Académique Documents
Professionnel Documents
Culture Documents
S ENIOR E DITORS
D USIT N IYATO K AI K IT W ONG X IANG -G EN X IA
Nanyang Technological Univ. Dept. of Electron. & Elect. Eng. Dept. of Elect. and Comput. Eng.
Singapore Univ. College London Univ. of Delaware
London, United Kingdom Newark, DE, USA
A SSOCIATE E DITORS
KOICHI A DACHI S WADES D E C HUAN H UANG B EHROOZ M AKKI TANELI R IIHONEN J IE X U
The Univ. of Electro-Communications Indian Institute of Technology Univ. of Electronic Science and Ericsson Tampere Univ. of Guangdong Univ. of
Tokyo, Japan Delhi Technology of China Stockholm, Sweden Technology Technology
New Delhi, India Chengdu, China Tampere, Finland Guangzhou, China
S ERGEY A NDREA PANOS P. M ARKOPOULOS
Tampere Univ. of Technology T OMASO DE C OLA YONGMING H UANG Rochester Institute of Technology C ONG S HEN KOJI YAMAMOTO
Tampere, Finland German Aerospace Center (DLR) Southeast Univ. Rochester, NY, USA Univ. of Sci. and Technol. of Kyoto Univ.
Wessling, Germany Nanjing, China China Kyoto, Japan
M OHAMAD A SSAAD Sheng, China
CentraleSupélec JAN M IETZNER
RODRIGO C. DE L AMARE H ARSHAN JAGADEESH Hamburg Univ. of Applied Sciences Y UAN S HEN S HAOSHI YANG
France
Pontifical Catholic Univ. Indian Institute of of Technology Delhi Hamburg, Germany Beijing Univ. of Posts and
Tsinghua Univ.
L IN BAI of Rio de Janeiro Delhi, India Telecommunications
Beijing, China
Beihang Univ. Rio de Janeiro, Brazil S AIF K HAN M OHAMMED Beijing, China
Beijing, China A BLA K AMMOUN Indian Institute of Technology M IN S HENG
H ARPREET D HILLON King Abdullah Univ. of Science Delhi Xidian Univ. H IROYUKI YOMO
C HAN -B YOUNG C HAE Virginia Tech and Technology New Delhi, India Xi’An, China Kansai Univ.
Yonsei Univ. Blacksburg, VA, USA Thuwal, Saudi Arabia Osaka, Japan
Seoul, Korea M OHAMMED NAFIE DANIEL KC S O
PANAGIOTIS D IAMANTOULAKIS M ARIOS KOUNTOURIS Cairo Univ. Univ. of Manchester G UANDING Y U
YANJIAO C HEN France Research Center
Aristotle Univ. of Thessaloniki Cairo, Egypt Manchester, United Kingdom Zhejiang Univ.
Wuhan Univ. Huawei Technologies Co. Ltd., France
Hubei, China Thessaloniki, Greece Hangzhou, China
L AKSHMI P RASAD NATARAJAN DANIELE TARCHI
I NGMAR L AND Univ. of Bologna
J INHO C HOI PAWEL D MOCHOWSKI Indian Institute of Technology J IANHUA Z HANG
France Research Center Bologna, Italy
Gwangju Institute of Science Victoria Univ. of Wellington Hyderabad Beijing Univ. of Posts and
Huawei Technologies Co. Ltd., France
and Technology Wellington, New Zealand Telangana, India Telecommunications
M ANUEL V ELEZ
Gwangju, Korea Beijing, China
J UNGWOO L EE Univ. of Basque Country
M IANXIONG D ONG H IEN N GO
K AEWON C HOI Seoul National Univ. Vizcaya, Spain
Muroran Institute of Technology Queen’s Univ. Belfast H AIBO Z HOU
Seoul National Univ. of Science Seoul, Korea
Muroran, Japan Belfast, UK RUI WANG Nanjing Univ.
and Technology Nanjing, China
PAN L I Southern Univ. of Sci. and
Seoul, Korea Y UE G AO AYÇA Ö ZÇELIKKALE
Case Western Reserve Univ. Technol.
C HUN T UNG C HOU Queen Mary Univ. of London Uppsala Univ. Guangdong, China S HENG Z HOU
Cleveland, OH, USA
Univ. of New South Wales London, United Kingdom Uppsala, Sweden Tsinghua Univ.
RUI WANG Beijing, China
Sydney, Australia A N L IU
WALAA H AMOUDA P RZEMYSLAW PAWELCZAK Tongji Univ.
Hong Kong Univ. of Shanghai, China
X IAOLI C HU Concordia Univ. TU Delft X IANGYUN Z HOU
Science and Technology
Univ. of Sheffield Montreal, Canada Delft, Netherlands The Australian National
Clear Water Bay, Hong Kong SAR C HAO -K AI W EN
Sheffield, United Kingdom Univ.
National Sun Yat-sen
D INH T HAI H OANG VASANTHAN R AGHAVAN Canberra, Australia
J USTIN C OON S HAODAN M A Univ.
Univ. of Technol. Sydney Qualcomm Flarion
Univ. of Oxford Univ. of Macau Kaohsiung, Taiwan
Sydney, Australia Technologies, Inc. Y U Z HU
Oxford, United Kingdom Macau SAR
Bridgewater, NJ, USA D IRK W ÜBBEN Fudan Univ.
Univ. of Bremen Shanghai, China
Bremen, Germany
IEEE O FFICERS
J OS É M. F. M OURA , President W ITOLD M. K INSNER , Vice President, Educational Activities
T OSHIO F UKUDA , President-Elect H ULYA K IRKICI , Vice President, Publication Services and Products
K ATHLEEN K RAMER , Secretary F RANCIS B. G ROSZ , J R ., Vice President, Member and Geographic Activities
J OSEPH V. L ILLIE , Treasurer ROBERT S. F ISH , President, Standards Association
JAMES A. J EFFERIES , Past President K. J. R AY L IU , Vice President, Technical Activities
T HOMAS M. C OUGHLIN , President, IEEE-USA
V IJAY K. B HARGAVA, Director, Division III-Communications Technology
IEEE W IRELESS C OMMUNICATIONS L ETTERS (ISSN 2162-2337) is published bi-monthly by the Institute of Elect. and Electronics Engineers, Inc. Responsibility for the
contents rests upon the authors and not upon the IEEE, the Society/Council, or its members. IEEE Corporate Office: 3 Park Avenue, 17th Floor, New York, NY 10016-5997.
IEEE Operations Center: 445 Hoes Lane, P.O. Box 1331, Piscataway, NJ 08855-1331. NJ Telephone: +1 732 981 0060. Price/Publication Information: Individual copies:
IEEE Members $20.00 (first copy only), nonmembers $91.00 per copy. Member and nonmember subscription prices available upon request. Available in microfiche and
microfilm. Copyright and Reprint Permissions: Abstracting is permitted with credit to the source. Libraries are permitted to photocopy for private use of patrons, provided
the per-copy fee indicated in the code at the bottom of the first page is paid through the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. For all
other copying, reprint, or republication permission, write to Copyrights and Permissions Department, IEEE Publications Administration, 445 Hoes Lane, P.O. Box 1331,
Piscataway, NJ 08855-1331. Copyright c 2019 by Institute of Elect. and Electronics Engineers, Inc. All rights reserved. IEEE prohibits discrimination, harassment and
bullying. For more information visit http://www.ieee.org/nondiscrimination. Printed in U.S.A.
Full-Duplex Energy-Harvesting Enabled Relay Networks in Generalized Fading Channels . . . . . . . . . . Khaled Rabie,
Bamidele Adebisi, Galymzhan Nauryzbayev, Osamah S. Badarneh, Xingwang Li, and Mohamed-Slim Alouini 384
Connectivity and Blockage Effects in Millimeter-Wave Air-To-Everything Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kaifeng Han, Kaibin Huang, and Robert W. Heath, Jr. 388
Receiver Design for OOK Modulation Over Turbulence Channels Using Source Transformation . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mohammad Taghi Dabiri and Seyed Mohammad Sajad Sadough 392
Handover Probability Analysis of Anchor-Based Multi-Connectivity in 5G User-Centric Network . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hongtao Zhang, Wanqing Huang, and Yi Liu 396
Traffic-Aware Relay Vehicle Selection in Millimeter-Wave Vehicle-to-Vehicle Communication . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bo Fan, Hui Tian, Shushan Zhu, Yanyan Chen, and Xuzhen Zhu 400
Decentralized Precoding for Cache-Enabled Ultra-Dense Radio Access Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . Shiwen He, Yiyun Chen, Ju Ren, Yongming Huang, Luxi Yang, and Yaoxue Zhang 404
LoRa Throughput Analysis With Imperfect Spreading Factor Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Antoine Waret, Megumi Kaneko, Alexandre Guitton, and Nancy El Rachkidy 408
An Adaptive Optimal Mapping Selection Algorithm for PNC Using Variable QAM Modulation . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tong Peng, Yi Wang, Alister G. Burr, and Mohammad Shikh-Bahaei 412
Deep Learning-Based CSI Feedback Approach for Time-Varying Massive MIMO Channels . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tianqi Wang, Chao-Kai Wen, Shi Jin, and Geoffrey Ye Li 416
A Dynamic Pricing Strategy for Vehicle Assisted Mobile Edge Computing Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Di Han, Wei Chen, and Yuguang Fang 420
Spectral-Energy Efficiency Pareto Front in Cellular Networks: A Stochastic Geometry Framework . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Marco Di Renzo, Alessio Zappone, Thanh Tu Lam, and Mérouane Debbah 424
Rician K-Factor-Based Analysis of XLOS Service Probability in 5G Outdoor Ultra-Dense Networks . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hatim Chergui, Mustapha Benjillali, and Mohamed-Slim Alouini 428
Power Splitting-Based SWIPT Systems With Decoding Cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mohsen Abedi, Hamed Masoumi, and Mohammad Javad Emadi 432
On Iterative Compensation of Clipping Distortion in OFDM Systems . . . . . . Shansuo Liang, Jun Tong, and Li Ping 436
Automatic Modulation Classification Using Cyclic Correntropy Spectrum in Impulsive Noise . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jitong Ma and Tianshuang Qiu 440
Vertical and Horizontal Building Entry Loss Measurement in 4.9 GHz Band by Unmanned Aerial Vehicle . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kentaro Saito, Qiwei Fan, Nopphon Keerativoranan, and Jun-ichi Takada 444
Multi-Slot Allocation Protocols for Massive IoT Devices With Small-Size Uploading Data . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tsung-Yen Chan, Yi Ren, Yu-Chee Tseng, and Jyh-Cheng Chen 448
On the Sum-Rate of Heterogeneous Networks With Low-Resolution ADC Quantized Full-Duplex Massive
MIMO-Enabled Backhaul . . . . . . . . . Prince Anokye, Roger K. Ahiadormey, Changick Song, and Kyoung-Jae Lee 452
Efficient Computation of Multivariate Rayleigh and Exponential Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Reneeta Sara Isaac and Neelesh B. Mehta 456
Adaptive Frequency Band and Channel Selection for Simultaneous Receiving and Sending in Multiband
Communication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ayaka Hanyu,
Yuichi Kawamoto, Hiroki Nishiyama, Nei Kato, Naoto Egashira, Kazuto Yano, and Tomoaki Kumagai 460
Localization Using Blind RSS Measurements . . . . . . . . . . . . . . . . . . . . Yongchang Hu, Jiani Liu, and Bingbing Zhang 464
Fast Analog Transmission for High-Mobility Wireless Data Acquisition in Edge Learning . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yuqing Du and Kaibin Huang 468
On Achieving the Maximum Streaming Rate in Hybrid Wired/Wireless Overlay Networks . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jianwei Zhang, Xinchang Zhang, Meng Sun, and Chunling Yang 472
Interleave-Division Multiple Access in High Rate Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yang Hu, Chulong Liang, Lei Liu, Chunlin Yan, Yifei Yuan, and Li Ping 476
To Establish a Secure Channel From a Full-Duplex Transmitter to a Half-Duplex Receiver: An Artificial-Noise-Aided
Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xinyue Hu, Caihong Kai, Shengli Zhang, Zhongyi Guo, and Jun Gao 480
Hybrid Precoding for Single Carrier Wideband Multi-Subarray Millimeter Wave Systems . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wei Huang, Zhaohua Lu, Yongming Huang, and Luxi Yang 484
Threshold Setting for Multiple Primary User Spectrum Sensing via Spherical Detector . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xi Yang, Kejun Lei, Shengliang Peng, Li Hu, Shu Li, and Xiuying Cao 488
High-Accuracy Entity State Prediction Method Based on Deep Belief Network Toward IoT Search . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Puning Zhang, Xuyuan Kang, Dapeng Wu, and Ruyan Wang 492
High Rate CCK Modulation Design for Bandwidth Efficient Link Adaptation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Han Wang, Lianyou Jing, Chengbing He, and Zhi Ding 496
Sequential 0/1 for Cooperative Spectrum Sensing in the Presence of Strategic Byzantine Attack . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jun Wu, Yue Yu, Tiecheng Song, and Jing Hu 500
Spherical Wave Positioning Based on Curvature of Arrival by an Antenna Array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . Siwei Zhang, Thomas Jost, Robert Pöhlmann, Armin Dammann, Dmitriy Shutin, and Peter Adam Hoeher 504
Molecular Communication: The First Arrival Position Channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nilay Pandey, Ranjan K. Mallik, and Brejesh Lall 508
Spatial Correlations of a 3-D Non-Stationary MIMO Channel Model With 3-D Antenna Arrays and 3-D Arbitrary
Trajectories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . Qiuming Zhu, Ying Yang, Cheng-Xiang Wang, Yi Tan, Jian Sun, Xiaomin Chen, and Weizhi Zhong 512
Protograph-Based Folded Spatially Coupled LDPC Codes for Burst Erasure Channels . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Inayat Ali, Hyunjae Lee, Ayaz Hussain, and Sang-Hyo Kim 516
Distribution of the Number of Users per Base Station in Cellular Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Geordie George, Angel Lozano, and Martin Haenggi 520
Joint Power, Altitude, Location and Bandwidth Optimization for UAV With Underlaid D2D Communications . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wenhuan Huang, Zhaohui Yang,
Cunhua Pan, Lu Pei, Ming Chen, Mohammad Shikh-Bahaei, Maged Elkashlan, and Arumugam Nallanathan 524
OTFS-Based Multiple-Access in High Doppler and Delay Spread Wireless Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Venkatesh Khammammetti and Saif Khan Mohammed 528
Optimal Hybrid Beamforming for Multiuser Massive MIMO Systems With Individual SINR Constraints . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . Guangda Zang, Ying Cui, Hei Victor Cheng, Feng Yang, Lianghui Ding, and Hui Liu 532
On the Error Rate Analysis of Coded OFDM Over Multipath Fading Channels . . . . . . . . . . . . . . . . . . . . . . Jinho Choi 536
Adaptive AoA and Polarization Estimation for Receiving Polarized mmWave Signals . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hang Li, Thomas Q. Wang, Xiaojing Huang, and Y. Jay Guo 540
Antieigenvalue-Based Spectrum Sensing for Cognitive Radio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chen Guo, Ming Jin, Qinghua Guo, and Youming Li 544
MRB Decoding of LT Codes Over AWGN Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Valerio Bioglio 548
Spectral and Energy Efficient Resource Allocation for Massive MIMO HetNets With Wireless Backhaul . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bo Huang and Aihuang Guo 552
Practical User Selection With Heterogeneous Bandwidth and Antennas for MU-MIMO WLANs . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sulei Wang, Zhe Chen, Yuedong Xu, Xin Wang, and Qingsheng Kong 556
Coupling Information Transmission With Window Decoding . . . . . . . . . . . . . . . Alireza Karami and Dmitri Truhachev 560
Secure UAV-to-UAV Systems With Spatially Random UAVs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jia Ye, Chao Zhang, Hongjiang Lei, Gaofeng Pan, and Zhiguo Ding 564
Flexible-Rate SIC-Free NOMA for Downlink VLC Based on Constellation Partitioning Coding . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chen Chen, Wen-De Zhong, Helin Yang, Pengfei Du, and Yanbing Yang 568
Meta Distribution of Downlink Non-Orthogonal Multiple Access (NOMA) in Poisson Networks . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Konpal Shaukat Ali, Hesham ElSawy, and Mohamed-Slim Alouini 572
PAPR Reduction Based on Parallel Tabu Search for Tone Reservation in OFDM Systems . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yajun Wang, Renjie Zhang, Jun Li, and Feng Shu 576
Resource Allocation in UAV-Assisted M2M Communications for Disaster Rescue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Xilong Liu and Nirwan Ansari 580
Coded Redundant Message Transmission Schemes for Low-Power Wide Area IoT Applications . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Samuel Montejo-Sánchez, Cesar A. Azurdia-Meza,
Richard Demo Souza, Evelio Martin Garcia Fernandez, Ismael Soto, and Arliones Hoeller, Jr. 584
On Optimizing Effective Rate for Random Linear Network Coding Over Burst-Erasure Relay Links . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Huangnan Wu, Ye Li, Yingdong Hu, Bin Tang, and Zhihua Bao 588
Different Power Adaption Methods on Fluctuating Two-Ray Fading Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hui Zhao, Zhedong Liu, and Mohamed-Slim Alouini 592
Optimal Dynamic Capacity Allocation for High Throughput Satellite Communications Systems . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . Anargyros J. Roumeliotis, Charilaos I. Kourogiorgas, and Athanasios D. Panagopoulos 596
Learning-Based Wireless Powered Secure Transmission . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dongxuan He, Chenxi Liu, Hua Wang, and Tony Q. S. Quek 600
New Analytical Approach in the SER Evaluation of CSIN-Assisted AF Dual-Hop Wireless Systems . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Yazid M. Khattabi 604
DF-CSPG: A Potential Game Approach for Device-Free Localization Exploiting Joint Sparsity . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sixing Yang, Yan Guo, Ning Li, and Dagang Fang 608
Fundamentals on Base Stations in Urban Cellular Networks: From the Perspective of Algebraic Topology . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ying Chen, Rongpeng Li, Zhifeng Zhao, and Honggang Zhang 612
Modified Conjugate Beamforming for Cell-Free Massive MIMO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Masoud Attarifar, Aliazam Abbasfar, and Angel Lozano 616
Divergence-Optimal Fixed-to-Fixed Length Distribution Matching With Shell Mapping . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Patrick Schulte and Fabian Steiner 620
SCR-Based Tone Reservation Schemes With Fast Convergence for PAPR Reduction in OFDM System . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jingqi Wang, Xin Lv, and Wen Wu 624
Average Age of Information in Wireless Powered Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ioannis Krikidis 628
User Cooperation in Wireless-Powered Backscatter Communication Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bin Lyu, Dinh Thai Hoang, and Zhen Yang 632
Tag Cardinality Estimation Using Expectation-Maximization in ALOHA-Based RFID Systems With Capture Effect
and Detection Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chuyen T. Nguyen, Van-Dinh Nguyen, and Anh T. Pham 636
Pilot Allocation and Computationally Efficient Non-Iterative Estimation of Phase Noise in OFDM . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ville Syrjälä, Toni Levanen, Tero Ihalainen, and Mikko Valkama 640
Energy-Perceptive MAC for Wireless Power and Information Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Youngil Cho, Yunmin Kim, and Tae-Jin Lee 644
Abstract—In this letter, we explore user pairing in a downlink with more-user NOMA. For analytical tractability, we consider
non-orthogonal multiple access (NOMA) network. As power allo- 2-user NOMA in this letter [3], [5]–[7], while the generaliza-
cation inherently intertwines with user pairing, a joint user tion of 2-user NOMA to p-user NOMA is also considered in
pairing and power allocation problem is considered to optimize
the achievable sum rate (ASR) with minimum rate constraint Section III-D.
for each user, which is a mixed integer programming problem. When applying the 2-user NOMA to mobile cellular,
To solve this non-convex problem, we first obtain the optimal the achievable benefit is highly dependent on user pair-
power allocation in an NOMA system with only 2 users; then ing [3], [7]. In [3], user pairing under two cases, i.e., NOMA
analyze the user pairing problem in a simplified situation, i.e., an with fixed power allocation (F-NOMA) and cognitive-radio-
NOMA system with four users. Finally, we obtain the closed-form
globally optimal solution in a general NOMA system. Extensive inspired NOMA (CR-NOMA), was studied. A general criteria
performance evaluations are conducted to compare the ASRs of was given to design distributed approaches for dynamic user
the NOMA and OMA systems. Results show that the performance pairing/grouping, but explicit user pairing strategy was not
of the NOMA system with the proposed optimal user pairing is given [3]. User pairing in the CR-NOMA system was further
significantly better than that of the OMA system, as well as the studied in [7], where the conventional distributed matching
performance of the NOMA system with random user pairing.
algorithm (DMA) was adopted to maximize the achievable
Index Terms—Non-orthogonal multiple access (NOMA), user sum rate (ASR). In this letter, we also explore the user pair-
pairing, power allocation. ing problem, but with a minimum rate constraint for all the
NOMA users, so as to guarantee the users’ quality of service
(QoS). With this system model, we obtain the closed-form
I. I NTRODUCTION globally-optimal solution.
ON-ORTHOGONAL multiple access (NOMA) is con-
N sidered as a key candidate technology for the fifth
generation (5G) networks [1]–[4]. The basic idea of
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
A. System Model
NOMA is to serve multiple users in the same resource Without loss of generality, we consider a downlink NOMA
(time/frequency/code) block (RB). As the signals of different system. There are 2K users uniformly deployed in a disc,
users are superimposed in power domain, the receivers exploit namely D with the radius d. The base station (BS) is located at
successive interference cancellation (SIC) to distinguish each the center of D. For the sake of improving the spectrum effi-
other. Thus both the number of users and the spectrum effi- ciency, the users are paired into K clusters. The paired users
ciency can be improved manyfold. In general, 2-user NOMA transmit information in the same RB, but users from different
is a typical scenario, where the number of users perform- pairs should transmit in different RBs. For each pair of users,
ing NOMA in a single RB is 2. The decoding complexity the BS transmits a superimposed signal as
and delay at the receivers are lower and shorter compared
s = αm P sm + αn P sn , (1)
Manuscript received June 4, 2018; accepted June 28, 2018. Date of pub- where sk (k = m,n) is the signal for User-k, and E (|sk |2 ) = 1.
lication July 9, 2018; date of current version April 9, 2019. This work was P is the total transmission power for each pair of users.
supported in part by the National Key Research and Development Program
under Grant 2016YFB1200100, in part by the National Natural Science αk (k = m,n) denotes the coefficient of signal power for User-k,
Foundation of China under Grant 61571025 and Grant 91538204, and in and αm + αn = 1.
part by the Open Research Fund of Key Laboratory of Space Utilization, The received signals at the users are
Chinese Academy of Sciences under Grant LSU-DZXX-2017-02. The asso- √ √
ciate editor coordinating the review of this paper and approving it for ym = hm (√ αm P sm +√ αn P sn ) + n̂m ,
publication was R. C. de Lamare. (Corresponding author: Jun Zhang.) (2)
L. Zhu, Z. Xiao, and X. Cao are with the School of Electronic and yn = hn ( αm P sm + αn P sn ) + n̂n ,
Information Engineering, Beihang University, Beijing 100191, China, also where hk (k = m,n) denotes the channel gain between the
with the Key Laboratory of Advanced technology of Near Space Information
System, Ministry of Industry and Information Technology of China, Beijing BS and User-k, which is assumed to be Rayleigh distributed.
100191, China, and also with the National Engineering Laboratory for Big Without loss of generality, we assume that the channel gain
Data Application Technologies for Comprehensive Traffic, Beijing 100191, obeys the standard Rayleigh distribution at the node 100m
China.
J. Zhang is with the Advanced Research Institute of Multidisciplinary away from the BS. Additionally, the pathloss is defined as
Science, Beijing Institute of Technology, Beijing 100081, China (e-mail: 1/d2ab , where dab is the distance between node a and node b.
buaazhangjun@vip.sina.com). n̂k denotes the white noise at User-k with power σ 2 .
D. O. Wu is with the Department of Electrical and Computer Engineering,
University of Florida, Gainesville, FL 32611 USA. Without loss of generality, assume the channel gains of the
Digital Object Identifier 10.1109/LWC.2018.2853741 users are sorted as |h1 |2 ≤ |h2 |2 ≤ · · · |h2K |2 . In particular,
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHU et al.: OPTIMAL USER PAIRING FOR DOWNLINK NOMA 329
User-m and User-n are two of the 2K users with |hm |2 ≤ III. S OLUTION OF THE P ROBLEM
|hn |2 . The channel gain of User-n is higher, so User-n can In order to solve the original problem, we commence from
first decode and remove sm , and then decodes sn . By contrast, the cases with relatively fewer number of users, i.e., 2K = 2,4,
User-m directly decodes sm , and meanwhile sn is treated as which is easier and can be instructive to investigate the general
noise. The SIC technique is only employed at the user with case.
higher channel gain in each pair. Then the achievable rates of
the users are A. Power Allocation for NOMA With 2 Users
(m,n) |h |2 (1−α )
Rm = log2 (1 + |h m|2 α +γn−1 ), In a NOMA system with 2 users, we do not need to consider
m n (3)
(m,n) user pairing. Problem (6) is simplified as
Rn = log2 (1 + |hn |2 αn γ),
(m,n) (m,n)
where γ = σP2 , and αm is replaced by (1-αn ). Maximize Rm + Rn
{αn }
The achievable rate of the user in an OMA system is (m,n) (OMA)
given by Subject to Rm ≥ Rm ,
(m,n) (OMA)
(OMA) 1 Rn ≥ Rn ,
Rk = log (1 + |hk |2 γ), (4)
2 2 0 ≤ αn ≤ 1. (7)
1
where the factor is due to the fact that conventional OMA
2 There is only one variable, αn , in Problem (7). We give the
results in a multiplexing loss of 12 . optimal solution in the following theorem.
Theorem 1: In a NOMA system with 2 users, the optimal
B. Problem Formulation power allocation of Problem (7) is
First, it is necessary to define a 2K-dimension matrix to rep-
(m,n) 1 + |hm |2 γ − 1
resent the pairing relationship among the users. The formation αn = . (8)
|hm |2 γ
rule of U is
Proof: The differential of the objective function in
1, User-m pairs User-n;
um,n = (5) Problem (7) is
0, others;
(m,n) (m,n)
where um,n is the m-th row and j-th column element. It is d (Rm + Rn ) 1 (|hm |2 αn + γ −1 )2
= . (9)
obvious that UT = U due to the definition, and the diagonal d (αn ) ln2 (|hn |2 − |hm |2 )(|hm |2 + γ −1 )
element of U is zero because one user cannot pair itself. Each Note that we have assumed |hm |2 ≤ |hn |2 , thus the objective
user can pair one and only one user, so the summation of the function is nondecreasing for αn . The range of αn can be
elements in each row or column of U is 1. Furthermore, there obtained from the constraints directly.
are minimal rate constraints for the users, i.e., the achievable
(m,n) (OMA)
rate of the user in the NOMA system should be no less than Rm ≥ Rm
that of this user in the OMA system. Then the optimization |hm |2 (1 − αn ) 1
⇔ log2 (1 + 2 −1
) ≥ log2 (1 + |hm |2 γ)
problem is formulated as |h | αn + γ 2
m
2K
2K
1 + |hm |2 γ − 1
Maximize um,n (Rm
(m,n) (m,n)
+ Rn ) ⇔ αn ≤ . (10)
|hm |2 γ
{αn ,um,n }
m=1 n=m+1 (m,n) (OMA)
(m,n) (OMA) Rn ≥ Rn
Subject to Rm ≥ um,n Rm , 1
(m,n) (OMA) ⇔ log2 (1 + |hn |2 αn γ) ≥ log2 (1 + |hn |2 γ)
Rn ≥ um,n Rn , 2
0 ≤ αn ≤ 1, 1 ≤ n ≤ 2K , 1 + |hn |2 γ − 1
⇔ αn ≥ . (11)
|hn |2 γ
um,n ∈ {0, 1}, 1 ≤ m, n ≤ 2K ,
um,n = un,m , 1 ≤ m, n ≤ 2K , Thus the optimal solution of αn is the upper bound, i.e.,
um,m = 0, 1 ≤ m ≤ 2K , (m,n) 1 + |hm |2 γ − 1
αn = . (12)
2K
|hm |2 γ
um,n = 1, 1 ≤ n ≤ 2K , (m,n)
m=1 Finally, we verify the range of αn . As√ |hm | 2γ > 0,
2
1 +|h m | γ−1
2K
we have 1 + |hm |2 γ > 1. Thus, 0 < |hm |2 γ
<
um,n = 1, 1 ≤ m ≤ 2K . (6) (1+|hm |2 γ)−1
n=1 |h |2 γ
= 1.
m
(m,n)
Problem (6) is a mixed integer programming problem. The The superscript of αn denotes the indices of the pair-
complexity of direct search for the optimal user pairing is ing users. It is worthy to note that the achievable rate of
(m,n) (OMA)
O ((2K-1)!!), which is rather high. In this letter, we will the user with lower channel gain is Rm = Rm when
obtain the closed-form solution of Problem (6) by analytical the power allocation is optimal. That is to say, only nec-
derivation. essary power is allocated to the user with a worse channel
330 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
1 (1+β x )2
to satisfy the rate constraint, and all the remaining power Define ψ(x ) = log2 ( 1+x ) and it is easy to find that
is allocated to the user with a better channel to maximize
ψ (x) ≥ 0 when x ≥ |h1 | γ. Thus we have
2
the ASR.
Rcase2 − Rcase1
(1,3) (2,4) (OMA) (1,2) (3,4) (OMA)
= (R3 + R4 + R2 ) − (R2 + R4 +R3 )
B. User Pairing for NOMA With 4 Users
(1 + |h3 |2 β1 γ)(1 + |h4 |2 β2 γ) 1 + |h2 |2 γ
In a NOMA system with 4 users, if User-1 pairs one user, = log2 ( )
then the other two users must be a pair. For this reason, there (1 + |h2 |2 β1 γ)(1 + |h4 |2 β3 γ) 1 + |h3 |2 γ
are three cases of user pairing, which are (1 + |h3 |2 β1 γ) 1 + |h2 |2 γ
≥ log2 ( )
Case 1: User-1 pairs User-2, i.e., {u1,2 = 1, u3,4 = 1}. (1 + |h2 |2 β1 γ) 1 + |h3 |2 γ
The ASR with the optimal power allocation in (12) is 1 (1 + β1 |h3 |2 γ)2 (1 + β1 |h2 |2 γ)2
= [log2 ( 2
) − log2 ( )] ≥ 0.
(1,2) (1,2) (3,4) (3,4) 2 1 + |h3 | γ 1 + |h2 |2 γ
Rcase1 = R1 + R2 + R3 + R4 (20)
(OMA) (1,2) (OMA) (3,4)
= R1 + R2 + R3 + R4 . (13)
√ i.e., u1,2K = 1.
Define φ(x ) = 1+x x
−1
and it is easy to find that φ (x)<0 2) When i = k (k ≥ 1), we assume that u1,2K =
when x > 0. For this reason, according to the assumption that 1, u2,2K −1 = 1, . . . , uk ,2K −k +1 = 1, and we also use
|h1 |2 ≤ |h2 |2 ≤ |h3 |2 , the order of βk (k = 1, 2, 3) is contradiction to prove that uk +1,2K −k = 1. Assume
User-(k + 1) pairs User-m (k + 2 ≤ m ≤ 2K − k − 1)
β3 ≤ β2 ≤ β1 . (18) and User-(2K-k) pairs User-n (k +2 ≤ n ≤ 2K −k −1).
User-(k + 1), User-m, User-n and User-(2K-k) consti-
Then, we have tute a NOMA system with 4 users, and the user pairing
among them is Case 1 (m < n) or Case 2 (m > n). Both
Rcase3 − Rcase2 of them are worse than Case 3. Thus the assumption
(1,4) (2,3) (1,3) (2,4)
= (R4 + R3 ) − (R3 + R4 ) does not hold. In other words, User-(k + 1) inevitably
(1 + |h4 |2 β1 γ)(1 + |h3 |2 β2 γ) pairs User-(2K-k), i.e., uk +1,2K −k = 1.
= log2 ( ) 3) Finally, we can conclude that u1,2K = 1, u2,2K −1 =
(1 + |h3 |2 β1 γ)(1 + |h4 |2 β2 γ)
1, . . . , uK ,K +1 = 1.
(|h4 |2 − |h3 |2 )(β1 − β2 )γ
= log2 (1 + ) ≥ 0. (19) Hereto, the closed-form solution of Problem (6) is obtained
(1 + |h3 |2 β1 γ)(1 + |h4 |2 β2 γ) in (12) and (21), which is globally optimal. Given the order
ZHU et al.: OPTIMAL USER PAIRING FOR DOWNLINK NOMA 331
Fig. 1. Comparison of ASRs between NOMA with optimal user pairing, Fig. 2. Comparison of ASRs between NOMA with optimal user pairing,
NOMA with random user pairing and OMA with varying number of users, NOMA with random user pairing and OMA with varying total power to noise
where P/σ 2 = 20 dB and d = 500 m. ratio, where K = 16 and d = 500 m.
of the channel gains, the optimal user pairing is determined power to noise ratio, respectively. Each point in these figures
according to the closed-form expression of (21). The main is the average performance based on 103 user distributions and
computational complexity is the calculation of power allo- channel realizations. It can be observed from the two figures
cation, which is shown in (12). Thus, the computational that the ASR for optimal user pairing is significantly better
complexity of the proposed approach is O(K ). than that for random user pairing and OMA.
Abstract—The capacity and optimal signaling over a fixed power budget and where each antenna is equipped with its own
Gaussian MIMO channel are considered under the joint total and power amplifier. The capacity of fixed Gaussian MISO chan-
per-antenna power constraints. While the general case remains nel under the joint TPC and PAC has been established in [10],
an open problem, a closed-form full-rank solution is obtained
along with its sufficient and necessary conditions. The conditions
where it was shown that the optimal signaling is a combination
for each constraint to be inactive are established. The high and of EGT and maximum ratio transmission (MRT), with phase
low-SNR regimes are studied. Isotropic signaling is shown to be shifts adjusted to compensate channel-induced phase shifts.
optimal in the former case while rank-1 signaling (beamforming) Following the remark in [9, Sec. II-B], the MISO result can be
is not necessarily optimal in the latter case. Unusual properties of also adapted to any rank-1 MIMO channel. This result was fur-
optimal covariance under the joint constraints are pointed out. ther extended to fading MIMO channels, where it was shown
Index Terms—MIMO, channel capacity, power constraint, that isotropic signaling is optimal if the fading distribution is
optimal signalling. right-unitary-invariant [10].
While an iterative algorithm to compute an optimal covari-
ance was developed in [11] for the general MIMO case under
I. I NTRODUCTION
the joint constraints, it provides limited insights (due to its
ULTI-ANTENNA (MIMO) systems have been widely
M accepted by both academia and industry due to their
high spectral efficiency. Presently, MIMO systems experience
iterative nature) and no closed-form solution is yet known.
The key difficulty is the fact that, unlike the TPC only case,
the feasible set of Tx covariance matrices is not isotropic any-
a re-surge of interest in the form of massive MIMO, which more (due to the PAC) and hence the tools developed under the
is considered a key technology for future 5G systems to meet TPC (which exploit this symmetry) cannot be used anymore.
ever-increasing traffic demand when a limited bandwidth is New tools are needed.
available [1]. The capacity of a fixed Gaussian MIMO chan- This letter partially closes this gap by obtaining a closed-
nel and its optimal signaling strategy are well-known under the form full-rank solution for the optimal signaling (Tx covari-
total transmit (Tx) power constaint (TPC): its is on the channel ance matrix) and respective capacity of a fixed full-rank
eigenmodes with power allocation given by the water-filling Gaussian MIMO channel under the joint constraints when the
(WF) procedure [2], [3]. While the TPC is motivated by a constraint powers exceed certain thresholds, thus extending
limited power (energy) supply, individual per-antenna powers earlier analytical results in [9] and [10]. Sufficient and nec-
can also be limited when each antenna is equipped with its essary conditions for optimal signaling to be of full rank are
own amplifier (of limited power), in either collocated or dis- also established. Optimal signaling under the joint constraints
tributed implementations, hence motivating per-antenna power is shown to have properties significantly different from those
constraint (PAC) for single-user as well as multi-user systems, under the TPC only, see Section VI. It is the inter-play between
as in [4]–[9]. While a number of iterative optimization algo- the TPC and PAC that induces these unusual properties. The
rithms have been proposed [4]–[6], closed-form solutions are conditions when either TPC and PAC are inactive are given.
known only in some special cases. The capacity and optimal The high and low-SNR regimes are studied. Isotropic signal-
signaling for a fixed Gaussian MISO channel under the PAC ing is shown to be optimal under the joint constraints in the
has been established in [7], which is significantly different former case while rank-1 signaling (beamforming) is not nec-
from the standard WF solution and is equivalent to the equal- essarily optimal in the latter case (in contrast to the standard
gain transmission (EGT) with phases adjusted to compensate WF signaling).
for the channel phase shifts. This problems remains open in Notations: Bold lower-case letters denote column vectors
the general MIMO case while a numerical algorithm was while bold capital denote matrices; R+ is Hermitian conjuga-
proposed in [8] and a closed-form full-rank solution was tion of R; rii denotes i-th diagonal entry of R; (R)ij is ij-th
obtained in [9]. entry of R, λi (R) is i-th eigenvalue of R, unless indicated oth-
The joint constraints, i.e., the TPC and the PAC simulta- erwise, eigenvalues are in decreasing order: λ1 ≥ λ2 ≥, . . . ,;
neously, are motivated by the scenario with limited overall R ≥ 0 means that R is positive semi-definite; |R| is the
determinant of R, I is identity matrix of appropriate size.
Manuscript received June 7, 2018; revised August 10, 2018; accepted
August 27, 2018. Date of publication August 30, 2018; date of current version
April 9, 2019. The associate editor coordinating the review of this paper and II. C HANNEL M ODEL
approving it for publication was L. P. Natarajan.
The author is with the School of Electrical Engineering and Computer Let us consider a discrete-time model of a fixed Gaussian
Science, University of Ottawa, Ottawa, ON K1N 6N5, Canada (e-mail: MIMO channel:
sergey.loyka@ieee.org).
Digital Object Identifier 10.1109/LWC.2018.2867858 y = Hx +ξ (1)
2162-2345
c 2018 Crown Copyright
LOYKA: ON CAPACITY OF GAUSSIAN MIMO CHANNELS UNDER JOINT POWER CONSTRAINTS 333
where y, x, ξ and H are the received and transmitted signals, entries set to zero), and D̄(W ) = W − D(W ) retaining
noise and channel, respectively; m is the number of trans- off-diagonal entries only; di = (W −1 )ii , where, without loss
mit antennas. The noise is Gaussian circularly-symmetric with of generality, d1 ≤ d2 ≤, . . ., i.e., in decreasing order.
zero mean and unit variance, so that power is also the SNR. Theorem 1: Let the channel matrix in (1) be of full column
The channel H is fixed and known to the transmitter and the rank, W = H + H > 0, and let the per-antenna and total
receiver (Rx). Under the Tx power constraint(s), Gaussian sig- transmit constraint powers be sufficiently high,
naling is known to be optimal in this setting [2], [3] so that
finding the channel capacity C and optimal signaling amounts P > λ−1 −1
m (W ), PT > mλm (W ) − trW
−1
(6)
to finding an optimal Tx covariance matrix R: Then, the (unique) optimal Tx covariance R∗ in (2) under the
C = maxR∈SR ln |I + W R| (2) TPC and PAC in (5) is of full-rank and is given by
where W = H + H , SR is the constraint set. In the R∗ = min(P I , λ−1 I − D(W −1 )) − D̄(W −1 ) (7)
case of the TPC constraint only, it takes the form SR = −1 −1 −1 −1
= λ I −W − ((λ − P )I − D(W ))+ (8)
{R : R ≥ 0, trR ≤ PT }, where PT is the maximum
total Tx power, and the optimal covariance is well-known: the where the operator min applies entry-wise; λ = 0 if
optimal signaling is on the channel eigenmodes with optimal mP ≤ PT , otherwise,
power allocation via the water-filling, which can be compactly ⎛ ⎞
m
expressed as 1 ⎝
λ−1 = PT + di − kP ⎠, (9)
m −k
R∗WF = (λ−1 I − W −1 )+ (3) i=k +1
where (A)+ retains positive eigenmodes of Hermitian matrix, k is the number of active PACs, determined as the largest
integer satisfying
(A)+ = λi (A)u i u +
i (4)
i : λi (A)>0 αk ≥ mP − PT (10)
m
where u i is i-th eigenvector of A; λ > 0 is determined from where αk = i=k (di − dk ). The capacity is
the TPC trR∗WF = PT . m
Under the PA constraints, SR = {R : R ≥ 0, rii ≤ P },
C = ln |W | + ln min(λ−1 , P + di ) (11)
where rii is i-th diagonal entry of R (the Tx power of i-
i=1
th antenna), P is the PA power constraint. No closed-form
solution is known for the optimal covariance in the general Proof: See the Appendix.
case under this constraint, while such solutions are available Note that αk is decreasing in k so that (10) can be efficiently
in the MISO case [7] and in the MIMO case when the optimal solved (even for massive MIMO) by verifying the condition
covariance is of full-rank [9]. for k in increasing order and stopping at the largest k satisfying
The joint power constraints, i.e., TPC and PAC, are moti- it. The number k of active PACs decreases with mP − PT .
vated by practical designs where each antenna has its own The expression in (8) has the following interpretation:
amplifier (and hence PAC) while limited total power/energy its first part λ−1 I − W −1 is the standard full-rank WF
supply motivates TPC. The optimal signaling and capacity solution under the TPC only, and its 2nd part (λ−1 I −
have been obtained in closed form under the joint constraints D(W −1 ) − P I )+ is a correction term accounting for
for the MISO channel in [10], while the general MIMO case the PAC.
remains an open problem. The next section provides a closed- It follows from (7) that per-antenna powers are as follows:
form full-rank solution for the MIMO case as well as sufficient
rii = min(P , λ−1 − (W −1 )ii ) > 0 (12)
and necessary conditions for this solution to hold and some
related properties. which also has an insightful interpretation: these powers are
the minimum of those under the PAC and TPC individually
III. O PTIMAL S IGNALING AND C APACITY (1st and 2nd term in the min operator, respectively).
Following the standard arguments, see [2], [3], Gaussian Next, we observe that the solution in Theorem 1 reduces to
signaling is still optimal under the joint constraints and the known solutions in some special cases.
channel capacity C is as in (2), where the constraint set SR Corollary 1: In Theorem 1, if the per-antenna constraint
is as follows: power P is sufficiently high,
and PT , P are the total and per-antenna constraint powers. then all PACs are inactive and (8) reduces to the
Unfortunately, no closed-form solution is known for the standard WF solution, R∗ = λ−1 I − W −1 , where
optimal covariance in (2) under the constraints in (5) in the λ−1 = m −1 (PT + trW −1 ).
general case. The following Theorem partially closes this gap Corollary 2: In Theorem 1, if the TPC power PT is suf-
and gives a closed-form full-rank solution for optimal signal- ficiently high, PT ≥ m P, then all PACs are active, the TPC
ing in this setting. To this end, let D(W) be the diagonal matrix is inactive and (7) reduces to the PAC-only full-rank solution
retaining only diagonal entries of W (with all off-diagonal in [9]: R∗ = P I − D̄(W −1 ).
334 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Corollary 3: In Theorem 1, i-th PAC is active if and only if where C (R) = ln |I + W R|, since R = PI is feasible under
−1 −1 the PAC. Next,
(W )ii < λ −P (14)
CPAC ≤ CTPC (mP ) (19)
It should be pointed out that while (6) are sufficient for the
optimal signaling to be of full-rank, they are not necessary, where CTPC (mP ) is the capacity under the TPC with the
i.e., there are cases where the optimal signaling is of full- total power PT = mP , since any feasible R under the PAC,
rank even when (6) does not hold. The following proposition rii ≤ P, is also feasible under the TPC with tr R ≤ mP.
gives necessary conditions for an optimal covariance to be of Using (16), one obtains at high SNR CTPC (mP ) ≈ C (P I )
full-rank. and hence CPAC ≈ C (P I ) and (17) follow.
Proposition 1: Let W > 0. The necessary conditions for We are now in a position to establish the optimality of
optimal covariance R∗ to be of full rank are as follows: isotropic signaling under the joint (TPC + PAC) constraints
P > λ1 (D̄(W −1 )), λ < λm (W ) (15) at high SNR.
Proposition 4: Consider a full column-rank channel. Let
where 1st condition is also sufficient if mP ≤ PT (inactive P ∗ = min(P , PT /m). Isotropic signaling is optimal in this
TPC), and λ is determined from (9). channel under the joint constraints (TPC + PAC) in the
Proof: Using (7), R∗ ≤ P I − D̄(W −1 ), so that R∗ > 0 high-SNR regime, i.e., when P ∗ λ−1 m (W ),
implies P I > D̄(W −1 ) and hence 1st condition in (15). 2nd
condition is obtained from 0 < R∗ ≤ λ−1 I − W −1 . R∗ ≈ P ∗ I , C ≈ ln |W | + m ln P ∗ (20)
Based on this, the following procedure can be used to
Proof: First, observe that C ≥ C (P ∗ I ) since R = P ∗ I is
establish whether optimal covariance is of full-rank in general:
feasible under the joint constraints: trR ≤ PT and rii ≤ P .
1. If PT ≥ mP, then 1st condition in (15) is both sufficient
Next,
and necessary for R∗ > 0 and (7) applies.
2. If PT < mP, define R∗ (λ) for a given λ > 0 from (7) and C ≤ min(CTPC , CPAC ) (21)
find λ from (9). If R(λ)∗ > 0, then it is a solution; otherwise,
optimal covariance is rank-deficient. and, at high SNR, CTPC ≈ C (PT I /m), CPAC ≈ C (P I ),
This procedure gives an exhaustive characterization of all and hence C ≈ C (P ∗ I ), as desired. The inequality P ∗
cases when R∗ is of full rank for a full-rank channel (since λ−1
m (W ) comes from the approximation ln(1 + x ) ≈ ln x ,
it follows from KKT conditions which are necessary for which holds if x 1.
optimality). It is remarkable that, for any of the constraints considered
In the following, we characterize the conditions when some here, isotropic signaling is optimal at high SNR. This sim-
constraints are inactive for a full-rank channel (even when plifies the system design significantly as no feedback and no
optimal covariance is not of full rank). elaborate precoding are necessary for this signaling strategy.
Proposition 2: Let W > 0. If the TPC is inactive, then all This also complements the respective result in [10] obtained
PACs are active. Hence, (i) when at least one PAC is inac- for the right-unitary-invariant fading channel.
tive, the TPC is active; (ii) the TPC is inactive if and only if
mP ≤ PT . V. L OW-SNR R EGIME
Proof: Follows from the stationarity condition in (24).
In this section, we consider the behaviour of optimal
It should be noted that this Proposition does not hold if the
covariance in the low-SNR regime, namely, when
channel is rank-deficient, as the example below demonstrates.
min(mP , PT ) λ−1
1 (W ) (22)
IV. H IGH -SNR R EGIME
It is well-known that, for the standard WF solution (under
It is well-known that isotropic signaling is optimal at high
the TPC only), the optimal signaling is beamforming (rank-1)
SNR for the standard WF solution (under the TPC only) in a
at low SNR, R∗WF ≈ PT u 1 u + 1 , where u 1 is the eigenvector
full-rank channel,
of W corresponding to its largest eigenvalue. As the following
PT
R∗WF ≈ I (16) example shows, this does not necessarily hold under the joint
m constraints.
when PT mλ−1 Example: Let PT = 1.5 · 10−2 , P = 10−2 , and W =
m (W ). In this section, we establish the
optimality of isotropic signaling under the joint constraints. diag{2,1}. It is straightforward to see that the optimal covari-
As a first step, the following proposition shows that isotropic ance is R∗ = 10−2 · diag{1, 0.5} in this case, i.e., full-rank
signaling is optimal at high SNR under the PAC. and beamforming is not optimal, does not matter how low
Proposition 3: Consider a full column-rank channel (W the SNR is. If, however, the per-antenna constraint power is
> 0). Isotropic signaling is optimal in this channel under the increased to P ≥ 1.5 · 10−2 , all PACs become inactive and
PAC in the high-SNR regime, i.e., when P λ−1 beamforming is optimal: R∗ = 10−2 · diag{1.5, 0}.
m (W ),
Hence, we conclude that it is the interplay between the TPC
R∗PAC ≈ P I , CPAC ≈ ln |W | + m ln P (17) and the PAC that makes a significant difference at low SNR
Proof: First, observe that while having negligible impact at high SNR: while the optimal
signaling under the TPC, the PAC and the joint constraints are
CPAC ≥ C (P I ) (18) all isotropic at high SNR, they are quite different at low SNR.
LOYKA: ON CAPACITY OF GAUSSIAN MIMO CHANNELS UNDER JOINT POWER CONSTRAINTS 335
A PPENDIX R EFERENCES
P ROOF OF T HEOREM 1 [1] M. Shafi et al., “5G: A tutorial overview of standards, trials, challenges,
Since the problem in (2) is convex and Slater’s condition deployment, and practice,” IEEE J. Sel. Areas Commun., vol. 35, no. 6,
pp. 1201–1221, Jun. 2017.
holds (as long as P, PT > 0), its KKT conditions are both suf- [2] B. S. Tsybakov, “Capacity of vector Gaussian memoryless channel,”
ficient and necessary for optimality [12]. The KKT conditions Prob. Inf. Transm., vol. 1, no. 1, pp. 26–40, 1965.
for this problem are as follows: [3] I. E. Telatar, “Capacity of multi-antenna Gaussian channels,” Eur. Trans.
Telecommun., vol. 10, no. 6, pp. 1–28, Dec. 1999.
− (I + W R)−1 W − M + λI + Λ = 0 (24) [4] W. Yu and T. Lan, “Transmitter optimization for the multi-antenna down-
link with per-antenna power constraint,” IEEE Trans. Signal Process.,
M R = 0, λ(trR − PT ) = 0, λi (rii − P ) = 0 (25) vol. 55, no. 6, pp. 2646–2660, Jun. 2007.
[5] A. Wiesel, Y. C. Eldar, and S. Shamai, “Zero-forcing precoding and
trR ≤ PT , rii ≤ P , R ≥ 0 (26) generalized inverses,” IEEE Trans. Signal Process., vol. 56, no. 9,
M ≥ 0, λ ≥ 0, λi ≥ 0 (27) pp. 4409–4418, Sep. 2008.
[6] S. Shi, M. Schubert, and H. Boche, “Per-antenna power constrained rate
where λ, λi are Lagrange multipliers (dual variables) respon- optimization for multiuser MIMO systems,” in Proc. Int. ITG Workshop
Smart Antennas, Vienna, Austria, 2008, pp. 270–277.
sible for the TPC and PAC, M is the (matrix) Lagrange [7] M. Vu, “MISO capacity with per-antenna power constraint,” IEEE Trans.
multiplier responsible for R ≥ 0, Λ = diag{λi }. The key diffi- Commun., vol. 59, no. 5, pp. 1268–1274, May 2011.
culty in solving analytically these conditions is that they are a [8] M. Vu, “MIMO capacity with per-antenna power constraint,” in Proc.
IEEE Globecom, Kathmandu, Nepal, Dec. 2011, pp. 1–5.
system of non-linear matrix equalities and inequalities, and the [9] D. Tuninetti, “On the capacity of the AWGN MIMO channel under
PACs make the feasible set SR non-isotropic so that standard per-antenna power constraints,” in Proc. ICC, Sydney, NSW, Australia,
tools (e.g., Hadamard inequality) cannot be used. However, Jun. 2014, pp. 2153–2157.
[10] S. Loyka, “The capacity of Gaussian MIMO channels under total and
when R is of full rank, the stationarity condition simplifies to per-antenna power constraints,” IEEE Trans. Commun., vol. 65, no. 3,
(R + W −1 )−1 = λI + Λ
pp. 1035–1043, Mar. 2017.
(28) [11] P. L. Cao and T. J. Oechtering, “Optimal transmit strategy for MIMO
channels with joint sum and per-antenna power constraints,” in Proc.
since M = 0 (from MR = 0), so that ICASSP, New Orleans, LA, USA, Mar. 2017, pp. 3569–3573.
[12] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
R = (λI + Λ)−1 − W −1 (29) Cambridge Univ. Press, 2004.
336 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Sparse code multiple access (SCMA) is a promising Physical layer security (PLS) has been widely studied in
air interface candidate technique for next generation mobile recent years [2]–[5], which is regarded as a promising supple-
networks. By introducing the Tent map in the Chaos theory, we ment to the cryptographic techniques. The first NOMA scheme
propose a novel physical layer transmission scheme with code-
word level interleaving at the transmitter in this letter, which
was defined on the power domain, and its secrecy outage
is termed as interleaver-based SCMA (I-SCMA). Simulation probability was derived in [6]. In [7], a secure transmission
results and analysis show that I-SCMA can provide high security scheme that maximizes the minimum confidential informa-
performance without any loss in performance and transmission tion rate among users was proposed. In [8], we propose a
rate, thus constitutes a viable solution for the next generation secure transmission scheme for downlink SCMA system with
wireless networks to provide secure communications. extra phase rotations in the design of codebook, called ran-
Index Terms—5G, SCMA, secure transmission, interleaving. domized constellation rotation based SCMA (RCR-SCMA).
However, in that work, extra uplink and downlink communi-
I. I NTRODUCTION cations are necessary, and the randomized codebooks are not
fully optimized, thus the secure communication is achieved by
CMA [1] is a code domain non-orthogonal multiple access
S (NOMA) scheme that is considered to be a promising
5G candidate due to its excellent ability to support massive
satisfying transmission rate with possible performance loss.
In this letter, we mainly focus on improving the com-
putational complexity for the eavesdropper to recover the
quantities of users under heavily loaded conditions. transmitted data, and thus achieving security for the system. In
In the possible application scenarios for SCMA, such as consequent, a novel physical layer secure transmission scheme
massive machine type of communication (mMTC), millions of with chaotic map and codeword level interleaving, which is
nodes must be accessed. Since ubiquitous mobile devices are denoted as I-SCMA, is proposed. To ensure the security,
required to be accessed to Internet of Things (IoT) in mMTC; interleavers that are constructed according to the channel phase
hence, unprecedented amount of private and sensitive data is with the association of Tent map in the Chaos theory is intro-
transmitted over wireless channels in an SCMA network. From duced. The basic idea of I-SCMA is to amplify the randomness
this perspective, the research on the secrecy issue of SCMA is of limited information that can be extracted from the chan-
of significant importance. Moreover, owing to the requirements nel state information (CSI), and thus leading to asymmetric
such as ultra low latency and low power consumption for the knowledge between the legitimate users (LUs) and eavesdrop-
IoT, it is also challenging to ensure the security in such a per in interleavers since they disperse the order of codewords
network. for each LU. From this perspective, the proposed scheme
Manuscript received August 10, 2018; accepted September 3, 2018. Date can be regarded as a physical layer encryption, which is a
of publication September 13, 2018; date of current version April 9, 2019. combination of conventional ciphers and physical layer secu-
This work was supported in part by the National Natural Science Foundation
of China under Grant 61502518, Grant 61372098, and Grant 61702536, in rity. Therefore, the security of I-SCMA is guaranteed by the
part by the National Defense Technology Foundation under Grant 3101168, unacceptable computational complexity at the eavesdropper
in part by the Hunan Natural Science Foundation under Grant 2017JJ2303, side.
and in part by the Natural Science Research Project of National University
of Defense Technology Research of New Multiple Access Techniques Based
The rest of this letter is organized as follow. Section II
on Joint Sparse Graph. The associate editor coordinating the review of this describes the system model of I-SCMA. In Section III,
paper and approving it for publication was R. Wang. (Corresponding author: the proposed I-SCMA transmission scheme is dis-
Lei Wen.) cussed. Numerical results and conclusion are presented
K. Lai and J. Lei are with the Department of Communication Engineering,
College of Electronic Science and Engineering, National University of in Sections IV and V, respectively.
Defense Technology, Changsha 410072, China.
L. Wen is with the Department of Communication Engineering, College
of Electronic Science and Engineering, National University of Defense II. S ECURE T RANSMISSION W ITH I NTERLEAVER
Technology, Changsha 410072, China, and also with the Institute for The system model of I-SCMA along with the construction
Communication Systems, Home of the 5G Innovation Centre, University of
Surrey, Guildford GU2 7XH, U.K. (e-mail: newton1108@126.com). of interleaver are presented in this section.
G. Chen is with the Department of Engineering, University of Leicester,
Leicester LE1 7RH, U.K. (e-mail: gaojie.chen@leicester.ac.uk).
P. Xiao is with the Institute for Communication Systems, Home of the 5G
A. Interleaver Based SCMA
Innovation Centre, University of Surrey, Guildford GU2 7XH, U.K. (e-mail: As shown in Fig. 1, we consider the uplink transmissions
p.xiao@surrey.ac.uk). where J single-antenna LUs transmit signal to the same base
A. Maaref is with Huawei Technologies Canada Co. Ltd., Ottawa,
ON K2K 3J1, Canada (e-mail: amine.maaref@huawei.com). station (BS) in the presence of an eavesdropper. We assume
Digital Object Identifier 10.1109/LWC.2018.2869792 that the eavesdropper is an external node, and the BS can
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
LAI et al.: SECURE TRANSMISSION WITH INTERLEAVER FOR UPLINK SCMA SYSTEM 337
B. Interleaver Construction
The key principle of I-SCMA is that the πj is derived
from the CSI of each LU to BS; hence, πj (j ∈ 1, . . . , J )
are unique for different users. It should be noted that the
interleaver in this letter has different usage with the ones of
Turbo codes or interleave-division multiple-access (IDMA), it
is utilized to disperse the order of SCMA codewords and thus
make the detected bits unpredictable for the eavesdropper. As
such, the eavesdropper cannot recover the original messages
without the knowledge of interleaver, therefore, the security
Fig. 1. Illustration of I-SCMA system with an external eavesdropper.
can be ensured. As the channels are assumed to be indepen-
dently, and the Tent map can generate quasi-random sequences
that are irreversible, thus πj are independent and random. In
distinguish it from LUs, which can be simply realized via
consequent, the transmitted data can be fully scrambled by
authentication in practice. As for the SCMA transmitter, each
the interleaver, i.e., the eavesdropper is unable to obtain any
function element is allocated to df users, and each user occu-
information even though it can receive the signals transmit-
pies du function elements. The encoder of I-SCMA is the
ted by LUs since the sequences of the data is different from
same as conventional SCMA (C-SCMA), which is defined by a
the original ones. It is obviously that the key of I-SCMA is
codebook Xj that maps log2 (C) binary bits to a K dimensional
the construction of πj , which requires sufficient randomness
complex codeword xj selected from the dedicated codebook
subject to very limited information that can be used in φr .
Xj corresponding to user j, where |Xj | = C, and C is the size
To construct an interleaver that can satisfy our demands, we
of constellation.
employ a Tent map in the Chaos theory [9], which can generate
At any transmission time, the received signal between single
a random sequence with a few initial parameters. The original
LU and BS can be expressed as
Tent map is expressed as follows:
yr (t) = hr (t)x (t) + n(t), (1) fμ := μ min{x , 1 − x } (2)
where x(t) is the transmitted signal, hr (t) is the channel gain For the values of the parameter μ ∈ [0, 2], fμ maps the unit
and n(t) is the additive white Gaussian noise. Without loss of interval [0, 1] into itself, thus defining a recurrence relation.
generality, hr (t) can be modeled as a complex Gaussian ran- In particular, iterating a given point x0 in [0, 1] gives rise to
dom variable, and its polar form is hr (t) = |hr (t)|e j φr (t) , a sequence xn :
where φr (t) ∈ [0, 2π]. At each transmission time instant,
μxn for xn < 12
the BS and each LU perform channel estimation in the same xn+1 = fμ (xn ) = (3)
μ(1 − xn ) for xn ≥ 12
coherent interval and thus they can obtain the same φr from
the CSI. Therefore, once the initial parameter μ and x0 are given,
In I-SCMA, bj and bˆj denote the coded information bits a random sequence take values that range from 0 to 1 can
and candidate decoded bits of user j, respectively, where be obtained. This is mainly due to the Tent map holds the
the code length is N. After encoding by the SCMA code- initial parameters within the given range define a dynamical
book mapper, there are N / log2 (C) codewords for each user. system and thus make the output from predictable to chaotic.
Subsequently, the codewords are conveyed to the interleaver πj It should be noted that such a Tent map is irreversible, i.e., the
to scramble the sequences of codewords. Note that each user original input cannot be obtained even the generated sequence
utilizes a different interleaving pattern and each data block is known to the eavesdropper.
can be encrypted with different πj at different transmission Considering the range of φr is [0, 2π] while the range of
time so that the security can be enhanced. The construction of μ and x0 are [0, 2] and [0, 1], respectively; hence, a mapping
interleaver will be demonstrated in the next subsection. After from φr to μ and x0 have to be constructed. Note that there
the transmitted chips are received, a de-interleaver is utilized exists numbers of such transformations and they are performed
before forward error correction (FEC) encoder. As the BS esti- locally and silently at BS and each LU such that no signals will
mates the channel in the same interval of each user, it can be radiated. Consequently, the eavesdropper cannot intercept
recover the interleaving patterns utilized by LUs and further f and g. However, due to the sensitivity of μ and x0 in the
attain the original transmitted data. It should be noted that as a Tent map, i.e., very minor variation of μ and x0 can generate
codeword level interleaver is applied and the detection at the a totally different sequence, and the robustness of I-SCMA,
receiver is codeword by codeword; hence, the received sig- f and g should have the ability to enlarge the digits of φr ,
nals can be directly detected. Subsequently, the de-interleaver which can further exploit the sensitivity of Tent map.
should be utilized to the detected bits of each user since the For the ease of implementation, we select two basic func-
codewords are scrambled. tions, which can be written as:
As can be observed from Fig. 1, the security of I-SCMA is
| sin φr | + | cos φr | = f : φr → μ (4)
based on the channel interdependence, moreover, the channel
independence is amplified via Tent map for the eavesdropper and
while the BS can recover the data since the CSI is assumed 1
to be unchanged within the coherent internal. (sin φr + 1) = g : φr → x0 (5)
2
338 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
Algorithm 1: Interleaver Construction of I-SCMA S EARCHING S PACE C OMPARISON P ER F RAME
Output: Interleavers πj
1 foreach transmission time do
2 LUs and BS start the channel estimation process;
3 LUs and BS obtain the channel phase φr (t) from the
CSI;
4 Each LU uses their φr (t) to calculate the initial
parameters in Tent mapping according to (4) and (5); channel phase with low margin. (iii) The interleavers can be
5 Each LU uses the generated μ and x in last step to changed over a period time as the CSI is time-varying, which
generate a Chaos sequence according to (3); can further enhance the security of I-SCMA.
6 Labling each block from 1 to N / log2 C; Therefore, the eavesdropper cannot recover the interleavers
7 Sorting the generated Chaos sequence and rearanging to intercept the transmitted data as BS do even the CSI
the lable of each block according to the sorted order; can be estimated with certain accuracy since the pilots are
8 Each user generates their own interleaver πj . broadcasted
To demonstrate the advantages of I-SCMA, we also com-
pare the decryption complexity of RCR-SCMA and C-SCMA
with I-SCMA by using brute force searching. In Table I, the
It is obviously that (4) and (5) can map φr to the range frame length equals to the code length N, C is the com-
of initial parameter in (2). After the μ and x0 are obtained, plexity of MPA to detect a data block, denotes the step
substituting them into (2) and (3) yields a sequence with cer- size of searching for a correct codebook. As reported in [8],
tain length (equal to the number of symbols of each user). the security cannot be enhanced by further reducing and
Therefore, each value of the generated sequence corresponds thus I-SCMA can achieve a better secrecy performance than
to an index of the transmitted codewords. For simplicity, the C-SCMA and RCR-SCMA in terms of decryption complexity.
construction of interleaver is given in Algorithm 1: B. Entropy Analysis of Post-Processing Attack
From the discussion above, inaccurate estimation will result
in catastrophic results. However, since the digits of channel Aiming at evaluating the performance of the post-processing
phase can be enlarged by f and g, the requiring accuracy of attacker (PPA) [11], we analyze the entropy of the received
extracted channel phase can be decreased. Furthermore, num- signal at the PPA side (HPPA ), and make comparison to the
bers of secret key negotiation and correction techniques are entropy of transmitted messages (HT ).
reported [10]. We assume that each user is independent as the codewords
are interleaved to be uncorrelated, according to the defini-
III. A NALYSIS AND D ISCUSSION tion of entropy, for a message with N bits, HT equals to N.
A. Decrypted Complexity of I-SCMA Considering various sizes of codebook in C-SCMA, as for a
PPA, the codebook size should be judged at first; hence, the
The decryption complexity for an eavesdropper is a vital
calculation of the entropy is:
secrecy performance metric. From the previous discussion, if
K Ik
the eavesdroppers intend to intercept the data by guessing the 1 N 1 1
interleavers of each user, then the complexity equals to J · HPPA = − · log2 (7)
K K log2 Ck k =1 i=1 KIk KIk
(N / log2 (C))!. Note that log2 (C) equals to 2 and 4 typically.
Consequently, the search space S can be approximated as: where K is the possible codebook size that the LUs can use;
Ck is the size of kth candidate codebook; Ik is the number
S = J · (N / log2 (C))! ≈ 10n·(ln n−1) , (6) of all possible interleavers. For simplicity and considering the
where n = N / log2 (C). As can be observed from (6), even if worse scenario, we assume that the codebook size is known
the code length N is moderate, the computational complexity to the eavesdropper, then (7) degenerates to:
is certainly unaffordable for the eavesdroppers. N 1 1 N
HPPA = − · log2 = · log2 (Ik ) (8)
In contrast to enumerate the interleavers as a random log2 C Ik Ik log2 C
Ik
attacker with brute force attack, another attacker model con-
sidered in this letter, called intelligent attacker [11] is able By applying the Stirling’s approximation, (7) can be further
to estimate the CSI with certain accuracy. However, they still written as:
encounter the following difficulties: (i) According to the fea- 2 √ 3
N 2π N 2
ture of Tent map in the Chaos theory, the generated sequences HPPA ≈ · log2 (9)
are very sensitive to the initial parameters μ and x0 . As log2 C e log2 C
reported in [9], the variation of 10 digits after decimal point √
2π
can even generate totally different sequences. Therefore, the Note that e ≈ 1, therefore,
intelligent attacker should have the ability to approximate the 2
3 N N
CSI with extremely high accuracy. (ii) The mappings from HPPA ≈ · log2 (10)
channel phase φr to μ and x0 are diverse, and the mapping f 2 log2 C log2 C
and g are unknown to the eavesdroppers; hence, it is impossi- In general, N C, thus (10), HPPA HT , which indicates
ble for the eavesdropper to intercept the initial parameter that that the entropy of secret keys is larger than plaintext; hence,
used to generate the interleavers even though it can guess the the proposed I-SCMA can reach the perfect secrecy for PPA.
LAI et al.: SECURE TRANSMISSION WITH INTERLEAVER FOR UPLINK SCMA SYSTEM 339
V. C ONCLUSION
Fig. 2. Error rate performance comparison of legitimate users and an
eavesdropper for I-SCMA. In this letter, we propose a novel secure transmission
scheme for uplink SCMA system, which is called I-
TABLE II SCMA. As indicated by the simulation results and analysis,
BER P ERFORMANCE OF E AVESDROPPER U NDER C ERTAIN
E STIMATED ACCURACY σ 2 = 0.04 I-SCMA can achieve a good secrecy performance with-
out any performance loss. Furthermore, I-SCMA does not
impose any overhead in terms of extra uplink and down-
link communications compared to the existing physical layer
security transmission schemes. In conclusion, our scheme
serves as a valuable supplement to conventional cryptographic
technologies.
IV. S IMULATION R ESULTS AND D ISCUSSION
The simulation results and security analysis of the I-SCMA
R EFERENCES
are discussed in this section.
As noted in [8] and [12], the error rate performance can [1] H. Nikopour and H. Baligh, “Sparse code multiple access,” in Proc.
IEEE 24th Int. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC),
be used to assess the secrecy performance of a system, thus London, U.K., Sep. 2013, pp. 332–336.
we evaluate the error rate performance from the perspective of [2] J. Qiao, H. Zhang, X. Zhou, and D. Yuan, “Joint beamforming and time
LUs and eavesdropper, respectively, to demonstrate the validity switching design for secrecy rate maximization in wireless-powered FD
of the proposed scheme. relay systems,” IEEE Trans. Veh. Technol., vol. 67, no. 1, pp. 567–579,
Jan. 2018.
Average error rate comparisons of each user for 4-point, [3] J. Qiao, H. Zhang, F. Zhao, and D. Yuan, “Secure transmission and
16-point I-SCMA and C-SCMA are shown in Fig. 2 (simula- self-energy recycling with partial eavesdropper CSI,” IEEE J. Sel. Areas
tion parameters K = 4; J = 6; df = 3; du = 2; max(Niter ) Commun., to be published, doi: 10.1109/JSAC.2018.2825541.
= 6; λ = 150%; N = 256 and code rate r = 0.5). As [4] G. Chen, Y. Gong, P. Xiao, and J. A. Chambers, “Dual antenna selection
in secure cognitive radio networks,” IEEE Trans. Veh. Technol., vol. 65,
can be seen from the figures, the performance of I-SCMA no. 10, pp. 7993–8002, Oct. 2016.
is slightly better than C-SCMA especially in the high SNR [5] G. Chen, J. Coon, and M. D. Renzo, “Secrecy outage analysis for down-
region, which indicates that I-SCMA will not suffer from link transmissions in the presence of randomly located eavesdroppers,”
performance loss compare to the existing secure transmission IEEE Trans. Inf. Forensics Security, vol. 12, no. 5, pp. 1195–1206,
May 2017.
for SCMA in [8]. This is mainly because the interleavers [6] Y. Liu, Z. Qin, M. Elkashlan, Y. Gao, and L. Hanzo, “Enhancing
disperse the coded sequences so that the adjacent blocks the physical layer security of non-orthogonal multiple access in large-
are approximately uncorrelated. Furthermore, it is clear that scale networks,” IEEE Trans. Wireless Commun., vol. 16, no. 3,
pp. 1656–1672, Mar. 2017.
eavesdroppers cannot obtain any information as they cannot
[7] B. He, A. Liu, N. Yang, and V. K. N. Lau, “On the design of secure
estimate φr with very high accuracy and the mappers are non-orthogonal multiple access systems,” IEEE J. Sel. Areas Commun.,
unknown to them. Note that the symbol error rate (SER) vol. 35, no. 10, pp. 2196–2206, Oct. 2017.
of I-SCMA for eavesdroppers approximate to 0.75 and 0.94 [8] K. Lai et al., “Secure transmission with randomized constellation rota-
tion for downlink sparse code multiple access system,” IEEE Access,
for 4-point and 16-point I-SCMA, respectively, which follows vol. 6, pp. 5049–5063, 2018.
from the fact that each symbol can be wrongly detected with [9] P. Collet and J. P. Eckmann, Iterated Maps on the Interval As Dynamical
probability: Systems. Boston, MA, USA: Birkhäuser, 1980.
[10] Y. Liu, H.-H. Chen, and L. Wang, “Physical layer security for next
C−1 generation wireless networks: Theories, technologies, and challenges,”
Pr{xi = x } = (11)
C IEEE Commun. Surveys Tuts., vol. 19, no. 1, pp. 347–376, 1st Quart.,
2017.
In Table II, the error rate performance of LUs under the [11] S. Althunibat, V. Sucasas, and J. Rodriguez, “A physical-layer secu-
condition that the eavesdropper can approximate the CSI with rity scheme by phase-based adaptive modulation,” IEEE Trans. Veh.
low error margin is presented. We assume that the estimated Technol., vol. 66, no. 11, pp. 9931–9942, Aug. 2017.
φr of eavesdropper follows the distribution φ˜r ∼ N (φr , σ 2 ). [12] I. M. Kim, B.-H. Kim, and J. K. Ahn, “BER-based physical layer
security with finite codelength: Combining strong converse and error
As can be observed from the table, although the eavesdrop- amplification,” IEEE Trans. Commun., vol. 64, no. 9, pp. 3844–3857,
per can estimate the CSI with certain accuracy, the BER and Sep. 2016.
340 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In conventional M-ary differential chaos shift keying communications. The idea of a hybrid combining modulation
modulation (DCSK) systems, the distance between constellation using PPM was first proposed in optical communications [11].
points gets closer as M increases, resulting in poor performance. A hybrid pulse PPM-BPSK add space for transmitted refer-
A hybrid modulation scheme based on pulse position modulation
(PPM) and DCSK is proposed in this letter to improve bit-error- ence pulse cluster (TRPC) is proposed to improve performance
rate (BER) performance. In this scheme, one part of the bits is in [12]. Similarly, DCSK also can be combined with PPM
modulated by PPM while the other part is modulated by DCSK. for better performance in spread spectrum communication
Thus, information bearing signals are simultaneously modulated systems.
by the information bit and the selected pulse position of PPM To improve the BER performance of conventional M-ary
which is determined by extra information bits. Analytical BER
performance of the proposed scheme is derived and verified by DCSK, a hybrid PPM-DCSK modulation scheme is proposed
simulations. Results show that the considered scheme outper- in this letter. The main contribution of the brief are summa-
forms conventional M-DCSK, code index modulation DCSK, and rized as follows. Firstly, based on a hybrid two-dimensional
commutation code index DCSK in additive white Gaussian noise modulation scheme, a mixed modulation scheme with PPM
and multipath Rayleigh fading channels. and DCSK is proposed which the PPM parts bearing mc
Index Terms—Differential chaos shift keying modulation bits signal. Secondly, the bit error rate (BER) expression for
(DCSK), pulse position modulation (PPM), hybrid modulation, the considered system is obtained analytically and validated
bit error rate (BER). then by simulations. Results show that the proposed system
I. I NTRODUCTION has better BER performance than conventional M-DCSK,
multidimensional CIM-DCSK and CCI-DCSK schemes.
HAOTIC communication has gained increasing attention
C as it can be used widely in spread spectrum com-
munication systems due to its low-power, low complexity,
This letter is organized as follows. Section II presents
the system model of PPM-DCSK system. The calculation of
the BER expressions of the proposed system is derived in
and excellent anti-fading capabilities [1], [2]. Differential
Section III. Simulation results and discussions are presented
chaos shift keying (DCSK), a modulation scheme proposed
in Section IV. Section V concludes this letter.
for chaotic communications [3], is characterized by its
simple transceiver configuration and excellent performance
II. S YSTEM M ODEL
over multipath fading channels [4], only requiring a simple
non-coherent demodulator without channel estimators and The transmitter of proposed system is shown in Fig. 1,
equalization [5]. Some variants of DCSK, such as permuta- where the total transmitted bits are mc +1 and symbol duration
tion index DCSK(PI-DCSK) [6] and code index modulation is (2mc +1)R. In the proposed system, mc bits are mapped into
DCSK (CIM,CCI-DCSK) [1], [7] which used index modu- a PPM position. The transmitted signal sl can be expressed as
lation [8], [9] have also been proposed. A multi-resolution
sl = [ cx , bl sPPM cx ] (1)
M-ary DCSK with M-ary phase-shift-keying (MPSK) con-
stellation is proposed based on quadrature chaotic shift reference information−bearing
keying (QCSK) [10] which offers better BER performance
where cx is a R-length
chaotic signal, bl ∈ {−1, 1} is
by changing the structure of constellation.
information bit, is Kronecker operator. In the proposed
Due to its simplicity and high performance, pulse position
system, the information bit is transmitted on one position in
modulation (PPM) is also widely used in spread spectrum
the PPM frame. The position is determined by mapping bits.
Manuscript received August 29, 2018; accepted September 10, 2018. Date sPPM = [0, 0, . . . , 1al , . . . , 0]1×P (P = 2mc ) in the PPM sig-
of publication September 19, 2018; date of current version April 9, 2019. This nal, where 1al represents that the alth position of sPPM is 1.
work was supported by the National Natural Science Foundation of China
under Grant 61671395. The associate editor coordinating the review of this al is a position index modulation symbol which is converted
paper and approving it for publication was J. Coon. (Corresponding author: by mapping bits.
Lin Wang.) Assuming that the transmitted signal is corrupted by a
M. Miao, L. Wang, and W. Xu are with the Department of
Communication Engineering, Xiamen University, Xiamen 361005, multipath Rayleigh fading channel, the received signal can be
China (e-mail: meiyuanmiao@foxmail.com; wanglin@xmu.edu.cn; written as
xweikai@xmu.edu.cn).
L
M. Katz is with the Centre for Wireless Communications, University of
Oulu, 90014 Oulu, Finland (e-mail: marcos.katz@ee.oulu.fi). rl = αl δ(t − τl ) ⊗ sl + nl , (2)
Digital Object Identifier 10.1109/LWC.2018.2871137 l=1
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
MIAO et al.: HYBRID MODULATION SCHEME COMBINING PPM WITH DCSK MODULATION 341
Fig. 1. Block diagram of PPM-DCSK transmitter. Fig. 2. Block diagram of PPM-DCSK receiver.
μ2 calculated as
2σ12 − 2σ12 −μ1 %
μ|Im | = e 1 − μ1 erf Es N0 γ, (14) ( 1
π 2σ12
= 1 4 2R − 2
Pe = erfc + 2 . (19)
2 rs rs
where
⎛ ⎞ total system Finally, substituting these above into (10), we
⎛ ⎞
−⎝ 4
1
2R
⎠ √ obtain the instantaneous BER of PPM-DCSK over Rayleigh
1 R rs + r 2 rs 1
γ= + e s − erf ⎝− 4 2R
⎠, fading channel, and the multipath Rayleigh fading channel can
2π 4πrs 2 rs + rs2 be given as
(15) ∞
here and Pmul = Pe · f (rs )drs , (20)
0
2
σ|Im|
= μ21 + σ12 − μ2|Im |
where f (rs ) is the PDF of rs which can be found in [5].
rs 1 R
= Es N0 ( + + − γ2) (16)
4 2 4rs IV. S IMULATION R ESULTS AND D ISCUSSION
ρ In this section, simulations are carried out to evaluate
Assume that X1 = max{|Im |}, m = 1, 2, . . . , P − 1. And the performance of the proposed system over AWGN and
the position of PPM detection error probability is calculated multipath Rayleigh fading channels. These results are then
by, contrasted with the theoretical results derived in previous
analysis on both AWGN and multipath Rayleigh fading chan-
Ped = 1 − Pr {Y ≥ X } nel. In all figures, SF represents the spreading factor and
∞ mc is the number bits mapped into one position of the
= [1 − Pr {y ≥ X }]f|Im | (y)dy PPM modulation. In multipath Rayleigh fading channel, three
0 paths L = 3, are considered, having equal average power
⎡ ⎡ ⎛ ⎞⎤P −1 ⎤ gain E[α21 ] = E[α22 ] = E[α32 ] = 1/3, with path delays
∞
1 ⎢ y ⎠⎦ ⎥ τ1 = 0, τ2 = 1, τ3 = 2. To quickly illustrate, mc in PPM-
= ⎣1 − ⎣erf ⎝ ⎦ DCSK means modulated bits number of PPM parts and the
2
2πσ|I | 2
2σ 2
⎧
m 0
⎫ whole number of modulated bits is mc + 1 while the whole
⎪
⎨ −
(y−μ |Im | ) 2 (y+μ 2
|Im | ) ⎪
⎬ number of M-DCSK is only mc .
−
2σ 2 2σ 2 The analytical and simulated results for PPM-DCSK are
× e |Im | +e |Im | dy. (17)
⎪
⎩ ⎪
⎭ shown in Fig. 3, where the analytical results match the simula-
tions at high SNR. As can be also seen, BER always decreases
Let μ = √ y , then the position detection error probability as mc increases and as SF decreases in AWGN while less
Es N0
is derived as influence in multipath fading. The reason for this behavior
∞$ '(P −1 )
is that more bits in the same symbol energy are transmitted
% &
1 μ with increasing mc , which means each transmitted bit needs
Ped = √ 1 − erf √ less energy. Fig. 4 shows performance comparisons of PPM-
2πρ 2λ
0 DCSK, CIM-DCSK, CCI-DCSK and M-DCSK over AWGN
* (y−γ)2 (y+γ)2
+
− 2ρ − 2ρ channel. Note that the PPM-DCSK exhibits lower BER than
× e +e dμ, (18)
the others, moreover the BER of M-DCSK increases as mc
MIAO et al.: HYBRID MODULATION SCHEME COMBINING PPM WITH DCSK MODULATION 343
V. C ONCLUSION
In this letter, a novel hybrid modulation scheme com-
bining PPM and DCSK is considered. The scheme avoids
the problem of distinguishing inaccurately the constellation
points in M-DCSK systems especially in large values of M.
Fig. 4. Performance comparisons of PPM-DCSK, CIM-DCSK, CCI-DCSK,
M-DCSK over AWGN channel SF = 64.
In the proposed system, information bearing signal is car-
ried by PPM, which carries both DCSK modulated bit and
extra red mc bits by position modulation. The theoretical
BER expression were derived and validated by simulations. By
comparing the BER performance with conventional M-DCSK,
CIM-DCSK and CCI-DCSK, the numerical results show that
the proposed hybrid PPM-DCSK system has superior BER
performance over AWGN and multipath Rayleigh fading
channel.
R EFERENCES
[1] Y. Tan, W. Xu, T. Huang, and L. Wang, “A multilevel code shifted
differential chaos shift keying scheme with code index modula-
tion,” IEEE Trans. Circuits Syst. II, Exp. Briefs, to be published,
doi: 10.1109/TCSII.2017.2764916.
Fig. 5. BER performance comparisons of PPM-DCSK, CIM-DCSK, CCI- [2] W. Xu, Y. Tan, F. C. M. Lau, and G. Kolumbán, “Design and
DCSK, M-DCSK over Rayleigh multipath fading channel with SF = 64, optimization of differential chaos shift keying scheme with code index
L = 3. modulation,” IEEE Trans. Commun., vol. 66, no. 5, pp. 1970–1980,
Mar. 2018.
[3] G. Kolumbán, B. Vizvári, W. Schwarz, and A. Abel, “Differential chaos
shift keying: A robust coding for chaos communication,” in Proc. NDES,
Seville, Spain, 1996, pp. 87–92.
[4] M. Dawa, G. Kaddoum, and Z. Sattar, “A generalized lower bound on
the bit error rate of DCSK systems over multi-path Rayleigh fading
channels,” IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 65, no. 3,
pp. 321–325, Mar. 2018.
[5] G. Cheng, L. Wang, W. Xu, and G. Chen, “Carrier index differen-
tial chaos shift keying modulation,” IEEE Trans. Circuits Syst. II, Exp.
Briefs, vol. 64, no. 8, pp. 907–911, Aug. 2017.
[6] M. Herceg, G. Kaddoum, D. Vranješ, and E. Soujeri, “Permutation
index DCSK modulation technique for secure multiuser high-data-rate
communication systems,” IEEE Trans. Veh. Technol., vol. 67, no. 4,
pp. 2997–3011, Apr. 2018.
[7] M. Herceg, D. Vranješ, G. Kaddoum, and E. Soujeri, “Commutation
code index DCSK modulation technique for high-data-rate communica-
tion systems,” IEEE Trans. Circuits Syst. II, Exp. Briefs, to be published,
Fig. 6. Effect of the number of modulated bits on BER at PPM-DCSK, CIM- doi: 10.1109/TCSII.2018.2817930.
DCSK, CCI-DCSK, M-DCSK over AWGN and Rayleigh multipath fading [8] G. Kaddoum, M. F. A. Ahmed, and Y. Nijsure, “Code index
channel with L = 3, SF = 64. modulation: A high data rate and energy efficient communica-
tion system,” IEEE Commun. Lett., vol. 19, no. 2, pp. 175–178,
Feb. 2015.
[9] G. Kaddoum, Y. Nijsure, and H. Tran, “Generalized code index
increases, while PPM-DCSK decreases as mc increases. The modulation technique for high-data-rate communication systems,”
gain of PPM-DCSK over M-DCSK is at least about 4 dB as IEEE Trans. Veh. Technol., vol. 65, no. 9, pp. 7000–7009,
Sep. 2016.
mc increase, and over CIM-DCSK and CCI-DCSK is about [10] L. Wang, G. Cai, and G. R. Chen, “Design and performance analysis
0.3 dB, 2.5 dB at 10−4 . The performance of PPM-DCSK is of a new multiresolution M -ary differential chaos shift keying com-
significantly improved over M-DCSK. As shown in Fig. 5, munication system,” IEEE Trans. Wireless Commun., vol. 14, no. 9,
pp. 5197–5208, Sep. 2015.
BER performance of PPM-DCSK on multipath fading chan- [11] L. Bosotti and G. Pirani, “A PAM-PPM signalling format in optical
nel outperforms those of the other schemes when SF = 64. fibre digital communications,” Opt. Quantum Electron., vol. 11, no. 1,
All schemes have similar increasing BER behavior, PPM- pp. 71–86, Jan. 1979.
[12] Y. Dai and X. Dong, “Hybrid PPM-BPSK for transmitted reference pulse
DCSK has less increasing tendency then others when mc cluster systems in UWB and 60-GHz channels,” IEEE Wireless Commun.
increasing. Note the good performance at high mc of the Lett., vol. 3, no. 6, pp. 657–660, Dec. 2014.
344 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In this letter, we consider a Rician fading envelope contributions brought about by shadowing. Hence, several
which is impacted by dual shadowing processes. We conveniently composite fading models have been proposed which address
refer to this as the double shadowed Rician fading model which these shortcomings. The shadowing in these is LOS if
can appear in two different formats, each underpinned by a dif- the dominant component of the envelope is shadowed, and
ferent physical signal reception model. The first format assumes
a Rician envelope where the dominant component is fluctu- multiplicative when the total power of the dominant (if
ated by a Nakagami-m random variable (RV) which is preceded present) and scattered components are shadowed. A number of
(or succeeded) by a secondary round of shadowing brought about multiplicative shadow fading models were proposed in [3]–[6].
by an inverse Nakagami-m RV. The second format considers These include the Nakagami-m/gamma [3], κ-μ/gamma [4],
that the dominant component and scattered waves of a Rician η-μ/gamma [5], κ-μ/inverse gamma and η-μ/ inverse gamma
envelope are perturbed by two different shadowing processes. models [6].
In particular, the dominant component experiences variations Here, we focus on the shadowed Rician model [7] which
characterized by the product of a Nakagami-m and an inverse
Nakagami-m RV, whereas the scattered waves are subject to fluc- considers a Rician envelope in which the LOS is perturbed by
tuations influenced by an inverse Nakagami-m RV. Using the shadowing shaped by a Nakagami-m random variable (RV).
relationship between the shadowing properties of the two for- This model has good analytical properties [8] and provides an
mats, we develop unified closed-form and analytical expressions excellent fit to data obtained from land mobile satellite and
for their probability density function, cumulative distribution underwater acoustic channels [7], [9]. Motivated by this, we
function, moment-generating function and moments. All derived introduce the double shadowed Rician fading model which
expressions are validated through Monte Carlo simulations and
can appear in two formats. The first format considers a Rician
reduction to a number of special cases.
signal in which the LOS undergoes variations influenced by
Index Terms—Composite fading, fading channels, inverse a Nakagami-m RV. It also assumes that the root mean square
Nakagaim-m distribution, shadowed Rician model. (rms) power of the dominant component and scattered waves
undergo a secondary round of shadowing shaped by an inverse
Nakagami-m RV. The second format assumes that the domi-
I. I NTRODUCTION nant component of a Rician envelope undergoes fluctuations
EVERAL statistical distributions have been proposed to characterized by the product of a Nakagami-m and an inverse
S characterize fading in wireless channels [1]. Shadowing
is commonly modeled using the lognormal distribution [1]
Nakagami-m RV, whilst the scattered waves are fluctuated by
an inverse Nakagami-m RV. The PDF, cumulative distribu-
whilst multipath fading is described by the Rayleigh, Rice, tion function (CDF), moment generating function (MGF) and
Nakagami-m, and more recently κ-μ and η-μ [2] distribu- moments are derived which are coincidentally identical for
tions. Nevertheless, these models are unable to account for both formats, differing only in the interpretation of the under-
fluctuations of the line-of-sight (LOS) or scattered signal lying physical phenomena. These results are then used to
obtain amount of fading (AF) and outage probability (OP).
Manuscript received June 13, 2018; revised September 11, 2018; accepted
September 13, 2018. Date of publication September 24, 2018; date of II. T HE P HYSICAL M ODEL
current version April 9, 2019. This work was supported in part by the
U.K. Engineering and Physical Sciences Research Council under Grant The PDF of the shadowed Rician fading model [7] is
EP/L026074/1, in part by the Department for the Economy, Northern Ireland md
2
2σ md −x 2
x 2 2
d x
under Grant USI080, and in part by Khalifa University of Science and fX(x ) = e 2σ2 1F1 md ; 1; 2 2
Technology under Grant 8474000122 and Grant 8474000137. The associate 2σ 2 md +d 2 σ2 2σ (2σ md +d 2 )
editor coordinating the review of this paper and approving it for publication
was A. Kammoun. (Corresponding author: Simon L. Cotton.)
(1)
N. Simmons and S. L. Cotton are with the Institute of Electronics, where, 1F1 (·; ·; ·) denotes the confluent hypergeometric func-
Communications and Information Technology, Queen’s University
Belfast, Belfast BT3 9DT, U.K. (e-mail: nsimmons01@qub.ac.uk;
tion [10, eq. (9.210.1)], md denotes the shape parameter of
simon.cotton@qub.ac.uk). the Nakagami-m RV, 2σ 2 is the average power of the scat-
C. R. N. da Silva and M. D. Yacoub are with the Wireless tered component, and d 2 is the average power of the LOS
Technology Laboratory, School of Electrical and Computer Engineering,
component. Here, k = 2σ d 2 is the Rician k parameter, and
University of Campinas, Campinas 13083-970, Brazil (e-mail: √ 2
carlosrn@decom.fee.unicamp.br; michel@decom.fee.unicamp.br). x̂ = E[X 2 ] = 2σ 2 + d 2 represents the rms power of x,
P. C. Sofotasios is with the Department of Electrical and where E[ · ] is the expectation operator. We rewrite (1) as
Computer Engineering, Khalifa University of Science and Technology,
Abu Dhabi 127788, UAE, and also with the Department of Electronics and mdmd 2x (1+k ) −(1+k ) x 22 k (1+k )x 2
Communications Engineering, Tampere University of Technology, 33720 fX(x ) = e x̂ 1F1 md ; 1; 2 .
Tampere, Finland (e-mail: p.sofotasios@ieee.org). (k +md )md x̂ 2 x̂ (md +k )
Digital Object Identifier 10.1109/LWC.2018.2871677 (2)
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
SIMMONS et al.: DOUBLE SHADOWING THE RICIAN FADING MODEL 345
The first format of the double shadowed Rician model where 2 F1 (·, ·; ·; ·) is the Gauss hypergeometric function [10].
assumes a Rician fading channel which undergoes LOS Proof: See Appendix A.
shadowing followed by a secondary round of composite shad- Letting γ represent the instantaneous signal-to-noise-ratio
owing or vice versa. Physically, this may arise when the signal (SNR) of the double shadowed Rician fading model, the PDF
power delivered through the direct path between the transmit- of its instantaneous SNR, fγ (γ), is obtained from the envelope
ter and receiver is subject to varying levels of shadowing, PDF in (8) via a transformation of variables (r = γ r̂ 2 /γ̄),
whilst further shadowing of the received power (combined md
scattered multipath and LOS) is due to obstacles moving in γ̄ ms ms (ms − 1)ms (1 + k ) md
fγ (γ) =
the vicinity of either the transmitter or receiver. Its signal (γ(1 + k ) + (ms − 1)γ̄)ms +1 md + k
envelope, R, is
k (1 + k )γ
2 × 2F1 md , ms + 1; 1;
R 2 = A2 (I + ξμx )2 + Q + ξμy (3) (md + k )(γ(1 + k ) + (ms − 1)γ̄)
(9)
where I and Q are mutually independent Gaussian random
processes
2
with mean E[I ] = E[Q] = 0 and variance E I 2 = where γ̄ = E[γ] denotes the corresponding average SNR.
E Q = σ 2 , μx and μy are the mean values of the in-phase Lemma 1: For k, md , γ̄, γ ∈ R+ , and ms > 1 the CDF
ξ is
and the quadrature phase components, respectively. In (3), of the double shadowed Rician fading model, Fγ (γ), can be
a Nakagami-m RV with shape parameter md where E ξ 2 = 1, obtained such that Fγ (γ) = 0γ fγ (t)dt, as
and A is an inverse
Nakagami-m RV with shape parameter ms ,
where E A2 = 1, whose PDF is given by ∞ i
md md k (md )i (i + 1)ms
2(ms − 1)ms − (ms −1) Fγ (γ) =
fA (α) = e α2 . (4) md + k md + k Γ(ms )Γ(i + 2)
i=0
Γ(ms )α2ms +1
γ(1+k ) i+1 γ(1+k )
Here, Γ(·) represents the Gamma function [10, eq. (8.310.1)]. × 2 F1 i +1, i +ms +1; i +2; − .
The second format of the double shadowed Rician model (ms − 1)γ̄ (ms − 1)γ̄
assumes a Rician faded signal in which the dominant com- (10)
ponent and scattered waves are subject to two different Γ(x +n )
shadowing processes. More precisely, the dominant compo- where (x )n = Γ(x ) denotes the Pochhammer sym-
nent experiences variations characterized by the product of a bol [10]. When (ms − 1)γ̄ > γ(1 + k ),
Nakagami-m and an inverse Nakagami-m RV, whilst the scat-
tered waves are subject to fluctuations influenced by an inverse mdmd ms γ(1 + k )
Fγ(γ) =
Nakagami-m RV. Its signal envelope, R, is (md +k )md (ms −1)γ̄
2
R 2 = (AI + B μx )2 + AQ + B μy (5) 2,1,0 ms + 1, 1; md ; −; k (1+k )γ −(1+k )γ
×F1,1,0 ,
2; 1; −; (md +k )γ̄(ms −1) γ̄(ms −1)
where B = Aξ, A and ξ are as defined above. It is worth high-
(11)
lighting that as shown in [11], B 2 follows a Fisher-Snedecor F
distribution [11]. Now, substituting for B in (5), we obtain (3). ·,·,· ·, ·; ·; ·;
where F·,·,· ·, · denotes the Kampé de Fériet
Note that although (3) and (5) are mathematically identical, ·; ·; ·;
their physical meanings differ as explained above. function [12]. On the contrary, when (ms − 1)γ̄ < γ(1 + k ),
md
III. S TATISTICAL C HARACTERISTICS md 1; md ; 0; k −(ms −1)γ̄
Fγ(γ) = F 1,1,1 ,
Exploiting the mathematical relationship above and select- (md +k )md 0,1,1 −; 1; 1 − ms ; md +k γ(1+k )
ing (3) as our starting point, the distribution of the received 1,1,1 ms + 1; md ; ms ; k −(ms −1)γ̄
− ζ F0,1,1 , (12)
signal envelope, R, in a double shadowed Rician channel can −; 1; 1 + ms ;md +k γ(1+k )
be obtained by determining the conditional probability
((m −1)γ̄)ms Γ(m +1)
∞ where ζ = m s(γ(1+k ))ms Γ(m
s
s s)
fR (r ) = fR|A (r |α)fA (α)d α (6) Proof: See Appendix A.
0 Lemma 2: For k, md , γ̄, γ ∈ R+ , and ms > 1 the MGF
where of the double shadowed Rician fading model,
Mγ (s), can be
r2
obtained such that Mγ (s) E[e −sγ ] = 0∞ e−sγ fγ (γ)d γ,
−(1+k ) k (1+k )r 2
2r(1+k )e α2 r̂ 2 1F1 md ; 1; α2 r̂ 2 (m +k )
fR|A(r |α) = d
.
md−md (md +k )md α2 r̂ 2 md md k
(7) Mγ(s) = ψ1 1, md, 1, 1−ms ; ,ζ
md + k md +k
Theorem 1: For k, md , r̂ 2 , r ∈ R+ and ms > 1 the PDF of ζ ms Γ(−ms ) k
+ ψ1 1+ms , md , 1, 1 +ms ; ,ζ (13)
the double shadowed Rician fading model can be written as B(ms , 1) md +k
md
2r r̂ 2ms ms (ms − 1)ms (1 + k ) md γ̄(m −1)s
s
fR (r ) = m +1 where ζ = 1+k and ψ1 (·, ·, ·, ·; ·, ·) is the Humbert ψ1
(r 2 (1 + k ) + (ms − 1)r̂ 2 ) s md + k
function [13].
k (1 + k )r 2 Proof: See Appendix B.
× 2 F1 md , ms + 1; 1;
(md + k )(r (1 + k ) + (ms − 1)r̂ 2 )
2 Lemma 3: For k, md , γ̄, γ ∈ R+ , and ms > 1 the n-th
(8) order moment of the double shadowed Rician fading model,
346 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
E[γ n ], can be obtained such that E[γ n ] 0∞ γ n fγ (γ)d γ, as
Γ(ms −n)Γ(1+n)2 F1 md , n + 1; 1; m k+k
E[γ n ] = −m
d
.
md d (md +k )md Γ(ms )(1+k )n [(ms −1)γ̄]−n
(14)
Proof: See Appendix B.
B. Outage Probability
Corollary 2: For k, md , γ̄ ∈ R+ and ms > 1 the OP of the
double shadowed Rician fading model, can be obtained such
that POP (γth ) P [0 ≤ γ ≤ γth ] = Fγ (γth ), as
∞ i
md md k (md )i (i + 1)ms
POP (γth ) =
md + k md + k Γ(ms )Γ(i + 2)
i=0
i+1
γth (1+k ) γth (1+k )
× 2 F1 i +1, i +ms +1; i +2; −
(ms − 1)γ̄ (ms − 1)γ̄
(16) Fig. 2. The AF in double shadowed Rician fading channels for a range of
ms and md when k = 0.5 and 20.6.
where γth is the threshold SNR.
Proposition 1: For (ms − 1)γ̄(md + k ) > γth k (1 + k ), the
truncation error, T , for the infinite series in (16) is given as
γ (1+k ) γth (1+k )
T ≤ 2 F1 T0 +1, T0 +ms +1; T0 +2; − th
(ms −1)γ̄ (ms −1)γ̄
γth k (1+k )
×ms 2 F1 md , 1+ms ; 2, . (17)
(ms − 1)γ̄(md +k )
Proof: See Appendix C.
∞ md i model,” in Proc. IEEE Int. Conf. Wireless Inf. Technol. Syst., Honolulu,
md k (md )i (i + 1)ms HI, USA, Aug. 2010, pp. 1–4.
Mγ (s) = [6] S. K. Yoo et al., “The κ-μ/inverse gamma and η-μ/inverse
md + k md + k Γ(ms )i!
i=0 gamma composite fading models: Fundamental statistics and
ms empirical validation,” IEEE Trans. Commun., to be published,
γ̄(ms −1)s γ̄(ms −1)s
B (ms , i +1) 1F1 i +1; 1−ms ; + doi: 10.1109/TCOMM.2017.2780110.
1+k 1+k [7] A. Abdi, W. C. Lau, M.-S. Alouini, and M. Kaveh, “A new simple model
for land mobile satellite channels: First- and second-order statistics,”
γ̄(ms −1)s
Γ(−ms ) 1F1 i +ms +1; 1+ms ; . (18) IEEE Trans. Wireless Commun., vol. 2, no. 3, pp. 519–528, May 2003.
1+k [8] J. F. Paris, “Closed-form expressions for Rician shadowed cumulative
distribution function,” Electron. Lett., vol. 46, no. 13, pp. 952–953,
By substituting the Kummer confluent hypergeometric func- Jun. 2010.
tion with [14, eq. (07.20.02.0001.01)] in (18), using the [9] F. Ruiz-Vega, M. C. Clemente, P. Otero, and J. F. Paris, “Ricean shad-
Pochhammer symbol identities, and the definition of the owed statistical characterization of shallow water acoustic channels for
Humbert ψ1 function [13], we obtain (13). wireless communications,” arXiv preprint arXiv:1112.4410, 2011.
[10] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
The n-th order moment is obtained by first substituting [14, Products, 7th ed. New York, NY, USA: Academic, 2007.
eq. (07.23.02.0001.01)] in (9), then substituting the resul- [11] S. K. Yoo et al., “The Fisher–Snedecor F distribution: A simple and
tant expression in E[γ n ] = 0∞ γ n fγ (γ)d γ to obtain an accurate composite fading model,” IEEE Commun. Lett., vol. 21, no. 7,
integral identical to [10, eq. (3.194.3)]. Now, using [14, pp. 1661–1664, Jul. 2017.
[12] Listing of Mathematical Notations, Wolfram Res., Inc., Champaign,
eq. (07.23.02.0001.01)] and simplifying, yields (14). IL, USA, 2017, Accessed: Aug. 31, 2018. [Online]. Available:
Substituting n = 1 and 2 into (14), we obtain the first http://functions.wolfram.com/Notations/5
and second moments of the double shadowed Rician model. [13] P. Humbert, “Sur les fonctions hypercylindriques,” Comptes Rendus
Hebdomadaires Des Séances De L’Académie Des Sci., vol. 171,
Utilizing these in the AF formulation (see corollary 1), pp. 490–492, 1920.
followed by using [14, eq. (07.23.02.0001.01)] and finally [14] (2016). Wolfram Research, Inc. Accessed: Feb. 3, 2017. [Online].
simplifying the resultant expression, yields the AF in (15). Available: http://functions.wolfram.com/id
348 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Backscatter communications have been widely encoding) and Miller modulation have been widely adopted
applied in wirelessly powered networks. Energy constraint forces in the RFID systems [2]. We found that there exists Energy
the nodes to only backscatter a few bits once, causing backscat- Consumption Disparity (ECD) between transmitting/receiving
ter communications unable to be applied in the applications that
bit 0 and bit 1 under the backscatter communications adopt-
require to deliver more data (e.g., an image). It is significant to
save energy for backscatter communications so that the nodes ing FM0 due to an additional mid-symbol phase inversion in
can deliver more data with the limited energy. In this letter, the backscattering bit 0 [4]. That is, the energy consumed in deliv-
energy-efficient code based backscatter communication (CBBC) ering a single bit 0 almost suffices to deliver two bit 1 s. In fact,
is proposed, which makes use of the energy consumption dis- ECD is also present in Miller modulation. To cope with the
parity between transmitting/receiving bit 0 and bit 1 in the ECD, the Energy-Efficient Data Delivery Scheme (EEDDS)
existing backscatter communications. The energy-efficient pre- was proposed in our previous work [4] that reduces energy
fix codebook is derived from the formulated energy consumption
minimization problem. In the CBBC, the codebook is shared consumption by using a codebook shared by the sender and
by the sender and the receiver of a backscatter link, and the the receiver. Under the EEDDS, the sender breaks original
sender breaks the original bit stream into equal-length blocks bit stream into multiple m-bit data blocks, finds the corre-
and delivers the energy-consuming blocks by using their corre- sponding codewords of the blocks, and then transmits the
sponding codewords from which the receiver decodes the original codewords from which the receiver recovers the original data.
data. The experiments show that the proposed CBBC can save There exist the following shortcomings in the EEDDS. Firstly,
energy for backscatter communication.
for a smaller ECD, say less than 2, the EEDDS almost does
Index Terms—Backscatter communication, wirelessly powered not save energy when m<12. That is, it requires a greater m to
networks, energy conservation, prefix code. gain energy saving. Thus, the number of the codewords (i.e.,
codebook size) in the EEDDS may be too large to be contained
in the nodes’ tightly-constrained memory since the codebook
I. I NTRODUCTION size is 2m , which makes the EEDDS hard to be implemented.
HE NODES in wirelessly powered networks are powered In fact, it is unnecessary to encode all the m-bit data blocks
T by dedicated wireless chargers or harvest energy from
the environment. Backscatter communications, emerged as a
because some of them contain many bit 1 s, which can be
directly delivered with less energy consumption. Secondly,
promising solution to achieve green communication for future each codeword has length greater than m, which causes each
Internet of Things (IoT) [1], have been widely applied in deliv- data block to be delivered with a longer delay. Hence, it is
ering data in wirelessly powered networks. The applications critical to develop an energy-efficient and easy-implementation
of backscatter communications include radio frequency iden- scheme for the backscatter communications used in the wire-
tification (RFID) systems [2], the backscatter sensors [3], and lessly powered networks. This is the main motivation of this
more. In the RFID systems, tags are powered by a reader and letter.
transmit their data to the reader through backscatter communi- The main contributions of this letter are as follows. 1) We
cations. Energy constraint causes the nodes to only backscatter propose the prefix code based backscatter communication
a few bits once. This prevents backscatter communications (CBBC) scheme, which supports to directly deliver the origi-
from being applied in the applications that require to deliver nal data blocks with number of bit 0 s less than an encoding
more data, such as backscattering image. Therefore, it is threshold in addition to delivering codewords. To inform the
important to let the energy-restricted nodes deliver more data receiver of whether a data block is backscattered using its
through backscatter communications. codeword or not, we let the sender adopt two different data
In backscatter communications, data are transmitted bit rates. 2) We formulate the energy consumption minimization
by bit. FM0 baseband encoding (a bi-phase space baseband problem to derive the codebook, in which only the blocks with
the number of energy-consuming bits being equal to or greater
Manuscript received July 31, 2018; revised September 15, 2018; accepted than the threshold are encoded. This leads to a smaller code-
September 24, 2018. Date of publication September 28, 2018; date of current book size so that the codebook can be entirely stored in the
version April 9, 2019. This work was supported by the National Natural nodes’ memory. That is, we emphasize the number of energy-
Science Foundation of China under Grant 61432015, Grant 61772470, Grant
61872322, and Grant 61672465. The associate editor coordinating the review
consuming bits in designing the codebook and only encode
of this paper and approving it for publication was H. T. Dinh. (Corresponding the blocks that bring in energy saving. 3) We experiment the
author: Yi-Hua Zhu.) CBBC on wireless identification sensing platform (WISP) [5],
The authors are with the School of Computer Science and Technology, a programmable, sensing and computationally enhanced plat-
Zhejiang University of Technology, Hangzhou 310023, China (e-mail:
yhzhu@zjut.edu.cn; kkchi@zjut.edu.cn; txz@zjut.edu.cn). form compatible with GS1 RFID protocols [2]. The CBBC
Digital Object Identifier 10.1109/LWC.2018.2872538 extremely extends the EEDDS and outperforms the EEDDS
2162-2345 c 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/
redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHANG et al.: ENERGY-EFFICIENT PREFIX CBBC FOR WIRELESSLY POWERED NETWORKS 349
in terms of the number of bits transmitted by the WISP with the ECD between transmitting/receiving bits 0 and 1 such that
the same amount of energy. et0 = αet1 and er 0 = αer 1 . Denoting the energy consump-
The remainder of this letter is organized as follows. We tions of delivering bits 0 and 1 by ε0 and ε1 , respectively, we
outline the CBBC in Section II, design the prefix codebook have ε0 = et0 + er 0 , ε1 = et1 + er 1 , and
in Section III, evaluate the performance of the CBBC in
Section IV, and conclude this letter in Section V. ε0 = αε1 . (1)
For a given m, there are 2m different m-bit data blocks, from
II. T HE CBBC which we choose and encode the ones that have the number
of bit 0 s being equal to or greater than the encoding threshold
The proposed CBBC consists of procedures of Proc_sender
z. Therefore, the codebook contains the number of codewords
and Proc_receiver, which are applied in the sender and the
as follows:
receiver, respectively. The CBBC delivers m bits from the
m
sender to the receiver once a time. An m-bit block is transmit- m
Nc = . (2)
ted using a codeword in the designed codebook if it contains
u=z
u
z or more bit 0 s, where z is a preset constant called encod-
ing threshold. Under the CBBC, the sender uses two different Necessarily, there exists one-to-one corresponding between
data rates r1 and r2 to transmit codewords and uncoded m-bit the encoded m-bit blocks and the codewords in the codebook.
blocks, respectively. The sender and the receiver share the Then, the number of m-bit blocks without being encoded is
designed codebook, which will be derived in Section III. 2m − Nc . Accordingly, the original m-bit blocks can be classi-
The procedure Proc_sender contains the following steps: fied into encoded and uncoded ones. Assume the probabilities
Step 1. The sender sets the optimal z, denoted by z ∗ , which of bit 0 and bit 1 occurring in the original data are identical.
is determined by the optimization problem (OP) in (11). Then, the probabilities of the sender transmitting a codeword
Step 2. The sender divides the original data into multiple and an uncoded m-bit block can be expressed in (3) and (4),
m-bit blocks. respectively.
Step 3. For each m-bit block, if it contains the number of m
Nc 1 m
bit 0 s equal to or greater than z ∗ , the sender searches for its Pc = m = m , (3)
corresponding codeword in the designed codebook, and then 2 2 u=z u
transmits the found codeword using data rate r1 . Otherwise, m
1 m
it transmits the original m-bit block using data rate r2 . Puc = 1 − Pc = 1 − m . (4)
2 u=z u
The procedure Proc_receiver is as follows. When the
receiver receives a packet, it considers the received packet as
a codeword if the packet is received with data rate r1 and as A. Energy Consumption of the Common Backscatter
an original m-bit block otherwise. For a received codeword, Communication
the receiver searches in the codebook for its corresponding We first calculate the energy consumption for the common
original m-bit block. backscatter communication (ComBC). For a given m, there are
The CBBC is illustrated in Fig. 1. In the figure, the sender totally 2m m-bit blocks, in which the numbers of bit 0 s and
can deliver data either in the coding mode or in the non-coding bit 1 s are each equal to (2m m)/2 = m2m−1 . Hence, the
mode. For instance, blocks 1 and 4 are directly delivered, average energy consumption of the sender and the receiver in
whereas blocks 2, 3, and 5 are delivered via the corresponding delivering one m-bit block under the ComBC is
codewords in the designed codebook.
m2m−1 ε0 + m2m−1 ε1 1
Ecbc = m = m(1 + α)ε1 . (5)
2 2
III. C ODEBOOK D ESIGN
We use et0 and et1 to represent the energy consumptions of B. Energy Consumption of Delivering Uncoded Data Blocks
transmitting bit 0 and bit 1, respectively. In addition, er 0 and Under the CBBC
er 1 stand for the energy consumptions of receiving bit 0 and Before deriving the codebook, we consider the energy con-
bit 1, respectively. Considering ECD is a constant for a given sumption under the CBBC without using any codebook. With
backscatter communication, we use constant α(>1) to reflect the CBBC, there are totally 2m − Nc uncoded m-bit blocks
350 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
deliveredwithout using codebook. These uncoded −1blocks We use OP (11) to generate prefix codebook C, where
−1 m
include zu=1 u u bit 0 s and (2m − Nc )m − zu=1 u mu notation ⊕ stands for the bitwise exclusive-OR (i.e., XOR)
bit 1 s. Thus, using (1), we obtain the average energy con- operation. In the OP, Constraint (11a) is obtained by substi-
sumption of the sender and the receiver in delivering one m-bit tuting (5) and (8) into (9). Eq. (11b) reflects prefix property,
block as follows where c i,j = min{|ci |, |cj |}. Eq. (11c) is the Kraft inequality.
−1 m −1 m
ε0 zu=1 u u + ε1 [m(2m − Nc ) − zu=1 u u ] Eqs. (11b) and (11c) are required in designing prefix code [6].
Euc = Constraint (11d) comes from tb ≤ θ/r using (10), where θ >0
(2m − Nc )
is a preset constant and r = max{r1 , r2 }. Here, 1/r repre-
z −1
ε1 (α − 1) m sents the duration of a bit transmitted using the higher rate in
= ε1 m + m u . (6)
(2 − Nc ) u r1 and r2 . Hence, Constraint (11d) aims to limit bit duration
u=1
tb so that it does not exceed θ multiples of the bit duration 1/r.
C. Average Energy Consumption of Delivering Encoded Aiming to accelerate delivering the original data over the link,
Data Blocks Under the CBBC we set r1 = r if N c
i=1 |c i | > m(2m − Nc ) (i.e., the num-
ber of bits in the codebook is larger than that in the uncoded
The CBBC applies an energy-efficient prefix codebook C
blocks), and r2 = r otherwise. In Constraint (11e), m̂ is the
that includes Nc prefix codewords. Here, prefix codewords are
upper bound of m to ensure that the derived codebook can be
also called prefix-free codewords, which have the property that
any codeword is not a prefix of the other ones. stored in nodes’ restricted memory. In (11f), z = 0 causes all
For a codeword ci ∈ C, we use |ci | to represent the the m-bit data blocks to be encoded, and z = m + 1 makes
number of the bits in ci such that the codeword can be no data block encoded.
expressed as ci = bi,1 bi,2 · · · bi,|ci | , where bi,j ∈ {0, 1}, i =
min Ecbbc
1, 2, . . . , Nc , j = 1, 2, . . . , |ci |. Thus, codeword ci has ni,1 =
|ci | |ci | w .r .t.: m, z , bi,k (i = 1, 2, . . . , Nc ; k = 1, 2, . . . , |ci |) (11)
j =1 bi,j bit 1 s and ni,0 = |ci | − j =1 bi,j bit 0 s. Hence,
delivering ci causes the sender and the receiver to expend s.t.:
energy of ni,1 ε1 + ni,0 ε0 = [|ci |α + (1 − α)ni,1 ]ε1 . As a ⎧ ⎡ ⎤
⎨ 1 Nc |ci |
result, energy consumption per codeword is
⎣|ci |α + (1 − α) bi,j ⎦
⎡ ⎤ ⎩ Nc
Nc |ci | i=1 j =1
ε1 ⎫
Ec = ⎣|ci |α + (1 − α) bi,j ⎦. (7) z −1 ⎬m
Nc α−1 m m
i=1 j =1 − m u −m
2 − Nc u ⎭ u
Therefore, the expected energy consumption in the CBBC u=1 u=z
using the prefix codebook is Ecbbc = Pc Ec + Puc Euc , i.e., z
−1
2m (α − 1) m
N ⎡ ⎤ + u < 2m−1 m(α − 1); (11a)
m |ci | 2m − Nc u
ε1 m c
⎣|ci |α + (1 − α)
u=1
Ecbbc = m bi,j ⎦ c i,j
2 Nc u=z u i=1 j =1
m m
bi,k ⊕ bj ,k = 0, ∀i , j ∈ {1, 2, . . . , Nc }(i = j );
m z
−1
2 − u=z u ε1 (α − 1) m k =1
+ m
ε1 m + m u . (8) (11b)
2 2 − Nc u=1 u
Nc
1
Clearly, to benefit from using codebook, we need ≤ 1; (11c)
i=1
2|ci |
Ecbbc < Ecbc . (9)
Nc
1 1 1 r1 θ
In the case when the original data blocks are delivered + m |ci | − Nc ≤ ; (11d)
r2 2 r1 m r2 r
through codewords, each m-bit block corresponds to a unique i=1
one of the Nc codewords in C. Hence, weobtain the aver- 1 ≤ m ≤ m̂; (11e)
age delivery time of an m-bit block as r11 ( N i=1 |c
c
i |/Nc ) so bi,k ∈ {0, 1}, i = 1, 2, . . . , Nc ; k = 1, 2, . . . , |ci |;
that the average delivery time per bit is mr1 Nc N1
i=1 |ci |,
c
z ∈ {0, 1, . . . , m + 1}. (11f)
where r1 is the data rate applied in the sender and the
receiver. Additionally, in the case when the original data blocks Noticing a prefix codebook corresponds to a binary tree, we
are delivered directly (the blocks are uncoded), the delivery use the pruning and expanding (PEO) operation [6] on binary
time per bit in the m-bit block is 1/r2 . Hence, the aver- trees to find the solution of OP (11). It has been proved in [6]
age transmission time per bit in the original block under the that the PEO is with complexity of O(Nc4 ) in the worst case. In
CBBC is solving the OP, the PEO is conducted for each pair of z and m,
Nc where z ∈ {0, 1, . . . , m + 1}, m ∈ {1, 2, . . . , m̂}. That is, the
1 1
tb = Pc |ci | + Puc times of the PEO being conducted is (m + 2)m̂ ≤ m̂(m̂ + 2).
mr1 Nc r2 Thus, the complexity of solving the OP is O(m̂ 2 Nc4 ). This
i=1
Nc
complexity is acceptable because m̂ is usually very small in
1 1 1 r1
= + m |ci | − Nc . (10) practice and the OP can be solved offline without strictly time
r2 2 r1 m r2 constraint.
i=1
ZHANG et al.: ENERGY-EFFICIENT PREFIX CBBC FOR WIRELESSLY POWERED NETWORKS 351
Fig. 2. PoINTB and PoRT vs θ when m̂ = 12. Fig. 3. PoINTB and PoRT vs r when m̂ = 12 and θ = 1.6.
Abstract—We investigate outage constrained robust multigroup communication systems cast into a multigroup multicast beam-
multicast beamforming for multi-beam satellite communica- forming optimization problem. In addition, the channel phase
tion systems with full frequency reuse. Based on a satellite uncertainty due to, e.g., the long propagation delays in satellite
downlink beam domain channel model with channel phase uncer- communication systems [6] motivates a robust beamforming
tainty taken into account, we first investigate robust multigroup design. Moreover, the per-beam power constraints should be
multicast beamforming with the aim to maximize the worst-case
outage signal-to-interference-plus-noise ratio under the outage
taken into account due to the limitation of on-board inter-beam
and the per-beam power constraints. We then cast the out- power sharing.
age constrained robust beamforming design into the convex Motivated by the above practical issues, we investigate
optimization framework with some approximation techniques. robust multigroup multicast beamforming for multi-beam
Simulation results show that the proposed robust multigroup satellite communications in this letter. Most of the previous
multicast beamformer can provide significant performance gains works on multigroup multicast precoding, see [7], usually
in terms of multicast rate and outage probability over the assumed perfect channel state information at the transmitter
conventional approach. (CSIT), which, however, is difficult to obtain in practical satel-
Index Terms—Multi-beam satellite communication systems, lite communication systems. For the imperfect CSIT case,
robust transmission, multigroup multicast beamforming, outage the expectation-based robust precoding designs were inves-
probability, channel state information (CSI). tigated in unicast [8] and multicast [9] multi-beam satellite
transmissions, respectively. In addition, the outage constrained
robust precoding designs were investigated in unicast [6]
I. I NTRODUCTION and single-group multicast [10] transmissions, respectively.
ULTI-BEAM satellite communication has received Satellite relaying systems were investigated in [17] and [18].
M extensive research interest recently due to its potential
to increase the satellite transmission rate and provide seam-
In this letter, we investigate outage-based robust multigroup
multicast beamforming for multi-beam satellite communica-
less connectivity in a wide coverage area [1]. Meanwhile, tion systems. Based on a satellite downlink beam domain
aggressive spectrum reuse among beams is desirable to achieve channel model with the channel phase uncertainty taken into
high transmission rates in multi-beam satellite communication account, we first investigate outage constrained robust multi-
systems [2], [3]. As a result, interference mitigation tech- group multicast beamforming subject to the per-beam power
niques become mandatory in multi-beam satellite systems constraints. We further cast the outage constrained robust
with aggressive frequency reuse to reduce the inter-beam beamforming design into the convex optimization framework
interference. Inter-beam interference mitigation can be per- with some formulation. Simulation results demonstrate the
formed at either the satellite side or the user side. In this letter, performance gains of our proposed robust approach over the
we focus on investigating linear beamforming performed at conventional approach.
the satellite side as it can effectively manage the inter-beam
interference with a relatively low complexity. II. S ATELLITE C HANNEL M ODEL
For beamforming design in multi-beam satellite communi- We consider a multi-beam satellite communication system
cation systems, several practical issues should be taken into with full frequency reuse, where Nt beams are utilized
account. Firstly, one beamformer is applied to several users in to simultaneously serve Nu single-antenna users. Consider
the same frame to cope with the framing structure of the cur- the multigroup multicast transmission where the number of
rent satellite standards, e.g., DVB-S2 [4] and DVB-S2X [5], multicast clusters is K = Nt and each user belongs to only
which leads the beamforming design in multi-beam satellite one cluster [11]. We denote K = {1, 2, . . . , K } as the total
Manuscript received August 30, 2018; accepted September 25, 2018. Date
cluster set and B = {1, 2, . . . , Nt } as the total beam set.
of publication October 1, 2018; date of current version April 9, 2019. This Denote Uk as the kth multicast cluster. We focus on the
work was supported in part by the National Natural Science Foundation of signal model in the beam domain. The signal received by the
China under Grant 61320106003, Grant 61761136016, Grant 61801114, Grant ith user in cluster Uk can be expressed as
61471113, Grant 61631018, and Grant 61521061, in part by the Natural
Science Foundation of Jiangsu Province under Grant BK20170688, in part yi = h H
i wk sk + hH i w s + ni , i ∈ Uk , (1)
by the National Science and Technology Major Project of China under Grant
2017ZX03001002-004, and in part by the Civil Aerospace Technologies =k
CNt ×1
Research Project under Grant D010109. The associate editor coordinating
the review of this paper and approving it for publication was J. Coon. where hi ∈ is the downlink beam domain channel
(Corresponding author: Wenjin Wang.) vector from the Nt beams to the ith user, wk ∈ CNt ×1 is the
The authors are with the National Mobile Communications beam domain precoder for cluster Uk , sk is the signal intended
Research Laboratory, Southeast University, Nanjing 210096, China for users in cluster Uk with unit power, and ni ∼ CN (0, N0 )
(e-mail: liyou@seu.edu.cn; ao_liu@seu.edu.cn; wangwj@seu.edu.cn;
xqgao@seu.edu.cn).
is the additive noise. Note
that theHradiation power from the
Digital Object Identifier 10.1109/LWC.2018.2872710 nth beam is given by [ K k =1 wk wk ]n,n .
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
YOU et al.: OUTAGE CONSTRAINED ROBUST MULTIGROUP MULTICAST BEAMFORMING FOR MULTI-BEAM SATELLITE COMMUNICATION SYSTEMS 353
The downlink channel vector between the satellite and the The outage constrained robust multigroup multicast beam-
ith user in the beam domain can be modeled as [2], [8] forming design formulated in (8) takes several practical issues
1 1 in multi-beam satellite communication systems into account.
hi = ψi bi2 ri2 exp{jθ i }, (2) Firstly, the multigroup multicast beamforming formulation nat-
urally embraces the framing structure of the existing satellite
where denotes the Hadamard product, ψi is the large communication standards [4], [5]. The per-beam power con-
scale fading coefficient, bi denotes the far-field beam radi- straints are also taken into account in the problem formulation.
ation pattern [12], ri represents the rain attenuation with In addition, a feasible solution to the problem F can guarantee
elements obeying the lognormal distribution [13], and θ i the quality of service of all users under the channel uncertainty,
denotes the channel phase vector with elements independently which is practically meaningful for satellite communications.
and uniformly distributed between 0 and 2π [3]. The problem F is in general difficult to handle. We first
For the ith user, the channel vector is estimated at instant decompose the problem F into a sequence of outage con-
t0 and fed back to the gateway. Then the CSIT is used at strained power minimization problem. In particular, for the
instant t1 after the propagation delays plus the processing predetermined SINR thresholds {γi }i , the outage constrained
delays [6], [14]. Due to the temporary invariance of amplitude, power minimization robust multigroup multicast beamforming
we model the channel phase at t1 as follows design can be formulated as follows
θ i (t1 ) = θ i (t0 ) + ei , (3)
Q: min r
{Wk }K
where ei ∼ N (0, σi2 I)
is the channel phase error and is σi2 k =1
the variance of the phase error vector [8], [14]. We denote the s.t. pi Pr{SINRi ≥ γi } ≥ αi ,
estimated channel at t0 and the actual channel at t1 by ĥi and K
hi , respectively. Then hi can be modeled as [6] 1
Wk ≤ r , ∀n ∈ B, ,
Pn
hi = ĥi qi = diag (ĥi )qi , (4) k =1 n,n
Wk 0, rank(Wk ) = 1, ∀k ∈ K. (9)
where qi exp{jei }. Let Qi qi qH i , then the long term
correlation matrix of qi can be expressed as Note that the feasible set of the problem Q will decrease
as min{γi } increases. Then, the optimum objective value of
Ai = E{qi qH
i } = E{Qi }. (5) i
Q is monotonically non-decreasing in min{γi }. Thus, using
It is not difficult to obtain the elements of Ai as follows [9] i
a classic bisection search approach [15], the problem F can
be solved via iteratively solving the problem Q with different
[Ai ]m,n =
1, m = n, (6)
exp −σi2 , otherwise. SINR thresholds {γi }i . In the following, we will focus on the
outage constrained power minimization robust beamforming
III. ROBUST M ULTIGROUP M ULTICAST B EAMFORMING design problem Q.
We first investigate the expression of the non-outage prob-
In this section, we investigate outage constrained robust ability pi = Pr{SINRi ≥ γi }. Define some auxiliary
multigroup multicast beamforming for frame-based satellite variables Zk Wk − γi =k W , and Ri hi hH i =
communication systems. From the signal model in (1), the H
diag(ĥi )Qi diag(ĥi ), then the non-outage probability pi
signal-to-interference-plus-noise ratio (SINR) at the ith user in (9) can be expressed as
in cluster Uk can be represented as [6] ⎧ ⎫
⎨ ⎬
hH
i Wk hi
SINRi , ∀i ∈ Uk , k , ∈ K, (7) pi = Pr Tr(Ri Wk ) ≥ γi Tr(Ri W ) + γi N0
H
=k i W hi + N0
h ⎩ ⎭
=k
where Wk wk wkH . = Pr{Tr(Ri Zk ) ≥ γi N0 }. (10)
Note that for the imperfect CSI case, it is difficult to design
Define xi Tr(Ri Zk ) and xi can be rewritten as
a beamformer to guarantee a target SINR all the time due to
the channel uncertainty. Thus, we are interested in the robust
xi = Tr diag(ĥi )Qi diag(ĥH
i )Zk = Tr(Ck Qi ), (11)
beamforming design and we aim to maximize the worst-case
outage SINR of all users in a high probability, which can be
formulated as the following problem where Ck = diag(ĥi )Zk diag(ĥH i ). We can observe from (11)
that xi is real-valued and is the sum of statistically independent
F : max min γi random variables. Thus, xi can be approximated by a real-
{Wk }K i
k =1 valued Gaussian distribution when Nt is sufficiently large from
s.t. pi Pr{SINRi ≥ γi } ≥ αi , the central limit theorem. The mean and variance of the real
K valued random variable xi can be obtained as follows
Wk ≤ Pn , ∀n ∈ B, μi = E{xi } = Tr(Ck E{Qi }) = Tr(Ck Ai ), (12)
k =1 n,n
vi2 = E{(Tr(Ck Qi ))2 } − μ2i
Wk 0, rank(Wk ) = 1, ∀k ∈ K, (8)
= vecH (CH H H 2
k )E{vec(Qi )vec (Qi )}vec(Ck ) − μi
where αi is the non-outage probability threshold for user i H
and Pn is the power budget of the nth beam. We focus on the = vecH (CH ∗ ∗ H 2
k )E{(qi ⊗ qi )(qi ⊗ qi ) }vec(Ck ) − μi
case where αi > 0.5, which is of more practical interest [6]. = vecH (CH T H 2
k )E{Qi ⊗ Qi }vec(Ck ) − μi , (13)
354 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
where vec(·) denotes the vectorization operation and ⊗ S IMULATION S ETUP PARAMETERS
denotes the Kronecker product.
