load is positively
array has the potential to significantly improve the spectral correlated with the numbers of the BS antennas and users,
efficiency and power efficiency. However, the technique the huge amount of feedback would occupy a great part
also faces many challenges, such as pilot contamination, of spectral resources, which, in turn, decreases the system
the channel modelling and the huge feedback in user spectral efficiency when the numbers of the BS antennas
scheduling [1], [2]. and users increase sharply. In addition to the huge feedback
In many user scheduling algorithms [3], [4], the BS overhead, it is also extremely difficult for the BS to select
needs the channel state information (CSI) of all the users in the optimal user set from a large number of users. Nowadays,

2169-3536
2018 IEEE. Translations and content mining are permitted for academic research only.
21910 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 6, 2018
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
X. Chen et al.: CUS in Massive MIMO Systems

many researchers have come up with a great variety of solu- systems. According to the performance analysis, we con-
tions to ease the problems of high computational complexity clude that the cooperative user scheduling scheme can
and huge feedback [3], [6]–[10]. Razi et al. [3] propose a reduce the computational complexity at the BS and the
greedy user scheduling algorithm based on the rate allocation system feedback overhead without greatly decreasing the
in vector perturbation precoding systems, which reduces the sum rate performance. The main contributions of this paper
computational complexity through removing the insignificant include:
users from the candidate user set. Considering that the prin- • To ensure the low feedback and efficient transmission
ciple of user scheduling is similar with that of antenna selec- in massive MIMO systems, we propose a cooperative
tion, Benmimoune et al. [6] propose a joint strategy which user scheduling scheme based on the cooperation among
simultaneously performs antenna selection and schedules the users. Through filtering out the non-significant users
optimal users with the objective of maximizing the system and simplifying the candidate user set before CSI feed-
sum rate and reducing the computational complexity of user back, the cooperative user scheduling (CUS) scheme can
scheduling and antenna selection. Owing to the limited space greatly reduce the number of feedback users. As far as
of the BS antenna array, the adjacent antennas are unavoid- we know, it is the first time to apply the concept of
ably correlated with each other when the number of antennas ‘‘user filtering’’ in user scheduling. Besides the advan-
increases sharply. Lee et al. [7], [8] utilize the correlations tage of reducing the system feedback, the user filtering
among the adjacent antennas to group the correlated antennas, is accomplished at users, so the proposed scheme can
thus the dimensionality of the antennas at the BS and the reduce the search complexity at the BS. Furthermore,
feedback overhead are reduced. Similarly, taking advantage through combining SUS algorithm with the user filter-
of the correlations among the users to group the users with ing, the CUS scheme can achieve excellent sum rate
the similar covariance eigenvectors [9], [10], the joint spa- performance through selecting the appropriate user set
tial division and multiplexing (JSDM) algorithm can reduce from the permitted feedback users.
the dimensionality of the effective channels, simplify the • To avoid occupying huge system resources during the
system operations, and reduce the system feedback through feedback phase, we design an efficient CSI feedback
randomly selecting part of users to feedback their CSIs. strategy to improve the system spectral efficiency.
For the articles above-mentioned, the user scheduling, By utilizing the cooperation among users, each user
the compression and recovery of CSI and the user grouping can obtain the correlation coefficient with the strongest
are accomplished at the BS, which would increase the feed- user, which is found through a timer. Through compar-
back overhead of the uplink channel and the computational ing these coefficients with a filtering threshold, each
complexity at the BS. If part of the task of user scheduling user can locally determine which users can feed back
is allocated to users, the computational complexity at the their CSIs.
BS and the feedback load would be effectively decreased. • Through analyzing the probability that each user is
Motived by the above viewpoints, we propose a cooperative selected and the probabilities under different numbers of
user scheduling scheme for Device-to-Device (D2D) commu- selected users, we obtain the lower bound of the system
nication systems [11]. The key of the proposed scheduling sum rate of the proposed CUS scheme. According to the
scheme is to utilize the D2D communication among users to simulation results, we verify the validity of the lower
filter out those non-significant users and decrease computa- bound of sum rate under different antenna numbers, and
tional complexity at the BS via the distributed computation. conclude that the lower bound would get tighter when
The proposed cooperative scheduling scheme consists of two the number of BS antennas increases.
phases: user filtering and user scheduling. During the user The remaining parts of this paper are organized as follows.
filtering phase, the cell users are divided into several groups In Section II, a multi-user MIMO system is considered and
firstly. Through a timer, the user with the best channel condi- the section also quantifies the effect of zero-forcing (ZF)
tion in each group is found, then the optimal user broadcasts beamforming on effective signal to interference plus noise
its CSI to the other users in the same group. Thus, the residual ratio (SINR). In Section III, the proposed cooperative user
users can calculate their correlation coefficients with the best scheduling scheme is discussed in detail. Section IV gives
user locally. Only the user with the best channel condition an elaborate theoretical analysis on the sum rate of the
and the users whose correlation coefficients are greater than scheduling algorithm. The simulation results are shown in
the filtering threshold can feed back their CSIs to the BS. Section V and the final concluding remarks are provided in
During the second phase, the BS selects the user set through Section VI.
the popular user scheduling algorithms. Considering the sim- As for notations, the uppercase boldface and lowercase
ilarity that both SUS algorithm [5] and the proposed user boldface are used to denote matrices and vectors. (·)H denotes
filtering algorithm simplify the candidate user set based on the conjugate transpose of a matrix or vector. k·k denotes the
the correlation coefficients, we combine SUS algorithm with Euclidean vector norm. card(·) is the number of the elements
user filtering for maximizing the system sum rate. in set. E (·) denotes the mathematical expectation operator.
In this paper, we analyze the system sum rate and feed- Ak,j denotes the element in the k-th row and j-th column of
back overhead of the proposed scheme in massive MIMO matrix A.

VOLUME 6, 2018 21911

 X. Chen et al.: CUS in Massive MIMO Systems

II. SYSTEM AND CHANNEL MODELS interfering the k-th user’s receiving signal, and n ∼ CN(0, I),
A single cell massive MIMO broadcast system is considered I ∈ CK ×K denotes the unitarymatrix. The power constraint
in this paper, where a BS transmits space information to N for the transmitted signal is E xH x = P̄T . Since the noise
users, as shown in Fig. 1. In the system, the BS is equipped variance is unit, P̄T also is the transmitting signal-to-noise
with Nt transmitting antennas, and each user is equipped ratio (SNR).
with one antenna. We consider frequency division duplex- In order to effectively eliminate the interference between
ing (FDD) in this paper, where the downlink and uplink users, the original signal s needs to be preprocessed. We adopt
channels occupy different frequency bandwidths, and the BS the zero-forcing beamforming (ZFBF) transmitter [13] for
obtains the downlink CSI via the uplink feedback channel. data preprocessing, so the post-processed signal x can be
It is assumed that the downlink (DL) channel is quasi-static expressed as
Rayleigh block fading channel, which means that the chan-
K q
nel condition changes from one frame until the next frame X
x= P̄k wk sk , k ∈ S, (2)
but remains constant within one frame. The uplink (UL)
k=1
channel is assumed to be perfect and free from the channel
noise. where sk is the original transmitting signal of the k-th user,
and s = [s1 , s2 , . . . , sk , . . . , sK ]T . wk ∈ CNt ×1 is the zero-
forcing beamforming vector of the k-th user. P̄k is the power
PK
allocated to the k-th user, and P̄k = P̄T . Let W̃ =
k=1
−1
HH HHH = w̃1 , w̃2 , . . . , w̃K [14] be the beamforming

