Int. J. Electron. Commun.

(AEÜ) 114 (2020) 153028

Contents lists available at ScienceDirect

International Journal of Electronics and

Communications (AEÜ)
journal homepage: www.elsevier.com/locate/aeue

Regular paper

Max-min fairness in downlink non-orthogonal multiple access with

short packet communications
Fateme Salehi, Naaser Neda, Mohammad-Hassan Majidi ⇑
Faculty of Electrical and Computer Engineering, University of Birjand, Birjand, Iran

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, the performance of non-orthogonal multiple access (NOMA) technique for communication
Received 28 August 2019 based on finite blocklength (FBL) codes is studied. NOMA, by co-serving multi users has great potential
Accepted 4 December 2019 for achieving low-latency and better spectral efficiency rather than conventional orthogonal multiple
access (OMA) techniques. Our goal is to have optimal power allocation to attain fair throughput in a
DL-NOMA scenario. The optimization problem is formulated and an analytical approach using optimal
Keywords: solution search algorithm is introduced. Performance of the proposed NOMA system is compared with
URLLC
an OMA scheme which is benefiting from optimal power and blocklength allocation as well. Numerical
Finite blocklength
Short packet communication
results show that NOMA significantly improves the users’ fair throughput in FBL regime.
NOMA Ó 2019 Elsevier GmbH. All rights reserved.
Max-min fairness

1. Introduction to short blocklength [4]. In [5], Polyanskiy and et al. calculated

an exact approximation of transmission rate in the FBL regime
The fifth generation (5G) of mobile communication systems, for the AWGN channel. Triggered by that, research in this context
like its predecessor, supports human type communication for developed to multiple-input multiple-output quasi-static fading
enhanced mobile broadband (eMBB) services such as telephony, channels [6], and quasi-static fading channels with retransmission
multimedia, and mobile internet. Furthermore, two types of [7] as well. Also, the effect of short packets in spectrum sharing and
machine communications are considered in the 5G. The first one relaying networks were examined in [8] and [9] respectively.
is massive machine type communications (mMTC) to connect Scheduling of latency-critical packets was addressed in [10]. Fur-
everything that is benefiting from being connected and the next thermore, radio resource management in URLLC networks was
one is ultra-reliable low-latency communications (URLLC) for investigated in [11–16].
enabling mission-critical services [1,2]. Use cases of such services On the other hand, it’s well known that NOMA, as an efficient
are industrial automation, smart transportation system, smart grid, radio access technique, by serving multiple users simultaneously,
e-health, virtual/augmented reality, and tactile internet [3]. Each of has a great capability of supporting low-latency communications
these applications has a specific latency and reliability require- and in comparison with conventional OMA, significantly improves
ment which was compared together in [3]. In some applications the spectral efficiency. Performance of NOMA along with tech-
such as industrial automation, latency requirement is even less niques like beamforming and cognitive radio was addressed in
than 1 ms and reliability is about 99.9999999% (i.e., block error [17] and [18], respectively. A multiple-input single-output NOMA
rate (BLER) 10
9) [3]. cognitive radio network relying on simultaneous wireless informa-
To realize low-latency communications, short packets with tion and power transfer were studied in [19] and [20]. In [19], in
finite block length codes should be used to decrease the transmis- order to improve the security of the primary network, an
sion delay. This will result in a fundamentally different system artificial-noise-aided cooperative jamming scheme was proposed.
design and performance analysis. In fact the conventional Shan- In [20], robust beamforming and power splitting ratio were jointly
non’s information capacity, which assumes infinite blocklength of designed for minimizing the transmission power of the cognitive
data, is no longer applicable. This is because in the FBL regime, base station and maximizing the total harvested energy of the sec-
decoding error probability in the receiver can’t be negligible due ondary users. In [21], NOMA capability in a decode-and-forward
relaying system with energy harvesting nodes was considered.
⇑ Corresponding author. Furthermore OFDM-NOMA was recognized as a promising tech-
E-mail addresses: [email protected] (F. Salehi), [email protected] (N. Neda), nique for next generation wireless communication in [22], where
[email protected] (M.-H. Majidi).

https://doi.org/10.1016/j.aeue.2019.153028
1434-8411/Ó 2019 Elsevier GmbH. All rights reserved.
2 F. Salehi et al. / Int. J. Electron. Commun. (AEÜ) 114 (2020) 153028

