Quality of Service Based NOMA Group D 2 D Communications

Non-orthogonal multiple access (NOMA) provides superior spectral efficiency and is considered as a promising multiple access scheme for fifth generation (5G) wireless systems. The spectrum efficiency can be further enhanced by enabling device-to-device (D2D) communications. In this work, we propose quality of service (QoS) based NOMA (Q-NOMA) group D2D communications in which the D2D receivers (DRs) are ordered according to their QoS requirements. We discuss two possible implementations of proposed Q-NOMA group D2D communications based on the two power allocation coefficient policies. In order to capture the key aspects of D2D communications, which are device clustering and spatial separation, we model the locations of D2D transmitters (DTs) by Gauss–Poisson process (GPP). The DRs are then considered to be clustered around DTs. Multiple DTs can exist in proximity of each other. In order to characterize the performance, we derive the Laplace transform of the interference at the probe D2D receiver and obtain a closed-form expression of its outage probability using stochastic geometry tools. The performance of proposed Q-NOMA group D2D communications is then evaluated and benchmarked against conventional paired D2D communications.


Introduction
With the advancement in mobile communication research, the usage of cellular technology has spread beyond voice and simple data transfer to high data rate, delay sensitive, and loss tolerant multimedia applications.Despite the fast growth of fourth generation (4G) systems, the current spectrum resources are still scarce to meet the ever increasing subscribers' demands for bandwidth and resource hungry applications, with vigorous requirements of seamless connectivity, anywhere and anytime.These trends compelled wireless researchers from the academia and industry to define new paradigm technologies and structures to achieve the goals of fifth generation (5G) systems [1,2].
In order to realize the concept of 5G into reality, many enabling technologies are proposed, among which millimetre waves, massive multiple-input multiple-output (MIMO), full-duplex (FD), heterogeneous deployments and software-defined networks have captured the attention of both academia and industry [2,3].Nevertheless, the role of multiple access scheme always remains a vital factor in cellular networks in order to enhance the system capacity in a cost effective manner, while utilizing the bandwidth in such a way that overall spectral efficiency will be increased [4].
Non-orthogonal multiple access (NOMA) is considered as a promising multiple access candidate for future fifth generation (5G) wireless systems due to its potential of improving spectrum efficiency [5,6].Unlike conventional orthogonal multiple access (OMA), NOMA superimposes message signals of different users in power domain and send this conglomerate signal using the same time, frequency or code resource [7,8].Successive interference cancellation (SIC) technique is employed at each receiver to cancel the intra-user interference [9].
Apart from NOMA, another emerging technique to enhance the spectral efficiency is device-to-device (D2D) communications [10], which has the ability to improve the spectral efficiency of the conventional cellular network by sharing same spectrum resources among cellular users and D2D pairs [11][12][13].
By introducing aforementioned two concepts and their potential to improve spectrum efficiency, it is natural to investigate the application of NOMA scheme to D2D communications.

