低延遲虛擬細胞車用通信網路下行鏈路多用戶檢測技術之研究

全文

(1)國立臺灣大學電機資訊學院電信工程學研究所博士論文 Graduate Institute of Communication Engineering College of Electrical Engineering and Computer Science. National Taiwan University Doctoral Dissertation. 低延遲虛擬細胞車用通信網路下行鏈路多用戶檢測技術之研究 Downlink Multiuser Detection in the Ultra-Low Latency Virtual Cell-Based Vehicular Networks. 曾智修 Chih-Hsiu Zeng. 指導教授：陳光禎博士 Advisor: Kwang-Cheng Chen, Ph.D.. 中華民國 108 年 7 月 July 2019 doi:10.6342/NTU201902343.

(2) 誌謝在軍旅生涯任官服役12年之際，我選擇退伍，返回學校攻讀博士。說起來這是一個大膽又有點不負責任的決定，畢竟我不是年輕小伙子。感謝父、母親理解、支持我任性的選擇。念博士感覺好像是一段自我修練、成長的過程。感謝指導教授陳光禎博士的訓練。他通常不會很具體地告訴我要做什麼，只是給一個方向、建議，引導我思考問題，實際上研究怎麼做、要做什麼，他都是讓我自己想，並說服他為什麼要做這個或這麼做。念博士的前1、2年，跟老師開會討論其實壓力挺大的，他喜歡挑戰學生的觀點、想法，他會說：「你講的東西是claim、是信仰，不是theorem、真理!」我想他真正想告訴學生的是「任何結論、陳述，都應該要提供證據」。我最喜歡老師的一句話是，當我們怕犯(說)錯，而不敢提出自己的觀點時，他會說：「如果你不敢犯錯，那你永遠不會做對!」我很慶幸我博士班的老闆是KC，如果重來一次，我仍會選KC當我指導教授，雖然他很“硬“。感謝瑞昱半導體柳德政處長，在每次計畫報告時給我的建議，他會很細心地審視我的模擬結果、並挑出報告中的typo。謝謝電信所所辦趙姐、康小姐行政上、經費結報的幫助。謝謝曾志成學長，在我做Clustering Coefficient的研究時，耐心聽我報告並給予建議。謝謝口試委員給我的建議及肯定，特別要感謝闕志達教授，在正式口試前給我的寶貴意見。很慶(榮)幸能夠結識實驗的學長(弟、妹)們。奕志真的幫我很多，尤其是博班第1、 2年;謝謝人豪在口試及上海 ICC 的幫助; 謝謝Eisaku在美國的時候煮飯給我吃，讓我不至於天天吃垃圾食物; 也很謝謝羅恩臨，在美國時的照顧、帶我們出去玩;謝謝紹洲、徐祥、奕丞在我當“MTK計畫’’ Coordinator時的鼎力幫助，跟你們討論讓我獲益良多。最後謝謝林茂昭教授實驗室的“大學長“蔡華龍博士，感謝你的關心及建議。. i. doi:10.6342/NTU201902343.

(3) 中文摘要為普及未來自動駕駛車之應用，須建構可靠、低延遲的無線網路通信系統。近期研究成果指出，引進霧或邊緣運算(Fog or Edge Computing)有利對自動駕駛之即時管理與控制。進而整合虛擬細胞 (Virtual Cell) 概念、開路通訊 (Open-loop Communications)、主動式網路鏈結(Proactive Network Association)，可將通信延遲時間降低至1毫秒(ms)，但鄰近細胞間(Virtual Cells)間的相互干擾卻難以避免。在開路通訊(Open-loop Communications)中，為提升頻寬使用效率，通道資訊(Channel State Information, CSI) 不會反饋至傳送端，傳統波束賦形 (Beamforming) 或干擾對齊 (Interference Alignment)將無法使用，故下行鏈路的干擾須由接收端運用多用戶檢測 (Multiuser Detection, MUD)予以處理。我們發現當使用最大似然多用戶檢測(Maximum-likelihood MUD, ML-MUD)時，位元錯誤率(Bit Error Rate, BER)對干擾源的調變技術(Modulation)非常敏感。若干擾源使用低階調變(Low-Order Modulation)，接收信號之位元錯誤率(Bit Error Rate, BER)仍可接近理論上之理想值。但是，當干擾源採用高階調變(High-Order Modulation)時，位元錯誤率則明顯地變差，我們的研究發現運用多天線技術可降低錯誤率對干擾源調變技術的敏感程度。我們也提出兩個方法降低多用戶檢測運算複雜度。第一個方法係利用下行鏈路的特性縮小所有可能解的信號空間(Solution Space)，稱為低運算最大似然多用戶檢測(Reduced-Computation. ML-MUD,. R-ML-MUD)。第二個方法是一新型投影. 接收機(Projection Receiver)，稱為一般化線性最小均方誤差等化法(Generalized Linear Minimum Mean Square Error. Equalizer, GLMMSE)，其與傳統投影接收機(Projection. Receiver)，相比，明顯有較佳的訊噪比(Signal-to-Noise Ratio, SNR)。來自不同接取點(Access Points, APs)振盪器有著不相同的載波頻率偏移(Carrier Frequency Offset, CFO)，導致嚴重的子載波相互干擾(Inter-Carrier Interference, ICI)，訊號干擾的情況更為糟糕。運用非同步多用戶檢測(Asynchronous MUD)及子載波干擾白化技術(ICI Whitening)可獲得良好系統效能，然而白化技術需要干擾信號之共變異矩陣，對下行鏈路接收機而言，此項資訊實際上難以獲得或估測。鑑於此，我們發展一套兩階段干擾訊號抑制方法。第一階段為虛擬子載波干擾白化技術 ii. doi:10.6342/NTU201902343.

(4) (Pseudo-ICI-Whitening, P-ICI-W)，該技術不須估測子載波干擾信號之共變異矩陣，故適用於下行鏈路接收機。第二階段即為一般化線性最小均方誤差等化法(GLMMSE)，對子載波干擾做更進一步的處理。此外我們將所發展的技術運用於時空編碼信號的干擾處理，其中所考慮的時空編碼技術為Alamouti編碼和複雜交織正交設計(Complex Interleaved Orthogonal Design)，比較兩種編碼經由我們發展的信號處理後的效能。最後，我們假設虛擬細胞的存取點(APs)運用合作式頻域編碼服務自駕車。即使子載波干擾(ICI)可完美消除，載波頻率偏移(CFO)所仍會造成嚴重效能損失，我們提出一簡單的存取點索引原則(AP Indexing Principle)解決此問題。我們也改善原本的編碼技術，使得存取點(APs)在任意索引方式下，位元錯誤率(Bit Error Rate, BER)均不會明顯提升，更有效地對抗頻率偏移(CFO)問題。. 關鍵字：多用戶檢測、干擾抑制、開路通訊、車用網路、虛擬細胞、超可靠和低延遲通信、第五代移動通信技術. iii. doi:10.6342/NTU201902343.

(5) Abstract To achieve ultra-low latency mobile networking, recent efforts to integrate virtual cell with open-loop communications and proactive network association suggest the facilitation of new technological paradigm, but the interference from different co-locating virtual cells is hard to handle. Open-loop transmissions make beam-forming/interference alignment (IA) infeasible due to the need of channel state information (CSI) feedback. Multiuser detection (MUD) is therefore employed to address downlink interference. We note that the bit error rate (BER) of maximum-likelihood MUD (ML-MUD) is sensitive to the modulation of interference. As the interferer uses low-order modulation, the BER of desired signal can approach the ideal case without interference. But if the interferer adopts high-order modulation, the resultant BER is significantly degraded. Our study shows that such modulation sensitivity can be eased by multiantenna technique. We also propose two methods to reduce the notorious computational complexity of MUD, particularly involving higher-order modulations. The first scheme is termed reduced-computation ML-MUD (R-ML-MUD) that exploits the characteristic of downlink to shrink the ML solution space, consequently leading to lower detection complexity. The second scheme is a new projection receiver, called generalized linear minimum mean square error equalizer (GLMMSE) resulting in notable signal-to-noise ratio (SNR) gain over the conventional projection method. Nevertheless, losing perfect synchronization creates difficulty in tackling multiple access interference (MAI). Multiple carrier frequency offsets (CFOs) due to different oscillators at different access points (APs) incur serious inter-carrier interfer-. iv. doi:10.6342/NTU201902343.

(6) v ence (ICI) to complicate downlink MAI. Asynchronous MUD with ICI-Whitening was shown leading to satisfactory performance, but the whitening scheme needs the covariance matrix of ICI that is practically hard to obtain for downlink receivers. We therefore develop a two-stage ICI suppression method to resolve this challenge. The first-stage processing is Pseudo-ICI-Whitening (P-ICI-W), which does not rely on the estimation of ICI covariance and is suitable for asynchronous downlink. In terms of post-processing signal-to-interference-plus-noise ratio (SINR) and BER, our proposed mechanism can approach ICI-Whitening. The second-stage processing is based on GLMMSE to further cancel some ICI terms. We also apply our scheme to space-time-block-coded signals, considering Alamouti coding and Complex Interleaved Orthogonal Design. Finally, we assume that APs can coordinately allocate radio resource for the served vehicles and enforce frequency-domain cooperative data encoding. Our analysis shows that CFOs will still noticeably worsen the BER, even if ICI is well-addressed. Such problem can be resolved by indexing APs according to the order of CFOs. Furthermore, we propose a robust encoding scheme that achieves satisfactory performance and allows random AP indexing, thus CFO feedback can be avoided.. Keyword: Multiuser detection, interference suppression, open-loop communications, vehicular networks, virtual cell, uRLLC, 5G. doi:10.6342/NTU201902343.

