2294 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO.

3, THIRD QUARTER 2018

A Survey of Non-Orthogonal
Multiple Access for 5G
Linglong Dai , Senior Member, IEEE, Bichai Wang , Student Member, IEEE, Zhiguo Ding , Member, IEEE,
Zhaocheng Wang , Senior Member, IEEE, Sheng Chen , Fellow, IEEE, and Lajos Hanzo , Fellow, IEEE

Abstract—In the fifth generation (5G) of wireless N OMENCLATURE

communication systems, hitherto unprecedented require-
ments are expected to be satisfied. As one of the promising
1G First generation mobile system
techniques of addressing these challenges, non-orthogonal 2G Second generation mobile system
multiple access (NOMA) has been actively investigated in recent 3G Third generation mobile system
years. In contrast to the family of conventional orthogonal 4G Fourth generation mobile system
multiple access (OMA) schemes, the key distinguishing feature of 5G Fifth generation mobile system
NOMA is to support a higher number of users than the number AWGN Additive white Gaussian noise
of orthogonal resource slots with the aid of non-orthogonal
resource allocation. This may be realized by the sophisticated BDM Bit division multiplexing
inter-user interference cancellation at the cost of an increased BER Bit error rate
receiver complexity. In this paper, we provide a comprehensive BLER Block error ratio
survey of the original birth, the most recent development, BPSK Binary phase shift keying
and the future research directions of NOMA. Specifically, the BS Base station
basic principle of NOMA will be introduced at first, with
the comparison between NOMA and OMA especially from CDMA Code division multiple access
the perspective of information theory. Then, the prominent CIR Channel impulse response
NOMA schemes are discussed by dividing them into two CoMP Coordinated multi-point
categories, namely, power-domain and code-domain NOMA. CS Compressive sensing
Their design principles and key features will be discussed in CSI Channel state information
detail, and a systematic comparison of these NOMA schemes
will be summarized in terms of their spectral efficiency, system
CTU Contention transmission unit
performance, receiver complexity, etc. Finally, we will highlight DEA Differential evolution algorithm
a range of challenging open problems that should be solved DF Decision feedback
for NOMA, along with corresponding opportunities and future EA Evolutionary algorithm
research trends to address these challenges. ESE Elementary signal estimator
Index Terms—5G, non-orthogonal multiple access (NOMA), EXIT Extrinsic information transfer-function
multi-user detection (MUD), spectral efficiency, massive connec- FDMA Frequency division multiple access
tivity, overloading, low latency. GA Genetic algorithm
i.i.d independent and identically distributed
Manuscript received October 11, 2017; revised March 20, 2018; accepted IDMA Interleave-division multiple access
May 4, 2018. Date of publication May 11, 2018; date of current ver- IoT Internet of Things
sion August 21, 2018. This work was supported in part by the National IrCC Irregular convolutional code
Natural Science Foundation of China for Outstanding Young Scholars under
Grant 61722109, in part by the National Natural Science Foundation of China ISI Inter-symbol-interference
under Grant 61571270, in part by the Royal Academy of Engineering under JMPA Joint message passing algorithm
the U.K.–China Industry Academia Partnership Programme Scheme under KPI Key performance indicators
Grant U.K.-CIAPP\49. The work of Z. Ding was supported in part by the U.K.
EPSRC under Grant EP/ L025272/1 and in part by H2020-MSCA-RISE-2015 LDPC Low-density-parity-check
under Grant 690750. The work of L. Hanzo was supported by the European LDS Low-density spreading
Research Council’s Advanced Fellow Grant QuantCom. (Corresponding LDS-CDMA Low-density spreading code division multiple
author: Bichai Wang; Lajos Hanzo.)
L. Dai, B. Wang, and Z. Wang are with Tsinghua National Laboratory for access
Information Science and Technology, Department of Electronic Engineering, LDS-OFDM Low-density spreading aided orthogonal fre-
Tsinghua University, Beijing 100084, China (e-mail: [email protected]; quency division multiplexing
[email protected]; [email protected]).
Z. Ding is with the School of Electrical and Electronic Engineering,
LMS Least mean square
University of Manchester, Manchester M13 9PL, U.K. (e-mail: Log-MPA Logarithmic-domain message passing
[email protected]). algorithm
S. Chen is with the Electronics and Computer Science, University of
Southampton, Southampton SO17 1BJ, U.K., and also with the Electrical
LTE Long term evolution
and Computer Engineering Department, King Abdulaziz University, Jeddah MAP Maximum a posteriori probability
21589, Saudi Arabia (e-mail: [email protected]). MBER Minimum bit error rate
L. Hanzo is with the Electronics and Computer Science, University of
Southampton, Southampton SO17 1BJ, U.K. (e-mail: [email protected]).
MC-CDMA Multi-carrier code division multiple access
Digital Object Identifier 10.1109/COMST.2018.2835558 MIMO Multiple input multiple output
1553-877X  c 2018 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.
DAI et al.: SURVEY OF NOMA FOR 5G 2295

ML Maximum likelihood code division multiple access (CDMA), and orthogonal fre-
MMSE Minimum mean square error quency division multiple access (OFDMA) have been used as
MPA Message passing algorithm the corresponding key multiple access technologies, respec-
MRC Maximum ratio combining tively [3], [4]. From the perspective of their design princi-
MU-SCMA Multi-user sparse code multiple access ples, these multiple access schemes belong to the category
MUD Multi-user detection of orthogonal multiple access (OMA), where the wireless
MUSA Multi-user shared access resources are orthogonally allocated to multiple users in the
NOMA Non-orthogonal multiple access time-, frequency-, code-domain or according in fact based
OFDMA Orthogonal frequency division multiple access on their combinations. We might collectively refer to these
OMA Orthogonal multiple access domains as “resources”. In this way the users’ information-
PAPR Peak-to-average-power ratio bearing signals can be readily separated at a low complexity
PDMA Pattern division multiple access by employing relatively cost-efficient receivers. However, the
PIC Parallel interference cancelation number of supported users is limited by the number of avail-
PSO Particle swarm optimization able orthogonal resources in OMA. Another problem is that,
QAM Quadrature amplitude modulation despite the use of orthogonal time-, frequency- or code-domain
QPSK Quaternary phase shift keying resources, the channel-induced impairments almost invariably
RWBS Repeated weighted boosting search destroy their orthogonality. More specifically, considering that
SAMA Successive interference cancellation amenable two signals s1 (t) and s2 (t), which are orthogonal either in
multiple access time-, frequency- or code-domain, are transmitted over dis-
SC-FDMA Single carrier-frequency division multiple persive channels h1 (t) and h2 (t), separately, the received
access signals x1 (t) = s1 (t) ⊗ h1 (t) and x2 (t) = s2 (t) ⊗ h2 (t)
SCMA Sparse code multiple access are usually non-orthogonal due to time or frequency expan-
SD Sphere detector sion. Hence, some high-complexity “orthogonality restoring
SDMA Spatial division multiple access measures”, such as multi-user equalizers, have to be invoked.
SE Spectral efficiency Consequently, it remains a challenge for OMA to satisfy the
SIC Successive interference cancellation radical spectral efficiency and massive connectivity require-
SINR Signal-to-interference-and-noise ratio ments of 5G.
SNR Signal-to-noise ratio The innovative concept of non-orthogonal multiple access
ST-DFE Space-time decision feedback equalization (NOMA) has been proposed in order to support more users
STE Space time equalizer than the number of available orthogonal time-, frequency-,
TDMA Time division multiple access or code-domain resources. The basic idea of NOMA is
UE User to support non-orthogonal resource allocation among the
URC Unity rate code users at the ultimate cost of increased receiver complexity,
UWB Ultra-wideband. which is required for separating the non-orthogonal sig-
nals. Recently, several NOMA solutions have been actively
investigated [5]–[13], which can be basically divided into
I. I NTRODUCTION two main categories, namely power-domain NOMA [14]–[63]
HE RAPID development of the mobile Internet and the and code-domain NOMA [64]–[97], including multiple access
T Internet of Things (IoT) leads to challenging require-
ments for the 5th generation (5G) of wireless communication
solutions relying on low-density spreading (LDS) [64]–[73],
sparse code multiple access (SCMA) [74]–[95], multi-user
systems, which is fuelled by the prediction of 1000-fold data shared access (MUSA) [96], successive interference cancel-
traffic increase by the year 2020 [1]. Specifically, the key lation amenable multiple access (SAMA) [97], etc. Some
performance indicators (KPI) advocated for 5G solutions can other closely-related multiple access schemes, such as spatial
be summarized as follows [2]: 1) The spectral efficiency is division multiple access (SDMA) [98]–[111], pattern divi-
expected to increase by a factors of 5 to 15 compared to sion multiple access (PDMA) [112], [113] and bit division
4G; 2) To satisfy the demands of massive connectivity for multiplexing (BDM) [114] have also been proposed. The
IoT, the connectivity density target is ten times higher than milestones of NOMA techniques are summarized in Fig. 1.
that of 4G, i.e., at least 106 /km2 ; 3) 5G is also expected to Note that most the existing survey papers on NOMA [5]–[10]
satisfy the requirements of a low latency (radio latency ≤ only focus on the power-domain NOMA. Some code-domain
1 ms), low cost (≥ 100 times the cost efficiency of 4G), and NOMA schemes are briefly introduced in [11] and [12]. On the
the support of diverse compelling services. In order to satisfy contrary, in this paper, both power-domain NOMA and code-
these stringent requirements, advanced solutions have to be domain NOMA, as well as other NOMA schemes proposed
conceived. as part of the Rel-14 3GPP NR Study Item shown in Table I
Over the past few decades, wireless communication sys- are introduced to provide a more comprehensive review of the
tems have witnessed a “revolution” in terms of their multiple most recent progress on NOMA. Furthermore, the comparison
access techniques. Specifically, for 1G, 2G, 3G, and 4G of 15 NOMA schemes proposed for Rel-14 3GPP NR Study
wireless communication systems, frequency division multi- Item are also provided to shed some insights on the unified
ple access (FDMA), time division multiple access (TDMA), design and implementation of NOMA. Moreover, in addition
2296 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

Fig. 1. The milestones of multiple access technologies.

to the basic principle and theoretical analysis, prototype eval-

uations and field test results are also presented to verify the
actual performance gain of NOMA in this paper, which are
not discussed in the existing papers.
In this article, we will discuss the basic principles as well
as pros and cons of NOMA in Section II. In Section III, the
design principles and key features of these dominant NOMA
solutions as well as the user grouping and resource allo-
cation will be discussed, and a systematic comparison of
these NOMA schemes will be provided in terms of their
spectral efficiency, system performance, receiver complexity,
etc. Performance evaluations and transmission experiments of
NOMA are also introduced to verify the analytical results.
In Section IV, we will highlight a range of challenging open
problems that should be solved for supporting NOMA, such
as the theoretical analysis, the sophisticated transmitter design,
and the tradeoff between the attainable system performance
versus receiver complexity. Accordingly, the corresponding
opportunities and future research trends will be highlighted in
order to provide some insights into potential future research
in this promising field. Finally, our conclusions are offered in
Section V. The structure of this article is shown in Fig. 2 at a
glance. Fig. 2. The structure of this article.

II. BASIC P RINCIPLES AND A DVANTAGES OF NOMA

In this section, we will firstly compare the basic principles FDMA, hence the receiver can readily detect all users’ data in
of OMA and NOMA. Then, the pros and cons of NOMA are their corresponding frequency bands, respectively. Similarly,
contrasted to those of OMA in detail. in TDMA, an exclusive time slot is allocated to each user,
In conventional OMA schemes, such as FDMA, TDMA, hence it is easy to distinguish the different users’ signals at
CDMA and OFDMA used for 1G, 2G, 3G, and 4G, the receivers in the time domain. In CDMA, multiple users can
respectively, multiple users are allocated to orthogonal radio share the same time-frequency resources, while the transmit-
resources in the time-, frequency-, code-domain or to their ted symbols of different users may be mapped to orthogonal
combinations. More specifically, each user transmits a unique, spreading sequences, such as Walsh-Hadamard codes. Hence,
user-specific signal over its unique frequency resource in a low-complexity decorrelation receiver can be used for
DAI et al.: SURVEY OF NOMA FOR 5G 2297

TABLE I
NOMA S CHEMES P ROPOSED FOR THE R EL -14 3GPP NR S TUDY I TEM

multi-user detection (MUD). OFDMA can be regarded as same time-frequency-code resources are shared among multi-
a smart integration of FDMA and TDMA, where the radio ple users. At the receiver side, power-domain NOMA exploits
resources are orthogonally partitioned in the time-frequency the users’ power-difference, in order to distinguish differ-
grid. Theoretically, as a benefit of orthogonal resource alloca- ent users based on successive interference cancellation (SIC).
tion, there is no interference among users in OMA systems, Code-domain NOMA is similar to CDMA or multi-carrier
hence low-complexity detectors with linear complexity can be CDMA (MC-CDMA), except for its preference for using low-
used to separate the different users’ signals [115]. However, density sequences or non-orthogonal sequences having a low
the maximum number of supportable users is rigidly restricted cross-correlation.
by the number of orthogonal resources available in conven-
tional OMA schemes, which becomes a hard limit when A. Channel Capacity Comparison of OMA and NOMA
massive connectivity is required for 5G. Additionally, it has From the perspective of information theory, for the capacity
been theoretically shown that OMA cannot always achieve the of multiple access channels in both additive white Gaussian
maximum attainable sum-rate of multi-user wireless systems, noise (AWGN) and fading scenarios, we have the following
while NOMA is capable of achieving the multi-user capacity results for OMA and NOMA (applicable to both power-
with the aid of time-sharing or rate-splitting if necessary [116], domain NOMA and code-domain NOMA):
which will be explained in detail in the following subsection. • AWGN channel: In the uplink of an AWGN channel sup-
In order to circumvent the above limitation of OMA, porting K users (K can be larger than 2), the capacity of
NOMA has been recently investigated as a design alterna- the multiple access channel can be formulated as [117]
tive. The key distinguishing feature of NOMA is to support  
K K
a higher number of users than the number of orthogonal P i
resource slots, with the aid of non-orthogonal resource allo- Ri ≤ W log 1 + i=1 , (1)
N0 W
cation. This may be achieved by the sophisticated inter-user i=1
interference cancellation at the cost of an increased receiver where W is the bandwidth, Pi is the transmitted
complexity, e.g., computational complexity with polynomial power, and N0 is the power spectral density of
or exponential order. Furthermore, it is worth bearing in mind Gaussian noise. More particularly, according to the capac-
that even if we opt for OMA schemes, the time-domain signals ity analysis found in the pioneering contribution of
are smeared by their convolution with the dispersive chan- Tse and Viswanath [116], Fig. 3 and Fig. 4 from [5]
nel impulse response (CIR). In addition, a historic concept of provided the channel capacity comparison of OMA and
NOMA is constituted by the family of CDMA systems rely- NOMA, where a pair of users communicating with a base
ing on more non-orthogonal sequences than the number of station (BS) over an AWGN channel is considered as an
chips Nc in a sequence to support more than Nc users at the example without loss of generality. Fig. 3 showed that
cost of imposing inter-user interference. Such interference can the uplink of NOMA is capable of achieving the capac-
only be efficiently mitigated with the aid of powerful multi- ity region, while OMA is suboptimal in general, except
user detectors. Hence, the NOMA concept is appealing. The at one point. However, at this optimal point, rate-fairness
family of NOMA schemes can be basically divided into two is not maintained, since the rate of the low-power user
categories: power-domain NOMA and code-domain NOMA. is much lower than that of the higher-power user when
In power-domain NOMA, different users are assigned differ- the difference of the received powers of the two users is
ent power levels according to their channel quality, while the high. Note that the results for the simple two-user case
2298 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018


which is valid for all possible splits P = K k =1 Pk of
total power at the BS. The optimal points can be achieved
by NOMA with the aid of superposition coding at the
transmitter and SIC at each of the receivers [116]. More
particularly, Fig. 4 showed that the boundary of the rate
pairs of NOMA is in general beyond that of OMA in
asymmetric channels.
• Fading channels: In fading channels, the sum capacity in
the uplink - provided that the channel state information
(CSI) is only known at the receiver - can be represented as
 K 
k =1 |hk |2 Pave
Csum = E log 1 + , (3)
N0

where we assume that each user has the same average

power Pave . In this way, OMA is strictly subopti-
mal in the uplink, while NOMA relying on MUD is
optimal [116].
Fig. 3. Channel capacity comparison of OMA and NOMA in the uplink • MIMO-NOMA and MIMO-OMA: NOMA can be also
AWGN channel: (a) Symmetric channel; (b) Asymmetric chaa [5] IEEE.
c extended to MIMO scenarios, where the BS has M
antennas and each user is equipped with N antennas.
Additionally, multiple users can be randomly grouped
into M clusters with two users in each cluster. It has
been shown in [119] that MIMO-NOMA is strictly better
than MIMO-OMA in terms of the sum channel capacity
(except for transmission to a single user in MIMO sys-
tems), i.e., for any rate pair achieved by MIMO-OMA
schemes, there is a power split for which MIMO-NOMA
is capable of achieving rate pairs that are strictly higher.

B. Advantages of NOMA Compared to OMA

We can see from the capacity analysis that it is feasible
for NOMA to achieve a higher transmission rate than OMA.
Specifically, the main advantages of NOMA compared to the
classical OMA can be summarized as follows:
• Improved spectral efficiency and cell-edge through-
put: The time-frequency resources are shared non-
orthogonally among users both in the power-domain
NOMA and in the code-domain of NOMA. As described
above, in the uplink of AWGN channels, although both
OMA and NOMA are capable of achieving the maxi-
mum attainable sum capacity, NOMA supports a more
Fig. 4. Channel capacity comparison of OMA and NOMA in the downlink
AWGN channel: (a) Symmetric channel; (b) Asymmetric channel [5] IEEE.
c equitable user fairness. Additionally, the capacity bound
of NOMA is higher than that of OMA in the downlink
of AWGN channels. In multi-path fading channels sub-
jected to inter-symbol-interference (ISI), although OMA
can be extended to the general case of an arbitrary num-
is indeed capable of achieving the maximum attainable
ber of users [116]. Explicitly, it is shown in [116] that
sum capacity in the downlink, NOMA relying on MUD
there are exactly K! corner points when the K-user sce-
is optimal, while OMA remains suboptimal, if the CSI is
nario is considered and the K-user NOMA system can
only known at the downlink receiver.
achieve the same optimal sum rate at all of these K! cor-
• Massive connectivity: Non-orthogonal resource alloca-
ner points. In the downlink, the boundary of the capacity
tion in NOMA indicates that the number of supportable
region is given by the rate tuples [118]:
users/devices is not strictly limited by the number of
⎛ ⎞ orthogonal resources available. Therefore, NOMA is
P |h |2 capable of significantly increasing the number of simul-
Rk = W log⎝1 + k k ⎠, (2) taneous connections in rank-deficient scenarios, hence
K 2
N0 W + j =k +1 Pj |hk | it has the potential to support massive connectivity. Of
DAI et al.: SURVEY OF NOMA FOR 5G 2299

