PAPER Joint Maximum Likelihood Detection in Far User of Non-Orthogonal Multiple Access

Transcription

1 IEICE TRANS. COMMUN., VOL.E100 B, NO.1 JANUARY PAPER Joint Maximum Likelihood Detection in Far User o Non-Orthogonal Multiple Access Kenji ANDO a), Student Member, Yukitoshi SANADA b), and Takahiko SABA c), Senior Members SUMMARY Non-orthogonal multiple access (NOMA) enables multiple mobile devices to share the same requency band. In a conventional NOMA scheme, the receiver o a ar user detects its desired signal without canceling the signal or a near user. However, the signal or the near user acts as intererence and degrades the accuracy o likelihood values or the ar user. In this paper, a joint maximum likelihood detection scheme or the ar user o the NOMA downlink is proposed. The proposed scheme takes the intererence signal into account in calculating the likelihood values. Numerical results obtained through computer simulation show that the proposed scheme improves the perormance by rom 0.2 db to 3.1 db or power allocation coeicients o 0.2 to 0.4 at a bit error rate (BER) o 10 2 relative to the conventional scheme. key words: non-orthogonal multiple access, joint maximum likelihood detection, successive intererence cancelation 1. Introduction Due to the rise in the penetration rate o mobile devices such as smartphones and tablet PC, demands or higher data rates and larger capacity have been increasing in cellular systems. The report released by ITU-R has shown that the amount o mobile data traic will increase 1000 times by 2020 [1]. Technical solutions to resolve the problems caused by this explosive traic growth have been investigated, and candidate system concepts or uture radio access beyond LTE- Advanced have been discussed. One solution or increasing the capacity is non-orthogonal multiple access (NOMA) [2] [16]. NOMA enables multiple mobile users to share the same requency band. By sending signals with dierent transmission powers, a receiver can extract its desired signal. It has been shown that the total channel capacity increases i a receiver can remove the overlaid intererence signal [17]. In NOMA, mobile users that are located near and ar rom a base station are assigned the same requency resource and allowed to share it through a scheduling algorithm. Since the amount o propagation loss depends on the distance rom the base station, the base station transmits the signal or the ar user with larger power than the signal or the near user. Manuscript received March 11, Manuscript revised June 17, Manuscript publicized July 29, The authors are with the Dept. o Electronics and Electrical Engineering, Keio University, Yokohama-shi, Japan. The author is with the Dept. o Computer Science, Chiba Institute o Technology, Narashino-shi, Japan. The part o this paper is submitted to IEEE APWCS a) kando@snd.elec.keio.ac.jp b) sanada@elec.keio.ac.jp c) saba@cs.it-chiba.ac.jp DOI: /transcom.2016EBP3098 The receiver o the near user cancels the signal or the ar user by means o successive intererence cancelation (SIC) and then detects its desired signal [6] [15]. On the other hand, the receiver o the ar user demodulates the signal without canceling the signal or the near user and employs maximum likelihood detection (MLD). This is because the signal or the near user is expected to be attenuated and it is not regarded as intererence by the receiver o the ar user. However, the signal or the near user actually intereres with the signal or the ar user and degrades the accuracy o likelihood values in the MLD. Thus, a joint detection scheme in the receiver o the ar user has been proposed [16]. In [16], the sum o the constellation constrained capacities (CCCs) o a NOMA downlink with the joint detection scheme in both the near and the ar users is derived. In deriving the CCCs, the signal or the near user is treated as probabilistic intererence. However, the derivation o CCCs does not assume the superposition o the constellation points o the signals or the near and the ar users. Thus, this paper investigates direct eects o the joint MLD in the ar user o the NOMA system on the link-level perormance with speciic settings o transmission parameters including modulation schemes, coding rates, and power allocation coeicients. These numerical results show that the joint MLD in the ar user o the NOMA system works eectively with various transmission parameter values. This paper is organized as ollows. Section 2 describes the system model and the proposed scheme. In Sect. 3, the bit error rate (BER) curves obtained through computer simulation are presented. Sect. 4 gives our conclusions. 2. System Model 2.1 Non-Orthogonal Multiple Access Model In this paper, a NOMA downlink is assumed as a system model, in which two mobile users share the same requency band. A base station and each o mobile terminals are equipped with a single antenna. A block diagram o the base station is presented in Fig. 1. In the LTE standard, a systematic parallel concatenated convolutional code is adopted [18]. In a turbo encoder, two 8-state encoders and one interleaver are used. With the input o N s bits, the turbo encoder generates three length-n s streams, d n (0), d n (1), and d n (2). They are reerred as the Systematic, Parity 1, and Parity 2 streams, respectively. At the end o each stream, our tail bits or trellis termination are appended. Copyright 2017 The Institute o Electronics, Inormation and Communication Engineers

