Non-Orthogonal Multiple Access: Common Myths and Critical Questions

Non-Orthogonal Multiple Access: Common Myths and Critical Questions

Mojtaba Vaezi, Senior Member, IEEE, Robert Schober, Fellow, IEEE,
Zhiguo Ding, Senior Member, IEEE, and H. Vincent Poor Fellow, IEEE
M. Vaezi is with the Department of Electrical and Computer Engineering, Villanova University, PA, USA (e-mail: mvaezi@villanova.edu).
R. Schober is with Friedrich-Alexander University of Erlangen-Nuremberg, Germany, (e-mail: robert.schober@fau.de).
Z. Ding is with the School of Electrical and Electronic Engineering, the University of Manchester, UK (e-mail: zhiguo.ding@manchester.ac.uk).
H. V. Poor is with the Department of Electrical Engineering, Princeton University, Princeton, NJ, USA (e-mail: poor@princeton.edu).
Abstract

Non-orthogonal multiple access (NOMA) has received notable attention for the design of radio access techniques for fifth generation (5G) wireless networks and beyond. The basic concept behind NOMA is to serve more than one user in the same resource block, e.g., a time slot, subcarrier, spreading code, or space. With this, NOMA promotes massive connectivity, lowers latency, improves user fairness and spectral efficiency, and increases reliability compared to orthogonal multiple access (OMA) techniques. While NOMA has gained significant attention from the communications community, it has also been subject to several widespread misunderstandings, such as

NOMA is based on allocating higher power to users with worse channel conditions. As such, cell-edge users receive more power in NOMA and due to this biased power allocation toward cell-edge users inter-cell interference is more severe in NOMA compared to OMA. NOMA also compromises security for spectral efficiency.
The above statements are actually false, and this paper aims at identifying such common myths about NOMA and clarifying why they are not true. We also pose critical questions that are important for the effective adoption of NOMA in 5G and beyond and identify promising research directions for NOMA, which will require intense investigation in the future.

I Introduction And Background

Multiple access techniques allow multiple users to share the same communication resource, and lie at the heart of cellular communication systems [1]. Previous generations of cellular networks have adopted one or more of the following multiple access methods:

  • Frequency division multiple access (FDMA)

  • Time division multiple access (TDMA)

  • Code division multiple access (CDMA)

  • Orthogonal frequency division multiple access (OFDMA)

  • Space division multiple access (SDMA)

Despite their very different approach to sharing the wireless resources, the above schemes have been designed with one common theme in mind: to generate orthogonal signals for different users at the receiver side. In particular, in OFDMA, which has been adopted in the fourth generation (4G) of cellular systems, users’ signals are orthogonal in the frequency and/or time domains. One resource block (RB), which occupies 180 kHz in the 4G long-term evolution (LTE) standard, cannot be allocated to more than one user. Orthogonality of the physical (PHY) layer is the underlying design principle of today’s standards.

The insistence on orthogonality poses significant challenges to 5G systems in which a massive number of devices111According to the ITU 5G performance requirements for IMT-2020, the minimum connection density is 1,000,000 devices per km, which is 100 times more than that in 4G [2]. with diverse data rate and latency requirements are to be connected in each cell. A large percentage of these connections are from devices which may only sporadically require transmission of very low-rate data. Allocating an entire RB to each of these connections is neither efficient nor feasible. The former is because such low-rate devices do not fully utilize the RB, and the latter is because the number and density of such devices is excessively high in 5G networks. In fact, one RB may be used to carry the data of many such low-rate devices.

\setkeys

Ginwidth=.7\OVP@calc

Fig. 1: Illustration of downlink NOMA via power domain multiplexing for two users (or user equipments (UEs), equivalently) with messages and . Let and be the channel gains for user 1 (UE1) and user 2 (UE2), respectively. In this figure, and throughout this paper, without loss of generality, it is assumed that

As a potential multiple access technique for 5G and beyond networks, non-orthogonal multiple access (NOMA) has been proposed to address the above issue. The underlying concept is to serve more than one user in the same wireless resource, be it a time-slot in TDMA, a frequency band in FDMA (or a subcarrier in OFDMA), a spreading code in CDMA, or space in SDMA. NOMA can be realized in different ways, e.g., via the power, code, and other domains [1], but in this paper we focus on power-domain NOMA in the downlink [3].

