416 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 5, NO.

4, AUGUST 2016

Relay Selection for Cooperative NOMA

Zhiguo Ding, Senior Member, IEEE, Huaiyu Dai, Senior Member, IEEE, and H. Vincent Poor, Fellow, IEEE

Abstract—This letter studies the impact of relay selection (RS) probability achieved by the two-stage relay selection strat-
on the performance of cooperative non-orthogonal multiple egy, which shows that this two-stage scheme can realize the
access (NOMA). In particular, a two-stage RS strategy is pro- maximal diversity gain. Furthermore, analytical results are
posed, and analytical results are developed to demonstrate that developed to demonstrate that the two-stage strategy is also
this two-stage strategy can achieve the minimal outage proba-
bility among all possible RS schemes, and realize the maximal
outage-optimal, i.e., it achieves the optimal outage probabil-
diversity gain. The provided simulation results show that coop- ity among all possible relay selection schemes. On the other
erative NOMA with this two-stage RS scheme outperforms that hand, the max-min relay selection criterion can achieve the
based on the conventional max–min approach, and can also yield same performance as the two-stage one, i.e., realizing the min-
a significant performance gain over orthogonal multiple access. imal outage probability, for a special symmetrical case, but it
Index Terms—Non-orthogonal multiple access (NOMA), coop- suffers a loss in outage probability in general.
erative NOMA, relay selection, and max-min relaying. II. S YSTEM M ODEL
I. I NTRODUCTION Consider a downlink scenario with one base station (BS),
two users, and N relays. Each node is equipped with a sin-
ON-ORTHOGONAL multiple access (NOMA) has been
N recognized as a promising enabling technology to
improve the spectral efficiency of the fifth generation (5G)
gle antenna. Assume that there is no direct link between
the BS and the users, and the BS-relay and relay-user chan-
nels experience independent and identically Rayleigh fading.
mobile network [1]–[3]. The key idea of NOMA is to serve Unlike [1]–[3], users are not ordered by their channel con-
multiple users in the same frequency band, but with differ- ditions, but rather are categorized by their quality of ser-
ent power levels. The application of cooperative transmission vice (QoS) requirements. Particularly, assume that user 1 is to
to NOMA is important since spatial degrees of freedom can be served for small packet transmission, i.e., quickly connected
still be harvested even if each node is equipped with a single with a low data rate, and user 2 is to be served opportunis-
antenna. tically [9]. For example, user 1 can be a vehicle which is
A few different forms of cooperative NOMA have been pro- to receive safety critical information containing a few bytes,
posed in the literature. The work in [4] relies on cooperation such as a road flood warning or incident avoidance alerts. As
among NOMA users, i.e., users with strong channel conditions a result, this user’s QoS requirements, such as its targeted data
act as relays. A dedicated relay has been used in [5] to improve rate, should be given higher priority. On the other hand, user
the transmission reliability for users with poor channel con- 2 is to download a movie (or perform background tasks), and
ditions. Similarly, a dedicated relay has been used in [6] to hence it will be served in a more opportunistic manner.
serve multiple users equipped with multiple antennas. Wireless During the first time slot, the BS will transmit the superim-
power transfer has been applied to cooperative NOMA in [7], posed mixture, (α1 s1 + α2 s2 ), where si denotes the signal to
as an incentive for user cooperation. user i, and αi denotes the power allocation coefficient. Note
This letter characterizes the impact of relay selection on that α12 +α22 = 1 and α1 ≥ α2 since user 1’s QoS requirements
the performance of cooperative NOMA, and we focus on two are given higher priority [9]. Relay n, 1 ≤ n ≤ N, observes
types of relay selection criteria. The first one is based on
conventional max-min relay selection [8]. The second one is yrn = hn (α1 s1 + α2 s2 ) + wrn , (1)
carried out in a two-stage strategy, in which the first stage is where hn denotes the channel gain between the BS and relay
to ensure that one user’s targeted data rate is realized, and n, and wrn denotes additive Gaussian noise.
the second is to maximize the other user’s rate opportunis- The relays will decode s1 first, and then remove it from their
tically. We obtain a closed form expression for the outage observations, following the principle of successive interference
cancellation (SIC) [10]. The conditions for a relay to decode
Manuscript received May 9, 2016; accepted May 18, 2016. Date of pub- the two signals, s1 and s2 , are given by
lication June 1, 2016; date of current version August 19, 2016. The work 1 |hn |2 α12 1  
of Z. Ding was supported in part by the H2020-MSCA-RISE-2015 under log 1 + ≥ R1 , log 1 + ρ|hn |2 α22 > R2 , (2)
Grant 690750, and in part by the U.K. EPSRC under Grant EP/N005597/1. 2 |hn |2 α22 + ρ1 2
The work of H. Dai was supported by the U.S. National Science Foundation
under Grants ECCS-1307949 and EARS-1444009. The work of H. V. Poor
where ρ denotes the transmit signal-to-noise ratio (SNR) and
was supported by the U.S. National Science Foundation under Grants CNS- Ri is the targeted data rate for user i.
1456793 and ECCS-1343210. The associate editor coordinating the review of During the second time slot, assume that relay n can decode
this paper and approving it for publication was H. H. Nguyen. the two signals and is selected to send (α1 s1 +α2 s2 ). Therefore,
Z. Ding and H. V. Poor are with the Department of Electrical Engineering,
Princeton University, Princeton, NJ 08544 USA, and also with the School of
user i receives the following:
Computing and Communications, Lancaster University, Lancaster LA1 4WA, ydn,i = gn,i (α1 s1 + α2 s2 ) + wdn,i , i ∈ {1, 2}, (3)
U.K.
H. Dai is with the Department of Electrical and Computer Engineering,
where gn,i denotes the channel gain between relay n and user
North Carolina State University, Raleigh, NC 27695 USA. i and wdn,i denotes additive Gaussian noise. User 1 decodes its
Digital Object Identifier 10.1109/LWC.2016.2574709 message with signal-to-interference-plus-noise ratio (SINR),
2162-2345 c 2016 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.
DING et al.: RS FOR COOPERATIVE NOMA 417