2 178 IEICE TRANS. COMMUN., VOL.E100 B, NO.1 JANUARY 2017 Fig. 1 Block diagram o base station. In order to support a higher data rate, bit puncturing is implemented in the turbo codes. In the LTE standard, puncturing is conducted in a rate matching (RM) block [18]. In the RM block, each o the three output streams is rearranged in a sub-block interleaver. Aterward, a single output buer with the length o 3(N s + 4) is illed by placing the rearranged Systematic bit stream in the beginning, ollowed by the two rearranged bit streams o Parity 1 and Parity 2 interlaced in a bit-by-bit ashion. The contents o the buer are then passed to a circular buer or bit selection and puncturing. The output o the circular buer orms a single bit stream with the length o 3(N s + 4). Ater bit puncturing, the RM block generates a single bit stream with the length o (N s + 4)/r, where r is the code rate and the minimum o r is 1/3. The outputs o the RM blocks or the near and the ar users are mapped to 2 M n QAM and 2 M QAM symbols through Gray coding, where M n and M are the numbers o bits per symbol o the near and the ar users, respectively. Ater symbol mapping, they are put into serial-toparallel (S/P) converters and summed together on each subcarrier. Here, S n [k] and S [k] are the symbols to be transmitted or the near and the ar users on the kth subcarrier. The transmit power is divided with the coeicients o α and (1 α). The symbol to be transmitted on the kth subcarrier is composed o the sum o the symbols o the near and the ar users as ollows: S αs n [k] + 1 αs [k] (0 k N 1), where N is the size o the IDFT. Ater the IDFT and the parallel-to-serial (P/S) conversion, the non-orthogonal OFDM signal is generated. The last part o each OFDM symbol is replicated and inserted at the beginning o the symbol as a guard interval (GI). The generated OFDM signal is put into a transmit ilter and is transmitted. At the receiver side, the received signal irst goes through a receive ilter. In an analog-to-digital (A/D) converter, the received signal is digitized to discrete samples. The signal on the kth subcarrier ater the removal o the GI and the N-sample DFT is given as Y N 1 n=0 ( y[n] exp j 2πnk ) N (1) = H[k]S[k] + W[k], (2) Fig. 2 Conventional detection scheme in ar user. where y[n] is the nth received signal in the time domain, Y[k] is the received signal, H[k] is the requency response, and W[k] is the noise in the requency domain on the kth subcarrier, respectively. The output o the DFT on the kth subcarrier is sent to a detector to obtain sot inormation or a turbo decoder. At the detector, the likelihood values or turbo decoding are calculated. Ater the likelihood calculation in the detector, bit depuncturing is carried out. The likelihood sequence is then passed to an inverse circular buer. Aterward, the likelihood sequence is divided into three streams that correspond to Systematic, Parity 1, and Parity 2 streams, respectively. Each o the three output streams is rearranged in a sub-block deinterleaver and is then put into the turbo decoder to obtain inormation bits. 2.2 Conventional Detection Scheme In the NOMA system, the signals or the near and the ar users are superposed on each subcarrier. Since the signal power o the ar user is larger than that o the near user, the conventional detection scheme in the ar user employs MLD. In the conventional detection scheme, the signal or the near user is expected to be attenuated and it is not regarded as intererence by the receiver o the ar user. The block diagram o the conventional detection scheme in the ar user is shown in Fig. 2. The likelihood calculations by the MLD are given as, M LD,1 Ŝ [k] {S } b Y[k] H[k] 1 αŝ [k] 2 ), (3)