course, it should be noted that some practical implemen- end of this section, performance evaluations and transmission
tation issues in NOMA systems, such as its hardware experiments of NOMA will be discussed.
imperfections and computational complexity, may hin-
der the realization of massive connectivity, which will
be detailed in Section IV. A. Power-Domain NOMA
• Low transmission latency and signaling cost: In conven- In this subsection, we will discuss the first category of
tional OMA relying on access-grant requests, a user first NOMA, namely, power-domain NOMA. In [14]–[17], the con-
has to send a scheduling request to the base station (BS). cept and key features of power-domain NOMA have been
Then, upon receiving this request, the BS schedules the described in detail. In contrast to the multiple access schemes
user’s uplink transmission by responding with a clear- relying on the time-, frequency-, code-domain or on their
to-send signal in the downlink channel. Thus, a high combinations, NOMA can be realized in a recently emerged
transmission latency and a high signaling overhead will new domain, namely in the power domain. At the transmit-
be imposed, which becomes unacceptable in the case of ter, different signals generated by different users are directly
massive 5G-style connectivity. Specifically, the access- superimposed on each other after classic channel coding and
grant procedure in LTE takes about 15.5 ms before the modulation. Multiple users share the same time-frequency
data is transmitted [120]. In this way, the radical require- resources, and then are detected at the receivers by MUD
ment of maintaining a user delay below 1 ms cannot be algorithms such as SIC. In this way, the spectral efficiency
readily satisfied [112]. By contrast, dynamic scheduling can be enhanced at the cost of an increased receiver complex-
is not required in some of the uplink NOMA schemes. ity compared to conventional OMA. Additionally, it is widely
To elaborate a little further, in the uplink of a SCMA sys- recognized based on information theory that non-orthogonal
tem, grant-free multiple access can be realized for users multiplexing using superposition coding at the transmitter and
associated with pre-configured resources defined in the SIC at the receiver not only outperforms classic orthogo-
time- and frequency-domain, such as the codebooks, as nal multiplexing, but it is also optimal from the perspective
well as the pilots. By contrast, at the receiver blind detec- of achieving the capacity region of the downlink broadcast
tion and compressive sensing (CS) techniques can be channels [116].
used for performing joint activity and data detection [91]. Some practical considerations for power-domain NOMA,
Hence again, beneficial grant-free uplink transmission such as multi-user power allocation, signalling overhead, SIC
can be realized in NOMA, which is capable of signif- error propagation and user mobility, were discussed in [14].
icantly reducing both the transmission latency and the To achieve a further enhancement of its spectral efficiency,
signaling overhead. Note that in some NOMA schemes the authors of [14]–[25] discussed the combination of NOMA
using SIC receivers, the SIC process may impose extra with MIMO techniques. Particularly, the capacity compar-
latency. Therefore, the number of users relying on SIC ison between MIMO-NOMA and MIMO-OMA has been
should not be excessive, and advanced MIMO techniques investigated in [18] and [19], and the superiority of MIMO-
can be invoked for serving more users, as discussed in NOMA over MIMO-OMA in terms of both sum channel
Section III. capacity and ergodic sum capacity was proved analytically.
• Relaxed channel feedback: The requirement of chan- Furthermore, in [20] and [21], the potential gains of MIMO-
nel feedback will be relaxed in power-domain NOMA, NOMA were shown based on both link-level as well as
because the CSI feedback is only used for power alloca- on system-level simulations and using the NOMA test-bed
tion. Hence there is no need for accurate instantaneous developed, where a hardware SIC receiver was used taking
CSI knowledge. Therefore, regardless whether fixed or into account the realistic hardware impairments quantified
mobile users are supported, having a limited-accuracy in terms of the error vector magnitude (EVM), the num-
outdated channel feedback associated with a certain max- ber of quantization bits in the analog/digital (A/D) converter,
imum inaccuracy and delay will not severely impair the etc. The simulation results and the measurements obtained
attainable system performance, as long as the channel showed that under multiple configurations, the cell through-
does not change rapidly. put achieved by NOMA is about 30% higher compared to
Given the above prominent advantages, NOMA has been OFDMA. Furthermore, some open implementation issues were
actively investigated, with a views for employment in 5G as also discussed in [20] and [21], including the granularity of
a promising solution. In the next section, we will discuss and the multi-user power allocation both in time and frequency,
compare the dominant NOMA solutions. as well as the signaling overhead, feedback enhancements
and receiver design. Additionally, the receiver design was dis-
cussed in [26]–[28]. A novel NOMA transmitter and receiver
III. D OMINANT NOMA S OLUTIONS design was proposed in [26], where the signals of multiple
In this section, we will discuss the families of prominant users are jointly modulated at the transmitter side and detected
NOMA schemes by dividing them into two categories, namely at the receiver side. In this scheme, the desired signal of the
power-domain and code-domain NOMA. Their design princi- cell center user can be directly detected without detecting the
ples and key features will be highlighted, respectively. We will signal of the cell edge user, i.e., without SIC processing. Thus,
also provide their comparison in terms of their spectral effi- a low complexity is achieved. Furthermore, the associated
ciency, system performance, receiver complexity, etc. At the simulation results have shown that compared to the ideal SIC,
2300 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

the downlink NOMA link-level performance depends both on a cooperative NOMA transmission scheme will also be pre-
the actual receiver design and on the difference in the power sented. Finally, the networking aspects of NOMA solutions
ratio split between the cell edge user and cell center user. will be discussed.
Besides, the design and performance of the SIC receiver for 1) Basic NOMA Relying on a SIC Receiver: Firstly, we
downlink NOMA combined with 2-by-2 open-loop SU-MIMO consider the family of single antenna systems relying on a
based on LTE TM3 (Transmission mode 3) were investigated single BS and K users.
in [27], where different receiver weight generation schemes In the downlink, the total power allocated to all K users is
were introduced both before SIC and after SIC according to limited to P, and the BS transmits the signal xi to the ith user
the transmission rank combination between the users. The link- subjected to the power-scaling coefficient pi . In other words,
level simulation results showed that the codeword level SIC the signals destined for different users are weighted by differ-
achieves higher performance than the symbol level SIC and ent power-scaling coefficients and then they are superimposed
in fact approaches the performance of ideal SIC. The impact at the BS according to:
of applying the SIC receiver for cell-edge users in downlink

K
√
NOMA using SU-MIMO was investigated in [28]. The sim- x= pi x i , (4)
ulation results showed that there is an improvement of the i=1
NOMA gains over OMA in conjunction with SIC process-
ing for the cell-edge users. Furthermore, in order to increase where E[|xi |2 ] = 1(i = 1, 2, . . . , K ) denotes the
 normalized
the attainable performance of the SIC receiver, cooperative power of the user signals, and we have P = K i=1 pi . The
NOMA transmission has been proposed in [29]–[31]. A range received signal yi of the ith user is
of investigations related to multi-cell NOMA schemes were yi = hi x + vi , (5)
carried out in [32]–[34]. Moreover, since having an increased
number of cell-edge users typically degrades the efficiency of where hi denotes the channel gain between the BS and
coordinated multi-point (CoMP) transmissions, this limitation the ith user, while vi associated with the power den-
was circumvented by promising NOMA solutions proposed sity Ni represents the Gaussian noise plus the inter-cell
for a CoMP system in [35] and [36]. Additionally, the perfor- interference.
mance of NOMA techniques supporting randomly distributed At the receiver, SIC is used for MUD. The optimal order of
users was evaluated in [37]. These simulation results demon- SIC detection relies on detecting the strongest to the weakest
strated that the outage performance of NOMA substantially user (determined by |hi |2 /Ni , i = 1, 2, . . . , K ). Based on this
depended both on the users’ target data rates and on their optimal SIC-detection order, any user can detect its informa-
allocated power. In [38]–[47], the system-level performance of tion without substantial interference-contamination imposed
power-domain NOMA was evaluated and the associated simu- the other users whose normalized channel gain is smaller than
lation results showed that both the overall cell throughput and that of this user. In this way, the user with the strongest normal-
the cell-edge user throughput, as well as the degree of propor- ized channel gain can cancel interference from all the other
tional rate-fairness of NOMA were superior to those of OMA. users, and thus is least contaminated. It is intuitive that the
Furthermore, the impact of the residual interference imposed users associated with small normalized channel gains should
by realistic imperfect channel estimation on the achievable be allocated higher power levels in order to improve their
throughput performance was investigated in [48]–[50]. On received signal-to-interference and noise ratio (SINR), so that
the one hand, the channel estimation error results in resid- a high detection reliability can be guaranteed. More particu-
ual interference in the SIC process, which hence reduces the larly, it has been validated in [121] that, to maximize the sum
achievable user throughput. On the other hand, the channel rate, it is optimal for each user to decode the signals of users
estimation error causes error in the transmission rate control with poorer normalized channel gains first. Although the users
for the respective users, which may result in decoding errors having larger normalized channel gains require less power,
not only at the destination user terminal but also at other user they are capable of correctly detecting their data with a high
terminals owing to the error propagation imposed by the SIC probability, as a benefit of SIC. Without loss of generality,
process. A simple transmission rate back-off algorithm was we assume that |h1 |2 /N1 ≥ |h2 |2 /N2 ≥ · · · ≥ |hK |2 /NK ,
considered in [48] and [49], and the impact of the channel and a descending-order-based power allocation p1 ≤ p2 ≤
estimation error was effectively mitigated. Simulation results · · · ≤ pK can be considered. Assuming perfectly error-free
showed that NOMA achieves beneficial user throughput gains decoding of the interfering signals, the achievable rate of user
over OMA in a scenario subject to channel estimation errors, i (i = 1, 2, . . . , K ) can be written as
which is similar to the case associated with perfect channel ⎛ ⎞
2
estimation. p |h |
Ri = W log⎝1 +  i i ⎠. (6)
Let us now elaborate on the power-domain NOMA tech- i−1 2
Ni W + j =i j p |h i |
niques in this subsection. Firstly, the basic principle of power-
domain NOMA relying on a SIC receiver will be discussed. In the case of two users as shown in Fig. 5 (a), we assume
Then, a promising extension relying on integrating NOMA that the normalized channel gain of the second user is lower
with MIMOs will be discussed for the sake of increasing its than that of the first one, i.e., |h1 |2 /N1 > |h2 |2 /N2 , and thus
attainable spectral efficiency. Another compelling extension to p1 < p2 . The second user detects its signal by regarding the
DAI et al.: SURVEY OF NOMA FOR 5G 2301

Fig. 6. Comparison of NOMA and OMA [14] IEEE.

c

BS detects the i-th user’s signal, it decodes the j-th (j < i )

user’s signal first and then removes (i − 1) users’ signals from
the observation y. The remaining (K − i) signals are regarded
Fig. 5. Power-domain NOMA: (a) Basic NOMA relying on a SIC receiver; as interference. As a result, the achievable data rate of the i-th
(b) NOMA in MIMO systems; (c) NOMA in CoMP. user becomes
 
pi |hi |2
Ri = W log 1 +   2 . (10)
signal of the first user as interference. The first user firstly N0 W + K hj  pj
j =i+1
detects the signal of the second user, and then subtracts its
remodulated version from the received signal, so that the first As illustrated in Section II-A, this NOMA scheme is capa-
user can detect its own signal without interference from the ble of achieving the maximum attainable multi-user capac-
second user. ity in AWGN channels both in the uplink and downlink.
Assuming that the transmission bandwidth is normalized to Furthermore, this NOMA scheme has the potential of strik-
1Hz, the data rates of the first user and the second user can ing a more attractive tradeoff between the spectral efficiency
be represented as and user-fairness.
  When the number of users is sufficiently high, the SIC-
p1 |h1 |2 induced error propagation may have a severe effect on the error
R1 = log2 1 + , (7)
N1 probability in the absence of preventative measures. However,
  some advanced user pairing and power allocation methods,
p2 |h2 |2 as well as powerful channel coding schemes can be used for
R2 = log2 1 + , (8)
p1 |h2 |2 + N2 reducing the error probability. Indeed, it has been shown that
error propagation only has a modest impact on the NOMA
respectively. Thus, by tuning power allocation coefficients, the
performance even under the worst-case scenario [14], [39].
BS can adjust the data rate of each user. More particularly, it
2) NOMA in MIMO Systems: Although the same time-
has been shown in Fig. 6 [14] that this NOMA scheme is capa-
frequency resources can be shared by multiple users in the
ble of achieving higher rates than OFDMA. On the other hand,
basic NOMA employing SIC, the improvement of spectral
this NOMA scheme makes a full use of the natural difference
efficiency still remains limited, hence may not satisfy the
of channel gains among the users, which implies that the near-
expected spectral efficiency improvements of 5G. An appeal-
far effect is effectively harnessed to achieve higher spectral
ing solution is the extension of the basic NOMA using SIC by
efficiency. As a result, both the attainable sum capacity and
amalgamating it with advanced MIMO techniques [16], [17].
the cell-edge user data rate can be improved [15].
As illustrated in Fig. 5 (b), in downlink NOMA of MIMO
In the uplink, the signal received at the BS is given by
systems, MBS BS antennas are used for generating B differ-

K
√ ent beams in the spatial domain with the aid of beamforming.
y= hi pi x i + v , (9) Within each beam, the signals of multiple users may be trans-
i=1 mitted by superimposing them, hence leading to the concept
where pi and xi are the transmit power and signal transmitted of intra-beam superposition modulation, which is similar to
by the i-th user, respectively. Furthermore, v associated with the basic NOMA using SIC, as discussed above. The bth
the power density N0 represents the Gaussian noise plus the (1 ≤ b ≤ B ) transmitter beamforming vector is denoted as
inter-cell interference at the BS. SIC is used for reliable signal mb . Let us assume that the number of users in the bth beam
detection at the BS. Without loss of generality, we assume is kb , the transmitted symbol of the ith user in the bth beam is
that p1 |h1 |2 ≥ p2 |h2 |2 ≥ · · · ≥ pK |hK |2 , and accordingly the xb,i , and the corresponding power-scaling coefficient is pb,i .
optimal decoding order for SIC is x1 , x2 , . . . , xK . Before the Then, by accumulating all signals of the B different beams,
2302 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

the MBS -dimensional transmitted downlink signal vector at

the BS can be formulated as

B 
kb
√
x0 = mb pb,i xb,i . (11)
b=1 i=1
Assuming that each user has Nr receiver antennas, the
Nr -dimensional received signal vector of the ith user encap-
sulated in the bth beam can be represented as
yb,i = Hb,i x0 + vb,i , (12)
where Hb,i denotes the channel matrix of size Nr × MBS
between the BS and the ith user in the bth beam, and vb,i
denotes the Gaussian noise plus the inter-cell interference.
Fig. 7. System-level performance for NOMA applying opportunistic random
At the receiver, a pair of interference cancellation beamforming in downlink [15] IEEE.
c
approaches are used for removing the inter-beam interference
and the intra-beam interference, respectively. The inter-beam
interference can be suppressed by spatial filtering, which is in the context of MIMO systems is higher than that of the
similar to the signal detection algorithm of spatial division basic NOMA using a single antenna at the BS. Additionally, in
multiple access (SDMA) systems. Assuming that the spatial this NOMA scheme, the number of reference signals required
filtering vector of the ith user in the bth beam is fb,i , the signal is equal to the number of transmitter antennas, regardless
zb,i after spatial filtering can be represented as of the number of non-orthogonal user-signals. In this way,
H when the number of users is increased beyond the number of
zb,i = fb,i yb,i
transmitter antennas, the number of orthogonal downlink ref-

kb
√ erence signals is not increased beyond the number of transmit
H
= fb,i Hb,i mb pb,j xb,j antennas.
j =1 Obviously, we can see from (14) and (15) that the design

B 
kb of beamforming vectors {mb } and spatial filtering vectors
H √ H
+ fb,i Hb,i mb  pb  ,j xb  ,j + fb,i vb,i . (13) {fb,i } is crucial for interference cancellation, which should be
b  =1 j =1 carefully considered in MIMO NOMA systems. Particularly,
b  =b
the beamforming and spatial filtering optimization problem
By normalizing the aggregated power of the inter-beam inter- is usually formulated as the maximization of the sum rate
ference and the receiver noise plus inter-cell interference to problem for perfect CSI scenarios and is formulated as the
unity, we can rewrite (13) as minimization of the maximum outage probability (MMOP)
problem for imperfect CSI scenarios [122]. For the perfect
√ 
kb
√ CSI scenarios, an iterative minorization-maximization algo-
zb,i = ab,i pb,j xb,j + qb,i , (14)
rithm (MMA) was proposed for beamforming design [123],
j =1
where a second-order cone program (SOCP) convex problem
where qb,i is the normalized term representing the sum of should be solved in each iteration. In addition, the duality
the inter-beam interference and receiver noise plus inter-cell scheme has been introduced in [124], which can be regarded as
interference, while ab,i is formulated as the quasi-degraded solution [125] for the sum rate maximiza-
|fb,i Hb,i mb |2 tion problem, and a quadratic constrained quadratic programs
ab,i = ⎧ B  kb ⎫. (15) (QCQP) convex problem should be solved. Furthermore, to
⎨ b  =1 j =1 pb  ,j |fb,i Hb,i mb  |2 ⎬ realize low-complexity beamforming, a multiple-user CSI-
b  =b based singular value decomposition (MU-CSI-SVD) algorithm
⎩ + fb,i
H E [v vH ]f ⎭
b,i b,i b,i was proposed [122] to solve the sum rate maximization prob-
After spatial filtering, the system model (14) becomes sim- lem in perfect CSI scenarios. What’s more, this algorithm can
ilar to that of the basic NOMA combined with SIC, as be easily used to simplify the outage probability expressions
described above. Therefore, the inter-beam interference can be in imperfect CSI scenarios. Therefore, for imperfect CSI sce-
suppressed, and then intra-beam SIC is invoked for removing narios, the MU-CSI-SVD algorithm can be also used with low
the inter-user interference imposed by superposition coding complexity.
within a beam. In the uplink systems associated with K users, where the
Naturally, more users can also be simultaneously supported, BS is equipped with MBS antennas and each user has a sin-
because more than two users can share a single beamforming gle transmitter antenna, the signal received at the BS can be
vector. To elaborate a little further, observe in Fig. 7 [15] that represented as
both the basic NOMA combined with SIC and the extended

K
√
NOMA relying on MIMO are capable of achieving a higher y= hi pi xi + v, (16)
sum-rate than OFDMA. Furthermore, the sum-rate of NOMA i=1
DAI et al.: SURVEY OF NOMA FOR 5G 2303

where hi is the MBS -dimensional channel vector between user employment of cooperation is indeed capable of enhancing
i and the BS. Furthermore, pi and xi are the transmit power the reception reliability. Note that it can only be invoked at
and the signal transmitted by the i-th user, respectively, while low user loads, because the above-mentioned regime requires
v represents the Gaussian noise plus the inter-cell interference potentially excessive resources for cooperation. This might
vector. reduce the extra gain of NOMA.
At the receiver, MMSE-SIC receiver can be used to realize To elaborate a little further, in practice the participation of
signal detection [5]. It has been shown in [5] that for any all users in the cooperative NOMA cannot be readily realized
decoding order, the sum throughput for all K users is equal to due to the extra requirement of short-range communication
the maximum of the total user throughput given the received resources as well as owing to the complex signal processing
signal vector in (16). In practice, the decoding order can be associated with a high signaling overhead. A promising solu-
adjusted according to the actual requirements, such as user tion to this problem is to reduce the number of cooperating
fairness. users. Without loss of generality, we consider the appealingly
3) Cooperative NOMA: Recently, a cooperative NOMA simple case of having only two cooperative users as an exam-
transmission scheme was proposed in [29]. Similar to the ple. Let us assume that the users are sorted in the order of
basic NOMA, cooperative NOMA also uses a SIC receiver improving channel qualities and that the mth and nth users
for detecting the multi-user signal. Therefore, the users asso- are paired together, where we have m < n. It has been shown
ciated with better channel conditions can be relied upon as in [29] that the worst choice of m and n is n = m + 1, while the
relays in order to improve the reception reliability of the users optimal choice is to group two users experiencing significantly
suffering from poor channel conditions. For cooperative trans- different channel qualities.
mission, for example short-range communication techniques - Again, the cooperative NOMA further exploits the specific
such as Bluetooth and ultra-wideband (UWB) schemes - can feature that the users having better channel qualities have the
be used for delivering signals from the users benefitting from knowledge of other users’ signals, whose channel qualities are
better channel conditions to the users with poor channel con- poor. In this way, the maximum diversity gain can be achieved
ditions, which is the key difference with respect to the basic for all users by transmitting the signals of those specific users
NOMA associated with SIC. who have better channel qualities to other users.
Without loss of generality, let us now consider a down- 4) NOMA in CoMP: NOMA also exhibits its own benefits
link cooperative NOMA system relying on a single BS for in multi-cell applications, leading to the concept of NOMA
supporting K users, where the K users are ordered based on in CoMP. However, by directly applying a single-cell NOMA
their channel qualities, with the first user having the worst design to multi-cell scenarios, NOMA in CoMP may result
channel condition, while the Kth user having the best chan- in severe inter-cell interference. As an example, a downlink
nel condition. Cooperative NOMA relies on the following two cellular system having two cells and four users is depicted in
phases [29]. Fig. 5 (c), where a two-user NOMA scheme is considered,
The first phase, also termed the broadcast phase, represents with user 1 and user 2 being served by BS 1, while user 3
the direct transmission. In this phase, the BS sends down- and user 4 being served by BS 2. However, at user 1, strong
link messages to all the K users based on the principle of interference may be imposed by the signals transmitted from
basic NOMA relying on SIC, where the superimposed infor- BS 2 (similarly, user 3 also suffers from severe interference
mation of the K users obeys the total power constraint. The caused by the signals transmitted from BS 1), which poten-
SIC process is implemented at the user side. As a result, tially leads to a significant performance degradation of the
the users having better channel conditions have the knowl- NOMA in CoMP scheme.
edge of the signals intended to the users having poor channel To mitigate the inter-cell interference in the downlink, joint
conditions [126]. transmit-precoding of all NOMA users’ signals can be uti-
The second phase represents the cooperative transmission. lized. However, all users’ data and channel information should
During this phase, the cooperating users transmit their sig- be available at the BSs involved, and finding the optimal
nals via their short range communication channels, such as transmit-precoder is not trivial. Moreover, multi-user trans-
Bluetooth or UWB. Particularly, the second phase includes mit precoding of single-cell NOMA may not be efficient in
(K − 1) time slots. In the first time slot, the Kth user broad- a NOMA in CoMP setting, since a beam generated via geo-
casts the superposition of the (K − 1) signals destined for graphically separated BS antennas may not be capable of
the remaining users. Then the SIC process is invoked again at covering more than one angularly separated user for intra-
these (K − 1) users. The (K − 1)th user combines the signals beam NOMA. By exploiting that the CIRs of different users
received during both phases by using maximum ratio combin- are likely to be rather different in the multi-cell scenario, a
ing (MRC), and it detects its own information at a higher SNR reduced-complexity transmit precoding scheme was proposed
than that of the traditional SIC. Similarly, in the kth time slot, for NOMA in CoMP [32], where the precoder is applied only
where 1 ≤ k ≤ K − 1, the (K − k + 1)st user also broadcasts to the signals of the cell edge users, such as users 1 and 3 of
the (K − k) superposed signals for the remaining (K − k) Fig. 5 (c). Additionally, a multi-cell uplink NOMA system has
users, whose channel conditions are worse than that of this been considered in [33], and the rate coverage probability (the
user. Then the (K − k + 1)st user combines the observations probability that a given user’s achievable rate remains above
gleaned from both phases and it detects its own informa- the target data rate) of a user who is at rank m (in terms of the
tion at a higher SNR than the traditional SIC. Therefore the distance from its serving BS) among all users in a cell and the
2304 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

systems is compared. It is observed that LDM offers better

performance than TDM/FDM in all scenarios, and the higher
the SNR threshold of the fixed service, the larger the advantage
of the LDM systems.