|gn,1 |2 α12
, and user 2 decodes its own message with SNR, probability is always one, a phenomenon also observed in [2].
|gn,1 |2 α22 + ρ1
By using the fact that all channels are assumed to be Rayleigh
ρα22 |gn,2|2 , provided that user 2 can decode s1 , i.e., log(1 + fading, we have
|gn,2 |2 α12 N  
) ≥ 2R1 . Note that fixed power allocation is used
|gn,2 |2 α22 + ρ1 P(O1 ) = 1 − e−3ξ1 . (10)
in this paper. Optimizing the power allocation coefficients can
n=1
further improve the performance of cooperative NOMA, which The term P(O2 ) can be calculated as follows:
is beyond the scope of this paper.  
P(O2 ) = P(E1 , |Sr | > 0) + P E2 , Ē1 , |Sr | > 0 , (11)
A. Relay Selection Strategies ∗
where E1 denotes the event that relay n cannot decode s2 , Ē1
1) Max-Min Relay Selection: The criterion for this type of denotes the complementary event of E1 , and E2 denotes the
relay selection can be obtained as follows
 [8]:  event that user 2 cannot decode s2 . The first term in the above
max min |hn |2 , |gn,1|2 , |gn,2 |2 , n ∈ Sr , (4) equation can be expressed as
   
which selects a relay with the strongest P(E1 , |Sr | > 0) = P log 1 + ρ|hn∗ |2 α22 < 2R2 , |Sr | > 0 , (12)
min{|hn |2 , |gn,1 |2 , |gn,2 |2 }. where ξ2 = 22R2 −1
.
2) Two-Stage Relay Selection: The aim of this relay selec- ρα22
tion strategy is to realize two purposes simultaneously. One is The second term in (11) can be expressed as 
 
to ensure user 1’s targeted data rate is realized, and the other P E2 , Ē1 , |Sr | > 0 = P log 1 + ρ|gn∗ ,2 |2 α22
is to serve user 2 with a rate as large as possible. Specifically,  
this two-stage user selection strategy can be described as fol- < 2R2 , log 1 + ρ|hn∗ |2 α22
lows. The first stage is to build the following subset of the 
relays by focusing on user 1’s targeted data rate:  > 2R2 , |Sr | > 0 .

1 |hn|2 α12 Therefore, the
 probability
 P(O2) can be calculated as follows:
Sr = n : 1 ≤ n ≤ N, log 1 +
2 |hn |2 α22 + ρ1 P(O2 ) = P log 1 + ρ|hn∗ |2 α22 < 2R2 , |Sr | > 0
    