Apart from the ability to serve multiple devices in one RB, which is particularly beneficial for addressing the increasing demand for massive machine type communication (mMTC), there are several other good reasons for using NOMA in 5G and beyond. NOMA can improve spectral efficiency and user fairness. Grant-free NOMA in the uplink can reduce latency, signaling overhead, and terminal power consumption, particularly for light traffic. The combination of NOMA with other emerging technologies, such as massive multiple-input multiple-output (MIMO) and millimeter wave communication, can effectively address the requirements for enhanced mobile broadband and mMTC.

Owing to the above potentials, NOMA has received significant attention in academia, industry, and standardization bodies during the past a few years. Nonetheless, there are several widespread myths and misunderstandings surrounding the basic NOMA concepts. In this overview paper, we first illustrate the basic NOMA principle with two users, and then consider its extension to a more general setting with an arbitrary number of users in a multi-cell setting. We use this theoretical basis to support our claims in the remainder of the paper where we present and discuss several myths and misunderstandings about NOMA in terms of resource allocation, interference management, etc. Finally, we also pose questions that are critical for the successful adoption of NOMA in practice and discuss potential research challenges.

Ii Downlink NOMA Basics: A Review

Downlink cellular communication is modeled by the broadcast channel (BC). The capacity region of the Gaussian BC is obtained via superposition coding at the base station (BS), as illustrated in Fig. 1, in which the codewords of two users are added up and one signal is transmitted to both users. Such a transmission is non-orthogonal since both users’ signals are transmitted at the same time and frequency. An alternative approach would be to divide time or frequency into two different slots and let each user transmit its signal in one of those orthogonal slots without interfering with the other user. The resulting scheme is TDMA or FDMA and is referred to as OMA in this paper. The achievable rates for NOMA (BC) and OMA are compared for the two-user case in Fig. 2. To gain more insight, we describe how the above regions are obtained.

Ii-a Oma

Ii-A1 Two-user single-cell network

For OMA, assuming a TDMA scheme where a fraction of time () is dedicated to user 1 and a fraction of time is dedicated to user 2, the users can achieve rates and , respectively, where , and are the received signal-to-noise ratio (SNR) and the channel gain for user , , is the BS transmit power, and the noise power is normalized to unity.

Ii-A2 -user single-cell network

The solution is very similar to the two-user case except that the available resource (time or frequency) is divided into orthogonal resources and each user gets , , .

Ii-A3 -user multi-cell network

With different frequencies in the adjacent cells, the solution in each cell is similar to the -user single-cell network.

Ii-B Noma

Ii-B1 Two-user single-cell network

NOMA enlarges OMA’s rate region by using superposition coding (SC) at the transmitter (BS) and successive interference cancellation (SIC) at the receiver. In particular, the BS allocates fractions , , and of its power to the signals of user 1 and user 2, respectively. For decoding, the user with the stronger channel uses SIC to cancel interference and decode its signal free of interference at a rate of , whereas the user with the weaker channel treats the other user’s signal as noise and decodes its own signal at a rate of . By varying from to , any rate pair () on the boundary of the capacity region of the BC (NOMA) can be achieved. For each () on the boundary of the capacity region there is one and only one such that and are the optimal powers for user 1 and user 2, respectively. Conversely, every results in a rate pair on the boundary of the capacity region.

The above discussion implies that NOMA can improve user fairness by efficient and flexible resource allocation. While in OMA a user may not be served for a long time due to the limited number of RBs, such a constraint does not apply to NOMA since, theoretically, NOMA can serve as many users as required in just one RB. In practice, only a few users may be served in one RB for complexity reasons. Nevertheless, NOMA increases the chance that a user is scheduled which, in turn, improves user fairness. Additionally, since the BS can flexibly change the fraction of power allocated to each NOMA user, it can smoothly cope with user fairness issues by increasing the power of the weaker user (the user with smaller channel gain) in order to increase its rate. Increasing the rate of such a user can be realized by maximizing the weighted sum-rate where a weight is given to the weaker user. It is straightforward to prove that in the above maximization problem there exists an optimal power allocation corresponding to every , and vice versa. Then, if user 2 is weaker than user 1, setting results in an that improves the user fairness while will make the matter worse.

