152 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 9, NO.

2, FEBRUARY 2020

A Quasi-Optimal Clustering Algorithm for MIMO-NOMA Downlink Systems

Fabio Saggese , Student Member, IEEE, Marco Moretti , Member, IEEE,
and Andrea Abrardo , Senior Member, IEEE

Abstract—In this letter we consider a resource allocation in three steps: beaforming, power allocation and clustering. For
problem for multi-user MIMO non orthogonal multiple access a given number N of users, we employ block-diagonalization
(MU-MIMO-NOMA) downlink transmissions. Under the NOMA (BD) beamforming to separate the users in N/2 clusters,
paradigm, users are organized in clusters of strong/weak pair and
our aim is to find an optimal clustering, beamforming and power each composed by a strong and a weak user [11]. Following
allocation scheme to minimize the power transmitted subject to a the NOMA paradigm, the strong user in the cluster per-
rate constraint for each user. Since the joint optimization problem forms successive interference cancellation (SIC) to remove the
is intractable, we split it in three sub-problems: clustering, which interference given by the paired weak user. For each possi-
is formulated as a mixed integer linear programming (MILP) ble strong/weak pair, the optimal beamforming matrices are
problem, beamforming and power allocation. Simulations results
show that the our proposed scheme greatly outperforms both the computed according to a minimum power criterion subject
classical OMA scheme and state-of-the-art NOMA techniques. to rate constraints. Unlike the other work in literature, the
clustering technique we propose is optimal being the solu-
Index Terms—MIMO, non-orthogonal multiple access, cluster-
ing, weighted bipartite matching. tion of a mixed integer linear programming (MILP) problem.
Simulations results show that the proposed NOMA approach
allows to clearly outperform classical OMA approach based
I. I NTRODUCTION on block diagonalization and state-of-the-art alternatives for
MISO communications.
ON-ORTHOGONAL multiple access (NOMA) is one of
N the key radio access technologies envisioned to meet the
heterogeneous demands on low latency, high reliability, mas-
sive connectivity, improved fairness, and high throughput of II. S YSTEM M ODEL
fifth generation (5G) mobile networks [1]. The key idea of We focus on a NOMA-MIMO downlink transmission sce-
NOMA is to exploit the power domain for allowing multiple nario with N users uniformly distributed in a cell of radius
users to be served concurrently at the same time on the same R. Each mobile user is equipped with Nr antennas. The
frequency channel. NOMA, initially developed for a SISO set- number of transmit antennas at the base station (BS) is
ting [2], faces several challenges in MIMO systems and only Nt ≥ NNr , so that there are enough spatial degrees of freedom
recent works have shown the potential to outperform classi- to ideally multiplex all users in the cell without introducing
cal MIMO-OMA systems [3]–[10]. Choosing the users that intra-cell interference. In the following, we make the non-
are matched together on the same channel has a very rele- restrictive assumption that N is even. To implement the NOMA
vant impact on the performance of any implementation of the paradigm, the N users are grouped in NC = N /2 clusters
NOMA paradigm. In a MIMO setting, where users are sepa- composed by couples of mutually interfering users, gener-
rated by a beamforming precoder, user clustering, the problem ally referred to as strong and weak users, so labelled on the
of choosing the users belonging to the same beam, becomes base of the quality of their propagation channel. Assuming
even more important because it has consequences also on the perfect channel state information at the BS, we sort in ascend-
shape of the beams. Because of the complexity of the task, ing order the users employing the Frobenius norm of the
most of the works in literature on clustering are either based channel gain matrix as quality indicator. Accordingly, weak
on heuristic algorithms [3]–[7], or require a large number of and strong users have indexes w ∈ W = {1, . . . , NC } and
antennas at the mobile terminals (MT) [8], or are theoretic s ∈ S = {NC + 1, . . . , N }, respectively.
analysis of the capacity of the systems [9]. The message for the generic user i, classified either as strong
The contribution of this letter is a novel design algorithm or weak, is spatially multiplexed on Nr data streams, so that
√ √ √
that solves the optimization problem for downlink transmis- its vector is xi = [ p1,i s1,i , p2,i s2,i , . . . , pNr ,i sNr ,i ]T
sions in a general MU-MIMO-NOMA scenario by splitting it where pj ,i and sj ,i are the power and the i.i.d. unitary symbol
transmitted on the j-th data stream of user i, respectively, so
Manuscript received August 2, 2019; accepted October 4, 2019. Date of
publication October 10, 2019; date of current version February 7, 2020. The
that it is
  1 if  = j and i = m
associate editor coordinating the review of this article and approving it for
publication was W. Zhang. (Corresponding author: Fabio Saggese.) 
∗
F. Saggese and M. Moretti are with the Dipartimento di Ingegneria E sj ,i s,m = (1)
dell’Informazione, University of Pisa, 56122 Pisa, Italy (e-mail:
0 otherwise.
[email protected]; [email protected]).
A. Abrardo is with the Dipartimento di Ingegneria dell’Informazione,
University of Siena, 53100 Siena, Italy (e-mail: [email protected]). and Pi = E{xi xH i } is the diagonal covariance matrix with
Digital Object Identifier 10.1109/LWC.2019.2946548 the vector pi = {p1,i , p2,i , . . . , pNr ,i } on the main diagonal.
2162-2345 c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Higher College of Technology. Downloaded on March 23,2021 at 07:09:29 UTC from IEEE Xplore. Restrictions apply.
SAGGESE et al.: QOC ALGORITHM FOR MIMO-NOMA DOWNLINK SYSTEMS 153