1 |gn,1|2 α12 + P log 1 + ρ|gn∗ ,2 |2 α22
≥ R1 , log 1 +   
2 |gn,1|2 α22 + ρ1
  < 2R2 , log 1 + ρ|hn∗ |2 α22 > 2R2 , |Sr | > 0 .
1 |gn,2|2 α12 (13)
≥ R1 log 1 + ≥ R1 . (5)
2 |gn,2|2 α22 + ρ1 Assuming |Sr | > 0, define   
Therefore all the relays in Sr can decode s1 by treating s2 as xn = min log 1 + ρ|hn |2 α22 , log 1 + ρ|gn,2 |2 α22 , (14)
noise. The second stage of relay selection is to select a relay
in Sr that can maximize and


therate for user 2,
i.e.,
∗ 1 2 2
xn∗ = max{xi , ∀i ∈ Sr }. (15)
n = arg max min log 1 + ρ|hn | α2 , The probability P(O ) can now be expressed as
n 2   2 
1   
P(O2 ) = P min log 1 + ρ|hn∗ |2 α22 ,
2 2
log 1 + ρ|gn,2 | α2 , n ∈ Sr . (6)    
2
log 1 + ρ|gn∗ ,2 |2 α22 < 2R2 , |Sr | > 0
III. P ERFORMANCE A NALYSIS
= P(xn∗ < 2R2 , |Sr | > 0). (16)
In this section, we will characterize the outage probability The above probability can further expressed as
achieved by the two-stage relay selection scheme. Note that
the overall outage event can be categorized as follows: N
 P(O2 ) = P(xn∗ < 2R2 , |Sr | = l)
O = O1 O2 , (7) l=1
∗
where O1 denotes the event that relay n cannot decode s1 , or 
N
either of the two users cannot decode s1 successfully, and O2 = P(xn∗ < 2R2 ||Sr | = l)P(|Sr | = l). (17)
denotes the event that s2 cannot be decoded correctly either l=1
by relay n∗ , or by user 2, while s1 can be decoded correctly For a relay randomly selected from Sr , denoted by relay n,
by the three nodes. the cumulative distribution function (CDF) of xn can be found
Therefore, the outage probability can be written as as follows:   2  2x − 1   
 

P(O) = P(O1 ) + P(O2 ). (8) F(x) = P min |hn |2 , gn,2  < n ∈ Sr , Sr   = 0
The term P(O1 ) can be calculated as follows: ρα22
  
2   
2 2 − 1 
N  x
 
= P |hn |2 > gn,2 , gn,2  < 
P(O1 ) = P(|Sr | = 0) = 1 − P |hn |2 > ξ1 ρα22
n=1  2 
   
|hn |2 > ξ1 , gn,2 > ξ1
× P |gn,1 |2 > ξ1 P |gn,2|2 > ξ1 ,
(9)   2 2x − 1 
1 + P |hn |2 < gn,2  , |hn |2 < 
where |Sr | denotes the size of Sr , ξ1 = ρ
and 1 = ρα22
α12 −1 α22  2 
22R1 − 1. It is assumed that α12 > 1 α22 , otherwise the outage |hn |2 > ξ1 , gn,2 > ξ1 . (18)
418 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 5, NO. 4, AUGUST 2016

Define the two probabilities on the right hand side of the above Theorem 1: The overall outage probability achieved by the
equation by Q1 and Q2 , respectively. The probability Q1 can two-stage relay selection scheme can be expressed as
be expressed as follows: N    N−l
  N
P |hn | > |gn,2|2 , |gn,2 |2 < y, |hn |2 > ξ1 , |gn,2 |2 > ξ1
2 P(O) = (F(2R2 ))l e−3lξ1 1 − e−3ξ1 , (27)
Q1 =   l=0
l
P |hn |2 > ξ1 , |gn,2 |2 > ξ1
 2    if α12 > 1 α22 ; otherwise P(O) = 1.
P |hn | > max ξ1 , |gn,2 |2 , ξ1 < |gn,2|2 < y
=   , (19)
P |hn |2 > ξ1 , |gn,2|2 > ξ1 A. High SNR Approximations
2x −1 At high SNR, ρ approaches infinity, and ξ1 approaches
where y = The constraint on y, y ≥ ξ1 , will be explained
.
ρα22
later. By using the Rayleigh assumption, we have zero. Therefore, the function F(2R2 ) can be approximated as
 y follows:   2R2  
Q1 = e 2ξ1
e− max{ξ1 ,z}−z dz 2 −1 γ
ξ1 F(2R2 ) ≈ 2 − 2ξ1 = , (28)
  ρα22 ρ
1
= e2ξ1 e−2ξ1 − e−2y . (20) where γ = 2 (2 2−1) − 2 2 1 2 . By using the above
2R2
2 α2 α1 −1 α2
Q2 can be obtained similarly, and therefore, the CDF can be approximation, P(O) can be approximated as follows:
N   l
expressed as