it suggests that combining NOMA and short packets, is a promising  For comparison purposes, we consider a two user URLLC sce-
approach for low latency systems. nario with different reliability but the same delay requirements.
In [23], the performance of NOMA in short-packet communica- It demonstrates that the NOMA scheme compared to OMA with
tions was investigated and the optimal power allocation and block- same constraints, achieves higher fair throughput.
length were obtained asymptotically for high SNR scenarios. Also, a
closed form of achievable transmission latency reduction of NOMA The rest of the paper is organized as follows. In Section 2, the
compared to conventional OMA was characterized in [23]. In [24], system model and transmission strategies with NOMA are pro-
the transmission rates and power allocation of NOMA scheme were vided. Problem formulation in two forms of NOMA and OMA
optimized to make a trade-off between system throughput and schemes is presented in Section 3. Section 4 solves the optimiza-
user fairness. In that study, the effective throughput of the user tion problems. In Section 5, numerical results are provided. Finally,
with a higher channel gain was maximized while a predefined Section 6 concludes this work.
effective throughput target for the other user was guaranteed. In
[25], the energy of the transmitter was minimized subject to 2. System model and transmission strategies with NOMA
heterogeneous reliability and latency constraints at receivers. It
was shown that in contrast to conventional NOMA with homoge- As shown in Fig. 1(a), a DL communication system with two
neous constraints, successive interference cancellation (SIC) maybe URLLC users is considered. Since the FBL regime consists of short
unfeasible and therefore different interference management tech- packets transmission, it’s assumed that the channel is constant
niques were introduced accordingly. during one frame and it varies independently from one frame to
In this paper, we merge short packet communications with the next one. According to [5], the achievable data rate Ri at user
NOMA and investigate its advantage in achieving higher fair i, i 2 f1; 2g, for a finite blocklength mi symbols (mi P 100), and
throughput over its OMA counterpart subject to the same reliabil- an acceptable BLER ei , has an exact approximation as
ity, blocklength, and total transmit energy constraints. In contrast sffiffiffiffiffiffi
to the previous works, here we target to guarantee the QoS V i Q
1 ðei Þ log2 mi
Ri  C i
þ ð1Þ
requirements of URLLC, i.e., achieving ultra-reliability and low mi ln2 mi
latency and providing fair throughput among users at the same
time. Fair throughput is an important criterion that, to the best where C i ¼ log2 ð1 þ ci Þ is the Shannon capacity, ci is the SNR/SINR
of our knowledge, has not been considered in the short packet ratio of user i, Q
1 ðÞ is inverse of the Gaussian Q-function i.e.,
R1  
communication, so far. It should be noted that the blocklength is Q ðxÞ ¼ x p1ffiffiffiffi exp
t2 dt, and V i ¼ 1
ð1 þ ci Þ
2 is the channel
2

2p
directly related to the latency. We consider a typical MTC scenario
dispersion of a quasi-static Rayleigh fading channel. Therefore, the
in local area communications [26] with two users and one BS
decoding error probability at user i is given by
which is assumed has access to perfect channel state information
!
(CSI). The two users are assumed to be in one cluster and share ðC i
Ri þ log2 mi =mi Þln2
one resource block. To fully exploit the benefit of NOMA, the chan- ei Q pffiffiffiffiffiffiffiffiffiffiffiffiffi ,Q ðci ; Ri ; mi Þ ð2Þ
V i =mi
nel gains of the users in the same cluster must be significantly dif-
ferent. Unlike the conventional communication with infinite It’s assumed that the BS and each of the users have one antenna,
blocklength where SIC could be assumed errorless, in the FBL and two users are grouped into one cluster. The NOMA scheme
regime, it is not possible to have a perfect SIC. This leads to chal- allows the BS to simultaneously serve the users in one cluster by
lenges in optimal design of the NOMA scheme which is addressed using the entire blocklength via a superposition coding. It means
in this paper. Since the successful connection of each user can be that user multiplexing is performed in the power domain. In the
very critical, providing fairness between the users will be an receiver side, the user with stronger channel decodes and removes
important metric in URLLC and 5G systems [27]. the signal of the user with weaker channel by SIC technique and
In this work, like related works in [23–25], we consider a two- extracts its own data. In this way, user i can cancel the interference
user single antenna scheme which serves as the basic scenario of of the weaker users and its achievable rate related to the decoding
more general cases, like [28] and [29]. The researchers in [28] of xk is as follows, i; k 2 f1; 2g,
and [29] considered more general multiple antenna two-way relay sffiffiffiffiffiffiffi

1  
networks to design beamforming and energy-efficient optimiza- V i;k Q ei;k log2 mk
Ri;k  C i;k
þ ð3Þ
tion respectively. However they assumed infinite blocklength mk ln2 mk
codes which are not applicable in URLLC. Therefore, extending
our study to more general cases remains as a subject for future where ei;k is the decoding error of xk at user i. According to power
research. domain NOMA principle, in the two users scenario, BS transmits
P2 pffiffiffiffi
Our main contributions in this work are summarized as follows: i¼1 pi xi , where xi and pi are the message and the allocated power
of user i respectively. The received signal at user i is given by
 We specify the optimal design of the NOMA scheme by power pffiffiffiffiffi pffiffiffiffiffi
yi ¼ hi ð p1 x1 þ p2 x2 Þ þ ni ð4Þ
allocation between two users to achieve fairness. Maximizing
the minimum throughput is considered as the fairness criterion. where ni is the complex additive white Gaussian noise with vari-
To the best of our knowledge, max-min fairness problem in the ance r2i , and hi shows the channel coefficient between BS and user
FBL regime has not been addressed yet. A moderate approach to 2 2
i. Without loss of generality, it’s assumed that jh1 j =r21 > jh2 j =r22 .
solve the complicated formulation of achievable rate in pres- User 2 directly detects x2 by considering x1 as interference, and
ence of imperfect SIC is introduced. the received SINR of x2 at user 2 is given by
 To show the advantages of the NOMA scheme in the context of
2
FBL regime, its optimal OMA counterpart is considered. To this p2 jh2 j
c2;2 ¼ 2
ð5Þ
end, users’ power and blocklength are jointly optimized and p1 jh2 j þ r22
optimal transmission rate is obtained.
 The computational complexity of max-min fairness design algo- The decoding error probability of x2 at user 2 is denoted by e2;2 ,
rithms in NOMA and OMA schemes are determined. where based on (2) is approximated by
 F. Salehi et al. / Int. J. Electron. Commun. (AEÜ) 114 (2020) 153028 3