B. Code-Domain NOMA
The NOMA schemes discussed in the previous subsec-
tion realize multiplexing in the power domain. By contrast,
in this subsection, we introduce the other main category of
NOMA schemes, which achieves multiplexing in the code
domain. The concept of code-domain NOMA is inspired by
the classic CDMA systems, in which multiple users share the
same time-frequency resources, but adopt unique user-specific
spreading sequences. However, the key difference compared to
Fig. 8. Channel capacity advantage of LDM vs. TDM/FDM [127] IEEE.
c CDMA is that the spreading sequences are restricted to sparse
sequences or non-orthogonal low cross-correlation sequences
in NOMA. In this subsection, we first present the initial form
mean rate coverage probability of all users in a cell were ana- of NOMA based on sparse spreading sequences, i.e., LDS-
lyzed using the theory of order statistics and poisson cluster CDMA [64]–[67]. Then, the family of LDS aided multi-carrier
process. It has been shown that the average rate coverage of a OFDM systems (LDS-OFDM) [68]–[73] will be discussed,
NOMA cluster is better than that of its counterpart OMA clus- which retains all the benefits of OFDM-based multi-carrier
ter for higher number of users per cell and for higher target transmissions in terms of its ISI avoidance, together with
rate requirements. Furthermore, an up-to-date literature review MUD-assisted LDS-CDMA operating at a lower complexity
of interference management techniques that apply NOMA in than that of the optimal maximum a posteriori probability
multi-cell networks has been provided in [34], including both (MAP) detector. Another important extension of LDS-CDMA
NOMA using joint processing and NOMA relying on coor- is SCMA [74]–[95], which still enjoys the benefit of low-
dinated scheduling/beamforming. The major practical issues complexity reception, but has a better performance than
and challenges that arise in the implementation of multi-cell LDS-CDMA. A suite of other improved schemes and special
NOMA have also been highlighted in [34], such as the SIC forms of CDMA, such as MUSA [96] and SAMA [97] will
implementation issues, imperfect CSI, as well as multi-user also be discussed in this subsection.
power allocation and clustering. 1) Low-Density Spreading CDMA (LDS-CDMA):
5) Application of Power-Domain NOMA: Recently, the Developed from the classic concept of CDMA, LDS-CDMA
concept of power-domain has been successfully applied to is designed for limiting the amount of interference imposed
ATSC 3.0 [127], which is a new next-generation broadcasting on each chip of conventional CDMA systems by using
standard in U.S., and this physical-layer non-orthogonal mul- LDS instead of conventional spreading sequences. The basic
tiplexing technology is named layered-division-multiplexing principle of LDS-CDMA has been discussed in [64]–[66].
(LDM). Additionally, [64] and [65] also discussed the iterative MUD
Specifically, a two-layer LDM structure consisting of the based on the message passing algorithm (MPA) imposing a
upper layer (UL) and the lower layer (LL) is accepted by lower complexity than that of the optimal MAP detector.
ATSC 3.0 to improve spectral efficiency and provide more Specifically, in [64], the performance of LDS-CDMA com-
versatile broadcasting services. The UL with higher power municating over memoryless Gaussian channels using BPSK
allocation is used to deliver mobile services to indoor, portable modulation was analyzed. The simulation results showed that
and handheld receivers, while the LL is designed to deliver the performance of LDS-CDMA is capable of approaching
high data rate services, such as UHDTV or multiple HDTV the single-user performance for a normalized user-load as
services to fixed reception terminals, where the operational high as 200%. However, the performance of LDS-CDMA
SNR is usually high due to the large and possibly direc- operating in multipath fading channels is still under inves-
tional receive antennas [127]. At the transmitter, the data of tigation at the time of writing. The challenge is that the
each layer is firstly processed by its own physical-layer signal multipath fading channels will destroy the original LDS struc-
processing modules, including channel encoding, interleav- ture. On the other hand, a structured approach of designing
ing, modulation, etc., and then the signals from all layers are LDS codes for LDS-CDMA has been proposed in [66], where
superimposed over the same time-frequency resources. At the the basic idea is to map the signature constellation elements
receivers, to decode the UL signal, the lower-power LL ser- to the spreading matrix hosting the spreading sequences.
vice is treated as an additional interference. To decode the LL Furthermore, the capacity region of LDS-CDMA was cal-
signal, the receiver firstly needs to cancel the UL signal, which culated using information theoretic analysis in [67], and the
is referred as SIC procedure in power-domain NOMA. accompanying simulation results showed how the attainable
As shown in Fig. 8 from [127], the channel capacity for capacity depended on the spreading sequence density factor
the mobile and fixed services of the LDM and TDM/FDM as well as on the maximum number of users associated with
DAI et al.: SURVEY OF NOMA FOR 5G 2305

each chip, which provided insightful theoretical guidelines for

practical LDS system designs.
Let us now consider a classic synchronous CDMA system
operating in the uplink and supporting K users with the aid
of Nc chips (Nc equals to the number of observations at the
receiver). The transmitted symbol xk of user k is firstly gener-
ated by mapping a sequence of independent information bits
to a constellation alphabet, i.e., xk is taken from a complex-
valued constellation set X. Then, the transmitted symbol xk is
mapped to a spreading sequence sk , such as the set of widely
used PN sequences, which is unique for each user. The signal
received during chip n can be represented by

K
Fig. 9. Illustration of LDS-CDMA: 6 users only employ 4 chips for
yn = gn,k sn,k xk + wn , (17) transmission, which implies that a normalized user-load of 150% can be
k =1 achieved.

where sn,k is the nth component of the spreading sequence

sk , gn,k is the channel gain of user k on chip n, and means that the interference imposed on each chip will be effi-
wn is a complex-valued Gaussian noise sample with a zero ciently reduced, hence mitigating the multi-user intereference
mean and a variance of σ 2 . When we combine the signals by carefully designing the spreading sequences. Therefore,
received during all the Nc chips, the received signal vector the received signal at chip n in LDS-CDMA systems can be
y = [y1 , y2 , . . . , yNc ]T is formulated as rewritten as
 
y = Hx + w, (18) yn = gn,k sn,k xk + wn = hn,k xk + wn ,
where x = [x1 , x2 , . . . , xK ]T , H is the channel matrix of k ∈N (n) k ∈N (n)
size (Nc × K ), and the element hn,k in the nth row and (19)
the kth column of H is denoted by gn,k sn,k . Finally, w = where N (n) denotes the set of users whose sparse spreading
[w1 , w2 , . . . , wNc ]T , and w ∼ CN (0, σ 2 I). sequences have a non-zero element at chip n, namely, N (n) =
In classic CDMA systems, the elements of the spreading {k |sn,k = 0}.
sequences sk (k = 1, . . . , K ) are usually non-zero, i.e., the At the receiver, MUD based on message passing algo-
spreading sequences are not sparse. Consequently, the sig- rithm may be performed. Given the joint probability func-
nals received from all the active users are overlaid on top tion p(x1, x2 , . . . , xE ) for random variables x1, x2 , . . . , xE , the
of each other at each chip, and every user will be subjected message-passing algorithm is capable of simplifying the calcu-
to inter-user interference imposed by all the other users. If lation of the marginal probability distribution for each variable
the spreading sequences are orthogonal, it is straightforward as follows:
to eliminate the interferences, hence the information of all   
users can be accurately detected by a low-complexity corre- p(xe ) = p x1, x2 , . . . , xE , (20)
lation receiver. However, the classical orthogonal spreading ∼{xe }
sequences can only support as many users as the number of where ∼ {xe } represents all variables except for xe . We
chips. By contrast, the above-mentioned PN-sequence fam- assume that the joint probability function can be decomposed
ily has many more codes than the number of chips in a into the product of some positive functions, namely,
sequence, but since the codes are non-orthogonal, they impose
1 
D
interference even in the absence of non-dispersive channels.  
Hence they require more complex MUDs. Another natural p x1, x2 , . . . , xE = fd (Xd ), (21)
Z
idea, which leads to LDS-CDMA, is to use sparse spreading d=1
sequences instead of the classic “fully-populated” spreading where Z is a normalized constant, Xd is a subset of
sequences to support more users, where the number of non- {x1, x2 , . . . , xE }, and f1 (X1 ), f2 (X2 ), . . . , fD (XD ) are positive
zero elements in the spreading sequence is much lower than functions which are not necessarily the probability functions.
Nc for the sake of reducing the interference imposed on each Then, we can translate this form into the factor graph, which
chip. Therefore, LDS-CDMA is potentially capable of improv- is a bipartite graph, as shown in Fig. 10 [128], where the cir-
ing the attainable system performance by using low-density cles represent variable nodes corresponding to x1, x2 , . . . , xE ,
spreading sequences [64], which is the key distinguishing while the squares indicate observation nodes corresponding
feature between conventional CDMA and LDS-CDMA. to f1 (X1 ), f2 (X2 ), . . . , fD (XD ). An edge is present between a
In LDS-CDMA, all transmitted symbols are modulated onto variable node xe and an observation node fd (Xd ) if and only
sparse spreading sequences. In this way, each user will only if xe ∈ Xd .
spread its data over a small number of chips, as shown in In general, the message passing algorithm relies on the fac-
Fig. 9. As a result, the number of the superimposed signals at tor graph representation of the problem as its input and returns
each chip will be less than the number of active users, which the marginal distribution of all variable nodes. Messages can
2306 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

reformulated as
 
Nc
x̂k = arg max p(yn |x[n] ), (26)
a∈X
∼{xk }xk =a n=1

where
⎧ ⎫
 1 ⎨ 1  ⎬
p yn |x[n] = √ exp − 2 (yn − hn,k xk )2 .
2πσ ⎩ 2σ ⎭
k ∈N (n)
Fig. 10. Factor graph representation of MPA. (27)

N Observe that (26) represents a marginal process of

n=1 p(yn |x[n] ), which is similar to the form of the decom-
c
be passed between the variable node and the observation node
through the edge between them, and the message can be inter- posable joint probability function of the message passing
preted as the soft-value that represents the reliability of the algorithm, apart from a normalization constant Z. To elabo-
variable associated with each edge. The marginal distribution rate a little further, we can regard each term p(yn |x[n] ) as
of a variable node can be interpreted as a function of the mes- a positive function fd (Xd ) in the message passing algorithm.
sages received by that variable node. The iterative form of the Then, the factor graph shown in Fig. 9 can be constructed just
message passing algorithm can be represented as like that of Fig. 10. Therefore, we can rewrite the iterative
  Equations (22) and (23) as follows:
(t) (t−1)
md→e (xe ) ∝ fd (Xd ) mi→d (xi ),  1 1
(t)
{xi |i∈N (d)\e} i∈N (d)\e mn→k (xk ) ∝ √ exp{− 2 (yn − hn,k xk
2πσ 2σ
(22) {xi |i∈N (n)\k }
(t)
 (t−1)
  (t−1)
me→d (xe ) ∝ mi→e (xe ), (23) − hn,i xi )2 } mi→n (xi ),
i∈N (e)\d i∈N (n)\k i∈N (n)\k
(28)
(t)
where md→e (xe ) denotes the message transmitted from the (t)
 (t−1)
mk →n (xk ) ∝ mi→k (xk ). (29)
observation node fd (Xd ) to the variable node xe at the tth
(t) i∈N (k )\n
iteration. Similarly, me→d (xe ) presents the message trans-
mitted from the variable node xe to the observation node Finally, the (approximate) marginal probability distribution
fd (Xd ). If the maximum number of iterations is T, the of each variable after T iterations can be calculated by (24).
marginal probability distribution for each variable can be In the case of LDS, the number of edges in the factor graph
finally calculated as is relatively low, hence less and longer loops can be expected
 (T )
based on a meritorious design of the factor graph based on
p(xe ) ∝ md→e (xe ). (24) a beneficial sparse spreading sequence design. Additionally,
d∈N (e) assuming that the maximum number of users superimposed at
the same chip is w, the receiver complexity is on the order
It has been theoretically shown that the marginal distribution of O(|X|w ) instead of O(|X|K ) (K > w ) for conventional
can be accurately estimated with the aid of a limited number CDMA.
of iterations, provided that the factor graph does not have sys- The performance of LDS-CDMA and direct sequence-
tems [128]. However, in many practical situations the presence CDMA (DS-CDMA), which is adopted by the 3G WCDMA
of loops cannot be avoided. Fortunately, the message pass- systems, have been compared in [64]. More specifically,
ing algorithm is quite accurate for “locally tree like” graphs, as shown in Fig. 11 [64], LDS-CDMA outperforms DS-
which implies that the length of the shortest loop is restricted CDMA using the best-found spreading sequences, where the
to O(log(E )). Therefore, in most practical applications asso- MMSE-based partial parallel interference cancellation (PPIC)
ciated with a sparse structure, we can obtain an accurate receiver has been adopted by both schemes. Furthermore,
marginal distribution estimate by an appropriate design of the when LDS-CDMA relies on an MPA receivers, its perfor-
factor graph. mance approaches the single user bound within a small margin
In LDS-CDMA, the optimum MAP detection of x in (18) of 1.17 dB at a BER of 10−4 .
can be formulated as: 2) Low-Density Spreading Aided OFDM (LDS-OFDM):

x̂k = arg max p(x|y). (25) OFDM and MC-CDMA are close relatives, especially when
a∈X considering frequency-domain-spreading, which spreads each
∼{xk }xk =a
user’s symbols across all the OFDM subcarriers, provided that
Without loss of generality, we assume that the transmitted the number of spreading-code chips is identical to the num-
symbols and noise are identically and independently dis- ber of subcarriers. Then multiple users may be supported by
tributed (i.i.d), and the transmitted symbols obey the uniform overlaying the unique, user-specific spreading sequences of all
distribution. Then according to Bayes’ rule, (25) can be users on top of each other across all subcarriers. As always, the
DAI et al.: SURVEY OF NOMA FOR 5G 2307

Fig. 12. SCMA encoding and multiplexing.

LDS-OFDM was compared to both SC-FDMA and OFDMA

Fig. 11. Comparison of LDS-CDMA and DS-CDMA [64] IEEE.
c in terms of the peak-to-average-power ratio (PAPR), as well
as the link-level and system-level performance. The associ-
ated simulation results showed that LDS-OFDM is capable of
significantly improving the attainable system performance at
spreading sequences may be chosen to be orthogonal Walsh- a given transmission power, spectral efficiency and fairness.
Hadamard codes or non-orthogonal m-sequences, as well as Furthermore, in order to improve the achievable performance
LDSs, for example. of LDS-OFDM, a joint subcarrier and power allocation method
Hence LDS-OFDM can be interpreted as an integrated ver- was proposed in [72], with the objective of maximizing the
sion of LDS-CDMA and OFDM, where for example each weighted sum-rate using an efficient greedy algorithm. As a
user’s symbol is spread across a carefully selected num- further result, a pair of PAPR reduction techniques have been
ber of subcarriers and overlaid on top of each other in the proposed for LDS-OFDM in [73].
frequency-domain. To elaborate a little further, in the conven- 3) Sparse Code Multiple Access (SCMA): The recently pro-
tional OFDMA system, only a single symbol is mapped to a posed SCMA technique constitutes another important NOMA
subcarrier, and different symbols are transmitted on different scheme, which relies on code domain multiplexing devel-
subcarriers, which are orthogonal and hence do not interfere oped from the basic LDS-CDMA scheme. In [74], SCMA
with each other. Therefore, the total number of transmitted was extensively discussed in terms of its transmission and
symbols is restricted by the number of orthogonal subcarri- multiplexing aspects, as well as in terms of its factor graph
ers. By contrast, in the LDS-OFDM system, the transmitted representation and receiver architecture relying on the mes-
symbols are firstly multiplied with LDS sequences, whose sage passing algorithm. In contrast to the basic LDS-CDMA,
length is equal to the number of subcarriers and the resultant as illustrated in Fig. 12, the bit-to-constellation mapping and
chips are transmitted on different subcarriers. When using LDS spreading operations in SCMA are intrinsically amalgamated,
spreading sequences, each original symbol is only spread to a hence the original bit streams are directly mapped to differ-
specific fraction of the subcarriers. As a result, each subcar- ent sparse codewords, where each user has its own codebook.
rier carries chips related to a fraction of the original symbols. Without loss of generality, we assume that there are J code-
Suffice to say that apart from the already accentuated bene- books, where each codebook contains M codewords of length
fits, frequency-domain spreading is particularly advantageous Kl , and the number of non-zero elements in every codeword
in strongly frequency-selective channels, which would often is Nnz . For example, in Fig. 12 we have J = 6, M = 4,
obliterate some of the subcarriers and their information, whilst K l = 4, and Nnz = 2. We consider the rank-deficient sce-
in the presence of frequency-domain spreading they would nario of Kl < J , which is capable of supporting the massive
only affect some of the chips conveying the original symbols. connectivity expected in 5G. All codewords in the same code-
This is likely to allow us to still recover the original symbols. book contain zeros in the same (Kl −Nnz ) dimensions, and the
We note in closing that the family of MUDs designed using positions of zeros in the different codebooks are unique and
the message passing algorithm for LDS-CDMA can also be distinct for the sake of facilitating collision avoidance for any
used for LDS-OFDM in order to separate the overlaid symbols pair of users. Therefore, the maximum number of codebooks
at the receiver. is  by the selection of Kl and Nnz , which is equal to
 restricted
At the time of writing a number of insightful LDS-OFDM Kl
. For each user, log2 M bits are mapped to a complex
investigations have already been disseminated in the literature. Nnz
For example, the system model and properties of LDS-OFDM, codeword. The codewords of all users are then multiplexed
including its frequency diversity order, receiver complexity, onto Kl shared orthogonal resources, such as the OFDM sub-
and its ability to operate under rank-deficient conditions in carriers. Due to the sparsity of codewords, the signal received
the presence of more users than chips have been presented on subcarrier k can be represented by
in [68]. An upper limit was imposed on the number of users 
per subcarrier, in order to control the receiver complexity yk = hkj xkj + wk , (30)
in [69]. Additionally, in [70] and [71], the performance of j ∈N (k )
2308 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