 N γ −3lξ1  N−l
(2x −1)  P(O) ≈ e 1 − e−3ξ1
−2 l ρ l
F(x) = e2ξ1 e−2ξ1 − e ρα22 . (21) l=0
   N−l
1  N l
N
31
≈ N γ , (29)
It is important to point out the following:
    ρ l α12 − 1 α22
l=0
x = log 1 + ρ min |hn |2 , |gn,2 |2 α22 which results in the following corollary.
  Corollary 1: The two-stage RS scheme can realize a diver-
≥ log 1 + ρξ1 α22 , (22) sity gain of N, the maximal diversity gain for the addressed
scenario with N relays.
which is due to the fact that both |hn |2 and |gn,2|2 should be
larger than ξ1 , since relay n is in Sr . With this constraint, one B. Case Study When ξ1 = ξ2
can easily verify that
   Simulation results show that the two-stage relay selection
F log 1 + ρξ1 α22 = 0, (23) scheme outperforms the max-min scheme. However, for a spe-
and F(∞) = 1. With this CDF, the probability for O2 can be cial case with ξ1 = ξ2 , we can show that the two schemes
calculated as follows: achieve the same performance. The overall outage probability
N can be expressed
 asfollows:
 
P(O2 ) = P(xn∗ < 2R2 ||Sr | = l)P(|Sr | = l) P(O) = P |hn |2 < ξ1 + P |hn |2 < ξ2 , |hn |2 > ξ1
l=1  

N + P |gn,1 |2 < ξ1 , |hn |2 > ξ2 , |hn |2 > ξ1
= (F(2R2 ))l P(|Sr | = l). (24)  
+ P |gn,2 |2 < ξ1 , |gn,1 |2 > ξ1 , |hn |2 > ξ2 , |hn |2 > ξ1
l=1 
On the other hand, the probability of having l relays in Sr can + P |gn,2 |2 < ξ2 , |gn,2 |2 > ξ1 , |gn,1 |2 > ξ1 , |hn |2 > ξ2 ,
be calculated as follows: 
  N−l |hn |2 > ξ1 .
N    (30)
P(|Sr | = l) = 1 − P |hπ(n) |2 > ξ1 When ξ1 = ξ2 , we have
l  
n=1
    Po = P |hn |2 < ξ1 + P |gn,1|2 < ξ1 , |hn |2 > ξ1
× P |gπ(n),1 |2 > ξ1 P |gπ(n),2 |2 > ξ1  
+ P |gn,2 |2 < ξ1 , |gn,1 |2 > ξ1 , |hn |2 > ξ1 . (31)

N   
× P |hπ(n) |2 > ξ1 Note
 that the following
  equality holds: 
n=N−l+1 P |hn| < ξ1 + P |gn,1 |2 < ξ1 , |hn |2 > ξ1
2
       
× P |gπ(n),1 |2 > ξ1 P |gπ(n),2 |2 > ξ1 , = P min |gn,1|2 , |hn |2 < ξ1 . (32)
(25) By using this equality, the outage probability achieved by the
where π(·) denotes a random permutation of the relays. max-minapproach
 is given by  
Following steps similar to those used to obtain (9), the above Po = P min |gn,2|2 , |gn,1|2 , |hn |2 < ξ1 , ∀n ∈ {1, . . . , N}
probability can be obtained as     N
  N−l
P(|Sr | = l) =
N
1 − e−3ξ1 e−3lξ1 . (26) = 1 − P min |gπ(1),2|2 , |gπ(1),1|2 , |hπ(1)|2 > ξ1
l  N
By combining (8), (10), (24), and (26), and also applying some = 1 − e−3ξ1 , (33)
algebraic manipulations, the overall outage probability can be which is exactly the same as the expression given in Theorem 1
obtained as in the following theorem. by applying ξ1 = ξ2 .
DING et al.: RS FOR COOPERATIVE NOMA 419

Fig. 1. Comparison between cooperative OMA and NOMA with different Fig. 2. The outage probabilities achieved by the max-min relay selection
relay selection (RS) strategies. R1 = 0.5 bit per channel use (BPCU), R2 = 2 scheme and the two-stage one. N = 10.
BPCU, and α2 = 14 . The analytical results are based on Theorem 1.

scheme achieves the minimal outage probability. When the

C. Relay Selection Optimality number of relays is small, the performance gap between the
two relay selection schemes is small, and the use of more
The optimality of the proposed two-stage scheme is asserted
relays can increase this gap. One can also observe that the
in the following theorem.
simulation results perfectly match the analytical results devel-
Theorem 2: For the addressed cooperative NOMA scenario,
oped in Theorem 1, which demonstrates the accuracy of the
the two-stage relay selection scheme minimizes the overall
developed analytical results. Furthermore, when ξ1 = ξ2 , the
outage probability.
two relay selection schemes achieve the same performance, as
Proof: The theorem can be proved by contradiction. If there
discussed in the previous section.
exists a better strategy achieving a lower outage probability, an
event such that the use of relay n∗ causes outage, but no outage
occurs with the relay selected by the new strategy, denoted by V. C ONCLUSION
n̄∗ , n̄∗ = n∗ , should happen. Recall that for any relay selection In this paper, we have studied the impact of relay selec-
scheme, the outage event can be categorized as follows: tion on cooperative NOMA. The developed analytical results