Related Work and Motivation
Recently, a FD D2D aided cooperative NOMA is proposed in [14].The base station (BS) sends a NOMA signal to one NOMA-strong and one NOMA-weak user, where the strong user is equipped with full-duplex ability.By invoking D2D communication between strong and weak NOMA user pair, the authors proposed to improve the outage performance of the weak user with the aid of D2D aided direct and cooperative transmissions.However, they only considered a single-cell scenario where NOMA is conducted at BS while strong and weak users communicate via conventional paired D2D communication.The authors in [15] developed an analytical framework based on stochastic geometry to analyse cellular networks with underlay D2D communications.The D2D users are also equipped with FD transceivers and can operate in FD mode.The authors proposed criteria to select between FD and D2D modes of operation.They derived closed-form expression for outage probability to evaluate the performance of cellular and D2D users.However, they modeled spatial topology of D2D users by a Poisson point process (PPP), which may not be realistic distribution choice for D2D users.The reason is that PPP cannot capture the features of device clustering and spatial separation of D2D communications due to its completely random nature [16].Furthermore, NOMA protocol is not applied for both cellular and D2D communications.A relay assisted diversity communication is proposed in [17].The proposed analytical framework analyses the frame error probability performance by considering the effects of node locations, link characteristics, power allocation, diversity methods and distributed coding and constellation signaling.However, they considered an OMA based communication between single source and destination assisted by three relay nodes.The authors in [18] considered a downlink multiuser MIMO NOMA celullar network with underlaid D2D communications.They proposed two beamforming schemes in order to eliminate the inter-beam interference and the one caused to D2D users by BS transmission.In addition, they formulated an optimization problem to jointly study the performance of both cellular and D2D users.A potential limitation to their approach is that the considered system model is limited to single cell and no specific random distribution is utilised to model the spatial topologies of cellular and D2D users.Hence, it is not straightforward to generalize the results for a case of multi-cell network.Furthermore, they considered a paired D2D communication where D2D users do not apply NOMA protocol to communicate with each other.
In [19], the authors considered a NOMA-based D2D communications and introduced the concept of D2D group, where a D2D transmitter (DT) is communicating with multiple D2D receivers (DRs) using NOMA protocol.The authors proposed an optimal resource allocation strategy for interference management that enables to realize the NOMA-based D2D group communications.Although the concept of NOMA group D2D was introduced in [19], their system model was comprised of single-cell and lack of interference characterization at DR. Furthermore, it requires interference modeling and performance evaluation at the DR in order to extend the concept of NOMA group D2D to a general scenario, where DTs and DRs are distributed in the entire network.The authors in [20] further proposed cooperative hybrid automatic repeat request (HARQ) assisted NOMA in large-scale D2D networks.They studied the outage and throughput performance of the D2D users under the considered network setting and demonstrated that cooperative HARQ assisted NOMA achieves lower outage probability than non-cooperative case and OMA scheme.However, their NOMA based D2D network model is restricted to the two-user case only i.e., they only considered two-user NOMA transmission from D2D sources.Furthermore, they considered a significant difference between channels of two D2D users by assuming that one user is always closer to D2D source compared to the other user.This assumption may not always hold, particularly in the scenario of NOMA based D2D communications because DRs are clustered around DTs and are located in proximity of each other and hence DRs may have very similar channel conditions.
In the light of the abovementioned discussion, very little attention was paid to investigate NOMA group D2D communications.Hence, the aforementioned gaps and shortcomings motivated us to investigate and analyse the performance of NOMA based group D2D communications under interference limited scenarios.In order to capture the key features of D2D communications, i.e., device clustering and spatial separation, the DTs are considered to be randomly distributed over R 2 according to Gauss-Poisson process (GPP), while the DRs are assumed to be randomly clustered around DTs.The reason for choosing GPP is that it is a relatively simpler cluster point process that maintains good trade-off between modeling accuracy and analytical tractability.Therefore, in the context of D2D communications, GPP provides more realistic modeling approach against PPP case by capturing the clustering behavior of D2D devices [21].
Furthermore, the current approach to order users in NOMA group D2D communications is based on the channel gains of the DRs.This ordering approach may not be suitable to D2D communication scenario under which the DRs in the same group are clustered around a common DT and are located in proximity of each other.Hence, the channel conditions of the DRs located in the same D2D group would be very similar.Consequently, this ordering strategy may result in very similar power allocation, which could limit the gains of applying NOMA to D2D communications [7].In the context of NOMA, there are few works that use quality of service (QoS) based ordering [22][23][24].They mainly focused on cellular networks and are limited to the two-user case only.In this paper, we propose and analyse QoS based NOMA (Q-NOMA) group D2D communications and make an attempt to fill the aforementioned gaps in literature.To the best of our knowledge, it is the first time that Q-NOMA is proposed and analysed to realize group D2D communications under interference limited scenario.

Contributions
The main contributions of this work are as follows:

•
We propose Q-NOMA group D2D communications in which D2D users are randomly distributed over the entire two-dimensional plane.Unlike the existing proposals, we order the DRs according to their QoS requirements, which is more appropriate for the D2D communications scenario.Furthermore, in contrast to PPP, which is most commonly used to model D2D users (both DTs and DRs), we model the spatial topology of DTs by GPP and DRs are considered to be randomly clustered around DTs.These spatial distributions of DTs and DRs are suitable to analyse the proposed network with any number of D2D users.In addition, based on the QoS ordering, we propose two policies to compute power allocation coefficients that could lead to two implementations of the proposed Q-NOMA group D2D communications.

•
We derive the interference distribution at the probe DR by utilizing the results from stochastic geometry.The Laplace transform of interference over GPP is derived in [21], which involves complex double integrals.In order to obtain useful insights, a major step in characterizing the interference is the approximation of integrals in the interference Laplace transform by applying Gaussian-Chebyshev and Gauss-Laguerre quadratures.This approximation results in an interference Laplace transform expression, which is easy to implement.

•
Based on the interference approximation results, we further derive the closed-form expression for outage probability of the DRs in the proposed Q-NOMA group D2D communications.

•
We present numerical results to validate the accuracy of the derived outage results and compare the performance of the proposed Q-NOMA group D2D with conventional paired D2D communications using OMA.

