Exploiting Multiple-Antenna Techniques for Non-Orthogonal Multiple Access

I. INTRODUCTION

• The paper designs a multiple-antenna NOMA downlink framework covering user clustering, CSI estimation, superposition coding, beamforming, and SIC in both TDD and FDD modes.The study considers heterogeneous downlink channels and a base station with arbitrarily many antennas and different CSI accuracies across channels.Source: §I · p. 4
• The analysis derives exact average transmission-rate expressions for each user in an arbitrary cluster as functions of transmit power, CSI accuracy, and transmission mode.Source: §I · p. 4
• The paper optimizes transmit power, feedback bits, and transmission mode, deriving closed-form power allocation and feedback distribution and proposing a low-complexity joint scheme.Source: §I · p. 4
• With moderate or high CSI accuracy, using two users per cluster is optimal in practical cases, theoretically supporting two-user setups studied in prior work –, –,,.Source: §I · p. 4
• Imperfect CSI leaves residual inter-cluster interference, creating a performance gap from perfect-CSI NOMA that grows with information-signal power and shrinks with CSI accuracy.Source: §I · p. 4
• To maintain constant performance as information-signal power grows, TDD pilot-sequence transmit energy and FDD spatial resolution should increase.Source: §I · pp. 4–5
• For sum rate, one cluster containing all users is optimal with no BS CSI, whereas one user per cluster is best with perfect BS CSI.Source: §I · p. 5
• In interference-limited conditions, the strongest user’s average rate grows linearly with feedback bits and equal power allocation asymptotically approaches optimal performance; under noise limitation, average rate is independent of CSI accuracy.Source: §I · p. 5

• The paper presents system design guidelines, uses Section V simulations to validate the proposed schemes, and concludes in Section VI.Source: §I · p. 5

A. User Clustering

• The proposed spatial-direction clustering is presented as a simple alternative to exhaustive-search clustering schemes with high implementation complexity.Source: A. User Clustering · p. 6
• Users with similar spatial directions but distinct propagation distances are grouped together, while users with similar channel gains may be separated by increasing spatial resolution.A shared beam can align users in a cluster; distance differences help avoid severe intra-cluster interference and support SIC.Source: A. User Clustering · p. 6

B. CSI Acquisition

• In TDD, the BS estimates downlink CSI from uplink pilots using MMSE estimation, with accuracy improvable by increasing pilot power or sequence length.Source: B. CSI Acquisition · p. 7
• In FDD, users quantize channel direction using a codebook and return its index; the BS recovers phase information, and larger codebooks improve CSI accuracy.Source: B. CSI Acquisition · pp. 7–8

C. Superposition Coding and Transmit Beamforming

• The BS uses zero-forcing beamforming to mitigate or cancel inter-cluster interference, constructing each cluster beam from the null space of other clusters’ estimated channels.Source: C. Superposition Coding and Transmit… · p. 8
• Within each cluster, the transmitted signal superimposes users’ signals with allocated powers, which distinguish users in the power domain.Source: C. Superposition Coding and Transmit… · p. 9
• With perfect CSI at the BS, inter-cluster interference can be completely canceled.Source: C. Superposition Coding and Transmit… · p. 9

D. Successive Interference Cancellation

• Users are ordered by effective channel gain so stronger users decode and subtract signals intended for weaker users before decoding their own signals.Source: D. Successive Interference Cancellation · p. 9
• The SINR accounts for residual intra-cluster interference after SIC, residual inter-cluster interference after beamforming, and noise; the analysis assumes perfect SIC.The paper notes practical SIC may be imperfect because of limited MU computational capability, but its impact is left for future work.Source: D. Successive Interference Cancellation · p. 10

III. PERFORMANCE ANALYSIS AND OPTIMIZATION

• For imperfect CSI, the paper derives closed-form average transmission rates for the first MU and other MUs, then proposes separate and joint optimization of power, feedback bits, and transmission mode.Source: §III · p. 10

A. Average Transmission Rate

• The derived rate expressions allow evaluation across system parameters and channel conditions, including transmit power, CSI accuracy, and transmission mode.Source: A. Average Transmission Rate · p. 13

B. Power Allocation

• The paper notes that the sum-rate power optimization is nonconvex, making a direct closed-form optimum difficult to obtain.Source: B. Power Allocation · p. 14
• The method separates allocation into two steps: distribute total power among clusters, then allocate each cluster’s assigned power among its users.Source: B. Power Allocation · p. 14
• The proposed low-complexity allocation distributes power across clusters in inverse proportion to their interference coefficients to mitigate inter-cluster interference.The scheme uses long-term path-loss information available at the BS, resulting in low overhead and implementation complexity.Source: B. Power Allocation · pp. 14, 16
• Within a cluster, power allocation balances sum-rate considerations with the power ordering needed to facilitate SIC.The proposed rule equally distributes powers to fulfill both criteria.Source: B. Power Allocation · p. 15