Ii-B2 -user single-cell network

Similar to the two-user BC above, the most efficient way to transmit users’ data over a given subband is to use SC at the BS and SIC decoding at the users. The order of SIC is obviously critical for optimal decoding (see [1, Chapter 5] for details).

Ii-B3 Multi-cell network

If different frequencies are allocated at the border of adjacent cells, the problem in each cell reduces to -user single-cell NOMA. However, if universal frequency reuse is used, the problem becomes much more involved, and capacity-achieving schemes are not known. The best achievable strategy for this multi-cell network (a.k.a. the interference channel) decodes part of the inter-cell interference (ICI) while treating the remaining part as noise, and is based on a combination of NOMA and OMA [1, Chapter 5].

Fig. 2: Achievable regions for the two-user OMA and NOMA (downlink) with and Points , and on the NOMA rate region are obtained for the specific values of as shown in the figure (see Table I, too). The powers allocated to user 1 (the stronger user) and user 2 (the weaker user) are obtained as and , respectively, as illustrated in Fig. 1.

Ii-C Mimo-Noma

By creating spatial dimensions, multi-antenna systems open the door to SDMA where multiple users can communicate at the same time and frequency but in different beams (spaces). MIMO-NOMA overloads SDMA by allocating a group (cluster) of users to each beam and using SC-SIC within each group. The interference between the clusters is managed by allocating a different beam to each of the clusters.222This is not however the theoretically optimal solution for the MIMO-BC. Interested readers may refer to [1, Chapter 5] for further information. MIMO-NOMA differs from multi-user MIMO in that a cluster of users, rather than just one user, share one spatial dimension. Hence, it can serve a higher number of users, and paves the way for massive connectivity.

Iii Myths and Misunderstandings About NOMA

Although NOMA is grounded in a well-established theory, its literature has been subject to several widespread myths and misunderstandings. In this section, we inspect several such beliefs and explain why they are incorrect.

TABLE I: Achievable rates () corresponding to the points marked on Fig. 2.
Point UEs served
A Only UE2 0 0 0.79 0.79
B UE1 and UE2 0.025 1.29 0.76 2.05
C UE1 and UE2 0.5 3.33 0.29 3.62
D UE1 and UE2 0.8 3.67 0.10 3.77
E Only UE1 1 3.83 0 3.83

Iii-a Myth 1: NOMA Always Allocates More Power to Users with Poor Channels

There is a common misunderstanding that NOMA always allocates more power to users with poor channels. In the case of two-user NOMA, this implies we should always allocate more power to the user with the weaker channel. Along the same line, in the case of NOMA with three or more users, many papers assume that power allocation should be in reverse order of the users’ channel gains; i.e., the amount of power allocated to a user with a stronger channel is less than that of a user with a weaker channel. But, should we always do so? In other words, does the user with the higher channel gain always get less power in NOMA? A myriad of papers assume this is always the case while the answer to these questions is “no” in general.

As described in Section II, power allocation depends on what point (rate pair) in the capacity region is being targeted, and depending on that specific point the amount of power allocated to the user with the weaker channel can be higher than, equal to, or less than that of the other user. That is, alone does not imply ; i.e., we should not necessarily allocate less power to the user with a stronger channel.

Example: Assume (equivalently ). As shown in Fig. 2, points , , and on the capacity region are obtained for , , and , respectively. Consequently, the power allocated to the strong user () would be less than, equal to, and more than that of the weak user () if we are targeting to achieve points , , and , respectively. Each point corresponds to a different rate pair (). From Table I, it is seen that the achievable sum rates () for these points are , , and bps/Hz, respectively. This clearly shows that the sum rate increases by allocating more power to the stronger user. Indeed, if achievable sum rate is the only system performance metric, the stronger user must receive all power; i.e., is optimal.