since the different layers are multiplexed with the aid of dif-
ferent codebooks. However, the best design criterion to be
used for solving the multi-dimensional constellation problem
is unknown at the time of writing. Having said this, it is antic-
ipated that using the powerful semi-analytical tool of extrinsic
information transfer-function (EXIT) charts for jointly design-
ing the channel code and the constellation would lead to
near-capacity performance. As a further solution, a multi-stage
design approach has been proposed for finding a meritori-
ous sub-optimal solution in [74]. More details concerning the
codebook design can be found in [87].
Specifically, in order to simplify the optimization prob-
lem of the multi-dimensional constellation design, a mother-
Fig. 13. Performance comparison between SCMA and LDS [74] IEEE.
c constellation can be generated first by minimizing the average
alphabet energy for a given minimum Euclidian distance
between any two constellation points. More particularly,
an optimized design of the mother-constellation based on
where xkj is the kth component of the codeword xj for user the classic star-QAM signaling constellation has been pro-
j, hkj is the channel gain of user j at the kth subcarrier, posed in [88]. The resultant simulation results showed that
and wk denotes the complex-valued Gaussian noise with zero the star-QAM based codebooks are capable of significantly
mean and variance σ 2 . Similar to LDS-CDMA, the message enhancing the BER performance of the square-QAM based
passing algorithm can also be used for MUD at the SCMA codebooks. Once the mother-constellation has been obtained,
receiver. However, the receiver complexity may become exces- the codebook-specific operation can be applied to the mother-
sive. To circumvent this problem, improved variants of the constellation in order to obtain specific constellations for each
message passing algorithm have been proposed in [75]–[83]. codebook. More specifically, the codebook-specific operations,
Specifically, a low-complexity logarithmic-domain message such as phase rotation, complex conjugation and dimensional
passing algorithm (Log-MPA) was proposed in [84]. The asso- permutation, can be optimized for introducing correlation
ciated simulation results showed that the performance degrada- among the non-zero elements of the codewords, which is ben-
tion of Log-MPA over the full-complexity MPA was negligible eficial in terms of recovering the codewords contaminated by
in practical applications, despite the fact that the Log-MPA the interference imposed by other tones. Additionally, differ-
achieved over 50% complexity reduction. For Log-MPA, the ent power can be assigned to the symbols superimposed over
conditional channel probability calculation imposes up 60% the same time-frequency index for ensuring that the message
of the total computational complexity in the whole decod- passing algorithm can operate more efficiently by mitigat-
ing procedure. In [75], a dynamic search algorithm based on ing the interferences between the paired layers. Furthermore,
classic signal uncertainty theory was proposed for eliminating inspired by the family of irregular low density parity check
any unnecessary conditional channel probability calculation (LDPC) codes, an irregular SCMA structure has been pro-
without degrading the decoding performance. On the other posed in [89], where the number of non-zero elements of the
hand, in order to improve the BER performance of SCMA, the codewords can be different for different users. In this way,
powerful turbo-principle has been invoked in [85] for exchang- users having different QoS requirements can be simultaneously
ing extrinsic information between the SCMA detector and the served.
channel decoder. By contrast, in [86], a low-complexity turbo- Again, in the uplink of a SCMA system, grant-free multiple
like combination of iterative detection and iterative decoding access can be realized by carefully assigning the codebooks
was conceived for striking a compelling performance versus and the pilots to the users based on [90]. As mentioned in
complexity balance. Section II, a user does not have to send a scheduling request to
Quantitatively, the performance of SCMA relying on a the BS in the grant-free transmission scheme, thus a significant
multi-dimensional constellation having four points was com- latency- and signaling overhead-reduction can be expected. As
pared to that of LDS using QPSK modulation in Fig. 13 [74]. shown in Fig. 14, the pre-configured resource to be assigned
These simulation results show that SCMA outperforms LDS to the users may be referred to as a contention transmission
in terms of its block error ratio (BLER). The key difference units (CTU). There are J codebooks defined over a time-
between LDS and SCMA is that SCMA relies on multi- frequency resource, and L pilot sequences are associated with
dimensional constellations for generating its codebooks, which each codebook. The grant-free multiple access regime allows
results in the so-called “constellation shaping gain” [74]. This contentions to occur, when multiple users are assigned to the
gain is unavailable for other NOMA schemes. More explic- same CTU. The network detects the uplink packets by attempt-
itly, the “shaping gain” terminology represents the average ing their reception using all possible access codes assigned to
symbol energy gain, when we change the shape of the mod- the predefined contention region. Then a random back-off pro-
ulation constellation. In general, the shaping gain is higher cedure can be invoked, when collisions occur. At the receiver,
when the shape of the constellation becomes similar to a blind detection and compressive sensing (CS) techniques can
sphere. However, the SCMA codebook design is complex [74], be used for performing joint activity and data detection, e,g.,
DAI et al.: SURVEY OF NOMA FOR 5G 2309

Fig. 15. Uplink MUSA system.

results in beneficial interference averaging). Then, all symbols

Fig. 14. Definition of a contention transmission unit (CTU) [90] IEEE.
c after spreading are transmitted over the same time-frequency
resources, such as OFDM subcarriers. Without any loss of
generality, we assume that each user transmits a single symbol
every time, and that there are K users as well as N subcarriers.
with the aid of the joint message passing algorithm (JMPA) Rank-deficient scenarios can also be supported by MUSA, i.e.,
of [91]. K > N, which will impose interference amongst the users. At
SCMA can also be used in the downlink in order to improve the receiver, linear processing and SIC are performed in order
the system throughput, leading to the multi-user SCMA to separate the different users’ data according to their channel
(MU-SCMA) concept [92]. Based on a limited knowledge of conditions.
the channel conditions of different users, the BS simply pairs In the downlink of the MUSA system the users are separated
the users, where the transmit power is appropriately shared into G groups. In each group, the different users’ symbols are
among multiple users. This regime is hence reminiscent of the weighted by different power-scaling coefficients and then they
NOMA scheme relying on the previously mentioned power- are superimposed. Orthogonal sequences of length G can be
domain multiplexing. Compared to MU-MIMO, MU-SCMA used as spreading sequences in order to spread the superim-
is more robust to channel-quality variations, and indeed, the posed symbols from G groups. More specifically, the users
provision of near-instantaneous CSI feedback is unnecessary from the same group employ the same spreading sequence,
for this open-loop multiple access scheme [92]. In [93], the while the spreading sequences are orthogonal across the dif-
concept of single-cell downlink MU-SCMA is extended to ferent groups. In this way, the inter-group interferences can
an open-loop downlink CoMP solution, which was termed be removed at the receiver. Then, SIC can be used for carry-
as MU-SCMA-CoMP. In this scheme, the SCMA layers and ing out intra-group interference cancellation by exploiting the
transmits power are shared among multiple users within a associated power difference.
CoMP cluster. The analysis and simulation results in [93] In MUSA, the spreading sequences should have low cross-
demonstrated its robustness both to user mobility and channel correlation in order to facilitate near-perfect interference can-
aging. Furthermore, the capacity of downlink Massive MIMO cellation at the receiver. The MUSA technique is capable of
MU-SCMA was analyzed in [94] based on random matrix the- improving the downlink capacity, which is an explicit ben-
ory and it has been shown that compared to Massive MIMO efit of the associated SINR difference and SIC. As a further
MU-OFDMA systems, Massive MIMO MU-SCMA is capable compelling benefit, MUSA is capable of guaranteeing fairness
of achieving a higher sum rate. amongst the multiplexed users without any capacity loss. In
In a nutshell, the efficiency of SCMA has been verified both a nutshell, with the advent of advanced spreading sequences
by simulations and real-time prototyping in [95]. Both the lab and powerful state-of-the-art SIC techniques, substantial gains
tests and the field tests demonstrated that SCMA is capable can be obtained by MUSA, even for a normalized user-load
of supporting upto three times more users than the number of as high as say 300%, which is shown in Fig. 16 [96].
resource-slots, whilst still maintaining a link-integrity close to 5) Successive Interference Cancellation Aided Multiple
that of orthogonal transmissions. Access (SAMA): Let us consider an uplink SAMA system
4) Multi-User Shared Access (MUSA): MUSA is another supporting K users with the aid of N orthogonal OFDM
NOMA scheme relying on code-domain multiplexing, which subcarriers, where we have K > N, i.e., when the system
can be regarded as an improved CDMA-style scheme. is rank-deficient. The system model of SAMA is similar to
In the uplink of the MUSA system of Fig. 15, all transmit- that of MUSA, but in SAMA, the non-zero elements of any
ted symbols of a specific user are multiplied with the same spreading sequence bk for user k are equal to one, and the
spreading sequence (Note that different spreading sequences spreading matrix B = (b1 , b2 , . . . , bK ) is designed based on
can also be used for different symbols of the same user, which the following principles [97].
2310 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

CDMA, based on the following philosophy. Even if orthog-

onal Walsh-Hadamard spreading sequences are employed for
distinguishing the users in CDMA systems, when they are
transmitted over dispersive channels, their orthogonality is
destroyed by their convolution with the CIR even in the
absence of co-channel interference. Hence we end up with
a potentially infinite variety of received sequences. This leads
to the appealing concept of simply using the unique, user-
specific CIRs for distinguishing the users, instead of unique,
user-specific spreading sequences. Naturally, when the users
transmitting in the uplink are close to each other, their CIRs
Fig. 16. The performance of MUSA with different-length spreading
sequences at different normalized user-loads [96] IEEE.
c
become quite similar, which aggravates the task of the MUD
in separating their signals. The beneficial properties of this
family of solutions have attracted substantial research efforts,
• The number of groups with different number of 1’s in the as detailed in [98]–[111].
spreading sequence should be maximized. To elaborate a little further, given the potentially infi-
• The number of the overlapped spreading sequences which nite variety of CIRs, these sophisticated SDMA systems are
have the same number of 1’s should be minimized. capable of operating under highly rank-deficient conditions,
Then the maximum number of user supported with the aid of namely when the number of mobile users transmitting in
N orthogonal subcarriers can be calculated as the uplink is much higher than the number of BS uplink-
      receiver antennas. This would avoid the hard-limited user-load
N N N
+ + ··· + = 2N − 1. (31) of the Walsh-Hadamard code based CDMA systems, since
1 2 N the system-performance would only gracefully decay upon
For example, spreading matrices for N = 2, K = 3, N = 3, increasing the number of users. The resultant SDMA systems
K = 7, and N = 4, K = 15 can be designed as follows: tend to exhibit a similar performance to their rank-deficient
  CDMA counterparts relying on m-sequences for example.
1 1 0
B2,3 = , (32) Since these SDMA systems rely on the CIR for distinguish-
1 0 1
⎛ ⎞ ing the users, they require accurate CIR estimation, which
1 1 0 1 1 0 0 becomes extremely challenging, when the number of users
B3,7 = ⎝ 1 1 1 0 0 1 0 ⎠, (33) is much higher than the number of BS receiver antennas.
1 0 1 1 0 0 1 This logically leads to the concept of joint iterative chan-
⎛ ⎞
1 1 0 1 1 1 1 1 0 0 0 1 0 0 0 nel and data estimation, which attracted substantial research
⎜1 1 1 0 1 1 0 0 1 1 0 0 1 0 0⎟ interests [102], [108]. These high-end solutions often rely
B4,15 = ⎜
⎝ 1 1 1 1 0 0 1 0 1 0 1 0 0 1 0 ⎠.
⎟
on powerful non-linear bio-inspired MUDs exchanging their
1 0 1 1 1 0 0 1 0 1 1 0 0 0 1 soft-information with the channel estimator.
(34) Below we will elaborate further on a variety of powerful
solutions in a little more detail. In [98] the family of minimum
At the receiver, the message passing algorithm is invoked for bit error rate (MBER) MUDs was shown to be capable of out-
separating the signals of the different users. The design objec- performing the classic minimum mean-squared-error (MMSE)
tive used for determining the spreading matrices in SAMA is MUD in term of the achievable BER owing to directly mini-
to facilitate convenient interference cancellation [97]. Consider mizing the BER cost function. In this paper, genetic algorithms
B4,15 for example. The spreading sequence of the first user (GAs) were invoked for finding the optimum weight vectors
has 4 non-zero elements, hence the resultant diversity order of the MBER MUD in the context of multiple antenna aided
is 4. Thus the first user’s symbol is the most reliable one. multi-user OFDM. It was shown that the MBER MUD is
Therefore, the first user’s symbol can be readily determined capable of supporting more users than the number of receiver
in a few iterations, which is beneficial for the convergence antennas available in highly rank-deficient scenarios.
of the symbol detection process of all the other users having A novel parallel interference cancelation (PIC) based turbo
lower diversity orders. space time equalizer (STE) structure was designed in [99]
for multiple antenna assisted uplink receivers. The proposed
C. Other NOMA Schemes receiver structure allowed the employment of non-linear type
Apart from the prominent power-domain NOMA and code- of detectors such as the Bayesian decision feedback (DF)
domain NOMA solutions discussed in the previous two assisted turbo STE or the MAP STE, while operating at a mod-
subsections, recently a range of alternative NOMA schemes erate computational complexity. The powerful receivers based
have also been investigated, which will be discussed in this on the proposed structure tend to outperform the linear turbo
subsection. detector benchmarker based on the classic MMSE criterion,
1) Spatial Division Multiple Access (SDMA): SDMA is even if the latter aims for jointly detecting all transmitters’
one of the powerful NOMA schemes, and the philosophy signals. Additionally, the PIC based receiver is also capa-
of SDMA may be deemed to be related to that of classic ble of equalizing non-linear binary pre-coded channels. The
DAI et al.: SURVEY OF NOMA FOR 5G 2311

performance difference between the presented algorithms was channel estimation and turbo MUD/decoding in the context of
discussed using the powerful semi-analytical tool of extrinsic SDMA/OFDM systems is a challenging problem, which has
information transfer-function (EXIT) charts. to consider both the channel estimation problem formulated
Wang et al. [100] demonstrated that the iterative exchange over a continuous search space and the MUD optimization
of extrinsic information between the K-best sphere detector problem defined over a discrete search space. Hence the capa-
(SD) and the channel decoder is appealing, since it is capable bility of the GA, RWBS, PSO, and DEA to achieve optimal
of achieving a near MAP performance at a moderate com- solutions at an affordable complexity was investigated in this
plexity. However, the computational complexity imposed by challenging application by Zhang et al. [102]. Their study
the K-best SD significantly increases when using a large value demonstrated that the EA-assisted joint channel estimation
of K for the sake of maintaining a near-MAP performance in and turbo MUD/decoder are capable of approaching both
a high-throughput uplink SDMA/OFDM system supporting a the Cramer-Rao lower bound of the optimal channel esti-
large number of users and/or a high number of bits/symbols. mation and the BER performance of the idealized optimal
This problem is further aggravated when the number of Maximum likelihood (ML) turbo MUD/decoder associated
users/MSs exceeds that of the receive antennas at the BS, with perfect channel estimation, respectively, despite impos-
namely, in the challenging scenario of rank-deficient systems. ing only a fraction of the idealized turbo ML-MUD/decoder’s
It was demonstrated that the iterative decoding convergence complexity.
of this two-stage system may be improved by incorporating a From the discussions above, we can see that the concept of
unity rate code (URC) having an infinite impulse response, NOMA has already existed in various systems, such as SDMA,
which improves the efficiency of the extrinsic information where users are distinguished using the unique, user-specific
exchange. Although this results in a slightly more com- CIRs. Actually, these systems require accurate CIR estima-
plex three-stage system architecture, it allows us to use a tion to successfully realize MUD, which becomes extremely
low-complexity SD having a significantly reduced detection challenging when the number of users is much higher than
candidate list size. Alternatively, a reduced SNR is required. that of receiver antennas. Solving this CIR estimation prob-
For example, given a target BER of 10−5 and a candidate lem logically leads to the concept of joint channel and data
list size of 32 for the SD, the three-stage receiver is capable estimation, and these high-end solutions often rely on pow-
of achieving a performance gain of 2.5 dB over its two-stage erful non-linear MUDs. In fact, most of the studies focus on
counterpart in a rank-deficient SDMA/OFDM 4-QAM system MUD design, and a series of non-linear MUD algorithms such
supporting eight co-channel users and employing for receive as parallel interference cancellation (PIC) [99], and space-
antennas at the BS, namely, in an (8 × 4) rank-deficient system time decision feedback equalization (ST-DFE) [101] have been
having a normalized user-load of two. For the sake of further proposed. Further, with the development of evolutionary algo-
enhancing the three-stage concatenated receiver, the proposed rithms, algorithms like genetic algorithm (GA), and particle
iterative center-shifting SD scheme and the so-called irregular swarm optimization (PSO) may be explored to acquire accu-
convolutional codes (IrCCs) were intrinsically amalgamated, rate CIR estimation [102]. In contrast, power-domain NOMA
which led to an additional performance gain of 2 dB. transmits the superposition of multi-user signals with differ-
Chen et al. [101] proposed a space-time decision feedback ent power-allocation coefficients, and usually SIC is used at
equalization (ST-DFE) assisted MUD scheme for multiple the receiver to detect multi-user signals. As the channel gain
receiver antenna aided SDMA systems. Again, a sophisticated difference among users is translated into different multiplex-
MBER MUD design was invoked, which was shown to be ing gains [5], both user grouping and resource allocation have
capable of improving the achievable BER performance and substantial effects on the achievable throughput. As a result,
enhancing the attainable system capacity over that of the stan- most of the studies concerning power-domain NOMA focus
dard MMSE design. An appealing adaptive implementation of on user grouping, resource (power) allocation, and perfor-
the MBER ST-DFE assisted MUD was also proposed using a mance analysis. Recently, with the development of mmWave
stochastic gradient-based least bit error rate algorithm, which communication and massive MIMO, combining power-domain
was demonstrated to consistently outperform the classical least NOMA with mmWave and massive MIMO has become a
mean square (LMS) algorithm, while imposing a lower com- promising technique [129]–[131].
putational complexity than the LMS algorithm for the binary 2) Pattern Division Multiple Access (PDMA): Apart from
signalling scheme considered. It was demonstrated that the the SDMA scheme mentioned above, the family of PDMA
MBER ST-DFE assisted MUD is more robust to channel esti- schemes [112], [113] constitutes another promising NOMA
mation errors as well as to potential error propagation imposed class that can be implemented in multiple domains. At the
by decision feedback errors, than the MMSE ST-DFE assisted transmitter, PDMA employs non-orthogonal patterns, which
MUD. are designed by maximizing the diversity and minimizing the
The development of evolutionary algorithms (EAs) [102], correlations among the users. Then, multiplexing can be real-
such as GAs, repeated weighted boosting search (RWBS), ized in the code-, power- or spatial-domains, or in fact in
particle swarm optimization (PSO), and differential evolu- their combinations. Multiplexing in the code domain is remi-
tion algorithms (DEAs) stimulated wide interests in the niscent of SAMA [97]. Multiplexing in the power domain has
communication research community. However, the quan- a system model similar to multiplexing in the code domain,
titative performance-versus-complexity comparison of GA, but power-scaling has to be considered under the constraint
RWBS, PSO, and DEA techniques applied to the joint of a given total power. Multiplexing in the spatial domain
2312 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

Fig. 18. The schematic of IGMA transmitter [136] IEEE.