(7) Contents. Acknowledgements. i. Abstract in Chinese. ii. Abstract. iv. Contents. viii. List of Figures. xvi. List of Tables. xvii. 1 Introduction. 1. 1.1. Ultra-Low Latency Vehicular Networking . . . . . . . . . . . . . . . .. 1. 1.2. Signal Detection Schemes. . . . . . . . . . . . . . . . . . . . . . . . .. 4. 1.2.1. Single-User Detection . . . . . . . . . . . . . . . . . . . . . . .. 5. 1.2.2. Multiuser Detection . . . . . . . . . . . . . . . . . . . . . . . .. 6. 1.2.3. ZF/LMMSE Detection . . . . . . . . . . . . . . . . . . . . . .. 8. 1.2.4. Projection Receiver . . . . . . . . . . . . . . . . . . . . . . . .. 10. 1.3. 1.4. Joint Detection to Address Interference in the Virtual Cell: Feasibility and Possible Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 11. OFDMA-Based Virtual Cell Networks. . . . . . . . . . . . . . . . . .. 13. 1.4.1. Signal Model in Perfect Synchronization . . . . . . . . . . . .. 14. 1.4.2. CFO-Induced ICI . . . . . . . . . . . . . . . . . . . . . . . . .. 16. vi. doi:10.6342/NTU201902343.

(8) CONTENTS 1.4.3 1.5. vii Challenges in Asynchronous Downlink . . . . . . . . . . . . .. 18. Organization and Contributions of Dissertation . . . . . . . . . . . .. 21. 2 Modulation Sensitivity in Multiuser Detection. 25. 2.1. Signal Model and Preliminaries . . . . . . . . . . . . . . . . . . . . .. 26. 2.2. The Impact of Modulation of Interference on BER . . . . . . . . . . .. 28. 2.3. Comparison with LMMSE . . . . . . . . . . . . . . . . . . . . . . . .. 33. 2.4. Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 34. 2.5. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 37. 3 Low-Complexity Multiuser Detection. 39. 3.1. Comparison between SUD and ML-MUD . . . . . . . . . . . . . . . .. 39. 3.2. Reduced-Computation ML-MUD . . . . . . . . . . . . . . . . . . . .. 43. 3.3. Generalized LMMSE . . . . . . . . . . . . . . . . . . . . . . . . . . .. 49. 3.4. Case Study by Simulations . . . . . . . . . . . . . . . . . . . . . . . .. 55. 3.5. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 59. 4 Two-Stage Inter-Carrier Interference Suppression 4.1. 4.2. 4.3. 4.4. 63. Asynchronous Modelling . . . . . . . . . . . . . . . . . . . . . . . . .. 64. 4.1.1. Probabilistic Analysis of TDOA-Induced ISI . . . . . . . . . .. 65. 4.1.2. Signal Model for Multi-CFO Issue . . . . . . . . . . . . . . . .. 68. First-Stage Processing . . . . . . . . . . . . . . . . . . . . . . . . . .. 71. 4.2.1. Pseudo Whitening . . . . . . . . . . . . . . . . . . . . . . . .. 72. 4.2.2. Joint Detection . . . . . . . . . . . . . . . . . . . . . . . . . .. 78. Second-Stage Processing . . . . . . . . . . . . . . . . . . . . . . . . .. 81. 4.3.1. ICI Suppression by Projection Method . . . . . . . . . . . . .. 81. 4.3.2. Compare Different ICI-Suppression Alternatives by Simulations. 82. Generalization to the Case with STBC . . . . . . . . . . . . . . . . .. 90. 4.4.1. 90. Complex Interleaved Orthogonal Design . . . . . . . . . . . .. doi:10.6342/NTU201902343.

(9) CONTENTS. 4.5. viii. 4.4.2. Alamouti Coding . . . . . . . . . . . . . . . . . . . . . . . . .. 92. 4.4.3. Performance Comparison . . . . . . . . . . . . . . . . . . . . .. 94. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 98. 5 Cooperative Coding in Frequency Domain. 99. 5.1. Cooperative Encoding and MRC . . . . . . . . . . . . . . . . . . . . . 100. 5.2. Benchmark Analysis: Asynchronous MRC . . . . . . . . . . . . . . . 106. 5.3. AP Indexing Principle . . . . . . . . . . . . . . . . . . . . . . . . . . 111. 5.4. Robust Cooperative Encoding Against CFO . . . . . . . . . . . . . . 116. 5.5. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118. 6 Conclusion. 119. 6.1. Dissertation Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 119. 6.2. Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120. Bibliography. 123. Appendix A Proof of (2.23). 133. Appendix B Proof of (2.32). 135. Appendix C Proof of (4.66). 137. Appendix D Proof of (5.26). 139. Appendix E Proof of (5.29). 141. doi:10.6342/NTU201902343.

(10) List of Figures 1.1. (Left) Each vehicle is served by multiple APs to form a virtual cell. Without central optimization of radio resource allocation, each AP may randomly select RRUs to provide downlink services, incurring cochannel interference. Even worse, such interference could also be from a HPN. (Right) Co-channel MAI in channel–1.. 1.2. . . . . . . . . . . . .. 3. Downlink MAI scenario. Without loss of generality, we look at the channel where AP–N relays the desired data of uRLLC applications from the AN to the AV, and AP–1∼AP–(N − 1) are interferers. . . .. 1.3. (Left) Each AP in the virtual cell allocates a RRU to serve a vehicle. (Right) Co-channel MAI in RRU–1.. 1.4. 5. . . . . . . . . . . . . . . . . . .. 14. The signals from different APs reach the vehicle with various delays, and ∆tn represents the discrete timing offset for AP–n’s signal relative to the first-arriving signal from some AP, say AP–1. (a) If Nch < Ncp − ∆tn , there is no ISI. (b) If Nch ≥ Ncp − ∆tn , ISI and ICI occur.. 1.5. 16. CFOs cause ICI. At the receiver’s qth FFT output, the desired data symbol from the qth subcarrier of AP–1 are coupled with the signals. 1.6. from all the subcarriers of APs–1∼3. . . . . . . . . . . . . . . . . . .. 19. Organization and contributions of dissertation.. 24. . . . . . . . . . . . .. ix. doi:10.6342/NTU201902343.

(11) LIST OF FIGURES 2.1. x. √ A2 h2 with the √ spacings of adjacent constellation points becoming ∆1 , 2 A1 r11 d1 √ and 2 A2 ∥h2 ∥2 d2 . (b) The receiver searches the ML solution from. (a) M1 and M2 are rotated and scaled by. √. A1 h1 and. a composite constellation under 2-PAM (Left) / 4-PAM (Right) interference. When x is the transmitted sequence, the most likely ere, and the distance between them is DML . The ror is decoding x as x constellation points of AP–1 act like quantization levels to quantize √ 2 A2 r12 d2 q1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2. (Left) Nrx = 2. (Right) Nrx = 3. AP–2 uses QPSK to transmit the desired data while. A2 A1. = 0 dB. The average BERs under different. modulated interference are plotted. . . . . . . . . . . . . . . . . . . . 2.3. 29. 35. (Left) Nrx = 2. (Right) Nrx = 3. AP–2 uses QPSK to transmit the desired data being interfered by AP–1’s 16-QAM modulated signal. The average BERs are plotted for. 3.1. A2 A1. = ±3 dB. . . . . . . . . . . . .. 36. f1 M1 and M2 are rotated and scaled by channel gains, becoming M f2 with the spacings of adjacent constellation points, ∆1 , and M √ √ 2 A1 ∥h1 ∥2 d1 and 2 A2 ∥h2 ∥2 d2 . MRC projects interference to span(h2 ), deviating the transmitted symbol from its original position. DSUD = √ e 1 , and I1 = 4. . . . . . . . . . . . . . . . . 2 A2 ∥h2 ∥2 d2 − (I1 − 1) ∆. 3.2. 40. The receiver searches the ML solution from a composite constellation M. When x is the transmitted sequence, the most likely error is e, and the distance between them is DML . DSUD = decoding x as x √ √ e 1 and DZF = 2 A2 r22 d2 account for the 2 A2 ∥h2 ∥2 d2 − (I1 − 1) ∆ detection performance of SUD and ZF. . . . . . . . . . . . . . . . . .. 42. doi:10.6342/NTU201902343.