 

matrix. For satisfying the power constraint, the beamforming

vector of each user has to be normalized, i.e., wi = kw̃ i
w̃i ki
,
h
so W = [w1 , w2 , . . . , wK ] = kw̃ 1
, w̃2 , . . . , kw̃
w̃1 k kw̃2 k
K
w̃K k
.
Thus, the k-th user’s received signal yk can be denoted
as
q X K q
yk = P̄k hk wk xk + hk P̄j wj xj + nk , k, j ∈ S, (3)
j=1,
j6=k

where nk is the noise received by the k-th user. According

FIGURE 1. System model of the massive MIMO system.
to (3), the effective receiving SINR of the k-th user is
P̄k |hk wk |2
As shown in Fig. 1, the BS transmits the data x to these γk = . (4)
PK  2
users belonging to the user set S. All the scheduled users’ 1+ P̄j hk wj
 
received data y can be expressed as j=1,
j6=k
 
h1 
n1

Without loss of generality, equal power allocation is assumed,
 h2  i.e., P̄k = P̄KT .
 
 n2 
 
 . 
 
 ..   . 
 
.
.  , k ∈ S,

y = Hx + n =  x +  (1) III. COOPERATIVE USER SCHEDULING (CUS) SCHEME
 
 hk  
 nk 

  In this section, we would introduce the proposed cooperative
 .   .. 
 
 .  user scheduling scheme in detail. The CUS scheme is divided
 .   . 
into two parts: user filtering based on D2D communication
h nK
K and user scheduling.
where y ∈ CK ×1is the received signal, in which K is the
number of the selected users in the user set S, that is, K = A. USER FILTERING BASED ON THE
card(S) ≤ N .x ∈ CNt ×1 denotes the transmitted T signal from D2D COMMUNICATION
the BS. H = hT1 , hT2 , . . . , hTk , . . . , hTK denotes the channel The first step of the proposed CUS scheme is to filter
gain matrix, in which hk ∈ C1×Nt is the channel condition out some users via the cooperation among users, which
vector of the k-th user. The j-th entry in hk is the chan- means that not all users in cell can feed back their
nel fading coefficient between the j-th BS antenna and the CSIs. Thus the scheme can decrease the system feedback
k-th user, and it obeys Rayleigh distribution. As [12] con- overhead.
sidered, the antennas of BS are independent. n ∈ CK ×1 For reducing the computational complexity and saving the
is the additive white noise vector, where nk is the noise hardware cost, we need to assign the ultra-high dimension

21912 VOLUME 6, 2018

X. Chen et al.: CUS in Massive MIMO Systems

matrix computation of BS to multiple user groups. Consid- flag packet also contains the best user’s channel state vector.
ering that most of the existing grouping algorithms have All the users in the same group, while waiting for their own
strong dependency on CSI or channel correlation informa- timers to reach zero, are in listening mode (each user cannot
tion [15]–[17], and need extra function evaluations, we adopt overhear the flag packet of the users in other groups). As soon
random user grouping algorithm [15] to ease the computa- as they hear one user in their own group flagging its presence
tional complexity and reduce the time delay at BS. Further- (the best user), they back off immediately. In this way, we can
more, In the proposed scheme, we cannot adopt the grouping easily find the user with the strongest channel in each group
algorithms which depend on the instant channel information, (the user with the greatest channel norm).
since the grouping is accomplished at the users before the The other users in each group, obtain the best user’s chan-
CSI feedback. The cell users are randomly divided into M nel vector hb . In the following parts, we assume that the
groups. For simplicity, it is assumed that the  N number
 of users index of the best user is b. Then they calculate the correla-
in each group is same, namely, NG = M . The users in tion coefficients between their own channel vectors and the
different groups work in the different time slots, so these b-th user’s channel vector, that is
groups are independent with each other, and there is no inter-  H
hs h 
group interference. αsb = b
, 1 ≤ s, b ∈ Z ≤ NG , s 6 = b, (6)
khs k khb k
Focusing on the optimal path selection in relay networks,
Bletsas et al. [18] propose a cooperative protocol to find the where s and b are the indices of the users which are in
best relay, and its idea of using a timer can be extended to the same group. Z denotes the integer set. Only the user
our scheme. Before feedback, each user is set a timer and with the best channel condition and the users satisfying the
the timer’s start point is directly correlated with the user’s filtering threshold can feed back theirs CSIs to the BS. That is,
CSI [18]. The timer of the user having the best end-to-end if αsb ≤ α, user s can feed back its CSI to the BS, otherwise,
channel condition will expire firstly. The start point of the user s is filtered out and will not communicate with the BS in
timer can be set as this frame. α is a given filtering threshold, which is same for
λ all the groups. Thus, after the user filtering, the user set can
Ts = , s = 1, 2, . . . , NG , (5) be denoted as
khs k
M
where Ts is an initial value of the s-th user’s timer. λ is a time
[
Sf = Sf ,m
constant. The start point Ts is inversely proportional to the m=1
channel state vector norm khs k, so the user with the strongest
M
(   H )
channel would expire firstly. Here, We assume that all the [ hs h 
s, b ∈ Gm  max khb k, αsb = b
≤ α, b 6= s ,

timers in the same group start at the same time. = 
b∈Gm khs k khb k
m=1
(7)
where Sf ,m denotes the permitted user set of the m-th group,
which can feed back its CSI to the BS. Gm is the initial user
set containing all the users of the m-th group before the user
filtering, and card(Gm ) = NG .

B. USER SCHEDULING
After the user filtering phase, the BS can get the most wanted
users’ CSIs from all the groups. Through combining with the
semi-orthogonal user scheduling algorithm [19], the appro-
priate user set can be selected from all the feedback users for
data transmission.
As shown in Tab. 1, nf denotes the total of the feed-
back users, M ≤ nf ∈ Z ≤ N . Sc,i and Si respectively
denote the candidate user set and the selected user set at the
FIGURE 2. User timer and packet broadcast in one group, the user that
expires firstly broadcasts a flag packet to other users in the same i-th iteration. π (i), i = 1, ..., Nt , denotes the index of the
group. selected user at the i-th iteration. Step 2 is carried out to select
the user with the strongest channel as the first scheduled
Based on the assumption that the users can estimate their user, and Step 3 is used for obtaining the projection vectors
CSIs exactly through channel estimation techniques, each of the candidate users on the orthogonal complement space
user calculates its timer start point according to (5) inde- of the selected user set. In Step 4, the candidate user with
pendently. As shown in Fig. 2, the timer of the user with the strongest projection vector joins in the selected user set.
the strongest channel will expire first. Then the best user Step 5 is used to simplify the candidate user set, and in this
transmits a short flag packet [18] to signal its presence. The step the correlation coefficients between the selected users