Mathematical Preliminaries on Gauss-Poisson Process
The generalized GPP [21] is defined as the Poisson cluster process with homogeneous independent clustering.Denote λ as the parent process intensity.Then, each cluster in GPP can be classified as single-point or two-point cluster.Let 1 − a and a denote the probabilities that a cluster is single-point, and two-point, respectively.When the cluster is single-point, the point is located at the position of the parent.When the cluster is two-point, one point is located at the position of the parent while the other is randomly distributed around the parent with some probability density function (PDF) f u (•), where u is the inter-point distance in two-point cluster.

System Model
Consider inband D2D communications with an overlay cellular network.We consider a frequency reuse factor of one among D2D users to achieve better spectrum efficiency.With this setting, every D2D transmission by a DT is subjected to interference from other active DTs.We consider a composite fading and path loss channel model between every DT and DR.In this work, we assume that the power fading coefficients are independent and identically distributed (i.i.d) with exponential distribution of unit mean, and adopt a path loss model of d −α , where d is the distance between the probe DR and test DT, and α is the path loss exponent.

Spatial Distribution of D2D Users
Consider that D2D users are randomly distributed over R 2 .At any time realization, the D2D users are classified as transmitters or receivers.We consider the group D2D scenario, where each DT is communicating with multiple DRs via a NOMA scheme.We refer to a DT communicating to multiple D2D devices as a group transmitter (GT).Any D2D user can take a role of GT.We assume that the selection of GTs is performed by a BS, and it can select multiple GTs in a given cell to improve overall system capacity.We allow multiple GTs to exist in proximity of each other, where each GT is communicating to its own group of receivers.Hence, at any time realization, each selected GT forms a group/cluster containing DRs.In order to capture both inband and device clustering, we model the spatial topology of the GTs by a stationary and isotropic GPP defined on R 2 , denoted by Φ GT with parent process intensity λ GT .Furthermore, we model the coverage of each GT by a disc D with radius R D .We consider that the DRs are clustered around each GT and are uniformly distributed inside coverage of GTs.An illustration of NOMA group D2D communication is presented in Figure 1.For a quick reference, a list of commonly used abbreviations is given in Table 1.

Q-NOMA Group D2D Communication
Since DRs are clustered around GTs and are located in proximity of each other, therefore, the DRs connected to same GT would have very similar channel conditions.Consequently, in the context of D2D communications, ordering DRs according to their channel conditions to apply NOMA at GT may not achieve the desired multiplexing gains and fairness among DRs.Hence, in this work, we propose ordering the DRs of the GT according to QoS requirements, which are determined by their targeted data rates.
Let there be a total of M DRs distributed inside coverage of a test GT.The probe DR is assumed to be located at the origin, with a desired test GT at x 0 = (d, 0) with d = 0. Without loss of generality, we assume that the DRs are ordered as R 1 ≤ ... ≤ R M , where R i is the targeted data rate of DR i, 1 ≤ i ≤ M. Correspondingly, the power allocation coefficients are sorted as The aforementioned procedure to order DRs of the GT and compute their power allocation coefficients according to users' targeted rates is termed as "Q-NOMA".When it is applied to D2D communications, we refer to the communication as "Q-NOMA group D2D communication".
Consider that the NOMA DR m is the probe receiver, then, the received signal from the test transmitter at the probe DR is given as: where P GT is the transmission power of test GT, s i is the message signal of DR i and n m is the additive white Gaussian noise (AWGN) with zero mean and variance σ 2 .

Power Allocation Coefficients Policies
The optimal power and resource allocation improve overall performance and utilise the system resources efficiently.However, the optimum power allocation strategies proposed in existing literature, for example, [25], cannot be directly applied to the current work because of significant difference in system model or underlying transmission method.Therefore, in this sub-section, we discuss two simple methods to compute power allocation coefficients {β i } M i=1 that would lead to two possible implementations of Q-NOMA group D2D communications.

Policy I
The Policy I utilises targeted rates of DRs to compute the power allocation coefficients.Similar to [7], the power allocation coefficient for DR i under Policy I is computed as: The intuition behind Equation ( 2) is that power allocation coefficients could be utilised to maintain fairness among DRs [3,26].In case they are computed in proportion with users' targeted rates, then the highest ordered DR M would result in the highest SIC decoding order with maximum signal-to-interference-plus-noise ratio (SINR) threshold among all ordered DRs.Consequently, all of the lower ordered DRs always require the maximum SINR threshold dictated by the user M, which would result in a biased treatment of lower ordered users (with lower targeted rates).Hence, in order to avoid this biasness, we propose to compute the power allocation coefficients as given by Equation (2).