(12) LIST OF FIGURES 3.3. xi. (a) The receiver treats the 4–PAM interference as being 2–PAM modulated. (b) The composite constellation space gets smaller while some extra interference is introduced to shift the transmitted sequence along ±q1 . DRML and eR,qua are defined similarly. The extra interference makes the transmitted sequence (−2d1 , d2 ) closer to or further away from the dashed purple/green line. . . . . . . . . . . . . . . . . . . .. 3.4. 44. AP–3 uses BPSK to transmit the desired signal. The average BERs of R-ML-MUD with different detection complexity are plotted. The 64-QAM interference from AP–1 is ignored or treated as being 16– QAM/QPSK modulated. Nrx = 3,. 3.5. A3 A2. = −5 dB, and. A3 A1. = 5 dB. . . .. 46. AP–3 uses BPSK to transmit the desired signal. The average BERs of R-ML-MUD with different detection complexity are plotted. The 16-QAM interference from AP–1 is ignored or treated as being QPSK modulated. Nrx = 3,. 3.6. A3 A2. = −5 dB, and. A3 A1. = 5 dB.. . . . . . . . . . .. 48. The average BERs are plotted for PR and GLMMSE. The desired signal from AP–5 is QPSK-modulated while APs–1∼4 use BPSK. Nrx = 6, and. 3.7. A5 A4. =. A5 A3. =. A5 A2. =. A5 A1. = 0 dB. . . . . . . . . . . . . . . . . . . .. 52. The average BERs are plotted for PR and GLMMSE. The desired signal from AP–5 is QPSK-modulated while APs–1∼4 use BPSK. Nrx = 6, and. 3.8. A5 A4. =. A5 A3. =. A5 A2. =. A5 A1. = 5 dB. . . . . . . . . . . . . . . . . . . .. 53. The average BERs are plotted for PR and GLMMSE. The desired signal from AP–5 is QPSK-modulated while APs–1∼4 use BPSK. Nrx = 6, and. 3.9. A5 A4. = 0 dB, and. A5 A3. =. A5 A2. =. A5 A1. = 0, 3, 6 dB. . . . . . . . . . . .. 54. The average BERs are plotted for different detection schemes. The desired signal from AP–3 is QPSK modulated while both AP–1 and AP–2 use 16-QAM. Nrx = 3,. A3 A2. = −5 dB, and. A3 A1. = 5 dB. . . . . . .. 55. doi:10.6342/NTU201902343.

(13) LIST OF FIGURES. xii. 3.10 The average BERs is plotted for different detection schemes. The desired signal from AP–4 is QPSK modulated while both AP–1/2 and A4 A3. AP–3 use 64-QAM and BPSK, respectively. Nrx = 4, A4 A2. = 3 dB, and. A4 A1. = 6 dB.. = −3 dB,. . . . . . . . . . . . . . . . . . . . . . . .. 56. 3.11 The average BERs are plotted for different detection schemes. The desired signal from AP–4 is QPSK modulated while both AP–1/2 and AP–3 use 64-QAM and BPSK, respectively. Nrx = 4, A4 A2. = −3 dB, and. A4 A1. = 6 dB.. A4 A3. = 3 dB,. . . . . . . . . . . . . . . . . . . . . . .. 57. 3.12 The average BERs are plotted for different detection schemes. The desired signal from AP–5 is QPSK modulated while both APs–1∼2 and APs–3∼4 adopt BPSK and 64-QAM, respectively. Nrx = 4, A5 A3. = 0 dB,. A5 A2. = 6 dB, and. A5 A1. A5 A4. =. = 8 dB. . . . . . . . . . . . . . . . . .. 58. 3.13 The average BERs are plotted for different detection schemes. The desired signal from AP–5 is QPSK modulated while both APs–1∼2 and APs–3∼4 adopt BPSK and 64-QAM, respectively. Nrx = 4, A5 A3. 4.1. = 0 dB,. A5 A2. = 6 dB, and. A5 A1. =. = 8 dB. . . . . . . . . . . . . . . . . .. 59. (Left) tcp = 0.59µs. (Right) tcp = 1.19µs. The probability of U(N ) − U(1) > ∆d is plotted for different N and λap . We set that tch =. 4.2. A5 A4. tcp . 2. .. 67. (Left) β = 1. (Right) β = 2. The average SINR loss of ICI-W and P-ICI-W versus CFO difference ∆ε is plotted. The parameters are set as. 4.3. A2 2 σV. =. A1 2 σV. = 10 dB, and. Gn An. = 0, ±2, ±5 dB for n = 1, 2.. . . . . . .. 75. (Left) β = 1. (Right) β = 2. In 2-AP case, the average BERs of the qth subcarrier of AP–2 are plotted w.r.t ∆ε. The parameters are set as ε2 = 0, ∆ε = ε1 − ε2 , q = 20, Nrx = 2, Gn An. = 0 dB for n = 1, 2.. A2 2 σV. =. A1 2 σV. = 13 dB, and. . . . . . . . . . . . . . . . . . . . . . . . . .. 79. doi:10.6342/NTU201902343.

(14) LIST OF FIGURES 4.4. xiii. (Left) α = 2. (Right) α = 6. In 2-AP case, the average BERs of the qth subcarrier of AP–2 are plotted w.r.t ∆ε. The parameters are set as ε2 = 0, ∆ε = ε1 − ε2 , q = 20, Nrx = 2, Gn An. 4.5. = 0, ±5 dB for n = 1, 2.. A2 2 σV. =. A1 2 σV. = 13 dB, and. . . . . . . . . . . . . . . . . . . . . . . .. 80. In 3-AP case, the average BERs of the qth subcarrier of AP–2 are plotted w.r.t ∆ε. The parameters are set as α = 10, Nrx = 3, ε2 = 0, ∆ε = ε1 −ε2 = ε2 −ε3 ,. An 2 σV. = 10 dB, and. Gn An. = 0 dB for n = 1, 2, 3. The. two-stage processing is implemented as PT-ICI-W + GLMMSE/PR. 4.6. 84. (Upper) β = 2. (Lower) β = 3. In 3-AP case, the average BERs of the qth subcarrier of AP–2 are plotted w.r.t ∆ε. The parameters are set as α = 10, Nrx = 3, ε2 = 0, ∆ε = ε1 − ε2 = ε2 − ε3 , and. Gn An. An 2 σV. = 10 dB,. = 0, ±5 dB for n = 1, 2, 3. One stage-processing is based on. PT-ICI-W, and two-stage processing is implemented as PT-ICI-W + GLMMSE. 4.7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 85. In 3-AP case, the average BERs of the qth subcarrier of AP–2 are plotted w.r.t ∆ε. The parameters are set as α = 10, Nrx = 3, ε2 = 0, ∆ε = ε1 − ε2 = ε2 − ε3 ,. An 2 σV. = 10 dB, and. Gn An. = 0 dB for n = 1, 2, 3.. The two-stage processing is implemented as PT-ICI-W + GLMMSE. 4.8. 86. In 3-AP case, the average BERs of the qth subcarrier of AP–2 are plotted w.r.t ∆ε. The parameters are set as α = 10, Nrx = 3, ε2 = 0, ∆ε = ε1 − ε2 = ε2 − ε3 ,. An 2 σV. = 10 dB, and. Gn An. = 0 dB for n = 1, 2, 3.. The two-stage processing is implemented as PT-ICI-W + PR. . . . . 4.9. 87. (Upper) β = 2. (Lower) β = 3. In 3-AP case, the average BERs of the qth subcarrier of AP–2 are plotted w.r.t ∆ε. The parameters are set as α = 10, Nrx = 3, ε2 = 0, ∆ε = ε1 − ε2 = ε2 − ε3 , dB, and. Gn An. An 2 σV. = 10. = 0, ±5 dB for n = 1, 2, 3. The two-stage processing is. implemented as PT-ICI-W + GLMMSE with Alternative–2. . . . . .. 88. doi:10.6342/NTU201902343.

(15) LIST OF FIGURES. xiv. 4.10 (Left) β = 2. (Right) β = 3. In 4-AP case, the average BERs of the qth subcarrier of AP–2 are plotted w.r.t ∆ε. The two-stage processing is implemented as PT-ICI-W + GLMMSE with α = 10 and and. Gn An. =0. dB for n = 1, 2, 3, 4. The parameters are set as , Nrx = 4, ε2 = 0.35, ε1 = 0.45, ε3 = ε2 − ∆ε, ε4 = ε2 − ∆ε − 0.02, 3 dB,. A3 A2. = −3 dB,. A4 A2. A2 2 σV. = 7 dB,. A1 A2. =. = −5 dB. Alternative–5 means suppressing. neighboring, far- and self-adjacent ICI. . . . . . . . . . . . . . . . . .. 89. 4.11 (Left) Alternative–5. (Right) Alternative–6. In 2-AP case, the average BERs of the qth subcarrier of AP–2 are plotted w.r.t ∆ε. The twostage processing is implemented as PT-ICI-W (α = 8, β = 2, and Gn An. = 0dB for n = 1, 2) + GLMMSE. Other parameters are ε2 = 0,. ∆ε = ε1 − ε2 ∈ [0, 0.5], Nrx = 2,. A2 2 σV. = 13 dB, and. A1 A2. = −3 dB.. . . .. 95. 4.12 (Upper-Left) ∆ε = 0. (Upper-Right) ∆ε = 0.2. (Lower-Left) ∆ε = 0.3. (Lower-Right) ∆ε = 0.5. In 2-AP case, the average BERs of Alternative–6 versus SNR. A2 2 σV. are plotted. The two-stage processing is. implemented as PT-ICI-W (α = 8, β = 2, and. Gn An. = 0 dB for n = 1, 2). + GLMMSE. Other parameters are ε2 = 0, ∆ε = ε1 − ε2 , Nrx = 2, and A1 A2. 5.1. = −3 dB.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 97. The serving APs in virtual cell are coordinated by the AN to allocate the same RRU for the green car and enforce cooperative transmissions. The blanks between different RRUs are intended for guard bands or pilots. Due to CFO issue, the ICI from other RRUs will interfere with green car’s data receiving from RRU–1.. 5.2. . . . . . . . . . . . . . . . . 100. Illustration of frequency domain cooperative encoding. . . . . . . . . 101. doi:10.6342/NTU201902343.