VOLUME 6, 2018 21913

 X. Chen et al.: CUS in Massive MIMO Systems

TABLE 1. Semi-orthogonal user scheduling algorithm in massive MIMO A. PROBABILITY ANALYSIS ABOUT THE DIFFERENT
systems.
SELECTED USER NUMBERS
1) USER FILTERING
After the user filtering, the probability Pnf that the remaining
user number is nf can be expressed as

M
1 Y
P{Nfilt = nf } = Pnf = P{Nm = nm } (11)
m=1
M  
Y NG − 1
NG −nm
= Pnfiltm −1 1 − Pfilt ,
nm − 1
m=1
(12)

where Pfilt denotes the probability that the correlation coef-

ficient between user s and the best user is less than α.
P {Nm = nm }, m = 1, 2, . . . , M , denotes the probability that
the remaining user number of the m-th group is nm , which can
be calculated as
 
NG − 1
P {Nm = nm } = Pnfiltm −1 (1 − Pfilt )NG −nm .
nm − 1
(13)

After the user filtering, the g-th (1 ≤ g ≤ M ) group’s user

set can be expressed as
(   )
and the candidate users are obtained through the projected  h hH 
s b
Wg = Sf ,g /b = s  ≤ α, s, b ∈ Gg . (14)

vector gn and g(i) . After the SUS scheduling phase, the BS  khs k khb k
obtains the final user set S for data transmission.
The probability that user s belongs to the the set Wg can be
IV. SUM RATE ANALYSIS OF THE COOPERATIVE described as [5]
USER SCHEDULING
In this section, we analyze the achievable sum rate of the Nt − j α 2
 
 1
proposed scheme with ZFBF transmitters. The sum rate of P sg ∈ Wg = Pfilt = F2j,2(Nt −j)
j 1 − α2
the massive MIMO system can be denoted as = Iα 2 (j, Nt − j) , (15)
Nt K
where sg denotes the user s in the g-th group. Fn,m (x) is
X X
PS (K ) E log2 (1 + SIN Rk )

CCUS = (8)
K =1 k=1
the cumulative density  function (CDF) of the F distribution.
Nt K Z ∞
Iz (a, b) = Bz (a, b) B (a, b) is the regularized incomplete
beta function, in which Bz (a, b) is the incomplete beta func-
X X
= PS (K ) log2 (1 + γk ) Pk (γk )dγk
K =1 k=1 0 tion, and B(a, b) is the complete beta function. Note that
(9) j = 1 in our case, since we only need to calculate the
Nt correlation coefficient with one vector. Due to the fact that
X
= PS (K ) Czf ,SUS (K ), (10) each group is i.i.d., we have, P {s1 ∈ W1 } = P {s2 ∈ W2 } =
K =1 · · · = P {sM ∈ WM } = Pfilt .
Let the number of all the remaining users, after the user
where PS (K ) is the probability that the number of the selected M
P
users is K . Pk (γk ) is the probability that the k-th user’s filtering, be nf , that is, nm = nf . (12) can be further
m=1
SINR is γk . As ZFBF can effectively suppress the interfer- denoted as
ence, SINRk = SNRk , so according to (4), we have γk =
P̄k |hk wk |2 M 
= P̄k |hk wk |2 . Czf ,SUS (K ) denotes the system

Y NG −1
NG −nm
K
1+ P̄j |hk wj |
P 2 Pnf = Pnfiltm −1 1−Pfilt
j=1,
nm −1
j6=k
m=1
sum rate with the ZFBF transmitters and the SUS algorithm. M  
nf −M Y NG −1
According to (10), the sum rate analysis is divided into two = Pfilt (1−Pfilt ) N −nf
. (16)
nm −1
parts, one is about PS (K ), the other is about Czf ,SUS (K ). m=1

21914 VOLUME 6, 2018

X. Chen et al.: CUS in Massive MIMO Systems

2) SEMI-ORTHOGONAL USER SCHEDULING (SUS) where, similarly, P{card(Sc,k−1 ) = ns,k−1 , s ∈ Sc,k−1 }

As Step 5 shown in Tab. 1, the probability that the absolute denotes the probability that, at the k − 1-th iteration, the can-
value of the correlation coefficient between two vectors is less didate users number is ns,k−1 and user s is in the candidate
than β is defined as the ‘‘survival probability’’. The ‘‘survival user set Sc,k−1 . The details of the final probability of each
probability’’ at the i-th iteration of SUS can be expressed user being selected are given in Appendix A.
as [19] As (19)-(22) shown, we can obtain the exact probability
of each user being selected at the first iteration of SUS
Pi (β) = 1 − (1 − β)Nt −i .
p
(17) algorithm, and the lower bound of the probability that user s is
selected at the k-th (k ≥ 2) iteration. Thus, according to (18),
The final probability Ps that each cell user is selected for we obtain the lower bound of the probability of each user
data transmission can be calculated as being selected.
Nt
X
Ps = Ps,i , (18)
i=1

where Ps,i , 1 ≤ i ≤ Nt , denotes the probability that user s is

selected at the i-th iteration.
The probability Ps,1 that user s is selected at the first
iteration of the SUS algorithm is
N M
1 X X
Ps,1 = Pf 1 , nm = nf , 1 ≤ nm ≤ nf , (19)
M
nf =M m=1

M N −1
 
G
Pnfiltm −1 (1 − Pfilt )NG −nm . Pf ,g1
Q
where Pf 1 = Pf ,g1
m=1, nm − 1
m6=g
denotes the probability that user s is the best user of the
g-th group and the remaining user number of the g-th group
FIGURE 3. The probability of each user being selected.
is ng . That is, Pf ,g1 = P(Ng = ng , s ∈ Sf ,g , s ∈
/ Wg ), it’s
details will be given in the Appendix A. The probability Ps,2
that user s is selected at the second iteration of the SUS Fig. 3 depicts the probability density function (PDF) that
algorithm is expressed as each user is selected. In this simulation, the BS is equipped
with 4 antennas, and there are 20 users which are divided
Ps,2 = Ps,21 + Ps,22 , (20) into 2 groups. The channels between the BS and the users
obey Rayleigh fading. The filtering threshold α is 0.5, and the
where Ps,21 denotes the probability that user s is selected correlation coefficient β is in the range of (0, 0.45). In reality,
at the second iteration and it is the best user of the for achieving a better scheduling user set, the parameters
g-th group. Ps,22 denotes the probability that user s is selected α and β should be smaller than the values given in the
at the second iteration and it is not the best user of the g-th simulation (it will be proved in Section V), which means that
group. The probability Ps,3 that user s is selected at the third the selected users would be less correlated. Here, we just
iteration of the SUS algorithm is consider the conventional antenna array instead of massive
ns,1 − 1 MIMO systems due to its huge computational complexity of
Ps,3 ≥ P{card(Sc,1 ) = ns,1 , s ∈ Sc,1 } Fig. 3 and Fig. 4.
ns,1 ns,2
  Just as Fig. 3 shown, we obtain a lower bound of the proba-
ns,1 − 2
· P2 ns,2 (β)(1 − P2 (β))ns,1 −ns,2 −1 , bility that each user is selected, which coincides with the anal-
ns,2 − 1
ysis of (18)-(22). The theoretical lower bound has an excellent
(21) performance. Especially, when the coefficient efficient β is
in the range of (0, 0.33), the result of theoretical analysis is
where P{card(Sc,1 ) = ns,1 , s ∈ Sc,1 } denotes the probability
almost equivalent to the simulation result. In reality, when
that, at the first iteration of SUS algorithm, the candidate user
there are thousands of users in cell, the optimal value of β is
number is ns,1 and user s is in the candidate user set Sc,1 . ns,2
just in this interval (0, 0.33) [5], thus the theoretical lower
is the candidate user number at the second iteration of SUS
bound (18) that each user is selected can be regard as the
algorithm.
actual probability.
The probability Ps,k , 3 ≤ k ≤ Nt that user s is selected at
Based on the probability of each user being selected,
the k-th iteration of the SUS algorithm is
we give a probability analysis about the different selected user
1 numbers after the SUS algorithm. According to (16) and (42),
Ps,k ≥ P{card(Sc,k−1 ) = ns,k−1 , s ∈ Sc,k−1 }, (22)
ns,k−1 the probability PS (1) of the selected user number being one