Define Gi E{QT i ⊗ Qi }, it is not difficult to show that
the (m, n)th element of Gi with m = (m1 − 1)K + m2 and
n = (n1 − 1)K + n2 is given by
⎧
⎨ 1, m1 = n1 and m2 = n2
2 , m
= n and m
= n
[Gi ]m,n = exp−2σi 1 1 2 2
⎩
exp −σi2 , otherwise.
With the above derivations, we can then obtain the expres-
sion of the non-outage probability pi = Pr{xi ≥ γi N0 } as
follows
1 1 μi − γ i N0
pi = + erf √ , for γi N0 ≤ μi , (14)
2 2 2vi
where erf(·) is the Gaussian error function. Note that in the K
evaluation of the non-outage probability, we only focus on the 1
[ Wk ]n,n ≤ r , ∀n ∈ B,
case where pi > 0.5, which is of interest for practical satellite Pn
k =1
communication links. Wk 0, ∀k ∈ K. (17)
From (12), (13) and (14), the outage constraint pi ≥ αi can
be rewritten as Problem Qf is a convex optimization problem with
1 2 a linear objective and second-order cone and semi-
G 2 vec(CH ) definite programming constraints and thus can be efficiently
i k
solved [15]. We observe from extensive simulation that the
2
⎛ ⎞2 problem Qf formulated in (17) yields rank-one solutions
1 a ai2 in most of the cases. For other cases, classic randomiza-
≤ 2⎝ bi2 + 1Tr(Ck Ai ) − i ⎠ +
2
,
bi bi2 + 1 bi + 1 tion approaches [16] can be adopted to address the rank
issue.
(15)
√ IV. S IMULATION R ESULTS
where ai = γi N0 and bi = 2erf −1 (2αi − 1). Note that (15)
can be transformed into a convex constraint when we drop the We present simulation results to illustrate the performance
term ai2 /(bi2 + 1). Then problem Q can be reformulated as of the proposed outage constrained robust multigroup multicast
beamforming for frame-based multi-beam satellite commu-
Qs : min r nication systems. Conventional approach which adopts the
{Wk }K
k =1 outdated CSI directly as the true CSI is also considered for
1
performance comparison [6]. The simulation results are based
H
s.t. Gi vec(Ck )
2
on 106 channel realizations.
⎛ 2 ⎞ The simulation setup is presented as follows. We adopt the
1 a channel model presented in Section II, and detailed param-
≤ ⎝ bi2 + 1Tr(Ck Ai ) − i ⎠, eter values are listed in Table I. Assume that the channel
bi bi2 + 1 phase error variances for different users are identical and
K given by σi2 = σ 2 . The non-outage probability and SINR
1 thresholds for all users are set to be equal, i.e., αi = α,
[ Wk ]n,n ≤ r , ∀n ∈ B,
Pn γi = γth . Assume that there are 7 beams covering the
k =1
target area in which a total of 35 users are uniformly
Wk 0, rank(Wk ) = 1, ∀k ∈ K. (16)
distributed.
As the term ai2 /(bi2 + 1) > 0, then the feasible set of the We first evaluate the achievable sum multicast rate given by
transformed problem Qs is a subset of the beamformer set K
that satisfies (15). Rsum = Nk · r k , (18)
Using a semidefinite relaxation technique [16], we relax the k =1
non-convex rank-one constraint, and then the problem Qs can
be rewritten as follows where Nk is the number of users in cluster k, and rk is the
multicast rate in cluster k given by
Qf : min r
{Wk }K
k =1 rk = E log2 (1 + min SINRi ) . (19)
1 i∈Uk
s.t. G
i
2
vec(CH
k ) We can observe from Fig. 1 that the proposed outage con-
⎛ 2 ⎞ strained robust approach shows multicast rate performance
1⎝ a gains over the conventional approach. In addition, the
≤ bi2 + 1Tr(Ck Ai ) − i ⎠, performance gain becomes larger as the non-outage prob-
bi 2
bi + 1 ability threshold α increases or the SINR threshold γth
YOU et al.: OUTAGE CONSTRAINED ROBUST MULTIGROUP MULTICAST BEAMFORMING FOR MULTI-BEAM SATELLITE COMMUNICATION SYSTEMS 355
R EFERENCES
[1] M. Á. Vázquez et al., “Precoding in multibeam satellite communica-
tions: Present and future challenges,” IEEE Wireless Commun., vol. 23,
no. 6, pp. 88–95, Dec. 2016.
[2] G. Zheng, S. Chatzinotas, and B. Ottersten, “Generic optimization
of linear precoding in multibeam satellite systems,” IEEE
Fig. 1. Comparison of the achievable sum multicast rate between the Trans. Wireless Commun., vol. 11, no. 6, pp. 2308–2320,
proposed robust and conventional beamforming approaches. Results are shown Jun. 2012.
versus the SINR threshold γth for different values of non-outage probability [3] V. Joroughi, M. Á. Vázquez, and A. Pérez-Neira, “Generalized
threshold α with σ 2 = 30◦ . multicast multibeam precoding for satellite communications,” IEEE
Trans. Wireless Commun., vol. 16, no. 2, pp. 952–966, Feb. 2017.
[4] Digital Video Broadcasting (DVB); Second Generation Framing
Structure, Channel Coding and Modulation Systems for Broadcasting,
Interactive Services, News Gathering and Other Broad-Band Satellite
Applications (DVB-S2), V1.4.1, ETSI Standard EN 302 307-1,
Nov. 2014.
[5] Digital Video Broadcasting (DVB); Second Generation Framing
Structure, Channel Coding and Modulation Systems for Broadcasting,
Interactive Services, News Gathering and Other Broad-Band Satellite
Applications. Part 2: DVB-S2 Extensions (DVB-S2X), V1.1.1, ETSI
Standard EN 302 307-2, Oct. 2014.
[6] A. Gharanjik, M. R. B. Shankar, P.-D. Arapoglou, M. Bengtsson, and
B. Ottersten, “Robust precoding design for multibeam downlink satellite
channel with phase uncertainty,” in Proc. IEEE ICASSP, Brisbane, QLD,
Australia, Apr. 2015, pp. 3083–3087.
[7] D. Christopoulos, S. Chatzinotas, and B. Ottersten, “Weighted fair
multicast multigroup beamforming under per-antenna power con-
straints,” IEEE Trans. Signal Process., vol. 62, no. 19, pp. 5132–5142,
Oct. 2014.
Fig. 2. Comparison of the outage probability between the proposed robust and [8] A. Gharanjik, M. R. B. Shankar, P.-D. Arapoglou, M. Bengtsson, and
conventional beamforming approaches. Results are shown versus the SINR B. Ottersten, “Precoding design and user selection for multibeam satel-
threshold γth for different values of non-outage probability threshold α with lite channels,” in Proc. IEEE SPAWC, Stockholm, Sweden, Jun. 2015,
σ 2 = 30◦ . pp. 420–424.
[9] W. Wang et al., “Robust multigroup multicast transmission for
frame-based multi-beam satellite systems,” IEEE Access, vol. 6,
pp. 46074–46083, 2018.
increases. For the case with α = 0.8 and γth = [10] M.-C. Yue, S. X. Wu, and A. M.-C. So, “A robust design
−4 dB, the proposed robust approach can provide approx- for MISO physical-layer multicasting over line-of-sight chan-
imately 46% multicast rate gain over the conventional nels,” IEEE Signal Process. Lett., vol. 23, no. 7, pp. 939–943,
approach. Jul. 2016.
In Fig. 2, the outage performance of the proposed robust [11] D. Christopoulos, S. Chatzinotas, and B. Ottersten, “Multicast multi-
group precoding and user scheduling for frame-based satellite communi-
and conventional approaches are depicted. We can observe cations,” IEEE Trans. Wireless Commun., vol. 14, no. 9, pp. 4695–4707,
that the proposed robust approach can provide significant Sep. 2015.
reduction in outage probability than the conventional beam- [12] C. Caini, G. E. Corazza, G. Falciasecca, M. Ruggieri, and F. Vatalaro,
forming approach.1 In particular, for the case with α = 0.8 “A spectrum- and power-efficient EHF mobile satellite system to be
integrated with terrestrial cellular systems,” IEEE J. Sel. Areas Commun.,
and γth = −3 dB, the proposed robust approach can pro- vol. 10, no. 8, pp. 1315–1325, Oct. 1992.
vide approximately 65% outage performance gain over the [13] T. Maseng and P. Bakken, “A stochastic dynamic model of rain atten-
conventional approach. uation,” IEEE Trans. Commun., vol. COM-29, no. 5, pp. 660–669,
May 1981.
[14] G. Taricco, “Linear precoding methods for multi-beam broadband satel-
V. C ONCLUSION lite systems,” in Proc. 20th Eur. Wireless Conf., Barcelona, Spain,
May 2014, pp. 1–6.
In this letter, we have investigated outage constrained [15] S. Boyd and L. Vandenberghe, Convex Optimization. New York, NY,
robust multigroup multicast beamforming for multi-beam USA: Cambridge Univ. Press, 2004.
satellite communication systems. With the channel phase [16] Z.-Q. Luo, W.-K. Ma, A. M.-C. So, Y. Ye, and S. Zhang, “Semidefinite
uncertainty taken into account, we formulated a robust beam- relaxation of quadratic optimization problems,” IEEE Signal Process.
former design with the aim to maximize the worst-case Mag., vol. 27, no. 3, pp. 20–34, May 2010.
[17] M. K. Arti and M. R. Bhatnagar, “Two-way mobile satellite relaying:
outage SINR of all users subject to the outage and the A beamforming and combining based approach,” IEEE Commun. Lett.,
vol. 18, no. 7, pp. 1187–1190, Jul. 2017.
1 Note that for larger values of γ , infeasibility might arise in the consid- [18] M. K. Arti, “Imperfect CSI based multi-way satellite
th
ered problem due to, e.g., the channel uncertainty. For these scenarios, joint relaying,” IEEE Wireless Commun. Lett., to be published,
design of beamforming and admission control should be considered. doi: 10.1109/LWC.2018.2833471.
356 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Differential spatial modulation (DSM) is a multi- of the previous matrix only, not a whole matrix. By doing
antenna technique which transmits additional data bits by so, the complexity is roughly proportional to the number of
selecting indexes of antennas and avoids pilot overhead. However, transmitter antennas only.
DSM is complicated in two aspects and we aim to reduce the
On the other hand, DSM with complex-valued antenna-
complexity of DSM in this letter. First, the complexity of the
noncoherent maximum-likelihood (ML) detector increases expo- index matrices in [5] has better error performance than DSM
nentially with the number of transmitter antennas. For this in [4]. But the matrices of full-rate DSM are obtained by
problem, we propose a new symbol-based ML detector whose random searches, and there are unlimited possibilities of trans-
complexity is roughly proportional to the number of transmit- mitted signals after differential encoding. This also increases
ter antennas. The other problem is that the signal constellation the hardware complexity [10]. Complex space-time matrices
of the transmitted signal has unlimited points due to complex- for reduced-rate (with increased diversity) DSM were designed
valued antenna-index matrices. We propose a systematic design
in [9] and [10]. In this letter, we propose a systematic design of
of complex-valued antenna-index matrices for which the con-
stellation of the transmitted signal has a few signal points only. complex space-time matrices for full-rate DSM using MPSK
Both the proposed techniques decrease the complexity without signal. Be properly selecting the complex value, the constel-
sacrificing error performance. lation of the transmitted signal only contains a few signal
Index Terms—Spatial modulation, differential encoding, dif-
points.
ferential detection. Notation: (.)T , (.)† and . denote the transpose, the
conjugate transpose and the Frobenius norm of a matrix,
respectively. diag{.} represents the operation from a row
I. I NTRODUCTION vector to a diagonal matrix. denotes the floor func-
ECAUSE of growing demand for wireless commu- tion. CN (0, σ 2 ) denotes the zero-mean, σ 2 -variance, complex
B nications, various techniques of wireless communica-
tions have been proposed. Among them, spatial modulation
Gaussian distribution.
symbols x(t) = [x1 (t), x2 (t), . . . , xNT (t)]. An NT × NT data s2 (t − 1)x2 (t)e j θ4 . For this A(t), Y(t − 1)X(t) = [y3 (t −
matrix X(t) is calculated by 1)x3 (t)e j θ1 , y1 (t − 1)x1 (t)e j θ2 , y4 (t − 1)x4 (t)e j θ3 , y2 (t −
1)x2 (t)e j θ4 ], so [y3 (t − 1), y1 (t)], [y1 (t − 1), y2 (t)], [y4 (t −
X(t) = diag{x(t)}A(t). (2)
1), y3 (t)] and [y2 (t −1), y4 (t)] are utilized for detecting x3 (t),
Note that ∀q ∈ {1, 2, . . . , Q}, Aq should satisfy the two x1 (t), x4 (t) and x2 (t), respectively.
(k )
restrictions, and the definitions of nonzero entries of Aq Let pq represent the position of the nonzero
in [4] and [5] are different. In [4], nonzero elements are entry, e j θq,k , in the kth column of Aq where
always 1, while in [5], to improve the error performance, k ∈ {1, 2, . . . , Nt } and q ∈ {1, 2, . . . , Q}. With
nonzero elements are complex numbers with unit absolute (k )
pq and X̃ = diag{[x̃1 , x̃2 , . . . , x̃NT ]}Aq , the met-
values. NR
ric in (4) becomes Y(t) − Y(t − 1)X̃2 = i=1
Because X(t) is a unitary matrix, differential encoding and NT j θq,k |2 . According to the
differential detection of DSTM can be applied to DSM and |y
k =1 ik (t) − y (k ) (t − 1)x̃ (k ) e
ipq pq
thus DSM is a special case of DSTM. At the transmitter, S(t) discussion above, the detected value of x (k ) (t) for Aq is
pq
is determined by
NR
(q)
S(t) = S(t − 1)X(t). (3) x̂ (k ) (t) = arg min |yik (t) − y (k ) (t − 1)x̃ e j θq,k |2 (5)
pq x̃ ipq
i=1
The initial reference matrix X(0) is the identity matrix, so
and the determined data symbols of Aq is represented by
the transmitted matrix S(t) is unitary and satisfies the two (q) (q) (q)
restrictions. At the receiver, the noncoherent ML detection is x̂q (t) = [x̂1 (t), x̂2 (t), . . . , x̂NT (t)]. The metric of Aq is
NT
NR
X̂(t) = arg min Y(t) − Y(t − 1)X̃2 (4) (q)
X̃∈X mq (t) = |yik (t) − y (k ) (t − 1)x̂ (k ) (t)e j θq,k |2 (6)
ipq pq
k =1 i=1
where X denotes the set of all possible values of X(t). To
and the detected value of A(t) is
obtain X̂(t), the receiver has to try all Q × M NT elements
in X . The signal constellation for x(t) is MPSK, but the signal Â(t) = arg min mq (t). (7)
Aq ∈A
constellation for S(t) obtained from x(t) and A(t) by (2) and (3)
may contain extremely many signal points if nonzero elements Based on the detected value of q corresponding to Â(t),
of A(t) are complex-valued. denoted by q̂, the detection of x(t) is x̂(t) = x̂q̂ (t).
The minimization of (5) and (7) needs to try Q ×(MNT +1)
III. P ROPOSED ML D ETECTION times, which is much less than the number for (4). For exam-
ple, in the case of 8PSK and NT = 4 which has Q = 16, the
In (4), data symbols x1 (t), x2 (t), . . . , xNT (t) are jointly number of test for (4) is 65536 but the number for (5) and (7) is
detected. In this section, we propose to separatively detect 528 only. The complex reduction of this example exceeds 99%.
data symbols instead. Let p (k ) ∈ {1, 2, . . . , NT } represent the For comparison, an example of the sphere decoder for NT = 4
position of the only nonzero entry in the kth column of A(t) in [8] has complex reduction less than 75%. In the case of
where k ∈ {1, 2, . . . , NT }, and define the permutation order 8PSK and NT = 8, the number of test for (4) is 224 Q while
p = (p (1) , p (2) , . . . , p (NT ) ). According to the differential the number for (5) and (7) is only 65Q. To further reduce the
encoding (3), in the kth time slot of the tth block, the acti- complexity, the complexity can be independent of constellation
vated antenna is the antenna used in the p (k ) th time slot of the size M with the aid of hard-limiting [14].
t − 1th block, and the transmitted data symbol is xp (k ) (t). In
other words, the transmitted symbol in the kth time slot of the IV. P ROPOSED D ESIGN OF C OMPLEX
tth block, denoted by sk (t), is sk (t) = sp (k ) (t −1)xp (k ) (t)e j θk S PACE -T IME M ATRICES
where θk is the phase of the nonzero element in the kth col-
Let X = diag{x}A and X = diag{x }A represent two
umn of A(t). Therefore, the receiver can use yip (k ) (t − 1) and
data matrices in (2) where x = [e j φ1 , e j φ2 , . . . , e j φNT ] and
yik (t) only to detect xp (k ) (t) where i ∈ {1, 2, . . . , NR }. j φ
1: Assume N x = [e j φ1 , e j φ2 , . . . , e NT ]. There are two possible cases for
⎛ Example ⎞T = 4, NR = 1, and A(t) = X = X : (i) A = A and x = x (ii) A = A . The transmitter
0 e j θ2 0 0
⎜ 0 diversity in case (i) is one, independent of the design of A. In
⎜ 0 0 e j θ4 ⎟ ⎟ whose p = (3 1 4 2), so X(t) =
⎝ e j θ1 0 this section, we only consider case (ii).
0 0 ⎠
There are at least two different elements of the permuta-
0 0 e j θ3 0 tion order p between A and A in case (ii). If all nonzero
⎛ ⎞
0 x1 (t)e j θ2 0 0 entries of A and A are 1, then the transmitter diversity
⎜ 0 0 0 x2 (t)e j θ4 ⎟
⎜ ⎟. If S(t − 1) order in case (ii) is only one. Therefore, we aim to design
⎝ x3 (t)e j θ1 ⎠
0 0 0 two matrices A and A with only two different elements in
0 0 x4 (t)e j θ3 0 p in the following. Without
⎛
s2 (t − 1)
⎞ ⎛ j θ1 loss of the generality, they
⎞ are
0 0 0 e 0 0 ··· 0
⎜ 0 0 0 s4 (t − 1) ⎟ ⎜ 0 e j θ2 0 · · · 0 ⎟
is ⎝ ⎠, then in ⎜ ⎟
0 0 s3 (t − 1) 0 ⎜ 0 0 e j θ3 · · · 0 ⎟
s1 (t − 1) represented by A = ⎜ ⎟ and
0 0 0 ⎜ . .. .. .. .. ⎟
S(t), we have s1 (t) = s3 (t − 1)x3 (t)e j θ1 , s2 (t) = s1 (t − ⎝ .. . . . . ⎠
1)x1 (t)e j θ2 , s3 (t) = s4 (t − 1)x4 (t)e j θ3 and s4 (t) = j θNT
0 0 0 ··· e
358 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
e j (1.041π+φ) 0
This example in [11] has X = j (1.571π+φ)
0 e
j (1.503π+φ )
Fig. 1. The values of 1 − cos(2θ), 1 − cos(2θ − 2π 4π 0 e
M ), 1 − cos(2θ − M ), and X = j (1.609π+φ ) , or
1 − cos(θ) and 1 − cos(θ − 2π
M ) with M = 8 for 0 ≤ θ ≤ 2π .
M j (0.633π+φ) e 0
e 0
X = and X =
⎛ ⎞ 0 e j (1.182π+φ)
0
e j θ2 0 ··· 0
⎜ e j θ1 ⎟ 0 e j (1.304π+φ )
⎜ 0 0 ··· 0 ⎟ j (0.011π+φ ) with M = 4.1 Note
⎜ ⎟ e 0
A = ⎜ 0 0 e j θ3 ··· 0 ⎟. According to the that (9) is invalid for reduced-rate DSM. For this example, (8)
⎜ . .. .. .. ⎟
⎜ . .. ⎟ becomes
⎝ . . . . . ⎠
j θ
0 0 0 · · · e NT |(X − X )(X − X )† | = 2[1 − cos(2φ − 2φ + θ1 + θ2 − θ1 − θ2 )].
design guideline of DSTM in [13], the rank of X − X which (10)
represents the transmitter diversity should be maximized first
and then the determinant of (X − X )(X − X )† should be Due to 2φ − 2φ = 4k π
M where k is an arbitrary integer, (10)
maximized [15]. It can be shown that this determinant is has the maximum value when
2π π
NT
θ1 + θ2 − θ1 − θ2 = ± =± . (11)
|(X − X )(X − X )† | = 2NT −1 [1 − cos(ψ1 + ψ2 )] (1 − cos ψi ) M 2
i=3
We find that in A of the example, θ1 + θ2 − θ1 − θ2
(8)
is (1.041 + 1.571 − 1.503 − 1.609)π = −0.5π or
where ψi = φi + θi − φi − θi ∀i ∈ {1, 2, . . . , NT }. (0.633 + 1.182 − 1.304 − 0.011)π = 0.5π which coincide
is θi = 0 and θ
Because in general A and A may differ in any two with (11). In the proposed design
which i =
columns, all nonzero entries of A (A ) should be the same. 1 0 0 e j π/4
θ∀i = 1, 2, A is simply be , .
Therefore, in the proposed design, we choose θi = θ and 0 1 e j π/4 0
θi = 0 ∀i ∈ {1, 2, . . . , NT }. Because φi and φi are phases Unlike [11] whose constellation for S(t) has extremely many
of MPSK symbols, the value of θ which maximizes the points, this A has the same coding gain but the signal con-
minimum value of the determinant, denoted by θopt , should stellation for S(t) is 8PSK only. Note that the unbounded
maximize [1 − cos(2θ + 2k π 2lπ
M )][(1 − cos(θ + M ))]
NT −2 where differential constellation size issue makes the employment of
k and l are arbitrary integers. The values of θ which minimize a high-resolution digital-to-analog converter (DAC) impera-
cos(2θ + 2k π 2lπ π π
M ) and cos(θ + M ) are 2M and M , respectively.
tive, which is both expansive and power hungry [10]. For the
Curves of 1 − cos(2θ + M ) and 1 − cos(θ + 2lπ
2k π
M ) for k,
proposed design, low-cost DACs are sufficient to implement
l = 0,−1,−2 are shown in Fig. 1, which indicates that θopt is the transmitter.
π ≤ θ ≤ π ) maximizing
the value of θ ( 2M Next, we propose a systematic construction for A of full-
M
rate DSM. The initial element in A is the identity matrix
f (θ) = [1 − cos(2θ −
2π
)](1 − cos θ)NT −2 . (9) denoted by A1 . Define Ω0 = {A1 }. First, N2T matrices in
M Ω1 = {A2 , A3 , . . . , A(NT )+1 } are obtained by interchanging
π and 2
For NT = 2, we have f (θ) = (1 − cos 2θ) so θopt = 2M any two columns of A1 and replacing all entries 1 with e j θ .
π ).
the minimum value of |(X − X )(X − X )† | is 2(1 − cos M Then matrices in Ω2 are obtained by interchanging any two
For NT > 2, the values of θopt can be found by math- columns of A2 , A3 , . . . , A(NT )+1 and replacing all entries e j θ
ematical software. With M = 4, the determinant of (8) is 2
with 1. If an element in Ω2 has the same permutation order p
0.0892 at θopt = 22.925◦ (In simulations we replace it by as any element in Ω0 , Ω1 , delete it. If there are two identical
22.5◦ for simplicity.) for NT = 3, and 0.0192 at θopt = 30◦ elements in Ω2 , remove one of them. ∀i ≥ 1, matrices in Ωi
for NT = 4. are obtained by interchanging any two columns of matrices
For the design of [4] whose nonzero entries of antenna- in Ωi−1 , and replacing 1 with e j θ for odd i or replacing e j θ
index matrices are all 1 (θ = 0), the diversity between X with 1 for even i. Delete elements in Ωi whose permutation
and X is 1. Consequently, increasing NT cannot increase
diversity order. For the proposed design of A and A , the 1 The DSM is different from conventional DSM. To avoid confusion or a
diversity between X and X is NT because X − X has full lot of explanation, we modify its form of presentation herein.
WEI AND LIN: LOW-COMPLEXITY DSM 359
Fig. 4. The curves of the maximum flow versus the given time period T Fig. 5. The curves of the maximum flow versus the different satellite caching
for 6 satellites. with given time period T = 6 hours.
be changed into that of the multiple-source multiple- We present the impact of the nodes’ caching on the
destination under the modified TEG, which is the impor- network’s max-flow in Fig. 5, where the parameters for the
tant characteristic of the TEG. We can add super-source second scenario is utilized. It can be seen from Fig. 5 that
and super-destination nodes into the modified TEG to the network flow increases within the small range of the
solve the max-flow problem with the Ford-Fulkerson caching size. However, in the big region, we can not obtain
algorithm [9]. the performance gain in terms of the network flow through
augmenting the caching. This phenomenon can be of great
IV. S IMULATIONS significance for the guidance of network design.
We conduct our simulations over a predictable low orbital
satellite network, which consists of 6 orbital planes with V. C ONCLUSION
one satellite fixed on each plane. All the satellites are arbi- In this letter, we propose a TEG-based transceiver resource
trarily selected from the Iridium constellation, and are fixed allocation scheme for satellite networks to maximize the
at a height of 780 km and inclination of 86.4◦ . Moreover, network flow and resource utilization. We construct one effec-
we utilize the Satellite Tool Kit(STK) simulator to gener- tive representation rule of the network’s transceiver resources,
ate the contact topology. The small time interval τ is set which is based on adding both the virtual nodes and edges.
as 1 minute. With this rule, we obtain the modified TEG, and imple-
To explore the performance of our scheme, we conduct ment joint multi-node resource allocation through solving the
two simulation cases. In the first scenario, the time T is max-flow problem over the modified TEG.
set as 12 hours. And each satellite possesses infinite stor-
age size, with transceiver resources set as 30 Mbps. In
the second case, T = 6 hours, and the storage size of R EFERENCES
any satellite varies from 0 MB to 4500 MB. Moreover, [1] C. Caini, H. Cruickshank, S. Farrell, and M. Marchese, “Delay-
in this case, three transceiver resources equipment, i.e., and disruption-tolerant networking (DTN): An alternative solution for
future satellite networking applications,” Proc. IEEE, vol. 99, no. 11,
20 Mbps, 30 Mbps, and 40 Mbps, will be examined. pp. 1980–1997, Nov. 2011.
Moreover, both the source and the destination will be randomly [2] G. Araniti et al., “Contact graph routing in DTN space networks:
selected. Overview, enhancements and performance,” IEEE Commun. Mag.,
In Fig. 4, we compare our proposed TEG-based allocation vol. 53, no. 3, pp. 38–46, Mar. 2015.
[3] S. E. Alaoui and B. Ramamurthy, “Routing optimization for DTN-based
scheme with the random and the average allocation schemes. space networks using a temporal graph model,” in Proc. IEEE ICC,
For the random scheme, the nodes randomly allocate the May 2016, pp. 1–6.
transceiver resources to their outgoing links with (3) estab- [4] Z. Zhang, C. Jiang, S. Guo, Y. Qian, and Y. Ren, “Temporal centrality-
lished and the incoming transmission links when (4) holds, balanced traffic management for space satellite networks,” IEEE Trans.
Veh. Technol., vol. 67, no. 5, pp. 4427–4439, May 2018.
while the average scheme allocates the transceiver resources [5] X. Zhu, C. Jiang, L. Kuang, N. Ge, and J. Lu, “Non-orthogonal multiple
to the corresponding links evenly. As shown in Fig. 4, the access based integrated terrestrial-satellite networks,” IEEE J. Sel. Areas
amounts of the network maximum flow of all methods gradu- Commun., vol. 35, no. 10, pp. 2253–2267, Oct. 2017.
ally increase, which is attributed to the fact that more contacts [6] L. R. Ford and D. R. Fulkerson, Flows in Networks. Princeton, NJ, USA:
Princeton Univ. Press, 1962.
can be generated with the increasing of time. Then it can be [7] Y. Li, C. Song, D. Jin, and S. Chen, “A dynamic graph optimization
checked that our proposed scheme has a larger network-flow framework for multihop device-to-device communication underlaying
than the other two schemes, which is expected and can be cellular networks,” IEEE Wireless Commun., vol. 21, no. 5, pp. 52–61,
explained as follows: our proposed scheme jointly considers Oct. 2014.
[8] A. V. Goldberg and R. E. Tarjan, “Efficient maximum flow algorithms,”
the transceiver constraints at multi-nodes, while both the ran- Commun. ACM, vol. 57, no. 8, pp. 82–89, Aug. 2014.
dom and the average scheme independently deal with these [9] L. R. Ford and D. R. Fulkerson, “Constructing maximal dynamic flows
constraints from one node to another. from static flows,” Oper. Res., vol. 6, no. 3, pp. 419–433, 1958.
364 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In band full duplex (IBFD) is an emerging −101.5 dBm [4]. For an assumed transmit-receive RF path iso-
transceiver technology that facilitates simultaneous transmission lation of 15 dB [2], SI power at the receiver will be 132.5 dB
and reception on the same frequency to double spectral efficiency. above BS’s REFSENS. Thus, for a FD receiver to achieve the
Self-interference (SI) from its own transmitter is the biggest
same link level Signal to Noise Ratio (SNR) as that of a HD
challenge in IBFD radios causing significant degradation to its
receiver performance. SI cancellation (SIC) techniques proposed device, SI power should be suppressed by at least 133 dB to
in literature involve considerable hardware and software com- make IBFD systems realistic.
plexities for practical realization of IBFD radios. In this letter, a Recent breakthroughs in SI cancellation (SIC) techniques
hybrid cellular architecture is proposed composing of IBFD base have demonstrated the feasibility of IBFD by suppressing SI
station and legacy half duplex user equipments (UEs), thus avoid- to the tolerable limit. SIC generally happens at three stages:
ing complex SIC requirement at UE. A novel frequency allocation Propagation domain, analog domain and digital domain [5].
scheme is proposed for this hybrid architecture that allows shar-
ing of carrier frequencies among UEs based on distance criteria.
Additional hardware and intensive processing required for SIC
As the number of UEs increases, chances of finding UEs that sat- result in increased cost, size and power consumption. These
isfy the distance criteria also increase steadily. Simulation results complexities make it difficult to realize an IBFD radio on every
reveal that the probability of 100% frequency reuse exceeds 0.9 User Equipment (UE) and may not be possible in near future.
for as few as 8 UEs itself, demonstrating the potential of the However, making a BS IBFD capable is quiet possible since
proposed idea. the restrictions on cost, size and power are not as stringent for
Index Terms—In band full duplex, frequency reuse, spectral BS. SIC up to 110 dBm is achieved so far as mentioned in [6].
efficiency, data rate, 5G network. Considering the total cancellation requirement of 133 dB for
LTE-A scenario, gap of 23 dB could be achieved using higher
TX/RX antenna spacing and other shielding mechanisms pos-
I. I NTRODUCTION sible at BS. Taking above factors into account, a hybrid cellular
MPROVING Spectral Efficiency (SE) is a preferred architecture composing of IBFD capable BS and legacy HD
I approach to meet the ever increasing data rate require-
ment for rapidly emerging applications and there is a special
UEs is proposed [7].
Frequency allocation scheme for the hybrid architecture
focus for the same in upcoming 5G [1]. Traditional cellular must consider the varying capabilities of BS and UEs to
systems being Half Duplex (HD) require two different carrier maximize SE. A hybrid architecture composing of IBFD relay
frequencies for uplink (UL) and downlink (DL) respectively. and HD UE is discussed in [8]. The relay operates either in
An emerging In Band Full Duplex (IBFD) or Full Duplex FD or HD mode based on the instantaneous Channel State
(FD) technique proposes the usage of same carrier frequency Information (CSI). The allocation scheme chooses a suitable
for both UL and DL at the same time [2]. This can effectively mode locally for the relay to improve SE. Another hybrid
double SE and thus gained significant attention recently. It is architecture discussed in [9] divided all HD UEs into two
a promising alternative to legacy HD systems and a potential groups and assumed that the two groups are sufficiently sep-
candidature for upcoming 5G [3]. arated to avoid the in-band interference between them. Each
Self-Interference (SI) is the biggest challenge that an IBFD group is configured to operate either in UL only or DL only
radio must overcome. SI is caused by its own strong UL signal mode and carrier frequencies were allocated accordingly.
interfering to a relatively weaker DL signal present in the same As a first of its kind, a novel frequency allocation scheme
frequency band at the same time. To understand the magnitude is proposed in this letter for hybrid cellular architecture. This
of SI, a practical example from LTE-A is considered, where a allocation algorithm runs at BS and works generically for
Wide Area (Macro) BS can transmit at a maximum power of all random distribution scenarios of HD UEs. The proposed
46 dBm and its receiver sensitivity power limit (REFSENS) is scheme allocates a frequency for UL of UEi and DL to UEj
and another frequency for UL of UEj and DL to UEi (two
Manuscript received August 15, 2018; accepted September 25, 2018. Date frequencies for two HD UEs) to improve frequency reuse for
of publication October 1, 2018; date of current version April 9, 2019. The better SE, while legacy HD system requires four frequencies
associate editor coordinating the review of this paper and approving it for for two HD UEs. To prevent the potential in-band interference,
publication was W. Zhang. (Corresponding author: Parthiban Annamalai.)
P. Annamalai is with Intel, Bengaluru 560103, India, and also a Research UEi and UEj must be sufficiently separated. It should be
Scholar with the International Institute of Information Technology Bangalore, noted that each frequency can be allocated only twice between
Bengaluru 560100, India (e-mail: parthiban.annamalai@intel.com). a pair of UEs and 100% frequency reuse is achieved in a BS
J. Bapat and D. Das are with the International Institute of Information when all UEs share the frequencies. Simulation results reveal
Technology Bangalore, Bengaluru 560100, India (e-mail: jbapat@iiitb.ac.in;
ddas@iiitb.ac.in). that, for as less as 8 UEs itself, probability of 100% frequency
Digital Object Identifier 10.1109/LWC.2018.2873316 reuse (p100% ) exceeds 0.9. As the number of UEs increases,
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ANNAMALAI et al.: NOVEL FREQUENCY ALLOCATION SCHEME FOR IBFD SYSTEMS IN 5G NETWORKS 365
TABLE I
F REQUENCY A LLOCATION P OSSIBILITIES FOR T WIN AND T RIPLET S HARING FOR N = 2 AND N = 3
A PPENDIX
E XAMPLE C OMPUTATION OF dmin
COST HATA path loss model (between two UEs) given
in (A.1) is used to compute the received power at the
LTE-A UE.
where,
PTi , GTi - Transmit power and antenna gain (0 dB) of UEi
PRj , GRj - Receive power and antenna gain (0 dB) of UEj
PLij - Path Loss between UEi and UEj as defined in (A.1)
Minimum value of d in (A.1) for which the UE’s received
power (A.2) should reach its REFSENS, after substituting for
path loss from (A.1), is termed as dmin . For a LTE-A UE oper-
ating at 10 MHz bandwidth in E-UTRA Band 2, its REFSENS
is −97.7 dBm for QPSK and maximum transmit power is
Fig. 4. Twin-Triplet Sharing Contribution. 23 dBm [11]. To pair UEi with UEj , dmin between them
should be such that the UL power of 23 dBm transmitted
by UEi should attenuate below its REFSENS of −97.7 dBm
while received at UEj and the same is estimated as 137 meters.
V. C ONCLUSION AND F UTURE W ORK
The frequency allocation scheme proposed in this letter R EFERENCES
allows BS to allocate a single carrier frequency to two UEs
to potentially double SE. BS identifies such UEs based on [1] “IMT vision—Framework and overall objectives of future develop-
ment of IMT for 2020 and beyond,” Int. Telecommun. Union, Geneva,
their inter-UE distances for frequency sharing. Grouping in Switzerland, ITU-Recommendation M.2083-0, Sep. 2015.
smaller units (either two or three) of UEs is found to be [2] A. Sabharwal et al., “In-band full-duplex wireless: Challenges
most efficient and accordingly twin/triplet sharing is proposed. and opportunities,” IEEE J. Sel. Areas Commun., vol. 32, no. 9,
pp. 1637–1652, Sep. 2014.
Achievable frequency reuse depends upon the geographical
[3] S. Hong et al., “Applications of self-interference cancellation in 5G and
distribution of UEs, better reuse is achieved when UEs are beyond,” IEEE Commun. Mag., vol. 52, no. 2, pp. 114–121, Feb. 2014.
randomly distributed within the good coverage area of BS. [4] Base Station (BS) Radio Transmission and Reception, V15.3.0, 3GPP
Proposed solution has the potential to reach 100% frequency Standard TS 36.104, Jul. 2018.
[5] Z. Zhang, K. Long, A. V. Vasilakos, and L. Hanzo, “Full-duplex wireless
reuse as the number of UEs (N) increases. It is important communications: Challenges, solutions, and future research directions,”
to note that these improvements are achieved while retaining Proc. IEEE, vol. 104, no. 7, pp. 1369–1409, Jul. 2016.
legacy HD UEs that are less complex than IBFD UEs and [6] D. Bharadia, E. McMilin, and S. Katti, “Full duplex radios,” in Proc.
requiring IBFD capability only at BS. Additional complexity ACM SIGCOMM, pp. 375–386, Aug. 2013.
[7] D. Korpi et al., “Full-duplex mobile device: Pushing the limits,” IEEE
and cost to realize IBFD capability at BS can be justified Commun. Mag., vol. 54, no. 9, pp. 80–87, Sep. 2016.
by the SE improvement gained by the proposed approach. [8] T. Riihonen, S. Werner, and R. Wichman, “Hybrid full-duplex/half-
Since the costs of acquiring additional spectrum are far higher, duplex relaying with transmit power adaptation,” IEEE Trans. Wireless
proposed approach will be more attractive to service providers Commun., vol. 10, no. 9, pp. 3074–3085, Sep. 2011.
[9] A. C. Cirik, K. Rikkinen, R. Wang, and Y. Hua, “Resource allocation in
despite the additional costs required for IBFD capability in BS. full-duplex OFDMA systems with partial channel state information,” in
In this letter, the combinatorial analysis based on inter- Proc. IEEE China Summit Int. Conf. Signal Inf. Process., pp. 711–715,
UE distances to allow frequency sharing was done using Jul. 2015.
[10] Stage 2 Functional Specification of User Equipment (UE) Positioning
exhaustive search algorithm for up to 8 UEs. This needs to in E-UTRAN, V 14.3.0, 3GPP Standard TS 36.305, Oct. 2017.
be extended systematically to obtain similar metrics for any [11] User Equipment (UE) Radio Transmission and Reception, V14.7.0,
higher values of N. In addition, the mobility of UEs must 3GPP Standard TS 36.101, Apr. 2018.
368 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—We describe a scheduling protocol for multimodal how to optimally schedule the transmissions of the available
networks of relatively limited size, whose nodes encompass var- PHYs.
ious underwater communication technologies. For such a case, Different from multi-band scheduling, which aims mostly
we show that significant improvement in the network operations
is possible when the transmission schedule is set to jointly utilize at interference avoidance, multimodal networks enable diver-
all communication technologies. Our solution is based on per- sity in terms of the PHY itself. For example, low-frequency
technology time-division multiple access frames, whose time slots acoustic communications achieve kbit/s transmission rates over
are determined optimally to maximize the overall channel utiliza- ranges of a few kilometers, high-frequency acoustics tops tens
tion while preserving flow limitations and maintaining fairness of kbit/s over up to a few hundred meters, whereas opti-
in resource allocation. Our numerical simulations and experi-
mental results for multimodal networks with several acoustic cal communications yield Mbit/s links over ranges of a few
technologies show that, while maintaining a fair resource alloca- meters. Different PHYs face different challenges: for exam-
tion, our scheduling solution provides both high throughput and ple, acoustic communications are sensitive to time-varying
low packet delivery delay. multipath, whereas water turbidity and ambient light ham-
Index Terms—Underwater communication networks, under- per optical communications [1]. Hence, efficient multimodal
water acoustic communications, multi-modal systems, transmis- networking requires a specific scheduling solution with differ-
sion scheduling. ent properties than the schemes in single-technology network
domains.
The approaches designed for multi-radio or multi-channel
I. I NTRODUCTION AND R ELATED W ORK wireless radio networks rely on frequent communications or
feedback, and target the management of voice calls rather
NDERWATER communications have gradually become
U the enabler of several types of submerged operations.
Submarines, divers, autonomous underwater vehicles (AUVs),
than data transmission [2]. Scheduling poses different require-
ments in multimodal underwater networks, where the available
PHYs have widely different communication capabilities, and
and floaters are often endowed with underwater communica- the scheduler must account for PHY-dependent adjacency and
tion capabilities. This has progressively led to the formation of interference matrices while avoiding bottlenecks.
underwater networks, where devices share their measured data For underwater multimodal operations, it has been proposed
and act to collaborate with other sensors. The increase of com- to use visual image processing [3] or signaling to identify the
munication systems and underwater operations will soon result fastest available PHY [4]. The MURAO protocol [5] organizes
in multimodal networks, where a node incorporates multiple the nodes in clusters, using optics for intra-cluster commu-
physical layer (PHY) technologies, e.g., acoustic, optical and nications and acoustics for cluster management. The above
radio-frequency (RF). This will enable new applications, such approaches are tailored to specific scenarios, or offer solutions
as data muling and wireless telemetry for hybrid vehicles. for stable networks. However, they may suffer from bottle-
With several heterogeneous nodes, employing a different necks and delays in realistic multimodal networks, where PHY
sub-network for each technology may result in disconnections performance changes over time and space due to mobility and
and poor data transfer performance. Instead, it would be pos- environmental conditions, and does so in different ways for
sible to increase the throughput, decrease the communication different PHYs. A scheduling mechanism that can optimize
delay via simultaneous transmissions, and reduce the occur- the use of the multimodal network’s resources is therefore
rence of bottlenecks by properly leveraging the full set of needed.
PHYs. The main challenge, and the focus of this letter, is To address the above challenges, we propose the optimal
Manuscript received July 25, 2018; revised September 10, 2018; accepted multimodal scheduling (OMS) protocol. OMS manages trans-
September 17, 2018. Date of publication October 1, 2018; date of current missions through any set of PHYs by jointly setting trans-
version April 9, 2019. This work was supported by the NATO Science for mission time slots in a per-technology time-division multiple
Peace and Security Programme under Grant G5293. The associate editor coor-
dinating the review of this paper and approving it for publication was S. De. access (TDMA) fashion, and divides the data load among the
(Corresponding author: Filippo Campagnaro.) PHYs to optimize link utilization and transmission delay. In
F. Campagnaro and M. Zorzi are with the Department of Information addition, OMS organizes transmission slots to favor packet
Engineering, University of Padova, 35131 Padua, Italy (e-mail: campagn1
@dei.unipd.it; zorzi@dei.unipd.it). routing and enforce a fair number of transmission opportu-
P. Casari is with IMDEA Networks, 28198 Madrid, Spain (e-mail: nities per node. We tested the performance of OMS against
paolo.casari@imdea.org). benchmark schemes in numerical simulations and in a sea
R. Diamant is with the Department of Marine Technologies, University of
Haifa, Haifa 3498838, Israel (e-mail: roeed@univ.haifa.ac.il). experiment using multimodal nodes encompassing different
Digital Object Identifier 10.1109/LWC.2018.2873329 acoustic PHYs. The results show that OMS achieves better
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
CAMPAGNARO et al.: OPTIMAL TRANSMISSION SCHEDULING IN SMALL MULTIMODAL UNDERWATER NETWORKS 369
throughput, packet delivery delay, and fairness in resource PHY n. The transmission slot indices are arranged in a vec-
allocation. tor tTx
i,n = {r : Si,r ,n = 1} for node i over technology n.
Our objective is to maximize channel utilization, measured via
the total number of transmissions over a given time period.
II. T HE OMS A LGORITHM The schedule also considers collisions among neighboring
A. System Model nodes and facilitates the forwarding of packets across multiple
Our system consists of N network nodes equipped with one hops. OMS requires the knowledge of the adjacency matrix M
or more of T underwater PHYs. The set of PHYs is arranged obtained, e.g., via [6]. This includes the existing connections
in the N × T technology matrix T such that Ti,n = 1 if node i and the available per-node PHY technologies.
has PHY n. Call M the adjacency matrix, where Mi,j ,n = 1 Let ∨ be the logical “or” and ∧ the logical “and” operators.
if node i is connected to node j via PHY n. The number of The optimal schedule S with time frame length τ fr is the
neighbors of i through technology n is Di,n = j Mi,j ,n . We solution of the following problem:
assume that T is given, and that M can be obtained via prelim-
inary link probing [6]. We remark that the difference between S , τfr = arg min max ci Si,t,n (1a)
τ fr i t n
the communication and interference range is limited in under-
water networks, due to the very fast power decay incurred for s.t. Ti,n = 0 =⇒ Si,t,n = 0 (1b)
increasing range by any PHY technology [4]. Hence, to har- Si,t,n = 1 ∧ Sj ,t,n = 1 =⇒
ness spatial reuse for performance gain, we allow collisions (Mi,p,n + Mj ,p,n = 0)
in OMS.
∨ (Mi,p,n + Mj ,p,n = 1 ∧ Sp,t,n = 0) (1c)
OMS organizes orthogonal multimodal PHYs via per-
technology TDMA frames. As different PHYs are character- ∨ (Mi,p,n + Mj ,p,n = 2 ∧ Sp,t,n = 1) ∀p = i , j
ized by diverse transmission rates and may incur different Si,t,n
≥ ci ≥ C ∀i , n (1d)
propagation delays (e.g., optics vs. acoustics), the duration of n
Di,n
t
the time slots is also set per-technology. We choose TDMA ∃ t s.t. Si,t,n = 1, Sj ,t,n = 0 ∀i , j , n s.t. Mi,j ,n = 1
since it allows a simple time slot alignment via guard intervals.
(1e)
This is specifically important in multimodal systems, where
different PHYs have a diverse outage capacity. Additionally, ∃ n s.t. Mp,j ,n = 1 ∧ Mj ,i,n = 1 ∧ Mp,i,n = 0
in TDMA-based schemes the transmission delay is known ∧ max(tTx Tx
j ,n ) > min(tp,n ). (1f)
in advance, making it possible to plan the load allocated to
each PHY. This is in contrast to handshake-based schemes The solution can be obtained via branch-and-bound, which
(where the delay depends also on the receiver) and to fully completes in polynomial time on average [7].
random access (where collisions may trigger an unpredictable Constraint (1b) prevents transmissions on technology n if
number of retransmissions). The synchronization of the low node i does not have it. Constraint (1c) allows simultaneous
latency technologies can be achieved either via atomic clocks transmissions by two nodes i and j in the same slot using
or through the network time protocol (NTP). For acoustics, technology n only if: ∀p = i, j, the links i ↔ p and j ↔ p
however, we can simply rely on guard times: as the time slot do not exist for technology n; or if one of the two links exists
duration is at least as long as the maximum propagation delay, and p does not transmit in the same slot (lest p would be deaf
such guard times are negligible. to i’s or j’s transmission); or otherwise, if both links exist, p
We impose traffic constraints by allowing a node i to trans- also transmits in the same slot (so that i and j’s transmissions
mit in at least ci > C time slots. Calling Ri the number of would not collide at p). Constraint (1d) specifies that more
bits transmitted in each time slot of the slowest communication slots are given to nodes with more neighbors, so that the total
technology of node i, constraint ci ensures the transmission of number of slots per neighbor is at least ci ≥ C . Constraint (1e)
at least ci Ri bits within a given frame. Fairness then results imposes that node i be the only transmitter in at least one slot
from setting ci such that nodes with lower Ri receive a higher t over each technology n. Hence, although (1c) allows primary
ci value. conflicts, there exists at least one slot for each node to transmit
free from interference. Finally, constraint (1f) facilitates that
the same packet can propagate further than one hop within the
B. OMS Scheduling Solution same frame, and is achieved by allowing a node j located at
Our solution allocates transmission time slots, organized an intermediate position between two nodes p and i to have
in n TDMA frames of N slots and duration τnsl : one frame at least one transmission slot later than node p’s slot, i.e.,
for each PHY. We synchronize transmissions by considering max(tTx Tx
j ,n ) > min(tp,n ).
a TDMA super-frame of length τ fr , such that for some PHYs The formalization in (1) shows that OMS optimizes link uti-
several (not necessarily full) TDMA cycles are possible every lization by allocating the data flows across all PHYs of a node,
τ fr seconds. The input to OMS is the PHY matrix T, the while considering possible bottlenecks and packet delays due
adjacency matrix M, the per-PHY communication capacity, to the different capabilities of the various PHYs. However,
and the number of slots N and time slot duration τnsl . The OMS is a centralized solution and thus fits the case of small
output is the minimum allowed value of τ fr , and a matrix networks. Still, by sharing the adjacency matrix M, OMS
S, where Si,t,n = 1 if node i can transmit in slot t via avoids the use of a centralized hub.
370 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
S IMULATIONS : C HARACTERISTICS OF THE PHY T ECHNOLOGIES
Fig. 4. Topology A: PDR (2), PDR fairness (3), and throughput (4).
Fig. 5. Topology B: PDR (2), PDR fairness (3), and throughput (4).
Abstract—Recently, coded caching scheme was proposed as the placement delivery array (PDA) was proposed creatively to
ability of alleviating the load of networks. Especially, the place- describe placement and delivery phase [5].
ment delivery array (PDA) used for characterizing the coded PDA construction is completely equivalent to coded caching
caching scheme has attracted vast attention. In this letter, a
scheme design, because a PDA can indicate what should be
deep neural architecture is first proposed to learn the construc-
tion of PDAs for reducing the computational complexity. The cached by users and what should be sent by the server in
problem of variable size of PDAs is solved using mechanism a single array. Numerous methods have been proposed to
of neural attention and reinforcement learning. Different from address the construction of PDAs. The method proposed in [5]
previous works using combined optimization algorithms to get significantly decreases F, while only suffering from a slight
PDAs, our proposed deep neural architecture uses sequence-to- sacrifice. Yan et al. [6] found the connection between strong
sequence model to learn construct PDAs. Numerical results are edge coloring of bipartite graphs and PDA construction and
given to demonstrate that the proposed method can effectively
implement coded caching meanwhile reducing the computational they proposed a placement delivery array design algorithm
complexity. using graph theory presented in [7].
However, finding an optimal strong edge coloring for bipar-
Index Terms—Coded caching, placement delivery array, deep
learning, neural attention.
tite graph is a NP-hard problem [8]. Although new PDAs
can be discovered through finding strong edge coloring, for
the general case, only some sporadic results exist [7], [8]. In
I. I NTRODUCTION this letter, we revisit PDA design problem in a simpler per-
UE TO the exponential growth in the number of smart
D mobile equipments and innovative high-rate mobile data
services (such as videos streaming for mobile gaming and
spective, i.e., sequence-to-sequence (Seq2Seq) learning model,
which is widely used in natural language processing [9], [10].
This inspires us to present a deep neural architecture to
road condition monitoring), 5G networks should accommo- devise coded caching schemes. The architecture has three key
date the overwhelming wireless traffic demands. Deploying technologies, i.e., Seq2Seq learning [9], content based input
intelligent caching is an efficient strategy and able to satisfy attention [11] and reinforcement learning.
the rate requirements of users for the ability of information The main contributions of this letter are summarized as
quick acquisition [1]. follows:
The gain from traditional (uncoded) caching approaches (i) A deep neural architecture is first proposed to learn the
derives from making content available locally, and it is con- construction of PDAs, and it allows us to realize coded caching
strained by the limited memory available at each individual using deep learning technology.
user. In the seminal work [2], a coded caching scheme was (ii) Attention model is used to deal with the fundamental
proposed for the centralized caching system by Maddah-Ali problem of representing variable size of PDAs. Reinforcement
and Niesen which is referred to AN scheme in this letter. signals are used to accelerate deep neural network training.
The AN scheme can create multicast opportunities depend- Our results demonstrate that this approach can achieve approx-
ing on the cumulative memory available at all users. It has imate solutions to the problems of construction of PDAs that
been used in many scenarios, for example, finite file size are computationally intractable.
caching system [3] and wireless networks with unequal link The rest of this letter is organized as follows. In Section II,
rates [4]. However, in order to implement the coded caching we introduce the background of PDA to elicit our Seq2Seq
scheme proposed in [2], each file must be split into F file learning problem. In Section III we propose the attention-based
packages. The number of packages generally increases expo- Seq2Seq learning algorithm of the construction of PDAs.
nentially with the number of users. In order to reduce F, Finally, numerical results and a discussion are presented in
Section IV, and a conclusion is reached in Section V.
Manuscript received September 3, 2018; accepted September 23, 2018. Date
of publication October 1, 2018; date of current version April 9, 2019. This
work was supported by the National Natural Science Foundation of China
under Grant 61671144. The associate editor coordinating the review of this II. S YSTEM M ODEL AND BACKGROUND
paper and approving it for publication was C. Shen. (Corresponding authors:
Yongming Huang; Luxi Yang.)
We consider a caching system composing of one server and
The authors are with the National Mobile Communications Research N files W = {W1 , W2 , . . . , WN }. This server is connected
Laboratory, School of Information Science and Engineering, Southeast to K users through an error-free shared link, and the set of all
University, Nanjing 210096, China (e-mail: zmzhang@seu.edu.cn; users is denoted by K = {1, 2, . . . , K } (N > K). We assume
mhua@seu.edu.cn; chunguoli@seu.edu.cn; huangym@seu.edu.cn;
lxyang@seu.edu.cn). that each file has equal size, and each user is equipped with a
Digital Object Identifier 10.1109/LWC.2018.2873334 cache of size M. The caching system is parameterized by K, M
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHANG et al.: PDA DESIGN VIA ATTENTION-BASED SEQ2SEQ MODEL WITH DEEP NEURAL NETWORK 373
TABLE I
S IMULATION PARAMETERS To construct a (K, F, Z, S) PDA (K × F >4) which corre-
sponding to a colored bipartite graph G(K , F , E ) (its degree
Δ ≥ 2), we compare the following schemes:
1) Scheme 1: Use coloring scheme in [6]. The complexity is
O(min{Cm a+b−λ C λ a−λ b−λ
m−(a+b−λ) , Cm Cm−(a−λ) }|E |(a log a+
b log b)), where Cm a = K , C b = F , 0 ≤ λ ≤ min{a, b}, |E |
m
is the size of the edge set.
2) Scheme 2: Use strong coloring scheme in [8]. The col-
oring of G whose degree no greater than 3, can be obtained
in polynomial time O((K + F )2 |E |).
3) Scheme 3: Use strong coloring scheme in [15]. The
complexity is O((K + F )Δ|E |2 ).
4) Scheme 4: Use our method. If the attention-based
Seq2Seq neural network has been trained, the complexity of
our PDA design method is O(|E | log(|E |)).
V. C ONCLUSION
Fig. 2. Training Loss. In this letter, we established the connection between the
placement delivery array in coded caching and the sequence-
to-sequence learning. We first proposed a learning method to
construct PDAs using attention model and reinforcement learn-
ing. Then a new coded caching scheme is constructed based on
the deep neural architecture. Numerical results demonstrated
that the proposed method can effectively implement coded
caching and the complexity is low.
R EFERENCES
[1] S. Wang et al., “A survey on mobile edge networks: Convergence
of computing, caching and communications,” IEEE Access, vol. 5,
pp. 6757–6779, 2017.
Fig. 3. Training Accuracy. [2] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,”
IEEE Trans. Inf. Theory, vol. 60, no. 5, pp. 2856–2867, May 2014.
TABLE II [3] S. Jin, Y. Cui, H. Liu, and G. Caire, “Order-optimal decentralized coded
N UMERICAL C OMPARISONS caching schemes with good performance in finite file size regime,” in
Proc. IEEE Glob. Commun. Conf. (GLOBECOM), Dec. 2016, pp. 1–7.
[4] A. Tang, S. Roy, and X. Wang, “Coded caching for wireless backhaul
networks with unequal link rates,” IEEE Trans. Commun., vol. 66, no. 1,
pp. 1–13, Jan. 2018.
[5] Q. Yan, M. Cheng, X. Tang, and Q. Chen, “On the placement delivery
array design for centralized coded caching scheme,” IEEE Trans. Inf.
Theory, vol. 63, no. 9, pp. 5821–5833, Sep. 2017.
[6] Q. Yan, X. Tang, Q. Chen, and M. Cheng, “Placement delivery array
design through strong edge coloring of bipartite graphs,” IEEE Commun.
listed in Table I. Seq2Seq [9] and Seq2Seq with attention [10] Lett., vol. 22, no. 2, pp. 236–239, Feb. 2018.
are used as benchmarks for comparison. We also compare the [7] J. J. Quinn and A. T. Benjamin, “Strong chromatic index of subset
graphs,” J. Graph Theory, vol. 24, no. 3, pp. 267–273, 1997.
complexity with [6], [8], and [15]. [8] J. Bensmail, A. Lagoutte, and P. Valicov, “Strong edge-coloring of (3, δ)-
Fig. 2 shows the convergence behavior of the loss function. bipartite graphs,” Discr. Math., vol. 339, no. 1, pp. 391–398, Jan. 2016.
Fig. 3 shows the training accuracy of our model. We can find [9] I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to sequence learn-
that no matter how (K, F, M, N) is, as long as L is less than 60 ing with neural networks,” Adv. Neural Inf. Process. Syst., vol. 2,
pp. 3104–3112, 2014.
or 120, our algorithm shows good convergence performance [10] J. Su et al., “A hierarchy-to-sequence attentional neural machine trans-
and the training accuracy can reach more than 90%. lation model,” IEEE/ACM Trans. Audio Speech Lang. Process., vol. 26,
To use Seq2Seq or Seq2Seq with attention methods solve no. 3, pp. 623–632, Mar. 2018.
[11] D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation
the learning problem for PDA design, the length of input and by jointly learning to align and translate,” in Proc. Int. Conf. Learn.
output should be fixed. Notice that this is not necessary for Representat. (ICLR), 2015, pp. 1–15.
our method. The experimental results in Table II show that our [12] O. Vinyals, M. Fortunato, and N. Jaitly, “Pointer networks,” in Proc.
method can achieve accuracy of 88.96% and 82.86% when L is Int. Conf. Neural Inf. Process. Syst., 2015, pp. 2692–2700.
[13] L. Yu, W. Zhang, J. Wang, and Y. Yu, “SeqGAN: Sequence generative
less than 60 and 120. It demonstrates that our proposed method adversarial nets with policy gradient,” in Proc. AAAI Conf. Artif. Intell.,
is obviously better than the other two methods. This is due to San Francisco, CA, USA, Feb. 2017, pp. 2852–2858.
the plausible combination of B-GRU [11], neural attention [12] [14] R. S. Sutton, D. McAllester, S. Singh, and Y. Mansour, “Policy gradient
and reinforcement learning [14]. In fact, pre-trained network methods for reinforcement learning with function approximation,” in
Proc. Adv. Neural Inf. Process. Syst., vol. 12, 1999, pp. 1057–1063.
already has a certain ability to work. Reinforcement learning [15] G. J. Chang and N. Narayanan, “Strong chromatic index of 2-degenerate
has fine-tuned the pre-trained network, which is very helpful. graphs,” J. Graph Theory, vol. 73, no. 2, pp. 119–126, 2013.
376 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Massive multiple-input multiple-output orthogonal angular domain and a unified channel estimation for TDD
frequency division multiplexing (mMIMO-OFDM) channel esti- and FDD system is proposed in [3]. Moreover, to reduce the
mation is considered recently utilizing compressed sensing (CS) required pilot and accurate channel estimation in mMIMO
based methods. Here, we proposed to use the joint sparsity of
mMIMO-OFDM channels using stage-wise forward backward systems, exploiting spatio-temporal sparsity, an adaptive struc-
pursuit (StFBP) algorithm. In order to increase the speed of tured subspace pursuit (ASSP) is introduced in [1]. ASSP
convergence and accuracy of estimation, we proposed to gather algorithm estimates the sparsity of the channel and channel
multiple good atoms in each step and to exploit common sparsity coefficients, simultaneously. Choi et al. [5] utilize the tempo-
in the system model, respectively. Furthermore, the backward ral correlation of the channels to develop the locally common
steps improve the accuracy by omitting bad previously gath-
ered atoms. Simulation results represent the superiority of the support (LCS) algorithm for channel estimation. The Bayesian
proposed StFBP approach rather than the conventional CS-based estimation was used via marginal based SCS method in order
and non-CS-based approaches. to improve the channel estimation accuracy in [6]. Stage-wise
Index Terms—Channel estimation, compressed sensing, mas- OMP (StOMP) was utilized in block sparsity mode to esti-
sive multiple input multiple output-orthogonal frequency division mate the channel state information of MIMO-OFDM rapidly
multiplexing (mMIMO-OFDM), stage-wise forward backward and accurately in [7]. In each step of OMP algorithm, the
pursuit (StFBP). estimation is based on just one of the atoms while in StOMP
the estimation works using multiple atoms. Hence, StOMP
based algorithms converge more rapid than the former. The
I. I NTRODUCTION
OMP and its derivatives the same as StOMP suffers from an
ASSIVE Multiple-Input Multiple-Output (mMIMO)
M is one the promising approaches for the future 5G
telecommunication systems. By employing a large number of
inherent drawback which is caused by not eliminating bad
atoms which are gathered in previous steps. This drawback is
compensated in Forward-Backward Pursuit (FBP) algorithm
antennas in Base Station (BS), it facilitates the implementa- introduced in [9].
tion of high-throughput wireless systems. In order to increase In this letter, we have proposed to utilize the common spar-
the data availability, accurate channel estimation is mandatory. sity together with the fading channel large scale behavior. First
Hence, researchers devote lots of their studies on the channel of all, we have considered all the channel ensembles com-
estimation in MIMO-OFDM systems, recently [1]–[6]. mon sparsity by using the μ-norm. Then, in order to deploy
The sparse behavior of the mMIMO-Orthogonal Frequency the channel large scale behavior, the channel is modeled by a
Division Multiplexing (OFDM) channel ensembles is uti- weighted cost function based on the distance of the channel
lized in Compressed Sensing (CS)-based channel estimation coefficients from the origin. Hence, the channel coefficients
methods which are the result of a small number of sig- will be extracted from the random measurements optimally.
nificant scatterers in the medium. Moreover, the small dis- In order to solve the resultant optimization problem which
tance of the antenna pairs on the mMIMO node rather than address the common sparsity and large scale behavior, we
the considerable conveyed distance by the wave leads to proposed a Stage-wise FBP (StFBP) algorithm mode wherein
unique and common support for all the channel ensembles forward steps multiple good atoms are gathered while in the
between two mMIMO nodes. Hence, using common spar- conventional FBP method only one atom is selected in each
sity channels, the compressed channel estimation is improved forward step. Consequently, the proposed method increases the
significantly [1]–[8]. A channel estimation scheme based speed of convergence. Moreover, exploiting backward steps
on training sequence (TS) design and optimization with makes it possible to omit bad atoms gathered previously, as
high accuracy and spectral efficiency is investigated in the a consequence it would improve the accuracy of estimation
framework of structured compressive sensing in [2]. Channel rather than StOMP and OMP. To exploit the proposed StFBP,
estimation for indoor mMIMO systems is decomposed in we design the channel matrix and measurement matrix to
exploit the common sparsity using μ-norm which is intro-
Manuscript received September 3, 2018; accepted September 21, 2018. Date duced. The system model is formulated using matrix repre-
of publication October 1, 2018; date of current version April 9, 2019. The
associate editor coordinating the review of this paper and approving it for sentation to exploit the common sparsity of the channel. Then,
publication was J. Choi. (Corresponding author: Mehrdad Ardebilipour.) a StFBP algorithm-based channel estimation is developed to
The authors are with the Department of Electrical and Computer estimate the channel coefficients in the time domain.
Engineering, K. N. Toosi University of Technology, Tehran 19697, Iran
(e-mail: mehrdad@eetd.kntu.ac.ir). The remainder of this letter is as follows. In Section II the
Digital Object Identifier 10.1109/LWC.2018.2873339 system model is represented, and the problem is formulated.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
AKBARPOUR-KASGARI AND ARDEBILIPOUR: mMIMO-OFDM CHANNEL ESTIMATION VIA DISTRIBUTED CS 377
Section III the proposed channel estimation method is scatterers where K << L. Consequently, the channel could be
introduced. Numerical results and Concluding remarks are modeled using sparse vectors. Moreover, the signal conveyed
represented in Sections IV and V, respectively. distance is large relative to the transmit-receive antenna spac-
ing in each terminal. Hence, the encountered scatterers in each
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION chip period are identical between different antennas. In other
A. System Model words, the delays of different paths are the same in all the
channel ensembles between two terminals. Thus, the sparsity
Consider a mMIMO BS A and a user terminal (UT) B pattern of varying channel pairs could be assumed to be the
which are equipped by NA and NB transmit-receive antennas, same while the channel attenuation is different. Since each
respectively. Each terminal employs OFDM signaling on path channel attenuation is composed of multiple distinct scat-
each of u-th antenna with u = 0, 1, . . . , NA (NB ) − 1 to terers with zero-mean and identically independent distributed
exchange data with the other terminal. The emitted OFDM (i.i.d.) subpaths, it is assumed to be CN (0, σ 2 ). Hence, the
symbol is constructed using N subcarriers which are spaced channel coefficient could be represented as
by Δ f in hertz while prefaced using NG cyclic prefix (CP). I
−1
The received signal by v-th antenna of the receiver with huv (l ) = αuv (i )g(lT − τ (i )) (3)
v = 0, 1, . . . , NB (NA ) − 1, is passed through a frequency i=0
selective channel which can be determined by a Finite
where l ∈ [0, 1, . . . , L] is the channel path index, τ (I − 1) ≥
Impulse Response (FIR) filter and constructed of multiple
· · · ≥ τ (1) ≥ τ (0) are the respective paths’ delay, αuv is
paths. Ts = 1/(N Δf ) is used as the sampling period in the
the corresponding paths’ gain, and g(.) is the shaping pulse
receiver. Ignoring the CP fallen chips and transforming other
in the continuous domain. The shaping pulse is zero out-
chips into Fourier transform domain, the received signal is
side the interval [0, Tg ], where Tg is the integer multiple
ready to be selected on the pilot subcarriers. This selection is
of chip time T. Without loss of generality, we assumed that
made on each of the antennas and respective pilot sequence.
τ (i) are integer multiples of T. Thus, the number of chan-
As a consequence, the received signal on the n-th pilot
nel paths, caused by channel itself and shaping filter is
symbol can be derived as
derived by L = τ (I − 1)/T + Tg /T + 1. Furthermore, we
L−1
assume that L is lower than TG /T . Using the mentioned nota-
yuv (n) = huv (l )e(l , u, n) + wuv (n) (1) tions, we can represent the channel impulse response using
l=0 huv = [huv (0), huv (1), . . . , huv (L − 1)]T .
where n = 0, 1, . . . , Np − 1 is the pilot index,
l = 0, 1, . . . , L − 1 is the paths index and C. Problem Formulation
e(l , u, n) = e −j 2πlpu (n)/N . Moreover, pu (n) represents the Up to now, the channel model and the system model
n-th pilot of u-th transmit antenna with n = 0, 1, . . . , Np . are represented. In order to estimate the channel impulse
wuv (n) denotes the received Additive White Gaussian Noise response, we have developed a formulation to utilize the
(AWGN) sample between the u-th transmit antenna, and v-th channel common sparsity together with the large scale charac-
received antenna in the n-th sample of pilot which obeys the teristic. To develop the model we consider Multiple Input and
CN (0, σw 2 ). Besides, h (l ) is the l-th channel tap between
uv Single Output (MISO) since it is practical according to the
the u-th transmitter antenna, and v-th received antenna. The mMIMO systems. As mentioned earlier, the channel impulse
vector representation of (1) can be formulated as response could be represented as huv wherein MISO chan-
yuv = Φu huv + wuv (2) nels u = 0, 1, . . . , NB − 1, and v = 0. Using the channel
ensembles in MISO system, we can represent the channel
where yuv = [yuv (0), yuv (1), . . . , yuv (Np − 1)]T is the
matrix as H = [h00 , h10 , . . . , h(NA −1)0 ] where the size of
received pilot vector, Φu = diag{xu }Fu is the mea-
H is L × NA . By collecting all the received pilot sequences,
surement matrix with xu = [xu (0), xu (1), . . . , xu (Np −
we can represent the Np × NA received pilot matrix as
1)]T as the transmitted pilot vector and Fu as the par-
Y = [y00 , y10 , . . . , y(NB −1)0 ]. Hence, the extension of (2)
tial Fourier matrix with Np rows corresponding to the pilot
could be represented in matrix form for MISO case as
sequence of transmit antenna u and L first columns, huv =
[huv (0), huv (1), . . . , huv (L − 1)]T is the channel coefficient Y = ΦH + W (4)
vector and wuv = [wuv (0), wuv (1), . . . , wuv (Np − 1)]T is the
where W is the AWGN matrix with corresponding columns
received noise vector. Besides, in this formulation the super-
according to the wu0 . Φ is the measurement matrix with
script T is the sign of transpose and diag{.} denotes the
the size of Np × L. For the sake of simplicity, we drop the
diagonal representation of the corresponding vector.
index 0 for the UT antenna. Defining μ norm for matrices
as μ(H) = card{Hu 2 = 0} where Hu is the u-th column
B. Channel Model
of H, and card{S } is the number of elements in the set S,
As mentioned, the channel between transmit-receive pair estimating the channel could be accomplished by following
which is denoted by huv is consist of L resolvable paths. These optimization criterion.
resolvable paths result from L scatterers which are encoun-
tered by the signal conveying from the transmitter antenna u min Y − ΦH22
to the receiver antenna v. K of these L scatterers are significant s.t. μ(H) ≤ K . (5)
378 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Apparently, K is the maximum sparsity of the columns of Algorithm 1 StFBP-Based Channel Estimation
H. We called the objective function as F(H) = Y − ΦH22 . 1: R(0) = Y, λ0 = φ and t = 0
Utilizing μ(H), we can exploit the joint sparsity of the 2: while stop criterion not met do
channel ensembles in the system. The objective function 3: Proxy of the signal is formed by P = ΦH R(t−1)
in (5) represents the error of channel estimation method, 4: r = Pθ
and the constraint controls the sparsity order of the channel 5: Select good atoms according to {i (t) :ri ≥ τ }
6: Merging supports of the previous iteration and the present one
ensembles in H. Moreover, using μ-norm definition, the block
λ(t) = λ(t−1) ∪ i (t)
sparsity of the channels are exploited. 7: Calculate H(t) by solving Eq. (10)
(t)
8: δF = F(H(t−1) ) − F(H(t) )
III. P ROPOSED S TAGE -W ISE F ORWARD -BACKWARD 9: while 1 do
(t)
P URSUIT 10: j (t) = arg minj ∈λ(t) F(H(t) − Hj )
(t) (t)
To solve the optimization problem in Eq. (5), pseudo-inverse 11: δB = F(Hj ) − F(H(t) )
of the matrix Φ is needed. Calculating the channel coefficients (t) (t)
12: if δB ≥ 0.5δF /|i (t) | then
without any consideration of channel characteristics, would
13: Update residual R(t) = Y − Φλ(t) H(t)
cause some difficulties in inversion of the matrix. Actually,
14: break
this inversion didn’t consider any characteristics of channels. 15: end if
In order to consider the channel impulse response, we have 16: Exclude bad atom λ(t) = λ(t) − j (t)
considered the large scale fading phenomena. Each path of the 17: Update H(t) by solving Eq. (10).
channel is changed in the power, according to the exponential 18: Update residual R(t) = Y − Φλ(t) H(t)
distribution based on the distance of the origin. In fact, each 19: end while
path could be defined by the following equation 20: end while
L−1
Z (H ) = hl 22 ωl (6)
l=0
a greedy algorithm is a particular case of StFBP which the
forward selection is present, but the backward fixing is absent,
where consequently, it cannot fix its own mistakes in previous steps.
1, if l = 0 Moreover, FBP algorithm constructs the new subspace by
ωl = σ (7)
lα , otherwise adding just one atom to the previous subspace, and in the back-
ward steps, it reconstructs the subspace by omitting bad atoms.
where 0 ≤ l ≤ L-1, σ is the coefficient of the pathloss model
To increase the speed of convergence, we have proposed to add
and α is the environment factor. Accordingly, by defining the
multiple elite atoms in each forward step and in the backward
new vector gl = [h1 (l ), h2 (l ), . . . , hNA (l )] and consequently
step it excludes various atoms, too. In this case, the speed
the matrix G = [g0T , g1T , . . . , gL−1
T ]T , the Z (H ) could be
of convergence would increase, while the accuracy is guar-
considered as the anteed using backward steps. As a consequence, the proposed
Z(H) = Tr(GΩGH ) (8) StFBP algorithm could be compared with its greedy one called
StOMP, where StOMP is the particular case of StFBP without
where Ω = diag{ω0 , ω1 , . . . , ωL−1 }. Thus, in order to backward steps to increase the estimation accuracy.
consider the channel ensembles, we should change the In Algorithm 1, |i (t) | denotes the number of selected atoms
(t)
optimization problem in Eq. 5 as in forwarding step, and Hj represents the H(t) while j (t) -th
min Y − ΦH22 + λTr(GΩGH ) column of it is omitted. Furthermore, τ is the threshold of
selection in forwarding steps and θ is NB × 1 ones vector
s.t. μ(H) ≤ K (9)
(i.e., all the elements equal to one).
where λ is the regulation factor. In order to solve the problem, In algorithm 1 each step is included of different
we try to consider the gradient of the optimization problem multiplications which will be mentioned in the following. The
equal to zero. proxy matrix costs 4NA Np real multiplications. Moreover, In
9th step 4NA Np real multiplications is encountered. In each
ΦH ΦH + λ/2HΩ = ΦH Y (10) step t the cardinality of λ (t) could be defined by |λ λ(t) |; hence,
th
the overall cost of 10 step is evaluated by 4|λ λ(t) | + 4NB .
which is a Lyapunov equation [10]. Hence, by solving a
(t)
Lyapunov equation, we can calculate the channel coefficients. In backward stages, calculating j (t) and δB cost 4(|λ λ(t) | −
In order to handle the optimization in (9), we have proposed 1) + 4NB . Furthermore, defining residual in 15 and 20th
th
a Stage-wise Forward-Backward Pursuit (StFBP) based on [9] step costs 4|λ λ(t) | real multiplications. Finally, updating H(t)
where 0 norm was used. The algorithm is represented in in each step consumes 4NA Np real multiplications.
details in Algorithm 1. Eq. (9) could be solved using three
different methods as convex relaxation, greedy processes, and
message-passing algorithms. StFBP which is based on the IV. N UMERICAL R ESULTS
Message-Passing algorithm is used here because of its forward The simulation results are represented in this section to
selection and backward fixing. Specifically, StOMP which is compare the performance of the proposed method with other
AKBARPOUR-KASGARI AND ARDEBILIPOUR: mMIMO-OFDM CHANNEL ESTIMATION VIA DISTRIBUTED CS 379
R EFERENCES
known approaches. In the simulations, the number of subcarri-
ers were N = 2048 while there were 15 kHz spacing between [1] Z. Gao, L. Dai, W. Dai, B. Shim, and Z. Wang, “Structured compres-
sive sensing-based spatio-temporal joint channel estimation for FDD
them. The BS was equipped with large number of antennas. massive MIMO,” IEEE Trans. Commun., vol. 64, no. 2, pp. 601–617,
Moreover, in each transmitter, the interleaved bits are modu- Feb. 2016.
lated using 16 Quadrature-Amplitude-Modulation (16-QAM). [2] X. Ma, F. Yang, S. Liu, J. Song, and Z. Han, “Design and
optimization on training sequence for mmWave communications: A
Transmit power of both data and pilot subcarriers are the same. new approach for sparse channel estimation in massive MIMO,”
Besides, the total available power is divided among transmit- IEEE J. Sel. Areas Commun., vol. 35, no. 7, pp. 1486–1497,
ting antennas. Channels are distributed according to Extended Jul. 2017.
Pedestrian A (EPA) available in LTE-A standard channel [3] D. Fan, F. Gao, G. Wang, Z. Zhong, and A. Nallanathan, “Angle domain
signal processing aided channel estimation for indoor 60GHz TDD/FDD
model [11]. Moreover, the channel taps are scaled to support massive MIMO systems,” IEEE J. Sel. Areas Commun., vol. 35, no. 9,
NA NB total channel power. In the detection, after channel pp. 1948–1961, Sep. 2017.
estimation, the zero-forcing equalizer omits the impact of the [4] Y. Nan, L. Zhang, and X. Sun, “Efficient downlink channel estimation
scheme based on block-structured compressive sensing for TDD mas-
channel. Furthermore, the simulation results are averaged on sive MU-MIMO systems,” IEEE Wireless Commun. Lett., vol. 4, no. 4,
2000 independent runs. pp. 345–348, Aug. 2015.
At first, the proposed channel estimation is compared with [5] J. W. Choi, B. Shim, and S.-H. Chang, “Downlink pilot reduction for
massive MIMO systems via compressed sensing,” IEEE Commun. Lett.,
the other well-known DCS-based channel estimation methods vol. 19, no. 11, pp. 1889–1892, Nov. 2015.
called ABSP from [4]. The results are shown in Fig. 1. This [6] M. Masood, L. H. Afify, and T. Y. Al-Naffouri, “Efficient coordinated
simulation is carried on utilizing Np = 32 pilot subcarriers recovery of sparse channels in massive MIMO,” IEEE Trans. Signal
Process., vol. 63, no. 1, pp. 104–118, Jan. 2015.
for each of the transmitting antenna. As represented in Fig. 1, [7] D. Lee, “MIMO OFDM channel estimation via block stagewise
the proposed StFBP approach outperforms the other method. orthogonal matching pursuit,” IEEE Commun. Lett., vol. 20, no. 10,
From NMSE perspective, the proposed approach is superior to pp. 2115–2118, Oct. 2016.
[8] A. Akbarpour-Kasgari and M. Ardebilipour, “Probability-based pilot
ABSP method almost 7 dB. Moreover, increasing SNR more allocation for MIMO relay distributed compressed sensing based chan-
than 30 dB is seemingly useless for ABSP. The main advantage nel estimation,” EURASIP J. Adv. Signal Process., vol. 2018, no. 1,
of the proposed StFBP approach to the well-known ABSP is pp. 1–18, Mar. 2018.
the fixing process which is absent in ABSP. As a consequence, [9] T. Zhang, “Adaptive forward-backward greedy algorithm for learn-
ing sparse representations,” IEEE Trans. Inf. Theory, vol. 57, no. 7,
previous steps selected bad atoms are omitted, and the NMSE pp. 4689–4708, Jul. 2011.
is getting lower while in ABSP the previous steps selected bad [10] A. Jameson, “Solution of the equation AX + XB = C by inversion
atoms are not omitted and remain in the last estimated channel of an M M or N N matrix,” SIAM J. Appl. Math., vol. 16, no. 5,
pp. 1020–1023, Sep. 1968.
impulse response. Hence, adjusting a reasonable threshold for [11] User Equipment Radio Transmission and Reception (Rel. 12), V12.9.0,
selecting atoms in each step is essential in ABSP while in 3GPP Standard TS 36.101, Oct. 2015.
380 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—We propose a modeling framework composed of 802.15.4. To fill this gap and based on [5], we propose here
a Markov chain and the related coupling equations to evalu- a Markovian model to evaluate the performance of hetero-
ate the performance of unslotted IEEE 802.15.4 wireless sensor geneous unslotted 802.15.4 networks with different “classes”
networks based on CSMA/CA medium access control. Different of nodes (each class generating traffic according to a class-
from the related literature, the proposed model is able to cap-
specific rate). The main contributions of this letter with respect
ture heterogeneous classes of nodes with class-specific traffic
generation rate and includes a more refined calculation of the to [5] are: first, instead of using a different Markov chain to
probability of finding the channel busy during the carrier sens- model each node, we use a Markov chain for each class of
ing process. The proposed model is used to derive class-specific nodes thus decreasing the complexity of the model; second, all
performance figures including the probability of a successful the aforementioned related work assume that the probability of
transmission and the average delay for a successful/unsuccessful finding the channel busy during a Clear Channel Assessment
transmission. (CCA) does not depend on the backoff stage of the node;
Index Terms—Wireless sensor networks, IEEE 802.15.4, although this effect is not relevant when all nodes are saturated,
medium access control (MAC). the accuracy of the model is compromised when saturated and
unsaturated nodes coexist in the same network; to this extent,
we show how to keep track of the backoff stage when deriving
the probability of finding the channel busy. Third, we derive
I. I NTRODUCTION a more accurate expression for the collision probability expe-
HE DIFFUSION of wireless sensor networks based on the
T IEEE 802.15.4 standard has stimulated research efforts
on the performance evaluation of the its Medium Access
rienced by the nodes. Specifically, we consider that if there is
a collision, the probability that any other node performs CCA
must be conditioned by the fact that the channel is not busy.
Control (MAC) scheme. For instance, a Markov chain model
for unslotted, acknowledged 802.15.4 CSMA/CA is proposed
in [1] for unsaturated nodes considering a deterministic idle II. P ROPOSED M ODEL
time after every transmission. Markov chains are also used to We assume a scenario with M classes of nodes, each class l
model the prioritized contention access for industrial appli- formed by Nl nodes generating packets according to a Poisson
cation of IEEE 802.15.4 [2] or to analyze the performance process of rate λl . All the nodes access the medium accord-
of a modified MAC mechanism assuming homogeneous and ing to the unslotted IEEE CSMA/CA 802.15.4 protocol; when
Poisson traffic [3]. A different approach is followed in [4], a node tries to transmit a new packet, it waits for a random
where the modeling of unslotted CSMA/CA is carried out number of backoff slots in the range [0, 2BE −1], being BE the
using Event Chains Computation instead of Markov chains, backoff exponent that is initialized to mmin . When the back-
showing that this approach better models the case where nodes off counter is 0, the node performs CCA to determine whether
start reporting data simultaneously when an event is detected. the transmission channel is empty. If not, BE is increased by 1
The number of works on heterogeneous 802.15.4 networks until it reaches the limiting value mmax and the node waits for
is more limited. The main works modeling heterogeneous a new random backoff period generated with the new value of
traffic in IEEE 802.15.4 networks are [5] for acknowl- BE. This process is repeated until the number of failed CCAs
edged unslotted 802.15.4 and [6] for unacknowledged slotted exceeds the parameter m. In that case, the packet is discarded
due to a channel access failure. On the contrary, if the chan-
Manuscript received September 12, 2018; accepted September 24, 2018.
Date of publication October 1, 2018; date of current version April 9, 2019. nel is empty, the node switches from the listening mode to
This work was supported in part by the Spanish Government from the Mobility the transmitting mode, transmits the packet and waits for the
Program of the Ministerio de Educacion, Cultura y Deporte under Grant reception of the ACK. If the ACK is not received, then the
CAS17/00624, in part by the Ministerio de Ciencia e Innovación under Project packet is retransmitted following the CSMA/CA mechanism
TEC2014-52969-R, in part by the Universidad de Zaragoza under Project
UZ2018-TEC-04, and in part by the Centro Universitario de la Defensa under
described above. This process can be repeated up to n times.
Project CUD2017-18. The associate editor coordinating the review of this When this value is exceeded, the packet is discarded due to a
paper and approving it for publication was H. Zhou. (Corresponding author: collision failure.
J. Ortín.) To model the backoff, sensing and transmitting states of the
J. Ortín is with the Centro Universitario de la Defensa, 50090 Zaragoza,
Spain (e-mail: jortin@unizar.es).
nodes, we rely on the Markov chain model shown in Fig. 1 and
M. Cesana and A. E. C. Redondi are with the Dipartimento di Elettronica, proposed in [5]. A state in the chain is the tuple (i, j, r), being
Informazione e Bioingegneria, Politecnico di Milano, 20133 Milan, Italy i the backoff stage, j the backoff counter and r the retransmis-
(e-mail: matteo.cesana@polimi.it; alessandroenrico.redondi@polimi.it). sion counter. The backoff stage and the retransmission counter
M. Canales and J. R. Gállego are with the Instituto de Investigación
en Ingeniería de Aragón, Universidad de Zaragoza, 50018 Zaragoza, Spain
are limited by the parameters m and n respectively. Similarly, j
(e-mail: mcanales@unizar.es; jrgalleg@unizar.es). ranges from 0 to Wi = 2BEi − 1, with BEi the backoff expo-
Digital Object Identifier 10.1109/LWC.2018.2873347 nent corresponding to the backoff stage i. In the states with
2162-2345 c 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/
redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ORTÍN et al.: ANALYSIS OF UNSLOTTED IEEE 802.15.4 NETWORKS WITH HETEROGENEOUS TRAFFIC CLASSES 381
where pl (0, 0, 0) is the steady state probability of state (0, 0, 0) where Ni = Ni if i = l and Ni = Ni − 1 if i = l . On
for nodes of class l and yl = Pc,l (1 − αm+1 ). The expression the other hand, αl,ack is the probability of finding the channel
for pl (0, 0, 0) is given in Eq. (2) at the top of the next page. busy because of the successful transmission of a packet, which
happens when only one packet is being transmitted
1 This value is L = L + t
s ack + Lack + IFS , with L the total transmission
N
time of a packet, tack is the ACK waiting time, Lack is the transmission αl,ack = Lack Ni τi (1 − τi )Ni −1 1 − τj j . (8)
time of the ACK and IFS is the Inter-Frame Spacing. i∈M j ∈M
2 This value is L = L + t
c m,ack , with tm,ack the timeout of the ACK. j =i
382 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
⎧
⎪
m+1 1−αm+1 1−yln+1 m+1 1−yl
n+1
⎪ 12 1−(2α
⎪ 1−2α
l)
W 0 + 1−α
l
1−y + Ls (1 − P c,l ) + L c P c,l 1 − αl 1−yl
⎪
⎪ l l l
⎪
⎪ −1
⎪
⎪ m+1
1−qcf ,l αl (1−yln+1 ) 1−qcr ,l n+1 1−qsuc,l (1−αm+1 )(1−yln+1 )
⎪
⎨ + ql + ql y l + (1 − Pc,l ) l
, if m < m̂ = mmax − mmin
1−yl ql 1−yl
pl (0, 0, 0) =
(2)
⎪
⎪ 1 1−(2αl )
m̂+1 1−αlm̂+1
mb +1 m̂+1 1−αl
m−m̂
1−yl n+1
m+1
⎪
⎪ 2 1−2αl W 0 + 1−αl + 2 + 1 αl 1−αl 1−yl + Ls (1 − P c,l ) + Lc P c,l 1 − αl
⎪
⎪
⎪
⎪ −1
⎪
⎪ 1−y n+1
1−q αm+1
(1−y n+1
) 1−q 1−q (1−α m+1
)(1−y n+1
)
⎩×
1−yl +
l
ql
cf ,l l
1−yl
l
+ qlcr ,l yln+1 + ql
suc,l
(1 − Pc,l ) l
1−yl
l
, otherwise
(1)
The term αl can be computed as We show now the expressions for the average delay expe-
rienced by a packet in a successful transmission and when it
(1) (0) is discarded due to a channel access failure or a retry limit.3
αl ≈ 1 · P (E0 ) + αl (1 − P (E0 )), (9)
Let Tsuc,l be the delay of a packet transmitted success-
where E0 is the event that the packet that has caused the CCA fully, Cj the event of having a successful transmission after
failure in the first attempt is still being transmitted when the (j )
j previous collisions, and Tsuc,l the delay experienced by a
second CCA is performed. From the Markov chain of Fig. 1, packet when the event Cj occurs. Following [5],
the probabilities of the states (−1, j, r), with 0 ≤ j < Ls
and a fixed r, are all equal. Therefore, if a device performs
n
(j )
CCA and the channel is busy with a successful transmission, E Tsuc,l = P (Cj )E Tsuc,l , (15)
the node that is transmitting can be at any of the states of j =0
the form (−1, j, r) with equal probability (i.e., it follows a with
discrete uniform distribution in [0, Ls − 1]). Likewise, if the j
j
channel is busy with a collided transmission, it can be at any 1 − Pc,l 1 − αlm+1 Pc,l 1 − αlm+1
of the states of the form (−2, j, r) with equal probability P (Cj ) = n+1 (16)
(i.e., it follows a discrete uniform distribution in [0, L−1]). 1 − Pc,l 1 − αlm+1
On the other hand, the first stage of the backoff also follows and
a discrete uniform distribution in [0, W0 − 1]. Therefore, if j
(W0 −1)/2+L−W0
, if L > W0
P (C > B0 ) = L (11) E[Tb ] = P (Di )E Tb,i , (18)
L−1 , otherwise,
2W0 i=0
where P (Di ) is the probability of finding the channel idle at
and
the i + 1th attempt, given that the channel has been found
(W0 −1)/2+Ls −W0 busy in the preceding i attempts and the packet has not been
Ls , if Ls > W0
P (S > B0 ) = Ls −1 (12) discarded due to a channel access failure; and E[Tb,i ] is the
2W0 , otherwise. expected time a node spends in backoff or sensing states given
the event Di . P (Di ) can be calculated as
Eqs. (1), (3) and (5) form a system of coupled nonlinear
equations with variables τl , αl and Pc,l that can be solved αi 1 − αl
P (Di ) = m l k = αli , (19)
numerically to obtain the point of operation of the network.
k =0 αl 1 − αlm+1
From them, different performance metrics can be obtained.
First, we derive the probabilities of a discarded packet due while
to a collision failure, Pcr ,l , and due to a channel access failure, i
W −1
Pcf ,l . From the Markov chain of Fig. 1, we have E Tb,i = (i + 1)tCCA + tb k , (20)
2
k =0
n+1
αlm+1 1 − Pc,l 1 − αlm+1 with tb and tCCA the durations of a backoff slot and CCA.
Regarding the delay suffered by a packet when it is dis-
Pcf ,l = (13) carded due to a channel access failure, Tcf ,l , it can be derived
1 − Pc,l 1 − αlm+1
following the same approach used to compute Tsuc,l
and
n
(j )
n+1 E Tcf ,l = P (Fj )E Tcf ,l , (21)
Pcr ,l = Pc,l 1 − αlm+1 . (14) j =0
3 We consider only the time from the instant the packet is ready to be
With this, the probability of a successful transmission is transmitted until an ACK is received or until it is discarded because of the
Psuc,l = 1 − Pcf ,l − Pcr ,l . aforementioned failures (i.e., we do not include queuing time in this analysis).
ORTÍN et al.: ANALYSIS OF UNSLOTTED IEEE 802.15.4 NETWORKS WITH HETEROGENEOUS TRAFFIC CLASSES 383
Fig. 2. Performance metrics for a network with 1 saturated stations and 50 unsaturated stations as a function of the un traffic rate of unsaturated stations.
where Fj is the event of having a channel access failure after j λ increases, both probabilities converge since the unsaturated
(j ) nodes tend to behave like the saturated one.
previous collisions and Tcf ,l is the delay suffered by a packet
on the occurrence of the event Fj . It can be easily derived Fig. 2(b) shows the average delay incurred by a packet in
a successful transmission, whereas Fig. 2(c) shows the aver-
that P (Fj ) = P (Cj ) and
age delay suffered by a packet when it is discarded due to a
j
−1 m
Wk − 1
channel access failure or a retry limit. This delay is
(j )
E Tcf ,l = E[Tb ] + jTc + (m + 1)tCCA + tb . pcf ,l pcr ,l
2
h=0 k =0 Tunsuc,l = T + Tcr ,l . (24)
(22) pcf ,l + pcr ,l cf ,l pcf ,l + pcr ,l
j −1 In both cases, the delay is higher for unsaturated nodes as
For j = 0, the term h=0 E[Tb ] is 0.
they will find the channel busy more frequently and will have
The delay suffered by a packet when it is discarded due to
to perform more backoffs.
a retry failure Tcr ,l is
The validity of the proposed model has also been tested
varying the rates of each class as well as the number of nodes
E Tcr ,l = (n + 1)Lc + E[Tb ]. (23)
per class, obtaining similar results in terms of accuracy.
Finally, the probabilities of having a packet ready to be
transmitted in idle state, after a successful transmission, after
IV. C ONCLUSION
a channel access failure and after a retry limit failure are ql =
1−e −λl tb , qsuc,l = λE[Tsuc,l ], qcf ,l = λE[Tcf ,l ] and qcr ,l = We have studied the performance of unslotted 802.15.4 with
λE[Tcr ,l ]. A detailed explanation of their derivation can be heterogeneous classes of nodes and class-specific packet gen-
found in [5]. Note that in case a station is saturated, qsuc,l = eration rate. We have validated our model in a scenario with a
qcr ,l = qcf ,l = 1 and the idle state in the Markov chain of saturated node and a fixed number of unsaturated nodes with
Fig. 1 is removed. varying traffic rate. The results show that the proposed model
reflects the simulated performance of the reference network
scenario better than previous state-of-art approaches.
III. R ESULTS AND M ODEL VALIDATION
To stress-test the proposed model, we have considered a
network scenario with 50 sensor nodes generating packets at R EFERENCES
a low packet rate, and one node with a backlog of saturated [1] K. Govindan, A. P. Azad, K. Bynam, S. Patil, and T. Kim, “Modeling
traffic. The results obtained through the model are validated and analysis of non beacon mode for low-rate WPAN,” in Proc. 12th
Annu. IEEE Consum. Commun. Netw. Conf. (CCNC), Jan. 2015,
against a system-level, discrete-event simulator of the IEEE pp. 549–555.
802.15.4 PHY/MAC layers. All the simulated results represent [2] M. P. R. S. Kiran and P. Rajalakshmi, “Performance analysis
the average of 108 generated packets. The MAC parameters of CSMA/CA and PCA for time critical industrial IoT applica-
used in the simulations are mmin = 4, mmax = 7, m = 4, tions,” IEEE Trans. Ind. Informat., vol. 14, no. 5, pp. 2281–2293,
n = 0, L = 7, Lack = tm,ack = 2, IFS = tack = 0, tb = May 2018.
[3] S. R. Pattanaik, P. K. Sahoo, and S.-L. Wu, “Performance analysis of mod-
20 · 16 μs, tCCA = 8 · 16 μs, and tta = 12 · 16 μs. ified IEEE 802.15.4e MAC for wireless sensor networks,” in Proc. 14th
Fig. 2 shows the performance of the different classes of ACM Symp. Perform. Eval. Wireless Ad Hoc Sensor Ubiquitous Netw.,
nodes as the traffic generation rate of the unsaturated nodes Nov. 2017, pp. 25–31.
varies, comparing the results of our model with those of the [4] D. D. Guglielmo, F. Restuccia, G. Anastasi, M. Conti, and S. K. Das,
“Accurate and efficient modeling of 802.15.4 unslotted CSMA/CA
one proposed in [5]. Fig. 2(a) depicts the probability of a suc- through event chains computation,” IEEE Trans. Mobile Comput., vol. 15,
cessful transmission for the saturated and unsaturated nodes. no. 12, pp. 2954–2968, Dec. 2016.
As expected, this probability is close to 1 for the saturated [5] P. D. Marco, P. Park, C. Fischione, and K. H. Johansson, “Analytical
node when the traffic of the unsaturated nodes is very low as modeling of multi-hop IEEE 802.15.4 networks,” IEEE Trans. Veh.
Technol., vol. 61, no. 7, pp. 3191–3208, Sep. 2012.
it finds the channel idle most of the time. On the contrary, [6] J. Zhu, Z. Tao, and C. Lv, “Performance evaluation for a beacon enabled
this probability starts approximately at 0.82 for the unsatu- IEEE 802.15.4 scheme with heterogeneous unsaturated conditions,” Int.
rated ones as they have to contend with the saturated one. As J. Electron. Commun., vol. 66, no. 2, pp. 93–106, 2012.
384 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—This letter analyzes the performance of a full-duplex two-way relaying for SWIPT. More specifically, AF relay-
decode-and-forward relaying network over the generalized κ-μ ing with multiple antennas and zero-forcing was deployed
fading channel. The relay is energy-constrained and relies entirely in this letter. Other studies exploiting jamming signals for
on harvesting the power signal transmitted by the source based on energy-harvesting (EH) have recently appeared in [6] and [7].
the time-switching relaying protocol. A unified analytical expres- None of the aforementioned works considered FD
sion for the ergodic outage probability is derived for the system EH-enabled relay networks over generalized κ − μ fading
under consideration. This is then used to derive closed-form
channels. In contrast, and motivated by this lack of analytical
analytical expressions for three special cases of the κ-μ fading
model, namely, Rice, Nakagami-m and Rayleigh. Monte Carlo analysis, we present in this letter a thorough performance eval-
simulations are provided throughout to validate our analysis. uation of FD EH-enabled relay networks over such generalized
fading channels. Specifically, DF and time-switching relaying
Index Terms—Decode-and-forward (DF) relaying, energy har- (TSR) protocols are deployed at the relay. The motivations of
vesting, full-duplex (FD), generalized κ − μ fading. our work come from the following two factors. Firstly, the
κ-μ model is a small-scale fading model and is able to char-
I. I NTRODUCTION acterize the scattering cluster in homogeneous communication
environments including Rice (κ = k, μ = 1), Nagakami-m
IMULTANEOUS wireless information and power trans-
S fer (SWIPT) in full-duplex (FD) relaying networks
has recently attracted a great deal of research attention.
(κ → 0, μ = m) and Rayleigh (κ → 0, μ = 1) [8]–
[10]. Secondly, FD communication allows devices to operate
on the same frequency, which potentially doubles the spec-
For instance, Zhong et al. [1] considered a dual-hop FD tral efficiency, and, because of this, FD has become a viable
SWIPT network with both amplify-and-forward (AF) and option for next generation wireless communication networks.
decode-and-forward (DF) relaying protocols equipped with a Thus, the network studied herein is meaningful and valuable
single-antenna. Several analytical expressions of the achiev- for consideration.
able throughput were derived. Instead of the single-antenna The main contribution of this letter resides in deriving
relay, Mohammadi et al. [2] extended the work in [1] to a novel unified analytical expression for the ergodic outage
include multiple-input multiple-output (MIMO) FD relaying. probability of the proposed system over the generalized κ-μ
Unlike [1] and [2], which focused on Rayleigh fading chan- fading channel. In addition, closed-form expressions for the
nels, the work in [3] analyzed the performance of a FD SWIPT aforementioned special cases of the κ-μ fading scenario are
system in indoor environments characterized by log-normal presented. The derived expressions were used to investigate
fading. The study in [4] considered the outage probability of the impact of several system and fading parameters on the
FD SWIPT networks over α−μ fading channels. Furthermore, performance. Results show that the system performance can
Okandeji et al. [5] studied physical layer security in FD be enhanced considerably as the fading parameters κ and μ are
increased. It is also shown that as the loop-back interference,
Manuscript received June 6, 2018; revised August 4, 2018; accepted associated with FD relaying, increases the outage probability
August 31, 2018. Date of publication October 1, 2018; date of current ver-
sion April 9, 2019. The associate editor coordinating the review of this performance deteriorates drastically.
paper and approving it for publication was L. Bai. (Corresponding author:
Khaled Rabie.)
II. S YSTEM M ODEL
K. Rabie and B. Adebisi are with the School of Engineering, The considered network consists of a source (S), a relay (R)
Manchester Metropolitan University, Manchester M15 6BH, U.K. (e-mail:
k.rabie@mmu.ac.uk; b.adebisi@mmu.ac.uk).
and a destination (D). The end nodes are equipped with a
G. Nauryzbayev is with the Division of Information and single-antenna whereas R, based on DF, has two antennas
Computing Technology, College of Science and Engineering, Hamad and operates in the FD mode. It is assumed that there is no
Bin Khalifa University, Qatar Foundation, Doha, Qatar (e-mail:
nauryzbayevg@gmail.com).
direct link between the end nodes due to severe shadowing and
O. S. Badarneh is with the Electrical and Communication Engineering path-loss effect [1]–[3]; hence, all communication is accom-
Department, School of Electrical Engineering and Information plished over two phases. The S-to-R, R-to-D and loop-back
Technology, German Jordanian University, Amman 11180, Jordan (e-mail:
osamah.badarneh@gju.edu.jo).
interference channel coefficients, denoted as h1 , h2 and h3 ,
X. Li is with the School of Physical and Electronics Engineering, respectively, are assumed to be independent but not necessar-
Henan Polytechnic University, Zhengzhou 454000, China (e-mail: ily identical following the κ-μ distribution with a probability
lixingwang@hpu.edu.cn).
density function (PDF)
M.-S. Alouini is with the Computer, Electrical and Mathematical Science
and Engineering (CEMSE) Division, King Abdullah University of Science and μi −1 φi z φi z
Technology, Thuwal 23955, Saudi Arabia (e-mail: slim.alouini@kaust.edu.sa). fh 2 (z ) = Υi z 2 exp − Iμi −1 2μi , (1)
Digital Object Identifier 10.1109/LWC.2018.2873360 i Ωi Ωi
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
RABIE et al.: FD EH ENABLED RELAY NETWORKS IN GENERALIZED FADING CHANNELS 385
where i ∈ {1, 2, 3}, Ωi = E[hi2 ], φi = μi (1 + κi ), arrive at a tractable expression, we use the series representation
μ +1
μi (1+κi )
i
2 of Iμ3 −1 (·) [12, eq. (8.445)]; that is
Υi = μi −1 μi +1 , Ip [·] is the modified Bessel func- ∞
1
exp(μi κi )κi 2 Ωi 2
Iμ3 −1 (2Λ) = Λμ3 −1+2q , (8)
tion of the first kind with arbitrary order p [11, eq. (9.6.20)], Γ(μ3 + q)q!
μi represents the number of the multipath clusters and κi > 0 q=0
denotes the ratio between the total powers of the domain com- where Λ = μ3 κ3 (1 + κ3 )z and Γ(·) is the Gamma function
ponents and the scattered waves. The path-loss exponents for [12, eq. (8.310.1)].
the S-to-R and R-to-D links are denoted by ξ1 and ξ2 , respec- Now, replacing (8) in (1) and then integrating, we can
tively. It is assumed that perfect channel state information express the cumulative distribution function (CDF) of Z,
(CSI) is available at all receiving nodes, and that R has no FZ (υ), as
power supply and operates by harvesting the RF signal com-
μ3 −1+2q
ing from S. The energy used for information processing at R is ∞
μ3 κ3 (1 + κ3 )
negligible and hence all the harvested energy will be utilized FZ (υ) = Υ3 J0 , (9)
to forward the source information. Γ(μ3 + q)q!
q=0
As mentioned earlier, the TSR protocol is used for EH at R, where
in which the time frame T is divided into two consecutive
b
time slots: α T and (1 − α)T, which are used for EH and J0 =
υ
z μ3 +q−1 exp(−φ3 z )dz
S-to-R / R-to-D information transmissions, respectively; where 0
0 ≤ α ≤ 1 is the EH time factor.
(β) −μ3 −q φ3 b
For the sake of brevity, we omit the mathematical modeling = φ3 γ μ3 + q, , (10)
of the received signals at R and D and present only the cor- υ
responding signal-to-noise ratios (SNRs). Readers may refer where b = 1−αηα and γ(·, ·) is the lower incomplete Gamma
to [1] and [3] for more details. The SNRs at R and D nodes function [12, eq. (8.350.1)]. Note that (β) in (10) is obtained
are expressed, respectively, as with the help of [12, eq. (3.351.1)].
Ps h12 1 Substituting (10) into (9), along with some straightforward
γr = = , and (2) manipulations, yields
ξ
Pr d1 1 h32 ζh32
∞
1 (μ3 κ3 )q φ3 b
Pr h22 ζPs h12 h22 FZ (υ) = γ μ3 + q, . (11)
γd = = , (3) exp(κ3 μ3 ) Γ(μ3 + q)q! υ
d2ξ2 σd2 d1ξ1 d2ξ2 σd2 q=0
Definition 1: For any two independent RVs U and V, the
where Ps is the source transmit power, Pr = ηαPs h12 /((1 −
CDF ofx the product of them is defined as P (UV ≤ x ) =
α)d1ξ1 ) is the relay transmit power, η is efficiency of the energy FV ( u )fU (u)du.
ηα
harvester, σd2 is the noise variance at D, ζ = 1−α , d1 and d2 Using this definition, we can determine the CDF of W as
∞ r
are the S-to-R and R-to-D distances, respectively.
The instantaneous capacity of the first, Cr , and second, Cd , FW (r ) = FX f (u)du, (12)
0 u Y
links can be given by
where fY (u) has the distribution as in (1) and FX ( ur ) can be
Ci = (1 − α)log2 (1 + γi ), i ∈ {r , d }. (4) obtained from (11) with the appropriate change of notations.
With this in mind, the ergodic outage probability, which With this in mind, (12) can be expressed as
is defined as the probability that the instantaneous capacity ∞
Υ2 (κ1 μ1 )n
falling below a certain threshold (Cth ), can be calculated as FW (υ) = Γ(μ1 + n)J1 − J2 , (13)
exp(κ1 μ1 ) Γ(μ1 + n)n!
n=0
Pout = Pr(min{Cr , Cd } < Cth ). (5)
where
∞ μ −1 √
2
III. P ERFORMANCE A NALYSIS J1 = z 2 exp(−φ2 z ) Iμ2 −1 2Δ2 z dz , (14)
0∞
In this section, we derive analytical expressions of the μ2 −1 √ φ υ
ergodic outage capacity in generalized κ-μ fading and its spe- J2 = z 2 exp(−φ
2 z )Iμ2 −1 2Δ2 z Γ μ1 + n, 1 dz .
0 az
cial cases. To begin with, we substitute (2) and (3) into (4) (15)
and then into (5), with some mathematical manipulations, to
ηαPs
obtain
Δ2 = μ2 κ2 (1 + κ2 ), a = (1−α)d m m 2 and Γ(·, ·)
1 d2 σd
1 ζPs W indicates the upper incomplete Gamma function
Pout (υ) = Pr min , <υ , (6)
ζZ d ξ1 d ξ2 σ 2 [12, eq. (8.350.2)].
1 2 d To the best of the authors’ knowledge, there is no ana-
Cth
where Z = h32 , W = XY, X = h12 , Y = h22 and υ = 2 1−α −1. lytical solution for the integral J1 . Hence, to solve this
Because the random variables (RVs) Z and W are indepen- integral, we use the infinite series representation of Iμ2 −1 (·)
dent, we can calculate the probability in (6) as [12, eq. (8.445)]. Thus, J1 can be rewritten as
ξ ξ ∞ μ +2m−1
∞
d1 1 d2 2 σd2 Δ2 2
Pout (υ) = 1 − F̄Z
1
F̄W υ , (7) J1 = z μ2 +m−1 exp(−φ2 z )dz
ζυ ζPs Γ(μ2 + m)m! 0
m=0
μ2 +2m−1
where F̄Z (·) is the complementary cumulative distribu- ∞ μ 2 κ (κ
2 2 + 1)
(β)
tion function (CCDF) of Z, which can be obtained by = μ +m , (16)
integrating (1) with appropriate notation changes. In order to m=0 φ2 2 m!
386 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
where (β) is obtained with the help of [12, eq. (3.351.3)] Thus, the resultant closed-form expression of the ergodic
followed with some basic algebraic manipulations. outage probability in Nakagami-m fading can be given as
Similarly, to solve the integral J2 , we first replace
{Nak} 2 b
Iμ2 −1 (·) and Γ(·, ·) with their series representations using [12, Pout =1− γ m3 , m3
Γ(m2 )Γ(m3 ) υ
eq. (8.445)] and [12, eq. (8.352.2)], respectively, as follows m 1 −1 m2 +k
∞
√ μ2 −1+2l 1 m1 m2 υ 2 υ
1 × Km2 −k 2 m1 m2 . (24)
Iμ2 −1 (2Δ) = Δ2 z , (17) k! a a
Γ(μ2 + l )l ! k =0
l=0
ψ k
φ υ φ υ 1 φ1 υ When m1 = m2 = m3 = 1, the expression in (24) reduces to
Γ μ1 + n, 1 = ψ!exp − 1 , (18) the Rayleigh fading scenario as
az az k! az
k =0
where ψ = μ1 + n − 1. {Ray} υ b υ
Pout =1−2 1 − exp − K1 2 . (25)
Using (17) and (18), we can rewrite J2 as a υ a
∞ ψ μ +2l−1 We know that K1 (z ) can be approximated by 1/z when
φ1 υ k Δ2 2
J2 = ψ J3 , (19) z 1. Based on this, at high SNR, (25) can be simplified
a Γ(μ2 + l ) l ! k ! {Ray}
l=0 k =0 to Pout ≈ exp(− 1−αηαυ ), which indicates that the system
where
performance improves as we increase η and/or decrease υ.
∞ φ1 υ
J3 = z μ2 +l−k −1 exp −φ2 z − dz
0 az IV. N UMERICAL R ESULTS AND D ISCUSSIONS
1 All our evaluations in this section, unless we specify other-
(β) φ2 a 2 (k −l−μ2 ) υ
= 2 Kk −l−μ2 φ1 φ2 , (20) wise, are based on: ξ1 = ξ2 = ξ3 = 2.7, d1 = d2 = 4 m,
φ1 υ a σd = 0.01 W, η = 1 and Cth = 0.2 bits/s/Hz [13]. To
where Kp [ · ] is the modified Bessel function of the second begin with, in Fig. 1 we show a 3D plot for the analytical
kind with arbitrary order p [11, eq. (9.6.22)]. Note that (β) and simulated ergodic outage probability as a function of the
is accomplished by means of [12, eq. (3.471.12)], along with fading parameters κ and μ. Note that these analytical results
some mathematical manipulations. are obtained using (22) while considering the first 20 terms
Substituting (20) into (19) and then (16) and (19) into (13), of all series. It is clear that the performance improves as κ
we obtain an expression for FW (·) given in (21), shown at the and/or μ is increased. This is because increasing κ indicates
bottom of this page. Finally, using (7), (11) and (21), along an increase in the ratio between the total powers of the dom-
with basic algebraic manipulations, we obtain an accurate and inant components and the scattered waves, and increasing μ
unified expression for the ergodic outage probability of the implies increasing the number of multipath clusters.
dual-hop FD-DF relaying system over the generalized κ-μ fad- Now, to illustrate the influence of the EH time factor we
ing channel. This is given by (22), shown at the bottom of this present in Fig. 2 the ergodic outage probability with respect
page. to α for the two special cases of the κ-μ fading model: Rice
Now, substituting μ1 = μ2 = μ3 = 1 in (22), we get an ana- fading in Fig. 2(a) with different values of κi and Nakagami-m
lytical expression of the outage probability for the Rice fading fading in Fig. 2(b) for various values of μi where i ∈ {1, 2, 3}.
scenario, given in (23), shown at the bottom of this page. To Note that the numerical results in Figs. 2(a) and 2(b) are
obtain a mathematical expression for the Nakagami-m fading obtained from (23) and (24), respectively. It is clear that, for
case, we start from (22) and substitute κ1 = κ2 = κ3 → 0, all fading scenarios, when α is either too high or too small,
μ1 = m1 , μ2 = m2 and μ3 = m3 . Note that due to the fact the performance degrades significantly; hence, this parameter
that κ1 = κ2 = κ3 → 0, only the first terms of all infinite must be selected carefully to minimize the outage probability.
series will have non-zero values, except the last summation. It is worthwhile pointing out that the results represented by
μ2 +2m−1
μ2+1
μ2 (1 + κ2 ) 2
∞
∞
μ2 κ2 (1 + κ2 ) (κ1 μ1 )n ∞
ψ
∞
FW (υ) = −2
((1 + κ2 )μ2 )μ2 +m m! n!
μ2 −1
κ2 2 exp(κi μi ) n=0 m=0 n=0 l=0 k =0
i∈{1,2}
μ2 +2l−1
n μ κ (1 + κ ) k − 1 (μ2 +l−k )
(κ1 μ1 ) (μ1 + n − 1)! 2 2 2 φ1 υ aφ2 2 υ
Kk −l−μ2 2 φ1 φ2 (21)
n! k ! l ! Γ(μ1 + n)Γ(μ2 + l ) a υφ1 a
∞ μ2+1 ∞ ∞
Δμ2 +2m−1
1 (μ3 κ3 )q φ3 b μ2 (1 + κ2 ) 2 2
Pout (υ) = 1 − γ μ3 + q, 1 − μ −1 μ +m
exp(κ3 μ3 ) Γ(μ3 + q)q! υ 2 φ2 2
q=0 κ2 2 exp(κi μi ) n=0 m=0
i∈{1,2}
n ∞
ψ
∞ n μ +2l−1 k
− 1 (μ2 +l−k )
(κ μ ) (κ1 μ1 ) Δ2 2
φ1 υ ψ!
φ2 a 2 υ
× 1 1 −2 Kk −l−μ2 2 φ1 φ2 (22)
m! n! Γ(μ1 + n)Γ(μ2 + l )n! k ! l ! a φ1 υ a
n=0 l=0 k =0
∞ ∞ ∞
{Ric} 1 κq3 b 1 κn κm
1 2
Pout (υ) = 1 − γ q + 1, (1 + κ3 ) 1−
exp(κ3 ) Γ(q + 1)q! υ exp(κ1 + κ2 ) n! m!
q=0 n=0 m=0
l
υ 2 (l+k +1)
∞ ∞ n 1
κn1 κ2 Cth
−2 (κ1 + 1)(κ1 + 1) Kk −l−1 2 (1 + κ1 )(1 + κ2 ) (23)
n! l ! k ! l ! a a
n=0 l=0 k =0
RABIE et al.: FD EH ENABLED RELAY NETWORKS IN GENERALIZED FADING CHANNELS 387
Abstract—Millimeter-wave (mmWave) offers high data rate success probability of establishing a backhaul link as well as
and bandwidth for air-to-everything (A2X) communications backhaul data rate. In [9], the multiple-input multiple-output
including air-to-air, air-to-ground, and air-to-tower. MmWave (MIMO) non-orthogonal multiple access (NOMA) techniques
communication in the A2X network is sensitive to buildings were used in UAV network and the outage probability and
blockage effects. In this letter, we propose an analytical frame-
work to define and characterise the connectivity for an aerial
ergodic rate of network were studied based on a stochastic
access point (AAP) by jointly using stochastic geometry and geometry model. In [4], [7], and [8], the blockage effects were
random shape theory. The buildings are modelled as a Boolean characterized by a statistical model where the link-level line-
line-segment process with fixed height. The blocking area for an of-sight (LOS) probability is approximated as a simple sigmoid
arbitrary building is derived and minimized by optimizing the function. The parameters of sigmoid function are determined
altitude of AAP. A lower bound on the connectivity probability by the buildings’ density, sizes, and heights’ distribution. The
is derived as a function of the altitude of AAP and different model is unsuitable for mmWave A2X networks since it fails to
parameters of users and buildings including their densities, capture the fact that multiple nearby links could be simultane-
sizes, and heights. This letter yields guidelines on practical
mmWave A2X networks deployment. ously blocked by the same building and does not consider the
diversity in user types (e.g., their different heights). In [10], a
Index Terms—A2X communications, mmWave networks, mathematical framework was proposed for studying mmWave
blockage effects, network connectivity, stochastic geometry, ran- A2A networks, in which multiple aerial-users are equipped
dom shape theory.
with antenna arrays. Blocking effects were not included since
A2A scenario was assumed to be well above the blockages.
I. I NTRODUCTION In this letter, we develop an analytical framework for char-
acterizing the blockage effects and connectivity of a mmWave
IR-TO-EVERYTHING (A2X) communications can
A leverage aerial access points (AAPs) mounted on
unmanned aerial vehicles (UAVs) to provide seamless wire-
A2X network covered by a single AAP. The 3D buildings are
modelled as a Boolean line-segment process with fixed height.
Given an arbitrary building, the corresponding blocking area
less connectivity to various types of users [1] (see Fig. 1).
is derived as a function of altitude of AAP, users and build-
Millimeter-wave (MmWave) communication is one way to
ings’ parameters including their density, sizes, and heights.
provide high data rate for aerial platforms [2]. Unfortunately,
Based on the model, the AAP coverage area is maximized (or
mmWave communication is sensitive to building block-
equivalently the blocking area is minimized) by optimizing the
ages [3], which are widely expected in urban deployments of
altitude of AAP. Furthermore, both upper and lower bounds on
AAPs. In this letter, we define and characterize the connec-
the blocking area and a suboptimal result of AAP’s altitude
tivity for an AAP, using tools from stochastic geometry and
are derived in closed-form. The spatial average connectivity
random shape theory.
probability of a typical A2X network is obtained, which is
Leveraging UAVs as AAPs has been studied in [4]–[10]. A
maximized by optimizing the AAP’s altitude.
single-UAV network was proposed in [4], where the network
coverage was maximized by optimizing the UAVs’ altitudes.
The coverage performance can also be maximized via optimiz- II. S YSTEM M ODEL AND P ERFORMANCE M ETRIC
ing the placement of UAVs [5]. In [6], the coverage probability Consider a mmWave A2X network as illustrated in Fig. 1.
of a finite 3D multi-UAV network was calculated via a stochastic In this letter, we focus on the downlink communication from
geometric approach. Both network coverage and the sum-rate a typical low-altitude AAP to users with different heights.
of a hybrid A2G-D2D network were investigated in [7]. An
analytical framework that UAV uses ground-BS for wireless A. Channel Model Between AAP and Users
backhaul was proposed in [8] with providing the analysis for
The mmWave channel between the AAP and the different
Manuscript received August 1, 2018; revised September 5, 2018; accepted
types of users is assumed to be LOS or blocked by a build-
September 20, 2018. Date of publication October 1, 2018; date of current ver- ing. For simplicity, we assume the non-LOS (NLOS) signals
sion April 9, 2019. The work of K. Han and K. Huang was supported by Hong are completely blocked due to severe propagation loss from
Kong Research Grants Council under Grant 17209917 and Grant 17259416. penetration and limited reflection, diffraction, or scattering [3].
The work of R. W. Heath was supported by the National Science Foundation
under Grant ECCS-1711702 and Grant CNS-1731658. The associate editor
For the LOS case, the channel is assumed to have path-loss
coordinating the review of this paper and approving it for publication was without small-scale fading [11]. We assume perfect 3D beam
S. Zhou. (Corresponding author: Robert W. Heath, Jr.) alignment between AAP and users for maximal directivity
K. Han and K. Huang are with the Department of EEE, University of gain [10]. For the path-loss model, we assume the reference
Hong Kong, Hong Kong (e-mail: kfhan@eee.hku.hk; haungkb@eee.hku.hk).
R. W. Heath is with the University of Texas at Austin, Austin, TX 78712
distance is 1 m. The AAP transmission with power P and
USA (e-mail: rheath@utexas.edu). propagation distance r is attenuated modelled as r −α where
Digital Object Identifier 10.1109/LWC.2018.2873361 α is the path-loss exponent [11]. Let σ 2 be the thermal noise
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HAN et al.: CONNECTIVITY AND BLOCKAGE EFFECTS IN mmWAVE A2X NETWORKS 389
dS cos β
where θ = arccos( H −H ) and β = disk with radius diameter (i.e., a cylinder in 3D), the block-
(1− Hb −Hu ΛH )
d
a u aging area is recalculated as Sb = 2θ Λ2H − (dx + 18 2 (θ +
cos θ− dS π)) − 1(Ha > Hb )Sgain (x ), where Sgain (x ) is lower bounded
arctan( sin θ L ). Then, Sb (x ) is calculated as follows. (−)
Lemma 1 (Size of Sb (x)): The blocking area is by Sgain (x ) = 2θ [Λ2H − ( H1b −Hu (dx + 12 ))2 ]+ . The case
1− H
a −Hu
1 2 that building has a rectangular shape in 2D can be analyzed
Sb (x ) = θΛH − dS dL sin θ − 1(Ha > Hb )Sgain (x ), similarly (e.g., [14]).
2
(4)
B. Network Connectivity Probability
where 1(·) denotes the indicator function and Sgain (x ) is given
in (3). In this section, we calculate the connectivity probability
The calculations follow from geometry, the detailed proof defined in (1). Notice that the spatial correlation between
is omitted due to limited space. To simplify the result in different buildings exsits such as the blocking areas of
Lemma 1 and obtain more insights therein, both upper and multiple buildings may overlap with each other. For analyt-
lower bounds of Sgain (x ) are derived. By assuming the dis- ical tractablility, we ignore the spatial correlation of buildings
tance between any point on building’s line-segment and o has due to overlap in the blockaging area of multiple buildings.
the same value dL or dS given in (2), the lower and upper This assumption is accurate when density of buildings is not
(−) (+) very high, which has been validated in [3]. We derive a lower
bounds of Sgain (x ), denoted by Sgain (x ) and Sgain (x ), are
bound of pc by jointly using Campbell’s theorem, random
derived as
shape theory, with the results given in Lemmas 1 and 2.
⎡ 2 ⎤+ Theorem 1 (Connectivity Probability of AAP): The connec-
(−) θ dL (−)
Sgain (x ) = ⎣Λ2H − ⎦ , (5) tivity probability pc is lower bounded by pc as
2 1− H b −Hu
Ha −Hu 2
ΛH
(+) (−) (−) πλ θ
and Sgain (x ) is obtained by replacing dL in Sgain (x ) with pc = 1 − 2b F (r , , ω)r drfΘ (ω)dωfL ()d, (7)
ΛH
dS . The result for the bounds of Sgain (x ) is summarized as L Θ 0
follows. where
Lemma 2 (Bounds of Sgain (x)): The coverage gain +
Sgain (x ) can be upper or lower bounded as follows. θ 2 2 2
F(r , , ω) = Λ − 1(Ha > Hb ) ΛH − (dL + ΩH )
2 H
(−) (+)
Sgain (x ) ≤ Sgain (x ) ≤ Sgain (x ), (6) 1
− dL2 sin θ, (8)
2
where ΛH is specified in Lemma 1 and [A]+ = max [0, A].
The bounds for Sgain can be treated as the bounds for Sb via ΩH = (1 − H b −Hu −1 2
Ha −Hu ) , and ΛH is specified in Lemma 1.
(−) (+) Proof: See the Appendix.
substituting (6) into (4): Sb (x ) ≤ Sb (x ) ≤ Sb (x ), where (−)
(−) (+) Remark 3: The lower bound pc becomes tighter when
Sb (x ) = 12 (θΛ2H − dS dL sin θ) − 1(Ha > Hb )Sgain (x ) and density of buildings, i.e., λb , becomes smaller. This is because
(+) (+) (−)
Sb (x ) is obtained via replacing Sgain (x ) in Sb (x ) with sparsely deployed buildings result in less spatial correlation.
(−) Remark 4: Based on the discussion in Remark 1 and
Sgain (x ). expression of F(r , , ω), the connectivity probability pc can
Remark 1 (Optimal Altitude of AAP): A larger AAP’s also be maximizing by optimizing the APP’s altitude Hu .
altitude Ha can effectively increase the coverage (LOS)
area, while shrinking the radius of effective coverage disk.
So there exists an optimal Ha∗ to maximize the size of IV. S IMULATION R ESULTS
coverage gain Sgain , which can be calculated by solving In this section, we validate the analytical results via Monte
Ha∗ = arg maxHa Sgain (x ). To obtain a simple result with Carlo simulation. The radius of maximal coverage sphere
closed-form, we characterize this behavior by optimizing is Rmax = 100 m. The height of building is Hb = 30 m
(−) and that of user is Hu = 2 m. The density of buildings is
instead Sgain (x ) to obtain a suboptimal solution for Ha . When
dL 2 (−) λb = 2×10−4 m2 . The length and orientation ω of building’s
( H −H ) < Λ2H , we have H̃a∗ = arg maxHa Sgain (x ) = line-segments follow independently and uniformly distribu-
1− Hb −Hu
a u
1 tions. Specifically, is uniformly distributed in (0, 15 m] and
(dL2 (Hb −Hu )) 3 +Hb . Substituting H̃a∗ into (3) gives the sub- ω is uniformly distributed in (0, π].
optimal solution of Sb (x ). It will be shown in Fig. 3 that the Fig. 3 shows the coverage gain Sgain calculated via (3) and
derived suboptimal AAP altitude H̃a∗ is close to the optimal (+) (−)
one Ha∗ via numerical calculation. its bounds Sgain , Sgain calculated via Lemma 2 versus the
Remark 2: Extending the current building model to any altitude of AAP Ha . It is observed that Sgain is well bounded
(+) (−)
model that each building has a random size in 2D projec- by Sgain and Sgain . The lower bound becomes tighter when
tion, such as rectangle [3] or disk, follows a similar analytical Ha is small and the upper bound becomes tighter when Ha is
structure. The main difference is that the area of buildings large. This agrees with the intuition because larger or smaller
should be included into blocking area Sb . Also, Sgain needs altitude of AAP results in larger or smaller coverage gain,
to be recalculated based on different building’s shape. For respectively, which makes the bound tighter. Moreover, the
instance, if the 2D projection of a building is modelled as a AAP’s altitude that maximizes Sgain given by Ha∗ and the
HAN et al.: CONNECTIVITY AND BLOCKAGE EFFECTS IN mmWAVE A2X NETWORKS 391
A PPENDIX
P ROOF OF T HEOREM 1
By omitting the spatial correlations between {Sb (x )}, con-
nectivity probability pc defined in (1) is lowered bounded as
follows.
⎛ ⎞
x ∈{Φ∩O(ΛH )} Sb (x )
pc ≥ 1 − E ⎝ ⎠
|O(ΛH )|
2
Fig. 3. The effect of the AAP’s altitude on coverage gain Sgain . The ΛH
coverage gain is shown to be a concave function of AAP’s altitude. The (a) πλb θ
parameters of building are set as {dx , , ω} = {25 m, 6 m, π/4}. The upper = 1− Sb (r )r drfΘ (ω)dωfL ()d. (9)
and lower bounds are plotted based on (6). It is observed that Sgain is well Λ2H
bounded and the lower bound becomes tighter when AAP’s altitude is small L Θ 0
and upper bound becomes tighter when AAP’s altitude is large. Moreover, both where (a) follows the Campbell’s theorem and ramdom shape
optima and suboptimal altitudes of APP, Ha∗ and H̃a∗ given in Remark 1, are
highlighted. theory [12]. Based on (4), the blocking area Sb (r ) can be
(−)
upper bounded by 12 [θΛ2H − dS2 sin θ] − 1(Ha > Hb )Sgain (r ).
(−)
Substituting the result above into (9) gives pc .
R EFERENCES
[1] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications
with unmanned aerial vehicles: Opportunities and challenges,” IEEE
Commun. Mag., vol. 54, no. 5, pp. 36–42, May 2016.
[2] Z. Xiao, P. Xia, and X.-G. Xia, “Enabling UAV cellular with millimeter-
wave communication: Potentials and approaches,” IEEE Commun. Mag.,
vol. 54, no. 5, pp. 66–73, May 2016.
[3] J. G. Andrews et al., “Modeling and analyzing millimeter wave cellular
systems,” IEEE Trans. Commun., vol. 65, no. 1, pp. 403–430, Jan. 2017.
[4] A. Al-Hourani, S. Kandeepan, and S. Lardner, “Optimal LAP altitude
for maximum coverage,” IEEE Wireless Commun. Lett., vol. 3, no. 6,
pp. 569–572, Dec. 2014.
[5] J. Lyu, Y. Zeng, R. Zhang, and T. J. Lim, “Placement optimization of
Fig. 4. The effect of building density on connectivity probability pc . The UAV-mounted mobile base stations,” IEEE Commun. Lett., vol. 21, no. 3,
exact value of pc is plotted via Monte Carlo simulation. The lower bound pp. 604–607, Mar. 2017.
(−) [6] V. V. Chetlur and H. S. Dhillon, “Downlink coverage analysis for a
pc is plotted based on (7). It is observed that connectivity probability
decreases with building density and the lowered bound becomes tighter when finite 3-D wireless network of unmanned aerial vehicles,” IEEE Trans.
building density is small. Commun., vol. 65, no. 10, pp. 4543–4558, Oct. 2017.