C. Feedback Distribution

• Because CSI accuracy affects inter-cluster interference and the feedback link has a limited bit budget, the paper minimizes average summed inter-cluster interference through feedback allocation.Source: C. Feedback Distribution · p. 16
• The feedback-bit allocation is formulated as an integer programming problem, relaxed for optimization, then rounded down to an integer for practical use.The relaxed allocation is determined from channel statistics and transmit power allocation.Source: C. Feedback Distribution · p. 17
• More feedback bits go to MUs that generate more average inter-cluster interference, supporting more accurate ZFBF to minimize total interference.The allocation is based on each MU's average inter-cluster interference relative to that of the other MUs.Source: C. Feedback Distribution · p. 18

D. Mode Selection

• With no downlink CSI, the optimal mode is one cluster; with perfect CSI, one MU per cluster is best.Without CSI, ZFBF cannot mitigate inter-cluster interference, while perfect CSI lets ZFBF completely eliminate it.Source: D. Mode Selection · p. 18
• Partial CSI leaves residual interference, and increasing users per cluster can reduce inter-cluster interference while increasing intra-cluster interference.Source: D. Mode Selection · p. 18
• For partial CSI, the paper dynamically selects the number of clusters and MUs per cluster to maximize the sum of average transmission rates.Source: D. Mode Selection · pp. 18–19
• Mode selection is an integer programming problem solved by numerical search with O(NK) complexity, which is acceptable when each cluster has few MUs.The paper gives K = 2 as an example of a small cluster size.Source: D. Mode Selection · p. 19

E. Joint Optimization Scheme

• Transmit power, feedback bits, and transmission mode are coupled, so the joint scheme allocates power, distributes feedback bits, and selects the mode with the largest sum rate.For a given mode, it uses the power allocation in (22) and feedback distribution in (27); complexity is mainly determined by mode selection.Source: E. Joint Optimization Scheme · p. 19

A. Interference Limited Case

• At high transmit power, the average transmission rate is independent of total transmit power, giving a performance ceiling.The result follows because the expressions in (32) and (33) are independent of total transmit power.Source: A. Interference Limited Case · pp. 20–21
• With perfect CSI, the first MU's average rate grows proportionally to log2(P^S_tot), while each k > 1 MU remains rate-bounded and cannot raise its bound through power allocation.Source: A. Interference Limited Case · p. 22
• For the first MU, imperfect CSI leaves a fixed rate gap from the ideal rate when training power in TDD or feedback bits in FDD scale with total transmit power.The stated FDD condition is B_n,1 = (M − 1) log2(P^S_tot/ε).Source: A. Interference Limited Case · p. 23
• At high power with many feedback bits, the first MU's average rate increases linearly with the number of feedback bits.Source: A. Interference Limited Case · p. 24

B. Noise Limited Case

• When interference is negligible relative to noise at low transmit power, the system is in the interference-free case.Source: B. Noise Limited Case · p. 24
• The average transmission rate is independent of CSI accuracy, so channel estimation and CSI feedback are unnecessary in this scenario.Source: B. Noise Limited Case · p. 24
• With intra- and inter-cluster interference negligible, ZFBF and SIC are not required, and all optimization schemes asymptotically approach the same performance.Source: B. Noise Limited Case · p. 24

Rate-expression validation

• The theoretical expressions closely match simulations across the SNR region, while both users’ average rates saturate at high SNR.The first MU performs better than the second, consistent with the stated NOMA principle.Source: Figure 2 · p. 25, §V · p. 25

Power allocation

• The proposed power allocation outperforms equal and fixed allocation, especially at medium SNR; at high SNR, it and equal allocation reach the same saturated sum rate, while fixed allocation loses performance.The fixed allocation uses a 1:4 power ratio between the two MUs to facilitate SIC.Source: §V · pp. 25–26, Figure 3 · p. 26

Large-feedback asymptotics

• With many feedback bits, the first MU’s average rate is nearly linear in feedback bits and exceeds the second MU’s rate.This behavior reconfirms Lemma 3.Source: §V · p. 26, Figure 5 · p. 27

Transmission mode

• At low CSI accuracy, mode 4 (N=1, K=6) achieves the largest sum rate; at medium accuracy, mode 2 (N=3, K=2) is optimal; at high accuracy, mode 1 (N=6, K=1) performs best.Dynamic mode selection can always obtain the maximum sum rate.Source: §V · p. 27

Joint optimization

• At high CSI accuracy, the proposed joint optimization scheme performs much better than the two baseline schemes, with gains of about 3 b/s/Hz at ρ = 0.8 and up to more than 5 b/s/Hz at ρ = 0.9.Source: §V · p. 29
• The fixed scheme beats TDMA at low and high CSI accuracy but is slightly worse in the medium regime; joint optimization performs much better than both baselines.The joint scheme distributes power, allocates feedback based on that power, and then selects the transmission mode.Source: Figure 7 · p. 28, §V · pp. 28–29

• The proposed framework covers a general multiple-antenna NOMA multiuser downlink in both TDD and FDD and is paired with exact closed-form average-rate expressions.Source: §VI · p. 29
• The paper optimizes transmit power, feedback bits, and transmission mode.Source: §VI · p. 29
• Asymptotic performance analysis yields insights into system performance and design guidelines.Source: §VI · p. 29