Policy II
The concept of fixed NOMA (F-NOMA) was proposed in [27], where the power allocation coefficients are fixed and are computed based on the given user ordering i.e., it does not utilise the actual channel gains to compute {β i } M i=1 .Similar to [27], we adopt a Policy II to compute {β i } M i=1 for conventional NOMA that does not utilise the actual targeted rates of DRs.The power allocation coefficient for DR i under Policy II is computed as: where β F m represents the power allocation coefficient of DR m in Policy II and µ is selected in such a way that ∑ M i=1 β F i = 1.

Interference Distribution
The reception at the probe DR from the test GT is interfered by the other GTs.The interference at probe DR is given as, x , where g x and d x represent the Rayleigh fading channel gain and distance between probe DR and interferer at x 0 , respectively.The following lemma provides the Laplace transform of the interference at probe DR.
Lemma 1.Consider a GPP Φ GT with parent process intensity λ GT modeling spatial topology of the GTs in a Q-NOMA group D2D communications.Then, the Laplace transform of the interference at the probe DR conditioned at the location of test GT is given by: where Ω p = ω p e r p , ω p = Γ(P+2)r p P!(P+1) 2 (LP+1(rp)) 2 , L P (•) is the Laguerre polynomial of degree P, r p are the roots of L P (•), X 1 (•) and Λ 2 (•) are given in Equations (A7) and (A10), respectively.
Proof.See Appendix A.

Outage Probability Analysis
In this section, we focus on the outage probability for the DRs in the considered Q-NOMA group D2D communication.Let τ m and R m denote the SINR threshold and targeted rate of DR m, respectively, where τ m = 2 R m − 1.Since each DR employs SIC, the outage at DR m occurs if it does not meet the targeted rate of any higher order DR j, where 1 ≤ j < m.Denote ζ m→j = Rm,j < R j as the outage event at DR m due to decoding of DR j, where Rm,j is the achievable rate of user j at DR m.The outage event ζ m→j can be expressed as: where is the transmit signal-to-noise ratio (SNR) and P I is the maximum received interference power at the probe DR.
Next, define ϕ max m = max {ϕ 1 , ..., ϕ m }.Based on Equation (3), the outage probability at the DR m can be given as: where F h m is the cumulative distribution function (CDF) of h m .Note that the set {h i } , i = 1, ..., M, of channel gains is not ordered because the users are sorted in ascending order of their targeted rates.Since the channel gains are i.id.random variables with common CDF F h , Equation ( 4) can be re-written as: Consequently, the outage probability of DR m is provided in the following theorem.
Theorem 1.The outage probability of DR m in the Q-NOMA group D2D communications is derived as: and L is the complexity-accuracy trade-off parameter.

Proof. See Appendix B.
Note that, due to the presence of interferers in the network, similar to [20,28], the derived outage probability in Equation ( 8) is a function of variables ϕ max m and interference Laplace transform L I .This is different from existing works that analyse NOMA wireless systems under no interference where outage probability is mainly a function of ϕ max m .For reference, please see [14,27,29].Furthermore, it is worthy to note that the current outage analysis approach remains valid if more complex cluster models (Poisson cluster process, etc.) are adopted for modeling spatial distributions of DTs and DRs.In that case, the L I term will be replaced by the Laplace transform of the interference for the adopted model.Intuitively, L I is performing a form of scaling in Equation ( 8) and hence the conclusions are expected to remain the same if the spatial distribution model(s) of the DTs and DRs is changed.However, the exact impact on outage probability when more complex cluster processes model is adopted for modeling DT and DR locations would require further study, and we plan to investigate them in our future work.

Numerical Results and Discussion
This section presents the numerical results to evaluate the performance of the considered network as well as to validate the accuracy of the derived expression in Equation ( 8) of Section 3. As shown in Table 2, simulation parameters used are similar to those in [29], unless otherwise stated.Furthermore, we adopt Policy I as a default policy to compute power allocation coefficients, unless otherwise stated.Outage Probability Several observations can be made from the results in Figure 2: (1) increasing the coverage radius of GT results in a higher outage probability because of a larger path loss; (2) different ordered users have distinct decreasing slopes of outage probability because of different targeted rates; (3) the higher order DRs in our proposed Q-NOMA group D2D communications achieve better outage performance because they have smaller targeted rates.This is different from conventional NOMA that orders users based on channel conditions, and where the higher ordered users have larger outage probabilities due to poor channel conditions [27,28].

Impact of d on Outage Probability
The impact of varying distance between probe DR and test GT on the outage performance is investigated in Figure 3.It can be observed that varying d has a larger impact on the average achievable outage probability at lower SNR value of 5 dB.This is intuitively plausible because increasing transmission power results in improved SINR at the receiver and hence better outage performance.