[7] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Unmanned aerial
vehicle with underlaid device-to-device communications: Performance
and tradeoffs,” IEEE Trans. Wireless Commun., vol. 15, no. 6,
suboptimal H̃a∗ are close, which confirms the discussion given
pp. 3949–3963, Jun. 2016.
[8] B. Galkin, J. Kibilda, and L. DaSilva. (2017). A Stochastic Geometry
in Remark 1. Model of Backhaul and User Coverage in Urban UAV Networks.
In Fig. 4, we validate the lower bound of connectivity prob- [Online]. Available: https://arxiv.org/abs/1710.03701
(−) [9] T. Hou, Y. Liu, Z. Song, X. Sun, and Y. Chen. (2018). Multiple Antenna
ability pc , i.e., pc , given in Theorem 1 by comparing it with Aided NOMA in UAV Networks: A Stochastic Geometry Approach.
the exact value via Monte Carlo simulation. It is observed that [Online]. Available: https://arxiv.org/abs/1805.04985
(−)
both pc and pc decrease with building density λb . More [10] T. Cuvelier and R. W. Heath, Jr. (2018). MmWave MU-MIMO for Aerial
(−) Networks. [Online]. Available: https://arxiv.org/abs/1804.03295
importantly, pc becomes tighter when buildings are sparsely [11] K. Han, Y. Cui, Y. Wu, and K. Huang, “The connectivity of mil-
deployed, which aligns with the discussion in Remark 3. limeter wave networks in urban environments modeled using random
lattices,” IEEE Trans. Wireless Commun., vol. 17, no. 5, pp. 3357–3372,
May 2018.
[12] A. K. Gupta, J. G. Andrews, and R. W. Heath, Jr., “Macrodiversity in cel-
V. C ONCLUSION AND F UTURE W ORK lular networks with random blockages,” IEEE Trans. Wireless Commun.,
In this letter, we propose an analytical framework to define vol. 17, no. 2, pp. 996–1010, Feb. 2018.
[13] X. Li, T. Bai, and R. W. Heath, Jr., “Impact of 3D base station antenna
and characterize the connectivity in a mmWave A2X network. in random heterogeneous cellular networks,” in Proc. IEEE WCNC,
Based on the blockage model that buildings are modelled by Apr. 2014, pp. 2254–2259.
the Boolean line-segment process with fixed height, we cal- [14] T. Bai, R. Vaze, and R. W. Heath, “Using random shape theory to model
culate the blocking area due to an arbitrary building and the blockage in random cellular networks,” in Proc. IEEE SPCOM, 2012,
pp. 1–5.
connectivity probability of an AAP. Moreover, the AAP’s alti- [15] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to
tude can be optimized to maximize the coverage area as well as coverage and rate in cellular networks,” IEEE Trans. Commun., vol. 59,
the network connectivity. Future work will focus on studying no. 11, pp. 3122–3134, Nov. 2011.
392 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—For coherent detection of ON-OFF keying (OOK) and its generalizations are advocated. However, implementing
symbols in free-space optical (FSO) communications, the receiver these works involves computation of cumbersome and time-
requires the instantaneous channel fading coefficients. To accu- consuming integrals which compromises real-time exploita-
rately detect OOK symbols over FSO channels without requiring
transmission of pilot symbols, in the first part of this letter,
tion. As an alternative to MLSD detection methods, gener-
we propose an expectation-maximization (EM)-based sequence alized likelihood ratio test (GLRT) sequence detection (SD)
detection (SD) method with low implementation complexity which is proposed in [8]. The method in [8] uses GLRT as a met-
is shown to be particularly suitable for fast FSO communication ric inside the Viterbi algorithm to detect the sequence of OOK
systems. In the second part of this letter, to remove the error floor symbols. Recently, a GLRT-based SD was proposed in [15] for
and to improve the accuracy of the proposed detector, by utilizing the more general scenario of non-return-to-zero (NRZ)-OOK
a source information transformation we propose an EM-based
SD for this state which significantly outperforms than classical
symbols where the variances of transmitted one-bit and zero-
EM-based method. bit are not necessarily equal. The GLRT-based SD method
proposed in [15] achieves performance close to the MLSD
Index Terms—Free-space optics (FSO), data detection,
receiver for a sufficiently large length of the observation
expectation-maximization (EM) algorithm, ON-OFF keying
(OOK) modulation, source transformation. window, with a considerably less computational complexity
compared to MLSD.
Notice that due to the very high data rate of FSO communi-
I. I NTRODUCTION cation, the speed of opto-electronic devices is the main limiting
REE space optical (FSO) communications have recently factor to implement an FSO link. In this letter, to detect
F attracted a great part of research activity with numerous
advantages over conventional radio-frequency (RF) transmis-
OOK symbols over FSO channel, we use the expectation-
maximization (EM) algorithm. The use of EM-based detection
sion such as very high optical bandwidth, low implementation in the context of optical wireless systems was first proposed
cost and high security [1]. However, despite its advantages, the in [6] and then analyzed more in depth in [16]. Here, without
main source of impairment in FSO communications is fading resorting to the transmission of pilot symbols, we first pro-
due to the atmosphere turbulence [2]. pose an EM-based SD for the more general case of NRZ-OOK
Intensity modulation with direct detection (IM/DD) by using symbols where the variances of transmitted one-bit and zero-
pulse-position modulation (PPM) or on-off keying (OOK) has bit are not necessarily equal. We will show that the proposed
attracted a great deal of attention from both academia and EM-based SD can achieve performance very close to those
industry in the most current commercial FSO systems [2]. achieved with perfect channel state information (CSI) for a
The main advantage of using PPM is that there is no need to sufficiently large length of the observation window, while
perform threshold adjustment at the receiver for signal demod- the processing load is even faster than the GLRT-based SD
ulation. On the other hand, OOK modulation offers a better recently introduced in [15]. The main challenge of using EM-
bandwidth efficiency but it requires adaptive threshold setting and GLRT-based methods is the error floor due to the transmis-
under channel fading conditions [3]. Although data detection sion of all-zero sequences [16]. To overcome this limitation,
issues have been widely addressed in the context of classical in the second part of this letter, by utilizing a specified source
RF wireless communication [4], these results are not directly information transformation (SIT), we propose an EM-based
applicable to OOK-based optical systems. Recently, different SD for this state that by removing all-zero sequence, achieve
techniques for FSO receiver design have been proposed and a significantly better performance compared to the similar
analyzed in the context of optical wireless communication methods.
systems [5]–[17].
As transmission of pilot symbols reduces the bandwidth
II. S YSTEM M ODEL
efficiency of FSO links [6], in most of the aforementioned
references, maximum-likelihood sequence detection (MLSD) Similar to the system model of [15], we consider an IM/DD
FSO link with NRZ-OOK modulation over atmospheric tur-
Manuscript received September 6, 2018; accepted September 24, 2018. bulence channel. The received signal denoted rk , can be
Date of publication October 1, 2018; date of current version April 9, 2019.
The associate editor coordinating the review of this paper and approv-
expressed at any discrete time k as rk = Pt sk h + nk , where
ing it for publication was P. P. Markopoulos. (Corresponding author: h is the fading channel coefficient that is assumed constant
Seyed Mohammad Sajad Sadough.) over a large number of transmitted bits, sk ∈ {αe , 1} is the
The authors are with the Department of Electrical Engineering, transmitted NRZ-OOK symbol, sk = αe represents the digital
Shahid Beheshti University G. C., Tehran 1983969411, Iran (e-mail:
s_sadough@sbu.ac.ir). symbol sk = 0 while sk = 1 represents the digital symbol
Digital Object Identifier 10.1109/LWC.2018.2873382 sk = 1 [1]. Without loss of generality, we assume that the
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
DABIRI AND SADOUGH: RECEIVER DESIGN FOR OOK MODULATION OVER TURBULENCE CHANNELS USING SOURCE TRANSFORMATION 393
R EFERENCES
[1] Z. Ghassemlooy, W. Popoola, and S. Rajbhandari, Optical Wireless
Communications. Boca Raton, FL, USA: CRC Press, 2012.
[2] M. A. Khalighi and M. Uysal, “Survey on free space optical communi-
Fig. 3. BER of proposed EM-based SD with SIT versus SNR for L = 4, cation: A communication theory perspective,” IEEE Commun. Surveys
compared to the BER achieved with simple proposed EM-based SD and Tuts., vol. 16, no. 4, pp. 2231–2258, 4th Quart., 2014.
GLRT-based method of [15]. [3] M. T. Dabiri, S. M. S. Sadough, and M. A. Khalighi, “FSO channel
estimation for OOK modulation with APD receiver over atmospheric
turbulence and pointing errors,” Opt. Commun., vol. 402, pp. 577–584,
Nov. 2017.
and Pe,tar = 10−12 , the computational complexity of EM- [4] Y. Liu, Z. Tan, H. Hu, L. J. Cimini, and G. Y. Li, “Channel estimation for
based method is proportional to 1800 and 3200, respectively, OFDM,” IEEE Commun. Surveys Tuts., vol. 16, no. 4, pp. 1891–1908,
and the computational complexity of GLRT-based method is 4th Quart., 2014.
[5] X. Zhu and J. M. Kahn, “Free-space optical communication through
proportional to 3.2 × 1010 and 4.4 × 1013 , respectively. atmospheric turbulence channels,” IEEE Trans. Commun., vol. 50, no. 8,
According to the methodology of SIT algorithm, 2L pos- pp. 1293–1300, Aug. 2002.
sible information sequences of length L is mapped to 2L
[6] N. D. Chatzidiamantis, M. Uysal, T. A. Tsiftsis, and G. K. Karagiannidis,
different sequences of length L = L +1 for which each
“Iterative near maximum-likelihood sequence detection for MIMO opti-
cal wireless systems,” J. Lightw. Technol., vol. 28, no. 7, pp. 1064–1070,
sequence of length L, contains at least one bit “1” [12]. For Apr. 1, 2010.
such method, the transceiver requires an one-to-one mapping [7] H. Moradi, H. H. Refai, and P. G. LoPresti, “Thresholding-based optimal
with 2L input and 2L output. Hence, the complexity of detection of wireless optical signals,” J. Opt. Commun. Netw., vol. 2,
no. 9, pp. 689–700, Sep. 2010.
SIT method is proportional to OS ∝ 2L = 2L−1 . Now, [8] T. Song and P.-Y. Kam, “A robust GLRT receiver with implicit channel
considering the joint effect of SIT and EM, the computa- estimation and automatic threshold adjustment for the free space optical
channel with IM/DD,” J. Lightw. Technol., vol. 32, no. 3, pp. 369–383,
tional complexity of EM-SIT algorithm is proportional to Feb. 1, 2014.
OES = OE + 2OS ∝ 2 × L2 + 2L . More precisely, the [9] T. Song and P.-Y. Kam, “Robust data detection for the photon-counting
overall computational complexity of EM-SIT algorithm is pro- free-space optical system with implicit CSI acquisition and back-
portional to OES = OE + 2OS ∝ 2 × L2 + 2L where the ground radiation compensation,” J. Lightw. Technol., vol. 34, no. 4,
pp. 1120–1132, Feb. 15, 2016.
complexity of SIT is multiplied by two as SIT is performed at [10] M. M. Abadi, Z. Ghassemlooy, M.-A. Khalighi, S. Zvanovec, and
both transmitter and receiver sides. On the other hand, accord- M. R. Bhatnagar, “FSO detection using differential signaling in outdoor
ing to the results depicted in Fig. 3, EM-SIT algorithm has a correlated-channels condition,” IEEE Photon. Technol. Lett., vol. 28,
no. 1, pp. 55–58, Jan. 1, 2016.
performance comparable to the receiver with perfect CSI for [11] M. M. Abadi et al., “Impact of link parameters and channel correlation
L = 4 and consequently its computational complexity is pro- on the performance of FSO systems with the differential signaling tech-
portional to 2 × 42 + 24 = 48 which is much lower than nique,” IEEE/OSA J. Opt. Commun. Netw., vol. 9, no. 2, pp. 138–148,
the complexity of EM- and GLRT-based methods at practical Feb. 2017.
[12] L. Yang, X. Song, J. Cheng, and J. F. Holzman, “Free-space optical
target BERs. communications over lognormal fading channels using OOK with finite
extinction ratios,” IEEE Access, vol. 4, pp. 574–584, 2016.
[13] L. Yang, J. Cheng, and J. F. Holzman, “Maximum likelihood estimation
V. C ONCLUSION of the lognormal-rician FSO channel model,” IEEE Photon. Technol.
Lett., vol. 27, no. 15, pp. 1656–1659, Aug. 1, 2015.
In the first part of this letter, we proposed an EM-based [14] L. Yang, B. Zhu, J. Cheng, and J. F. Holzman, “Free-space optical
detector over FSO channels for the general case of NRZ OOK communications using on–off keying and source information trans-
formation,” J. Lightw. Technol., vol. 34, no. 11, pp. 2601–2609,
symbols where the variances of transmitted one-bit and zero- Jun. 1, 2016.
bit are not necessarily equal. The proposed detector has a [15] M. T. Dabiri, S. M. S. Sadough, and H. Safi, “GLRT-based sequence
high spectral efficiency since it does not require transmis- detection of OOK modulation over FSO turbulence channels,” IEEE
sion of any pilot symbols for channel estimation. We showed Photon. Technol. Lett., vol. 29, no. 17, pp. 1494–1497, Sep. 1, 2017.
[16] M. T. Dabiri and S. M. S. Sadough, “Performance analysis of EM-based
that the proposed EM-based detector is particularly suitable blind detection for ON–OFF keying modulation over atmospheric optical
for employing in fast FSO communications because it has channels,” Opt. Commun., vol. 413, pp. 299–303, Apr. 2018.
BER performance close to those achieved with perfect CSI. [17] M. T. Dabiri and S. M. S. Sadough, “Generalized blind detection of OOK
modulation for free-space optical communication,” IEEE Commun. Lett.,
Moreover, it was shown that EM-based SD outperforms GLRT vol. 21, no. 10, pp. 2170–2173, Oct. 2017.
as it provides higher quality of service in terms of target BER. [18] L. C. Andrews and R. L. Phillips, Laser Beam Propagation Through
Although compared to GLRT the proposed EM-based method Random Media, vol. 1. Bellingham, WA, USA: SPIE Press, 2005.
requires larger values of L to achieve performance close to the [19] M. T. Dabiri, S. M. S. Sadough, and M. A. Khalighi, “Channel
modeling and parameter optimization for hovering UAV-based free-
perfect CSI case, its computational complexity is lower than space optical links,” IEEE J. Sel. Areas Commun., to be published,
that of GLRT-based method. The main challenge of using EM- doi: 10.1109/JSAC.2018.2864416.
396 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
I. I NTRODUCTION
ENSIFYING access points (APs) is deemed as a
D promising solution to improve system capacity in 5G [1].
Extensive studies have been conducted to quantize the han-
Fig. 1. A UE connect with the nearest 3 APs. Handover in CP occurs when
AP C is leaving the AS.
update the APs in AS (defined as a UP handover) according IV. H ANDOVER P ROBABILITY IN U SER -P LANE
to the filtered RSRP, i.e., DL-RSS, to avoid the ping-pong The UP handover event is represented by Hu , and Fj
effect. Hence, the AS is always formed by the nearest M denotes the event that AP j (1 ≤ j ≤ M) becomes the fur-
APs. Especially, before the serving AP C leaving the AS, it thest serving AP in the AS after the movement of the UE.
will choose a target AP C among all the serving APs and According to the system model, the UP HOP is given by
send handover request (defined as a CP handover). During the
handover, other APs in the AS keep serving the UE. Fig. 1 P(Hu )
illustrates the handover when the AS size is 3, where AP 2 M
is leaving the AS after the movement of UE. Especially, if = P Hu |Fj , rM +1 , rj , θ P Fj |rM +1 , rj , θ
AP 2 takes charge of the control-plane, it will choose the best j =1 (Ω)
serving AP as the target AP C, leading to CP handover process.
·fj ,M +1 rj , rM +1 |θ fθ (θ)d rM +1 d rj d θ
III. A NALYTICAL M ODEL AND P RELIMINARY A NALYSIS M
(a) 1
In this letter, AP j represents the j-th nearest APs to the UE. = P Hu |Fj , rM +1 , rj , θ P Fj |rM +1 , rj , θ
π
The UE moves a distance v in a unit of time at angle θ with j =1 (Ω)
respect to the direction of the connection with the AP j from
the initial location l1 , to a new location l2 . Note that, AP M is ·fj ,M +1 rj , rM +1 d rM +1 d rj d θ, (3)
not always the furthest serving AP after the movement of the where (a) uses the assumption that θ is independent of
UE. For the sake of unity in analyses, AP (M + 1) is taken rj , and its distribution is uniform whose PDF is fθ (θ) =
into account. rj (R) is the distance between l1 (l2 ) and AP j. 1/π due to the symmetry. The integration domain Ω and
C (Cj ) denotes the circle with its center at l1 and its radius fj ,M +1 (rj , rM +1 ) are given by Proposition 1 and 2, respec-
rM +1 (rj ), and A denotes the circle with its center at l2 and tively.
its radius R. It is obvious that R 2 = rj2 + v 2 + 2v rj cos θ. Fj is satisfied if there is no APs left in the region
In our model, the positional relation between circles A and C\A ∩ C. C\A ∩ C have to be divided into two indepen-
C depends on the value of rj , rM +1 and θ. However, the dent regions: SI : Cj \Cj ∩ A with at most (j − 1) APs, and
handover occurs iif circle A intersects with circle C, which
SII : (C \A ∩ C )\SI with at most (M − j) APs. The loca-
limits the the integration domain used in following sections
from {(rj , rM +1 , θ)|0 < rj < rM +1 , 0 < θ < π} to Ω, given tions of the APs then follows two independent Binomial point
by the following proposition. processes (BPPs) on both the region SI and the region SII [10].
Proposition 1: Based on the system model in Section II, Thus, the conditional probability of Fj is given by
the necessary condition of the handover is that the circle A
P(Fj |rM +1 , rj , θ) = P(N (|SI |) = 0)P(N (|SII |) = 0)
intersects with the circle C. Let AP j be the furthest serving
M −j
AP in AS after the movement of the UE. Therefore, rj , rM +1 |SI | j −1 |SII |
= 1− 1−
and θ are limited by the domain Ω, given by |Cj | |C \Cj |
j −1 M −j
S∩ (rj , R, v ) S∩ (rM +1 , R, v ) − S∩ (rj , R, v )
(Ω) = (Ω1 ) + (Ω2 ) + (Ω3 ), (1) = , (4)
πrj2 πrM2
+1 − πrj
2
where
⎧ ⎫ where |S| denotes the measure of S, N(·) is the number of APs
⎨ 0 < rj < v , ⎬ in the specified area, and the function S∩ (r , R, v ) represents
(Ω1 ) = rj , rM +1 , θ rj < rM +1 < 2v − rj , ,
⎩ 0 < θ < π ⎭ the common area between the two intersecting circles with
⎧ ⎫ radii r and R, and the central distance v, is given by
⎪ 0 < rj < v , ⎪
⎪
⎨ ⎪
⎬ r 2 + v 2 − R2 R2 + v 2 − r 2
2v − rj < rM +1 < 2v + rj , 2
S∩ (r , R, v ) = r cos
−1 2
+ R cos
−1
(Ω2 ) = rj , rM +1 , θ 2
, 2vr 2vR
⎪
⎪
2
0 < θ < cos−1 rM +1 −2v rM +1 −rj ⎪
⎪
⎩ 2v rj
⎭ 1
− (r + R − v )(r + R + v )(v + r − R)(v − r + R). (5)
⎧ ⎫ 2
⎪ rj > v , ⎪
⎪
⎨ ⎪
⎬ Conditioned on Fj , the UP handover does not occur only if
rj < rM +1 < rj + 2v ,
(Ω3 ) = rj , rM +1 , θ 2
. two independent conditions are fulfilled: (a) no new APs is in
⎪
⎪
2
0 < θ < cos−1 rM +1 −2v rM +1 −rj ⎪
⎪ the region A\A ∩ C which is denoted by SIII , and (b) the arc
⎩ 2v rj
⎭
of circle C which is included in area A have to be exclusive
Proof: See Appendix A. of AP (M + 1), which is denoted by the notation Narc . The
The joint distribution of user-to-APs distances is also con- probability of condition (a) can be expressed by
sidered in the analysis. The following proposition gives the
P(N (|SIII |) = 0) = e −λB [πR −S∩ (rM +1 ,R,v )] .
2
joint probability density function (PDF) fj ,M (rj , rM ). (6)
Proposition 2: Consider the distance between a typical user The event Narc is the necessary condition for the event that
and its j-th nearest APs is rj in a network with APs distributed the UP handover does not occur. Since the positions of the APs
according to PPP. The joint PDF of rj and rM is given by follows a PPP, the AP (M + 1) follows a uniform distribution
on the circle C. Based on the relation of the central angle and
(πλB )M rM rj2j −1 M −j −1
fj ,M (rj , rM ) = 4 2
rM − rj2
2
e −λB πrM , (2) 2π, the probability of Narc is given by
Γ(j )Γ(M − j ) 2
1 −rM +1 + rj2 + 2v rj cos θ
where Γ(·) is the gamma function and rj ≤ rM . P(Narc ) = cos −1
. (7)
Proof: See Appendix B. π 2v rM +1
398 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
(8)
P(Hc )
M
= P Hc |Cj , rj , rM +1 , θ P Cj |rj , rM +1 , θ
j =1 (Ω)
·fj ,M +1 rj , rM +1 fθ (θ)drj drM +1 dθ
M
1 Fig. 3. HOPs versus the AP density for different AS size M when user
= P Hc |Cj , rj , rM +1 , θ fj ,M +1 rj , rM +1 drj drM +1 dθ,
Mπ velocity is 5 m/s: a) UP HOP; b) CP HOP.
j =1 (Ω)
(9)
n (M −1)!
CM −1 = n!(M −n−1)! is Binomial Coefficient. According to
where the integration domain Ω and fj ,M +1 (rj , rM +1 ) are
the property of PPP, we have
given by proposition 1 and 2, respectively, and we assume
P(Cj ) = 1/M for all j. For single connection, i.e., M = 1, the P(N (|SIII |) = n)
CP HOP is equal to the UP HOP. In this section, only the case n
πR 2 − S∩ (rM +1 , R, v ) λB 2
that M ≥ 2 is under our consideration. After the movement = e −(πR −S∩ (rM +1 ,R,v ))λB
. (12)
n!
of the UE, if the number of APs in the region A\A ∩ C (and
on its periphery) is more than that in the region C\A ∩ C, the Due to the complementarity between P(Hc |Cj , rj , rM +1 , θ)
CP handover occurs. Hence, we have and (10), the expression for conditional CP HOP can be
solved. So far, all the factors for CP HOP are derived.
P H̄c |Cj , rj , rM +1 , θ
= P(Narc )P(N (|SIII |) ≤ N (|SI |) + N (|SIII |)) VI. N UMERICAL R ESULTS
+P N̄arc P(N (|SIII |) ≤ N (|SI |) + N (|SIII |) − 1) The simulations in Figs. 2 and 3 validate the theoretical
M −1 n−1
analysis, where (A) and (S) in the legends denote analytical
and simulation results, respectively. Compared with [2], HOP
= P(N (|SI |) + N (|SII |) = n) P(N (|SIII |) = k )
n=1
in CP for MC shows a decrease of at least 40% over the model
k =0
M −1
in [2] when λM = 0.1λB , where λM is the macro cell density.
Especially, when λM = 0, the scenario in [2] is as same as
+P(Narc ) P(N (|SIII |) = n)P(N (|SI |) + N (|SII |) = n),
the single connectivity (i.e., M = 1) in this letter.
n=0
(10) Fig. 2 shows the HOPs versus the user velocity. Before
HOPs reach 0.5, which implies that the HOP is equal to
the number of handovers in a unit time approximately, they
and the APs in SI and SII follow independent BPP, thus
are almost proportional to the user velocity. The increases of
UP HOP and the decreases of CP HOP brought by larger
P(N (|SI |) + N (|SII |) = n)
min(n,j −1)
i j −i−1 AS size are diminishing, since the radius of AS rM follows
|SI | |SI | Gamma distribution whose difference between peak values of
= Cji−1 1−
i=max(0,n−M +j )
πrj2 πrj2 rM diminishing with M.
⎛ ⎞n−i ⎛ ⎞M −n−j +i Fig. 3 illustrates the HOP versus the AP density in logarith-
n−i ⎝ |SII | |SII | mic coordinates. The HOPs are proportional to the logarithm
·CM −j
⎠ ⎝1 − ⎠ , of AP density approximately before AP density achieves 0.01
2
π rM 2 2 2
+1 − rj π rM +1 − rj
cells/m2 . This is because the HOP is equal approximately to
(11) the number of handovers in a unit time in that case, and the
logarithm of the number of APs in the specified area is pro-
where the upper and lower boundaries of i are set to guar- portional to the AP density. The HOPs converge to 1 for larger
antee the number of APs in SI and SII are positive, and AP density, which is consistent with our expectations.
ZHANG et al.: HANDOVER PROBABILITY ANALYSIS OF ANCHOR-BASED MC IN 5G UCN 399
Abstract—Line-of-sight (LOS) blockage is a crucial problem Results indicate that the probability of V2V LOS blockage
in millimeter-wave vehicle-to-vehicle communication due to its is notably increased as the increasing traffic density improves
severe penetration loss and high vehicle mobility. In order to the blockage probability of mobile and static objects (vehicles,
overcome the LOS blockage problem, in this letter, we propose
using neighbor vehicles as relays to forward the blocked traf-
buildings, trees, etc).
fic flows. Specifically, a traffic-aware relay vehicle selection is To overcome the blockage effect, a self-organized V2V
investigated. First, analytic hierarchy process (AHP) is adopted association strategy is proposed in [5]. When blockage inter-
to handle the situation that different traffic types have different rupts the data transmission, either a new data packet queue
preferences on the performances of rate, delay and data dropping or a new V2V association is triggered, leaving the unfin-
ratio. Second, we introduce a coalitional game (CG) to evaluate ished data packets dropped. Through the proposed strategy,
the relative performances of the relay vehicles. Finally, by com-
bining the results of AHP and CG, a heuristic relay selection the performance of data rate and delay can be improved,
scheme is devised by selecting the relay vehicle with the best but the reliability can not be guaranteed, especially when the
rationality degree. Simulation results show that the proposed dropped data packets have high priority. To improve the relia-
scheme can adapt to the requirements of different traffic types bility, Wu et al. [6] propose using mmWave relays to forward
and improve the performances of rate, delay and data dropping the blocked data packets. Considering the vehicle mobility
ratio when LOS blockage occurs.
and mmWave propagation conditions, Deng et al. [7] pro-
Index Terms—LOS blockage, Millimeter-wave V2V communi- pose a low-complexity network control framework for fast
cation, Relay selection. relay selection and resource allocation. The relay is selected
by opportunistically discovering the relays which meets a
predefined signal-to-interference-noise (SINR) threshold.
In existing studies, however, the requirements of the
I. I NTRODUCTION specific traffic services in V2V communications are ignored.
EHICLE-TO-VEHICLE (V2V) communication is an For example, intelligent driving services are designed to
V important technology to support future automotive ser-
vices including intelligent driving, in-vehicle infotainment, etc.
help drivers avoid collisions and increase road safety [8].
Infotainment services provide audio and video entertainment
Many of these services require high data rate. For example, to vehicle users. That is to say, intelligent driving needs
self-driving vehicles can generate up to 1 TB data per driv- high-reliable communications while infotainment services are
ing hour [1] and vehicle laser radars produce high-resolution more sensitive to rate and delay. Therefore, when determining
maps requiring 10x Mbps data rate [2]. In order to satisfy the the relay selection, the preference of different traffic types
high data rate requirement, the adoption of millimeter-wave to different network performances is important. However,
(mmWave) in V2V communications has grasped considerable exiting studies in [6] and [7] only consider single performance
attention [3]. without the actual traffic type.
In mmWave V2V communications, one significant chal- In this letter, we devise a traffic-aware relay vehicle selection
lenge is the high possibility of inter-vehicular line-of-sight to handle the multi-traffic and multi-performance relay selection
(LOS) blockages, which are caused by the dynamic topology problem and meet the preference of different traffic types.
of V2V networks. Boban et al. [4] analyze the LOS block- First, analytic hierarchy process (AHP) is adopted to model
age for V2V communications via realistic traffic information. the preference of different traffic types to different network
performances. AHP provides an efficient method for quantifying
Manuscript received August 27, 2018; revised September 26, 2018;
accepted September 27, 2018. Date of publication October 3, 2018; date of the traffic preference through pre-defined pair-wise comparison
current version April 9, 2019. This work was supported by the National Key matrix and evaluating the performance weights according to
Research and Development Program of China under Grant 2017YFC0803903. specific traffic types [9]. However, the subjectiveness of the pair-
The associate editor coordinating the review of this paper and approving it
for publication was H. Q. Ngo. (Corresponding author: Yanyan Chen.)
wise comparison weakens the accuracy of the AHP method. To
B. Fan and Y. Chen are with the Beijing Key Laboratory of Traffic further improve the accuracy of the AHP method, a coalitional
Engineering, Beijing University of Technology, Beijing 100022, China game (CG) is introduced where relay vehicles are modeled as
(e-mail: fanbo@bjut.edu.cn; cdyan@bjut.edu.cn). players. CG provides a comparatively fair method to evaluate
H. Tian and X. Zhu are with the State Key Laboratory of
Networking and Switch Technology, Beijing University of Posts and the relative contribution of the performance capabilities for
Telecommunications, Beijing 100000, China (e-mail: tianhui@bupt.edu.cn; each players in the coalition [10]. Finally, we combine the
zhuxuzhen@bupt.edu.cn). results of AHP and CG to assess the rationality degree of the
S. Zhu is with the School of Mechanical Engineering, Beijing Institute of
Technology, Beijing 100081, China (e-mail: 13520870832@126.com). relay vehicles and choose the most rational one. Simulation
Digital Object Identifier 10.1109/LWC.2018.2873585 results show that the proposed solution can overcome the LOS
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
FAN et al.: TRAFFIC-AWARE RELAY VEHICLE SELECTION IN mmWAVE V2V COMMUNICATION 401
s
dj ,k = αn,j qns /γj ,k + djalignment
,k (6b)
III. P ROBLEM F ORMULATION AND S OLUTION
Since the antenna sector ID has already been learned via
Let J {1, 2, . . . , J } denote the set of candidate relay the relay vehicle discovery, only fine-granularity beam-level
vehicles with cardinality J. qns represents the size of the nth alignment is performed within the sector. Thus the delay
traffic flow with service type s. Each relay vehicle j is equipped of beam alignment can effectively reduced, which can be
s be a binary relay asso-
with a limited buffer size Qj . Let αn,j estimated by [5]
s
ciation variable such that αn,j = 1 means relay vehicle j is
alignment
s = 0. Due to
selected to transmit traffic flow n, otherwise αn,j di,j = ψitx ψjrx /ϕti,j
x
ϕri,jx Tp (7a)
the fact that different traffic types have different preferences
alignment
on different communication performances, we formulate the dj ,k = ψjtx ψkrx /ϕtjx,k ϕrj x,k Tp (7b)
402 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
alignment alignment
where di,j and dj ,k are the beam alignment delay modeled as players with different capabilities in terms of the
of link i,j and j ,k respectively. Within (7a), ψitx and ψjrx optimization objectives in (9). When a coalition of players
denote the sector-level beamwidths of vTx i and vRx j, and cooperates, a certain overall gain can be obtained from the
Tp denotes the pilot transmission duration. ψjtx and ψkrx in coalition. Shapley value calculates the contribution, or the rela-
(7b) denote the sector-level beamwidths of vTx j and vRx k. tive importance that each player makes to the overall gain [10].
Data dropping ratio: If the relay’s buffer size is smaller First, we model relay vehicles as players and define the
than the traffic size qns , the exceeding data will be dropped. coalitional set as S ⊆ J . To evaluate the capabilities to trans-
Hence the data dropping ratio is mit the nth traffic flow, we define the game by (S, Φu ), u ∈
{R, D, Γ}. Φu is the characteristic function of coalition S
Γn,j = 1 − min qns , αn,js
Qj /qns (8) which have three different formations
To this end, we have obtained the multiple performances ΦR (S) = Rn,j (11a)
j ∈S
of the relay transmission. Next we write the relay selection
ΦD (S) = maxj ∈S (Dn,j ) (11b)
problem as a MADM problem
ΦΓ (S) = minj ∈S (Γn,j ) (11c)
max R, min D, min Γ
s ΦR , ΦD , ΦΓ respectively represent the transmission rate,
s.t. C1: αn,j ∈ {0, 1}, ∀n, j
J delay and data dropping ratio when relay vehicles in S simul-
s taneously transmit the same traffic flow. The Shapley value
C2 : αn,j = 1, ∀n (9)
j =1 calculates the average marginal contribution (relative impor-
J N J
where R = N n=1 j =1 Rn,j , D = n=1 j =1 Dn,j , Γ =
tance) of a certain player in the coalitional game as follows
N J s (J − |S|)!(|S| − 1)!
n=1 j =1 Γn,j . The decision variable is αn,j . The objec- φu,j = (Φu (S) − Φu (S − {j }))
tives include maximizing the relay transmission rate, minimiz- J!
S(j ∈S)
ing the delay and data dropping ratio. Constraint C2 ensures
that one traffic flow can be allocated to at most one relay (12)
vehicle. where |S| denotes the cardinality of set S. The Shapley value
is expressed as a weighted sum of the player j’s marginal
A. Evaluating Objective Weights by AHP contribution Φu (S) − Φu (S − {j }), in the assumption that
We introduce AHP to evaluate the weights of different the player join the coalition by a completely random order.
(J −|S|)!(|S|−1)!
optimization objectives. First, the importance level of each J! is a weighting factor that assigns equal share
objective is compared with that of another according to the of the generated contribution to each coalition of interest.
importance numerical scales listed in [9]. Thus, the compar- To this end, we can get a characteristic matrix [φu,j ]3×J . In
ison matrix A = [au,v s ] s
3×3 is constructed as follow: au,v order to insure that all the elements are the-smaller-the-better,
denotes the relative importance of objective u comparing to we convert the first row of the matrix by φR,j = 1/φR,j .
s = 1/a s and a s = 1 (u, v ∈ {R, D, Γ}).
objective v, au,v v ,u u,u Further, to make the values of different characteristic types
The largest eigenvalue λmax of A and its corresponding comparable, we normalize the elements in matrix [φu,j ]3×J
normalized eigenvector W can be calculated as follows by φu,j = φu,j /maxj (|φu,j |). To determine the most suitable
relay vehicle, we introduce the definition of rationality degree
w̄us = 3 a s , u ∈ {R, D, Γ} (10a) (RD) by combining the results of AHP and CG as follows
v u,v
3
wus = w̄us / w̄vs , u ∈ {R, D, Γ} (10b) RDj =
u=1
wu φu,j (13)
v
s s
W = [wR , wD , w s ]T (10c) where for each traffic flow, the relay vehicle with the minimum
Γ
λmax = ( a s w s /3wus ) (10d) RD value is selected.
u v u,v v
Then we can use wus to represent the weight of the corre- IV. S IMULATION AND A NALYSIS
sponding objective u ∈ {R, D, Γ}. Since AHP is a subjective The simulation parameters are listed in Table I. We con-
evaluating method, the consistency of A should be checked to duct 1000 simulation periods and take the average value. Two
guarantee accuracy. First, the consistency index C.I. is calcu- benchmarks are used: transmission without relay and oppor-
lated as (λmax − 1)/(|A| − 1) (|A| denotes the order of the tunistic relay selection in [7]. Intelligent driving and infotain-
comparison matrix). The random consistency index R.I. is 0.58 ment traffics are generated with equal probability (50%) in
when |A| = 3 [9]. The consistency constraint is C.I./R.I. < 0.1. Poisson distribution. The traffic arrival rate is 0.8 Mits/ms in
Figs. 2(a), 3(a), 4(a) and the halfpower beamwidth is 15◦ in
B. Modeling the Relay Selection As CG Figs. 2(b), 3(b), 4(b).
Though AHP has helped us to evaluate the objective weights From Figs. 3–4, we can observe that comparing to the
according to different traffic types, the weakness is that the relay selection in [7], the delay and data dropping ratio of the
choose of the comparison matrix is subjective. Therefore, we proposed scheme are reduced by 0.43ms and 8.82% on aver-
introduce the Shapley value in coalitional game (CG) to reduce age for infotainment and intelligent driving respectively. While
the subjectiveness of AHP. In the game, relay vehicles are Fig. 2 (a), (b) indicate that the benchmark in [7] achieves the
FAN et al.: TRAFFIC-AWARE RELAY VEHICLE SELECTION IN mmWAVE V2V COMMUNICATION 403
TABLE I
S IMULATION PARAMETERS
R EFERENCES
[1] A. D. Angelica. (2013). Google’s Self-Driving Car Gathers Nearly
1 GB/Sec. [Online]. Available: http://www.kurzweilai.net/googles-self-
driving-car-gathers-nearly-1-gbsec
[2] C. Perfecto, J. Del Ser, M. Bennis, and M. N. Bilbao, “Beyond
WYSIWYG: Sharing contextual sensing data through mmWave V2V
communications,” in Proc. Eur. Conf. Netw. Commun. (EuCNC), Oulu,
Finland, 2017, pp. 1–6.
[3] J. Choi et al., “Millimeter-wave vehicular communication to support
massive automotive sensing,” IEEE Commun. Mag., vol. 54, no. 12,
pp. 160–167, Dec. 2016.
[4] M. Boban, X. Gong, and W. Xu, “Modeling the evolution of line-of-
Fig. 3. Delay versus (a) Source vTx/Destination vRx beamwidth, (b) Average sight blockage for V2V channels,” in Proc. IEEE Veh. Technol. Conf.
traffic arrival rate. (VTC), Montreal, QC, Canada, 2016, pp. 1–7.
[5] C. Perfecto, J. Del Ser, and M. Bennis, “Millimeter-wave V2V com-
munications: Distributed association and beam alignment,” IEEE J. Sel.
best rate performance. The reason is that the proposed scheme Areas Commun., vol. 35, no. 9, pp. 2148–2162, Sep. 2017.
[6] S. Wu, R. Atat, N. Mastronarde, and L. Liu, “Improving the cov-
puts emphasis on different performances according to different erage and spectral efficiency of millimeter-wave cellular networks
traffic types, whereas benchmark in [7] only considers SINR using device-to-device relays,” IEEE Trans. Commun., vol. 66, no. 5,
and ignores the traffic-to-performance preference. pp. 2251–2265, May 2018.
[7] J. Deng, O. Tirkkonen, R. Freij-Hollanti, T. Chen, and N. Nikaein,
An interesting observation drawn from Fig. 3(a) is that “Resource allocation and interference management for opportunis-
the proposed scheme tends to achieve the minimum delay at tic relaying in integrated mmWave/sub-6 GHz 5G networks,” IEEE
15◦ − 20◦ beamwidth, where the alignment delay decrease Commun. Mag., vol. 55, no. 6, pp. 94–101, Jun. 2017.
[8] M. C. G. Quintero and P. A. C. Cuervo, “Intelligent driving assistant
and transmission delay increase approach comparable values. based on accident risk maps analysis and intelligent driving diagnosis,”
Wider beamwidths lead to lower alignment complexity but in Proc. IEEE Intell. Veh. Symp. (IV), Los Angeles, CA, USA, 2017,
increase the transmission delay. This tradeoff between align- pp. 914–919.
[9] R. W. Saaty, “The analytic hierarchy process—What it is and how it is
ment delay and transmission delay directs our future work used,” Math. Model., vol. 9, no. 35, pp. 161–176, 1987.
towards fine-grained beamwidth optimization joint with relay [10] M. Pulido, J. S. Soriano, and N. Llorca, “Game theory techniques for
selection. Note that the delay is upper-bounded by a maximum university management: An extended bankruptcy model,” Ann. Oper.
Res., vol. 109, nos. 1–4, pp. 129–142, 2002.
threshold of 5 ms set in the simulation. [11] J. Wildman, P. H. J. Nardelli, M. Latva-Aho, and S. Weber, “On the joint
In Fig. 4(a), the beamwidth has minimal effect on data impact of beamwidth and orientation error on throughput in directional
dropping ratio because the LOS relay transmission can guar- wireless Poisson networks,” IEEE Trans. Wireless Commun., vol. 13,
antee successful data delivery within stringent delay threshold, no. 12, pp. 7072–7085, Dec. 2014.
[12] A. Yamamoto, K. Ogawa, T. Horimatsu, A. Kato, and M. Fujise, “Path-
whereas the first baseline renders severe data dropping ratio loss prediction models for intervehicle communication at 60 GHz,” IEEE
due to LOS blockages. However, the data dropping ratio of the Trans. Veh. Technol., vol. 57, no. 1, pp. 65–78, Jan. 2008.
404 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—This letter studies the fully decentralized access nodes (RANs). In the last decades, some decentral-
optimization of the precoding matrices for the downlink ized optimization methods were designed for the conventional
of multiple-input-multiple-output cache-enabled ultra-dense network MIMO system to maximize certain criterion, such
radio access network. To avoid signalling exchange between
coordinated edge radio access nodes (RANs), we first construct a as the weighted sum-rate maximization (WSRMax), and to
virtual downlink cache-enabled ultra-dense radio access network. reduce the amount of signalling exchange between coordinated
Following, a new concept of signal-to-interference-leakage-plus- RANs [3]–[5]. In [6], the high SINR approximation method
noise ratio is defined. Based on this concept, a virtual weighted was also used to decouple the WSR maximization problem.
sum rate maximization problem is formulated to optimize the Edge caching and signal processing at the networks is
precoders. We then propose a computationally efficient decen-
tralized optimization algorithm that achieves stationary point to regarded as a promising way to reduce the burden on the fron-
the problem. Numerical results show that the proposed algorithm thaul links and the delivery latency via caching files at the
can achieve more than 96.15% of the performance of the cen- local cache in close proximity to the mobile terminals [7]–[9].
tralized optimization without any signalling exchange between Park et al. [7] investigated the joint design of cloud and edge
coordinated edge RANs for multiple-input single-output cases. processing for cache-enabled radio access network. Cache-
Index Terms—Interference network, multiple-input multiple- enabled physical layer security video transmission was studied
output, edge caching, decentralized optimization. and was illustrated to achieves simultaneously a low secrecy
outage probability and a high power efficiency [8]. Hybrid
content caching design that does not require the knowledge of
I. I NTRODUCTION content popularity was investigated for hierarchical network
HE INCREASING number of mobile devices and surging architecture [9]. One common point of the aforementioned
T user demands for multi-media and delay-sensitive service
have triggered massive amounts of data in wireless networks,
methods is that the signalling exchange per-iteration among
coordinated RANs is needed [3]–[5], [7]–[9].
which brings about huge burden on fronthaul/backhaul links. In addition, in ultra-dense networks, the lack of direct
Recently, caching of popular content during off-peak traffic link (wireline or wireless link) among RANs prohibits
periods at various levels of the wireless network architecture the signalling exchange per-iteration between RANs [10].
and network densification have emerged as promising ways Furthermore, signalling exchange per-iteration between RANs
to satisfy low latency demands [1]. From the perspective of requires strict synchronization and causes a considerable delay
network architecture, cache-enabled multiple-input multiple- and signalling overhead that are determined by the conver-
output (MIMO) coordinated design is very similar with the gence speed. It also implies that the conventional network
traditional network MIMO and coordinated beamforming [2]. MIMO technologies without considering edge caching and sig-
Unfortunately, to obtain the performance gain provided by nal processing have no ability to satisfy the requirements of
network MIMO and coordinated beamforming, a large amount the low latency and low overhead of fronthaul/backhaul in the
of information, such as the channel state information (CSI) or wireless communication systems.
data streams of users, need to be exchanged between radio With aiming to reduce delivery latency, edge caching and
signal processing generally limits the operation of edge nodes
Manuscript received September 1, 2018; revised September 28, 2018; to non-cooperative transmission strategies. How to design fully
accepted September 29, 2018. Date of publication October 3, 2018; date decentralized transmission scheme has become a key problem
of current version April 9, 2019. This work was supported in part by the
National Natural Science Foundation of China under Grant 61471120 and for cache-enabled ultra-dense communication system. In this
Grant 61711540305, and in part by the National Science and Technology letter, we investigate the fully decentralized optimization of
Major Project of China under Grant 2018ZX03001002-003. The associate the precoding matrices for the downlink of cache-enabled
editor coordinating the review of this paper and approving it for publication
was X. Chu. (Corresponding author: Yongming Huang.) ultra-dense radio access network. To avoid signalling exchange
S. He, J. Ren, and Y. Zhang are with the School of Information Science among coordinated edge RANs causing additional delay, we
and Engineering, Central South University, Changsha 410083, China, and also construct firstly a virtual coordinated downlink of cache-
with the School of information Technology, Jiangxi University of Finance
and Economics, Nanchang 330032, China (e-mail: shiwen.he.hn@csu.edu.cn; enabled ultra-dense radio access network and introduce a
renju@csu.edu.cn; zyx@csu.edu.cn). new concept of signal-to-interference-leakage-plus-noise ratio,
Y. Chen, Y. Huang, and L. Yang are with the School of Information which is only related with local CSI. Using local CSI infor-
Science and Engineering, Southeast University, Nanjing 210096, China
(e-mail: 220160742@seu.edu.cn; huangym@seu.edu.cn; lxyang@seu.edu.cn). mation, the optimization of the precoding matrices is for-
Digital Object Identifier 10.1109/LWC.2018.2873671 mulated as a series of parallel WSRMax problems. Finally,
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HE et al.: DECENTRALIZED PRECODING FOR CACHE-ENABLED ULTRA-DENSE RADIO ACCESS NETWORKS 405
a fully decentralized optimization algorithm that is indepen- which needs the signalling exchange per-iteration among coor-
dently implemented at each edge RAN without any signalling dinated RANs [3]–[5]. The signalling exchange per-iteration
exchange is proposed to address the considered problem. is prohibitive in practical communication systems, especially
for delay sensitive traffic in ultra-dense radio access network.
II. S YSTEM M ODEL
Consider the downlink K-cell multiuser MIMO time divi- III. D ECENTRALIZED O PTIMIZATION OF P RECODERS
sion duplex (TDD) cache-enabled ultra-dense radio access In this section, we focus on exploring a fully decentral-
network where each cell includes one cache-enabled multiple ized optimization method to design the precoding matrices
antennas RAN at the edge of the networks and a plurality at each RAN independently by defining a virtual WSRMax
of users equipped with multiple antennas. Each RAN has (VWSRMax) problem. We assume that local CSI is available
cached the requested contents at its local cache and each for each user, namely, RAN k knows the local channel matri-
user associates with and served by only one RAN. Let k ces Hjl ,k , ∀j ∈ Kul , ∀l ∈ Kb , which can be obtained by
and l be the RAN index, i and j be the user index. RAN k, exploiting the downlink-uplink reciprocity in TDD system.
k ∈ Kb = {1, . . . , K }, is equipped with Mk transmit antennas To realize the fully decentralized optimization of the
and serves Ik users located in cell k. Let us define ik to be precoding matrices and avoid the signalling exchange among
the i-th user in cell k and Nik be the number of antennas of coordinated edge RANs, we construct a virtual downlink
user ik . The received signal of user ik is given by: cache-enabled ultra-dense radio access network by regard-
yik = Hik ,l Wjl xjl + zik . (1) ing RAN k as Ik virtual RANs, ∀k ∈ Kb . The cor-
responding channel coefficients are defined as, Gik ,ik =
l∈Kb j ∈Kul
[HH H
ik ,k , 0Mk ×(N ik −Nik ) ] , Gjk ,ik = Gjk ,jk , ∀j ∈ Kuk \ {i },
where Kuk is the set of the users located in cell k, ∀k ∈ Kb . and the channel coefficient from virtual RAN ik , ∀i ∈ Kuk ,
Hjl ,k ∈ CNjl ×Mk denotes the flat fading channel coeffi- ∀k ∈ Kb , to user jl served by virtual RAN l is defined as
cient between RAN k and user jl . Wik ∈ CMk ×dik and dik H
denote the precoding matrix and the number of data streams Gjl ,ik = HH jl ,k , 0M ×(N i −Nj ) , ∀j ∈ Kul , ∀l ∈ Kb \ {k },
k k l
for user ik , respectively. xik ∈ Cdik ×1 denotes the informa-
tion signal intended for user ik with distribution CN (0, Idi ). where 0M ×N is an M × N zero matrix and N ik =
k max(Nik , max max Njl ). The virtual received signal at
zik ∈ CNik ×1 denotes the additive white Gaussian noise with l∈Kb \{k } j ∈Kul
distribution CN (0, σi2k INi ) and Id is d × d identity matrix. user ik can be defined as:
k
In this letter, we assume that there are no direct link among
yik = Gik ,ik Wik xik + Gik ,ik Wjk xjk
edge RANs, i.e., the signalling exchange among coordinated
j ∈Kuk \{i}
edge RANs is impossible. Therefore, we need to explore the
fully decentralized design of precoders without any signalling + Gjl ,ik Wik xik + zik , (5)
exchange among coordinated edge RANs for cache-enabled l∈Kb \{k } j ∈Kul
ultra-dense communication systems. To design the precoding
matrices Wik ∈ CMk ×dik , ∀i ∈ Kuk , k ∈ Kb , the WSRMax where zik = [zH H
ik , 01×(N ik −Nik ) ] . Then, for RAN k, ∀k ∈
problem is formulated as Kb , a VWSRMax is defined as
Ik
max αik Rik , s.t. Tr Wik WiHk ≤ Pk , i , s.t. H
{Wk } max αik R k
Tr W i k
W i k ≤ Pk . (6)
k ∈Kb i∈Kuk i∈Kuk Wk
i∈Kuk i=1
(2)
Note that problem (6) is not equivalent to problem (2). The
where Wk denotes the set of all precoding matrix Wik , the i of user ik is defined as:
virtual achievable rate R k
weight αik is used to denote the priority of user ik , ∀i ∈ Kuk ,
i = ln det I H H −1
∀k ∈ Kb . The achievable rate Rik of user ik is given by: R k N i + Gik ,ik Wik Wik Gik ,ik Ωik . (7)
k
−1
Rik = ln det INi + Hik ,k Wik WiHk HH ik ,k Ωik . (3) i denotes the interference-plus-noise matrix includ-
In (7), Ω
k k
ing the intra-cell inter-user interference, inter-cell leakage
In (3), Ωik denotes the interference-plus-noise matrix includ-
interference, and the noise covariance, which is different from
ing intra-cell inter-user interference, inter-cell inter-user
the signal-to-leakage-noise ratio defined in [12], given by
interference, and the additive white Gaussian noise, given by
i =
Ω Gi ,i Wj WH GH
Ωik = Hik ,k Wjk WjHk HH
ik ,k
k k k k jk ik ,ik
j ∈Kuk \{i}
j ∈Kuk \{i}
+ Gjl ,ik Wik WiHk GH 2
+ Hik ,k Wjk WjHk HH + σi2k INi . (4) jl ,ik + σik Φik , (8)
ik ,k k l∈Kb \{k } j ∈Kul
l∈Kb \{k } j ∈Kuk
It is well known that problem (2) is non-convex and is difficult where matrix Φik ∈ CN ik ×N ik is the covariance of zik .
to obtain the global optimal solution. Furthermore, problem (2) When the inter-cell leakage interference is not considered,
is in general solved in a centralized or decentralized way i.e., removing the second item in the right side of (8) called
406 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
as Proposed Algorithm (wo), problem (6) is simplified to Algorithm 1 Decentralized Optimization Algorithm
maximize per-cell sum rate. Different from the research on 1: Initialize Wk such that the transmit power constraint is
the maximization of SLNR, the power allocation in this let- met. Compute R i with Wk , ∀i ∈ Ku and let flag = 1.
k k
ter needs to be optimized for each user, i.e., Tr(Wik WiHk ), 2: while flag == 1 do
−1
i ∈ Kuk , k ∈ Kb , under per-cell transmit power constraint. 3: Uik ← Ωik Gik ,ik Wik .
−1
Due to the coupling between optimization variables in
4: Vik ← I − UH G W
ik ik ,ik ik .
the objective of problem (6), it is non-convex and is dif-
ficult to obtain its global optimum solution. Furthermore, 5: Wik ← Ξ−1 H
ik Gik ,ik Uik Vik with
optimal λk .
unlike the optimization problem investigated in [5], we cannot i ≤ η, let flag = 0, other-
i and the minimum mean
directly use the relation between R 6: If ln det Vik − R
k i∈Ku k
square errors (MMSE) of xik to simplify the optimization k
wise let R i = ln det Vi and go to Step 3.
problem (6). Therefore, to obtain a tractable form, we firstly k k
7: end while
introduce a virtual optimization problem, i.e.,
min Tr Eik , s.t. Tr Wik WiHk ≤ Pk , (9)
Wk ,Uk
i∈Kuk i∈Kuk where λk ≥ 0 is the Lagrange multiplier associated with the
where virtual MMSE matrix Eik is defined as: transmit power constraint of RAN k. The first-order optimality
condition of L(Wk , λk ) with respect to each Wk yields
Eik = Idi − UH H H H
ik Gik ,ik Wik − Wik Gik ,ik Uik + Uik Ωik Uik ,
k Wi∗k = αik Ξ−1 H
ik Gik ,ik Uik Vik . (16)
where Uik ∈ CN ik ×dik is an auxiliary variable, ∀i ∈ Kuk , In (16), λk should be chosen such that the complementarity
and Uk is the set of all auxiliary variables. Ωik is given by slackness condition of the power constraint is satisfied, and
Ωi = Gi ,i Wi WH GH + Ω i . (10) Ξik is
k k k k ik ik ,ik k
In problem (9), fixing precoding matrices Wk , the optimal Ξik = λk IMk + αjk GH H
jk ,jk Ujk Vjk Ujk Gjk ,jk
solution of Uik is j ∈Kuk
−1
U∗ik = Ωik Gik ,ik Wik . (11) + αik GH H
jl ,ik Uik Vik Uik Gjl ,ik . (17)
Subsituting (11) into (10), we have l∈Kb \{k } j ∈Kul
−1
E∗ik = Idi − WiHk GH
ik ,ik Ωik Gik ,ik Wik . simplicity, let Wik (λHk ) be the right side of (16). When
(12) For
k
Tr(Wik (0)Wik (0) ) ≤ Pk and the matrix Ξik −λk IMk
Applying the Woodbury matrix identity to (7), we have R i = i∈Kuk
k
ln det((E∗ik )−1 ) and can obtain the following conclusion. is invertible, then Wi∗k = Wik (0), otherwise we must have
Proposition 1: Let Vik ∈ Cdik ×dik 0 be a weight matrix Tr(Wik (λk )Wik (λk )H ) = Pk . Let Qik Λik QH
ik be the
for user ik . Problem (6) is equivalent to the following problem: i∈Kuk
eigenvalue decomposition of Ξik − λk IMk , then we have
min αik Tr Vik Eik − ln det Vik , −2
Wk ,Uk ,Vk Tr Λik + λk IMk Ψik = Pk , (18)
i∈Kuk
i∈Kuk
s.t. Tr Wik WiHk ≤ Pk . (13)
i∈Kuk where Ψik = αi2k QH H H H
ik Gik ,ik Uik Vik Vik Uik Gik ,ik Qik . We fur-
ther have
where Vk denotes the set of Vik for cell k. Mk
In what follows, we focus on addressing problem (13) Ψik m,m
2 = Pk , (19)
instead of solving directly problem (6). The objective func-
i∈Kuk m=1 Λik m,m + λk
tion of problem (13) is not joint convex with respect to
optimization variables Wk , Uk , and Vk . but it is convex in where [A]m,n denotes the (m, n)-th of matrix A. Note that the
each of them. Therefore, we propose to use the block coor- left side of (19) is a decreasing function in λk . Hence, the
dinate descent method to solve problem (13). Specifically, we optimal λk can be found by using one dimension search.
address problem (13) sequentially by fixing two of the three The detailed steps used to optimize problem (13) is sum-
variables Wk , Uk , Vk , and updating the third. Fixing the other marized as Algorithm 1 for RAN k, ∀k ∈ Kb , where η
variables, the optimal solution of Vik is is a predefined stop threshold. The similar procedure can
Vi∗k = E−1
ik (14) be used to extending the algorithm in [12] to multi-cell
multiuser power allocation systems, referred as Extended
and the optimal solution of Uik is given by (11). Now, we
leakage Algorithm.
focus on optimizing problem (13) with fixed Uk and Vk via
Following the similar procedure [5, Th. 3], we have:
Lagrange duality theory. The Lagrange function is given by
Proposition 2: Any limit point Wk∗ , Uk∗ , and Vk∗ of the
L(Wk , λk ) = αik Tr Vik Eik − ln det Vik iterations generated by Algorithm 1 is stationary point of
i∈Kuk problem (13), and the corresponding Wk∗ is a stationary point
⎛ ⎞
of problem (6). Conversely, if Wk∗ is a stationary point of
+ λk ⎝ Tr Wik WiHk − Pk ⎠, (15) problem (6), then Uk∗ calculated by (11), Vk∗ is given by (14),
i∈Kuk and Wk∗ given by (16), are stationary point of problem (13).
HE et al.: DECENTRALIZED PRECODING FOR CACHE-ENABLED ULTRA-DENSE RADIO ACCESS NETWORKS 407
Abstract—LoRa is one of the promising techniques for enabling operate in the same channel and hence boost the achievable
low power wide area networks for future Internet-of-Things system throughput. Thus, a number of works have consid-
devices. Although LoRa allows flexible adaptations of coverage ered the effect of co-SF interference only, where end-devices
and data rates, it is subject to intrinsic types of interferences: using the same SF on the same channel are subject to colli-
co-SF interferences due to collisions among end-devices with the sions [2], [3]. In particular, the outage probability of a LoRa
same spreading factor (SF), and inter-SF interferences due to col-
lisions among devices with different SFs. Most current works have
system under co-SF interference was analyzed in [2], where
considered perfect orthogonality among different SFs. We thus a signal could be captured if its Signal-to-Interference-plus-
provide a theoretical analysis of the achievable LoRa through- Noise Ratio (SINR) was higher than 6 dB. As the number of
put in uplink, where all LoRa-specific capture conditions are devices increased, it was shown that those co-SF interferences
included. Results show the accuracy of our analysis despite were causing a scalability limit. However, recent studies have
approximations, and the throughput losses from imperfect SF pointed out the fact that SFs were not perfectly orthogonal
orthogonality, under different SF allocations. Our analysis will among themselves [4]. Thus, the effect of inter-SF collisions
enable the design of specific SF allocation mechanisms, in view was investigated through computer simulations and/or exper-
of further throughput enhancements. iments. Namely, [4], [5] showed that inter-SF interferences
Index Terms—LoRa, spreading factor, uplink throughput, could considerably decrease LoRa performance, especially for
imperfect orthogonality. high SFs where frames have a greater time on air.
In this letter, we propose a theoretical analysis of the
achievable throughput on the uplink of a LoRa network,
I. I NTRODUCTION encompassing the effects of co- and inter-SF interferences. To
ensure a successful transmission, a packet must thus satisfy
S THE amount of mobile data traffic will rapidly increase
A during the upcoming years (studies forecast 50 billion
Internet of Things (IoT) devices by 2020), new spectrum
three conditions: 1) its SNR is above the reception threshold,
2) its SINR is above the co-SF capture threshold if there is
co-SF interference, and 3) its SINR is above the inter-SF cap-
access strategies adapted to high device densities are ever ture threshold if there is inter-SF interference. Considering two
more crucial. LoRa [1] is one of the prominent candidates different types of SF allocations, we theoretically derive the
for Low Power Wide Area Networks (LPWANs), providing achievable throughput expressions for both perfect and imper-
wide communication coverage with low power consumption, fect SF orthogonality. Simulation results show the accuracy
at the expense of data rate. Operating in license-free ISM of our analytical expressions despite approximations, as well
bands (i.e., 868MHz in Europe), the LoRa PHY layer uses a as the impact of the various types of interferences and SF
chirp spread-spectrum modulation where different Spreading allocations on the overall system performance.
Factors (SFs) tune the chirp modulation rates. Lower SFs such
as SF7 allow for higher data rates but reduced transmission
range, whereas higher SFs such as SF12 provide longer range II. S YSTEM M ODEL
at lower data rates. On top of the LoRa PHY layer, the higher We consider one cell of radius R with one gateway located
layers were defined by the LoRa Alliance and referred as at its center, as depicted in Fig. 1. There are N end-devices
LoRaWAN. In particular, the MAC protocol is based on a pure uniformly distributed within the cell. We denote by di the
ALOHA access with duty cycle limitations. The LoRaWAN distance from end-device i to the gateway. Since the goal of our
network architecture is a star-like topology where end-devices analysis is to derive the achievable rate by LoRa, we assume
communicate with gateways over several channels. that all end-devices transmit in a single channel of bandwidth
Most studies on LoRa scalability so far assumed a perfect BW = 125 kHz and that they all have packets to transmit.
orthogonality among SFs, thereby creating virtual channels This corresponds to the pure ALOHA access as in LoRaWAN
where multiple users with different SFs could simultaneously with saturated traffic.1 We consider M = 6 SFs, for m ∈ M,
M = {mmin , . . . , mmax }, with mmin = 7 and mmax = 12,
Manuscript received June 20, 2018; revised September 22, 2018; accepted with symbol times Tm = BW 2m . The bit-rate R of SF [1] is
m m
September 24, 2018. Date of publication October 3, 2018; date of current
version April 9, 2019. This work was supported in part by the Grant-in-Aid m × CR
for Scientific Research (Kakenhi) from the Ministry of Education, Science, Rm = 2m
, (1)
Sports, and Culture of Japan under Grant 17K06453, and in part by the NII BW
Research Grant. The associate editor coordinating the review of this paper
and approving it for publication was T. De Cola. (Corresponding author: where CR is the coding rate defined as 4/(4+n) with
Megumi Kaneko.) n ∈ {1, 2, 3, 4}. Lower SFs allow higher data rate but lower
A. Waret and M. Kaneko are with the National Institute of communication range whereas higher SFs provide longer
Informatics, Tokyo 101-8430, Japan (e-mail: antoine.waret@grenoble-inp.org; range at the expense of data rate (see Table I).
megkaneko@nii.ac.jp).
A. Guitton and N. El Rachkidy are with CNRS, LIMOS,
Two types of SF allocation will be investigated. In the first
University Clermont Auvergne, 63000 Clermont-Ferrand, France one, the SFs are uniformly distributed, i.e., every end-device
(e-mail: alexandre.guitton@uca.fr; nancy.el_rachkidy@uca.fr).
Digital Object Identifier 10.1109/LWC.2018.2873705 1 Our analysis can be easily applied to multiple channels and duty cycles.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
WARET et al.: LoRa THROUGHPUT ANALYSIS WITH IMPERFECT SF ORTHOGONALITY 409
c
with equal SFs transmitting on the same frequency simulta- + (Rm + Rl )P (Sm , Sl ; S{∀m =m,l} )
c c
P (Sm ; S{∀m =m} ) + l=m P (Sm , Sl ; S{∀m =m,l} ) + . . . + Marginalizing over γ1 , . . . , γj −1 and making the change of
P (S1 , S2 , . . . , SM ), i.e., the sum probabilities of all events variable γi = rcα , we get by independency of user channels,
i
with a success for SFm , but any state for all other SFm . (i,m)
Psucc (SFm ) is hence the marginal probability P (Sm ; Sm ∪ PcapcoSF (j )
c , ∀m = m),3 detailed in the following sections. −1 lm
j
Sm
lm qcoSF riα
h(rp )drp
Finally, the uplink throughput τ can be expressed as = e− c α h(ri )dri . (14)
lm−1 p=1 lm−1 1 + qcoSF rrpi
τ= Rm × Psucc (SFm ), (8) p=i
m h(r )dr
m∈M We define I (ri ) = llm−1 r . Since all user
1+qcoSF ( ri )α
where Rm is the bit-rate of SFm and Psucc (SFm ), the positions, i.e., distance random variable rp for user p, are sta-
probability of a successful transmission. tistically equivalent as they follow
−1the same probability density
Next, we analyze the throughput under perfect and imperfect function h(r), we can write jp=1 I (ri ) = [I (ri )]j −1 . After
SF-orthogonality, for the SF-distance case.4 similar derivations as [8], the expression becomes
lm
(i,m) q rα
A. Perfect Orthogonality PcapcoSF (j ) = exp − coSF i [I (ri )]j −1 h(ri )dri . (15)
lm−1 c
We assume first that SFs are perfectly orthogonal: no end-
In particular, for α = 4, the primitive function of I(ri ) is
device suffers inter-SF interferences. Hence, the probability of
a successful transmission Psucc (SFm ) of (8) is given by r 2 r 2√ 1
i
J (ri , r ) = − qcoSF arctan 2√ .
N
R R ri
qcoSF
N r
Psucc (SFm ) = Pj (caprx , capcoSF ), (9) (16)
j
j =1
In case of co-SF interferences, the interferers are the end-
where j denotes the total number of end-devices at SFm among devices with the same SF as end-device i, i.e., with the same
N and Pj (caprx , capcoSF ) is the joint probability for reception distance boundaries. The expression of I(ri ) becomes
condition and co-SF capture.
I (ri ) = J (ri , lm ) − J (ri , lm−1 ). (17)
1) For j = 1: there are no co-SF interferences, thus only
the reception condition for device i with SFm among N−1 Therefore, (9) can be written for SF-distance allocation with
devices with different SFs needs to be satisfied, perfect SF orthogonality as,
(i,m)
P1 (caprx , capcoSF ) = P1 (caprx ) = (1 − pm )N −1 Pcaprx . Psucc (SFm ) =
N (i,m)
(1 − pm )N −1 × Pcaprx
(10) 1
N
(i,m)
N (i,m)
We determine Pcaprx for the SF-distance case. Given our + (1 − pm )N −j × PcapcoSF (j ), (18)
assumptions, the SNR γi is modeled as an exponential random j
j =2
variable with mean γi . Therefore, (i,m) (i,m)
(i,m)
with Pcaprx given in (12) and PcapcoSF (j ) in (15).
Pcaprx = P (γi ≥ qSFm |γi ) × P (γi ), (11)
where qSF m is the specific threshold of SFm . Defining γi = B. Imperfect Orthogonality
c P0 ×A(fc ) In reality, SFs are not perfectly orthogonal, so all three
riα , where c = σn2
is the path-loss constant, we obtain
capture conditions are to be satisfied to achieve a successful
lm
(i,m) qSFm riα 2r transmission. Thus, Psucc (SFm ) of (8) becomes
Pcaprx = exp − × 2i dri . (12) N
lm−1 c R
N
Psucc (SFm ) = Pj (caprx , capcoSF , capiSF ), (19)
Although this integral cannot be expressed in closed form, it j
j =1
can be efficiently determined by numerical methods.
2) For j ≥ 2: both the reception and co-SF conditions where Pj (caprx , capcoSF , capiSF ) is the joint probability for
must be fulfilled. As qSFm ≤ 1 in linear for all SFs whereas reception condition, co-SF capture and inter-SF capture, when
qcoSF = 4 (6 dB) for all SFs as explained in Section II, if there are j devices among N with SFm .
co-SF capture is satisfied, so is the reception condition, hence 1) For j = 1: the device is only subject to inter-SF capture
and reception conditions. From Table I, the dominant condition
Pj (caprx , capcoSF ) = Pj (capcoSF ) depends on each SFm . Thus, we may approximate
(i,m)
= (1 − pm )N −j PcapcoSF (j ). (13) P1 (caprx , capcoSF , capiSF ) = P1 (caprx , capiSF )
(i,m) (i,m)
In case of co-SF interferences, there are j−1 interfer- ≈ min (1 − pm )N −1 Pcaprx , PcapiSF (1) . (20)
(i,m)
ers, PcapcoSF (j ) = P ( j −1γi ≥ qcoSF ), which (i,m) (i,m)
k =1 γk +1 The expression of PcapiSF (j ) is similar to PcapcoSF (j ), but
is developed using random instantaneous SNR variables
γk and random average SNR (position) variables γk as with different thresholds and number of interferers. Inter-SF
P ( j −1γi ≥ qcoSF |γ1 , . . . , γi , . . . , γj −1 , γ1 , . . . , γj −1 ) × interferences are caused by the devices that are not in the
k =1 γk +1 annulus corresponding to SFm , i.e., they have different SFs.
P (γ1 , . . . , γi , . . . , γj −1 , γ1 , . . . , γi−1 , γi+1 , . . . , γj −1 ). If j is the number of devices at SFm , there are N−j devices
with other SFs. Therefore,
3 This can be also easily verified by direct calculation.
lm
(i,m) qiSF m riα
[I (ri )]N −j h(ri )dri . (21)
4 Analysis for SF-random case are omitted due to lack of space; they will
PcapiSF (j ) = exp −
be detailed in an extended journal version. lm−1 c
WARET et al.: LoRa THROUGHPUT ANALYSIS WITH IMPERFECT SF ORTHOGONALITY 411
to use wireless communications, whilst the backhaul link is a the NCVs will be forwarded to the CPU from each AP, and
lossless but capacity-limited ‘bit-pipe’. Note that the backhaul then by concatenating them and multiplying by the inverse
link can be wireless or fixed and the proposed scheme can be of the binary PNC mapping matrix, the original data from
implemented in both cases. each MT will be recovered. Note that the global mapping
A 2m -QAM scheme is employed at each MT with a modu- matrix G [G1 G2 · · · Guj ]T must be non-singular in order
lation function of M : F2m −→ Ω, where m stands for a mod- to unambiguously recover the source data [12]. Maximum
ulation order, in bits per symbol, and Ω stands for the set of likelihood (ML) detection can also be used for PNC decoding.
all possible modulated symbols. In the multi-access stage, uj
MTs are served by an AP where the MTs transmit their signals
III. D ESIGN C RITERION FOR B INARY PNC
simultaneously, yielding a received signal at the j th AP of
Singular fading in the multi-access stage is a serious
uj
problem which affects detection performance at the CPU.
rj = hj ,i si + zj , for j = 1, 2, . . . , n, (1) We give a simple example here with 2 MTs to illustrate
i=1 this problem and the singular fading is defined as a situa-
where hj ,i denotes the complex Gaussian distributed random tion in which different pairs of transmitted signals cannot be
channel coefficient between the j th AP and the i th MT, and distinguished at the receiver, mathematically given by
si = M (bi ) is the 2m -QAM signal, where bi denotes the
binary data vector with a dimension of m × 1. zj denotes the hj ,1 s1 + hj ,2 s2 = hj ,1 s1 + hj ,2 s2 , (5)
additive white Gaussian noise with zero mean and variance σ 2 . where si and si stand for the QAM signals at the i th
For the purpose of industrial applicability, binary mapping MT, and si = si . The special channel coefficient vector
matrices are employed in the proposed PNC encoder and
hsf [hj ,1 , hj ,2 ] is defined as a singular fade state (SFS).
decoder. Define a mapping function Gj at the j th AP, the
Note that the solution of (5) is not unique, since there is more
PNC encoding is then given by
than one SFS for each QAM scheme. (5) implies that super-
xj = Gj ⊗ b, (2) imposed constellation points corresponding to two different
MT data combinations coincide, and hence this data cannot
where b [b1 , b2 , . . . , buj ]T denotes the ms × 1 binary joint be unambiguously decoded at this AP.
uj
message vector with b ∈ F2ms ×1 and ms = i=1 mi , where Besides the coincident symbols, there is a set of superim-
mi denotes the modulation order at the i th MT. Gj denotes posed symbols that would map to the same NCV; this set is
a binary matrix with dimensions of lj × ms , where lj denotes (1) (2)
defined as a cluster and denoted by scl = [sj ,sc , sj ,sc , . . . ]. We
the number of rows of the mapping matrix used at the j th then define the minimum distance between different clusters as
l ×1
AP, and ⊗ denotes matrix multiplication over F2 . xj ∈ F2j (τ ) (τ )
i k 2
is the network codeword vector (NCV) detected by the AP dmin = min |sj ,sc − sj ,sc | ,
(τi ) (τ )
k )
which consists of lj linear combinations of the original binary Nj (sj ,sc )=Nj (sj ,sc
data. (τ )
i (τ ) (τ ) (τ )
∀sj ,sc ∈ scl , ∀sj ,sc
k
∈ scl , for i = 1, 2, . . . , k = 1, 2, . . . . (6)
We define scj as a set which contains all possible combina-
tions of the modulated signals from the MTs, where uj denotes In the coincident symbols cases dmin = 0. The design
the number of MTs served by the j th AP. Given a 1×uj chan- criterion is to employ a mapping function that labels the
nel coefficient vector hj [hj ,1 , hj ,2 , . . . , hj ,uj ] at the j th AP, constellation points within a clash to the same NCV with
the vector containing all ujms possible superimposed signals maximised dmin in order to achieve unambiguous decoding
can be calculated by at the CPU. A detailed mathematical proof is derived in [12]
and the design criterion holds in multiple-MT case.
(1) (2) (u ms )
sj ,sc [sj ,sc , sj ,sc , . . . , sj ,scj ] = hj scj , for j = 1, 2, . . . , n.
(3) IV. A DAPTIVE M APPING F UNCTION
S ELECTION A LGORITHM
According to the definition, the modulation function M (·) is a
one-to-one bijective mapping function which maps all possible In this section, we describe an adaptive PNC mapping
combinations of binary data to the complex symbols s ∈ Ω. matrix selection algorithm based on the design criterion intro-
In (2), the joint message vector b can be mapped to an NCV duced in the previous section. According to the criterion, the
x by a binary matrix G. Then with the expression in (3), we separate mapping matrices used at each AP should encode
can always find a surjective PNC mapping function Nj that the constellation points within one cluster to the same NCV
(k ) and additionally, the global mapping matrix G formed by the
maps the superimposed constellation point sj ,sc to an NCV, concatenation of the selected mapping matrices at each AP
mathematically given by should be invertible for unambiguous decoding at the CPU.
(k )
Nj (sj ,sc ) = xj , for k = 1, 2, . . . , ujms . (4) We define an L × uj matrix Hsfs = [hsfs1 , . . . , hsfsL ]T whose
rows contain all special channel vectors that cause different
At each AP, an estimator calculates the conditional prob- SFSs for uj modulation scheme combinations. Like the work
ability of each possible NCV given the mapping function in [6], our proposed algorithm comprises two procedures, the
Nj and the channel coefficients. The estimator returns the first of which is an Off-line search and the second is an
log-likelihood ratio (LLR) of each bit of xj which is then On-line search. The proposed Off-line and On-line algorithms
applied to a soft decision decoder. In the backhaul stage, are described in Algorithm 1.
414 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Algorithm 1 Binary Matrices Selection (Off/On-Line Search When the channel qualities between the MTs and the j th AP
Algorithm) are poor, the value of dmin may be much smaller than that at
Off-line Search the other APs, and this leads to performance degradation. Our
1: Define bs = [bs1 · · · bsu ] as a u ms ×ms matrix containing solution is to employ lj × ms mapping matrices at the j th AP
all possible combinations of binary data with lj ∈ [1, ms ]. An l × ms matrix with l > ms /2 at one AP
2: Define Gofl as a set that contains all l ×ms binary matrices and l < ms /2 at the other may give a greater overall mini-
for l ∈ [1, ms ], where the number of matrices is Nofl mum inter-cluster distance dmin over both APs than choosing
3: sc = [M1 (bs1 ), · · · , M2 (bsu )]T l = ms /2 at both.
4: for i = 1 : L do each SFS Our proposed algorithm provides a solution to reduce back-
(1) (u ms ) haul load in N-MIMO networks with a high user density by
5: si,sfs [si,sfs , · · · , si,sfs ] = hsfsi sc
(n ) (n )
employing PNC technique. However, the calculation of singu-
1 2 2
6: d = |si,sfs − si,sfs | , ∀n1 , n2 , n1 = n2 lar fading with multiple MTs in PNC is still an open question
7: Find Ncl different clashes scl containing the si,sfs with in binary systems. In order to apply the proposed algorithm in
d = 0 and the corresponding binary data combinations bcl practical scenarios, orthogonal frequency division multiplexing
8: for nofl = 1:Nofl do each binary matrix (OFDM) [15] could be utilised by allocating a pair of MTs
9: for ncl = 1:Ncl do each clash to the same frequency in the multiple access stage. In [13],
(ncl ) (nofl ) (ncl )
10: xcl = Gofl ⊗ bcl NCV calculation the proposed algorithm in an OFDM scheme using multiple
11: end for Universal Software Radio Peripherals (USRPs) is presented.
(n )
12: if all xcl cl are the same then We focused on network design and hardware implementation,
(n ) such as the work in [11], and studies of how to address the
13: Store Goflofl as a matrix candidate
14: end if synchronisation, channel estimation and interference issues in
15: end for the perspective of hardware design.
16: end for According to [12], the number of SFSs increases exponen-
On-line Search tially with increasing modulation order index m which leads to
17: Define Gonl as the set that contains the Nsel selected an increase in the number of matrices returned by the Off-line
mapping matrix candidates search and makes the On-line search algorithm impractical.
18: for j = 1 : J do each AP However, the 2m -QAM schemes are all defined in F2m and
19: sj ,sc = hj sc All possible superimposed signals the same constellation points can be found in different constel-
20: for nsel = 1:Nsel do each selected matrix lation books. This property allows us to use the same mapping
(n ) (n ) matrix to resolve different SFSs, and the number of mapping
21: xselsel = Gonlsel ⊗ bs NCV calculation matrices being stored reduces dramatically. We term SFSs
22: Find the set of clusters from the positions of which can be resolved by the same matrix as another SFS,
(n )
repeated columns in xselsel , and among all pairs of image SFSs. Then the original number of SFSs in quadrature
clusters: phase shift keying (QPSK) is 13 and that in 16-QAM is 389
(cl ) (cl )
23: dmin = min |sj ,sci − sj ,sck |2 , but the number is reduced to 5 in QPSK and 169 in 16-QAM
∀i,k ,i=k
where cli and clk denote the cluster indexes after the image SFSs are removed. A detailed study of the
24: end for impact on the network performance with fewer SFSs is derived
25: end for in [12] with computational complexity analysis.
26: The non-singular global mapping matrix selection:
G = [G1 · · · GJ ]T with maximum dmin V. N UMERICAL R ESULTS AND D ISCUSSION
In this section, we illustrate the simulation results of the
proposed algorithm in the 5-node system shown in Fig. 1. We
assume the multi-access links are wireless and the backhaul is
The main differences between the proposed design and the a capacity-limited wired link. For simulation simplicity, two
work in [6] and [12] are as follows. In [6], MTs employ MTs are served by two APs and we assume a 3dB difference
the same modulation scheme and the channels have the same in average path loss between each AP and the two MTs. The
average path loss, and the dimension of the mapping matri- MT with better channel quality employs a higher order modu-
ces stored at each AP is equal to ms /2 so that the sizes of lation scheme, corresponding to a higher rate. A convolutional
the NCVs at each AP are the same. In [12], a design guide- code is used in the simulation while other channel codes, such
line for an arbitrary number of cooperating users employing as low-density parity-check (LDPC) code [14], may also be
the same QAM scheme and transmission power is presented. employed. As benchmarks we use ideal CoMP with an unre-
The approach presented in this letter focuses on the situation stricted number of bits transmitted in the backhaul network and
in which the modulated signals from MTs are received at an non-ideal CoMP in which the soft information on the backhaul
AP with different powers, then in order to maintain a good is quantized to a total of 12, 6 and 2 bits per symbol, resulting
performance in terms of error probability, different modulation in the total backhaul load of 48, 24 and 8 respectively (since 4
schemes should be used. In order to achieve a higher degree LLRs are calculated at each AP).
of freedom, mapping matrices with different dimensions may In Fig. 2, the BER performances of different approaches
be utilised. The non-singular global mapping matrix used for are illustrated. We assume block fading in the multi-access
PNC decoding at the CPU is formed by concatenating the stage and that perfect channel information can be obtained to
mapping matrices used at each AP with the maximum dmin . obtain accurate SFSs. As shown in the figure, a 2.5dB and 3dB
PENG et al.: ADAPTIVE OPTIMAL MAPPING SELECTION ALGORITHM FOR PNC USING VARIABLE QAM MODULATION 415
VI. C ONCLUSION
In this letter, an engineering applicable approach is
presented to implement adaptive PNC in binary N-MIMO
systems when different MTs transmit with different powers.
Unlike our work in [12] and [13], we focus on applying the
Fig. 2. BER of the proposed algorithm with 3dB difference in channel proposed algorithm in practical scenarios in which multiple
quality. MTs could employ different modulation schemes. We also
illustrate how the estimated channel affects the optimal matrix
selection accuracy. The simulation results illustrate the bene-
fits of the proposed PNC approach in terms of backhaul load
reduction and BER improvement compared to the non-ideal
CoMP with quantized bits in the backhaul channel.
R EFERENCES
[1] M. V. Clark et al., “Distributed versus centralized antenna arrays in
broadband wireless networks,” in Proc. Spring IEEE Veh. Technol. Conf.,
May 2001, pp. 33–37.
[2] D. Lee et al., “Coordinated multipoint transmission and reception in
LTE-advanced: Deployment scenarios and operational challenges,” IEEE
Commun. Mag., vol. 50, no. 2, pp. 148–155, Feb. 2012.
[3] L. Zhou and W. Yu, “Uplink multicell processing with limited backhaul
via per-base-station successive interference cancellation,” IEEE J. Sel.
Areas Commun., vol. 31, no. 10, pp. 1981–1993, Oct. 2013.
[4] Y. Wang, Z. Chen, and M. Shen, “Compressive sensing for uplink
cloud radio access network with limited backhaul capacity,” in Proc. 4th
ICCSNT, Harbin, China, Dec. 2015, pp. 898–902.
Fig. 3. Non-optimal mapping matrices selection probability with estimated [5] Q. T. Sun, J. Yuan, T. Huang, and K. W. Shum, “Lattice network codes
channel. based on Eisenstein integers,” IEEE Trans. Commun., vol. 61, no. 7,
pp. 2713–2725, Jul. 2013.
[6] A. G. Burr and D. Fang, “Linear physical-layer network coding for
improvement can be observed when employing the proposed 5G radio access networks,” in Proc. 1st Int. Conf. 5GU, Äkäslompolo,
algorithm with QPSK+BPSK and 16-QAM+QPSK compared Finland, Nov. 2014, pp. 116–221.
[7] Y. Wang et al., “A multilevel framework to lattice network coding,”
to employing the algorithm in [12] using QPSK+QPSK and IEEE Trans. Inf. Theory, submitted for publication. [Online]. Available:
16-QAM+16-QAM, respectively. The ideal CoMP is the best https://arxiv.org/abs/1511.03297
among the approaches due to the availability of unrestricted [8] T. Koike-Akino, P. Popovski, and V. Tarokh, “Optimized constellations
bits in the backhaul network; while quantization with limited for two-way wireless relaying with physical network coding,” IEEE J.
Sel. Areas Commun., vol. 27, no. 5, pp. 773–787, Jun. 2009.
precision results in a performance degradation in non-ideal [9] D. Fang and A. G. Burr, “Uplink of distributed MIMO: Wireless network
CoMP. In the QPSK+BPSK case, the proposed algorithm coding versus coordinated multipoint,” IEEE Commun. Lett., vol. 19,
outperforms CoMP with 24 bits, while requiring a total of no. 7, pp. 1229–1232, Jul. 2015.
only 9 backhaul bits. [10] B. Nazer and M. Gastpar, “Compute-and-forward: Harnessing
interference through structured codes,” IEEE Trans. Inf. Theory, vol. 57,
The channel coefficients obtained at each AP are impor- no. 10, pp. 6463–6486, Oct. 2011.
tant to the proposed algorithm because its performance relies [11] G. Chiurco, M. Mazzotti, F. Zabini, D. Dardari, and O. Andrisano,
on the accurate identification of the closest SFS. In practi- “FPGA design and performance evaluation of a pulse-based echo
canceller for DVB-T/H,” IEEE Trans. Broadcast., vol. 58, no. 4,
cal communication networks, the trade-off between the length pp. 660–668, Dec. 2012.
of information symbols and that of pilot symbols for chan- [12] T. Peng et al., “Adaptive wireless network coding in network MIMO:
nel estimation has been discussed in [15]. We have illustrated A new design for 5G and after,” IEEE Trans. Commun., submitted for
the effect of channel estimation on the proposed algorithm publication. [Online]. Available: https://arxiv.org/abs/1801.07061
[13] Y. Chu et al., “Implementation of uplink network coded modulation for
in terms of the probability of selecting a non-optimal matrix, two-hop networks,” IEEE Access, submitted for publication. [Online].
which is shown in Fig. 3. In the simulation, different num- Available: https://arxiv.org/abs/1808.07354
bers of pilot symbols have been employed to estimate the [14] A. G. Burr, Modulation and Coding for Wireless Communications.
Harlow, U.K.: Prentice-Hall, 2000.
channel, which affects the accuracy of identification of the [15] H. D. Taun, H. H. Kha, H. H. Nguyen, and V.-J. Luong, “Optimized
SFS calculated at each AP. As illustrated in Fig. 3(a), when training sequences for spatially correlated MIMO-OFDM,” IEEE Trans.
a short pilot sequence is employed, there is a maximum of Wireless Commun., vol. 9, no. 9, pp. 2768–2778, Sep. 2011.
416 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Massive multiple-input multiple-output (MIMO) CSI. These methods sparsify CSI under certain bases to
systems rely on channel state information (CSI) feedback to apply CS for feedback and reconstruction [2] or distributed
perform precoding and achieve performance gain in frequency compressive channel estimation [4]. In reality, CSI is only
division duplex networks. However, the huge number of anten- approximately sparse under elaborate base selection or
nas poses a challenge to the conventional CSI feedback reduction sparsity modeling. Many existing CS algorithms experience
methods and leads to excessive feedback overhead. In this letter,
we develop a real-time CSI feedback architecture, called CsiNet-
difficulty in CSI compression and recovery if there is a model
long short-term memory (LSTM), by extending a novel deep mismatch.
learning (DL)-based CSI sensing and recovery network. CsiNet- Time correlation property of slow-varying channels has been
LSTM considerably enhances recovery quality and improves considered in [2] to further reduce feedback quantity. This
tradeoff between compression ratio (CR) and complexity by method reuses the previously retained channel information
directly learning spatial structures combined with time corre- for subsequent CSI recovery if the error is under a certain
lation from training samples of time-varying massive MIMO threshold. However, the reused information only provides an
channels. Simulation results demonstrate that CsiNet-LSTM estimate and is hard to update in real time. As a result, reso-
outperforms existing compressive sensing-based and DL-based lution degrades and the feedback overhead cannot reduce any
methods and is remarkably robust to CR reduction. more in fast-changing channels.
Index Terms—Massive MIMO, FDD, CSI feedback, compres- Recently, deep learning (DL) methods has been successfully
sive sensing, deep learning. applied in wireless communications [5]–[7]. A CSI feedback
network, called CsiNet [8], uses an autoencoder-like architec-
ture to mimic the CS and reconstruction processes. It uses
I. I NTRODUCTION an encoder to obtain compressed representation (codewords)
ASSIVE multiple-input multiple-output (MIMO) by directly learning channel structures from the training
M systems have been recognized as a critical develop-
ment for future wireless communications. With downlink
data and a decoder to recover CSI via one-off feedforward
multiplication. CsiNet remarkably outperforms the CS-based
channel state information (CSI), a base station (BS) with methods. But it reconstructs CSI independently, and ignores
massive antennas can use channel-adaptive techniques to time correlation in time-varying channels.
eliminate inter-user interference and increase channel capac- In this letter, we improve the architecture by considering
ity. In frequency division duplex (FDD) networks, downlink time correlation. This letter is motivated by the recurrent
CSI can only be estimated at user equipment (UE) and fed convolutional neural network (RCNN) architecture that has
back to the BS. The excessive overhead has motivated many been successfully used in video representation and recon-
feedback reduction techniques, such as vector quantization struction [9]. The basic idea is to use a convolutional neural
and codebook-based approaches [1]. However, quantization network (CNN) and a recurrent neural network (RNN) to
errors pose a challenge to CSI-sensitive applications, whereas extract spatial features and interframe correlation, respec-
the huge number of antennas complicates the codebook tively. Our contribution in this letter is summarized as
design and accordingly increases feedback overhead. follows.
The compressive sensing (CS)-based CSI feedback • We propose an DL-based CSI feedback protocol for FDD
approaches proposed recently address the aforementioned MIMO systems by extending CsiNet with a long short-
problems by using the spatial and temporal correlation of term memory (LSTM) network, which is a classic type
of RNN. The proposed network, called CsiNet-LSTM,
Manuscript received July 16, 2018; revised September 8, 2018; accepted modifies the CNN-based CsiNet for CSI compression and
September 24, 2018. Date of publication October 5, 2018; date of current initial recovery and uses LSTM to extract time correlation
version April 9, 2019. This work was supported in part by the National for further improvement in resolution.
Science Foundation (NSFC) for Distinguished Young Scholars of China under
• The experiment results demonstrate that CsiNet-LSTM
Grant 61625106, and in part by NSFC under Grant 61531011. The work of
C.-K. Wen was supported in part by the Ministry of Science and Technology achieves the best recovery quality and outperforms state-
of Taiwan under Grant MOST 106-2221-E-110-019 and in part by ITRI in of-the-art CS methods in terms of complexity. CsiNet-
Hsinchu, Taiwan. The associate editor coordinating the review of this paper LSTM exhibits remarkable robustness to compression
and approving it for publication was A. Liu. (Corresponding author: Shi Jin.)
ratio (CR) reduction and enables real-time and extensible
T. Wang and S. Jin are with the National Mobile Communications
Research Laboratory, Southeast University, Nanjing 210096, China CSI feedback applications without considerably increas-
(e-mail: wangtianqi@seu.edu.cn; jinshi@seu.edu.cn). ing overhead compared with CsiNet.
C.-K. Wen is with the Institute of Communications Engineering, National
Sun Yat-sen University, Kaohsiung 80424, Taiwan (e-mail: ckwen@ieee.org).
G. Y. Li is with the School of Electrical and Computer II. S YSTEM M ODEL
Engineering, Georgia Institute of Technology, Atlanta, GA 30332 USA
(e-mail: liye@ece.gatech.edu). Consider an FDD downlink massive MIMO-orthogonal
Digital Object Identifier 10.1109/LWC.2018.2874264 frequency division multiplexing (OFDM) system with Nc
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
WANG et al.: DL-BASED CSI FEEDBACK APPROACH FOR TIME-VARYING MASSIVE MIMO CHANNELS 417
subcarriers. The BS deploys a uniform linear array (ULA) with 2D-DFT is performed to obtain the original spatial frequency
Nt transmit antennas. In a time-varying channel caused by UE channel matrix.
mobility, the received signal at time t on the nth subcarrier for
UE with a single receiver antenna can be modeled as, III. C SI N ET-LSTM
The CsiNet in [8] demonstrates remarkable performance
yn,t = hT
n,t vn,t xn,t + zn,t , (1) in CSI sensing and reconstruction. However, the resolution
degrades at low CR because the it only focuses on angular-
where hn,t ∈ CNt ×1 , xn,t ∈ C, and zn,t ∈ C denote the delay domain sparsity (Observation 1) and ignores the time
instantaneous channel vector in the frequency domain, trans- correlation (Observation 2) of time-varying massive MIMO
mit data symbol, and additive noise, respectively, vn,t ∈ channels. The two observations in Section II are similar to
CNt ×1 is the beamforming or precoding vector designed by the spatial structure and interframe correlation of videos,
the BS based on the received downlink CSI. We denote respectively. Motivated by RCNN that excels in extracting
the CSI matrix at time t in the spatial-frequency domain spatial-temporal features for video representation [9], we will
as Ht = [h1,t , . . . , hNc ,t ]T ∈ CNc ×Nt . In practice, the extend CsiNet with LSTM to improve CR and recovery quality
UE continuously estimates and feeds instantaneous CSI (i.e., trade-off. We will also introduce the multi-CR strategy in [9]
Ht , Ht+1 , . . .) back to the BS to track the time-varying char- to implement variable CRs on different channel matrices.
acteristics of the channel. To reduce feedback overhead, we The proposed CsiNet-LSTM is illustrated in Fig. 1(b) with
can exploit the following observations. CsiNet shown in Fig. 1(a). Our model includes the following
Observation 1 (Angular-Delay Domain Sparsity): Ht can two steps: angular-delay domain feature extraction, correlation
be transformed into an approximately sparsified matrix Ht in representation and final reconstruction.
the angular-delay domain via 2D discrete Fourier transform 1) Angular-Delay Domain Feature Extraction: We apply
(2D-DFT) [8] by Ht = Fd Ht Fa , where Fd ∈ CNc ×Nc and CsiNet with two different CRs to {Ht }T t=1 to learn the
Fa ∈ CNt ×Nt are two DFT matrices. First, due to limited angular-delay domain structure and perform sensing and ini-
multipath time delay, performing DFT on frequency domain tial reconstruction. A high-CR CsiNet transforms the first
channel vectors (i.e., column vectors of Ht ) can transform Ht channel H1 into an M1 × 1 codeword vector that retains suf-
into a sparse matrix in the delay domain, with only the first ficient structure information for high resolution recovery. A
Nc (< Nc ) rows having distinct non-zero values. Secondly, low-CR CsiNet encoder performs on the remaining T − 1
as proved in [10], the channel matrix is sparse in a defined channel matrices to generate a series of M2 × 1 codewords
angle domain by performing DFT on spatial domain channel (M1 > M2 ), given that less information is required due to
vectors (i.e., row vectors of Ht ) if the number of transmit channel correlation. The T − 1 codewords are all concate-
antennas, Nt → +∞, is very large. Usually, Ht is only nated with the first M1 × 1 codeword before being fed into
approximately sparse for finite Nt , which challenges conven- the low-CR CsiNet decoder to fully utilize feedback informa-
tional CS methods. Therefore, we will propose a DL-based tion. Each CsiNet outputs two matrices with size Nc × Nt as
feedback architecture without sparsity prior constraint. We per- extracted features from the angular-delay domain.
form sparsity transformation to decrease parameter overhead All low-CR CsiNets shown in Fig. 1(b) share the same
and training complexity. We retain the first Nc non-zero rows network parameters, i.e., weights and bias, because they per-
and truncate Ht to a Nc × Nt matrix, Ht , which reduces the form the same work. which dramatically reduces parameter
total number of parameters for feedback to N = Nc Nt . overhead. Furthermore, the architecture can be easily rescaled
Observation 2 (Correlation Within Coherence Time): UE to perform on channel groups with different T if the value of
motion during communication results in a Doppler spread, T changes to adapt to the channel-changing speed and feed-
i.e., time-varying characteristics of wireless channels. With the back frequency. In practice, a low-CR CsiNet will be reused
maximum movement speed denoted as v, coherence time can T − 1 times instead of making T − 1 copies. The grey blocks
be calculated as in Fig. 1(b) load parameters from the original CsiNets as pre-
c training before end-to-end training with the entire architecture.
Δt = , (2) This method can alleviate vanishing gradient problems due to
2vf0
long paths from CsiNets to LSTMs.
where f0 is the carrier frequency and c is the speed of light [3]. 2) Correlation Representation and Final Reconstruction:
The CSI within Δt is considered correlated with one other. We use LSTMs to extend the CsiNet decoders for time correla-
Therefore, instead of independently recovering CSI, the BS tion extraction and final reconstruction. LSTMs have inherent
can combine the feedback and previous channel information memory cells and can keep the previously extracted informa-
to enhance the subsequent reconstruction. We set the feed- tion for a long period for later prediction. In particular, the
back time interval as δt and place T adjacent instantaneous outputs of the CsiNet decoders form a sequence of length T
angular-delay domain channel matrices into a channel group, before being fed into three-layer LSTMs. Each LSTM has
i.e., {Ht }T
t=1 = {H1 , . . . , Ht , . . . , HT }. The group exhibits 2Nc Nt hidden units, which is the same as the output dimen-
correlation property, as long as T satisfies 0 ≤ δ t · T ≤ Δ t. sion. The final outputs are then reshaped into two Nc × Nt
In this letter, we design an encoder, st = fen (Ht ), at the matrices as the final recovered Ĥt . The spatial frequency
UE to compress each complex-valued Ht of {Ht }T t=1 into an domain CSI can then be obtained via inverse 2D-DFT. At each
M-dimensional real-valued codeword vector st (M < N). If time step, the LSTMs implicitly learn time correlation from the
two real number matrices are used to represent the real and previous inputs and then merge them with the current inputs
imaginary parts of Ht , then CR will be M/2N. We also design to increase low CR recovery quality. Correlation information
a decoder with a memory that can extract time correlation from is updated after each step due to the nature of LSTM. The
the previously recovered channel matirces, Ĥ1 , . . . , Ĥt−1 and experimental results show that the highly compressed T − 1
combine them with the received st for current reconstruction, matrices can achieve better recovery accuracy than H1 as a
Ĥt = fde (st ; Ĥ1 , . . . , Ĥt−1 ), where 1 ≤ t ≤ T. Then, inverse benefit from LSTMs.
418 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Fig. 1. (a) CsiNet architecture presented in [8]. It comprises an encoder with a 3 × 3 conv layer and an M-unit dense layer for sensing and a decoder with
a 2Nc Nt -unit dense layer and two RefineNet for reconstruction. Each RefineNet contains four 3 × 3 conv layers with different channel sizes. (b) Overall
architecture of CsiNet-LSTM. H1 and the remaining T − 1 channel matrices are compressed by high-CR and low-CR CsiNet encoders, respectively. Codewords
are concatenated before being fed into the low-CR CsiNet decoder, and final reconstruction is performed by three 2Nc Nt -unit LSTMs.
We use end-to-end learning to obtain all parameters for the places with two uniform velocities. Therefore, Δt is 30s
encoder and the decoder denoted as Θ = {Θen , Θde }. Notably, and 0.56s, respectively. Compressed CSI is fed back every
Ht are normalized with all elements scaled into the [0, 1] δt = 0.04 s. We set the channel group size T = 10, which sat-
range before being fed into the network. This normalization is isfies δ t · T < Δ t in both scenarios. We perform experiments
required for CsiNet. For details, we refer to [8]. Let f denote at CR values of 1/16, 1/32, and 1/64, with the first channel
the final trained network defined as H1 compressed under 1/4.
Ĥt = f (Ht ; Θ) = fde (fen (H1 ; Θen ), . . . , fen (Ht ; Θen ); Θde ). Training, validation, and testing sets have 75,000, 12,500,
and 12,500 samples, respectively, for offline training. Some
(3)
parameters are preloaded from the CsiNet for initialization.
We select ADAM as the optimization algorithm and use mean- The epochs are adjusted for a convergence situation ranging
squared error (MSE) as the loss function, which is defined as, from 500 to 1,000. The batch size is 100 and the learning
M T
1 rates are 0.001 and 0.0001 for the former and latter epochs,
L(Θ) = f (Ht ; Θ) − Ht 22 , (4) respectively.
M We compare our architecture with three state-of-the-
m=1 t=1
where M is the total number of samples in the training set and art CS-based algorithms, namely, LASSO 1 −solver [12],
· 2 is the Euclidean norm. TVAL3 [13], and BM3D-AMP [14], and the DL-based
The procedure for CsiNet-LSTM is described as follows. CsiNet [8]. LASSO uses simple sparsity priors but achieves
Multiple CR CsiNet encoders are deployed at UE, whereas the good performance. TVAL3 is a minimum total variation
CsiNet decoders and LSTMs are deployed at the BS. Each side method that provides remarkable recovery quality but with
has a counter. At the beginning, H1 is compressed with high high computing efficiency. BM3D-AMP achieves the most
CR at the UE and recovered by a high-CR CsiNet decoder and accurate recovery performance on natural images and is 10
initialized by the LSTMs at the BS. In the subsequent time times faster than other iterative methods.
step t (2 ≤ t ≤ T), Ht is transformed into a lower-dimensional We use the default configuration in the open source codes
codeword st at the UE, which is expected to contain the of the aforementioned methods for simulation. When compar-
learned correlation information. The lower-dimensional code- ing with CsiNet, we consider the slight difference between
word, st is then concatenated with the first one s1 and inversely datasets and refine the CsiNet parameters on our training set
transformed by the LSTMs at the BS. After each time step, for several epochs for fairness. We run the conventional CS-
the counters add by one. Similar operations continue until the based methods on an Intel Core i7-6700 CPU due to the lack
counters accumulate to T and the LSTMs are reset for the of a GPU solution. CsiNet and CsiNet-LSTM are trained and
subsequent channel group recovery. tested on Nvidia GeForce GTX 1080 Ti GPU.
Normalized MSE (NMSE) is used to evaluate the recovery
IV. S IMULATION R ESULTS AND A NALYSIS performance, which is defined as follows:
T
We use the COST 2100 model [11] to simulate time-varying 1
MIMO channels and generate training samples. We set the NMSE = E Ht − Ĥt 22 /Ht 22 . (5)
MIMO-OFDM system to work on a 20 MHz bandwidth with T
t=1
Nc = 256 subcarriers and use ULA with Nt = 32 antennas at
the BS. The angular-delay domain channel matrix is truncated To compare with CsiNet, the following cosine similarity is
to a size of 32 × 32. Two scenarios are considered: the indoor also calculated:
scenario at 5.3 GHz with UE speed v = 0.0036 km/h and the T Nc
1 1 |ĥH
n,t hn,t |
outdoor scenario at 300 MHz with UE speed v = 3.24 km/h. ρ=E , (6)
T Nc
Data sets are generated by randomly setting different start t=1 n=1 ĥn,t 2 hn,t 2
WANG et al.: DL-BASED CSI FEEDBACK APPROACH FOR TIME-VARYING MASSIVE MIMO CHANNELS 419
R EFERENCES
[1] D. J. Love et al., “An overview of limited feedback in wireless com-
munication systems,” IEEE J. Sel. Areas Commum., vol. 26, no. 8,
pp. 1341–1365, Oct. 2008.
[2] P.-H. Kuo, H. T. Kung, and P.-A. Ting, “Compressive sensing
based channel feedback protocols for spatially-correlated massive
antenna arrays,” in Proc. IEEE WCNC, Shanghai, China, Apr. 2012,
pp. 492–497.
[3] A. Paulraj, R. Nabar, and D. Gore, Introduction to Space-Time Wireless
Communications. Cambridge, U.K.: Cambridge Univ. Press, 2003.
[4] X. Rao and V. K. N. Lau, “Distributed compressive CSIT estimation
and feedback for FDD multi-user massive MIMO systems,” IEEE Trans.
Signal Process., vol. 62, no. 12, pp. 3261–3271, Jun. 2014.
[5] T. Wang et al., “Deep learning for wireless physical layer: Opportunities
and challenges,” China Commun., vol. 14, no. 11, pp. 92–111,
Nov. 2017.
[6] H. Ye, G. Y. Li, and B.-H. Juang, “Power of deep learning for chan-
nel estimation and signal detection in OFDM systems,” IEEE Wireless
where ĥn,t denotes the reconstructed channel vector of the nth Commun. Lett., vol. 7, no. 1, pp. 114–117, Feb. 2018.
[7] H. He, C.-K. Wen, S. Jin, and G. Y. Li, “Deep learning-
subcarrier at time t. ρ can measure the quality of beamforming based channel estimation for beamspace mmwave massive
vector when the vector is set as vn,t = ĥn,t /ĥn,t 2 since MIMO systems,” IEEE Wireless Commun. Lett., to be published,
the UE will achieve the equivalent channel ĥH n,t ĥn,t /ĥn,t 2 .
doi: 10.1109/LWC.2018.2832128.
The performance comparison of NMSE, ρ, and runtime [8] C.-K. Wen, W.-T. Shih, and S. Jin, “Deep learning for massive
MIMO CSI feedback,” IEEE Wireless Commun. Lett., to be published,
are summarized in Table I. From the table, the DL-based doi: 10.1109/LWC.2018.2818160.
CsiNet and CsiNet-LSTM considerably outperform all CS- [9] K. Xu and F. Ren, “CSVideoNet: A real-time end-to-end learning frame-
based methods. Fig. 2 shows a reconstruction result of the work for high-frame-rate video compressive sensing,” in Proc. IEEE
5th channel matrix of a certain channel group in outdoor sce- WACV, Mar. 2018, pp. 1680–1688.
[10] C.-K. Wen, S. Jin, K.-K. Wong, J.-C. Chen, and P. Ting, “Channel
nario as an example, which represents the average performance estimation for massive MIMO using Gaussian-mixture Bayesian learn-
at different CRs. Apparently, CsiNet and CsiNet-LSTM con- ing,” IEEE Trans. Wireless Commun., vol. 14, no. 3, pp. 1356–1368,
tinue to offer adequate beamforming gain at low CRs, where Mar. 2015.
CS-based methods fail to work. In particular, CsiNet-LSTM [11] L. Liu et al., “The COST 2100 MIMO channel model,” IEEE Wireless
achieves the lowest NMSE at all CRs and is multiple times Commun., vol. 19, no. 6, pp. 92–99, Dec. 2012.
[12] I. Daubechies, M. Defrise, and C. D. Mol, “An iterative threshold-
lower than CsiNet, especially when CR is low. ing algorithm for linear inverse problems with a sparsity constraint,”
Notebly, CsiNet-LSTM has the least performance loss as Commun. Pure Appl. Math., vol. 75, no. 11, pp. 1412–1457, Nov. 2004.
CR decreases, with only 8% and 10% for indoor and outdoor, [13] C. Li, W. Yin, and Y. Zhang, “User’s guide for TVAL3: TV minimization
respetively. The simulation results indicate that the remain- by augmented Lagrangian and alternating direction algorithms,” CAAM
ing channel matrices {Ht }T t=2 recovered from a low CR
Rep., vol. 20, pp. 46–47, 2009.
[14] C. A. Metzler, A. Maleki, and R. G. Baraniuk, “From denoising to com-
exhibit similar recovery quality and are better than the first pressed sensing,” IEEE Trans. Inf. Theory, vol. 62, no. 9, pp. 5117–5144,
channel matrix H1 from a high CR, which is −14.74 dB Sep. 2016.
420 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
k
Markov chain. Rm,n and qm,n (Δm, Δn) could be obtained By substituting Eq. (15) into problem (14), the steady-state
according to Eqs. (1-4) in Section II and details are condition can be expressed as constraint (16.c) and a matrix
omitted due to space limit. Based on this, the transi- equation Qy = 0 with ym,n k k } as variables,
= {πm,n fm,n
tion matrix of the Markov chain H could be obtained. where the constant matrix is denoted by Q and can be derived
Denote the steady-state distribution of this Markov chain from H. In this way, the optimization (14) is converted into a
by π = {π0,0 , π0,1 , . . . , π0,N0 , . . . , πn+n0 ,n . . . , πN0 +n0 ,N0 } linear programming which is summarized as follow:
which satisfies N0 n+n K
T 0
1 π = 1, k
(10) min nck ym,n (16.a)
k
π ,fm,n
H π = π. n=0 m=0 k =1
In the t-th time-slot, when system state (M[t], N[t]) = (m, n), N0 n+n
0 K
1 k k th
the cost and expectation of packets dropped are given by nck s.t Rm,n ym,n ≤ Plos (16.b)
k k . Thus, the average cost and λu
and Rm,n with probability fm,n n=0 m=0 k =1
average expectation of tasks dropped are given by n+n
N0 0 K
k
N0 n+n K
ym,n =1 (16.c)
0 n=0 m=0 k =1
k
Cava = nck πm,n fm,n (11)
Qy = 0 (16.d)
n=0 m=0 k =1
k
and ym,n ≥ 0 ∀m, n, k (16.e)
N0 n+n
K
0 This problem can be solved efficiently in polynomial time
k k
Elos = Rm,n πm,n fm,n . (12) using interior-point method [10]. After the optimal solu-
n=0 m=0 k =1 k ∗ of the linear programming (16) is obtained, the
tion ym,n
Furthermore, the expectation of number of the tasks arriving corresponding steady-state distribution can be represented as
in the t-th time-slot is the arrival rate λu . Thus, the packet loss K
∗ k ∗
rate, which is the probability that a computing task is dropped πm,n = ym,n . (17)
before it is completed, is given by k =1
E k ∗
ava
Plos = los . (13) To obtain the cost-optimal strategy, we can derive fm,n
λu k ∗
from ym,n , which is given below.
∗
Case 1: When πm,n = 0, the optimal strategy is given by
IV. O PTIMAL C OST-Q O S T RADEOFF
k ∗
In typical systems, the QoS of users should be guarantee
k ∗ ym,n
ava cannot exceed the toler-
firstly, i.e., the packet loss rate Plos fm,n = . (18)
∗
πm,n
ance Plos th . Furthermore, it is necessary to reduce the cost of
Case 2: When πm,n ∗ = 0, which means that the state (m, n)
the MEC system, i.e., average cost Cava , as much as possible.
Thus, we have the following optimization problem: is a transient state. Then, a simple strategy can be used, i.e.,
N0 n+n
K
0 k ∗ 1
k fm,n = . (19)
min nck πm,n fm,n (14.a) k
k
π ,fm,n
n=0 m=0 k =1 k ∗k ∗
The time complexity of deriving fm,n from ym,n is
N0 n+n
0 K
1 k k th also polynomial. In conclusion, the cost-optimal dynamic
s.t Rm,n πm,n fm,n ≤ Plos (14.b)
λu pricing strategy could be obtained in polynomial time.
n=0 m=0 k =1
Based on this result, the optimal cost-QoS tradeoff can be
1T π = 1 (14.c) achieved.
Hπ = π (14.d) Remark 1: Our proposed approach could be extended and
K applied to a more generalized scenario where the priority of
k
fm,n =1 ∀m, n (14.e) tasks and vehicles with different computing capability are con-
k =1 sidered. By modeling each type of task and vehicle into a
k queue, the system could be formulated into a Markov chain
fm,n≥ 0 ∀m, n, k (14.f)
with more dimensions. Then, the cost-optimal dynamic pricing
πm,n ≥ 0 ∀m, n, (14.g) strategy could be obtained using the proposed approach.
where constraints (14.b) and (14.c-d) denote the constraints
on QoS tolerance and the steady-state condition, respectively. V. N UMERICAL R ESULTS
The objective function and constraints in optimization (14) In this section, we validate our theoretical results
k }, {f k }, or {π
are linear combinations of {πm,n fm,n m,n m,n }. via simulation studies, and explain the outcomes in a
k } in Eq. (8),
By recalling the normalization condition of {fm,n more comprehensive way. Throughout this section, we set
πm,n can also be expressed as n0 = 3, K = 7, μu = 0.2, and other parameters
K
K
are summarized in Table I. By solving the optimization
k k
πm,n = πm,n fm,n = ym,n . (15) problem (16), the optimal dynamic pricing strategy can be
k =1 k =1 obtained.
HAN et al.: DYNAMIC PRICING STRATEGY FOR VEHICLE ASSISTED MEC SYSTEMS 423
TABLE I
PARAMETERS U SED FOR S IMULATION into consideration. Thus, the performance improvement of the
proposed dynamic pricing strategy has been verified.
Fig. 4 presents the optimal cost versus the size of the park-
ing lot with different task arrival rates λu . The packet loss rate
th is set to 0.15. As expected, the cost decreases
constraint Plos
when the size of parking lot N0 increases and then approaches
to different asymptotes, where the line with higher λu exhibits
higher average cost. For a given N0 , the system with higher
λu requires higher cost. Therefore, for a busy MEC system
with higher λu , a larger parking lot is required or otherwise
higher cost has to be incurred.
VI. C ONCLUSION
In this letter, we have investigated a dynamic pricing
strategy for vehicle assisted MEC systems. By adjusting
the price dynamically to control the arrival and depar-
ture rates of vehicles, the cost of the MEC system will
be minimized under a given constraint on QoS, which is
evaluated by the packet drop rate. The system is mod-
Fig. 3. Optimal cost-QoS tradeoffs between different strategies. eled as a two-dimensional Markov chain. Then the average
cost and QoS could be obtained by analyzing the steady-
state distribution of Markov chain. Based on these, the
optimization problem is formulated and solved. Moreover, the
cost-optimal dynamic pricing strategy could be obtained to
minimize the cost of the MEC system and achieve optimal
cost-QoS tradeoff, which has a significant performance
improvement compared with the fixed pricing strategy.
Finally, our theoretical results is validated by comprehensive
simulations.
R EFERENCES
[1] H. Ding, C. Zhang, Y. Cai, and Y. Fang, “Smart cities on wheels:
A newly emerging vehicular cognitive capability harvesting network
for data transportation,” IEEE Wireless Commun., vol. 25, no. 2,
pp. 160–169, Apr. 2018.
Fig. 4. Optimal cost versus parking lost size N. [2] S. Liu, J. Tang, Z. Zhang, and J.-L. Gaudiot, “Computer architectures
for autonomous driving,” IEEE Comput., vol. 50, no. 8, pp. 18–25,
Aug. 2017.
[3] M. T. Beck, M. Werner, S. Feld, and T. Schimper, “Mobile edge comput-
ing: A taxonomy,” in Proc. Int. Conf. Adv. Future Internet, Nov. 2018,
Fig. 3 compares the numerical results between the proposed pp. 48–55.
dynamic pricing strategy, one another dynamic pricing strategy [4] P. Mach and Z. Becvar, “Mobile edge computing: A survey on architec-
ture and computation offloading,” IEEE Commun. Surveys Tuts., vol. 19,
based on Lyapunov optimization, and fixed pricing strategy no. 3, pp. 1628–1656, 3rd Quart., 2017.
with N0 = 4 and λu = 0.8. The basic idea of the Lyapunov [5] N. Zheng and N. Geroliminis, “Modeling and optimization
optimization [11] is to minimize its Lyapunov drift-plus- of multimodal urban networks with limited parking and
penalty function, whose objective is to stabilize the virtual dynamic pricing,” Transp. Res. B Methodol., vol. 83, pp. 36–58,
Jan. 2016.
queue N [t] + n0 − M [t] while optimizing the average cost. [6] G. R. Newsham and B. G. Bowker, “The effect of utility time-varying
In Fig. 3, simulation results of the proposed dynamic pricing pricing and load control strategies on residential summer peak elec-
strategy are given by Monte-Carlo simulation, which match tricity use: A review,” Energy Policy, vol. 38, no. 7, pp. 3289–3296,
Jul. 2010.
perfectly well with the optimization results. It can be seen [7] X. Hou, Y. Li, M. Chen, D. Wu, and S. Chen, “Vehicular fog comput-
that with the decrease of packet loss rate constraint Plosth , i.e.,
ing: A viewpoint of vehicles as the infrastructures,” IEEE Trans. Veh.
higher QoS requirement, the required cost rises in all strate- Technol., vol. 65, no. 6, pp. 3860–3873, Jun. 2016.
[8] J. Liu, Y. Mao, J. Zhang, and K. B. Letaief, “Delay-optimal computation
gies. Additionally, when the cost is large enough, the packet task scheduling for mobile-edge computing systems,” in Proc. IEEE
loss rate in all strategies will decrease to a minimum, which ISIT, Jul. 2016, pp. 1451–1455.
verifies the performance improvement by utilizing the com- [9] M. Jia, W. Liang, Z. Xu, and M. Huang, “Cloudlet load balancing
in wireless metropolitan area networks,” in Proc. IEEE INFOCOM,
puting units of parked vehicles as temporary MEC servers. Apr. 2016, pp. 1–9.
Moreover, for any Plos th , the average cost in the fixed pricing
[10] S. P. Boyd and L. Vandenberghe, Convex Optimization. New York, NY,
strategy is always higher than that in the proposed dynamic USA: Cambridge Univ. Press, 2004.
pricing strategies. The average cost of the strategy based on [11] W. Sun, J. Liu, Y. Yue, and H. Zhang, “Double auction-based resource
allocation for mobile edge computing in industrial Internet of Things,”
Lyapunov optimization is also higher since it does not take IEEE Trans. Ind. Informat., vol. 14, no. 10, pp. 4692–4701, Oct. 2018,
the distribution of arrival and departure of the task and vehicle doi: 10.1109/TII.2018.2855746.
424 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
Abstract—We compute the spectral-energy efficiency Pareto 2 γ , f˙ : 1 ST D ERIVATIVE )
N OTATION (α = 3.5, δ = 2/β, β > 2, η = κσN A
front in Poisson cellular networks, by formulating a spectral-
energy efficiency bi-objective optimization problem as a function
of either the transmit power or the density of the base sta-
tions. Capitalizing on fundamental theoretical results on weighted
Tchebycheff optimization problems applied to strictly quasi-
concave functions, we derive analytical expressions of the unique
Pareto-optimal solution of the bi-objective problem. We prove
that the Pareto front is constituted by a subset of the spectral-
energy efficiency tradeoff and that it can be formulated in
analytical terms. We identify new functional relations between the
Pareto-optimal transmit power and the density of base stations.
Index Terms—Cellular network, Pareto front, point process.
I. I NTRODUCTION
HE SPECTRAL Efficiency (SE) and Energy Efficiency
T (EE) are important performance metrics that guide the
optimization of cellular networks. Under typical operating con-
trade-off but in characterizing the SE-EE Pareto front, which
is the solution of a bi-objective optimization problem [2].
ditions, however, they are conflicting objective functions [1]: Recently, Aydin et al. [8] and Hao et al. [9] have computed
There exists no single solution that simultaneously optimizes the SE-EE Pareto front in cellular networks with the aid of
each of them. There exist, on the other hand, several optimal multi-objective optimization theory. Therein, however, numer-
solutions for which none of the two objective functions can ical methods are used, and, thus, no analytical formulation of
be improved without degrading the other objective. The SE- the SE-EE Pareto front is given.
EE pairs that fulfill the latter optimality condition are referred Against this background, we derive an explicit analyt-
to as Pareto-optimal solutions, and the corresponding SE-EE ical formulation of the SE-EE Pareto front in cellular
curve is known as the Pareto front [2, Definition 2.2.1]. networks, which is obtained by solving a SE-EE bi-objective
The aim of this letter is to derive a complete and explicit optimization problem, as a function of either the transmit
formulation of the SE-EE Pareto front in cellular networks. power or the density of the BSs. This major contribution
We focus our attention on analytically formulating the SE-EE is obtained by capitalizing on the approach recently intro-
Pareto front from the system-level standpoint, i.e., by tak- duced in [10] for computing, in closed-form, the SE and EE
ing the average with respect to the irregular deployments in Poisson cellular networks, and on fundamental theoreti-
of the cellular Base Stations (BSs) and the random loca- cal results on the existence and uniqueness of Pareto-optimal
tions of the Mobile Terminals (MTs) within the cells. Several solutions in weighted Tchebycheff optimization problems [2].
authors have studied the SE-EE trade-off in wireless networks,
e.g., [3] and [4], and some of them have recently analyzed the II. B I -O BJECTIVE P ROBLEM F ORMULATION
SE-EE trade-off in cellular networks from the system-level We consider a cellular network whose BSs and MTs are dis-
standpoint [5]–[7]. The contribution of this letter is, how- tributed according to two mutually independent homogeneous
ever, different: We are not interested in analyzing the SE-EE Poisson Point Processes (PPPs) of density λBS and λMT ,
respectively. The same system model as in [10] is assumed.
Manuscript received August 8, 2018; accepted September 29, 2018. Date By using the notation in Table I, the SE (bit/sec/m2 ) and the
of publication October 9, 2018; date of current version April 9, 2019. This
EE (bit/Joule) can be formulated, respectively, as follows:
work was supported in part by the EC through the Projects 5Gwireless,
BESMART, and CacheMire. The associate editor coordinating the review of λBS L(λMT /λBS ) λMT
this paper and approving it for publication was M. Sheng. (Corresponding SE = BW ρD Q λBS , Ptx , (1)
author: Marco Di Renzo.) 1 + ΥL(λMT /λBS ) λBS
M. Di Renzo, A. Zappone, and T. T. Lam are with the Laboratoire SE
des Signaux et Systèmes, CNRS, CentraleSupélec, Univ Paris EE = (2)
Sud, Université Paris-Saclay, 91192 Gif-sur-Yvette, France (e-mail:
λBS (Ptx − Pi )L(λMT /λBS ) + λMT Pc + λBS Pi
marco.direnzo@l2s.centralesupelec.fr). where the denominator in (2), i.e., Pgrid , is the power
M. Debbah is with the Laboratoire des Signaux et Systèmes, consumption (Watt/m2 ) of the cellular network [10].
CNRS, CentraleSupélec, Univ Paris Sud, Université Paris-Saclay, 91192
Gif-sur-Yvette, France, and also with the Mathematical and Algorithmic In Table II, we summarize important properties of the SE
Sciences Lab, Huawei Technologies, 92100 Boulogne-Billancourt, France. and EE, as a function of Ptx and λBS , that are used next. For
Digital Object Identifier 10.1109/LWC.2018.2874642 generality, we use the symbol ξ to denote either Ptx or λBS .
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
DI RENZO et al.: SPECTRAL-EE PARETO FRONT IN CELLULAR NETWORKS: STOCHASTIC GEOMETRY FRAMEWORK 425
TABLE II
ξ = Ptx OR ξ = λBS , ξ (min) & ξ (max) A RE THE M IN & M AX OF ξ, Proposition 1: Let f (ξ) and g(ξ) be two strictly quasi-
Ω = [ξ (min) , ξ (max) ], Ω(min) = [ξ (min) , ξ (o) ], concave functions in ξ ∈ Ω. Then, the optimization problem:
Ω(max) = [ξ (o) , ξ (max) ] Pmax − min : maxξ∈Ω {min{f (ξ), g(ξ)}} (5)
has a unique solution in ξ ∈ Ω.
Proof: It follows from Lemma 1 and Lemma 2.
Let f (ξ) be a strictly quasi-concave function in ξ ∈ Ω.
Similar to the notation in Table II, the unique maximizer
of f (·) and its corresponding maximum objective value are
denoted by ξ (f ,opt) and fopt = f (ξ (f ,opt) ), respectively.
A. Weighted Tchebycheff Method
Unless otherwise stated, we assume f (ξ) = SE(ξ) and
g(ξ) = EE(ξ) in ξ ∈ Ω. From Table II, we have fopt = SEm
and gopt = EEo . Let μ ∈ [0, 1], the weighted Tchebycheff
optimization problem is defined as follows [2, Sec. 3.4]:
PT : maxξ∈Ω {min{μF (ξ), (1 − μ)G(ξ)}} (6)
where F (ξ) = f (ξ)/fopt − 1 and G(ξ) = g(ξ)/gopt − 1.
In particular, the unique maximizer of EE, ξ (o) , is [10]: Lemma 3: Let f (ξ) and g(ξ) be two strictly quasi-concave
functions in ξ ∈ Ω. Then, PT in (6) has, for each μ ∈ [0, 1],
ξ (o) = ξ (EE,opt) = max ξ (min) , min ξ (∗) , ξ (max) (3) a unique solution that is (strong) Pareto-optimal.
Proof: It follows from Prop. 1 and [11, Corollary 3.4.4].
where ξ (∗) is the unique unconstrained maximizer of the EE.
Proposition 2: According to Lemma 3, let ξ (μ) be the
We note that the SE and EE are continuous functions in ξ ∈ Ω.
unique solution of PT for μ ∈ [0, 1]. The Pareto front of
We are interested in solving, as a function of ξ = Ptx or
ξ = λBS , the following bi-objective optimization problem [2]: P is given by the pairs (f (ξ (μ) ), g(ξ (μ) )) obtained varying μ
in [0, 1].
P : maxξ∈Ω [SE(ξ), EE(ξ)] (4) Proof: It follows from [11, Th. 3.4.5].
Definition 1: The bi-objective optimization problem in (4) Lemma 4: According to Lemma 3, let ξ (μ) be the unique
is said to be non-trivial or meaningful if ξ (o) < ξ (max) . solution of PT for μ ∈ [0, 1]. Then, ξ (μ) ∈ [ξ (o) , ξ (max) ].
Remark 1: If ξ (o) ≥ ξ (max) , P in (4) is trivial because, Proof: In the range ξ (μ) < ξ (o) , the SE and EE are increas-
based on Table II, both the SE and EE are increasing functions ing functions. Based on the definition of Pareto-optimality, the
in ξ. Thus, they are not conflicting objectives and are both pairs (f (ξ (μ) ), g(ξ (μ) )) cannot be Pareto-optimal.
maximized for ξ = ξ (max) , i.e., the (trivial) solution of (4). Lemma 5: According to Lemma 3, let ξ (μ) ∈ [ξ (o) , ξ (max) ]
Remark 2: Strictly monotonically increasing and unimodal be the unique solution of PT for μ ∈ [0, 1]. Then: i) ξ (μ) =
functions are strictly quasi-concave functions [11]. ξ (g,opt) = ξ (o) if and only if μ = 0, ii) ξ (μ) = ξ (f ,opt) =
From Remark 2, we evince that the SE and EE are continu- ξ (m) if and only if μ = 1, iii) ξ (μ) is the unique solution of
ous and strictly quasi-concave functions. We focus our atten-
μF (ξ (μ) ) = (1 − μ)G(ξ (μ) ) if and only if μ ∈ (0, 1).
tion only on non-trivial bi-objective optimization problems.
Proof: If μ = 0 and μ = 1, PT in (6) is equivalent to max-
imizing G(·) and F (·), respectively, since, by definition, they
III. T CHEBYCHEFF S CALARIZATION are both negative functions. If μ ∈ (0, 1), the SE is monotoni-
To solve P and compute the Pareto-front, we employ cally increasing and the EE is monotonically decreasing in the
the scalarization approach [2]. Two methods are used: 1) range [ξ (o) , ξ (max) ]. Also, μF (ξ (o) ) < (1 − μ)G(ξ (o) ) = 0
the conventional weighted Tchebycheff optimization (PT ) and (1 − μ)G(ξ (max) ) < μF (ξ (max) ) = 0. So, the func-
[2, Sec. 3.4], and 2) the simplified weighted Tchebycheff tions μF (ξ) and (1 − μ)G(ξ) cross each other exactly once
optimization (PST ). PST is introduced in this letter and proved in [ξ (o) , ξ (max) ]. By definition of max-min optimization, this
to be equivalent to PT , but shown to be instrumental to analyt- unique crossing point is the solution of PT in (6).
ically formulate the Pareto front. PT is studied first in order to Lemma 6: Let f (ξ) and g(ξ) be two continuous and strictly
mathematically prove the equivalence with PST . We start with quasi-concave functions in ξ ∈ Ω. Then, there exists a con-
some preliminary results on strictly quasi-concave functions. tinuous and strictly decreasing function, C, that expresses the
Lemma 1: Let f (ξ) and g(ξ) be two strictly quasi-concave objective function g in terms of the objective function f , i.e.,
functions in ξ ∈ Ω. Then, the point-wise minimum function, g = C(f ), where f (ξ (g,opt) ) ≤ f ≤ fopt = f (ξ (f ,opt) ) and
X (ξ) = min{f (ξ), g(ξ)}, is strictly quasi-concave in ξ ∈ Ω. g(ξ (f ,opt) ) ≤ g ≤ gopt = g(ξ (g,opt) ).
Proof: Let us define ξa = aξ1 + (1 − a)ξ2 , Xf (ξ1 , ξ2 ) = Proof: It follows from [12, Ths. 2.1 and 2.2], since f (·) and
min{f (ξ1 ), f (ξ2 )}, Xg (ξ1 , ξ2 ) = min{g(ξ1 ), g(ξ2 )}. By g(·) are continuous and strictly quasi-concave functions, and
virtue of strict quasi-concavity, f (ξa ) > Xf (ξ1 , ξ2 ), g(ξa ) > Ω is a non-empty, compact, and convex set.