(a) System model

P2
P1 P2
P1
m m1 m2
(b) NOMA frame structure (c) OMA frame structure
Fig. 1. System model and frame structures.

 
e2;2  Q c2;2 ; R2;2 ; m2 ð6Þ fairness. The throughput of user i, T i , is defined as the average bits
per each complex symbol (or channel use), which is decoded cor-
Since x2 is directly detected, e2;2 is the overall decoding error rectly at the receiver;
probability at user 2, i.e., e2 ¼ e2;2 . On the other hand, user 1 per- mi
forms SIC. It first decodes x2 while x1 treats as interference and Ti, Ri ð1
ei Þ ð12Þ
Dmax
after that, x1 is detected normally. So, the received SINR of x2 at
user 1 is given by where Ri  Ri;i and 1
ei is defined as the reliability of user i, and it’s
2 acceptable amount is defined by URLLC application.
p2 jh1 j
c1;2 ¼ 2
ð7Þ
p1 jh1 j þ r21 3.1. Optimization problem in NOMA
Similarly, the decoding error probability of x2 at user 1 is
approximated by In the NOMA scheme, the superposition coding is performed at
  the transmitter such that the BS enables to transmit users’ signals
e1;2  Q c1;2 ; R1;2 ; m2 ð8Þ with different power simultaneously, so m1 ¼ m2 ¼ m. In Fig. 1(b) a
frame of NOMA is presented. Therefore, the desired optimization
It’s a reasonable assumption that x1 can’t be detected correctly problem is formulated as follows
if x2 is in error. However, if user 1 successfully decode and remove
x2 , following (4), the received SNR of x1 is given by max minfT 1 ; T 2 g ð13aÞ
p1 ;p2

2
p1 jh1 j
c1;1 ¼ ð9Þ s:t: p1 þ p2 6 Pave ð13bÞ
r 2
1

Accordingly, the decoding error probability of x1 at user 1, i.e., pi > 0; i 2 f1; 2g ð13cÞ
e1;1 , is denoted by
  ei 6 ei th ; i 2 f1; 2g ð13dÞ
e1;1  Q c1;1 ; R1;1 ; m1 ð10Þ

Based on the above discussion, it’s clear that the overall decod- m ¼ Dmax ð13eÞ
ing error probability at user 1 can be approximated as From (13a) it’s clear that the optimization variables are the allo-
e1 ¼ e1;2 þ ð1
e1;2 Þe1;1  e1;2 þ e1;1 ð11Þ cated powers of two users subject to some constraints. Constraints
(13b) and (13c) are to guarantee the practical power allocations,
where we took into account the fact that in URLLC services ei;k is in where Pave shows the upper limit of average power of transmitter.
order of 10
5 ~ 10
9. (13d) guarantees that the preferred reliability of the user i, doesn’t
violate than ei th . Moreover, finite blocklength of the two users is
3. Problem formulation stated by (13e).

In the considered scenario, BS should serve two users within 3.2. Optimization problem in OMA
Dmax symbols period to provide fairness subject to the total trans-
mit energy. If a channel feedback is available at the transmitter As a benchmark for comparison, we formulate the optimization
side, users’ rates can be set according to their instantaneous chan- problem for an equivalent OMA scheme as well, where the two
nel conditions. In this case, a suitable criterion is the max-min users are served in orthogonal (i.e., different) channel uses.
4 F. Salehi et al. / Int. J. Electron. Commun. (AEÜ) 114 (2020) 153028

Fig. 1(c) demonstrates the OMA frame structure. Now, the received To obtain the optimal solution of problems (13) and (17), we
signal at user i, i 2 f1; 2g, is given by propose an approach according to the above propositions and
pffiffiffiffi corollary.
yi ¼ hi pi xi þ ni ð14Þ
In (14), because of orthogonality, the users have no interference 4.1. Design of Max-Min fairness in NOMA
on each other, so, the SNR for each user is given by
2 Based on the discussion presented above, one should note the
pi jhi j following points:
ci ¼ ð15Þ
r 2
i

As a result, following (2), the decoding error probability of user i & According to Corollary 1; c2;2 , c1;2 , and c1;1 presented in (5), (7),
is approximated by and (9) respectively are just functions of p1 .
& Based on Proposition 2, the optimal solution only can be
ei  Q ðci ; Ri ; mi Þ ð16Þ achieved while T 1 ¼ T 2 . Using (12) and the fact that in the
 
Finally, the optimization problem of the OMA transmission 1
eth
NOMA scheme m1 ¼ m2 ¼ Dmax , this leads to R1;1 ¼ 1
e2th R2;2 .
scheme is formulated as 1

& A certain amount of achievable bit rate for user 2 is considered

max minfT 1 ; T 2 g ð17aÞ
fmi ;pi gi¼1;2 at the receiver 1 or 2. This means that R1;2 ¼ R2;2 .