Denoting with Wi the beamforming matrix for the generic Implementing the NOMA paradigm [12], we need to define
user i, the transmitted signal is: an optimal strategy for grouping the N users into clusters of
strong/weak users. The partition of all users into disjoint clus-
yi = Wi xi . (2)
ters of two users is called a clustering. In order to properly
The signal at the i-th receiver is formulate the clustering problem, we introduce the allocation
  variable ρm,n ∈ {0, 1}, m ∈ W, n ∈ S, which is ρm,n = 1, if
 
ri = Hi Ww xw + Ws xs + ni , (3) users m ∈ W and n ∈ S are clustered together, and ρm,n = 0
w ∈W s∈S
otherwise. Note that the number of allocation variables is Nc2 ,
equal to the number of possible clusters. Since each user can
where Hi ∈ CNr ×Nt is the channel matrix and ni ∼ belong at most to a cluster and each cluster must contain
CN (0, σn2 INr ) is the AWGN vector at the receiver. As for a strong and a weak user only, any feasible clustering must
the channel model, we assume uncorrelated antennas, i.e., the satisfy the two following constraints
entries of Hi are i.i.d. circularly symmetric, complex Gaussian 
random variables. ρw ,s = 1, ∀s ∈ S,
Let us denote by C = {w , s} a generic cluster composed w ∈W

by w ∈ W and s ∈ S and denote by k ∈ {w , s} a generic ρw ,s = 1, ∀w ∈ W, (9)
element of the cluster. The inter-cluster interference covariance s∈S
matrix seen by user k can be expressed as: Accordingly, the number of feasible clustering, i.e., the num-
⎛ ⎞
ber of combination of Nc clusters fulfilling (9) is Nc !. Let us
⎜  ⎟ H now denote by P, W and ρ the vectors and matrices stacking
Jk = Hk ⎜
⎝ Wi Pi WiH + Wj Pj WjH ⎟
⎠ Hk . (4) all the optimization variables, i.e., the transmitting powers,
i∈W j ∈S the beamforming matrices and the cluster allocation vari-
i=w j =s
ables, respectively. We can formulate the power minimization
In each cluster the strong NOMA user is able to remove problem as
the interference of the weak NOMA user by performing    
H
interference cancellation, while the weak user is detected in min ρw ,s tr{Ww Pw Ww }+ tr{Ws Ps WsH }
P,W,ρ
the presence of the interference caused by the strong user. This w ∈W s∈S
implies that the strong users have knowledge of the codebook subject to Ri ≥ ηi , ∀i ∈ {W, S}
of its weak associated user, after optimal clustering scheme has Pi  0, ∀i ∈ {W, S}
been found. Accordingly, the noise covariance matrix, obtained ρw ,s ∈ {0, 1}, ∀w ∈ W, ∀s ∈ S
by summing the covariance of thermal noise, intra-cluster and 
intra-cluster interference is: ρw ,s = 1, ∀s ∈ S,
 2 w ∈W
σn INr + Hk Ws Ps WsH HH k + Jk , if k = w (5)