3 ANDO et al.: JOINT MAXIMUM LIKELIHOOD DETECTION IN FAR USER OF NON-ORTHOGONAL MULTIPLE ACCESS M LD,0 Ŝ [k] {S } b Y[k] H[k] 1 αŝ [k] 2 ), (4) 179 where M,1 LD [k] and M,0 LD [k] are the sums o the likelihood values or the mth (1 m M ) bit being 1 and 0 in the symbol or the ar user S [k] on the kth subcarrier, N 0 is the power spectrum density o the white Gaussian noise, Ŝ [k] is the symbol candidate or the ar user on the kth subcarrier, and {S } b and {S } b are the sets o symbols o which the mth bit is equal to 1 and 0, respectively. From Eqs. (3) and (4), the log likelihood ratio (LLR) is calculated as L(b m Y[k]) = log M LD,1 [k] M LD,0 [k], (5) where L(b m Y[k]) is the LLR o the mth bit o the symbol obtained rom the received signal Y[k] on the kth subcarrier. The LLR calculated as above is passed to the turbo decoder and then the inormation bits are recovered. Fig. 3 Codeword level SIC in ar user. 2.3 Codeword Level SIC in the Far User The conventional detection scheme with the MLD in the ar user neglects the signal or the near user. However, when the signal power or the near user is larger than the noise power at the ar user, intererence owing to the signal or the near user degrades the accuracy o the likelihood values. Meanwhile, in [6] [15], the codeword level SIC is implemented in the receiver o the near user. Thus, it is assumed here that the codeword level SIC is also applied to the ar user. The block diagram o the codeword level SIC in the ar user is presented in Fig. 3. The receiver o the ar user employs the MLD and then obtains the decoded bits or the ar user. From the decoded bits, the receiver generates a replica signal or the ar user. The receiver then subtracts the replica signal or the ar user rom the received signal. The received signal ater the intererence cancelation is given as Y SIC n Y[k] H[k] 1 α S [k], (6) where Yn SIC [k] is the received signal ater the codeword level SIC and S [k] is the replica signal or the ar user on the kth subcarrier. The receiver o the ar user carries out the MLD against the remaining signal. The likelihood calculations or the signal or the near user are given as, n,1 M LD Yn SIC [k] Ŝ n [k] {S n } b n M LD n,0 H[k] αŝ n [k] 2 ), (7) Yn SIC Ŝ n [k] {S n } b n [k] H[k] αŝ n [k] 2 ), (8) where n,1 M LD [k] and n,0 M LD [k] are the sums o the likelihood values or the mth (1 m M n ) bit o S n [k] or the near user on the kth subcarrier, Ŝ n [k] is the symbol candidate or the near user on the kth subcarrier, and {S n } b n and {S n } b n are the sets o symbols o which the mth bit is equal to 1 and 0, respectively. From the sum o the likelihood values, the LLR is calculated. The LLR is passed to the decoder and then the decoded bits or the near user are obtained. The codeword level SIC generates the replica signal or the near user. Applying the codeword level SIC again, the received signal ater the intererence cancelation is given as Y SIC Y[k] H[k] α S n [k], (9) where Y SIC [k] is the received signal ater the codeword level SIC or the ar user and S n [k] is the replica signal or the near user on the kth subcarrier. The receiver o the ar user carries out the MLD against the remaining signal to obtain the decoded bits or the ar user. The likelihood calculations or the ar user are given as, D M LD,1 D M LD,0 Ŝ [k] {S } b Ŝ [k] {S } b Y SIC [k] H[k] 1 αŝ [k] 2 ), (10) Y SIC [k] H[k] 1 αŝ [k] 2 ), (11)

4 180 IEICE TRANS. COMMUN., VOL.E100 B, NO.1 JANUARY 2017 where D M,1 LD [k] and D M,0 LD [k] are the sums o the likelihood values or the mth bit being 1 and 0 in the symbol S [k] on the kth subcarrier, respectively. From the sum o the likelihood values, the LLR is calculated as the same manner as Eq. (5). The decoded bits or the ar user are then decoded. 2.4 Proposed Joint MLD Scheme In the conventional scheme, the receiver o the ar user employs the MLD by neglecting the signal or the near user. Alternatively, as described in Sect. 2.3, the receiver may employ the codeword level SIC by treating the signal or the near user as the intererence. On the other hand, in the proposed scheme, the receiver o the ar user applies joint MLD to the received signal. The joint MLD treats the received signal as the superposed signal o the near and the ar users with a larger number o constellation points. In the joint MLD, the signal or the near user is taken into account in the likelihood calculation. The LLR is calculated with the coordinates o the candidate constellation points and the received signal point in each symbol. It does not depend on the code rate o the signal or the near user because the ar user does not decode the received signal or the near user. Thus, the perormance o the ar user depends only on the code rate o the signal or the ar user and the modulation schemes o the signals or the near and the ar users. The block diagram o the proposed detection scheme in the ar user is shown in Fig. 4. The likelihood calculations o the joint MLD are given as, D J M LD,1 Ŝ n [k] {S n },Ŝ [k] {S } b exp ( 1 σ 2 Y[k] H[k]( αŝ n [k]) D J M LD,0 Ŝ n [k] {S n },Ŝ [k] {S } b exp ( 1 σ 2 Y[k] H[k]( αŝ n [k]) + 1 αŝ [k]) 2), (12) + 1 αŝ [k]) 2), (13) where D J M LD,1 [k] and D J M LD,0 [k] are the sums o the likelihood values or the mth bit being 1 and 0 in the symbol Fig. 4 Proposed joint MLD scheme in ar user. S [k] on the kth subcarrier, respectively. Since the signals or the near and the ar users are taken into account at the same time, the likelihood values are more accurate through the joint MLD. From Eqs. (12) and (13), the LLR is calculated as L J M LD (b m Y[k]) = log DJ M LD,1 [k] D J M LD,0 [k], (14) where L J M LD (b m Y[k]) is the LLR o the mth bit in the received symbol o the ar user obtained rom the received signal Y[k] on the kth subcarrier. The LLR calculated as above is passed to the decoder and then the inormation bits or the ar user are recovered. 