c

4) Interleaver-Based NOMA: Interleave-grid multiple

Fig. 17. The LSSA transmitter structure [132] IEEE.
c
access (IGMA) is an interleaver-based multiple access
scheme, which can distinguish the different users based on
their different bit-level interleavers, different grid mapping
patterns, or the combinations of these two techniques [136].
leads to the concept of spatial PDMA, which relies on multi- The typical transmitter structure of IGMA is shown in
antenna aided techniques. In contrast to MU-MIMO, spatial Fig. 18. Specifically, the channel coding process can be either
PDMA does not require joint precoding for realizing spa- simple repetition coding (spreading) of moderate coding rate
tial orthogonality, which significantly reduces the system’s for classic forward error correction (FEC) or low-rate FEC.
design complexity. Additionally, multiple domains can be By contrast, the grid mapping process of Fig. 18 may vary
combined in PDMA to make full use of the various wireless from sparse mapping based on zero padding to symbol-level
resources available. The simulation results of [113] demon- interleaving, which could provide another dimension for user
strated that compared to LTE, PDMA is potentially capable of multiplexing. Whilst we need well-designed FEC codes and
achieving a 200% normalized throughput gain in the uplink, spreading code sequences, the design of bit-level interleavers
and more than 50% throughput gain may be attained in the and/or grid mapping patterns in Fig. 18 is somewhat more
downlink. related. They provide scalability to support different con-
3) Signature-Based NOMA: Signature-based NOMA nection densities, whilst striking a trade off between the
schemes are also proposed as promising candidates for 5G. channel coding gain and the benefit gleaned from sparse
Low code rate and signature based shared access (LSSA) resource mapping. Moreover, the symbol-level interleaving
is one of them, and the transmitter structure conceived for of Fig. 18 randomizes the symbol sequence order, which
uplink massive machine-type communication (mMTC) is may bring about further benefits in terms of combating the
depicted in Fig. 17 [132]. LSSA [132] multiplexes each user’s deleterious effects of frequency selective fading and inter-cell
data either at bit or symbol level with the aid of a specific interference. Besides, a relatively low-complexity multi-user
signature pattern, which consists of a reference signal (RS), detector can be applied and the employment of a sparse grid
complex/binary sequence, and permutation pattern of a short mapping pattern could further reduce the detection complexity
length vector. All the users’ signatures share the same short imposed.
vector length, which can be chosen randomly by the mobile Another interleaver-based multiple access scheme, namely
terminal or assigned to the user by the network. Moreover, interleave-division multiple access (IDMA), has also been
LSSA can be optionally modified to have a multi-carrier proposed. Explicitly, IDMA interleaves the chips after the
variant in order to exploit the frequency diversity provided symbols have been multiplied by the spreading sequences.
by a wider bandwidth, and to reduce the latency. It can also Hence, IDMA is effectively chip-interleaved CDMA. As
support asynchronous uplink transmissions, because the BS is shown in [137], compared to CDMA, IDMA is capable of
capable of distinguishing/detecting the overlaid user signals achieving about 1 dB Eb /N0 gain at a BER of 10−3 in highly
by correlating them with the signature patterns, even if the loaded systems having a normalized user-load of 200%. The
transmission timing is different from each other. gain is mostly attributable to the fact that chip-interleaving
Similar to LSSA, resource spread multiple access results in an increased diversity-gain compared to conventional
(RSMA) [133], [134] also assigns unique signatures to sepa- bit-interleaving.
rate the different users and spreads their signals over all the 5) Spreading-Based NOMA: There are also many other
available time and frequency resources. The unique signa- NOMA schemes based on spreading codes, which are con-
tures may be constituted by the power, spreading/scrambling sistent with the concept of the aforementioned code-domain
codes with good correlation properties, interleavers, or their NOMA, and non-orthogonal coded multiple access (NCMA)
combinations, and interference-cancellation type receivers can is one of them [138]. NCMA is based on resource spread-
be utilized. Depending on the specific application scenarios, ing by using non-orthogonal spreading codes having a low
RSMA may include [135]: correlation. These codes can be obtained by finding the solu-
• single carrier RSMA: It is optimized for battery power tions of a Grassmannian line packing problem [139]. By
consumption and link budget extension by utilizing sin- imposing additional layers using superposition coding, it can
gle carrier waveforms and very low peak to average provide an increased throughput and improve connectivity
power ratio (PAPR) modulations. It allows grant-free at a low block error ratio (BLER). Furthermore, since the
transmission and potentially allows asynchronous access. receiver of the NCMA system adopts parallel interference
• Multi-carrier RSMA: It is optimized for low-latency cancellation (PIC), it has a scalable performance vs complex-
access and allows for grant-free transmissions. ity. Consequently, NCMA is eminently suitable for a larger
DAI et al.: SURVEY OF NOMA FOR 5G 2313

(LDS-SVE) proposed by Fujitsu [149], as well as the fre-

quency domain spreading (FDS) [150] and the low code rate
spreading (LCRS) [150] schemes proposed by Intel. They
spread the user symbols to multiple RBs in order to attain an
increased diversity gain. Repetition division multiple access
(RDMA) proposed by MTK [141] belongs to the family of
interleaver-based NOMA schemes, which can readily separate
the different users’ signals, whilst exploiting both time- and
Fig. 19. NOCA transmitter structure [140] IEEE.
c frequency-diversity with the aid of the cyclic-shift repetition
of the modulated symbols.

number of connections exchanging small packets in massive

machine-type communication (mMTC), or for reducing the D. User Grouping and Resource Allocation
collision probability in contention based multiple access. In power-domain NOMA, the channel gain difference
Non-orthogonal coded access (NOCA) is also a spreading among users is translated into different multiplexing gains [5].
based multiple access scheme [140]. Similar to other spread- Therefore, both the user grouping and resource allocation
ing based schemes, the basic idea of NOCA is that the data techniques have substantial effects on the achievable through-
symbols are spread using non-orthogonal sequences before put [151]. In order to optimize the user grouping and resource
transmission, which can be applied both in the frequency allocation, a reasonable optimization criterion has to be found
domain and/or in time domain based configurations. The basic first. Then, a compelling performance versus complexity trade-
transmitter structure is shown in Fig. 19, where SF denotes off has to be struck. Specifically, the classic proportional fair
the spreading factors and Cj is the spreading sequence of the (PF) scheduler is known to strike an attractive tradeoff between
jth user. Specifically, the original modulated data sequence is the capacity and fairness attained. Hence it has been used
first converted into P parallel sequences, then each of the P both in the orthogonal 3G and 4G multiple access systems.
sequences is spread across SF number of subcarriers. In order Therefore, it has also been widely considered as a benefi-
to meet the requirements of different scenarios, the spreading cial optimization criterion [5], [54]–[57] in NOMA systems.
factors can be adaptively varied. Specifically, both multi-user scheduling and power allocation
6) Bit Division Multiplexing (BDM): The basic concept per frequency block can be realized by maximizing the prod-
of BDM [114] relies on hierarchical modulation, but the uct of the average user throughput of all the users within a
resources of multiplexed users are partitioned at the bit level cell [5], [54]. In the uplink, the transmission power is usually
instead of the symbol level. Strictly speaking, the resource independently determined for all users, thus scheduling a user
allocation of BDM is orthogonal in the bit domain, but multi- set can be configured at a given power level. In the downlink,
user signals can share the same constellation, which implies under the constraint of a fixed total power, user scheduling
that they are superimposed in the symbol domain. and power allocation should be jointly optimized. More par-
7) Compressive Sensing (CS)-Based NOMA: CS may be ticularly, for a given scheduling user set, the classic iterative
readily combined with NOMA schemes for exploiting either water-filling based power allocation algorithm [5], [54] can
the user activity sparsity or data sparsity [143], [144]. A range be used, which achieves the maximum weighted sum of the
of CS-based random access schemes have been conceived user throughput, when exploiting the uplink-downlink dual-
recently, such as the family of asynchronous random access ity. Given the optimal power allocation for each scheduling
protocols [145] and compressive random access arrangement user set, the scheduling user set is selected by maximizing the
of [146]. Additionally, in [147], random multiple access rely- optimization criterion.
ing on CS was invoked for maximizing the system’s total More particularly, matching theory has been shown an effec-
throughput. Furthermore, the attainable throughput associated tive tool for user grouping in NOMA. Matching theory can
with different amount of channel knowledge was discussed, be used to efficiently solve the combinatorial problem of
which provided useful insights into the quantitative bene- matching players from two distinct sets according to players’
fits of CS in the context of throughput maximization in individual information and their preferences [152], [153]. Due
random multiple access schemes. Furthermore, the joint detec- to the low complexity, matching theory has been widely used
tion of both node activity and of the received data was to address the resource optimization problems in NOMA wire-
proposed in [148] for machine-type communication, which less networks [154]–[157]. The resource allocation in NOMA
exploited the sporadic nature of the expected communica- systems, such as user grouping and subchannel allocation, can
tion. All in all, CS is expected to play an important role in be modeled as classical matching structures (one-to-one [154],
NOMA schemes. many-to-one [155], [156], and many-to-many [157]). The sim-
8) Miscellaneous NOMA Schemes: Apart from the NOMA plest matching model is one-to-one matching, where each
schemes introduced above, there are four other schemes pro- user from one set can match with at most one user from
posed by different companies as part of the Rel-14 3GPP the opposite set. In a downlink cognitive radio inspired
NR Study Item shown in Table I. Among these four NOMA NOMA networks, a one-to-one matching problem was for-
schemes, three are spreading-based NOMA arrangements, mulated to optimize the user pairing to improve the system
namely the low density spreading-signature vector extension throughput [154]. Two users in the matched user pair can
2314 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

share the same spectrum to improve each user data rate In addition to spectral efficiency, energy efficiency is also
and the entire system sum rate. Regarding NOMA hetero- a KPI for 5G. In [156], with the assumption of the known of
geneous networks, a many-to-one matching was modeled perfect CSI at BS, the energy efficiency maximization problem
to address the spectrum allocation problem [155], where a was investigated in a downlink NOMA system by jointly opti-
swap-operation enabled matching algorithm was proposed to mizing the subchannel assignment and power allocation. Due
match the small base stations with the suitable resource block to the fact that the perfect CSI is challenging to obtain, [159]
aiming at maximizing small cell users throughout. In many- considered the energy efficiency maximization problem in a
to-many matching, at least one player in one set can be downlink NOMA system with imperfect CSI. Reference [160]
matched with multiple players in the opposite set. In [157], studied the energy efficiency optimization for the fading
in order to maximize the system sum rate, subchannel alloca- MIMO NOMA systems with statistical CSI at the transmit-
tion was formulated as a two-sided many-to-many matching ter, where the energy efficiency is defined by ergodic capacity
process for a downlink NOMA network. Sequentially, a under unit power consumption. Reference [161] investigated
two-to-one matching was utilized to improve the energy effi- the energy efficiency optimization problem in a cognitive radio
ciency of a downlink NOMA network [156]. In this research inspired multiuser downlink NOMA system subject to an indi-
work, a low complexity algorithm was proposed to allo- vidual quality of service constraint for each primary user,
cate multiple users on subchannels to maximize the system and an efficient algorithm based on the sequential convex
energy efficiency. The subchannel allocation was considered approximation method was presented to solve the formulated
as dynamic matching process between user set and subchan- non-convex fractional programming problem. From another
nel set. According to the predefined preference lists, each hand, the wireless power transfer can also integrated into
user can send matching request to its most preferred sub- NOMA system to further improve the energy efficiency of
channel. However, the subchannel can accept or reject the the NOMA systems, e.g., [162] considered the application of
user depending on the energy efficiency the user can pro- simultaneous wireless information and power transfer into a
vide on this subchannel. The matching process will terminate cooperative NOMA system to improve the energy efficiency
when there is no user left to match. Besides, for uplink of the system.
NOMA, user grouping based on CSI was investigated con- Considering the inter-cell interference, fractional frequency
sidering some predefined power allocation schemes in [158], reuse (FFR), which allows the users under different channel
where the optimization problem of user grouping to achieve conditions to rely on different frequency reuse factors, has
the maximum sum rate was analyzed in the large system been employed in [54], [59], and [61] for further enhanc-
limit for various scenarios, and some optimum and sub- ing the performance of the cell-edge users. FFR-based power
optimum algorithms with a polynomial-time complexity were allocation strikes a tradeoff between the frequency bandwidth
proposed. utilization per cell and the impact of inter-cell interference.
Additionally, the max-min fairness criterion associated with
instantaneous CSI knowledge, i.e., maximizing the minimum
user rate, and the min-max fairness criterion relying on the E. Comparison of NOMA Solutions
average-CSI knowledge, i.e., minimizing the maximum out- Based on the discussion above, Table II summarizes the
age probability, have been considered in [52] for deriving the comparison of existing dominant NOMA schemes.
power allocation. Low-complexity algorithms have also been From a theoretical perspective, code-domain NOMA is
developed for solving the associated non-convex problems. capable of achieving a beneficial “spreading gain” with
The max-min fairness criterion has also been used in MIMO the aid of using spreading sequences, which may also be
NOMA systems [58], where a dynamic user allocation and termed as codewords. However, this benefit cannot be readily
power optimization problem was investigated. Specifically, a reaped by the above-mentioned power-domain NOMA regime.
sub-optimal two-step method has been proposed. In the first Achieving a “spreading gain” is an innate benefit of classic
step, the power allocation is optimized by fixing a specific CDMA, which may also be viewed as a low-rate repetition-
combination of user allocation according to the max-min fair- style channel coding scheme, where the code-rate is given
ness, while the second step considered all the user allocation by the spreading factor. In simple plausible terms, when for
combinations. Furthermore, the joint power and channel allo- example one of the chips is corrupted by a high noise- or
cation optimization has been shown to be NP-hard in [59], interference-sample, the specific spreading code may still be
and an algorithm combining Lagrangian duality and dynamic recovered by a matched filter or correlator based receiver.
programming was proposed for delivering a competitive sub- In contrast to CDMA, SCMA is capable of achieving an
optimal solution. Furthermore, in [60], power allocation has extra “shaping gain” due to the optimization of the associated
been conceived for the NOMA downlink supporting two users multi-dimensional constellation [87].
when practical modulation schemes are employed. To elabo- We can also compare the dominant NOMA schemes in
rate a little further, the mutual information metric rather than terms of their signalling techniques and complexity. In power-
the Shannon-theoretic throughput metric has been used for domain NOMA, SIC constitutes one of the popular inter-
deriving a more accurate result, and it has been shown that ference cancellation techniques. The complexity of the SIC-
the power allocation problem formulated for maximizing the MMSE is O(K 3 ), where K is the number of users supported.
total mutual information depends on the modulation schemes Therefore, the complexity of SIC is significantly lower than
employed. that of the optimal full-search-based MUD, especially when
DAI et al.: SURVEY OF NOMA FOR 5G 2315

TABLE II
C OMPARISON OF NOMA S CHEMES P ROPOSED FOR THE R EL -14 3GPP NR S TUDY I TEM [142]

K is high, which is expected to be the case in practical 5G this part, we intend to present some performance evaluations
systems. Note that the implementation of NOMA in CoMP and transmission experiments of NOMA, so as to verify the
imposes a relatively high complexity, and a promising con- analytical results.
genial solution to this problem has been discussed in [32]. To assess the efficiency of NOMA, NTT DOCOMO
However, side-information has to be transmitted in order to performed performance evaluations and transmission exper-
signal the associated power assignment, which imposes a sig- iments using prototype equipment [20], [21], [164]–[167].
nalling overhead. Additionally, in code-domain NOMA, the Specifically, in their experiments, the radio frame configura-
specific spreading sequences or codebooks have to be known at tion was designed based on LTE Release 8, and the targets
the receiver in order to support the MUD, which will increase of these evaluations were Transmission Mode 3 (TM3) and
the signalling cost, especially when the receiver does not know Transmission Mode 4 (TM4), operating without and with feed-
which users are active. On the other hand, in LDS-CDMA, ing back the user’s precoding matrix index to the base station,
LDS-OFDM, SCMA, and SAMA, the complexity of the MPA- respectively [168].
based receiver is proportional to O(|X|w ), where |X| denotes Researchers also performed experiments in an indoor radio-
the cardinality of the constellation set X, and w is the maxi- wave environment using the prototype equipment in [168]. In
mum number of non-zero signals superimposed on each chip this experiment, both UE1 and UE2 are stationary, and the
or subcarrier. As a result, the complexity becomes high in former was near the base station, while the latter was at a
typical scenarios of massive connectivity. The BER perfor- point about 50 meters from the base station. It has been shown
mance of three typical code-domain NOMA schemes is also that NOMA was capable of obtaining a throughput gain of
compared in [163]. approximately 80% over OFDMA when a 2-by-2 SU-MIMO
is adopted.
Apart from NTT DOCOMO, Huawei Technologies have
F. Performance Evaluations and Transmission Experiments also developed a SCMA-based multi-user uplink prototype to
of NOMA verify the advantages of the SCMA technology in real com-
We have provided some theoretical analysis of NOMA in munication systems [169]. The Huawei demo system consists
the previous sections, which shows that NOMA yields a better of one base station using two antennas for diversity combined
performance than traditional OMA schemes, and this makes reception, and 12 single-antenna aided users for uplink access
it a promising candidate for 5G wireless communication. In and data transmission.
2316 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

TABLE III
S PECIFICATION FOR SCMA P ROTOTYPE OF H UAWEI [169] IEEE
c

Fig. 20. SCMA throughput gain over OFDM in field testing [169] IEEE.
c

The basic system configurations of the demo are aligned designed, and UEs are deployed at different locations to evalu-
with the current LTE TDD system. In particular, the ate the performance of SCMA under different conditions. The
researchers use the LTE TDD Configuration 1 and use four test cases are as follows:
OFDMA as the baseline for their performance comparison. • Case 1: 12 UEs closely located in an area without
The specifications of the prototype system are shown in mobility.
Table III. • Case 2: 12 UEs located in an area with distant separation
The in-lab prototype system relies on a software-defined but no mobility
baseband, which means that all the baseband processing is • Case 3: 12 UEs moving along a road about 120 meters
realized by a CPU instead of FPGA/DSP. At the base station away from the BS (open-loop power control maintains
side, a single server (Huawei Tecal RH2288) is responsible mediocre transmit power at UE))
for all the baseband processing, which is then connected to • Case 4: 12 UEs are moving along a road about 180
standard commercial radio frequency components (Huawei meters away from the BS (open-loop power control
product RRU3232). At the user side, the CPU of a laptop provides a comparatively high power for the UE))
(MacBook Pro ME294CH/A) is used for modeling the base- In all field trial tests, typical small packets of 20 bytes
band processing for two users, which is connected to two (METIS definition) are used as payload for both LTE and
mobile RF modules for testing. A user interface (UI) is devel- SCMA, and the scheduling resources in the whole system are
oped to show the real-time throughput of each UE, supporting limited to 4 RBSs in each subframe. The comparative test
the real-time change of both the user status and of the system results of OFDM and SCMA are shown in Fig. 20, which
operational modes as well. indicates that SCMA achieves a nearly 300% throughput gain
The prototype can run in either OFDMA or SCMA mode, over OFDM.
and supports real-time switching from one to the other. To
ensure a fair comparison, same data rate is maintained for IV. C HALLENGES , O PPORTUNITIES , AND
each user to guarantee the same quality of service. It is shown F UTURE R ESEARCH T RENDS
by the test that compared to the orthogonal multiple access Existing NOMA schemes relying on either power-domain
baseline of 4G LTE, SCMA technology attains up to 300% or code-domain multiplexing are capable of improving the
throughput gain. For instance, 150% throughput gain can be spectral efficiency with the aid of non-orthogonal resource
observed by considering the fact that, if each user requires sharing. What’s more, NOMA techniques are capable of oper-
12 physical resource blocks (RBs), a system having a total ating in rank-deficient scenarios, which facilitates the support
of 48 RBs can serve at most 4 users using orthogonal LTE of massive connectivity. Therefore, NOMA solutions are con-
OFDMA. However, with SCMA, the codebook design sup- sidered as potentially promising 5G candidates. However, there
ports 6 users with the same throughput to simultaneously share are still numerous challenging problems to be solved. Hence
48 RBs, thus the equivalent delivered amount of data is actu- below some of the key challenges of NOMA designs will
ally 12 × 6 = 72 RBs rather than 12 × 4 = 48 RBs, which be highlighted, along with opportunities and future research
results in the throughput gain of about 72/48 = 150%. The trends addressing these challenges.
300% gain can be calculated in a similar way, but needs a
different codebook associated with a larger spreading factor
and larger number of data layers. In the prototype, a 24-by-8 A. Theoretical Analysis
SCMA codebook is used to support 12 users (each having 2 To elaborate a little further, in-depth theoretical analysis is
data streams) to transmit simultaneously. By contrast, for LTE required to provide additional insights to guide and inform the
OFDMA, only 4 users out of 12 can transmit data. associated system design. The attainable capacity of multiple
Apart from the fading simulator-based test in the Lab, a access schemes constitutes one of the most essential system
SCMA prototype has also been deployed in field trials to eval- performance criteria. Specifically, the capacity bounds of code-
uate the performance. Specifically, four different test cases are domain NOMA relying sophisticated spreading sequences has
DAI et al.: SURVEY OF NOMA FOR 5G 2317