VOLUME 6, 2018 21915

 X. Chen et al.: CUS in Massive MIMO Systems

number being ns,k can be expressed as

P{card(Sc,k ) = ns,k }
≈ P{card(Sc,k−1 ) = ns,k−1 }
 
ns,k−1 − 1 n
· Pk s,k (β)(1 − Pk (β))ns,k−1 −ns,k −1 .
ns,k
(26)

Here, we obtain an approximated result based on the assump-

tion that each user is i.i.d. Since the relationships would
become weak and complicated after the second iteration due
to the projection in Step 3 of Tab. 1. According to (24), (25)
and (26), we can obtain the probability PS (k) of the selected
user number being k, 2 ≤ k ≤ Nt .
FIGURE 4. The probability analysis of different selected user numbers. Fig. 4 depicts the PDF that the numbers of the final selected
users are one, two, three, and four. The simulation conditions
are same as those of the Fig. 3. As seen from Fig. 4, an exact
can be calculated as result is obtained when the selected user number is one. When
the user numbers are two, three or four, we only obtain the
N M M
1 X X
n −1 Y approximated results, while the theoretical results are very
PS (1) = Pnf 1 − P1,c i {1 − P1 (β)}nm ,
M close to the simulation results. Concluding from Fig. 3 and
nf =M i=1 m=1,
m6=i Fig. 4, we can verify the correctness of the theoretical analysis
(23) about the probability of different selected numbers, which
accomplishes the first part of (10).
Following this, the probability PS (k) of the selected user
number being k, 2 ≤ k ≤ Nt can be expressed as B. SUM RATE ANALYSIS
PS (k) ≈ P{card(Sc,k−1 ) = ns,k−1 }{1−Pk (β)} ns,k−1 −1
, (24) Theorem 1: The lower bound of the sum rate of the pro-
posed CUS algorithm can be denoted as
where P{card(Sc,k−1 ) = ns,k−1 }, ns,k−1 ≥ 1 represents the Nt
probability that the candidate user number at the k − 1-th
X
CCUS = PS (K )Czf ,SUS (K ) (27)
iteration is ns,k−1 , in which P{card(Sc,1 ) = ns,1 } can be cal- K =1
culated as
Nt

X eK P̄T
K Nt −K +2
≥ log(e) PS (K ) (28)


P card Sc,1 = ns,1 t −K +1
K =1 P̄N
T
N
1 X −K +1 
Nt X n 
= Pnf

P̄T K
M
nf =M
· 0 n − Nt + K − 1, ,
K P̄T
n=1
M
·
X
P card(Sc,1m ) = ns,11 P card(S̄c,1m ) = ns,12 ,
  R∞
where 0(a, x) = t a−1 e−t dt is the upper incomplete gamma
m=1 x
function [13].
(25) Proof: The effective SINR of user k, 1 ≤ k ≤ K with

where P card(Sc,1m ) = ns,11 denotes the probability that ZFBF can be expressed as
the m-th group’s remaining user number is ns,11 , and P̄k P̄T
γk = h i ,
 
 nm −1 ns,11
n −n −1 i = h (29)
P card(Sc,1m ) = ns,11 = P1,c 1 −P1,c m s,11 . (HHH )−1 K (HHH )−1
 n s,11 kk kk
The probability P card(S̄c,1m ) = ns,12 denotes that the
where K is the total number of the effective users. Set
other groups’ (except the m-th group) remaining user num-
ber is ns,12 , and the probability  can
 be expressed as Z = HHH , (30)
nf −nm ns,1
P1 2 (β) (1 − P1

P card(S̄c,1m ) = ns,12 =
ns,12 where Z is termed as a complex Wishart distribution [13], and
(β))nf −nm −ns,12 . We have ns,11 + ns,12 = ns,1 , 1 ≤ ns,1 ≤ Z ∼ WK (Nt , 3), 3 = I.
nf , Sc,1m ∪ S̄c,1m = Sc,1 . At the k-th iteration (2 ≤ k ≤ According to [20], γk in (29) is a Chi-squared distribution
Nt ), the probability P{card(Sc,k ) = ns,k } of the candidate user random variable with degrees of freedom 2(Nt −K +1), so the

21916 VOLUME 6, 2018

X. Chen et al.: CUS in Massive MIMO Systems

PDF of γk can be calculated as to (10), (24) and (34), which can be calculated as
Nt
K e−K γk /P̄T
 Nt −K
K
X
fγk (γk ) = γk , γk ≥ 0. (31) CCUS = PS (K )Czf ,SUS (K )
P̄T (Nt − K )! P̄T K =1
Nt
Note that the constraint Nt ≥ K (the constraint of Wishart X eK /P̄T K Nt −K +2
≥ log(e) PS (K )
distribution) is satisfied because of the SUS algorithm. P̄N t −K +1
K =1 T
Given the user number K , the sum rate [21]–[23] with the −K +1 
Nt X n  
ZFBF transmitters when the cell users are independent is P̄T K
· 0 n − Nt + K − 1, .
∞
K P̄T
K K Z n=1
(35)
X X
Czf (K ) = E{log(1+γk )} = fγk (γk ) log(1+γk )dγk
k=1 k=1 0 So we obtain the sum rate expression in Theorem 1. Here,
Z∞ we give a detailed performance analysis of the proposed algo-
rithm step by step. Yoo and Goldsmith [5] just give an asymp-
=K fγk (γk ) log(1 + γk )dγk
totic sum rate expression when the users number approaches
0 infinity, and do not give an exact probability analysis about
eK /P̄T K Nt −K +2 the selected user number. Except obtaining the probabilities
= log(e) of different selected user numbers, our formulation obtains
t −K +1
P̄N
T an elaborate lower bound of the sum rate. And the theoretical
−K +1 
Nt X n   lower bound would get closer to the simulation results as the
P̄T K
· 0 n − Nt + K − 1, , (32) BS antenna number increases.
K P̄T
n=1

where (32) is obtained from the integral identity [13]

Z∞
log(1 + t)e−µt t n−1 dt
0
n
0(i − n, µ)
= (n − 1)!eµ
X
, n = 1, 2, . . . . (33)
µi
i=1