Xg (ξ1 , ξ2 ) for ξ1 = ξ2 ∈ Ω and a ∈ (0, 1). Then, X (ξa ) = Remark 3: Proposition 2, Lemma 5, and Lemma 6 provide
min{f (ξa ), g(ξa )} > min{Xf (ξ1 , ξ2 ), Xg (ξ1 , ξ2 )} = us with fundamental properties of the SE-EE Pareto front.
min{X (ξ1 ), X (ξ2 )}, where the inequality holds true because They, however, have the following limitations: i) Proposition 2
the point-wise minimum is an increasing function, and f (·) does not yield an explicit analytical formulation of the Pareto
and g(·) are strictly quasi-concave functions. front, which is parameterized as a function of μ, ii) if
Lemma 2: Let f (ξ) be a strictly quasi-concave function in μ ∈ (0, 1), it is not straightforward to obtain a closed-form
ξ ∈ Ω. Then, f (ξ) has a unique maximizer in ξ ∈ Ω. expression of ξ (μ) from Lemma 5, and iii) Lemma 6 asserts
Proof: It follows from [11, Corollary 2.5.1]. the existence of the curve C but does not provide an explicit
426 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
formula for it. Also, no insight for system design is obtained Remark 6: From Proposition 3, we evince that the SE-EE
from them. Pareto front of PT in (6) can be obtained from PST in (7). It
These limitations are overcome with the aid of PST . is constituted, in particular, by: i) the point (SEm , EE(ξ (m) ))
if w = wl (extreme right value), ii) the point (SE(ξ (o) ), EEo )
B. Simplified Weighted Tchebycheff Method if w = wu (extreme left value), and iii) the continuous set of
Unless otherwise stated, we assume f (ξ) = SE(ξ) and points obtained by varying w in the range (wl , wu ).
g(ξ) = EE(ξ) in ξ ∈ Ω. We introduce the simplified weighted Compared with PT in (6), the advantage of PST in (7) is
Tchebycheff problem as follows (for any w ∈ (0, 1)): that, with the exception of the extreme left and right points of
the SE-EE Pareto front that are known, the other points are
PST : maxξ∈Ω(max) min w F (ξ), (1 − w )G(ξ) (7) the unique solution of w F (ξ (w ) ) = (1 − w )G(ξ (w ) ) for w ∈
where F (ξ) = SE(ξ)/SEm and G(ξ) = EE(ξ)/EEo . (wl , wu ) and ξ (w ) ∈ Ω(max) . This is the fundamental result
Remark 4: The optimization problem in (7) is restricted that allows us to compute an explicit analytical formulation of
over the set ξ ∈ Ω(max) by virtue of Lemma 4, i.e., only the SE-EE Pareto front, as proved in the next section.
the values ξ ≥ ξ (o) are admissible in order for the pairs
(SE(ξ), EE(ξ)) not to contradict the definition of Pareto- IV. SE-EE PARETO F RONT
optimality. The following sections provide an explicit formulation of
Remark 5: PST in (7) fulfills the conditions stated in the SE-EE Pareto front for ξ = Ptx and ξ = λBS . We prove,
Proposition 1: It has a unique solution in ξ ∈ Ω(max) for every in particular, that the SE-EE Pareto front can be computed by
w ∈ (0, 1). No conclusion, however, can be drawn about its knowing only ξ (m) = ξ (max) and ξ (o) available from [10].
Pareto-optimality, i.e., Lemma 3 is, in general, not true.
The following lemma and proposition provide us with suf- A. Case Study ξ = Ptx
ficient conditions under which PT and PST describe the same
SE-EE Pareto front as a function of μ and w, respectively. Lemma 8: Let ξ = Ptx and w ∈ [wl , wu ], where wl and
(o)
Lemma 7: Define wl = (1 + EEo /EE(ξ (m) ))−1 ∈ (0, 1) wu are defined in Lemma 7. Define Ptx = ξ (o) . The unique
(w ) (w ) (o)
and wu = (1 + SE(ξ (o) )/SEm )−1 ∈ (0, 1). The unique solu- solution, ξ (w ) = Ptx , of PST in (7) is Ptx = Ptx if w =
tion, ξ (w ) , of PST in (7) is: i) ξ (w ) = ξ (m) if and only if (w )
wu and Ptx = Ptx
(max)
if w = wl . If w ∈ (wl , wu ), it is:
w ≤ wl , ii) ξ (w ) = ξ (o) if and only if w ≥ wu , and iii) the
(w ) (1/w − 1)(SE m /EEo ) − T (o) (max)
unique solution of the equation w F (ξ (w ) ) = (1 − w )G(ξ (w ) ) Ptx = ∈ Ptx , Ptx (8)
in ξ (w ) ∈ Ω(max) if and only if w ∈ (wl , wu ). λBS L(λMT /λBS )
Proof: In Ω(max) = [ξ (o) , ξ (max) ], the SE and EE are where T = λMT Pc + λBS Pi (1 − L(λMT /λBS )).
monotonically increasing and decreasing, respectively. For any Proof: The cases w = wl and w = wu follow from
w ∈ (0, 1), three cases are possible. i) If w F (ξ (max) ) ≤ Lemma 7. As for w ∈ (wl , wu ), (8) follows by insert-
(1 − w )G(ξ (max) ), i.e., the maximum of w F (ξ) in ξ ∈ ing (1) and (2) in w F (ξ (w ) ) = (1 − w )G(ξ (w ) ). Also,
(w ) (o) (max)
Ω(max) is less than the minimum of (1 − w )G(ξ) in ξ ∈ Ptx ∈ [Ptx , Ptx ] by virtue of the continuity of the Pareto
Ω(max) , we have min{w F (ξ), (1 − w )G(ξ)} = F (ξ). This front (Lemma 6).
occurs if w ≤ wl . ii) If (1 − w )G(ξ (o) ) ≤ w F (ξ (o) ), Remark 7: From (8), the following remarks can be made:
(w )
i.e., the maximum of (1 − w )G(ξ) in ξ ∈ Ω(max) is less i) Ptx is given in an analytical form. This is not pos-
than the minimum of w F (ξ) in ξ ∈ Ω(max) , we have sible, in general, from PT in (6); and ii) the functional
(w )
min{w F (ξ), (1 − w )G(ξ)} = G(ξ). This occurs if w ≥ wu . relation between Ptx and λBS is different compared with
iii) If w F (ξ (o) ) < (1 − w )G(ξ (o) ) and w F (ξ (max) ) > the case study where the coverage probability is maximized.
(1 − w )G(ξ (max) ), the functions w F (ξ) and (1 − w )G(ξ) In, e.g., the highly-loaded regime, i.e., L(λMT /λBS ) ≈ 1, (8)
(w )
cross each other exactly once in Ω(max) , since the SE and EE yields Ptx ∝ (λBS )/λBS , where (λBS ) = SEm /EEo =
are monotonically increasing and decreasing, respectively, in SEm (λBS )/EEo (λBS ) implicitly depends on λBS . If the cov-
Ω(max) . By definition of max-min optimization, this unique erage probability is maximized, on the other hand, the func-
crossing point is the solution of PST in (7). By definition, tional relation between transmit power and density of the BS
−β/2
EEo /EE(ξ (m) ) > 1, SE(ξ (o) )/SEm < 1. So, wl < wu . is Ptx ∝ λBS , where β is the path-loss slope [10].
Proposition 3: Let ξ (w ) be the unique solution of PST Theorem 1: As a function of Ptx , the SE-EE Pareto front
in (7) according to Lemma 7. The pairs (f (ξ (w ) ), g(ξ (w ) )) can be obtained from (2) by setting Ptx as follows:
β/2
obtained by varying w in [wl , wu ] describe the same SE-EE η 2/β S(λMT /λBS )SE
Pareto front as PT in (6). Given w, (f (ξ (w ) ), g(ξ (w ) )) Ptx = − ln 1 − (9)
is obtained from PT in (6) by choosing 1/μ = 1 + S(λBS ) L(λBS )
(w ) (w ) (w )
(1/w − φ+ )(1/φ− ), where φ± = 1 ± 1/G(ξ (w ) ). where S(λMT /λBS ) = 1 + ΥL(λMT /λBS ), S(λBS ) =
Proof: From Lemma 5 and Lemma 7, PT and PST yield πλBS S(λMT /λBS ), L(λBS ) = BW ρD λBS L(λMT /λBS ), and
the same (f (ξ (w ) ), g(ξ (w ) )) = (f (ξ (μ) ), g(ξ (μ) )) pairs if and SE lies in the range SE ∈ [SE(ξ (o) ), SEm ].
only if the two equations μF (ξ (μ) ) = (1 − μ)G(ξ (μ) ) and (w )
Proof: It follows by inserting Ptx in (8) into (1) and (2),
w F (ξ (w ) ) = (1 − w )G(ξ (w ) ) are simultaneously satisfied. and by expressing w as a function of the SE from (1). The
By imposing this condition, we obtain 1/μ = 1 + range of values SE ∈ [SE(ξ (o) ), SEm ] follows by virtue of
(w ) (w )
(1/w − φ+ )(1/φ− ). From this latter formula, we evince the continuity of the Pareto front (as stated in Lemma 6).
that: i) μ = 1 if and only if w = wl , ii) μ = 0 if and only if In conclusion, the SE-EE Pareto front is obtained by
w = wu , and iii) μ ∈ (0, 1) if and only if w ∈ (wl , wu ), since inserting (9) into (2) and by plotting the curve for SE ∈
μ decreases monotonically as w increases. [SE(ξ (o) ), SEm ], which is decreasing in SE (Lemma 6).
DI RENZO et al.: SPECTRAL-EE PARETO FRONT IN CELLULAR NETWORKS: STOCHASTIC GEOMETRY FRAMEWORK 427
Abstract—In this letter, we introduce the concept of Rician (DL) would urge the adoption of decoupling strategies. In that
K-factor-based radio resource and mobility management for case, we may set a K-factor threshold to associate the UL to a
fifth generation (5G) ultra-dense networks (UDN), where the LOS/OLOS gNB where, thanks to the minimal path-loss, the
information on the gradual visibility between the new radio UE can reduce its transmit power allowing the reduction of UL
node B and the user equipment (UE)—dubbed X-line-of-sight signal to interference plus noise ratio (SINR) variance, which
(XLOS)—would be required. We therefore start by presenting translates into more efficient and effective UL schedulers and
the XLOS service probability as a new performance indicator;
performance gains [4]. Zooming out from the applications, a
taking into account both the UE serving and neighbor cells. By
relying on a lognormal K-factor model, a parametric expression mathematical characterization of XLOS is yet to be established.
of the XLOS service probability in a 5G outdoor UDN is derived, In this letter, we propose the XLOS service probability
where the link between network parameters and the availability as a performance indicator, and start by introducing a
of an XLOS condition is established. The obtained formula is broader definition of the concept thereof; accommodating
given in terms of the multivariate Fox H-function, wherefore we the monitoring of both the UE serving and neighbor cells.
develop a fast graphical processing unit-enabled MATLAB code. Under the general framework of 5G large-scale parameters
Residue theory is then applied to infer the relevant asymptotic (LSPs) [5], we invoke a lognormal K-factor model for outdoor
behavior and show its practical implications. Finally, numeri- UEs [6] to conduct a closed-form analysis of the XLOS
cal results are provided for various network configurations, and service probability in a 5G multi-tier heterogeneous network
underpinned by extensive Monte-Carlo simulations. (HetNet). The obtained formula is including the different
Index Terms—5G, GPU, multivariate Fox H-function, Rician network parameters such as gNB density, height and antennas
K-factor, UDN, XLOS service probability. beamwidth, and expressed in terms of the multivariate Fox
H-function [7, A.1], wherefore we provide a fast GPU-enabled
I. I NTRODUCTION MATLAB code. Finally, we study the asymptotic behavior
HE EMERGENCE of 5G ultra-dense networks [1] will highlighting the effect of different network and channel
T certainly prompt the reshaping of radio resource and mobil-
ity management algorithms, wherefore a new set of measured
parameters on the availability of LOS conditions.
Using the total probability theorem as well as the indepen- 1 This theorem can be viewed as a generalization of the well-established
dence between γnm , m = 1, . . . , M , the definition (5) can be Gauss-Hermite representations of the lognormal PDF and CDF (see [11]).
430 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
XLOS S ERVICE P ROBABILITY A SYMPTOTIC E XPRESSIONS By plugging (14) into (12), we recognize that integral I1 can
be re-expressed in terms of the multivariate Fox H-function [7,
A.1] as given by (15) on bottom of this page, where parameter
Λnm Ωnm /πλT is encompassing network density, power
and shadowing effects. Finally, a closed-form expression for
PXLOS is deduced by substituting (15) in (9).
B. Asymptotic Behavior
As depicted in Table I, the two asymptotic regimes of the
α/2
ratio Kth Λnm /ωl,m Knm (in W α/ν m 2 ) reflect many prac-
tical scenarios, wherefore it is interesting to establish the
corresponding XLOS service probability expressions; denoted
P XLOS in the sequel. Let H stand for the multivariate Fox
where the multidimensional integral I1 is expressed as
zn2 M znM
+∞
α/2
M
⎡ α/2
⎛ ⎞
⎤
Kth Λn1
M −times
⎢ ωl,1 Kn1
2i ⎜ ⎟
⎥
α ⎢ .
1− ; 11≤i , . . . , 1M ≤i 1≤i≤M −1 , ⎝1 − M ; α ,..., α
2 ⎠
(1, 1)
(1, 1)⎥
I1 = (πλT )−M H0,M : 0,1 : ... : 0,1 ⎢ .
α 2
. . .
⎥ (15)
M ,M −1 : 1,1 : ... : 1,1 ⎢ .
(0, 1)
(0, 1)⎥
2 ⎣
⎦
α/2
Kth Λn
2i
M −times M
− α ; 11≤i , . . . , 1M ≤i 1≤i≤M −1
ωl,M KnM
CHERGUI et al.: RICIAN K-FACTOR-BASED ANALYSIS OF XLOS SERVICE PROBABILITY IN 5G OUTDOOR UDN 431
TABLE II
N ETWORK AND T RANSMISSION S ETTINGS Stroud monomial cubature [12] to (20), and recalling that
0,0
δ(γm − a) = H0,0 [ γam ], we get
L M
1 0,0 γm
p = M /2 wl H0,0 , (21)
π ωl,m
l=1 m=1
with wl and (ul,1 , . . . , ul,M ) are respectively the l th weight
and√abscissas of the M-dimensional cubature, and ωl,m =
10( 2σm ul,m +μm )/10 . Finally, by invoking [7, eq. (2.53)], the
integration of (21) with respect to γm from 0 to γth,m (m =
1, . . . , M ) leads to (7).
A PPENDIX B
P ROOF OF C OROLLARY 1
It immediately follows from applying the superposition
) the-
orem [9, eq. (1.3.3)] to the equivalent PPP ΦT = N
Φ
n=1 n ,
and performing a PDF transformation to the joint distance
distribution of the first M neighbors given by theorem
[13, Appendix].
Fig. 1. XLOS probability versus Kth for UDN, HetNet (Macro/Femto) and R EFERENCES
LDN. Path-loss exponent ν = 3 and shadowing mean μn = 0 for all tiers
n = 1, . . . , N . [1] S. Sun, T. S. Rappaport, R. W. Heath, A. Nix, and S. Rangan,
“MIMO for millimeter-wave wireless communications: Beamforming,
spatial multiplexing, or both?” IEEE Commun. Mag., vol. 52, no. 12,
To that end, we make use of Stenger’s tabulations [16] to pp. 110–121, Dec. 2014.
update the MATLAB code in [17]. Moreover, we introduce [2] L. Bernadó, T. Zemen, F. Tufvesson, A. F. Molisch, and
in [18] an efficient GPU-oriented MATLAB routine to C. F. Mecklenbräuker, “Time- and frequency-varying K -factor of
non-stationary vehicular channels for safety relevant scenarios,” IEEE
calculate the multivariate Fox H-function. Trans. Intell. Transp. Syst., vol. 16, no. 2, pp. 1007–1017, Apr. 2015.
Fig. 1 shows that, in a UDN with light shadowing, LOS [3] N. Garcia, H. Wymeersch, E. G. Larsson, A. M. Haimovich, and
links are easily established (e.g., Kth = 12 dB is obtained with M. Coulon, “Direct localization for massive MIMO,” IEEE Trans. Signal
probability 1). Conversely, the low density network (LDN) Process., vol. 65, no. 10, pp. 2475–2487, May 2017.
scenario unfolds in NLOS situations with non-negligible [4] H. Elshaer et al., “Decoupled uplink and downlink access in heteroge-
neous networks,” in 5G Wireless Technologies. London, U.K.: IET Digit.
probability (e.g., K < −5 dB with probability 0.5). By Library, 2017, ch. 8.
considering the neighboring cells (M = 2, 3) in the HetNet [5] “Study on channel model for frequencies from 0.5 to 100 GHz (release
(Macro/Femto) case for instance, we remark that a substantial 14),” 3GPP, Sophia Antipolis, France, Rep. TR 38.901, May 2017.
increase of the XLOS probability is achieved only in the [6] L. J. Greenstein, S. S. Ghassemzadeh, V. Erceg, and D. G. Michelson,
non-asymptotic regime. Indeed, a high Kth requirement can “Ricean K-factors in narrow-band fixed wireless channels: Theory,
be fulfilled merely by the serving cell, since the K-factors experiments, and statistical models,” IEEE Trans. Veh. Technol., vol. 58,
no. 8, pp. 4000–4012, Oct. 2009.
of neighbor cells become limited by the corresponding [7] A. M. Mathai, R. K. Saxena, and H. J. Haubold, The H-Function: Theory
path-losses as implied by (2). and Applications. New York Ny, USA: Springer, 2010.
[8] NR: Physical Channels and Modulation (Release 15), 3GPP Standard
V. C ONCLUSION TS 38.211, Sep. 2017.
[9] F. Baccelli and B. Blaszczyszyn, Stochastic Geometry and Wireless
In this letter, we have introduced the XLOS service prob- Networks, Volume I–Theory. Boston, MA, USA: Now, 2009.
ability as a new K-factor-based performance indicator, and [10] P. Madhusudhanan et al. (2012). Downlink Performance Analysis
provided its analytical and asymptotic expressions that unveil for a Generalized Shotgun Cellular Systems. [Online]. Available:
arxiv.org/abs/1002.3943
the effect of the variation of 5G network and transmission [11] F. Yilmaz and M.-S. Alouini, “A novel unified expression for the capac-
parameters on the gradual visibility condition of radio links. ity and bit error probability of wireless communication systems over
By tweaking a K-factor threshold Kth , the XLOS metric can generalized fading channels,” IEEE Trans. Commun., vol. 60, no. 7,
be used by network optimization algorithms as the probabil- pp. 1862–1876, Jul. 2012.
ity of e.g., reconfiguring the uplink in a LOS/OLOS gNB. As [12] A. H. Stroud, Approximate Calculation of Multiple Integrals. Englewood
Cliffs, NJ, USA: Prentice-Hall, 1971.
a perspective, the adopted K-factor model from [6] can be [13] H. R. Thompson, “Distribution of distance to Nth neighbour in a pop-
extended to the vehicular case in future works. ulation of randomly distributed individuals,” Ecology, vol. 37, no. 2,
pp. 391–394, Apr. 1956.
A PPENDIX A [14] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
Products, 7th ed. Amsterdam, The Netherlands: Academic, 2007.
P ROOF OF T HEOREM 1 [15] A. A. Kilbas and M. Saigõ, H-Transforms: Theory and Applications
First, by making a simple variablechange, the prod- (Analytical Methods and Special Functions). Boca Raton, FL, USA:
M Chapman & Hall, 2004.
uct of lognormal PDFs pγm , p = m=1 pγm , can be [16] F. Stenger, “Tabulation of certain fully symmetric numerical integra-
reformulated as
tion formulas of degree 7, 9 and 11,” Math. Comput., vol. 25, no. 116,
pp. 935 and S58–S125, Oct. 1971.
1 2 2
p= e −(u1 +...+uM ) Q(u1 , . . . , uM )du1 . . . duM , (20) [17] J. Burkardt. (2010). Stroud Numerical Integration in M Dimensions.
π M /2 [Online]. Available: people.sc.fsu.edu/∼jburkardt/m_src/stroud/stroud.
RM
html
M [18] H. Chergui, M. Benjillali, and M.-S. Alouini, GPU-Enabled Multivariate
where
√ Q(u1 , . . . , uM ) = m=1 δ(γm − Fox H-Function Code, Zenodo, Geneva, Switzerland, Aug. 2018,
10( 2σm um +μm )/10 ). By applying the Gaussian-weight doi: 10.5281/zenodo.1400403.
432 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—We study optimal power splitting-based simultane- not been studied in the literature. In this letter, we analyze
ous wireless information and power transfer (PS-SWIPT) for PtP and two-hop cooperative communication systems wherein
two communication systems wherein all receiving nodes are sub- the destination (and relay) are PS-SWIPT based nodes with
ject to a decoding cost. First, for a point-to-point (PtP) channel, inherent powers and subject to a non-decreasing DCF. From
the optimal achievable rate is obtained, and then, for a two- implementation points of view, increasing the rate does not
hop half-duplex (HD) decode-and-forward (DF) relay channel, necessarily mean the exponential power consumption for ID.
the achievable data rate is maximized. Finally, performances of
Hence, unlike the existing works which assume an increasing
the PtP system and the relay channel with DF and amplify-and-
forward (AF) relaying schemes are discussed where we observe convex DCF for optimization simplicity, we assume a more
performance superiority of the DF relay channel over AF relay- general non-decreasing DCF. The main contributions of this
ing and the PtP channel for sufficiently less decoding cost at the letter are twofold; First, by considering the decoding cost at
relay than the destination. the receiving nodes, the achievable rates are presented and
then, for the non-convex optimization problem, we analytically
Index Terms—PS-SWIPT, PtP, two-hop DF relay, decoding derive closed-form expressions of the optimal data rates. Our
cost, power splitting.
numerical evaluations for the PS-SWIPT systems subject to the
decoding cost reveal the outperformance of the DF relaying
I. I NTRODUCTION compared to the AF one and PtP transmission.
NE OF the promising aspects of the evolving fifth gen-
O eration of wireless networks is to reduce and manage
the power consumption at inexpensive energy harvesting (EH)
II. P OINT TO P OINT C HANNEL W ITH D ECODING C OST
A. System Model
nodes constrained by the amount of battery power, while Assume a PS-SWIPT PtP transmission where the destina-
achieving higher data rate. Thus, the SWIPT and utilizing the tion splits the received signal into two parts; One for EH, and
cooperative relaying are the two key enabling techniques to the other for ID. Since decoding the signal consumes power
prolong the lifetime and throughput of the system, respectively. which is named the decoding cost, the destination utilizes this
By employing the SWIPT, each EH node can adaptively harvested energy for ID. So, this node must optimally har-
manage its received signal for both EH and or information vest energy to maximize the achievable data rate. The required
decoding (ID) [1]. Recently SWIPT-enabled cooperative com- powers for ID, i.e., decoding cost, is given by a non-decreasing
munication systems have gained many research interests from function ϕd (R) where R indicates the achievable data rate.
different aspects; Optimal PS-SWIPT schemes for two-way The received signal at the destination is given by
DF, and full-duplex (FD) DF relaying strategies are analyzed
in [2] and [3], respectively. Reference [4] studies two-way Yd = hsd Xs + Zd , (1)
massive multi-antenna relay channel wherein the users employ where hsd indicates the source-destination channel coeffi-
power splitting technique to harvest energy for data transmis- cient. Throughout this letter, it is assumed that, the local
sion. Bit error rate of the amplify-and-forward (AF) SWIPT channel state information (CSI) is available at the destina-
relaying with direct link over Nakagami-m channels is derived tion (and the relay).1 Xs is the transmitted symbol with
in [5]. In [6], SWIPT two-way AF relay network is consid-
ered to maximize the secrecy sum-rate subject to attaining E{|Xs |2 } = Ps , where E{.} denotes the statistical expec-
minimum energy requirement at the EH node. tation and Zd ∼ CN (0, Nd ) is an independent complex
On the other hand, in contrast to the conventional com- Gaussian noise. Thus, the two splitted signals for EH and ID
munication system, at the low-power and EH nodes, power are respectively given by YEH d , and YID d as follows
√
consumption for ID, named as decoding cost, may become YEH d = ρd Yd , YID d = 1 − ρd Yd + Zp,d , (2)
non-negligible. Recently, energy efficiency (EE) of SWIPT
in Internet of Things (IoT) system subject to constant circuit where ρd ∈ [0, 1] denotes the power splitting ratio and, Zp,d ∼
power consumption is studied in [7] and [8]. Moreover, opti- CN (0, Np,d ) represents the processing noise at the destination.
mal power management for EH wireless networks with decod- Thus, the harvested power is
ing cost are studied in [9]–[11], wherein the decoding cost
function (DCF) is assumed to be an increasing convex function PEH d = ηd ρd (|hsd |2 Ps + Nd ), (3)
of the data rate which simplifies the optimization problems. where ηd ∈ [0, 1] denotes the energy conversion efficiency.
To the best of our knowledge, the impact of decoding cost
in a SWIPT-enabled cooperative communication system has B. Optimization Problem
Manuscript received August 28, 2018; accepted October 2, 2018. Date The optimization problem is formulated as
of publication October 9, 2018; date of current version April 9, 2019. The maximize Rt
associate editor coordinating the review of this paper and approving it for pub- ρd
lication was K. W. Choi. (Corresponding author: Mohammad Javad Emadi.)
The authors are with the Department of Electrical Engineering, Amirkabir s.t. Rt = min Rsd , Rϕ , (4)
University of Technology (Tehran Polytechnic), Tehran 1591634311,
Iran (e-mail: mohsenabedi@aut.ac.ir; hamed_masoomy@aut.ac.ir; 1 By assuming imperfect CSI, using similar steps as presented in this letter,
mj.emadi@aut.ac.ir). one can obtain the optimal solutions, and achievable rate decreases as the
Digital Object Identifier 10.1109/LWC.2018.2874886 channel estimation quality degrades.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ABEDI et al.: PS-SWIPT SYSTEMS WITH DECODING COST 433
Fig. 3. Energy efficiency vs. data rate. Fig. 6. Rt versus source-relay distance and Pr0 .
V. C ONCLUSION
The achievable data rates of PS-SWIPT-based point-to-point
Fig. 4. Rt versus βr for fixed βd . and two-hop DF relay channels subject to the non-decreasing
decoding costs at the receiving nodes are derived and opti-
mized. The numerical evaluations presented for the DF/AF
relay channel, and the PtP system indicate that the two-hop
DF relay channel outperforms the AF relaying and the PtP
channel when the decoding cost at relay is sufficiently less
than that of the destination, subject to optimal placement of
the relay.
R EFERENCES
[1] I. Krikidis et al., “Simultaneous wireless information and power trans-
fer in modern communication systems,” IEEE Commun. Mag., vol. 52,
no. 11, pp. 104–110, Nov. 2014.
[2] C. Peng et al., “Optimal power splitting in two-way decode-and-forward
Fig. 5. Achievable rate versus hardware qualities βr , βd . relay networks,” IEEE Commun. Lett., vol. 21, no. 9, pp. 2009–2012,
Sep. 2017.
[3] H. Liu et al., “Power splitting-based SWIPT with decode-and-forward
besides considering the exponential DCF, linear and constant full-duplex relaying,” IEEE Trans. Wireless Commun., vol. 15, no. 11,
ones are studied in Fig. 4 to analyze effects of different types pp. 7561–7577, Nov. 2016.
of DCFs on data rates for different schemes with fixed hard- [4] X. Wang et al., “Wireless power transfer-based multi-pair two-way relay-
ware quality at the destination, i.e., βd = 5. For i ∈ {r , d }, ing with massive antennas,” IEEE Trans. Wireless Commun., vol. 16,
let the DCFs as; ϕi (R) = βi (2R − 1) for the exponential, no. 11, pp. 7672–7684, Nov. 2017.
[5] Y. Lou et al., “Performance of SWIPT-based differential AF relay-
ϕi (Rt ) = mβi Rt for the linear, and ϕi (Rt ) = ci for the ing over Nakagami-m fading channels with direct link,” IEEE Wireless
constant ones, and let m = 0.5, cd = 2.5, cr = βr . For the Commun. Lett., vol. 7, no. 1, pp. 106–109, Feb. 2018.
given parameters, all schemes with the linear DCF outper- [6] Q. Li et al., “Secure relay beamforming for SWIPT in amplify-and-
form the other cost functions. Note that the superiority of the forward two-way relay networks,” IEEE Trans. Veh. Technol., vol. 65,
performance strongly depends on the values of m, and ci . no. 11, pp. 9006–9019, Nov. 2016.
[7] Y. Huang et al., “Energy-efficient SWIPT in IoT distributed antenna
Fig. 5 investigates joint effects of βr and βd on the optimal systems,” IEEE Internet Things J., vol. 5, no. 4, pp. 2646–2656,
achievable rate for different schemes. For fixed βd , increasing Aug. 2018.
βr increases cost of decoding at the relay, therefore results in [8] J. Tang et al., “Energy efficiency optimization with SWIPT in MIMO
a performance degradation only for the DF relaying. Moreover broadcast channels for Internet of Things,” IEEE Internet Things J.,
for fixed βr , increasing βd increases decoding cost at the vol. 5, no. 4, pp. 2605–2619, Aug. 2018.
[9] C. Qin, W. Ni, H. Tian, R. P. Liu, and Y. J. Guo, “Joint beamforming and
destination which causes the performance loss for all schemes. user selection in multiuser collaborative MIMO SWIPT systems with
Note that the previous figures are obtained for the optimum nonnegligible circuit energy consumption,” IEEE Trans. Veh. Technol.,
placement of the relay. Fig. 6 addresses the impact of the vol. 67, no. 5, pp. 3909–3923, May 2018.
relay’s inherent power and location for βr = βd . For small [10] A. Arafa and S. Ulukus, “Optimal policies for wireless networks with
values of Pr0 , scarcity of relay’s inherent power forces this energy harvesting transmitters and receivers: Effects of decoding costs,”
IEEE J. Sel. Areas Commun., vol. 33, no. 12, pp. 2611–2625, Dec. 2015.
node to move as close as possible towards the source, so it can [11] A. Arafa et al., “Energy harvesting two-way channels with decoding
harvest more energy to support higher data rates. By increas- and processing costs,” IEEE Trans. Green Commun. Netw., vol. 1, no. 1,
ing Pr0 , relay has enough power for the decoding and signal pp. 3–16, Mar. 2017.
436 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—We consider an iterative compensation method to recent progress of orthogonal approximate message passing
treat clipping distortion in orthogonal frequency division multi- (OAMP) [11].
plexing systems with clipping. A conventional approach is using
extrinsic information (EI) from a decoder to estimate and can-
cel clipping distortion iteratively. In this letter, feedback in the II. I TERATIVE C OMPENSATION
form of a posteriori information (API) from the decoder is A. Transmitter Structure
investigated. Correlation may be introduced among messages in
iterative processing when API is directly used as feedback. We Fig. 1 (a) illustrates a coded OFDM system with N sub-
show that carefully choosing a linearization model can suppress carriers. At the transmitter, information bits are first encoded
such correlation. Simulation results show that the API approach by a binary encoder (ENC) and then randomly interleaved.
outperforms the EI one.
Coded bits are mapped to a sequence of N modulated
Index Terms—OFDM, clipping, linearization, iterative com- symbols X = [X1 , X2 , . . . , XN ]T using a constellation
pensation, decorrelation. S ≡ {sj , j = 1, 2, . . . , 2B } with average power normalized to
I. I NTRODUCTION one. An inverse discrete Fourier transform (IDFT) is then
applied as
IGH peak-to-average power ratio (PAPR) is a well-
H known problem in orthogonal frequency division mul-
tiplexing (OFDM) systems. Various methods have been inves-
x = F †X , (1)
tigated to handle the problem [1], [2]. Among them, clipping is where F is the N × N unitary discrete Fourier transform (DFT)
a simple way to achieve very low PAPR [1], [3]. Compared to matrix and (·)† denotes conjugate transpose. The entries of
other alternatives, such as selective mapping and tone reser- x are approximately complex Gaussian distributed and thus
vation, clipping incurs no rate loss and has low transmitter have high PAPR [12]. Clipping can be employed to reduce
implementation complexity [4], [5]. However, clipping causes PAPR [1]. Specifically, a clipping function is defined by
nonlinear distortion in the transmitted signals. Iterative com- Axi /|xi |, |xi | > A
f (xi ) = , (2)
pensation (IC) has been shown to be effective in treating the xi , |xi | ≤ A
clipping distortion at the expense of an increased receiver com-
plexity [6]–[8]. Various symbol detection approaches can be where A > 0 is the clipping threshold. For simplicity, we
used in iterative processing: hard decision is used in [6], while will omit the standard OFDM operations involving adding and
soft decision is used in [8]. Compared to hard decision, soft stripping off cyclic prefixes. The time-domain received signal
decision offers noticeable performance improvement [8]. r is given by
Correlation among messages is a main problem in iterative r = F † ΛF · f (x ) + w , (3)
processing. The standard treatment is using extrinsic informa-
tion (EI) from a soft-input soft-output decoder [8], [9]. It has where w ∼ CN (0, σ 2 I) is a sequence of independent
been reported that a posteriori information (API) may outper- identically distributed (IID) Gaussian noise and Λ =
form EI in some iterative systems [10]. Most related works diag{H1 , H2 , . . . , HN } is a diagonal matrix with Hi being
are experimental. There is still a lack of theoretical treatments the channel coefficient on the i-th subcarrier. Applying DFT
for the API approach. to r, we obtain a frequency-domain signal vector R as
In this letter, we provide a mathematical justification for
R = ΛF · f (F † X ) + W , (4)
using API in iterative compensation. We show that correla-
tion in the iterative process can be suppressed by carefully where W = Fw contains IID Gaussian noise samples.
choosing a linearization model. Our results are based on the
Manuscript received August 18, 2018; accepted October 2, 2018. Date B. Linearization of the Clipping Function
of publication October 9, 2018; date of current version April 9, 2019. This
work was supported by the University Grants Committee of the Hong Kong
The clipping function f(·) in Fig. 1 (a) is nonlinear, which
SAR, China under Project CityU 11280216 and Project CityU 11216817. makes symbol detection difficult. To circumvent this difficulty,
The associate editor coordinating the review of this paper and approving it we linearize f (x ) as [8]
for publication was J. Mietzner. (Corresponding author: Shansuo Liang.)
S. Liang and L. Ping are with the Department of Electronic f (x ) = αx + d , (5)
Engineering, City University of Hong Kong, Hong Kong (e-mail:
ssliang3-c@my.cityu.edu.hk; eeliping@cityu.edu.hk).
J. Tong is with the School of Electrical, Computer and Telecommunications
where α is a constant scalar and d = f (x) −αx is the clipping
Engineering, University of Wollongong, Wollongong, NSW 2522, Australia distortion. Substituting (5) to (4), we have
(e-mail: jtong@uow.edu.au).
Digital Object Identifier 10.1109/LWC.2018.2874935 R = ΛF · (αF † X + d ) + W . (6)
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
LIANG et al.: ON IC OF CLIPPING DISTORTION IN OFDM SYSTEMS 437
2B
Fig. 1. (a) A coded OFDM system with clipping. (b) Diagram of iterative
compensation at the receiver, where ext and APP represent the extrinsic and X̄n = sj · Pr(Xn = sj )API , (10a)
a posteriori information, respectively. j =1
2B
Defining D = Fd in (6), we have V [Xn ] = |sj − X̄n |2 · Pr(Xn = sj )API . (10b)
j =1
R = αΛX + ΛD + W . (7)
Since x = F† X, the a posteriori mean and variance of x,
We aim at estimating X based on (7). Since Λ is diagonal,
denoted as x̄ and Vx respectively, are given by
we can perform symbol-by-symbol detection. The difficulty is
that D might in general be correlated with X. Such correlation N
1
could originate from d in the linearization model (5). We will x̄ = F X̄†
and Vx = V [Xn ]. (11)
address this problem in Section II-D. N
n=1
Fig. 1 (b) shows the proposed receiver structure. The basic
idea is estimating X and d iteratively according to (7) and (5), Now we write (5) in a symbol-by-symbol form as
respectively, as detailed next.
dn = f (xn ) − αxn , n = 1, 2, . . . , N . (12)
C. Iterative Process
Given the mean and variance of x in (11), we model
Assumption 1 below underpins the proposed iterative {p(xn |x̄n ) = CN (x̄n , Vx ), ∀n} by a Gaussian approximation,
receiver shown in Fig. 1 (b). which can be justified due to the IDFT operation [12]. Then
Assumption 1: The entries of D can be approximated by the conditional mean and variance of each dn are, respectively,
Gaussian random variables with mean D̄ and equal variance given by
VD . Further, the entries of D are independent of those of X
and W. d̄n = Exn |x̄n [dn ] = Exn |x̄n [f (xn )] − αx̄n , (13a)
Since D = Fd, the entries of D are weighted sums of 2
many uncorrelated items. The Gaussian approximation for D V [dn ] = Exn |x̄n [|dn − d̄n | ]. (13b)
in Assumption 1 can be justified by the central limit theo-
rem [12]. The independence assumption will be verified in Since D = Fd, the mean and variance of D are, respectively,
Section II-D. computed as1
The iterative receiver in Fig. 1 (b) involves iterations N
between the following steps. 1
D̄ = F d̄ and VD = V [dn ]. (14)
Step 1 (Estimating X via (7)): Since Λ is diagonal, the N
n=1
system model (7) is decoupled into parallel linear channels
as Overall Iterative Process: Steps 1 and 2 above consti-
tute one iteration. In the next iteration, D̄ computed in (14)
Rn = αHn Xn + Hn Dn + Wn , n = 1, 2, . . . , N . (8) is fed back to (9) as the a priori mean and {Pr(Xn =
The modulated symbols {Xn } are assumed to be IID random sj )EI } outputted from the DEM-DEC are used to update
variables taken from the constellation S = {sj } with a priori {Pr(Xn = sj )} in (9). The process then continues iteratively
probabilities {Pr(Xn = sj ), ∀j , n}. At the beginning of the (see Fig. 1(b)). Note that (9) in Step 1 relies on the indepen-
iterative process, we set {Pr(Xn = sj ) = 1/2B , ∀j , n} when dence in Assumption 1. It does not necessarily hold during the
there is no decoding feedback. We also set D̄ = 0 and VD iterative process since d̄ is a function of x̄ as seen from (13).
can be evaluated numerically. More details in the subsequent This issue will be examined carefully in Section II-D.
iterations will be given in Step 2. Based on (8), the a posteriori
probabilities (APP) of the modulated symbols can be evaluated 1 Following the method in [8], V can be evaluated using a look-up table
D
by (9) (at the top of the next page). method. A two-dimension table (obtained by the Monte Carlo Method) can
be built beforehand to characterize the relationship of VD and Vx for the
The a posteriori probabilities {Pr(Xn = sj |Rn )} are trans- clipping function f(·) with a specific clipping threshold. The detailed table
formed into bit log-likelihood ratios (LLRs) by a demapper generation and analysis can be found in [8] and [14].
438 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
|R −Hn D̄n −αHn sj |2
exp − n σ2 +|H 2 · Pr(Xn = sj )
n | VD
Pr(Xn = sj |Rn ) = B 2
, j = 1, 2, . . . , 2B , n = 1, 2, . . . , N . (9)
2 |R n −H D̄
n n −αH s
n j |
j =1 exp − 2
σ +|H | V2 · Pr(X n = s j )
n D
D. The Correlation Problem = Ex ;x̄ f (x ) − α (19b)
The following three propositions establish the orthogonality = Ex f (x ) − α, (19c)
between D̄ and X̄. We will comment on its implication at the
where the last step holds since f(x) is not a function of x̄ . For
end of this sub-section.
α = E[f (x )], we have Ex̄ [η (x̄ )] = 0 from (19c).
Proposition 1: Let f(x) be a continuous function with x ∼
We model X̄ in (10) as [15]
N (εμx , σx2 ). We treat μx as a variable and treat ε and σx2 as
constant values. Then we have
X̄ = · X + Z , (20)
d E[f (x )] df (x )
=ε·E . (15) where = E[X † X̄ ]/E[X 2 ]. It can be verified that, with
d μx dx
such , Z is statistically uncorrelated with X. Defining z = F† Z,
Proof: Take the derivative of E [f (x)] as and according to (11), we have
d E[f (x )] d
= f (x )N (x ; εμx , σx2 )dx (16a) x̄ = · x + z . (21)
d μx dμ
x
d
Since F† is unitary, z and x remain statistically uncorrelated
= f (x ) N (x ; εμx , σx2 ) dx (16b) to each other.
d μx
d
The following approximation is widely used for the treat-
= −ε f (x ) N (x ; εμx , σx2 ) dx (16c) ment of mutually independent signals after DFT or IDFT [12].
dx
∞ Approximation 1: The entries of x, x̄, z, D and D̄ are
= −ε f (x )N (x ; εμx , σx2 ) approximately Gaussian distributed.
−∞
Since z and x are statistically uncorrelated, Approximation 1
2
+ε N (x ; εμx , σx )df (x ) (16d) implies that the entries of z are independent of those of x.
Proposition 3: Under Approximation 1 and α = E[f (x )],
df (x )
=ε N (x ; εμx , σx2 )dx (16e) we have
dx
df (x ) E[(x̄ − x )∗ · (d̄ − d )] = 0, (22)
= ε·E . (16f)
dx where the expectation is taken over the joint distribution of
x̄ and x.
Return to (13), d̄n is a function of x̄n if we treat Vx as Proof: Based on Proposition 2, we have Ex̄ [η (x̄ )] = 0 if
a parameter. Since {xn } have the same variance of Vx , we α = E[f (x )]. Then Proposition 3 is a direct consequence of
denote d̄ = η(x̄ ) with η being an entry-wise function of x̄. [11, Definition 2 and Proposition 2].3
Proposition 2: Let α = E[f (x )] in (5). Then we have We can further show that E[(x̄ − x )∗ · (d̄ − d )] = 0 using
the generalized Stein’s lemma [17], provided that (x̄ −x ) and
E η (x̄ ) = 0, (17)
(x̄ −x ) are jointly Gaussian distributed. Rigorous proof of this
where the expectation is taken on the distribution of x̄ . result requires more dedicated analysis, which is beyond the
Proof: For simplicity, we drop the subscript n in (13) and scope of this letter. Nevertheless, we conjecture that d̄ remains
take the derivative of η(·) as statistically orthogonal to x̄ during the iterative process. Since
d d F is unitary, D̄ and X̄ are also statistically orthogonal to each
η(x̄ ) = Ex |x̄ [f (x )] − αx̄ (18a)
d x̄ d x̄ other:
d
= E [f (x )] − α (18b) E[(X̄ − X )∗ · (D̄ − D)] = 0. (23)
d x̄ x |x̄
d By Approximation 1, the entries of (D̄ − D) can be approx-
= Ex |x̄ f (x ) − α, (18c)
dx imated by zero-mean Gaussian random variables. Combining
with (23), we can see that D̄ and X̄ are independent of each
where in the last step we have applied Proposition 1.2
other if X is Gaussian distributed. The Gaussian requirement
The expectation in (18) is taken over the distribution of x
for X can be approximately ensured using, e.g., superposition
conditional on fixed x̄ . Now we consider all entries in x̄ and
coded modulation [14]. So far, we are still unable to provide
treat x̄ as a random variable. We take the expectation of both
a rigorous justification for Assumption 1 if X is not Gaussian.
sides in (18) over the distribution of x̄ :
Nevertheless, we observed numerically that API works well
d d in general.
Ex̄ η(x̄ ) = Ex̄ Ex |x̄ f (x ) − α (19a)
d x̄ dx
3 The proof in [11] is for real Gaussian inputs. A similar proof for complex
2 It can be verified that Proposition 1 holds for complex Gaussian inputs. Gaussian inputs can be found in [16].
LIANG et al.: ON IC OF CLIPPING DISTORTION IN OFDM SYSTEMS 439
IV. C ONCLUSION
We investigated an iterative API approach to treat clip-
ping distortion in clipped OFDM systems. We showed that
Fig. 2. The BER performance of API and EI in different scenarios. The
frame length is 4096 and CR = 0 dB. For PDF, we consider a downlink
correlation among messages can be suppressed by a proper
OFDMA system with four users and only one target user performs decoding. linearization model. Simulation results showed that the API
The number of iterations is 5. We observed numerically that more than 5 approach outperforms the EI one, especially for PDF.
iterations leads to marginal improvement.
R EFERENCES
[1] G. Wunder, R. F. H. Fischer, H. Boche, S. Litsyn, and J.-S. No, “The
III. A PPLICATIONS AND S IMULATION R ESULTS PAPR problem in OFDM transmission: New directions for a long-lasting
A. Applications to Various OFDM Scenarios problem,” IEEE Signal Process. Mag., vol. 30, no. 6, pp. 130–144,
Nov. 2013.
The proposed method is flexible for different performance- [2] Y. Rahmatallah and S. Mohan, “Peak-to-average power ratio reduction
in OFDM systems: A survey and taxonomy,” IEEE Commun. Surveys
complexity tradeoff. We consider the following strategies Tuts., vol. 15, no. 4, pp. 1567–1592, 4th Quart., 2013.
(from high to low complexity): [3] H. Ochiai and H. Imai, “Performance analysis of deliberately clipped
• Full Decoding Feedback (FDF): All information bits are OFDM signals,” IEEE Trans. Wireless Commun., vol. 50, no. 1,
pp. 89–101, Jan. 2002.
decoded. [4] S. Y. Le Goff, S. S. Al-Samahi, B. K. Khoo, C. C. Tsimenidis, and
• Partial Decoding Feedback (PDF): Only some infor- B. S. Sharif, “Selected mapping without side information for PAPR
reduction in OFDM,” IEEE Trans. Wireless Commun., vol. 8, no. 7,
mation bits are decoded. For the remaining symbols, pp. 3320–3325, Jul. 2009.
the posteriori mean in (10) is calculated by replacing [5] A. Behravan and T. Eriksson, “Tone reservation to reduce the envelope
{Pr(Xn = sj )API } with {Pr(Xn = sj |Rn )} in (9). PDF fluctuations of multicarrier signals,” IEEE Trans. Wireless Commun.,
vol. 8, no. 5, pp. 2417–2423, May 2009.
is suitable for a multi-user downlink OFDMA system, [6] H. Chen and A. M. Haimovich, “Iterative estimation and cancellation of
where a target user does not want to decode other users’ clipping noise for OFDM signals,” IEEE Commun. Lett., vol. 7, no. 7,
pp. 305–307, Jul. 2003.
messages. [7] D. Hao and P. Hoeher, “Iterative estimation and cancellation of clipping
• No Decoding Feedback (NDF): The a posteriori noise for multi-layer IDMA systems,” in Proc. 7th Int. ITG Conf. Source
Channel Coding, Jan. 2008, pp. 1–6.
mean (10) is calculated using {Pr(Xn = sj |Rn )} in (9). [8] J. Tong, L. Ping, Z. Zhang, and V. K. Bhargava, “Iterative soft com-
Then after iterative processing, the DEC only performs pensation for OFDM systems with clipping and superposition coded
decoding once at the final stage. modulation,” IEEE Trans. Commun., vol. 58, no. 10, pp. 2861–2870,
Oct. 2010.
[9] C. Berrou and A. Glavieux, “Near optimum error correcting coding
and decoding: Turbo-codes,” IEEE Trans. Commun., vol. 44, no. 10,
B. Simulation Results pp. 1261–1271, Oct. 1996.
Consider an OFDM system with N = 128 sub-carriers. [10] A. Movahed, M. C. Reed, N. Aboutorab, and S. E. Tajbakhsh,
“EXIT chart analysis of turbo compressed sensing using message pass-
We use a rate-1/2 convolutional code (23, 35)8 and 16-QAM ing dequantization,” IEEE Trans. Signal Process., vol. 64, no. 24,
signaling scheme with Gray mapping. The system rate is pp. 6600–6612, Dec. 2016.
[11] J. Ma and L. Ping, “Orthogonal AMP,” IEEE Access, vol. 5,
R = 2 bits/symbol. Define Eb /N0 in decibel as Eb /N0 = pp. 2020–2033, 2017.
10log10 (E[|f (z )|2 ]/Rσ 2 ) and clipping ratio (CR) in decibel [12] T. Araujo and R. Dinis, “On the accuracy of the Gaussian approximation
as CR = 10 log10 (A2 /E[|x |2 ]). The transmitted signals go for the evaluation of nonlinear effects in OFDM signals,” IEEE Trans.
Commun., vol. 60, no. 2, pp. 346–351, Feb. 2012.
through Rayleigh fading channels with L = 4 taps. [13] A. Chindapol and J. A. Ritcey, “Design, analysis, and performance eval-
In Fig. 2, we compare the bit error rate (BER) performance uation for BICM-ID with square QAM constellations in Rayleigh fading
channels,” IEEE J. Sel. Areas Commun., vol. 19, no. 5, pp. 944–957,
of API and EI in the scenarios discussed in Section III-A. “No May 2001.
treatment” in Fig. 2 means directly detecting and decoding by [14] J. Tong, L. Ping, and X. Ma, “Superposition coded modulation with
peak-power limitation,” IEEE Trans. Inf. Theory, vol. 55, no. 6,
ignoring the clipping effect. We can see that its performance is pp. 2562–2576, Jun. 2009.
quite poor. From Fig. 2, we have the following observations: [15] X. Wang and H. V. Poor, “Iterative (turbo) soft interference cancellation
• For FDF, the advantage of API over EI is small since EI and decoding for coded CDMA,” IEEE Trans. Commun., vol. 47, no. 7,
pp. 1046–1061, Jul. 1999.
performance is already close to the ideal unclipped one. [16] K. Takeuchi, “Rigorous dynamics of expectation-propagation-based sig-
• For PDF, the advantage of API over EI is noticeable since nal recovery from unitarily invariant measurements,” in Proc. IEEE Int.
Symp. Inf. Theory (ISIT), Jun. 2017, pp. 501–505.
EI, without the decoder feedbacks for the unwanted users, [17] C. M. Stein, “Estimation of the mean of a multivariate normal distribu-
performs poorly in this case. tion,” Ann. Stat., vol. 9, no. 6, pp. 1135–1151, Nov. 1981.
440 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Automatic modulation classification (AMC) plays an However, in real communication scenario, due to the influence
important role in many military and civilian communication from naturally or man-made signal sources, such as multiuser
applications. However, it remains a challenging task to sup- interference, car ignitions, and atmospheric noise, the channel
port such AMC mechanisms under impulsive noise environments. noise is non-Gaussian and has impulsive characteristics [6].
Aiming at improving the classification performance in impul-
Although there have been lots of AMC methods designed
sive noise, in this letter, a novel modulation classification method
is proposed by using the cyclic correntropy spectrum (CCES). for optimal performance in Gaussian noise, these algorithms
In the proposed method, CCES is introduced into AMC for typically exhibit significantly worse performance when impul-
effectively suppressing impulsive noise. Specifically, it is veri- sive noise is present. Effective modulation classification in
fied that modulation types can be distinguished through CCES. impulsive noise remains a challenge.
Then, multi-slices are extracted at different cycle-frequencies In order to deal with impulsive noise, Chavali and
from CCES as the original features for AMC. Following the da Silva [7] made use of the composite hypothesis testing
extraction, the principal component analysis is applied to these approach to estimate the impulsive noise parameters. The per-
slices to further optimize the original features. Finally, the radial formance of this LB method depends on the estimation of
basis function neural network is used as a classifier to perform
modulation classification. Monte Carlo simulations demonstrate unknown parameters, while it is difficult to be modeled pre-
that the proposed algorithm outperforms other existing schemes cisely. In [8], Kolmogorov-Smirnov test was used to make a
in impulsive noise cases, especially with a low generalized signal decision to get faster performance in impulsive noise. A vari-
to noise ratio. ational mode decomposition technique was proposed in [9]
to identify the modulation types. Low order wavelet packet
Index Terms—Automatic modulation classification (AMC),
cyclic correntropy spectrum (CCES), impulsive noise. decomposition technique was proposed in [10] to extract char-
acteristic parameters for modulation classification in impulsive
noise. In [11], sparse signal decomposition was employed
to remove the impulsive noise. Then cyclostationary features
I. I NTRODUCTION were extracted for AMC. However, in low GSNR, the above
UTOMATIC modulation classification (AMC) is usu-
A ally utilized as the intermediate process between signal
detection and demodulation to identify the modulation type
methods were unable to perform AMC in a high accuracy, and
their performance still need to be further improved.
Recently, a novel signal analysis concept, named cyclic
of an unknown received signal even with little prior knowl- correntropy (CCE), has just emerged to supress impulsive
edge about the transmitted signal [1]. Over the past few years, noise in [12]–[14]. CCE innovatively combines the cyclosta-
AMC has gained widespread attention due to its wide appli- tionarity method and the concept of correntropy. It has been
cations in many military and civilian fields [2]. Generally, demonstrated that CCE can achieve better performance in the
the typical AMC approaches can be divided into two cate- suppression of impulsive noise than other conventional meth-
gories: likelihood-based (LB) methods and feature-based (FB) ods [14]. Although CCE has advantages in cyclostationarity
methods [1]. Compared with LB methods, FB methods are signal processing, its applications have yet to be developed and
more prominent in the practical implementations [3]. In FB reinforced. To the best of our knowledge, CCE had only been
methods, there are mainly two crucial steps, namely fea- used in frequency estimation and spectrum sensing and never
ture extraction and classification. Some discriminating features been explored for AMC. In this letter, we attempt to extend its
have been adopted in pervious AMC methods. For example, application to AMC in impulsive noise so that the advantage
in [4], different slices were taken from spectral coherence of CCE can be taken to get a promising performance.
function as the features for AMC. Meanwhile, various clas- Inspired by CCE, we propose a novel AMC method
sifiers have also been widely studied, such as support vector based on CCES for impulsive noise environment. In this
machine, genetic algorithm and RBF networks [5]. In these method, at first, the theoretical analysis is carried out to
AMC methods, the wireless communication channel noise is verify that different modulation types have different CCES
often assumed to be AWGN (additive white Gaussian noise). patterns even in impulsive noise environment. Hence, CCES
is introduced into AMC for effectively suppressing impul-
Manuscript received August 29, 2018; accepted October 1, 2018. Date of sive noise. Through CCES, the received signals are trans-
publication October 9, 2018; date of current version April 9, 2019. This work
was supported by the National Natural Science Foundation of China under
formed into the cyclic-frequency domain. Then, multi-slices
Grant 61671105, Grant 61139001, Grant 61172108, and Grant 81241059. The at different cycle-frequencies are extracted from CCES as
associate editor coordinating the review of this paper and approving it for original features of these received signals. Particularly, PCA
publication was R. C. de Lamare. (Corresponding author: Tianshuang Qiu.) is adopted to optimize these original features. Finally, RBF
The authors are with the Faculty of Electronic Information and Electrical
Engineering, Dalian University of Technology, Dalian 116024, China (e-mail:
neural network is employed as a classifier to perform the
mjt@mail.dlut.edu.cn; qiutsh@dlut.edu.cn). modulation classification. The simulation results show that the
Digital Object Identifier 10.1109/LWC.2018.2875001 proposed method yields better performance than other methods
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
MA AND QIU: AMC USING CCES IN IMPULSIVE NOISE 441
V. C ONCLUSION
In this letter, a novel AMC method based on cyclic corren-
tropy spectrum is proposed. It is verified that modulation types
can be distinguished by CCES, and features extracted from
CCES are more effective in AMC, especially for impulsive
noise environment. Besides, PCA algorithm and RBF neu-
ral network are utilized to optimize features and classify the
modulation types respectively. Simulation results indicate that
the proposed method can lead to higher accuracy and robust
Fig. 4. Classification performance in different α. modulation classification in impulsive noise.
R EFERENCES
A. Performance Comparisons in Different GSNR [1] O. A. Dobre, A. Abdi, Y. Bar-Ness, and W. Su, “Survey of auto-
matic modulation classification techniques: Classical approaches and
In this subsection, we conduct the experiment with GSNR new trends,” IET Commun., vol. 1, no. 2, pp. 137–156, Apr. 2007.
in a range from -5dB to 10dB. In addition, 2000 samples are [2] M. W. Aslam, Z. Zhu, and A. K. Nandi, “Automatic modulation clas-
sification using combination of genetic programming and KNN,” IEEE
generated for every individual modulation type in each decibel Trans. Wireless Commun., vol. 11, no. 8, pp. 2742–2750, Aug. 2012.
and there are 2000 symbols in each sample. Then half samples [3] X. Yan, G. Feng, H.-C. Wu, W. Xiang, and Q. Wang, “Innovative robust
are used as training set to train RBF network and the others modulation classification using graph-based cyclic-spectrum analysis,”
are treated as the testing set. Besides, the characteristics expo- IEEE Commun. Lett., vol. 21, no. 1, pp. 16–19, Jan. 2017.
[4] A. Fehske, J. Gaeddert, and J. H. Reed, “A new approach to signal clas-
nent α is set to be 1.8. In Fig. 3(a), the correct recognition sification using spectral correlation and neural networks,” in Proc. 1st
percentage is shown for different modulation types. It is appar- IEEE Int. Symp. New Front. Dyn. Spectr. Access Netw. (DySPAN), 2005,
ent from Fig. 3(a) that the proposed algorithm outperforms the pp. 144–150.
existing algorithms under low GSNR conditions. The correct [5] J. Li, C. He, J. Chen, and D. Wang, “Automatic digital modulation
recognition based on Euclidean distance in hyperspace,” IEICE Trans.
recognition percentage of the proposed algorithm can converge Commun., vol. 89, no. 8, pp. 2245–2248, Aug. 2006.
to 1 when GSNR > 0dB, while the existing algorithm requires [6] Y. Liu, T. Qiu, and H. Sheng, “Time-difference-of-arrival estima-
GSNR to be at least 4dB so as to achieve the same perfor- tion algorithms for cyclostationary signals in impulsive noise,” Signal
mance. Furthermore, the probability of correct classification Process., vol. 92, no. 9, pp. 2238–2247, Sep. 2012.
(Pcc) for the aforementioned four modulation types is illus- [7] V. G. Chavali and C. R. C. M. da Silva, “Classification of digi-
tal amplitude-phase modulated signals in time-correlated non-Gaussian
trated in Fig. 3(b). It shows that our AMC method performs channels,” IEEE Trans. Commun., vol. 61, no. 6, pp. 2408–2419,
better than other contrastive methods. Thus, our algorithm is Jun. 2013.
more suitable for AMC in low GSNR condition. [8] F. Wang and X. Wang, “Fast and robust modulation classification via
Kolmogorov–Smirnov test,” IEEE Trans. Commun., vol. 58, no. 8,
pp. 2324–2332, Aug. 2010.
[9] T. Dutta, U. Satija, B. Ramkumar, and M. S. Manikandan, “A novel
method for automatic modulation classification under non-Gaussian
B. Performance Comparisons in Different α noise based on variational mode decomposition,” in Proc. IEEE 22nd
Fig. 4(a) shows the classification performance of the pro- Nat. Conf. Commun. (NCC), Mar. 2016, pp. 1–6.
posed and existing algorithms with α = 1.3. As shown in [10] Y. Hu, M. Liu, C. Cao, and B. Li, “Modulation classification in alpha
stable noise,” in Proc. IEEE 13th Int. Conf. Signal Process. (ICSP),
Fig. 4(a), the proposed method performs better than other Nov. 2016, pp. 1275–1278.
contrastive methods, especially in low GSNR conditions. In [11] U. Satija, M. Mohanty, and B. Ramkumar, “Cyclostationary features
comparison with Fig. 3(a), both in α = 1.8 and α = 1.3, the based modulation classification in presence of non Gaussian noise using
proposed AMC method achieves higher accuracy than other sparse signal decomposition,” Wireless Pers. Commun., vol. 96, no. 4,
pp. 5723–5741, Oct. 2017.
contrastive methods. In order to further explain the perfor- [12] S. Luan, T. Qiu, Y. Zhu, and L. Yu, “Cyclic correntropy and its spec-
mance of the proposed method, the classification accuracies trum in frequency estimation in the presence of impulsive noise,” Signal
over α in GSNR = 1 dB are shown in Fig. 4(b). Clearly, com- Process., vol. 120, pp. 503–508, Mar. 2016.
pared with other contrastive methods, the proposed method [13] A. I. R. Fontes, J. B. A. Rego, A. de Medeiros Martins, L. F. Q. Silveira,
and J. C. Principe, “Cyclostationary correntropy: Definition and appli-
has higher accuracy, especially in lower characteristic expo- cations,” Expert Syst. Appl., vol. 69, pp. 110–117, Mar. 2017.
nent. Thus, the proposed AMC method can be used to solve [14] T. Liu, T. Qiu, and S. Luan, “Cyclic correntropy: Foundations and
AMC problem in impulsive noise environments. theories,” IEEE Access, vol. 6, pp. 34659–34669, Jul. 2018.
444 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—User traffic of mobile wireless communication is BEL characteristics, typically through building windows, were
rapidly increasing in urban areas. Thus, service cell planning investigated by up to the 2 GHz band [4], [5], and 3.5 GHz
becomes an important issue for efficient radio resource usage. band [6], 5 GHz band [7], [8], 8 GHz band [9], 10 GHz
Since users are also inside high buildings in those areas, it is
important to know the building entry loss (BEL) characteristics band [10], and 38 GHz band [11]. The results showed that
from the outside base stations of various locations for the pur- the BEL characteristics approximately follow these models.
pose. In this letter, we measured the penetration loss of buildings However, in most of the literature, only the horizontal or ver-
with large windows in the 4.9 GHz band by using a system with tical domain characteristics were investigated, and combined
an unmanned aerial vehicle. Our contribution is to clarify the characteristics in the vertical-horizontal domain were investi-
vertical-horizontal BEL characteristics through exhaustive mea-
surements. We also extended the COST 231 BEL model for the gated by using a simulation such as a knife-edge diffraction
vertical-horizontal BEL characteristics, and estimated the model model [12] because of the limited variety of measurement con-
parameters from the measurements. The results showed that the ditions. This limitation comes from the physical difficulties of
vertical-horizontal domain BEL characteristics were quite dif- having to place the BS antenna outside.
ferent, but that they were not independent. The proposed model We thought that exhaustive BEL measurements are needed
improved the root mean squared error of the BEL prediction
by approximately 3 dB. This is expected to be utilized for the to construct the vertical-horizontal domain BEL model to
improvement of cell planning efficiency. enable 3D cell planning. Recently, there has been much
progress in robotic technologies such as unmanned aerial
Index Terms—Building entry loss, outdoor-to-indoor propaga-
tion, propagation loss measurement, unmanned aerial vehicle, vehicles (UAVs). This has attracted interest even in the com-
vertical and horizontal domain loss, window penetration loss. munications research area [13], [14]. We developed a radio
measurement system using a multicopter to solve the physi-
cal limitation problem of BS placement [15]. The contribution
of this letter is to clarify the vertical-horizontal domain pen-
I. I NTRODUCTION etration loss characteristics of buildings with large windows
WING to the widespread use of various application through exhaustive outdoor-to-indoor (O2I) measurements by
O services such as video streaming and cloud services,
user traffic in cellular networks is rapidly increasing. Thus,
using the system. The measurement sites were in front of
a research building and a cafeteria, and the frequency was
for efficient radio resource usage, service cell planning is an the 4.9 GHz band for the low super high frequency (SHF)
important issue. In particular, cell planning becomes complex band communication of the 5G cellular system. The results
in urban areas because many users are also inside high build- showed that the vertical-horizontal domain BEL characteristics
ings, which are sometimes higher than the base stations (BSs). were quite different, but they were not independent. We also
Therefore, three-dimensional (3D) cell planning is considered extended the COST 231 BEL model for the vertical-horizontal
in those areas [1]. domain BEL characteristics by using the effective azimuth
Knowledge of the building entry loss (BEL) characteris- and elevation angles from the window center, and estimated
tics of radio waves is important for this purpose. It is known the model parameters from the measurements. The proposed
that the BEL characteristics change according to the incident model improved the root mean squared error (RMSE) of the
angle of the radio wave to the building. The COST 231 BEL BEL prediction by approximately 3 dB.
model [2] and ITU-R P.2109 model [3] specified the azimuth
and elevation angle characteristics, respectively, and various
II. 3D E XTENSION OF COST 231 BEL M ODEL
radio measurements were conducted for the validation. The
The COST 231 BEL model [2] is the baseline model for
Manuscript received August 18, 2018; accepted October 3, 2018. Date of this letter. In the model, the propagation loss Ltotal from the
publication October 9, 2018; date of current version April 9, 2019. This work
was supported in part by the Fujikura Foundation and in part by the Support outdoor BS to the indoor mobile station (MS) is modeled by
Center for Advanced Telecommunications Foundation. The associate editor the outdoor path loss Lout , the indoor path loss Lin , and the
coordinating the review of this paper and approving it for publication was building penetration loss Ltw as follows:
W. Zhang. (Corresponding author: Kentaro Saito.)
The authors are with the School of Environment and
Society, Tokyo Institute of Technology, Tokyo 152-8550, Japan Ltotal = Lout + Lin + Ltw (1)
(e-mail: saitouken@tse.ens.titech.ac.jp; fan.q.aa@m.titech.ac.jp; Lout = 32.44 + 20log(f ) + 20log(S + d ) (2)
keerativoranan.n.aa@m.titech.ac.jp; takada@ide.titech.ac.jp).
Digital Object Identifier 10.1109/LWC.2018.2875003 Lin = αd (3)
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
SAITO et al.: VERTICAL AND HORIZONTAL BEL MEASUREMENT IN 4.9 GHz BAND BY UAV 445
Fig. 2. Measurement maps and photos of (a) research building (map), (b) research building (outdoor view), (c) research building (indoor view), (d) UAV
station, (e) cafeteria (map), (f) cafeteria (outdoor view), and (g) cafeteria (indoor view).
TABLE I
M EASUREMENT PARAMETERS loss at the perpendicular angle We ranged from 5 dB to 6 dB,
and these results corresponded to those in [2] and [7] for a
glass window case. Although the azimuth angular dependent
loss coefficient WGe,azm was similar to that in [7] for the
research building environment, it was 0 dB in the cafeteria
environment. This is because the cafeteria was a glass-walled
circular room that did not have a specific azimuth angular
dependency. The elevation angular dependent loss coefficient
WGe,elv was larger than WGe,azm in our measurements. The
reason is thought to be that the window was horizontally long
and caused a more significant diffraction loss in the verti-
cal plane. On the other hand, these BEL characteristics could
not be modeled by a single parameter WGe of the COST
231 model. As a result, We tended to increase to compensate
for the large estimation error of the angular characteristics.
Fig. 3(c) shows the cumulative distribution function (CDF) of
the modeling error. In the proposed method, the modeling error
was significantly improved from that of the COST 231 model.
The median error was improved by approximately 3 dB.
The measurement results showed that the vertical-horizontal
domain BEL characteristics were different, and that the
horizontal domain characteristics, the receiving power changed proposed 3D BEL model improved the modeling accuracy
more significantly in the high altitude area compared to the compared to the COST 231 model.
low altitude area, as described in Section II. A similar trend
was observed in the vertical domain profile. In Fig. 3(b), these
characteristics were modeled well in the proposed model by V. S UMMARY
introducing the effective incident angles in both domains. In this letter, we presented vertical-horizontal BEL mea-
All estimated model parameters are summarized in Table II. surement results from the developed system using an UAV.
The 5.25 GHz band measurement result of another litera- The 4.9 GHz band penetration loss characteristics of build-
ture [7] is also shown. In the proposed model, the penetration ings with large windows were clarified in research building
SAITO et al.: VERTICAL AND HORIZONTAL BEL MEASUREMENT IN 4.9 GHz BAND BY UAV 447
TABLE II
M ODEL PARAMETER E STIMATION R ESULTS
Fig. 3. Measurement Results:(a) receiving power profile (research building, measurement data), (b) receiving power profile (research building, proposed
model), and (c) modeling error CDF.
and cafeteria environments. The results showed that the verti- [6] K. L. Chee, A. Anggraini, T. Kaiser, and T. Kürner, “Outdoor-to-indoor
cal domain loss was more significant because of the window propagation loss measurements for broadband wireless access in rural
areas,” in Proc. 5th Eur. Conf. Antennas Propag. (EUCAP), Apr. 2011,
shapes. In addition, the horizontal domain loss varied more pp. 1376–1380.
drastically as the difference between the BS height and the [7] M. Alatossava, E. Suikkanen, J. M. H. Veli-Matti, and J. Ylitalo,
window height decreased. We also extended the COST 231 “Extension of COST 231 path loss model in outdoor-to-indoor envi-
ronment to 3.7 GHz and 5.25 GHz,” in Proc. Int. Symp. Wireless Pers.
BEL model for the vertical-horizontal BEL characteristics and Multimedia Commun., Sep. 2008, pp. 1–4.
estimated the model parameters from the exhaustive measure- [8] J. Medbo, J. Furuskog, M. Riback, and J.-E. Berg, “Multi-frequency
ments. The median modeling error was improved by 3 dB by path loss in an outdoor to indoor macrocellular scenario,” in Proc. 3rd
Eur. Conf. Antennas Propag., Mar. 2009, pp. 3601–3605.
our proposal. The further extension of the BEL model such [9] H. Okamoto, K. Kitao, and S. Ichitsubo, “Outdoor-to-indoor prop-
as the indoor angular characteristics and the detailed angular agation loss prediction in 800-MHz to 8-GHz band for an urban
profile shapes are the future work. This letter is expected to area,” IEEE Trans. Veh. Technol., vol. 58, no. 3, pp. 1059–1067,
Mar. 2009.
be utilized in 3D service cell planning for future work. [10] A. Roivainen, V. Hovinen, N. Tervo, and M. Latva-Aho, “Outdoor-
to-indoor path loss modeling at 10.1 GHz,” in Proc. 10th Eur. Conf.
Antennas Propag. (EuCAP), Apr. 2016, pp. 1–4.
R EFERENCES [11] T. Imai et al., “Outdoor-to-Indoor path loss modeling for 0.8 to 37 GHz
[1] H. Omote, M. Miyashita, and R. Yamaguchi, “Measurement of time- band,” in Proc. 10th Eur. Conf. Antennas Propag. (EuCAP), Apr. 2016,
spatial characteristics between indoor spaces in different LOS buildings,” pp. 1–4.
in Proc. Int. Symp. Antennas Propag. (ISAP), Nov. 2015, pp. 1–4. [12] I. Rodriguez et al., “A novel geometrical height gain model for line-
[2] COST Action 231—Digital Mobile Radio Towards Future Generation of-sight urban micro cells below 6 GHz,” in Proc. Int. Symp. Wireless
Systems—Final Report. Luxembourg City, Luxembourg: Office Official Commun. Syst. (ISWCS), Sep. 2016, pp. 393–398.
Publ. Eur. Commun., 1999. [13] D. W. Matolak and R. Sun, “Unmanned aircraft systems: Air-ground
[3] “Prediction of building entry loss,” Int. Telecommun. Union, Geneva, channel characterization for future applications,” IEEE Veh. Technol.
Switzerland, ITU-Recommendation P.2109, 2017. [Online]. Available: Mag., vol. 10, no. 2, pp. 79–85, Jun. 2015.
https://www.itu.int/rec/R-REC-P.2109/en [14] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communi-
[4] Y. L. C. de Jong, M. H. J. L. Koelen, and M. H. A. J. Herben, cations with unmanned aerial vehicles: Opportunities and
“A building-transmission model for improved propagation prediction challenges,” IEEE Commun. Mag., vol. 54, no. 5, pp. 36–42,
in urban microcells,” IEEE Trans. Veh. Technol., vol. 53, no. 2, May 2016.
pp. 490–502, Mar. 2004. [15] K. Saito, Q. Fan, N. Keerativoranan, and J. Takada, “Outdoor-to-
[5] D. I. Axiotis and M. E. Theologou, “An empirical model for predicting indoor radio propagation loss measurement by using an unmanned
building penetration loss at 2 GHz for high elevation angles,” IEEE aerial vehicle in 4.9 GHz band,” in Proc. IEICE TC-SRW, Aug. 2017,
Antennas Wireless Propag. Lett., vol. 2, pp. 234–237, 2003. pp. 25–30.
448 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—The emergence of Internet of Things applications User Equipment (UE) terminals intending to join RA. The BS
introduces new challenges such as massive connectivity and small broadcasts a parameter (e.g., probability) such that UEs can
data transmission. In traditional data transmission protocols, an perform RA probabilistically. In dynamic resource allocation,
ID (i.e., IP address or MAC address) is usually included in a the BS can dynamically allocate resources of RA channel and
packet so that its receiver is able to know who sent the packet.
However, this introduces the big head-small body problem for
data channel. It also derives an optimal trade-off problem to
light payload. To address this problem, the Hint protocols have maximize the Machine-to-Machine (M2M) throughput. In slot-
been proposed. The main idea is to “encode” information in a ted access, each Machine-Type Communication (MTC) device
tiny broadcast Hint message that allows devices to “decode” their is assigned a dedicated RA slot to access RA. The dilemma
transmission slots. Thus, it can significantly reduce transmission is that short RA cycles may lead to collision while long RA
and contention overheads. In this letter, we extend eHint to sup- cycles may lead to long delay.
port multi-slot data transmissions. Several efficient protocols are To address the big head-small body problem, a series of
proposed. Our simulation results validate that the protocols can Hint protocols [8]–[10], which remove IDs from data packets,
significantly increase the number of successfully transmitted
devices, channel utilization, and payload of transmitted devices have been proposed. In these protocols, a tiny Hint message is
compared with eHint. broadcast to allow IoT devices to decode their assigned trans-
mission resources. Interestingly, the assigned location of these
Index Terms—Internet of Things (IoT), machine-to-machine resources also imply the sender’s ID, thus eliminating a large
(M2M) communication, multi-slot allocation, random access,
wireless networks.
part of packet header. Therefore, even the device’s ID is not
transmitted, its receiver (usually a BS) is still able to know its
I. I NTRODUCTION identity. In [8], we proposed a set of Hint-based frameworks
for small data transmission. Later, a Chinese remainder theo-
NTERNET of Things (IoT) traffic characterized by small
I data transmission introduces new challenges to research
community. There are various types of IoT devices rang-
rem based Hint protocol [9] is applied to LTE-A networks
to reduce the signaling cost in random access procedures.
However, the both Hint protocols [8], [9] are based on a
ing from small tags, sensors, to complicated actuators and strong assumption that the small data has the same size and
machines [1]. Statistics show that 50% of IoT packets are is carried by the same size channel resource, i.e., a time
less than 100 bytes [2]. Collecting small data from such mas- slot. To release the assumption, the Hint protocols are then
sively connected IoT devices introduces new challenges. In enhanced by supporting multi-slot data transmission in [10].
most protocol designs, an ID (e.g., IP address or MAC address) However, when the number of transmitting devices is large, the
is included in a packet so that its receiver is able to identify multi-slot Hint protocol requires intensive computation capac-
the sender. When IoT data is small, however, the overhead of ity to find a satisfactory seed, which may lead to transmission
its ID becomes relatively large, leading to the big head-small latency. To address this issue, in this letter we use a novel
body problem. iterative approach to reduce computation overhead and enable
Connection-oriented network architectures have been stud- more devices to transmit. Through extensive simulations, we
ied in [3]–[6]. When a device intends to send data to the Base demonstrate that the proposed iterative approach significantly
Station (BS), a connection is established by using the Random outperforms eHint [10] in terms of the number of successfully
Access (RA) procedure. As the number of devices increases, transmitted devices, channel utilization, and the total payload
it may cause significant collision in RA [7], leading to a long of transmitted devices.
delay. To solve the problem, several solutions have been pro-
posed, including Access Class Barring (ACB) [3], dynamic
resource allocation [4], slotted access [5], deep learning II. M ULTI -S LOT A LLOCATION P ROBLEM
scheduling [6]. The idea of ACB is to control the number of We consider a set D = {d1 , d2 , . . . , dm } of m IoT devices
Manuscript received September 5, 2018; accepted October 4, 2018. Date
covered by a BS. Each IoT device di , i = 1..m, needs to
of publication October 11, 2018; date of current version April 9, 2019. The report its data at a regular pattern to the BS. Our goal is to
associate editor coordinating the review of this paper and approving it for allocate radio resources to D to transmit data to the BS. We
publication was M. Dong. (Corresponding author: Tsung-Yen Chan.) make the following assumptions:
T.-Y. Chan, Y.-C. Tseng, and J.-C. Chen are with the Department
of Computer Science, College of Computer Science, National Chiao
1) The value of m is quite large.
Tung University, Hsinchu 30010, Taiwan (e-mail: tychan@cs.nctu.edu.tw; 2) Each di switches between two modes, active and sleep.
yctseng@cs.nctu.edu.tw; jcc@cs.nctu.edu.tw). When intending to transmit data, di goes to the active
Y. Ren was with the Department of Computer Science, National Chiao mode; otherwise, it switches to the sleep mode.
Tung University, Hsinchu 30010, Taiwan. He is now with the School of
Computing Science, University of East Anglia, Norwich NR4 7TJ, U.K.
3) The active pattern of di is denoted by a binary period-
(e-mail: e.ren@uea.ac.uk). ical function Pi (t), where t is time (by frame count).
Digital Object Identifier 10.1109/LWC.2018.2875455 Pi (t) = 1 if di intends to transmit data at the t-th frame;
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
CHAN et al.: MULTI-SLOT ALLOCATION PROTOCOLS FOR MASSIVE IoT DEVICES WITH SMALL-SIZE UPLOADING DATA 449
where v is a vector. The length of v is |v | ≥ di ∈M (t) ni ,
where M (t) = {di |Pi (t) = 1}. The i-th element of v (resp.,
Alloc), is denoted by v[i] (resp., Alloc[i]). The value of v[i]
falls in {0, 1, 2}. v is for devices to decode whether they can
transmit or not. If it is safe to transmit in ni continuous slots
in Alloc, the corresponding elements in v will be: “1 2 ·
· · 2”,
ni −1
where “1” means “starting slot” and “2” means “continuous
slot”. If it is not safe to transmit or a slot is not assigned to
any device, a “0” is used.
Next, we review the VF briefly:
Fig. 1. The proposed frame structure.
1) The BS randomly picks a seed s and computes
h(s, di ) mod |v | for each device di ∈ M (t). Next,
the BS selects a subset M ⊆ M(t) of devices that can
otherwise, Pi (t) = 0. For example, if di is attached
correctly decode v for safe transmission.
to a temperature sensor which needs to report in every
2) The BS executes Step 1 a few times with different seeds
3 minutes, Pi (t) has a period of 3 minutes. If one more
and then selects the s from the iteration leading to the
humidity sensor which needs to report in every 5 minutes
largest number of safe transmission slots.
is attached to di , Pi (t) is a function of the combina-
3) After selecting, the BS encodes v.
tion of two periodical functions with periods = 3 and
4) The BS broadcasts <s, v> in Bcast.
5 minutes, respectively.
5) Each device di ∈ M (t) checks whether it is allowed to
4) Whenever Pi (t) = 1, device di needs to send ni slots
transmit or not.
of data to the BS at t. Note that ni is also quite small
6) If di ∈ M (t) can transmit safely, it uploads its data
(such as less than 3 or 5 slots).
at the corresponding Alloc. Otherwise, di can use
5) When our Hint protocol starts, the BS is informed of its
random access to contend for transmission in Rand.
Pi (t) and ni , for each di ∈ D.
Discussion: In VF, some devices may not receive slots for
To solve the multi-slot allocation problem, the communica-
transmission due to collision in v. Given a fixed v, more
tion channel is divided into a sequence of fixed-length frames.
devices may lead to higher collisions. This not only reduces
Each frame consists of three parts: (1) Broadcast (Bcast):
transmission opportunities in Alloc, but also lacks of flex-
It is for the BS to broadcast and announce resource alloca-
ibility. These shortcomings motivate us to design two new
tion information (i.e., Hint) to devices. (2) Allocated (Alloc):
protocols, 2VF and IVF.
It consists of multiple slots for uplink data transmission (no
transmitter’s ID required in our case). (3) Random (Rand): It
is for any unscheduled/unpredicted transmission not arranged B. Two Virtual Frame (2VF)
in Alloc. Our goal is to design an efficient access protocol
for such transmission. To reduce computation overhead and enable more devices to
Fig. 1 shows the frame structure. In this example, the active transmit in Alloc, the 2VF protocol divides Alloc into two
periods of d1 , d2 , and d3 are T1 , T2 , and T3 , respectively. parts, Alloc_1 and Alloc_2, and uses two seeds s1 and s2
Their requirements are n1 = 2, n2 = 1, and n3 = 3 slots, to achieve this goal. Two vectors v1 and v2 will be used. Also,
respectively. At frame t, all devices will transmit. At t + 1, the length of v1 and v2 are set to |v1 | and |v2 |, respectively,
only d1 will transmit. At t + 2, d1 and d2 will transmit. The where |v1 | ≥ di ∈M (t) ni and |v2 | ≥ di ∈M (t)−M ni .
BS will schedule their transmissions in Alloc, through the It works as follows:
announcement in Bcast. Details will be discussed later. For 1) The BS repeats the following steps a few times.
exceptions (such as transmission errors or emergency traffics), • Randomly pick a seed s1 and compute
devices can use Rand. h(s1 , di ) mod |v1 | for each device di ∈ M (t).
• Select a subset M ⊆ M(t) of devices such that
III. P ROPOSED P ROTOCOLS only devices in M can correctly decode v1 for safe
transmission.
The main idea of the Hint protocol is to arrange a com- • A seed s2 is randomly picked. Compute
mon function shared by the BS and all devices. The function h(s2 , di ) mod |v2 | for each device di ∈ M (t)−M .
takes two inputs: (i) a small piece of information computed • Select a subset M ⊆ M(t) − M’ of devices such
by the BS, and (ii) a device’s ID. The BS will broadcast the that only devices in M can correctly decode v2 for
information computed in (i) by using Bcast. Each device safe transmission.
then can use the broadcast information and its ID to retrieve 2) The BS then selects the s1 and s2 from the iteration
the transmission slots allocated to it through the function. The in Step 1, leading to the largest number of safe trans-
broadcast Hint is thus more efficient than typical 1-by-1 noti- mission slots. (Note: once s1 and s2 are selected, the
fications. Next, we first review the VF proposed in [10]. We corresponding M and M are also selected.)
then propose two more efficient protocols called 2VF and IVF. 3) v1 and v2 are encoded as follows:
• (Collision-free) For each di ∈ M such that k =
A. Review: Virtual Frame (VF) h(s1 , di ) mod |v1 |, set v1 [k ] = “1” and v1 [k +
Let h(s, ID) be a hash function which takes a device ID and 1 : k + ni − 1] = “2 ·· · 2”. For each di ∈ M such
a seed s as inputs. In frame t, the Hint message = <s, v>, ni −1
450 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Fig. 4. The different effects of the number of successfully transmitted devices, |Alloc|, and CU with different parameters.
2) Upon receiving Bcast, a device di with Pi (t) = 1 can payload to transmit. The values of CU gradually converge for
transmit in one of the l spaces (Alloc_1, Alloc_2, VF, 2VF, and IVF with the increase of |M(t)|.
. . . , Alloc_l ) according to its hash results h(sk , di ) Overall, Figs. 4(a)-(f) demonstrate that the iteration
and vectors vk . If it is not allowed to transmit in Alloc, based Hint protocol has much better performance compared
it can contend in Rand. with [10] in terms of the number of successfully transmitted
devices, the total payload of transmitted devices, and channel
utilization.
IV. P ERFORMANCE E VALUATION
We evaluate 2VF and IVF through simulations in terms of V. C ONCLUSION
three metrics, the number of successfully transmitted devices In this letter, we propose two advanced Hint protocols, 2VF
in M(t), the length of Alloc (i.e., |Alloc|), and Channel and IVF, which enable more devices to upload their data in
Utilization (CU). The results are compared with VF and tra- a collision-free manner. Compared to the eHint [10], the pro-
ditional polling protocol which collects data from each device posed 2VF and IVF provide more flexibility and address the
one-by-one. Each slot size is assumed to be 32 bits and the issue when the number of devices intending to transmit is
length of s is set to 16 bits. Device address is assumed to be large. We demonstrate, through extensive simulations, that the
64 bits. The value of ni is small and it randomly falls in the proposed 2VF and IVF outperform both eHint [10] and tra-
range of [1∼3] or [1∼5] slots. ditional polling protocol in terms of CU and slot allocation
payload
The CU is defined as: CU = payload+ packet header . For capability. As to future work, designing a smart seed with
|Alloc| low computation cost is worth of exploring.
2VF and IVF, CU =
|Alloc| +2×16+2×(|v1 |+|v2 |)
, and
|Alloc|
CU = |Alloc|+l×16+2×(|v |+|v |+···+|v |) , respectively. For
1 2 l
|Alloc| R EFERENCES
VF, CU = |Alloc|+16+2×|v | . For traditional polling proto-
[1] H. Li, K. Ota, and M. Dong, “Energy cooperation in battery-free wireless
|Alloc|
cols, CU = |Alloc|+64×|M (t)| . For IVF, we execute VF until communications with radio frequency energy harvesting,” ACM Trans.
Embedded Comput. Syst., vol. 17, no. 2, p. 44, 2018.
all the devices are allocated in Alloc successfully. [2] W. John and S. Tafvelin, “Analysis of Internet backbone traffic and
Fig. 4(a) and Fig. 4(b) demonstrate the number of success- header anomalies observed,” in Proc. ACM SIGCOMM Conf. Internet
fully transmitted devices in M(t) with ni = [1 ∼ 3] and Meas., 2007, pp. 111–116.
[1∼5], respectively. Both the figures show that IVF and 2VF [3] Z. Wang and V. W. S. Wong, “Optimal access class barring for stationary
machine type communication devices with timing advance informa-
outperform VF significantly. Considering Fig. 4(b) as an exam- tion,” IEEE Trans. Wireless Commun., vol. 14, no. 10, pp. 5374–5387,
ple, when |M(t)| = 200, IVF and 2VF gain 44% and 30% Oct. 2015.
transmitted devices compared with VF, respectively. [4] D. T. Wiriaatmadja and K. W. Choi, “Hybrid random access and data
Fig. 4(c) and Fig. 4(d) demonstrate the impacts of |M(t)| transmission protocol for machine-to-machine communications in cellu-
lar networks,” IEEE Trans. Wireless Commun., vol. 14, no. 1, pp. 33–46,
on |Alloc| by varying ni = [1 ∼ 3] and [1∼5], respectively. Jan. 2015.
We observe that both 2VF and IVF outperform VF with large [5] “Study on RAN improvements for machine-type communications,
margins, and IVF performs the best. Specifically, |Alloc| (Release 11),” 3GPP, Sophia Antipolis, France, Rep. TR 37.868,
increases as the number of devices intending to transmit grows. Sep. 2011.
[6] H. Li, K. Ota, and M. Dong, “Learning IoT in edge: Deep learning for
We also observe that the three curves are linear. This means the Internet of Things with edge computing,” IEEE Netw., vol. 32, no. 1,
that the proposed protocols are scalable and resilient to the pp. 96–101, Jan./Feb. 2018.
increasing of |M(t)|. [7] “Study on new radio (NR) access technology,” 3GPP, Sophia Antipolis,
Fig. 4(e) and Fig. 4(f) show the impacts of |M(t)| on CU France, Rep. TR 38.912, 2017.
[8] Y. Ren, R.-J. Wu, T.-W. Huang, and Y.-C. Tseng, “Give me a hint: An
when ni = [1 ∼ 3] and [1∼5], respectively. Overall, the Hint ID-free small data transmission protocol for dense IoT devices,” in Proc.
protocols (i.e., VF, 2VF, and IVF) outperform the traditional IEEE Wireless Days, 2017, pp. 121–126.
protocol significantly since we use a Hint message instead of [9] T.-W. Huang, Y. Ren, K. C.-J. Lin, and Y.-C. Tseng, “r-Hint: A message-
notifying devices by 64 ×|M(t)| times. The CU of the tradi- efficient random access response for mMTC in 5G networks,” in Proc.
IEEE PIMRC, 2017, pp. 1–6.
tional scheme is around 0.5 and 0.6 in Fig. 4(e) and Fig. 4(f), [10] T.-Y. Chan, Y. Ren, Y.-C. Tseng, and J.-C. Chen, “eHint: An efficient
respectively, when |M(t)| increases. The traditional protocol protocol for uploading small-size IoT data,” in Proc. IEEE WCNC, 2017,
with [1∼3] has higher CU than [1∼5] since [1∼5] has more pp. 1–6.
452 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—This letter analyzes the sum-rate of a heterogeneous only a natural consequence. Massive MIMO FD is studied
network with wireless backhaul supported by low-resolution by [2] in a decode-and-forward relay where it is shown that SI
analog-to-digital converters (ADCs) quantized full-duplex mas- effects diminish with a large number of antennas. An SI-aware
sive multiple-input multiple-output. Communication is achieved ZF precoder is proposed for massive MIMO FD in [5].
in two phases. First phase: the macro-cell base station (BS) Recently, massive MIMO has been introduced into HetNets
is equipped with massive receive antennas and a few transmit to provide wireless backhaul support. A HetNet describes a
antennas, and the small-cell BSs deploy massive receive anten- system where a macro-cell (MC) is overlaid with low-powered
nas and a single transmit antenna. Second phase: the roles of the small-cells (SCs) serving stationary UTs. By densifying the
antennas are switched using a circulator. We derive closed-form
network, the distance to BSs is lessened leading to reduced
expressions for the uplink/downlink rates by assuming imperfect
channel state information. It is shown that the quantization noise
path losses, and improved SE [6]. Sanguinetti et al. [7] studied
(QN) due to the use of low-resolution ADCs, degrades the rate. interference management of a massive MIMO-enabled back-
This rate loss is efficiently minimized by using massive receive haul in a two-tier HetNet. Anokye et al. [8] analyzed the
antennas in the first phase. However, in the second phase, the QN sum-rate of a two-tier HetNet where the backhaul is supported
is of the same order as the desired signal. Therefore the massive by FD massive MIMO in both the MC BS and the SC BSs.
number of transmit antennas is unable to effectively suppress The use of a high number of receive antennas in massive
the QN. MIMO leads to significant increase in the power consump-
tion due to the high-resolution analog-to-digital converters
Index Terms—Massive multiple-input multiple-output, full-
(ADCs). To elucidate this, an ADC with d-bit resolution and
duplex, heterogeneous network, quantization noise.
sampling frequency f performs 2d · f computations per sec-
ond. This implies that the power increases linearly with the
I. I NTRODUCTION sampling frequency and exponentially with the resolution.
ASSIVE multiple-input multiple-output (MIMO), full- Thus, if high-resolution ADCs are employed, massive MIMO
M duplex (FD) and heterogeneous networks (HetNets)
have attracted attention as enabling technologies for next-
would be financially infeasible. It is therefore imperative to
research quantized massive MIMO where the receivers use
generation communication. In massive MIMO, base stations low-resolution ADCs. Meanwhile, the aforementioned papers
(BSs) are equipped with antennas such that the number of do not consider the impact of quantization noise (QN) which
antennas is much higher than the user terminals (UTs) [1]. occurs due to low-resolution ADCs. Fan et al. [9] studied
The degree of freedom offered by the massive antennas excel- uplink (UL) rate of quantized HD massive MIMO by assum-
lently mitigates fast fading, non-coherent interference and ing perfect channel state information (CSI). To extend [9], [10]
noise while improving spectral efficiency (SE) [1]. Also, with assumes imperfect CSI in Rician channels. FD massive MIMO
massive MIMO, simple processing such as maximum ratio with low-resolution ADCs in an amplify-and-forward relay
combining/transmission (MRC/MRT) achieves high SE [1]. was studied in [11].
Of paramount interest also is in-band FD where BS This letter studies a HetNet where the backhaul is supported
transceivers simultaneously transmit and receive on the same by low-resolution quantized FD massive MIMO. The MC BS
frequency band [2]. Thus FD systems can double the SE possesses a dedicated wired backbone connection but supports
of half-duplex (HD) at the expense of self-interference (SI). the SCs through FD wireless links. Data communication is
Multi-antenna techniques such as minimum mean square error achieved in two phases. During the first phase, the MC BS
(MMSE), zero-forcing (ZF) and null space projection have deploys a large number of receive antennas and a few trans-
been applied to cancel SI [3], [4]. Considering the advantages mit antennas. The SC BSs have massive receive antennas and
of massive MIMO and FD, combining both technologies is a single transmit antenna. For the second phase, the roles of
the receive and transmit antennas are switched using a circula-
Manuscript received September 12, 2018; accepted October 6, 2018. Date tor [12]. This configuration enables the BSs to serve UL UTs
of publication October 15, 2018; date of current version April 9, 2019. This during the first phase and downlink (DL) UTs in the second
work was supported in part by the IITP through the Korea Government phase. A similar scenario was studied in [8] but the influence
(MSIT) under Grant 2018-0-00218, and in part by the NRF through Korea of QN was ignored. Here, the aim is to equip receive antennas
Government (MSIT) under Grant NRF-2016R1C1B2011921 and Grant NRF- with low-resolution ADCs and then study the combined effects
2018R1D1A1B07049824. The associate editor coordinating the review of this
paper and approving it for publication was W. Hamouda. (Corresponding of SI, and QN. Analytic solutions are derived for the UL/DL
author: Kyoung-Jae Lee.) rate and the impact of low-resolution ADCs is characterized.
P. Anokye, R. K. Ahiadormey, and K.-J. Lee are with the Department
of Electronics and Control Engineering, Hanbat National University,
Daejeon 34158, South Korea (e-mail: princemcanokye@yahoo.com; II. S YSTEM M ODEL
rogerkwao@gmail.com; kyoungjae@hanbat.ac.kr).
C. Song is with the Department of Electronic Engineering, Korea
Consider a two-tier HetNet where an MC is overlaid with
National University of Transportation, Chungju 27469, South Korea (e-mail: K SCs [8]. For the first phase, the MC BS is equipped with
c.song@ut.ac.kr). massive receive antennas Mrx and a few transmit antennas
Digital Object Identifier 10.1109/LWC.2018.2875907 Mtx fixed such that Mtx = K (i.e., Mrx Mtx ). This allows
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ANOKYE et al.: ON SUM-RATE OF HETEROGENEOUS NETWORKS WITH LOW-RESOLUTION ADC QUANTIZED FD MASSIVE MIMO-ENABLED BACKHAUL 453
TABLE I
the MC BS to send independent streams to the SCs. Each SC α FOR D IFFERENT d-B IT ADC R ESOLUTION
BS deploys massive receive antennas Nrx and a single transmit
antenna. In the second phase, the antenna roles are switched,
i.e., the MC BS now has massive transmit antennas Mtx and
a few receive antennas Mrx = K and the SC BSs possess
massive transmit antennas Ntx and a single receive antenna. the outputs of the MC BS and the SC BSs are written,
The received signals are quantized using low-resolution ADCs. respectively, as
For a coherence time T, τ (< T) slots are assigned for channel
estimation and the remaining slots are used for the first and K
(1) √ √
second phases. Assuming a time division duplex protocol, the yq,k = ε1 ps ĥH
k hk xk + ε1 ps ĥH
k hj xj
same channel estimate employed for decoding in the first phase j =1,j =k
is used to precode in the second phase. √
+ ε1 pm ĥH H H
k Qs + ε1 ĥk n + ĥk nq , (5)
(1) (1)
A. First Phase rq,k = ĝkH rq,k , (6)
Each MC BS transmit antenna sends data to a corresponding
SC BS in the DL and the k-th SC BS transmits to the MC BS where ĥk , and ĝk are the estimates of hk , and gk , respectively.
in the UL, simultaneously. The MC BS receives the desired
signal plus its own transmit signals causing SI. The k-th SC B. Second Phase
BS receives the desired signal from the MC BS together with The MC BS precodes data and transmits to the K SC BSs.
its own signal and unintended transmissions from other SC Simultaneously, the k-th SC BS precodes data and sends to
BSs, i.e., SI and SC-to-SC interference. The received signals the k-th MC BS receive antenna. Assuming MRT, the received
at the MC BS and k-th SC BS are, respectively, given by [8]
signals at the MC BS and the k-th SC BS are, respectively,
√ √
y(1) = ps Hx + pm Qs + n, (1) represented by
K K
(1) √ √ √ (2) √ H √ H
rk = pm gk sk + pm gj sj + ps qs,k xk yk = ρs,k gk ĝk xk + ρs,j gk ĝj xj
j =1,j =k j =1,j =k
√
K + ρm zm,k Ĥs + nk , (7)
√
+ ps qc,kj xj + vk , (2) (2) √
K
√ √
j =1,j =k rk = ρm hH
k ĥk sk + ρm hH
k ĥj sj + ρs,k zs,k ĝk xk
j =1,j =k
where H = [h1 , . . . , hK ] ∈ C Mrx ×K , gk ∈ C Nrx ×1 , Q ∈ K
C Mrx ×Mtx , qs,k ∈ C Nrx ×1 , and qc,kj ∈ C Nrx ×1 denote the √
+ ρs,j zc,kj ĝj xj + vk , (8)
SC BS to the MC BS channel, MC BS to the k-th SC BS chan- j =1,j =k
nel, SI channel at the MC BS, SI channel at the k-th SC BS,
and the SC-to-SC interference channel from the j-th SC BS where zm,k ∈ C 1×Mtx , zs,k ∈ C 1×Ntx , and zc,kj ∈ C 1×Ntx
to the k-th SC BS, respectively. x = [x1 , . . . , xK ]T ∈ C K ×1 , indicate the SI channel at the k-th receive antenna of the
s = [s1 , . . . , sK ]T ∈ C K ×1 , n ∈ C Mrx ×1 , vk ∈ C Nrx ×1 , MC BS, SI channel at the k-th SC BS, and SC-to-SC inter-
pm , and ps indicate the transmit signals from the SC BSs, ference channel from the j-th SC BS to the k-th SC BS,
the MC BS transmit signals, noise at the MC BS, noise at respectively. Again, we have Ĥ = [ĥ1 , . . . , ĥK ], ρm =
the k-th SC BS, the transmit power of the MC BS, and the pm K /E{tr (ĤĤH )}, and ρs,k = ps /E{tr (ĝk ĝkH )}. The ele-
SC BSs, respectively. The elements of n, vk , x, and s are 2 ), CN (0, η 2 ),
ments of zm,k , zs,k , and zc,kj are CN (0, ηm,k s,k
modeled by CN (0, 1). The channels hk ∼ CN (0, βk IMrx ), 2 ), respectively. After quantization, the signals
and CN (0, ηc,kj
and gk ∼ CN (0, ζk INrx ) consider both small-scale and large- at the MC BS and the k-th SC BS are given, respectively, as
scale fading where βk , and ζk describe the large-scale fading
(2) (2)
which are assumed to remain unchanged over a coherence yq,k = ε2 yk + nq,k , (9)
interval. The above channels depend on the favorable prop- (2) (2)
agation assumption [2]. The elements of Q, qs,k , and qc,kj rq,k = θ2 rk + vq,k , (10)
are distributed as CN (0, σm 2 ), CN (0, σ 2 ), and CN (0, σ 2 ),
s,k c,kj ε2 , and θ2 indicate ADC resolution at the MC BS, and the
respectively [8]. SC BSs, respectively. nq,k , and vq,k denote the QN at the
For tractability, we assume the additive quantization noise MC BS, and the k-th SC BS with covariances Nq,k = ε2 (1 −
model [9]. After quantizing, the signals at the MC BS, and (2) (2)
the k-th SC BS are, respectively, expressed as ε2 )E{|yk |2 }, and Vq,k = θ2 (1−θ2 )E{|rk |2 }, respectively.
(1)
yq = ε1 y(1) + nq , (3) C. Channel Estimation
(1) (1) Each MC BS transmit antenna sends mutually orthogonal
rq,k = θ1 rk + vq,k , (4) pilot signal to the corresponding SC BS for the DL chan-
where ε1 = 1 − α, and θ1 = 1 − α indicate the ADC res- nel estimation and the SC BSs send orthogonal sequences of
olution at the MC BS, and SC BSs, respectively. nq , and equal length to the MC BS for the UL channel estimation.
vq,k denote the QN of the MC BS, and the k-th SC BS This dictates that the pilot length τ ≥ 2K [8].1 By using
with covariances Nq = ε1 (1 − ε1 )diag(E{y(1) y(1)H }), and 1 During the channel estimation phase, the SC BSs can transmit their mutu-
(1) (1)H
Vq,k = θ1 (1 − θ1 )diag(E{rk rk }), respectively. α val- ally orthogonal pilot sequences to the MC BS while the MC BS keeps silent
and vice-versa as in [11]. Another technique is that both the MC BS and the
ues for a d-bit resolution
√ ADC are given in Table I. For SC BSs transmit the pilot signals simultaneously as explained in [2]. However,
d > 5, α = π 2 3 · 2−2d [9]. Using an MRC decoder, both approaches achieve equal performance.
454 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
MMSE estimation, the k-th SC BS to MC BS true channel is the receive antennas fixed at Mrx = K . For the k-th SC BS,
decomposed as hk = ĥk + h̃k , where h̃k ∼ CN (0, β̃k2 IMrx ) Ntx → ∞ and a single receive antenna are considered.
denotes the error vector for ĥk ∼ CN (0, β̂k2 IMrx ) with Theorem 1: With the MRC receive filter in the first phase
ε τ p β2 and under the assumption of imperfect CSI, the UL/DL rates
β̂k2 = 1+τ
1 τ k 2 2
pτ βk , and β̃k = βk − β̂k [11]. pτ is the pilot power. of the wireless backhaul link supported by a low-resolution
Similarly, the MC BS to the k-th SC BS channel is given as ADC quantized FD massive MIMO are approximated by
gk = ĝk + g̃k , where g̃k ∼ CN (0, ζ̃k2 INrx ) indicates the error,
θ1 τ pτ ζk 2 (1) ε1 ps β̂k2 Mrx
and ĝk ∼ CN (0, ζ̂k2 INrx ) for ζ̂k2 = 1+τ 2 2 Rm,k = log2 1 + (1) , (15)
pτ ζk , ζ̃k = ζk − ζ̂k . κ̃m,k + μ̃m,k
(1)
Abstract—We propose an efficient approach for the computa- covariance matrices, but severely limits the number of corre-
tion of cumulative distribution functions of N correlated Rayleigh lated random variables (RVs) N. For example, infinite series
or exponential random variables (RVs) for arbitrary covariance representations are derived for N = 2 in [5] and N = 3 and 4
matrices, which arise in the design and analysis of many wireless
systems. Compared to the approaches in the literature, it employs
in [6]–[8] for the arbitrary correlation model.
a fast and accurate randomized quasi-Monte Carlo method that A second class of approaches limits itself to special forms
markedly reduces the computational complexity by several orders of the covariance matrix. For example, for the constant cor-
of magnitude as N or the correlation among the RVs increases. relation model, the CDF FE (a) of N exponential RVs E =
Numerical results show that an order of magnitude larger values (E1 , . . . , EN ) evaluated at a = (a1 , . . . , aN ) is written in
of N can now be computed for. Its application to the performance terms of an infinite series as follows [9]:
analysis of selection combining is also shown.
Index Terms—Multivariate, Rayleigh, exponential, correlated ∞ n
1−ρ ρ
fading, cumulative distribution function. FE (a) =
1 + (N − 1)ρ 1 + (N − 1)ρ
n=0
N
n ai
I. I NTRODUCTION × γ , li + 1 , (1)
l1 , . . . , lN c(1 − ρ)
(l1 ,...,lN ) i=1
ULTIVARIATE or correlated Rayleigh and exponen-
M tial distributions arise in the design and analysis of
many wireless systems. An example of this is a multiple input
l1 ≥0,...,lN ≥0
l1 +···+lN =n
multiple output (MIMO) system when the transmit or receive where ρ2 is the correlation coefficient between Ei and Ej , for
antennas are spaced close together, which leads to the chan- i = j , c is the mean of Ei , for 1 ≤ i ≤ N, and γ(x , k )
nel gains of the different transmit-receive antenna pairs being is the lower incomplete gamma function [10, eq. (6.5.2)].
correlated [1, Ch. 6], [2, Ch. 13]. Another example is a sensor Computing the nth term of this series requires enumerating
network, in which the measurements of different sensor nodes
N +n −1
are correlated. Alternately, when the sensor nodes or the relay all combinations of the vector of non-negative
N −1
nodes in a cooperative relay system are close to each other
integers (l1 , . . . , lN ) that satisfy Ni=1 li = n. This entails a
or when there is a common scatterer in the propagation envi- N −1
computational complexity of O(n ), which increases expo-
ronment, the channel gains seen by the different nodes are
nentially with N. This significant complexity challenge also
correlated [3], [4]. These distributions also arise in orthogonal
arises while evaluating the CDF of correlated Nakagami-m
frequency division multiplexing (OFDM) systems, where the
RVs that is derived in [11].
subchannel gains of a user are correlated [2, Ch. 19].
When the inverse of the covariance matrix of the underly-
Performance analyses of communication techniques
ing Gaussian RVs has a tridiagonal structure, the multivariate
employed in these systems often lead to expressions that
Rayleigh CDF simplifies [12]. However, it is still in the form
involve the multivariate cumulative distributive function
of an infinite series, evaluating which entails O(n N −1 ) com-
(CDF) of such correlated channel gains. For example, the
plexity. The one exception is the specialized correlation model
outage probability of an N-branch selection combiner (SC) in
of [13] in which the multivariate PDF and CDF are given
a multi-antenna system [1, Ch. 9] or of an N-relay cooperative
in single-integral forms. These can be numerically integrated
system [4] is written in terms of the multivariate CDF of the
using Gauss-Laguerre quadrature [10, eq. (25.4.45)].
channel gains.
A third class of approaches considers arbitrary covari-
Considerable effort has been made in the literature to
ance matrices and applies to any N. In [14], the multivariate
derive tractable mathematical representations for the multi-
Nakagami-m CDF for an arbitrary correlation matrix is given
variate probability density functions (PDFs) and CDFs of
as a multi-dimensional integration of the PDF, which itself is
these distributions. One class of approaches considers arbitrary
in the form of an infinite series. In [15], the Green’s approach
Manuscript received July 20, 2018; accepted October 7, 2018. Date of is used to approximate any arbitrary covariance matrix to a
publication October 15, 2018; date of current version April 9, 2019. This matrix whose inverse has a tridiagonal form. However, the
work was supported by the Department of Telecommunications, Ministry multivariate Rayleigh CDF again involves (N − 1) nested
of Communications, India, as part of the Indigenous 5G Test Bed Project
and the Qualcomm Innovation Fellowship. The associate editor coordinating
infinite series, which entails a computational complexity of
the review of this paper and approving it for publication was A. Kammoun. O(n N −1 ) if the series is truncated after n terms. Therefore,
(Corresponding author: Neelesh B. Mehta.) results for N > 5 are seldom available in the literature for
The authors are with the Department of Electrical Communication general correlation models. However, larger values of N are
Engineering, Indian Institute of Science, Bengaluru 560012, India (e-mail:
reneetaisaac@gmail.com; nbmehta@iisc.ac.in). of interest in practice. For example, more antennas for next
Digital Object Identifier 10.1109/LWC.2018.2875999 generation systems and wireless systems with more relays and
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ISAAC AND MEHTA: EFFICIENT COMPUTATION OF MULTIVARIATE RAYLEIGH AND EXPONENTIAL DISTRIBUTIONS 457
sensors are being considered. In many systems, the covariance Converting to polar coordinates, let
matrix need not take a specialized form.
Ei = Xi2 + Yi2 and φi = tan−1 (Yi /Xi ). (4)
Contributions: In this letter, we first propose a novel
approach to numerically compute the multivariate CDF of It can be seen that Ei is an exponential RV with mean 2σi2 and
N Rayleigh or exponential RVs with an arbitrary covariance φi is uniformly distributed between 0 and 2π. Furthermore, Ei
matrix. Its key ideas are as follows. It employs a series and φj are mutually independent for all i and j (including i = j)
of transformations to first convert the multivariate PDF of and (E[Ei Ej ] − 4σi2 σj2 )/( var(Ei )var(Ej )) = ρ2kj .
Rayleigh or exponential RVs to the multivariate PDF of cor- The multivariate CDF FE (·) of E = (E1 , . . . , EN )T
related Gaussian RVs. The CDF expression is then accurately evaluated at a = (a1 , . . . , aN )T is then
and quickly computed using the randomized quasi-Monte 2π 2π a1 aN
Carlo (RQMC) method [16, Ch. 4]. In contrast to the above
FE (a) = ··· ··· fE,Φ (e, θ ) d e d θ , (5)
approaches, we show that it can compute the CDF easily even 0 0 0 0
for N = 50, which is much larger than the values considered
in [5]–[9], [11], [12], [14], and [15]. Notably, it can do so where Φ = (φ1 , . . . , φN )T
and fE,Φ (·) is the joint CDF of E
even when the RVs are highly correlated. This is unlike the and Φ.
above approaches, which require more terms to be computed Using the rules governing transformation of variables, we
as the RVs become more correlated since the rate of conver- get the following from (3) and (4):
u1 u2N
gence of the series decreases. We then illustrate the utility of fX,Y (x, y)
our approach by presenting results for the performance anal- FE (a) = ··· d xd y, (6)
ysis of an N-branch SC, which is a popular receive diversity
l1 l2N x12 + y12 · · · xN 2 + y2
N
combining technique [1, Ch. 9], for a range of values of N. √ √
where
l1 = − a1 , . . . , lN = − aN , lN +1 =
2 , u = √a , . . . , u =
Outline: In Section II, we present the proposed approach. 2
In Section III, we compare its accuracy and computational − a1 − x1 , . . . , l2N = − aN − xN 1 1 N
complexity with existing approaches. In Section IV, we ana- √ 2 2
aN , uN +1 = a1 − x1 , . . . , u2N = aN − xN .
lyze the outage probability of SC. Our conclusions follow in Next, we show how to efficiently compute the integral in (6)
Section V. using the RQMC method of [16, Ch. 4].
Notations: The probability of an event A is denoted by
Pr(A). The expectation of an RV X is denoted by E[X ] and
B. RQMC Method
its variance is denoted by var(X). The joint CDF of a random
vector X = (X1 , . . . , XN ) evaluated at x = (x1 , . . . , xN ) is It transforms the integral in (6) into an integral over a unit
denoted by FX (x) = Pr(X1 ≤ x1 , . . . , XN ≤ xN ). The joint hyper-cube. This is then numerically computed using quasi-
PDF of X is denoted by fX (·). The transpose of a vector x is Monte Carlo techniques. It is as follows.
denoted by xT . 1) Take the Cholesky decomposition of Σ = CCT , where
C = [cij ] is a lower triangular matrix with cii > 0.
T
Using the variable transformation z = C−1 xT yT
II. C OMPUTATION OF M ULTIVARIATE E XPONENTIAL
changes (6) to
AND R AYLEIGH CDF
1
We first consider the multivariate exponential distribution FE (a) =
and then the multivariate Rayleigh distribution. (2π)2N
u u z12 2
z2N
1 2N
× ··· e − 2 · · · e − 2 g(z) d z, (7)
A. Multivariate Exponential CDF l1
l2N
Consider two real Gaussian random vectors X =
where li = (li − ji−1
=1 cij zj )/cii and ui = (ui −
(X1 , . . . , XN )T and Y = (Y1 , . . . , YN )T with zero mean i−1
whose second-order moments are given in general by N j =1 cij
i
zj )/cii , for 1 ≤ i ≤ 2N, and g(z) =
N +i
[( 2 2 −1/2 .
i=1 j =1 cij zj ) + ( j =1 c(N +i)j zj ) ]
E Xk2 = E Yk2 = σk2 , for 1 ≤ k ≤ N , 2) Using vi = Ψ(zi ), where Ψ(z )
z 2 √
E Xk Xj = E Yk Yj = ρkj σk σj , for k = j , − θ2
−∞ e d θ/( 2π), yields
E Xk Yj = 0, for 1 ≤ k , j ≤ N . (2) ⎡⎛ ⎞2
e1 e2N N i
⎢⎝
Thus, ρkj is the correlation coefficient between Xk and Xj FE (a) = ··· ⎣ cij Ψ−1 (vj )⎠
or Yk and Yj . The multivariate Gaussian PDF fX,Y (·, ·) of X b1 b2N i=1 j =1
and Y is ⎛ ⎞2 ⎤− 12
N +i
1 1
x ⎥
fX,Y (x, y) = exp − xT yT Σ−1 , (3) +⎝ c(N +i)j Ψ−1 (vj )⎠ ⎦ d v, (8)
|Σ|(2π)2N 2 y
j =1
T T
x = (x1
, . . . , xN ) , y = (y1 , . . . , yN ) , Σ =
where
where bi = Ψ([li − ji−1 −1
X T T
i−1 =1 cij Ψ (vj )]/cii ) and ei =
E X Y , and | · | denotes determinant. Ψ([ui − j =1 cij Ψ−1 (vj )]/cii ), for 1 ≤ i ≤ 2N.
Y
458 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
3) Applying the transformation wi = (vi − di )/(ei − C OMPARISON OF CDF S C OMPUTED AT P RE -S PECIFIED P OINTS (G IVEN
di ) changes (8) to the following integral over a unit IN [11, T BL . 1]) U SING P ROPOSED A PPROACH AND ISA, AND
hyper-cube: C OMPUTATIONAL C OMPLEXITY OF ISA
1 1
FE (a) = ··· f (w)d w, (9)
0 0
where
2N
f (w) = (ek − bk )
k =1
⎡⎛ ⎞2
N
i
⎢⎝
× ⎣ cij Ψ−1 (bj + wj (ej − bj ))⎠
i=1 j =1
V. C ONCLUSION
The complexity of computing the multivariate CDF with
an arbitrary covariance matrix using conventional approaches
increased exponentially as N increased. We presented a novel
approach for computing the multivariate Rayleigh and expo-
nential CDFs for any N and for an arbitrary correlation
Fig. 1. Exponential correlation model: Outage probability of SC as a function
structure. It had a much lower computational complexity that
of N for different ρ (γth = 3 dB and ωs /N0 = 1). no longer increased as N or ρ increased. We then demonstrated
the utility of our approach by analyzing the outage probability
of an N-branch SC receiver.
R EFERENCES
[1] M. Simon and M.-S. Alouini, Digital Communication Over Fading
Channels, 2nd ed. Hoboken, NJ, USA: Wiley-Intersci., 2005.
[2] A. F. Molisch, Wireless Communications. Chichester, U.K.: Wiley, 2005.
[3] A. Attarkashani and W. Hamouda, “Throughput maximization using
cross-layer design in wireless sensor networks,” in Proc. ICC, May 2017,
pp. 1–6.
[4] B. V. Nguyen, R. O. Afolabi, and K. Kim, “Dependence of outage
probability of cooperative systems with single relay selection on chan-
nel correlation,” IEEE Commun. Lett., vol. 17, no. 11, pp. 2060–2063,
Nov. 2013.
[5] C. C. Tan and N. C. Beaulieu, “Infinite series representations of
the bivariate Rayleigh and Nakagami-m distributions,” IEEE Trans.
Commun., vol. 45, no. 10, pp. 1159–1161, Oct. 1997.
[6] K. D. P. Dharmawansa, R. M. A. P. Rajatheva, and C. Tellambura,
“Infinite series representations of the trivariate and quadrivariate
Nakagami-m distributions,” in Proc. ICC, Jun. 2007, pp. 1114–1118.
[7] K. Peppas and N. C. Sagias, “A trivariate Nakagami-m distribu-
Fig. 2. Arbitrary correlation model: Outage probability of SC as a function tion with arbitrary covariance matrix and applications to generalized-
of γth for different N. selection diversity receivers,” IEEE Trans. Commun., vol. 57, no. 7,
pp. 1896–1902, Jul. 2009.
output SNR γSC is given by γSC = max{γ1 , . . . , γN }. The [8] Y. Chen and C. Tellambura, “Infinite series representations of the trivari-
outage probability of SC for an SNR threshold γth is given by ate and quadrivariate Rayleigh distribution and their applications,” IEEE
! Trans. Commun., vol. 53, no. 12, pp. 2092–2101, Dec. 2005.
" " [9] R. K. Mallik, “On multivariate Rayleigh and exponential distributions,”
γth N0 γth N0
Pout = Pr(γSC ≤ γth ) = FR ,..., . (14) IEEE Trans. Inf. Theory, vol. 49, no. 6, pp. 1499–1515, Jun. 2003.
ωs ωs [10] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions:
With Formulas, Graphs, and Mathematical Tables, 9th ed. New York,
Exponential Correlation Model: In this model, ρkj = ρ|k −j | NY, USA: Dover, 1964.
[11] G. K. Karagiannidis, D. A. Zogas, and S. A. Kotsopoulos, “On the mul-
and σk = 1, for 1 ≤ k, j ≤ N. It arises in equi-spaced diversity tivariate Nakagami-m distribution with exponential correlation,” IEEE
antennas [1]. Fig. 1 plots the outage probability of SC as a Trans. Commun., vol. 51, no. 8, pp. 1240–1244, Aug. 2003.
function of N for different ρ. To compute the multivariate CDF [12] L. E. Blumenson and K. S. Miller, “Properties of generalized Rayleigh
in (11), we use M = 8 and P = 1000. Also shown are results distributions,” Ann. Math. Stat., vol. 34, no. 3, pp. 903–910, Sep. 1963.
[13] N. C. Beaulieu and K. T. Hemachandra, “Novel simple represen-
from simulations, in which 106 realizations of the vector of tations for Gaussian class multivariate distributions with generalized
N correlated Rayleigh RVs R = (R1 , . . . , RN ) are generated correlation,” IEEE Trans. Inf. Theory, vol. 57, no. 12, pp. 8072–8083,
and the average number of outage occurrences is measured. Dec. 2011.
[14] R. A. A. de Souza and M. D. Yacoub, “On the multivariate Nakagami-m
The analysis and simulation results match well even for N as distribution with arbitrary correlation and fading parameters,” in Proc.
large as 50. For a given ρ, Pout decreases as N increases. IMOC, Oct. 2007, pp. 812–816.
However, as ρ increases, Pout increases. [15] G. K. Karagiannidis, D. A. Zogas, and S. A. Kotsopoulos, “An effi-
cient approach to multivariate Nakagami-m distribution using Green’s
Arbitrary Correlation Model: To evaluate the utility for matrix approximation,” IEEE Trans. Wireless Commun., vol. 2, no. 5,
an arbitrary correlation model, the correlation coefficients are pp. 883–889, Sep. 2003.
taken to be linearly spaced between ρmin and ρmax with [16] A. Genz and F. Bretz, Computation of Multivariate Normal and t
σk = 1, for 1 ≤ k ≤ N. Specifically, ρkj = ρmax − (|k − Probabilities, 1st ed. Heidelberg, Germany: Springer, 2009.
[17] Q. T. Zhang, “Maximal-ratio combining over Nakagami fading channels
j | − 1)(ρmax − ρmin )/(N − 2). It applies to linear arrays with with an arbitrary branch covariance matrix,” IEEE Trans. Veh. Technol.,
antennas spaced unevenly [11], [17]. vol. 48, no. 4, pp. 1141–1150, Jul. 1999.
460 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
I. I NTRODUCTION
Fig. 1. Overview of the existing and proposed network.
N RECENT years, a demand for device-to-device (D2D)
I communication, which enables communication between
mobile devices, has been increasing. In order to realize the to its destination node after re-encoding and modulation, as
D2D communication, several methods have been studied, such shown in Fig. 1 [6]. This relay transmission method is called
as the maximization of the time-average total throughput of “Decode-and-Forward (DF)” and adopted by wireless local
network [1], efficient performance of the D2D network [2], area network (WLAN) standards, such as IEEE 802.11s [7].
maximization of the bounds on the capacity [3], signifi- The DF method causes a large delay because of the recep-
cant performance gains of average energy efficiency [4]. In tion and forwarding process in the relay node (Fig. 1). The
particular, the relay-transmission technology, which conveys delay can be reduced by performing data reception and for-
information to a destination node through multiple devices, warding at the same time. We propose a relay transmission
has attracted much attention for the D2D communication [5]. method using multiple bands (Fig. 1) to realize simultaneous
In one of existing relay-transmission technology, a relay data reception and forwarding at the relay node [8], [9]. Using
node that receives a frame from a source node performs multiple bands, the relay method performs the relay transmis-
demodulation and decoding, then forwards the decoded data sion in another frequency band, while the relay node receives
a data frame. Moreover, channels used for reception and for-
Manuscript received June 22, 2018; revised August 29, 2018; accepted warding can be chosen among multiple bands, and spectral
October 8, 2018. Date of publication October 16, 2018; date of current ver- resources can be efficiently utilized.
sion April 9, 2019. A part of this letter was performed by the research contract However, for the relay using multiple bands, channel condi-
entrusted by the Ministry of Internal Affairs and Communications “Research tions, such as propagation loss and fading, differ in each band.
and Development of Spectral-Efficiency Improvement Technology Employing
Simultaneous Transmission over Multiple License-Exempt Bands” and the This causes a mismatch of the data rates between reception
national program, Cross-ministerial Strategic Innovation Promotion Program and forwarding at the relay node. Buffer overflow and packet
(SIP), supported by the Cabinet Office, Japan. The associate editor coordi- loss will occur if the data rate between the source and relay
nating the review of this paper and approving it for publication was Y. Gao. nodes is greater than that between the relay and destination
(Corresponding author: Ayaka Hanyu.)
A. Hanyu, Y. Kawamoto, H. Nishiyama, and N. Kato are with the Graduate nodes. Therefore, this letter proposes an adaptive frequency
School of Information Sciences, Tohoku University, Sendai 980-8579, band and channel selection method that minimizes the trans-
Japan (e-mail: ayaka.hanyu@it.is.tohoku.ac.jp; youpsan@it.is.tohoku.ac.jp; mission delay while avoiding buffer overflow. In the proposal,
hiroki.nishiyama.1983@ieee.org; kato@it.is.tohoku.ac.jp).
N. Egashira and K. Yano are with the Wave Engineering Laboratories,
the relay node calculates the sending and receiving rates based
Advanced Telecommunications Research Institute International, Kyoto on the signal-to-interference plus noise power ratio (SINR) of
619-0288, Japan (e-mail: egashira.naoto@atr.jp; kzyano@atr.jp). each channel and selects an appropriate channel combination
T. Kumagai was with the Wave Engineering Laboratories, Advanced at which the sending and receiving rates become as high as
Telecommunications Research Institute International, Kyoto 619-0288, Japan.
He is now with the Enterprise Solutions Business Headquarters, NTT
possible.
Advanced Technology Corporation, Kawasaki 212-0014, Japan. The remainder of this letter is organized as follows:
Digital Object Identifier 10.1109/LWC.2018.2876258 Section II explains the relay transmission using the multiple
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HANYU et al.: ADAPTIVE FREQUENCY BAND AND CHANNEL SELECTION FOR SIMULTANEOUS RECEIVING AND SENDING 461
B. Problem Formulation
Fig. 1 demonstrates a simplified relay transmission system
in which the source, relay, and destination nodes are indicated
Fig. 2. Proposed algorithm for selecting the appropriate channel combination.
by S, R, and D, respectively. The communication link between
the source and relay nodes is represented by S-R, while that
between the relay and destination nodes is denoted by R-D. As with the largest transmission rates of R-D is selected, and the
mentioned earlier, we consider the simultaneous data receiving sending rates is r RD . In order to select S-R channel, we make
and sending in different bands. The receiving (S-R) rate rSR a list of channels arranged in descending order of SINR, and
and the sending (R-D) rate r RD depend on several factors, the selected channel for R-D is removed from the list because
such as propagation loss. The data accumulated in the relay of loop interference avoidance. Next, we select S-R channel
node may exceed its buffer size when rSR is larger than r RD . with the largest transmission rates of S-R is selected, and the
The data overflowed from the buffer is lost, then leads to the receiving rates on the selected channel is r RD .
increase of the transmission delay caused by the retransmission We assume that the size of the transmitted data is F .
of the lost data. The amount of data that accumulates in the relay node
when the S-R transmission is finished is represented as
(F /rSR ) · (rSR − r RD ). The data that overflows from the
III. P ROPOSED A DAPTIVE F REQUENCY BAND AND
buffer will be lost when it exceeds the buffer size b of the
C HANNEL S ELECTION M ETHOD
relay node. To avoid this situation, the channel used for S-R
A. Algorithm is re-selected until the condition (F /rSR ) · (rSR − r RD ) ≤ b
We propose herein an algorithm that selects the appropriate is satisfied. If no channel matches the condition, the rate
channels used for receiving and sending based on its SINR is adjusted according to the MCS. Furthermore, if no MCS
to avoid data loss caused by the difference between rSR and matches the condition, we use DF method using the chan-
r RD . This letter considers the employment of spatial reuse by nels with the largest transmission rates of S-R and R-D. After
adjusting a threshold of acceptable interference as discussed the channels for S-R and R-D are determined, the relay node
for IEEE 802.11ax wireless LAN [10]. In the proposed algo- informs the source and destination nodes which channels are to
rithm, the channels are selected such that the receiving and be used and starts the data transmission. In addition, when all
sending rates are as high as possible, and the transmission channels are not available or do not have communicable SINR
delay is reduced. Fig. 2 shows the flowchart of our proposal. on either S-R or R-D, this algorithm is executed again after a
In the algorithm, the SINR of each channel is first calculated lapse of time because the transmission cannot be executed.
for S-R and R-D (how to calculate the SINR is described in
Section III-B), and the modulation and coding scheme (MCS)
is chosen by looking up on the MCS table that contains an B. SINR Calculation
appropriate MCS corresponding to SINR. The receiving and The SINR in dB of each channel is calculated as follows:
sending rates are calculated as follows based on the MCS:
PS
m · NSD · rc SINR = 10 log10 , (2)
r= , (1) PN + PI
T
where T is the symbol length of an orthogonal frequency- where PS , PI , and PN represent the desired signal power,
division multiplexing (OFDM) signal [11]; m is the number interference signal power, and noise power, respectively, in
of bits included in a complex symbol; NSD is the number of each channel. We assume the free-space propagation model;
data subcarriers; and rc is the coding rate. Next, the channels hence, both PS and PI can be calculated as follows, and PS
462 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
S IMULATION PARAMETERS
2 2
c c
P S = Pt , (3) PIβ = Pt nallβm · ρβm (x , y) dS . (7)
m 4πdfβ
4πdf Ω
The total interference power PIβ can be expressed as follows
where Pt is the transmission power; c is the light velocity; j
d is the distance between the nodes; and f is the frequency of using PIβ :
m
using channel. jMAX
We now explain how to obtain PI . PI is the total interference
PIβ = olf |j −m| · PIβ , (8)
signal power from the nodes that use the focused channel, j m
m=1
except for S, R, and D. We consider an infinitesimal area dS
in the circles centered both at R and D (Fig. 3). The former and where jMAX is the maximum channel index in the 2.4 GHz
the latter correspond to the area where some devices interfere band; olf h is a coefficient that defines how much the adjacent
with the S-R and R-D links, respectively. We set nall as the channel affects the channel of interest; and h represents the
total number of nodes distributed around the nodes focusing channel difference between the channel of interest and the
on. Moreover nall is separately set for each channel. Among adjacent channel [12].
them, the nodes within the communicable distance dMAX from
the focused nodes are interfering when the focused nodes use IV. P ERFORMANCE E VALUATION
the channel. We define ρ (x, y) as the probability density func- A. Assumptions
tion of the interfering nodes in the area. The total amount
of interference is given as the expectation of the sum of the The assumed environment in this performance evaluation is
interference arriving from within dMAX . first described. While many nodes perform a D2D communica-
tion, we pay attention to three nodes, namely a single source,
2 a relay, and a destination node. Table I shows the simulation
c
PI = Pt nall · ρ(x , y) dS , parameters. The two-dimensional (2D) normal distribution is
Ω 4πdf used herein to represent the distribution of interfering nodes;
Ω : (x − x )2 + (y − y )2 = d 2 ≤ dMAX2
(4) hence, nμ , which is the number of nodes at (μx , μy ), is ran-
domly set for each channel between 2 and 3. Here, (μx , μy ) is
where (x , y ) is the center of the assumed circle. the mean coordinate. nμ is used to calculate the total number
The channels in the 920 MHz, 2.4 GHz, and 5 GHz bands of interfering nodes as nall = 2πσ 2 nμ (a maps a to the
are denoted as αi , βj , and γk , respectively. i, j, and k are greatest integer less than or equal to a), where σ is the vari-
the channel number of each frequency band. The channels ance. In this evaluation, S is located the origin. R and D are
are arranged without overlapping in the 920 MHz and 5 GHz randomly located within a circle with a radius of 100 m from
bands; thus, we neglect the impact of the adjacent-channel S and R, respectively. Table II shows the MCS table used in
interference in these bands. Therefore, the interference signal this evaluation.
power is expressed as follows when the transmission power is We assume that in the environment, the node dis-
Pt in all frequency bands: tribution between S-R and R-D is different. Therefore,
2 in this evaluation, the probability density of the nodes
c is represented by the 2D normal distribution given as
PIαi = Pt nallαi · ραi (x , y) dS , (5) follows:
Ω 4πdfα
2
c 1 (x − μx )2 + (y − μy )2
PIγ = Pt nallγ · ργk (x , y) dS . (6) ρ(x , y) = exp − . (9)
k
Ω 4πdfγ k 2πσ 2 2σ 2
In this evaluation, (μx , μy ) is the coordinate within a circle
Meanwhile, in the 2.4 GHz band, the interference from the
with a radius of 300 m from the origin. (μx , μy ), σ, and nμ are
nodes using adjacent channel with overlapping is considered
uniformly selected at random within the range of the specified
because we assume the IEEE 802.11-based WLAN. To cal-
value, and each parameter is separately determined for each
culate PIβ , which includes the adjacent channel interference,
j channel. The number of channels in 920 MHz, 2.4 GHz, and
we first calculate the interference PIβ from the nodes on 5 GHz bands is 11, 13, and 19, respectively. The bandwidth
m
HANYU et al.: ADAPTIVE FREQUENCY BAND AND CHANNEL SELECTION FOR SIMULTANEOUS RECEIVING AND SENDING 463
TABLE II
MCS TABLE Fig. 4(b) shows the performance of the data transmis-
sion success rate from S to D. In the proposed and DF
methods, the success rate was 1.0 because the optimal
channel was always selected. In contrast, the random-
channel selection method significantly degraded the success
rate because it selected the channels, regardless of the
channel conditions, and, therefore, cannot avoid the buffer
overflow.