2.5 Detection Complexity and Processing Delay The dierence o the detection complexity between the MLD and the joint MLD is the number o the candidate constellation points to be considered. The number o the constellation points in the MLD is equal to the modulation order o the signal or the ar user whereas the number o the constellation points in the joint MLD is the product o the modulation orders o the signals or the near and the ar users. Meanwhile, the MLD is also employed in the codeword level SIC. Even though it is applied three times, the detection complexity o the MLD in the codeword level SIC is smaller than that o the joint MLD. However, the codeword level SIC also requires the turbo decoding three times and the replica generation two times. These processes at least cause a delay which is larger than that caused by the proposed scheme though the amount o delay depends largely on the implementation o the decoding process. 3. Numerical Results 3.1 Simulation Conditions Numerical results obtained through computer simulation are presented in this section. Simulation conditions are shown in Table 1. An 8-state memory turbo code is employed and the size o the interleaver is 4800 [18]. The code rates o the signals or the near and the ar users are selected rom 1/3, 1/2, 2/3, and 5/6 through the RM unction. Encoded bits are modulated with QPSK, 16QAM or 64QAM or the near user and QPSK or the ar user on each subcarrier. The power allocation coeicient has been examined in a range o up to 0.4 in [15]. Thus, the link-level perormance o the receiver o the ar user is evaluated under power allocation coeicients o 0.1 to 0.4. In this paper, the power allocation coeicient, α, is set to 0.2 or 0.35 unless it is speciied. The signals or the near and the ar users are superposed on the subcarrier to orm the non-orthogonal signal. Other speciications such as the channel bandwidth, the subcarrier spacing, the number Dierent combinations o modulation orders in the near and the ar users are also examined, but the observed results show the same tendency as those presented in the ollowing sections.

5 ANDO et al.: JOINT MAXIMUM LIKELIHOOD DETECTION IN FAR USER OF NON-ORTHOGONAL MULTIPLE ACCESS 181 Table 1 Simulation conditions. Forward error coding turbo code Interleaver size 4800 Code rate 1/3, 1/2, 2/3, 5/6 Modulation scheme QPSK/16QAM/64QAM or near user +OFDM Modulation scheme QPSK or ar user +OFDM Power allocation α = 0.2, 0.35 coeicient Channel Bandwidth 2.5 MHz Subcarrier spacing 15 khz Number o subcarriers 256 Number o 151 data subcarriers Sampling requency 3.84 MHz GI length 5.21 µs (1st symbol) 4.69 µs (2nd-7th symbols) Decoding algorithm Log-MAP algorithm Decoding iterations 8 Channel model 6 Taps GSM-TU Model Channel estimation Ideal Number o trials bits o DFT/IDFT points, and the sampling requency ollow those o the LTE standard. The channel bandwidth is set to 2.5 MHz and the subcarrier spacing is set to 15 khz. The number o DFT/IDFT points is set to 256 and 151 subcarriers are used or data transmission. The sampling requency is set to 3.84 MHz. The guard interval is set to 5.21 µs or the irst symbol and 4.69 µs or the ollowing six symbols. The Log-MAP algorithm is employed or decoding and the number o decoding iterations is eight. The channel model is assumed to be the 6-tap GSM typical urban (TU) model [19]. The channel response on each subcarrier is assumed to be ideally estimated. As detection schemes, MLD, codeword level SIC, and joint MLD are employed. The number o trials is set to bits or each plot. Furthermore, the perormance o an orthogonal multiple access (OMA) scheme without the superposition o the signal or the near user is also presented or comparison [15]. Fig. 5 BER vs. E b /N 0 in ar user (near user: 16QAM, ar user: QPSK, code rate o ar user: 1/3, power allocation coeicient α = 0.2). 