(16) LIST OF FIGURES 5.3. xv. (a) The constellations Mk ’s are rotated and scaled by heff,k ’s. Mk = {±1} for k = q, q + 1, q + 2. The green, red, and blue arrows point the directions of heff,q , heff,q+1 , and heff,q+2 . (b) The receiver searches the ML solution from the composite constellation Mq,η .. 5.4. APs–1 ∼ N respectively use only the qth ∼ (q + N − 1)th subcarriers ¯ qz . At the receiver, MRC is performed. to transmit X. 5.5. . . . . . . . . . 102. . . . . . . . . . 103. (Left) Nrx = 1. (Right) Nrx = 2. The average BER of ideal MRC and encoding scheme are plotted. A1 = A2 , η = 3, and BPSK is adopted.. 5.6. (Left) Nrx = 1. (Right) Nrx = 2. The average BER of ideal MRC and encoding scheme are plotted. A1 = A2 , η = 3, and QPSK is adopted.. 5.7. 105. 106. (Upper) For asynchronous MRC of 2-AP case, the approximated and simulated values of average BERs are plotted w.r.t ∆ε = ε2 − ε1 . The parameters are. A1 2 σV. =. A2 2 σV. = 5 dB, Nrx = 2. (Lower) The correlation. between cq (ε1 ) and cq+1 (ε2 ) is plotted w.r.t ∆ε. . . . . . . . . . . . . 109 5.8. The approximated and simulated values of average BERs are plotted for different CFO orders. The values of CFOs are all different, and ε1 , ε2 , ε3 ∈ {−0.3, 0.1, 0.45}, and other parameters are Nrx = 1, dB,. 5.9. A3 A2. A1 A2. =2. = −3 dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110. (Upper) For the cooperative encoding of 2-AP case with η = 3, the simulated values of average BERs are plotted w.r.t ∆ε = ε2 −ε1 . BPSK is adopted, and the parameters are. A1 2 σV. =. A2 2 σV. = 5 dB, Nrx = 2. (Lower). The correlations respectively between cq (ε1 ) and cq+1 (ε2 ), cq+1 (ε1 ) and cq+2 (ε2 ), cq+2 (ε1 ) and cq (ε2 ) are plotted w.r.t ∆ε . . . . . . . . . . . 111. doi:10.6342/NTU201902343.

(17) LIST OF FIGURES. xvi. ¯ z . (Right) X ¯ z . For the cooperative encod¯ z . (Middle) X 5.10 (Left) X q q+1 q+2 ing of 3-AP case with η = 3, the simulated values of average BERs are plotted. The values of CFOs are all different, and ε1 , ε2 , ε3 ∈ {−0.3, 0.1, 0.45}. BPSK is adopted, and the parameters are dB,. A3 A2. A1 A2. = 2. = −3 dB, Nrx = 1. . . . . . . . . . . . . . . . . . . . . . . . . 112. 5.11 Illustration of frequency domain cooperative encoding. The scheme in (a) requires proper AP indexing such that ε1 < ε2 < ε3 . The coding schemes in (b) and (c) are robust in the sense that the AN can index APs randomly. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 ¯ z . (Upper-Right) X ¯ z . (Lower-Left) X ¯ z . (Lower5.12 (Upper-Left) X q q+1 q+2 ¯ z . For the cooperative encoding of 3-AP case with η = 4, Right) X q+3 the simulated values of average BERs are plotted. The values of CFOs are all different, and ε1 , ε2 , ε3 ∈ {−0.3, 0.1, 0.45}. BPSK is adopted, and the parameters are. A1 A2. = 2 dB,. A3 A2. = −3 dB, Nrx = 1.. . . . . . . 115. 5.13 (Left) 2-AP case cooperative encoding with η = 3. (Right) 2-AP case ¯ z ’s are robust encoding with η = 4. the average BERs of different X k collectively evaluated. the CFOs are different with ε1 , ε2 ∈ {±0.35}. BPSK is adopted, and the parameters are A1 = A2 , Nrx = 2.. . . . . 116. 5.14 (Left) 3-AP case cooperative encoding with η = 4. (Right) 2-AP ¯ z ’s case robust encoding with η = 6. the average BERs of different X k are collectively evaluated. The values of CFOs are all different, and ε1 , ε2 , ε2 ∈ {−0.3, 0.1, 0.45}. BPSK is adopted, and the parameters are A1 A2. = 2 dB,. A3 A2. = −3 dB, Nrx = 1.. . . . . . . . . . . . . . . . . . . . 117. doi:10.6342/NTU201902343.

(18) List of Tables 1.1 The Maximum absolute value of CFO normalized w.r.t subcarrier spacing for ±20 ppm oscillator frequency mismatch . . . . . . . . . . . . .. 17. 3.1. The reduced-order constellation of 16-QAM and 64-QAM . . . . . . .. 47. 4.1. Parameters/Notations . . . . . . . . . . . . . . . . . . . . . . . . . .. 75. xvii. doi:10.6342/NTU201902343.

(19) Chapter 1 Introduction 5G networks contain three major components [1]: enhanced mobile broadband (eMBB), massive machine-type communication (mMTC), and ultra-reliable low-latency communication (uRLLC). eMBB aims at high data rates, and mMTC accommodates connectivity of massive devices to enable the Internet of Things (IoT); meanwhile uRLLC is expected to support a wide range of new services from industrial automation to augmented reality, which require both low latency and high reliability [2, 3]. Undoubtedly, due to safety concerns, vehicular networking is one of uRLLC applications with the most stringent latency and reliability restriction.. 1.1. Ultra-Low Latency Vehicular Networking. To widely deploy autonomous vehicles (AVs), the common agreement of target endto-end networking latency should be at the order of 1 ms [4, 5], far below the latency of 100 ms in 4G systems. Due to the high mobility of AVs, the required technologies can be even more challenging than Tactile Internet [6]. According to [7–10], such kind of uRLLC can be promisingly realized by introducing fog/edge computing in a heterogeneous network architecture, consisting of low-power access points (APs) to enhance spectrum efficiency and high-power nodes (HPNs) to ensure ubiquitous services. A group of APs are governed by an anchor node (AN) serving fog/edge computing to warrant real-time management and control of AVs. 1. doi:10.6342/NTU201902343.

(20) CHAPTER 1. INTRODUCTION. 2. To greatly curtail latency in radio access, the network architecture [10] revolutionarily utilizes open-loop communications together with proactive network association and anticipatory mobility management. In open-loop communications, receivers do not provide transmitters with CSI and acknowledgment messages for physical layer (PHY) transmissions to forward packets to higher layer as fast as possible [8]. For better reliability, AVs communicate with the AN through multiple APs (multiple paths). In the uplink, each AV proactively associates with appropriate APs to access and proceed transmissions [9] while the grant-free transmissions [11, 12] are carried out. The reliability can be guaranteed given that at least one of the multi-path transmissions succeeds. On the other hand, the AN with fog/edge computing can execute anticipatory mobility management to predict the APs with which each vehicle is going to associate [13]. Consequently, the downlink packets can be sent from the AN to appropriate APs and subsequently relayed to the vehicle in time. The packet size is typically small in uRLLC [14]. Regarding short-packet transmissions, the achievable capacity has been derived in [15], and [16] gives a review of recent information-theoretic works in communication with short packets. In [17], the blocklength design to minimize the decoding error probability is proposed to support uRLLC-enabled unmanned aerial vehicle (UAV) system. As to the candidate channel coding schemes including polar code, low density parity check (LDPC) code, convolutional code, etc., the comparative investigations can be found in [14, 18, 19]. In traditional wireless networks, each mobile node (a vehicle in our context) is served by one AP or base station (BS) with complicated closed-loop control signaling, and handover is performed when nodes move across the coverage boundaries of APs or BSs. Differently, in the discussed uRLLC vehicular network, each AV is served by multiple APs, which virtually form a large cell, in addition, the conventional time-consuming handover process is greatly simplified through proactive network association and anticipatory mobility management. That is, in this vehicle-. doi:10.6342/NTU201902343.

(21) CHAPTER 1. INTRODUCTION. 3. &ORXG $QFKRU 1RGH. &RUH QHWZRUN +31 . . $3. )RJ(GJH &RPSXWLQJ. . . . . . . . 85//&. . . $3. . . +LJKGDWDUDWH VHUYLFH H0%%

(22). . &KDQQHO. +31. $3. . . . $3. $3. Figure 1.1: (Left) Each vehicle is served by multiple APs to form a virtual cell. Without central optimization of radio resource allocation, each AP may randomly select RRUs to provide downlink services, incurring co-channel interference. Even worse, such interference could also be from a HPN. (Right) Co-channel MAI in channel–1. centric networking, there is only one AV in each virtual cell, and multiple APs serve this mobile node. Each AP is designated as a network and radio slice [20, 21] to this virtual cell, and simultaneously serves multiple virtual cells using other radio slices. In order to facilitate this concept, APs and subsequently the AN must run network virtualization in software defined networking (SDN) [7]. However, the holistic design of resource management remains technically challenging in uRLLC [22]. Due to the broadcast nature of wireless medium to incur interference, it seems that radio resource allocation should be globaly optimized across APs and HPNs, requiring information from AVs and other user equipments (UEs). Nevertheless, the overall processing time for global optimization plus collecting information from different network entities could be too long to satisfy the dynamic vehicle network. To ultimately save overhead and latency, each AP may just randomly se-. doi:10.6342/NTU201902343.