Then, what is the reason for the common myth that power allocation in NOMA should be in reverse order of the users’ channel gains? The first answer to this question is that it is intuitive to assign more power to a user with a weaker channel to compensate for the higher channel loss. Such a mechanism, known as power control [4], has been adopted in 2G-4G cellular networks. Power control is important for the efficient and fair operation of cellular systems. This intuition leads to a more concrete answer to the above question. Allocating a higher power to users with weaker channels is motivated by supporting a certain quality of service (QoS) or improving user fairness. QoS is usually quantified by where is the minimum required rate for user . For weaker users this ‘usually’ implies allocating more power to compensate for the channel loss, but in general it depends on the value of , the minimum required rate. As an example, for and any is acceptable in Fig. 2. Clearly then, it is not necessary to allocate a higher power to the weaker user as any satisfies the QoS requirements while giving less power to the weaker user. In contrast, user fairness ‘commonly’ improves if more power is allocated to the weaker user. An example of this is moving from to in Fig. 2 which is equivalent to increasing the weaker user’s power from to . This, however, does not imply that allocating more power to the weak user is better in terms of fairness. A vivid example of this is which is an extremely unfair power allocation in view of user fairness.

In short, NOMA per se does not imply allocating a higher power to the user with the worse channel. Power allocation depends on the targeted point on the capacity region of the users scheduled in one cluster.

Iii-B Myth 2: SIC Decoding Order in NOMA Varies with Power Allocation

From Section II, we know that the stronger user first decodes the weaker user’s signal and next it decodes its own signal after cancelling the interference (i.e., the signal of the weak user). Obviously, the order of SIC is crucial for achieving the capacity region. One might, however, think power allocation affects the order of decoding. Let and be the powers allocated to user 1 and user 2, respectively, and suppose that . The question is whether the value of affects the order of SIC decoding at the receivers?

The answer to this question is “no”. The order of SIC decoding merely depends on the order of SNR at the receivers (), or equivalently, the magnitude of the channel gains. More specifically, with , to achieve the capacity region, regardless of the amount of power allocated to the users, user 1 needs to decode user 2’s signal first, and apply SIC to decode its own message free of interference. Further, user 2 has to treat user 1’s signal as noise when decoding its own message. This decoding (and SIC order) is optimal for any , including and even in the extreme case where ().

Misinterpretation of the SNR at the receivers could be a possible reason for the myth that “power allocation can affect the order of SIC decoding.” Specifically, at first glance, one might think the SNR at user 1 and user 2 is and , respectively. This is, however, wrong because both users receive one superimposed signal whose power is , which results in and as SNR at user 1 and user 2, respectively. This result extends to the case of , , users; i.e., in -user NOMA (-user BC), power allocation does not affect the SIC order, and the optimal decoding in general.

Iii-C Myth 3: Although the Weak User Does not Use SIC, the Impact of Interference Is Small due to Power Allocation

This misunderstanding is based on two incorrect assumptions. First, it is assumed that NOMA necessarily allocates a higher power to the weaker user, which is not, however, correct as elaborated in Myth 1. Second, even when the power allocation is very biased towards the weak user (), the effect of inter-user interference (caused by the strong user’s signal) may not be small depending on the value of . The term in will be negligible when . At dB, for example, this will be true only for an extremely biased power allocation (). Such a power allocation is usually very inefficient in terms of sum rate because it is allotting a tiny fraction of the power (less than ) to the strong user which has the biggest contribution to the achievable sum rate. That is, with we sacrifice sum rate, unless . On the other hand, if is very small (), we will have even if . Hence, in this case, inter-user interference will be negligible regardless of power allocation.

Iii-D Myth 4: The Main Reason Behind Using NOMA is to Improve Spectral Efficiency

This statement is not correct as illustrated with the following example.

Example: Assume that there are two users to be served but only one RB is available. Using the parameters listed in the caption of Fig. 2, we compare the achievable rates for the following three scenarios:

  1. only user 1 is served (OMA)

  2. only user 2 is served (OMA)

  3. both users are served (NOMA)

From Table I, it is seen that the sum rate is maximized when all power is allocated to UE1, which implies an OMA scheme since only one user has non-zero power. That is, OMA achieves a better sum rate than NOMA. This is not surprising as UE1 has a stronger channel than UE2 () and it is intuitive to allocate all power to UE1 if the goal is to maximize the sum rate. Since the achievable sum rate (network capacity from the mobile operators’ point of view) is an important metric for spectral efficiency, it is clear that NOMA is not as efficient as OMA in this sense.