to be investigated. Similar methods and tools can also be more efficiently exchanged between the decoder and demod-
conceived for MC-CDMA. On the other hand, the maximum ulator used at the receiver for improving the signal detection
normalized user-load that may be supported is limited both by performance.
the achievable interference cancellation capability and by the For an SIC-based receiver, the associated error propagation
affordable receiver complexity, which is related to the specific may degrade the performance of some users. Therefore, at
design of both the spreading sequences and the receiver. each stage of SIC, a high-performance non-linear detection
algorithm can be invoked for alleviating the influence of error
propagation.
B. Design of Spreading Sequences or Codebooks
D. Channel Estimation
In LDS systems, due to the non-orthogonal resource allo-
cation, there is mutual interference amongst the users. The In most NOMA contributions, perfect CSI is assumed
maximum number of superimposed symbols at each orthogo- for resource allocation or multi-user detection. However, it
nal resource “index” is determined by the particular spreading is not practical to obtain perfect CSI in practical systems,
sequences or codewords of the users, which has a direct impact hence channel estimation errors exist in NOMA. The impact
on the interference cancellation capability achieved at the of the residual interference imposed by realistic imperfect
receiver. Therefore, the factor graph of the message passing channel estimation on the achievable throughput of NOMA
algorithm should be optimized to strike a compelling tradeoff systems has been investigated in [48] and [49], and a low-
between the normalized user-load supported and the receiver complexity transmission rate back-off algorithm was con-
complexity imposed. ceived for mitigating the impact of the channel estimation
In addition, it has been shown that the message passing errors. Furthermore, the design of the practical channel estima-
algorithm is capable of determining the exact marginal distri- tors conceived for NOMA was investigated in [170] and [171],
bution in case of an idealized cycle-free factor graph, and an and some optimization algorithms have been proposed for
accurate solution can be obtained with the aid of “locally tree reducing the channel estimation error. Nevertheless, with
like” factor graphs, which implies that the cycle girth should the increase of the number of users in future 5G sys-
be sufficiently high. Graph theory can be used to design cycle- tems, more grave inter-user interference will be caused,
free or “locally tree like” factor graphs for NOMA without any which in turn may result in severe channel estimation error.
loss of spectral efficiency. On the other hand, realistic factor Therefore, more advanced channel estimation algorithms are
graphs exhibiting cycles can be decomposed into cycle-free required to achieve accurate channel estimation in NOMA
graphs in some practical applications. In this way, the mes- systems.
sage passing based receiver is capable of attaining the optimal
performance at the cost of a moderately increased receiver E. Grant-Free NOMA
complexity. Additionally, the classic matrix design principle As illustrated in Section II, a high transmission latency
and low-density-parity-check (LDPC) code based design meth- and a high signaling overhead are encountered by an access-
ods can be invoked for constructing the factor graph of NOMA grant based transmission scheme due to the uplink schedul-
solutions. ing requests and downlink resource assignments required.
Apart from the challenge of factor graph design, we should It is expected that NOMA is capable of operating with-
also consider how to choose the non-zero values for each out grant-free transmissions at a low transmission latency, at
sequence. The non-zero values superimposed at the same a small signaling overhead, whilst supporting massive con-
resource indices should be distinct. A promising technique is nectivity, especially in case transmitting short packets, as
to select different values from a complex-valued constellation expected in 5G. Hence contention-based NOMA schemes
for these non-zero elements in order to maintain the maximum constitute a promising solution, in which one or more pre-
possible Euclidean distance. configured resources are assigned to the contending users. On
the other hand, integrated protocols - including random back-
off schemes - can be considered as a technique of resolving
C. Receiver Design non-orthogonal collisions, whilst reducing the packet drop-
The complexity of an MPA-based receiver may still become ping rates. Additionally, without relying on any access-grant
excessive for massive connectivity in 5G. Therefore, some procedure, the BS cannot obtain any information on the asso-
approximate solutions of the MPA can be used for reduc- ciated user activity, which however can be fortunately detected
ing receiver complexity, such as a Gaussian approximation by CS-aided recovery algorithms due to the sparsity of user
of the interference, which models the interference-plus-noise activity.
as a Gaussian distribution. This approximation becomes more
accurate, when the number of connections becomes high, F. Resource Allocation
as expected in 5G. Additionally, the MPA can be used to In power-domain NOMA, the interference cancellation
jointly detect and channel-decode the received symbols, where capability of receivers is closely related to the accuracy of
the constructed graph consists of variable nodes, observation the power allocation scheme. On the other hand, the accu-
nodes and check nodes corresponding to the check equations racy of allocating the power of each user directly affects the
of the LDPC code. In this way, extrinsic information can be throughput of both power-domain NOMA and of code-domain
2318 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

NOMA. By carefully adjusting the power allocation under better connections to the BS. By exploiting the appealing con-
a specific total power constraint, the BS becomes capable cept of cognitive radio networks we can readily illustrate the
of flexibly controlling the overall throughput, the cell-edge performance gain of NOMA over conventional OMA, which
throughput and the rate-fairness of the users. The optimal significantly simplifies the design of NOMA systems [172].
resource allocation scheme has to search through the entire For example, for the scenarios associated with MIMO schemes
search space of legitimate solutions, and thus the complexity or in the presence of co-channel interference, the design of
may become excessive. Both dynamic programming algo- optimal power allocation is difficult, since it is challenging to
rithms and greedy algorithms may be considered for realizing decide the quality-order of the users’ channel conditions. The
a near-optimal power allocation operating at a low complexity. exploitation of cognitive radio networks may impose new con-
Additionally, in order to support various applications, dynamic straints on the power allocation, which has to strike a similar
power allocation constitutes a promising research topic for throughput and fairness tradeoff as conventional NOMA.
future work.
I. Further Challenges
G. Extension to MIMO Some further challenges should also be resolved in the con-
It is desirable to extend the existing NOMA schemes text of NOMA systems, including the associated reference
to their MIMO-aided counterparts, especially to large-scale signal design and channel estimation, the reduction of the
MIMO systems, in order to further improve the attainable PAPR in multi-carrier NOMA systems, such as LDS-OFDM,
spectral efficiency by exploiting the spatial diversity gain maintaining system scalability, the issues of channel-quality
and/or the multiplexing gain of MIMO systems. However, the feedback design, BS cooperation, etc. Additionally, the exist-
design of MIMO-aided NOMA techniques is by no means ing multiple access design routinely assumes the employment
trivial. Consider the power-domain NOMA as an example. of a single scheme for all applications, regardless of their
Recall that the key idea of power-domain NOMA is to allo- diverse requirements. Accordingly, various system design fac-
cate the transmission power to users inversely proportionally tors have to be considered in order to accommodate the
to their channel conditions. For scenarios associated with worst-case condition, which leads to inefficient multiple access
single-antenna nodes, it is possible to compare the users’ chan- design in many applications. Therefore, the software defined
nel conditions, since channel gains/attenuations are scalars. multiple access technology is expected to support the flexible
However, in MIMO scenarios the channels are represented by configuration of multiple access schemes, and thus differ-
a matrix. Hence it becomes difficult to decide, which user’s ent services as well as applications can be supported in
channel is better. This dilemma leads to implementional diffi- 5G. It is expected that NOMA solutions will achieve further
culties for NOMA-solutions. This is still a promising open area performance improvements by addressing these challenges.
at the time of writing, with very few solutions proposed in the
open literature. A possible solution is to request the BS to form V. C ONCLUSION
multiple beams, where NOMA techniques are invoked for sup- In this article, we have discussed the key concept and
porting the users covered by the same directional beam and advantages of NOMA techniques, which constitute one of
MIMO precoding/detection is used to cancel the inter-beam the promising technologies for future 5G systems. The domi-
interference [172]. Another possible solution is to assign dif- nant NOMA schemes have been introduced together with their
ferent beams to different users individually, where the NOMA comparison in terms of their operating principles, key features,
power allocation constraint has to be taken into consideration receiver complexity, pros and cons, etc. We also highlighted
for the design of beamforming [123]. a range of key challenges, opportunities and future research
trends related to the design of NOMA, including the theoreti-
H. Cognitive Radio Inspired NOMA cal analysis, the design of spreading sequences or codebooks,
the receiver design, the design issues of access-grant-free
The advantage of NOMA techniques can be simply illus- NOMA, resource allocation schemes, extensions to massive
trated by exploiting the concept of cognitive radio networks. MIMO systems and so on. It is expected that NOMA will
Specifically, the user associated with poorer channel conditions play an important role in future 5G wireless communication
in a NOMA system can be viewed as a primary user in the con- systems supporting massive connectivity and low latency.
text of cognitive radio networks. If conventional OMA is used,
the bandwidth resources assigned to this primary user, such as
the time slots or frequency slots are solely occupied by this R EFERENCES
user, and no other users can access these bandwidth resources, [1] A. Osseiran et al., “Scenarios for 5G mobile and wireless communica-
even if the primary user has a poor connection to the BS. The tions: The vision of the METIS project,” IEEE Commun. Mag., vol. 52,
benefit of using NOMA is analogous to that of cognitive radio no. 5, pp. 26–35, May 2014.
[2] F. Boccardi, R. W. Heath, Jr., A. Lozano, T. L. Marzetta, and
networks, where additional secondary users can be admitted P. Popovski, “Five disruptive technology directions for 5G,” IEEE
into the specific beam occupied by the primary user. While Commun. Mag., vol. 52, no. 2, pp. 74–80, Feb. 2014.
these secondary users may impose performance degradations [3] A. W. Scott and R. Frobenius, “Multiple access techniques: FDMA,
TDMA, and CDMA,” in RF Measurements for Cellular Phones
on the primary user, the overall system throughput can be sig- and Wireless Data Systems. Hoboken, NJ, USA: Wiley, Jan. 2008,
nificantly improved, particularly if the secondary users have pp. 413–429.
DAI et al.: SURVEY OF NOMA FOR 5G 2319

[4] H. Li, G. Ru, Y. Kim, and H. Liu, “OFDMA capacity analysis in MIMO [26] C. Yan et al., “Receiver design for downlink non-orthogonal multi-
channels,” IEEE Trans. Inf. Theory, vol. 56, no. 9, pp. 4438–4446, ple access (NOMA),” in Proc. IEEE Veh. Technol. Conf. (IEEE VTC
Sep. 2010. Spring), Glasgow, U.K., May 2015, pp. 1–6.
[5] K. Higuchi and A. Benjebbour, “Non-orthogonal multiple access [27] K. Saito, A. Benjebbour, Y. Kishiyama, Y. Okumura, and T. Nakamura,
(NOMA) with successive interference cancellation for future radio “Performance and design of SIC receiver for downlink NOMA with
access,” IEICE Trans. Commun., vol. E98-B, no. 3, pp. 403–414, open-loop SU-MIMO,” in Proc. IEEE Int. Conf. Commun. Workshop
Mar. 2015. (IEEE ICCW), London, U.K., Jun. 2015, pp. 1161–1165.
[6] S. M. R. Islam, N. Avazov, O. A. Dobre, and K.-S. Kwak, “Power- [28] X. Chen, A. Beiijebbour, A. Li, H. Jiang, and H. Kayama,
domain non-orthogonal multiple access (NOMA) in 5G systems: “Consideration on successive interference canceller (SIC) receiver at
Potentials and challenges,” IEEE Commun. Surveys Tuts., vol. 19, no. 2, cell-edge users for non-orthogonal multiple access (NOMA) with SU-
pp. 721–742, 2nd Quart., 2017. MIMO,” in Proc. IEEE Annu. Int. Symp. Pers. Indoor Mobile Radio
[7] Z. Wei, J. Yuan, D. W. K. Ng, M. Elkashlan, and Z. Ding, “A sur- Commun. (IEEE PIMRC), Aug. 2015, pp. 522–526.
vey of downlink non-orthogonal multiple access for 5G wireless [29] Z. Ding, M. Peng, and H. V. Poor, “Cooperative non-orthogonal mul-
communication networks,” ZTE Commun., vol. 14, no. 4, pp. 17–23, tiple access in 5G systems,” IEEE Commun. Lett., vol. 19, no. 8,
2016. pp. 1462–1465, Aug. 2015.
[8] M. S. Ali, H. Tabassum, and E. Hossain, “Dynamic user clustering [30] Z. Ding, H. Dai, and H. V. Poor, “Relay selection for cooperative
and power allocation for uplink and downlink non-orthogonal multiple NOMA,” IEEE Wireless Commun. Lett., vol. 5, no. 4, pp. 416–419,
access (NOMA) systems,” IEEE Access, vol. 4, pp. 6325–6343, 2016. Aug. 2016.
[9] M. S. Ali, E. Hossain, and D. I. Kim, “Non-orthogonal multiple [31] D. Wan, M. Wen, F. Ji, H. Yu, and F. Chen, “Non-orthogonal multi-
access (NOMA) for downlink multiuser MIMO systems: User clus- ple access for cooperative communications: Challenges, opportunities,
tering, beamforming, and power allocation,” IEEE Access, vol. 5, and trends,” IEEE Wireless Commun., vol. 25, no. 2, pp. 109–117,
pp. 565–577, 2016. Apr. 2018.
[10] S. M. R. Islam, M. Zeng, and O. A. Dobre, “NOMA in 5G systems: [32] S. Han, C.-L. I, Z. Xu, and Q. Sun, “Energy efficiency and spectrum
Exciting possibilities for enhancing spectral efficiency,” IEEE 5G Tech efficiency co-design: From NOMA to network NOMA,” IEEE MMTC
Focus, vol. 1, no. 2, pp. 1–6, Jun. 2017. E-Lett., vol. 9, no. 5, pp. 21–24, Sep. 2014.
[11] Z. Ding et al., “A survey on non-orthogonal multiple access for 5G [33] H. Tabassum, E. Hossain, and M. J. Hossain, “Modeling and analysis
networks: Research challenges and future trends,” IEEE J. Sel. Areas of uplink non-orthogonal multiple access (NOMA) in large-scale cellu-
Commun., vol. 35, no. 10, pp. 2181–2195, Oct. 2017. lar networks using Poisson cluster processes,” IEEE Trans. Commun.,
[12] Y. Cai, Z. Qin, F. Cui, G. Y. Li, and J. A. McCann, “Modulation vol. 65, no. 8, pp. 3555–3570, Aug. 2017.
and multiple access for 5G networks,” IEEE Commun. Surveys Tuts., [34] W. Shin et al., “Non-orthogonal multiple access in multi-cell networks:
vol. 20, no. 1, pp. 629–646, 1st Quart., 2018. Theory, performance, and practical challenges,” IEEE Commun. Mag.,
[13] L. Dai et al., “Non-orthogonal multiple access for 5G: Solutions, vol. 55, no. 10, pp. 176–183, Oct. 2017.
challenges, opportunities, and future research trends,” IEEE Commun. [35] J. Choi, “Non-orthogonal multiple access in downlink coordinated two-
Mag., vol. 53, no. 9, pp. 74–81, Sep. 2015. point systems,” IEEE Commun. Lett., vol. 18, no. 2, pp. 313–316,
Feb. 2014.
[14] A. Benjebbour et al., “Concept and practical considerations of non-
[36] M. S. Ali, E. Hossain, A. Al-Dweik, and D. I. Kim, “Downlink power
orthogonal multiple access (NOMA) for future radio access,” in Proc.
allocation for CoMP-NOMA in multi-cell networks,” IEEE Trans.
IEEE Intell. Signal Process. Commun. Syst. (IEEE ISPACS), Naha,
Commun., to be published.
Japan, Nov. 2013, pp. 770–774.
[37] Z. Ding, Z. Yang, P. Fan, and H. V. Poor, “On the performance of
[15] Y. Saito et al., “Non-orthogonal multiple access (NOMA) for future
non-orthogonal multiple access in 5G systems with randomly deployed
radio access,” in Proc. IEEE Veh. Technol. Conf. (IEEE VTC Spring),
users,” IEEE Signal Process. Lett., vol. 21, no. 12, pp. 1501–1505,
Jun. 2013, pp. 1–5.
Dec. 2014.
[16] K. Higuchi and Y. Kishiyama, “Non-orthogonal access with random [38] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, “System-
beamforming and intra-beam SIC for cellular MIMO downlink,” in level performance evaluation of downlink non-orthogonal multiple
Proc. IEEE Veh. Technol. Conf. (IEEE VTC Fall), Las Vegas, NV, USA, access (NOMA),” in Proc. IEEE Pers. Indoor Mobile Radio Commun.
Sep. 2013, pp. 1–5. (IEEE PIMRC), London, U.K., Sep. 2013, pp. 611–615.
[17] N. Nonaka, Y. Kishiyama, and K. Higuchi, “Non-orthogonal multiple [39] A. Benjebbovu et al., “System-level performance of downlink NOMA
access using intra-beam superposition coding and SIC in base station for future LTE enhancements,” in Proc. IEEE Glob. Commun.
cooperative MIMO cellular downlink,” in Proc. IEEE Veh. Technol. Conf. Workshops (IEEE Globecom Workshops), Atlanta, GA, USA,
Conf. (IEEE VTC Fall), Vancouver, BC, Canada, Sep. 2014, pp. 1–5. Dec. 2013, pp. 66–70.
[18] M. Zeng, A. Yadav, O. A. Dobre, G. I. Tsiropoulos, and H. V. Poor, [40] Y. Saito et al., “System-level evaluation of downlink non-orthogonal
“Capacity comparison between MIMO-NOMA and MIMO-OMA with multiple access (NOMA) for non-full buffer traffic model,” in Proc.
multiple users in a cluster,” IEEE J. Sel. Areas Commun., vol. 35, IEEE Conf. Stand. Commun. Netw. (IEEE CSCN), Tokyo, Japan,
no. 10, pp. 2413–2424, Oct. 2017. Oct. 2015, pp. 94–99.
[19] M. Zeng, A. Yadav, O. A. Dobre, G. I. Tsiropoulos, and H. V. Poor, [41] A. Benjebbour, A. Li, Y. Kishiyama, H. Jiang, and T. Nakamura,
“On the sum rate of MIMO-NOMA and MIMO-OMA systems,” IEEE “System-level performance of downlink NOMA combined with SU-
Wireless Commun. Lett., vol. 6, no. 4, pp. 534–537, Aug. 2017. MIMO for future LTE enhancements,” in Proc. IEEE Glob. Commun.
[20] A. Benjebbour, K. Saito, A. Li, Y. Kishiyama, and T. Nakamura, “Non- Conf. Workshops (IEEE Globecom Workshops), Austin, TX, USA,
orthogonal multiple access (NOMA): Concept, performance evaluation Dec. 2014, pp. 706–710.
and experimental trials,” in Proc. IEEE Int. Conf. Wireless Netw. Mobile [42] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, “System-
Commun. (IEEE WINCOM), Marrakesh, Morocco, Oct. 2015, pp. 1–6. level performance of downlink non-orthogonal multiple access
[21] A. Benjebbour et al., “NOMA: From concept to standardization,” in (NOMA) under various environments,” in Proc. IEEE Veh. Technol.
Proc. IEEE Conf. Stand. Commun. Netw. (IEEE CSCN), Tokyo, Japan, Conf. (IEEE VTC Spring), May 2015, pp. 1–5.
Oct. 2015, pp. 18–23. [43] M. Kimura and K. Higuchi, “System-level throughput of NOMA with
[22] B. Kim et al., “Uplink NOMA with multi-antenna,” in Proc. IEEE SIC in cellular downlink under FTP traffic model,” in Proc. IEEE
Veh. Technol. Conf. (IEEE VTC Spring), Glasgow, U.K., May 2015, Int. Symp. Wireless Commun. Syst. (IEEE ISWCS), Brussels, Belgium,
pp. 1–5. Aug. 2015, pp. 1–5.
[23] Z. Ding, F. Adachi, and H. V. Poor, “The application of MIMO to non- [44] Y. Endo, Y. Kishiyama, and K. Higuchi, “Uplink non-orthogonal access
orthogonal multiple access,” IEEE Trans. Wireless Commun., vol. 15, with MMSE-SIC in the presence of inter-cell interference,” in Proc.
no. 1, pp. 537–552, Jan. 2016. IEEE Int. Symp. Wireless Commun. Syst. (IEEE ISWCS), Paris, France,
[24] Z. Ding, L. Dai, and H. V. Poor, “MIMO-NOMA design for small Aug. 2012, pp. 261–265.
packet transmission in the Internet of Things,” IEEE Access, vol. 4, [45] P. Sedtheetorn and T. Chulajata, “Spectral efficiency evaluation for non-
pp. 1393–1405, 2016. orthogonal multiple access in Rayleigh fading,” in Proc. IEEE Int.
[25] Y. Lan, A. Benjebboiu, X. Chen, A. Li, and H. Jiang, “Considerations Conf. Adv. Commun. Technol. (IEEE ICACT), Jan. 2016, pp. 747–750.
on downlink non-orthogonal multiple access (NOMA) combined with [46] Z. Ding, F. Adachi, and H. V. Poor, “Performance of MIMO-NOMA
closed-loop SU-MIMO,” in Proc. IEEE Signal Process. Commun. Syst. downlink transmissions,” in Proc. IEEE Glob. Commun. Conf. (IEEE
(IEEE ICSPCS), Dec. 2014, pp. 1–5. GLOBECOM), San Diego, CA, USA, Dec. 2015, pp. 1–6.
2320 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