V. C ONCLUSION
TABLE III This letter introduced the relay transmission that performed
C HANNEL - TO -C HANNEL OVERLAP FACTORS
simultaneous data receiving and sending to reduce the trans-
mission delay at the relay node using multiple frequency
bands. An issue remains to be solved in the relay transmis-
sion, that is, data loss from buffer overflow is caused by the
difference between the receiving and sending rates at the relay
node. Therefore, we proposed herein an algorithm for balanc-
ing the sending and receiving rates by selecting an appropriate
channel according to its SINR to avoid buffer overflow. The
evaluation results showed that the proposed method reduces
the transmission delay compared with two conventional meth-
ods and attains a transmission success rate of 1.0 by avoiding
buffer overflow.
Abstract—Localization using received signal strength (RSS) Most RSS-based localization approaches are based on an
measurements becomes popular due to the simplicity of practi- underlying assumption that signal can be successfully demod-
cal implementation. Traditional RSS measurements are obtained ulated. Hence, the impact of the background noise (BGN),
after successful demodulation such that the impact of the back- which stems from the interference, the environmental noise
ground noise (BGN) is ignored. However, critical information for
demodulation might be expensive or difficult to obtain in hos- and etc., is significantly mitigated and reasonably ignored.
tile or harsh environments. In this case, the RSS measurements However, this requires some critical configuration informa-
need to be blindly collected without demodulation and hence tion such as the details of modulation, which might be very
characterized by a recent model with the BGN power (already expensive or difficult to access in unattended or hostile envi-
validated by real-life data). This kind of measurement is referred ronments. For instance, in military scenarios, it is hard to
to as “blind RSS measurement”. In this letter, we introduce four demodulate the signal from adversaries if used for localiza-
models for the localization using the blind RSS measurements, tion. An alternative solution has been reported in [7] that the
respectively considering the BGN power and the transmit power
to be known or unknown. A general semi-definite programming
blind RSS measurement may be collected by directly integrat-
solution that applies to all these models is proposed. The corre- ing the observed power spectral density (PSD), where we use
sponding Cramér–Rao lower bounds are presented, indicating a the term “blind” rather than “non-cooperative” as in [7] in case
significant impact of the BGN power on the estimation accuracy. of the confusion with the cooperative/non-cooperative localiza-
Numerical results show the proposed method yields a good and tion with multiple/single target(s). The drawback though is that
reliable performance with different models. both the signal and the BGN contribute to the measurement,
Index Terms—Localization, received signal strength, back- i.e., the BGN power has to be taken into account. Thus, the tra-
ground noise, semidefinite programming, Cramér-Rao lower ditional log-normal fading model is not applicable to this kind
bound. of RSS measurement. Fortunately, based on real-life experi-
ment data, R Martin et al. reported a sound and solid model for
I. I NTRODUCTION the blind RSS measurement that includes the BGN power in
OCATION awareness nowadays has been ubiquitously [7, eq. (17)]. Based on this model, an ML solution for local-
L required in many aspects of commercial, public service,
military sectors. Thus, source localization techniques have
ization was accordingly presented therein. Nevertheless, this
ML solution suffers from the same disadvantages as mentioned
become much prevalent in recent years. The commonly used before and, more importantly, there is still no other method
measurements are time-of-arrival (TOA), time-difference-of- that can blindly estimate target location without demodulating
arrival (TDOA), angle-of-arrival (AOA) and received signal the signal. Therefore, the topic of blind RSS-based localization
strength (RSS) [1]. Owing to the simplicity of practical is still in its infancy.
implementation, source localization using the RSS measure- This letter is aimed at enriching the research of this topic.
ment has drawn much attention, thus resulting into many First, Section II present the model for the blind RSS mea-
notable contributions on this topic. Some maximum likeli- surement that includes the BGN power. In Section III, four
hood (ML) solutions were introduced in [2], but they require linear models are introduced, respectively considering the
an appropriate initialization and high computational com- BGN power and the transmit power to be known or unknown.
plexity. In [3], they proposed some least squares estimators, In this section, a general SDP approach that applies to all these
which however are very susceptible to a large measurement models is also proposed and the corresponding Cramér-Rao
noise. Furthermore, some semidefinite programming (SDP) lower bounds (CRLBs) are derived. Finally, to evaluate those
approaches were also reported in [4] and [5]. This kind of models and the SDP solution, numerical simulations have been
solution has gradually become the new favourite at this time conducted, leading to some conclusions in Section IV.
due to its very robust localization performance [6].
Manuscript received July 19, 2018; revised September 10, 2018; accepted
October 12, 2018. Date of publication October 16, 2018; date of current II. B LIND RSS M ODEL
version April 9, 2019. The associate editor coordinating the review of this
paper and approving it for publication was D. So. (Corresponding author: This section presents the blind RSS model that includes the
Yongchang Hu.) BGN and is used throughout this letter. Assume the received
Y. Hu is with the Cloud Core Network Product Line/Research and signal y(t) with the time instance t can be expressed as
Development Department, Huawei Technologies Company, Ltd., Xi’an
710000, China (e-mail: hycforever2000@gmail.com). y(t) = x (t) h(t) + n(t), where denotes the convolu-
J. Liu is with the Faculty of Electrical Engineering, Mathematics, and tion operator, x(t) is the transmitted signal, h(t) indicates the
Computer Science, Delft University of Technology, 2628 CD Delft, The channel response and n(t) is the zero-mean white Gaussian
Netherlands (e-mail: j.liu-1@tudelft.nl).
B. Zhang is with the School of Electronic Science, National University of
BGN. After the signal is successfully demodulated, the impact
Defense Technology, Changsha 410073, China (e-mail: zbbzb@nudt.edu.cn). of n(t) can be largely alleviated, turning into relatively triv-
Digital Object Identifier 10.1109/LWC.2018.2876319 ial demodulation errors. Since the collection of the traditional
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HU et al.: LOCALIZATION USING BLIND RSS MEASUREMENTS 465
RSS measurement estimates the expected value of the instan- distance term di2 for convenience of unfolding the Euclidean
taneous signal power r (t)2 |x (t) h(t)|2 by averaging norm. Thus, (3) is further reformulated into
the samples (see [6, Appendix A]), the demodulation errors − γ2 P 2
−γ
can further be mitigated. Therefore, it is very reasonable that P0 ||x − si ||2 = ( i − τbg ) . (4)
the traditional RSS measurement is modelled by the famous χi
log-normal fading that ignores the impact of n(t) [8]. To be For a sufficiently small χi , the right-hand side (RHS) of (4)
specific, for localization, suppose the target node is located at can be approximated by applying the first-order Taylor expan-
x ∈ Rd and N anchors are pre-deployed with known locations. P 2 2
)− γ ≈ (P − τ )− γ + ln(10) P (P −
sion as ( χi − τbg
The traditional RSS measurement in dB scale associated with i
i bg 5γ i i
− γ2 −1
the i-th anchor located at si ∈ Rd is expressed as )
τbg χi , resulting into
Pi = P̄i + χi (1) 2
− γ − 2 ln(10) 2
− γ −1
||x − si ||2 ≈ Pi − τbg
γ
P0 + Pi (Pi − τbg ) χi . (5)
5γ
P̄i P0 −10γlog10 (di ) with di ||x−si ||, P0 is the transmit
power, γ is the path-loss exponent (PLE) and χi ∼ N (0, σ 2 ) Then, rearranging (5) and stacking both sides lead to our first
is the shadowing effect. linear model with both known τbg and P0 as
However, when the signal is very difficult or expensive to y1 = F1 θ 1 + D1 χ , (6)
demodulate, we can do nothing but integrating the observed
PSD of y(t) to obtain the blind RSS measurement (in watts) where θ 1 [x, ||x||2 ]T ,
χ stacks all values of χi , D1
2
as Pr T1 |y(t)|2 dt, where T is the time window, which
ln(10)
diag([ · · · , − 5γ Pi (Pi − τbg )− γ −1 , · · · ]T ) with diag(·)
⎡ ⎤
obviously includes the BGN power. Note that the model for .. ..
this kind of measurement, which we will introduce next, has − 2⎢ . . ⎥
been reported before and already evaluated by real-life exper- the diagonal matrix, F1 P0 γ ⎢ ⎣−2si
T 1⎥ and y
⎦ 1
iments in literature. Here, we will not add more discussions .. ..
. .
here and if readers are interested in details, please refer to [7] − 2 T
2
and the references therein. −
· · · , P − τ − P γ ||si ||2 , · · · .
γ
Accordingly, the blind RSS measurement in dB scale can i bg 0
reasonably be modelled as 2) P0 is Unknown: We rewrite (5) as
− 2 2
−2sT 2
P̄i τbg γ
Pi = 10log10 (10 10 + 10 10 ) + χi , (2) i x + ||x|| − Pi − τbg P0 γ
2 ln(10) 2
− γ −1
where τbg indicates the BGN power in dB scale. Here, we ≈ −||si ||2 + P0γ Pi (Pi − τbg ) χi (7)
assume the PLE is known a priori, since the PLE awareness 5γ
has become an indispensable and possible feature for effi- and observe that the unknown transmit power P0 issue can
2
ciently designing wireless systems owing to the self-estimation
easily be tackled by estimating an unknown P0γ . Thus, stack-
techniques in [9]. The impact of the BGN power τbg can be
ing both sides of (7) leads to our second linear model with
readily seen from the model in (2) that there exists a power
unknown P0 as
floor for E (Pi ), i.e., E (Pi ) > τbg , where E(·) indicates
the expectation operator. Additionally, the BGN also affects y2 = F2 θ 2 + D2 χ , (8)
the localization performance, which will be studied later in 2 2
Section IV. where θ 2 [x, ||x||2 , P0γ ]T , D2 P0γ D1 , y2
[ · · · , −||si ||2 , . . . ]T and
III. P ROPOSED L OCALIZATION A PPROACHES ⎡ .. .. .. ⎤
. . .
This section introduces four linear models for blind RSS- ⎢ − 2 ⎥
based localization with known/unknown P0 or τbg . Because ⎢ γ⎥
F2 ⎢−2sT 1 − Pi − τbg
⎥.
P0 may be known as the standard configuration or otherwise ⎣ i ⎦
.. .. ..
due to the power control for energy saving or security pur- . . .
poses. Similarly, τbg may be constantly detected or unknown
in the dynamic environments. Then, a general SDP solution is 3) τbg Is Unknown: We equivalently consider an unknown
and notice that τ is still not linear in the RHS of (5).
τbg
proposed, which applies to all these models. The Cramér-Rao bg
P in watts and
To cope with that, we use the fact that τbg
lower bounds (CRLBs) are also studied. 0
introduce a new unknown variable α to replace the unknown
as τ = αP . Since α is sufficiently small, we can
τbg
A. Linear Models bg 0
approximate the following terms in (5) using their first order
1) Both τbg and P0 Are Known: We first convert the blind Taylor expansions around α = 0, i.e.,
RSS measurement Pi in (2) back into watts as
− 2 − 2 2 − 2 −1
P̄i Pi − αP0 γ ≈ Pi γ + P0 Pi γ α (9)
Pi = χi (10 10
+ τbg ) ⇒ Pi = χi (P0 di−γ + τbg
), (3) γ
Pi χi P0 τbg and
where Pi 10 10 , χi 10 10 , P0 10 10 and τbg
10 10 .
− 2 −1 − 2 −1 2 + γ − 2+2γ
Recalling di ||x − si ||, we need to collect the squared Pi − αP0 γ ≈ Pi γ + P0 Pi γ α. (10)
γ
466 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Substituting (9) and (10) into the RHS of (5) and ignoring the Then, considering the dependence of the elements in θ I ,
high order small term O(αχi ) result in we can formulate our weighted optimization problem for all
− 2 the introduced linear models as
ln(10) 2
− γ −1
Pi − τbg
γ
+ Pi (Pi − τbg ) χi min (yI − FI θ I )T WI (yI − FI θ I )
5γ θ
(16a)
I
− 2 2 − 2 −1 ln(10) − γ2
≈ Pi γ + P0 Pi γ α + Pi χi . (11) s.t. [θ I ]T
1:d [θ I ]1:d = [θ I ]d+1 (16b)
γ 5γ
Plugging (11) into (5), we similarly obtain our third linear Using the Schur complement and constructing some linear
model with an unknown τbg as matrix inequalities (LMIs) result into our SDP optimization
problem
y3 = F3 θ 3 + D3 χ , (12)
min tI (17a)
where the new parameter vector becomes θ 3 [x, ||x||2 , α]T , θ I ,tI
ln(10) − 2 IN
1/2
WI (yI − FI θ I )
D3 diag([ · · · , − 5γ Pi γ , · · · ]T ), y3 s.t. 1/2 0 (17b)
T WI (yI − FI θ I )T tI
− γ2 − γ2
· · · , Pi 2
− P0 ||si || , · · · and
Id [θ I ]1:d
Θ(θ I ) (17c)
⎡ ⎤ [θ I ]T [ θ I ]d+1
.. .. .. 1:d
⎢ . . . ⎥ Θ(θ I ) 0, (17d)
⎢ − 2 − 2 2
⎥
2 P P − γ −1 ⎥.
F3 ⎢−2P γ sT P γ 1/2
⎣ 0 i 0 γ 0 i ⎦ where tI is an auxiliary slack variable, WI = D−1
1 , I = 1, 2
.. .. .. 1/2
. . . and WI = D−1
3 , I = 3, 4. Note that (17) drops the rank
constraint, i.e., rank(Θ(θ I )) = d , in order to guarantee a con-
4) Both τbg and P0 Are Unknown: Considering the approx- vex set for θ I . This is called the semidefinite relaxation (SDR).
imations (9) and (10) and omitting the high order small term Finally, the proposed SDP optimization problem in (17) can
O(αχi ), we rewrite (7) into be efficiently solved by CVX and the estimate of the tar-
− γ2 2 2 − γ2 −1 γ2 +1 get location is hence given by x̂ = [θˆI ]1:d . Considering the
−2sT 2
i x + ||x|| − Pi P0 γ − P αP0 worst-case complexity, i.e., the interior-point algorithm, the
γ i
complexity for each iteration
√ is O[d 2 N ] and the iteration num-
2 ln(10) − γ2
≈ −||si ||2 + P0γ P i χi , (13) ber is bounded by O[ N ln(1/ξ)], where ξ is the iteration
5γ tolerance, thus the total complexity is O[d 2 N 1.5 ln(1/ξ)].
which readily leads to our fourth linear model with unknown
τbg and P0 as C. CRLBs
y4 = F4 θ 4 + D4 χ , (14) According to (2), we stacks all the measurements Pi
as p ∼ N (μ, Σχ ), where p [ · · · , Pi , . . . ]T , μ
γ2 γ2 +1 γ2 [ · · · , μi , . . . ]T with μi 10log10 (P̄i + τbg
) and P̄
where θ 4 [x, ||x||2 , P0 , αP0 ]T , D4 P0 D3 , y4 i
P̄i
y2 , and 10 10 , and the covariance matrix of the measurement noise
⎡ ⎤ is Σχ = σ 2 IN . When the elements of the parameter vec-
.. .. .. ..
⎢ . . . . ⎥ tor θ are mutual independent, the Fisher information matrix
⎢ − 2 − γ2 −1 ⎥ (FIM) can be computed as [J]n,m = [ ∂θ ∂ μ T −1 ∂ μ
] Σχ [ ∂θm ] =
F4 ⎢−2sT 1 −P γ 2
− γ Pi ⎥. n
⎣ i i ⎦ 1 [ ∂ μ ]T [ ∂ μ ], where the parameter vector θ could be
.. .. .. .. σ 2 ∂θn ∂θm
. . . . θ = x, θ = [xT , P0 ]T , θ = [xT , τbg ]T or θ =
∂μ ∂[μ]
[xT , P0 , τbg ]T , and ∂θ n
[ · · · , ∂θni , . . . ]T . Letting
B. Semidefinite Programming Solution
x = [x1 , . . . , xd ]T and si = [si,1 , . . . , si,d ]T , we obtain
For convenience, our four linear models in (6), (8), (12) ∂[μ]i 10γ P̄ i i,k x −s ∂[μ]
and (14) are formulated into the following general form ∂xk = − ln(10) i
||x−s ||2 ,
P̄i +τbg
k = 1, . . . , d , ∂P0i =
i
P̄i ∂[μ] bg τ
yI = FI θ I + DI χ, I = 1, . . . , 4, (15) and ∂τ i = P̄ +τ . Now, we can easily calculate
P̄i +τbg
bg i bg
where the dimensions of FI and θ I change according to the FIMs for different θ and hence the corresponding CRLBs.
the number of unknown parameters to estimate. Obviously, In this letter, the main concern is estimating the target loca-
the model noise DI χ is colored as its covariance matrix is tion x, forwhich the CRLBs with different θ are given by
d
ΣDI χ = σ 2 D2I IN . In order to whiten the model noise, we CRLB = −1
k =1 [J ]k ,k .
can optimally construct the weight matrix WI as WI D−2 I ,
1/2
since the covariance matrix of WI DI χ becomes (scaled) IV. N UMERICAL R ESULTS
2 Monte Carlo simulations with 1000 trials have been con-
identity. Notice that, since both D2 P0γ D1 and D4
2 2 ducted to study the CRLBs and to evaluate the introduced
P0γ D3 are simply scaled by an unknown P0γ , we choose models and the SDP solution, where the target node is ran-
their weight matrices as W2 = W1 and W4 = W3 . domly placed in a 50 m × 50 m field and 10 anchors are
HU et al.: LOCALIZATION USING BLIND RSS MEASUREMENTS 467
B. Localization Performance
The proposed SDP estimators associated with our models
in (6), (8), (12) and (14) are respectively denoted as SDP1,
SDP2, SDP3 and SDP4. A recent constrained weighted least
squares (CWLS) estimator for the traditional RSS-based local-
ization that ignores the BGN power is also considered for
comparison [3]. As shown in Fig. 2, our proposed SDP estima-
tors significantly outperform the traditional CWLS estimator
with at least 1 m better in the RMSE, which again strongly
proves that the BGN power cannot be ignored. In stark con-
trast, even the performance degradation caused by unknown
parameters does not seem that significant. Furthermore, with
small τbg or σ, the proposed methods yield the performance
Fig. 1. The blue solid curves are the CRLBs for the traditional RSS-based that are close to the corresponding CRLBs. Admittedly, under
localization, while the red dashed curves represent the CRLBs for the blind large τbg or σ, the approximations in the derivations become
RSS-based localization.
less effective, thus resulting into a gap to the CRLBs, even
though the proposed methods are still better than the tradi-
tional CWLS estimator. Actually, this is a common issues for
most localization problems [6]. However, due to the page limit,
we will leave this issue for the future research challenges.
V. C ONCLUSION
In this letter, we consider the localization using the blind
RSS measurements that include the BGN power. As a kick-
off of this topic, we introduce four models with the transmit
power and the BGN power to be known or unknown. We also
propose a general SDP solution and the corresponding CRLBs
that indicates the importance of considering the BGN power.
The numerical results show that the proposed models and the
Fig. 2. Performance of the proposed SDP estimators compared with a general SDP solution are very reliable and suffice to serve this
traditional LLS estimator that ignores the BGN. new kind of localization. This letter casts light on this new kind
of localization and our future challenges include coping with
pre-deployed with known locations (50, 50), (50, 0), (0, 50), an unknown PLE and system parameter errors.
(0, 0), (25, 7), (25, 43), (12, 33), (12, 16), (37, 33) and (33, 16).
The PLE is set to 4 and the transmit power is 30 dBm. The
root mean squares error (RMSE) is used to evaluate the esti- R EFERENCES
mation accuracy and the average CRLB is calculated from [1] F. Gustafsson and F. Gunnarsson, “Mobile positioning using wireless
different target locations. Due to the page limit, we will only networks: Possibilities and fundamental limitations based on available
focus on the localization performance, not the estimation for wireless network measurements,” IEEE Signal Process. Mag., vol. 22,
other unknown parameters. no. 4, pp. 41–53, Jul. 2005.
[2] X. Li, “RSS-based location estimation with unknown pathloss model,”
IEEE Trans. Wireless Commun., vol. 5, no. 12, pp. 3626–3633,
Dec. 2006.
A. CRLBs [3] Z. Li, “Constrained weighted least squares location algorithm using
As shown in Fig. 1a, the blind source localization based received signal strength measurements,” China Commun., vol. 13, no. 4,
on the RSS measurement with the BGN power cannot outper- pp. 81–88, Apr. 2016.
[4] R. M. Vaghefi, M. R. Gholami, R. M. Buehrer, and E. G. Strom,
form the traditional one, which however can be understood as “Cooperative received signal strength-based sensor localization with
a trade-off of no signal demodulation. Obviously, the impact unknown transmit powers,” IEEE Trans. Signal Process., vol. 61, no. 6,
of the BGN is rather significant especially with a large BGN pp. 1389–1403, Mar. 2013.
power. Nonetheless, when the BGN power is small, the blind [5] X. Guo, L. Chu, and X. Sun, “Accurate localization of multiple sources
using semidefinite programming based on incomplete range matrix,”
RSS-based localization can yield almost the same accuracy IEEE Sensors J., vol. 16, no. 13, pp. 5319–5324, Jul. 2016.
as the traditional one, where no signal demodulation now [6] Y. Hu and G. Leus, “Robust differential received signal strength-
becomes a very remarkable advantage especially for military based localization,” IEEE Trans. Signal Process., vol. 65, no. 12,
scenarios. Next, let us focus on the blind RSS-based localiza- pp. 3261–3276, Jun. 2017.
[7] R. K. Martin et al., “Modeling and mitigating noise and nuisance
tion and observe that, with more unknown parameters τbg or parameters in received signal strength positioning,” IEEE Trans. Signal
P0 , the localization performance deteriorates. Particularly, the Process., vol. 60, no. 10, pp. 5451–5463, Oct. 2012.
unknown P0 yields a more severe impact than the unknown [8] T. S. Rappaport, Wireless Communications: Principles and Practice
τbg . Interestingly, although considering the BGN power is (Prentice Hall Communications Engineering and Emerging Technologies
Series), London, U.K.: Dorling Kindersley, 2009.
very important, its knowledge is relatively not that significant. [9] Y. Hu and G. Leus, “Self-estimation of path-loss exponent in wireless
Finally, from Fig. 1b, all the CRLBs become worse with a networks and applications,” IEEE Trans. Veh. Technol., vol. 64, no. 11,
large shadowing effect. pp. 5091–5102, Nov. 2015.
468 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
C OMPARISON OF D IFFERENT MIMO T RANSMISSION S CHEMES
Fig. 4. The clustering structure of a binary MoG data set (a) before and (b)
after GAE.
Fig. 6. Learning performance comparison for two cases: (a) a varying
Doppler shift with the average transmit SNR equal to 15 dB; (b) a varying
average transmit SNR with the normalized Doppler shift fixed at 0.01.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHANG et al.: ON ACHIEVING MAXIMUM STREAMING RATE IN HYBRID WIRED/WIRELESS OVERLAY NETWORKS 473
TABLE I
N OTATIONS Based on Eq. (2) and u(·) and c(·) defined in Table I, the
following notations are introduced for ease of presentation:
⎧
⎨u(L, A) = u(A) + |A|−1 u(B )A , A ⊂ L
|A|
. (3)
⎩c(L, A) = c(A) + |A|−1 c(B ), A⊂L
|A|+1
close they are, must go through the core network, i.e., there Proof: The proof can be illustrated by Fig. 1.
are no direct communications among wireless nodes connect- Step 1: We prove that Eq. (4) is an upper bound. Obviously,
ing to the same AP or base station (BS). The bottleneck lies in rmax is bounded by rA and u0 . Thus, we only prove how rmax
the access network rather than the core network, which usually is bounded by the third term, which can be intuitively under-
has abundant bandwidth resources [1], [5], [6]. stood as the average aggregate bandwidths that the system
There are two node sets A, L ⊆ N, where N denotes the uni- provides to all nodes in A.
versal node set. Each of them may contain wired and wireless 1) Case A ⊂ L: For a wired node v or a wireless node w
nodes. B denotes the difference of A and L. The server sends in A, it forwards a substream from the server to all other
a stream to the overlay at the rate equal to its UL bandwidth |A|− 1 nodes in A. Suppose that all nodes in A receive the
u0 , referred to as the original streaming rate. All nodes in A substream at rmax . The upper bounds of the total available UL
want to concurrently receive the stream. All nodes in L always bandwidths of Awd and Awl are u(A) and c(A) − |Awl |rmax ,
contribute their UL bandwidths and forward the stream. There respectively.
are two cases A ⊂ L and A ⊇ L. We refer to the former as For a wired node x in B, it forwards a substream from the
the inflow differentiated service, i.e., all the bandwidth of L server to all |A| nodes in A. Similar to Eq. (2), the available
is aggregated into A, and the latter as the outflow differenti- UL bandwidth of node x is thus limited by min{|A|dx , ux }.
ated service, i.e., all the bandwidth of L is spread across A. To fully exploit its UL bandwidth, the server should provide
Without loss of generality, we incorporate the equal service 1 min{|A|d , u } of the UL bandwidth to node x. Summing
|A| x x
case A = L into the outflow case. The first objective is to cal- up all wired nodes in B, u(B )A of their UL bandwidths can be
culate the maximum (achievable) streaming rate rmax for A 1 u(B ) of the server’s UL bandwidth
used at the cost that |A| A
when L and A are fixed.
is consumed.
Further, note that rmax may not always reach u0 . In live
For a wireless node y in B, all its UL bandwidth can be
streaming, how many nodes at most can reach u0 is a cru-
utilized when it forwards a substream from the server at rate
cial performance metric. Suppose that L is fixed and that all 1 c to all nodes in A. Summing up all wireless nodes
nodes in N are arranged in some order. The second objective |A|+1 y
|A|
is to calculate the largest integer |A|, defined as the hybrid in B, |A|+1 c(B ) of their UL bandwidths can be used while
multiplicity Mh , such that the first |A| nodes can reach u0 . 1 c(B ) of the server’s UL bandwidth is consumed.
|A|+1
Since all nodes in A receive the stream at the same rate, Combining the wired and wireless cases, rmax satisfies:
using dmin (·) and cmin (·) defined in Table I, the maximum
streaming rate that A can achieve is bounded by: 1
rmax ≤ u0 + u(A) + c(A) − |Awl |rmax
|A|
|A| − 1 |A| − 1
rA = min{dmin (A), cmin (A)}. (1)
+ u(B )A + c(B ) . (5)
|A| |A| + 1
When A ⊂ L, suppose that a wired node i in B wants to Plugging Eq. (3) into the above equation and moving rmax
forward a substream from the server to all nodes in A. The to the left-hand side, we have:
substream that it receives from the server is at most at the
rate equal to di . Thus, the total UL bandwidth that node i can 1
rmax ≤ u0 + u(L, A) + c(L, A) . (6)
offer to A is bounded by min{|A|di , ui }. Similarly, the total |A|+|Awl |
UL bandwidth that Bwd can offer to A is calculated as: 2) Case A ⊇ L: For a wired node v or a wireless node w in
L, it forwards a substream from the server to all other |A|− 1
u(B )A = min{|A|di , ui }, A ⊂ L. (2) nodes in A. Consequently, all nodes in B can reach the same
i∈Bwd streaming rate as the nodes in A, without contributing their UL
474 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
bandwidths. Similar to the previous case, the upper bounds on of nodes in Bwd are sufficiently large, i.e., A ⊂ L, c(L,A) = 0,
the total available UL bandwidths of Lwd and Lwl are u(L) Awl = 0, and |A|di ≥ ui , i ∈ Bwd , Eq. (4) is reduced to the
and c(L) − |Lwl |rmax , respectively. case in [1]. When there are only wireless nodes using the equal
Combining the wired and wireless cases, rmax satisfies: service, i.e., A = L and |A| = |Lwl |, Eq. (4) is reduced to the
1 case in [6].
rmax ≤ u0 +u(L) + c(L) − |Lwl |rmax . (7)
|A|
B. Hybrid Multiplicity
Moving rmax to the left-hand side, we can obtain:
Based on Theorem 1, we can easily calculate the hybrid
rmax ≤
1
u0 + u(L) + c(L) . (8) multiplicity Mh as follows.
|A|+|Lwl | Theorem 2: Suppose that all nodes in N are numbered from
1 to |N|. L contains the first |L| nodes and A the first |A| nodes.
Step 2: We prove that the upper bound can be achieved.
u +u(L,A)+c(L,A) Define the hybrid multiplicity function as:
1) Case A ⊂ L: Denote r = 0 |A|+|A | . There ⎧
wl
are six cases according to the relationship among r , u0 and ⎪
⎪ +∞, |A| = 1
⎨ u(L,A)+c(L,A)
rA . We only prove the fundamental case r ≤ u0 ≤ rA f (A) = , |A| ∈ (1, |L|) . (11)
|A|+|Awl |−1
where the performance bottleneck lies in the whole overlay ⎪
⎪
⎩ u(L)+c(L) , |A| ∈ [|L|, |N |]
rather than the server or individual nodes; other cases can |A|+|Lwl |−1
be proven similarly using the methods in [1]. As shown in
Then the first |A| nodes can reach u0 if and only if:
Fig. 1, we construct five types of substreams sent by the
server: u0 ≤ min{f (A), rA }. (12)
T1: A substream traverses a wired node v in A and is
copied to other |A|− 1 nodes in A. The streaming rate is: The hybrid multiplicity Mh can be computed as the largest
1 u .
sv1 = |A|−1 integer |A| such that Eq. (12) holds. If Eq. (12) cannot be
v
T2: A substream traverses a wireless node w in A and is satisfied in any case, Mh equals zero.
copied to other |A|− 1 nodes in A. The streaming rate is: Proof: Since the case |A| = 1 is trivial, we only prove the
1 (c − r
sw2 = |A|−1 other two cases.
w max ).
T3: A substream traverses a wired node x in B and is 1) Case |A| ∈ (1, |L|). Eq. (11) and Eq. (12) yield:
copied to |A| nodes in A. The streaming rate is: sx3 = 1
1 min{|A|d , u }.
x x
u0 ≤ u0 + u(L, A) + c(L, A) . (13)
|A| |A|+|Awl |
T4: A substream traverses a wireless node y in B and is
copied to all nodes in A. The streaming rate is: sy4 = |A|+1 1 c . 2) Case |A| ∈ [|L|, |N |]. Eq. (11) and Eq. (12) yield:
y
T5: A substream is directly sent to all nodes in A. The
1 u −
1
streaming rate is: s05 = |A| 0 sv1 − w ∈Awl sw2 − u0 ≤ u0 + u(L) + c(L) . (14)
v ∈A wd |A|+|Lwl |
3 4
x ∈Bwd sx − y∈Bwl sy . Eqs. (13) and (14) imply that the average aggregate band-
Each node t in A receives the original stream at the rate:
widths that the system provides to the first |A| nodes are larger
r= sv1 + sw2 + sx3 + sy4 + s05 . (9) than or equal to u0 . Specifically, as long as Eq. (12) holds,
v ∈Awd w ∈Awl x ∈Bwd y∈Bwl the streaming rate of the first |A| nodes can reach u0 .
Remark 2: Once the node order is given, |A| can uniquely
Through a simple calculation, we can obtain r = rmax . determine the nodes contained by A. Therefore, given |L| and
u +u(L)+c(L)
2) Case A ⊇ L: Denote r = 0 |A|+|L | . For the same u0 , the node order uniquely determines Mh .
wl
reason as in the previous case, we only consider the funda- The model in this letter has wide applicability in media
mental case r ≤ u0 ≤ rA . As shown in Fig. 1, we construct streaming and file sharing. For inflow, the nodes in B can serve
three types of substreams sent by the server: as helpers to accelerate the streaming rate in A. For outflow, the
T1: A substream traverses a wired node v in L and is premium users in B, without sharing any bandwidths to oth-
copied to |A|− 1 other nodes in A. The streaming rate is: ers, can reach the same streaming rate as the common users
1 u .
sv1 = |A|−1 v in L. With the hybrid multiplicity, the service provider can
T2: A substream traverses a wireless node w in L and is measure how many users at most it can cover with the highest
copied to |A|− 1 other nodes in A. The streaming rate is: quality service. In practice, multiple media streams may be
1 (c − r
sw2 = |A|−1 concurrently requested by multiple node classes. Fortunately,
w max ).
T3: A substream is directly sent to all nodes in A. The the corresponding n-class problem can often be decomposed
1 u −
1−
streaming rate is: s03 = |A| 0 s
v ∈Lwd v
2
w ∈Lwl sw .
into a series of two-class problems investigated in this let-
Each node t in A receives the original stream at the rate: ter [6]; the hybrid multiplicity can play a crucial role in file
sharing [2], [3].
r= sv1 + sw2 + s03 . (10)
v ∈Lwd w ∈Lwl IV. N UMERICAL E XAMPLES
Through a simple calculation, we can obtain r = rmax . To analyze the obtained results in the previous section, we
Remark 1: When there are only wired nodes using the set u0 ∈ [3, 12], |N| = 10, |L| = 6, and |A| ∈ [1, 10] in the
inflow differentiated service and assuming the DL bandwidths three networks. The bandwidth profile is given in Table II.
ZHANG et al.: ON ACHIEVING MAXIMUM STREAMING RATE IN HYBRID WIRED/WIRELESS OVERLAY NETWORKS 475
Fig. 2. rmax and Mh . (a) Wired network. (b) Wireless network. (c) Hybrid network. (d) Scalability, u0 = 8.
Fig. 3. Effect of node combinations. u0 = 50. The horizontal dashed line separates the inflow case and outflow case. (a) Wired network, with infinite DL
bandwidths di . (b) Wired network. (c) Wireless network. (d) Hybrid network.
TABLE II
P ROFILE FOR W IRED , W IRELESS , AND H YBRID N ETWORKS possibly larger rmax , not necessarily always so, because the
node combination has a significant effect.
V. C ONCLUSION
In this letter, we have derived explicit formulas for the
maximum streaming rate and the hybrid multiplicity in a
fully-connected hybrid wired/wireless overlay. Since the band-
widths of all nodes are heterogeneous, the obtained results are
practical and generalizable. In the future, this letter can be
extended to more flexible duplex modes, or can be adapted to
other networking paradigms with similar topologies and data
Figs. 2a–2c visualize the two theorems. When u0 = 10, transmission processes.
Mh of the three networks are 1, 4, and 2. When u0 > 10, Mh
becomes zero in Fig. 2a, while it is always positive in Fig. 2b. R EFERENCES
Mh in Fig. 2c falls in the middle compared to the other two [1] R. Kumar and K. Ross, “Peer-assisted file distribution: The minimum
networks. There is a tradeoff between the performance metrics distribution time,” in Proc. IEEE HOTWEB, Boston, MA, USA, 2006,
{rmax , Mh } and the system parameters {u0 , |L|, |A|}. From pp. 1–11.
the perspective of the server, the smaller the value of u0 , the [2] G. M. Ezovski, A. Tang, and L. L. H. Andrew, “Minimizing average
finish time in P2P networks,” in Proc. IEEE INFOCOM, Rio de Janeiro,
more the nodes which can reach it. The larger the size of A, the Brazil, 2009, pp. 594–602.
smaller the maximum streaming rate that the nodes in A can [3] M. Mehyar, W. Gu, S. H. Low, M. Effros, and T. Ho, “Optimal strategies
reach. Mh can be viewed as the critical point of the system. for efficient peer-to-peer file sharing,” in Proc. IEEE ICASSP, Honolulu,
HI, USA, 2007, pp. 1337–1340.
When |A| > Mh , no nodes in A can reach u0 , although the [4] S. Sengupta et al., “Peer-to-peer streaming capacity,” IEEE Trans. Inf.
bandwidths are fully utilized. When |A| ≤ Mh , all nodes in A Theory, vol. 57, no. 8, pp. 5072–5087, Aug. 2011.
can reach u0 , but some bandwidths may be wasted. [5] J. Zhang, X. Zhang, and C. Yang, “Towards the multi-request mecha-
In Fig. 2d, we increase the network size by replicating each nism in pull-based peer-to-peer live streaming systems,” Comput. Netw.,
vol. 138, pp. 77–89, Jun. 2018.
node in Table II from 1 to 100 times (scaling factor) while [6] X. Meng, “Bandwidth partition strategies for minimizing peer-to-
keeping their order unchanged. When |L| increases proportion- peer multi-file distribution time,” M.S. thesis, Dept. Elect. Electron.
ally, Mh grows linearly as the network size increases. Eng., Univ. Hong Kong, Hong Kong, 2013. [Online]. Available:
In Fig. 3, we fix L and study the effect of node combina- http://hub.hku.hk/handle/10722/192851
|L| [7] A. Ghaderzadeh, M. Kargahi, and M. Reshadi, “ReDePoly: Reducing
tions within A. There are |A| combinations for inflow and delays in multi-channel P2P live streaming systems using distributed
|N |−|L| intelligence,” Telecommun. Syst., vol. 67, no. 2, pp. 231–246, Feb. 2018.
|A|−|L|
combinations for outflow. Fig. 3a shows that when [8] Y. Gao and L. Dai, “Optimal downlink/uplink throughput allocation for
there is no limitation on di (or similarly, di ui ), a smaller IEEE 802.11 DCF networks,” IEEE Wireless Commun. Lett., vol. 2,
no. 6, pp. 627–630, Dec. 2013.
|A| leads to a strictly larger rmax than does a greater |A|, [9] Q. Liao, “Dynamic uplink/downlink resource management in flexible
regardless of the node combinations. In the other three fig- duplex-enabled wireless networks,” in Proc. IEEE ICC Workshops, Paris,
ures where di are limited, a smaller |A| can only guarantee a France, 2017, pp. 625–631.
476 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Interleave-division multiple access (IDMA) is a nonfading additive white Gaussian noise (AWGN) channels.
multiple access scheme that has been considered in several recent We outline several practical matching techniques including
proposals for the 5th generation cellular system. In this letter, modulation, power control, repetition coding and zero padding.
basing on evolution analysis, we show that the performance of
IDMA can be enhanced using the transfer function matching
Incidentally, we also show that zero padding together with
principle. Such matching can be realized by superposition coded cyclic shifting can reduce the implementation cost related to
modulation, power control, repetition coding, and zero padding. user-specific interleaving in IDMA. Our analysis is based on
Zero padding together with cyclic shifting also leads to reduced AWGN channels and we will provide experimental results for
implementation complexity. Our analysis is based on additive fading channels. We will show that the proposed techniques
white Gaussian noise channels and we show by simulations that can provide noticeable performance enhancement.
the matching techniques can also provide impressive performance
in fading channels.
II. S YSTEM M ODEL AND E VOLUTION A NALYSIS
Index Terms—IDMA, evolution technique, system design.
A. Transmitter Principles
I. I NTRODUCTION Consider a K-user up-link multiple access system with
NTERLEAVE-DIVISION multiple access (IDMA) [1] is received symbols:
I inspired by the success of low-density parity-check (LDPC)
codes [2]. Recently, IDMA has been discussed for the 5th y(j ) =
K
hk xk (j ) + η(j ), j = 1, 2, . . . , J , (1)
generation (5G) cellular system [3]–[5]. k =1
For LDPC codes, decoding performance can be optimized where hk is the channel coefficient of user k, xk (j ) a transmit-
by matching the transfer functions of local decoders [2]. This ted symbol, η(j ) a complex AWGN sample with mean zero
matching principle was later extended to different iterative and variance σ 2 , and J the frame length. We assume an under-
systems [6]–[8]. An IDMA receiver also involves two local lying orthogonal frequency division multiplexing (OFDM)
processors named as, respectively, an elementary signal esti- layer that resolves the intersymbol interference problem and a
mator (ESE) and a decoder (DEC). (See Section II.) As quasi-static channel that remains unchanged over a frame.
shown in [9], the performance of IDMA can be improved The principle of IDMA is illustrated graphically in Fig. 1.
by tuning an underlying LDPC code for better matching The graph is randomized with user-specific interleaving, which
between ESE and DEC. There are, however, some obstacles is illustrated in Fig. 1 by the shuffled edge connections
for this strategy. First, in 5G, the LDPC code used has already between {ck (j )} and {xk (j )}. Fig. 1 can be seen as a graphic
been specified [10] so other alternatives, instead of alter- extension of a single-user LDPC code to a multiuser system.
ing code structure, should be used for system optimization. The randomness resulting from interleaving reduces short
Second, there is also a lack of efficient matching method when cycles in the graph, which facilitates low-cost message passing
high order modulation is involved for high rate applications. decoding. More details can be found in [1] and [11].
Third, matching for multiuser systems is generally a difficult Fig. 1 involves a user-specific interleaver for each user. This
problem. Very limited progress is made in this direction. interleaver can be combined with the inherent interleaver in
In this letter, we consider IDMA system design in high the LDPC code involved. This is equivalent to the scheme
sum-rate situations. We first derive the achievable rate for in [12], in which each user employs a unique interleaver for
IDMA using the matching principle. We show that, with its code. Later we will show that such an interleaver can be
perfect matching, IDMA is potentially capacity approaching in realized by cyclic shifting (see Fig. 5 below), which further
Manuscript received July 19, 2018; revised October 5, 2018; accepted
reduces the hardware implementation cost for IDMA.
October 9, 2018. Date of publication October 17, 2018; date of current version
April 9, 2019. This work was supported in part by the Research Fund of ZTE B. Receiver Principles
Corporation, and in part by the University Grants Committee of the Hong
Kong Special Administrative Region, China, under Project CityU 11280216 We divide an iterative detector for the system in Fig. 1 into
and Project CityU 11216817. The associate editor coordinating the review of two local processors: an ESE and a DEC. The iterative process
this paper and approving it for publication was W. Zhang. (Corresponding
author: Lei Liu.)
is outlined below.
Y. Hu, C. Liang, L. Liu, and L. Ping are with the Department of Initialization: Assume that the modulation constellation of
Electronic Engineering, City University of Hong Kong, Hong Kong (e-mail: {xk (j )} is with zero mean and unit average power. Then
yhu228-c@my.cityu.edu.hk; chuliang@cityu.edu.hk; leiliu@cityu.edu.hk; E(xk (j )) and Var(xk (j )) are respectively initialized to 0 and
eeliping@cityu.edu.hk).
C. Yan and Y. Yuan are with the Algorithm Department, ZTE 1, ∀k , j .
Corporation, Shenzhen 518057, China (e-mail: yan.chunlin@zte.com.cn; ESE Operations: We rewrite (1) as
yifei.yuan@ztetx.com).
Digital Object Identifier 10.1109/LWC.2018.2876538 y(j ) = hk xk (j ) + ζk (j ), (2a)
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HU et al.: IDMA IN HIGH RATE APPLICATIONS 477
Fig. 1. A factor graph of a 2-user IDMA system with LDPC coding. J = 8. {ck (j ), j = 1, 2, . . . , J } is a codeword that is interleaved and modulated to
produce {xk (j )}. Circles represent variables and squares constraints. Three types of constraints are presented: a white square for an LDPC coding constraint,
a square with “×” for a modulation constraint and a square with “+” for a multiple access constraint defined in (1).
R EFERENCES
[1] L. Ping, L. Liu, K. Wu, and W. K. Leung, “Interleave division multiple-
access,” IEEE Trans. Wireless Commun., vol. 5, no. 4, pp. 938–947,
Apr. 2006.
[2] T. Richardson and R. Urbanke, Modern Coding Theory. Cambridge,
U.K.: Cambridge Univ. Press, 2008.
[3] LLS Results for RDMA, GOCA, RSMA and IDMA Schemes, docu-
Fig. 5. An example of CSZP. K = 4. Two SCM layers are used. ment R1-167536, 3GPP TSG RAN WG1 Meeting #86, MediaTek Inc.,
Gothenburg, Sweden, Aug. 2016
[4] On the Performance of IDMA and RSMA Multiple Access Schemes,
document R1-167561, 3GPP TSG RAN WG1 Meeting #86, InterDigit.
Commun., Gothenburg, Sweden, Aug. 2016.
[5] C. Yan, Z. Yuan, W. Li, and Y. Yuan, “Non-orthogonal multiple
access schemes for 5G,” ZTE Commun., vol. 14, no. 4, pp. 11–16,
Oct. 2016.
[6] K. Bhattad and K. R. Narayanan, “An MSE-based transfer chart
for analyzing iterative decoding schemes using a Gaussian approx-
imation,” IEEE Trans. Inf. Theory, vol. 53, no. 1, pp. 22–38,
Jan. 2007.
[7] K. Wu, K. Anwar, and T. Matsumoto, “BICM-ID-based IDMA using
extended mapping,” IEICE Trans. Commun., vols. E97–B, no. 7,
pp. 1483–1492, Jul. 2014.
[8] X. Yuan, L. Ping, C. Xu, and A. Kavcic, “Achievable rates of
MIMO systems with linear precoding and iterative LMMSE detec-
tion,” IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7073–7089,
Nov. 2014.
Fig. 6. IDMA vs. SCMA in TDL-C channels. 3GPP NR LDPC code is used [9] Y. Zhang, K. Peng, and J. Song, “Enhanced IDMA with rate-compatible
for both IDMA and SCMA. IDMA involves SCM (with power levels of 0.865 raptor-like quasi-cyclic LDPC code for 5G,” in Proc. IEEE GC Wkshps,
and 0.135 for both cases) and CSZP. performances of SCMA are from [17]. Singapore, Dec. 2017, pp. 1–6.
[10] Technical Specification Group Radio Access Network; NR; Multiplexing
and Channel Coding (Release 15), V2.0.0, 3GPP Standard TS 38.212,
Dec. 2017.
that a random interleaver is used in the underlying LDPC [11] Y. Hu, C. Liang, J. Hu, and L. Ping, “Low-cost implementation tech-
code. The user-specific shifting in Fig. 5 provides an equiv- niques for interleave division multiple access,” IEEE Wirelee Commun.
Lett., to be published. [Online]. Available: https://ieeexplore.ieee.org/
alent realization for the user-specific interleaving in Fig. 1, abstract/document/8398463
since a shifted version of a random interleaver can be seen [12] R. Zhang, L. Xu, S. Chen, and L. Hanzo, “Repeat accumulate code
as a different random interleaver [11]. Fig. 5 also involves a division multiple access and its hybrid detection,” in Proc. IEEE ICC,
Beijing, China, May 2008, pp. 4790–4794.
two-layer SCM scheme. To distinguish the two layers of each [13] C. Berrou, A. Glavieux, and P. Thitimajshima, “Near Shannon
user, each layer is further independently and cyclically shifted limit error-correcting coding and decoding: Turbo-codes. 1,”
within its active frame duration. Such shifting technique, both in Proc. IEEE ICC, vol. 2, Geneva, Switzerland, May 1993,
user-specific as well layer-specific, significantly reduces the pp. 1064–1070.
[14] S. T. Brink, “Convergence behavior of iteratively decoded parallel con-
hardware implementing cost for interleavers in IDMA. catenated codes,” IEEE Trans. Commun, vol. 49, no. 10, pp. 1727–1737,
Fig. 6 shows the FERs with SCM and CSZP in the TDL- Oct. 2001.
C fading channels [17]. Two antennas are assumed at the [15] D. Guo, S. Shamai, and S. Verd, “Mutual information and minimum
mean-square error in Gaussian channels,” IEEE Trans. Inf. Theory,
receiver. We consider the following two settings: vol. 51, no. 4, pp. 1261–1282, Apr. 2005.
• Case 1: Jinfo = 616, K = 10 and sum-rate = 7.13; [16] L. Ping, J. Tong, X. Yuan, and Q. Guo, “Superposition coded modulation
• Case 2: Jinfo = 496, K = 12 and sum-rate = 6.89. and iterative linear MMSE detection,” IEEE J. Sel. Areas Commun.,
The power levels of the two SCM layers given in the caption vol. 27, no. 6, pp. 995–1004, Aug. 2009.
[17] Preliminary LLS Evaluation Results for mMTC Scenario, docu-
of Fig. 6 are obtained by exhaustive searching. Iterative linear ment R1-1803665, 3GPP TSG RAN WG1 Meeting #92bis, Huawei,
MMSE detection [11] is used at the receiver. As reference, HiSilicon, Shenzhen, China, Apr. 2018.
480 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—This letter considers a new secure communication artificial-noise-aided secure scheme (ANAS) for the practical
scenario where a full-duplex transmitter (Alan) needs to trans- secure communication scenario where a FD transmitter (Alan)
mit confidential information to a half-duplex receiver (Bob), with needs to transmit confidential information to a HD receiver
an eavesdropper (Eve) that tries to overhear the confidential (Bob), with an eavesdropper (Eve) that tries to overhear the
information. For realizing the secure communication, we design information.
an effective artificial-noise-aided secure scheme (ANAS) which is Specifically, an effective two-phase ANAS scheme is
composed of two phases’ transmissions: in Phase 1, Alan and Bob
proposed: in Phase 1, Alan transmits an AN nA and Bob sends
transmit two independent artificial noises (ANs) simultaneously,
while in Phase 2, Alan superimposes the AN received in Phase 1 another independent AN nB simultaneously, while in Phase 2,
with its confidential signal and sends out the mixed signal. Since Alan superimposes the signal received in Phase 1 with the con-
the superimposed AN by Alan in Phase 2 can be effectively can- fidential signal and sends it to Bob. After Phase 2, recall that
celled by Bob while remains an interference to Eve, a secrecy rate Bob knows nB , it could cancel nB from the received signal.
could then be achieved. Importantly, we derive the approximate By contrast, in Phase 1 Eve only receives the mixture of nB
closed-form solutions of the average secrecy rate and secrecy out- and nA , and thus could not cancel nB from the signal received
age probability of ANAS under a Rayleigh block-fading channel. in Phase 2, leading to severer more interference than Bob. By
Numerical results show that the secrecy rate of ANAS is about doing so, a secure communication channel can be estblished.
twice higher than the benchmark scheme, even though in ANAS As can be seen, our ANAS scheme is simple to implement and
half of the time is used to transmit ANs. has no risk of the AN-leakage problem in which the injected
Index Terms—Physical-layer security, artificial noise, secrecy AN could be overheard by the eavesdropper [9], [10].
rate, outage probability, full-duplex. Importantly, we move forward to analyze the average
secrecy rate and the secrecy outage probability of ANAS under
Rayleigh block-fading channel. Through Gauss Laguerre
I. I NTRODUCTION quadrature, the approximate closed-form expressions of the
ECENTLY, with the progress of full-duplex (FD)
R communications, the adoption of the FD technique to
realize secure communications at the physical layer has been
average secrecy rate and the secrecy outage probability are
derived. Numerical results show that ANAS could achieve
good secrecy performance in terms of both secrecy rate and
widely studied [1]–[8]. However, most prior work considered outage probability, and thus provides a good solution to the
a scenario in which either the relay or the receiver has the secure communication from a FD transmitter to a HD receiver.
FD communication capability, and then the relay or receiver
could receive the information signals and transmit artificial II. S YSTEM M ODEL AND THE ANAS S CHEME
noises (AN) simultaneously to puzzle the eavesdropper. A. System Model
It is important to note that in practical downlink transmis- We consider a wireless communication system with three
sions, the transmitter (e.g., the base station in cellular network communication nodes: a transmitter with FD capability (Alan),
or the access point in WLAN) usually has more powerful capa- a HD intended receiver (Bob) and a HD eavesdropper (Eve).
bilities (e.g., supporting the FD mode) than the user devices. We assume all nodes are equipped with single antenna. In the
Thus, for providing secure downlink services it is also non- system, Alan needs to send confidential information to Bob,
trivial to establish a secure channel from a FD transmitter and Eve tries to eavesdrop the information.
to a half-duplex (HD) receiver. Observing that most prior Let h1 and h3 be the channel coefficients of the main
schemes such as [1]–[8] cannot be applied to the downlink channel (between Alan and Bob) and the eavesdropper chan-
communication scenario, in this letter we design an effective nel (between Alan and Eve), respectively, h2 be the channel
coefficient between Bob and Eve, and hSI be the residual
Manuscript received September 10, 2018; revised October 9, 2018; accepted
October 11, 2018. Date of publication October 17, 2018; date of current self-interference (RSI) channel coefficient of Alan, which is
version April 9, 2019. This work was supported by the National Natural modeled as Rayleigh block-fading [8]. We assume the reci-
Science Foundation of China under Grant 61571178 and Grant 61771315. procity of the forward and backward channels and a Rayleigh
The associate editor coordinating the review of this paper and approving it block-fading channel model where each channel coefficient
for publication was X. Zhou. (Corresponding author: Caihong Kai.) remains unchanged for a time-slot duration of T seconds. In
X. Hu and C. Kai are with the School of Computer Science and Information
Engineering, Hefei University of Technology, Hefei 230009, China, and also the system, since Alan and Bob are active nodes, by listen-
with the Anhui Province Key Laboratory of Industry Safety and Emergency ing to the pilot signal Alan (Bob) transmits, Bob (Alan) could
Technology, Hefei 230601, China (e-mail: huxinyue@mail.hfut.edu.cn; obtain h1 , and similarly Eve could obtain h2 and h3 . However,
chkai@hfut.edu.cn). whether h2 and h3 can be obtained by Alan and Bob depends
S. Zhang is with the School of Information Engineering, Shenzhen on whether Eve is a silent node.
University, Shenzhen 516080, China (e-mail: zsl@szu.edu.cn).
Z. Guo and J. Gao are with the School of Computer Science and Information B. The ANAS Scheme
Engineering, Hefei University of Technology, Hefei 230009, China (e-mail: √
guozhongyi@hfut.edu.cn; gaojun@hfut.edu.cn). Phase 1: Bob sends an artificial Gaussian noise PB nB
Digital Object Identifier 10.1109/LWC.2018.2876540 to Alan and Alan synchronously sends another independent
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HU et al.: TO ESTABLISH SECURE CHANNEL FROM FD TRANSMITTER TO HD RECEIVER: ANAS 481
√
Gaussian noises PA1 nA to prevent Eve from over-
artificial√ B. The Rate of the Eavesdropper Channel
hearing PB nB .1 We have nB , nA ∼CN (0, 1), PB and PA1 For achieving higher achievable rate, we assume Eve is
are the transmit power of Bob and Alan. smart enough and has two methods to jointly process yE 1
After Phase 1,2 denote the signals received by Alan and Eve in (2) and yE 2 in (5): 1) maximum interference cancellation;
by yA and yE 1 , respectively. We have
2) joint decoding. The details of the two methods are given
yA = h1 PB nB + hSI PA1 nA + nA1 , (1) as follows:
1) Maximum Interference Cancellation: Eve uses yE 1 in
yE 1 = h2 PB nB + h3 PA1 nA + nE 1 , (2)
(2) to reduce the interference in yE 2 as much as possible.
where nA1 and nE 1 are the additive white Gaussian
√ noises Specifically, Eve multiplies yE 1 with a complex coefficient
(AWGN) with variance N0 and the term hSI PA1 nA in (1) hx , then removes hx yE 1 from yE 2 . After cancelling the
is introduced by the RSI. √ interference, the signal obtained by Eve is
Phase 2: Instead of decoding PB nB 3 from √ yA , Alan
superimposes its confidential information signal Ps sA with yES = h3 Ps sA + βh3 h1 PB nB + βh3 nA1 + nE 2
yA , where sA ∼CN (0, 1) and Ps is the transmit power of
sA . We assume sA and yA have the same length in the time N
domain. The superimposed signal is then − hx h2 PB nB + h3 PA1 nA + nE 1 .
(8)
xA = Ps sA + βyA , (3)
hx yE 1
Δ 2 2
where β = PA2 /(|h1 | PB + |hSI | PA1 + N0 ). As can be
seen in (3), the transmit power of yA is scaled into a prede- As regarding to how hx is determined to maximize the signal-
termined power PA2 by β. For synchronizing, Alan then adds to-interference-and-noise ratio (SINR) of Eve, we have the
a pilot signal ahead of xA and sends it to Bob. Denote the following lemma:
received signals at Bob and Eve in Phase 2 by yB and yE 2 . Lemma 1: To achieve the highest SINR, Eve can select6
We have βPB |h1 h2 h3 |
yB = h1 Ps sA + βh1 h1 PB nB + βh1 hSI PA1 nA ĥx = e (θ1 +θ3 −θ2 )i (9)
g2 PB + g3 PA1 + N0
+ βh1 nA1 + nB2 , (4)
in which θ1 , θ2 and θ3 are the phases of h1 , h2 and h3 ,
yE 2 = h3 Ps sA + βh3 h1 PB nB + βh3 hSI PA1 nA respectively.
+ βh3 nA1 + nE 2 , (5) Proof: For achieving the highest SINR, ĥx must be
where nB2 and nE 2 are the AWGNs with variance N0 .
Following [4], we assume the adopted FD technique could well ĥx = argmaxhx SIN RyES , (10)
2
√ (i.e., |hSI | is very small) 4and we
suppress the self-interference
in which
neglect the term βh3 hSI PA1 nA in the rate analysis. Also,
since |h1 |2 PB is much larger than |hSI |2
PA1 and N0 in most SIN RyES =
g3 Ps
, (11)
cases, to ease analysis, we assume β ≈ PA2 /(|h1 | PB ). 2 Var (N − hx yE 1 )
√
Since
√ Bob knows PB nB , it could cancel the term EN
βh1 h1 P√ B nB from yB . By contrast, it is difficult for Eve and we have
√ detect PB nB from yE 1 because of the interference
to √ of
PA1 nA . Thus, Eve cannot cancel the βh3 h1 PB nB term EN = |βh3 h1 − hx h2 |2 PB + |hx |2 g3 PA1 + |hx |2 N0
from yE 2 , then suffers more interference than Bob, and ANAS
+ β 2 g3 N0 + N0 . (12)
thus builds up a secure channel from Alan to Bob.
From (11), we have ĥx = argminhx (EN ). It is easy to see
III. A NALYSIS ON S ECRECY R ATE AND that when the phase of βh3 h1 is equal to that of hx h2 , the
O UTAGE P ROBABILITY |βh3 h1 − hx h2 | term in (12) is the smallest. That is, θ̂x =
A. The Rate of the Main Channel θ1 + θ3 − θ2 .
√
Since Bob knows √PB nB and h1 , it can cancel the After θ̂x is determined, we can write EN as
interference term βh1 h1 PB nB from yB and obtain yBS 5
EN = (g2 PB + g3 PA1 + N0 )|hx |2 − 2β|h1 ||h2 ||h3 |PB |hx |
yBS = h1 Ps sA + βh1 hSI PA1 nA + βh1 nA1 + nB2 . (6)
+ β 2 g1 g3 PB + β 2 g3 N0 + N0 , (13)
Let g1 =|h1 |2 , g2 =|h2 |2 , g3 =|h3 |2 and gSI = |hSI |2 . The
achievable rate ofBob is then which is a convex function of |hx | and it is easy to verify that
g1 Ps βPB |h1 h2 h3 |
|ĥx | = g2 PB +g .
RB = log2 1 + P PA2
. (7) 3 PA1 +N0
PB gSI PA1 + PB N0 + N0
A2 To ease expression, we define M as
PB g2 g3
1 It is important to note that PB nB and PA1 nA are only known by M = . (14)
Bob and Alan, respectively. g2 PB + g3 PA1 + N0
2 For synchronizing, in Phase 1, a pilot signal is added ahead of
P B nB .
3 After Phase 1, both Alan and Eve will not try to decode PB nB . Instead, By submitting ĥx into (8), the achievable rate of Eve is
Alan uses (1) to randomize the confidential information, and Eve records (2)
for cancelling the interference as much aspossible. g3 Ps
RE = log2 1 + 2 . (15)
4 We note that compared with βh h
3 1 PB nB , βh3 hSI PA1 nA has β g3 N0 + N0 + PA2 g3 − PA2 M
little effect on yE 2 in (5).
5 We cannot neglect the term h h
1 SI PA1 nA in (6), since its power may
be with the same level of the power of h1 nA1 . 6 Eve could estimate h by correlating y
1 E 1 and yE 2 , and obtain ĥx .
482 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
2) Joint Decoding: We write the two observations yE 1 and Lemma 2: The average secrecy rate and the secrecy outage
yE 2 at Eve as a signal vector, and the received signal vector probability of ANAS can be approximately computed by
at Eve is E (RS ) ≈
yE 1 σ xi +g1L (σ2 xj ,σ3 xk ,σSI xl ,0)
2 2 2 2
yE = n
n n
n
− 1 2
yE 2 0.5 ωi ωj ωk ωl e σ1
•
√ √
0 √ h PB nB + h3 PA1 nA + nE 1 i=1 j =1 k =1 l=1
= Ps sA + 2 √ . (16)
RS σ12 xi + g1L σ22 xj , σ32 xk , σSI
2
xl , 0 , σ22 xj , σ32 xk , σSI
2
xl , (26)
h3 βh1 h3 PB nB + βh3 nA1 + nE 2
Eqn. (16) is a MIMO Gaussian channel model. The informa- and
tion covariance matrix of (16) is n
n n (2 x ,σ 2 x ,σ 2 x ,r
g1L σ2 i 3 j SI k s )
Pout (rs ) ≈ 1 − ωi ωj ωk e
−
σ1 2
, (27)
0 0
Cinfo = , (17) i=1 j =1 k =1
0 g3 Ps
and the interference plus noise covariance matrix of (16) is respectively, where xi isx then ith root of n-order Laguerre
polynomial Ln (x ) = en! dx d (e −x x n ), the weight ω =
n i
g2 PB + g3 PA1 + N0 βh1∗ h2 h3∗ PB xi
Cipn = . (18) 2 2 [12] and we set n = 15.
βh1 h2∗ h3 PB β 2 g1 g3 PB + β 2 g3 N0 + N0 (n+1) (Ln+1 (xi ))
Proof: The average secrecy rate, E (RS ) and the secrecy
Consequently, the achievable rate of Eve is given by outage probability, Pout (rs ) = 1 − P (g1 ≥ g1L ) of ANAS
RE 2 = log2 det I2 + Cinfo C−1ipn . (19) can be computed by
∞ ∞ ∞ ∞ g
− 12
g
− 22
g
− 32
g
− SI
σ1 σ2 σ3 σ 2
Remark 1: It is interesting and also important to note that E (RS ) = 0.5 e e e e SI •
0 0 0 g1L
after further determinant operation, the right hand side of (19)
1
and that of (15) are exactly the same. That is, no matter RS (g1 , g2 , g3 , gSI )dg1 dg2 dg3 dgSI
σ12 σ22 σ32 σSI
2
which method Eve adopts, it will obtain the same overhearing g g g g
∞ ∞ ∞ ∞ − z2 − 22 − 32 − 1L − SI
capability. = 0.5 e σ1
e σ2
e σ3
e
2
σ1
e
2σ
SI •
0 0 0 0
C. Average Secrecy Rate and Secrecy Outage Probability 1
RS (z + g1L , g2 , g3 , gSI )dzdg2 dg3 dgSI , (28)
In this section, we analyze the average secrecy rate and the σ12 σ22 σ32 σSI
2
(PA2 gSI PA1 /PB + PA2 N0 /PB + N0 ) + Ps PA2 g3 N0 /PB , (24) D. An ON/OFF Scheme for Reducing Outage Probability
and Since g2 and g3 could be unknown to Alan, we next
C = (1 − 2rs )(PA2 gSI PA1 /PB + PA2 N0 /PB + N0 ) • design a simple ON/OFF scheme similar to [13] to further
reduce the outage probability. Specifically, we set a thresh-
(N0 + PA2 g3 − PA2 M )g3 N0 . (25) old as gth (rs ) = min (g1L ) for a given rs . Only when
g2 ,g3 ,gSI
That is, g1L (g2 , g3 , gSI , rs ) defined in (22) is the lower bound g1 ≥ gth (rs ), Alan transmits confidential information to Bob
of g1 to guarantee the achievable secrecy rate is equal or at the rate of rs . The threshold is the required minimum chan-
greater than an expected secrecy rate rs . For expressing briefly, nel gain of the main channel under which Alan has chances
we use g1L instead of g1L (g2 , g3 , gSI , rs ) in the rest of this to achieve the secure rate rs , regardless of g2 and g3 . In other
letter. words, when g1 < gth (rs ) Alan can never achieve the secrecy
7 Alan needs to know g , g , g and g
1 2 3 SI for determining the instantaneous
rate rs for any g2 and g3 .
secrecy rate RS calculated by (21) in each time block and then obtain the From (22), we can examine that g1L increases with g2 and
average secrecy rate E (RS ) over a long period of time. In practical systems, gSI , and argming2 ,g3 ,gSI (g1L ) must be (g2 = 0, g3 , gSI =
if Eve is an active node, Alan and Bob could estimate g3 and g2 respectively, 0). When g2 = 0, we can prove that g1L increases with g3 ,
and Bob could periodically feedback g2 to Alan. Also, Alan could estimate
gSI by sensing the received power. However, if Eve keeps silent, Alan could thus the minimum g1L is g1L (0, 0, 0, rs ). However, g3 = 0
hardly obtain g3 and g2 . In this case, giving an expected secrecy rate rs , we is equivalent to that Eve does not exist in the system, which
then compute the secrecy outage probability Pout (rs ). does not make sense. Fortunately, although g1L increases with
HU et al.: TO ESTABLISH SECURE CHANNEL FROM FD TRANSMITTER TO HD RECEIVER: ANAS 483
2 , in which P
Fig. 1. Average secrecy rate versus Ps with different σSI A1 = Fig. 2. Outage probability versus rs with different transmit power of ANs.
PA2 = PB = 200.
interference caused by ANs at Eve becomes more effective.
g3 , there has an upper bound when g3 → ∞, and thus we Last but not least, in the ON/OFF scheme, since Alan does
set gth (rs ) = lim g1L (0, g3 , 0, rs ) as the threshold of the not transmit confidential information when g1 is below the
g3 →∞ threshold as described in Section III-D, the outage probability
ON/OFF scheme. The outage probability with the ON/OFF can be further reduced.
scheme can be approximately calculated by the Bayes’ formula
ON /OFF V. C ONCLUSION
Pout (rs ) = P (g1 ≤ g1L |g1 ≥ gth (rs ))
This letter has proposed a two-phase transmission scheme,
P (rs ) − (1 − P (g1 ≥ gth (rs )))
≈ out . (30) ANAS, to achieve secure transmission from a FD transmitter
P (g1 ≥ gth (rs )) to a HD receiver. The average secrecy rate and the outage
probability of ANAS under Rayleigh fading channel have
IV. N UMERICAL R ESULTS been analyzed and numerical results verified ANAS’s secrecy
We next present both numerical results and Monte Carlo performance. We note in this letter, we assume all the ANs are
simulations to evaluate our ANAS scheme. Monte Carlo transmitted with the same power. As a future work, it would be
simulations are based on 105 independent channel realiza- interesting to investigate the power allocation for transmitting
tions with hj ∼ CN (0, 1), j ∈ {1, 2, 3}. To serve as a ANs and the secrecy performance could be further improved.
benchmark, we give the average secrecy rate of the tradi-
R EFERENCES
tional scheme [14], in which, Alan knows g1 and g3 , and
only when g1 > g3 , Alan transmits confidential informa- [1] T. Guo, B. Wang, W. Wu, and P. Deng, “Secrecy-oriented antenna
tion to Bob with power Pbm . For fair comparison of the assignment optimization at full-duplex receiver with self-interference,”
IEEE Wireless Commun. Lett., vol. 7, no. 4, pp. 562–565, Aug. 2018.
two schemes in terms of power consumption, we set Pbm [2] S. Yan, X. Zhou, N. Yang, T. D. Abhayapala, and A. L. Swindlehurst,
to be Ps +PA1 + PA2 + PB .8 Then, the secrecy rate of the “Secret channel training to enhance physical layer security with a full-
benchmark is log2 (1 + g1 Pbm ) − log2 (1 + g3 Pbm ). duplex receiver,” IEEE Trans. Inf. Forensics Security, vol. 13, no. 11,
Fig. 1 presents the secrecy capacities versus Ps with respect pp. 2788–2800, Nov. 2018.
2 in which P
to different σSI [3] C. Liu, L.-L. Yang, and W. Wang, “Secure spatial modulation with
A1 = PA2 = PB = 200, where
9
a full-duplex receiver,” IEEE Wireless Commun. Lett., vol. 6, no. 6,
the numerical results are computed by (28), the GLQ approx- pp. 838–841, Dec. 2017.
imate results are given by (26). As can be seen, our ANAS [4] G. Chen, Y. Gong, P. Xiao, and J. A. Chambers, “Dual antenna selection
achieves much higher average secrecy rate than the bench- in secure cognitive radio networks,” IEEE Trans. Veh. Technol., vol. 65,
mark, even though in ANAS half of the time is used to transmit no. 10, pp. 7993–8002, Oct. 2016.
[5] B. Zhong and Z. Zhang, “Secure full-duplex two-way relaying networks
ANs. Also, our analytical results perfectly agree with Monte with optimal relay selection,” IEEE Commun. Lett., vol. 21, no. 5,
Carlo simulations, and the GLQ approximate computation pp. 1123–1126, May 2017.
keeps only about 0.02 bps/Hz lower than the numerical results, [6] T.-X. Zheng, H.-M. Wang, J. Yuan, Z. Han, and M. H. Lee, “Physical
also revealing high accuracy. Moreover, as σSI 2 increases, the layer security in wireless ad hoc networks under a hybrid full-/half-
2 corresponds
average secrecy rate decreases, since a larger σSI
duplex receiver deployment strategy,” IEEE Trans. Wireless Commun.,
vol. 16, no. 6, pp. 3827–3839, Jun. 2017.
to a severer residual interference in Alan’s FD transmission. [7] T.-X. Zheng, H.-M. Wang, Q. Yang, and M.-H. Lee, “Safeguarding
It is also worthwhile to note that when Ps is small, ANAS decentralized wireless networks using full-duplex jamming receivers,”
is worse than the benchmark, because the ratio of the noise IEEE Trans. Wireless Commun., vol. 16, no. 1, pp. 278–292, Jan. 2017.
power caused by RSI to Ps is large. However, as long as Ps [8] B. V. Nguyen, H. Jung, and K. Kim, “Physical layer security
schemes for full-duplex cooperative systems: State of the art
is not too small, our ANAS keeps better then the benchmark. and beyond,” IEEE Commun. Mag., to be published. [Online].
Fig. 2 shows the outage probability with respect to rs , where Available: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=
Δ 2 = 0.0001. 8337816, doi: 10.1109/MCOM.2017.1700588.
Ps = 400, PAN = PA1 = PA2 = PB and σSI [9] Y. Zou, “Physical-layer security for spectrum sharing systems,” IEEE
The Monte Carlo simulations validate our analysis and the Trans. Wireless Commun., vol. 16, no. 2, pp. 1319–1329, Feb. 2017.
GLQ approximate computation keeps high accuracy. It is [10] R. Zhang, L. Song, Z. Han, and B. Jiao, “Physical layer security for
straightforward to see that as rs increases, the outage proba- two-way untrusted relaying with friendly jammers,” IEEE Trans. Veh.
bility increases. Furthermore, as the transmit power of ANs, Technol., vol. 61, no. 8, pp. 3693–3704, Oct. 2012.
[11] A. D. Wyner, “The wire-tap channel,” Bell Syst. Tech. J., vol. 54, no. 8,
PAN , increases, the outage probability decreases, since the pp. 1355–1387, Oct. 1975.
[12] M. Abramowitz, I. Stegun, and A. Donald, Handbook of Mathematical
8 From Section II-B, by ignoring the RSI and AWGN power of Alan, the
Functions, New York, NY, USA: Dover, 1970.
average transmit power of ANAS is (Ps +PA1 + PA2 + PB )/2. In the [13] S. Yan, N. Yang, I. Land, R. Malaney, and J. Yuan, “Three artificial-
benchmark we have P (g1 > g3 ) = 0.5. Thus, for fair comparison, we set noise-aided secure transmission schemes in wiretap channels,” IEEE
Pbm to be Ps +PA1 + PA2 + PB . Trans. Veh. Technol., vol. 67, no. 4, pp. 3669–3673, Apr. 2018.
9 The transmit powers are normalized with AWGN power N = 1. That is, [14] P. K. Gopala, L. Lai, and H. E. Gamal, “On the secrecy capacity of fad-
0
the transmit power P means the transmit power is 10log10 P dB higher than ing channels,” IEEE Trans. Inf. Theory, vol. 54, no. 10, pp. 4687–4698,
the AWGN power. Oct. 2008.
484 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—This letter studies digital/analog hybrid precoding antenna elements via the same amount of phase shifters. This
for wideband millimeter wave (mmWave) systems with multi- results in high energy consumption for a large-scale antenna
subarray architecture. By exploiting the mmWave propagation array. Hence, hybrid precoding with multi-subarray structure
property that the dominate signal energy arrives at the receivers where each RF chain is only connected with a pre-determined
through some concentrated directions, we first present the sin-
gle carrier transmission with time delay compensation technique collection of antennas, tends to be more attractive for some
performed at the base station, which can efficiently suppress mmWave communication scenarios [5].
inter-symbol interference. Then, in order to address the number Studies on hybrid precoding in wideband mmWave
of radio-frequency chains limitation, we propose a joint subarray frequency selectivity systems are relatively limited.
selection and precoding design scheme by using the group sparse In [6] and [7], hybrid precoding over mmWave multiple
approach. Numerical results also demonstrate the effectiveness input multiple output with orthogonal frequency division
of the proposed scheme. multiplexing (MIMO-OFDM) systems was considered.
Index Terms—Millimeter wave systems, single-carrier, com- However, in the OFDM mode, it is difficult to design a
pressive sensing, time delay compensation. common analog precoder shared across all the subcarriers.
Besides, OFDM transmission usually has high peak to average
I. I NTRODUCTION power (PAPR) and is power inefficient compared with single
ILLIMETER wave (mmWave) communication has been carrier (SC) transmission mode..
M viewed as a promising candidate for future cellular
communication systems [1], [2], owing to its abundant spec-
SC transmission schemes were proposed for lens [8], [9]
and fully connected array systems [10] where path and time
trum. Although mmWave signals typically suffer from severe delay compensation strategy are proposed to suppress the ISI
path loss and absorption, a large-scale antenna array can be in the mmWave frequency selective channels, respectively.
packed into an area with small form. This makes the imple- Compared with the subarray structures, the beamforming capa-
mentation of large-scale antenna array at the transceivers bility of lens systems is weaker due to its physical structure
possible, which can provide enough array gains to over- constraints, and the power consumption in fully connected
come the severe free-space pathloss in mmWave frequencies. architecture is much larger. Different from [8] and [9], where
Nevertheless, realizing large-scale array mmWave communi- only the multi-path sparsity of mmWave MIMO channels is
cation is non-trival. Equipping one radio-frequency (RF) chain applied for the path compensation technique, in this letter,
for each antenna element (the conventional fully-digital array by exploiting the channel characteristic of some mmWave
systems) generally means high energy and cost consumption. scenarios that the line-of-sight (LoS) component of signals
Therefore, a hybrid architecture with fewer RF chains than is dominate and the non-LoS (NLOS) component has much
antennas, has been proposed to overcome the issue. lower energy, we study the hybrid precoding for wideband
Recently, hybrid architecture based two-layer processing has mmWave multi-subarray systems with frequency selective
been studied for mmWave systems [3], [4]. For the fully con- MIMO channels. Specifically, we propose the SC with time
nected architecture, hybrid precoding is to divide the precoding delay pre-compensation approach to suppress ISI for LoS
into analog and digital ones so as to balance the system paths between all mobile stations (MS) and BS. For the resid-
performance and RF chain overhead, where the analog and ual ISI and inter-user interference (IUI) due to the arrival of
digital precoding are generally implemented by phase shifters weak energy in NLOS directions at the receivers, we formulate
and dedicated digital baseband hardware, respectively. In the an optimization problem to maximize the minimum signal-
fully-connected architecture, each RF chain is connected to all to-interference-plus-noise-ratio (SINR) of all users. Different
from the two layer hybrid precoding designs in [10], for
Manuscript received October 5, 2018; accepted October 14, 2018. Date of the considered subarray structure, the extra subarray selection
publication October 22, 2018; date of current version April 9, 2019. This
work was supported by the National Natural Science Foundation of China matrix results in a nonlinear mixed integer problem. To tackle
under Grant 61720106003. The associate editor coordinating the review of the issue, we transform the problem into the sparse precoding
this paper and approving it for publication was R. Wang. (Corresponding optimization design problem and solve it by relaxing l0 norm
author: Yongming Huang.) by the l1,∞ norm.
W. Huang, Y. Huang, and L. Yang are with the School of
Information Science and Engineering, Southeast University, Nanjing Notations: In this letter, boldface lower- and upper-case
210096, China (e-mail: huangwei2013@seu.edu.cn; huangym@seu.edu.cn; letters denote vectors and matrices, respectively. CM ×N
lxyang@seu.edu.cn). denotes the space of M × N complex-valued matrices. ||x||0 ,
Z. Lu is with the Algorithm Department, Wireless Product Research
and Development Institute, Wireless Product Operation, ZTE Corporation,
||x||1 , ||x||1,∞ and ||x|| denote the l0 , l1 , l1,∞ and l2
Shenzhen 518000, China (e-mail: lu.zhaohua@zte.com.cn). norm of vector x, respectively. Re(·) denotes the real part
Digital Object Identifier 10.1109/LWC.2018.2877358 of variable.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HUANG et al.: HYBRID PRECODING FOR SC WIDEBAND MULTI-SUBARRAY mmWAVE SYSTEMS 485
derived SINR expression in (9), we formulate to maximize the then, the sparse optimization problem (13) can be equivalently
minimum SINR received by any of K users via joint subarray written as
selection and precoding design under the limited RF chains
environment, which can be expressed as max t
{w̄k }K
k =1 ,t
max min SINRk s.t. SINRk ≥ t, ||w̄k ||2 ≤ P , ||ẅ||0 ≤ MRF .
F,{Λk ,wk }K
k =1
K
(14)
s.t. Λk ∈ {0, 1}, F ∈ Farray , ||FΛk wk ||2 ≤ P , Problem (14) is non-convex since both SINR and l0 norm con-
k =1 straints are non-convex. Nevertheless, when the target value
(10) t ≥ 0 is fixed, we obtain the following feasibility problem
where Farray denotes the beam set transmitted by all arrays Find : {w̄k }K
k =1
and P denotes the maximum transmit power at the BS. The
s.t. SINRk ≥ t, ||w̄k ||2 ≤ P , ||ẅ||0 ≤ MRF . (15)
switch network makes the formulated problem a mixed inte-
ger nonlinear programming problem, which is a challenging When the target value t is attainable, problem (15) is feasible
problem because of several difficulties, including non-convex and the optimal value t can be obtained by bisection searching
objective function, combinatorial characteristic of subarray method. Therefore, optimal solution of problem (14) can be
selection matrix and non-linear coupling of subarray selection derived by solving the following feasibility problem:
matrix, digital and analog precoders.
Usually, the analog precoder F is selected from a predefined min ||ẅ||0 s.t. SINRk ≥ t, ||w̄k ||2 ≤ P . (16)
{w̄k }K
k =1
codebook [11]. In this letter, the number of subarrays is larger
than that of RF chains and the beams transmitted by subarrays Because of the non-convexity of l0 norm, problem (16) is still
only align at LoS paths of all MSs so the analog precoding difficult to be directly solved. A regularization norm can be
matrix can use the DFT codebook F̄. Furthermore, we use used to promote sparsity for all entries of w̃n . We can approx-
a large dimension precoding vector w̄k = Λk wk ∈ CN ×1 imate the ||ẅ||0 norm by ||ẅ||1 . According to the definition
instead of the original MRF baseband precoding vector. Then, of ẅ in (12), we have
we rearrange the original precoding matrix W̄ ∈ CN ×K N N
W̄ = [w̄1 , . . . , w̄k ] = [ w̄11 , . . . , w̄1K ; . . . ; w̄N 1 , . . . , w̄NK ], (11) ||ẅ||1 = ||w̃n ||∞ = max |w̄nk |. (17)
k
n=1 n=1
w̃1 w̃N
Note that in (17), the l1,∞ norm of the original precoding
where w̃n ∈ C1×K denotes the precoding coefficient corre- matrix W̄ defined in (11) is able to be replaced by the group
sponding to the nth codeword. Next, we define the N × 1 sparse inducing norm of vector ẅ. Therefore, optimization
vector problem (16) can be further approximated as
ẅ = [||w̃1 ||q , ||w̃2 ||q , . . . , ||w̃n ||q , . . . , ||w̃N ||q ], (12) min ||ẅ||1 (18a)
{w̄k }K
k =1
where q ≥ 2, in this letter, we set p = ∞. Obviously, we do ⎛ ⎞
not wish to select the nth codeword from the codebook, so the Lk
corresponding precoding coefficients for all entries of the vec- s.t. t ∗ ⎝ |h̄H 2
k ,l w̄k | + |h̄H 2 2⎠
k ,l w̄k | + σk
tor w̃n need to be set to zero simultaneously and problem (10) l=LoS k =k l=1
is recast into the sparse precoding design problem ≤ |h̄k ,LoS w̄k |2 ,
H
(18b)
max min SINRk K
{wk }K ||w̄k ||2 ≤ P. (18c)
k =1 k =1
K
Up to now, the cost function in (18) is convex. Observing
s.t. ||ẅ||0 ≤ MRF , ||w̄k ||2 ≤ P , (13) that there exists an optimal precoding vector {w̄k }K
k =1 to (18)
k =1 that makes h̄H w̄
k ,LoS k
a real number. Consequently, the SINR
Note that in (13), the l0 norm constraint ||ẅ||0 ≤ MRF guar- constraint in (18b) is able to be written as a second order cone
antees that the number of the selected subarrays is less than form. Then, problem (18) can be equivalently rewritten as
that of the total RF chains at the BS. Moreover, the SINRk is
defined as min ||ẅ||1
{w̄k }K
k =1
|h̄H
k ,LoS w̄k |
2
SINRk = , s.t. GH H H H
kk wk ; Gk 1 w1 ; Gk 2 w2 ; . . . ; Gk (k −1) wk −1 ;
Lk
|h̄H 2 |h̄H 2 2
k ,l w̄k | + k ,l w̄k | + σk . . . GHk (k +1) wk +1 ; GH
kK w K ; σ k
l=LoS k =k l=1
1
where the effective channel h̄H H ≤ Re h̄H k ,LoS w̄k ,
k .LoS = hk ,LoS F̄ can be viewed t
as beamspace channel [12]. Problem (13) is also hard to solve K
because of the non-convexity of objective function and l0 norm ||w̄k ||2 ≤ P , (19)
constraint. To tackle the issue, we bring in a slack variable t, k =1
HUANG et al.: HYBRID PRECODING FOR SC WIDEBAND MULTI-SUBARRAY mmWAVE SYSTEMS 487
Algorithm 1 Joint Subarrays Selection and Precoding Design paths become even worse for the proposed scheme, but MISO-
1: Initialize tmin ,tmax ; OFDM scheme can completely eliminate the ISI. In addition,
2: Repeat we also found that the proposed sparse precoding schemes
3: t = (tmin + tmax )/2; based on time delay pre-compensation in the limited number
4: Solve problem (19) and obtain optimal precoding vector of RF chains case can achieve the performance as full RF
{w̄k }K
k =1 corresponding to ẅ in (12); chains environment.
5: Judging: if ||ẅ||0 ≤ MRF In Fig. 2(b), by assuming that the BS is equipped with
then set tmin = t, MRF = 5 RF chains, it reveals that the convergence of
else Algorithm 1 for the sparse optimization precoding design.
then set tmax = t; We observe that the Algorithm. 1 rapidly converges to the
6: Until tmax − tmin ≤ ; total number of RF chains at the BS, which indicates that
7: Output: Solution {ŵk }Kk =1 . the proposed l1,∞ norm can accurately approximate the l0
norm. Furthermore, the convergence rate is faster in the fewer
number of antennas case.
V. C ONCLUSION
In this letter, we investigated the wideband mmWave com-
munication system with subarray architecture and presented
the SC with time delay compensation transmission scheme.
Then, by considering the residual ISI and IUI, we derived the
resulting SINR expression. According to the SINR, we trans-
formed the joint subarray selection and precoding problem
Fig. 2. Max-min rate versus SNR and convergence curve of algorithm 1. into the sparse precoding design problem, which overcomes
the limited number of RF chains. Additionally, simulation
results demonstrated the significant effectiveness of the SC
where GH
kk ∈ C
(Lk −1)×N and GH ∈ CLk ×N are defined as
kk
scheme.
GH H H H H H
kk = [h̄k ,1 ; h̄k ,2 ; . . . ; h̄k ,LoS−1 ; h̄k ,LoS+1 ; . . . h̄k ,Lk ], R EFERENCES
GH
kk = [h̄H H H
k ,1 ; h̄k ,2 ; . . . ; h̄k ,Lk ].
[1] Y. Huang, J. Zhang, and M. Xiao, “Constant envelope hybrid precoding
for directional millimeter-wave communications,” IEEE J. Sel. Areas
Commun., vol. 36, no. 4, pp. 845–859, Apr. 2018.
Note that problem (19) is convex and the complete [2] C. Zhang, Y. Huang, Y. Jing, S. Jin, and L. Yang, “Sum-rate anal-
algorithm is summarized in Algorithm 1, where the ysis for massive MIMO downlink with joint statistical beamforming
tmax and tmin represent the upper and lower bound, and user scheduling,” IEEE Trans. Wireless Commun., vol. 16, no. 4,
respectively. pp. 2181–2194, Apr. 2017.
[3] X. Zhang, A. F. Molisch, and S.-Y. Kung, “Variable-phase-shift-based
RF-baseband codesign for MIMO antenna selection,” IEEE Trans. Signal
Process., vol. 53, no. 11, pp. 4091–4103, Nov. 2005.
IV. S IMULATION R ESULTS [4] Z. Xiao, P. Xia, and X.-G. Xia, “Codebook design for millimeter-
In this section, we verify the performance of our developed wave channel estimation with hybrid precoding structure,” IEEE Trans.
Wireless Commun., vol. 16, no. 1, pp. 141–153, Jan. 2017.
scheme. We assume that each MS is equipped with only one [5] S. Park, A. Alkhateeb, and R. W. Heath, “Dynamic subarrays for hybrid
antenna, whereas the total number of antennas at the BS is precoding in wideband mmWave MIMO systems,” IEEE Trans. Wireless
M = 64 and further assume that the number of subarrays is Commun., vol. 16, no. 5, pp. 2907–2920, May 2017.
equal to that of RF chains N = MRF = 8. Besides, the carrier [6] A. Alkhateeb and R. W. Heath, “Frequency selective hybrid precoding
for limited feedback millimeter wave systems,” IEEE Trans. Commun.,
frequency of mmWave system is set to the 28 GHz and the vol. 64, no. 5, pp. 1801–1818, May 2016.
number of channel paths for each MS has Lk = 3. Moreover, [7] C. G. Tsinos, S. Maleki, S. Chatzinotas, and B. Ottersten, “On the
the time delay index is uniformly distributed in [0, Tm ], where energy-efficiency of hybrid analog–digital transceivers for single- and
Tm = 100 ns denotes the maximum time delay and the total multi-carrier large antenna array systems,” IEEE J. Sel. Areas Commun.,
vol. 35, no. 9, pp. 1980–1995, Sep. 2017.
available bandwidth 500 MHz. Besides, the initial upper bound [8] Y. Zeng, L. Yang, and R. Zhang, “Multi-user millimeter wave
tmax and lower bound tmin are set to 50 and 0 in Algorithm 1, MIMO with full-dimensional lens antenna array,” IEEE Trans. Wireless
respectively. Commun., vol. 17, no. 4, pp. 2800–2814, Apr. 2018.
Fig. 2(a) reveals the average minimun rate for differ- [9] W. Huang, Y. Huang, Y. Zeng, and L. Yang, “Wideband millimeter
wave communication with lens antenna array: Joint beamforming and
ent schemes by assuming that the number of MS K = 5. antenna selection with group sparse optimization,” IEEE Trans. Wireless
It is shown that the proposed SC with time delay com- Commun., vol. 17, no. 10, pp. 6575–6589, Oct. 2018.
pensation scheme has a better average minimum rate than [10] W. Huang, Y. Huang, R. Zhao, S. He, and L. Yang, “Wideband mil-
the conventional MISO-OFDM scheme, that is because the limeter wave communication: Single carrier based hybrid precoding
with sparse optimization,” IEEE Trans. Veh. Technol., vol. 67, no. 10,
SC scheme can save the cyclic prefix overhead in OFDM pp. 9696–9710, Oct. 2018.
mode and effectively suppress the ISI and IUI after time [11] H. Seleem, A. I. Sulyman, and A. Alsanie, “Hybrid precoding-
delay pre-compensation and precoding processing. However, beamforming design with hadamard RF codebook for mmwave large-
scale MIMO systems,” IEEE Access, vol. 5, pp. 6813–6823, 2017.
in the high SNR area, the MISO-OFDM scheme is supe- [12] P. V. Amadori and C. Masouros, “Low RF-complexity millimeter-wave
rior to the proposed SC scheme. This is expected, with the beamspace-MIMO systems by beam selection,” IEEE Trans. Commun.,
increase of SNR, the ISI and IUI caused from the NLoS vol. 63, no. 6, pp. 2212–2223, Jun. 2015.
488 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—It is known that the spherical detector (SD) is very Recently, a multiple PU spectrum sensing method based on
efficient for the multiple primary user spectrum sensing in cogni- the spherical detector (SD) has been proposed in [2]. The
tive radios. However, numerical results showed that the decision SD performs spectrum sensing by evaluating whether the
threshold obtained by the existing approximation methods has population covariance matrix differs from a matrix propor-
high accuracy in large sample scenarios, but it is not the case tional to the identity matrix. The SD belongs to a blind
in small sample situations. In this letter, we develop SD for the
practical applications with small observations. Here, two methods
detection scheme, which performs spectrum sensing without
to calculate the decision threshold of the SD are proposed. In the any prior information on the noise power, channel gains, sig-
first approach, the accurate expression, in the form of Meijer’s nals powers and the number of PU signals. Moreover, the SD
G-function, for the probability of false-alarm is derived. Based is the optimal detector in the generalized likelihood ratio test
on this, the accurate threshold can be obtained via the numerical (GLRT) sense when the covariance matrix of the received pri-
solving when a target false-alarm probability is given. In the sec- mary signal is positive definite [1]. Although the SD is very
ond method, an approximation method is presented to lead to a efficient for the multiple PU detection, the numerical results
simple expression for the decision threshold, which enables us to showed that the decision threshold obtained by the existing
directly compute the threshold with low computation complexity approximation methods has high accuracy in large sample sce-
and high accuracy. Simulation results verify the effectiveness of narios, but it is not the case in small sample situations [3]. The
the two proposed methods.
latter makes the whole decision-making process of the SD
Index Terms—Cognitive radio (CR), spectrum sensing, inaccurate when only small observations are available due to
multiple primary users, spherical detector (SD), decision time-varying channel and/or detection in the shortest possible
threshold. time [4].
In this letter, we focus on the SD for the more practi-
cal applications with small observations. Here, two methods
are proposed to compute the theoretical decision threshold
I. I NTRODUCTION for the SD. In the first approach, the accurate expression
PECTRUM sensing is one fundamental and key technol- using Meijer’s G-function is derived for the probability of
S ogy in cognitive radio networks (CRNs). Existing work
on spectrum sensing is mainly based on the assumption that
false-alarm. Based on this, the accurate threshold for any
parameter setting can be obtained via the numerical solving.
only one active primary user (PU) shares the licensed spec- In the second method, an approximate but simple expression
trum. However, the single PU assumption may be too simple for the sensing threshold is provided, which can be utilized
in some practical CRNs, where multiple primary users may to directly calculate the threshold for any given false-alarm
share the same frequency band [1], [2]. probability. The second approach has much less computa-
Using existing single PU spectrum sensing algorithms in tion complexity than the first one, which is very attractive
the multiple PU detection scenarios cannot obtain optimal for real-time applications. Simulation results show that, com-
performance. As a result, the multiple PU spectrum sens- pared with the classical Beta and chi-square approximation
ing problem has attracted the attention of many researchers. methods [2], [5], our approximation method yields more accu-
rate decision threshold for the sensing scenarios with small
Manuscript received May 12, 2018; revised July 30, 2018 and observations.
September 22, 2018; accepted October 18, 2018. Date of publication
October 22, 2018; date of current version April 9, 2019. This work
was supported in part by the National Natural Science Foundation of II. S YSTEM M ODEL
China under Grant 61861019 and Grant 61362018, in part by the
Post-Doctoral Scientific Research Project of Jiangsu Province under Grant Consider K sensors cooperatively detect the presence of
1402041B, and in part by the Project of Hunan Provincial Department P (P ≥ 1) PU signals [2]. The received signal is denoted
of Education under Grant 16A174. The associate editor coordinating the as x = Hs + η, where x ∈ CK , H = [h1 , h2 , . . . , hP ] is the
review of this paper and approving it for publication was P. Pawelczak. K × P channel gain matrix between the K sensors and the
(Corresponding author: Kejun Lei.)
X. Yang is with the College of Information Science and Engineering,
P PUs, s = [s1 , s2 , . . . , sP ]T is zero mean PU signal vec-
Jishou University, Jishou 416000, China, and also with the National Mobile tor, and η is the complex Gaussian noise vector with zero
Communications Research Laboratory, Southeast University, Nanjing 210096, mean and covariance matrix σ 2 I. Without loss of general-
China (e-mail: ynkej@163.com). ity, we assume that PU signals and noise are independent
K. Lei, L. Hu, and S. Li are with the College of Information of each other. Let H1 (H0 ) denote the presence (absence)
Science and Engineering, Jishou University, Jishou 416000, China (e-mail:
leikejun-123@163.com; huli2000neu@163.com; dawning2008@qq.com). of PU signals. Denote X = [x1 , x2 , . . . , xN ] as the received
S. Peng is with the School of Information Science and Engineering, Huaqiao data matrix. Here, N denotes the number of observations. Note
University, Xiamen 361021, China (e-mail: peng.shengliang@hqu.edu.cn). that the population covariance matrix can be written
as Σ
X. Cao is with the National Mobile Communications Research Laboratory, E{XX† }/N = σ 2 I under H0 , and Σ = σ 2 I + P p=1 p hp hp
γ †
Southeast University, Nanjing 210096, China (e-mail: cao_xy@seu.edu.cn).
†
Digital Object Identifier 10.1109/LWC.2018.2877361 under H1 , where γp = E{sp sp } is the transmission power of
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
YANG et al.: THRESHOLD SETTING FOR MULTIPLE PU SPECTRUM SENSING VIA SD 489
the p-th PU. Invoking the results of [2], the decision rule can where the last equality results from the fact that 0 < TSD ≤ 1.
be written as Substituting (9) into (10), we get
H0 γ
det(R̂x ) K −1,0 a1 , . . . , aK −1
K 1
m,n a1 ,...,an ,an+1 ,...,ap
where R̂x = XX† /N is the sample covariance matrix, TSD Using the result of Gp,q b1 ,...,bm ,bm+1 ,...,bq
x dx =
and γ represent the test statistic and the threshold, respectively. m,n+1 1,a1 +1,...,an +1,an+1 +1,...,ap +1
Here, the condition N > K should be satisfied to guarantee G p+1,q+1 b1 +1,...,bm +1,0,bm+1 +1,...,bq +1 x [7], the definite
the positive definiteness of R̂x [5]. integration in (11) can be further calculated as
K −1,1 1, a1 + 1, . . . , aK −1 + 1
Pf = Ω(N , K ) GK ,K γ
III. ACCURATE T HRESHOLD S ETTING FOR SD b1 + 1, . . . , bK −1 + 1, 0
n Γ(KN )K Kn Γ̃K (N + n)
E(TSD )= , (2) Plugging ai = N − 1 + K1 and bi = N − 1 − i into (12) yields
Γ̃K (N )Γ(KN + Kn)
where Γ(α) = 0∞ x α−1 e −x dx and Γ̃K (α) is the complex K −1,1 1, N + K , . . . , N + K
1 K −1
Pf = Ω(N , K ) GK ,K γ
multivariate Gamma function. Using the fact that N − 1, . . . , N − K + 1, 0
K −1,1 1, N + K , . . . , N + K
1 K −1
Γ̃K (α) = π K (K −1)/2 Γ(α)Γ(α − 1) · · · Γ(α − K + 1) (3) − GK 0 . (13)
,K N − 1, . . . , N − K + 1, 0
i=0
fX (x ) = Ω(N , K )GK −1,K −1 x (9)
b1 , . . . , bK −1
(15)
−1 Γ(N + Ki ) where ϑ(h) = ln Γ(KN + KNh), gi (h) = ln Γ(N + Nh − i ).
with Ω(N , K ) K i=1 Γ(N −i) , ai = N − 1 + i /K and Expanding ϑ(h) and gi (h) by using the Stirling’s formula for
bi = N − 1 − i . Thus the exact expression of the probability the log-Gamma function [8], we obtain the series expressions
of false-alarm is obtained as as (16) and (17)
γ (shown
at the top of the next page),
where Bn (x ) = nk=0 nk Bk x n−k and Bk is the Bernoulli
Pf = P (TSD < γ|H0 ) = P (X < γ) = fX (x )dx , (10)
0 number [6]. Usually, L = 7 is chosen to guarantee the
490 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
L
KN (1+h)− 1 1 (−1)l+1 Bl+1 (0) 1
ϑ(h) = ln KN (1 + h) 2 − KN (1 + h) + ln(2π) 2 + + O (16)
l=1 (KN (1 + h))l l (l + 1) (KN )L+1
l+1
N (1+h)−i− 1 1 L (−1) Bl+1 (−i ) 1
gi (h) = ln N (1 + h) 2 − N (1 + h) + ln(2π) +
2 +O (17)
l=1 (N (1 + h))l l (l + 1) N L+1
L
accuracy in practical applications. Plugging (16) and (17) − fχ2 (−2N ln γ) β Φ (−2N ln γ). (27)
K 2 −1 l=1 l l
into (15) yields
K 2 −1 L It is not difficult to be seen from (22) that Y converges to a
Nh
ln E(TSD ) ln(1 + h)− 2 + βl (1 + h)−l − 1 , limiting distribution Fχ2 (y) as N increases. As a result, we
l=1 K 2 −1
(18) can use the result of [10] to determine the relation between the
quantiles of FY (y) and Fχ2 (y). Let γ0 and y0 respectively
where the terms O( (KN1)L+1 ) 1
and O( N L+1 ) are omitted and K 2 −1
be the Pf quantiles of Fχ2 (y) and FY (y). Then we have
−1 K 2 −1
(−1)l Bl+1 (0) − K l K B l+1 (−i )
βl =
i=0
. (19) Pf = P (χ2K 2 −1 > γ0 ) = 1 − Fχ2 (γ0 ), (28)
l K 2 −1
l (l + 1)(KN )
Pf = P (Y > y0 ) = 1 − FY (y0 ). (29)
From (18), the approximate moment of TSD can be given as
L Using the result of [10, eq. (23)], we obtain
2
Nh − K 2−1 −l
E(TSD ) (1 + h) exp βl (1 + h) − 1 . y0 γ0 − (β1 Φ1 (γ0 ) + β2 Φ2 (γ0 ) + · · · + βL ΦL (γ0 )), (30)
l=1
(20) where the term of O(N −2 ) is omitted. Comparing (27)
Further, the above equation can be approximated as with (29), the threshold γ corresponding to Pf can be given
L
by γ = exp(−y0 /2N ). According to (30), we have
K 2 −1
Nh
E(TSD ) (1 + h)− 2 1+ βl (1 + h)−l − 1 . L
−1
l=1 γ exp 0.5N −γ0 + βl Φl (γ0 ) . (31)
l=1
(21)
Nh )| The computational complexity of the approximated method
Note that E(TSD h=−2jt = E(exp(jt(−2N ln TSD ))) is the to compute the threshold mainly comes from the computa-
characteristic function (CF) of Y −2N ln TSD , and (1 + tion of βl and Φl (γ0 )(l = 1, . . . , L). The former requires
h)−l |h=−2jt is the CF of the χ2 -distribution with 2l degrees of 3KL(L+1)
+ (4K + 3)L multiplications and
KL(L+1)
+
2 2
freedom. From (21), the cumulant distribution function (CDF) (2K + 1)L additions, and the latter requires L(L + 1) − 1
of Y can then be written as multiplications and L(L + 1) − 1 additions. As mentioned ear-
L lier, L = 7 is usually chosen to obtain a good approximation
FY (y) Fχ2 (y) + βl Fχ2 (y) − Fχ2 (y) ,
K 2 −1 l=1 K 2 −1+2l K 2 −1 result. Obviously, the computation cost for the approximated
(22) threshold is much lower than the exact one in Section III.
where Fχ2 (y) is the CDF of χ2K 2 −1 .
Compared with
K 2 −1 V. S IMULATION R ESULTS
the classical chi-square approximation in [5], the approx-
imation accuracy of FY (y) is raised from the order of Fig. 1 shows the decision threshold of the SD versus N when
O( N12 ) to O( N17 ), which helps to obtain a high-precision K = 6, the target Pf = 0.05 and 0.1. It can be seen that the
threshold in small sample scenarios. Using the properties of accurate thresholds via the numerical solution of (13) match
χ2 -distribution [9], we can verify that the Monte-Carlo results pretty well, which verifies the effec-
tiveness of our theoretical analysis in Section III. In addition,
Fχ2 (y) − Fχ2 (y) = fχ2 (y)Φl (y), (23) we can also observe that there is a little loss in accuracy of (31)
K 2 −1+2l K 2 −1 K 2 −1
compared to the accurate threshold. However, as mentioned
where earlier, the computational burden of the new approximation
2
1 −y K −3 method is relieved due to avoiding the calculation of Meijer’s
fχ2 (y) = K 2 −1
2 exp y 2 , (24) G-function. On the other hand, both the classical Beta [2]
K 2 −1 2
2 2 Γ K 2−1 and chi-square [5] approximation methods produce accurate
2 enough decision thresholds when the number of samples N is
Φl (y) = Φl−1 (y) − l (25) relatively large, however, when N is small, the error of the
q=1 (K 2 + 2q − 3)/y threshold calculation via them cannot be ignored.
with Φ0 (y) = 0. Substituting (23) into (22) yields As analyzed before, the problem of computation complexity
L and numerical stability for the accurate threshold calculation
FY (y) Fχ2 (y) + fχ2 (y) βl Φl (y). (26) via (13) is very difficult to solve when K increases. Thus,
K 2 −1 K 2 −1 l=1 the approximation method to calculate the threshold with
Note that the probability of false-alarm can be written as Pf = high accuracy is very attractive. Fig. 2 compares the decision
P (TSD < γ) = P (Y > −2N ln γ). Thus, we have thresholds obtained by the proposed approximation method
and the Beta and chi-square approximation methods when K
Pf = 1 − FY (−2N ln γ) 1 − Fχ2 (−2N ln γ) increases to 20. It is shown that the thresholds obtained using
K 2 −1
YANG et al.: THRESHOLD SETTING FOR MULTIPLE PU SPECTRUM SENSING VIA SD 491
Fig. 1. Threshold versus N when K = 6: (a) Pf = 0.05; (b) Pf = 0.1. Fig. 4. Actual Pd and Pf versus N (K = P = 20, SNR = −5 dB).
Fig. 4 shows that the actual Pd and Pf obtained using the new
approximated threshold match the Monte-Carlo results very
well, however, the gaps between the actual Pd and Pf curves
via the approximation methods in [2] and [5] and the Monte-
Carlo results are further increased. For instance, compared
with the actual performance, the detection probability via the
Beta approximation is reduced by about 36.79% when N = 25,
however, the detection probability via the chi-square approx-
imation has been falsely raised by about 14.85%. Obviously,
both of the two classical approximation methods produce very
Fig. 2. Threshold versus N when K = 20: (a) Pf = 0.05; (b) Pf = 0.1. unreliable decision results.
VI. C ONCLUSION
In this letter, two methods to calculate the threshold of the
SD have been proposed. In the first method, the accurate false-
alarm probability expression using the Meijers G-function
was derived. Based on this, the accurate threshold can be
obtained via the numerical solution. In the second method,
an approximation method with high accuracy but low com-
putation complexity was proposed to directly calculate the
threshold. Compared with the classical Beta and chi-square
Fig. 3. Actual Pd and Pf versus N (K = P = 6, SNR = 0 dB).
approximation methods, the proposed approximation method
improves the threshold accuracy and can obtain the reliable
(31) match the Monte-Carlo results very well even in the case sensing results even when the number of observations is small.
with a small number of samples. On the contrary, both the Numerical results verified the effectiveness of our methods.
classical Beta and chi-square approximation methods cannot
give the decision threshold with high accuracy when N is rel-
atively small. Furthermore, by comparing Fig. 1 and Fig. 2, it R EFERENCES
is easy to find that with increasing of K and N, the accuracy of [1] L. Wei, P. Dharmawansa, and O. Tirkkonen, “Multiple primary user
the proposed method has been further improved. The reason is spectrum sensing in the low SNR regime,” IEEE Trans. Commun.,
that the accuracy of the approximated distribution of the test vol. 61, no. 5, pp. 1720–1731, May 2013.
statistic TSD increases with increasing of K and N, which can [2] L. Wei and O. Tirkkonen, “Spectrum sensing in the presence of multiple
primary users,” IEEE Trans. Commun., vol. 60, no. 5, pp. 1268–1277,
be seen from (16) and (17). By contrast, the accuracy of the May 2012.
classical Beta and chi-square approximation methods decrease [3] D. A. Guimarães, R. A. A. de Souza, and G. P. Aquino, “Multiantenna
when K and N increase. spectrum sensing in the presence of multiple primary users over fading
Fig. 3 and Fig. 4 illustrate the effect of threshold setting on and nonfading channels,” Int. J. Antennas Propag., vol. 2015, pp. 1–14,
Jun. 2015.
the actual sensing performance of the SD. Here, we assume
[4] X. Yang, K. Lei, S. Peng, and X. Cao, “Blind detection for primary
that the received SNRs of all sensors keep the same and the tar- user based on the sample covariance matrix in cognitive radio,” IEEE
get Pf is 0.05. It is noted that the more accurate the threshold Commun. Lett., vol. 15, no. 1, pp. 40–42, Jan. 2011.
is, the closer the corresponding Pf and the detection probabil- [5] R. J. Muirhead, Aspects of Multivariate Statistical Theory. New York,
ity (Pd ) are to the simulation curves. From Fig. 3, we can see NY, USA: Wiley, 1982.
that the actual Pd and Pf are in very good agreement with the [6] I. S. Gradšhteyn, I. Ryžhik, A. Jeffrey, and D. Zwillinger, Table of
Integrals, Series, and Products, Waltham, MA, USA: Academic, 2007.
Monte-Carlo results when the accurate threshold is used. We [7] Wolfram Research, Inc. MeijerG Function. Accessed: Jul. 10, 2018.
can also see that, when N is relatively small, the new approx- [Online]. Available: http://functions.wolfram.com/PDF/MeijerG.pdf
imated threshold obtained via (31) results in that the practical [8] T. W. Anderson, An Introduction to Multivariate Statistical Analysis,
Pf is slightly lower than the target value, with a slight loss in 3rd ed. Hoboken, NJ, USA: Wiley, 2003.
the detection probability; By contrast, the decision thresholds [9] K. Krishnamoorthy, Handbook of Statistical Distributions With
Applications. Boca Raton, FL, USA: CRC Press, 2006.
via the Beta and chi-square approximation methods produce a [10] G. W. Hill and A. W. Davis, “Generalized asymptotic expansions of
non-ignorable false-alarm probability deviation, which leads Cornish-Fisher type,” Ann. Math. Stat., vol. 39, no. 4, pp. 1264–1273,
to the unreliable sensing results. When K increases to 20, Dec. 1968.
492 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—The state of physical entity in the Internet of Things to predict its state at the search moment. Then Dyser esti-
(IoT) has an obvious time-varying characteristic. Preliminarily mates the matching probability of an entity with the search
selecting candidate entities by predicting their current state when request based on the predicted state. Then Dyser would access
searching match entities from massive ones can effectively reduce
the communication overhead of IoT search system. The existing
partial entities in descending order of matching probability
methods are all based on shallow learning theories whose per- rather than all of them. Whereas Dyser can only search for
formances are very limited. Thus, a high-accuracy entity state qualitative state of entity. To meet the search needs for quan-
prediction method (HESPM) based on deep learning theory is titative state of entity, [11] presented CSS, which was based
proposed. The model of HESPM is built by utilizing the deep on the fuzzy set theory to estimate the matching probability
belief network. Then the contrastive divergence algorithm is of an entity by utilizing the quantitative sensor output that
adopted to train the model. Therefore, the dynamic evolution
trend of entity state can be accurately perceived and the future
was non-periodic. MSE, which is based on Grey system, and
entity state can be precisely predicted. Simulation results demon- MPM, which is based on LS-SVM, were respectively proposed
strate the effectiveness of HESPM in enhancing the prediction in [12] and [13] to predict the future entity state by exploring
accuracy and communication overhead performances. the temporal correlations between entity state data. However,
Index Terms—IoT, IoT search, entity state prediction, DBN. these three methods are all based on shallow learning theory,
which result in limited prediction precision of entity quantita-
tive state. It further leads to the large communication overhead
I. I NTRODUCTION and energy consumptions of IoT sensors.
ITH the in-depth applications of IoT, users have In this letter, a high-accuracy entity state prediction method
W increasing demands for conveniently and effectively
obtaining entity state information in the physical world [1]–[3],
(HESPM) is proposed. The model of HESPM is built based
on DBN [14] and the CD [15] algorithm is adopted to train
which gives birth to the IoT search service [4]. The objects the model, so as to accurately predict the future entity state
of IoT search are physical entities whose state is dynamically and effectively reduce the communication overhead of IoT
changing, while the traditional Internet search engines focus on sensor networks. To our knowledge, this is the first research
static virtual information resources in the cyberspace [5]–[7]. on entity state prediction method based on deep learning
Thus, the traditional Internet searching methods can’t be theory in IoT search. Remaining parts are organized as fol-
directly applied in IoT search. lows. Section II describes the search mechanism. Section III
Typical IoT search prototype systems, like Microsearch [8], proposes the HESPM. Section IV discusses the performance
and Snoogle [9], stored static keyword descriptions of a evaluation. Section V finally concludes this letter.
physical entity in its associated sensor, which required access-
ing all sensors in the search system, according to keyword
II. S EARCH M ECHANISM D ESCRIPTION
constraints provided by the user, to search for desired enti-
ties that meet keyword requirements. As a result, it brings As shown in Fig. 1, the IoT search system consists of client,
tremendous communication overhead to the IoT search sys- server, IoT gateway, database, sensor and entity. The entity is
tem, which will definitely consume large of energy resources the target being searched. One sensor is associated with one
of IoT sensors. Besides, it cannot search for entity dynamic entity and is responsible for observing the entity state. The
state. To reduce the communication overhead and support the server responds to the search request and issue it to appropriate
search for entity dynamic state, [10] proposed Dyser, which gateway. The IoT gateway is used for managing sensor net-
perceived the transition periodicity of entity dynamic state works, collecting entity state data reported by the sensor, and
predicting the future entity state based on HESPM method pro-
Manuscript received August 18, 2018; accepted October 10, 2018. Date of posed in Section III. The database stores the predicted future
publication October 23, 2018; date of current version April 9, 2019. This entity state. The user can publish search requests via the client.
work was supported in part by the National Natural Science Foundation
of China under Grant 61771082, in part by the Program for Innovation The IoT entity search mechanism is divided into two stages,
Team Building at Institutions of Higher Education in Chongqing under offline predicting (OP) stage and online searching (OS) stage.
Grant CXTDX201601020, and in part by the Science and Technology In the OP stage, the gateway receives entity state data and
Research Program of Chongqing Municipal Education Commission under
Grant KJQN201800615. The associate editor coordinating the review of
predicts future state. Multiple search requests can be initi-
this paper and approving it for publication was M. Velez. (Corresponding ated during the OS stage. Different from the traversal search
author: Puning Zhang.) method discussed above, our search mechanism finds match
The authors are with the School of Communication and Information entities by preliminarily predicting the entity state at the search
Engineering, Chongqing University of Posts and Telecommunications,
Chongqing 400065, China (e-mail: zhangpn@cqupt.edu.cn). moment rather than accessing all entities, which will reduce
Digital Object Identifier 10.1109/LWC.2018.2877639 vast communication overhead.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHANG et al.: HESPM BASED ON DEEP BELIEF NETWORK TOWARD IoT SEARCH 493
Fig. 1. IoT entity search mechanism. The key to predict the future state of a given entity lies
in solving the probability distribution of the entity state data
x (t − τ ), x (t − 2τ ), . . . , x (t − nτ ) as
−E (v,h|θ)
e
P (v|θ) = P (v, h|θ) = h (3)
Z (θ)
h
The computation complexity of normalization factor Z (θ)
is extremely high, so P (v|θ) cannot be calculated by solv-
ing Z (θ). However, since neurons in adjacent layers are fully
connected and contactless in the same layer, if the state vec-
tor of neurons in the visible layer v is known, the activation
probability of the j-th neuron in hidden layer is defined as
P hj = 1|v, θ = 1/ 1 + exp(−bj − vi wij ) (4)
i
As RBM has a symmetrical structure, when the state vector
of hidden layer neurons h is given, the probability of the i-th
Fig. 2. The model of HESPM. neuron in the visible layer being activated can be computed as
P (vi = 1|h, θ) = 1/ 1 + exp(−ai − hj wij ) (5)
j
III. H ESPM D ESCRIPTION
Thus the probability distribution of the entity state data can
In this part, we propose HESPM method to perceive the be estimated and the entity state data can be reconstructed.
dynamic evolution feature of entity state and accurately predict The entity state data are taken as the input of visible layer
the future entity state. neurons in RBM1 and then tune the parameters in θ of RBM1
to reconstruct the input. The output of hidden layer in RBM1
are also taken as the input of RBM2. Continue this process
and then the entity state at time t, x(t), can be predicted in the
A. Model Building
output layer as x (t). Obviously, the fine-tuning method for
The HESPM method is based on DBN theory with several solving θ in each RBM is the key to make predictions.
Restricted Boltzmann Machines (RBMs) to build its model,
which consists of the input layer, RBMs, and the output layer,
B. Model Training
as shown in Fig. 2.
The RBM is composed of the visible layer and the hid- In this part, we adopt the CD algorithm to solve θ. In
den layer. vi is the i-th neuron of visible layer and hj general, the optimal solution of θ can be obtained by max-
is the j-th neuron of hidden layer. The entity state data imizing the log-likelihood function (θ) of entity state data
x (t − τ ), x (t − 2τ ), . . . , x (t − nτ ) are used as the input of reconstructed by RBM as
visible layer where τ is the sample interval and n is the num- T
ber of neurons in the visible layer. The hidden layer can be θ∗ = arg max (θ) = arg max log P v(t) , h|θ
regarded as a feature extractor of entity state data. Neurons in θ θ t=1 h
adjacent layers are fully connected and no connections in the T
(t)
same layer. = arg max log e[−E (v ,h|θ )]
Denote the state vector of RBM visible layer neurons as v = θ t=1 h
[v1 , v2 , . . . , vn ]. The state vector of neurons in the hidden layer
is h = [h1 , h2 , . . . , hm ] where m is the number of neurons in − log e[−E (v,h|θ )] (6)
the hidden layer. Then for a given state combination (v, h), v h
494 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
Assume ϕ to be one of the parameters in θ, then take the S IMULATION PARAMETERS S ETTING
partial derivative of log-likelihood function for ϕ as
T
∂(θ) ∂ (t)
= log e[−E (v ,h|θ )]
∂ϕ ∂ϕ
t=1 h
[−E (v,h| θ )]
− log e
h v
⎛
T (t)
∂ −E v , h|θ
= ⎝
∂ϕ
t=1 P (h|v(t) , θ)
∂(−E (v, h|θ ))
− (7)
∂ϕ P (v,h|θ )
Fig. 4. Prediction performance under different predictive steps. Fig. 6. CO performances under different predictive steps.
R EFERENCES
[1] D. P. Wu, S. S. Si, S. E. Wu, and R. Y. Wang, “Dynamic trust rela-
tionships aware data privacy protection in mobile crowd-sensing,” IEEE
Internet Things J., vol. 5, no. 4, pp. 2958–2970, Aug. 2018.
[2] C. Y. Zhang and W. Zhang, “Spectrum sharing for drone networks,”
IEEE J. Sel. Areas Commun., vol. 35, no. 1, pp. 136–144, Jan. 2017.
Fig. 5. CO performances under different query ranges. [3] P. Zhang and J. Ma, “Channel characteristic aware privacy protection
mechanism in WBAN,” Sensors, vol. 18, no. 8, p. 2403, Aug. 2018.
[4] N. K. Tran, Q. Z. Sheng, M. A. Babar, and L. Yao, “Searching the Web
of Things: State of the art, challenges, and solutions,” ACM Comput.
4.18, and superior to MSE, 4.85. It indicates that HESPM has Surveys, vol. 50, no. 4, pp. 1–34, Nov. 2017.
[5] D. P. Wu, J. J. Yan, H. G. Wang, D. L. Wu, and R. Y. Wang, “Social
better multi-step prediction performance than MPM and MSE. attribute aware incentive mechanism for device-to-device video dis-
tribution,” IEEE Trans. Multimedia, vol. 19, no. 8, pp. 1908–1920,
Aug. 2017.
B. Comparisons About Search Performances [6] D. P. Wu, Q. R. Liu, H. G. Wang, D. L. Wu, and R. Y. Wang, “Socially
In this part, we compared the performances of HESPM- aware energy-efficient mobile edge collaboration for video distribution,”
IEEE Trans. Multimedia, vol. 19, no. 10, pp. 2197–2209, Oct. 2017.
based, MPM-based, and MSE-based search mechanisms under [7] P. Zhang, X. Kang, Y. Liu and H. Yang, “Cooperative willingness aware
different query range and predictive step conditions. As men- collaborative caching mechanism towards cellular D2D communication,”
tioned above, the proposed search mechanism includes the IEEE Access, to be published, doi: 10.1109/ACCESS.2018.2873662.
entity state validation process to ensure the reliability of [8] C. C. Tan, B. Sheng, H. D. Wang, and Q. Li, “Microsearch: A search
engine for embedded devices used in pervasive computing,” ACM Trans.
search results. Thus the efficiency of search mechanism is Embedded Comput. Syst., vol. 9, no. 4, pp. 1–29, Mar. 2010.
directly related to the number of entities that needs to be [9] H. D. Wang, C. C. Tan, and Q. Li, “Snoogle: A search engine for
verified, na , when Rq is given. Therefore, we define the com- pervasive environments,” IEEE Trans. Parallel Distrib. Syst., vol. 21,
no. 8, pp. 1188–1202, Aug. 2010.
munication overhead (CO) as follows. The smaller of CO, [10] K. Romer, B. Ostermaier, F. Mattern, M. Fahrmair, and W. Kellerer,
CO = 1 − Rq /na indicates the higher efficiency. “Real-time search for real-world entities: A survey,” Proc. IEEE, vol. 98,
In Figure 5, as the query range increases, the CO of the three no. 11, pp. 1887–1902, Nov. 2010.
[11] C. Truong and K. Römer, “Content-based sensor search for the
mechanisms all show gradual downward trends. Compared Web of Things,” in Proc. IEEE Glob. Commun. Conf., Dec. 2013,
with MPM and MSE, the proposed HESPM reduces the aver- pp. 2654–2660.
age CO by about 25.4% and 35.5% respectively under different [12] P. N. Zhang, Y.-A. Liu, F. Wu, and B. H. Tang, “Matching state esti-
query ranges. As seen from Figure 6, with the increase of mation scheme for content-based sensor search in the Web of Things,”
Int. J. Distrib. Sensor Netw., vol. 2015, pp. 1–15, Jan. 2015.
predictive step, the CO of these three gradually rise. We also [13] P. N. Zhang, Y. A. Liu, F. Wu, S. Y. Liu, and B. H. Tang, “Low-
see that HESPM can improve the average CO by 14.6% and overhead and high-precision prediction model for content-based sensor
21.7%, respectively, in comparison with MPM and MSE. search in the Internet of Things,” IEEE Commun. Lett., vol. 20, no. 4,
pp. 720–723, Apr. 2016.
[14] G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm
for deep belief nets,” Neural Comput., vol. 18, no. 7, pp. 1527–1554,
V. C ONCLUSION Jul. 2006.
For reducing the communication overhead of IoT search [15] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representa-
system, an entity state prediction method, HESPM. is proposed tions by back-propagating errors,” Nature, vol. 323, no. 9, pp. 533–536,
Oct. 1986.
in this letter to sense the dynamic evolution characteristic of [16] Intel Lab Data. (Oct. 2016). [Online]. Available:
entity state and predict its future state with high accuracy. http://db.lcs.mit.edu/labdata/labdata.html
496 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
High Rate CCK Modulation Design for Bandwidth Efficient Link Adaptation
Han Wang , Lianyou Jing, Chengbing He, and Zhi Ding , Fellow, IEEE
Abstract—This letter presents a novel design of high rate is to improve the bandwidth efficiency of CCK modulation in
complementary code keying (CCK) modulation scheme (named wireless communication links.
QAM-CCK) that is suitable for bandwidth efficient wireless link Another challenge in designing a new high bandwidth
adaptation. The proposed QAM-CCK can increase the bandwidth
efficiency of CCK signaling by simultaneously implementing
efficiency CCK modulation is the corresponding detection
quadrant and phase modulations. We introduce an optimal design performance. Under additive white Gaussian noise (AWGN)
criterion based on weighted minimum Euclidean distance to channels, the optimum receiver based on maximum a pos-
optimize the QAM-CCK mapping with minimum average bit teriori (MAP) detection principle achieves the minimum
error probability. In order to utilize the inherent coding gain probability of error for data symbols [6]. As is com-
of QAM-CCK, we propose an iterative soft maximum likeli- monly known, the MAP detection is simplified to a min-
hood receiver to enhance the decoding performance. Compared
with traditional QPSK of the same bandwidth efficiency, the imum distance detection under AWGN channels [7]. Thus,
proposed QAM-CCK achieves a 4-dB SNR gain under additive Euclidean distance between codewords is an important code
white Gaussian noise channels when used in conjunction with designing metric against AWGN channels. We are motivated
proposed iterative receiver. to optimize the new CCK-based modulation by maximiz-
Index Terms—Complementary code keying (CCK), modulation ing the weighted Euclidean distance between codewords.
design, link adaption, bandwidth efficiency. Moreover, inspired by the concept of turbo decoding by
Berrou et al. [8], soft information exchange has demonstrated
significant performance improvement of many MAP based
communication receivers [9]. In particular, the turbo receiver
I. I NTRODUCTION is a high performance joint equalization and decoding scheme
that successfully utilizes the principle of iterative soft infor-
N RECENT decades, there have been only a few advances
I in modulation technology to achieve high bandwidth
efficiency in wireless communications despite the tremen-
mation exchange [10]. To take advantage of the coding gain in
CCK, we are motivated to adopt the soft information exchange
in our receivers.
dous growth of data network traffic. At the start of WiFi
In this letter, we propose a high rate CCK modulation
network standardization, complementary code keying (CCK)
scheme that is bandwidth efficient to increase the information
was adopted as a high rate spread spectrum technology
bit load conveyed in each transmitted symbol as much as the
for indoor wireless local area networks (WLANs) [1] with
basic QPSK. The proposed CCK modulation is named QAM-
substantial success. Although CCK is not among the most pop-
CCK as it adopts the same constellation diagram of 16-QAM.
ular modulation techniques, practical applications in [2]–[4]
Furthermore, we propose a weighted minimum Euclidean dis-
demonstrate its advantages, such as flexible rate and inherent
tance as the design criterion to optimize the QAM-CCK
coding or spreading gain, under distorted channels.
performance. By utilizing the proposed iterative turbo receiver,
However, the achievable bandwidth efficiency of traditional
which is capable of soft-information exchange, the optimized
CCK modulation is lower than that of the commonly applied
QAM-CCK design demonstrates superior performance in both
PSK or QAM modulations without spreading. For example,
analytical and numerical simulation results.
IEEE 802.11b provided a CCK algorithm offering a rate of
The rest of this letter is organized as follows. Section II
1 bit/symbol and a higher rate of 1.5 bits/symbol was reported
describes a novel QAM-CCK modulation and correspond-
in [5]. Nevertheless, this rate is still below the 2 bits/symbol
ing turbo-structure iterative receiver. Section III presents the
rate of plain QPSK signaling. Therefore, one of our objectives
details of QAM-CCK design optimization based on a weighted
minimum Euclidean distance. Section IV provides simula-
Manuscript received August 13, 2018; accepted October 12, 2018. Date
of publication October 23, 2018; date of current version April 9, 2019. tion results of the proposed QAM-CCK modulation. Finally,
The work of H. Wang and C. He was supported by the National Natural concluding remarks are given in Section V.
Science Foundation of China under Grant 61771396 and Grant 61471298.