(23) CHAPTER 1. INTRODUCTION. 4. lect radio resource units (RRUs) to serve AVs, making multiple access interference (MAI) inevitable. From the receiver’s angle, in each of downlink channels, there will be only one AP in the virtual cell to transmit the desired signal interfered by other APs’ transmissions. Even worse, such MAI could be also from a HPN, as an inherent challenge in heterogeneous networks [23,24]. For example, in the left part of Fig. 1.1, there are two virtual cells centered at different colored cars. APs–1∼3 respectively allocate channels–1∼3 to the green car, while these APs and the HPN serve other vehicles/UEs via different RRUs to interfere with the green car’s packets receiving in channel–1, as shown in the right part of Fig. 1.1. Although beam-forming [25,26] and interference alignment (IA) [24,27] are known to suppress inter-cell interference, such schemes do not fit open-loop communications due to the need of CSI at transmitters.. 1.2. Signal Detection Schemes. As just indicated, MAI is hard to manage/suppress from network/transmitter side because the required processing time may violate the latency constraint. Thus MAI is supposed to be handled at the receiver side, and a more sophisticated receiver is desirable. In this section, some well-known detection schemes against MAI are introduced. To enhance receiving performance, the diversity technique shall be adopted. Here we assume spatial diversity by multiple receiving antennas. Suppose that the vehicle associates with N ≥ 2 APs to form a virtual cell. Using multi-path transmissions, the AN sends the packets to the vehicle through the N APs, which may randomly select radio channels to serve the vehicle. In the worst case, the vehicle receives the packets from different APs and different channels. Like the situation mentioned at Section 1.1 (Fig. 1.1), in each channel, only one AP sends the desired signal, and other APs cause co-channel interference (CCI). Without loss of generality, we look at the channel where AP–N relays the desired data of uRLLC applications from the AN to the AV, and AP–1∼AP–(N − 1) are interferers (refer. doi:10.6342/NTU201902343.

(24) CHAPTER 1. INTRODUCTION. 5. )RJ(GJH &RPSXWLQJ. $QFKRU1RGH. $3. $3. $3. Figure 1.2: Downlink MAI scenario. Without loss of generality, we look at the channel where AP–N relays the desired data of uRLLC applications from the AN to the AV, and AP–1∼AP–(N − 1) are interferers. to Fig. 1.2). For n = 1, · · · , N , Xn and Mn represent the transmitted symbol and [ ] signal constellation of AP–n, and E |Xn |2 = 1. Assume that each AP is equipped with a single antenna, and the vehicle is equipped with Nrx antennas for Nrx ≥ N , then the received signal y is N √ ∑ y= An hn Xn + v.. (1.1). n=1. An is AP–n’s mean signal power received at the vehicle, hn = [H1n . . . HNrx n ]T is the fading gain from AP–n to the vehicle, and v = [V1 . . . VNrx ]T is the white noise at the receiving antennas with v ∼ CN (0, σV2 INrx ). Next, we give a brief summary of conventional detection schemes in the following subsections.. 1.2.1. Single-User Detection. A single-user detection for XN treats the interference term. ∑N −1 √ An hn Xn in (1.1) n=1. as noise, and estimate XN according to. 2 √. b XN = arg min y − AN hN xN . xN ∈MN. (1.2). 2. The detection works well only when the interfering signals from other APs are much weaker than the desired signal from AP–N . If the interference from some AP, say. doi:10.6342/NTU201902343.

(25) CHAPTER 1. INTRODUCTION. 6. AP–1, is stronger than the desired signal, the receiver can decode X1 first, subtract √ b1 from y, and then decode XN by A1 h1 X. 2 √ √. b b XN = arg min y − A1 h1 X1 − AN hN xN ,. (1.3). 2 √. b X1 = arg min y − A1 h1 x1 .. (1.4). xN ∈MN. x1 ∈M1. 2. 2. Without losing generality, we assume that the interference from APs–1∼ K is weaker than AP–N ’s signal, can be estimated according to the descending signal power order A1 > A2 > · · · > AK by (1.4) and (1.5).. 2 n−1. ∑ √ √. bn = arg min b X y − A h X − A h x for n = 2, . . . , K.. g g g n n n. xn ∈Mn g=1. (1.5). 2. Then the estimate of XN is obtained via. 2 K √. ∑ √. b b XN = arg min y − An hn Xn − AN hN xN .. xN ∈MN n=1. (1.6). 2. The method described in (1.4)∼(1.6) is called successive interference cancellation (SIC), its effectiveness heavily depends on the significant differences in signal strengths.. 1.2.2. Multiuser Detection. In MUD, the receiver may adopt individually optimum or jointly optimum strategies [28]. The individually optimum decision is made according to maximum a posterior { } bN ̸= XN . (MAP) rule given in (1.7) to minimize the error probability Pr X. 2  N √. ∑ y −  − A h x n n n   ∑  n=1 2 Pr {X1 = x1 , . . . , XN = xn } exp  2 σV   ∏N −1 . bN = arg min X xN ∈MN. (x1 ,...,xN −1 )∈. n=1. Mn. 2  N √. ∑.  y − − A h x n n n    n=1 2 exp . σV2   . = arg min xN ∈MN. ∑ (x1 ,...,xN −1 )∈. ∏N −1 n=1. Mn. (1.7). (1.8). doi:10.6342/NTU201902343.

(26) CHAPTER 1. INTRODUCTION. 7. The equality in (1.8) holds when the constellation points xn in Mn are equiprobable for any n. On the other hand, the jointly optimum decisions are the maximum-likelihood ( ) b b (ML) decisions X1 , . . . , XN obtained by (. ). b1 , . . . , X bN = X. arg min (x1 ,...,xN )∈. N ∏. Mn. 2. N √. ∑. An hn Xn . y −. n=1. (1.9). 2. n=1. The operation of (1.9) is also called maximum likelihood sequence estimation (MLSE) or ML-MUD. In (1.8), if the SNR is not extremely low, the sum is dominated by the largest term, the following approximation can be used [29, 30].  2  N √. ∑  y − A h x − n n n   ∑  n=1 2 exp  σV2   ∏N −1 (x1 ,...,xN −1 )∈. n=1. Mn. 2 . N √. ∑. An hn xn  − y −   n=1 2 ≈ max∏ exp , 2 N −1 σ   (x1 ,...,xN −1 )∈ n=1 Mn V . (1.10). whereby (1.8) is simplified as (1.9), both type of decisions will agree with very high probability [28], and the difference between their performance is small, particularly at high SNR [29]. SIC and MUD require the knowledge of modulation format of interferers and interfering channel gains. The modulation information is usually encoded in the preamble, so it can be decodable at the receiver even though the signals from APs– 1 ∼ (N − 1) are not intended for the vehicle in the discussed scenario (Fig. 1.2), and the channel estimation is made possible by appropriate placement of pilots.. doi:10.6342/NTU201902343.

(27) CHAPTER 1. INTRODUCTION. 1.2.3. 8. ZF/LMMSE Detection. The complexity of MUD grows exponentially with the number of jointly detected symbols N . In this subsection, we introduce two suboptimal detectors with the complexity that grows linearly with N , and the method of SIC derived from them. 1) Zero-forcing (ZF) detection [√ ] √ Let HA = A1 h1 · · · An hN and x = [X1 · · · XN ]T . The received signal y in (1.1) is rewritten as y = HA x + v.. (1.11). Using ZF [31, 32] to entirely eliminate MAI, the received signal is pre-multiplying by the Moore-Penrose pseudo inverse of the channel matrix HA , leading to ( )−1 H ( )−1 H e = HH HA v. HA y = x + HH y A HA A HA. (1.12). e by YeN , the desired data symbol is detected via Denote the N th component of y

(28)

(29) 2 bN = arg min

(30)

(31) YeN − XN

(32)

(33) , X. (1.13). e = wH YeN = eH Ny ZF y. (1.14). wZF = HA C−1 eN ,. (1.15). C = HH A HA .. (1.16). XN ∈MN. where. In the above equations, eN is a N × 1 vector with the N th entry replaced with one. 2) Linear minimum mean square error (LMMSE) detection The LMMSE detection [31–33] is to estimate the desired signal by linearly combining the weighted versions of received signals. The weighting vector is derived based. doi:10.6342/NTU201902343.

(34) CHAPTER 1. INTRODUCTION. 9. on mean square error minimization (1.17) or SINR maximization (1.18) criterion. [

(35)

(36) 2 ] √ wMMSE = arg min = E

(37) wH y − XN

(38) = AN C−1 (1.17) y hN , w

(39) 2

(40) √ AN

(41) wH hN

(42) ] = wMaxSINR arg max [

(43) AN C−1 (1.18) (∑ )

(44) 2 I hN . √

(45) H

(46) w N −1 E

(47) w An hn Xn + v

(48) n=1 where [ ] Cy = E yyH = CI + AN hN hH N, CI =. N −1 ∑. 2 An hn hH n + σV INrx. (1.19) (1.20). n=1. Using the matrix inversion lemma [28], √ ( )−1 AN CI + AN hN hH hN N ( H −1 ) √ C−1 −1 I hN hN CI = AN CI − −1 hN −1 AN + hH N CI hN ( ) H −1 √ C−1 −1 I hN hN CI hN = AN CI hN − −1 −1 AN + hH N CI hN ( ) √ A−1 N = AN C−1 I hN H −1 −1 AN + hN CI hN ) ( A−1 N wMaxSINR , = H −1 A−1 N + hN CI hN. wMMSE =. (1.21). which shows that wMMSE and wMaxSINR are proportional to each other. Since multiplicative constants in weighting vectors do not change the resultant SINR, the two criteria are equivalent. Thus, the LMMSE detection is to execute

(49) ( )

(50) 2 √

(51) bN = arg min

(52)

(53) wH y − X A h x N N N

(54) . MMSE xN ∈MN. (1.22). 3) LMMSE/ZF-Successive interference cancellation (LMMSE/ZF-SIC) The receiver applies successive interference cancellation to recover the streams (interfering signals or desired signals) one by one. At each step, the receiver estimates one stream according ZF or LMMSE detection, then subtracts the estimated component from the received signal. For example, after ℓ steps, the received signal is bℓ = y − y. ℓ ∑ √ b ns , Ans hns X. (1.23). s=1. doi:10.6342/NTU201902343.