Nk = ρw ,s = 1, ∀w ∈ W,
σn2 INr + Jk , if k = s.
s∈S
We are now in the position of deriving the achievable rate of Rw ≤ Rw ,s , ∀w ∈ W, ∀s ∈ S, ρw ,s = 1,
user k as:
  (10a)
−1
Rk = log2 det INr + Hk Wk Pk WkH HH N
k k . (6)
Problem (10) is a mixed-integer non-convex problem, and its
NOMA interference cancelation is possible under the con- global optimum solution is unknown. In order to obtain a solu-
dition that the strong user s can correctly decode (and cancel!) tion of this problem, we follow a sub-optimal approach and
the signal of the weak user w. This condition can be translated we split (10) into three different steps.
into the relation 1) For each of the possible Nc2 clusters of users, we
adopt block diagonalization beamforming so that the
Rw ≤ Rw ,s (7) spatial precoder D is employed to nullify the interference
where Rw ,s is the so called weak-strong rate, i.e., the rate of between clusters.
the weak user at the receiver of the strong user 2) Having set to zero the inter-cluster interference, we can
  now compute for each cluster the optimal beamformer
H H −1
Rw ,s = log2 det INr + Hs Ww Pw Ww Hs Nw ,s (8) B and optimally distribute the power on the various
streams so that the rate constraints are met.
with Nw ,s = σn2 INr + Hs Ws Ps WsH HH
s + Js . 3) We select the feasible clustering strategy that yields the
minimum transmit power.
III. P ROBLEM F ORMULATION
In this letter we address the problem of minimizing the IV. B EAMFORMING D ESIGN
transmit power in the presence of per-user rate constraints ηi , Implementing the beamforming strategy described in the
so that each user gets a portion of the available spectrum. previous section, determines that the spatial precoder for
Unlike rate maximization, the problem we study is oriented to user i is computed as the cascade of two spatial filters, i.e.,
a fair resource allocation between users. Wi = Di Bi .

Authorized licensed use limited to: Higher College of Technology. Downloaded on March 23,2021 at 07:09:29 UTC from IEEE Xplore. Restrictions apply.
154 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 9, NO. 2, FEBRUARY 2020

A. Block Diagonalization C. Weak User’s Beamforming

The inter cluster interference can be set to zero by employ- Given the above, the signal at weak user receiver is
ing block diagonalization (BD) beamforming [13] to separate
the signals of the various clusters. Let us consider a generic rw = H̄w Bw xw + H̄w Bs xs + nw , (18)
cluster C = {w , s} and introduce H−(w ,s) ∈ CdNOMA ×Nt as where H̄w = Hw Dw ,s . The optimal beamforming matrices
the matrix obtained by stacking the channel matrices of all for the weak users is obtained by exploiting the knowledge
N−2 users in the system different from w and s, i.e., of the covariance matrix Nw of the interference plus noise
vector [14]. At the weak receiver the correlation matrix Nw
H−(w ,s) = [HH H H H
1 , . . . , Hw −1 , Hw +1 , . . . , HNC in (5) becomes
HH H H H H
NC +1 , . . . , Hs−1 , Hs+1 , . . . , HN ] , (11) Nw = H̄w Bs Ps BH H 2
s H̄w + σn INr . (19)
where dNOMA = (N − 2)Nr . The singular value decomposi- The noise at the weaker receiver can be whitened by the
tion (SVD) of H−(w ,s) is following operation:
(1) (0) −1 −1
H−(w ,s) = U−(w ,s) Σ−(w ,s) [V−(w ,s) , V−(w ,s) ]H . (12) Nw 2 rw = Nw 2 H̄w Bw xw + ñw , (20)
− 12
where ñw = Nw (H̄w Bs xs + nw ) ∼ CN (0, INr ). The
where the columns of V−(w ,s) ∈ CNt ×(Nt −dNOMA ) span the
(0)
optimal transmit and receive filter are then obtained by
null space vector of H−(w ,s) , so that by employing the spatial computing the SVD of the filtered equivalent weak channel
filter
−1 H
(0) Nw 2 H̄w = Ūw Σ̄w V̄w . (21)
Dw ,s = V−(w ,s) , (13)
and setting
the inter-cluster interference generated by w and s is com-
pletely eliminated. Bw = V̄w ∈ C(Nt −dNOMA )×Nr ,
−1
Cw = Nw 2 Ūw ∈ CNr ×Nr . (22)
B. Strong User’s Beamforming The combining operation at the receiver yields:
Let us focus on the strong user s and assume that condi-
tion (7) is fulfilled, i.e., the interference of weak user can be zw = CH
w rw = Σ̄w xs + n̄w , (23)
canceled. The cases where (7) is not satisfied is treated sep- where n̄w = ŪH w ñw ∼ CN (0, INr ). Hence, the achievable
arately in Section IV-E. Because of BD precoding, the inter rate of weak user w is
cluster interference is zero and the received signal is