Comparison between Two Implementations of Q-NOMA Group D2D Communications
In this section, we compare the performance of two possible implementations of Q-NOMA group D2D communications based on power allocation coefficients Policies I and II.The results for Q-NOMA group D2D communication under Policy I are also obtained by utilising Equation ( 8), with the exception that {ϕ max m } M m=1 are computed by using fixed power allocation coefficients β F m M m=1 .We consider five different cases for users' targeted rates and compared the outage probabilities achieved by the Q-NOMA group D2D communications under Policies I and II at an SNR of 25 dB.Table 3 summarizes the outage comparison between these two implementations of the group D2D communications.The power allocation coefficients for Policy I are derived as β F m = {0.5, 0.33, 0.17} M m=1 , whereas those for Policy II are calculated using Equation (2) based on users' targeted rates and are shown in Table 3.It is observed from the results in Table 3 that Q-NOMA under Policy I achieves lower outage probability in cases 1 and 2, whereas it obtains better outage performance in cases 4 and 5 for Policy II.In addition, the results in case 3 indicate the importance of proper power and rate allocation to avoid situations of complete outage.Moreover, it can be observed that the Q-NOMA group D2D communication under Policy I performs better in cases where the users' targeted rates are significantly apart.This is more suitable to D2D communication scenarios where different users may have diverse QoS requirements.
In order to extend the analysis beyond the case of M > 3 users, we further present in Table 4 the average outage comparison between paired D2D and Q-NOMA group D2D communications (based on power allocation coefficients Policies I and II).We consider M = 5, 7, 9 with R m ∼ U (0.1, 2) , ∀m ∈ M and ρ t = 25 dB, where U (•, •) represents the random uniform distribution function.It can be observed that the results in Table 4 are consistent with those presented in Figure 4, for example, Q-NOMA group D2D communications consistently achieve lower outage probability than paired D2D communications.

Conclusions
In this paper, we have proposed Q-NOMA group D2D communications.In order to study the performance of the proposed network, we first derive the Laplace transform of the interference expression, based on which we further derive the closed-form expression for outage probability to analyse the performance of the DRs in the proposed Q-NOMA group D2D communications.The results show that the proposed Q-NOMA group D2D achieves overall lower outage probability than its counterpart paired D2D communication.Furthermore, based on two power allocation coefficient policies, we have presented the comparison between two possible implementations of the proposed Q-NOMA group D2D communications.Due to the similar channel conditions and diverse QoS requirements of DRs, the results show that the proposed Q-NOMA implementation based on Policy I is more realistic and suitable than one based on Policy II for group D2D communications.As future work, we plan to investigate and analyse the performance of proposed Q-NOMA group D2D communications in underlay cellular mode, where additional interferences from base stations and cellular users will be considered.Another interesting future direction is to extend the current network model to a framework of MIMO systems.This would require a robust precoder and MIMO-NOMA design to improve the system performance and capacity while mitigating the inter and intra beam interference efficiently.
Based on Equation (A12), P m in Equation ( 5) can be expressed as:

Figure 2
Figure 2 presents the impact of varying coverage radius R D of test GT on the outage probability of ordered DRs as a function of SNR.The derived outage results in Equation (8) are shown to be in good agreement with the Monte Carlo simulations.

Figure 2 .
Figure 2. Impact of R D on outage probability.

Figure 3 .
Figure 3. Impact of d on outage probability.

4. 3 .
Comparison between Paired and Grouped D2D CommunicationsThe average outage probability achieved by Q-NOMA group D2D communications under different path loss exponents is shown in Figure4.The performance of paired D2D communication based on OMA is also presented in the figure as a benchmark for comparison.It can be observed that Q-NOMA group D2D achieves overall lower outage probability than the paired D2D communication for different values of path loss.This is because, as opposed to paired D2D, Q-NOMA group D2D communication uses only single transmission, which results in better SINR at the DRs under an interference limited scenario.

Table 4 .
Average outage comparison between paired device-to-device (D2D) and quality of service based non-orthogonal multiple access (Q-NOMA) group D2D communications (M > 3).Total Users, M Q-NOMA Policy I Q-NOMA Policy II Paired D2D f I (x) is the PDF of interference I and the last step follows from the definition of Laplace transform.This proves the result in Theorem 1.

Table 1 .
List of commonly used abbreviations.

Table 3 .
Average outage comparison between two implementations of quality of service based non-orthogonal multiple access (Q-NOMA) group device-to-device (D2D) communications.