3.2 Comparison o BER Perormance The BER perormance versus E b /N 0 in the ar user is presented in Figs Encoded bits are modulated with 16QAM or 64QAM or the near user and QPSK or the ar user on each subcarrier. The power allocation coeicient, α, is set to 0.2 or 0.35 in Figs The BER perormance versus E b /N 0 in the ar user is shown in Fig. 5. The code rate o the signal or the ar user is set to 1/3. The proposed scheme outperorms the MLD by 0.2 db at a BER o This is because the joint MLD takes the signal or the near user into account when calculating the likelihood values, whereas the MLD neglects it. On the other hand, the perormance o the codeword level SIC with a code rate o 1/3 in the near user is worse by 0.2 db at a BER o 10 2 as compared with that o the MLD. Dierent code rates o the signals or the near user such as 1/2 and 5/6 are also examined and the same BER tendency has been observed. Fig. 6 BER vs. E b /N 0 in ar user (near user: 16QAM, ar user: QPSK, code rate o ar user: 1/3, power allocation coeicient α = 0.35). Although the MLD is also employed in the codeword level SIC, the likelihood values are not accurate, which results in incorrect replica signals. This is due to the attenuation o the signal power as well as the residual intererence ater cancelation. Thus, the cancelation o the signal or the near user increases the intererence to the signal or the ar user. The BER perormance versus E b /N 0 in the ar user with a power allocation coeicient o 0.35 is shown in Fig. 6. The

6 182 IEICE TRANS. COMMUN., VOL.E100 B, NO.1 JANUARY 2017 Fig. 7 BER vs. E b /N 0 in ar user (near user: 16QAM or 64QAM, ar user: QPSK, code rate o ar user: 1/2, power allocation coeicient α = 0.2). code rate o the signal or the ar user is set to 1/3. For a power allocation coeicient o 0.35, the proposed scheme outperorms the MLD by 1.5 db at a BER o The improvement in E b /N 0 due to the proposed scheme with a power allocation coeicient o 0.35 is larger than that with a power allocation coeicient o 0.2. This is because when the power allocation coeicient increases, the intererence to the signal or the ar user becomes large. The receiver with the MLD suers rom the intererence while the receiver with the joint MLD can calculate the likelihood values accurately. On the other hand, the perormance o the codeword level SIC with a power allocation coeicient o 0.35 is worse by 2.8 db at a BER o 10 2 as compared with that o the MLD. The amount o the degradation with a power allocation coeicient o 0.35 is larger than that with a power allocation coeicient o 0.2. This is because the MLD is also employed at the irst stage in the codeword level SIC and then decoding errors occur due to the large intererence. In the codeword level SIC, decoding errors increase the residual intererence ater cancelation. The BER perormance versus E b /N 0 in the ar user with a code rate o 1/2 in the signal or the ar user is shown in Fig. 7. For a code rate o 1/2, the proposed scheme outperorms the MLD by 0.5 db at a BER o On the other hand, the perormance o the codeword level SIC with a code rate o 1/2 in the signal or the near user is worse by 0.8 db at a BER o 10 2 as compared with that o the MLD. The perormance o the codeword level SIC or the case o 64QAM in the signal or the near user is worse by 0.6 db at a BER o 10 2 as compared with that o the MLD. The BER is worse or the case o 16QAM in the signal or the near user. The residual intererence ater cancelation due to decoding Fig. 8 BER vs. E b /N 0 in ar user (near user: 16QAM, ar user: QPSK, code rate o ar user: 5/6, power allocation coeicient α = 0.2). errors is smaller or the case o 64QAM than or the case o 16QAM though the BERs o the signal or the near user ater the MLD are almost equivalent as shown in Fig. 13 in Sect This is because the minimum distance between the signal constellation points is smaller or the case o 64QAM in terms o the same symbol power. Thus, the inluence o decoding errors or the case o 64QAM is smaller than that or the case o 16QAM. The BER perormance versus E b /N 0 in the ar user with a code rate o 5/6 is also shown in Fig. 8. For a code rate o 5/6, the proposed scheme outperorms the MLD by 0.5 db at a BER o The perormance o the codeword level SIC is worse and the BER curve shows the error loor with the increase in E b /N 0. The reason is that the error correction capability is too low or a code rate o 5/6 and bit errors remain ater decoding o the signal or the near user. 3.3 Eect o Code Rate The required E b /N 0 at a BER o 10 2 versus the code rate is shown in Fig. 9. As the code rate increases, the conventional detection scheme requires more E b /N 0 than the proposed scheme. This is because when the code rate increases, the error correction capability becomes low. As shown in Sect. 3.2, the perormance o the codeword level SIC is worse than that o the MLD. 3.4 Eect o Power Allocation Coeicient The required E b /N 0 at a BER o 10 2 versus the power allocation coeicient, α, is shown in Fig. 10. When the power allocation coeicient is large, the signal or the near user intereres with the signal or the ar user more signiicantly and the required E b /N 0 increases. The improvement in the