Nr
 
Rw = log2 1 + pj ,w Λj ,w , (24)
rs = Hs Dw ,s Bs xs + Hs Dw ,s Bw xw +ns
   j =1
canceled by SIC
where Λj ,w = σj2,w /σn2 and σj ,w are the diagonal elements
= H̄s Bs xs + ns , (14)
of Σ̄w .
where H̄s = Hs Dw ,s . The precoding and combining matrices
are obtained by the SVD of the H̄s = Ūs Σ̄s V̄sH , by setting D. Power Allocation Algorithm
From (17) and (24), one can formulate the minimum power
Bs = V̄s ∈ C(Nt −dNOMA )×Nr , problem subject to rate constraints for a generic user k ∈
Cs = Ūs ∈ CNr ×Nr . (15) {w , s} belonging to the cluster as
min tr{Pk }
The combining operation at the strong receiver yields:

Nr
 
zs = Cs rs = Σ̄s xs + n̄s , (16) subject to log2 1 + pj ,k Λj ,k ≥ ηk ,
j =1
where the noise vector n̄s = Cs ns has the same statis- pj ,k ≥ 0, j = 1, . . . , Nr . (25)
tics of ns . The achievable rate of strong user s can be now
Problem (25) is convex and differentiable and it can be effi-
expressed as
ciently solved in the Lagrangian dual domain. The minimum
is obtained by the well known water-filling solution [15]

Nr
 
Rs = log2 1 + pj ,s Λj ,s (17)  +
μk 1
j =1 pj∗,k = − , ∀j , k . (26)
ln 2 Λj ,k
where Λj ,s = σj2,s /σn2 , where σj ,s are the diagonal elements where μk are chosen so that the rate constraints in (25) for
of Σ̄s . user k is met.

Authorized licensed use limited to: Higher College of Technology. Downloaded on March 23,2021 at 07:09:29 UTC from IEEE Xplore. Restrictions apply.
SAGGESE et al.: QOC ALGORITHM FOR MIMO-NOMA DOWNLINK SYSTEMS 155

Algorithm 1: QOC
1 Initialize: sort in increasing order the user channel gains;
2 Set W = {1, . . . , NC }, S = {NC + 1, . . . , N };
3 for w ∈ W do
4 for s ∈ S do
5 Compute Dw ,s according to Section IV-A;
6 Compute Bs and Cs according to Section IV-B;
7 Evaluate power allocation for Ps solving problem (25);
8 Compute Bw and Cw according to Section IV-C;
9 Evaluate power allocation for Pw solving problem (25);
10 if Rw > Rw ,s then
11 Perform BD of s and w;
12 Pw ,s = tr{Pw } + tr{Ps }

13 Solve problem (28) for optimal clustering;

Fig. 1. Ptot as a function of ηi for the NOMA proposed algorithm and for
E. NOMA Constraint the OMA scheme.
In the case constraint (7) is not met in a specific cluster,
we adopt for that cluster a conventional OMA scheme, where
the two users in the cluster are orthogonalized by means of
BD. In this case, to implement BD the number of columns of
the effective channel H̄ matrix is reduced by a factor Nr and
the users will experience a smaller degree of spatial diversity.
This situation typically occurs for those clusters where the
strong and the weak users are characterized by similar channel
quality.