(55) CHAPTER 1. INTRODUCTION. 10. where ns is the index of cancelled signal at the sth step. In other words, this type of detector tries to generate an identical copy of MAI by ZF/LMMSE-estimate of Xn ’s from interfering APs and cancel the associated interference successively until the desired symbol is detected.. 1.2.4. Projection Receiver. Multiuser projection receiver (PR) [35, 36] partially eliminates MAI, then jointly detects the remaining interference and desired signal in the following manner. ( ) 2 N. ( ) ∑ √. bK+1 , . . . , X bN = X arg min An hn xn , P y −. N ∏ (xK+1 ,...,xN )∈. Mn. n=K+1. (1.24). 2. n=K+1. where ( )−1 H P = INrx − HI HH HI , I HI [√ ] √ HI = A1 h1 · · · AK hK .. (1.25) (1.26). The operation of (1.24) means that the receiver jointly decodes XK+1 , . . . , XN after cancelling the interfering signals from APs–1 ∼ K. When K = N − 1 in (1.26), we have [. ] [ ] [( √ √ ) ( ) ]T −1 wZF = HI AN hN C eN = HI AN h N C−1 1N · · · C−1 N N , (1.27) [√ ] √ HI = A1 h1 · · · AK hN −1 , (1.28) [ ] √ HH AN HH H I hN I HI √ C = HA HA = . (1.29) AN hH AN hH N HI N hN By the block-wise matrix inversion formula [34], [(. C−1. ). ]T √ ( )−1 H ( ) ( −1 ) AN HH · · · C = − HI hN C−1 N N . I HI 1N (N −1)N. (1.30). Substituting (1.30) into (1.27) leads to (. wZF = C. −1. ) NN. (. INrx − HI. (. )−1 HH I HI. HH I. ) ) (√ AN hN .. (1.31). From (1.31) and (1.25), ZF equalizer can be deemed as a special case of PR.. doi:10.6342/NTU201902343.

(56) CHAPTER 1. INTRODUCTION. 1.3. 11. Joint Detection to Address Interference in the Virtual Cell: Feasibility and Possible Issues. Please recall that ultra-low latency is achieved mainly by system architecture innovation, networking protocols simplification, etc. As a matter of fact, the significant gain in end-to-end networking latency does not come from signal detection in PHY; however, proactive and open-loop communication contributes such latency deduction but also creates a new technology challenge to reliably detect and demodulate signals [10]. In other words, the concern “how to receive the low-latency data transmissions reliably?” naturally arises in this innovative cross-layer system design, where we adopt MUD to achieve the best possible BER performance that is closely related to “reliability”. MUD, the joint detection of multiple simultaneous transmissions, has been long investigated in code division multiple access (CDMA) [28, 30] and multiple-input multiple-output (MIMO) communications [32]. In CDMA, ML-MUD can effectively overcome the near-far problem [37], in which a near user causes excess interference to a far user. But the detection complexity grows exponentially as the number of involved users increases, as a major obstacle to practical implementation. Likewise, MIMO receivers suffer from high detection complexity for high-order modulation. However, in our scenario, the number of neighboring APs is limited. In addition, high-order modulation might not be necessary since the ultimate goal of our system design is to achieve ultra-low latency rather than pursuing high data rate [14]. In other words, we care about uRLLC traffic instead of eMBB traffic in this research. High instantaneous data rate does not necessarily imply good uRLLC throughput if the time for running networking protocols is considered. A great amount of latency can be saved in layer-1 and layer-2 via proactive open-loop communications, the average throughput can possibly meet the requirement of uRLLC by low-order modulation, say BPSK or QPSK.. doi:10.6342/NTU201902343.

(57) CHAPTER 1. INTRODUCTION. 12. Unfortunately, owing to lacking sufficient processing time for global radio resource allocation, some high-order modulated signals generated from eMBB and other traffic can still interfere with uRLLC in the PHY of air interface (see Fig. 1.1) [14, 21]. The threat from MAI particularly of high-order modulation is twofold: degrading BER performance more and increasing detection complexity. By ML-MUD, it is possible to approach the ideal detection performance as if no interferer was present, but the BER is very sensitive to the modulation schemes of interferers. It is well-known that MLMUD in CDMA and general multiuser communication scenarios is “asymptotically” near-far resistant and thus insensitive to interference power [37]. Nonetheless, in our context that is not possible to enjoy the asymptotic behavior, ML-MUD is found still suffering from the sensitivity to the modulations of interfering signals that is seldom discussed in the traditional MUD literatures. Interferers using high-order modulations obviously further deteriorate the joint detection especially for comparable received power of desired signal and interference at the receiver. Luckily, we discover that such sensitivity can be easily relieved by multi-antenna technique [38, 39]. When the interfering signal with high-order modulation has weaker received power than the desired signal (due to path loss and/or shadowing), the receiver may just ignore it, but the BER will be poor if the interference is not weaker enough. MLMUD suffers from high-complexity concern, and ignoring the interference may lead to bad performance. This dilemma suggests something in-between. Different from typical uplink MUD, the purpose of downlink MUD is to counter MAI, not to correctly decode all the received signals from designated transmitter and interferers. Instead of disregarding interference, we exploit this downlink distinction to propose a unique realization of MUD, R-ML-MUD, treating high-order modulated interfering signal as being lower-order modulated (which is equivalent to partially ignoring weak interference), whereby the ML solution space shrinks substantially [38, 40]. The resulting BER can be acceptable even when LMMSE does not ideally function. Although the. doi:10.6342/NTU201902343.

(58) CHAPTER 1. INTRODUCTION. 13. complexity still grows exponentially with the increased number of interfering sources, R-ML-MUD can practically lower the computation load down to a reasonable level because of the limited number of neighboring APs, particularly if smart arrangement of RRUs in use. To entirely avoid the high complexity of ML-MUD, using ZF/LMMSE to cancel all the interference is common, but leads to diversity loss [31,41]. Conventional projection receiver (PR) [35, 36] to cancel just a portion of interference serves a compromise between complexity and performance. The received signal is projected towards the orthogonal complement of subspace spanned by some portion of interfering signals, which can be totally removed. Nonetheless, the energy of desired signal could be lost a lot after projection. Towards mitigating the drawback, we propose a generalization of LMMSE (GLMMSE) [38] as a new type of projection detection structure to find a subspace where there is some residue of suppressed interference after projection, but more amount of energy will be retained in the desired signal. Compared to the conventional PR, GLMMSE has noticeable SNR gain for multiple weaker interfering signals being suppressed.. 1.4. OFDMA-Based Virtual Cell Networks. Our research is conducted based on Orthogonal Frequency Division Multiple Access (OFDMA), which has been widely used in many applications such as wireless local area Networks (WLAN), The Third Generation Partnership Project LongTerm-Evolution (3GPP-LTE). In OFDMA-based virtual cell networks, the scenario in Fig. 1.1 can be depicted as Fig. 1.3, where each AP allocates a RRU composed of several resource elements (REs) for the served vehicle, and each RE is 1 subcarrier × 1 OFDM symbol. In this section, we introduce the signal model of OFDM [42, 43], and account for the challenge in asynchronous downlink.. doi:10.6342/NTU201902343.

(59) CHAPTER 1. INTRODUCTION. 14. ''( ''( ''( ''(.

(60)

(61) !" #$ . .

(62) . . . . . . )

(63)

(64) ''( . . . %!

(65)

(66) &" $. ''( ''( ''(. . . . .

(67) . Figure 1.3: (Left) Each AP in the virtual cell allocates a RRU to serve a vehicle. (Right) Co-channel MAI in RRU–1.. 1.4.1. Signal Model in Perfect Synchronization. The discrete time-domain transmitted signal from AP–n during the zth OFDM symbol duration is xzn,ℓ. Nfft −1 j2πk(ℓ−Ncp −(z−1)(Nfft +Ncp )) 1 ∑ z Nfft = , (z − 1) (Nfft + Ncp ) ≤ ℓ ≤ z (Nfft + Ncp ) , Xn,k e Nfft k=0 (1.32). z where ℓ is the sample index of time-domain signal, Xn,k is the kth subcarrier data. symbol of AP–n, Nfft and Ncp are the size of fast Fourier transform (FFT) and cyclic prefix (CP) length. CP is a copy of the last Ncp samples appended in front of each OFDM symbol. The channel impulse response (CIR) between AP–n and the vehicle’s mth antenna is denoted by hm n,τ , which is assumed to be time-invariant over several OFDM symbol periods in uRLLC, where the latency requirement is typically shorter than the channel coherence time [22, 44]. In addition, the CIR length Nch is smaller than CP length, i.e. Nch < Ncp . The signals from different APs reach the vehicle with different propagation delays,. doi:10.6342/NTU201902343.