s:t: m1 p1 þ m2 p2 6 Dmax P ave ð17bÞ Therefore, the optimization problem for NOMA in (13), can be
rearranged as follows
0 < pi 6 jPave ; i 2 f1; 2g ð17cÞ max R2;2 ð18aÞ
p1

ei 6 ei th ; i 2 f1; 2g ð17dÞ    
s:t: Q c1;2 ; R1;2 ; Dmax þ Q c1;1 ; R1;1 ; Dmax ¼ eth
1 ð18bÞ
m1 þ m2 ¼ Dmax ð17eÞ
 
where Ri in (12) defined by (1). Now the optimization parameters Q c2;2 ; R2;2 ; Dmax ¼ eth
2 ð18cÞ
are the blocklength and power allocated to the two users. Con-
straints (17b)–(17d) are similar to (13b)–(13d) respectively. The  
1
eth
coefficient j in (17c) shows the peak to average power ratio (PAPR) R1;2 ¼ R2;2 ; R1;1 ¼ 2
R2;2 ð18dÞ
1
eth
factor. Also, (17e) is to guarantee that the two scenarios have the 1

same overall latency.

ð5Þ; ð7Þ; and ð9Þ ð18eÞ
4. Problem solving Although it is not possible to obtain p1 in a closed-form, how-
ever by searching algorithms like bisection method [30], problem
In this section we are going to introduce a step by step approach (18) and equivalently the optimization problem (13) can be solved.
to solve the optimization problems in (13) and (17). To simplify the
approach, we have to first analyze the constraints and determine Proposition 3:. which is defined as
their optimal status. Let’s first consider the constraint (13d) and    
(17d) on acceptable BLER of two users. Since each URLLC applica- ERðp1 Þ,Q c1;2 ; R1;2 ; Dmax þ Q c1;1 ; R1;1 ; Dmax ð19Þ
tion needs a specific reliability, allocating more resources to get
is a convex function with respect to p1 . (See Appendix D for
BLER lower than the required eth i , wastes the resources. Therefore, proof.)
ei ¼ eth is an optimal selection. In addition, it’s easy to show that
i
  Proposition 3 indicates that there are at most two p1 values that
the throughput presented in (12), that is T i ¼ 1
eth i Ri mi =Dmax ,
is a monotonic function of SNR/SINR ci . This is stated by Proposi- guarantee ER0 ðp1 Þ 6 eth1 . The smaller value which satisfies

tion 1, as follows: ER0ðp1 Þ 6 0, is the optimal solution p1 that assures T 1 ¼ T 2 . Note
that in NOMA scenario, it is always assumed that the two users
Proposition 1:. The acceptable data rate (i.e., Ri > 0) in (1), where which share the same resources, experience different channel
2 2
ei ¼ eth gains and since we suppose that jh1 j =r21 > jh2 j =r22 , then we
i , is a monotonically increasing function of SNR/SINR ci . (See
Appendix A for proof.) should have p1 < p2 . The power allocation algorithm in the NOMA
Based on the Proposition 1, the inequality in (13b) or (17b) can scheme is demonstrated in Table 1.
be changed to equality for the optimal solution using the following Table 1
corollary. Power allocation algorithm in the NOMA scheme.

Input: total blocklength Dmax , overall BLER of user i eth i , BS average power
Corollary 1:. To maximize minfT 1 ; T 2 g, equality in the energy P ave , required accuracy 2 .
constraint, i.e., m1 p1 þ m2 p2 ¼ Dmax Pave , is always guaranteed. (See Output: optimum power p1 , p2 , and fair throughput T 1 ¼ T 2 ¼ T 0  .
Appendix B for proof.) 1. Set p1 ¼ 0.
Corollary 1 significantly simplifies the power allocation at the 2. while ER > eth 1 do
3. Set p1 ¼ minðp1 þ Dp; P ave Þ.
BS, since p2 is now related to p1 directly.
4. Calculate c2;2 , c1;2 , and c1;1 presented in (5), (7), and (9) respectively.
Finally, Proposition 2 is essential for simplifying the solutions 5. Calculate R2;2 , R1;2 , and R1;1 using (18c) and (18d) respectively.
which is expressed here. 6. end while
7. Set plb 1
Dp,

1 ¼ p 1 ¼ p1 .
and pub
Proposition 2:. The optimum resource allocation in (13) and (17) 8. Find p1 2 plb 1 ; p1
ub
that satisfies ER ¼ eth 1 via bisection method with
required accuracy 2 .
should always get T 1 ¼ T 2 . (See Appendix C for proof.)  
9. return p1 ¼ p1 , p2 ¼ P ave
p1 and T 0 ¼ 1
eth
2 R2;2 .
 F. Salehi et al. / Int. J. Electron. Commun. (AEÜ) 114 (2020) 153028 5

4.2. Design of max-min fairness in OMA where ERð jÞ is the result obtained after j iterations. According to
[11,27], power allocation for two NOMA users through a bisection
In the OMA scheme, based on the aforementioned discussion, procedure constraint to desired accuracy 2 (Table 1), has computa-
  
the following points are considered tional complexity of O log2 eth1 =e . On the other hand, power and
blocklength allocation among two OMA users through the algo-
& According to Corollary 1 and (17e); c1 and c2 determined in (15) rithm in Table 2 subject to desired accuracy 2, has computational
are functions of only p1 and m1 .   
1 =e . It reveals that the computational
complexity of O Dmax log2 eth
& Based on Proposition 2 and (12), one can claim in the optimal
   complexity of the OMA scheme is Dmax times of those with the
1
eth
design R1 ¼ 1
e2th m 2
m1
R2 . NOMA scheme.
1

Therefore, the optimization problem for OMA in (17), can be 5. Numerical results
rearranged as follows
max R2 ð20aÞ In this section, performance of the proposed NOMA scheme and
m1 ;p1
its OMA counterpart are evaluated through the numerical results.