Then, we will consider the effect of correlation on the sum

rate. The sum rate Czf (K ) in (32) is obtained based on the
assumption of independent users. However, in the proposed
CUS scheme, users are selected based on the correlation coef-
FIGURE 5. The sum rate comparisons between the simulation results and
ficients. The selected users are semi-orthogonal with each the theoretical results.
other instead of being randomly selected. So, the elements off
the diagonal of Z in our scheme are smaller compared with Fig. 5 depicts the sum rate comparisons of the simulation
those of independent users. results and theoretical results. In the simulation, there are
Given Z0k,k = Zk,k and Z0k,j > Zk,j , k 6 = j, we have 32 users, they are randomly divided into two groups. For ana-
 
Z0 −1 > Z−1 k,k . Thus, according to (29), the effective


lytical tractability, the numbers of the BS transmit antennas
k,k
γk would be greater after the SUS procedure. So, are set equal to 4, 8, and 16. The filtering threshold α and
correlation coefficient β are 0.3 and 0.25. The simulation
Czf ,SUS (K ) ≥ Czf (K ). (34) results of the sum rate are obtained over 105 independent
channel realizations. Concluding from Fig. 5, we obtain the
Thus, we obtain the lower bound of the sum rate with the SUS lower bound of the system sum rate, which coincides with the
algorithm and the ZFBF transmitters, which accomplishes analysis in (35).
the second part of (10). According to the known channel- As shown in Tab. 2, given the receiving SNR, the normal-
hardening [24] property in massive MIMO systems, as the ized relative differences between the simulation results and
number of the BS antennas increases sharply, the different the theoretical results decrease obviously with the increase of
users’ channels are nearly orthogonal. Thus the non-diagonal antenna number, in which the normalized relative difference
|C −C |
elements in Z would get smaller further, and the lower bound is calculated as simCsimtheory , where Csim and Ctheory respec-
will get tighter. Combining the sum rate Czf ,SUS and the tively denote the simulation sum rate and the theoretical sum
probability of different selected user numbers, we obtain the rate. Concluding from Tab. 2, the lower bound would get
lower bound of the sum rate of the CUS scheme according tighter when the antenna number increases, which coincides

VOLUME 6, 2018 21917

 X. Chen et al.: CUS in Massive MIMO Systems

TABLE 2. The normalized relative differences between simulation results and theoretical results.

with our previous analysis. So it is reasonable to believe that given the number of groups, the filtering threshold and the
the theoretical analysis would obtain a closer approximation, correlation coefficient, the proposed CUS scheme only has
when the theoretical results are applied in massive MIMO slight sum rate degradation compared with the full feedback
systems. According to Fig. 3, Fig. 4, and Fig. 5, we can SUS algorithm. Especially, with α = β = 0.22, the sum rate
verify the correctness of the theoretical lower bound of sum degradation is nearly negligible. Benefited from filtering out
rate (35). the non-significant users, the proposed scheme can greatly
In the performance analysis, we give the figures about the save the system feedback resources. The details of reducing
theoretical results and simulation results in the conventional the feedback user number are depicted in Tab. 3.
MIMO systems. Since the computation complexity of the
theoretical analysis would have an exponential growth when TABLE 3. The average numbers of transmitting users and filtering users
with different filtering thresholds α.
the number of antennas increase sharply due to the step-
by-step theoretical analysis. While, as stated in our paper,
we analyze the results in the massive MIMO systems, and
conclude that, in the massive MIMO scenario, the theoretical
results would get closer to the simulation results. That is,
in the massive MIMO, the theoretical results would obtain
a better performance. Except that, in the Section V, we give
the figures of performance comparison in the massive MIMO According to Tab. 3, through setting α = β, the system
systems. feedback load can be greatly saved. Especially, with α =
β = 0.15, only about half of the cell users need to feed back
V. SIMULATION AND DISCUSSION their CSIs, and the sum rate only has a slight degradation as
In this part, we mainly focus on the sum rate performance depicted in Fig. 6.
and the system feedback overhead of the proposed CUS
scheme. The noise variance of each user is one. The sum
rate is obtained over 104 channel realizations. The channel
in the paper obeys the Rayleigh distribution with mean 0 and
variance 1.

FIGURE 7. The relationship between the system sum rate and filtering
threshold α.

Fig. 7 depicts the effect of filtering threshold α on the

FIGURE 6. The system sum rate with different filtering thresholds α and system sum rate. The simulation conditions are same as those
correlation coefficients β. of Fig. 6. As seen in Fig. 7, the sum rate of the CUS scheme
increases with the filtering threshold α given the correlation
Fig. 6 shows the system sum rate performance of the CUS coefficients β. But if the threshold α is too large, it cannot
scheme. In this simulation, the BS is equipped with 32 anten- effectively reduce the feedback users, and if α is too small,
nas, and there are 100 users which are randomly divided into it would greatly degrade the system sum rate. Considering
M = 10 groups. α is equal to β. The problem about how to this factor, we can set β − α = 0.04 according to curves
set filtering threshold α and correlation coefficient β will be in Fig. 7. Since, in this interval, the scheme would not
discussed in detail in the following parts. As shown in Fig. 6, greatly decrease system sum rate (less than 15%), and it can

21918 VOLUME 6, 2018

X. Chen et al.: CUS in Massive MIMO Systems

effectively save the system resources. Thus the system can

get a good trade-off between the sum rate and the system
feedback load. As for the realization in practice, the system
can determine the values of α and β beforehand according to
the system configuration [5], then BS informs all the users in
the way similar to the realization of SINR threshold in limited
feedback [25].

FIGURE 9. The sum rate comparison with two popular algorithms.

into 10 groups and the receiving SNR is 10 dB. Seen from

Fig. 9, the proposed user scheduling scheme achieves higher
sum rate compared with greedy zero-forcing dirty-paper
(gZF-DP) algorithm [14]. According to the feedback schemes
of the two algorithms, our scheme occupies less system
FIGURE 8. The average numbers of transmitting user with different user resources. The zero-forcing with user scheduling (ZFS) [14]
groups.
algorithm obtains the highest sum rate, since it exhaustively
searches the optimal user set among all the users. Thus ZFS
Fig. 8 plots the curves of the average numbers of trans-
algorithm has the highest computational complexity, which
mitting users with different user groups. The 100 users are
is infeasible as the numbers of users and BS antenna are
randomly divided into M = 10 groups and M = 20 groups.
large. Comparably, our proposed user scheduling scheme can
The average receiving SNR is 10 dB. According to the results
decrease the complexity of scheduling through reducing the
of Fig. 7, α is ranging from 0.11-0.18 when β is in the range
number of feedback users. So, according to the figure, our
of [0.15, 0.22]. As seen from Fig. 8, the average number of
scheme can achieve an excellent sum rate, save the system
transmitting users is in the range of [35, 68] when 0.11 ≤
resources and decrease the computational complexity at the
α ≤ 0.18, Nt = 32, and M = 10. So the scheme can
same time.
save about 32% − 65% system feedback resources. Mean-
while, the reduction would not greatly affect the system sum
VI. CONCLUSION
rate. Due to the channel-hardening characteristic, the filtering
In this paper, through making use of the cooperation among
threshold α and correlation coefficient β would decrease
users to filter out the users whose correlation coefficients are
with the increasing BS antennas in order to obtain higher
greater than the filtering threshold, a new cooperative user
sum rate, so more users would be filtered out when the BS
scheduling scheme was proposed based on the D2D com-
antenna increases further. According to Fig. 8, the number
munication among users. The cooperative user scheduling
of transmitting users increases with M slightly. So, we can
scheme could effectively ease the problem of huge feedback
slightly decrease the number of the groups M to reduce the
in massive MIMO systems and decrease the computational
system feedback overhead. As shown in Tab. 4, M has little
complexity of user scheduling at the BS, which can be
influence on the system sum rate (the simulation conditions
applied to the maritime communication systems. Meanwhile,
are same as those of Fig. 6).
we analyzed the influence of filtering threshold, correlation
TABLE 4. The relationship between the sum rate and user group
coefficient, and the number of groups on the system sum
number M. rate performance and feedback load. According to the sim-
ulations, the sum rate and feedback overhead could achieve
a good balance, while the corresponding values of these
parameters were changeable according to specific needs of
different systems. Through exhaustive analysis on the CUS
scheme, we obtained the probability of each user being
In Fig. 9, we compare the proposed CUS scheme with selected, the probabilities of different selected user numbers
two popular user scheduling algorithms to give more insights and the lower bound of the sum rate. For analytical tractabil-
on the proposed scheme’s performance. In the simulation, ity, we just plotted the curves of the theoretical results in
we assume the BS has 32 antennas, and the users are divided conventional MIMO systems. Whilst, as analyzed, when the