(68) CHAPTER 1. INTRODUCTION. 15. and ∆tn represents the discrete timing offset from AP–n’s signal relative to the first arriving signal from some AP, say AP–1. For any n, suppose that Nch < Ncp − ∆tn and the receiver’s FFT window is precisely aligned with AP–1’s OFDM symbols, as shown in Fig. 1.4(a). At the qth FFT output, AP–n’s signal received at the vehicle’s mth antenna is m Yz,n,q =. N∑ fft −1. [N −1 ch ∑. ] z hm n,τ xn,ℓ−τ −∆tn e. τ =0 ℓ=0 [ −1 N∑ fft −1 N∑ ch. −j2πqℓ Nfft. (. )] N∑ fft −1 j2πk(ℓ−τ −∆tn ) −j2πqℓ 1 z Nfft Nfft e X e = hm n,k n,τ N fft τ =0 k=0 ℓ=0 [ (N −1 ) ] N∑ −1 N∑ fft ch fft −1 ∑ j2π(k−q)ℓ −j2πk∆tn −j2πkτ 1 z Nfft Xn,k e Nfft e Nfft = hm n,τ e N fft τ =0 k=0 ℓ=0 ) ( N∑ N∑ fft −1 fft −1 j2π(k−q)ℓ 1 m z m z = Hn,q Xn,q , e Nfft = Hn,k Xn,k N fft ℓ=0 k=0 where m Hn,k. =e. −j2πk∆tn Nfft. N∑ ch −1. hm n,τ e. −j2πkτ Nfft. .. (1.33). (1.34). τ =0. Taking the mean received signal power An into account, the qth FFT outputs across different receiving antennas can be expressed as yz,q =. N √ ∑. z An hn,q Xn,q + vq ,. (1.35). n=1. [ 1 ] Nrx T where hn,q = Hn,q · · · Hn,q and vq is the noise term. Without loss of generality, we can consider only one OFDM symbol and one FFT output. Dropping the indices m = Hmn , we obtain the signal model in (1.1). z, q in (1.35) and letting Hn,q. However, if the time differences of arrival (TDOA) is not small enough such that Nch ≥ Ncp − ∆tn (see Fig. 1.4(b)) to cause inter-symbol interference (ISI) and intercarrier interference (ICI) [45], the signal model of (1.33) and (1.35) will not be applicable. Whether such propagation delay-induced ISI/ICI occurs depends on the virtual cell size (the number of serving AP in a virtual cell), AP density, and the CP length. Using the well-known stochastic geometry modeling [46], we conduct a probabilistic. doi:10.6342/NTU201902343.

(69) CHAPTER 1. INTRODUCTION. 16. ))7:LQGRZ $3 &3. 7KH V\PERO. WK 2)'0. &3. WK 2)'0. 7KH V\PERO. 7KH WK 2)'0 V\PERO 7KH WK 2)'0 V\PERO. $3. D

(70) . ))7:LQGRZ $3 &3. 7KH V\PERO. WK 2)'0. 7KH V\PERO. &3. WK 2)'0. 7KH WK 2)'0 V\PERO 7KH WK 2)'0 V\PERO. $3. E

(71) . Figure 1.4: The signals from different APs reach the vehicle with various delays, and ∆tn represents the discrete timing offset for AP–n’s signal relative to the firstarriving signal from some AP, say AP–1. (a) If Nch < Ncp − ∆tn , there is no ISI. (b) If Nch ≥ Ncp − ∆tn , ISI and ICI occur. analysis in Section 4.1.1, showing that the occurrence probability of TDOA-related ISI/ICI can be little in small cell deployment if the CP length and the size of virtual cell are properly designed.. 1.4.2. CFO-Induced ICI. Refer to Fig. 1.4(a) and consider carrier frequency offset (CFO) resulting from oscillator mismatch to incur ICI. Let εn stand for the normalized CFO (w.r.t the subcarrier spacing) between the vehicle and AP–n. If the oscillator precision tolerance is specified to be less than ±20 ppm, the CFO ranges between −40 ppm and +40 ppm. For different subcarrier spacings and carrier frequencies, the maximum absolute values of normalized CFO |εn | are listed in Table 1.1. ICI is caused by the fractional part of. doi:10.6342/NTU201902343.

(72) CHAPTER 1. INTRODUCTION. 17. Table 1.1: The Maximum absolute value of CFO normalized w.r.t subcarrier spacing for ±20 ppm oscillator frequency mismatch Sucarrier spacing. 15 kHz. 30 kHz. 60 kHz. 120 kHz. Carrier frequency 1.9 GHz. 5.0667. 2.5333. 1.2667. 0.6333. Carrier frequency 2.4 GHz. 6.4. 3.2. 1.6. 0.8. Carrier frequency 5.9 GHz. 15.7333. 7.8667. 3.9333. 1.9667. m CFO, thus we simply assume that |εn | ≤ 0.5, and derive the expression of Yz,n,q in. the following. m Yz,n,q =. √. An. N∑ fft −1. [(N −1 ch ∑. ) z hm n,τ xn,ℓ−τ −∆tn. e. j2πεn [(z−1)(Nfft +Ncp )+Ncp +ℓ] Nfft. ] e. −j2πqℓ Nfft. τ =0. ℓ=0. ) ] [N −1 ( N∑ ch fft −1 ∑ j2πk(ℓ−τ −∆tn ) j2πεn ℓ −j2πqℓ 1 z Nfft e Nfft e Nfft = ej((z−1)ϕn +ρn ) An Xn,k e hm n,τ Nfft k=0 τ =0 ℓ=0 ) ] [ (N −1 N∑ ch fft −1 fft −1 ∑ √ N∑ j2π(k−q+εn )ℓ −j2πkτ 1 j((z−1)ϕn +ρn ) z −jγn,k m Nfft =e An Xn,k e hn,τ e Nfft e N fft τ =0 k=0 ℓ=0 ( ) N∑ N∑ fft −1 √ fft −1 j2π(k−q+εn )ℓ 1 m z Nfft = ej(z−1)ϕn e An Hn,k Xn,k (1.36) N fft k=0 ℓ=0 jπεn (Nfft −1) sin(πεn ) √ m z ( ) An Hn,q = ej(z−1)ϕn e Nfft Xn,q πεn Nfft sin Nfft ∑ jπ(k−q+εn )(Nfft −1) sin(π (k − q + εn )) √ m z Nfft ) An Hn,k ( e + ej(z−1)ϕn Xn,k (1.37) π(k−q+εn ) N sin k̸=q fft Nfft √. with ϕn =. 2π(Nfft +Ncp )εn , Nfft. N∑ fft −1. ρn = m Hn,k. 2πNcp εn , Nfft. =e. γn,k =. j (ρn −γn,k ). 2πk∆tn , Nfft. N∑ ch −1. and. hm n,τ e. −j2πkτ Nfft. .. τ =0. The second term of (1.37) denotes the ICI, namely the signals from other subcarriers. From the first term of (1.37), we also see that another effect of CFO is to cause attenuation in magnitude, phase shift of the signal conveyed by the qth subcarrier. Define the CFO matrix as C(εn ) = FD(εn ) FH. (1.38). doi:10.6342/NTU201902343.

(73) CHAPTER 1. INTRODUCTION. 18. ( ) j 2π ε j 2π (N −1)εn where D(εn ) = diag 1, e Nfft n , . . . , e Nfft fft , and F is the Nfft × Nfft unitary DFT matrix with (F)ql =. √1 e Nfft. −j2πql Nfft. . For the matrices C(εn ) , F, and D(εn ), the. row/column/entry indices begin with zero. The (q, k) entry of C(εn ) is (C(εn ))qk =. N∑ fft −1. N∑ N∑ fft −1 fft −1 [ ] ( ) H (F)qs (D(εn ))sl FH lk (F)qs D(εn ) F sk = s=0. s=0. =. 1 Nfft. N∑ fft −1. e. −j2πqs Nfft. e. j2πεn s Nfft. e. j2πks Nfft. =. s=0. 1 Nfft. l=0 N∑ fft −1. e. j2π(k−q+εn )ℓ Nfft. .. (1.39). ℓ=0. By (1.36) and (1.39), AP–n’s frequency-domain signal received from the vehicle’s mth antenna can be expressed as N∑ fft −1 √ [ m ]T m j(z−1)ϕn m z Yz,n,0 · · · Yz,n,Nfft −1 = e An Hn,k Xn,k ck (εn ) .. (1.40). k=0. In (1.40), ck (εn ) is the kth column of C(εn ), which is also known as the signature waveform of AP–n’s kth subcarrier. Therefore, the overall frequency-domain received signal can be written as ym z,F. =. N N∑ fft −1 ∑. ej(z−1)ϕn. √ m z An Hn,k Xn,k ck (εn ) + vm z,F .. (1.41). n=1 k=0. Here vm z,F is the noise term.. 1.4.3. Challenges in Asynchronous Downlink. Go back to Fig. 1.3, and let q denote the index of subcarrier on which the desired data is transmitted from AP–1 to the green car. If synchronization is ideally achieved, at the qth output of receiver’s FFT module, only the signals from different APs’ qth subcarriers are superimposed without any interference from other subcarriers. The receiver only needs to deal with co-channel interference (CCI) by performing MUD or other interference cancellation schemes individually on each subcarrier. Unfortunately, ICI rising from CFO makes the downlink MAI in Fig. 1.3 more disastrous, as portrayed in Fig. 1.5, where we depict the superimposition of CCI and ICI on the desired signal at the receiver’s qth FFT output. Different APs have different. doi:10.6342/NTU201902343.