7
A heterogeneous network is supposed with users BLER 1 ¼ 10
eth
s:t: Q ðc1 ; R1 ; m1 Þ ¼ eth
1 ð20bÞ
5
2 ¼ 10 . PAPR factor is considered as j ¼ 1:2. Furthermore,
and eth
the noise power at each receiver is assumed as r21 ¼ r22 ¼ r2 .
Q ðc2 ; R2 ; m2 Þ ¼ eth
2 ð20cÞ
Numerical results are provided based on fixed channel gains and
   required accuracy 2 ¼ 10
15 .
1
eth m2
R1 ¼ 2
R2 ð20dÞ In order to realize the relation between the objective function
1
eth
1
m1
and parameters of interest we consider the normalized channel
2
m1 þ m2 ¼ Dmax ; and ð15Þ ð20eÞ gains of two users, i.e., jhi j =r2 , set to be fixed. For instance
2 2
jh1 j =r2 ¼ 0:8 and jh2 j =r2 ¼ 0:1. Also, it is assumed that
We consider the decoding error of user 1, as Pave ¼ 10 W and Dmax ¼ 200 channel uses, unless otherwise stated.
ERðm1 ; p1 Þ,Q ðc1 ; R1 ; m1 Þ. The above optimization problem con- In Fig. 2, the plot of user 1 decoding error, i.e., ERðp1 Þ in (19), is
sists of two independent variables p1 and m1 . Because of nonlinear indicated versus p1 . It’s observed there are two p1 values that satisfy
property of constraints, the feasible solutions which guarantee
7
1 ¼ 10 , where according to the outcome of Proposition 3, the
eth
Q ðc1 ; R1 ; m1 Þ ¼ eth
1 are introduced using a searching algorithm via smaller value is optimal. It should be noted that due to the channel
taking m1 as a constant in each iteration and attaining p1 . gains of the two users and limited power, the reliability of user 1
maybe never satisfied. In that case the solution does not exist.
Remark:. It’s clear that m2 ¼ Dmax
m1 and In Fig. 3, the two users’ throughput achieved by the NOMA

  
 
p2 ¼ Dmax Pave
m1 p1 =m2 . However, after determining the opti- scheme, defined in (12) and (3), is plotted versusp1 . It is clear that
mal solution, if p2 > jPave , we choose the nearest solution to p2 T 1 and T 2 are monotonically increasing and decreasing functions,
among other feasible solutions, which doesn’t violet the power respectively in the reasonable range (i.e., Ti > 0). Their intersection
bound. is the solution of the max-min fairness optimization problem. This
In Table 2, the power and blocklength allocation algorithm in means that with power allocation in about p1 ¼ 1 W and p2 ¼ 9W,
the OMA scheme is demonstrated. both users have equal throughput as T 0 ¼ 0:52 bits=ch:use.
The two users’ throughput defined in (12) and (1), are also plot-
4.3. Computational complexity ted versus p1 and m1 in Figs. 4 and 5 respectively. From these fig-
ures, it is clear that in the reasonable range, T 1 and T 2 are
The complexity of a searching algorithm depends on the stop- respectively monotonically increasing and decreasing functions
ping criterion. For both algorithms presented in Tables 1 and 2, in both p1 and m1 . The intersection of the two curves in Fig. 4 with
 
the iterations stop, if the condition ERð jÞ
ERð jþ1Þ  < e is satisfied,

Table 2
Power and blocklength allocation algorithm in the OMA scheme.

Input: total blocklength Dmax , overall BLER of user i eth i , BS average power
P ave , required accuracy 2 .
Output: optimum power pi and blocklength mi allocated to user i, and fair
throughput T 1 ¼ T 2 ¼ T 0  .
1. Set p1 ¼ jP ave .
2. form1 ¼ 1 : Dmax
3. if ER > eth 1 do nothing
4. else
5. while ER < eth 1 do
6. Set p1 ¼ maxðp1
Dp; 0Þ.
7. Calculate c1 , and c2 presented in (15).
8. Calculate R2 , and R1 using (20c) and (20d) respectively.
9. end while
1 ¼ p 1 and p
1 ¼ p1 þ Dp.
Set plb ub
10.
11. Find p1 2 plb 1 ; p1
ub
that satisfies ER ¼ eth
1 via bisection method with
required accuracy 2 .
12. end if
13. end for
  