VOLUME 6, 2018 21919

 X. Chen et al.: CUS in Massive MIMO Systems

number of BS antennas increased, the theoretical analysis Since the selected user at the first iteration must be one of
could obtain better performance, and the lower bound would the best users of M groups, so we do not need to consider
get tighter. the case that user s is not the best user of the g-th group. The
probability Ps,21 that user s is selected at the second iteration
APPENDIX A
and is the best user of the g-th group, can be expressed as
PROOF OF EQUATIONS (18)–(22)
1) USER FILTERING 1
Ps,21 ≥ P{card(Sc,1 ) = ns,1 , s ∈ Sc,1 } (39)
After the user filtering phase, the probability that the remain- ns,1
ing user number is nf and user s belongs to the set Sf can be N nf −1
M −1X X 1
calculated as = Pf 1 P{Ni,s1 = ni,s1 } (40)
M ns,1
nf =M ns,1 =1
P Nfilt = nf , s ∈ Sf


M
M Y
·P{Ng,s1 = ng,s1 } P{Nm,s1 = nm,s1 },
Y
= Pf ,m
m=1,
m=1 m6=i,
M m6=g

Y
= P Ng = ng , s ∈ Sf ,g P (Nm = nm ) where P{Nm,s = nm,s1 }, m = 1, 2, · · · , M , denotes the prob-
m=1,
m6=g
ability that, after the first iteration of the SUS algorithm,
M   the remaining user number of the m-th group is nm,s1 .

Y NG −1 nm −1
N −n M
= Pf ,g1 + Pf ,g2 Pfilt 1−Pfilt G m P
nm,s1 = ns,1 , and card(Sc,1 ) = ns,1 . Here, it is assumed
nm −1
m=1, m=1
m6=g
that the selected user at the first iteration is coming from the
= Pf 1 + Pf 2 , (36)
i-th group, i 6 = g.
where Pf ,m = P (Nm = nm ) , m = 1, 2, . . . , M , m 6 =
P{Nm,s1 = nm,s1 }
g is the probability that the m-th group’s remaining user   
 ni − 1 ni,s


n −n −1
number is nm , that is, Pf ,m = P{card Sf ,m = nm } =  P1,c1 1 − P1,c i i,s1 , m=i
  ni,s1

M
  
NG − 1 nm −1
N −n
1 − Pfilt G m , and
P  
Pfilt nm = nf . 
ng − 1 ng,s
nm − 1 m=1 = P1 1 (β){1 − P1 (β)}ng −ng,s1 , m = g
Sf ,m is the m-th group’s user set
after the user filter-  ng,s1 −1

nm
 nm,s
ing. Pf ,g = P Ng = ng , s ∈ Sf ,g denotes the proba- P1 1 (β){1 − P1 (β)}nm −nm,s1 , others



nm,s1

bility that the g-th group’s remaining user number is
ng and user s is in the set Sf ,g . Pf ,g = Pf ,g1 + (41)
Pf ,g2 , in which Pf ,g1 denotes the probability that the g-th when m = i, the ‘‘survival probability’’ P1,c of the candidate
group’s remaining user number is ng and user s is the users in the i-th group, at the first iteration, is a conditional
P Ng = ng , s ∈ Sf ,g , s ∈ / Wg ,


best user. That is,  Pf ,g1 =  probability instead of P1 (β). Since the candidate users from
NG − 1 n −1 the i-th group are already satisfied with the coefficient con-
and Pf ,g1 = N1G Pfiltg (1 − Pfilt )NG −ng . Pf ,g2
ng − 1 straint (14). Thus the ‘‘survival probability’’ can be expressed
is the probability that the g-th group’s remaining user
as
number is ng and user s is not the
best one. That (     )
hs hH   h hH 
is, Pf ,g2 = P Ng = ng , s ∈ Wg , and Pf ,g2 = s b
P1,c = P b
≤β ≤α

M
 
NG −1 N G − 2 n −1 khs k khb k  khs k khb k
Pfiltg (1 − Pfilt )NG −ng . Pf 1 = Pf ,g1 Pf ,m ,
Q
NG ng − 2    
m=1, hs hH 
m6=g P khs kkhb k ≤ β
b
M P1 (β)
Pf 2 = Pf ,g2
Q
Pf ,m . =   H  = . (42)
hs h  Pfilt
m=1,
m6=g
P khs kkhb k ≤ α
b

2) SEMI-ORTHOGONAL USER SCHEDULING (SUS)

The probability that user s is selected for data transmission Here, α ≥ β (when α ≤ β, P1,c = 1). The inequality (39)
can be calculated as holds since the candidate users which are in the same group
with the first selected user satisfy (14), which would lead to a
Nt
X high selection probability. While, in this analysis, we assume
Ps = Ps,i , (37)
that all the remaining users’ selection probability are P1 (β),
i=1
which causes that the theoretical probability are lower than
where Ps,i denotes the probability that user s is selected at the the actual one.
i-th iteration of the SUS algorithm. Ps,1 can be expressed as The probability Ps,22 that user s is selected at the second
N M iteration and is not the best user of the g-th group is calculated
1 X X
Ps,1 = Pf 1 , nm = nf , 1 ≤ nm ≤ nf . (38) in (43), as shown at the top of the next page, in which the
M case 1 represents the selected user and user s are in the
nf =M m=1

21920 VOLUME 6, 2018

X. Chen et al.: CUS in Massive MIMO Systems

1 
P card(Sc,1 ) = ns,1 , s ∈ Sc,1

Ps,22 ≥ (43)
ns,1
N n0f −1 M
M −1X X 1  n oY 
P Ni,s1 = n0i,s1 P Ng,s1 = n0g,s1 P Nm,s1 = n0m,s1

= Pf 2
M 0 ns,1
nf =M ns,1 =1 m=1,
m6=i,
m6=g
| {z }
case1
N f n00 −1 M
1 X X 1 n oY
P Ng,s1 = n00g,s1 P Nm,s1 = n00m,s1 .