This example indicates that spectral efficiency (if measured by sum rate) cannot be the only, or main, reason for adopting NOMA. Instead, the main motivation is to increase the number of users served in a limited number of RBs. Nonetheless, when other metrics such as user fairness and QoS (and weighted sum-rate in general) are considered, OMA is not advantageous in general and NOMA is the optimal solution. Therefore, by allowing both users (and in general users) to share one RB, NOMA sacrifices sum rate to increase the number of users or to improve QoS and user fairness.

Fig. 3: Universal and fractional frequency reuse in a network with total available bandwidth . (a) Universal frequency reuse where the same (total) bandwidth is used in all cells. This causes sever ICI at the cell-edge regions as shown in the figure. (b) FFR where total bandwidth is divided into four subbands: the same frequency () is used in all cell-centers while cell-edge regions in different cells use different frequencies (, , and ) to avoid ICI. (c) An example of NOMA-FFR in which , , and are used both at the cell-edge and cell-center to help pair NOMA users with different channel gains while the same is use at the cell-center regions to increase the reuse factor.

Iii-E Myth 5: ICI is More Severe in NOMA-based Networks due to the Biased Power Allocation Towards Cell-edge Uses

When a user is moving far away from the BS, its signal-to-interference-plus-noise ratio (SINR) generally reduces mainly for two reasons: 1) the received signal power becomes lower due to attenuation, and 2) the interference power from the adjacent cells, or ICI, becomes higher because the user gets closer to an adjacent BS. ICI arises due to simultaneous transmissions over the same frequency in adjacent cells. With universal frequency reuse in recent cellular networks, cell-edge users usually suffer from worse QoS due to ICI. This, in turn, reduces the overall system spectrum efficiency. Intuitively, ICI is negatively affected by increasing the signal transmission power for cell-edge users. Because of this, there is a misunderstanding that ICI is more severe in NOMA-based networks due to biased power allocation towards cell edge users. However, both NOMA users (in general, all users in the same cluster) receive one superimposed signal whose power is ,333We assume the total transmission power is , and the same power is used for OMA and NOMA for a fair comparison. as described in Myth 2. That is, cell-edge users receive the same power in NOMA and OMA, and ICI is not affected by power allocation to the individual users.

Iii-F Myth 6: NOMA is not Compatible with FFR

The basic idea behind fractional frequency reuse (FFR) is to allocate different (orthogonal) frequencies to the cell-edge regions of adjacent cells while allocating the same frequency band to the cell-interior regions of all cells, as shown in Fig. 3. Hence, with FFR, cell-interior and cell-edge users utilize different frequency bands in each cell. On the other hand, the canonical example of NOMA is to pair cell-interior and cell-edge users on the same frequency band to maximize the gain over OMA. These two concepts seem to be clashing. The former orthogonalizes the bandwidth allocated to cell-interior and cell-edge users whereas the latter tries to avoid orthogonalization for its suboptimality. Because of this and the fact that FFR is an effective ICI management technique in LTE networks, some researchers are skeptical about using NOMA in multi-cell networks.

Clearly, pairing cell-interior and cell-edge users in the cell-interior band contradicts the definition of FFR as otherwise the same frequencies would be used universally in the cell-edge region of all cells. In contrast, we can pair cell-interior and cell-edge users in the cell-edge band as a solution to combine NOMA and FFR, see Fig. 3(c). This implies that a higher number of users will share the cell-edge band and requires a larger fraction of the total bandwidth for the cell-edge bands. Another solution is to pair cell-interior users together and cell-edge users together, each in their own bands. Those users are very likely to have similar channel gains due to their comparable distance from the BS. Then, the spectral efficiency of NOMA, compared to OMA, reduces. To overcome this, scheduling becomes important and other techniques can be applies (see Myth 9).

Iii-G Myth 7: Decoding Complexity of NOMA Is Prohibitively High for UEs

Rooted in the BC theory, the basic idea behind NOMA is not particularly new; it has been established several decades ago. One main reason that this concept has not been used in practice is the fact that UEs have had limited processing power making interference cancellation prohibitively complex. However, recent advances have made the implementation of interference cancellation at UEs practical. For example, in 3GPP LTE-A, a category of relatively complex UEs, known as network-assisted interference cancellation and suppression (NAICS) terminals, has been adopted to mitigate interference in multi-cell networks [5]. NAICS leverages UEs’ interference cancelation capability to improve cell-edge users’ and consequently system throughput. Recent experimental trials of NOMA [1, Chapter 18] have shown that the complexity of NOMA is within the capabilities of current user terminals. Thanks to Moore’s law, the processing power is doubled approximately every two years. In light of this and the previous experience with NAICS, current and new generations of UEs are/will be capable of decoding NOMA.