14. return m1 ; p1 ¼ argmaxR2 , m2 ¼ Dmax
m1 , p2 ¼ Dmax P ave
m1 p1 =m2 ,
   
and T 0 ¼ 1
eth
2 m 2 =D max R2 .
Fig. 2. User 1 decoding error, ERðp1 Þ, versus p1 in the NOMA scheme.
6 F. Salehi et al. / Int. J. Electron. Commun. (AEÜ) 114 (2020) 153028

fixed and optimal blocklength value, shows the p1 . Similarly, the
intersection of the two curves in Fig. 5 with fixed and optimal
power value, shows the m1 .
In Fig. 6, we plot the feasible solution set which guarantees the
reliability requirement of user 1 in the OMA scheme (i.e.,
ERðm1 ; p1 Þ¼ eth 1 ). Since the blocklength could only be an integer
within the range of 1 to Dmax , the feasible solutions are limited.
Among the feasible solutions, that pair that results to maximum
throughput and simultaneously does not lead to power limit viola-
tion is optimal, which is demonstrated by a red point in Fig. 6. We
notice that this point which is equal to p1 ¼ 5:5 W and m1 ¼ 62 is
the same cross point of the curves in Figs. 4 and 5.
In Figs. 7 and 8, the throughput T 1 ¼ T 2 ¼ T 0 achieved by the
OMA scheme versus p1 and m1 is illustrated respectively. It’s seen
that both are concave and have one global maximum, T 0 ¼ 0:5,
which determines the optimal values of p1  and m1 . These optimal
values, i.e., p1  and m1 , are determined by red points in Figs. 7 and 8
respectively.
In Fig. 9, the effect of maximum blocklength, Dmax , on the
Fig. 3. The two users’ throughput achieved by the NOMA scheme versus p1 . achievable fair throughput in the NOMA scheme is assessed and
compared with the results of optimal OMA scheme. It’s observed
the proposed NOMA scheme effectively outperforms the OMA,
regardless of blocklength.

Fig. 4. The two users’ throughput achieved by the OMA scheme versus p1 .

Fig. 6. The feasible solution set achieved by the OMA scheme.

Fig. 5. The two users’ throughput achieved by the OMA scheme versus m1 . Fig. 7. Throughput T 0 achieved by the OMA scheme versus p1 .
 F. Salehi et al. / Int. J. Electron. Commun. (AEÜ) 114 (2020) 153028 7

In Fig. 10, the effect of average power, Pave , on the achievable

fair throughput is investigated. The superiority of the NOMA
scheme with respect to the OMA is notable in this graph as well.

6. Conclusion

In this paper, the combination of NOMA with short packets was

considered in critical-IoT, where low latency is a requirement. On
the other hand, achieving fairness in URLLC is essential to guaran-
tee the QoS for all users of critical services. For a DL-NOMA system
with two users, an optimization problem was formulated to
address the fairness by optimizing power and rate subject to total
energy, reliability and latency constraints. To simplify the design of
optimal power allocation and users transmission rate, optimal
states of inequality constraints were determined first and then,
an analytical approach with an optimal solution search algorithm
was proposed. The OMA was considered as an evaluation scheme
of the proposed system. An analytical approach for optimal power
and blocklength allocation was introduced for OMA, as well.
Fig. 8. Throughput T 0 achieved by the OMA scheme versus m1 .
Numerical results show that the proposed NOMA scheme improves
the users’ fair throughput compared to the OMA with the same
constraints, significantly. In addition, the computational complex-
ity of the proposed algorithms was studied. It was indicated that
design of NOMA with respect to OMA has lower complexity,
because unlike the OMA scheme it doesn’t need the blocklength
optimization. However, some concepts like jointly design of users
clustering and transmission strategy in a multi-user scenario as
well as extension of the proposed approach to a relaying enabled
system remain for the future studies.

Declaration of Competing Interest

We declare that we have no conflict of interest.

Appendix A: Proof of proposition 1

Proof. To prove that the reasonable Ri in (1) is monotonically

increasing function of the corresponding SNR/SINR, i.e., ci , i 2 f1; 2g,
the partial derivative of Ri with respect to ci is derived as
Fig. 9. Maximum achievable fair throughput by the NOMA and OMA schemes 0 1
versus Dmax . e
@Ri B
log2 d C
¼ @1
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiA ðA:1Þ
@ ci 1 þ ci 2
ð1 þ ci Þ ð1 þ ci Þ
1
pffiffiffiffiffiffi
where d ¼ Q
1 ðei Þ= mi . From @Ri =@ ci ¼ 0, the answer is given by
  pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1=2
c^i ¼ 12 1 þ 1 þ 4d2 ^i , we have

1. In the range of ci > c
@Ri =@ ci > 0, so Ri is monotonically increasing function, and in the
range of 0 < ci < c ^i , we have @Ri =@ ci < 0, i.e., Ri is monotonically
decreasing function. On the other hand, for ci ¼ 0, we have
Ri  log2 ðmi Þ=mi  0 for mi P 100. As a result, in the range of
0 < ci < c ^i , we would have Ri < 0 which is unacceptable. So one
can suppose that Ri is a monotonically increasing function of ci . It
should be noted that in general Ri can be Ri;k , and also ci can be
ci;k . h

Appendix B: Proof of corollary 1 (according to [16, Appendix B])

Proof:. We first suppose that optimal power allocation py1 and py2 ,
satisfying m1 py1 þ m2 py2 < Dmax Pave , gives the maximum of
minfT 1 ; T 2 g which is denoted by T y0 .
We increase and by py1 py2
 
Fig. 10. Maximum achievable fair throughput by the NOMA and OMA schemes multiplying in a scalar value a ¼ Dmax P ave = m1 py1 þ m2 py2 to attain
versus P ave .
8 F. Salehi et al. / Int. J. Electron. Commun. (AEÜ) 114 (2020) 153028