The associate editor coordinating the review of this paper and approving it
for publication was C. Huang. (Corresponding author: Han Wang.) II. S YSTEM M ODEL
H. Wang and C. He are with the School of Marine Science and
Technology, Northwestern Polytechnical University, Xi’an 710072, China A. Basic CCK Modulation
(e-mail: whan@mail.nwpu.edu.cn; hcb@nwpu.edu.cn).
L. Jing is with the School of Information and Communication
In the conventional CCK modulation technique, a sequence
Engineering, Dalian University of Technology, Dalian 116024, China (e-mail: of binary bits for transmission is first grouped into subse-
lyjing@dlut.edu.cn). quences of length 8 and each subsequence is encoded inde-
Z. Ding is with the Department of Electrical and Computer Engineering, pendently afterwards. Denote a vector corresponding to such a
University of California at Davis, Davis, CA 95616 USA (e-mail:
zding@ucdavis.edu). subsequence as b = [b1 , b2 , . . . , b8 ]. Each information vector
Digital Object Identifier 10.1109/LWC.2018.2877648 b is then encoded as one CCK codeword c = [c1 , c2 , . . . , c8 ]
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
WANG et al.: HIGH RATE CCK MODULATION DESIGN FOR BANDWIDTH EFFICIENT LINK ADAPTATION 497
θ = Gφ + ϕ, (2)
in which the generating matrix G is defined as 5dB. Thus, the selection and arrangement of c Q and c S are
⎡ ⎤ crucial for satisfactory performance. The detailed principle for
1 1 1 1 1 1 1 1
⎢1 0 1 0 1 0 1 0⎥ selecting c Q and c S will be presented in Section III.
GT = ⎢ ⎣1 1
⎥. (3)
0 0 1 1 0 0⎦
1 1 1 1 0 0 0 0 C. Iterative Soft Maximum Likelihood Detection
The cover code phase vector ϕ = [0 0 0 π 0 0 π 0]T rotates the QAM-CCK inherits the block coding property of conven-
fourth and seventh chips by π, respectively. ϕ optimizes the tional CCK modulation, hence the modulation/demodulation
sequence correlation properties and minimizes the DC offset process can be regarded as a coding/decoding process. After
within the codewords [11]. The mapping from the payload introducing channel encoding before QAM-CCK modulation,
information bits b to the phase vector φ = [φ1 , φ2 , φ3 , φ4 ]T the resulted system can be treated as a serially concatenated
can also be found in [11]. coding procedure. A commonly applied decoder for concate-
nated code is the iterative detector based on turbo structure.
Thus, we propose an iterative soft maximum likelihood detec-
B. Proposed QAM-CCK Modulation
tor as shown in Fig. 1 to exploit the coding property. Under
In order to increase the rate and improve the bandwidth- an AWGN channel, the received signal can be denoted as
efficiency of CCK modulation without expanding the code set
C , it is intuitive to enlarge the information bits conveyed in y = x + n, (5)
a single chip. Unlike QPSK chips adopted by conventional
CCK, we propose a new CCK modulation algorithm whose where n is the additive white Gaussian noise with variance σ 2 .
chips are QAM symbols. This new modulation technique is The iterative soft detector communicates reliability or
named QAM-CCK because of the same constellation dia- soft information between the demodulator and the decoder.
gram as 16-QAM. The proposed modulation has a rate of 2 Compared with the hard decision method, the soft iterative
bits/symbol as a result of compacting the original 8 chips into detection can elevate the confidence level by using the soft
4 chips instead. We now describe the specific modulation steps. information at the cost of additional computations, delay and
The input bits b are first divided into subsequences of length storage. The generally adopted soft information for exchange
8 and then be modulated into CCK codewords c = [c1 , . . . , c8 ] is the log likelihood ratio (LLR). The specified posteriori LLR
through a conventional CCK modulator defined in Eqs. (1)-(3). for QAM-CCK is denoted by LP (bn |y) in (6),
It is noted that c has the constellation with QPSK modulation.
P (bn = 0|y)
QAM improves the bandwidth-efficiency by simultaneously LP (bn |y) = ln
applying quadrant and phase modulations. This strategy is P (bn = 1|y)
inherited in our QAM-CCK modulation to elevate the number ∀x:bn =0 ∀φ :φ →x p(y|x, φ , bn = 0) · p(x|φ , bn = 0)
of bits conveyed in each chip. Hence, the 8 chips of each orig- = ln
∀x:bn =1 ∀φ :φ →x p(y|x, φ , bn = 1) · p(x|φ , bn = 1)
inal CCK codeword are equally divided into two groups: c Q
and c S , used for quadrant and phase modulation, respectively. ·p(φ |bn = 0) · p(bn = 0)
Because each chip has 4 possible values, c Q first defines the ·p(φ |bn = 1) · p(bn = 1)
new √ centroid from 4 different points on the circle with a radius
of 2 2 from the origin. Then c S determines one of the four ∀x:bn =0 ∀φ :φ →x p(y|x, φ , bn = 0) · p(x|φ , bn = 0)
= ln
possible phase rotations on the unit circle with respect to the ∀x:bn =1 ∀φ :φ →x p(y|x, φ , bn = 1) · p(x|φ , bn = 1)
new circle centroid selected. ·p(φ |bn = 0)
The QAM-CCK codeword x is jointly modulated by c Q + L(bn )
·p(φ |bn = 1)
and c S . This procedure can be described mathematically as
= LE (bn |y) + L(bn ), (6)
follows,
√ √
xn = 2 2 · cnQ + 2 · cnS , n = 1, 2, 3, 4, (4) where LE (bn |y) and L(bn ) represent the extrinsic LLR
and prior LLR, respectively. Under AWGN channels,
where cnQ and cnS are the chips of original CCK codewords. It p(y|x, φ, bn ) ∼ N (0, σ 2 ) is a Gaussian random variable,
is worth noting that different divisions of c Q and c S result in
diverse bit error performance. Our simulations suggest that the 1 −|x − y|2
p(y|x, φ, bn ) = √ exp( ). (7)
performance loss among different divisions can be as much as 2πσ 2σ 2
498 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Note that p(x|φ), p(φ|bn ) and L(bn ) are fixed once the code
design method is determined. LP (bn |y) is mainly affected
by p(y|x, φ, bn ), which is decided by the Euclidean distance
|x − y|2 with certain noise variance σ 2 . We are able to increase
the reliability of LP (bn |y) by enlarging the Euclidean dis-
tance between different codewords. Consequently, maximizing
Euclidean distance is beneficial for improving the detection
performance.
Fig. 3. Experimental results: (a) EXIT chart analysis (I represents the mutual information); (b) BER after three iterations for SISO system under AWGN
channels; (c) FER after three iterations for SISO system under AWGN channels.
IV. N UMERICAL R ESULTS detection and decoding to enhance the detection accuracy.
In this section, we investigate the bit-error-rate (BER) and In order to understand the convergence property of proposed
frame-error-rate (FER) performance of proposed QAM-CCK iterative receiver, we use the EXIT chart to verify that our
under AWGN channels. We set the length of each data frame receiver is able to converge with 3 iterations. Furthermore,
to 1024 and apply a rate 1/2 convolution code for forward error our simulation results demonstrate substantially improved
correction (FEC). Our Monte Carlo simulation uses 10000 BER and FER performance for the proposed QAM-CCK
transmitted frames. For performance comparison, we shall pro- design when compared against traditional QPSK modulation
vide the receiver performances of both QAM-CCK modulation or QAM-CCK without codeword optimization. Although this
and the conventional QPSK modulation. We also apply the letter only focuses on optimized design under AWGN channel,
same FEC code in both systems. the proposed methodology can be similarly applied to fading
We use the extrinsic information transfer (EXIT) chart [13] channels using different PEP upper bound.
as shown in Fig. 3a to demonstrate the convergence of
proposed iterative receiver. The arrows in the figure indi- R EFERENCES
cate the direction of the exchange of extrinsic information [1] C. Andren, “CCK modulation delivers 11Mbps for high rate IEEE
between the demodulator and decoder. Although different val- 802.11 extention,” in Proc. Wireless Symp. Portable Design Conf. Spring,
ues of Es/N0 may require different numbers of iterations 1999.
[2] M. V. Clark, K. K. Leung, B. McNair, and Z. Kostic, “Outdoor IEEE
to converge, it can be seen that the detector typically con- 802.11 cellular networks: Radio link performance,” in Proc. IEEE Int.
verges within 3 iterations with Es/N0 = 7dB. Thus, we set Conf. Commun., vol. 1, 2002, pp. 512–516.
the iteration number to 3 in the following BER and FER [3] V. Arneson, K. Ovsthus, O. I. Bentstuen, and J. Sander, “Field trials
with IEEE 802.11b-based UHF tactical wideband radio,” in Proc. IEEE
comparison. Mil. Commun. Conf., vol. 1, 2005, pp. 493–498.
Fig. 3b and Fig. 3c demonstrate the receiver output BER [4] L. Jing, H. Wang, C. He, and Z. Ding, “Spatial CCK modulation
and FER performance after 3 iterations, respectively. In order and iterative detection over frequency-selective fading channels,” IEEE
Wireless Commun. Lett., vol. 6, no. 4, pp. 506–509, Aug. 2017.
to demonstrate the effectiveness of the proposed QAM-CCK [5] C. He, J. Huang, and Z. Ding, “A variable-rate spread-spectrum system
design criterion in Eq. (12), we provided the performance of for underwater acoustic communications,” IEEE J. Ocean. Eng., vol. 34,
the best selected QAM-CCK along with a random mapping no. 4, pp. 624–633, Oct. 2009.
QAM-CCK as a benchmark. It is clear that the optimized [6] J. G. Proakis, Digital Communications. New York, NY, USA:
McGraw-Hill, 1995.
QAM-CCK outperforms the random codeword design both in [7] T. K. Moon, Error Correction Coding: Mathematical Methods and
terms of BER and FER. Furthermore, the optimized QAM- Algorithms. New York, NY, USA: Wiley, 2005
CCK achieves approximately 4 dB gain at 10−5 level after [8] C. Berrou, A. Glavieux, and P. Thitimajshima, “Near Shannon limit
error-correcting coding and decoding: Turbo-codes,” in Proc. IEEE Int.
the third iteration compared with plain QPSK modulation. Conf. Commun., vol. 2, May 1993, pp. 1064–1070.
[9] S. ten Brink, “Convergence behavior of iteratively decoded parallel con-
catenated codes,” IEEE Trans. Commun., vol. 49, no. 10, pp. 1727–1737,
V. C ONCLUSION Oct. 2001.
[10] C. Douillard, M. Jezequel, C. Berrou, A. Picart, and P. Didier, “Iterative
This letter proposes a novel QAM-CCK modulation scheme correction of intersymbol interference: Turbo-equalization,” in Proc. Eur.
to achieve high bandwidth-efficiency by optimizing the code- Trans. Telecommun., vol. 6, no. 5, pp. 507–511, 1995.
[11] IEEE Standard for Information Technology—Telecommunications and
word selection based on maximized weighted minimum Information Exchange Between Systems—Local and Metropolitan
inter-codeword Euclidean distance. The proposed QAM-CCK Networks—Specific Requirements—Part 11: Wireless LAN Medium
scheme inherits the spread spectrum coding gain of CCK Access Control (MAC) and Physical Layer (PHY) Specifications: Higher
Speed Physical Layer (PHY) Extension in the 2.4 GHz Band, IEEE
while providing a higher bandwidth-efficiency. We also intro- Standard 802.11b-1999, pp. 11–30, Jan. 2000.
duce a weighted minimum Euclidean distance based on [12] D. Divsalar and M. K. Simon, “The design of trellis coded MPSK for
both the Hamming distance of the data sequence and the fading channels: Performance criteria,” IEEE Trans. Commun., vol. 36,
Euclidean distances between codewords to optimize the bit no. 9, pp. 1004–1012, Sep. 1988.
[13] M. Tüchler, R. Koetter, and A. C. Singer, “Turbo equalization: Principles
error performance. We further present an iterative receiver and new results,” IEEE Trans. Commun., vol. 50, no. 5, pp. 754–767,
based on soft-information exchange between QAM-CCK May 2002.
500 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
user (PU), has been considered as a solution to improve spec- II. S YSTEM M ODEL
trum utilization. Cooperative spectrum sensing (CSS) is a key A. Spectrum Sensing
technology of cognitive radio (CR) system to enhance the PU Consider a network of N SUs detecting cooperatively the PU
detection performance by exploiting SUs’ spatial diversity. signal by the local spectrum sensing, the attack ratio is ρ in N
However, the process of CSS requires the large commu- collaborative SUs. In general, the local spectrum sensing can
nication resource for reporting sensing results, particularly, be formulated as a binary hypothesis test. Let H0 and H1 be
in a large cognitive radio network (CRN), which limits or the two hypotheses with P (H0 ) and P (H1 ) denoting the asso-
even compromises the available cooperative gain [1], [2]. ciated a priori probabilities. The local sensing performance, the
On this account, the authors studied the performance of false alarm and miss detection probabilities are assumed to be
hard-decision CSS with sequential reporting based on K-out- the same for every SU irrespective of whether they are reliable
of-N in [2]–[4], resulting in less samples for decision-making or malicious, denoted by Pf and Pm , respectively. Otherwise,
to avoid consuming too much communication resource, but the detection probability is derived from Pd = 1 − Pm .
none of them consider Byzantine attack, which significantly After each SU independently performs local spectrum sens-
degrades CSS performance. To protect CSS from Byzantine ing, the fusion center (FC) is responsible for the global
attack, Chen et al. [5], [6] combined the nodes reputation and decision by collecting sensing results from SUs to cooper-
used sequential probability ratio test (SPRT) and depended on ate sensing and aggregate sensing results via a fusion rule.
careful threshold to identify attackers. Furthermore, [7]–[10] Multiple SUs usually access the control channel in time divi-
moved forward with weighted sequential probability ratio test sion multiple access mode, and their sensing results need to be
(WSPRT) against Byzantine attack. All of them reduce the collected one by one not in parallel [4]. The communication
sample size to some extent while extremely increase the channels between SUs and the FC are error-free in this letter.
computational complexity. Unfortunately, they have strong
assumptions (such as there exist only a few attackers which
act in a simplified strategy) and require a priori information
B. Byzantine Sophistication
Manuscript received September 3, 2018; accepted October 15, 2018. Date of In the process of CSS, attackers intentionally falsify own
publication October 23, 2018; date of current version April 9, 2019. This work sensing results before submitting them to the FC in an attempt
was supported in part by the National Natural Science Foundation of China to mislead the global decision regarding the phenomenon of
under Grant 61771126, and in part by the Key Research and Development Plan
of Jiangsu Province under Grant BE2018108. The associate editor coordinat-
presence, thereby undermining the premise of CR technology.
ing the review of this paper and approving it for publication was W. Hamouda. Previous malicious detection and suppression algorithms focus
(Corresponding author: Jun Wu.) on simple “always attack” only (i.e., attackers do not always
The authors are with the National Mobile Commutation Research report the false information to the FC), since “always attack”
Laboratory, Southeast University, Nanjing 211189, China (e-mail:
wojames2011@163.com; yuyue@seu.edu.cn; songtc@seu.edu.cn;
strategy can be easily identified. In fact, attackers may act in a
louy@seu.edu.cn). strategic manner, such as, misleading the network occasionally
Digital Object Identifier 10.1109/LWC.2018.2877665 but behaving correctly during the rest of the time, as depicted
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
WU et al.: SEQUENTIAL 0/1 FOR CSS IN PRESENCE OF STRATEGIC BYZANTINE ATTACK 501
⎛ ⎛ ⎛ ⎞⎞⎞
N
ρN N
−ρN
N ρN km N −ρN k
Qf ,cvr = ⎝
i
⎝
km P (1 − Pfa )ρN −km ⎝
fa P n (1 − Pf )N −ρN −kn ⎠⎠⎠
kn f (4)
i=K km =0 kn =K −kn
⎛ ⎛ ⎛ ⎞⎞⎞
N
ρN N
−ρN
N ρN km N −ρN
Qm,cvr = 1 − ⎝
i
⎝ P (1 − Pda )ρN −km ⎝
km da P kn (1 − Pd )N −ρN −kn ⎠⎠⎠
kn d (5)
i=K km =0 kn =K −kn
in Fig. 1, a strategic Byzantine attack model is described as In CVR, N report results are required to submit to the FC.
follows Such a fixed sample size is inefficient to proceed with fus-
ing sensing information. For this reason, we first consider a
P (r = 1|s = 0) = α sequential approach using CVR to reduce the sample size in
(1)
P (r = 0|s = 1) = β the following section.
where s is the sensing result and r is the report result, α is the
III. S EQUENTIAL 0/1 FOR C OOPERATIVE
false alarm attack probability, the false alarm attack implies
S PECTRUM S ENSING
that the attacker has detected the absence of PU but submits
the presence decision 1 to the FC, its aim is to prevent reliable A. Sequential Voting Rule
SUs from using the idle channel. β is the miss detection attack Encouraged by sequential decision rule, we start with
probability. The miss detection attack implies that the attacker an objective to develop SVR. In SVR, SUs’ report results
has detected the PU’s presence but submits the absence deci- sequentially arrive at the FC until the decision condition is
sion 0, thereby alluring them to access the channels in use satisfied. That is to say, the FC receives K report result 1s or
and causing excessive interference to the PU. In the proposed N−K+1 report result 0s, the rest of report results do not need
attack model, the attacker has a certain probability, varying to be submitted. Obviously, the decision condition of SVR
from 0 to 1, to conduct various attack strategies. Thus, the is consistent with CVR, therefore, the cooperative sensing
false alarm and miss detection probabilities of the attacker performance of SVR is also identical to CVR.
can be represented as The next thing comes into consideration is the sample size.
Assume that P1 and P̂1 denote the probability of an individual
Pfa = P (s = 0|H0 )P (r = 1|s = 0) report result 1 for the reliable SU and the attacker, respectively,
+ P (s = 1|H0 )P (r = 1|s = 1) thus the average number of samples to satisfy the voting rule
= (1 − Pf )α + Pf (1 − β) (2) can be obtained by
Pma = P (s = 0|H1 )P (r = 0|s = 0) N
+ P (s = 1|H1 )P (r = 0|s = 1) ψ(N , K , P1 , P̂1 ) = ϕ(i , K , P1 , P̂1 )
i=1
= Pm (1 − α) + (1 − Pm )β (3)
+ ϕ(i , N − K + 1, 1 − P1 , 1 − P̂1 )
Besides, the detection probability of the attacker can be (6)
computed by Pda = 1 − Pma . Through above attack model, n−1
ρN ρN km N −ρN
if possible, attackers would want to make the FC completely where ϕ(n, t, p1 , p̂1 ) = t−1 km =0 km P̂1 ) t−km
t−km
ρN −km ρN −km km )(
n−t−(ρN −km )
unable to decide on a particular decision and the performance ·P1 (
km
(1 − P̂ 1 ) kn =0
km =0
at the FC can be no better than just a random guess of the state n−t−(ρN −k m)
k
of channel. In other words, the report results received by the kn (1 − P1 ) n denotes the probability of
FC are completely independent of the hypothesis test. That is, Negative Binomial (NB) distribution.1 When a global deci-
if the attack ratio and a pairwise of attack probabilities satisfy sion is made in SVR, there must be either K report result 1s
ρ ≥ 1/(α + β), which can make the FC blind. More details or N−K+1 report result 0s received at the FC, thus
of the blind scenario are given in [11]. N
ϕ(i , K , P1 , P̂1 )
C. Conventional Voting Rule i=1
Considering that voting rule (a.k.a. K-out-of-N) can be real- + ψ(i , N − K + 1, 1 − P1 , 1 − P̂1 ) = 1 (7)
ized in the low complexity without any prior knowledge on the LHS in (7) represents the sum of Qm,cvr and Qd,cvr , and
PU signal, it is utilized as a fusion rule in this letter. In con-
ventional voting rule (CVR), more than K in N report results certainly equals to one as RHS. Therefore, ψ(N , K , P1 , P̂1 ) ≤
are declared as the PU’s presence, then the FC broadcasts the N , which implies a smaller average number of samples
channel is busy, otherwise the channel is idle. Through sum- required for CSS with SVR.
ming the possibility of report results of satisfying K-out-of-N According to the Bayes theorem, the average number of
rule, when there exist ρN attackers in the network and ρN ≤ samples required at the FC regarding the primary signal can
K, the false alarm and miss detection probabilities of CVR can 1 ϕ(n, t, p , p̂ ) represents the probability of n trials that occur for a given
1 1
be obtained as (4) and (5) shown at the top of this page. As number of t successes, p1 and p̂1 refers to the “success” probability of two
for the case of ρN > K, similar results are also obtained. participants, respectively
502 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
be given by
N̄svr = ψ(N , K , Pd , Pda |H1 )P (H1 )
+ ψ(N , K , Pf , Pfa |H0 )P (H0 ) (8)
where ψ(N , K , Pd , Pda |H1 ) and ψ(N , K , Pf , Pfa |H0 ) can be
calculated by (6).
Overall, the main advantage of SVR is that it requires, on
an average, fewer samples to achieve the same performance
as a fixed sample size test.
B. Sequential 0/1
Though SVR reduces the sample size without performance
loss, it fails in considering the malicious presence. Under the Fig. 2. Correct sensing ratio for various voting rules.
assumption of a simple attack strategy, [5]–[10] have come up
with the sequential hypothesis test to defend against Byzantine
attack, but its advantage is attained at the expense of additional Byzantine attacks, but without any alteration of the decision
computation and prior knowledge of primary signal. computa- condition. Therefore, it is not surprising that S0/1 provides the
tion and prior knowledge of primary signal. Such limitations performance as well as CVR in the absence of attackers.
fuel the motivation in considering a sequential approach that As for the sample size, the FC only requires N−K+1 report
does not require any prior information to reduce the sample result 0s in support of the global decision 0 in S1 or requires
size and computational complexity. Therefore, we introduce K report result 1s in support of the global decision 1 in S0,
S0/1 using SVR to counteract strategic Byzantine attacks. hence, according to (6), the average number of samples of S0
In S0, only the report result 0 is allowed to submit to the FC
can be computed by
whereas the report result 1 is not. To be specific, the reports
result 0s from N report results are collected sequentially, if N̄s0 = ψ(N − K + 1, N − K + 1, 1 − Pd , 1 − Pda |H1 )P (H1 )
the cumulative number of 0s satisfies the decision condition,
+ ψ(N − K + 1, N − K + 1, 1 − Pf , 1 − Pfa |H0 )P (H0 )
i.e., the number of 0s is N−K+1, the remaining report result
0s are not required, the global decision is 0; Otherwise, if the (9)
FC has received N report results that still cannot satisfy the
For S1, the average number of samples are easily obtained
decision condition, then the global decision is automatically
as follows
declared as 1. In contrast with S0, S1 only allows the SU to
submit the report result 1. If the cumulative number of the N̄s1 = ψ(K , K , Pd , Pda |H1 )P (H1 )
report result 1s exceeds K, the remaining 1s are not required, + ψ(K , K , Pf , Pfa |H0 )P (H0 ). (10)
the global decision is 1; Otherwise, the cumulative number of
1s still cannot exceed K after N report results arrive at the FC,
then the global decision is announced as 0. IV. S IMULATION R ESULTS
From implementation point of view, although the PU has In this section, we simulate 1000 sensing intervals to cor-
no obligation to provide any channel information for SUs, the roborate the proposed S0/1. The related parameter settings are
network administrator basically knows the business circum- set as follows: N = 100, P (H0 ) = 0.8 and P (H1 ) = 0.2,
stances of the network coverage area, for example, the primary Pf = Pm = 0.1. In SPRT-based approaches, the toler-
network is busy during a certain period of time and idle dur- ated false alarm and miss detection probabilities are 10−3
ing another period. Before proceeding with S0/1, the network and 10−4 [8], [9], respectively. Considering the blind and
administrator can observe each SU’s performance over a sens- non-blind scenario, we simulate the correct sensing ratio (rep-
ing observation, then in the real operation of S0/1, the network resents the ratio of correct detecting the PU’s presence or
administrator will not tell SUs with the poor performance absence during a certain sensing interval) and sample size in
to choose S0/1 unless they return the normal performance the context of strategic attack (ρ, α, β) such as (0.5,0.5,0.5)
in the sequential spectrum sensing. Additionally, cooperative and (0.8,0.8,0.8).
spectrum sensing can be verified by an additional monitoring In Fig. 2, as anticipated, SVR and CVR has the same cor-
process. For a centralized CRN, the monitoring process is easy rect sensing ratio, despite which starts to deteriorate owing to
to perform by the FC itself. If the global decision is reached the malicious presence, particularly in the blind scenario. The
after X reports (X<N), the global decision will be sent along difference, however, is that SVR requires less samples than
with the stopp reporting message to all SUs. CVR. To be specific, the sample size of SVR increases at first
Our proposed S0/1 not only has a natural advantage in and decreases at last, but the high correct sensing ratio cannot
defending Byzantine attack, but reduces the sample size. In be guaranteed. Whereas the correct sensing ratio of S0/1 is
detail, if the report result 1 occurs during S0, this implies always consistent with that of CVR in the absence of attack-
that there must exist the attacker which has falsified the orig- ers, which is not affected by the blind problem. In addition,
inal report result 0. In S1, if the report result 0 appears, it along with the increasing K, the sample size of S0 reduces
represents the presence of the attacker falsifying the original while S1 increases, as depicted in Fig. 3.
report result 1. Subsequently, the attacker identification policy One interesting effect is that the sample size of S0 is greater
is easily implemented and does not depend on careful thresh- than that of S1 for K<75. This is because the low spectrum
old selection. This feature enables S0/1 to cope with various underutilization makes S1 more advantageous when P (H0 ) =
WU et al.: SEQUENTIAL 0/1 FOR CSS IN PRESENCE OF STRATEGIC BYZANTINE ATTACK 503
Fig. 3. Sample size for various voting rules. Fig. 5. Sample size versus attack ratio.
V. C ONCLUSION
In this letter, we analyze Byzantine attack behavior in the
process of CSS and develop a sophisticated attack model.
Subsequently, a simple yet effective S0/1 based SVR is pre-
Fig. 4. Correct sensing ratio versus attack ratio. sented to counteract strategic Byzantine attack. Simulation
results show that in contrast to existing approaches, the pro-
posed S0/1 not only strongly reduces the sample size also
0.8. K is usually used as an optimization variable to obtain the provides the higher correct sensing ratio, even in the blind
optimal performance, it can be observed that within K>75, K scenario.
report result 1s are required for S1 to make the global decision
while N−K+1 report result 0s are required for S0. Such as, in R EFERENCES
S0/1, a small or large value of K(such as, K<20 and K>80) [1] M. S. Khan, M. Usman, V.-V. Hiep, and I. Koo, “Efficient selec-
results in the low correct sensing ratio, the optimal K should tion of users’ pair in cognitive radio network to maximize throughput
using simultaneous transmit-sense approach,” IEICE Trans. Commun.,
be between 20 and 80. Simulation results provide a criterion vol. 100, no. 2, pp. 380–389, 2017.
for the optimal K to counteract Byzantine attack and reduce [2] H. Luan, O. Li, and X. Zhang, “Cooperative spectrum sensing with
the sample size. energy-efficient sequential decision fusion rule,” in Proc. IEEE 23rd
Wireless Opt. Commun. Conf. (WOCC), Newark, NJ, USA, 2014,
Through above analyses, it is known that the choice of S0 or pp. 1–4.
S1 depends on whether the PU is active or inactive for a certain [3] A. A. Alkheir and H. T. Mouftah, “Sequential hard-decision fusion for
agile cooperative spectrum sensing,” in Proc. IEEE Int. Conf. Commun.
period of time. Given the reliable performance, the smaller Workshop (ICCW), London, U.K., 2015, pp. 1014–1019.
K, the lower sample size in S1 while a larger K leads to a [4] S. Peng, W. Zheng, R. Gao, and K. Lei, “Fast cooperative energy detec-
smaller sample size in S0. Based on this criterion, Fig. 4 and 5 tion under accuracy constraints in cognitive radio networks,” Wireless
Commun. Mobile Comput., vol. 2017, pp. 1–8, Sep. 2017.
respectively compare the correct sensing ratio and sample size [5] R. Chen, J.-M. J. Park, and K. Bian, “Robustness against Byzantine
of S0/1 against SPRT, WSPRT, RWSPRT across a wide range failures in distributed spectrum sensing,” Comput. Commun., vol. 35,
of attack ratios when α = β = 1. Further detail on the SPRT, no. 17, pp. 2115–2124, 2012.
[6] C.-Y. Chen, Y.-H. Chou, H.-C. Chao, and C.-H. Lo, “Secure centralized
WSPRT, RWSPRT can be found in [8] and [9]. spectrum sensing for cognitive radio networks,” Wireless Netw., vol. 18,
In Fig. 4, the correct sensing ratios of other three schemes no. 6, pp. 667–677, 2012.
have declined to varying degrees as the attack ratio increases. [7] A. A. Sharifi and J. M. Niya, “Securing collaborative spectrum sensing
against malicious attackers in cognitive radio networks,” Wireless Pers.
Specifically, the correct sensing ratios of SPRT and WSPRT Commun., vol. 90, no. 1, pp. 75–91, 2016.
are approximately 0.5 after the attack ratio exceeds 50%, this [8] J. Wu et al., “Robust cooperative spectrum sensing against probabilis-
tic SSDF attack in cognitive radio networks,” in Proc. IEEE 86th Veh.
confirms that Byzantine attack makes FC once ρ ≥ 1/(α +β). Technol. Conf. (VTC-Fall), Toronto, ON, Canada, 2017, pp. 1–6.
As the attack ratio further increases, the correct sensing ratios [9] J. Wu, T. Song, Y. Yu, C. Wang, and J. Hu, “Sequential cooperative
of SPRT is gradually close to 0 and WSPRT coverages to spectrum sensing in the presence of dynamic Byzantine attack for mobile
networks,” PLoS ONE, vol. 13, no. 7, 2018, Art. no. e0199546.
αP (H0 ) + βP (H1 ) = 0.2. Although RWSPRT is better than [10] Z. Hu, Y. Bai, L. Cao, M. Huang, and M. Xie, “A sequential com-
SPRT and WSPRT, it also begins to decline when ρ = 70%. pressed spectrum sensing algorithm against SSDH attack in cognitive
While the attack ratio has no effect on the correct sensing ratio radio networks,” J. Elect. Comput. Eng., vol. 2018, no. 8, pp. 1–9, 2018.
[11] B. Kailkhura, Y. S. Han, S. Brahma, and P. K. Varshney, “Distributed
of S0/1, absolutely, this benefits from anti-attack property of Bayesian detection in the presence of Byzantine data,” IEEE Trans.
S0/1. Signal Process., vol. 63, no. 19, pp. 5250–5263, Oct. 2015.
504 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Array processing is a key technology for emerging must be known to the receiver. Additionally, synchronization
mobile networks, especially in short to moderate range and line- between the transmitter and the receiver must be ensured.
of-sight scenarios. In these scenarios, the incoming wavefront Large antenna arrays are widely foreseen for 5G applica-
can be modeled by a spherical wave. The wavefront curvature,
i.e., curvature of arrival (CoA), contains position information of tions, varying from dozens of antennas at mobile stations [5]
the transmitter and is observable by an antenna array poten- to over one hundred at local access points [7] and up to a few
tially asynchronous and non-coherent to the transmitter. We thousands at base stations [1]. These arrays provides communi-
derive a simplified expression of the spherical wave position- cation coverage, from their radiating near field (a few meters)
ing (SWP) Cramér–Rao bound for arbitrary centro-symmetric to the beginning of the Fraunhofer region (up to a hundred
arrays, which provides a geometrical inference about the achiev-
able performance. Additionally, a low complexity CoA positioning meters), to mobile devices under line-of-sight (LoS) condi-
algorithm is proposed. In contrast to conventional methods, tion. The signal wavefront received by the array is modeled
the proposed algorithm requires neither multiple anchors nor by a spherical wave. Under this model, not only DoA but also
coordination between devices. It also outperforms the Fresnel distance information of the transmitter is contained in the car-
approximation based SWP algorithms by overcoming the model rier phase, which enables SWP of the transmitter [7]–[10].
mismatch. Therefore, the proposed CoA positioning algorithm is
promising for precise positioning in future mobile networks. Most previous works apply the Fresnel approximation to
arrays with special geometries, e.g., uniform linear arrays
Index Terms—Joint DoA/distance estimation, 5G position- (ULAs) [8]–[10], and introduce a model mismatch. This mis-
ing, near field, spherical wavefront, antenna array, Cramér-Rao
bound. match has recently been noticed to jeopardize the achievable
positioning precision [11]. In [12] a lookup table is used
for ULA model correction. The maximum likelihood (ML)
I. I NTRODUCTION algorithm in [7] exploits the exact model, but includes a
BIQUITOUS realtime position information is envisaged computationally expansive recursion. In [13] the Cramér-Rao
U as a key feature of future mobile networks, for example
the 5th generation (5G) networks, due to emerging device-
bound (CRB) of SWP is analyzed for special array geometries.
In this letter, a simplified expression of the CRB of SWP is
centric applications [1]. Studies have been conducted on derived for arbitrary centro-symmetric arrays (CSAs), which
5G positioning, where 5G networks provide opportunities only depends on the relative geometry and characteristics of
for precise positioning in global navigation satellite system the antennas’ spatial distribution. This expression provides
(GNSS)-impaired environments [2]. Most traditional position- a geometric inference about the achievable performance and
ing techniques exploit either time of arrival (ToA) or direction brings further insights into array design. Additionally, we pro-
of arrival (DoA) estimates and demand multiple anchors [3]. pose a method dubbed CoA positioning, which extracts the
Recent research has focused on positioning with a single transmitter position information directly from the wavefront
anchor. The simultaneous localization and mapping (SLAM) curvature. CoA positioning overcomes the model mismatch
algorithm in [4] utilizes multipath components for position- introduced by Fresnel approximation, while maintaining low
ing. However, it requires memory-intensive storage and a complexity for realtime.
static environment. In combined DoA and ToA estimation, an
antenna array is used to position a transmitter. With the far- II. P OSITION I NFORMATION IN S PHERICAL WAVE
field assumption, distance information is obtained solely from
A single transmitter antenna is placed at point Ps , which
the propagation delay of the baseband pilot signal, whereas the
radiates a single-carrier signal at carrier frequency fc with
DoA is estimated from the carrier phase differences between
an unknown real-valued amplitude S and an unknown phase
antennas [5], [6]. For ToA estimation, the pilot signal structure
φδ . The signal propagates under LoS condition to a generic
Manuscript received August 31, 2018; revised October 11, 2018; accepted point P at distance d with speed of light c0 . The signal
October 20, 2018. Date of publication October 25, 2018; date of current phase at P, φ = φδ − ωc d /c0 , is a continuous function in
version April 9, 2019. This work was supported by the German Research space, where ωc = 2πfc . The transmitter’s position infor-
Foundation (DFG) under Contract FI 2176/1-1 and Contract HO 2226/17-1.
The associate editor coordinating the review of this paper and approving it mation w.r.t. an observation point Po can be extracted from
for publication was Y. Shen. (Corresponding author: Siwei Zhang.) the continuous wavefield. We define a two-dimensional (2D)
S. Zhang, T. Jost, R. Pöhlmann A. Dammann, and D. Shutin are with Cartesian coordinate system1 C (xy) that originates at point Po
the German Aerospace Center, Institute of Communications and Navigation,
82234 Wessling, Germany (e-mail: siwei.zhang@dlr.de). 1 The coordinates of a specific point P in a coordinate system C (ξψ) are
P. A. Hoeher is with the Chair of Information and Coding Theory,
(ξψ)
University of Kiel, 24143 Kiel, Germany (e-mail: ph@tf.uni-kiel.de). defined as p = [ξ , ψ ]T , where ξ and ψ are the two dimensions of
Digital Object Identifier 10.1109/LWC.2018.2877971 that coordinate system. The subscript is omitted for generic points P.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHANG et al.: SWP BASED ON CoA BY ANTENNA ARRAY 505
2σ 2 ds2
COV[p̂s ] CRB[ps ] =
S2
L
∂dsl ∂dsl 1 ∂dsl ∂dsm −1
L L
Fig. 1. Spherical wave positioning based on CoA. × − . (4)
∂ps ∂pT
s L ∂ps ∂pT
s
l=1 l=1 m=1
and includes point Ps . The corresponding polar coordinate
system is defined as C (dθ) . We consider the 2D position- To infer the geometry impacts on SWP, we first investigate
(dθ) the symmetric linear array (SLA) case. An SLA is deployed
ing problem, i.e., estimating ps = [ds , θs ]T , where ds is
along the x-axis, with L elements and an aperture length A.
the distance between points Ps and Po , and θs is the DoA
We define the k th moment of the normalized antennas’ spatial
w.r.t. the positive x-axis of C (xy) . For the rest of this letter,
distribution Mk = L k
l=1 l /A) /L, and the effective aperture
(d
we use C (dθ) as the default coordinate system and omit the
superscript (d θ) for simplicity. The transmission phase φδ and length à = A sin θs , to characterize the array geometry.
signal amplitude αs after propagation are not of interest, but Theorem 1 (CRB of SWP for SLA): For the SLA, assum-
need to be estimated jointly with ps . The total parameter vec- ing L 1 and ds A, the CRB of DoA estimate can be
tor to be estimated is χ = [pT T approximated by
s , φδ , αs ] . The spherical wave
intersects the xy-plane with co-phase circles centered at Ps . 2σ 2 ds2
CRB[θs ] ≈ , (5)
The curvature of these circles contains information about the S 2 LÃ2 M2
distance to the transmitter. We apply a coordinate transforma- whereas the distance CRB is approximated by
Po ,θs
tion C (xy) −−−−→ C (uv ) , where the new Cartesian coordinate 2σ 2 ds2 4ds4
system C (uv ) originates at Po and the u-axis is aligned with θs . CRB[ds ] ≈ . (6)
S 2 LÃ4 (M4 − M22 )
Definition 1 (Signal CoA): The signal CoA κo at point
Po is defined as the extrinsic curvature of −φc0 /ωc along Proof: See the Appendix.
the v-axis of C (uv ) . With the spherical wave model, CoA is The array’s aperture is often physically constrained by appli-
propotional to the absolute value of the phase’s second-order cations. For a fixed A, more elements can be deployed for
derivative and equals to the reciprocal of ds higher carrier frequencies, without resulting in severe antenna
c0 ∂ 2 φ
mutual coupling. Both CRBs in (5) and (6) linearly decrease
1
κo − 2 = . (1) with the number of elements L, which shows a benefit of
ωc ∂v Po ds higher fc , such as foreseen in 5G. The term 2σ 2 ds2 /S 2 shows
The distance information can be extracted from the second- the effect of the signal-to-noise ratio (SNR). The CRB for
order derivative of the signal phase. The DoA needs to be DoA decreases quadratically with à and linearly with M2 ,
estimated prior to coordinate transformation. In practice, a 2D the antennas’ spatial spread. The distance CRB experiences a
antenna array on the xy-plane, potentially asynchronous and quartic growth with the ratio ds /Ã, indicating a strong impact
non-coherent to the transmitter, with L elements centered at Po from the relative geometry. Additionally, it decreases with
is used to sample the continuous wavefield at discrete spatial M4 − M22 , i.e., the shape of the antennas’ spatial distribu-
points. An element l, l = 1, . . . , L, located at point Pl , receives tion. More importantly, when θs = 0◦ , both CRBs approach
the baseband signal2 infinity. Hence the array’s aperture expanded in u direction
contains no information of the transmitter’s position. With the
rl (t) = αs e j φδ e −j ωc dsl /c0 + nl (t), (2)
last observation, we extend Theorem 1 to arbitrary 2D CSAs.
where nl (t) ∼ CN (0, σ 2 )
is an i.i.d. circularly-symmetric Many typical arrays are centro-symmetric, e.g., uniform cir-
complex normally distributed noise process. According to the cular/linear arrays, the ones in [13], as well as the uniform
geometry under investigation, dsl can be expressed as rectangular array (URA) illustrated in Fig. 1.
Corollary 1: A CSA centered at P0 can be projected on
dsl = ds2 + dl2 − 2ds dl cos(θl − θs ). (3) the v-axis, forming a virtual SLA. The positioning CRB can
The received samples rl are acquired at an arbitrary time point, be approximated by applying Theorem 1 to the virtual SLA.
coherently at all elements, with a received sample phase φl . Proof: Since the aperture expanded in u direction does not
The concept of the SWP with CoA is illustrated in Fig. 1. contain position information, the projected virtual linear array
along v-axis is equivalent to the original CSA in the sense
III. F UNDAMENTAL L IMITS OF SWP of SWP. By the definition of centro-symmetry, for any non-
centered element l with position pl = [dl , θl ]T , there exists an
The Fisher information matrix (FIM) of χ can be calculated element m with position pm = [dl , θl +π]T . Elements l and m
from the given model (2) and (3), similarly as in [13]. The are projected on the v-axis at ±dl sin(θl −θs ) respectively and
2 We assume the array aperture to be much smaller than d . Therefore, the
s
are symmetric w.r.t. P0 . Hence the projected array is an SLA,
distance-related attenuation differences among elements are negligible. which meets the condition of Theorem 1.
506 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
L L 2
∂dsl ∂dsl al al bl
and ≈ .
∂ps ∂pTs al bl dl2 sin2 θs + bl2
l=1 l=1
(a) DoA estimation error The CRB of DoA in (5) can be directly obtained by taking the
second diagonal entity of (11). The distance CRB is derived
by taking the first diagonal entity of (11)
The value of D is determined by the temperature, the fluid Using (3), (8) and (9), the Green’s function can be written as
viscosity, and the molecule’s Stokes radius [3].
The solution of the convection-diffusion equation in (1) is 1 −(x − x0 − vt)2 − (y − y0 )2
G(x , y, x0 , y0 , t) = exp .
given in terms of the diffusion propagator G(r, r0 , t) (also 4πDt 4Dt
known as heat kernel or Green’s function). For an absorbing (10)
boundary condition, G(r, r0 , t) satisfies the initial condition
G(r, r0 , t)|t=0 = δ(r − r0 ), where δ(·) is the Kronecker To implement the condition of an absorbing receiver at
delta function, stating that r0 is the starting point of the x = 0, we use the method of images to construct the Green’s
Brownian motion; it also satisfies the Dirichlet boundary con- function for an absorbing wall. Using this method, many
dition, G(r, r0 , t)|t=0 = 0 at r ∈ S, where, S represents the boundary conditions can be automatically satisfied, and it
set of points forming the receiver surface. works particularly well for boundaries with flat surfaces [11].
The velocity field is assumed to remain constant near the In this method, it is assumed that for each particle released
boundary, such that the fluid velocity does not vanish near at (x0 , y0 ), an image source of negative mass is released at
the receiver. This leads to an outflow boundary condition. (−x0 , y0 ). These two particles diffuse freely, giving a zero
Here, the receiver acts like an imaginary surface where the resultant concentration at x = 0. The Green’s function in the
concentration becomes zero, without affecting the flow. case of an absorbing wall is then
Gab (x , y, x0 , y0 , t) = G(x , y, x0 , y0 , t)
III. C HANNEL C HARACTERIZATION − a(x0 )G(x , y, −x0 , y0 , t). (11)
For a 2-dimensional system, having a constant drift velocity
v from the transmitter to the receiver along the x-axis, (1) can The absorption condition at the receiver requires that
Gab (0, y, x0 , y0 , t) = 0, giving
be rewritten as
G(0, y, x0 , y0 , t) x v
∂c(x , y, t) ∂c(x , y, t) ∂ 2 c(x , y, t) ∂ 2 c(x , y, t) 0
+v =D ,
a(x0 ) = = exp − , (12)
∂t ∂x ∂x 2
+
∂y 2 G(0, y, −x0 , y0 , t) D
(2) which is independent of t and can be interpreted as the mass
of the negative particle. Thus, the Green’s function for the
where, c(x , y, 0) = δ(x − x0 )δ(y − y0 ) and c (x, y, t) = 0, convection-diffusion equation in (2) for an absorbing receiver
for x , y ∈ S.
Using the separation of variables, the concentration field at x = 0 can be written as
c (x, y, t) can be written as 1 (x − x0 − vt)2 + (y − y0 )2
Gab (x , y, x0 , y0 , t) = exp −
4πDt 4Dt
c(x , y, t) = c1 (x , t) · c2 (y, t). (3)
x0 v (x + x0 − vt)2 + (y − y0 )2
− exp − − .
Using (3), we get D 4Dt
∂2c ∂ 2 c1 ∂ 2 c ∂ 2 c2 (13)
= c2 , = c1 , (4)
∂x 2 ∂x 2 ∂y 2 ∂y 2 Using (13), the diffusive flux at (0, y) is then
and ∂Gab (x , y, x0 , y0 , t)
J (0, y, t) = −D
∂c ∂c2 ∂c1 ∂x x =0
= c1 + c2 . (5)
∂t ∂t ∂t 1 (a + vt) (x0 + vt)2 + (y − y0 )2
=− exp −
Using (4) and (5) in (2) and rearranging the terms gives 8πD t2 4Dt
∂c2 ∂ 2 c2 ∂c1 ∂c ∂ 2 c1 (x0 − vt) x0 v (x0 − vt) + (y − y0 )2
2
c1 −D + c2 +v 1 −D = 0. (6)
∂t ∂y 2 ∂t ∂x ∂x 2 − exp − −
t2 D 4Dt
Since c1 , c2 = 0, we have the relations
v (x0 + vt)2 + (y − y0 )2
∂c1 ∂c1 ∂ 2 c1 ∂c2 ∂ 2 c2 =− exp − . (14)
+v −D = 0, −D = 0. (7) 4πDt 4Dt
∂t ∂x ∂x 2 ∂t ∂y 2
The diffusive flux J acts as the joint probability density func-
Thus, c1 (x , t) is the solution of the 1-dimensional convection- tion (PDF) for the exit time and the exit position from the
diffusion equation, and c2 (y, t) is the solution of the half plane. Thus, the arrival position density of the Brownian
1-dimensional diffusion equation in (7). Using the Laplace particle irrespective of the time (also known as the harmonic
transform method as in [11] for the initial condition measure) can be calculated as the marginal density for position
c(x , y, 0) = δ(x − x0 )δ(y − y0 ), c1 (x , t) and c2 (y, t) are by integrating over time. Since the receiver surface is perfectly
obtained as absorbing, this arrival position will always be the first arrival
position. Let X = (x0 , y0 ) denote the point of release of the
1 (x − x0 − vt)2
c1 (x , t) = √ exp − , (8) information molecule and Y = (x, y) its arrival position at
4πDt 4Dt the receiver boundary. The conditional PDF of Y at a receiver
located at the origin (x = 0) is given as
and
∞
1 (y − y0 )2 fY |X (y|x ) = J (x , y, t) dt |x =0 . (15)
c2 (y, t) = √ exp − . (9)
4πDt 4Dt 0
510 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Fig. 3. SEP vs diffusion coefficient D for v = 5μm/s, 10μm/s. Fig. 4. Capacity vs diffusion coefficient D for v = 5μm/s, 10μm/s.
where λ is the Lagrange multiplier. Taking the partial deriva- figures, the change in SEP as well as the change in capacity
tives of the Lagrange function with respect to pi , 1 ≤ i ≤ M, is more pronounced for smaller values of D, and then the
and equating them to zero, we get the set of equations increase flattens out. Also, the SEP and the channel capac-
⎡ ity have opposite behavior with respect to M in the case of
M ⎢M an M-ary scheme, necessitating a proper trade-off depending
⎢ p(y|x )
⎢ p(y|i )log2 upon the application.
⎣ M
y=1 x =1 px p(y|x ) VI. C ONCLUSION
⎛x =1 ⎞⎤ In this letter, we investigate the FAP of the information
px ⎜ p(y|i ) ⎟⎥ molecule for an MC channel. We derive a closed form expres-
⎜ ⎟⎥ sion for the arrival position density for the case where the fluid
+ p(y|x )⎜ ⎟⎥ = −λ, (27)
ln 2 ⎝
M ⎠⎦ medium has a positive drift. We then consider an MC chan-
px p(y|x ) nel where the release position of the information molecules is
x =1
modulated to convey information. We call this the FAP chan-
where i = 1, . . . , M . nel and analyze this MC channel in terms of SEP and capacity
We solve the set of equations in (27) numerically to in the information theoretic sense.
obtain the optimum (p̄1 , p̄2 , . . . , p̄M , λ̄). At this optimum,
both the Lagrange multiplier as well as the mutual informa- R EFERENCES
tion (23) attain their respective maximum values. Using (23), [1] T. Suda, M. Moore, T. Nakano, R. Egashira, and A. Enomoto,
the capacity per channel use can be obtained as “Exploratory research on molecular communication between nanoma-
⎡ ⎛ ⎞ ⎤ chines,” in Proc. Genet. Evol. Comput. Conf. (GECCO), Washington,
DC, USA, Jun. 2005, pp. 25–29.
M
⎢ M M ⎜ ⎟ ⎥
⎢ p̄x ⎜ p(y|i) ⎟ ⎥ [2] I. F. Akyildiz, F. Brunetti, and C. Blázquez, “Nanonetworks: A new com-
C = p̄i ⎢ − p(y|x )⎜ ⎟ − λ̄⎥. (28) munication paradigm,” Comput. Netw., vol. 52, no. 12, pp. 2260–2279,
⎣ ln 2 ⎝
M ⎠ ⎦ Aug. 2008.
i=1 y=1 x =1 p̄x p(y|x )
x =1 [3] N. Farsad, H. B. Yilmaz, A. Eckford, C.-B. Chae, and W. Guo, “A
comprehensive survey of recent advancements in molecular communi-
cation,” IEEE Commun. Surveys Tuts., vol. 18, no. 3, pp. 1887–1919,
3rd Quart., 2016.
V. N UMERICAL R ESULTS [4] N.-R. Kim and C.-B. Chae, “Novel modulation techniques using iso-
In this section, numerical results are obtained for typ- mers as messenger molecules for nano communication networks via
ical parameter values for short-range molecular com- diffusion,” IEEE J. Sel. Areas Commun., vol. 31, no. 12, pp. 847–856,
munication [1]. We consider the diffusion coefficient Dec. 2013.
[5] M. S. Kuran, H. B. Yilmaz, T. Tugcu, and I. F. Akyildiz, “Modulation
D ∈ [10, 100]μm2 /s, drift velocity v = 5, 10μm/s, and distance techniques for communication via diffusion in nanonetworks,” in Proc.
d = 10μm. The different release points for the information IEEE Int. Conf. Commun., Kyoto, Japan, Jun. 2011, pp. 1–5.
molecules signifying different symbols of the M-ary scheme [6] K. V. Srinivas, A. W. Eckford, and R. S. Adve, “Molecular communica-
are given by xi = (d , i ) for i = 0, 1, . . . , M , (i.e., a = 1), tion in fluid media: The additive inverse Gaussian noise channel,” IEEE
with all the distances being in μ-meters. The variation of SEP Trans. Inf. Theory, vol. 58, no. 7, pp. 4678–4692, Jul. 2012.
[7] N. Farsad, W. Guo, C.-B. Chae, and A. W. Eckford, “Stable distri-
with diffusion coefficient D is shown in Fig. 1, for two values butions as noise models for molecular communication,” in Proc. IEEE
of the drift velocity v. At a constant v, the SEP increases with Glob. Commun. Conf. (GLOBECOM), San Diego, CA, USA, Dec. 2015,
increase in D. This behavior results because the drift and the pp. 1–6.
diffusion components start to interact for larger values of D. [8] N. Pandey, R. K. Mallik, and B. Lall, “Truncated Lévy statistics for dif-
For a given value of D, the SEP decreases with increase of v, fusion based molecular communication,” in Proc. IEEE Glob. Commun.
Conf. (GLOBECOM), Singapore, Dec. 2017, pp. 1–6.
since for higher v, the drift component dominates the diffu- [9] Y. Murin, M. Chowdhury, N. Farsad, and A. Goldsmith, “Diversity
sion component. Furthermore, the SEP increases with increase gain of one-shot communication over molecular timing channels,” in
of M. Proc. IEEE Glob. Commun. Conf. (GLOBECOM), Singapore, Dec. 2017,
The capacity decreases with increase in diffusion coeffi- pp. 1–6.
cient D for as illustrated in Fig. 2. The capacity increases with [10] Y. Huang, M. Wen, L. Yang, C. B. Chae, and F. Ji. (2018).
Spatial Modulation for Molecular Communication. [Online]. Available:
increase of v, as higher drift diminishes the impact of the ran- https://arxiv.org/abs/1807.01468
domness introduced due to D. Furthermore, the capacity for [11] S. Redner, A Guide to First-Passage Processes. Cambridge, U.K.:
an M-ary scheme increase with increase of M. In both the Cambridge Univ. Press, 2001.
512 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—By considering the 3-D antenna arrays and 3-D arbi- Meanwhile, despite difficulty and high price in realization,
trary trajectories of a mobile station, a generic non-stationary the three dimensional (3D) antenna array is a promising solu-
geometry-based stochastic model for multiple-input multiple- tion to improve directional performance of MIMO systems.
output channels is proposed. Under 3-D non-isotropic von
Mises-Fisher scattering scenarios, the theoretical and approx- However, insufficient antenna spacing or lack of scattering
imate expressions of time-variant spatial correlation function would reduce these benefits due to increased spatial correla-
(SCF) are also derived and analyzed. Simulation results show tion (SC). Therefore, the exploitation of SC is vital for design,
that the SCFs of proposed model match well with the ones of optimization, and performance evaluation of wireless MIMO
existing models for the special cases of 1-D linear and 2-D curve communication systems. Especially, the closed-form expres-
trajectories. In addition, the derived theoretical SCFs also have
good agreements with simulated and measured results. sions of spatial correlation functions (SCFs) are essential to
derive the theoretical results of system performance [2], [3],
Index Terms—Non-stationary MIMO channels, geometry- i.e., capacity, energy efficiency, and bit error rate (BER).
based stochastic model (GBSM), spatial correlation (SC), von
Mises-Fisher (VMF), arbitrary trajectories. Most of SCFs in [3] and [4] were investigated for the sta-
tionary MIMO channels with wide-sense stationarity (WSS)
I. I NTRODUCTION assumption. Measurement campaigns have proved that the
WSS assumption is only valid for short time intervals and
ULTIPLE-INPUT multiple-output (MIMO) technolo-
M gies have drawn attention for their ability to increase
spectral efficiency and system capacity significantly [1].
the non-stationarity should be considered [1]. Recently, a few
3D non-stationary MIMO channel models have been presented
in [5]–[10]. However, the models in [5]–[8] only considered
Manuscript received July 27, 2018; revised September 4, 2018 and 3D scattering environments but assumed the mobile station
October 13, 2018; accepted October 17, 2018. Date of publication (MS) moving with a constant velocity (or 1D linear trajec-
October 26, 2018; date of current version April 9, 2019. This work was tory). Bian et al. [9] and Borhani et al. [10] took the velocity
supported by EU H2020 ITN 5G Wireless Project under Grant 641985, in
part by EU H2020 RISE TESTBED Project under Grant 734325, in part variations of MS into account, but they only considered 2D
by EPSRC TOUCAN Project under Grant EP/L020009/1, in part by the curve trajectories on the azimuth plane. For simplicity pur-
National Key Scientific Instrument and Equipment Development Project under pose, the MS in [5]–[10] was configured with a uniform linear
Grant 2013YQ200607, in part by NSF of China under Grant 61631020 and
Grant 61827801, in part by the Open Foundation for Graduate Innovation of array or 2D antenna array. In addition, the corresponding SCFs
NUAA under Grant KFJJ20170405, in part by the Taishan Scholar Program of were usually analyzed numerically due to complex mathemat-
Shandong Province, and in part by Fundamental Research Funds of Shandong ical derivations. This letter aims to fill these research gaps.
University under Grant 2017JC009. The associate editor coordinating the
review of this paper and approving it for publication was V. Raghavan. Overall, the major contributions and novelties of this letter
(Corresponding author: Cheng-Xiang Wang.) are summarized as follows:
Q. Zhu is with the Key Laboratory of Dynamic Cognitive System of 1) This letter develops a new generic 3D non-stationary
Electromagnetic Spectrum Space, College of Electronic and Information
Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing geometry-based stochastic model (GBSM) for MIMO chan-
211106, China, and also with the Institute of Sensors, Signals and Systems, nels. The proposed GBSM allows for 3D scattering environ-
School of Engineering and Physical Sciences, Heriot-Watt University, ments, 3D antenna arrays, and 3D arbitrary trajectories of the
Edinburgh EH14 4AS, U.K. (e-mail: zhuqiuming@nuaa.edu.cn).
Y. Yang, X. Chen, and W. Zhong are with the Key Laboratory of MS, which makes it more realistic and suitable for a wide
Dynamic Cognitive System of Electromagnetic Spectrum Space, College of variety of communication scenarios.
Electronic and Information Engineering, Nanjing University of Aeronautics 2) We derive the theoretical and approximate expressions
and Astronautics, Nanjing 211106, China (e-mail: yingy@nuaa.edu.cn;
chenxm402@nuaa.edu.cn; zhongwz@nuaa.edu.cn). of time-variant SCFs for our proposed GBSM under 3D
C.-X. Wang is with the National Mobile Communications Research von Mises-Fisher (VMF) scattering scenarios which are flex-
Laboratory, School of Information Science and Engineering, Southeast ible and allow the dependence between azimuth angles and
University, Nanjing 210096, China, and also with the Institute of Sensors,
Signals and Systems, School of Engineering and Physical Sciences, Heriot- elevation angles.
Watt University, Edinburgh EH14 4AS, U.K. (e-mail: chxwang@seu.edu.cn). The remainder of this letter is organized as follows.
Y. Tan is with the Institute of Sensors, Signals and Systems, Section II gives a new generic 3D non-stationary GBSM.
School of Engineering and Physical Sciences, Heriot-Watt University,
Edinburgh EH14 4AS, U.K. (e-mail: yi.tan@hw.ac.uk). In Section III, the theoretical and approximate expressions
J. Sun is with the Shandong Provincial Key Laboratory of Wireless of SCF for our proposed GBSM under 3D VMF scattering
Communication Technologies, School of Information Science and scenarios are derived. In Section IV, the SCFs of proposed
Engineering, Shandong University, Qingdao 266237, China (e-mail:
sunjian@sdu.edu.cn). model for three typical trajectories are simulated and validated.
Digital Object Identifier 10.1109/LWC.2018.2878210 Finally, some conclusions are given in Section V.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHU et al.: SCs OF 3-D NON-STATIONARY MIMO CHANNEL MODEL WITH 3-D ANTENNA ARRAYS AND 3D ARBITRARY TRAJECTORIES 513
ρMS
u1 ,u2 ,n (t; Δd
MS
)= e n (t)) ·Rv (t)·Δd
M
1 j(ΦD L I
−π −π
h̃u,s,n (t) = lim e n,m (t)+Φn,m (t)+Φn,m ) (2) × pφMS MS (φMS MS MS MS
n (t), θn (t))dφn dθn
M →∞ M n,m (t),θn,m (t)
m=1
(10)
where M denotes the number of rays, and ΦIn,m is a ran-
dom initial phase uniformly distributed over [0, 2π). In (2), where ΔdBS = dBS s1 − ds2 denotes the space lag at the
BS
ΦD L
n,m (t) and Φn,m (t) denote the time-variant phases caused BS which does not change over time, ΔdMS (t) = RV (t) ·
by Doppler frequency variations and antenna location varia- ΔdMS,t0 denoted the time-variant space lag at the MS, and
MS,t MS,t
tions, respectively, and they can be further modeled as [12] ΔdMS,t0 = Δdu1 0 − du2 0 means the initial space lag.
t In (9) and (10), p BS/MS (φ
BS/MS BS/MS
(t), θn (t))
φn,m (t),θn,m (t) n
BS/MS
D
Φn,m (t) = k vMS (t ) − vZn (t ) s̃MS
n,m (t )dt
(3)
0 represents the time-variant joint probability density function
514 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
(PDF) of random angles, i.e., AAoDs φBS
n (t) and EAoDs PARAMETERS OF T HREE D IFFERENT T RAJECTORIES
θnBS (t), or AAoAs φMS
n (t) and EAoAs θn (t).
MS
R EFERENCES
Fig. 1. Absolute values of the theoretical SCFs of our proposed model [1] C.-X. Wang et al., “Recent advances and future challenges for massive
and other models for three paths at t = 4 s (UMi NLOS scenario, MIMO channel measurements and models,” Sci. China Inf. Sci., vol. 59,
κ = 65, v An and vZn ∼U(0,5) m/s, φAn and φZn ∼U(−π, π), θAn and no. 2, pp. 1–16, Feb. 2016.
θ Zn ∼U(−π/2, π/2)). [2] J.-A. Tsai, R. M. Buehrer, and B. D. Woerner, “BER performance of
a uniform circular array versus a uniform linear array in a mobile
radio environment,” IEEE Trans. Wireless Commun., vol. 3, no. 3,
pp. 695–700, May 2004.
[3] Q.-U.-A. Nadeem, A. Kammoun, M. Debbah, and M.-S. Alouini, “A
generalized spatial correlation model for 3D MIMO channels based
on the Fourier coefficients of power spectrums,” IEEE Trans. Signal
Process., vol. 63, no. 14, pp. 3671–3686, Jul. 2015.
[4] K. Mammasis, R. W. Stewart, and J. S. Thompson, “Spatial fad-
ing correlation model using mixtures of Von Mises Fisher distribu-
tions,” IEEE Trans. Wireless Commun., vol. 8, no. 4, pp. 2046–2055,
Apr. 2009.
[5] J. Bian et al., “A WINNER+ based 3-D non-stationary wideband
MIMO channel model,” IEEE Trans. Wireless Commun., vol. 17, no. 3,
pp. 1755–1767, Mar. 2018.
[6] S. Wu, C.-X. Wang, M. Aggoune, M. M. Alwakeel, and X. You, “A
general 3D non-stationary 5G wireless channel model,” IEEE Trans.
Commun., vol. 66, no. 7, pp. 3065–3078, Jul. 2018.
[7] A. G. Zajić, “Impact of moving scatterers on vehicle-to-vehicle narrow-
band channel characteristics,” IEEE Trans. Veh. Technol., vol. 63, no. 7,
pp. 3094–3106, Sep. 2014.
[8] Q. Zhu et al., “A novel 3D non-stationary wireless MIMO channel sim-
Fig. 2. Absolute values of the theoretical, approximated, simulated SCFs of
ulator and hardware emulator,” IEEE Trans. Commun., vol. 66, no. 9,
the proposed model for Path III at three time instants and the measured SCF
pp. 3865–3878, Sep. 2018.
in [14] (UMi NLOS scenario, κ = 65, v An and vZn ∼U(0,5) m/s, φAn and [9] J. Bian, C.-X. Wang, M. Zhang, X. Ge, and X. Gao, “A 3-D non-
φZn ∼U(−π, π), θAn and θ Zn ∼U(−π/2, π/2)). stationary wideband MIMO channel model allowing for velocity vari-
ations of the mobile station,” in Proc. ICC, Paris, France, Jul. 2017,
pp. 1–6.
The theoretical results of SCFs at t = 0 s, 4 s, and 8 s of [10] A. Borhani, G. L. Stüber, and M. Pätzold, “A random trajectory approach
proposed GBSM for path III are compared with the simulated for the development of nonstationary channel models capturing different
and approximate results in Fig. 2. It clearly shows that SCFs scales of fading,” IEEE Trans. Veh. Technol., vol. 66, no. 1, pp. 2–14,
Jan. 2017.
change over time due to the movement of the MS. Meanwhile, [11] H. Hofstertter, A. F. Molisch, and N. Czink, “A twin-cluster MIMO
the measurement results in [14] are also shown in Fig. 2. channel model,” in Proc. EuCAP, Nice, France, Nov. 2006, pp. 1–8.
The good agreement between measured, theoretical, approxi- [12] B. Boashash. Time-Frequency Signal Analysis and Processing: A
Comprehensive Reference. Amsterdam, The Netherlands: Academic,
mate, and simulated SCFs shows the correctness of both the 2015.
proposed model and theoretical derivations. [13] W. Dahech, M. Pätzold, C. A. Gutiérrez, and N. Youssef, “A non-
stationary mobile-to-mobile channel model allowing for velocity and
trajectory variations of the mobile stations,” IEEE Trans. Wireless
V. C ONCLUSION Commun., vol. 16, no. 3, pp. 1987–2000, Mar. 2017.
[14] S. Payami and F. Tufvesson, “Channel measurements and analysis
We have proposed a generic 3D non-stationary GBSM for very large array systems at 2.6 GHz,” in Proc. Antennas Propag.
incorporating 3D arbitrary trajectories and 3D antenna arrays (EUCAP), Prague, Czech Republic, Mar. 2012, pp. 433–437.
516 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In this letter, protograph-based folded spatially and asymptotic optimality is shown for burst erasure chan-
coupled (FSC) LDPC codes are proposed. The new folded-type nels. The splitting of the convolutional structure into multiple
structure is obtained by folding the spatial coupling chain of a bands at different location makes windowed decoding impos-
conventional spatially coupled (SC) LDPC protograph and inter-
sible; so the low latency continuous decoding property of SC
lacing the nodes at staggered spatial positions. The proposed
codes outperform the SC LDPC codes over single and multiple LDPC codes is lost by this splitting technique. Asymmetric
random-burst erasure channels. We extend the construction of spatially coupled LDPC codes are proposed in [7] to enhance
the folded-type structure by connecting multiple one-sided SC the burst erasure correcting capability.
LDPC chains for higher resilience to burst erasure channels. The In this letter, we propose a protograph-based design of
FSC LDPC codes are also compatible with windowed decoder folded-type SC LDPC codes, which are formed by simply
and outperform conventional SC LDPC codes. folding the spatial structure of the conventional SC LDPC
Index Terms—Spatially coupled LDPC codes, windowed codes. The sub-blocks, each of which is a subset of vari-
decoder, burst erasure channels, protograph-based codes. able nodes at a spatial position, are interlaced between each
other, which introduces an interleaving effect when codewords
are transmitted over the burst erasure channels. The newly
I. I NTRODUCTION structured codes outperform the conventional SC LDPC codes
PATIALLY coupled (SC) LDPC codes have capac- in the asymptotic analysis and simulations. With the folded
S ity achieving performance over binary-input memoryless
symmetric-output (BMS) channels [1]. The belief propagation
structure, the low latency windowed decoding can still be con-
ducted while keeping regularity in the protograph structure. It
(BP) threshold of these codes saturates to the maximum a should be noted that the idea of folding was also suggested
posteriori (MAP) threshold of same degree distribution uncou- in [11] to get one-sided SC LDPC codes, which have a smaller
pled block-LDPC codes when the coupling length ‘L’ tends rate loss. However, we use a different idea of folding that
to infinity, which was proven for BMS channels in [2]. Good comes with interlacing. The scheme is extended by construct-
performance of these codes is due to the wave-like propaga- ing generalized-FSC LDPC structure from multiple one-sided
tion of reliable information from two terminated sides of the SC LDPC codes for better resiliency to burst erasures.
graph during BP decoding. This wave-like progress of decod-
ing can halt when a burst of erasures is introduced. Therefore,
II. C ODE C ONSTRUCTION
it was shown that, despite their good performance in mem-
oryless channels, the SC LDPC codes perform far from the A. Folded SC LDPC Codes
Singleton bound for burst erasure channels [3]. An extensive Spatial coupling of multiple LDPC protographs results in a
finite-length analysis of SC LDPC codes over different burst base matrix with a convolutional diagonal band structure [9].
erasure channels was done in [8] and tight bounds on erasure Let us consider a regular (J, K) LDPC protograph with base
probability are shown. matrix B, where, J and K are variable and check node degrees,
To improve the burst erasure correcting capability of SC respectively. The diagonal band in the base matrix depends
LDPC codes, convolutional interleavers are used to scatter the on the edge spreading pattern. Different ensembles of SC
erasures in a burst [4]. Multidimensional SC LDPC codes are LDPC codes can be obtained by using different edge spreading
introduced in [5] to overcome burst erasures, but these codes rules [9]. To preserve the degree distribution of the underlying
have a higher rate loss than the conventional SC LDPC codes. (J, K) LDPC base matrix, the valid edge spreading condition
The complexity of designing finite-length multidimensional that must be satisfied is given by
SC LDPC codes is also higher. In [6], the band-splitting w
permutation technique for the SC LDPC codes is proposed,
Bi = B, (1)
Manuscript received August 17, 2018; accepted October 21, 2018. Date i=0
of publication October 30, 2018; date of current version April 9, 2019.
This work was supported by the Basic Science Research Program through where Bi ’s are component base matrices that define edge
the National Research Foundation of Korea (NRF) funded by the Ministry spreading. In this letter, two (3, 6) ensembles are considered
of Education under Grant (NRF-2015R1D1A1A01058975) and Grant (NRF- for performance evaluation. Ensemble A is defined by com-
2018R1A2B6004195). The associate editor coordinating the review of this
paper and approving it for publication was I. Land. (Corresponding author: ponent base matrices B0 = B1 = B2 = [1 1], whereas
Sang-Hyo Kim.) ensemble B by B0 = [2 2], B1 = [0 1], B2 = [1 0].
The authors are with the College of Information and Communication For constructing finite length codes, we follow the procedure
Engineering, Sungkyunkwan University, Suwon 16419, South Korea
(e-mail: inayat@skku.edu; dlguswo77@skku.edu; ayaz@skku.edu;
defined in [9] and [10].
iamshkim@skku.edu). The FSC LDPC codes are obtained by folding the chain
Digital Object Identifier 10.1109/LWC.2018.2878562 structure of SC LDPC codes as illustrated in Fig. 1. Two
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ALI et al.: PROTOGRAPH-BASED FSC LDPC CODES FOR BURST ERASURE CHANNELS 517
TABLE I
SC & FSC LDPC T HRESHOLDS FOR D IFFERENT lb & εb
Fig. 4. BEC threshold (ε∗ ) calculated for a burst at different locations in the
Fig. 3. BEC threshold (ε∗ ) vs burst erasure rate (εb ) for a single burst protograph (lb = 4, εb = 0.6 and 0.57 for ensemble A and B, respectively).
located at the center of the protograph (lb = 6).
TABLE II
AVERAGE T HRESHOLDS FOR R ANDOM B URSTS
In asymptotic settings (i.e., when M → ∞, here M is lifting
factor of graph), we assume that an erasure burst applies in
terms of only complete spatial positions. Therefore, lb is the
number of spatial positions that a burst spans. The erasure rate
threshold for a given burst erasure rate εb is defined as
i→∞ for FSC LDPC codes when the burst location is at the pivot
ε∗ = sup{ε ∈ [ 0, 1 ] | PiBP (ε, εb ) −−−−→ 0 , ε < εb }, (3) of the folded structure, this case does not dominate the over-
all performance due to the random location of a burst in the
where, PiBP (ε, εb ) is the resulting erasure probability at the
SBC. The average threshold ε∗avg for SC and FSC LDPC codes
i th iteration of DE. The thresholds of SC and FSC LDPC
are shown in Table II. From these observations, we note that
codes are compared for different lb and εb in Table I. To find
FSC LDPC codes perform better on average, but do not help
these thresholds, the position of the erasure burst is fixed at
in the worst case, i.e., the pivot location. Moreover, it can
the center of the protograph such that the starting position of
be noted that ensemble L(3, 6, 3, 51) has a relatively better
the burst is j = k = (L−lb )/2
. It can be noted that the FSC
threshold at pivot because of the degree of freedom to choose
LDPC codes have higher thresholds when lb and εb increase.
better connectivity among multiple chains. The thresholds of
Because of the interlacing, the disjoint nodes between the con-
convolutional interleaved SC LDPC codes [4] are also shown
nected nodes split the burst into two from the perspective of
in Fig. 3 and 4 for comparison. We selected the structure of
the original SC LDPC chain. However, if the erasure burst is
convolutional interleaver such that it has the same memory
located at the pivot of the FSC LDPC protograph, the split
as that of FSC LDPC codes. However, this interleaver incor-
bursts are not separated.
porates some delay due to its structural weakness and it also
Figure 3 shows the improvement in burst erasure correcting
requires more hardware block to interleave and deinterleave
capability of proposed FSC LDPC codes. The threshold ε∗ of
the data.
SC LDPC codes decreases to zero at low εb (i.e., εb = 0.5573
and 0.538 for ensemble A and B, respectively), whereas it can
be seen that the FSC LDPC codes outperform SC LDPC codes IV. N UMERICAL R ESULTS
by attaining the same threshold at a much higher burst era- The performance of FSC LDPC codes is evaluated and com-
sure rate. More considerable gain is attained with the ensemble pared with that of conventional SC LDPC codes over SBC and
L(3, 6, 3, 51). The lb = 6 and the burst location is at the cen- MBC. The finite length codes from ensemble A and B are con-
ter of the protograph for Fig. 3. The thresholds for different structed with the code parameters set as L = 50 and M = 512,
burst locations are shown in Fig. 4. It is observed that ε∗ gets so that we obtain code length N = 51200. The girth of the
higher when the burst is located at the terminated side for constructed codes was kept larger or equal to 10. For all the
both SC and FSC LDPC codes. Although ε∗ becomes lower constructed codes, the code rate R = 0.48 is obtained because
ALI et al.: PROTOGRAPH-BASED FSC LDPC CODES FOR BURST ERASURE CHANNELS 519
V. C ONCLUSION
We propose a family of windowed-decodable codes called
FSC LDPC codes, whose structure is formed by folding and
interlacing the protograph of conventional SC LDPC codes.
Fig. 5. FER performances for full block BP decoding of SC and FSC LDPC The DE analysis showed that the FSC LDPC codes are suitable
codes with m = 2 constructed from ensembles A and B over SBC. for transmission over burst erasure channels. We also proposed
a construction method of generalized structures based on the
idea of folding SC LDPC chains. The generality of the result is
confirmed by examining multiple protograph-based ensembles.
By comparing the simulated FER curves, a significant gain in
performance of FSC LDPC codes is observed over single and
multiple random-burst erasure channels under full block BP
and windowed decoding.
R EFERENCES
[1] S. Kudekar, T. Richardson, and R. L. Urbanke, “Spatially coupled
ensembles universally achieve capacity under belief propagation,” IEEE
Trans. Inf. Theory, vol. 59, no. 12, pp. 7761–7813, Dec. 2013.
[2] S. Kumar, A. J. Young, N. Macris, and H. D. Pfister, “Threshold satura-
tion for spatially coupled LDPC and LDGM codes on BMS channels,”
IEEE Trans. Inf. Theory, vol. 60, no. 12, pp. 7389–7415, Dec. 2014.
[3] A. R. Iyenger et al., “Windowed decoding of protograph-based LDPC
Fig. 6. FER performances for windowed decoding of SC and FSC LDPC convolutional codes over erasure channels,” IEEE Trans. Inf. Theory,
codes with m = 2 constructed from ensembles A and B over SBC (εb = 0.8). vol. 58, no. 4, pp. 2303–2320, Apr. 2012.
[4] R. A. Ashrafi and A. E. Pusane, “Spatially-coupled communication
system for the correlated erasure channel,” IET Commun., vol. 7, no. 8,
pp. 755–765, May 2013.
[5] R. Ohashi, K. Kasai, and K. Takeuchi, “Multi-dimensional spatially-
coupled codes,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Istanbul,
Turkey, Jul. 2013, pp. 2448–2452.
[6] H. Mori and T. Wadayama, “Band splitting permutations for spatially
coupled LDPC codes achieving asymptotically optimal burst erasure
immunity,” IEICE Trans. Fundam., vol. E100-A, no. 2, pp. 663–669,
Feb. 2017.
[7] Z. Zhang and Y. Li, “Asymmetric spatially coupled LDPC codes for
multiple-burst erasure channels,” IEEE Commun. Lett., vol. 21, no. 8,
pp. 1695–1698, Aug. 2017.
[8] V. Aref, N. Rengaswamy, and L. Schmalen, “Finite-length analysis of
spatially-coupled regular LDPC ensembles on burst-erasure channels,”
IEEE Trans. Inf. Theory, vol. 64, no. 5, pp. 3431–3449, May 2018.
[9] D. G. M. Mitchell, M. Lentmaier, and D. J. Costello, “Spatially coupled
LDPC codes constructed from protographs,” IEEE Trans. Inf. Theory,
vol. 61, no. 9, pp. 4866–4889, Sep. 2015.
Fig. 7. FER performances for full block BP decoding of SC and FSC LDPC [10] I. Ali, J.-H. Kim, S.-H. Kim, H. Kwak, and J.-S. No, “Improving win-
codes with m = 2 constructed from ensembles A and B over MBC with 6 dowed decoding of SC LDPC codes by effective decoding termination,
random bursts and εb = 0.7. message reuse, and amplification,” IEEE Access, vol. 6, pp. 9336–9346,
2017.
[11] M. R. Sanatkar and H. D. Pfister, “Increasing the rate of spatially-
coupled codes via optimized irregular termination,” in Proc. 9th Int.
of the termination. We apply the standard flooding schedule Symp. Turbo Codes Iterative Inf. Process. (ISTC), Brest, France,
for full block BP and windowed decoding [10]. The maximum Sep. 2016, pp. 31–35.
[12] P. M. Olmos, D. G. M. Mitchell, D. Truhachev, and D. J. Costello, “A
number of iterations (IMAX ) is set as 100. finite length performance analysis of LDPC codes constructed by con-
Figure 5 shows the frame error rate (FER) performances of necting spatially coupled chains,” in Proc. IEEE Inf. Theory Workshop
SC and FSC LDPC codes from ensemble A and B, respec- (ITW), Sevilla, Spain, Sep. 2013, pp. 1–5.
[13] I. Andriyanova, N. Ul Hassan, M. Lentmaier, and G. P. Fettweis,
tively, decoded with full block BP decoder over the SBC. “SC-LDPC codes over the block-fading channel: Robustness to a syn-
We simulated the transmission of the codes over the SBC chronisation offset,” in Proc. IEEE Int. Black Sea Conf. Commun. Netw.
with εb = 0.8, 0.9 and lb = 2048, 2560 to investigate the (BlackSeaCom), Constanta, Romania, May 2015, pp. 97–101.
520 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—We consider the number of users associating with multiple-input multiple-output transmission [9]. Even in a lat-
each base station in a cellular network. Extending and unify- tice network whose cells are of equal size, the shadowing and
ing the characterizations for certain settings available in the the stochastic nature of the user locations would induce dispar-
literature, we derive a result that is asymptotic in the strength
of the shadowing, yet otherwise universally valid: it holds for ities in the number of users associating with different BSs, and
every network geometry and shadowing distribution. We then such disparities are bound to increase in irregular networks.
illustrate how this result provides excellent representations in var- This letter addresses the stochastic modeling of the number
ious classes of networks and with realistic shadowing strengths, of users per BS. An approximate characterization available in
evidencing broad applicability. the literature is discussed, and a new asymptotic characteriza-
Index Terms—Cellular networks, user association, number of tion is provided and tested.
users, shadowing, stochastic geometry, multiuser MIMO, Poisson
point process, lattice networks, PPP networks.
II. M ODELING F EATURES
Our focus is on cellular networks where users associate with
the BS from which they enjoy the strongest large-scale channel
gain. Let us next describe the essential modeling features of
I. M OTIVATION
the networks to which the considerations in the sequel apply.
HE ANALYSIS of cellular networks via Poisson point
T process (PPP) modeling of the base station (BS) loca-
tions is a welcome complement, and sometimes even an
A. Geometries
In terms of the positions of BSs and users, virtually every
outright alternative, to the Monte-Carlo simulations that had
cellular scenario of relevance is encompassed. The BS loca-
long dominated system-level performance evaluations [1]–[3].
tions may conform to any stationary and ergodic point process
Such analysis, seemingly fitting only for ad hoc networks, hap-
Φb ⊂ R2 of density λb , or any realization thereof, say a lat-
pens to be highly relevant to cellular networks by virtue of the
tice network. This implies that the density of BSs within any
following result: regardless of the BS locations, provided only
region converges to λb > 0 as this region’s area grows [6].
that they are agnostic to the radio propagation, the distribu-
Meanwhile, the user positions Φu ⊂ R2 may belong to any
tion of powers received at any user converges (asymptotically
independent point process of density λu that is also stationary
in the strength of the shadowing) to what would be received
and ergodic.
from a PPP field of BSs [4]–[6]. This convergence, moreover,
Without loss of generality, a specific BS is set at the origin.
is very evident for practical strengths of the shadowing. In
In the random case, we condition on a BS to be located at the
hexagonal lattice networks, for instance, PPP-based analyses
origin; under expectation over Φb , this becomes the typical
are highly representative for shadowing standard deviations on
BS. In the deterministic case, we pick an arbitrary BS and
the order of 10 dB [6]–[8], well in line with the typical values
translate the coordinate system so that this BS is located at
encountered in macrocellular deployments.
the origin. In both cases, we label this BS at the origin as the
An issue that arises in the analysis of cellular networks
0th BS. Denoting by K the number of users associating with
is the modeling of the number of users associating with
such 0th BS, our purpose is to inspect the distribution of K
each BS. While hidden if each BS is assumed to commu-
under expectations over Φu and Φb .
nicate with a single user per signaling resource, this issue
becomes material once that assumption is removed, say in the
B. Large-Scale Gains
face of multiantenna BSs capable of implementing multiuser
The large-scale channel gains include path loss with expo-
Manuscript received August 1, 2018; revised September 28, 2018; accepted nent η > 2 and shadowing that is IID across links. Particularly,
October 22, 2018. Date of publication October 30, 2018; date of current between the th BS and the kth user served by the 0th BS, the
version April 9, 2019. This work was supported by MINECO/FEDER, UE large-scale gain is
under Project TEC2015-66228-P, in part by the European Research Council
through the H2020 Framework Programme/ERC under Grant 694974, and L
in part by the U.S. NSF under Award CCF 1525904. The associate editor G,(k ) = ηref χ,(k ) ∈ N0 , k ∈ {0, . . . , K − 1}, (1)
coordinating the review of this paper and approving it for publication was
r,(k )
C. Shen. (Corresponding author: Geordie George.)
G. George and A. Lozano are with the Department of Information and with Lref the path loss intercept at a unit distance, r,(k )
Communication Technologies, Universitat Pompeu Fabra, 08018 Barcelona, the link distance, and χ,(k ) the shadowing coefficient hav-
Spain (e-mail: geordie.george@upf.edu; angel.lozano@upf.edu). ing standard deviation σdB . The shadowing can be arbitrarily
M. Haenggi is with the Department of Electrical Engineering, University
of Notre Dame, Notre Dame, IN 46556 USA (e-mail: mhaenggi@nd.edu). distributed, with the only mild restriction that E χ2/η < ∞
Digital Object Identifier 10.1109/LWC.2018.2878579 to guarantee the asymptotic behavior advanced in Section I.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
GEORGE et al.: DISTRIBUTION OF NUMBER OF USERS PER BS IN CELLULAR NETWORKS 521
Fig. 1. In solid, histograms of K in a double-lattice network with η = 4 and σdB = 0, 10 and 14 dB (in shaded and in clear for the small and large cells,
respectively); users are PPP distributed with K̄ = 10. In dashed, a Poisson PMF with mean K̄.
Fig. 3. In solid, histograms of K in a PPP network with η = 4 and σdB = 0, 10 and 14 dB; users are PPP distributed with K̄ = 10. In dotted, the PMF
in (2). In dashed, a Poisson PMF with mean K̄.
R EFERENCES
[1] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to
coverage and rate in cellular networks,” IEEE Trans. Commun., vol. 59,
no. 11, pp. 3122–3134, Nov. 2011.
[2] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for
modeling, analysis, and design of multi-tier and cognitive cellular wire-
less networks: A survey,” IEEE Commun. Surveys Tuts., vol. 15, no. 3,
pp. 996–1019, 3rd Quart., 2013.
[3] B. Błaszczyszyn, M. Haenggi, P. Keeler, and S. Mukherjee, Stochastic
Geometry Analysis of Cellular Networks. Cambridge, U.K.: Cambridge
Univ. Press, 2018.
[4] H. P. Keeler, N. Ross, and A. Xia, “When do wireless network signals
appear Poisson?” Bernoulli, vol. 24, no. 3, pp. 1973–1994, Aug. 2018.
[5] N. Ross and D. Schuhmacher, “Wireless network signals with mod-
erately correlated shadowing still appear Poisson,” IEEE Trans. Inf.
Theory, vol. 63, no. 2, pp. 1177–1198, Feb. 2017.
[6] B. Błaszczyszyn, M. K. Karray, and H. P. Keeler, “Wireless networks
appear Poissonian due to strong shadowing,” IEEE Trans. Wireless
Commun., vol. 14, no. 8, pp. 4379–4390, Aug. 2015.
Fig. 4. In solid, variance of K as a function of σdB in a PPP network with [7] G. George, R. K. Mungara, A. Lozano, and M. Haenggi, “Ergodic
η = 4; users are PPP distributed with K̄ = 10. In dotted, variance of the PMF spectral efficiency in MIMO cellular networks,” IEEE Trans. Wireless
in (2). In dashed, variance of a Poisson PMF with mean K̄. Commun., vol. 16, no. 5, pp. 2835–2849, May 2017.
[8] G. George, A. Lozano, and M. Haenggi, “Massive MIMO for-
ward link analysis for cellular networks,” 2018 [Online]. Available:
https://arxiv.org/abs/1811.00110
[9] R. W. Heath and A. Lozano, Foundations of MIMO Communication.
predictable, is beneficial in terms of resource provisioning. In Cambridge, U.K.: Cambridge Univ. Press, 2018.
[10] M. Haenggi, Stochastic Geometry for Wireless Networks. Cambridge,
this respect, shadowing turns out to be operationally beneficial. U.K.: Cambridge Univ. Press, 2012.
We hasten to recall that all the foregoing observations [11] H. ElSawy, A. Sultan-Salem, M. S. Alouini, and M. Z. Win, “Modeling
rely on the premise that the BS locations are agnostic to and analysis of cellular networks using stochastic geometry: A tutorial,”
IEEE Commun. Surveys Tuts., vol. 19, no. 1, pp. 167–203, 1st Quart.,
the radio propagation, dominated by deployment opportunities 2017.
and restrictions associated with terrain, permits, infrastructure, [12] H. ElSawy and E. Hossain, “Two-tier HetNets with cognitive femtocells:
power supply, and backhaul. At the other extreme in terms Downlink performance modeling and analysis in a multichannel envi-
ronment,” IEEE Trans. Mobile Comput., vol. 13, no. 3, pp. 649–663,
of planning we would have networks whose BS locations are Mar. 2014.
optimized for coverage and service, with none of those restric- [13] Y. Wang, M. Haenggi, and Z. Tan, “The meta distribution of the SIR for
tions. Such networks are best modeled by regular lattices with cellular networks with power control,” IEEE Trans. Commun., vol. 66,
no. 4, pp. 1745–1757, Apr. 2018.
no shadowing [15], and again the number of users per BS [14] A. Guo and M. Haenggi, “Spatial stochastic models and metrics for the
would then be Poisson-distributed. structure of base stations in cellular networks,” IEEE Trans. Wireless
The broad suitability of a model as simple as the Poisson Commun., vol. 12, no. 11, pp. 5800–5812, Nov. 2013.
[15] A. Guo and M. Haenggi, “Joint spatial and propagation models for cel-
distribution is a welcome finding, especially if a further layer lular networks,” in Proc. IEEE Glob. Commun. Conf., San Diego, CA,
of modeling is to be overlaid so as to incorporate admission USA, Dec. 2015, pp. 1–6.
524 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Joint Power, Altitude, Location and Bandwidth Optimization for UAV With
Underlaid D2D Communications
Wenhuan Huang, Zhaohui Yang, Cunhua Pan , Lu Pei , Ming Chen, Mohammad Shikh-Bahaei,
Maged Elkashlan , and Arumugam Nallanathan , Fellow, IEEE
Abstract—In this letter, we aim to maximize the rate of a presented a cost function based multiple UAVs deployment
device-to-device (D2D) pair for a downlink unmanned aerial vehi- model. By taking beamwidth into account, a joint UAV altitude
cle (UAV)-aided wireless communication system, where D2D users and beamwidth optimization problem for UAV-aided multiuser
coexist in an underlaying manner. We jointly optimize the trans- communication systems was studied in [4]. Through jointly
mit power of the UAV and D2D users, the flying altitude and optimizing altitude, beamwidth and bandwidth, the sum power
location of the UAV and ground terminals’ allocated bandwidth.
To solve this problem, an iterative algorithm with low complexity
was further minimized in [5].
is accordingly proposed. Simulation results show that the altitude Apart from UAV-aided wireless communications, device-
of the UAV has an important impact on the system performance. to-device (D2D) communication has been regarded as one
of the crucial technologies in future wireless communication
Index Terms—UAV communications, D2D communications, networks [6]. D2D communication allows direct transmissions
altitude and location optimization, power allocation, bandwidth between users in proximity, which is helpful in offloading
allocation.
network traffic and reducing end-to-end delay. Compared to
the previous investigations on D2D communication underlay-
ing cellular networks [7], the coexistence of UAVs and under-
I. I NTRODUCTION laid D2D communications will introduce new interference
ECENTLY, wireless communication assisted by management challenges. Unlike fixed BSs, the altitude of the
R unmanned aerial vehicles (UAVs) has been regarded as a
promising technique which can provide economical wireless
UAV is adjustable and will influence channel characteristics.
Moreover, the impacts of the mobility of the UAV on D2D
access for mobile devices without deploying fixed network communications should be analyzed. In [8], a UAV flight
infrastructure [1]. Different from conventional terrestrial pattern selection problem for D2D communications in dis-
communications, UAVs act as flying base stations (BSs) aster areas was studied. Mozaffari et al. [9] focused on the
in UAV-aided wireless communications and bring plenty performance analysis of the coexistence between the UAV and
of benefits. Owing to their agility and mobility, UAVs can an underlaid D2D communication network in a downlink sce-
be deployed to support temporary or urgent events over a nario. However, to our best knowledge, there is no existing
wide area, which enhances the quality of service for ground work studying the performance of UAVs with underlaid D2D
terminals (GTs). Moreover, links between UAVs and GTs communications from the optimization point of view.
are dominated by line-of-sight (LoS) connections, leading to In this letter, we aim to maximize the rate of a D2D pair
enhanced data rate. for a downlink UAV-aided wireless communication system,
To fully reap the benefits of UAV-aided communications, it where D2D users coexist in an underlaying manner. We for-
is crucial to exploit the UAV mobility in a three-dimensional mulate the problem of joint power, altitude, location and
space. To address the UAV deployment challenge, an efficient bandwidth optimization. In order to solve this nonconvex rate
deployment approach based on the circle packing theory was maximization problem, we propose a low-complexity iterative
proposed in [2]. For capacity enhancement, Sharma et al. [3] algorithm. It turns out to have attractive closed-form solu-
tions for power allocation subproblem and altitude planning
Manuscript received October 1, 2018; accepted October 25, 2018. Date of subproblem.
publication October 30, 2018; date of current version April 9, 2019. This work
was supported in part by the National Natural Science Foundation of China
under Grant 61871128, Grant 61521061, and 61571125, and in part by the II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
Engineering and Physical Science Research Council through SENSE under
Grant EP/P003486/1. The work of M. Elkashlan was supported by the EPSRC We consider a downlink UAV-aided wireless communica-
under Project EP/N029666/1. The work of A. Nallanathan was supported tion system1 with one flying UAV, K GTs and one D2D
by EPSRC under Project EP/M016145/2. The associate editor coordinating pair which coexists in an underlaying manner. The horizontal
the review of this paper and approving it for publication was C. Huang.
(Corresponding author: Cunhua Pan.)
and vertical locations of the D2D receiver, D2D transmitter
W. Huang, L. Pei, and M. Chen are with the National Mobile and GT k are denoted by r = (0, 0), t = (t(1), t(2)) and
Communications Research Laboratory, Southeast University, Nanjing g k = (gk (1), gk (2)), respectively. The altitude of the D2D
210096, China (e-mail: huang_wenhuan@seu.edu.cn; peilu@seu.edu.cn; pair and GTs are assumed to be zero. The UAV is deployed
chenming@seu.edu.cn). as a flying BS at an altitude H with horizontal and vertical
Z. Yang and M. Shikh-Bahaei are with the Centre for Telecommunications
Research, King’s College London, London WC2B 4BG, U.K. (e-mail: location u = (x, y).
yang.zhaohui@kcl.ac.uk; m.sbahaei@kcl.ac.uk).
C. Pan, M. Elkashlan, and A. Nallanathan are with the School of 1 We consider the system where all terminals are equipped with a sin-
Electronic Engineering and Computer Science, Queen Mary University gle isotropic antenna. When each user is equipped with one antenna, the
of London, London E1 4NS, U.K. (e-mail: c.pan@qmul.ac.uk; optimization problem becomes more tractable. Moreover, it is appealing
maged.elkashlan@qmul.ac.uk; a.nallanathan@qmul.ac.uk). to equip each device with only one antenna to reduce the implementation
Digital Object Identifier 10.1109/LWC.2018.2878706 complexity.
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/
HUANG et al.: JOINT POWER, ALTITUDE, LOCATION AND BANDWIDTH OPTIMIZATION FOR UAV WITH UNDERLAID D2D COMMUNICATIONS 525
We consider the case that GTs and the D2D receiver are III. S OLUTION A PPROACH
located outdoors, and the channel between the UAV and each Due to (4a) and (4d), Problem (4) is a nonconvex problem.
GT (D2D receiver) is dominated by the LoS path. The down- It is difficult to obtain its globally optimal solution. In the
link channel power gain between the UAV and GT k is given following, we propose a low-complexity iterative algorithm.
by [4] Since (4a) is a monotonically decreasing function of pku , the
β0 data rate constraints in (4d) for GTs should hold with equality
hku = , (1) at the optimal point. As a result, we have
u − g k 2 + H 2
Rmin
where β0 denotes the channel power gain between the UAV 2 ak
−1
and GTs (D2D receiver) at the unit distance, and · is pku ∗ = (u − g k 2 + H 2 ) pd hkd + σ 2 , (5)
β0
the Euclid norm. Similarly, the downlink channel power gain
between the UAV and the D2D receiver is where pku ∗ denotes the optimal solution of pku .
β0 To ensure that Problem (4) is feasible, we employ the fea-
h0u = 2
. (2) sibility checking algorithm. The feasibility checking problem
u + H 2 is the minimization of transmit power of the UAV subject to
The downlink achievable rate of GT k can be expressed as constraints (4b), (4d)-(4g). If the minimal sum power of the
UAV is larger than Pumax , Problem (4) is infeasible. Since pku ∗
g pku β0 is the minimal value of pku according to (4d), the feasibility
rk = ak log2 1 + , (3)
pd hkd + σ 2 u − g k 2 + H 2 checking problem is equivalent to obtain the minimum value
v ∗ of K p
k =1 k
u ∗ , when p = 0, H = H
d min , and constrains
where ak denotes the allocated bandwidth proportion for GT (4f) and (4g) are satisfied. An exhaustive algorithm can be
k, pku denotes the transmit power from UAV to GT k, pd is adopted to solve it. With fixed u, the optimal bandwidth allo-
the transmit power of the D2D transmitter, hkd is the channel cation a can be obtained via the interior point method. The
power gain between the D2D transmitter and GT k, and σ 2 is optimal u is obtained via the two-dimensional (2D) exhaustive
the power of the additive white Gaussian noise. search. As a result, Problem (4) is feasible if and only if when
We aim for maximizing the rate of the D2D pair while Pumax ≥ v ∗ .
satisfying the minimal rate requirements of all GTs via power, Based on (5) and the fact that log2 (1+x ) is a monotonically
altitude, location and bandwidth optimization. Mathematically, increasing function, Problem (4) is equivalent to
the D2D pair achievable rate maximization problem is
pd h0d
⎛ ⎞ max
Rmin
(6a)
u−g k 2 +H 2
pd ,a,u,H K
ak 2 ak
−1 pd hkd + σ 2 + σ 2
⎜ pd h0d ⎟ H 2 +u2
⎜ ⎟ k =1
max log2 ⎜1 + ⎟ (4a) Rmin
pd ,pu ,a,u,H ⎝ K
β0 ⎠ K
2 ak
−1
ak pku + σ2 s.t. 0≤
2 2 d 2
(u − g k + H )(pd hk + σ ) ≤ Pu
max
,
H 2 +u2 β0
k =1 k =1
V. C ONCLUSION
In this letter, we aim to maximize the rate of a D2D pair
with the coexistence between the UAV and an underlaid D2D
communication network in a downlink scenario. Numerical
results show that the proposed algorithm always outperforms
the algorithms which partially optimize altitude, location or
bandwidth. Moreover, the altitude of the UAV cannot be too
high or too low because the altitude of the UAV has an impor-
tant influence on the UAV-aided networks with underlaid D2D
communications.
Fig. 2. D2D achievable rate versus minimal rate demand.
R EFERENCES
IV. N UMERICAL R ESULTS
[1] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications
We consider a scenario where there are K = 3 GTs and with unmanned aerial vehicles: Opportunities and challenges,” IEEE
one D2D pair randomly distributed in a circle area with Commun. Mag., vol. 54, no. 5, pp. 36–42, May 2016.
radius 300 m. The distance between the D2D transmitter and [2] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Efficient deploy-
ment of multiple unmanned aerial vehicles for optimal wireless cover-
the corresponding receiver is set to 30 m. We set the total age,” IEEE Commun. Lett., vol. 20, no. 8, pp. 1647–1650, Aug. 2016.
bandwidth of the system as 10 MHz, β0 = 1.42 × 10−4 , [3] V. Sharma, M. Bennis, and R. Kumar, “UAV-assisted heterogeneous
σ 2 = −169 dBm/Hz, Hmin = 50 m, Hmax = 500 m, networks for capacity enhancement,” IEEE Commun. Lett., vol. 20, no. 6,
Pmax
d =10 dBm and Pmax u = 30 dBm. pp. 1207–1210, Jun. 2016.
[4] H. He, S. Zhang, Y. Zeng, and R. Zhang, “Joint altitude and
Similar to [14], the channel power gain hkd is modeled beamwidth optimization for UAV-enabled multiuser communications,”
as β0 ηd −3 , k = 0, 1, . . . , K , where d represents distance IEEE Commun. Lett., vol. 22, no. 2, pp. 344–347, Feb. 2018.
and η represents the Rayleigh fading coefficient, which fol- [5] Z. Yang et al., “Joint altitude, beamwidth, location, and bandwidth
optimization for UAV-enabled communications,” IEEE Commun. Lett.,
lows the exponential distribution with unit mean. We compare vol. 22, no. 8, pp. 1716–1719, Aug. 2018.
the proposed algorithm with the following three algorithms: [6] M. N. Tehrani, M. Uysal, and H. Yanikomeroglu, “Device-to-device
fixed altitude and location algorithm with optimized bandwidth communication in 5G cellular networks: Challenges, solutions, and
(labeled as ‘FAL’), fixed bandwidth algorithm with optimized future directions,” IEEE Commun. Mag., vol. 52, no. 5, pp. 86–92,
altitude and location (labeled as ‘FB’), and the near globally May 2014.
[7] C.-H. Yu, K. Doppler, C. B. Ribeiro, and O. Tirkkonen, “Resource
optimal algorithm via running the proposed iterative algorithm sharing optimization for device-to-device communication underlaying
with 1000 initial points (labeled as ‘NGO’). cellular networks,” IEEE Tans. Wireless Commun., vol. 10, no. 8,
Firstly, we study the convergence behaviour of the proposed pp. 2752–2763, Aug. 2011.
algorithm. Fig. 1 illustrates the achievable rate of the D2D [8] E. Christy et al., “Optimum UAV flying path for device-to-device com-
munications in disaster area,” in Proc. IEEE Int. Conf. Signal Syst.,
pair versus the number of iterations when Hmin = 50 m. Denpasar, Indonesia, May 2017, pp. 318–322.
It shows that the achievable rate of the D2D pair increases [9] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Unmanned aerial
monotonically and converges rapidly. vehicle with underlaid device-to-device communications: Performance
The achievable rate of the D2D pair with differ- and tradeoffs,” IEEE Trans. Wireless Commun., vol. 15, no. 6,
ent approaches versus Rmin when Hmin = 50 m and pp. 3949–3963, Jun. 2016.
[10] Q. Wu, G. Y. Li, W. Chen, and D. W. K. Ng, “Energy-efficient D2D
Hmin = 200 m is illustrated in Fig. 2. We observe that the overlaying communications with spectrum-power trading,” IEEE Trans.
achievable rate of the D2D pair decreases when Hmin is Wireless Commun., vol. 16, no. 7, pp. 4404–4419, Jul. 2017.
increased for all algorithms. Moreover, the altitude of the [11] Z. Yang et al., “Downlink resource allocation and power control for
UAV cannot be too high or too low. If the altitude of the device-to-device communication underlaying cellular networks,” IEEE
Commun. Lett., vol. 20, no. 7, pp. 1449–1452, Jul. 2016.
UAV is too high, the coverage region of the UAV is enhanced [12] Z. Yang, C. Pan, W. Xu, H. Xu, and M. Chen, “Joint time allocation
while the channel gain is weak, which leads to high trans- and power control in multicell networks with load coupling: Energy sav-
mit power from the UAV to satisfy the rate demand of GTs ing and rate improvement,” IEEE Trans. Veh. Technol., vol. 66, no. 11,
and severe interference to the D2D pair. If the altitude of the pp. 10470–10485, Nov. 2017.
UAV is too low, the channel gains are high from the UAV [13] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
Cambridge Univ. Press, 2004.
to GTs and the D2D receiver, which will result in terrible [14] T. D. Hoang, L. B. Le, and T. Le-Ngoc, “Energy-efficient resource allo-
interference to the D2D pair. From Fig. 2, we observe that cation for D2D communications in cellular networks,” IEEE Trans. Veh.
the proposed algorithm respectively achieves 20% and 10% Technol., vol. 65, no. 9, pp. 6972–6986, Sep. 2016.
528 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—We propose a multiple-access (MA) method in the the delay and Doppler spreads are high, the information trans-
uplink of a orthogonal time frequency space modulation-based mitted on a Delay-Doppler resource block (DDRB) could
wireless communication system where the channel has high spread to neighbouring DDRB’s leading to severe multi-user
Doppler and delay spread. Each user terminal (UT) is allocated interference (MUI). In [9] guard bands have been proposed to
delay-Doppler (DD) resource blocks which are spaced at equal increase the separation between the DDRBs allocated to dif-
intervals in the DD domain. This limits the corresponding time- ferent UTs so as to reduce MUI. This however reduces the
frequency (TF) transmit signal to a sub-domain of the entire effective system capacity as guard bands are an overhead.
TF domain. By allocating non-overlapping portions to the TF
In this letter, we propose a novel MA method for OTFS
transmit signal of different UTs, multi-user interference (MUI)
is avoided. The proposed MA method is analytically shown to be based systems in high mobility and delay spread scenarios,
MUI free and therefore has a significantly higher sum spectral which multiplexes UTs in both the DD and the TF domains
efficiency when compared to other methods proposed in literature in such a manner that there is no MUI and no guard bands are
which utilize guard bands in the DD domain to reduce MUI. required. The DDRBs allocated to a UT are spaced at equal
intervals in both the delay and the Doppler domain. Due to
Index Terms—Doppler, delay spread, OTFS, multiple-access. this, the corresponding TF signal occupies only a portion of the
entire TF domain. This enables us to allocate non-overlapping
portions of the TF domain to different UTs which allows the
I. I NTRODUCTION BS to separate the received TF signals of different UTs due
CHIEVING high data rates in high mobility and delay to which there is no MUI. We derive the effective channel
A spread wireless channels1 is an objective of IMT-
2020 [1]. In Orthogonal Frequency Division Multiple Access
model for the proposed method and analytically show that
it is MUI free. Simulations show that under high delay and
(OFDMA), high mobility results in inter-carrier interference Doppler spread the sum spectral efficiency (SE) achieved by
which degrades channel capacity [3]. Recently, Orthogonal the method in [9] saturates with increase in the power transmit-
Time Frequency Space Modulation (OTFS) has been intro- ted by each UT and is significantly smaller than that achieved
duced, which has been shown to be more robust to Doppler by the proposed method.
spread as compared to OFDM based systems [4], [5]. OTFS
modulation spreads each information symbol in the Delay
Doppler (DD) domain over the entire time-frequency (TF) II. S YSTEM M ODEL
domain (unlike OFDMA systems) due to which full time- In OTFS modulation, the information symbols are trans-
frequency diversity is realizable [5]. Each information symbol mitted in the delay Doppler (DD) domain instead of the
sees the same constant channel gain, which greatly simplifies time-frequency domain (as in OFDM). The DD domain is T
the transmitter and receiver design [6]–[8]. The constant chan- seconds wide along the delay domain and Δf = 1/T Hz wide
nel gain across the entire DD domain also helps in reducing along the Doppler domain. The delay domain and the Doppler
the overhead of frequent channel estimation and feedback. domain are further sub-divided into M and N equal parts
In this letter, we consider OTFS based uplink multiple- respectively. A DD resource block (DDRB) is a combination
access (MA) with single antenna at the base station (BS) and at of a sub-division along the delay domain together with a sub-
the user terminals (UTs). As this topic is recent, there is little division along the Doppler domain (see Fig. 1a). Information
prior literature available. In [9], UTs have been multiplexed by symbols are transmitted on DDRBs, and hence they get
allocating non-overlapping regions in the DD domain. When delayed and Doppler shifted by known integer multiples of
T/M and Δf /N respectively, prior to transmission.2
Manuscript received June 13, 2018; revised September 26, 2018; accepted The received time domain signal is transformed back to the
October 24, 2018. Date of publication October 30, 2018; date of current ver- DD domain. Each multipath of the wireless channel induces
sion April 9, 2019. This work was supported in part by the Visvesvaraya delay and Doppler shift, and therefore an information symbol
Ph.D. Scheme of Ministry of Electronics and IT, Government of India, being transmitted on a single DDRB could be received at multi-
implemented by Digital India Corporation and in part by EMR Funding ple DDRBs. These received symbols can then be coherently
from the Science and Engineering Research Board, Department of Science
and Technology, Government of India. The associate editor coordinating
combined thereby achieving delay and Doppler diversity, and
the review of this paper and approving it for publication was W. Zhang. hence robustness towards Doppler shift. However, the maxi-
(Corresponding author: Saif Khan Mohammed.) mum possible delay spread τmax of the channel must be less
The authors are with the Department of Electrical Engineering, than T since otherwise two different paths having a delay dif-
Indian Institute of Technology Delhi, New Delhi 110016, India ference of T seconds would appear to the OTFS receiver as a
(e-mail: saifkmohammed@gmail.com).
Digital Object Identifier 10.1109/LWC.2018.2878740
1 One such scenario is that of an urban macro base station (BS) with a large 2 As the delay resolution is T/M, the transmitted time domain signal can
coverage area serving mobile UTs travelling at high speed. An example of significantly change in a time scale of the order of T/M seconds and hence
this is the standardized 3GPP Extended Typical Urban (ETU 300) channel its bandwidth is M/T = MΔf Hz. The minimum possible Doppler shift is
model where the delay spread is 5 μs and the maximum Doppler shift is Δf /N and hence for the receiver to identify such a small shift, the total time
300 Hz [2]. duration of the OTFS frame is N /Δf = NT seconds (see Fig. 1b).
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
KHAMMAMMETTI AND MOHAMMED: OTFS-BASED MA IN HIGH DOPPLER AND DELAY SPREAD WIRELESS CHANNELS 529
single multipath and similarly in the Doppler domain the max- III. P ROPOSED OTFS-MA S CHEME
imum possible Doppler shift νmax must be less than Δf , i.e.,
T is chosen such that Δf = 1/T > νmax and T > τmax .3 In this letter, we propose a novel MA scheme for Q = g1 g2
Let xq [k , l ] denote the information symbol transmitted UTs. We assume that g1 and g2 divide M and N respectively.
by the q-th UT on the (k, l)-th DDRB. Using the Inverse The q-th UT (0 ≤ q < g1 g2 ) is allocated DDRB’s in the set
Δ
Symplectic Finite Fourier Transform (Inverse SFFT), the DD Rq = (k , l ) | k = q/g1 + g2 u, l = (q)g1 + g1 v , 0 ≤ u <
information symbols are transformed to the TF domain (NT N /g2 , 0 ≤ v < M /g1 . The corresponding TF symbols are
sec × MΔf Hz). The modulated TF symbol transmitted by the given by
q-th UT on the (m, n)-th time-frequency RB (TFRB) is given
−j 2π ml
− nk 1 −j 2π ml
M − nk
N
1 N −1 M −1 M N Xq [n, m] = xq [k , l ] e
by Xq [n, m] = MN k =0 l=0 xq [k , l ] e , MN
m = 0, 1, . . . , (M − 1), n = 0, 1, . . . , (N − 1). (k ,l)∈Rq
N M
These are then converted to time domain and trans- −1 −1
M−1 N
−1 (a)
g
g2 1
−j 2π( Mmv − Nnu )
mitted [4], i.e., sq (t) = Xq [n, m]gtx (t − = λ(m, n) x̃q [u, v ] e /g /g 1 2
hq (τ, ν) = 0 , |τ | > τmax or |ν| > νmax (3) 4 Since the transformation of DD domain information symbols to the time-
domain signal and its inverse are both linear, the single-antenna channel model
and that the transmit (Tx) and receive (Rx) pulse satisfy in (5), can be extended to a scenario where both the UTs and BS have multiple
antennas and each UT can send distinct DD domain information symbols
∗
grx (t − τ )gtx (t)e −j 2πν(t−τ ) dt = 0 from each of its antennas. The DD domain symbol received on a DDRB at a
particular BS antenna will simply be a sum of the symbols received on that
DDRB at the same BS antenna from all the transmit antennas of all the UTs.
|τ − nT | ≤ τmax , |ν − mΔf | ≤ νmax , (m, n) = (0, 0) (4) 5 As the effective channel in the DD domain is a 2-D circular convolution
(see (5)), a symbol transmitted by a UT on a DDRB can interfere with another
3 Also, the transmission time NT of the OTFS frame should be sufficiently
symbol transmitted by another UT on an adjacent DDRB either along the
small in order that the change in the delay and Doppler shift of each multipath delay domain if the delay spread is larger than T/M, or along the Doppler
is insignificant compared to M T and Δf respectively, during this period. domain if the maximum Doppler shift is more than Δf /N , leading to MUI.
N
530 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
× (M /g1 )q/g2 Δf , (M /g1 )(q/g2 +1)Δf Hz of the TF For the q-th UT the BS transforms Yq [ñ, m̃] back to the
domain (see Fig. 1b). The corresponding time-domain signal DD domain through SFFT, i.e., for all 0 ≤ ũ < N /g2 and
transmitted from the q-th UT is 0 ≤ ṽ < M /g1 we have
N /g2 −1 M /g1 −1
M /g1 −1 N /g2 −1
Δ j 2π Mm̃/g
ṽ − ñ ũ
xq [ũ, ṽ ] = Yq [ñ, m̃]e 1 N /g 2 . (13)
sq (t) = Xq [n, m]gtx t − (n + (N /g2 )(q)g2 )T
ñ=0 m̃=0
m=0 n=0
Theorem 2: The output of the proposed SFFT in (13), i.e.,
e j 2π m+(M /g1 )q/g2 Δf t−(n+(N /g2 )(q)g2 )T
. (7) xq [., .] is related to the information symbols transmitted by the
Let the DD channel for the q-th UT be [6] q-th UT (i.e., x̃q [., .]) through the 2-D circular convolution
N /g2 −1 M /g1 −1
pq xq [ũ, ṽ ] = x̃q [u, v ]h̃q (ũ − u)N /g2 , (ṽ − v )M /g1 )
hq (τ, ν) = hq,i δ(τ − τq,i )δ(ν − νq,i ) (8) u=0 v =0
i=1 + wq [ũ, ṽ ] ,
pq
τ
−j 2π νq,i τq,i + q,i M q − νq,i N (q)
Δ g2
where δ(·) is the impulse function, τq,i and νq,i are the delay h̃q [k , l] = hq,i e T g1 g2 Δf g2
and Doppler along the i-th multipath and pq is the number of i=1
paths. Further, we assume that hq (τ, ν) satisfies (3), i.e., Dq,i [l]Eq,i [k ]
M /g1 −1
(q)g
|τq,i | ≤ τmax , |νq,i | ≤ νmax . (9) Δ 1 −j 2πm 1 l
− M /g
τ
q,i
+ T
Dq,i [l] = e M 1
M
m=0
At the BS, for the q -th UT the BS computes N /g2 −1 q/g
Δ 1
1 − k νq,i
Eq,i [k ] = e
j 2πn N N /g2
+ Δf
(14)
∗ N
Yq [ñ, m̃] = grx t − (ñ + (N /g2 )(q )g2 )T r (t) n=0
where wq [ũ, ṽ ] is additive noise.
m̃+(M /g1 )q /g2 Δf t−(ñ+(N /g2 )(q )g2 )T
e −j 2π dt. Proof: We substitute the expression of Xq [n, m] from (6)
ñ = 0, 1, . . . , (N /g2 − 1), m̃ = 0, 1, . . . , (M /g1 − 1). (10) into (11) of Theorem 1, which gives us an expression
for Yq [ñ, m̃] in terms of x̃q [., .]. We then substitute this
expression of Yq [ñ, m̃] into the R.H.S. of (13), which
Theorem 1: For the q -th UT, Yq [ñ, m̃] computed at the
gives us (14).6
BS is free from MUI and is given by
In (14), wq (ũ, ṽ ) ∼ CN (0, 1/g1 g2 ) since the summa-
tion in (13) is only over (MN /g1 g2 ) terms as compared
Yq [ñ, m̃] = Xq [ñ, m̃] H̃q [ñ, m̃] + Wq [ñ, m̃]
to MN terms in (2) (w [k , l ] ∼ CN (0, 1) in (5)). The
pq q x̃q +
Δ 2-D convolution in (14) can be written as xq = H
H̃q [ñ, m̃] = hq ,i e −j 2πνq ,i τq ,i e j 2πνq ,i ñ+(N /g2)(q )g2 T Δ
i=1 w̃q where x̃q = x̃q [0, 0], . . . , x̃q [ gN2 − 1, 0], . . . , x̃q [0, Mg1 −
MN
m̃+(M /g1 )q /g2 T ×1
e −j 2πτq ,i Δf (11) 1], . . . , x̃q [ gN2 − 1, M
g1 − 1] and similarly
xq ∈ C g1 g2 is
MN MN
related to xq [., .]. The matrix H q ∈ C g1 g2 × g1 g2 is derived
where Wq [ñ, m̃] is the additive noise.
from the 2-D convolution in (14). With i.i.d. CN (0, ρ) symbols
Proof: Using the expression for sq (t) from (7) and that of
x̃q [u, v ] and Gaussian distributed noise, the sum SE (bits/s/Hz)
hq (τ, ν) from (8) in (1), we get the expression for r(t) using
is given by
which in (10) we get
Q−1
Δ 1 qH H .
N /g2 −1 M /g1 −1 Q−1 pq
R= log2 I + ρg1 g2 H q (15)
MN
Yq [ñ, m̃] = Xq [n, m]hq,i q=0
n=0 m=0 q=0 i=1
IV. N UMERICAL R ESULTS
e −j 2πν
q,i τq,i e −j 2πτq,i Δf m+(M /g1 )q/g2 We consider an OTFS based communication system where
Fi (ñ, m̃, n, m, q, q ) + Wq [ñ, m̃] Δf = 1/T = 15 KHz, M = 36, N = 18 and the channel model
Δ ∗ is the same for all the UTs. We consider the 3GPP standardized
Fi (ñ, m̃, n, m, q, q ) = grx t − (ñ + (N /g2 )(q )g2 )T Extended Typical Urban (ETU 300) channel model [2], where
the maximum Doppler shift is νmax = 300 Hz.7 The channel
gtx (t − τq,i − nT − (N /g2 )(q)g2 T )
6 Using the fact that the effective single-antenna channel model in (5) can
j 2πt νq,i +Δf m−m̃+(M /g1 )(q/g2 −q /g2 )
be extended to the multi-antenna scenario (where UTs and BS have multiple
e dt. (12) antennas), the analysis in this section can be extended to show that the main
results of this letter, i.e., MUI free attribute of the proposed MA scheme in
Since n, ñ ∈ [0, N /g2 ), m, m̃ ∈ [0, M /g1 ), Theorem 1 and the 2-D circular convolution in Theorem 2 are preserved even
(τq,i , νq,i ) satisfy (9) and for q = q , ((q)g2 , q/g2 ) =
for the multi-antenna scenario.
7 The delay profile {τ Pq =9
((q )g2 , q /g2 ), from the property of Tx and q,i }i=1 is [0, 50, 120, 200, 230, 500, 1600,
Rx pulse in (4)
Pq =9
2300, 5000] ns and the corresponding power profile {E[|hq,i |2 ]}i=1 is
we get Fi (ñ, m̃, n, m, q, q ) = e j 2πνq ,i T ñ+(N /g2 )(q )g2 [−1, −1, −1, 0, 0, 0, −3, −5, −7] dB. We further normalize this power
δ(m − m̃)δ(n − ñ)δ(q − q ) using which in (12) 9
profile such that E[|hq,i |2 ] = 1.
we get (11). i=1
KHAMMAMMETTI AND MOHAMMED: OTFS-BASED MA IN HIGH DOPPLER AND DELAY SPREAD WIRELESS CHANNELS 531
V. C ONCLUSION
Fig. 2. MA Method in [9] (M = N = 9, Q = 3, G = 2).
In this letter, we have proposed a novel Multiple-Access
method for OTFS based communication systems which is ana-
lytically shown to be MUI free. Numerical simulations confirm
that the proposed method is able to achieve high sum spectral
efficiency in high Doppler and delay spread channels.
R EFERENCES
[1] “IMT vision—Framework and overall objectives of the future deploy-
ment of IMT for 2020 and beyond,” Int. Telecommun. Union, Geneva,
Switzerland, ITU-Recommendation M-2083-0, Sep. 2015. [Online].
Available: www.itu.int
[2] LTE, Evolved Universal Terrestrial Radio Access (E-UTRA); Base
Station (BS) Radio Transmission and Reception, Version 8.6.0, 3GPP
Standard ETSI TS 36.104, Jul. 2009.
[3] T. Wang, J. G. Proakis, E. Masry, and J. R. Zeidler, “Performance
degradation of OFDM systems due to doppler spreading,” IEEE Trans.
Wireless Commun., vol. 5, no. 6, pp. 1422–1432, Jun. 2006.
[4] R. Hadani et al., “Orthogonal time frequency space modulation,” in
Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), San Francisco,
CA, USA, Mar. 2017.
[5] R. Hadani and A. Monk, “OTFS: A new generation of modulation
addressing the challenges of 5G,” arXiv:1802.02623[cs.IT], Feb. 2018.
[Online]. Available: www.arxiv.org
Fig. 3. Sum Spectral Efficiency versus average received SNR. [6] K. R. Murali and A. Chockalingam, “On OTFS modulation for high-
doppler fading channels,” in Proc. Inf. Theory Workshop Appl. (ITA),
Feb. 2018, pp. 1–10.
[7] R. Raviteja, K. T. Phan, Q. Jin, Y. Hong, and E. Viterbo, “Low-
gains hq,i in (8) are modeled as independent Rayleigh faded complexity iterative detection for orthogonal time frequency space
random variables. The Doppler shift for the i-th path is taken modulation,” arXiv:1709.09402[cs.IT], Sep. 2017. [Online]. Available:
to be νq,i = νmax cos(θq,i ) where θq,i are independent and www.arxiv.org
uniformally distributed in [0, 2π) [7]. Note that νmax < Δf [8] A. Farhang, A. R. Reyhani, L. E. Doyle, and B. Farhang-Boroujeny,
“Low complexity modem structure for OFDM-based orthogonal time
and τmax < T , i.e., (3) is satisfied. frequency space modulation,” IEEE Wireless Commun. Lett., vol. 7,
We compare the average sum SE achieved by the proposed no. 3, pp. 344–347, Jun. 2017.
MA method (i.e., R in (15) averaged over the channel statis- [9] S. Rakib and R. Hadani, “Multiple access in wireless telecommunica-
tics) to that achieved by the method proposed in [9], as a tions system for high-mobility applications,” U.S. Patent 9 722 741 B1,
function of the received signal to noise ratio (SNR). The Aug. 2017.
[10] A. El Gamal and Y. H. Kim, Network Information Theory. Cambridge,
received SNR is the ratio of the received signal power from U.K.: Cambridge Univ. Press, 2011.
a UT (i.e., E[ |sq (t)|2 dt]/(NT )) to the noise power in the
total bandwidth of MΔf Hz. For the proposed method, from
the expressions derived in Section III it follows that the aver-
age received SNR is ρ/Q. The MA method in [9] allocates
non-overlapping contiguous RBs in the DD domain to each 8 With a fixed received SNR for each UT, although the number of DDRBs
UT, with DDRBs allocated to distinct UTs being separated
allocated to each UT decreases with increasing Q (number of UTs), the energy
by guard bands either in the delay domain or in the Doppler of each transmitted information symbol increases proportionately. This along
domain. For example, a guard band in the delay domain con- with the MUI free attribute of the proposed MA method explains our obser-
sists of a contiguous region in the DD domain which is G vation on the sum SE being roughly the same irrespective of Q. This also
RBs wide along the delay domain and N RBs wide along implies that the per-UT SE decreases with increasing Q, which is however a
the Doppler domain (see Fig. 2). For the method in [9], G is fundamental fact from information theory of multiple access channels where
the BS has a single antenna, which limits the degrees of freedom to one [10].
chosen so as to maximize the sum SE. 9 As the MUI free attribute of the proposed MA method holds also for
In Fig. 3 we plot the average sum SE versus the received the multi-antenna scenario, the proposed MA method is expected to achieve
SNR. For a fixed number of UTs, the sum SE achieved by the higher sum SE than that achieved by the MUI limited guard band based
method in [9] saturates with increasing SNR, and also that method even in the multi-antenna scenario.
532 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In this letter, we consider optimal hybrid multiuser massive MIMO systems. However, most previous
beamforming design to minimize the transmission power works on multiuser hybrid beamforming design fail to con-
under individual signal-to-interference-plus-noise ratio (SINR) sider individual SINR constraints. Geng et al. [4] consider
constraints in a multiuser massive multiple-input-multiple- a non-convex multiuser hybrid beamforming design problem
output (MIMO) system. This results in a challenging non-convex with individual SINR constraints and propose a semidefi-
optimization problem. We consider two cases. In the case where nite relaxation-based alternating (SDR-Alt) algorithm to obtain
the number of users is smaller than or equal to that of radio a feasible solution. In particular, in each iteration, a digi-
frequency (RF) chains, we propose a low-complexity method to
tal beamforming design problem is solved by computing a
obtain a globally optimal solution and show that it achieves
the same transmission power as an optimal fully-digital beam- semi-closed form solution, and an analog beamforming design
former. In the case where the number of users is larger than that problem is solved with complexity O(M 4.5 N 4.5 ) using stan-
of RF chains, we propose a low-complexity globally convergent dard techniques for semidefinite programming (SDP), where
alternating algorithm to obtain a stationary point. M denotes the number of antennas and N denotes the num-
ber of RF chains. Moreover, most of previous works (e.g., [4])
Index Terms—Multiuser massive MIMO, hybrid beamforming, focus on the case where the number of users is no greater than
power minimization, penalty method. that of RF chains, and hence cannot provide meaningful solu-
tions for the emerging massive connectivity applications. To
I. I NTRODUCTION our knowledge, hybrid beamformer optimizations with indi-
ITH a large number of antennas deployed in massive
W multiple-input-multiple-output (MIMO) systems, power
consumption and cost of devices increase significantly and
vidual SINR constraints in multiuser massive MIMO systems
have not been successfully solved.
In this letter, we consider a multiuser massive MIMO system
may not be affordable for practical implementation. To address with K users and N RF chains and assume perfect channel
these issues, hybrid analog/digital structure with a reduced state information (CSI). We study optimal hybrid beamforming
number of radio frequency (RF) chains has been regarded as a design to minimize the transmission power subject to individ-
promising solution. Analog beamforming refers to the analog ual SINR constraints. The resulting challenging non-convex
operations applied to a signal before being transmitted through problem is solved in two cases. In the case of K ≤ N, we
antennas, and digital beamforming refers to the baseband sig- propose a low-complexity method to obtain a globally optimal
nal processing applied to a signal before being sent to RF solution and show that it achieves the same transmission power
chains. as an optimal fully-digital beamformer with a reduced num-
Hybrid beamforming technologies have been widely stud- ber of RF chains, by connecting the original optimization
ied in both point-to-point and multiuser massive MIMO problem to a fully-digital beamforming design problem. In
systems [1]–[3]. It is desirable to consider individual signal-to- the case of K > N, we propose a low-complexity globally
interference-plus-noise ratio (SINR) constraints to guarantee convergent alternating algorithm to obtain a stationary point,
quality of service (QoS) requirements for different users in based on problem transformation and a penalty method. To the
best of our knowledge, the proposed solutions are so far the
Manuscript received September 1, 2018; revised October 18, 2018; accepted most promising ones in the two cases in terms of computa-
October 20, 2018. Date of publication October 30, 2018; date of current tional complexity and theoretical guarantee. Finally, numerical
version April 9, 2019. This work was supported in part by the National Natural
Science Foundation of China under Grant 61401272, Grant 61771309, Grant results show that the proposed solutions have much lower
61671301, Grant 61420106008, and Grant 61521062, in part by the Shanghai computational complexity than the SDR-Alt algorithm.
Key Laboratory Funding under Grant STCSM15DZ2270400, in part by the
CETC Key Laboratory of Data Link Technology Foundation under Grant
CLDL-20162306, and in part by the Medical Engineering Cross Research II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
Foundation of Shanghai Jiao Tong University under Grant YG2017QN47. Consider a downlink multiuser massive MIMO system with
The associate editor coordinating the review of this paper and approving it one multi-antenna base station (BS) and K single-antenna
for publication was P. D. Diamantoulakis. (Corresponding author: Feng Yang.)
G. Zang is with the Department of Electronic Engineering, users, denoted by K {1, . . . , K }. The BS has M (≥ K)
Shanghai Jiao Tong University, Shanghai 200240, China, also with antennas and N RF chains. To reduce hardware cost and power
the Department of New Technology of Communication, Shanghai consumption, we consider hybrid beamforming with a reduced
Microwave Research Institute, Shanghai 200331, China, and also number of RF chains (i.e., N < M). As illustrated in Fig. 1, we
with the CETC Key Laboratory of Data Link Technology, China
Electronics Technology Group Corporation, Xi’an 710068, China (e-mail:
adopt the widely used fully-connected structure, where each
zangguangda@sjtu.edu.cn). RF chain is connected to all M antennas. Thus, the output
Y. Cui, F. Yang, L. Ding, and H. Liu are with the Department of Electronic signal of each antenna can be seen as a linear combination
Engineering, Shanghai Jiao Tong University, Shanghai 200240, China of all RF signals. Let W [w1 , . . . , wK ] ∈ CN ×K denote
(e-mail: cuiying@sjtu.edu.cn; yangfeng@sjtu.edu.cn; lhding@sjtu.edu.cn; the digital beamformer, where wk ∈ CN ×1 denotes the dig-
ital beamforming vector for user k. Let V ∈ CM ×N denote
huiliu@sjtu.edu.cn).