Iii-H Myth 8: SIC Error Propagation Makes NOMA Inviable

Receiving multiple interfering users’ signal is not a new concept in cellular communications and most of the recent cellular systems have been dealing with this issue [6] due to the universal frequency reuse. CDMA receivers in 3G and NAICS UEs in 4G are notable examples of this. Albeit SIC used in those settings is different from that in downlink NOMA, there is genuine hope for widespread use of multi-user receivers and NOMA, at least under certain conditions.

In light of recent research, experimental results, and practical developments in various settings, see e.g. [1, 7, 8, 6], today it is known that SIC is amenable to implementation, and has been employed in commercial systems, such as CDMA and IEEE 802.15.4 [9]. Further, using stochastic geometry in random wireless networks, in [7] it is shown that SIC is highly beneficial with very low-rate codes and in environments where path loss is high. Similarly, [8] shows that channel disparity between the near and far users is important for successful decoding. Experimental results using universal software radio peripheral (USRP) hardware boards in [8] confirm the feasibility of SC-SIC in the two-user case. Appropriate channel and systems parameters such as channel disparity, modulation type, and power allocation are, however, crucial for successful operation.

Iii-I Myth 9: NOMA Users Must Have Different Channel Gains

This statement is not correct and NOMA users can even have exactly the same channel gains. However, with similar channel gains the spectral efficiency benefits of NOMA, compared to OMA, diminish, and for the NOMA rate region in Fig. 2 becomes the same as the OMA rate region. However, recall from Myth 4 that spectral efficiency is not the main reason for using NOMA. Further, there are other solutions to overcome this. In MIMO-NOMA, even if the users’ channel gains are similar, we can design the precoding matrix at the BS to degrade the effective channel gains of one user while enhancing that of the other user concurrently [10]. More sophisticated power allocation strategies, e.g., cognitive radio power allocation [11], can be used to strictly guarantee the users’ QoS requirements, even if they have similar channel gains.

Iii-J Myth 10: NOMA Compromises Security and Privacy

Since stronger users are capable of decoding the weaker users’ signal in NOMA, one might think that security and privacy of weaker users are compromised. But, this can even happen in OMA due to the broadcast nature of the wireless channel. On the other hand, being able to decode a user’s signal at the PHY layer does not imply decoding its message. There are upper-layer security measures to prevent this, e.g., scrambling bits based on a UE-specific code called cell-radio network temporary identifier (C-RNTI). Even when C-RNTI is needed to be shared with other UE’s, other encryption-based solutions can be used to avoid security/privacy issues [5]. Finally, physical layer security can guarantee security for NOMA (BC) even in the PHY layer [1, Chapter 5].

Iv Critical Questions and Future of NOMA

The important remaining challenges for NOMA are not theoretical, but rather related to system design and implementation as will be detailed below.

Iv-a What Benefits Does NOMA Offer under Practical Conditions?

The canonical NOMA problem, illustrated in Fig. 1, relies on several assumptions: there are only two users in each cluster, the channel state information (CSI) is known at the BS and the users, SIC can be performed perfectly, and user scheduling (clustering) is based on the users’ CSI. Although an increasing number of papers are pushing NOMA research ahead by going beyond those assumptions, still a majority of papers on NOMA, even those considering imperfect CSI, assume SIC can be performed perfectly and thus error propagation is negligible. A relevant question is then to what extent such imperfections affect the performance of NOMA, particularly when several users are clustered together. More specifically, can NOMA support several users in one OFDM RB in an efficient manner? Therefore, a critical question is: Can NOMA work efficiently in practical cellular networks?

Motivated by the above question, several research groups have proposed methods and made trials to evaluate the performance of NOMA under realistic conditions. A notable example is the recent experimental trials on NOMA elaborated in [1, Chapter 18]. This work assesses the link-level performance of a MIMO-NOMA system with different types of receivers in both indoor and outdoor environments. To achieve a block error rate of , the SNR gap between the experimental and simulation-based results is within 0.8 dB, as shown in [1, Fig. 18.8].