[2] Bhoyar P, Sahare P, Dhok SB, Deshmukh RB. Communication technologies and
p1 ¼ apy1 and p2 ¼ apy2 which satisfies m1 p1 þ m2 p2 ¼ Dmax Pave . We security challenges for Internet of things: a comprehensive review. AEU – Int J
note that since > 1, so p1 > py1 and p2
a > py2 . According to (9) and Electron Commun 2019;99:81–99.
[3] Sutton GJ, Zeng J, Liu RP, Ni W, Nguyen DN, Jayawickrama BA, et al. Enabling
(5), we have respectively 1 > y1 and c c technologies for ultra-reliable and low latency communications: from PHY and
MAC layer perspectives. IEEE Commun Surv Tutor 2019;21(3):2488–524.
py2 jh2 j py2 jh2 j
2 2 2
p2 jh2 j https://doi.org/10.1109/COMST.973910.1109/COMST.2019.2897800.
c2 ¼ 2
¼ r2
> y 2
¼ cy2 ðB:1Þ [4] Durisi G, Koch T, Popovski P. Toward massive, ultrareliable, and low-latency
p1 jh2 j þ r22 py1 jh2 j þ a2
2
p1 jh2 j þ r22 wireless communication with short packets. Proc IEEE 2016;104(9):1711–26.
[5] Polyanskiy Y, Poor HV, Verdú S. Channel coding rate in the finite blocklength
This shows that as py1 and py2 increase to p1  and p2  respectively, regime. IEEE Trans Inf Theory 2010;56(5): 2307–59.
[6] Yang W, Durisi G, Koch T, Polyanskiy Y. Quasi-static multiple antenna fading
ci increases which causes to Ri increase according to the Proposi- channels at finite blocklength. IEEE Trans. Inf. Theory Jul. 2014;60(7):4232–65.
tion 1. As a result, the answer increases from T y0 to T 0 , where con- [7] Wu P, Jindal N. Coding versus ARQ in fading channels: how reliable should the
PHY be?. IEEE Trans Commun 2011;59(12):3363–74.
tradicts with the claim of optimality T y0 that is the maximum of [8] Makki B, Svensson T, Zorzi M. Finite block-length analysis of the incremental
minfT 1 ; T 2 g. So we conclude that m1 p1 þ m2 p2 ¼ Dmax Pave is always redundancy HARQ. IEEE Wireless Commun Lett 2014;3(5):529–32.
[9] Hu Y, Schmeink A, Gross J. Blocklength-limited performance of relaying under
guaranteed in the optimal solution of problems (13) and (17). h
quasi-static rayleigh channels. IEEE Trans Wireless Commun; 2016. p. 1–1.
[10] Xu S, Chang T-H, Lin S-C, Shen C, Zhu G. Energy-efficient packet scheduling
Appendix C: Proof of proposition 2 with finite blocklength codes: convexity analysis and efficient algorithms. IEEE
Trans Wireless Commun 2016;15(8):5527–40.
[11] She C, Yang C, Quek TQS. Cross-layer optimization for ultrareliable and low-
Proof:. To prove in the optimal solution T 1 ¼ T 2 ¼ T 0  , we first latency radio access networks. IEEE Trans Wireless Commun 2018;17
(1):127–41.
prove in the allowed SNR/SINR which causes to Ri > 0, T 1 and T 2 [12] Sun C, She C, Yang C, Quek TQS, Li Y, Vucetic B. Optimizing resource allocation
are monotonically increasing and decreasing functions of p1 in the short blocklength regime for ultra-reliable and low-latency
respectively. As per the chain rule, @T 1 =@p1 ¼ ð@T 1 =@ c1 Þð@ c1 =@p1 Þ, communications. IEEE Trans Wireless Commun 2019;18(1):402–15. https://
doi.org/10.1109/TWC.2018.2880907.
which the partial derivative of c1 with respect to p1 is always
[13] She C, Yang C, Quek TQS. Joint uplink and downlink resource configuration for
positive according to (8). As well, pursuant to Proposition 1 one can ultra-reliable and low-latency communications. IEEE Trans Commun 2018;66
 