H. V. Cheng is with the Department of Electrical Engineering, Linköping
University, 581 83 Linköping, Sweden (e-mail: hei.cheng@liu.se). the analog beamformer. As in [5] and [6], we do not impose
Digital Object Identifier 10.1109/LWC.2018.2878766 modulus constraints on the analog beamformer. Note that an
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZANG et al.: OPTIMAL HYBRID BEAMFORMING FOR MULTIUSER MASSIVE MIMO SYSTEMS WITH INDIVIDUAL SINR CONSTRAINTS 533
W = UH Sw . The following result shows the relationship Algorithm 2 Solution for the Case of K > N
between Problem POri and Problem PEq . N
1: while trace(X) > i=1 λi (X) do
Theorem 1: If X is a globally optimal solution of 2: construct X(0) with random values and set i := 0
Problem PEq , (V, W) is a globally optimal solution of 3: repeat
Problem POri . Furthermore, if X is a stationary point
4: Obtain P(i+1) by solving the problem in (9);
of Problem PEq , (V, W) is a stationary point of Problem POri .
5: Obtain X(i+1) by solving the problem in (10);
Proof: See the Appendix.
6: i ← i + 1;
7: until convergence criterion is met;
B. Penalty Method 8: μ := 2μ;
Based on Theorem 1, we can solve Problem PEq instead of 9: end while
Problem POri . The rank-N constraint in (6) is non-convex and
non-smooth, and hence is hard to deal with. To address this
challenge, instead of (6), we consider the following constraint is monotonically non-increasing with i, the iterative alternat-
ing procedure for given μ converges to a limit point. As the
trace(X) − N i=1 λi (X) ≤ 0, (7)
constraint sets of the two problems are disjoint, the limit
where λi (·) denotes
the i-th largest eigenvalue of the argument. point is a stationary point of Problem PAlt [10]. A suffi-
As trace(X) ≥ N λ (X) holds for any X 0, (7) implies ciently largeμ (> μ0 ) can be found by increasing μ until
i=1 i
trace(X) = N λ (X), which means that X has at most N trace(X) − N i=1 λi (X)=0.
i=1 i
nonzero eigenvalues, i.e., (6) holds. Then we incorporate (7) The details are summarized in Algorithm 2. By Theorem 2
as a penalty for violation and obtain and by the equivalence between Problem PAlt and
Problem PPen , we know that a stationary point of
PPen : minX Sv XSw 2F + μ(trace(X) − N i=1 λi (X)) Problem POri can be obtained by Algorithm 2. As far as we
know, this is the first work providing a convergent stationary
s.t. (3), (4), (5). point of Problem POri in the case of K > N.
Using similar arguments in [8], we have the following result.
Theorem 2: There exists μ0 ∈ (0, +∞) such that for all V. N UMERICAL R ESULTS
μ > μ0 , trace(X) − N i=1 λi (X) = 0 and (V, W) is a sta- In this section, we provide numerical results to illustrate
tionary point of Problem POri , where X is a stationary point the performance of Algorithm 1 and Algorithm 2. In the sim-
of Problem PPen . ulations, the one-ring channel model is used by setting the
Based on Theorem 2, we first solve Problem PPen for any angular spread as Δ = 15◦ and assuming the azimuth angle
given μ. Let ΦM +K ,N {P ∈ SM +K , 0 P I, trace(P) = of arrival for user √ k as θk = −180◦ + Δ + (k − 1) 360
◦
K .
M +K −N } denote the convex hull of the rank-(M+K−N) We choose ηk = 2 − 1 and σk2 = 1. We consider four
projection matrices. As baselines for comparison. The first baseline is the hybrid beam-
N
former obtained using the SDR-Alt algorithm in [4] for solving
trace(X) − λi (X) = min trace(PT X) (8) Problem POri . The other three baselines are three typical fully-
P∈ΦM +K ,N digital beamformers (N = M), i.e., the optimal solution WD of
i=1
Problem PFD (optimal fully-digital beamformer), fully-digital
holds [9], Problem PPen can be rewritten as beamformer based on zero-forcing (ZF) and fully-digital
beamformer based on maximum-ratio-transmission (MRT),
PAlt : min min Sv XSw 2F + μ trace(PT X) which satisfy the SINR constraints in (2). In evaluating the
X P∈ΦM +K ,N
two proposed algorithms and the SDR-Alt algorithm in [4],
s.t. (3), (4), (5),
we use the same convergence criterion; we generate 30 ran-
which can be solved alternatively. Specifically, let X(i) denote dom channels (same for all schemes), and show the mean
the estimate of X at the i-th iteration. Then, the estimates of and standard deviation (see vertical bar at each point) of
the performance. We compare the normalized average power
P and X at the (i + 1)-th iteration are updated as
consumption which is unit-less.
P(i+1) = arg min trace(PT X(i) ) (9) Fig. 2 illustrates the average power versus the number of
P∈ΦM +K ,N users K. We can observe that, in the case of K≤N, Algorithm 1
achieves the same average power as the optimal fully-digital
X(i+1) = arg min Sv XSw 2F + μ trace((P(i+1) )T X) (10)
X beamformer. In the case of K>N, Algorithm 2 outperforms
s.t. (3), (4), (5). the fully-digital beamformers based on ZF and MRT, and
achieves similar average power compared to the optimal fully-
An optimal solution of the convex problem in (9) is given by digital beamformer. These indicate that hybrid beamforming
P(i+1) =QQH , where Q ∈ C(M +K )×(M +K −N ) is composed can achieve most of beamforming performance with reduced
of the M+K−N eigenvectors corresponding to the smallest hardware cost. In Fig. 2, we do not provide results for the
M+K−N eigenvalues of X(i) [9] and can be obtained by SDR-Alt algorithm, as its computational complexity at N = 36
standard matrix decomposition methods such as singular value and M = 96 is not acceptable. In Fig. 3, we compare the
decomposition. The convex SDP problem in (10) can be solved average power and simulation time (reflecting computational
with complexity O((M +K )4.5 ) using the standard interior- complexity) of the proposed algorithms and the SDR-Alt algo-
point toolboxes such as SeDuMi. Thus, it is clear that the rithm at small N and M. The proposed algorithms achieve
iterative alternating procedure for given μ has much lower the same average power as the SDR-Alt algorithm with much
computational complexity than the SDR-Alt algorithm [4]. lower computational complexity. In addition, the computa-
Since Sv X(i) Sw 2F + μ trace(PT X(i) ) is nonnegative and tional complexity of the proposed algorithms almost does
ZANG et al.: OPTIMAL HYBRID BEAMFORMING FOR MULTIUSER MASSIVE MIMO SYSTEMS WITH INDIVIDUAL SINR CONSTRAINTS 535
k gH Sv UUH d ≥ 0, k ∈ K.
k (13)
H
(gk Sv UUH Sw )H
≤ 1+ηk gH Sv UUH dk , k ∈ K. (14)
σk ηk k
2
On the Error Rate Analysis of Coded OFDM Over Multipath Fading Channels
Jinho Choi
Abstract—In this letter, an (approximate) closed-form back to the transmitter at the initial link setup (i.e., prior to
expression for the probability of codeword error is derived in data transmission) so that URLLC becomes possible with a
coded orthogonal frequency division multiplexing over multipath guaranteed target probability of codeword error. Note that in
fading channels. The derived error probability depends on the
channel realization and can be used to decide the values of key order to use such a scheme, it is expected that the variation
parameters of finite-length codes and the signal-to-noise ratio so of CSI is negligible within the feedback interval.
that a high reliability can be guaranteed in terms of the error The performance of coded OFDM over frequency-selective
probability for a given multipath fading channel. (or multipath) fading channels has been studied in the liter-
Index Terms—Ultra-reliable and low-latency communications ature. For example, a performance analysis of coded OFDM
(URLLC), orthogonal frequency division multiplexing, multipath is studied when hard-decision decoding is employed in [7].
diversity. In [8], based on an information-theoretic approach, an approx-
imate outage probability is studied to be used to estimate the
I. I NTRODUCTION probability of codeword error for coded signals. In [9], the
N ORDER to support mission-critical communications in performance of coded OFDM is studied with interferer with
I 5th generation (5G) systems (for industrial automation
and control, remote control of real-time operations, and so
an expression that takes into account all possible error patterns.
Unfortunately, existing approaches do not provide a reasonable
on), the notion of ultra-reliable and low-latency communica- estimate of the probability of codeword error in terms of key
tions (URLLC) has been considered [1], [2]. For URLLC, parameters for given channel realization when soft-decision
it is expected to use short packets based on finite-length decoding is employed at a receiver.
codes [3], [4]. While a good channel code has to be employed In this letter, we derive an upper-bound on the pairwise
for ultra-reliability, optimal or soft-decision decoding is desir- error probability (PEP) of a pair of codewords with a cer-
able to avoid any performance loss due to hard-decision. tain distance for a given channel realization. Using this
For low-latency, it might be desirable to transmit packets upper-bound, key parameters can be decided to guarantee a
without re-transmissions, where each packet can be transmit- certain target error rate. An estimate of the probability of
ted with a high reliability over fading channels. To this end, codeword error is also obtained as a closed-form expres-
we can consider orthogonal frequency division multiplexing sion of key parameters, which can be used for any block
(OFDM) where a set of subcarriers can be used simultane- codes. The closed-form expression is compared with simula-
ously to transmit a block of bits or a packet as an OFDM tion results that are obtained using the Bose, Chaudhuri, and
symbol. For a high reliability, an OFDM symbol can be a Hocquenghem (BCH) code [10] with a soft-decision decod-
codeword. In [5] and [6], it is shown that coded OFDM can ing technique. From the results, we can see that the derived
have a good coded performance by exploiting the frequency expression can be used to predict the probability of codeword
diversity gain that is due to multipath fading. error (in terms of key parameters) for given channel realiza-
Although coded OFDM can be an attractive approach tion, which is vital to guarantee a high reliability in terms
to URLLC, there are a few difficulties. For example, the of the probability of codeword error for coded OFDM over a
performance of coded OFDM is highly dependent on the multipath fading channel.
channel realization. Thus, in order to guarantee a certain relia- Notation: E[·] denotes the statistical expectation and vari-
bility, a closed-form expression for the probability of codeword ance. CN (a, R) represents the distribution of circularly sym-
error is highly desirable, which allows to decide the values of metric complex Gaussian (CSCG) random vectors with mean
key parameters such as the code rate and transmit power in vector a and covariance matrix R. The Q-function is given by
t2
advance, depending on the realization of a multipath fading Q(x ) = x∞ √1 e − 2 dt.
channel. That is, at a receiver, the channel state information 2π
(CSI) can be estimated, and using the expression for the prob-
ability of codeword error with the estimated CSI, the receiver II. S YSTEM M ODEL
can determine the values of key parameters and feed them Suppose that an OFDM symbol corresponds to a codeword
in coded OFDM. The length of an OFDM symbol is denoted
Manuscript received August 10, 2018; revised October 15, 2018; accepted
October 26, 2018. Date of publication October 30, 2018; date of current by L and the lth element is denoted by sl ∈ S, where S repre-
version April 9, 2019. The associate editor coordinating the review of this sents the signal constellation. Note that L is also the number
paper and approving it for publication was J. Mietzner. of subcarriers in the OFDM system. Then, the received signal
The author is with the School of Information Technology,
Deakin University, Burwood, VIC 3125, Australia (e-mail: at a receiver can be given by
jinho.choi@deakin.edu.au).
Digital Object Identifier 10.1109/LWC.2018.2878820 yl = Hl sl + nl , l = 1, . . . , L, (1)
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
CHOI: ON ERROR RATE ANALYSIS OF CODED OFDM OVER MULTIPATH FADING CHANNELS 537
V. C ONCLUDING R EMARKS
In this letter, we studied coded OFDM over multipath
fading channels and derived an upper-bound on the PEP
for given CSI. Based on the upper-bound, we also derived
an approximate (upper-bound on) probability of codeword
error, from which key parameters of the block code can be
decided to guarantee highly reliable transmissions. Simulation
results showed that the derived expression for the prob-
ability of codeword error can predict the performance of
(short-length) BCH codes with soft-decision decoding for a
wide range of target probability of codeword error (between
10−1 and 10−5 ).
R EFERENCES
[1] C. She, C. Yang, and T. Q. S. Quek, “Radio resource management for
ultra-reliable and low-latency communications,” IEEE Commun. Mag.,
vol. 55, no. 6, pp. 72–78, Jun. 2017.
[2] P. Popovski et al., “Wireless access for ultra-reliable low-latency com-
munication: Principles and building blocks,” IEEE Netw., vol. 32, no. 2,
pp. 16–23, Mar./Apr. 2018.
[3] G. Durisi, T. Koch, and P. Popovski, “Toward massive, ultrareliable, and
low-latency wireless communication with short packets,” Proc. IEEE,
vol. 104, no. 9, pp. 1711–1726, Sep. 2016.
[4] M. Shirvanimoghaddam et al., “Short block-length codes for ultra-
reliable low-latency communications,” CoRR, vol. abs/1802.09166,
Sep. 2018.
[5] H. Sari, G. Karam, and I. Jeanclaude, “Transmission techniques for dig-
ital terrestrial TV broadcasting,” IEEE Commun. Mag., vol. 33, no. 2,
pp. 100–109, Feb. 1995.
[6] J. Choi, Adaptive and Iterative Signal Processing in Communications.
Cambridge, U.K.: Cambridge Univ. Press, 2006.
[7] Y. H. Kim, I. Song, H. G. Kim, T. Chang, and H. M. Kim,
“Performance analysis of a coded OFDM system in time-varying
multipath Rayleigh fading channels,” IEEE Trans. Veh. Technol., vol. 48,
no. 5, pp. 1610–1615, Sep. 1999.
[8] J. Zheng and S. L. Miller, “Performance analysis of coded OFDM
Fig. 3. Average probability of codeword error of BCH codes (with the bound systems over frequency-selective fading channels,” in Proc. IEEE Glob.
from (13)) as a function of SNR when P = 5 and L = 64: (a) (127, 50, 27) Telecommun. Conf. (GLOBECOM), vol. 3, Dec. 2003, pp. 1623–1627.
BCH code; (b) (127, 36, 31) BCH code. [9] C. Snow, L. Lampe, and R. Schober, “Error rate analysis for coded
multicarrier systems over Quasi-static fading channels,” IEEE Trans.
Commun., vol. 55, no. 9, pp. 1736–1746, Sep. 2007.
[10] S. Lin and D. Costello, Error Control Coding: Fundamentals and
i.e., (N , K , dmin ) = (127, 50, 27) and (127, 36, 31), when Applications, 2nd ed. Upper Saddle River, NJ, USA: Prentice-Hall, 2004.
P = 5, L = 64, and N = 127, where the ordered statis- [11] A. Guillén i Fàbregas, A. Martinez, and G. Caire, “Bit-interleaved
tics based decoding approach with order 2 and 3 is used. coded modulation,” in Foundations and Trends on Communications and
Information Theory. vol. 5, Boston, MA, USA: Now, 2008, pp. 1–153.
We consider a number of different channel realizations (105 [12] M. Chiani and D. Dardari, “Improved exponential bounds and approxi-
different channel realizations) to obtain the average (code- mation for the Q-function with application to average error probability
word) error probability. For the bound, we use (13). While the computation,” in Proc. IEEE Glob. Telecommun. Conf. (GLOBECOM),
vol. 2, Nov. 2002, pp. 1399–1402.
decoding performance can be improved with a higher compu- [13] M. Mitzenmacher and E. Upfal, Probability and Computing:
tational complexity (from order 2 to order 3), it is not the ML Randomized Algorithms and Probability Analysis. Cambridge, U.K.:
performance [14]. As a result, the bound obtained from (13) Cambridge Univ. Press, 2005.
[14] M. P. C. Fossorier and S. Lin, “Soft-decision decoding of linear block
can be lower than simulation results at high SNRs. In general, codes based on ordered statistics,” IEEE Trans. Inf. Theory, vol. 41,
from Fig. 3, we can see that (13) can be used to predict the no. 5, pp. 1379–1396, Sep. 1995.
540 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—This letter proposes a novel hybrid dual-polarized on a partially connected array because of its lowest hardware
antenna array which exploits two orthogonally collocated dipoles complexity compared with other architectures.
to capture the full power of a polarized millimeter wave sig- The existing massive arrays typically employ single-
nal. To maximize the received signal-to-noise ratio (SNR), we polarized antennas to receive the electric field of an incident
study the adaptive angle-of-arrival and polarization state estima- wave. Since the microwave channels are rich in scatters, the
tion, and develop a differential beam tracking algorithm and a antennas can normally receive signals with sufficient power
cross-correlation-to-power ratio polarization tracking algorithm
regardless of polarization matching. This, however, is not
for interleaved hybrid dual-polarized arrays. Simulation results
verify the superior performance of the proposed algorithms, and applicable in mmWave channels where the line-of-sight (LOS)
confirm the significant improvement of SNR obtained by using component dominates [2] and thus necessitates the alignment
the proposed array and algorithms. of polarization directions between the antennas and the inci-
dent wave. To address this issue, dual-polarized antennas have
Index Terms—Hybrid dual-polarized antenna array, angle-of- been used to collect the entire signal power for the subse-
arrival estimation, polarization state estimation, and mmWave. quent processing [4]. Crossed dipole antennas are an effective
approach to providing dual polarization with a wide frequency
I. I NTRODUCTION range from RF to mmWave [5]. Therefore, we propose to
employ antenna elements with dual-polarized crossed dipoles
ILLIMETER wave (mmWave) communication is an
M enabling technology for future wireless systems [1].
Ranging from 30-300 GHz, the spectrum of mmWave pro-
for a hybrid antenna array.
The use of dual-polarized antennas in an mmWave hybrid
array is at an early stage where many open problems need
vides an unprecedented resource to support extremely fast data to be addressed. These include the angle-of-arrival (AoA)
transmission. In addition, the vast spectrum is an effective sup- and polarization state estimations which play critical roles
plement to the currently saturated radio frequency (RF) bands in optimizing the signal-to-noise ratio (SNR) at the decoder.
(700 MHz to 2.6 GHz) for wireless communications. As a In the existing literatures, the estimations have been studied
result, a wide range of applications based on mmWave, e.g., in two contexts: the fully digital arrays with dual-polarized
multiple-input multiple-output communications [2] and target antennas [6]–[8] and the hybrid arrays with single-polarized
tracking [3], have been increasingly studied in recent years. antennas [3]. For the former, the AoA and polarization state
The short wavelength of mmWave enables massive antenna are estimated by using ESPRIT and MUSIC-based algo-
arrays in the transceivers to overcome the path loss, and rithms. However, these algorithms require the computations
to provide large-scale spatial multiplexing and highly direc- of the covariance matrix and singular value decomposition
tional beamforming. A massive array, consisting of hundreds which grow cubically with the total number of antennas. As
of antenna elements, can be typically configured in the archi- a result, they can be only applied to the arrays with mod-
tectures of digital array and hybrid array. The digital array est sizes. For the latter, only AoA can be estimated because
has each antenna connect to a dedicated RF chain, allowing of the use of single-polarized antennas. Recent researches
the joint digital signal processing among different antennas. have reported a variety of methods in this context, includ-
This potentially leads to the best performance in terms of data ing cross-correlation [3], [9] and MUSIC-based [10] algo-
rate, but results in prohibitive costs in hardware implemen- rithms. It is shown that the cross-correlation based algorithm
tation and real-time signal processing [1]. Alternatively, the requires only scalar multiplications and additions, and thus
array can employ hybrid architectures to reduce the associ- have much lower computational complexity than its counter-
ated costs. For hybrid architectures, analog beamforming is parts based on ESPRIT and MUSIC. For polarized signals,
performed on top of the signal processing in digital domain. the accuracy of the estimation depends on the received
It is typically implemented by using phase shifters to connect power which may suffer a loss due to the polarization mis-
the RF chains to the antenna elements. Depending on the con- match. This issue can be addressed by using dual-polarized
nections, a hybrid array can be configured to be either fully antennas.
connected [1] or partially connected [2]. In this letter, we focus In this letter, we propose a hybrid dual-polarized adaptive
antenna array to immunize the degradation of SNR caused
Manuscript received September 18, 2018; revised October 22, 2018; by the polarization mismatch. We present the estimations
accepted October 26, 2018. Date of publication November 1, 2018; date of
current version April 9, 2019. This work was supported by the Australian of AoA and polarization state, and demonstrate the associ-
Research Council Discovery Project under Grant DP160101693. The associate ated mean square errors (MSEs). We show that by using
editor coordinating the review of this paper and approving it for publication the hybrid dual-polarized antenna array, improved accuracy
was Y. Gao. (Corresponding author: Hang Li.) of AoA estimation can be achieved compared with the exist-
The authors are with the Global Big Data Technologies Centre, ing work [3]. It is also shown that the polarization state of
University of Technology Sydney, Ultimo, NSW 2007, Australia (e-mail:
hang.li@uts.edu.au; qian.wang@uts.edu.au; xiaojing.huang@uts.edu.au; an arbitrarily polarized wave can be estimated, leading to an
jay.guo@uts.edu.au). enhanced SNR at the decoder after combining with the AoA
Digital Object Identifier 10.1109/LWC.2018.2879010 estimation.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
LI et al.: ADAPTIVE AoA AND POLARIZATION ESTIMATION FOR RECEIVING POLARIZED mmWAVE SIGNALS 541
where (·)∗ and |(·)| represent the conjugate and absolute value Algorithm 1 The AoA and Polarization State Estimation
of (·), respectively. The AoA is contained in both Rx and Ry (0) (0)
Initialization: û (0) , R̂ (0) , Q̂xy , P̂x and P̂y ;
(0)
in the form of u = 2π λ ds sin θ. For i = 1 : I (I is the number of iterations) do
The expectation with respect to (w.r.t.) the statistics of the (i)
1. Set α̂n = −nM û (i−1) , n = 0, . . . , N − 1;
signals shown in (2) can be approximated by the arithmetic (i)
average w.r.t. the time and subarrays. Using adaptive filter- 2. Update [sxm [i], sym [i]] using α̂n in (1);
ing theory, the cross-correlation evaluated based on the first i 3. Calculate û (i) = arg{R̂ (i) } using (3);
received symbols, R̂ (i) , can be expressed as a weighted sum of 4.
(i)
Update ŵm = M
(i)
1 e −jm û , m = 0, . . . , M − 1;
that on the first i−1 symbols, R̂ (i−1) , and an update, ΔR̂ (i) , (i)
depending on the ith symbol only, i.e., 5. Update [sx [i], sy [i]] using ŵm in (4);
(i) (i) (i)
6. Update Q̂xy , P̂x , P̂y using (8);
R̂ (i) = (1 − μ)R̂ (i−1) + μΔR̂ (i) , (3)
7. Determine γ̂ (i) and η̂ (i) using (6) or (7).
where μ denotes the updating coefficient, ranging from zero
End for
to one. The updates for x- and y-axis dipole signals can be
(i) −1 m ∗ m+1 (i)
expressed as ΔR̂ = M m=0 (sx [i ]) sx [i ] and ΔR̂y =
M −1 m ∗ xm+1
m=0 (sy [i ]) sy [i ], respectively [3]. They both consist
of a wanted component with a phase u (see (2)), and a noise where cos θ can be obtained from û, as cos θ = 1 − û 2 /π 2 .
component with random and different phases. Therefore, the As a result, it is referred to as cross-correlation-to-power ratio
(i) (i) polarization tracking (CPRPT).
sum of ΔR̂x and ΔR̂y will have the wanted components In the presence of noise component, the cross-correlation
added in phase, leading to enhanced SNR for the estimation and powers, [Qxy , Px , Py ], must be estimated. Denoting their
of AoA included in the phase. As a result, ΔR̂ (i) can be (i) (i) (i)
(i) (i) estimates after receiving i symbols by [Q̂xy , P̂x , P̂y ], they
formed as ΔR̂x + ΔR̂y which results in the estimated AoA can be expressed as
based on the first i symbols given by û = arg{R̂ (i) }. Note that
compared with a conventional array [3], the computation of the
(i) (i) (i)
cross-correlation involves approximately twice the computing Q̂xy , P̂x , P̂y
loads. However, we will show in simulation results that this
(i−1) (i−1) (i−1)
leads to improved estimation performance and SNRs in return. = (1 − β) Q̂xy , P̂x , P̂y
σ2 σ2
B. Cross-Correlation-to-Power Ratio Polarization Tracking + β sx [i ](sy [i ])∗ , |sx [i ]|2 − n , |sy [i ]|2 − n , (8)
M M
The AoA estimation is followed by the adjustment of digital
and analog beamformers. Using the adjusted weights given by σ2
1 e −jm û , m = 0, . . . , M −1, the outputs of the digital where β (0 < β < 1) is the updating coefficient, and Mn is
ŵm = M the power of noise component in (4). The noise component
beamformer, [sx [i ], sy [i ]], can be expressed as also results in different values for the estimates returned by
M −1
1 jm(u−û) (6) and (7). Their accuracies depend on the SNRs of sx [i ]
[sx [i], sy [i]] = [ex , ey ] · s̃[i]e j 2πfD Ti Ps (θ) · e and sy [i ], respectively, which can be estimated by using the
M (i) (i)
m=0
estimates, P̂x and P̂y obtained from (8). As identical power
signal component (i)
is assumed for noise component in sx [i ] and sy [i ], P̂x and
M −1
1 m (i)
P̂y can be used to indicate the SNRs. Therefore, (6) is used
+ [nx [i], nym [i]]e −jm û , (4)
M (i) (i)
m=0 for the estimation if P̂x ≥ P̂y , and (7), otherwise.
The AoA and polarization state estimation algorithms are
noise component
summarized in Algorithm I. It is noted that the proposed
which indicates that the polarization state information is con- algorithms are blind adaptive since no knowledge about
tained in the complex amplitude of the signal component, the the reference signal is assumed. As shown in (2) and (5),
cross-correlation and powers of which are given by the Doppler frequency shift fD does not affect the cross-
Qxy , Px , Py correlations [Rx , Ry ] in DBT, or the cross-correlation and
2 powers [Qxy , Px , Py ] in CPRPT, since the phase shift pro-
[ex ey∗ , |ex |2 , |ey |2 ] 2 M−1 duced by fD is the same for all subarray signals corresponding
2 jm(u−û)
= 2
σs̃ |Ps (θ)| e . (5)
M to both x- and y-axis dipoles. As a result, both AoA and
m=0 polarization state estimations are Doppler resilient. In addi-
Therefore, when γ = 0 and θ = − π2 ,
the polarization tion, the proposed algorithms can be applied for tracking
state η and γ, can be evaluated by using either one of ∗ the the AoA and polarization state. The selection of updating
Q e
two cross-correlation-to-power ratios: 1) Wx = Pxyx
= ey∗ = coefficients has an impact on both the convergence speed and
x
ej η Q variance of the estimated parameters. In general, a larger coef-
tan γ cos θand 2) Wy = Pxy y
= eexy = tan γ cos θe j η , which ficient will quicken convergence speed but lead to a larger
lead to the estimated polarization state respectively given by variance, whereas a smaller coefficient will decelerate the con-
vergence but reduce the variance. To maximize the received
|Wx |−1
γ = arctan , η = arg{Wx } (6) SNR, we coherently combine two beamformed output sig-
cos θ nals sx [i ] and sy [i ].
After applying the MRC coefficients
|Wy | [κx , κy ] = [ex∗ , ey∗ ]/ |ex |2 + |ey |2 to [sx [i ], sy [i ]], the SNR
and γ = arctan , η = arg{Wy }, (7)
cos θ of the combined signal s[i] equals the sum of two output SNRs.
LI et al.: ADAPTIVE AoA AND POLARIZATION ESTIMATION FOR RECEIVING POLARIZED mmWAVE SIGNALS 543
Fig. 2. Estimated parameters (θ̂, γ̂, η̂) versus the number of iterations. Fig. 4. The combined received SNR versus the number of iterations.
Fig. 3. MSEs versus the average SNR per antenna γs (dB). V. C ONCLUSION
A hybrid dual-polarized adaptive antenna array is proposed
to exploit the polarization diversity to improve the reliability
IV. S IMULATION R ESULTS
of the mmWave system. With the interleaved subarray con-
In this section, we present the simulation results of figuration, the DBT and CPRPT algorithms are developed to
the proposed algorithms. A linear hybrid array with four estimate the AoA and polarization state. Simulation results of
subarrays, each consisting of 16 dual-polarized antennas with MSEs show that the proposed algorithms are effective in terms
ds = λ2 is studied. An elliptically-polarized wave is assumed of both parameter estimation and diversity combining.
with polarization state and AoA given by γ = 3π/10 =
0.9425, η = −π/2 = −1.5708 and θ = 0, respectively. Let γs R EFERENCES
denote the equivalent average SNR per antenna. The received
signals are assumed to be Gaussian distributed because of the [1] R. W. Heath, Jr. et al. “An overview of signal processing techniques for
millimeter wave MIMO systems,” IEEE J. Sel. Topics Signal Process.,
absence of synchronization between the transceivers [3]. vol. 10, no. 3, pp. 436–453, Apr. 2016.
Fig. 2 shows some realizations of the estimations of AoA [2] J. A. Zhang et al. “Massive hybrid antenna array for millimeter-wave
and polarization state using the proposed array and algorithm cellular communications,” IEEE Wireless Commun., vol. 22, no. 1,
with γs = −5 dB. As a comparison, the AoA estimation pp. 79–87, Feb. 2015.
using a conventional array [3] with four subarrays each con- [3] X. Huang et al. “A hybrid adaptive antenna array,” IEEE Trans. Wireless
Commun., vol. 9, no. 5, pp. 1770–1779, May 2010.
sisting of 16 single (y-axis) dipole antennas is also plotted. [4] O. Jo et al. “Exploitation of dual-polarization diversity for 5G
The updating coefficient, μ, is set to be 0.01 in both arrays. millimeter-wave MIMO beamforming systems,” IEEE Trans. Antennas
As shown in the figure, the proposed array outperforms the Propag., vol. 65, no. 12, pp. 6646–6655, Dec. 2017.
conventional one in terms of the speed of convergence. It can [5] S. X. Ta et al. “Crossed dipole antennas: A review,”
be seen that the proposed array can achieve very consistent IEEE Antennas Propag. Mag., vol. 57, no. 5, pp. 107–122,
Oct. 2015.
estimate of AoA after a few iterations, whereas the estimate [6] J. Li and R. T. Compton, “Angle and polarization estimation using
converges slowly for the conventional one. The estimations ESPRIT with a polarization sensitive array,” IEEE Trans. Antennas
of γ and η are also shown in the figure with various updat- Propag., vol. 39, no. 9, pp. 1376–1383, Sep. 1991.
ing coefficients (β = 0.001 and 0.05). After converging, they [7] K. T. Wong and M. D. Zoltowski, “Uni-vector-sensor ESPRIT for mul-
return results with high precision. The estimates {γ̂, η̂} for tisource azimuth, elevation, and polarization estimation,” IEEE Trans.
Antennas Propag., vol. 45, no. 10, pp. 1467–1474, Oct. 1997.
β = 0.001 and 0.05 are given by {0.9231, −1.5860} and [8] L. Wan et al., “DOA and polarization estimation for non-circular signals
{0.9005, −1.6047}, respectively. The MSEs of the estimates in 3-D millimeter wave polarized massive MIMO systems.” Accessed:
are shown as a function of γs in Fig. 3, where the results are Dec. 15, 2017. [Online]. Available: https://arxiv.org/abs/1712.05587
obtained after 200 iterations averaged over 5000 independent [9] K. Wu et al. “Robust unambiguous estimation of angle-of-arrival in
simulations. As shown in the figure, the proposed array and hybrid array with localized analog subarrays,” IEEE Trans. Wireless
Commun., vol. 17, no. 5, pp. 2987–3002, May 2018.
algorithm outperform the conventional array in estimating the [10] F. Shu et al., “Low-complexity and high-resolution DOA estimation for
AoA. We can also see that the MSEs of AoA and polarization hybrid analog and digital massive MIMO receive array,” IEEE Trans.
state reduce with the increase of γs . Commun., vol. 66, no. 6, pp. 2487–2501, Jun. 2018.
544 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In this letter, we investigate the application of trace of sample covariance matrices), respectively. In [3], the
antieigenvalues to spectrum sensing for the first time and design expression for the decision threshold of MME was derived by
an antieigenvalue-based detector. A theoretical expression for the exploiting approximate independence between the maximum
decision threshold of the proposed detector is also derived. It
is shown that the mean-to-square extreme eigenvalue detector
and minimum eigenvalues. In [4], an asymptotic expression
is a special case of the proposed one. Numerical simulations for the decision threshold of MEAM was derived by using
validate the theoretical analysis, and demonstrate the superior an approximation with chi-square distribution. Without the
performance of the proposed detector. rank-one assumption, AGM, EMR and EW detectors, which
Index Terms—Cognitive radio, spectrum sensing, sample exploit all (weighted) eigenvalues, were proposed in [5]–[7].
covariance matrix. However, obtaining all eigenvalues requires high compu-
tational complexity in the eigen-decomposition of sample
covariance matrices especially with high dimensions. Recently,
I. I NTRODUCTION to avert the computational cost of AGM, the mean-to-square
OGNITIVE radio (CR), which allows secondary users extreme eigenvalue (MSEE) detector was proposed by using
C (SUs) to utilize spectrum holes, has been recognized as
a promising technology for alleviating the problem of shortage
only part of the eigenvalues [9], which is also the focus of
this letter.
of spectral resources. One of the most important requirements The above mentioned eigenvalue-based detectors employ
for CR is the capability of spectrum sensing so that SUs can either all eigenvalues or the maximum eigenvalue. Using all
probe spectrum holes and opportunistically use them without eigenvalues with full eigen-decomposition involves high com-
causing harmful interference to licensed primary users (PUs). putational complexity especially when the matrix dimension
A variety of spectrum sensing methods have been proposed is large [10]. While using only the largest eigenvalue induces
in [1]. Among these methods, eigenvalue-based detectors inferior performance as the covariance matrix of primary
have been extensively investigated due to their high detec- signals often has several large eigenvalues [11]. Moreover,
tion performance at low signal-to-noise ratio (SNR). The detectors employing multiple large eigenvalues are vulnerable
eigenvalue-based detection mainly includes maximum eigen- to noise uncertainty as the ME detector.
value (ME), maximum-minimum eigenvalue (MME), maxi- To circumvent these issues, in this letter, we investigate the
mum eigenvalue to arithmetic mean (MEAM), arithmetic to application of antieigenvalues to spectrum sensing, and design
geometric mean (AGM), eigenvalue-moment-ratio (EMR) and an antieigenvalue-based detector which employs multiple large
eigenvalue weighting (EW) detectors [2]–[8]. ME was derived and small eigenvalues with partial eigen-decomposition [10].
based on the rank-one assumption of primary signals, and the It is shown that the MSEE detector is a special case of the
theoretical expression for its decision threshold was analyzed proposed one. Moreover, a theoretical expression for the deci-
in [2]. However, it turns out that ME is sensitive to noise sion threshold of the proposed detector is derived. Numerical
uncertainty because of the dependence of its decision thresh- simulations are provided to validate the theoretical analysis
old on the noise power [2]. MME and MEAM can be regarded and demonstrate the superior performance of the proposed
as modified versions of ME by replacing the noise power detector compared to existing eigenvalue-based detectors. In
in ME with the minimum eigenvalue and the average of all particular, compared with AGM which requires all eigenval-
eigenvalues of sample covariance matrices (equivalently, the ues, the proposed detector delivers a similar or even better
performance but with lower computational complexity.
Manuscript received October 1, 2018; accepted October 30, 2018. Date of The rest of this letter is organized as follows. The signal
publication November 2, 2018; date of current version April 9, 2019. This model and the preliminaries about antieigenvalue are intro-
work was supported in part by the Natural Science Foundation of China under
Grant 61871246 and Grant 61571250, in part by the Zhejiang Natural Science duced in Section II. The new detector and the theoretical
Foundation under Grand LY18F010008, in part by the Ningbo Leading and expression for its decision threshold are derived in Section III.
Top-Notch Talented Training Project under Grand NBLJ201801001, in part by Simulation results are provided in Section IV, followed by
the Australian Research Council’s DECRA under Grant DE120101266, and
in part by the K. C. Wong Magna Fund in Ningbo University. The associate
conclusions drawn in Section V.
editor coordinating the review of this paper and approving it for publication
was K. W. Choi. (Corresponding author: Ming Jin.)
C. Guo, M. Jin, and Y. Li are with the Faculty of Electrical II. P RELIMINARIES AND S IGNAL M ODEL
Engineering and Computer Science, Ningbo University, Ningbo A. Signal Model
315211, China (e-mail: 1971018124@qq.com; jinming@nbu.edu.cn;
liyouming@nbu.edu.cn). Spectrum sensing is usually treated as a binary hypothesis
Q. Guo is with the School of Electrical, Computer and Telecommunications problem, and a decision needs to be made on whether pri-
Engineering, University of Wollongong, Wollongong, NSW 2522, Australia
(e-mail: qguo@uow.edu.au). mary signals are present or not. Let x(n) ∈ C M ×1 denote the
Digital Object Identifier 10.1109/LWC.2018.2879339 received discrete-time complex baseband signal vector from an
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
GUO et al.: ANTIEIGENVALUE-BASED SPECTRUM SENSING FOR CR 545
SU or multiple closely deployed SUs. Each SU is equipped Algorithm 1 Antieigenvalue-Based Spectrum Sensing
with one or multiple antennas. The signal vector x(n) can be Input: Sample covariance matrix R̂x and an estimated rank
obtained from either multiple antennas at one time instant [4] P̂ .
or one antenna by stacking multiple time instants [11]. The Output: Decision on whether priamry signals are present or
received signal x(n) can be expressed as not.
w(n), H0 1: Compute the P̂ largest and the P̂ smallest eigenvalues of
x(n) = (1) R̂x , i.e., λ̂1 , · · · , λ̂P̂ , λ̂M −P̂ +1 , · · · , λ̂M by using partial
s(n) + w(n), H1 , n = 0, 1, . . . , N − 1
eigen-decomposition methods [10];
where H0 and H1 represent the absence and presence of pri- 2: Compute the P̂ smallest antieigenvalues of R̂x as in (6),
mary users, respectively; s(n) denotes the observed primary i.e., ν̂1 , · · · , ν̂P̂ ;
signals, which is assumed to have mean zero and covariance
3: Compute the test-statistic T in (7) with P = P̂ ;
Rs [2]; and w(n) denotes the additive white circularly sym-
4: Set a decision threshold γ according to a false-alarm prob-
metric complex Gaussian noise with mean zero and covariance
2 I , i.e., w(n) ∼ CN (0, σ 2 I ). In addition, we assume ability Pf , and make a decision: if T < γ, primary signal
σw M w M exists; otherwise, primary signal does not exist.
that s(n) and w(n) are independent of each other [4]. It has
already been shown that the covariance matrix of real-world
signals has a small number of large eigenvalues [11]. Here we
assume that Rs has a rank of P which can be much smaller which is a scaled identity matrix under H0 while not under H1 .
than the size of Rs . According to (3), we can find that the antieigenvalues of Rx
are all equal to one under H0 , while the P small antieigenval-
B. Eigenvalue and Antieigenvalue ues of Rx are smaller than one under H1 . This characteristic
can be employed to distinguish whether Rx is a scaled iden-
In this section, we briefly introduce the definition of tity matrix or not, and further to decide whether x(n) contains
antieigenvalues and the relationship between antieigenvalues primary signals.
and eigenvalues. Interested readers can refer to [12] and [13] With N samples, the sample covariance matrix of the
for more details. Given an Hermitian positive definite matrix received signals is given by
A with size M × M, we use θ to denote the angle between a
N −1
non-zero vector x and Ax. There are total M pairs of eigen- 1
value and eigenvector (λk , xk ) satisfying Axk = λk xk (i.e., R̂x = x(n)xH (n). (5)
N
cos θ = 1 for xk and Axk ). In other words, eigenvalues are n=0
obtained by maximizing cos θ. Without loss of generality, we
Let λ̂1 ≥ λ̂2 ≥ · · · ≥ λ̂M be the ordered eigenvalues of R̂x .
assume that λ1 ≥ λ2 ≥ · · · ≥ λM > 0.
The antieigenvalues of R̂x are given by
Differently, antieigenvalues are obtained by minimizing
cos(θ), and are defined as [12], [13].
2 λˆk λ̂M −k +1
(x∗k )H Ax∗k M ν̂k = . (6)
νk = ∗min∗ λ̂k + λ̂M −k +1
∗ || , k ≤ 2 (2)
xk ⊥Sk ||x∗
k || · ||Ax k
It can be seen that the P small antieigenvalues are distin-
where x∗k denotes the antieigenvector corresponding to the kth guishable under H0 and H1 , while it is difficult to distinguish
smallest antieigenvalue, Sk∗ = {x∗1 , x∗2 , . . . , x∗k −1 } with S1∗ = other antieigenvalues under H0 and H1 . Hence, we take the
∅, and x∗k ⊥ Sk∗ denotes that x∗k is orthogonal to the elements summation of the P small antieigenvalues as the test-statistic1
in Sk∗ . The relationship between νk s and λk s is given by [13]
P P 2 λˆ λ̂
k M −k +1
2 λk λM −k +1 T = ν̂k = . (7)
νk = . (3) λ̂ + λ̂M −k +1
λk + λM −k +1 k =1 k =1 k
The use of the antieigenvalue-based detector is detailed in
III. A NTIEIGENVALUE -BASED D ETECTOR AND I TS Algorithm 1. It is noted that the perfect knowledge of rank P
D ECISION T HRESHOLD may not be available in practice, and it can be estimated by
In this section, we make use of antieigenvalues to distin- many classical methods, such as the Gerschgorin radii crite-
guish whether the population covariance matrix is a scaled rion [14]. Let P̂ denote the estimated rank of the covariance
identity one or not, and then propose a new detector based on matrix of primary signals. It is shown in Section IV that the
the antieigenvalues of sample covariance matrix. Moreover, we proposed detector is robust against the inaccurate knowledge
derive a theoretical expression for the false-alarm probability of the rank. We can see that the MSEE detector is a special
and the decision threshold of the proposed detector. case of the proposed one with P̂ = 1. In Algorithm 1, setting
the decision threshold γ is a key to control the false-alarm
A. Antieigenvalue-Based Detector
1 In [7], the optimal eigenvalues weighting (OEW) detection uses all eigen-
The population covariance matrix of x(n) is given by values which are weighted properly. When the covariance matrix of primary
2 signals is not full rank, the detector actually uses the summation of weighted
σ w IM , H0
Rx = 2I , (4) large eigenvalues. In contrast, the proposed detector in this letter employs the
Rs + σw M H 1 summation of P small antieigenvalues as the test-static.
546 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
V. C ONCLUSION
In this letter, we have proposed an antieigenvalue-based
detector, and derived the theoretical expression for the decision
threshold and false-alarm probability of the proposed detector.
Numerical simulations have been provided to verify the theo-
retical analysis and demonstrate the superior performance of
the proposed detector. It has been also shown that the proposed
Fig. 2. Antieigenvalues under H0 and H1 for (a) P = 5 and (b) P = 15 detector is robust to noise uncertainty (as the decision thresh-
when M = 40, N = 1000 and SNR = 0dB. old does not depend on noise power) and the estimation error
of the rank of the covariance matrix of primary signals.
R EFERENCES
[1] A. Ali and W. Hamouda, “Advances on spectrum sensing for cognitive
radio networks: Theory and applications,” IEEE Commun. Survey Tuts.,
vol. 19, no. 2, pp. 1277–1304, 2nd Quart., 2017.
[2] Z. Li, D. Wang, P. Qi, and B. Hao, “Maximum-eigenvalue-based
sensing and power recognition for multiantenna cognitive radio
system,” IEEE Trans. Veh. Technol., vol. 65, no. 10, pp. 8218–8229,
Oct. 2016.
[3] F. Penna, R. Garello, and M. A. Spirito, “Cooperative spectrum sensing
based on the limiting eigenvalue ratio distribution in Wishart matrices,”
IEEE Commun. Lett., vol. 13, no. 7, pp. 507–509, Jul. 2009.
[4] P. Wang, J. Fang, N. Han, and H. Li, “Multiantenna-assisted spectrum
Fig. 3. Pm versus SNR for detectors when M = 40, N = 1000 and P = 5 sensing for cognitive radio,” IEEE Trans. Veh. Technol., vol. 59, no. 4,
at Pf = 0.01. pp. 1791–1800, May 2010.
[5] R. Zhang, T. J. Lim, Y.-C. Liang, and Y. Zeng, “Multi-antenna based
spectrum sensing for cognitive radios: A GLRT approach,” IEEE Trans.
Commun., vol. 58, no. 1, pp. 84–88, Jan. 2010.
[6] L. Huang, J. Fang, K. Liu, H. C. So, and H. Li, “An eigenvalue-
moment-ratio approach to blind spectrum sensing for cognitive radio
under sample-starving environment,” IEEE Trans. Veh. Technol., vol. 64,
no. 8, pp. 3465–3480, Aug. 2015.
[7] C. Liu, H. Li, J. Wang, and M. Jin, “Optimal eigenvalue weighting detec-
tion for multi-antenna cognitive radio networks,” IEEE Trans. Wireless
Commun., vol. 16, no. 4, pp. 2083–2096, Apr. 2017.
[8] L. Wei and O. Tirkkonen, “Spectrum sensing in the presence of multiple
primary users,” IEEE Trans. Commun., vol. 60, no. 5, pp. 1268–1277,
May 2012.
[9] K. Bouallegue, I. Dayoub, M. Gharbi, and K. Hassan, “Blind spectrum
sensing using extreme eigenvalues for cognitive radio networks,” IEEE
Commun. Lett., vol. 22, no. 7, pp. 1386–1389, Jul. 2018.
Fig. 4. Pm versus SNR for detectors when M = 40, N = 1000 and P = 15 [10] W. Murase and M. Lindenbaum, “Partial eigenvalue decomposition of
at Pf = 0.01. large images using spatial temporal adaptive method,” IEEE Trans.
Image Process., vol. 4, no. 5, pp. 620–629, May 1995.
[11] P. Zhang and R. Qiu, “GLRT-based spectrum sensing with blindly
Fig. 3 shows miss-detection probability (Pm ) for various learned feature under rank-1 assumption,” IEEE Trans. Commun.,
detectors (including volume-based detector (VD) and covari- vol. 61, no. 1, pp. 87–96, Jan. 2013.
[12] K. E. Gustafson, “The geometrical meaning of the
ance absolute value (CAV) [18]) when M = 40, N = 1000 and Kantorovich–Wielandt inequalities,” Linear Algebra Appl., vol. 296,
P = 5 at Pf = 0.01. Note that Pm = Prob(T > γ|H1 ). For nos. 1–3, pp. 143–151, 1999.
the proposed detector, the results with different estimates of [13] R. Khattree, “On the calculation of antieigenvalues and antieigenvec-
tors,” J. Interdiscipl. Math., vol. 4, nos. 2-3, pp. 195–199, 2001.
P are given. It can be observed from Fig. 3 that the proposed [14] H.-T. Wu, J.-F. Yang, and F.-K. Chen, “Source number estimators using
detector outperforms other detectors. This is because the transformed Gerschgorin radii,” IEEE Trans. Signal Process., vol. 43,
proposed detector takes advantage of the rank information and no. 6, pp. 1325–1333, Jun. 1995.
[15] F. Ferrari, Antieigenvalues and Sample Coviarance Matrices, Linköping
exploits large eigenvalues effectively, while the other detectors Univ., Linköping, Sweden, Jul. 2016.
employ either only the largest or all eigenvalues. Moreover, we [16] M. A. Girshick, “On the sampling theory of roots of determinantal
can also see that inaccurate knowledge of the rank of primary equations,” Ann. Math. Stat., vol. 10, no. 3, pp. 203–224, Sep. 1939.
signals has small effect on the proposed detector. [17] B. Jóhannesson and N. Giri, “On approximations involving the beta dis-
tribution,” Commun. Stat. Simulat. Comput., vol. 24, no. 2, pp. 489–503,
Fig. 4 shows Pm versus SNR when M = 40, N = 1000 1985.
and P = 15 at Pf = 0.01. It again demonstrates the supe- [18] L. Huang, C. Qian, Y. Xiao, and Q. T. Zhang, “Performance analysis
rior performance of the proposed detector. It is observed that, of volume-based spectrum sensing for cognitive radio,” IEEE Trans.
Wireless Commun., vol. 14, no. 1, pp. 317–330, Jan. 2015.
when P is large, AGM has sightly better performance than [19] V. Tawil. (May 2006). 51 Captured DTV Signal. [Online].
the proposed one at the cost of high complexity in obtaining Available: http://grouper.ieee.org/groups/802/22/Meeting_documents/
full eigenvalues. This is because the distinguishability of the 2006_May/Informal_Documents
548 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
expected for LT codes; however, its high computational cost If binary symbols are transmitted through a BI-AWGN channel
and BPSK modulation, c̃i 2s̃ σ2
i
. LLRs are then passed to the
Manuscript received May 31, 2018; revised September 25, 2018; accepted soft-input/hard-output LT decoder. The result of the decod-
October 28, 2018. Date of publication November 2, 2018; date of current
version April 9, 2019. The associate editor coordinating the review of this ing is an estimation û of the transmitted message. Now a
paper and approving it for publication was W. Zhang. CRC check is performed: if passed, the decoder outputs the
The author is with the Mathematical and Algorithmic Sciences estimation û and the receiver disconnects, otherwise a novel
Laboratory Paris Research Center, Huawei Technologies SASU, 92100
Boulogne-Billancourt, France (e-mail: valerio.bioglio@huawei.com). decoding attempt is performed after the reception of new sym-
Digital Object Identifier 10.1109/LWC.2018.2879359 bols. In the rest of this letter, we will use the notation x, x̃
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
BIOGLIO: MRB DECODING OF LT CODES OVER AWGN CHANNELS 549
ũj Ll j . (8)
and x̂ to represent alphabet symbols, symbol LLRs and hard
l Ej
decisions on symbol LLRs respectively.
Decoding stops if the CRC is met, while decoding fails if
A. LT Encoding the CRC is not checked after imax iterations; a new decoding
attempt is then executed after the reception of new symbols.
The K information symbols u u1 , . . . , uK are encoded
into a stream of symbols c c1 , c2 , . . . , cN , . . . . Every
III. P ROPOSED A LGORITHM
encoded symbol cj is generated independently drawing a
degree dj N0 according to a degree distribution ρ d and
The goal of a most reliable basis (MRB) algorithm is to
selecting di distinct integers Di t1 , . . . , tdi
at random select the K most reliable linearly independent received sym-
with 1 t1
tdi K , such that bols and invert the matrix formed by their encoding equations
to perform the LT decoding. These two procedures, namely the
ci u t1
u td
i
ut . (2) symbols sorting and the matrix inversion, have a large compu-
t Di tational cost. Moreover, matrix inverse has to be recalculated
Robust Soliton Distribution (RSD) [1] is employed for erasure from the scratch in case of decoding failure, after the recep-
channels, with appropriate modifications for AWGN [5]. If we tion of more reliable symbols. In order to reduce the decoding
call encoding equation of ci the vector g i g1i , . . . , gK
i with
complexity of classical MRB for LT codes, we propose to
modify the On-the-Fly Gaussian Elimination (OFG) algorithm,
1 if j Di
gji (3) proposed in [11] for the decoding of LT codes over erasure
0 otherwise
channels, in order to handle soft inputs. In the following, we
then (2) becomes ci g i
u T , and the matrix G formed by show that an appropriate modification transforms OFG into a
vectors g i as rows is the generator matrix of the LT codes. MRB decoder, permitting to triangularize the matrix composed
by the encoding equations of the most reliable information on
B. LT Soft Decoding via BP the fly.
Soft decoding of LT codes can be performed through Belief
Propagation (BP) by exchanging messages, representing LLR A. sOFG Algorithm Description
estimates, between check nodes (CNs) and variable nodes The proposed sOFG algorithm is presented as Algorithm 1
(VNs). In LLR-based sum-product implementation of the BP and works as follows. We suppose the K K decoding matrix
algorithm, variable nodes are initialized with the correspond- G to be partially upper triangular after the reception of i1
ing received symbol LLR. A check to variable message Li j symbols, i.e., either a row G i has the leftmost nonzero element
represents the LLR estimation of the variable node uj made by on the diagonal or it is all-zero. Each row G i is connected
a check node ci , while a variable to check message Li j rep- to a symbol ṽi , representing the LLR of the encoded sym-
resents the LLR estimation of the variable node uj excluding bol having encoding equation G i ; ṽi 0 if G i is empty.
check node ci . This process is depicted in Figure 3. A new encoded symbol with LLR c̃i is then received, along
Messages flowing from check nodes to variable nodes are with the corresponding encoding equation g i . The position s
calculated first. According to (2), uj ci ut1
utα¡1 of the leftmost 1 of g i is located in order to insert g i in the
utα 1
utdj with j tα , hence Li j can be calculated s-th row of G, keeping it upper triangular. If the row is empty,
as the LLR of the xor of di independent symbols as g i is inserted, and its LLR c̃i is stored as ṽs . If the row is
t j t j
full, the following new swap heuristic is used to decide which
encoding equation will be stored: if the received symbol is
Li j L ci ut c̃i Li t . (4)
t Di t Di more reliable, i.e., c̃i ṽs , then the row and the received
550 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
V. C ONCLUSION
In this letter we presented sOFG, a novel decoder for
Fig. 5. BLER for K 100 and various N.
LT codes over noisy channels. Its MRB approach permits
to improve the BLER performance of the code by reducing
the impact of trapping sets on the decoding algorithm. The
incremental nature of sOFG mitigates the effect of decoding
failures. Finally, the decoding complexity is distributed along
the symbols reception, making this algorithm suitable for time-
demanding scenarios, where the decoding latency impacts the
quality of service. The proposed approach shows to outperform
iterative decoders for short LT codes.
R EFERENCES
[1] M. Luby, “LT codes,” in Proc. IEEE Symp. Found. Comput. Sci.,
Vancouver, BC, Canada, Nov. 2002, pp. 271–280.
[2] R. Palanki and J. S. Yedidia, “Rateless codes on noisy channels,” in
Proc. IEEE Int. Symp. Inf. Theory (ISIT), Chicago, IL, USA, Jun. 2004,
p. 37.
[3] H. Jenkač, T. Mayer, T. Stockhammer, and W. Xu,“Soft decoding of
LT-codes for wireless broadcast,” in Proc. IST Mobile Commun. Summit,
Dresden, Germany, Jun. 2005, pp. 1–5.
[4] V. L. Orozco and S. Yousefi, “Trapping sets of fountain codes,” IEEE
Fig. 6. BLER for K 500 and various N.
Commun. Lett., vol. 14, no. 8, pp. 755–757, Aug. 2010.
[5] J. H. Sorensen, T. Koike-Akino, P. Orlik, J. Ostergaard, and P. Popovski,
“Ripple design of LT codes for BIAWGN channels,” IEEE Trans.
impact of decoding failures on the decoding complexity, due Commun., vol. 62, no. 2, pp. 434–441, Feb. 2014.
[6] A. Kharel and L. Cao, “Decoding of short LT codes over BIAWGN chan-
to the reuse of the decoding matrix and hence of the com- nels with Gauss–Jordan elimination-assisted belief propagation method,”
putational effort spent in the matrix diagonalization. This is in Proc. IEEE Wireless Telecommun. Symp. (WTS), New York, NY, USA,
an advantage over BP, for which every decoding failure rep- Apr. 2015, pp. 1–6.
[7] M. P. C. Fossorier, “Iterative reliability-based decoding of low-density
resents a waste of computational power since the decoder has parity check codes,” IEEE J. Sel. Areas Commun, vol. 19, no. 5,
to be rerun from the scratch. pp. 908–917, May 2001.
[8] M. Baldi, N. Maturo, E. Paolini, and F. Chiaraluce, “On the applicability
of the most reliable basis algorithm for LDPC decoding in telecommand
IV. N UMERICAL R ESULTS links,” in Proc. IEEE Int. Conf. Inf. Commun. Syst. (ICICS), Amman,
Jordan, Apr. 2015, pp. 1–6.
In the following we present simulations of the block error [9] J. Van Wonterghem, A. Alloumf, J. J. Boutros, and M. Moeneclaey,
rate (BLER) performance of the proposed sOFG in comparison “Performance comparison of short-length error-correcting codes,” in
to classical BP for the decoding of short LT codes over AWGN Proc. IEEE Symp. Commun. Veh. Technol. (SCVT), Mons, Belgium,
Nov. 2016, pp. 1–6.
channels. In our simulations we used two LT codes of dimen- [10] J. Hagenauer, E. Offer, and L. Papke, “Iterative decoding of binary
sion K 100 and K 500 designed according to the RSD block and convolutional codes,” IEEE Trans. Inf. Theory, vol. 42, no. 2,
with parameters δ 0.001 and c 0.03, while imax 20 for pp. 429–445, Mar. 1996.
[11] V. Bioglio, M. Grangetto, R. Gaeta, and M. Sereno, “On the fly Gaussian
the BP decoder. Codes are assessed at fixed lengths, hence the elimination for LT codes,” IEEE Commun. Lett., vol. 13, no. 12,
addition of a CRC is not necessary to evaluate the decoders pp. 953–955, Dec. 2009.
552 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—This letter provides a spectral and energy effi- algorithm is developed by applying alternating optimization
ciency evaluation framework for massive MIMO heterogeneous (AO) to decouple original problem into the beamforming and
networks with wireless backhaul. This framework is to maximize power allocation subproblem and wireless backhaul bandwidth
the weighted summation of spectral efficiency and energy effi- allocation subproblem, and further solve them by applying
ciency with guaranteeing quality of service for users, interference successive convex approximation (SCA) and Lagrange dual,
mitigation, and sufficient capacity in wireless backhaul. To respectively.
solve this non-convex problem, we develop a novel alternating
optimization algorithm, which is a combination of Lagrange dual
with successive convex approximation. Numerical results verify II. S YSTEM M ODEL
the effectiveness of the framework in improving the energy and
spectral efficiency by comparing with the other schemes.
Consider a downlink HetNet consisting of one macro base
station (MBS) with N antennas and J single-antenna small
Index Terms—Energy efficiency, resource allocation, spectral base station (SBS). For convenience, we denote the set of
efficiency, heterogeneous networks, wireless backhaul. all BSs B0 ≡ B {0}, where B = {1, . . . , J } is the set of
SBSs and the index 0 is for the MBS. The MBS serves M
macro users (MUE) and each SBS serves the same number
I. I NTRODUCTION of K small cell users (SUE) within their coverage. Assume
UE TO the spectrum scarcity and growing attention to that the HetNet operates in the reverse time division duplex
D energy consumption, the joint optimization of spectral
efficiency (SE) and energy efficiency (EE) has become an
(RTDD) mode. Specifically, the uplink/downlink time slots
of the MBS are also the downlink/uplink time slots of the
essential issue in 5G communication networks [1]. Massive SBSs simultaneously [7]. With it, the interference from MBS
MIMO and heterogeneous networks (HetNets) are recognized to SUEs in downlink can be avoided. Meanwhile, the MBS
as two key technologies to improve SE and EE [2]. Moreover, provides in-band wireless backhaul to SBSs. Consequently,
wireless backhaul is preferred for 5G HetNets since its easy define β ∈ [0, 1] as the bandwidth allocation ratio of wire-
deployment and low cost. This letter will investigate the less backhaul, the rest (1 − β) is allocated to wireless access
tradeoff between SE and EE in massive MIMO HetNet with links of BSs.
wireless backhaul. Considering the downlink transmission of MBS, where
The joint design of downlink beamforming and power allo- it serves MUEs and SBSs. Let F = {B , M} =
cation in wireless backhaul HetNets to maximize system {(1, . . . , J ), (J + 1, . . . , J + M )} be the index of receivers
EE was investigated in [3], but the bandwidth allocation of for the MBS, the received signal at kth receiver is
wireless backhaul was not considered. Wang et al. [4] and
Liu et al. [5] jointly optimized user association and backhaul y0,k = hT 0,k w0,k x0,k +
hT0,k w0,k x0,k + n0,k (1)
k ∈F \k
bandwidth allocation to maximize SE, while none of them
involved EE. Although Zhang et al. [6] analyzed EE in the where w0,k ∈ CN ×1 and h0,k ∈ CN ×1 are the beamform-
massive MIMO enabled HetNet with wireless backhaul by ing vector and the channel vector to kth receiver, respectively.
jointly optimizing bandwidth and power allocation, the SE and x0,k is transmission information and n0,k is the white Gaussian
interference mitigation were not considered. noise. Regularized zero-forcing (RZF) precoding is employed
−1
In this letter, we provide an optimization framework to (HHT +αI) h0,k
by the MBS, which is given by u0,k = ,
maximize SE and EE simultaneously in massive MIMO (HHT +αI)−1 h0,k
HetNets with wireless backhaul while guaranteeing minimum where H = [h0,1 , . . . , h0,J +M ] ∈ CN ×(J +M ) is channel
quality of service (QoS) requirement of each user, interference matrix and α > 0 is regularization parameter. u0,k satis-
mitigation, and sufficient capacity wireless backhaul. To solve fies the power constraint k ∈F uT 0,k u0,k vk ≤ P
max , where
this non-convex problem, a new iterative resource allocation vk denotes the transmit power to kth receiver and the corre- √
Manuscript received September 12, 2018; revised October 22, 2018;
sponding beamforming can be computed by w0,k = u0,k vk .
accepted October 28, 2018. Date of publication November 5, 2018; date Thereby the SINR from MBS to kth receiver is r0,k =
vk g0,k 2
of current version April 9, 2019. This work was supported in part by the log(1 + ), where g0,k = hT0,k u0,k and
National Natural Science Foundation of China under Grant 61331009, in part k ∈F \k vk g0,k +N0
by the Open Project of State Key Laboratory of Millimeter Waves under N0 is the power of the noise, and the corresponding achievable
Grant K201935, and in part by the National Science and Technology Major rate is
Project under Grant 2017zx05005001-005. The associate editor coordinating 1−β
M r0,k , k ∈ M
the review of this paper and approving it for publication was A. Ozcelikkale.
(Corresponding author: Bo Huang.) R0,k = (2)
The authors are with the Department of Information and Communication
βr0,k , k ∈B
Engineering, Tongji University, Shanghai 201804, China (e-mail:
hb0533@tongji.edu.cn; tjgah@tongji.edu.cn). Denoting pj as the transmit power of jth SBS and hj ,k as
Digital Object Identifier 10.1109/LWC.2018.2879428 the channel from jth SBS to the kth SUE. The achievable rate
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HUANG AND GUO: SPECTRAL AND ENERGY EFFICIENT RESOURCE ALLOCATION FOR MASSIVE MIMO HetNets WITH WIRELESS BACKHAUL 553
(t)
Similarly, the lower bound L0,k (v) of r0,k can also be Algorithm 1 Joint Beamforming, Power and Bandwidth
obtained and is omitted for brevity. Allocation
Proof: At the Appendix. 1: Select feasible initial value for p(t) , v(t) ;
(t) (t) (t) 2: Set t = 0;
By applying the convex bounds Lj ,k (p), Uj ,k (p), L0,k (v)
3: repeat
into (8), the approximate convex program is given by 4: Compute optimum β ∗ according to (19);
⎛ 5: repeat
(1 − η)(1 − β) L(t)
0,k (v) 6: Obtain l∗ , p∗ and v∗ according to (14) with the increment
max (t)
˜ (p, v, β) =
f ⎝
p,v, R M t = t +1;
l≥0 k ∈M
⎞ 7: until convergence of p(t) , v(t) ;
L(t) j ,k (p) ⎠ η
8: until convergence of objective function (7)
+ − PC (p, v, β)
j ∈B k ∈K
K P TABLE I
S IMULATION PARAMETERS
(14a)
1 − β (t)
s.t. U (p) ≤ βlj , ∀j ∈ J (14b)
K k ∈K j ,k
(5b*), (5c*), (5e), (11) (14c)
B. Wireless Backhaul Bandwidth Allocation C. Joint Beamforming, Power and Bandwidth Allocation
Algorithm
When p and v are fixed, we can obtain the wireless backhaul
bandwidth subproblem as The proposed joint beamforming, power and bandwidth
allocation algorithm is shown in Algorithm 1, which only
(1 − η) requires polynomial complexity: the complexity of solving
max ψSE (p, v, β)
β R problems (14) and (17) are O(2JK + 3JN + 3MN )3 and
s.t. (5b) − (5d), (5f) (16) O(JK + 3J + M ), respectively [10].
Fig. 1. (a) Energy and spectrum efficient tradeoff for different QoS. (b) Comparison of energy and spectrum efficient at different strategies. (c) Energy and
spectrum efficient tradeoff in limited CSI.
Fig. 1(b) compares the performance of the proposed where gj ,k = gj ,k /N0 . According to [12], 2+x 2x log(1+x ) for
Algorithm 1 with the optimal algorithm and existing algo- all x > 0, it follows that
rithm [6]. Note that the optimal algorithm solves the original
⎛ ⎞
problem (7) by exhaustive search. We see that the proposed 2 pj gj ,k
algorithm performs similarly to optimal algorithm, and shows
j ∈B\j
SE-EE performance improvement in comparison with the algo- log⎝1 + pj gj ,k ⎠ > (25)
2 + pj gj ,k
rithm of [6]. The reason is that the lower and upper convex j ∈B\j j ∈B\j
bound of non-convex term we derived are closer to the optimal
solution. Equation (13) can be proved by combination (24) and (25).
In Fig. 1(c), we compare the performance of Algorithm 1 in Similarly, minor modifications to (24) and (25) can derive (12).
limited channel state information
(CSI). The imperfect CSI can
be modeled as [11], h0,k = 1 − τ0,k 2 ĥ
0,k +τ0,k e0,k , ∀k ∈ F
and hj ,k = 1 − τj2,k ĥj ,k + τj ,k ej ,k , ∀j ∈ B , k ∈ K , where R EFERENCES
ĥ0,k and ĥj ,k are the estimated CSI, e0,k and ej ,k are the chan- [1] K. N. R. S. V. Prasad, E. Hossain, and V. K. Bhargava, “Energy
efficiency in massive MIMO-based 5G networks: Opportunities and
nel noise, and τj ,k ∈ [0, 1] indicates the channel estimation challenges,” IEEE Wireless Commun., vol. 24, no. 3, pp. 86–94,
error of CSI. We show the SE-EE tradeoff at different val- Jun. 2017.
min = R min = 0.6. We
ues of error τj ,k = 0, 0.2, 0.3, and R0,k j ,k [2] J. G. Andrews et al., “What will 5G be?” IEEE J. Sel. Areas Commun.,
observe that the performance of both SE and EE are degraded vol. 32, no. 6, pp. 1065–1082, Jun. 2014.
as τj ,k increases. [3] T. M. Nguyen, A. Yadav, W. Ajib, and C. Assi, “Centralized and dis-
tributed energy efficiency designs in wireless backhaul HetNets,” IEEE
Trans. Wireless Commun., vol. 16, no. 7, pp. 4711–4726, Jul. 2017.
VI. C ONCLUSION [4] N. Wang, E. Hossain, and V. K. Bhargava, “Joint downlink cell
In this letter, we proposed a spectral and energy effi- association and bandwidth allocation for wireless backhauling in two-
tier HetNets with large-scale antenna arrays,” IEEE Trans. Wireless
ciency evaluation framework for massive MIMO HetNets with Commun., vol. 15, no. 5, pp. 3251–3268, May 2016.
wireless backhaul while guaranteeing user QoS, interference [5] Y. Liu, L. Lu, G. Y. Li, Q. Cui, and W. Han, “Joint user association and
mitigation and sufficient capacity in wireless backhaul. We spectrum allocation for small cell networks with wireless backhauls,”
developed an efficient algorithm to solve the non-convex IEEE Wireless Commun. Lett., vol. 5, no. 5, pp. 496–499, Oct. 2016.
optimization problem based on Lagrange dual and SCA. [6] H. Zhang, H. Liu, J. Cheng, and V. C. M. Leung, “Downlink energy effi-
Numerical results demonstrated the effectiveness of the frame- ciency of power allocation and wireless backhaul bandwidth allocation
in heterogeneous small cell networks,” IEEE Trans. Commun., vol. 66,
work in improving the energy and spectral efficiency. no. 4, pp. 1705–1716, Apr. 2018.
[7] L. Sanguinetti, A. L. Moustakas, and M. Debbah, “Interference manage-
A PPENDIX ment in 5G reverse TDD HetNets with wireless backhaul: A large system
analysis,” IEEE J. Sel. Areas Commun., vol. 33, no. 6, pp. 1187–1200,
P ROOF OF THE P ROPOSITION 1 Jun. 2015.
By simple algebraic manipulation, the rj ,k is rewritten as [8] O. Amin, E. Bedeer, M. H. Ahmed, and O. A. Dobre, “Energy
⎛ ⎞ ⎛ ⎞ efficiency–spectral efficiency tradeoff: A multiobjective optimization
approach,” IEEE Trans. Veh. Technol., vol. 65, no. 4, pp. 1975–1981,
rj ,k log⎝ N0 + pj gj ,k ⎠ − log⎝ N0 + pj gj ,k ⎠ (23) Apr. 2016.
j ∈B j ∈B\j [9] B. R. Marks and G. P. Wright, “A general inner approximation algo-
rithm for nonconvex mathematical programs,” Oper. Res., vol. 26, no. 4,
Since log(x ) is a concave function with x > 0, we approxi- pp. 681–683, 1978.
mate the right side first term of (23) by its first-order condition [10] A. Ben-Tal and A. Nemirovski, Lectures on Modern Convex
(t) Optimization: Analysis, Algorithms, and Engineering Applications,
around the point pj as vol. 2. Philadelphia, PA, USA: SIAM, 2001.
[11] T. K. Vu, M. Bennis, S. Samarakoon, M. Debbah, and M. Latva-Aho,
⎛ ⎞ ⎛ ⎞ gj ,k pj − pj
(t)
“Joint load balancing and interference mitigation in 5G heteroge-
(t) ⎠ j ∈B
log⎝1 + pj gj ,k⎠ ≤ log⎝1 + pj gj ,k + neous networks,” IEEE Trans. Wireless Commun., vol. 16, no. 9,
(t)
j ∈B j ∈B pj gj ,k + 1 pp. 6032–6046, Sep. 2017.
j ∈B [12] E. R. Love, “64.4 some logarithm inequalities,” Math. Gazette, vol. 64,
(24) no. 427, pp. 55–57, 1980.
556 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—User selection is one of the most important Although the above seminal approaches perform well, they
components for next generation multi-user multiple-input- neglect a crucial phenomenon, that is, user heterogeneity. The
multiple-output wireless local area networks. However, state- origin of user heterogeneity stems from the diversity of devices
of-the-art approaches neglect the heterogeneity of users in the accessing WLANs, e.g., laptops, cellphones, wearable as well
available bandwidth and the number of antennas, which dimin- as smart-home devices. The users possess different number of
ishes their performance considerably. To tackle this challenge, we antennas and support different maximum bandwidths. The user
formulate a novel integer optimization framework to select the
heterogeneity poses new challenges to the user selection in
antennas of heterogeneous users simultaneously. With estimated
signal-to-interference-and-noise ratio of users via channel vec- MIMO systems. For instance, grouping all the antennas of one
tor projection, we propose a low-complexity branch-and-prune multiple-antenna user does not always yield a high throughput.
algorithm to search for the near-optimal combinations of user If two users are selected whose bandwidths are 20MHz and
antennas. Our algorithm is compatible with legacy 802.11ac and 40MHz respectively, the AP will transmit in the 20MHz band-
is implemented on the software defined radio system. Extensive width. Therefore, the user selection strategy must be capable of
experiments show that our algorithm achieves around 95% of handling the user heterogeneity as it is genuinely remarkable
the optimal throughput and outperforms a benchmark scheme and coupled with the mitigation of inter-user interference.
with a 1.18× gain in realistic indoor environments. In this letter, we formulate the user selection as an inte-
Index Terms—MU-MIMO, user selection, heterogeneity, ger programming problem. The objective is to maximize the
branch-and-prune. aggregate throughput in each slot, constrained by the number
of transmit antennas, and the available bandwidth of users.
Finding the optimal set of users incurs a prohibitive compu-
I. I NTRODUCTION tational complexity, thus is infeasible for online processing.
U-MIMO is a key enabling technology to scale up To circumvent this difficulty, we first adopt a channel vector
M the capacity of 802.11ac wireless local area networks
(WLANs) [1]. Equipped with multiple antennas, an access
projection method [7] to estimate the SINR of user antennas.
A branch-and-prune algorithm [8] is further applied to select
point (AP) is capable of transmitting multiple data streams the user antennas incrementally. Specifically, our algorithm
to different users or receive antennas concurrently, achieving maintains multiple candidate user antenna combinations that
a spatial reuse gain up to the number of transmit anten- enlarge the search space for better throughput. The algorithm
nas [2]. This spatial reuse gain depends on the channel is compatible with legacy 802.11ac and advanced techniques
orthogonality among users and such orthogonality cannot be including CSI compression. We implement the proposed algo-
always preserved [3]. Hence, the AP needs to schedule the rithm on the software defined radio platform WARP [9], and
transmission of a group of users wisely in each slot so extensive experiments manifest the effectiveness.
as to reduce the inter-user interference, especially when the
candidate pool is large. II. P ROBLEM F ORMULATION
The design of user selection strategy has gripped much Suppose that the AP is equipped with S antennas and there
attention recently in 802.11ac WLANs. Xie and Zhang [4] are K users contending for transmission, where user k is
proposed an orthogonality probing based user selection equipped with nk antennas and supports the maximum avail-
scheme named OPUS. SIEVE [5] balanced the trade-off able bandwidth Bk . With the binary selection mask xs,k ,i
between performance and complexity with a scalable multiuser indicating whether antenna i at user k is selected and served
selection module. Recently, MUSE [6] was designed to by data stream s or not, we denote the receiving SINR as
perform user selection with limited CSI feedback on com- SINRs,k ,i . The corresponding throughput Us,k ,i is given by
modity Wi-Fi devices. Us,k ,i = B log(1 + SINRs,k ,i ) where B is the transmission
bandwidth. Thus the problem is formulated as follows:
Manuscript received September 7, 2018; revised October 26, 2018; accepted S K nk
October 29, 2018. Date of publication November 6, 2018; date of current max xs,k ,i Us,k ,i (1)
version April 9, 2019. This work was supported in part by the Natural Science x,B s=1 k =1 i=1
Foundation of China under Grant 61772139, in part by the Shanghai–Hong S K nk
s.t x ≤ S, (2)
i=1 s,k ,i
Kong Collaborative Project under Grant 18510760900, and in part by the
s=1 k =1
CERNET Innovation Project under Grant NGII20170209. The associate editor
coordinating the review of this paper and approving it for publication was B = min s,k ,i {xs,k ,i Bk }\{0}, (3)
C. T. Chou. (Corresponding author: Yuedong Xu.) xs,k ,i ∈ {0, 1}. (4)
S. Wang, Y. Xu, and Q. Kong are with the Research Center of Smart
Networks and Systems, School of Information Science and Technology, Eq. (2) means that the S-antenna AP can serve no more
Fudan University, Shanghai 200433, China (e-mail: wangsl16@fudan.edu.cn; than S user antennas concurrently. The users support diverse
ydxu@fudan.edu.cn; qskong@fudan.edu.cn). maximum bandwidths, while the AP can only transmit on
Z. Chen and X. Wang are with the School of Computer Science,
Fudan University, Shanghai 200433, China (e-mail: zhechen13@fudan.edu.cn; a central frequency with one channel bandwidth in each
xinw@fudan.edu.cn). transmission slot. Therefore, once the set of user antennas
Digital Object Identifier 10.1109/LWC.2018.2879668 are selected, the AP transmits using the lowest available
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
WANG et al.: PRACTICAL USER SELECTION WITH HETEROGENEOUS BANDWIDTH AND ANTENNAS FOR MU-MIMO WLANs 557
are retained, i.e., {A1 , A3 }, {A2 , AN } and {AN , A1 }, and modulation/coding, channel estimation and ZFBF. In the MAC
the remaining branches are pruned subsequently. layer, we implement the aforementioned CSI feedback mech-
The branch-and-prune process terminates automatically anism. Our experiments are conducted in a typical office
when the number of selected user antennas equals to the num- environment. The users are placed randomly and sometimes
ber of antennas at the AP. Since only one user antenna is move at a walking speed for evaluating the mobility scenario.
added in each incremental step, the algorithm will terminate
at Step S. We then select the user antenna combination yield- B. Evaluation
ing the maximum sum throughput in the final M choices Ωi 1) Effectiveness of Our Algorithm: To evaluate the effec-
(i = 1, . . . , M ). tiveness of our algorithm, we setup a MU-MIMO WLAN with
Note that a combination with more user antennas does not a four-antenna AP and a random number of heterogeneous
always lead to a higher total throughput. There are two reasons users that possess twelve antennas in total. Based on the SINR
accounting for this phenomenon: 1) the channel of a newly calculated from the received preambles, we obtain the system
added user antenna is highly correlated with the existing user throughput via the SNR-to-rate lookup table [2]. For compari-
antennas; 2) the newly added user antenna that is of lower son, we implement a greedy user selection algorithm (M = 1).
bandwidth may force the existing users to use the lower band- Besides, we implement MUSE [6], a heuristic user grouping
width, causing a downgraded throughput. When a candidate scheme for MU-MIMO with the consideration of bandwidth
user antenna combination outperforms all its branching combi- heterogeneity. Moreover, we calculate the optimal user selec-
nations, we mark it as a special combination and let it compare tion results in SU-MIMO and MU-MIMO modes offline. Our
with the top-M combinations in the final step. The branch- experiments are conducted in one hundred different scenarios,
and-prune loop remains unchanged. As shown in Fig. 2, and each experiment has a runtime of 100 rounds.
ΩM +1 = {AN , A1 } is such an user antenna combination, Fig. 4 plots the CDFs of the sum throughput. The optimal
and is put in the final candidate pool. result in SU-MIMO mode is the worst, verifying the necessity
Fairness Control: To maintain fairness among user antennas of MU-MIMO. The greedy algorithm is also not satisfactory,
in each scheduling round, we divide all the user antennas into because the AP has no knowledge of the inter-user interference
an active set and an inactive set. The user antennas in the active of unselected users when performing user selection incremen-
set are moved to the inactive set after being selected. The user tally. Our user selection algorithm achieves a total median
antennas in the inactive set are restored to the active set when throughput of 336.675 Mbps, outperforming MUSE by 18%.
it is empty. This mechanism is easy to be implemented and can Less than 20% of experiments have a throughput below
effectively ensure the fairness among different user antennas. 300 Mbps in our algorithm, while more than 40% of them have
Complexity: Our user antenna selection algorithm has a in MUSE. The proposed algorithm is also repeated without
polynomial-time computational complexity. In the initializa- considering bandwidth heterogeneity. The resulting throughput
tion step, N branches are created and maintained without is much worse than that with bandwidth knowledge, implying
any pruning. Therefore, N(N−1) candidate combinations are the significance of taking bandwidth heterogeneity into con-
branched at step 1. The selection of the highest M combi- sideration. Furthermore, our algorithm reaches around 95% of
nations yields a complexity order O(MN 2 ). After step 1, throughput of the optimal result in MU-MIMO mode.
only M branches are maintained until the end of the branch- 2) Impact of Search Space: We hereby evaluate the
ing of all the steps. Hence, there exist M(N−1) branches in proposed branch-and-prune algorithm with different scales of
each following step where the selection of the top-M com- search space. Recap that the search space is determined by M,
binations has a complexity order O(M 2 N ). Considering that the number of the candidate user combinations. Fig. 5 illus-
the above procedure is executed for S − 1 times since Step trates the CDF of sum throughput when M increases from one
2, our branch-and-prune algorithm has a complexity order of to six. A larger M yields a better sum throughput at the cost
O(MN 2 + SNM 2 ), which is lower than exhaustive searching. of increased computational complexity. When M changes from
Compatibility: Our algorithm does not require any modi- one to five, one can witness a nearly 20% throughput gain in
fication on vanilla 802.11ac medium access control (MAC) most of the experiments; when it increases from five to six,
protocols. The algorithm is executed based on the CSI feed- the throughput gain becomes negligible. Hence, an optimistic
back, and the data streams are precoded based on the selection message on our algorithm is that a small search space (e.g.,
results. It is compatible with techniques including CSI feed- M is chosen to be five) might be good enough.
back compression and frame aggregation, as long as effective 3) Execution Time: We record the execution time under
CSI feedback is provided. different numbers of user antennas and different search space.
As shown in Fig. 6, the execution time grows almost linearly
IV. I MPLEMENTATION AND E VALUATION with the number of user antennas, but fortunately with gentle
slopes. Meanwhile, the execution time is proportional to the
A. Implementation and Experimental Setup search space. Compared with the greedy algorithm (M = 1),
We implement our user selection algorithm on the soft- the proposed algorithm costs more runtime, but leads to better
ware defined radio platform WARP [9]. The PHY layer throughput performance. Note that this part of experiments
follows the 802.11ac specifications, consisting of OFDM, is conducted on a laptop configured with Intel Core i7-4500U
WANG et al.: PRACTICAL USER SELECTION WITH HETEROGENEOUS BANDWIDTH AND ANTENNAS FOR MU-MIMO WLANs 559
C. Large-Scale Simulation
To evaluate the scalability of our algorithm, we consider a
large scale WLAN that consists of an AP with eight to sixteen
antennas and a certain amount of users with a total number
of fifty receive antennas. Due to the restriction of SDR plat-
form (expensive price and up to four antennas on each board),
we collect realistic wireless transmit and receive traces using
Fig. 7. Fairness and throughput w/ and w/o fairness control. our system at different locations, and then emulate the selec-
tion procedure offline. Fig. 9 shows the mean throughputs
of MUSE, the optimal scheme and our algorithm. Our algo-
rithm outperforms optimal SU-MIMO and MUSE by 78.7%
and 40.9%, respectively. The performance gap between our
algorithm and the optimal MU-MIMO is nearly 10%, slightly
larger than that in Fig. 4. When the AP is equipped with more
antennas, e.g., sixteen, our algorithm still exhibits remarkable
gains and remains close to the optimality. This manifests that
our algorithm possesses an excellent scalability.
Fig. 8. Compatibility.
V. C ONCLUSION
In this letter, we address the user selection problem in MU-
MIMO WLANs with heterogeneous maximum bandwidth and
number of receive antennas. A novel branch-and-prune algo-
rithm is proposed to achieve the low-complexity selection
of user antennas. Our algorithm is compatible with legacy
802.11ac, and is implement on the software defined radio
platform WARP. Experimental results demonstrate that our
Fig. 9. Large-scale simulation.
algorithm outperforms a very recent counterpart by 1.18 times
and reaches around 95% of the optimal results.
1.80GHz CPU using MATLAB. We believe that a binary code
implementation on commercial AP will be much faster. R EFERENCES
4) Effectiveness of Fairness Control: We further test our
[1] IEEE Draft Standard for IT—Telecommunications and Information
user and antenna selection algorithm with fairness consid- Exchange Between Systems—LAN/MAN—Specific Requirements—Part
eration. The preceding experiments are repeated except that 11: Wireless LAN Medium Access Control and Physical Layer
the fairness control scheme is activated. For comparison, we Specifications—AMD 4: Enhancements for Very High Throughput for
implement the proportional fairness scheduling of PF-11ac+ Operation in Bands Below 6GHz, IEEE Standard P802.11ac/D3.0,
in [10]. We adopt Jain’s fairness index to quantify the fair- Jun. 2012.
[2] M. Gast, 802.11ac: A Survival Guide, 1st ed. Beijing, China: O’Reilly
ness of all the receive antennas where the CDFs of average Media Inc., 2013.
throughput and fairness index are shown in Fig. 7. One can [3] T. Yoo et al. “Multi-antenna downlink channels with limited feed-
observe that our fairness control method balances the trans- back and user selection,” IEEE J. Sel. Areas Commun., vol. 25, no. 7,
mission opportunities well among all the receive antennas. pp. 1478–1491, Sep. 2007.
PF-11ac+ achieves better fairness performance because it is [4] X. Xie and X. Zhang, “Scalable user selection for MU-MIMO
networks,” in Proc. IEEE INFOCOM, 2015, pp. 808–816.
designed for long-term fairness. However, enforcing fairness [5] W. Shen et al. “SIEVE: Scalable user grouping for large MU-MIMO
control usually reduces the sum throughput. Both our fairness systems,” in Proc. IEEE INFOCOM, 2015, pp. 1975–1983.
control method and PF-11ac+ degrades the sum throughput [6] S. Sur et al. “Practical MU-MIMO user selection on 802.11ac commod-
slightly, because some users with poor channel conditions or ity networks,” in Proc. ACM MobiCom, 2016, pp. 122–134.
low bandwidth are grouped for fairness guarantee. [7] W. Shen et al., “Rate adaptation for 802.11 multiuser MIMO network,”
in Proc. ACM MobiCom, 2012, pp. 29–40.
5) Compatibility: The proposed algorithm also works [8] J. Porta et al. “A branch-and-prune algorithm for solving systems of
under CSI compression mechanisms that intends to reduce distance constraints,” in Proc. IEEE ICRA, 2003, pp. 342–348.
the feedback overhead. We implement AFC, a representa- [9] WARP Project. Accessed: Oct. 26, 2018. [Online]. Available:
tive CSI compression mechanism proposed in [11] and fix http://warpproject.org
the configurations under three different compression lev- [10] K. Lee and C. Kim, “User scheduling for MU-MIMO transmission with
active CSI feedback,” EURASIP J. Wireless Commun. Netw., vol. 2015,
els for verifiable comparison: conservative CSI compression p. 112, Apr. 2015.
(sharing CSI across 10ms, one subcarrier and quantizing [11] X. Xie et al. “Adaptive feedback compression for MIMO networks,” in
numerical values into 8 bits), median CSI compression Proc. ACM MobiCom, 2013, pp. 477–488.
560 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—We consider a coupling information transmission a number of coupled systems including spatially-coupled
modulation format utilized for communication over a multiple- low-density parity-check codes (SC-LDPCs) [7]–[10], slid-
access channel. A multitude of packets is transmitted by a ing window superposition coding [11] and modulation [12]
number of sources to a common receiver. Each packet is modu-
lated via simple repetition, interleaving, and an application of a proposed for broadcast in 5G cellular networks to reduce co-
signature sequence. The packets are transmitted with time offsets channel interference. This letter focuses on a system operating
that initiate coupling of the transmitted signals into a received over a multiple-access channel (MAC) rather than broad-
signal that lends itself to an efficient window-based iterative esti- cast and on iterative sliding window decoding. We present
mation and interference cancellation decoding. We present the a window decoding algorithm, and determine the optimal
window decoding formulation, a theoretic analysis, and numerical
results for the achievable sum-rate, gap to the channel capacity window sizes using both theoretical analysis and system simu-
and the optimal window size. lation. Finally, we present simulation results for the data rates
achievable with window decoding of the coupled systems and
Index Terms—Multiuser detection, spatial graph coupling,
multiple access. compare them to the decoding in uncoupled regimes and the
channel capacity.
I. I NTRODUCTION
II. S YSTEM M ODEL
ULTIPLE-ACCESS communications has numerous
M practical applications starting from cellular systems to
ad-hoc and sensor type networks. A significant interest is ded-
Consider a communications scenario in which multiple
transmitters communicate packets to a common receiver. In
icated to Internet of Things (IoT) and machine-to-machine order to encode and modulate the kth packet, k = 1, 2, . . .,
(MtoM) communications in a non-orthogonal multiple- access a K-bit long information vector u k is encoded by an outer
(NOMA) framework for 5G. Systems that can sustain a large error-correction code to produce a vector v k which consists
number of simultaneously arriving packets, where each packet of N coded bits (code rate R = K/N). Each bit in v k is
may be encoded via a number of simple operations are desir- repeated M1 times. The resulting vector is then permuted by
able. The receiver, at the same time, may use sophisticated a permutor of size NM1 . Finally each bit is multiplied by a
techniques to extract the information from all arriving packets. signature sequence s k = (sk ,1 , sk ,2 , . . . , sk ,M2 ), sk ,j ∈ {±1},
We focus on coupling information transmission technique j = 1, 2, . . . , M2 producing a BPSK modulated MN-long vec-
in which each packet is encoded via repetition, interleav- tor ṽ k = (ṽk ,1 , ṽk ,2 , . . . , ṽk ,MN ), where M = M1 M2 . The
ing and signature sequence operations. Each packet can be above operations can be represented in a matrix form
represented as a bipartite graph that connects variable nodes ṽ k = Bk v k . (1)
and modulated symbol nodes, the format which is rooted to
sparse code-division multiple-access type techniques [1]–[3]. The matrix Bk = Sk Pk R, where R is the MN × N bit rep-
To get advantage of the spatial-graph coupling threshold sat- etition matrix repeating each bit M times, Pk is a binary
uration phenomenon [4], [5] the packets are transmitted with permutation matrix and Sk is the diagonal signature sequence
time offsets that enable spatial graph coupling of the resulting matrix that multiplies each repeated data bit by a bit of the
packet graphs at the receiver. The receiver performs iterative signature sequence. The kth data stream (packet) is transmit-
interference cancellation and symbol estimation exploiting the ted with the time offset θk . We consider transmission over a
coupling structure of the underlying message graph. As a result real-valued additive white Gaussian noise (AWGN) channel
the technique is capacity-achieving in a number of regimes [6]. and assume that the packets are symbol-synchronous at the
In this letter we study a window decoding approach, the receiver. The received signal r = (r1 , r2 , . . .) is given by
∞
method that can be used in a practical receiver implemen- 1
tation for such coupled system with optimized complexity rτ = √ ṽk ,τ −θk + nτ τ = 1, 2, . . . (2)
L k =1
and decoding delay. Window decoding has been applied to
where nτ are iid AWGN samples with zero mean and variance
Manuscript received September 28, 2018; accepted October 28, 2018. Date σ 2 . Here we assume that ṽk ,j = 0 for any j < 1 or j > MN.
of publication November 6, 2018; date of current version April 9, 2019.
This work was supported by the Natural Sciences and Engineering Research If, in average, L data streams arrive within one packet length
Council of Canada. The associate editor coordinating the review of this we have a system load of α = L/M. Hence, the total transmit
paper and approving it for publication was J. Choi. (Corresponding author: power is normalized to one.
Dmitri Truhachev.)
The authors are with the Department of Electrical and Computer The received signal can be expressed using the matrix
Engineering, Dalhousie University, Halifax, NS B3J 1Z1, Canada (e-mail: notation as
akarami@dal.ca; dmitry@dal.ca).
Digital Object Identifier 10.1109/LWC.2018.2879840 r = Av + n (3)
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
KARAMI AND TRUHACHEV: COUPLING INFORMATION TRANSMISSION WITH WINDOW DECODING 561
where v = (v 1 , v 2 , . . .) and n is the vector of the noise sam- data bits of the jth packet are then estimated via a conditional
ples. The matrix A is a band-diagonal matrix depicted in Fig. 1 expectation estimator of binary symbols (soft bits)
with the matrices of the transmitted packets B1 , B2 , . . . etc. on
(i) (i−1)
the diagonal. v̂ j = tanh RRT − IMN PT j Sj λj . (7)
(i−1)
III. W INDOW D ECODING The above equation demonstrates that the LLRs λj are first
multiplied by the signature sequence, then inverse-permuted
The decoder performs iterative data estimation and
PTj Sj . Finally for each of the M replica’s of each data bits we
interference cancellation operations on a sliding window that
sum up the LLRs, all except one, not to reuse self-information
covers LW received data packets or approximately W packet
in the iterative process, and apply the ‘tanh’ operation to
lengths of the received sequence. The kth decoding window is
compute the conditional expectation estimates of the data bits.
the received sequence segment
The next step is the interference cancellation operation that
rw
k = (rθk , rθk +1 , . . . , rθk +LW −1 +MN −1 ).
computes
(i)
(i−1)
that consists of received packets from k to k + LW − 1 where r j = r̃ j − Sj Pj v̂ j (8)
in matrix form j =j
rw
k = Aw w
k vk + nw
k + ṽ b + ṽ a , that is we subtract the impact of all packets except packet
vw = (v k , v k +1 , . . . , v k +LW −1 ) (4) j for the received sequence segment r̃ j that corresponds to
k
the jth packet. If we subtract the jth estimated packet from
is the data vector of the LW received packets contained in (i)
r j as well we can use the residual to compute the noise and
the current decoding window, and n w k is the correspond- interference power vectors σ 2j ,i .
ing vector of the noise samples. Fig. 1 shows the window At the last, in the Ith data estimation iteration all LLRs are
matrix Aw k , composed of the individual packet matrices Bj , used to compute the estimated data bits
j = k , k + 1, . . . , k + LW − 1. The vectors ṽ b and ṽ a are the
(i−1)
residual from the previous window and the beginning of the v̂ j = sign RT PT j Sj λj (9)
next window respectively. Once I decoding iterations are per-
formed the decoding window is shifted to accommodate the j = k , k + 1, . . . , k + L − 1 that are forwarded to the outer
next L received packets. error-correction decoder. Modulated bits Bj v̂ j for j = k , k +
(I )
The first iteration of the window decoder starts with 1, . . . , k + L − 1 as well as Bj v̂ j , j = k + L, k + L +
r̃ w w 1, . . . , k + WL − 1 are used to compute the initial interference
k = r k − v̂ b (5)
cancellation term v̂ b for the next window (see (5)).
where the impact of the data bits decoded and estimated in Instead of a decoding schedule with the fixed number of
the previous window are subtracted. Let’s define r̃ j to be the iterations I per decoding window we can let the number of
segment of the received sequence r̃ w k corresponding to the iterations vary from window to window and apply a stop-
jth packet. We also denote the same segment of the received ping criteria: the iterative process stops once the estimated
sequence after i − 1 interference cancellation iterations by signal-to-noise and interference ratio (SINR) computed based
(i−1) on σ 2j ,i ’s of packet k + L reaches the SINR for packet k.
r̃ j . Iteration i starts with computation of log-likelihood
ratios (LLRs) of the data bits of the jth packet An alternative decoding algorithm is the approximate mes-
(i−1) (i−1)
sage passing algorithm (AMP) [13] that has been derived for
λj = L−1/2 r̃ j /σ 2j ,i−1 (6) (un-windowed) coupled transmission in [14]. Instead of pass-
ing M and L messages from each channel and variable node
where σ 2j ,i−1 is the vector of noise and interference power
of the system’s graph at each iteration we only pass a sin-
(i−1)
computed for the jth packet based on r j (see below). The gle message from either node and include a correction factor
562 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
VI. C ONCLUSION
In this letter we propose a window-based decoding algo-
rithm or the coupling information transmission multi-user
communication format. The decoding window size and the
number of decoding iterations per window determine the com-
plexity of the decoding and the resulting achievable rate. Both
theoretic and numerical results demonstrate that the window of
size four is sufficient to achieve near-capacity communications
and at the same time allow for a larger number of simultaneous
arriving packets (higher system loads) than traditional multi-
user communication approaches, such as equal-power dense
Fig. 3. The achievable sum-rate for various values of M and L. The size of
the decoding window is W = 4.
CDMA with maximum load α = 1.49 [15] and α = 2.07 for
partition-spreading CDMA [2].
TABLE I
AVERAGE N UMBER OF I TERATIONS , N ORMALIZED C OMPLEXITY VS . W
R EFERENCES
[1] L. Ping, L. Liu, K. Wu, and W. K. Leung, “Interleave division multiple-
access,” IEEE Trans. Wireless Commun., vol. 5, no. 4, pp. 938–947,
Apr. 2006.
[2] D. Truhachev, C. Schlegel, and L. Krzymien, “Low-complexity capacity
V. S IMULATION R ESULTS achieving two-stage demodulation/decoding for random matrix chan-
nels,” in Proc. IEEE Inf. Theory Workshop, Lake Tahoe, CA, USA,
We start with an experiment that looks for the optimal win- Sep. 2007, pp. 584–589.
dow size and compares it to the findings suggested by the [3] P. A. Hoeher and T. Wo, “Superposition modulation: Myths and facts,”
IEEE Commun. Mag., vol. 49, no. 12, pp. 110–116, Dec. 2011.
analysis above. We consider repetition factor M = 25 and the [4] M. Lentmaier, A. Sridharan, D. J. Costello, and K. S. Zigangirov,
average number of L = 65 packets interfering at each time at “Iterative decoding threshold analysis for LDPC convolutional codes,”
the receiver giving systems load of α = 2.6 which is close to IEEE Trans. Inf. Theory, vol. 56, no. 10, pp. 5274–5289, Oct. 2010.
[5] S. Kudekar, T. J. Richardson, and R. L. Urbanke, “Threshold satura-
the limits for 20dB total system SNR (as we will see below). tion via spatial coupling: Why convolutional LDPC ensembles perform
We choose a decoding schedule with the stopping criteria (see so well over the BEC,” IEEE Trans. Inf. Theory, vol. 57, no. 2,
Section III) and estimate the average number of iterations per pp. 803–834, Feb. 2011.
[6] D. Truhachev, “Universal multiple access via spatially coupling data
window Iav and compute the normalized complexity η (see transmission,” in Proc. IEEE Int. Symp. Inf. Theory, Istanbul, Turkey,
Table I). The packet size is chosen to be N = 400 bits. We Jul. 2013, pp. 1884–1888.
notice that W = 3 is insufficiently long and results in high [7] M. Lentmaier, M. M. Prenda, and G. P. Fettweis, “Efficient message
passing scheduling for terminated LDPC convolutional codes,” in Proc.
complexity while window size W = 4 leads to the smallest IEEE Int. Symp. Inf. Theory, St. Petersburg, Russia, Jul./Aug. 2011,
complexity overall. pp. 1826–1830.
Fig. 3 shows the achievable system sum-rate as a function [8] A. R. Iyengar et al., “Windowed decoding of protograph-based LDPC
convolutional codes over erasure channels,” IEEE Trans. Inf. Theory,
of total system SNR (in dB) for the message passing and the vol. 58, no. 4, pp. 2303–2320, Apr. 2012.
AMP decoding algorithms with window size W = 4 and dif- [9] A. R. Iyengar, P. H. Siegel, R. L. Urbanke, and J. K. Wolf, “Windowed
ferent spreading factors. For each SNR point we choose the decoding of spatially coupled codes,” IEEE Trans. Inf. Theory, vol. 59,
no. 4, pp. 2277–2292, Apr. 2013.
highest system load α = L/M such that the system converges [10] N. U. Hassan, A. E. Pusane, M. Lentmaier, G. P. Fettweis, and
to the approximately interference-free case, compute the post D. J. Costello, Jr., “Non-uniform window decoding schedules for spa-
interference cancellation SNR and compute the achievable nor- tially coupled LDPC codes,” IEEE Trans. Commun., vol. 65, no. 2,
pp. 501–510, Feb. 2017.
malized sum-rate as LR/M = αR where the error-correction [11] L. Wang, E. Şaşoğlu, and Y.-H. Kim, “Sliding-window superposition
code rate R is selected equal the capacity on a binary-input coding for interference networks,” in Proc. IEEE Int. Symp. Inf. Theory
AWGN channel with the post interference cancellation SNR. (ISIT), Jun./Jul. 2014, pp. 2749–2753.
The achievable rate is compared to the capacity of the AWGN [12] K. T. Kim et al., “Interference management via sliding-window coded
modulation for 5G cellular networks,” IEEE Commun. Mag., vol. 54,
channel. For M = 250 AMP the achievable sum-rate closely no. 11, pp. 82–89, Nov. 2016.
follows the channel capacity curve with a gap of about 1.6dB. [13] D. L. Donoho, A. Maleki, and A. Montanari, “Message-passing algo-
The M = 50 message passing algorithm performance is simi- rithms for compressed sensing,” in Proc. Nat. Acad. Sci. USA, vol. 106,
no. 45, pp. 18914–18919, Jul. 2009.
lar for the SNR range from 15 to 22 dB. For smaller repetition [14] D. Truhachev and D. McNutt, “Coupling information transmission with
factors slight performance degradation with respect to capacity approximate message-passing,” IEEE Commun. Lett., vol. 20, no. 10,
occurs at high SNRs. This is in agreement with the theo- pp. 1995–1998, Oct. 2016.
[15] T. Tanaka, “A statistical-mechanics approach to large-system analysis of
retic evaluations suggesting that higher repetition factors are CDMA multiuser detectors,” IEEE Trans. Inf. Theory, vol. 48, no. 11,
required for near-capacity performance at high SNRs. pp. 2888–2910, Nov. 2002.
564 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In this letter, we investigate the secrecy performance through theoretical analysis and simulation results. In [13],
of an unmanned aerial vehicle (UAV)-to-UAV system, where a hybrid outage probability was derived to examine the security
UAV acts as the source (S) transmitting information to a legiti- issue of UAV-aided communication systems while the eaves-
mate UAV receiver while a group of UAVs trying to eavesdrop the dropper performing eavesdropping and malicious jamming
information delivery between S and legitimate UAV receiver. The simultaneously. The tight lower-bound of ergodic capacity was
locations of the legitimate UAV receiver and the eavesdropping obtained for UAV aided cellular communications systems [14].
UAVs are randomly distributed in the coverage space of S. We
However, most of the existing works mainly focus on the
first characterize the statistical characteristics of the signal-to-
noise ratio over S to the legitimate UAV receiver links; and then UAV-to-ground systems in 2-dimensional (2D) space which
the closed-form analytical expressions for secrecy outage prob- only has one direction of communication model from the
ability and the average secrecy capacity have also been derived ground to the air. The security of UAV-to-UAV (A2A) systems
accordingly. Finally, Monte-Carlo simulations are carried out to in 3-dimensional (3D) space will be more complicated since
verify our proposed analytical models. the receivers or eavesdroppers work in all directions, leading
to the difficulties in the mathematical derivations during the
Index Terms—Average secrecy capacity, secrecy outage prob-
performance modeling, as shown in Sections II–IV. It is obvi-
ability, stochastic geometry, unmanned aerial vehicles.
ous that the 3D security model has not been well investigated
and understood, still leaving an open issue.
I. I NTRODUCTION Furthermore, most of the researches only consider the fixed
ECENTLY, unmanned aerial vehicles (UAVs) hook more locations and fixed number of UAVs, while UAVs can move
R and more researchers’ interests as they have been
regarded as an effective complement to aerial communica-
freely and flexibly in the 3D space. However, in practical sce-
narios, the positions of UAVs may vary due to the conditions
tions, which can provide robust and reliable communication of airspace or the scheduled tasks to the UAVs. Especially, the
networks [1]. UAVs could not only play a key role in both malicious UAVs, who want to eavesdrop the transmitted infor-
military and civilian area [2], but also apply in wireless com- mation, will always change their positions to cover up their
munications [3]. Mozaffari et al. [4] analyzed the deployment eavesdropping behaviors. Thus, in this letter we will consider a
of and UAV as a flying base station in a given geographical general case with the randomness of the terminals’ positions,
communication area. Reference [5] studied the optimum relay as well as the random number of the eavesdropping UAVs.
UAV placement to maxmize reliability. The applicability of Moreover, Many previous works have considered 2D space
UAV networks was investigated in [6]. and are not suitable for A2A systems, because UAVs may be
However, most of the existing literatures have not consid- distributed anywhere in the airspace.
ered the security of the confidential information transmitted Motivated by the above observations, in this letter we inves-
from the transmitting UAV to a legitimate UAV. Actually, tigate the secrecy performance of an A2A communication
there is a promising method, that is, physical-layer (PHY) system with a legitimate UAV and a group of eavesdropping
security which can prevent information delivery from eaves- UAVs in the line of sight (LoS) 3D space. It is reasonable
dropping [7], [8]. Several techniques have been investigated to assume that channels mainly experience LoS fading in the
to achieve positive secrecy rates for UAV communication open space [15] because UAVs flying above buildings and
systems, such as artificial noise [9], power control [10], and so shadows are more likely to observe the radio path clearance
on. Zhang et al. [11] studied the maximum secrecy rate and from other UAVs in the surrounding areas. Also, in some
the UAV’s optimizing trajectory considering transmit power application scenarios that UAVs fly in stratosphere, there is
over a finite horizon. Reference [12] discussed how spatial no reflected signal and LoS fading plays the main role during
communication security was affected by the communication the transmissions [16].1 We also consider the randomness of
the locations of all UAVs and the number of the eavesdrop-
Manuscript received August 17, 2018; revised October 25, 2018; accepted ping UAVs by using the stochastic geometry theory. The main
November 1, 2018. Date of publication November 6, 2018; date of current ver- contributions of this letter are summarized as: 1) We charac-
sion April 9, 2019. This work was supported in part by the U.K. EPSRC under
Grant EP/P009719/2, and in part by H2020-MSCA-RISE-2015 under Grant terize the probability density function (PDF) and cumulative
690750. The work of H. Lei was supported by the Project of Fundamental distribution function (CDF) of the signal-to-noise-ratio (SNR)
Science and Frontier Technology Research Plan of Chongqing under Grant over the A2A links; 2) The closed-form analytical expressions
cstc2017jcyjAX0204. The associate editor coordinating the review of this for secrecy outage probability (SOP) and the average secrecy
paper and approving it for publication was P. A. Dmochowski. (Corresponding capacity (ASC) have been derived.
author: Gaofeng Pan.)
J. Ye, C. Zhang, and G. Pan are with the Chongqing Key Laboratory Notations: B(x , y) and f (x, y) denote Beta functions and
of Nonlinear Circuits and Intelligent Information Processing, Southwest joint PDF related x and y. 3 F2 (·) and 2 F1 (·) denote Gauss
University, Chongqing 400715, China (e-mail: gfpan@swu.edu.cn).
H. Lei is with the Chongqing Key Laboratory of Mobile Communications
Technology, Chongqing University of Posts and Telecommunications, 1 In Section V, we present some simulation results under case with both LoS
Chongqing 400065, China. and Non-line-of-sight (NLoS, e.g., Nakagami-m fading) fading to show the
Z. Ding is with the School of Electrical and Electronic Engineering, impact of NLoS fading on the secrecy performance of the considered system.
University of Manchester, Manchester M13 9PL, U.K. We leave the analysis of this case for future works due to its complexity and
Digital Object Identifier 10.1109/LWC.2018.2879842 the page limitation.
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
YE et al.: SECURE UAV-TO-UAV SYSTEMS WITH SPATIALLY RANDOM UAVs 565
x π 2π
3
Fdi (x ) = sin φi di2 dθi dφi d(di )
Fig. 1. System model. 4πD 3
0 0 0
⎧
⎨ 0, x <D
hypergeometric functions. ln(·) denote natural logarithms. = Dx3 , 0 ≤ x ≤ D . (2)
(N −1)! ⎩ 3
CNj −1 = j !(N −1−j ) with N > 1. 1, x >D
Therefore, we can achieve the CDF of γi as
II. S YSTEM M ODEL
In this letter, we consider a A2A communication system, γP
Fγi (x ) = Pr{γi ≤ x } = 1 − Pr di ≤ . (3)
as shown in Fig. 1, where a UAV with a single omni trans- x
mitting antenna acts as the source (S) trying to transmit its
information to a legitimate UAV.2 Furthermore, there are also Then, the proof is completed.
a group of UAVs distributed in the coverage space (in which Corollary 1: Accordingly, the PDF of γi can be expressed
the received signal strength is equal or above the threshold at as
the receiver to demodulate and decode the received signal) of ⎧ √ 3
⎨ 3( γP )
S trying to eavesdrop the information transmission between S 5 , x ≥ γP
D2 .
and the legitimate UAV. For tractability purpose, in this letter fγi (x ) = 2x 2 D 3 (4)
⎩ γP
we treat the coverage space of S as a sphere, V, with radius, 0, x < D2
D (D > 0) m, where S is located at the center of the sphere.
Without of loss generality, we assume that the legitimate UAV Proof: The PDF of γi is the derivative of (1).
and N (N ≥ 1) eavesdropping UAVs can be modeled as a Lemma 2: The CDF of γmax with N eavesdropping UAVs
set of independent and identical uniformly distributed points can be derived as
⎧
without cooperation with each other in the sphere V, denoted ⎪
⎨
√
( γP )
3 N
by W, in order to protect their eavesdropping activities. The 1 − , x ≥ γP
FγNmax (x ) = 3
x 2 D3 D2 . (5)
number of the eavesdropping receivers is Poisson distributed ⎪
⎩ 0, γP
with density λ, i.e., P {N = k } = (μkV /k !) exp(−μV ), where x < D2
3
μV = 4πD 3 λ is the mean measure. Proof: Using probability theory, we have Fz (z ) =
In the following, we name the legitimate receiver as the Pr{max{x1 , x2 , . . . , xM } ≤ z }, if x1 , x2 , . . . ,
xM are M
0th receiver to facilitate the following analysis. Therefore,
(M > 1) independent variables. Then, Fz (z ) = M i=1 Fxi (z ),
the distance between S and UAVs can be calculated from
where Fxi (x ) is the CDF of xi . As the eavesdropping chan-
W, the PDF of which can be given by using [17, eq. (1)] as
3 . In this letter, we assume that the communi- nels are independent with each other, we can derive the CDF
fW (w ) = 4πD 3 of γmax as (5).
cation channels from S to UAVs are dominated by LoS links
rather than other channel impairments in the open airspace, Corollary 2: Accordingly, the PDF of γmax with N eaves-
such as shadowing or small-scale fading, as treated in [3]. dropping UAVs can be expressed as
Thus, the channel power gain from S to the ith (0 < i ≤ N) ⎧
N −1 √
⎪ √
⎨ ( γP )3 3( γP )
3
γP
UAV follows the free-space path loss model, which can be N 1− 3 , x ≥
fγNmax (x ) = 5 D2
given by gi = βdi−2 , where β denotes the channel power at ⎪
⎩ 0,
x 2 D3 2x 2 D 3
γP
x <
the reference distance d = 1 m, whose value depends on the ⎧ D2
√ 3 N −1 √
carrier frequency, antenna gain, etc., and di is the link distance ⎨ N 3( γP )
(−1)j C j
3j
⎪ ( γP ) γP
2D 3 N −1 3 j+ 5 , x ≥ D2
between S and the ith UAV. = j =0 x 2 2 D 3j .
⎪
⎩ γP
Let P denote the transmit power at S. The received SNR 0, x < D2
at the ith UAV from S can be expressed as γi = Pδg2i = γP d2
, (6)
i
where δ 2 denotes the noise power, and γ = δβ2 represents the
Proof: The PDF of γmax is the derivative of (5).
2 In this letter, we only consider omni transmitting antenna to introduce the
analysis method for A2A systems and the system with directional antenna is a III. T HE S ECRECY O UTAGE P ROBABILITY
special case of the one with omni transmitting antenna: the coverage space of
directional antenna is a part of the one of omni antenna, leading to a portion In this letter, SOP is defined as the probability that the
of the sphere for eavesdroppers to distribute, as indicated in Fig. 1. instantaneous secrecy capacity is below a threshold secrecy
566 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
N = ∞ ln(1 + x )f (x ) x
rate, Cth (Cth ≥ 0). Therefore, the instantaneous secrecy where C̄s1 υmin γ0 υmin fγmax (y)dydx and
N = ∞ ln(1 + y)f
∞
capacity from S to the legitimate UAV is C̄s2 υmin γmax (y) y γ0 (x )dx dy.
f
N
Thus, we can calculate C̄s1 as
CS (γ0 , γmax ) = max{log2 (1 + γ0 ) − log2 (1 + γmax ), 0}, (7)
∞
where γmax = maxi∈{1,...,N } {γi }. N
C̄s1 = ln(1 + x )f0 (x )
Then, let ν = 2Cth . SOP can be written as vmin
√ 3 N −1 √ 3j
1 + γ0 Cth
x
3 γP γP
Pr (Cth ) = Pr{Cs ≤ Cth } = Pr ≤2 × N (−1) j j
CN dydx
1 + γmax 2D 3 −1 3 5
y 2 j + 2 D 3j
vmin j =0
= Pr{γ0 ≤ νγmax + ν − 1}. (8) √ 3j +6
N
−1 ∞
j 9N (−1)j γP 5
Theorem 1: Using the Corollary 1 and Corollary 2, we can = CN −1 3 j+ 3 x − 2 ln(1 + x )dx
j =0 2(3j + 3)vmin 2 2 D 3j +6
express Pr (Cth ) as follows: vmin
⎡ N
−1 j √ 3j +6 ∞
√ 3j +6 j 9N (−1) γP 3
∞ N −1 − CN x − 2 j −4 ln(1 + x )dx .
μV −μV ⎢ j j 3N γP −1
2(3j + 3)D 3j +6
Pr (Cth ) = e ⎣ (−1) CN −1 3 j =0 vmin
N =1
N ! j =0 (3j + 3)vmin 2 j +3 D 3j +6
(13)
N −1 √ 3j +6
j 3N γP
− (−1)j CN In orderto facilitate the analysis, we define a new function
g1 (m) = v∞
−1 3
j =0 2ν 2 D 3j +6 x m ln(1 + x )dx . Using [19, eq. (2.6.10.47)],
3 ⎤ min
B j + 3, 1 2 F1
3
+ 52 , 1; 32 j + 32 ;
j 3 we can easily calculate g1 (m) as
2 2 2vmin ⎥
× ⎦, (9)
3 5
vmin 2 j + 2 1 − ν1 + vmin g1 (m) = vmin m+1 B (1, −m − 1)[ln(vmin ) + ψ(−m)
− ψ(−m − 1)] + vmin m B (1, −m)
Proof: As the main and eavesdropping channels are inde-
1
pendent with each other, we can calculate Pr (Cth ) with N × 3 F2 −m, 1, 1; 2, −m + 1; − . (14)
vmin
eavesdropping UAVs as
N as
Then, using (14) in (13), we can easily derive C̄s1
∞
νy+ν−1
= Cs (γ0 , γmax )f (γ0 , γmax )dγ0 dγmax . Finally, ASC can be obtained by substituting (15) and (16)
υmin υmin into (12).
(11)
V. N UMERICAL R ESULTS
As the main and eavesdropping channels are independent In this section, Monte Carlo simulations are carried out to
with each other, we can rewrite ASC as validate our proposed analytical expressions for SOP and ASC
∞
under dominated LoS fading channels and Nakagami-m fad-
μV N N
C̄s (γ0 , γmax ) = exp(−μV )(C̄s1 − C̄s2 )/ ln 2, (12) ing. The main adopted parameters are set as γ = 80 dB [11],
N!
N =1 P = 10 dBW, Cth = 1 bits/s/Hz and the expectation of chan-
nel power gain of the Nakagami-m fading channel is 1 dB.
3 The ergodic secrecy capacity aims at a case that the source can gain the
Moreover, the coverage distance of the source UAV is set
channel state information (CSI) of the eavesdropping link and transmits the from hundreds of meters to tens of km to reflect the practical
information only when the legitimate channel outperforms the eavesdropping
channel. Therefore, the ergodic secrecy capacity is a special case of ASC, as scenarios of UAVs in civil and military applications.
ASC is calculated for the case that no matter the CSI of the eavesdropping Fig. 2(a) shows the SOP versus λ for various Cth with
channel is available or not at the source. D = 500 m. Since a high threshold secrecy rate means the
YE et al.: SECURE UAV-TO-UAV SYSTEMS WITH SPATIALLY RANDOM UAVs 567
VI. C ONCLUSION
In this letter, we have studied the secrecy performance of
a A2A communication system and derived the closed-form
analytical expressions for SOP and ASC. We consider the ran-
Fig. 2. SOP. domness of the number and the positions of all UAVs to make
our system more practical. From the numerical results, we can
obtain that the density of the eavesdropping UAVs and the
radius of the coverage space of S exhibit a negative impact on
SOP. However, the ASC can be improved by increasing the
radius of the coverage space of the source UAV.
R EFERENCES
[1] S. Hayat, E. Yanmaz, and R. Muzaffar, “Survey on unmanned aerial
vehicle networks for civil applications: A communications viewpoint,”
IEEE Commun. Surveys Tuts., vol. 18, no. 4, pp. 2624–2661, 4th Quart.,
2016.
[2] M. Erdelj, E. Natalizio, K. R. Chowdhury, and I. F. Akyildiz, “Help from
the sky: Leveraging UAVs for disaster management,” IEEE Pervasive
Comput., vol. 16, no. 1, pp. 24–32, Jan./Mar. 2017.
Fig. 3. ASC. [3] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications
with unmanned aerial vehicles: Opportunities and challenges,” IEEE
Commun. Mag., vol. 54, no. 5, pp. 36–42, May 2016.
[4] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Unmanned aerial
vehicle with underlaid device-to-device communications: Performance
system requires a large channel secrecy rate, we can see that and tradeoffs,” IEEE Trans. Wireless Commun., vol. 15, no. 6,
pp. 3949–3963, Jun. 2016.
the systems with a small Cth outperforms the one with a large [5] Y. Chen, W. Feng, and G. Zheng, “Optimum placement of UAV as
Cth . Fig. 2(b) shows the SOP versus D for various λ. Clearly, relays,” IEEE Commun. Lett., vol. 22, no. 2, pp. 248–251, Feb. 2018.
there are more eavesdropping UAVs when λ and/or D increase, [6] T. Andre et al., “Application-driven design of aerial communication
resulting in degraded SOP. Because the diversity gain of the networks,” IEEE Commun. Mag., vol. 52, no. 5, pp. 129–137, May 2014.
information eavesdropping will increase, while the number of [7] G. Pan, J. Ye, and Z. Ding, “On secure VLC systems with spatially
random terminals,” IEEE Commun. Lett., vol. 21, no. 3, pp. 492–495,
eavesdropping UAVs enlarging. Mar. 2017.
Fig. 3(a) presents the ASC versus D for various λ, in which [8] G. Pan, J. Ye, and Z. Ding, “Secure hybrid VLC-RF systems with
one can see that ASC can be improved in the low λ region and light energy harvesting,” IEEE Trans. Commun., vol. 65, no. 10,
degrades in large λ region when λ decreases or D increases. pp. 4348–4359, Oct. 2017.
Because low λ and large D can improve the probability that the [9] S. Goel and R. Negi, “Guaranteeing secrecy using artificial noise,” IEEE
Trans. Wireless Commun., vol. 7, no. 6, pp. 2180–2189, Jun. 2008.
distance between S and the legitimate UAV is shorter than the [10] P. K. Gopala, L. Lai, and H. E. Gamal, “On the secrecy capacity of fad-
ones among S and the eavesdropping UAVs. Then, the legiti- ing channels,” IEEE Trans. Inf. Theory, vol. 54, no. 10, pp. 4687–4698,
mate UAV will suffer weaker eavesdropping, compared to the Oct. 2008.
case with high λ and small D. However, it is exactly oppo- [11] G. Zhang, Q. Wu, M. Cui, and R. Zhang, “Securing UAV communi-
site for ASC in the high λ and/or low D regions, since there cations via trajectory optimization,” in Proc. GLOBECOM, Singapore,
2017, pp. 1–6.
will be more eavesdropping UAVs, leading to poorer secrecy [12] S.-W. Kim and S.-W. Seo, “Cooperative unmanned autonomous vehicle
performance. As shown in Fig. 3(b) with λ = −130 dB, control for spatially secure group communications,” IEEE J. Sel. Areas
we can also see that the reference SNR γ cannot play a Commun., vol. 30, no. 5, pp. 870–882, Jun. 2012.
positive role significantly, since the received SNR over the [13] C. Liu, T. Q. S. Quek, and J. Lee, “Secure UAV communication in
eavesdropping links will also be improved when γ increases. the presence of active eavesdropper (invited paper),” in Proc. WCSP,
Nanjing, China, 2017, pp. 1–6.
Moreover, while suffering both LoS and Nakagami-m fad- [14] S. Hu, J. Flordelis, F. Rusek, and O. Edfors, “Unmanned aerial vehicle
ing, we also present the SOP performance with m = 2 and √ assisted cellular communication,” arXiv:1803.05763, 2018.
m = 3 in Fig.
√ 2(a) and Fig. 2(b), and the ASC with m = 6 [15] K. Welch, “Evolving cellular technologies for safer drone operation,”
and m = 2 in Fig. 3(a) and Fig. 3(b), separately. One can San Diego, CA, USA, Qualcomm 5G, White Paper, Oct. 2016.
see that the SOP and ASC suffering both LoS propagation [16] Iskandar, “Wireless channel characteristic and its performance for strato-
spheric platform communication,” Ph.D. dissertation, Waseda Univ.,
and Nakagami-m fading outperforms the ones only experienc- Tokyo, Japan, 2007.
ing LoS propagation. This observation can be explained by [17] G. Pan, H. Lei, Z. Ding, and Q. Ni, “On 3-D hybrid VLC-RF systems
the fact: Benefiting from the diversity gain of multiple eaves- with light energy harvesting and OMA scheme over RF links,” in Proc.
dropping UAVs, the equivalent received SNR at eavesdroppers, GLOBECOM, Singapore, 2017, pp. 1–6.
γmax , will degrade slower compared with the received SNR [18] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series and
Products, 7th ed. San Diego, CA, USA: Academic, 2007.
at the legitimated UAV, γ0 , when the channels get worse. [19] A. P. Prudnikov, Y. A. Brychkov, and O. I. Marichev, Integrals and
Therefore, we can see that the SOP only suffering LoS fad- Series: Elementary Functions, vol. 1, 3rd ed. New York, NY, USA:
ing acts as the upper bound of the SOP, and the ASC only Gordon Breach Sci. Publ., 1992.
568 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Visible light communication (VLC) systems utilizing multiplexing gain achievement via multiple-input multiple-
conventional superposition coding/successive interference cancel- output (MIMO) transmission [8]–[10], etc.
lation (SPC/SIC)-based non-orthogonal multiple access (NOMA) As a promising candidate for 5G systems, power-domain
suffer from the error propagation effect due to imperfect SIC. In
this letter, we propose a flexible-rate SIC-free NOMA technique non-orthogonal multiple access (NOMA) has been consid-
for downlink VLC systems, based on constellation partitioning ered for capacity improvement in downlink VLC systems.
coding (CPC) and uneven constellation demapping (UCD). By Marshoud et al. [11] applied NOMA in VLC systems and
using CPC/UCD, SIC is not required and hence error propaga- proposed a gain ratio power allocation strategy. Advanced
tion can be eliminated in NOMA-based VLC systems. Moreover, power allocation strategies were proposed for NOMA-based
by selecting a proper bit allocation scheme, flexible-rate multiple
access can be supported in the VLC system applying CPC/UCD- VLC systems by considering user fairness and the quality
based NOMA. Proof-of-concept two-user VLC experiments verify of service (QoS) constraint [12], [13]. An in-depth evalua-
that, compared with conventional SPC/SIC-based NOMA, the bit tion of NOMA in VLC systems was further performed by
error rate performance of the near user can be greatly improved Yin et al. [14]. In [15], NOMA was applied in MIMO VLC
by using CPC/UCD-based NOMA, and hence the effective power systems. In these works, superposition coding and successive
allocation ratio range can be substantially extended.
interference cancellation (SPC/SIC) are adopted and perfect
Index Terms—Visible light communication, non-orthogonal SIC is generally assumed. However, perfect SIC cannot always
multiple access, constellation partitioning coding. be guaranteed in practical NOMA-based VLC systems. It has
been shown that imperfect SIC might cause error propagation
and hence degrade the bit error rate (BER) performance [16].
I. I NTRODUCTION To address this issue, Li et al. [17] proposed symmetric
SPC with symmetric SIC for error propagation mitigation.
N RECENT years, visible light communication (VLC)
I using white light-emitting diodes (LEDs) has gained
tremendous attention, due to the dual use of white LEDs
Nevertheless, the error propagation effect cannot be com-
pletely eliminated since SIC is still required and only fixed-rate
multiple access can be supported.
for simultaneous illumination and communication in indoor
Moreover, NOMA has also been considered in uplink VLC
environments [1]. As a complementary technology to tra-
systems. In [18], phase pre-distortion was applied to improve
ditional radio-frequency technologies such as Wi-Fi, VLC,
the BER performance of uplink NOMA-based VLC systems.
also known as Li-Fi [2], has many inherent advantages such
In [19], a joint detection scheme was presented, which is SIC-
as huge and unregulated spectrum, low-cost front-ends and
free and maximum likelihood optimal. Nevertheless, according
no electromagnetic interference emission [3]. Nevertheless,
to [20], joint detection requires bit-level joint maximum like-
the achievable capacity of VLC systems is far beyond the
lihood calculations and hence has relatively high complexity.
expectations, due to the limited modulation bandwidth of
In this letter, for the first time, we propose and demonstrate
off-the-self white LEDs [4]. To overcome the bandwidth
a novel NOMA technique based on constellation partitioning
limitation, many capacity-enhancing techniques have been
coding (CPC) and uneven constellation demapping (UCD) for
proposed so far, such as bandwidth extension based on
downlink VLC systems. By using CPC/UCD, user decoding
pre- or post-equalization in the frequency domain [5], [6],
can be realized without SIC and hence the error propagation
spectral efficiency improvement using orthogonal frequency
effect due to imperfect SIC can be eliminated, resulting in
division multiplexing (OFDM) with high-order quadrature
improved BER performance. Moreover, by selecting a proper
amplitude modulation (QAM) constellations [7], diversity or
bit allocation scheme, flexible-rate multiple access can also be
Manuscript received September 9, 2018; revised October 19, 2018; accepted achieved. The feasibility of applying the proposed flexible-rate
November 3, 2018. Date of publication November 8, 2018; date of current ver- SIC-free NOMA technique in practical VLC systems has been
sion April 9, 2019. This work was supported in part by the Delta Electronics
Inc., and in part by the National Research Foundation, Singapore, through the successfully verified through two-user VLC experiments.
Corp Lab@University Scheme. The associate editor coordinating the review
of this paper and approving it for publication was B. Makki. (Corresponding
author: Chen Chen.) II. T WO -U SER VLC U SING CPC/UCD-BASED NOMA
C. Chen, W.-D. Zhong, H. Yang, and P. Du are with the School of Electrical
and Electronic Engineering, Nanyang Technological University, Singapore In this section, we introduce a downlink VLC system with
639798 (e-mail: chen0884@e.ntu.edu.sg). two users using the proposed flexible-rate SIC-free NOMA
Y. Yang is with the College of Computer Science, Sichuan University,
Chengdu 610065, China (e-mail: yangyanbing@scu.edu.cn). technique based on CPC and UCD. Fig. 1 illustrates the block
Digital Object Identifier 10.1109/LWC.2018.2879924 diagram of the system. As we can see, the input bits of both
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
CHEN et al.: FLEXIBLE-RATE SIC-FREE NOMA FOR DOWNLINK VLC BASED ON CPC 569
the near and far users are fed into the CPC block to generate
the CPC-coded constellation and the detailed principle of CPC Fig. 3. Flexible-rate partitioning of Gray-coded 8-QAM constellation by
is described in Section II-A. Subsequently, inverse fast Fourier CPC for (a) bn = b2 b3 , bf = b1 and (b) bn = b2 , bf = b1 b3 .
transform (IFFT) is executed and Hermitian symmetry (HS) is
imposed to obtain a real-valued OFDM signal. The resultant effect of severe error propagation due to the non-orthogonality
digital signal is converted to an analog signal via digital-to- of the superposed constellation [17].
analog conversion (DAC) and a direct-current (DC) bias is When using 8-QAM constellation, totally three bits (b1 b2 b3 )
added to ensure the non-negativity of the LED-driving signal. can be transmitted per 8-QAM symbol. For a two-user VLC
After propagation over the indoor VLC channel, the light system, the three bits are allocated to the near user and the far
is converted into an electrical analog signal via photode- user. Specifically, two bit allocation schemes can be adopted
tection (PD) at each user. The obtained analog signals are according to users’ data rate requirements: (1) the near user is
converted back to digital signals via analog-to-digital conver- allocated with two bits while the far user is allocated with one
sion (ADC). In the following, fast Fourier transform (FFT) bit, i.e., Mn = 4 and Mf = 2; (2) the near user is allocated
and frequency-domain equalization (FDE) are performed to with one bit while the far user is allocated with two bits, i.e.,
obtain the respective constellations of both users. At each user, Mn = 2 and Mf = 4. Hence, flexible-rate multiple access can
uneven Gray-coded QAM demapping is first performed and be achieved by employing CPC/UCD-based NOMA.
then the desired bits for each user are extracted. The principle Let bn and bf denote the bit/bits allocated to the near and
of UCD with adaptive thresholds is discussed in Section II-B. far users, respectively. Fig. 3(a) illustrates the first bit alloca-
tion scheme with bn = b2 b3 and bf = b1 , where the 8-QAM
constellation is divided into two 4-QAM subconstellations.