[47] A. Li, A. Benjebbour, and A. Harada, “Performance evaluation of [68] R. Hoshyar, R. Razavi, and M. Al-Imari, “LDS-OFDM an efficient
non-orthogonal multiple access combined with opportunistic beam- multiple access technique,” in Proc. IEEE Veh. Technol. Conf. (IEEE
forming,” in Proc. IEEE Veh. Technol. Conf. (IEEE VTC Spring), Seoul, VTC Spring), Taipei, Taiwan, May 2010, pp. 1–5.
South Korea, May 2014, pp. 1–5. [69] M. Al-Imari, P. Xiao, M. A. Imran, and R. Tafazolli, “Uplink non-
[48] K. Yamamoto, Y. Saito, and K. Higuchi, “System-level throughput of orthogonal multiple access for 5G wireless networks,” in Proc. 11th
non-orthogonal access with SIC in cellular downlink when channel Int. Symp. Wireless Commun. Syst. (IEEE ISWCS), Barcelona, Spain,
estimation error exists,” in Proc. IEEE Veh. Technol. Conf. (IEEE VTC Aug. 2014, pp. 781–785.
Spring), Seoul, South Korea, May 2014, pp. 1–5. [70] M. Al-Imari, M. A. Imran, R. Tafazolli, and D. Chen, “Performance
[49] N. Nonaka, A. Benjebbour, and K. Higuchi, “System-level throughput evaluation of low density spreading multiple access,” in Proc. IEEE
of NOMA using intra-beam superposition coding and SIC in MIMO Wireless Commun. Mobile Comput. Conf. (IEEE IWCMC), Limassol,
downlink when channel estimation error exists,” in Proc. IEEE Int. Cyprus, Aug. 2012, pp. 383–388.
Conf. Commun. Syst. (IEEE ICCS), Nov. 2014, pp. 202–206. [71] M. Al-Imari, M. A. Imran, and R. Tafazolli, “Low density spreading
[50] Z. Yang, Z. Ding, P. Fan, and G. K. Karagiannidis, “On the perfor- for next generation multicarrier cellular systems,” in Proc. IEEE Future
mance of non-orthogonal multiple access systems with partial channel Commun. Netw. (IEEE ICFCN), Baghdad, Iraq, Apr. 2012, pp. 52–57.
information,” IEEE Trans. Commun., vol. 64, no. 2, pp. 654–667, [72] M. Al-Imari, M. A. Imran, R. Tafazolli, and D. Chen, “Subcarrier and
Feb. 2016. power allocation for LDS-OFDM system,” in Proc. IEEE Veh. Technol.
[51] Q. Sun, S. Han, C.-L. I, and Z. Pan, “On the ergodic capacity of Conf. (IEEE VTC Spring), Yokohama, Japan, May 2011, pp. 1–5.
MIMO NOMA systems,” IEEE Wireless Commun. Lett., vol. 4, no. 4, [73] M. Al-Imari and R. Hoshyar, “Reducing the peak to average power
pp. 405–408, Apr. 2015. ratio of LDS-OFDM signals,” in Proc. IEEE Wireless Commun. Syst.
[52] S. Timotheou and I. Krikidis, “Fairness for non-orthogonal multiple (IEEE ISWCS), York, U.K., Sep. 2010, pp. 922–926.
access in 5G systems,” IEEE Signal Process. Lett., vol. 22, no. 10, [74] H. Nikopour and H. Baligh, “Sparse code multiple access,” in Proc.
pp. 1647–1651, Oct. 2015. IEEE 24th Int. Symp. Pers. Indoor Mobile Radio Commun. (IEEE
[53] K. Yakou and K. Higuchi, “Downlink NOMA with SIC using uni- PIMRC), London, U.K., Sep. 2013, pp. 332–336.
fied user grouping for non-orthogonal user multiplexing and decoding [75] Y. Zhou, H. Luo, R. Li, and J. Wang, “A dynamic states reduction mes-
order,” in Proc. IEEE Int. Symp. Intell. Signal Process. Commun. Syst. sage passing algorithm for sparse code multiple access,” in Proc. IEEE
(IEEE ISPACS), Nov. 2015, pp. 508–513. Wireless Telecommun. Symp. (IEEE WTS), London, U.K., Apr. 2016,
[54] J. Umehara, Y. Kishiyama, and K. Higuchi, “Enhancing user fairness pp. 1–5.
in non-orthogonal access with successive interference cancellation for [76] Y. Du, B. Dong, Z. Chen, J. Fang, and X. Wang, “A fast convergence
cellular downlink,” in Proc. IEEE Int. Conf. Commun. Syst. (IEEE multiuser detection scheme for uplink SCMA systems,” IEEE Commun.
ICCS), Singapore, Nov. 2012, pp. 324–328. Lett., vol. 5, no. 4, pp. 388–391, Aug. 2016.
[55] N. Otao, Y. Kishiyama, and K. Higuchi, “Performance of non- [77] H. Mu, Z. Ma, M. Alhaji, P. Fan, and D. Chen, “A fixed low complex-
orthogonal access with SIC in cellular downlink using proportional ity message pass algorithm detector for up-link SCMA system,” IEEE
fair-based resource allocation,” in Proc. IEEE Int. Symp. Wireless Wireless Commun. Lett., vol. 4, no. 6, pp. 585–588, Dec. 2015.
Commun. Syst. (IEEE ISWCS), Paris, France, Aug. 2012, pp. 476–480. [78] Z. Jia, Z. Hui, and L. Xing, “A low-complexity tree search based quasi-
[56] X. Chen, A. Benjebbour, A. Li, and A. Harada, “Multi-user pro- ML receiver for SCMA system,” in Proc. IEEE Int. Conf. Comput.
portional fair scheduling for uplink non-orthogonal multiple access Commun. (IEEE ICCC), Chengdu, China, Oct. 2015, pp. 319–323.
(NOMA),” in Proc. IEEE Veh. Technol. Conf. (IEEE VTC Spring), [79] Y. Liu, J. Zhong, P. Xiao, and M. Zhao, “A novel evidence theory
Seoul, South Korea, May 2014, pp. 1–5. based row message passing algorithm for LDS systems,” in Proc. IEEE
[57] F. Liu, P. Mähönen, and M. Petrova, “Proportional fairness-based user Int. Conf. Wireless Commun. Signal Process. (IEEE WCSP), Nanjing,
pairing and power allocation for non-orthogonal multiple access,” in China, Oct. 2015, pp. 1–5.
Proc. IEEE 26th Annu. Int. Symp. Pers. Indoor Mobile Radio Commun. [80] D. Wei, Y. Han, S. Zhang, and L. Liu, “Weighted message passing
(IEEE PIMRC), Aug. 2015, pp. 1127–1131. algorithm for SCMA,” in Proc. IEEE Int. Conf. Wireless Commun.
[58] Y. Liu, M. Elkashlan, Z. Ding, and G. K. Karagiannidis, “Fairness Signal Process. (IEEE WCSP), Nanjing, China, Oct. 2015, pp. 1–5.
of user clustering in MIMO non-orthogonal multiple access systems,” [81] K. Xiao, B. Xiao, S. Zhang, Z. Chen, and B. Xia, “Simplified mul-
IEEE Commun. Lett., vol. 20, no. 7, pp. 1465–1468, Jul. 2016. tiuser detection for SCMA with sum-product algorithm,” in Proc. IEEE
[59] L. Lei, D. Yuan, C. K. Ho, and S. Sun, “Joint optimization of power and Int. Conf. Wireless Commun. Signal Process. (IEEE WCSP), Nanjing,
channel allocation with non-orthogonal multiple access for 5G cellular China, Oct. 2015, pp. 1–5.
systems,” in Proc. IEEE Glob. Commun. Conf. (IEEE GLOBECOM), [82] Y. Du, B. Dong, Z. Chen, J. Fang, and L. Yang, “Shuffled multiuser
San Diego, CA, USA, Dec. 2015, pp. 1–6. detection schemes for uplink sparse code multiple access systems,”
[60] J. Choi, “On the power allocation for a practical multiuser superposi- IEEE Commun. Lett., vol. 20, no. 6, pp. 1231–1234, Jun. 2016.
tion scheme in NOMA systems,” IEEE Commun. Lett., vol. 20, no. 3, [83] A. Bayesteh, H. Nikopour, M. Taherzadeh, H. Baligh, and J. Ma,
pp. 438–441, Mar. 2016. “Low complexity techniques for SCMA detection,” in Proc. IEEE Glob.
[61] Y. Hayashi, Y. Kishiyama, and K. Higuchi, “Investigations on power Commun. Conf. Workshops (IEEE Globecom Workshops), San Diego,
allocation among beams in non-orthogonal access with random beam- CA, USA, Dec. 2015, pp. 1–6.
forming and intra-beam SIC for cellular MIMO downlink,” in Proc. [84] S. Zhang et al., “Sparse code multiple access: An energy efficient
IEEE Veh. Technol. Conf. (IEEE VTC Fall), Las Vegas, NV, USA, uplink approach for 5G wireless systems,” in Proc. IEEE Glob.
Sep. 2013, pp. 1–5. Commun. Conf. (IEEE Globecom), Austin, TX, USA, Dec. 2014,
[62] K. Higuchi and Y. Kishiyama, “Non-orthogonal access with succes- pp. 4782–4787.
sive interference cancellation for future radio access,” in Proc. IEEE [85] Y. Wu, S. Zhang, and Y. Chen, “Iterative multiuser receiver in sparse
Veh. Technol. Soc. Asia Pac. Wireless Commun. Symp. (IEEE APWCS), code multiple access systems,” in Proc. IEEE Int. Conf. Commun.
Aug. 2012, pp. 1–5. (IEEE ICC), London, U.K., Jun. 2015, pp. 2918–2923.
[63] T. Takeda and K. Higuchi, “Enhanced user fairness using non- [86] B. Xiao et al., “Iterative detection and decoding for SCMA systems
orthogonal access with SIC in cellular uplink,” in Proc. IEEE Veh. with LDPC codes,” in Proc. IEEE Int. Conf. Wireless Commun. Signal
Technol. Conf. (IEEE VTC Fall), San Francisco, CA, USA, Sep. 2011, Process. (IEEE WCSP), Nanjing, China, Oct. 2015, pp. 1–5.
pp. 1–5. [87] M. Taherzadeh, H. Nikopour, A. Bayesteh, and H. Baligh, “SCMA
[64] R. Hoshyar, F. P. Wathan, and R. Tafazolli, “Novel low-density sig- codebook design,” in Proc. IEEE Veh. Technol. Conf. (IEEE VTC Fall),
nature for synchronous CDMA systems over AWGN channel,” IEEE Vancouver, BC, Canada, Sep. 2014, pp. 1–5.
Trans. Signal Process., vol. 56, no. 4, pp. 1616–1626, Apr. 2008. [88] L. Yu, X. Lei, P. Fan, and D. Chen, “An optimized design of SCMA
[65] D. Guo and C.-C. Wang, “Multiuser detection of sparsely spread codebook based on star-QAM signaling constellations,” in Proc. IEEE
CDMA,” IEEE J. Sel. Areas Commun., vol. 26, no. 3, pp. 421–431, Int. Conf. Wireless Commun. Signal Process. (IEEE WCSP), Nanjing,
Apr. 2008. China, Oct. 2015, pp. 1–5.
[66] J. Van De Beek and B. M. Popovic, “Multiple access with low-density [89] S. Zhang, B. Xiao, K. Xiao, Z. Chen, and B. Xia, “Design and analysis
signatures,” in Proc. IEEE Glob. Commun. Conf. (IEEE Globecom), of irregular sparse code multiple access,” in Proc. IEEE Int. Conf.
Honolulu, HI, USA, Nov./Dec. 2009, pp. 1–6. Wireless Commun. Signal Process. (IEEE WCSP), Oct. 2015, pp. 1–5.
[67] R. Razavi, R. Hoshyar, M. A. Imran, and Y. Wang, “Information the- [90] K. Au et al., “Uplink contention based SCMA for 5G radio access,”
oretic analysis of LDS scheme,” IEEE Commun. Lett., vol. 15, no. 8, in Proc. IEEE Glob. Commun. Conf. (Globecom), Austin, TX, USA,
pp. 798–800, Aug. 2011. Dec. 2014, pp. 900–905.
DAI et al.: SURVEY OF NOMA FOR 5G 2321

[91] A. Bayesteh, E. Yi, H. Nikopour, and H. Baligh, “Blind detec- [113] S. Kang, X. Dai, and B. Ren, “Pattern division multiple access for 5G,”
tion of SCMA for uplink grant-free multiple-access,” in Proc. IEEE Telecommun. Netw. Technol., vol. 5, no. 5, pp. 43–47, May 2015.
Wireless Commun. Syst. (IEEE ISWCS), Barcelona, Spain, Aug. 2014, [114] J. Huang, K. Peng, C. Pan, F. Yang, and H. Jin, “Scalable video
pp. 853–857. broadcasting using bit division multiplexing,” IEEE Trans. Broadcast.,
[92] H. Nikopour et al., “SCMA for downlink multiple access of 5G wire- vol. 60, no. 4, pp. 701–706, Dec. 2014.
less networks,” in Proc. IEEE Glob. Commun. Conf. (IEEE Globecom), [115] R. Zhang and L. Hanzo, “A unified treatment of superposition coding
Austin, TX, USA, Dec. 2014, pp. 3940–3945. aided communications: Theory and practice,” IEEE Commun. Surveys
[93] U. Vilaipornsawai, H. Nikopour, A. Bayesteh, and J. Ma, “SCMA for Tuts., vol. 13, no. 3, pp. 503–520, 3rd Quart., 2011.
open-loop joint transmission CoMP,” in Proc. IEEE Veh. Technol. Conf. [116] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
(IEEE VTC Fall), Boston, MA, USA, Sep. 2015, pp. 1–5. Cambridge, U.K.: Cambridge Univ. Press, 2005.
[94] T. Liu, X. Li, and L. Qiu, “Capacity for downlink massive MIMO MU- [117] T. M. Cover and J. A. Thomas, Elements of Information Theory.
SCMA system,” in Proc. IEEE Int. Conf. Wireless Commun. Signal Hoboken, NJ, USA: Wiley, 2006.
Process. (IEEE WCSP), Nanjing, China, Oct. 2015, pp. 1–5. [118] P. P. Bergmans, “A simple converse for broadcast channels with addi-
[95] L. Lu et al., “Prototype for 5G new air interface technology SCMA and tive white Gaussian noise,” IEEE Trans. Inf. Theory, vol. IT-20, no. 2,
performance evaluation,” China Commun., vol. 12, no. 1, pp. 38–48, pp. 279–280, Mar. 1974.
Dec. 2015. [119] Y. Liu, G. Pan, H. Zhang, and M. Song, “On the capacity compari-
[96] Z. Yuan, G. Yu, and W. Li, “Multi-user shared access for 5G,” son between MIMO-NOMA and MIMO-OMA,” IEEE Access, vol. 4,
Telecommun. Netw. Technol., vol. 5, no. 5, pp. 28–30, May 2015. pp. 2123–2129, 2016.
[97] X. Dai et al., “Successive interference cancelation amenable multiple [120] L. Wang, X. Xu, Y. Wu, S. Xing, and Y. Chen, “Sparse code
access (SAMA) for future wireless communications,” in Proc. IEEE multiple access-towards massive connectivity and low latency 5G com-
Int. Conf. Commun. Syst. (IEEE ICCS), Nov. 2014, pp. 222–226. munications,” Telecommun. Netw. Technol., vol. 5, no. 5, pp. 6–15,
[98] M. Y. Alias, S. Chen, and L. Hanzo, “Multiple-antenna-aided OFDM May 2015.
employing genetic-algorithm-assisted minimum bit error rate multiuser [121] Z. Yang, W. Xu, C. Pan, Y. Pan, and M. Chen, “On the optimal-
detection,” IEEE Trans. Veh. Technol., vol. 54, no. 5, pp. 1713–1721, ity of power allocation for NOMA downlinks with individual QoS
Sep. 2005. constraints,” IEEE Commun. Lett., vol. 21, no. 7, pp. 1649–1652,
[99] A. Wolfgang, S. Chen, and L. Hanzo, “Parallel interference cancellation Jul. 2017.
based turbo space-time equalization in the SDMA uplink,” IEEE Trans. [122] C. Chen, W. Cai, X. Cheng, L. Yang, and Y. Jin, “Low complexity
Wireless Commun., vol. 6, no. 2, pp. 609–616, Feb. 2007. beamforming and user selection schemes for 5G MIMO-NOMA sys-
[100] L. Wang, L. Xu, S. Chen, and L. Hanzo, “Three-stage irregular con- tems,” IEEE J. Sel. Areas Commun., vol. 35, no. 12, pp. 2708–2722,
volutional coded iterative center-shifting K -best sphere detection for Dec. 2017.
soft-decision SDMA-OFDM,” IEEE Trans. Veh. Technol., vol. 58, [123] M. F. Hanif, Z. Ding, T. Ratnarajah, and G. K. Karagiannidis,
no. 4, pp. 2103–2109, May 2009. “A minorization-maximization method for optimizing sum rate in non-
[101] S. Chen, L. Hanzo, and A. Livingstone, “MBER space-time decision orthogonal multiple access systems,” IEEE Trans. Signal Process.,
feedback equalization assisted multiuser detection for multiple antenna vol. 64, no. 1, pp. 76–88, Sep. 2015.
aided SDMA systems,” IEEE Trans. Signal Process., vol. 54, no. 8, [124] Q. Sun et al., “Sum rate optimization for MIMO non-orthogonal mul-
pp. 3090–3098, Aug. 2006. tiple access systems,” in Proc. IEEE Wireless Commun. Netw. Conf.
[102] J. Zhang, S. Chen, X. Mu, and L. Hanzo, “Evolutionary algo- (IEEE WCNC), New Orleans, LA, USA, Mar. 2015, pp. 747–752.
rithm assisted joint channel estimation and turbo multi-user detec-
[125] Z. Chen, Z. Ding, X. Dai, and G. K. Karagiannidis, “On the application
tion/decoding for OFDM/SDMA,” IEEE Trans. Veh. Technol., vol. 63,
of quasi-degradation to MISO-NOMA downlink,” IEEE Trans. Signal
no. 3, pp. 1204–1222, Mar. 2014.
Process., vol. 64, no. 23, pp. 6174–6189, Dec. 2016.
[103] S. Chen, A. Wolfgang, C. J. Harris, and L. Hanzo, “Symmetric RBF
[126] L. Li, L. Wang, and L. Hanzo, “Differential interference suppres-
classifier for nonlinear detection in multiple-antenna-aided systems,”
sion aided three-stage concatenated successive relaying,” IEEE Trans.
IEEE Trans. Neural Netw., vol. 19, no. 5, pp. 737–745, May 2008.
Commun., vol. 60, no. 8, pp. 2146–2155, Aug. 2012.
[104] A. Wolfgang, J. Akhtman, S. Chen, and L. Hanzo, “Reduced-
complexity near-maximum-likelihood detection for decision feedback [127] L. Zhang et al., “Layered-division-multiplexing: Theory and practice,”
assisted space-time equalization,” IEEE Trans. Wireless Commun., IEEE Trans. Broadcast., vol. 62, no. 1, pp. 216–232, Mar. 2016.
vol. 6, no. 7, pp. 2407–2411, Jul. 2007. [128] F. R. Kschischang, B. J. Frey, and H.-A. Loeliger, “Factor graphs and
[105] J. Akhtman, A. Wolfgang, S. Chen, and L. Hanzo, “An optimized- the sum-product algorithm,” IEEE Trans. Inf. Theory, vol. 47, no. 2,
hierarchy-aided approximate log-MAP detector for MIMO systems,” pp. 498–519, Feb. 2001.
IEEE Trans. Wireless Commun., vol. 6, no. 5, pp. 1900–1909, [129] Z. Ding, P. Fan, and H. V. Poor, “Random beamforming in millimeter-
May 2007. wave NOMA networks,” IEEE Access, vol. 5, pp. 7667–7681,
[106] S. Chen, A. Livingstone, H. Q. Du, and L. Hanzo, “Adaptive mini- 2017.
mum symbol error rate beamforming assisted detection for quadrature [130] B. Wang, L. Dai, Z. Wang, N. Ge, and S. Zhou, “Spectrum and energy
amplitude modulation,” IEEE Trans. Wireless Commun., vol. 7, no. 4, efficient beamspace MIMO-NOMA for millimeter-wave communica-
pp. 1140–1145, Apr. 2008. tions using lens antenna array,” IEEE J. Sel. Areas Commun., vol. 35,
[107] J. Zhang, S. Chen, X. Mu, and L. Hanzo, “Turbo multi-user detec- no. 10, pp. 2370–2382, Oct. 2017.
tion for OFDM/SDMA systems relying on differential evolution aided [131] A. S. Marcano and H. L. Christiansen, “Performance of non-orthogonal
iterative channel estimation,” IEEE Trans. Commun., vol. 60, no. 6, multiple access (NOMA) in mmWave wireless communications for 5G
pp. 1621–1633, Jun. 2012. networks,” in Proc. IEEE Int. Conf. Comput. Netw. Commun. (IEEE
[108] J. Zhang, S. Chen, X. Mu, and L. Hanzo, “Joint channel estimation ICNC), Jan. 2017, pp. 969–974.
and multi-user detection for SDMA/OFDM based on dual repeated [132] Low Code Rate and Signature Based Multiple Access Scheme for
weighted boosting search,” IEEE Trans. Veh. Technol., vol. 60, no. 7, New Radio, document TSG RAN1 #85, 3GPP, Nanjing, China,
pp. 3265–3275, Sep. 2011. May 2016.
[109] C.-Y. Wei, J. Akhtman, S.-X. Ng, and L. Hanzo, “Iterative near- [133] Discussion on Multiple Access for New Radio Interface, document TSG
maximum-likelihood detection in rank-deficient downlink SDMA sys- RAN WG1 #84bis, 3GPP, Busan, Korea, Apr. 2016.
tems,” IEEE Trans. Veh. Technol., vol. 57, no. 1, pp. 653–657, [134] Initial Views and Evaluation Results on Non-Orthogonal Multiple
Jan. 2008. Access for NR Uplink, document TSG RAN WG1 #84bis, 3GPP, Busan,
[110] A. Wolfgang, J. Akhtman, S. Chen, and L. Hanzo, “Iterative MIMO Korea, Apr. 2016.
detection for rank-deficient systems,” IEEE Signal Process. Lett., [135] Candidate NR Multiple Access Schemes, document TSG RAN WG1
vol. 13, no. 11, pp. 699–702, Nov. 2006. #84b, 3GPP, Busan, Korea, Apr. 2016.
[111] L. Xu, S. Chen, and L. Hanzo, “EXIT chart analysis aided turbo [136] Non-Orthogonal Multiple Access Candidate for NR, document
MUD designs for the rank-deficient multiple antenna assisted OFDM TSG RAN WG1 #85, 3GPP, Nanjing, China, May 2016.
uplink,” IEEE Trans. Wireless Commun., vol. 7, no. 6, pp. 2039–2044, [137] K. Kusume, G. Bauch, and W. Utschick, “IDMA vs. CDMA: Analysis
Jun. 2008. and comparison of two multiple access schemes,” IEEE Trans. Wireless
[112] C.-L. I, S. Han, Z. Xu, Q. Sun, and Z. Pan, “5G: Rethink mobile Commun., vol. 11, no. 1, pp. 78–87, Jan. 2012.
communications for 2020+,” Philosoph. Trans. Roy. Soc. A Math. Phys. [138] Considerations on DL/UL Multiple Access for NR, document
Eng. Sci., vol. 374, no. 2062, 2016, Art. no. 20140432. TSG RAN WG1 #84bis, 3GPP, Busan, Korea, Apr. 2016.
2322 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 20, NO. 3, THIRD QUARTER 2018