say that @T i =@ ci ¼ 1
ethi ðmi =Dmax Þ@Ri =@ ci > 0 and following
(5):2266–80.
[14] Hou Z, She C, Li Y, Quek TQS, Vuectic B. Burstiness aware bandwidth
@T 1 =@p1 > 0. reservation for ultra-reliable and low-latency communications in tactile
Similarly we have @T 2 =@p1 ¼ ð@T 2 =@ c2 Þð@ c2 =@p1 Þ. As for (5) and Internet. IEEE J Select Areas Commun 2018;36(11):2401–10.
alsop2 ¼ P ave
p1 , one can write [15] She C, Chen Z, Yang C, Quek TQS, Li Y, Vucetic B, lmproving network
availability of ultra-reliable and low-latency communications with multi-
  connectivity. IEEE Trans Commun 2018; 66(II): 5482–96.
2 2
@ c2
Pave jh2 j þ r2 jh2 j
2
[16] Hu Y, Ozmen M, Gursoy MC, Schmeink A. Optimal power allocation for QoS-
¼  2 <0 ðC:1Þ constrained downlink multi-user networks in the finite blocklength regime.
@p1 2
p1 jh2 j þ r22 IEEE Trans Wireless Commun 2018;17(9):5827–40.
[17] Chinnadurai S, Selvaprabhu P, Jeong Y, Sarker AL, Hai H, Duan W, et al. User
clustering and robust beamforming design in multicell MIMO-NOMA system
According to this and former result it can be concluded
for 5G communications. AEU - Int J Electron Commun 2017;78:181–91.
@T 2 =@p1 < 0. Now the amount of T 1 and T 2 in p1 ¼ 0 is calculated. [18] Emam S, Çelebi ME. Non-orthogonal multiple access protocol for overlay
If T 2 > T 1 , this two functions have one cross point that would be cognitive radio networks using spatial modulation and antenna selection. AEU
- Int J Electron Commun 2018;86:171–6.
the optimal solution. In p1 ¼ 0 we have R1  log2 m1 =m1 and
qffiffiffiffiffi Q
1 eth [19] Zhou F, Chu Z, Sun H, Hu RQ, Hanzo L. artificial noise aided secure cognitive
ð Þ 2
R2  log2 ð1 þ c2 Þ
m12 ln2 2 þ logm22m2 where c2 ¼ P ave jh2 j =r22 . beamforming for cooperative MISO-NOMA Using SWIPT. IEEE J Sel Areas
Commun 2018;36(4):918–31.
Since in the NOMA scheme m1 ¼ m2 , clearly R2 > R1 . Finally, due [20] Sun H, Zhou F, Hu RQ, Hanzo L. Robust beamforming design in a NOMA
cognitive radio network relying on SWIPT. IEEE J Sel Areas Commun 2019;37
1  e2 it can be derived that T 2 > T 1 . Therefore, minfT 1 ; T 2 g
to eth th
(1):142–55.
is maximized when T 1 ¼ T 2 ¼ T 0  . This would be provable for the [21] Hoang TM, Son VV, Dinh NC, Hiep PT. Optimizing duration of energy
OMA scheme too, where due to similarity and page limit has been harvesting for downhink NOMA full-duplex over Nakagami-m fading
channel. AEU – Int J Electron Commun 2018;95:199–206.
omitted. h
[22] Trivedi VK, Ramadan K, Kumar P, Dessouky MI, Abd El-Samie FE. Enhanced
OFDM-NOMA for next generation wireless communication: a study of PAPR
Appendix D: proof of proposition 3 reduction and sensitivity to CFO and estimation errors. AEU - Int J Electron
Commun 2019;102: 9–24.
[23] Yu Y, Chen H, Li Y, Ding Z, Vucetic B. On the performance of non-orthogonal
Proof:. ERðp1 Þ is sum of the two functions e1;1 and e1;2 . Asymp- multiple access in short-packet communications. IEEE Commun Lett 2018;22
(3):590–3.
totically when p1 ! 0, the decoding error probability of x1 data [24] Sun X, Yan S, Yang N, Ding Z, Shen C, Zhong Z. Short-packet downlink
stream by user 1 is at most, i.e., e1;1 ! 1. In contrast to this, the transmission with non-orthogonal multiple access. IEEE Trans Wireless
decoding error probability of x2 data stream by user 1 is at least, Commun 2018;17(7):4550–64.
[25] Xu Y, Shen C, Chang TH, Lin SC, Zhao Y, Zhu G. Energy-efficient non-orthogonal
i.e., e1;2 ! 0, because the most portion of P tot is allocated to user 2. transmission under reliability and finite blocklength constraints. In: 2017 IEEE
In other word, e1;1  e1;2 and can be said ERðp1 Þ  e1;1 . By Globecom Workshops (GC Wkshps); Dec. 2017.
increasing p1 , e1;1 decreases but e1;2 increases, till e1;1 ¼ e1;2 . Finally [26] Feng D, She C, Ying K, Lai L, Hou Z, Quek TQS, et al. Towards ultra-reliable low-
latency communications: typical scenarios, possible solutions, and open
at high p1 (i.e., when p2 ! 0), with the same argument e1;2  e1;1 issues. IEEE Vehicular Technology Magazine, early access; Jan. 2019.
and can be said ERðp1 Þ  e1;2 . As we know, e1;1 and e1;2 are [27] Timotheou S, Krikidis I. Fairness for non-orthogonal multiple access in 5G
monotonically decreasing and increasing functions of p1 respec- systems. IEEE Signal Process Lett 2015;22(10):1647–51.
[28] Li Q, Yang L. Robust optimization for energy efficiency in MIMO two-way relay
tively in range of ð0; 1 . As a result, ERðp1 Þ in terms of p1 is first networks with SWIPT. IEEE Syst J 2019; 1–12.
decreasing and then increasing, therefore it’s a convex function. h [29] Li Q, Yang L. Beamforming for cooperative secure transmission in cognitive
two-way relay networks. IEEE Trans Inform Forensic Secur 2020;15:130–43.
https://doi.org/10.1109/TIFS.1020610.1109/TIFS.2019.2918431.
References [30] Boyd S, Vandenberghe L. Convex optimization. Cambridge, U.K.: Cambridge
Univ. Press; 2009.
[1] Sachs J, Wikström G, Dudda T, Baldemair R, Kittichokechai K. 5G radio network
design for ultra-reliable low-latency communication. IEEE Network 2018;32
(2):24–31.