7 ANDO et al.: JOINT MAXIMUM LIKELIHOOD DETECTION IN FAR USER OF NON-ORTHOGONAL MULTIPLE ACCESS 183 Fig. 11 Constellation points o superposed signals or near and ar users (near user: 16QAM, ar user: QPSK, power allocation coeicient α = 0.3). Fig. 9 Required E b /N 0 at BER o 10 2 versus code rate (near user: 16QAM, ar user: QPSK, power allocation coeicient α = 0.2). Fig. 12 Required E b /N 0 at BER o 10 2 versus modulation schemes or near user (power allocation coeicient α = 0.2, code rate: 1/2). Fig. 10 Required E b /N 0 at BER o 10 2 versus power allocation coeicient α (near user: 16QAM, ar user: QPSK, code rate: 1/3). required E b /N 0 due to the joint MLD becomes larger when the power allocation coeicient increases. Concretely, the amount o the improvement is 0.2 db under the condition that the power allocation coeicient is set to 0.2 and it becomes 0.8 db, 1.5 db, and 3.1 db when the power allocation coeicient is set to 0.3, 0.35, and 0.4, respectively. It is clear that the proposed scheme is more eective when larger intererence is caused by the signal or the near user. The amount o the perormance deterioration o the codeword level SIC is especially large when the power allocation coeicient is at around 0.3. With a power allocation coeicient o 0.3, some o the constellation points in the superposed signal are close to each other as shown in Fig. 11. Thereore, decoding errors occur easily, which results in incorrect intererence cancelation. 3.5 Eect o Modulation Scheme o Near User The required E b /N 0 at a BER o 10 2 versus the various modulation schemes or the near user is presented in Fig. 12. The amount o the improvement due to the joint MLD or the case o 16QAM in the signal or the near user is larger than that or the case o QPSK. The amplitude o QPSK is constant while that o 16QAM varies. The receiver o the ar user occasionally suers rom larger intererence owing to 16QAM symbols even though the mean power o the intererence is the same. The variation o the intererence am-

8 184 IEICE TRANS. COMMUN., VOL.E100 B, NO.1 JANUARY 2017 Fig. 13 BER perormance ater irst codeword level SIC vs. E b /N 0 (ar user: QPSK, power allocation coeicient α = 0.2, code rate: 1/2). Table 2 Conditions o system-level simulation. Cell layout 19 hexagonal cell site Inter-site distance 500 m Minimum distance 35 m between users and cell site Distance dependent path loss log 10 (r) db r: Distance [km] Shadowing standard deviation 8 db Shadowing correlation 0.5 Channel model Six-path Rayleigh Number o users per cell U = 5, 10, 20, 30 Scheduling algorithm PF scheduling Time interval 100 Transmit power 42 dbm Power allocation coeicient (0.05 step) Receiver noise density 174 dbm/hz System bandwidth 4.32 MHz Resource block bandwidth 180 khz Number o RBs 24 Modulation scheme QPSK, 16QAM, 64QAM User drop 400 Trial per user drop 30 Number o symbols per trial 100 plitude deteriorates the BER perormance due to the MLD. This phenomenon is also observed in the case with 64QAM. The proposed scheme is then more eective when a higher order modulation scheme is employed or the near user. As or the codeword level SIC, the required E b /N 0 or the case o QPSK in the signal or the near user is smaller than those or the cases o 16QAM and 64QAM in Fig. 12. The BER perormance or the case o 64QAM in the signal or the near user is better than that or the case o 16QAM as shown in Fig. 7 and the required E b /N 0 or the case o 64QAM is smaller than that or the case o 16QAM in Fig. 12. The ollowing two actors aect the perormance o the required E b /N 0 versus modulation schemes in the signal or the near user in the codeword level SIC. The BER perormance ater the irst codeword level SIC versus E b /N 0 is shown in Fig. 13. As shown in Fig. 13, the BER perormance o the signal or the near user or the cases o 16QAM and 64QAM is worse than that or the case o QPSK. The amplitude o the QPSK signal is constant, whereas those o the 16QAM and the 64QAM signals vary. For the cases o 16QAM and 64QAM in the signal or the near user, the receiver o the ar user suers rom larger amplitude symbols as compared with the receiver or the case o QPSK even though the mean power o the signal is the same at the input o the codeword level SIC. Thereore, or the cases o 16QAM and 64QAM, decoding errors occur more easily than or the case o QPSK, which results in incorrect intererence cancelation. Furthermore, the minimum distance between the signal constellation points is another actor as explained in Sect The BER perormance is better or the case o 64QAM in the signal or the near user than or the case o 16QAM in Fig. 7. This is because the minimum distance between the signal constellation points is smaller or the case o 64QAM in terms o the same symbol power in average. Thus, the inluence o decoding errors or the case o 64QAM is smaller than that or the case o 16QAM. 3.6 System-Level Aspect The system-level simulation o the NOMA downlink with the joint detection scheme in both the near and the ar users is conducted as described in [20]. The conditions o the system-level simulation are presented in Table 