V. C LUSTER S ELECTION
In order to formulate the optimal clustering scheme for the
problem at hand, we denote by Pw ,s the total transmitting
power of cluster {w,s}, i.e.:
 
Pw ,s = pj ,w + pj ,s , (27) Fig. 2. Ptot as a function of the number of users, for ηi = 6.
j j
and the optimal clustering problem can be formulated as performance of the proposed NOMA scheme together with
  the benchmark OMA scheme based on BD proposed in [13],
min Ptot (ρ) = ρw ,s Pw ,s , referred to as OMA in the following.
ρ
w ∈W s∈S Figs. 1 and 2 show a comparison between NOMA and OMA
subject to ρw ,s ∈ {0, 1}, ∀w ∈ W, ∀s ∈ S, in terms of the overall power spent Ptot . For each user, we

ρw ,s = 1, ∀s ∈ S, have considered a number of receiving antenna Nr = 2, while
w ∈W the number of antennas at the BS is set to Nt = Nr N = 2N .

ρw ,s = 1, ∀w ∈ W. (28) Fig. 1, presents the total required power as a function of
the target rates for N = 32 and N = 64, where the same
s∈S
target rate is assumed for all users. NOMA clearly outperforms
Problem (28) is a particular type of MILP problem called OMA, thus demonstrating the effectiveness of the clustering
weighted bipartite matching (WBM). WBM problems have algorithm in exploiting the NOMA potential. As expected, in
the important property that the coefficient matrix is totally uni- all cases we observe an exponential increase (linear in log-
modular [16], which guarantees that the optimal solution of the scale) of the required power with the increase of the target
linear problem is integer, regardless of the solver employed. rate.
In particular WBM problems can be solved with very fast Fig. 2 shows the total required power as a function of the
algorithms [17]. number of users in the cell, for a fixed target throughput
The quasi-optimal clustering (QOC) algorithm is summa- ηi = 6. The figure shows the results for NOMA with Nr = 2
rized in Algorithm 1. and OMA with Nr = 2 and Nr = 4. For any number of users
N, NOMA outperforms the OMA scheme implemented with
VI. N UMERICAL R ESULTS the same number of receive antennas and, as the number of
We consider a single cell scenario with radius R = 100 m. users increases, the performance of our proposed scheme tends
The exponential path loss is γ = 4. The central carrier to the performance of OMA with a double number of receive
frequency is f0 = 2 GHz. The power of the additive Gaussian antennas, i.e., Nr = 4. Indeed, in the presence of a high num-
noise is σn2 = −125 dBm. In all figures, we plot the ber of users, the clustering algorithm exploits the multi-user

Authorized licensed use limited to: Higher College of Technology. Downloaded on March 23,2021 at 07:09:29 UTC from IEEE Xplore. Restrictions apply.
156 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 9, NO. 2, FEBRUARY 2020

clustering allocation is selected as the solution of a MILP

problem. Simulations results show that the proposed MIMO-
NOMA scheme outperforms both classical MIMO-OMA and
state-of-the-art MIMO-NOMA algorithms.