+ Pf 2 (44)
M 00 ns,1
nf =M ns,1 =1 m=1,
m6=g
| {z }
case2

P{card(Sc,1 ) = ns,1 , s ∈ Sc,1 }

N nf −1 M
M −1X X Y
= Pf 1 P{Ni,s1 = ni,s1 }P{Ng,s1 = ng,s1 } P{Nm,s1 = nm,s1 }
M
nf =M ns,1 =1 m=1,
m6=i,
m6=g

N f n0 −1 M
M −1X X Y
+ Pf 2 P{Ni,s1 = n0i,s1 }P{Ng,s1 = n0g,s1 } P{Nm,s1 = n0m,s1 }
M 0
nf =M ns,1 =1 m=1,
m6=i,
m6=g

N f n00 −1 M
1 X X Y
+ Pf 2 P{Ng,s1 = n00g,s1 } P{Nm,s1 = n00m,s1 }
M 00
nf =M ns,1 =1 m=1,
m6=g

N nf −1 M
a M −1 X X Y
= (Pf 1 + Pf 2 ) P{Ni,s1 = ni,s1 }P{Ng,s1 = ng,s1 } P{Nm,s1 = nm,s1 }
M
nf =M ns,1 =1 m=1,
m6=i,
m6=g

N f n0 −1 M
1 X X Y
+ Pf 2 P{Ng,s1 = n0g,s1 } P{Nm,s1 = n0m,s1 }. (50)
M 0
nf =M ns,1 =1 m=1,
m6=g

P card(Sc,k ) = ns,k , s ∈ Sc,k


 
ns,k−1 − 1 ns,k−1 − 2 n
≈ P{card(Sc,k−1 ) = ns,k−1 , s ∈ Sc,k−1 } Pk s,k (β)(1 − Pk (β))ns,k−1 −ns,k . (51)
ns,k−1 ns,k − 1

different groups. The case 2 represents the first selected user P{Nm,s1 = n00m,s1 }
is in the same group with the user s, in which  
nm
 00
nm,s n −n00
P1 1 (β)(1 − P1 (β)) m m,s1 , m 6 = g

 n00

P{Nm,s1 = n0m,s1 } =  m,s1  00
ng − 2 ng,s
n −n00 −1
P1,c 1 1 − P1,c g g,s1 , m = g


   0  00
ng,s1 − 1
 ni 0− 1 P i,s1 1 − P1,c ni −ni,s1 −1 ,
n
0

1,c m=i

 n
  i,s1
 (46)
 0
ng − 1 ng,s

n −n0
= P1 1 (β){1 − P1 (β)} g g,s1 , m = g M M
0 n0m,s1 = ns,1 , 0 ≤ n0m,s1 ≤ ns,1 . n00m,s1 =
P P

 ng,s1 −1 0 where
m=1 m=1
nm
 nm,s n −n0
ns,1 , 0

P1 1 (β){1 − P1 (β)} m m,s1 , others ≤ n00m,s1
≤ ns,1 . n0m,s1 and n00m,s1 respectively represent


 0
nm,s1
the remaining users number of the m-th group in case 1 and
(45) case 2, 1 ≤ m ≤ M .