2. A 19- hexagonal macrocell model is assumed. The cell radius o the macrocells is set to 289 m (inter-site distance = 500 m). Users are dropped randomly with a uniorm distribution. As a propagation model, distance-dependent path loss with a decay actor o 3.76 and lognormal shadowing with a standard deviation o 8 db are assumed. The shadowing correlation between the sites is set to 0.5. A six-path Rayleigh ading channel model with an exponential decay proile is assumed. The root-mean-square (RMS) delay spread is set to 1 µs and the maximum Doppler requency is set to 5.55 Hz. The transmit power is set to 42 dbm and the receiver noise density is set to 174 dbm/hz. The power allocation coeicient is selected rom between 0 and 1 with a step size o In the OMA scheme, it is set to 1.0. The throughputs o the conventional and proposed schemes are evaluated with CCC through Monte Calro simulation [16]. Proportional airness (PF) scheduling is applied and the time interval o the PF scheduling is set to 100 [21]. The number o resource blocks (RBs) is 24 and the number o subcarriers in one RB is 12. The system bandwidth is 4.32 MHz. A QPSK, 16QAM, or 64QAM symbol is transmitted on each subcarrier. The number o user drops is 400 and the number o trials per user

9 ANDO et al.: JOINT MAXIMUM LIKELIHOOD DETECTION IN FAR USER OF NON-ORTHOGONAL MULTIPLE ACCESS 185 Fig. 14 Number o users in each cell vs. system throughput. reason is that the required E b /N 0 reduces with the proposed scheme or the same BER in that range o the power allocation coeicient as presented in Fig. 10. The PF scheduling then selects the power allocation coeicient with higher probabilities at the range o 0.25 to Conclusions In this paper, a joint MLD scheme in the NOMA downlink or the ar user has been proposed and its BER perormance investigated. In the conventional NOMA system, the receiver o the ar user employs the MLD and neglects the signal or the near user even though its intererence degrades the accuracy o likelihood values. On the other hand, the proposed joint MLD scheme takes the signal or the near user into account when calculating the likelihood values. It improves the accuracy o the likelihood values or the signal or the ar user and thus improves its BER. Numerical results obtained through computer simulation have shown that the proposed scheme improves the perormance by rom 0.2 db to 3.1 db or power allocation coeicients o 0.2 to 0.4 at a BER o 10 2 relative to the conventional scheme the conventional scheme. In addition, the proposed scheme is more eective i the power allocation coeicient increases or a higher order modulation scheme is employed or the near user. Acknowledgments This work is supported in part by a Grant-in-Aid or Scientiic Research (C) under Grant No rom the Ministry o Education, Culture, Sport, Science, and Technology in Japan. Reerences Fig. 15 PDF o power allocation coeicient with the NOMA scheme (Transmit power: 42 dbm, 30 users). drop is 30. The channel response is renewed each trial. The number o transmitted symbols per trial is 100 and the last 80 symbols are used or throughput evaluation. The relationship between the number o users in each cell versus the system throughput per subcarrier is shown in Fig. 14. Here, 19 hexagonal cell sites are assumed and proportional airness scheduling is employed or user assignment. The numerical results obtained through system-level simulation show that the joint MLD in the ar user increases the system throughput by 0.3 bit/subcarrier. The probability distribution unction (PDF) o the power allocation coeicient, α, is presented in Fig. 15. The probability that the power allocation coeicient is within the range o 0.25 to 0.5 is higher in the ar user with the joint MLD whereas it is 0 in the ar user without joint MLD. The [1] Assessment o the global mobile broadband deployments and orecasts or international mobile telecommunications, Report ITU-R, M2243, Jan [2] Y. Kishiyama, A. Benjebbour, and T. Nakamura, Superimposed radio resource sharing or improving uplink spectrum eiciency, 14th Asia-Paciic Conerence on Communications, pp.1 5, Oct [3] H. Osada, H. Nishimura, M. Inamori, and Y. Sanada, Adjacent channel intererence cancelation in ractional sampling OFDM receiver, IEEE Vehicular Technology Conerence (VTC Fall), pp.1 5, [4] H. Osada, M. Inamori, and Y. Sanada, Non-orthogonal access scheme over multiple channels with iterative intererence cancellation and ractional sampling in OFDM receiver, IEICE Trans. Commun., vol.e95-b, no.12, pp , Dec [5] Y. Chida and Y. Sanada, LLR calculation based on intererence cancelation with channel estimation error or non-orthogonal multiple access, IEEE 80th Vehicular Technology Conerence (VTC2014- Fall), pp.1 5, [6] T. Takeda and K. Higuchi, Enhanced user airness using nonorthogonal access with SIC in cellular uplink, IEEE Vehicular Technology Conerence (VTC Fall), pp.1 5, [7] S. Tomida and K. Higuchi, Non-orthogonal access with SIC in cellular downlink or user airness enhancement, International Symposium on Intelligent Signal Processing and Communications Systems (ISPACS), pp.1 6, [8] Y. Kishiyama, A. Benjebbour, and T. Nakamura, Initial views on non-orthogonal multiple access based radio interace or uture radio