In academia, there has been a sensible movement towards NOMA research with more realistic assumptions. As an example, the impact of imperfect SIC due to imperfect CSI has recently been investigated in [12], and it is shown that imperfect SIC significantly degrades performance. However, [8] shows that although imperfect SIC can largely degrade the performance of multi-user detection, with well-designed codes, SC can still provide higher rates compared to OMA. More comprehensive studies are required to better understand the effect of different types of imperfect CSI on the performance of NOMA-based systems.

It is important to investigate, e.g. using tools from stochastic geometry [7], NOMA gains in large-scale wireless networks under practically relevant assumptions. From [7], it is known that SIC is beneficial only for very low-rate codes and successful decoding exponentially decreases with the number of users if high-rate codes are used. It is of great importance to understand the performance limits and benefits of NOMA in realistic settings (e.g., the effective number of users that can be clustered together) in terms of CSI and network size.

Iv-B Can NOMA Benefit from Machine Learning and Deep Learning?

Machine learning (ML) provides a data-driven approach to learn information and solve traditionally challenging problems without relying on predetermined models and equations. An emergent subfield of ML, namely deep learning (DL), has seen tremendous growth during the past years, and is being applied to almost every industry and research area, including different fields within communications, thanks to recent powerful DL software libraries and specialized hardware [13].

The viability of learning technologies, deep or shallow, in the field of communications has been confirmed by many independent researches. Notably, several works have recently used ML/DL for beamforming and power allocation. DL is also being applied to various NOMA problems such as encoding/decoding in sparse code multiple access [14], and power allocation in NOMA-based caching [15]. Given that complexity of NOMA clustering and power allocation grows exponentially with the number of users, and cellular networks are naturally dynamic in terms of topology and scheduling, it is of great interest to use learning-based approaches for user clustering, power allocation, and beamforming in the case of MIMO-NOMA systems. But the critical question is: Can learning-based approaches work effectively in dynamic networks with rapidly varying CSI?

To see the challenge, note that ML algorithms try to find data patterns to help make near-optimal decisions. Since, wireless channels can change as fast as every few milliseconds and the network topology is naturally very dynamic (mobile), learning the network and resource allocation via deep neural networks appears to be challenging but is an interesting research field.

References

  • [1] M. Vaezi, Z. Ding, and H. V. Poor, Multiple Access Techniques for 5G Wireless Networks and Beyond. Springer, 2019.
  • [2] ITU, “Minimum requirements related to technical performance for IMT-2020 radio interface(s).” https://www.itu.int/md/R15-SG05-C-0040/en, 2017. [Online; accessed August 2008].
  • [3] Y. Saito, Y. Kishiyama, A. Benjebbour, T. Nakamura, A. Li, and K. Higuchi, “Non-orthogonal multiple access (NOMA) for cellular future radio access,” in Proc. IEEE 77th Vehicular Technology Conference (VTC Spring), pp. 1–5, 2013.
  • [4] M. Chiang, P. Hande, T. Lan, C. W. Tan, et al., “Power control in wireless cellular networks,” Foundations and Trends® in Networking, vol. 2, no. 4, pp. 381–533, 2008.
  • [5] G. Zhou, W. Xu, and G. Bauch, “Network assisted inter-cell codeword cancellation for interference-limited LTE-A and beyond,” in Proc. IEEE Wireless Communications and Networking Conference Workshops, pp. 52–57, 2014.
  • [6] J. G. Andrews, “Interference cancellation for cellular systems: a contemporary overview,” IEEE Wireless Communications, vol. 12, no. 2, pp. 19–29, 2005.
  • [7] X. Zhang and M. Haenggi, “The performance of successive interference cancellation in random wireless networks,” IEEE Transactions on Information Theory, vol. 60, no. 10, pp. 6368–6388, 2014.
  • [8] S. Vanka, S. Srinivasa, Z. Gong, P. Vizi, K. Stamatiou, and M. Haenggi, “Superposition coding strategies: Design and experimental evaluation,” IEEE Transactions on Wireless Communications, vol. 11, no. 7, pp. 2628–2639, 2012.
  • [9] Y. Liu, Z. Qin, M. Elkashlan, Z. Ding, A. Nallanathan, and L. Hanzo, “Nonorthogonal multiple access for 5G and beyond,” Proceedings of the IEEE, vol. 105, no. 12, pp. 2347–2381, 2017.
  • [10] Z. Ding, L. Dai, and H. V. Poor, “MIMO-NOMA design for small packet transmission in the Internet of Things,” IEEE Access, vol. 4, pp. 1393–1405, 2016.
  • [11] Z. Ding, P. Fan, and H. V. Poor, “Impact of user pairing on 5G non-orthogonal multiple access downlink transmissions,” IEEE Transactions on Vehicular Technology, vol. 19, no. 8, pp. 1462–1465, 2015.
  • [12] X. Chen, Z. Zhang, C. Zhong, R. Jia, and D. W. K. Ng, “Fully non-orthogonal communication for massive access,” IEEE Transactions on Communications, vol. 66, no. 4, pp. 1717–1731, 2018.
  • [13] S. Dörner, S. Cammerer, J. Hoydis, and S. ten Brink, “Deep learning based communication over the air,” IEEE Journal of Selected Topics in Signal Processing, vol. 12, no. 1, pp. 132–143, 2018.
  • [14] M. Kim, N.-I. Kim, W. Lee, and D.-H. Cho, “Deep learning-aided SCMA,” IEEE Communications Letters, vol. 22, no. 4, pp. 720–723, 2018.
  • [15] K. D. Nguyen, M. Vaezi, W. Shin, H. V. Poor, H. Shin, and T. Q. Quek, “Power allocations in cache-aided NOMA systems: Optimization and deep learning approaches,” submitted to IEEE Journal of Selected Topics in Signal Processing (J-STSP), September 2018.