O = O1 O2 . (34) have demonstrated that the two-stage scheme can achieve not
only the optimal diversity gain, but also the minimal outage
We focus only on the cases with |Sr | = 0, otherwise outage probability. Compared to the two-stage scheme, the max-min
always occurs, no matter which relay is used. When |Sr | > criterion can result in an increase in the outage probability.
0, one can conclude that relay n̄∗ must be in Sr , i.e., n̄∗ ∈
Sr , since otherwise outage occurs for sure by using relay n̄∗ .
R EFERENCES
According to (5), relay n∗ will not cause the outage event O1
as well, if |Sr | = 0. Now by using the criterion in (4) and the [1] Y. Saito et al., “Non-orthogonal multiple access (NOMA) for cellular
future radio access,” in Proc. IEEE Veh. Tech. Conf., Dresden, Germany,
definition of O2 , one can conclude that it is not possible that Jun. 2013, pp. 1–5.
relay n∗ causes O2 but relay n̄∗ does not, since relay n∗ is the [2] Z. Ding, Z. Yang, P. Fan, and H. V. Poor, “On the performance of
optimal solution to avoid O2 . The theorem is proved. non-orthogonal multiple access in 5G systems with randomly deployed
users,” IEEE Signal Process. Lett., vol. 21, no. 12, pp. 1501–1505,
Dec. 2014.
IV. N UMERICAL S TUDIES [3] S. Timotheou and I. Krikidis, “Fairness for non-orthogonal multiple
access in 5G systems,” IEEE Signal Process. Lett., vol. 22, no. 10,
In this section, the performance of cooperative NOMA pp. 1647–1651, Oct. 2015.
with the two relay selection strategies is evaluated by using [4] Z. Ding, M. Peng, and H. V. Poor, “Cooperative non-orthogonal
computer simulations. In Fig. 1, the performance of cooper- multiple access in 5G systems,” IEEE Commun. Lett., vol. 19, no. 8,
ative NOMA is compared with that of orthogonal multiple pp. 1462–1465, Aug. 2015.
[5] J.-B. Kim and I.-H. Lee, “Non-orthogonal multiple access in coordinated
access (OMA). For OMA, four time slots are needed, and direct and relay transmission,” IEEE Commun. Lett., vol. 19, no. 11,
the max-min criterion is used for relay selection. As can pp. 2037–2040, Nov. 2015.
be observed from Fig. 1, cooperative NOMA can efficiently [6] J. Men and J. Ge, “Non-orthogonal multiple access for multiple-
reduce the outage probability, and hence the use of cooperative antenna relaying networks,” IEEE Commun. Lett., vol. 19, no. 10,
pp. 1686–1689, Oct. 2015.
NOMA can offer a significant performance gain over OMA in [7] Y. Liu, Z. Ding, M. Elkashlan, and H. V. Poor, “Cooperative non-
terms of reception reliability. The reason for this performance orthogonal multiple access with simultaneous wireless information
gain is that the use of NOMA can ensure that two users are and power transfer,” IEEE J. Sel. Areas Commun., vol. 34, no. 4,
served simultaneously, whereas twice the resources, such as pp. 938–953, Apr. 2016.
[8] Y. Jing and H. Jafarkhani, “Single and multiple relay selection schemes
time slots, are needed for OMA to serve the two users. and their achievable diversity orders,” IEEE Trans. Wireless Commun.,
The performance difference between the max-min relay vol. 8, no. 3, pp. 1414–1423, Mar. 2009.
selection scheme and the two-stage one is also illustrated in [9] Z. Ding, L. Dai, and H. V. Poor, “MIMO-NOMA design for small
Figs. 1 and 2. When ξ1 = ξ2 , the two-stage relay selection packet transmission in the Internet of Things,” IEEE Access, vol. 4,
pp. 1393–1405, Aug. 2016.
scheme outperforms the max-min scheme, and this observation [10] T. M. Cover and J. A. Thomas, Elements of Information Theory, 6th ed.
is consistent with Theorem 2 which shows that the two-stage New York, NY, USA: Wiley, 1991.