VOLUME 6, 2018 21921

 X. Chen et al.: CUS in Massive MIMO Systems

The probability Ps,2 that user s is selected at the second [9] Y. Xu, G. Yue, N. Prasad, S. Rangarajan, and S. Mao, ‘‘User grouping and
iteration of the SUS algorithm is scheduling for large scale MIMO systems with two-stage precoding,’’ in
Proc. IEEE Int. Conf. Commun. (ICC), Sydney, NSW, Australia, Jun. 2014,
pp. 5197–5202.
Ps,2 = Ps,21 + Ps,22 . (47) [10] J. Nam, A. Adhikary, J.-Y. Ahn, and G. Caire, ‘‘Joint spatial division and
multiplexing: Opportunistic beamforming, user grouping and simplified
According to (39), (43), and (47), we obtain the lower bound downlink scheduling,’’ IEEE J. Sel. Topics Signal Process., vol. 8, no. 5,
of the probability that user s is selected at the second iteration. pp. 876–890, Oct. 2014.
[11] K. Doppler, M. Rinne, C. Wijting, C. B. Ribeiro, and K. Hugl, ‘‘Device-to-
Similarly, the probability that user s is selected at the device communication as an underlay to LTE-advanced networks,’’ IEEE
k-th iteration of the SUS algorithm (3 ≤ k ≤ Nt ) can de Commun. Mag., vol. 47, no. 12, pp. 42–49, Dec. 2009.
described as [12] X. Liu, Y. Liu, X. Wang, and H. Lin, ‘‘Highly efficient 3-D resource allo-
cation techniques in 5G for NOMA-enabled massive MIMO and relaying
1 systems,’’ IEEE J. Sel. Areas Commun., vol. 35, no. 12, pp. 2785–2797,
Ps,k ≥ P{card(Sc,k−1 ) = ns,k−1 , s ∈ Sc,k−1 } (48) Dec. 2017.
ns,k−1
[13] D. A. Gore, R. W. Heath, Jr., and A. J. Paulraj, ‘‘Transmit selection
ns,k−2 − 1 in spatial multiplexing systems,’’ IEEE Commun. Lett., vol. 6, no. 11,
≥ P{card(Sc,k−2 ) = ns,k−2 , s ∈ Sc,k−2 } pp. 491–493, Nov. 2002.
ns,k−2 ns,k−1
  [14] G. Dimić and N. D. Sidiropoulos, ‘‘On downlink beamforming with greedy
ns,k−2 −2 ns,k−1 user selection: Performance analysis and a simple new algorithm,’’ IEEE
× P (β){1−Pk−1 (β)}ns,k−2 −ns,k−1 −1
ns,k−1 −1 k−1 Trans. Signal Process., vol. 53, no. 10, pp. 3857–3868, Oct. 2005.
[15] A. Song, Q. Yang, W.-N. Chen, and J. Zhang, ‘‘A random-based
(49) dynamic grouping strategy for large scale multi-objective optimization,’’
in Proc. IEEE Congr. Comput. (CEC), Vancouver, BC, Canada, Jul. 2016,
where ns,k−1 and ns,k−2 respectively represent the remaining pp. 468–475.
users number at the k − 1-th and k − 2-th iteration, that is, [16] S. K. Mohammed and E. G. Larsson, ‘‘Improving the performance of the
zero-forcing multiuser MISO downlink precoder through user grouping,’’
card(Sc,k−1 ) = ns,k−1 and card(Sc,k−2 ) = ns,k−2 . IEEE Trans. Wireless Commu., vol. 15, no. 2, pp. 811–826, Feb. 2016.
According to (40) and (44), as shown at the top of the [17] X. Wu, Z. Ma, and Y. Wang, ‘‘Joint user grouping and resource allocation
previous page, the probability P {card (Sc,1 ) = ns,1 , s ∈ Sc,1 for multi-user dual layer beamforming in LTE-A,’’ IEEE Commun. Lett.,
vol. 19, no. 10, pp. 1822–1825, Oct. 2015.
that the total survival user number, at the first iteration, is ns,1 [18] A. Bletsas, A. Khisti, D. P. Reed, and A. Lippman, ‘‘A simple cooperative
and user s is in the candidate user set Sc,1 can be obtained diversity method based on network path selection,’’ IEEE J. Sel. Areas
as (50), as shown at the top of the previous page, in which the Commun., vol. 24, no. 3, pp. 659–672, Mar. 2006.
a [19] J. Yang, S. Jang, and D. K. Kim, ‘‘Sum rate approximation of zero-forcing
equality = is used for simplifying the equation.
beamforming with semi-orthogonal user selection,’’ J. Commun. Netw.,
The probability P card(Sc,k ) = ns,k , 2 ≤ k ≤ Nt that, vol. 12, no. 3, pp. 222–230, Jun. 2010.
[20] C. Chen and L.-C. Wang, ‘‘Performance analysis of scheduling in multiuser
at the k-th iteration of the SUS algorithm, the remaining MIMO systems with zero-forcing receivers,’’ IEEE J. Sel. Areas Commun.,
users number is ns,k and user s is in the set Sc,k can be vol. 25, no. 7, pp. 1435–1445, Sep. 2007.
calculated as (51), as shown at the top of the previous page. [21] A. Forenza, M. R. McKay, A. Pandharipande, R. W. Heath, Jr., and
I. B. Collings, ‘‘Capacity enhancement via multi-mode adaptation in spa-
So, according to (38), (47), and (48), we can obtain the lower tially correlated MIMO channels,’’ in Proc. IEEE 16th Int. Symp. Pers.,
bound of the probability of each user being selected, which Indoor Mobile Radio Commun., Berlin, Germany, Sep. 2005, pp. 754–758.
also proves (18). [22] M. S. Alouini and A. J. Goldsmith, ‘‘Capacity of Rayleigh fading chan-
nels under different adaptive transmission and diversity-combining tech-
niques,’’ IEEE Trans. Veh. Technol., vol. 48, no. 4, pp. 1165–1181,
REFERENCES Jul. 1999.
[1] L. Lu, G. Y. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang, [23] I. S. Gradshteyn, I. M. Ryzhik, A. Jeffrey, and D. Zwillinger, Table of
‘‘An overview of massive MIMO: Benefits and challenges,’’ IEEE J. Sel. Integrals, Series, and Products, 6th ed. San Diego, CA, USA: Academic,
Topics Signal Process., vol. 8, no. 5, pp. 742–758, Oct. 2014. 2007.
[2] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, ‘‘Massive [24] T. L. Narasimhan and A. Chockalingam, ‘‘Channel hardening-exploiting
MIMO for next generation wireless systems,’’ IEEE Commun. Mag., message passing (CHEMP) receiver in large-scale MIMO systems,’’ IEEE
vol. 52, no. 2, pp. 186–195, Feb. 2014. J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 847–860, Oct. 2014.
[3] A. Razi, D. J. Ryan, I. B. Collings, and J. Yuan, ‘‘Sum rates, rate allocation, [25] W. Zhang and K. B. Letaief, ‘‘MIMO broadcast scheduling with limited
and user scheduling for multi-user MIMO vector perturbation precoding,’’ feedback,’’ IEEE J. Sel. Areas Commun., vol. 25, no. 7, pp. 1457–1467,
IEEE Trans. Wireless Commun., vol. 9, no. 1, pp. 356–365, Jan. 2010. Sep. 2007.
[4] Z. Shen, R. Chen, J. G. Andrews, R. W. Heath, Jr., and B. L. Evans,
‘‘Low complexity user selection algorithms for multiuser MIMO systems
with block diagonalization,’’ IEEE Trans. Signal Process., vol. 54, no. 9,
pp. 3658–3663, Sep. 2006.
[5] T. Yoo and A. Goldsmith, ‘‘On the optimality of multiantenna broadcast
scheduling using zero-forcing beamforming,’’ IEEE J. Sel. Areas Commun,
vol. 24, no. 3, pp. 528–541, Mar. 2006.
XIANG CHEN was born in Hubei, China, in 1992.
[6] M. Benmimoune, E. Driouch, W. Ajib, and D. Massicotte, ‘‘Joint transmit
He received the B.Sc. degree in communica-
antenna selection and user scheduling for Massive MIMO systems,’’ in
Proc. IEEE Wireless Commun. Netw. Conf., New Orleans, LA, USA,
tions engineering from Dalian Maritime Univer-
Mar. 2015, pp. 381–386. sity, Dalian, China, in 2014, and the M.Sc. degree
[7] B. Lee, L. Ngo, and B. Shim, ‘‘Antenna group selection based user in communications engineering from Xidian Uni-
scheduling for massive MIMO systems,’’ in Proc. IEEE Global Commun. versity, Xi’an, China, in 2014,where he is currently
Conf. (GLOBECOM), Austin, TX, USA, Dec. 2014, pp. 3302–3307. pursuing the Ph.D. degree in communication and
[8] B. Lee, J. Choi, J.-Y. Seol, D. J. Love, and B. Shim, ‘‘Antenna grouping information system. His research interests include
based feedback compression for FDD-based massive MIMO systems,’’ massive MIMO wireless communications, cooper-
IEEE Trans. Commun., vol. 63, no. 9, pp. 3261–3274, Sep. 2015. ative communications, and NOMA.

21922 VOLUME 6, 2018

X. Chen et al.: CUS in Massive MIMO Systems

FENG-KUI GONG (M’12) was born in Shandong, GUO LI (S’14) was born in Shaanxi, China,
China, in 1979. He received the M.S. and Ph.D. in 1989. He received the B.Sc. and M.Sc. degrees
degrees from Xidian University, Xi’an, China, in communications engineering from Xidian Uni-
in 2004 and 2007, respectively. From 2011 to versity, Xi’an, China, in 2011 and 2014, respec-
2012, he was a Visiting Scholar with the Depart- tively, where he is currently pursuing the Ph.D.
ment of Electrical and Computer Engineering, degree in communication and information system.
McMaster University, Hamilton, ON, Canada. His research interests include MIMO wireless
He is currently a Professor with the State Key communications, cooperative communications,
Laboratory of Integrated Services Networks, and large-scale antenna system.
Department of Communication Engineering,
Xidian University. His research interests include cooperative communi-
cation, distributed space–time coding, digital video broadcasting system,
satellite communication, and 4G/5G techniques.

HANG ZHANG was born in Hebei, China,

in 1983. He received the B.Sc. degree in com-
munications engineering from Xidian University,
Xi’an, China, in 2005, and the M.Sc. degree
in electromagnetic field and microwave technol-
ogy from the 54th Research institute of CETC,
Shijiazhuang, China, in 2008. He is currently
pursuing the Ph.D. degree in information and com-
munication engineering with Northwestern Poly-
technical University. Since 2008, he has been with
the Science and Technology on Information Transmission and Dissemina-
tion in Communication Networks Laboratory, China. His research interests
include microwave and millimeter wave communication wireless ad hoc
network.

VOLUME 6, 2018 21923