[139] A. Medra and T. N. Davidson, “Flexible codebook design for lim- [163] B. Wang, K. Wang, Z. Lu, T. Xie, and J. Quan, “Comparison study of
ited feedback systems via sequential smooth optimization on the non-orthogonal multiple access schemes for 5G,” in Proc. IEEE Int.
Grassmannian manifold,” IEEE Trans. Signal Process., vol. 62, no. 5, Symp. Broadband Multimedia Syst. Broadcast. (IEEE BMSB), Ghent,
pp. 1305–1318, Mar. 2014. Belgium, Jun. 2015, pp. 1–5.
[140] Non-Orthogonal Multiple Access for New Radio, document TSG RAN [164] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, “System-
WG1 #85, 3GPP, Nanjing, China, May 2016. level performance evaluation of downlink non-orthogonal multiple
[141] New Uplink Non-Orthogonal Multiple Access Schemes for NR, docu- access (NOMA),” in Proc. IEEE Annu. Symp. (IEEE PIMRC), London,
ment TSG RAN WG1 #86, 3GPP, Gothenburg, Sweden, Aug. 2016. U.K., Sep. 2013, pp. 611–615.
[142] Categorization and Analysis of MA Schemes, document TSG RAN [165] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, “System-
WG1 Meeting #86bis, 3GPP, Lisbon, Portugal, Oct. 2016. level performance evaluation of downlink non-orthogonal multiple
[143] B. Wang, L. Dai, Y. Yuan, and Z. Wang, “Compressive sensing access (NOMA) under various environments,” in Proc. IEEE Veh.
based multi-user detection for uplink grant-free non-orthogonal mul- Technol. Conf. (IEEE VTC-Spring), May 2015, pp. 1–5.
tiple access,” in Proc. IEEE Veh. Technol. Conf. (IEEE VTC Fall), [166] K. Saito, A. Benjebbour, A. Harada, Y. Kishiyama, and T. Nakamura,
Boston, MA, USA, Sep. 2015, pp. 1–5. “Link-level performance evaluation of downlink NOMA with SIC
[144] B. Wang, L. Dai, T. Mir, and Z. Wang, “Joint user activity and data receiver considering error vector magnitude,” in Proc. IEEE Veh.
detection based on structured compressive sensing for NOMA,” IEEE Technol. Conf. (IEEE VTC-Spring), Glasgow, U.K., May 2015, pp. 1–5.
Commun. Lett., vol. 20, no. 7, pp. 1473–1476, Jul. 2016. [167] K. Saito, A. Benjebbour, Y. Kishiyama, Y. Okumura, and T. Nakamura,
“Performance and design of SIC receiver for downlink NOMA with
[145] V. Shah-Mansouri, S. Duan, L.-H. Chang, V. W. S. Wong, and J.-Y. Wu,
open-Loop SU-MIMO,” in Proc. IEEE Int. Conf. Commun. Workshop
“Compressive sensing based asynchronous random access for wire-
(IEEE ICCW), London, U.K., Jun. 2015, pp. 1161–1165.
less networks,” in Proc. IEEE Wireless Commun. Netw. Conf. (IEEE
[168] A. Benjebbour, K. Saito, Y. Saito, and Y. Kishiyama, “5G radio access
WCNC), Apr. 2013, pp. 884–888.
technology,” NTT DOCOMO Tech. J., vol. 17, no. 4, pp. 16–28, 2015.
[146] G. Wunder, P. Jung, and C. Wang, “Compressive random access for [169] L. Lu et al., “Prototype for 5G new air interface technology SCMA
post-LTE systems,” in Proc. IEEE Int. Conf. Commun. Workshops and performance evaluation,” China Commun., vol. 12, pp. 38–48,
(IEEE ICC), Sydney, NSW, Australia, Jun. 2014, pp. 539–544. Dec. 2015.
[147] J.-P. Hong, W. Choi, and B. D. Rao, “Sparsity controlled random mul- [170] Y. Tan, J. Zhou, and J. Qin, “Novel channel estimation for non-
tiple access with compressed sensing,” IEEE Trans. Wireless Commun., orthogonal multiple access systems,” IEEE Signal Process. Lett.,
vol. 14, no. 2, pp. 998–1010, Feb. 2015. vol. 23, no. 12, pp. 1781–1785, Dec. 2016.
[148] Components of a New Air Interface—Building Blocks and Performance, [171] K. Struminsky, S. Kruglik, D. Vetrov, and I. Oseledets, “A new
document EU FP7 INFSO-ICT-317669-METIS/D2.3, Mobile Wireless approach for sparse Bayesian channel estimation in SCMA uplink sys-
Commun. Enablers Twenty-Twenty Inf. Soc., Mar. 2014. tems,” in Proc. IEEE Int. Conf. Wireless Commun. Signal Process.
[149] Initial LLS Results for UL Non-Orthogonal Multiple Access, document (IEEE WCSP), Yangzhou, China, Oct. 2016, pp. 1–5.
TSG RAN WG1 #85, 3GPP, Nanjing, China, May 2016. [172] Z. Ding, F. Adachi, and H. V. Poor, “The application of MIMO to non-
[150] Multiple Access Schemes for New Radio Interface, document TSG orthogonal multiple access,” IEEE Trans. Wireless Commun., vol. 15,
RAN WG1 #84bis, 3GPP, Busan, South Korea, Apr. 2016. no. 1, pp. 537–552, Jan. 2016.
[151] M. A. Sedaghat and R. R. Müller. (Jul. 2017). On User Pairing in
NOMA Uplink. [Online]. Available: https://arxiv.org/abs/1707.01846
[152] A. E. Roth and M. Sotomayor, Two Sided Matching: A Study in Game-
Theoretic Modeling and Analysis, 1st ed. Cambridge, U.K.: Cambridge
Univ. Press, 1989.
[153] Y. Gu, W. Saad, M. Bennis, M. Debbah, and Z. Han, “Matching theory
for future wireless networks: Fundamentals and applications,” IEEE
Commun. Mag., vol. 53, no. 5, pp. 52–59, May 2015.
[154] W. Liang, Z. Ding, Y. Li, and L. SongH, “User pairing for downlink
non-orthogonal multiple access networks using matching algorithm,”
IEEE Trans. Commun., vol. 65, no. 12, pp. 5319–5332, Dec. 2017.
[155] J. Zhao et al., “Spectrum allocation and power control for non- Linglong Dai (M’11–SM’14) received the B.S.
orthogonal multiple access in HetNets,” IEEE Trans. Wireless degree from Zhejiang University in 2003, the M.S.
Commun., vol. 16, no. 9, pp. 5825–5837, Sep. 2017. degree (Highest Hons.) from the China Academy
of Telecommunications Technology in 2006, and
[156] F. Fang, H. Zhang, J. Cheng, and V. C. M. Leung, “Energy-
the Ph.D. degree (Highest Hons.) from Tsinghua
efficient resource allocation for downlink non-orthogonal multiple
University, Beijing, China, in 2011. From 2011 to
access (NOMA) network,” IEEE Trans. Commun., vol. 64, no. 9,
2013, he was a Post-Doctoral Research Fellow with
pp. 3722–3732, Sep. 2016.
the Department of Electronic Engineering, Tsinghua
[157] B. Di, S. Bayat, L. Song, and Y. Li, “Radio resource alloca- University, where he was an Assistant Professor
tion for downlink non-orthogonal multiple access (NOMA) networks from 2013 to 2016 and has been an Associate
using matching theory,” in Proc. IEEE Glob. Commun. Conf. (IEEE Professor since 2016. He has co-authored the book
Globecom), San Diego, CA, USA, Dec. 2015, pp. 1–6. entitled mmWave Massive MIMO: A Paradigm for 5G (Academic Press,
[158] H. Tabassum, M. S. Ali, E. Hossain, M. J. Hossain, and D. I. Kim. Elsevier, 2016). He has published over 60 IEEE journal papers and over
(Aug. 2016). Non-Orthogonal Multiple Access (NOMA) in Cellular 40 IEEE conference papers. He also holds 15 granted patents. His current
Uplink and Downlink: Challenges and Enabling Techniques. [Online]. research interests include massive MIMO, millimeter-wave communications,
Available: https://arxiv.org/abs/1608.05783 NOMA, sparse signal processing, and machine learning. He was a recipient of
[159] F. Fang, H. Zhang, J. Cheng, S. Roy, and V. C. M. Leung, “Joint five conference Best Paper Awards at the IEEE ICC 2013, the IEEE ICC 2014,
user scheduling and power allocation optimization for energy-efficient the IEEE ICC 2017, the IEEE VTC 2017-Fall and the IEEE ICC 2018. He also
NOMA systems with imperfect CSI,” IEEE J. Sel. Areas Commun., received the Tsinghua University Outstanding Ph.D. Graduate Award in 2011,
vol. 35, no. 12, pp. 2874–2885, Dec. 2017. the Beijing Excellent Doctoral Dissertation Award in 2012, the China National
[160] Q. Sun, S. Han, C.-L. I, and Z. Pan, “Energy efficiency optimiza- Excellent Doctoral Dissertation Nomination Award in 2013, the URSI Young
tion for fading MIMO non-orthogonal multiple access systems,” in Scientist Award in 2014, the IEEE T RANSACTIONS ON B ROADCASTING
Proc. IEEE Int. Conf. Commun. (IEEE ICC), London, U.K., Jun. 2015, Best Paper Award in 2015, the Second Prize of Science and Technology
pp. 2668–2673. Award of China Institute of Communications in 2016, the Electronics
[161] Y. Zhang et al., “Energy efficiency optimization in cognitive radio Letters Best Paper Award in 2016, the IEEE Communications Letters
inspired non-orthogonal multiple access,” in Proc. IEEE Annu. Int. Exemplary Editor Award in 2017, the National Natural Science Foundation
Symp. Pers. Indoor Mobile Radio Commun. (IEEE PIMRC), Valencia, of China for Outstanding Young Scholars in 2017, and the IEEE ComSoc
Spain, Sep. 2016, pp. 1–6. Asia–Pacific Outstanding Young Researcher Award in 2017. He currently
[162] Y. Xu et al., “Joint beamforming and power-splitting control in serves as an Editor of the IEEE T RANSACTIONS ON C OMMUNICATIONS,
downlink cooperative SWIPT NOMA systems,” IEEE Trans. Signal the IEEE T RANSACTIONS ON V EHICULAR T ECHNOLOGY, and the IEEE
Process., vol. 65, no. 18, pp. 4874–4886, Sep. 2017. C OMMUNICATIONS L ETTERS.
DAI et al.: SURVEY OF NOMA FOR 5G 2323

Bichai Wang (S’15) received the B.S. degree in Sheng Chen (M’90–SM’97–F’08) received the
electronic engineering from Tsinghua University, B.Eng. degree from East China Petroleum Institute,
Beijing, China, in 2015, where she is currently Dongying, China, in 1982, and the Ph.D. degree
pursuing the Ph.D. degree with the Department from City University, London, in 1986, both in
of Electronic Engineering. Her research interests control engineering, the Higher Doctoral Degree
are in wireless communications, with the empha- and the Doctor of Sciences (D.Sc.) degree from
sis on nonorthogonal multiple access, mmWave the University of Southampton, Southampton, U.K.,
massive MIMO, and deep learning-based wire- in 2005. From 1986 to 1999, he held research
less communications. She was a recipient of the and academic appointments with the University of
Freshman Scholarship of Tsinghua University in Sheffield, U.K., University of Edinburgh, U.K., and
2011, the Academic Merit Scholarships of Tsinghua Portsmouth, U.K. Since 1999, he has been with the
University in 2012, 2013, and 2014, respectively, the Excellent Thesis School of Electronics and Computer Science, University of Southampton,
Award of Tsinghua University in 2015, the National Scholarship in 2016, where he holds the post of Professor in Intelligent Systems and Signal
the IEEE VTC’17 Fall Best Student Paper Award in 2017, and the IEEE Processing. His research interests include adaptive signal processing, wireless
T RANSACTIONS ON C OMMUNICATIONS Exemplary Reviewer Award in communications, modeling and identification of nonlinear systems, neural net-
2017. work and machine learning, intelligent control system design, and evolutionary
computation methods and optimization. He has published over 600 research
papers. He has 11 400 Web of Science citations and over 24 000 Google
Scholar citations. He is a fellow of the United Kingdom Royal Academy of
Engineering and IET, a Distinguished Adjunct Professor with King Abdulaziz
Zhiguo Ding (S’03–M’05) received the B.Eng. University, Jeddah, Saudi Arabia, and an ISI Highly Cited Researcher in
degree in electrical engineering from the Beijing engineering in 2004.
University of Posts and Telecommunications in
2000. From 2005 to 2018, he was with Queen’s
University Belfast, Imperial College, Newcastle
University, and Lancaster University. Since 2018,
he has been with the University of Manchester
as a Professor of communications. From 2012 to
2018, he was an Academic Visitor with Princeton
University. His research interests are 5G networks,
game theory, cooperative and energy harvesting net-
works, and statistical signal processing. He was a recipient of the Best
Paper Award in IET ICWMC-2009 and IEEE WCSP-2014, the EU Marie
Curie Fellowship from 2012 to 2014, the Top IEEE TVT Editor 2017, and
the IEEE Heinrich Hertz Award 2018. He is serving as an Editor for the
IEEE T RANSACTIONS ON C OMMUNICATIONS, the IEEE T RANSACTIONS
ON V EHICULAR T ECHNOLOGY , and the Journal of Wireless Communications
and Mobile Computing. He was an Editor for the IEEE W IRELESS
C OMMUNICATION L ETTERS and IEEE C OMMUNICATION L ETTERS from
2013 to 2016.

Zhaocheng Wang (M’09–SM’11) received the B.S.,

M.S., and Ph.D. degrees from Tsinghua University Lajos Hanzo (M’91–SM’92–F’04) received the
in 1991, 1993, and 1996, respectively. From 1996 to D.Sc. degree in electronics in 1976 and the
1997, he was a Post-Doctoral Fellow with Nanyang Doctorate degree in 1983, the Honorary Doctorate
Technological University, Singapore. From 1997 to degrees from the Technical University of Budapest
1999, he was a Research Engineer/Senior Engineer and the University of Edinburgh in 2009 and
with OKI Techno Centre Pte. Ltd., Singapore. From 2015, respectively. In 2016, he was admitted to
1999 to 2009, he was a Senior Engineer/Principal the Hungarian Academy of Science. During his
Engineer with Sony Deutschland GmbH, Germany. 40-year career in telecommunications he has held
Since 2009, he has been a Professor with the various research and academic posts in Hungary,
Department of Electronic Engineering, Tsinghua Germany, and the U.K. Since 1986, he has been with
University, where he is currently the Director of Broadband Communication the School of Electronics and Computer Science,
Key Laboratory, Beijing National Research Center for Information Science University of Southampton, U.K., where he holds the Chair in telecommuni-
and Technology. His research interests include millimeter wave communica- cations. He has successfully supervised 111 Ph.D. students, and co-authored
tions, optical wireless communications, and digital broadcasting. He holds 18 Wiley/IEEE Press books on mobile radio communications totalling in
34 U.S./EU granted patents (23 of them as the first inventor) and pub- excess of 10 000 pages, published over 1700 research contributions at IEEE
lished over 135 peer-reviewed international journal papers. He has authored or Xplore, acted both as a TPC and a General Chair of IEEE conferences, pre-
co-authored two books, which have been selected by IEEE Series on Digital sented keynote lectures and has been awarded a number of distinctions. He is
and Mobile Communication and published by Wiley-IEEE Press. He was a currently directing a 60-strong academic research team, working on a range
recipient of the ICC2013 Best Paper Award, the OECC2015 Best Student of research projects in the field of wireless multimedia communications spon-
Paper Award, the 2016 IEEE Scott Helt Memorial Award, the 2016 National sored by industry, the Engineering and Physical Sciences Research Council,
Award for Science and Technology Progress (First Prize), and the ICC2017 U.K., the European Research Council’s Advanced Fellow Grant and the Royal
Best Paper Award. He was an Associate Editor of the IEEE T RANSACTIONS Society’s Wolfson Research Merit Award. He is an enthusiastic supporter of
ON W IRELESS C OMMUNICATIONS from 2011 to 2015 and an Associate industrial and academic liaison and he offers a range of industrial courses.
Editor of the IEEE C OMMUNICATIONS L ETTERS from 2013 to 2016. He was He is also the Governor of the IEEE ComSoc and VTS. From 2008 to 2012,
also technical program co-chairs of many international conferences, includ- he was the Editor-in-Chief of the IEEE Press and a Chaired Professor with
ing ICC and GlobeSIP. He is a fellow of the Institution of Engineering and Tsinghua University, Beijing. He is fellow of Royal Academy of Engineering,
Technology. IET, and EURASIP.