R EFERENCES
[1] Z. Ding, X. Lei, G. K. Karagiannidis, R. Schober, J. Yuan, and
V. K. Bhargava, “A survey on non-orthogonal multiple access for 5G
networks: Research challenges and future trends,” IEEE J. Sel. Areas
Commun., vol. 35, no. 10, pp. 2181–2195, Oct. 2017.
[2] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and
K. Higuchi, “Non-orthogonal multiple access (NOMA) for cellular
future radio access,” in Proc. IEEE VTC Spring, Dresden, Germany,
Jun. 2013, pp. 1–5.
[3] B. Kimy et al., “Non-orthogonal multiple access in a downlink multiuser
beamforming system,” in Proc. IEEE MILCOM, San Diego, CA, USA,
Fig. 3. Ptot as a function of ηi , for the NOMA proposed algorithm, IBPA Nov. 2013, pp. 1278–1283.
+ WF, IBPA [6] and OMA, for N = 32 and for cell radius R = 100 m and [4] J. Choi, “Minimum power multicast beamforming with superposition
R = 500 m. coding for multiresolution broadcast and application to NOMA systems,”
IEEE Trans. Commun., vol. 63, no. 3, pp. 791–800, Mar. 2015.
[5] Z. Chen, Z. Ding, X. Dai, and G. K. Karagiannidis, “On the application
of quasi-degradation to MISO-NOMA downlink,” IEEE Trans. Signal
diversity of the system and is more likely to find a couple of Process., vol. 64, no. 23, pp. 6174–6189, Dec. 2016.
users which can efficiently share the same channel with a low [6] Z. Chen, Z. Ding, and X. Dai, “Beamforming for combating inter-cluster
power consumption. and intra-cluster interference in hybrid NOMA systems,” IEEE Access,
vol. 4, pp. 4452–4463, 2016.
Finally, Fig. 3 provides the comparison of the proposed [7] Z. Chen and X. Dai, “MED precoding for multiuser MIMO-NOMA
scheme with two other schemes: the current state-of-the-art downlink transmission,” IEEE Trans. Veh. Technol., vol. 66, no. 6,
algorithm [6] for NOMA beamforming and clustering, labelled pp. 5501–5505, Jun. 2017.
[8] Z. Ding, R. Schober, and H. V. Poor, “A general MIMO framework
as inversion based paired algorithm, ‘IBPA’, and the scheme for NOMA downlink and uplink transmission based on signal align-
obtained by combining IBPA clustering with the beamform- ment,” IEEE Trans. Wireless Commun., vol. 15, no. 6, pp. 4438–4454,
ing strategy presented here, labelled as ‘IBPA + WF’. The Jun. 2016.
[9] M. Zeng, A. Yadav, O. A. Dobre, G. I. Tsiropoulos, and H. V. Poor,
results for this second algorithm show the superiority of our “Capacity comparison between MIMO-NOMA and MIMO-OMA with
clustering algorithm with respect to the heuristic presented multiple users in a cluster,” IEEE J. Sel. Areas Commun., vol. 35, no. 10,
in [6]. The results are plotted for R = 100 m and R = 500 m pp. 2413–2424, Oct. 2017.
[10] M. Zeng, A. Yadav, O. A. Dobre, and H. V. Poor, “Energy-efficient
cell radius with no appreciable difference in the algorithms power allocation for MIMO-NOMA with multiple users in a cluster,”
behaviour. In order to compare the three algorithms, we con- IEEE Access, vol. 6, pp. 5170–5181, 2018.
sider a MISO transmission paradigm, i.e., Nr = 1, and we [11] M. Moretti and A. I. Perez-Neira, “Efficient margin adaptive scheduling
for MIMO-OFDMA systems,” IEEE Trans. Wireless Commun., vol. 12,
collect the results of the total power spent Ptot for the three no. 1, pp. 278–287, Jan. 2013.
algorithms in the same simulated scenarios, for N = 32. The [12] S. M. R. Islam, M. Zeng, O. A. Dobre, and K.-S. Kwak, “Resource allo-
different number of receive antenna is the reason for the differ- cation for downlink NOMA systems: Key techniques and open issues,”
IEEE Wireless Commun., vol. 25, no. 2, pp. 40–47, Apr. 2018.
ent results in Fig. 1 and Fig. 3. Due to the optimal clustering [13] Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, “Zero-forcing meth-
selection, the proposed algorithm largely outperforms both ods for downlink spatial multiplexing in multiuser MIMO channels,”
reference algorithms. IEEE Trans. Signal Process., vol. 52, no. 2, pp. 461–471, Feb. 2004.
[14] F. R. Farrokhi, G. J. Foschini, A. Lozano, and R. A. Valenzuela,
“Link-optimal space-time processing with multiple transmit and receive
antennas,” IEEE Commun. Lett., vol. 5, no. 3, pp. 85–87, Mar. 2001.
VII. C ONCLUSION [15] D. P. Palomar and J. R. Fonollosa, “Practical algorithms for a family
In this letter, the joint clustering, beamforming and power of waterfilling solutions,” IEEE Trans. Signal Process., vol. 53, no. 2,
pp. 686–695, Feb. 2005.
allocation problem for the downlink of multi-user MIMO- [16] M. Moretti, A. Abrardo, and M. Belleschi, “On the convergence and
NOMA cellular systems has been investigated. In particular, optimality of reweighted message passing for channel assignment prob-
we have proposed a scheme where strong and weak users lems,” IEEE Signal Process. Lett., vol. 21, no. 11, pp. 1428–1432,
Nov. 2014.
are paired together to form NOMA clusters and the vari- [17] B. Huang and T. Jebara, “Loopy belief propagation for bipartite maxi-
ous clusters are separated by BD beamforming. The optimal mum weight b-matching,” in Proc. Artif. Intell. Stat., 2007, pp. 195–202.

Authorized licensed use limited to: Higher College of Technology. Downloaded on March 23,2021 at 07:09:29 UTC from IEEE Xplore. Restrictions apply.