(74) CHAPTER 1. INTRODUCTION 558. 558. $3. 19 558. 558. $3. 7LPHGRPDLQ 2)'0V\PERO

(75) )UHTXHQF\GRPDLQ 6XEFDUULHU

(76). 7KH WK ))7RXWSXW. )UHTXHQF\GRPDLQUHFHLYHGVLJQDO 'HVLUHGVLJQDO &RFKDQQHO ,QWHUIHUHQFH &&,

(77). $3. ,&, 558. 558. 558. Figure 1.5: CFOs cause ICI. At the receiver’s qth FFT output, the desired data symbol from the qth subcarrier of AP–1 are coupled with the signals from all the subcarriers of APs–1∼3. oscillators, resulting in various CFOs that may not be simultaneously compensable at the vehicle even though their estimates are available by the schemes in [47–49] and the references therein. Resolving this issue at transmitter sides, i.e. APs might be infeasible because each AP probably belongs to multiple virtual cells, it is hard to find a frequency offset value to pre-compensate that is universally suitable for all vehicles in service. Multi-CFO problem inherently exists in OFDMA uplink [50–55] and cooperative communications [45, 56–61] with a brief summary of recent works given in the following. Reference [51] formulates a tri-linear signal model, whereby a subspace nulling approach is invented to address CFO-induced MAI in uplink MIMO-OFDM systems. For massive MIMO uplink, [54] proposes an angle-domain adaptive filtering scheme, in which the beamformers for each user are constructed to enable separate CFO estimation and data detection. In a downlink coordinated multi-point (CoMP) trans-. doi:10.6342/NTU201902343.

(78) CHAPTER 1. INTRODUCTION. 20. mission scenario, [56] proposes a frequency-domain data-encoding scheme by base stations (BSs), and derive a suboptimal CFO compensation value for the receiver to improve SINR. In a relay network, [58] designs a cooperative Alamouti-coding scheme, which nearly achieves full diversity under the consideration of oscillator frequency offsets and Doppler shifts. In spatial modulation OFDM system with multiple CFOs, a symbol-by-symbol-aided iterative detector and a LMMSE-based detector are developed in [62] and [63]. In terms of BER, the optimal countermeasure against ICI is to jointly detect the signals on all the subcarriers, but the complexity will be extremely high, not practical to implement. A reasonable compromise for the detection should include only the dominant ICI terms that are the signals on several nearest neighboring subcarriers from the one conveying the desired data. However, there will be an irreducible error floor incurred by the far (or residual) ICI, which is supposed to be whitened before signal detection to improve the performance [64]. Like general LMMSE equalization schemes [45, 55, 57, 59–61, 63], ICI-Whitening (ICI-W) must rely on the knowledge of ICI covariance matrix whose estimate is based on the channel conditions of the subcarriers, from which the residual ICI originates. In proactive uplink, it is normal for the receiver to perform channel estimation on all the subcarriers, the whitening scheme is practical. Nonetheless, in proactive open-loop downlink communications, it may be unreasonable and very difficult to estimate channel gains on all the subcarriers that do not belong to the radio slices allocated for the receiver, making ICI-W infeasible. Such a problem unique to downlink receiver implementation about network/resource slicing remains unknown in the existing literatures.. doi:10.6342/NTU201902343.

(79) CHAPTER 1. INTRODUCTION. 1.5. 21. Organization and Contributions of Dissertation. Our objective is to investigate the transceiver design for ultra-low latency virtual-cell network, and the research is conducted in the following steps: 1) Starting with the perfect synchronization assumption (i.e. no ISI/ICI), where the serving APs allocate the RRUs randomly, and the vehicle decodes the desired data from different paths (APs) separately (refer to Fig. 1.1 and Fig. 1.2), we look at how high-order modulated interference affects MUD performance, which is relevant to the coexistence of uRLLC and eMBB services. Two methods of lowering the MUD complexity in this scenario are also proposed. 2) After that, a more complicated situation with multi-CFO problem (Fig. 1.5) is considered. We investigate how to practically address ICI and perform MUD in downlink. 3) Finally, we discuss the design of APs’ cooperative transmissions that is made possible by the AN’s coordination, and the multi-CFO problem is still taken into consideration. The organization and contributions of this dissertation are sketched below (and also summarized in Fig. 1.6). • Chapter 2: The modulation sensitivity problem of MUD is analyzed under the scenario of Fig. 1.2. We discovery that the BER performance of ML-MUD is very sensitivity to the modulations of interfering signals, and such sensitivity can be lessened by increasing antennas of receiver. The comparison of MLMUD and LMMSE/ZE detection is also given. In addition to mathematical derivation and simulation study, we provide an easily understandable way to view the distinction between different detection schemes. • Chapter 3: This chapter proposes two new detection schemes based on the. doi:10.6342/NTU201902343.

(80) CHAPTER 1. INTRODUCTION. 22. model of Chapter 2. Different than typical uplink MUD, the purpose of downlink MUD is to combat against MAI, but not necessary to decode all the received signals from the designated transmitter and interferers, and thus the detection performance of packets from interfering APs is not worth attention. We therefore utilize this feature of downlink transmissions to propose a unique realization of MUD, R-ML-MUD to reduce the complexity of detection, which is different from traditional sphere decoding [65, 66] and deemed as a method scheme to “partially ignore weak interference”. Furthermore, we note the drawback of conventional multiuser projection receiver [35, 36] to propose GLMMSE as a new type of projection detection structure that partially suppresses interference. After that, we strike flexible trade-off between complexity and performance by combining the two proposed methods. • Chapter 4: Taking multi-CFO problem into account, we construct a two-stage ICI suppression method that is more practical to downlink receiver in the scenario described in Fig. 1.5. The first-stage processing is Pseudo-ICI-Whitening (P-ICI-W) [67], which does not rely on the estimation of ICI covariance and is thus suitable for asynchronous downlink. The proposed mechanism is shown to approach ICI-Whitening in terms of post-processing SINR and BER. The second-stage processing is to further cancel some ICI terms by GLMMSE developed in Chapter 3. Moreover, our proposed scheme is compatible with spacetime-block-coded signals, namely Alamouti [68] coding and Complex Interleaved Orthogonal Design (CIOD) [69] to yield more reliable proactive wireless communications. • Chapter 5: We shift our focus onto transmission design according to the asynchronous signal model used in Chapter 4. Under the coordination of AN, the serving APs of virtual cell are assumed to be able to cooperatively encode trans-. doi:10.6342/NTU201902343.

(81) CHAPTER 1. INTRODUCTION. 23. mitted data symbols across space-frequency domain. Our analysis reveals that the multi-CFO issue will still result in serious performance loss, even though ICI is perfectly addressed or does not exist. Such issue can be easily resolved by our AP indexing principle. Moreover, we propose a robust encoding scheme against CFOs that not only achieve better performance but also allow AN to index APs randomly, eliminating the necessity of CFO feedback. • Chapter 6: This chapter is devoted to the concluding remark. Additionally, some future works towards developing virtual-cell networks are also suggested.. doi:10.6342/NTU201902343.

(82) CHAPTER 1. INTRODUCTION. 24. 2)'0$%DVHG9LUWXDO&HOO 7UDQVFHLYHU'HVLJQ 1RQ&RRSHUDWLYH 7UDQVPLVVLRQV &KDSWHU. 0RGXODWLRQ6HQVLWLYLW\3UREOHP 08'SURSHUW\DQDO\VLV

(83). &RRSHUDWLYH 7UDQVPLVVLRQV 7UDQVPLVVLRQ'HVLJQ &KDSWHU $3,QGH[LQJ3ULQFLSOH. 5HFHLYHU'HVLJQ &KDSWHU. 5REXVW&RRSHUDWLYH(QFRGLQJ &RQVLGHUPXOWL&)2LVVXHV

(84). 50/08' 3DUWLDOO\LJQRUHLQWHUIHUHQFH

(85) */006( 3DUWLDOO\VXSSUHVVLQWHUIHUHQFH

(86) $VVXPHSHUIHFWV\QFKURQL]DWLRQ

(87). &KDSWHU )XWXUH:RUN. &RQVLGHUPXOWL&)2 LVVXHV. &KDSWHU 7ZR6WDJH,&,6XSSUHVVLRQ 3VHXGR:KLWHQLQJ*/006(

(88) . Figure 1.6: Organization and contributions of dissertation.. doi:10.6342/NTU201902343.

(89) Chapter 2 Modulation Sensitivity in Multiuser Detection In this chapter, we discuss how MUD works in our virtual cell-based vehicular networks. The analysis is conducted based on the situation of Fig. 1.2 with perfect synchronization. For better readability, the scenario is re-stated. The vehicle associates with N ≥ 2 APs to form a virtual cell. In multi-path transmissions, the AN sends the packets to the vehicle through the N APs, which randomly choose radio channels to serve the vehicle. In the worst case, the vehicle receives the packets from different APs and different channels. Without loss of generality, we look at the channel (or subcarrier ) where AP–N relays the desired data from the AN to the AV, and AP–1∼AP–(N − 1) are interferers. As a MAI countermeasure, we analyze the effectiveness of ML-MUD in terms of uncoded BER, comparing it with ZF/LMMSE. In traditional MIMO systems where the spatial streams are usually assumed to convey the data symbols with the same modulation format, and the detection performance is evaluated across the data symbols from all the transmit antennas. However, in our downlink scenario of Fig. 1.2, only the detection performance of AP–N ’s signal is the receiver’s concern. Lacking global resource allocation and precise power control, the signals from different APs may originate from different services, and are perhaps differently modulated. Due to this particularity, it is necessary to figure out how the characteristics of interfering signals such as modulation and power affect the BER of 25. doi:10.6342/NTU201902343.