10 186 IEICE TRANS. COMMUN., VOL.E100 B, NO.1 JANUARY 2017 access, IEICE Technical Report, RCS , July [9] K. Higuchi and Y. Kishiyama, Non-orthogonal access with successive intererence cancellation or uture radio access, IEEE Vehicular Technology Society Asia Paciic Wireless Communications Symposium, Aug [10] N. Otao, Y. Kishiyama, and K. Higuchi, Perormance o nonorthogonal access with SIC in cellular downlink using proportional air-based resource allocation, International Symposium on Wireless Communication Systems (ISWCS), pp , [11] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, Systemlevel perormance evaluation o downlink non-orthogonal multiple access (NOMA), IEEE 24th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC), pp , [12] A. Benjebbour, A. Li, Y. Saito, Y. Kishiyama, A. Harada, and T. Nakamura, System-level perormance o downlink NOMA or uture LTE enhancements, 2013 IEEE Globecom Workshops (GC Wkshps), pp.66 70, [13] K. Yamamoto, Y. Saito, and K. Higuchi, System-level throughput o non-orthogonal access with SIC in cellular downlink when channel estimation error exists, IEEE 79th Vehicular Technology Conerence (VTC Spring), pp.1 5, [14] Study on Network-Assisted Intererence Cancellation and Suppression (NAIC) or LTE, 3GPP TR V12.0.0, March [15] K. Saito, A. Benjebbour, A. Harada, Y. Kishiyama, and T. Nakamura, Link-level perormance o downlink NOMA with SIC receiver considering error vector magnitude, IEEE 81st Vehicular Technology Conerence (VTC Spring), pp.1 5, [16] T. Yazaki and Y. Sanada, Eect o joint detection and decoding in non-orthogonal multiple access, International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS), pp , [17] D. Tse and P. Viswanath, Fundamentals o Wireless Communication, Cambridge University Press, [18] Multiplexing and Channel Coding, 3GPP TS V11.4.0, Jan [19] Radio Transmission and Reception, 3GPP TS V11.4.0, Jan [20] T. Yazaki and Y. Sanada, Throughput perormance o joint detection in non-orthogonal multiple access with and without semi-orthogonal multiple access modulation, IEICE Technical Report, RCS , March [21] T. Seyama, T. Dateki, and H. Seki, Eicient selection o user sets or downlink non-orthogonal multiple access, IEEE 26th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC), pp , Yukitoshi Sanada was born in Tokyo in He received his B.E. degree in electrical engineering rom Keio University, Yokohama Japan, his M.A.Sc. degree in electrical engineering rom the University o Victoria, B.C., Canada, and his Ph.D. degree in electrical engineering rom Keio University, Yokohama Japan, in 1992, 1995, and 1997, respectively. In 1997 he joined the Faculty o Engineering, Tokyo Institute o Technology as a Research Associate. In 2000 he joined Advanced Telecommunication Laboratory, Sony Computer Science Laboratories, Inc., as an associate researcher. In 2001 he joined Faculty o Science and Engineering, Keio University, where he is now a proessor. He received the Young Engineer Award rom IEICE Japan in His current research interests are in sotware deined radio, cognitive radio, and OFDM systems. Takahiko Saba was born in Tokyo in He received his B.E., M.E., and Ph.D. degrees all in electrical engineering rom Keio University, Yokohama, Japan in 1992, 1994, and 1997, respectively. From 1994 to 1997, he was a Special Researcher o Fellowships o the Japan Society or the Promotion o Science or Japanese Junior Scientists. From 1997 to 1998, he was with the Department o Electrical and Computer Engineering, Nagoya Institute o Technology, Nagoya, Japan, as a Research Assistant. From 1998, he joined the Department o Computer Science, Chiba Institute o Technology, Narashino, Japan, where is now a Proessor. In 2008, he was a Visiting Associate Proessor o University o British Columbia, B.C., Canada. From 2015, he is a Vice President o Chiba Institute o Technology. He received Distinguished Contributions Award rom Communications Society o IEICE in 2009 and From 2013 to 2015, he was an Editor-in-Chie o IEICE Transactions on Communications (Japanese Ed.). His current research interests include wireless communications and physical layer security. He is a member o IEEE. Kenji Ando was born in Shizuoka, Japan in He received his B.E. degree in electronics engineering rom Keio University, Japan in Since April 2015, he has been a graduate student in School o Integrated Design Engineering, Graduate School o Science and Technology, Keio University. His research interests are NOMA system.