Mojtaba Vaezi [SM’18] (mvaezi@villanova.edu) received his Ph.D. in Electrical Engineering from McGill University in 2014. From 2015 to 2018, he held several research positions at Princeton University. He is currently an Assistant Professor at Villanova University. His research interests include the broad areas of wireless communications and information theory. Among his publications in these areas is the book Multiple Access Techniques for 5G Wireless Networks and Beyond (Springer, 2019). Dr. Vaezi was/is an Editor of IEEE Communications Magazine and IEEE Communications Letters, and an organizing member of four NOMA workshops at Globecom and ICC. He is a recipient of the IEEE Larry K. Wilson Student Activities Award in 2013 and NSERC Postdoctoral Fellowship in 2014.

Robert Schober [F’10] (robert.schober@fau.de) Robert Schober was a Professor and Canada Research Chair at the University of British Columbia (UBC), Vancouver, Canada, from 2002 to 2011. Since January 2012 he is an Alexander von Humboldt Professor and the Chair for Digital Communication at Friedrich-Alexander University Erlangen-Nuremberg (FAU), Germany. His research interests fall into the broad areas of Communication Theory, Wireless Communications, and Statistical Signal Processing.

Zhiguo Ding [SM] (zhiguo.ding@manchester.ac.uk) Zhiguo Ding received his B.Eng from BUPT in 2000, and his Ph.D degree from Imperial College in 2005. Currently, he is a professor in Communication at the University of Manchester. Dr Ding’ research interests are 5G networks, game theory, energy harvesting networks and statistical signal processing. He has been serving as an Editor for IEEE TCOM and IEEE TVT, and served as an editor for IEEE WCL and CL. He received the best paper award in IET-CWMC-2009 and IEEE-WCSP-2015, the EU Marie-Curie Fellowship 2012-2014, IEEE TVT Top Editor 2017, IEEE Heinrich Hertz Award 2018, and Jack Neubauer Memorial Award 2018.

H. Vincent Poor [F’87] (poor@princeton.edu) is the Michael Henry Strater University Professor of Electrical Engineering at Princeton University. His interests include information theory and signal processing, with applications in wireless networks and related fields. Dr. Poor is an IEEE Fellow, a Member of the National Academy of Engineering and the National Academy of Sciences, and a Foreign Member of the Royal Society. He received the Marconi and Armstrong Awards of the IEEE Communications Society in 2007 and 2009, respectively. Recent recognition of his work includes the 2017 IEEE Alexander Graham Bell Medal and honorary doctorates from several universities.

Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
   
Add comment
Cancel
Loading ...
278647
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description