A. Constellation Partitioning Coding (CPC) Fig. 3(b) shows the second bit allocation scheme with bn = b2
As shown in Fig. 1, the general rule of CPC for two-user and bf = b1 b3 , where the 8-QAM constellation is partitioned
NOMA consists of two parts. The first part is Gray-coded into four binary phase-shift keying (BPSK) subconstellations.
(Mn × Mf )-QAM mapping, where Mn and Mf denote the As shown in Figs. 3(a) and (b), for both bit allocation schemes,
orders of QAM constellations desired by the near user and the the electrical powers allocated to the far user and the near user
far user, respectively. The second part includes two steps: one are a 2 and b 2 , respectively. Hence, the power allocation ratio
is constellation partitioning which is performed to adaptively between the electrical powers allocated to the far user and the
partition a Gray-coded QAM constellation into multiple sub- near user for both bit allocation schemes is obtained by
constellations according to a pre-defined bit allocation scheme, a2
and the other is power allocation which is executed according ρ = 2. (1)
b
to a pre-defined power allocation strategy. In the following,
Therefore, the resultant (Mn × Mf )-QAM symbol after CPC
without loss of generality, we introduce CPC by taking the
can be represented by
8-QAM constellation as an example, i.e., Mn × Mf = 8.
The Gray-coded 8-QAM constellation for CPC is depicted P ρP
xCPC = xn + x , (2)
1+ρ f
in Fig. 2(a), where there is only a one-bit difference between
1+ρ
any two adjacent constellation points. However, for the non-
Gray-coded 8-QAM constellation generated by SPC, as shown where xn and xf are the symbols desired by the near and far
in Fig. 2(b), there is a two-bit difference between the constel- users, respectively, and P is the total input electrical power for
lation points representing bits “010” and “100”, and the ones the two users at the LED transmitter.
representing bits “011” and “101”. Therefore, the near user Although only 8-QAM constellation is considered here as
using SPC/SIC-based NOMA might suffer from the adverse an example, the general rule of CPC for two-user NOMA is
570 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
R EFERENCES
[1] T. Komine and M. Nakagawa, “Fundamental analysis for visible-
light communication system using LED lights,” IEEE Trans. Consum.
Electron., vol. 50, no. 1, pp. 100–107, Feb. 2004.
[2] H. Haas, L. Yin, Y. Wang, and C. Chen, “What is LiFi?” J. Lightw.
Technol., vol. 34, no. 6, pp. 1533–1544, Mar. 15, 2016.
[3] L. Grobe et al., “High-speed visible light communication systems,” IEEE
Commun. Mag., vol. 51, no. 12, pp. 60–66, Dec. 2013.
[4] Z. Ghassemlooy, L. N. Alves, S. Zvanovec, and M.-A. Khalighi, Visible
Light Communications: Theory and Applications. Boca Raton, FL, USA:
CRC Press, Jul. 2017.
[5] H. Le Minh et al., “100-Mb/s NRZ visible light communications using a
postequalized white LED,” IEEE Photon. Technol. Lett., vol. 21, no. 15,
pp. 1063–1065, Aug. 1, 2009.
[6] C. Chen, W.-D. Zhong, and D. Wu, “Indoor OFDM visible light com-
munications employing adaptive digital pre-frequency domain equaliza-
tion,” in Proc. Conf. Lasers Elect. Opt. (CLEO), Jun. 2016, pp. 1–2.
[7] M. Z. Afgani, H. Haas, H. Elgala, and D. Knipp, “Visible light com-
munication using OFDM,” in Proc. Int. Conf. Testbeds Res. Infrastruct.
Develop. Netw. Commun. (TRIDENTCOM), Mar. 2006, pp. 129–134.
[8] L. Zeng et al., “High data rate multiple input multiple output (MIMO)
Fig. 5. Measured BER vs. power allocation ratio for 8-QAM constellation optical wireless communications using white LED lighting,” IEEE J.
with (a) bn = b2 b3 , bf = b1 and (b) bn = b2 , bf = b1 b3 . Sel. Areas Commun., vol. 27, no. 9, pp. 1654–1662, Dec. 2009.
[9] T. Fath and H. Haas, “Performance comparison of MIMO techniques for
optical wireless communications in indoor environments,” IEEE Trans.
Commun., vol. 61, no. 2, pp. 733–742, Feb. 2013.
the effective power allocation ratio ρe as the value of ρ which [10] C. Chen, W.-D. Zhong, and D. Wu, “On the coverage of multiple-input
guarantees that both the near and far users can achieve BERs multiple-output visible light communications [Invited],” IEEE/OSA J.
below 3.8 × 10−3 . For 8-QAM constellation with bn = b2 b3 , Opt. Commun. Netw., vol. 9, no. 9, pp. D31–D41, Sep. 2017.
[11] H. Marshoud, V. M. Kapinas, G. K. Karagiannidis, and S. Muhaidat,
bf = b1 , ρe is in a very small range between 1.75 and 1.9 “Non-orthogonal multiple access for visible light communications,”
when using conventional SPC/SIC-based NOMA. However, IEEE Photon. Technol. Lett., vol. 28, no. 1, pp. 51–54, Jan. 1, 2016.
[12] X. Zhang, Q. Gao, C. Gong, and Z. Xu, “User grouping and power
when CPC/UCD-based NOMA is applied, the range of ρe allocation for NOMA visible light communication multi-cell networks,”
is extended to (1.68, 2.22). Similarly, for 8-QAM constella- IEEE Commun. Lett., vol. 21, no. 4, pp. 777–780, Apr. 2017.
tion with bn = b2 , bf = b1 b3 , the range of ρe is extended [13] Z. Yang, W. Xu, and Y. Li, “Fair non-orthogonal multiple access for
visible light communication downlinks,” IEEE Wireless Commun. Lett.,
from (6.85, 8.51) to (5.81, 9.29) when conventional SPC/SIC- vol. 6, no. 1, pp. 66–69, Feb. 2017.
based NOMA is replaced by CPC/UCD-based NOMA. From [14] L. Yin, W. O. Popoola, X. Wu, and H. Haas, “Performance evaluation of
the practical implementation point of view, a wider effective non-orthogonal multiple access in visible light communication,” IEEE
Trans. Commun., vol. 64, no. 12, pp. 5162–5175, Dec. 2016.
power allocation ratio range indicates the potential to support [15] C. Chen, W.-D. Zhong, H. Yang, and P. Du, “On the performance
higher data rates and the flexibility of overall system design. of MIMO-NOMA-based visible light communication systems,” IEEE
Photon. Technol. Lett., vol. 30, no. 4, pp. 307–310, Feb. 15, 2018.
[16] S. M. R. Islam, N. Avazov, O. A. Dobre, and K.-S. Kwak, “Power-
IV. C ONCLUSION domain non-orthogonal multiple access (NOMA) in 5G systems:
Potentials and challenges,” IEEE Commun. Surveys Tuts., vol. 19, no. 2,
In this letter, we have proposed and experimentally veri- pp. 721–742, 2nd Quart., 2017.
fied a novel CPC/UCD-based NOMA technique for downlink [17] H. Li, Z. Huang, Y. Xiao, S. Zhan, and Y. Ji, “Solution for error propaga-
tion in a NOMA-based VLC network: Symmetric superposition coding,”
VLC systems. By applying CPC with UCD in NOMA, the Opt. Exp., vol. 25, no. 24, pp. 29856–29863, Nov. 2017.
desired signals for all users in the VLC system can be [18] X. Guan, Q. Yang, Y. Hong, and C. C.-K. Chan, “Non-orthogonal
successfully decoded without SIC. Therefore, the error prop- multiple access with phase pre-distortion in visible light communica-
tion,” Opt. Exp., vol. 24, no. 22, pp. 25816–25823, Oct. 2016.
agation effect induced by imperfect SIC can be efficiently [19] X. Guan, Q. Yang, and C.-K. Chan, “Joint detection of visible light com-
mitigated. A two-user VLC system achieving a sum data rate munication signals under non-orthogonal multiple access,” IEEE Photon.
of 45 Mbit/s has been experimentally demonstrated to verify Technol. Lett., vol. 29, no. 4, pp. 377–380, Feb. 15, 2017.
[20] K. Ando, Y. Sanada, and T. Saba, “Joint maximum likelihood detection
the feasibility of the proposed CPC/UCD-based NOMA tech- in far user of non-orthogonal multiple access,” IEICE Trans. Commun.,
nique. The obtained results show that flexible rates can be vol. 100, no. 1, pp. 177–186, Jan. 2017.
572 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—We study the meta distribution (MD) of the cover- however, the joint decoding associated with SIC is not taken
age probability (CP) in downlink non-orthogonal-multiple-access into account.
(NOMA) networks. Two schemes are assessed based on the loca- This letter characterizes the MD in downlink cellular
tion of the NOMA users: 1) anywhere in the network and networks for two NOMA schemes, namely, everywhere
2) cell-center users only. The moments of the MD for both NOMA (E-NOMA) and cell-center NOMA (C-NOMA). E-
schemes are derived and the MD is approximated via the beta dis- NOMA utilizes NOMA for UEs located everywhere in the
tribution. Closed-form moments are derived for the first scheme;
network [5], [7], while C-NOMA restricts NOMA to cell-
for the second scheme exact and approximate moments, to sim-
plify the integral calculation, are derived. We show that restrict- center UEs only [2], [3]. We derive closed-form expressions
ing NOMA to cell-center users provides significantly higher for the moments of the MD in E-NOMA. Integral expressions
mean, lower variance and better percentile performance for are obtained for the moments in C-NOMA; consequently, we
the CP. propose accurate approximate moments to simplify the inte-
gral calculation. The MD is then approximated using the beta
Index Terms—Stochastic geometry, meta distribution, non- distribution via moment matching to characterize the UEs per-
orthogonal multiple access (NOMA).
centile performance. Different from [7] we derive and compare
the statistics of the MD for two NOMA schemes, and con-
sider joint decoding for all SIC phases. To the best of our
I. I NTRODUCTION
knowledge, NOMA works in the literature employ one scheme
ONVENTIONALLY, orthogonal multiple access (OMA)
C is used for transmissions to different users (UEs) served
by the same base station (BS). OMA assigns different time-
and do not compare different schemes. Our results show that
C-NOMA not only provides higher SCP, but also reduces
the variance of the CP across the UEs in the network when
frequency resource blocks (TF-RBs) to each UE to avoid compared to the E-NOMA.
intracell interference. However, spectrum scarcity and the
increasing capacity demand call for more efficient spectrum II. S YSTEM M ODEL
utilization. In this regard, non-orthogonal multiple access
(NOMA) is a technique that improves spectral efficiency by We consider a downlink cellular network where BSs are
superposing the messages of multiple UEs on one TF-RB. distributed according to a homogeneous PPP Φ with inten-
Successive interference cancellation (SIC) is used for NOMA sity λ. Each BS serves N UEs in one TF-RB by multiplexing
decoding. The superiority of NOMA over OMA schemes in a the signals for each UE with different power levels using a
noise-limited regime is well established from an information total power budget P = 1. A Rayleigh fading environment is
theoretic perspective [1]. assumed such that the fading coefficients are i.i.d. with a unit
Using stochastic geometry, the superiority of NOMA mean exponential distribution. A power-law path-loss model
has also been established for large-scale interference prone is considered where the signal decays at the rate r −η with
networks [2]–[5]. Such studies usually focus on the spatially distance r, η > 2 denotes the path-loss exponent and δ = η2 .
averaged coverage probability (SCP), which averages the cov- SIC requires ordering the UEs according to some measure of
erage probability (CP) over all fading, activity, and network link strength [2]. For i ∈ {1, . . . , N }, the i th strongest UE is
realizations. However, network operators are usually more referred to as UEi . In this letter, we order the UEs based on the
interested in the percentile performance of UEs, where the link distance R. The ordered link distance of UEi is denoted
fading and activity change while the network realization is by Ri ; consequently, UEi is nearer to the BS and therefore
kept constant. The CP given a fixed network realization is stronger than UEj for i<j (i.e., Ri < Rj ). Exploiting SIC,
defined as the conditional CP (CCP) [6]. The complementary UEi decodes and cancels messages intended for all weaker
cdf of the CCP, denoted as the meta distribution (MD), reveals UEs before decoding its own message. On the other hand,
the percentile performance across an arbitrary network realiza- messages for stronger UEs are treated as noise and contribute
tion. Reference [7] studies the MD for uplink and downlink to the intracell interference. We incorporate imperfect SIC into
NOMA with NOMA UEs located everywhere in the network; our analysis by considering a fraction β of residual intracell
interference from the canceled messages of weaker UEs. Let
Manuscript received September 12, 2018; accepted November 5, 2018. Date Pi and log(1 + θi ) denote the power allocated and target rate
of publication November 8, 2018; date of current version April 9, 2019. The for UEi ; the corresponding signal-to-interference ratio (SIR)
associate editor coordinating the review of this paper and approving it for threshold for the message of UEi is θi . Note that due to the
publication was M. Kountouris. (Corresponding author: Konpal Shaukat Ali.) power budget, N i=1 Pi = 1. For feasible SIC, proper resource
K. S. Ali and M.-S. Alouini are with the Computer, Electrical, and
Mathematical Sciences and Engineering Divison, King Abdullah University allocation (RA), i.e., power allocation and rate adaptation (e.g.,
of Science and Technology, Thuwal 23955-6900, Saudi Arabia (e-mail: Pi ≤ Pj and/or θi ≥ θj for i<j), for all UEs is required.
konpal.ali@kaust.edu.sa; slim.alouini@kaust.edu.sa). Lemma 1: For any ascending ordered statistic like Ri , based
H. ElSawy is with the Department of Electrical Engineering, King Fahd
on the statistics of the unordered counterpart R, the pdf is
University of Petroleum and Minerals, Dhahran 31261, Saudi Arabia (e-mail: i−1
fRi (r ) = N − 1 NfR (r ) FR (r ) (1 − FR (r ))N −i .
hesham.elsawy@kfupm.edu.sa).
Digital Object Identifier 10.1109/LWC.2018.2880210 i −1
(1)
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ALI et al.: MD OF DOWNLINK NOMA IN POISSON NETWORKS 573
In terms of components larger than i, (1) can be rewritten as For a fixed, yet arbitrary, realization of the network, the
N
CCP of UEi in a randomly selected cell, PCi , is
m −1 ⎡ ⎛ ⎞ ⎤
fRi (r ) = fR i (r ) + (−1)m−i fR
m (r ), (2)
i −1 (a)
m=i+1 PCi = P(Ci |Φ) = Egyi ⎣exp⎝−Riη Mi gyi yi −η ⎠ | Φ⎦
N −1 j −1 for i ≤ j ≤ N. In x∈Φ\x0
where fR j (r ) = j −1 NfR (r )(FR )
(b) 1
terms of components smaller than i, (1) can be rewritten as = , (6)
i−1 1 + Riη Mi yi −η
(N − m)!(−1)i−m
x∈Φ\x0
fRi (r ) = fR
(r ) + f (r ), (3) where (a) follows using the cdf of hi ∼ exp(1) and (b) follows
i (m − 1)!(i − m)! R m
m=1 from the MGF of the independent RVs gyi ∼ exp(1).
N −1
where fR j (r ) = j −1 NfR (r )(1 − FR )
N −j for 1 ≤ j ≤ i. Denote the b th moment of the CCP of UEi across all links
in an arbitrary fixed realization of the network by Mi,b . Then,
We denote the distance between a BS and its nearest ⎡ ⎤
neighboring BS by ρ. Since Φ is a PPP, the pdf of ρ is −b
η
Mi,b = E⎣ 1 + Ri Mi yi −η ⎦.
2
fρ (x ) = 2πλxe −πλx , x≥ 0. Consider a disk around each (7)
BS located at x with radius ρ/2, i.e., b(x,ρ/2); we refer to this x∈Φ\x0
as the in-disk. The in-disk is the largest disk centered at a BS
that fits inside its Voronoi cell. We study and compare NOMA Remark: If P̃j < 0, the CCP is zero. Henceforth we assume
for the following two schemes. RA such that P̃j ≥ 0.
1) Everywhere Noma (E-Noma): N UEs are distributed uni- Note: If b = 1 in (7), we obtain the SCP of UEi .
formly and independently in each Voronoi cell. Consequently, Through moment matching, the MD of UEi is approximated
the distribution of 2the unordered link distance R follows using the beta distribution [6] as follows
#
to the serving BS. UEs with relatively larger link distances are (a) Ri
= ERi exp −2πλ 1−f a da . (10)
better served in their own resource block without sharing [2]. Ri a
Using the ordered RDP for UEi , the expectation in (7) can
III. SIR A NALYSIS also be evaluated as ⎡ ⎤
SIC requires a UE to successfully decode all of the messages
intended for weaker UEs. Consider a randomly selected BS Mi,b = E⎣ (1 + Mi y η )−b ⎦. (11)
y∈Ri
located at x0 and its associated UEs; the SIR at UEi of the
message intended for UEj for i≤ j ≤ N is
hi Ri−η Pj A. E-NOMA Scheme
SIRij =
, We characterize the PGFL of the ordered RDPs and obtain
j
−1
N
hi Ri−η Pm + β Pk + gyi yi −η closed for expressions for Mi,b .
m=1 k =j +1 x∈Φ\x0 Lemma 2: The PGFL of Ri for 1≤ i ≤ N in E-NOMA is
where yi = x − ui , ui is the location of UEi , · denotes the i−1
(N − m)!(−1)i−m
Euclidean norm, and hi (gyi ) is the fading power gain from GRi [f ] = GR
i [f ] + G [f ], (12)
the serving (interfering) BS to UEi . (m − 1)!(i − m)! R m
m=1
Accordingly, due to SIC decoding, coverage at UEi is where for 1 ≤ j ≤ i
N −1
defined via the following joint event
⎧ ⎫ j −1 N
N ⎨ ⎬ GR
[f ] = . (13)
+ 1) + 2 1∞ (1 − f (y −1 ))y
N
θj j (N − j dy
Ci = SIRij > θj = hi > Riη gyi yi −η , (5)
⎩ P̃j ⎭ Proof: We obtain (12) using (3) in (10). Also using (10),
j =i j =i x∈Φ\x0
" ∞ " ∞ x
j
−1 N
where P̃j = Pj −θj ( Pm +β Pk ). We rewrite (5) as GR
[f ] = fR
(x ) exp −2πλ 1−f a da dx
j
0
j
Ri a
m=1 k =j +1
θ N − 1 " ∞
where (a) is obtained by changing variables and (13) is Lemma 4: Using A1 and A2, the PGFL of Ri conditioned
obtained using the MGF of m ∼ exp(πλ(N − j + 1)). on ρ for 1 ≤ i ≤ N in the C-NOMA scheme is
Corollary 1: Mi,b for 1 ≤ i ≤ N in E-NOMA is N
m −1
i−1
GRi |ρ [f ] = GR [f ] + (−1)m−i GR
m |ρ [f ],
$i,b + (N − m)!(−1)i−m $ i |ρ i −1
Mi,b = M Mm,b , (14) m=i+1
(m − 1)!(i − m)! (17)
m=1
where for 1 ≤ j ≤ i where for i ≤ j ≤ N
N −1
$j ,b = N − 1
M
N
. (15) j −1 Γ(j ) − Γ j , πλρ2 ∞
2 1 1 − f 1
y ydy
j − 1 N − j + 2 F1 (b, −δ, 1 − δ, −Mi ) j |ρ [f ] =
GR j .
1 ρ2 ∞ 1
Proof: (14) is obtained using (12), where we define using (11) N 2 πλ 1 1 − f y ydy
% &
1 (18)
M$j ,b = G
Rj
(1 + Mi y η )b Proof: We obtain (17) using (2) in (10). Also using (10),
N −1 " ∞ " ∞ x
(a) j −1 N GR
|ρ
[f ] = fR
(x ) exp −2πλ 1−f a da dx
= ∞ . j
0
j
Ri a
−η −b
" ρ2 −2πλm ∞(1−f (y −1 ))y
N −j +1+2 1 − (1 + Mi y ) ydy (a) N −1 4j 4 dy
1 = N 2j e 1 m j −1 dm
j −1 ρ 0
We obtain (a) using (13), and (15) follows by y → g −1 .
(a) follows by changing variables, and (18) by integration.
B. C-NOMA Scheme We approximate Mi,b by substituting the approximate
PGFL of Ri , conditioned on ρ, into (11) and averaging over ρ.
We obtain integral expressions for Mi,b . We also propose Corollary 2: Using A1 and A2, Mi,b for 1 ≤ i ≤ N in
approximate PGFLs of the ordered RDP and use these to
C-NOMA is
evaluate Mi,b in a simpler form. N
m − 1
Lemma 3: The b th moment of the CCP for UEi in the )i,b +
Mi,b = M )m,b ,
(−1)m−i M (19)
C-NOMA scheme is i −1
m=i+1
⎡
−b ⎤
∞ η where for i ≤ j ≤ N
rdr
−b
M R
−2πλ 1− 1+ ri η i
Mi Riη
⎡ ⎤
⎢ ρ−Ri ⎥
Mi,b ≈ Eρ,Ri ⎣e 1+ ⎦. πλρ2
ρη ⎢ Γ(j ) − Γ j , 4 (2 F1 (b, −δ, 1 − δ, −Mi ) − 1) ⎥
)j ,b
M = Eρ ⎢
⎣ j 2j
⎥.
⎦
(πλ) ρ j
j
(2 F1 (b, −δ, 1 − δ, −Mi ) − 1)
(16) N −1
j −1
N 4
V. C ONCLUSION
We study the MD of the CCP of NOMA UEs distributed
according to two models. Closed form expressions for the
moments of the MD in the E-NOMA scheme are derived.
The C-NOMA scheme requires a triple integral so we pro-
pose approximate moments that reduce to a single integration.
Fig. 3. SCP and variance of the MD vs. θ (identical target rate for all UEs) Our results show that employing NOMA for cell-center users
with P1 = 1/3 for both schemes. Solid lines are for β = 0, dashed for is significantly more beneficial than using it for all UEs in
β = 0.1, dash-dot for β = 0.2. Note: the weakest NOMA UE is unaffected a cell, thereby motivating the works of [2] and [3]. We also
by β.
emphasize the importance of RA in NOMA.
Fig. 2 plots the mean and variance of the MD for the NOMA R EFERENCES
UEs in the C-NOMA scheme. We compare using the moments
obtained with and without the approximations A1 and A2. [1] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
Cambridge, U.K.: Cambridge Univ. Press, 2004.
We observe that the approximation is tight for the SCP and [2] K. S. Ali et al., “Downlink non-orthogonal multiple access (NOMA) in
overestimates the variance, particularly for UE2 near the peak. Poisson networks,” IEEE Trans. Commun., to be published.
Fig. 3 plots the mean and variance of the MD of the UEs for [3] K. S. Ali et al., “Analyzing non-orthogonal multiple access (NOMA) in
both schemes using identical RA. We observe that C-NOMA downlink Poisson cellular networks,” in Proc. IEEE Int. Conf. Commun.
outperforms the E-NOMA scheme in terms of both SCP and (ICC), 2018, pp. 1–6.
[4] H. Tabassum et al.,“Modeling and analysis of uplink non-orthogonal
variance. Increasing β deteriorates performance of the non- multiple access (NOMA) in large-scale cellular networks using Poisson
weakest UEs, decreasing SCP and increasing variance. For cluster processes,” IEEE Trans. Commun., vol. 65, no. 8, pp. 3555–3570,
a given β, the higher SCP of the C-NOMA scheme can be Aug. 2017.
attributed to the fact that the UEs are closer to the BS on [5] K. S. Ali et al., “Non-orthogonal multiple access for large-scale
average than the E-NOMA scheme. The lower variance is also 5G networks: Interference aware design,” IEEE Access, vol. 5,
pp. 21204–21216, 2017.
due to the limited vicinity leading to lower disparity than the [6] M. Haenggi, “The meta distribution of the SIR in Poisson bipolar
E-NOMA model. Furthermore, σi2 peaks at high θ for the and cellular networks,” IEEE Trans. Wireless Commun., vol. 15, no. 4,
C-NOMA scheme (corresponding to low SCP); which is not pp. 2577–2589, Apr. 2016.
the case for the E-NOMA scheme. This implies the existence [7] M. Salehi et al. Meta Distribution of the SIR in Large-Scale Uplink and
of θ with high SCP and low σi2 in C-NOMA, thereby high- Downlink NOMA Networks. Accessed: Apr. 2018. [Online]. Available:
https://arxiv.org/abs/1804.02710
lighting its superiority with careful RA. The C-NOMA is also [8] R. K. Ganti and M. Haenggi, “Asymptotics and approximation of the
a more consistent scheme as both SCP and variance are better SIR distribution in general cellular networks,” IEEE Trans. Wireless
for UE1 than UE2 ; this is not the case for the E-NOMA. Commun., vol. 15, no. 3, pp. 2130–2143, Mar. 2016.
576 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In this letter, we focus on the high peak to average The PAPR reduction performance of TR-clipping-based
power ratio (PAPR) reduction problem in tone reservation-based schemes relies on how to select the peak reduction tone
orthogonal frequency division multiplexing systems. We first (PRT) set and the optimal clipping threshold. However, it’s
propose a parallel Tabu search (PTS)-based scheme to find a
sub-optimal peak reduction tone (PRT) set. After finding the a nondeterministic polynomial-time (NP)-hard issue [4] to
sub-optimal PRT set, we apply it in an adaptive iterative clipping find the optimal PRT set. Therefore, sub-optimal solutions
and filtering (AICF) method for PAPR reduction. Furthermore, are preferable, like the genetic algorithm (GA)-PRT [4],
PAPR reduction and bit error rate (BER) performances are com- cross entropy (CE)-PRT [5] and invasive weed optimiza-
pared among the AICF method, the adaptive scaling, the adaptive tion and particle swarm optimization (IWOPSO)-PRT [6].
amplitude clipping and the fast iterative shrinkage-thresholding
algorithm schemes. Simulation results verify that the PTS-based
These methods converge slowly and have high computation
PRT scheme can obtain better secondary peaks with lower com- complexity (CC).
putational complexity, and the AICF scheme effectively reduces To obtain a low PAPR signal, an adaptive scaling (AS)-
PAPR with a faster convergence speed while its BER performance TR algorithm [7] was proposed to reduce PAPR using a
is only slightly worse than for the existing methods. pre-determined clipping threshold. However, for the AS-TR
Index Terms—OFDM, PAPR, tone reservation, parallel tabu it is hard to select an optimal clipping threshold. An adap-
search. tive amplitude clipping (AAC)-TR [4] method was proposed
to improve AS-TR performance. A fast iterative shrinkage-
I. I NTRODUCTION thresholding algorithm (FISTA) scheme [8] was presented to
address PAPR with power control on magnitude of reserved
RTHOGONAL frequency division multiplexing
O (OFDM) has numerous advantages as its channels
are orthogonal to each other. This contributes to avoiding
tones. Although the AAC-TR and FISTA schemes acquire
better PAPR performance, the CCs of the AAC-TR and
FISTA ones are also higher. The equation-based approach
narrow-band interference and multi-path fading [1]. But
considers only in-band distortion at Nyquist sampling rate
OFDM has several drawbacks such as high peak to average
and does not take the effect of out-of-band noise into
power ratio (PAPR). High PAPR results in bad effects on
account [9].
the orthogonality of transmitted signals in conjunction with a
In this letter, we focus on the efficiency of finding the
non-linear power amplifier. To reduce high PAPR, numerous
PRT set and the CC decrease of methods for PAPR reduction.
conventional methods [2] have been proposed. Among the
Specifically, we first propose a novel method based on parallel
ways, the tone reservation (TR) method proposed by Tellado
tabu search (PTS) to search a sub-optimal PRT set. Then, we
firstly [3] is a simple and efficient one without requiring
propose an AICF algorithm to reduce PAPR after finding the
transmission of side information.
sub-optimal PRT set. We compare performances of the AICF
Manuscript received July 21, 2018; revised September 9, 2018 and October algorithm and traditional methods from two aspects of PAPR
6, 2018; accepted November 5, 2018. Date of publication November 9, 2018;
date of current version April 9, 2019. This work was supported in part by the
reduction and bit error ratio (BER). Simulation results validate
National Natural Science Foundation of China under Grant 61727802, Grant the effectiveness of our proposed scheme in PAPR reduction
61872184, and Grant 61501238, in part by the Jiangsu Provincial Science and BER performance.
Foundation under Project BK20150786, in part by the Specially Appointed
Professor Program in Jiangsu Province, 2015, in part by the Fundamental
Research Funds for the Central Universities under Grant 30916011205, and II. OFDM S YSTEMS AND T ONE R ESERVATION
in part by the Open Research Fund of National Mobile Communications
Research Laboratory, Southeast University, under Grant 2017D04. The asso- A. OFDM Systems and PAPR
ciate editor coordinating the review of this paper and approving it for
publication was J. Mietzner. (Corresponding author: Jun Li.)
In OFDM systems, N independent data symbols Xk are
Y. Wang is with the Department of Information and Computational modulated through phase shift keying (PSK) or quadrature
Sciences, Jiangsu University of Science and Technology, Zhangjiagang amplitude modulation (QAM) on a set of N orthogonal sub-
215600, China (e-mail: wangyj1859@just.edu.cn).
R. Zhang and F. Shu are with the School of Electronic and Optical
carriers with the oversampling factor J. The OFDM block
Engineering, Nanjing University of Science and Technology, Nanjing 210094, is expressed as X = [X0 , X1 , . . . , XN −1 ]T , where (·)T
China (e-mail: renjie.zhang@njust.edu.cn; shufeng@njust.edu.cn). means the transpose of a vector. After inverse fast Fourier
J. Li is with the School of Electronic and Optical Engineering, Nanjing transform (IFFT), the discrete time domain OFDM signal is
University of Science and Technology, Nanjing 210094, China, also with the
National Mobile Communications Research Laboratory, Southeast University, generated as
Nanjing, China, and also with the School of Computer Science and Robotics,
JN
−1
National Research Tomsk Polytechnic University, Tomsk 634050, Russia 1 j 2πmk
(e-mail: jun.li@njust.edu.cn). xm = √ Xk · e JN , m = 0, 1, . . . , JN − 1. (1)
Digital Object Identifier 10.1109/LWC.2018.2880432 JN k =0
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
WANG et al.: PAPR REDUCTION BASED ON PTS FOR TR IN OFDM SYSTEMS 577
The PAPR of x is defined as the ratio of maximum Algorithm 1 PTS Algorithm for PRT Set
instantaneous power to the average power, 1: Input N, M, U, K1 , pc , pm and K.
max |xm |2 2: Randomly generate four initial sequences. Do the TS for
0≤m<JN every sequence.
PAPR(x) = , (2)
E [|xm |2 ] 3: Update initial sequences and tabu list.
where x = [x0 , x1 , . . . , xJN −1 ]T . The complementary cumu- 4: Repeat K1 times and obtain four better sequences.
lative distribution function (CCDF) is used to measure the 5: Do crossover and mutation to the four better sequences
capability of PAPR reduction. The CCDF is the probability above. Select the best sequence of smallest SP.
that the PAPR of an OFDM symbol exceeds the predetermined 6: Repeat K cycles. Output the final PRT set with the
threshold PAPR0 , smallest SP.
TABLE I
Algorithm 2 AICF-Based Algorithm for PAPR Reduction S ECONDARY P EAK (SP) AND C OMPUTATION
1: Input OFDM symbols, initial γi , i , maximal iteration C OMPLEXITY (CC) C OMPARISON
number imax .
2: Calculate Pav and Ti .
3: Calculate vi , Vi , Hi and hi .
4: Get PAPR reduced signal xi+1 by (12).
5: Update the CR γi by (13).
6: i = i + 1. The algorithm ends until i = imax .
V. C ONCLUSION
In this letter, we propose a novel method based on parallel
tabu search (PTS) to find the sub-optimal PRT set for PAPR
reduction in OFDM systems. Compared with existing algo-
rithms, the PTS scheme can find better PRT sets with lower
Fig. 4. BER performance comparison among various methods. computational complexity. Then, we adopt the AICF algorithm
to decrease PAPR.Simulation results reveal that the proposed
AICF algorithm achieves better PAPR reduction, faster con-
CC of the PTS one is the lowest within 100 iterations among vergence rate and a comparable BER as the AS-TR and FISTA
the four algorithms. Above 100 iterations, The gap of the aver- algorithms.
age SPs for the four algorithms is small. Our aim is to fast
obtain a better PRT set with a lower computational complex- R EFERENCES
ity. So the PTS algorithm is the best choice among the four
[1] R. Van Nee and R. Prasad, OFDM for Wireless Multimedia
schemes. Communications. Boston, MA, USA: Artech House, Mar. 2000.
In Fig. 2, the PAPR reduction performances of various [2] T. Jiang and Y. Wu, “An overview: Peak-to-average power ratio reduction
schemes are compared with the identical PTS-PRT. The maxi- techniques for OFDM signals,” IEEE Trans. Broadcast., vol. 54, no. 2,
pp. 257–268, Jun. 2008.
mum number of iterations of the AICF-TR and FISTA schemes [3] J. Tellado, “Peak to average power reduction for multicarrier modula-
is 3, while that for the AAC-TR and AS-TR is 10. Other tion,” Ph.D. dissertation, Dept. Elect. Eng., Stanford Univ., Stanford,
parameters of the AAC-TR, AS-TR and FISTA are referred CA, USA, 2000.
to [4], [7], and [8], respectively. When the CCDF equals 10−4 , [4] Y. Wang, W. Chen, and C. Tellambura, “Genetic algorithm based nearly
optimal peak reduction tone set selection for adaptive amplitude clipping
the PAPR of the original OFDM signal, the AS-TR, AAC-TR, PAPR reduction,” IEEE Trans. Broadcast., vol. 58, no. 3, pp. 462–471,
FISTA and AICF-TR are 11.9 dB, 9.3 dB, 8.6 dB, 6 dB and Sep. 2012.
[5] J.-C. Chen and C.-P. Li, “Tone reservation using near-optimal peak
4.5 dB, respectively. Clearly, the PAPR reduction performance reduction tone set selection algorithm for PAPR reduction in OFDM
of the AICF-TR scheme is superior to others. systems,” IEEE Signal Process. Lett., vol. 17, no. 11, pp. 933–936,
In Fig. 3, the average PAPRs of the four schemes are Nov. 2010.
[6] H.-L. Hung, C.-H. Cheng, and Y.-F. Huang, “PAPR reduction of OFDM
compared as the iteration number increases. The initial CR using invasive weed optimization-based optimal peak reduction tone set
γ = 4.5 dB is the same for the AAC-TR, AS-TR and AICF- selection,” EURASIP J. Wireless Commun. Netw., vol. 244, pp. 1–12,
TR. When the iteration number is 2, the average PAPR of Oct. 2013.
the AICF-TR converges to 4.5 dB while the AAC-TR, AS-TR [7] L. Wang and C. Tellambura, “Analysis of clipping noise and
tone-reservation algorithms for peak reduction in OFDM sys-
and FISTA are still at the state of decrease. When the iter- tems,” IEEE Trans. Veh Technol., vol. 57, no. 3, pp. 1675–1694,
ation number is 10, the average PAPRs are 4.5 dB, 6.2 dB, May 2008.
7.2 dB and 4.2 dB for the AICF-TR, AAC-TR, AS-TR and [8] Y. J. Wang, S. Xie, and Z. B. Xie, “FISTA-based PAPR reduction method
for tone reservation’s OFDM system” IEEE Wireless Commun. Lett.,
FISTA, respectively. Fig. 2 and Fig. 3 show that the AICF-TR vol. 7, no. 3, pp. 300–303, Jun. 2018.
algorithm can acquire better PAPR decrease, lower CC and a [9] N. Bibi, A. Kleerekoper, N. Muhammad, and B. Cheetham, “Equation-
faster convergence rate than the AAC-TR, AS-TR and FISTA method for correcting clipping errors in OFDM signals,” SpringerPlus,
vol. 5, no. 1, p. 931, Jun. 2016, doi: 10.1186/s40064-016-2413-0.
schemes within 6 iterations. [10] F. Glover, “Future paths for integer programming and links to artifi-
To evaluate the BER performance of the whole OFDM sys- cial intelligence,” Comput. Oper. Res., vol. 13, no. 3, pp. 533–549,
tem, the processed signal passes through a solid-state power May 1986.
[11] J. Hou, C. Tellambura, and J. Ge, “Tone injection for PAPR reduction
amplifier (SSPA) model [1], using parallel Tabu search algorithm in OFDM systems,” in Proc. IEEE
Glob. Commun. Conf., Anaheim, CA, USA, 2012, pp. 4899–4904.
|Si | [12] K. Anoh, C. Tanriover, and B. Adebisi, “On the optimization of iterative
So = 1 e j Θ, (14) clipping and filtering for PAPR reduction in OFDM systems,” IEEE
|Si | 2p 2p
[1 + ( A ) ] Access, vol. 5, pp. 12004–12013, 2017.
580 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Internet of Things (IoT) promotes the awareness Recently, unmanned aerial vehicles (UAVs) have been incor-
about our world and eases human life. In IoT, machine- porated into the cellular system to assist mobile networks.
to-machine (M2M) communications empowers machine-type- UAV-mounted-BSs (UBSs) can provide coverage over a radius
devices (MTDs) to cooperatively exchange information and per- of 5km and enable high quality voice calls and real-time video
form actions. Current M2M communications primarily leverages
cellular networks to provision reliable services. When a disrup- streaming [4]. In addition, UBS provisions faster and flexible
tive disaster destroys the local cellular infrastructure, unmanned deployment and connectivity, and can boost network through-
aerial vehicle mounted base stations (UBS) can be deployed to put and enhance QoS, attributed to the presence of line-of-sight
assist rescue by facilitating M2M communications among the links.
human portable/wearable MTDs (HMTDs). The UBS network Since the conventional network access scheme (i.e., Random
access and resource allocation scheme is proposed to maximize Access Scheme (RAS)) in most cellular systems is not
the number of HMTDs to establish communications. We have designed for disaster rescue [5] and the resource alloca-
validated the proposed scheme through extensive simulations.
tion schemes in UAV-assisted M2M communications are not
Index Terms—Internet of Things, machine-to-machine commu- well investigated, in this letter, we propose to deploy UBS
nications, unmanned aerial vehicle mounted base station, disaster in the disaster area to assist disaster rescue by facilitating
rescue, network access and resource allocation. M2M communications for the human portable/wearable MTDs
(HMTDs). In this disaster rescue scenario, the objective is
I. I NTRODUCTION to enable the maximum number of HMTDs to establish con-
NTERNET of Things (IoT) promotes the level of awareness nectivity and send rescue messages with required data rates.
I about our world and helps run modern lives more effi-
ciently. Physical things (referred to as machine-type-devices
Hence, we propose the UBS Network Access and Resource
Allocation (UNARA) scheme to maximize the number of
(MTDs) in IoT) are embedded with sensing and trans- HMTDs in provisioning transmissions. As the UBS carries a
mission ability, and are thus enabled to gather and share limited capacity battery, the energy consumption of the UBS is
information. Home appliances, smart meters, vehicles, sen- a primary concern and the hovering time is a key factor for per-
sors, and human portable/wearable devices are examples of forming rescue, and so minimizing the transmission power for
such MTDs, which are expected to be the main users in relaying data at the UBS to tangibly extend the hovering time
IoT [1]. Machine-to-machine (M2M) communications empow- is critical and considered in this letter. We have also validated
ers various applications, such as health care, proximal social the proposed UNARA scheme through extensive simulations.
networking, facilities monitoring, transportation, security and
disaster rescue.
It has been predicted that by 2020, more than 50 billion II. S YSTEM M ODEL
MTDs will be connected for IoT [2] and in a smart city,
the number of MTDs per square kilometer will be in the The UBS deployment scenario is depicted in Fig. 1. Given
order of tens of thousands [3]. Owing to the pervasiveness the location (latitude and longitude coordinate) of the disaster
of MTDs and IoT applications, most current M2M communi- area, which loses local cellular infrastructure, the UBS first
cations relies on cellular (or licensed Low Power Wide Area reaches the airspace above the particular area. Once the alti-
(LPWA)) infrastructure because base stations (BSs) provision tude of the UBS is determined (the altitude selection is beyond
centralized scheduling and resource management, interference the scope of this letter), its effectively covered HMTDs are
mitigation, quality of service (QoS) guarantees and secured determined [6]. UBS will select the unused (empty) cellular
services. However, when a disaster destroys the existing cel- spectrum band in that area (to avoid the mutual interference if
lular infrastructure, local IoT services will be disrupted, thus there exist surrounding working BSs) to facilitate local M2M
hampering people from seeking help. communications among HMTDs. The UBS communication
system utilizes Long-Term Evolution (LTE) connectivity and
Manuscript received September 21, 2018; revised October 31, 2018; is time-slotted.
accepted November 5, 2018. Date of publication November 9, 2018; date Within the coverage of a UBS, we assume that the real-
of current version April 9, 2019. This work was supported by NSF under time channel state information (CSI) between the HMTDs,
Grant CNS-1814748. The associate editor coordinating the review of this
paper and approving it for publication was M. Nafie. (Corresponding author: and that between the HMTDs and the UBS are known by the
Xilong Liu.) UBS [7]. In every time slot, according to CSI, the HMTDs in
The authors are with the Advanced Networking Laboratory, Helen and the coverage are partitioned into two groups. If the channel
John C. Hartmann Department of Electrical and Computer Engineering,
New Jersey Institute of Technology, Newark, NJ 07102 USA (e-mail:
gain between the source and destination (SD) HMTD is better
xl249@njit.edu; nirwan.ansari@njit.edu). than that between the source HMTD and the UBS, we clas-
Digital Object Identifier 10.1109/LWC.2018.2880467 sify this SD pair into the direct M2M group (for simplicity, D
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
LIU AND ANSARI: RESOURCE ALLOCATION IN UAV-ASSISTED M2M COMMUNICATIONS FOR DISASTER RESCUE 581
B. Relay M2M Communications Model by increasing its transmission bandwidth WDSi , the data rate
will be increased accordingly. It can also be seen from Eq. (5)
Denote Rh as the index of SD pairs in the R group. Since that CDSD is a concave increasing function of W S for a fixed
the orthogonal channels are available, the transmission data i Di
PD S [10].
rate from the Rh th source HMTD to the UBS (SU) [8] is i
Suppose, at the UBS, the total available spectrum band-
S g SU
PR width is WM 2M . In order to enable the maximum number of
h Rh
CRSU = WRSh log2 (1 + ). (2) HMTDs to effectively transmit data in the D group, we pro-
h
N0 WRS
h pose the UBS Network access and Resource allocation (UNR)
Here, PRS is the R th source HMTD’s transmission power, algorithm (i.e., Algorithm 1).
h h
SU the channel gain between the R th source HMTD and By first allocating the spectrum to the HMTDs with rela-
gR h h tively small required bandwidth, the UNR algorithm enables
the UBS, and WRSh the bandwidth allocated to the Rh th source the maximum number of the HMTDs to transmit data.
HMTD. The computational complexity of the UNR algorithm is
U and W U are the transmission power and
Similarly, PR h Rh O(I 2 log I ), and thus the UNR algorithm is a polynomial time
bandwidth of the UBS for serving the Rh th destination HMTD algorithm, i.e., its solution can be obtained efficiently.
(UD), respectively; gRUD is the channel gain between the UBS
h
and the destination HMTD; the data rate is B. Resource Allocation for the Relay M2M Group
U g UD
PR h Rh
As the occupied bandwidth
in the D group is already
CRUD = WRUh log2 (1 + ). (3) obtained, WDirect = M W S
h
N0 WRU Dm =1 Dm , if WDirect < WM 2M ,
h the remaining bandwidth, i.e., WRelay = WM 2M − WDirect ,
For the relay M2M communications, in the first half time is allocated to the R group. According to Eq. (2), Eq. (4)
slot, the Rh th source HMTD sends the data to the UBS, and and the HMTDs’ required data rates, in the first half time
then, in the second half time slot, the UBS forwards the data slot, the UBS first executes Algorithm 1 (by replacing the
582 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
U WU
h
N0 WRUh
PR = (2 Rh
− 1) · UD
. (12)
h
gR
h
WM 2M in Line 6 with WRelay ) to maximize the number of
HMTDs to transmit data in R group and allocate the HMTDs Thus, the objective function Eq. (6) and its constraints can be
their required bandwidth. Suppose there are H HMTDs, each re-written as
UD
indexed by Rh , which are facilitated in the first half time slot; H
CR
H
Here, WRUh and CRUD h
are the only variables in Eq. (13).
U
min PR h
(6) By calculating the second-order derivatives of Eq. (12), its
U ,W U }
{PR
h R h Rh =1 Hessian matrix is positive semidefinite. That is, Eq. (12) is
s.t. CRSU ≤ UD
CRh (7) convex [10]. The summation of convex functions, Eq. (13), is
h still a convex function [10]. At the same time, the inequal-
H
ity constraints (14)-(18) are convex. Hence, this optimization
WRUh ≤ WRelay (8) problem is a convex problem.
Rh =1
H
U IV. S IMULATIONS
PR h
≤ PTotal (9)
Rh =1
Simulations are set up as follows. The radius of the UBS’s
coverage is 500m. The UBS is placed at 80m above the cen-
0 < WRUh (10) ter of the coverage area. The distribution of the HMTDs is
U
0< PR h
. (11) generated by a Poisson Point Process in this area. In simu-
lations, the distance-dependent channel model [7] is adopted.
Here, Eq. (9) constrains the total transmission power for relay- The HMTDs’ locations are fixed within each time slot 10ms.
ing data not to exceed the UBS’s hardware allowed maximum The transmission power of each HMTD is arbitrarily assigned
transmission power PTotal . It can be proved that this optimiza- among 8mW, 10mW and 20mW. In the coverage of the
tion problem is convex. The optimal solution can be obtained UBS, the pairing for the SD pairs are arbitrarily assigned.
efficiently. The SD pairs’ required data rates are arbitrarily assigned
LIU AND ANSARI: RESOURCE ALLOCATION IN UAV-ASSISTED M2M COMMUNICATIONS FOR DISASTER RESCUE 583
V. C ONCLUSION
Fig. 4. UBS’s transmission power vs. the number of HMTDs in the relay In this letter, we have proposed to adopt the UBS to assist
M2M group. M2M communications for disaster rescue. In order to enable
the maximum number of HMTDs to transmit data, we have
proposed the UNARA scheme to wisely select the HMTDs and
allocate them with network resource accordingly. Meanwhile,
among 150Kbps, 200Kbps and 300Kbps. The total available
the UNARA scheme has minimized the transmission power at
bandwidth at the UBS is 10MHz.
the UBS when it relays data for the HMTDs in the relay M2M
Fig. 2 shows the average number of HMTDs that can trans-
group to furthest prolong its hovering time. The performance
mit data, facilitated by our proposed UNARA scheme and
of the proposed UNARA scheme has been validated through
the conventional RAS [5] (i.e., HMTDs are randomly selected
extensive simulations.
to transmit), respectively, versus different amounts of UBS
total available bandwidth. In this simulation, the total num-
ber of active HMTD SD pairs is set as 620. When the UBS R EFERENCES
has more available bandwidth, a larger number of HMTDs [1] A. Rico-Alvarino et al., “An overview of 3GPP enhancements on
can transmit. In each setting of the available bandwidth, the machine to machine communications,” IEEE Commun. Mag., vol. 54,
no. 6, pp. 14–21, Jun. 2016.
UNARA scheme outperforms the conventional RAS because [2] D. Lake, A. Rayes, and M. Morrow, “The Internet of Things,” Internet
our proposed scheme preferentially allocates the spectrum to Protocol J., vol. 15, no. 3, pp. 10–19, 2012.
the SD pairs which require small bandwidth, thus enabling [3] Vodafone, “RACH intensity of time controlled devices,” 3rd Gener.
more HMTDs to access the spectrum to transmit. In contrast, Partnership Project, Sophia Antipolis, France, Rep. R2-102296, 2012.
[4] L. Zhang, Q. Fan, and N. Ansari, “3-D drone-base-station placement
with the conventional RAS, UBS randomly selects the active with in-band full-duplex communications,” IEEE Commun. Lett., vol. 22,
HMTDs in its coverage, and some of them may demand large no. 9, pp. 1902–1905, Sep. 2018.
bandwidth, and so the total number of HMTDs that can access [5] Y.-H. Chen, E. H.-K. Wu, C.-H. Lin, and G.-H. Chen, “Bandwidth-
the spectrum is not maximized. satisfied and coding-aware multicast protocol in MANETs,” IEEE Trans.
Mobile Comput., vol. 17, no. 8, pp. 1778–1790, Aug. 2018.
Fig. 3 shows the spectrum efficiency of the UBS network [6] M. Alzenad, A. El-Keyi, F. Lagum, and H. Yanikomeroglu, “3-D
with different amounts of UBS available bandwidth. Note that placement of an unmanned aerial vehicle base station (UAV-BS) for
with different amounts of UBS available bandwidth, the spec- energy-efficient maximal coverage,” IEEE Wireless Commun. Lett.,
trum efficiency of our proposed UNARA scheme and that of vol. 6, no. 4, pp. 434–437, Aug. 2017.
[7] F. Wang, C. Xu, L. Song, and Z. Han, “Energy-efficient resource
RAS is almost the same because the spectrum in both schemes allocation for device-to-device underlay communication,” IEEE Trans.
are all efficiently allocated once they select the HMTDs. Even Wireless Commun., vol. 14, no. 4, pp. 2082–2092, Apr. 2015.
though the data rate that can be transmitted over the avail- [8] C. E. Shannon and W. Weaver, The Mathematical Theory of
able bandwidth in both schemes are nearly the same, our Communication. Champaign, IL, USA: Univ. Illinois Press, 2002.
[9] G. Zhao, C. Yang, G. Y. Li, D. Li, and A. C. K. Soong, “Power and chan-
proposed UNARA scheme enables more HMTDs to trans- nel allocation for cooperative relay in cognitive radio networks,” IEEE
mit (Fig. 2) because the UNARA scheme first allocates the J. Sel. Topics Signal Process., vol. 5, no. 1, pp. 151–159, Feb. 2011.
spectrum to the SD pairs which require small bandwidth, thus [10] S. P. Boyd and L. Vandenberghe, Convex Optimization. Cambridge,
U.K.: Cambridge Univ. Press, 2004.
allowing more people to tap on the spectrum to send res- [11] L. Zhang and N. Ansari, “On the number and 3-D placement of in-
cue messages. In general, the spectrum efficiency of RAS is band full-duplex enabled drone-mounted base-stations,” IEEE Wireless
slightly lower than that of UNARA because in RAS, some of Commun. Lett., to be published, doi: 10.1109/LWC.2018.286750.
584 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—We propose a novel transmission scheme, suitable to low noise levels, resulting in high sensitivity and low power
Internet of Things, which sends coded redundant information in consumption [2].
independent packets. The results indicate that this scheme out- Reliability is always a concern in wireless communications,
performs, in terms of reliability, transmit power, and coverage, and it can be increased by retransmission and redundancy
the typical direct transmission strategy, schemes based on repli- schemes [3]–[6]. A feedback channel is usually employed to
cations, and methods that embed coded redundant information
in packets that contain new information. Moreover, we demon-
request retransmissions. However, LPWANs avoid the contin-
strate the impact of selecting the most appropriate transmission uous use of the downlink channel, since this link is often
scheme according to the target information outage probability near congested, as the BS covers several nodes at a low data
and node location. rate [1]–[5]. So, as in [3]–[5], this possibility is not con-
sidered in this letter. In contrast, redundancy-based schemes
Index Terms—Coded redundant message transmissions, energy do not need any feedback information since the devices
efficiency, Internet of Things (IoT), LPWAN.
transmit redundant packets without prior request. Remarkable
redundancy-based schemes are fountain codes [7], [8] and net-
I. I NTRODUCTION work coding [8], [9]. But since in LPWANs the one-hop link
HE INTERNET-OF-THINGS (IoT) aims at providing
T connectivity for thousands of devices [1]. An IoT base
station (BS) should cover many devices in a relatively simple
to the BS is essential to reduce energy consumption, solu-
tions based on cooperative diversity [8] are typically not well
suited [2]–[5].
manner, thus reducing infrastructure cost [2], [3]. Typically, Packet-level erasure coding on top of pure physical-layer
the BS collects information from many devices [1]–[5], which coding is known to be beneficial in block fading channels [6].
transmit a small amount of data with short duty cycle. In [3], by taking into account the randomness in time and
Ensuring energy efficient and reliable communications in frequency domains, a redundancy-based scheme using sim-
resource constrained networks is a primary concern in IoT [1]. ple packet replications has been considered in a UNB based
That is why Low Power Wide Area Networks (LPWANs), LPWAN. However, solutions based on coded packet redun-
such as LoRa and SigFox, attract so much interest [2]. For dancy [4], [5], [7] can bring increased reliability for the same
instance, by using ultra narrow band (UNB) and low data rate. In [4], inspired by network coding, we proposed a non-
rates, SigFox utilizes bandwidth efficiently and experiences cooperative transmission scheme based on coded redundant
information, while in [5], a convolutional fountain erasure
Manuscript received September 27, 2018; revised November 2, 2018; coding scheme for data recovery called DaRe is proposed.
accepted November 5, 2018. Date of publication November 12, 2018; date In this letter, first we extend the scheme addressed in [4] to
of current version April 9, 2019. This work was supported in part by
FONDECYT Post-Doctoral under Grant 3170021, in part by FONDECYT embed redundancy in more than just the next data packet and
Iniciación under Grant 11160517, in part by FONDEF under Grant to include more than just one coded message per packet. Then,
IT17M10012, in part by ERANet-LAC under Grant ELAC2015/T10-0761, and more importantly, we propose a novel coded transmission
in part by CNPq under Grant 304503/2017-7, and in part by the Araucaria scheme, which sends redundant information in new indepen-
Foundation, Brazil. The associate editor coordinating the review of this paper dent packets; unlike the aforesaid schemes [4], [5] which focus
and approving it for publication was A. Ozcelikkale. (Corresponding author:
Samuel Montejo-Sánchez.) on the case where redundancy is embedded in the next data
S. Montejo-Sánchez is with the Programa Institucional de Fomento a la packets. Besides, we establish a fair comparison between the
I+D+i, Universidad Tecnológica Metropolitana, Santiago 8940577, Chile transmission schemes, imposing a maximum delay constraint,
(e-mail: smontejo@utem.cl). taking into account the impact of the protocol overhead, and
C. A. Azurdia-Meza is with the Department of Electrical
Engineering, Universidad de Chile, Santiago 8370451, Chile (e-mail:
ensuring the same channel time utilization.
cazurdia@ing.uchile.cl). The main contribution of this letter is a novel transmission
R. D. Souza is with the Department of Electrical Engineering, Federal scheme, based on independent coded redundant informa-
University of Santa Catarina, Florianópolis 88034500, Brazil (e-mail: tion, that outperforms direct transmission, methods based on
richard.demo@ufsc.br). replications [3], and schemes based on embedded redundant
E. M. G. Fernandez is with the Department of Electrical Engineering,
Federal University of Paraná, Curitiba 81531-990, Brazil (e-mail: messages [4], [5], in terms of reliability, transmit power, and
evelio@ufpr.br). coverage, while ensuring the same channel time utilization.
I. Soto is with the Department of Electrical Engineering, Universidad de
Santiago de Chile, Santiago 9170124, Chile (e-mail: ismael.soto@usach.cl).
A. Hoeller is with the Department of Telecommunications Engineering, II. S YSTEM M ODEL
Federal Institute for Education, Science and Technology of Santa Catarina,
São José 88103-310, Brazil (e-mail: arliones.hoeller@ifsc.edu.br). We consider the uplink between node U and the BS
Digital Object Identifier 10.1109/LWC.2018.2880959 in an LPWAN. We assume additive white Gaussian noise
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
MONTEJO-SÁNCHEZ et al.: CODED REDUNDANT MESSAGE TRANSMISSION SCHEMES FOR LOW-POWER WIDE AREA IoT APPLICATIONS 585
TABLE I
an outage for different values of n, E VENTS FOR S UCCESSFUL D ECODING OF CTI-I W ITH n = 1
3 n−1 n
ICT−E,n = OCT -E,n (1 + OCT-E,n − OCT-E,n
2n−2 2n−1
+ OCT -E,n − OCT-E,n ). (5)
Note that the diversity order is not increasing with n, making
CT-E worse than RT-E, when n ≥ 2. DaRe [5] is a similar
embedded scheme, which uses linear combinations of previous
messages, but performs better with higher window size values,
limiting its application in scenarios with delay constraints.3 TABLE II
S YSTEM PARAMETER VALUES
B. Novel Independent Coded Transmission Scheme
Fig. 1-e illustrates the novel CT-I scheme with n = 1. The
k th message is sent in one independent data packet, but it is
also contained in two additional packets: i) the coded mes-
sage sent after the k th data packet, in a linear combination of
the (k − 1)th and the k th messages; ii) the coded message
sent after the (k + 1)th data packet, in a linear combina-
tion of the k th and the (k + 1)th messages. Table I lists the
events (one per row), between the (k − 3)th and (k + 3)th this idea, the number of complementary independent sets of
data packets, that allow the k th message decoding, according events that allow decoding of the k th message is 2n and
to the successful (S) or failed (F) decoding of that particular P(Ej ) = En : ∀j ∈ {1, 2, ..2n}. So, the information out-
packet, where M and R are new message and coded redun- age probability of CT-I can be formulated considering the
dant information, respectively. Thus, if the direct decoding of union of these independent sets of events [11,
the k th packet fails, the k th message can still be obtained eq. (2.15)],
ICT-I,n = 1 − [(1 − OCT-I,n ) + OCT-I,n P( 2n j =1 Ej )]. This
from the other events, which can be classified in two sets union can be rewritten using the binomial function, and then
containing: i) the events that depend on the linear combi- by an algebraic approach based on the binomial theorem [11,
nation with the (k − 1)th message, denoted as E1 and eq. (6.4)] we determine a closed form expression for the
highlighted in red; ii) the events that depend on the linear information outage probability of CT-I as
combination with the (k + 1)th message, denoted as E2
2n
and highlighted in green. From Table I we can determine the
2n
ICT-I,n = OCT-I,n − OCT-I,n (−1)j +1 Enj
probability of occurrence of both independent sets of events, j
j =1
which are P(E1 ) = P(E2 ) = E1 , 2n+1
= OCT-I,n − OCT-I,n (1 − (1 − En )2n ) = OCT 2n
-I,n Fn , (9)
2 3
E1 = (1 − OCT-I,1 ) + OCT-I,1 (1 − OCT-I,1 ) where
2 4
+ OCT -I,1 (1 − OCT-I,1 ) . (6) 2 3
En = (1 − OCT-I,n ) + OCT-I,n (1 − OCT-I,n )
Then, considering the union of the events in Table I, the 2 4
+ OCT-I,n (1 − OCT-I,n ) , (10)
information outage probability of CT-I with n = 1 is 2 3 4 5
Fn = 1 + OCT-I,n + OCT-I,n − 5OCT-I,n + 4OCT-I,n − OCT-I,n .
ICT-I,1 = 1 − (1 − OCT-I,1 ) + OCT-I,1 P E1 E2 (11)
2 3 2
= OCT-I,1 − OCT-I,1 2E1 − E1 = OCT-I,1 F1 , (7) Note that, from (9) the diversity order of CT-I is 2n + 1.
2 3 4 5
F1 = 1 + OCT-I,1 + OCT-I,1 − 5OCT-I,1 + 4OCT-I,1 − OCT-I,1 . (8)
IV. N UMERICAL R ESULTS
If n = 2 is used, the k th message is sent in one indepen-
dent data packet, but it is also contained in four additional We evaluate the proposed schemes in terms of information
packets, which leads to four independent sets of events that outage probability, transmit power consumption, and coverage.
All closed-form expressions were validated by Monte-Carlo
allow message decoding with probability of occurrence E2
simulations. Unless stated otherwise, we use the parameters in
and depending on: i) one coded message sent after the k th Table II, which coincide with those in the Sigfox standard [12].
data packet, in a linear combination of the (k − 2)th and Fig. 2 shows the information outage probability as a function
the k th messages; ii) another coded message sent after the of n, which is the number of replicas sent by RT or the number
k th data packet, in a linear combination of the (k − 1)th of coded messages sent by CT. Several interesting points arise:
and the k th messages; iii) the coded message sent after i) For all redundancy-based schemes, the optimum n depends
the (k + 1)th data packet, in a linear combination of the on the communication distance, except for CT-E whose opti-
k th and the (k + 1)th messages; iv) the coded message mum value is n = 1; ii) For the other schemes, the optimal
sent after the (k + 2)th data packet, in a linear combina- value of n decreases as d increases, n = 3 for RT-I and CT-I,
tion of the k th and the (k + 2)th messages. Generalizing as well as n = 5 for RT-E when d = 5 km (see Fig. 2-a),
while n = 2 for RT-I and CT-I, as well as n = 3 for RT-E
3 DaRe [5] with a code rate equal to 1/2 and W = 2 is equivalent to CT-E when d = 10 km (see Fig. 2-b); iii) For the same d, n of
with n = 1, while if W = 1, then it is equivalent to RT-E with n = 1. RT-E is greater than n of RT-I and CT-I, due to the protocol
MONTEJO-SÁNCHEZ et al.: CODED REDUNDANT MESSAGE TRANSMISSION SCHEMES FOR LOW-POWER WIDE AREA IoT APPLICATIONS 587
Fig. 2. Information outage as a function of the number of replicas or coded Fig. 4. (a) Minimum transmit power (Pt,min ) and (b) Maximum commu-
messages (n), for (a) d = 5 km and (b) d = 10 km. nication distance (dmax ) as a function of the information outage.
Fig. 3. Information outage as a function of (a) the transmit power (Pt ) and
V. C ONCLUSION
(b) the protocol overhead. We proposed CT-I, which outperforms direct transmission,
methods based on replications, and another coded transmission
efficiency of RT-E; iv) CT-I outperforms all other schemes in scheme based on embedded messages, in terms of reliabil-
terms of information outage with n . ity, transmit power, and coverage, with the same channel
Fig. 3-a shows the information outage probability as a time utilization. CT-I transmits redundancy with less addi-
function of the transmit power. Several interesting points tional packets than the other schemes. The optimum number
arise: i) All redundancy-based schemes outperform DT, despite of redundant messages depends on range, protocol overhead,
being their respective link outage probabilities greater than transmit power and information outage requirements.
ODT ; ii) CT-I with n = 2 outperforms all other schemes in
terms of information outage, when 12 dBm ≤ Pt ≤ 18 dBm; R EFERENCES
iii) The optimum n depends on Pt , for instance for CT-I if [1] G. A. Akpakwu, B. J. Silva, G. P. Hancke, and A. M. Abu-Mahfouz,
Pt < 12 dBm then n = 1, but if Pt > 18 dBm then n = 3. “A survey on 5G networks for the Internet of Things: Communication
technologies and challenges,” IEEE Access, vol. 6, pp. 3619–3647, 2018.
Fig. 3-b shows the information outage probability as a func- [2] U. Raza, P. Kulkarni, and M. Sooriyabandara, “Low power wide area
tion of the protocol overhead, defined as the ratio between the networks: An overview,” IEEE Commun. Surveys Tuts., vol. 19, no. 2,
header and the packet sizes. In this experiment, we set the DT pp. 855–873, 2nd Quart., 2017.
[3] Y. Mo, M.-T. Do, C. Goursaud, and J.-M. Gorce, “Optimization of the
packet size to 12 bytes, so the protocol overhead varies accord- predefined number of replications in a ultra narrow band based IoT
ing to lh and lm (e.g., if lh = 4 bytes, then lm = 8 bytes and the network,” in Proc. IEEE Wireless Days, 2016, pp. 1–6.
[4] S. Montejo-Sánchez et al., “An alternative non-cooperative transmis-
protocol overhead is 33%). Two additional points: i) The per- sion scheme based on coded redundant information,” in Proc. IEEE
formance of schemes based on embedded messages improves LATINCOM, 2017, pp. 1–6.
when the protocol overhead increases, due to (3); ii) CT-I with [5] P. J. Marcelis, V. S. Rao, and R. V. Prasad, “DaRe: Data recovery through
application layer coding for LoRaWAN,” in Proc. IEEE IoTDI, 2017,
n = 2 outperforms all other schemes in terms of information pp. 97–108.
outage, when the protocol overhead is less than 50 %, from [6] T. A. Courtade and R. D. Wesel, “Optimal allocation of redun-
which RT-E with n = 3 becomes the best option. dancy between packet-level erasure coding and physical-layer channel
coding in fading channels,” IEEE Trans. Commun., vol. 59, no. 8,
Fig. 4-a shows the minimum transmit power required to pp. 2101–2109, Aug. 2011.
sustain a target information outage probability. The proposed [7] J. W. Byers, M. Luby, and M. Mitzenmacher, “A digital fountain
scheme, CT-I with n = 2 and n = 1, as well as CT-E with approach to asynchronous reliable multicast,” IEEE J. Sel. Areas
Commun., vol. 20, no. 8, pp. 1528–1540, Oct. 2002.
n = 1, require half the transmit power of RT-I with n = 2 [8] R. R. Borujeny and M. Ardakani, “Fountain code design for the Y-
to meet Io = 10−3 . Requiring less Pt to meet the same Io network,” IEEE Commun. Lett., vol. 19, no. 5, pp. 703–706, May 2015.
[9] G. Angelopoulos, A. Paidimarri, A. P. Chandrakasan, and M. Médard,
and channel time utilization implies greater energy efficiency, “Experimental study of the interplay of channel and network cod-
since the additional computational cost of CT-I is minimal ing in low power sensor applications,” in Proc. IEEE ICC, 2013,
when compared to the RF power consumption and can thus pp. 5126–5130.
[10] A. Goldsmith, Wireless Communications. Cambridge, U.K.: Cambridge
be neglected. If information outage requirements are mild, then Univ. Press, 2005.
it is more convenient to use few replicas. Fig. 4-b shows the [11] T. T. Soong, Fundamentals of Probability and Statistics for Engineers.
Hoboken, NJ, USA: Wiley, 2004.
maximum range as a function of Io . For the case of a relatively [12] SigFox. (2018). Lightweight Protocol. [Online]. Available: https://www.
large Io = 10−2 , CT-E with n = 1 is the best option, while sigfox.com/en/sigfox-iot-technology-overview
588 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
βl < 1, respectively. The average packet loss rate of this model of an S-R transmission, D has received Y innovative packets,
Pgl
is Pel = PblP+P bl
gl
(1 − αl ) + Pgl +P bl
(1 − βl ) [4]. and R has X ≤ m packets buffered which are innovative with
Suppose that S has M source packets s 1 , . . . , s M to send respect to the Y innovative packets at D. Over the S-R link,
to D. Each packet contains K bytes. The relay node R connects the probability that R receives an innovative packet from S is
X +Y X +Y
S and D, and has a finite buffer of size m M packets. We then P1 = 1− q q M . The term q q M is the probability that a
assume that the oldest buffered packet would be replaced by uniformly randomly generated M-dimensional EV over Fq lies
the new incoming packet if the buffer is full. The two links in the span of the X + Y linearly independent M-dimensional
are transmitted in a time division manner. We assume that S-R vectors that have been received by R and D [9]. D receives an
transmits in odd time slots, and R-D transmits in even time innovative packet if and only if X > 0 before the R-D trans-
slots. The packet is received immediately by the downstream mission and the recoding coefficients are not all zero. Denote
node if it is not erased over the channel. the probability that a recoded packet from R is innovative as
WithRLNC, coded packets are sent from S in the form of P2 . We have P2 = 1 − q1X if the latest S-R transmission did
p = M i=1 gi s i , where the coding coefficients gi
s are uni-
not increase X, and P2 = 1− q X1+1 otherwise (i.e., X → X +1
formly randomly chosen from the finite field Fq of size q.
in the latest S-R transmission).
The vector [g1 , . . . , gM ] is the encoding vector (EV) of the
In the model, we have X ∈ {0, 1, . . . , m}, Y ∈
coded packet. Suppose that R has received k coded pack-
{0, 1, . . . , M }, and X + Y ≤ M. In addition, there might be
ets p 1 , . . . , p k , 0 < kk ≤ m, recoded k packets are sent from
M two states for each link, Good and Bad, abbreviated as G
R in the form of j =1 h j p j = j =1 h j ( i=1 gi,j s i ) = and B, respectively. The two-hop transmission process can
M k
i=1 ( j =1 gi,j hj )s i , where gi,j ’s are the coding coeffi- therefore be fully characterized by a Markov chain with state
cients of pj , hj ’s are uniformly randomly chosen from Fq , (X , Y , S1 , S2 ), where S1 and S2 denote the channel states of
k
and j =1 gi,j hj are the coding coefficients of the recoded
S-R and R-D, respectively, S1 ∈ {G, B }, S2 ∈ {G, B }. The
packet. The coding coefficients are delivered in the header total number of the (X , Y , S1 , S2 ) states is
of each coded/recoded packet. When D receives a sufficient
m(m + 1)
number of coded packets which contain M innovative (i.e., lin- η = 16 (m + 1)(M − m + 1) + , (2)
early independent) EVs, the source packets can be recovered 2
by Gaussian elimination. where 16 = 4 × 4 is because there are 4 different (S1 , S2 )
combinations and each of which may transit to 4 different
B. Performances Metrics states. Depending on the different numbers of innovative pack-
M , where E{T } ets at R and D, the state transition probabilities after every two
The achieved end-to-end rate is R = E{T } time slots can be determined as follows.
is the expected completion time, i.e., the expected num- For simplicity, here we only present the transition prob-
ber of time slots required for the decoder to recover all abilities for one possible change of the channel states,
the M source packets. The minimum expected completion (G, B) → (G, B). As mentioned, a total of 16 such channel state
time Tmin = 2M /(1 − max{Pe1 , Pe2 }) when the max-flow transitions are possible. The (X , Y , S1 , S2 ) transition proba-
capacity is achieved, where 2 is due to the time division trans- bilities corresponding to the other channel state transitions can
mission, and Pe1 and Pe2 are the average packet loss rates of be calculated similarly, and are omitted to save space.
the two links, respectively. The effective rate is defined as Case 1: X = 0, Y < M. Since no innovative packet is
K M buffered at R at the beginning of the S-R transmission, Y
Re = M log2 q
, (1)
K+ E{T } would not increase regardless of the R-D transmission if the
8
S-R link can not transmit an innovative packet successfully
where K is data payload size of each coded packet in bytes. (i.e., either the S-R transmission fails or the transmission suc-
Re measures the effects of coefficient overhead on the rate.
ceeds but the packet is non-innovative). When R receives
an innovative packet successfully, Y increases and X would
III. A NALYZING AND O PTIMIZING E FFECTIVE R ATE remain 0 if R-D transmits an innovative packet successfully.
We exploit a Markov chain model to analyze the expected Otherwise, Y would not increase and X would increase by 1.
completion time of using RLNC with finite field size q. The Therefore, we may have the following transition probabilities:
analysis is in part inspired by [8]. Based on the analysis, we
then optimize q to maximize the effective rate of RLNC. Pr{(X , Y , G, B) → (X , Y + 1, G, B)} = (1 − Pg1 )α1 P1 (1 − Pb2 )β2 P2 ,
Pr{(X , Y , G, B) → (X , Y , G, B)}
A. Analysis
= [(1 − Pg1 )α1 (1 − P1 ) + (1 − Pg1 )(1 − α1 )](1 − Pb2 ),
The model tracks the numbers of innovative packets at
Pr{(X , Y , G, B) → (X + 1, Y , G, B)}
R and D,1 which are then used to determine the expected
transition steps that are requried to achieve the transmission = (1 − Pg1 )α1 P1 [(1 − Pb2 )(1 − β2 ) + (1 − Pb2 )β2 (1 − P2 )],
completion.2 Specifically, we consider that at the beginning
where we recall that α1 and β2 are the probabilities of success-
1 We note that during the transmission we do not actually check the innova- ful transmission in Good and Bad states, respectively. Note that
tiveness of packets at R to avoid complicating the node. However, this model α2 and β1 do not appear here because the S-R link remains
simplifies the analysis, and will be shown to match the simulations. Good and the R-D link remains Bad.
2 The analysis can also be extended to a line network with multiple relays
Case 2: 0 < X < m, Y < M − X. On the one hand, when
by tracking the numbers of innovative packets at the relay nodes. However,
the state space would be significantly enlarged. Since this letter focuses on the S-R cannot transmit an innovative packet successfully, Y
the single-relay link, we omit the extension to the multi-relay scenario. would increase (and X would decrease) if the R-D transmits
590 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
an innovative packet successfully. Otherwise, both X and Y Algorithm 1 Search for the Optimal q
are not changed. On the other hand, when the S-R transmits 1: q ← 2, q̂ ← q
an innovative packet successfully, X would increase and Y 2: Calculate Re (q) according to equations (1) and (4)
are not changed if the R-D could not transmit an innovative 3: Re∗ ← Re (q) placeholder of the maximum Re
packet successfully. Otherwise, Y would increase and X is not 4: while 2q ≤ q ∗ do
changed. We may have the following transition probabilities: 5: q ← 2q
6: if Re (q) > Re∗ then
Pr{(X , Y , G, B) → (X − 1, Y + 1, G, B)} 7: Re∗ ← Re (q)
8: q̂ ← q optimal q
= [(1 − Pg1 )α1 (1 − P1 ) + (1 − Pg1 )(1 − α1 )](1 − Pb2 )β2 P2 ,
9: end if
Pr{(X , Y , G, B) → (X , Y , G, B)} = [(1 − Pg1 )α1 (1 − P1 ) 10: end while
+ (1 − Pg1 )(1 − α1 )][(1 − Pb2 )(1 − β2 ) + (1 − Pb2 )β2 (1 − P2 )],
Pr{(X , Y , G, B) → (X + 1, Y , G, B)}
expected number of times of staying in the corresponding state,
= (1 − Pg1 )α1 P1 [(1 − Pb2 )(1 − β2 ) + (1 − Pb2 )β2 (1 − P2 )],
given that the random process starts at (0, 0, G, G). E is an
Pr{(X , Y , G, B) → (X , Y + 1, G, B)} = (1 − Pg1 )α1 P1 (1 − Pb2 )β2 P2 . (η−4)×(η−4) identity matrix. The variance of the completion
time can also be obtained as
Case 3: X = m or X + Y = M. When X = m, R’s buffer is
full, and X would not increase after the S-R transmission. Y Var{T } = e 1,η−4 [(2H − E )t] − (e 1,η−4 t)2 , (5)
increases by 1 and X decreases by 1 accordingly if and only where e 1,η−4 is a row vector of length (η−4) and its elements
if the R-D transmits an innovative packet successfully. When are all-zero except that the 16th element being 1. t = H1 where
X + Y = M, R and D have obtained all the needed innovative 1 is a column vector of length (η − 4) with all 1 entries.
packets, and X would not increase anymore. Therefore, the
state transition probabilities are the same as X = m. We may
B. Optimizing Effective Rate
have the following transition probabilities:
It is clear that the completion time depends on the finite field
Pr{(X , Y , G, B ) → (X − 1, Y + 1, G, B )} size q. From (1), q also affects the coefficient overhead. We
= (1 − Pg1 )(1 − Pb2 )β2 P2 , can therefore optimize q based on the above analysis results to
Pr{(X , Y , G, B ) → (X , Y , G, B )} maximize the effective data rate. We consider that the nodes
have limited computation capability and storage space, and
= (1 − Pg1 )[(1 − Pb2 )(1 − β2 ) + (1 − Pb2 )β2 (1 − P2 )]. therefore the maximum supported finite field size is limited.
Note that we have 4 absorbing states, i.e., (0, M, G, G), Let q ∗ denote the maximum allowed finite field size. Without
loss of generality, the allowed finite fields are all extensions
(0, M, G, B), (0, M, B, G), (0, M, B, B). Let ζ = M −m +1 and
of F2 , i.e., q = 2k , k ∈ Z+ . We can use Algorithm 1 to
label the 4 channel state combinations (S1 , S2 ), i.e., (G, G), traverse all q s in [2, q ∗ ] to find the optimal q that maximizes
(G, B), (B, G), (B, B), as W = 0, 1, 2, 3, respectively. We can the effective data rate. The optimal q is denoted as q̂.
obtain the transition matrix of the Markov chain by ordering
all the four-tuple states. Define IV. S IMULATION R ESULTS
⎧
⎪ In this section, we evaluate the performance of the proposed
⎨16[(m + 1)Y + X + 1] − W ,
⎪ Y ≤ ζ,
−ζ−1 scheme. Throughout the coded transmission simulations, we
f (X , Y , S1 , S2 ) = 16[(m + 1)ζ + kY=0 (m − k )
⎪
⎪
⎩ use packets each containing K = 200 bytes given the fact that
+X + 1] − W , Y > ζ, many applications such as IoT use small payload sizes. The
where W is to identify the channels states. Let ΔX and results are based on averaging over 100,000 trials.
ΔY denote the change of X and Y after every two time We first validate the proposed analytic model. Suppose that
slots, respectively, ΔX , ΔY ∈ {−1, 0, 1}. By assigning index the number of source packets M = 64, and the channel state
λ = f (X , Y , S1 , S2 ) and ω = f (X + ΔX , Y + ΔY , S1 , S2 ), transition probabilities of the links are Pg1 = Pb1 = Pg2 =
we can write the η × η transition matrix of (X , Y , S1 , S2 ) as Pb2 = 0.2. The successful transmission probabilities on each
link in the Good and Bad states are α1 = α2 = 0.8 and β1 =
A Z β2 = 0.6, respectively. We observe that the simulation and
Π= . (3)
0 I 4×4 analysis results (see equations (4), (5)) are closely matched in
Fig. 2, where the vertical bars are the standard deviations. Non-
The (λ, ω)-th entry of matrix Π, [Π]λ,ω = surprisingly, the completion time decreases as the finite field
Pr{(X , Y , S1 , S2 ) → (X + ΔX , Y + ΔY , S1 , S2 )}. A size increases. The average completion time decreases more
is an (η − 4) × (η − 4) matrix and I 4×4 is an 4 × 4 identity significantly when the finite field size is small (i.e., q ≤ 4),
matrix which corresponds to the 4 absorbing states. whereas it remains almost unchanged when q is large (i.e.,
Given Π, and assuming that the transmission starts at the q ≥ 8). Note that there is a gap between the achieved and the
16-th state, i.e., (0, 0, G, G), the expected completion time minimum achievable completion times. The reason is that the
n−4 results are in the finite regime M = 64. The completion time
E{T } = [H ]16,b , (4) would achieve the minimum as M → ∞ and q → ∞.
Fig. 3 shows the effective rates of RLNC using q =
b=1
{2, 4, 8, . . . , 256}, respectively. The maximum rate is achieved
where H = (E − A)−1 is the fundamental matrix of the at q = 4. Although q = 256 has the smallest completion time
Markov chain [8]. Each entry in the 16-th row of H is the as shown in Fig. 2, the effective rate is the lowest due to the
WU et al.: ON OPTIMIZING EFFECTIVE RATE FOR RLNC OVER BURST-ERASURE RELAY LINKS 591
V. C ONCLUSION
In this letter, we showed that, by analyzing the completion
time of RLNC over finite-buffer burst-erasure relay links using
a Markov chain model and then optimizing the finite field size,
we can increase the effective rate of RLNC.
R EFERENCES
[1] X. Ge, S. Tu, G. Mao, C.-X. Wang, and T. Han, “5G ultra-dense cel-
lular networks,” IEEE Wireless Commun., vol. 23, no. 1, pp. 72–79,
Feb. 2016.
Fig. 3. Effective rates with various q. [2] T. Ho et al., “A random linear network coding approach to multicast,”
IEEE Trans. Inf. Theory., vol. 52, no. 10, pp. 4413–4430, Oct. 2006.
[3] Z. Liu, C. Wu, B. Li, and S. Zhao, “UUSee: Large-scale operational
on-demand streaming with random network coding,” in Proc. IEEE
high coefficient overhead. We observe that Re has only one INFOCOM, San Diego, CA, USA, 2010, pp. 1–9.
maximum in Fig. 3. This implies that if using Algorithm 1 [4] X. Xu, Y. L. Guan, and Y. Zeng, “Batched network coding with adaptive
recoding for multi-hop erasure channels with memory,” IEEE Trans.
to optimize q, we may actually exit the search after evalu- Commun., vol. 66, no. 3, pp. 1042–1052, Mar. 2018.
ating q = {2, 4, 8}, i.e., breaking out the loop after finding [5] J. Heide, M. V. Pedersen, F. H. P. Fitzek, and M. Medard, “On code
the first q such that Re ( q2 ) < Re (q) > Re (2q). This may parameters and coding vector representation for practical RLNC,” in
save the computational cost significantly. Through extensive Proc. IEEE Int. Conf. Commun. (ICC), 2011, pp. 1–5.
numerical evaluations and simulations for various M and q, [6] M. Nistor, D. E. Lucani, T. T. V. Vinhoza, R. A. Costa, and J. Barros,
“On the delay distribution of random linear network coding,” IEEE J.
we have confirmed that such a simplified algorithm is indeed Sel. Areas Commun., vol. 29, no. 5, pp. 1084–1093, May 2011.
effective.3 [7] X. Shi, M. Medard, and D. E. Lucani, “Whether and where to code in
In Fig. 4, we show the simulated effective rates for various the wireless packet erasure relay channel,” IEEE J. Sel. Areas Commun.,
M using q s found by Algorithm 1 and that found by the vol. 31, no. 8, pp. 1379–1389, Aug. 2013.
[8] Y. Li et al., “A low-complexity coded transmission scheme over finite-
simplified algorithm, respectively. The corresponding optimal buffer relay links,” IEEE Trans. Commun., vol. 66, no. 7, pp. 2873–2887,
q s are marked. As a comparison, we also plot the calculated Jul. 2018.
[9] O. Trullols-Cruces, J. M. Barcelo-Ordinas, and M. Fiore, “Exact decod-
3 Unfortunately, we are not able to mathematically prove that R has only
e ing probability under random linear network coding,” IEEE Commun.
one maximum since we cannot obtain T as a closed-form function of q. Lett., vol. 15, no. 1, pp. 67–69, Jan. 2011.
592 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In this letter, we consider a typical scenario where In practice, there is an average transmit power constraint at
the transmitter employs different power adaption methods, the transmitter, which will involve power adaption according to
including the optimal rate and power algorithm, optimal rate instantaneous channel state to enhance the EC. Optimal power
adaption, channel inversion and truncated channel inversion,
and rate algorithm (OPRA) and channel inversion (CI) are
to enhance the ergodic capacity (EC) with an average trans-
mit power constraint over fluctuating two-way fading channels. two comment methods of power adaption [5]–[7], where the
In particular, we derive exact closed-form expressions for the performance of OPRA is much better than that of CI, because
EC under different power adaption methods, as well as cor- a large amount of the transmit power is required to compensate
responding asymptotic formulas for the EC valid in the high for the deep channel fading. To improve the performance of
signal-to-noise ratio region. Finally, we compare the performance CI, [8] proposed a kind of truncated CI (TCI), where the cutoff
of the EC under different power adaption methods, and this also level can be selected to achieve a specified outage probability.
validates the accuracy of our derived expressions for the exact
In this letter, we derive exact and corresponding asymptotic
and asymptotic EC.
closed-form expressions for the EC over FTR fading channels
Index Terms—Asymptotic ergodic capacity, ergodic capacity, under different power adaption methods, namely OPRA, ORA,
fluctuating two-ray fading channel, power adaption. CI and TCI, and compare the EC among them by simulation.
I. I NTRODUCTION
UE TO the exponential increase in aggregate traffic, II. S YSTEM M ODEL
D millimeter-wave (mmWave) has been recently used to
overcome the wireless spectrum shortage. Although some con-
The EC is defined as
∞
ventional fading channels, such as Rayleigh and Rician fading C = E{ln(1 + γd )} = ln(1 + γd )fγd (γd )d γd , (1)
0
channels, has been verified to suit sometimes the mmWave
radio communications, the fluctuation suffered by the received where γd is the instantaneous SNR at the destination, and
signal cannot always be modeled accurately by conventional fγd (·) is the probability density function (PDF) of γd . In
fading models. In view of this issue, [1] introduced the fluctu- the following sections, we will investigate several adaptive
ating two-ray (FTR) fading model consisting of random phase transmission methods to improve the EC.
plus a diffuse component, which is the natural generalization The PDF and cumulative density function (CDF) of γd over
of the two-wave with diffuse (TWDP) model in [2] where FTR fading channels are given by [3]
the specular components of the TWDP model are just con-
stant amplitudes, and also can reduce to many conventional md ∞ jd γdjd exp − 2σγd
2
md Kd djd d
fading models, such as Rician and Nakagami-m fading mod- fγd (γd ) = 2 jd +1 , (2)
Γ(md ) j !j !
els (more special cases refer to [1, Table I]). Subsequently, jd =0 d d 2σd
[3], [4] extended the work of [1] in terms of elementary func- md ∞ jd
md Kd djd γd
tions and coefficients consisting of fading parameters, where Fγd (γd ) = Υ jd + 1, 2 , (3)
Γ(md ) j !j ! 2σd
the parameter m in FTR fading can be valued by an arbitrary j =0 d d
d
positive real number, rather than only positive integers in [1].
However, there is no derivation of asymptotic ergodic capacity respectively, where Γ(·) and Υ(·, ·) denote the Gamma func-
(AEC) and asymptotic ergodic secrecy capacity (AESC) in the tion and lower incomplete Gamma function [9], respectively.
high signal-to-noise ratio (SNR) region in [3] and [4], resulting Kd is the average power ratio of the dominant waves and
in missing some insights and less efficiency in calculation of remaining diffuse multipath, md is the parameter of Gamma
the ergodic capacity (EC) and ergodic secrecy capacity (ESC) distribution with unit mean, σd2 is the variance of the real (or
in the high SNR region. Further, the authors in [3] and [4] imaginary) diffuse component, and
only consider the optimal rate adaption (ORA) case, where jd
k
k
Δ jd Δd k
the transmit power is fixed over the whole transmission. djd = Γ(jd + md + 2l − k )
k 2 l
Manuscript received September 7, 2018; revised October 25, 2018; accepted k =0 l=0
November 8, 2018. Date of publication November 14, 2018; date of current
−(jd 2+md )
π(2l − k )i 2 2
version April 9, 2019. The associate editor coordinating the review of this · exp (md + Kd ) − (Kd Δd )
paper and approving it for publication was I. Krikidis. (Corresponding author: 2
⎛ ⎞
Hui Zhao.)
k −2l m d + Kd
The authors are with the Computer, Electrical, and Mathematical ·Pjd +md −1 ⎝ ⎠, (4)
Science and Engineering Division, King Abdullah University
of Science and Technology, Jeddah 23955-6900, Saudi Arabia (md + Kd )2 − (Kd Δd )2
(e-mail: hui.zhao@kaust.edu.sa; zhedong.liu@kaust.edu.sa;
slim.alouini@kaust.edu.sa). in which Δd , i, and P(·) denote a ratio defined by (4) in [3], the
Digital Object Identifier 10.1109/LWC.2018.2881158 imaginary unit, and the Legendre function of the first kind [9],
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHAO et al.: DIFFERENT POWER ADAPTION METHODS ON FTR FADING CHANNELS 593
respectively. From Fγd (∞) = 1 and [3, eqs. (7) and (8)], In view of the definition of upper incomplete Gamma function
md d ∞
m j
Kdd djd Γ(·, ·), we finally have the constraint condition as
we can easily derive Γ(m ) j d =0 jd ! = 1, and further
d
∞ j
rewrite Fγd (·) as mdmd Kdd djd
∞ j Γ(md ) j !j !
mdmd Kdd djd j =0 d d
Fγd (γd ) = 1 − d
Γ(md ) jd ! 1 γ0 1 γ0
jd =0 · Γ jd + 1, 2 − 2 Γ jd , 2 = 1. (12)
j n γ0 2σd 2σd 2σd
γd d
1 γd
· exp − 2 . (5)
2σd n! 2σd2 Let
n=0 ∞
1 fγd (γd )
f (γ0 ) = 1 − Fγd (γ0 ) − d γd . (13)
III. E RGODIC C APACITY U NDER OPRA γ0 γ0 γd
A. Exact EC Under OPRA Differentiating f (γ0 ) with respect to γ0 by using Leibniz Rule,
Our goal is to adjust the transmit power according to the we have
instantaneous channel state to maximize the EC subject to a ∂f (γ0 ) 1 − Fγd (γ0 )
certain average transmit power (P t ) by using OPRA, where =− . (14)
the EC in the integral form is given by (7) in [5] ∂γ0 γ02
∞ ∂ f (γ ) < 0. Thus, f (γ ) is monotonically
γ For γ0 > 0, ∂γ 0 0
C opra = ln d fγd (γd )d γd , (6) 0
γ0 γ0 decreasing over γ0 ∈ [0, ∞). When γ0 → ∞, (13) will
converge to 0, and when γ0 → 0, it will go to infinity. To sum-
where γ0 = λP t , in which λ ≥ 0 is the corresponding marize, there exists unique γ0 satisfying the identity of (12).
Lagrangian multiplier. Substituting the PDF of γd into (6), When 2σd2 → ∞, we have (15), shown on the top of the
we can derive the EC under OPRA as next page. Our numerical results show that γ0 increases as
mdmd ∞ j
Kdd djd γ D increases, and as such γ0 will lie in the interval [0, 1].
C opra = 2 jd +1
Γ(md )
jd =0 jd !jd ! 2σd B. Asymptotic EC Under OPRA
∞
Γ(s,x )
γd jd γd By using the limit identity lim xs = − 1s for Re(s) < 0
· ln γ exp − 2 d γd , (7) x →0
γ0 γ0 d 2σd and lim Γ(0, x ) = − ln x + ψ(1), we have
x →0
I1
Γ −p + l − 1, 1 2σd2 1
where lim 2 −p+l−1 = , (16)
2
1/(2σd )→0 p + 1−l
∞ 1 2σd
x + γ0 jd x + γ0
I1 = ln (x + γ0 ) exp − dx and
0 γ0 2σd2
jd −p p
jd
γ0 1 γ0
jd jd −p γ0 lim + Γ 0, 2
= γ exp − 2 1/(2σd2 )→0 2σd2 p+1−l 2σd
p 0 2σd l=1
p=0 ⎧
∞ ⎨0, p < jd ;
x x
p
· x ln 1 + exp − 2 dx . (8) =
⎩ψ(p + 1) − ln γ0
, p = jd ,
(17)
0 γ0 2σd 2σ 2 d
By using the integral identity derived in [5, Appendix B], I1 where ψ(·) denotes the digamma function [9]. In view of this
can be easily solved in closed-form as limit relationship, the AEC under OPRA can be derived by
jd p+1
jd
2 l j +1−l γ0 ∞ md
md ∞
j
Kdd djd
I1 = p! 2σd γ0 d
Γ −p − 1 + l , 2 , (9) C opra = ln 2σd2 − ln γ0 + ψ(jd + 1). (18)
p=0
p
l=1
2σd Γ(md )
jd =0
jd !
m ∞ j jd m ∞ j ∞
1 md d Kdd djd γ0 1 γ0 n md d Kdd djd jd −1 γd
exp − 2 − γ exp − d γd = 1 (11)
γ0 Γ(md ) j =0 jd ! 2σd n=0 n! 2σd2 Γ(md ) j =0 jd !jd !(2σ 2 )jd +1 γ0 d 2σd2
d d d
mdmd ∞
Kdjd djd 1 γ0 1 γ0
m ∞
1 md d Kdd djd
j
Γ jd + 1, 2 − 2 Γ jd , 2 = = 1 ⇒ γ0 = 1 (15)
Γ(md ) j !j ! γ0 2σd 2σd 2σd γ0 Γ(md ) jd !
jd =0 d d jd =0
=1
We can easily see that the slope of EC in high SNRs is unity The first derivative of the nth moment with respect to n is
with respect to ln γ D , and the power offset of EC in high SNRs
Kdd djd
2 n
∞ j
is independent of γ D , because when γ D → ∞, γ0 → 1, which ∂E γdn mdmd
= 2σd Γ(n + jd + 1)
means that ln γ0 → 0, and therefore, there is no impact of γ D ∂n Γ(md ) j !j !
jd =0 d d
on the power offset in high SNRs.
!
· ln 2σd2 + ψ(jd + n + 1) . (23)
∞ ⎜ 2σd2 Γ(md ) ⎟
C tci = ln⎜
⎝ ⎟
⎠
ORA in the low SNR region. It is interesting to note that the
m ∞ j
Kdd djd Γ(jd )
md d d0 ξ + jd =1 jd !jd !
EC under CI is zero, and this can be explained by the fact that
⎛ ⎞ Γ(jd ) → ∞ for jd = 0 while d0 is not equal to zero or Kd
∞
j
Γ(md ) Kdd djd Γ(jd ) is not infinity in (26).
= ln 2σd 2
+ ln m − ln⎝d0 ξ + ⎠. (31)
md d jd !jd ! From Fig. 1, our derived asymptotic results matches very
jd =1
well with the simulation and analytical results in the high SNR
From (31), we can see that the EC under TCI is not a line region, and the slope of EC under OPRA and ORA is always
function with respect to ln(2σd2 ) or ln(γ D ) in the high SNR unity with respect to ln γ D , regardless of parameter settings.
region. However, the slope of (31) with respect to ln(2σd2 )
changes very slowly in the high SNR region, which can be R EFERENCES
shown in Fig. 1 in the simulation section. [1] J. M. Romero-Jerez, F. J. Lopez-Martinez, J. F. Paris, and
A. J. Goldsmith, “The fluctuating two-ray fading model: Statistical char-
acterization and performance analysis,” IEEE Trans. Wireless Commun.,
VI. S IMULATION vol. 16, no. 7, pp. 4420–4432, Jul. 2017.
[2] L. Wang, N. Yang, M. Elkashlan, P. L. Yeoh, and J. Yuan, “Physical
In this section, we use Monte-Carlo simulation to validate layer security of maximal ratio combining in two-wave with diffuse
our derived closed-form expressions for the EC and AEC in power fading channels,” IEEE Trans. Inf. Forensics Security, vol. 9,
OPRA, ORA, CI, and TCI cases. In calculation of the infi- no. 2, pp. 247–258, Feb. 2014.
[3] J. Zhang, W. Zeng, X. Li, Q. Sun, and K. P. Peppas, “New results on
nite summation terms in the PDF and CDF of γd over FTR the fluctuating two-ray model with arbitrary fading parameters and its
fading channels, we can truncate the infinite summation terms applications,” IEEE Trans. Veh. Technol., vol. 67, no. 3, pp. 2766–2770,
into finite terms, where the resulting truncation error can be Mar. 2018.
evaluated by (6) shown on [4, Sec. II-B]. [4] W. Zeng, J. Zhang, S. Chen, K. P. Peppas, and B. Ai, “Physical
layer security over fluctuating two-ray fading channels,” IEEE
As shown in Fig. 1, we can easily see that the EC increases Trans. Veh. Technol., vol. 67, no. 9, pp. 8949–8953, Sep. 2018,
with γ D increasing, because of the improved channel state doi: 10.1109/TVT.2018.2842126.
between the transmitter and receiver. It is also obvious that [5] M.-S. Alouini and A. J. Goldsmith, “Capacity of Rayleigh fading
the EC is in decline as md decreases, due to heavier fading, channels under different adaptive transmission and diversity-combining
techniques,” IEEE Trans. Veh. Technol., vol. 48, no. 4, pp. 1165–1181,
which is reflected by the larger power offset (intercept on hor- Jul. 1999.
izontal axis). Moreover, the performance of EC under OPRA [6] A. Laourine, M.-S. Alouini, S. Affes, and A. Stephenne, “On the capac-
is better than that under ORA, because the transmitter adjusts ity of generalized-K fading channels,” IEEE Trans. Wireless Commun.,
vol. 7, no. 7, pp. 2441–2445, Jul. 2008.
its transmit power according to the instantaneous channel state [7] G. Pan, E. Ekici, and Q. Feng, “Capacity analysis of log-normal chan-
in the OPRA case, rather than just using the fix power (aver- nels under various adaptive transmission schemes,” IEEE Commun. Lett.,
age power) to transmit signal in another case. It is also worth vol. 16, no. 3, pp. 346–348, Mar. 2012.
to note that the EC in OPRA and ORA cases converges in the [8] A. J. Goldsmith and P. P. Varaiya, “Capacity of fading channel with
channel side information,” IEEE Trans. Inf. Theory, vol. 43, no. 6,
high SNR region, due to the fact that γ0 → 1 as γ D → ∞, pp. 1986–1992, Nov. 1997.
which means that the transmit power under OPRA case is [9] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
close to the average power. Products, 7th ed. San Diego, CA, USA: Academic, 2007.
[10] F. Yilmaz, O. Kucur, and M.-S. Alouini, “A novel framework on
The EC under TCI, where the cutoff level γ0 is 0.1, is exact average symbol error probabilities of multihop transmission over
smallest except the one under CI in the medium and high amplify-and-forward relay fading channels,” in Proc. 7th ISWCS, York,
SNR region, while this figure for TCI is larger than that under U.K., Sep. 2010, pp. 546–550.
596 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In this letter, a simple engineering scheme for The spatial diversity of multiple feeder links in a Smart
achieving optimal dynamic capacity allocation in geostationary Gateway Diversity (SGD) concept, for resource management
and medium Earth orbit multi-beam high throughput satellite in HTS systems, similar to this letter, has recently been applied
systems is presented. Exploiting smart gateway diversity setup
in [6] and [7] considering system’s capacity losses. In [6]
and considering users’ requested and gateways’ offered capaci-
ties, we propose a theoretically optimal, based on Monge arrays, three different allocation algorithms, namely rate matching,
capacity allocation scheme for minimization of both system’s load balance and fairness method are examined under a com-
capacity losses and rate matching performance metrics. This binatorial solution framework regarding the number of time
scheme has low complexity considering the appropriate sorting slots for GW-UE beam pairs’ connection. Moreover, in [7] a
of system’s offered and requested capacities. Finally, simulation GW-UE allocation scheme based on the deferred acceptance
results also confirm its optimality by presenting its identical per- algorithm from matching theory is proposed. Finally, in satel-
formance with the much more complicated and time consuming
lite downlinks, in [8] authors apply the Lagrangian approach
exhaustive mechanism.
and an optimal multi-beam power and spot-beam allocation
Index Terms—High throughput satellite systems, dynamic are presented.
resource allocation, Monge arrays, optimization. In this letter we investigate the optimal dynamic alloca-
tion of capacity to the user data demands by Medium Earth
I. I NTRODUCTION and Geostationary Orbit, i.e., MEO and GEO, HTS systems
applying the SGD framework. Moreover, we consider the
OWARDS the upcoming 5G networks, the combination
T of huge traffic demand with higher users’ quality of ser-
vice (QoS) requirements and the lack of available spectrum is
GWs’ offered capacities, originated by spatial-temporal total
atmospheric attenuation realistic MEO and GEO stochastic
dynamic channels presented in [9] and [10] respectively and
very challenging for the designers of wireless communication the UEs’ requested capacities and propose a very low complex-
systems. The demanding wireless environment is also stated ity optimal, scalable, dynamic resource allocation mechanism
by Cisco reporting that in 2021 Wi-Fi and mobile devices will between the GWs and UEs in terms of minimizing two dif-
account for 63% of IP traffic and the number of devices con- ferent performance metrics, i.e., the important objective of
nected to IP networks will be three times as high as the global system’s capacity losses and the rate matching cited in [6]. The
population [1]. optimality of the proposed resource allocation scheme is the-
Satellite networks constitute a crucial pillar to assist 5G oretically originated by the fact that both performance metrics
terrestrial infrastructure to satisfy the aforementioned chal- constitute Monge arrays [11], [12]. Finally, this optimality is
lenges [2]. Specifically multi-beam, where different user termi- confirmed in numerical results by the identical performance of
nals or user equipment (UE) beams, called also as UEs below, proposed mechanism with the corresponding time consuming
are served by different gateways (GWs), high throughput satel- exhaustive scheme.
lite (HTS) systems, that can achieve values of Terabit/s [3], For the rest of this letter, Section II presents the system
can satisfy the demanding high QoS users’ requirements. The model, dynamic capacity allocation problems and proposed
satellite systems exploit in the satellite-UE beam link, called optimal matching algorithm. In Section III, numerical results
as user link or downlink, the Ka Band (20/30GHz) and in of the performance of the proposed matching algorithm are
the gateway-satellite link, called as feeder link, the Q/V Band given and in Section IV conclusions are depicted.
(40/50GHz). The signal in these links is mainly deteriorated by
the atmospheric phenomena [4], including rain, clouds, gases
and tropospheric turbulence. The spatial diversity is the most II. O PTIMAL DYNAMIC C APACITY
efficient fade mitigation technique [5] concerning the usage A LLOCATION M ECHANISM
of multiple propagation paths for increasing the availability In this letter, we consider an HTS system with M GWs
of the satellite network [4] and can be implemented either in that can serve N UEs as presented in Fig. 1. The satellite is
the side of feeder links (uplink transmission) or as downlink considered non-regenerative and in a frame period, contain-
reception diversity [4]. ing a number of time-slots, the M-active time multiplexing
SGD context is investigated. In a time-slot basis each UE can
Manuscript received October 17, 2018; accepted November 7, 2018. Date
of publication November 16, 2018; date of current version April 9, 2019. The be served by one GW and each GW serves either only one
associate editor coordinating the review of this paper and approving it for pub- UE assuming M = N, or many UEs simultaneously in the
lication was C. Shen. (Corresponding author: Athanasios D. Panagopoulos.) case M < N and frequency and time multiplexing in a frame
The authors are with the School of Electrical and Computer Engineering,
National Technical University of Athens, 15780 Athens, Greece (e-mail:
period are considered. In the former case each GWk , with
aroumeliot@mail.ntua.gr; harkour@mail.ntua.gr; thpanag@ece.ntua.gr). k = 1, . . . , M , serves qk = 1 UE, while in the latter case
Digital Object Identifier 10.1109/LWC.2018.2881693 qk > 1 UEs are served simultaneously by GWk satisfying the
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ROUMELIOTIS et al.: OPTIMAL DYNAMIC CAPACITY ALLOCATION FOR HIGH THROUGHPUT SATELLITE COMMUNICATIONS SYSTEMS 597
N
pair, (b) N m m
j =1 xij = 1, ∀i and (c) i=1 xij = 1, ∀j declar-
ing (b)/(c) that each GW/UE is paired with one UE/GW
respectively and index m is L or RM representing problems
(1) or (2) respectively.
N
N
min
m
cijm xijm , s.t. (a), (b), (c), (3)
X
i=1 j =1
Fig. 3. CCDF of rate matching as the number of UEs increases for Sorting
Fig. 2. CCDF of losses as the number of UEs increases for Sorting and and Exhaustive mechanisms in: (a) MEO and (b) GEO HTS systems.
Exhaustive mechanisms in: (a) MEO and (b) GEO HTS systems.
RM
cij = RCj − OCi = 2 max RCj , OCi − RCj − OCi , (24) Propag., vol. 60, no. 11, pp. 5400–5411, Nov. 2012.
[11] R. E. Burkard, B. Klinz, and R. Rudolf, “Perspectives of Monge prop-
max RCj , OCi = max RCj − OCi , 0 + OCi . (25) erties in optimization,” Discrete Appl. Math., vol. 70, no. 2, pp. 95–161,
Sep. 1996.
Considering a same approach as before and the same sorting [12] P. Brucker, Scheduling Algorithms, 5th ed. Berlin, Germany:
as in (7), (8) the array CRM is a Monge array if and only if Springer-Verlag, 2007.
600 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In this letter, we propose a learning-based wireless constraint, considering that only the imperfect channel state
powered secure transmission, in which a source utilizes energy information (CSI) are available at the legitimate system. In [4],
harvested from a power beacon to communicate with a legitimate the use of artificial noise (AN) was considered to enhance
receiver, in the presence of an eavesdropper. In order to confuse
the eavesdropper, we assume that the source transmits the arti- physical layer security in wireless powered systems. The joint
ficial noise signals, in addition to the information signals. We design of the power allocation for the AN signals as well
first characterize the effective secrecy throughput of our system, as the time fraction between the power transfer (PT) phase
showing its dependence on the transmission parameters, includ- and the information transfer (IT) phase under the harvest-then-
ing the fraction of time allocated for wireless power transfer, the transmit protocol was examined in [5]. However, these works
fraction of power allocated to the information signals, as well as
the wiretap code rates. We then leverage the deep feedforward often require complicated algorithms, which may lead to large
neural network to learn how the optimal transmission parame- latency impact in real-world deployments.
ters that jointly maximize the effective secrecy throughput can In this letter, we propose a learning-based wireless pow-
be obtained. Through numerical results, we demonstrate that ered secure transmission, in which we exploit the potential of
our learning-based scheme can achieve almost the same secrecy machine learning in rapidly configuring the wireless powered
performance as the optimal solution obtained from the exhaustive
search, while requiring much less computational complexity. systems so as to maximize the effective secrecy throughput
(EST). In fact, machine learning has been shown to be effective
Index Terms—Wireless power transfer, artificial noise, physical in many wireless network applications with secrecy con-
layer security, deep feedforward neural network.
straints, such as anti-jamming [6] and secure transmit antenna
selection [7]. Compared to these works (e.g., [6] and [7]), our
I. I NTRODUCTION contributions are summarized as follows. First, we focus on
IRELESS power transfer has been envisaged as a the scenario of the wireless powered secure transmission, and
W promising solution to fulfill the ever-increasing demand
for energy in the fifth-generation (5G) and beyond wireless
characterize the impacts of various transmission parameters on
the EST. Second, we show how the deep feedforward neural
networks, since it can harvest energy from the radio frequency network (DFNN) can be utilized to efficiently determine the
signals without relying on the location or the climate [1]. On optimal transmission parameters that maximize the EST.
the other hand, security is another important issue in wireless
communications, due to the broadcasting nature of wire-
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
less medium. In particular, the decentralized modern wireless
networks have introduced significant challenges to traditional We consider a wireless powered secure communication
key-based cryptographic techniques, such as key generation, system, which consists of a power beacon (PB), a source
distribution, and management. To tackle this problem, physi- (Alice), a legitimate receiver (Bob), and an eavesdropper
cal layer security [2] has been proposed as an alternative for (Eve). In this system, Alice utilizes energy collected from PB
cryptographic techniques, since it can achieve the information- to communicate with Bob in the presence of Eve. We consider
theoretic secure communications without using secret keys. that Alice is equipped with Na antennas, while PB, Bob, and
Recently, wireless powered secure communication has also Eve are all equipped with a single antenna. We assume that
been receiving an increasing research attention [3]–[5]. In [3], the CSI between PB and Alice and the CSI between Alice
robust beamforming schemes were proposed to maximize the and Bob are perfectly known at PB and Alice, respectively,
throughput of wireless powered systems with the secrecy while only statistical information on Eve’s channel is available
to the legitimate nodes. We also assume that all the chan-
Manuscript received October 3, 2018; accepted November 14, 2018. Date nels are subject to identical and independent distributed (i.i.d)
of publication November 19, 2018; date of current version April 9, 2019. This Rayleigh fading. We denote hij and dij as the channel and dis-
work was supported in part by the China Scholarship Council ([2017] 3109)
and in part by the National Natural Science Foundation of China under
tance between node i, i ∈ {p, a}, and node j, j ∈ {a, b, e},
Grant 61471037, Grant 61771048, and Grant 61201181. The associate editor respectively.
coordinating the review of this paper and approving it for publication was
L. P. Natarajan. (Corresponding author: Chenxi Liu.)
D. He and H. Wang are with the School of Information and A. Wireless Powered Secure Transmission
Electronics, Beijing Institute of Technology, Beijing 100081, China
(e-mail: hdxbit@bit.edu.cn; wanghua@bit.edu.cn). We now detail the wireless powered secure transmission
C. Liu and T. Q. S. Quek are with the Information Systems Technology considered in this letter. The transmission between Alice and
and Design Pillar, Singapore University of Technology and Design, Singapore
487372 (e-mail: chenxi_liu@sutd.edu.sg; tonyquek@sutd.edu.sg). Bob consists of two phases, namely, the PT phase and the IT
Digital Object Identifier 10.1109/LWC.2018.2881976 phase. In the PT phase, Alice harvests energy from PB, and
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HE et al.: LEARNING-BASED WIRELESS POWERED SECURE TRANSMISSION 601
then utilizes all the energy harvested in the PT phase to com- As such, with the aid of [8], the secrecy outage probability of
municate with Bob in the IT phase. We denote τ as the fraction our system can be derived as
of time allocated to the PT phase. As such, the available
transmit power at Alice in the IT phase can be expressed as Pso (τ, α, Re ) = Pr(γe > κe )
(1 − α)κe −(Na −1) − αϕγ
κe
Pa = ϕP̂ , (1) = 1+ e e , (6)
α(Na − 1)
τ , P̂ = ξPd −η 2
where ϕ = (1−τ ) pa |hpa | , 0 < ξ < 1 denotes
the energy harvest efficiency at Alice, P denotes the transmit where κe = 2Re − 1.
power at PB, and η denotes the path loss exponent. In order to evaluate the secrecy performance of our system,
In the IT phase, we assume that Alice transmits the AN sig- we adopt the modified EST [8], [9] as the performance metric,
nals, in addition to the information signals, in order to confuse given by
the eavesdropper. As such, we express the transmitted signal
at Alice as Ts (τ, α, Re ) = (1 − τ )(Rb − Re )(1 − Pso (τ, α, Re )). (7)
xs = wts + Gtsan , (2) We can see that the EST in (7) is a function of the fraction
of the time allocated to the PT phase τ , the fraction of power
where w and G denote the Na × 1 beamforming vector used allocated to the information signals α, as well as the redun-
to transmit the information signal ts and the Na × (Na − 1) dancy rate of the wiretap code Re . The key goal of this letter is
beamforming matrix used to transmit the AN signal tsan , to find the optimal transmission parameter tuple (τ ∗ , α∗ , Re∗ )
respectively. In order to degrade the quality of the received that achieves the maximum EST Ts∗ per transmission block.
signal at Eve, while maintaining the quality of the received Mathematically, this problem can be formulated as
h†
signal at Alice, we choose w = hab and G as the projection
ab
matrix onto the null space of hab , respectively [8], [9]. As max Ts (τ, α, Re ), (8a)
τ,α,Re
such, we have hab G = 0, and w and the columns of G form s.t. 0 < τ < 1, 0 < α ≤ 1, 0 < Re ≤ Rb . (8b)
an orthonormal basis. In addition, we denote α as the fraction
of power allocated to the information signals. As such, we However, the maximization problem in (8) is non-convex.
(1−α)P
have E[|ts |2 ] = αPa and E[tsan t†san ] = Na −1 a INa −1 . Therefore, it is difficult to analytically obtain the optimal trans-
Based on (1) and (2), we express the received signal-to- mission parameter tuple (τ ∗ , α∗ , Re∗ ). Traditionally, (8) can
interference-plus-noise ratios (SINRs) at Bob and Eve in the be solved numerically via an exhaustive search, but requiring
IT phase, respectively, as a high computational complexity. To tackle this problem, we
propose a learning-based scheme that is capable of solving (8)
γb = αϕγ b |hab |2 , (3) in a far more efficient way compared to the exhaustive search.1
αϕγ e |hae w|2
γe = 1−α , (4)
2
Na −1 ϕγ e hae G + 1
III. P ROPOSED L EARNING -BASED S CHEME
−η 2 −η 2 2
where γ b = P̂ dab /σb , γ e
= P̂ dae /σe , σb and σe2
denote the In this section, we present details on our proposed learning-
variance of the additive white Gaussian noise at Bob and Eve, based scheme that solves (8) efficiently.2 Specifically, we
respectively. According to (3) and (4), the secrecy capacity of utilize the DFNN to learn the nonlinear mapping from hab
our system is expressed as to the optimal transmission parameter tuple (τ ∗ , α∗ , Re∗ ) that
achieves the maximum EST, which can be expressed as
Cs = {Cb − Ce }+ , (5)
(τ ∗ , α∗ , Re∗ ) = f ∗ (hab ). (9)
where {·}+ denotes max{0, ·}, Cb = log2 (1 + γb ) and Ce =
log2 (1 + γe ) denote the capacity of Bob’s channel and the We note that the DFNN has shown to be effective in approx-
capacity of Eve’s channel, respectively. imating any measurable function to any desired degree of
accuracy [11], and thus is an appropriate method to solve (8).3
B. Problem Formulation
1 Note that in practice the choice of transmission parameters is finite.
In this letter, we employ the well-known Wyner’s wiretap
We will show in Section IV that, even in such scenarios, our proposed
code [2] with the parameter pair (Rb , Re ) to perform secure learning-based scheme still outperforms the exhaustive search in terms of
transmissions, where Rb denotes the transmission rate of the the computational complexity.
wiretap code and Re denotes the redundancy rate of the wire- 2 Besides wireless powered communication systems, our learning-based
scheme can also be applied to enhance the security of various communication
tap code, representing the cost of preventing eavesdropping. systems. For example, in secure communication systems with the coopera-
When Rb > Cb , the transmitted signal from Alice cannot be tive jammers (as in [10]), the proposed learning-based scheme can be utilized
reliably decoded at Bob, and the transmission outage occurs. to determine the optimal power allocation at the cooperative jammers that
maximizes the secrecy rate.
When Re ≤ Ce , the information on the transmitted signal is 3 We note that the curve fitting method can also be applied to approximate
leaked to Eve, and the secrecy outage occurs. Since we assume the mapping from hab to (τ ∗ , α∗ , Re∗ ). However, it may not achieve the
that hab is perfectly known at Alice, we can set Rb = Cb . same degree of accuracy as our proposed learning-based scheme.
602 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
C OMPLEXITIES AND RUNNING T IMES OF D IFFERENT S CHEMES
where βi ∼ CN (0, σi2 ) is the i.i.d varying-component of techniques, (5) and (6) can be expressed as
the i th hop; and ρi = J0 (2πfc νi Ts /c) is its correlation- √
PSK a1 a2 ∞ −a2 x 2
parameter where J0 (.) is the zeroth-order Bessel function of P e (j ) = √ e Fγ (x 2 ) dx (7)
the first kind, νi is the relative speed between the i th hop’s π 0
√
communicating nodes, Ts is the symbol duration, fc is the
b1 √b2 ∞
M
−1
2
carrier frequency, and c is the speed of light [9]. According M−QAM 2
Pe (j ) = √ e −b2 x Fγ (x 2 ) dx . (8)
to the pilot-assisted estimation method [11], we assume that, π 0
for each frame, the system’s receivers estimate their fading m=0
coefficients only over the 1st signaling period, which are ĥ1 (1) We continue the system’s SER analyses using (7), where the
and ĥ2 (1). In addition, we assume imperfect CSI estimation, results could be then extended directly to the M-QAM case.
i.e., ĥi (1) = hi (1) + i (1) where i (1) ∼ CN (0, σe2i ) is the
estimation error [11]. Over the j th signaling period, S trans- A. SER Using Literature Method
mits the modulated symbol x(j) with energy Es to R, which Based on the approximate method commonly adopted in lit-
in turn amplifies the received signal by the CSIN-assisted erature to evaluated SER performance of dual-hop AF systems,
amplification gain [9, eq. (6)] like in [4]–[8], we can approximate the SNR in (3) as
γ γ1 (j )γ2 (j )
α(j ) = 2(j −1) 2(j −1) 2
(2) γeq (j ) ≈ γ̂eq (j ) = (9)
ρ1 2
|ĥ1 (1)| + (1 − ρ1 )σ1 γ + 1 γ1 (j ) + γ2 (j )
where γ = Es /No and No is the white noise power of the which has cdf obtained from (4) by replacing c(j ) = 0 as
first hop. After that, R transmits the amplified signal towards
1
1
D, which uses ĥ1 (1) and ĥ2 (1) in the detection process [9]. Fγ̂eq (j ) (γ) = 1 − 2γd(j ) e
2 −z(j ) γ
K1 2γd(j ) .
2
(10)
Based on this, the system’s equivalent end-to-end SNR over
the j th signaling period can be obtained at D as [9, eq. (15)]
By plugging (10) into (7) along with making use of [13,
γ1 (j )γ2 (j ) eq. (6.621.3)], and under the assumption of equiprobabale
γeq (j ) = (3)
γ1 (j ) + γ2 (j ) + c(j ) symbols in each frame, the PSK per-frame-average SER can
2(j −1)
be approximately obtained by
ρi γ|ĥi (1)|2
where γi (j ) = ωi (j )
, ωi (j ) = 1 + [(1 − 1
J 6π a12 a2 d(j )
A(j ) − 2d(j2 )
2(j −1) 2 2(j −1) 2 φ(j ) PSK 0.5
ρi )σi + ρi σei ]γ, and c(j ) = ω (j )ω (j ) with Pe
J j =1
a1 − 1 F 2.5, 1.5; 2; 1
1 2
2(j −1) 2 2 2(j −1) 2 2(j −1) 2 (A(j ) + 2d(j2 ) )2.5 A(j ) + 2d(j2 )
φ(j ) = (1 − ρ1 )σ1 γ [(1 − ρ2 )σ2 + ρ2 σe2 +
(11)
1 ] + (1 − ρ2(j −1) )σ 2 γ[ρ2(j −1) σ 2 γ + 1] + ρ2(j −1) σ 2 γ + 1 +
γ 2 2 2 e1 2 e2
2(j −1) 2 2(j −1) 2 2 where A(j ) = a2 + z(j ) , and F(a, b; c; z) is the Gauss’ hyper-
ρ1 σe1 ρ2 σe2 γ . If the nodes are static and the esti-
geometric function [13, eq. (9.100)]. Although (11) is obtained
mation is perfect, c(j ) = 1. As γi (j ) is exponential R.V with
ϑ (j ) 2(j −1) 2
based on the literature method, it is new and generalizes its
mean γ i (j ) = ωi (j ) , where ϑi (j ) = ρi σi γ, the exact literature counterparts under the impacts of nodes’ motion and
i
cdf of γeq (j ) can be obtained as incorrect CSI estimates. Unde these impacts, the system’s SER
−z(j ) γ
suffers from error floors. By taking the limit of (11) as γ → ∞,
Fγ (γ) = 1 − 2 (γ 2 + c(j ) γ)d(j ) e K1 2 (γ 2 + c(j ) γ)d(j )
eq (j ) these floors can be given as
(4)
J
√
floor a1 3πa1 a2 e1 e2 a2 + λ2
where K1 (·) is the 1st -order
modified-Bessel function of the Pe =
2
−
f1 f2 (a2 + λ1 )2.5
F 2.5, 1.5; 2;
a2 + λ1
(12)
2nd -type, d(j ) = 1/(γ 1 (j )γ 2 (j )), and z(j ) = d(j ) [γ 1 (j ) + j =1
∞
per-frame-average SER can be given by −A(j ) u k ++ 12 u
× e u ln 1 + du . (21)
J
√ 0 c(j )
a1 1 a1 a2
Pe = − √
2 J π I6
j =1
Upon solving the integral in (19) using [13, eq. (3.371)], we
∞ x 4 + c(j ) x 2 x 4 + c(j ) x 2
−A(j ) x 2
× 2e K1 2 dx . (14) have
0 γ 1 (j )γ 2 (j ) γ 1 (j )γ 2 (j )
+1 C k +1 d k +1 c k +1− 2(k + ) + 1 !!
√ k (j ) (j )
I1
I3 = π (22)
By denoting u = x 2 , we can write I1 as =0 2k + +1 Ak(j+ +1.5
)
=f (u) where (·)!! denotes the double-factorial operator. Alternatively
k ++1
I1 =
∞
e −A(j ) u 2
d(j ) (u + c(j ) )K1 2 d(j ) (u + c(j ) u) du. (15) writing (2(k + ) + 1)!! = 2 √π Γ(k + + 32 ) yields I3 to be
0 ultimately given as
We are unaware of a closed-form solution to I1 . However, k
+1 C k +1 d(jk +1 c k +1− Γ(k + + 32 )
exploiting the expansion of Kν (·) [13, eq. (8.446)] gives I3 =
) (j )
. (23)
k + +1.5
K
=∞
(d(j ) (u 2 + c(j ) u))k + 2
1
=0 A(j )
1
K1 f (u) = +
2 d(j ) (u 2 + c(j ) u) k !(k + 1)! Solving the integral in (20) using [13, eq. (4.352.1)] yields
k =0
1 k
+1 Ck +1 d(j
k +1 k +1−
c(j ) Γ(k + + 32 ) ψ(k + + 32 ) − ln A(j )
× ln d(j ) (u 2 + c(j ) u) − ψ(k + 1) + ψ(k + 2) . (16) I4 =
)
.
2 k ++ 3
=0 A(j ) 2
J
∞ √ k+1 k +1 1
a1 1 a12 a2 ϑ1 (j )ϑ2 (j ) a1 a2 C φ(j )k +1− (ω1 (j )ω2 (j )) (ϑ1 (j )ϑ2 (j )) + 2
Pe = − + √ 3 V(k , )
2 J 4χ(j ) 2 πk !(k + 1)! (χ(j ))k + + 2
j =1 k =0 =0
φ(j ) ω1 (j )ω2 (j )ϑ1 (j )ϑ2 (j )
1, 3 −( 21 +k + ),1,1
+ ln Γ(k + + 1.5) + G 3, 2 F(j ) , with F(j ) = (13)
χ(j ) 1,0 φ(j )χ(j )
−(n+ 12 ),1,1 πγ(n + 1.5, −1/F (j )) Γ(n + 1.5 + ε)
G 13 ,3
,2 F (j ) = √ + Γ(−ε) Γ(n + 1.5) − 1 F 1 (−ε; −(n + ε + 0.5); 1/F (j )) (26)
1,0 −1 (1/F (j ))ε
KHATTABI: NEW ANALYTICAL APPROACH IN SER EVALUATION OF CSIN-ASSISTED AF DUAL-HOP WIRELESS SYSTEMS 607
Fig. 1. M-PSK SER Vs γ with 65 mph speed, σ12 = σ22 = 1, and σe2 = Fig. 2. M-QAM SER Vs γ with static nodes, σ12 = σ22 = 1, and σe2 = 0.
0.001.
is clear that, similarly, the new exact results match the simu-
By expanding the right hand side of (27) using lation and outperform the old results. It is notable from Fig. 2
[13, eq. (9.304)] and then applying the facts that that because the nodes are static and the estimation is perfect,
2 F2 (a, 0; b, c; z ) = 1, 2 F2 (a, b; a, c; z ) = 1 F1 (b; c; z ), the SER plots behave normally and do not suffer from floors.
π
Γ(−x ) = − sin(πx )Γ(1+x , Γ(1 + x ) = x Γ(x ), and finally
)
using [13, eq. (8.351.2)] we obtain (26). V. C ONCLUSION
Computing (13) by adopting Corollary 1 reduces its running
time to up to 91% in case of 16-PSK as an example. In this letter, a new analytical arrangement has been
2) Proposed Approximate SER: proposed to derive novel exact as well as approximate expres-
sions for the SER performance of a Rayleigh fading CSIN-
Corollary 2: The per-frame-average SER can be evaluated assisted based dual-hop AF cooperative system with mobile
using the following closed-form approximate expression nodes and imperfect CSI estimates. The exact expression
J
√
involves a quickly convergent infinite summation, while the
a1 a2 a2 d(j ) c(j ) ln d(j ) /A(j )
Pe 1− − approximate one is in closed form and shows very satisfying
2J j =1 4A(j ) 2 A1.5
(j )
results. As a future work, this proposed arrangement could be
0.114c(j ) 3 0.78 + ln d (j ) /A (j ) easily extended considering other network models.
+ + . (28)
A1.5
(j )
2 A2.5
(j )
R EFERENCES
Proof: By taking only the 1st term (k = 0) from the infi- [1] J. N. Laneman, D. N. C. Tse, and G. W. Wornell, “Cooperative diversity
nite sum in (17), along with letting ln(d(j ) (u + c(j ) ) ≈ in wireless networks: Efficient protocols and outage behavior,” IEEE
ln(d(j ) u/c(j ) ) in I5 , I1 can be approximately obtained as Trans. Inf. Theory, vol. 50, no. 12, pp. 3062–3080, Dec. 2004.
[2] M. O. Hasna and M.-S. Alouini, “A performance study of dual-hop
∞ transmissions with fixed gain relays,” IEEE Trans. Wireless Commun.,
π d(j )
I1 + ln(d(j ) ) − ψ(1) − ψ(2) e −A(j ) u vol. 3, no. 6, pp. 1963–1968, Nov. 2004.
4A(j ) 2 0 [3] M. O. Hasna and M.-S. Alouini, “End-to-end performance of transmis-
∞ ∞
sion systems with relays over Rayleigh-fading channels,” IEEE Trans.
3 √ √
× u 2 du + c(j ) e −A(j ) u udu + 2 e −A(j ) u c(j ) u Wireless Commun., vol. 2, no. 6, pp. 1126–1131, Nov. 2003.
0 0
[4] Y. Xiao et al., “Forwarding strategy selection in dual-hop NOMA relay-
3 ing systems,” IEEE Commun. Lett., vol. 22, no. 8, pp. 1644–1647,
× ln(u) + u 2 ln(u) du . (29)
Aug. 2018.
[5] L. Yang, M. O. Hasna, and I. S. Ansari, “Unified performance analy-
Solving the first two integrals using [13, eq. (3.371)] and sis for multiuser mixed η-μ and M - distribution dual-hop RF/FSO
the last one using [13, eq. (4.352.1)], and then substitut- systems,” IEEE Trans. Commun., vol. 65, no. 8, pp. 3601–3613,
ing the result into (14), along with some simplifications, Aug. 2017.
gives (28). [6] N. S. Ferdinand and N. Rajatheva, “Unified performance analysis of two-
hop amplify-and-forward relay systems with antenna correlation,” IEEE
Trans. Wireless Commun., vol. 10, no. 9, pp. 3002–3011, Sep. 2011.
IV. N UMERICAL R ESULTS [7] D. Senaratne and C. Tellambura, “Unified exact performance analysis
of two-hop amplify-and-forward relaying in Nakagami fading,” IEEE
Fig. 1 shows that, as compared to exact simulation, SER Trans. Veh. Technol., vol. 59, no. 3, pp. 1529–1534, Mar. 2010.
results obtained using (11) (the old method) show unaccept- [8] G. Farhadi and N. C. Beaulieu, “A general framework for symbol error
able tightness especially for low modulation orders and over probability analysis of wireless systems and its application in amplify-
the low and medium SNR regions; while the new exact results and-forward multihop relaying,” IEEE Trans. Veh. Technol., vol. 59,
no. 3, pp. 1505–1511, Mar. 2010.
computed using (13) perfectly match with simulation results. [9] Y. M. Khattabi and M. M. Matalgah, “Performance analysis of multiple-
The best performance improvement achieved by the proposed relay AF cooperative systems over Rayleigh time-selective fading
approach over the old approximate is about 2.1 dB and 1.2 dB channels with imperfect channel estimation,” IEEE Trans. Veh. Technol.,
over the low and medium SNR regions, respectively. The num- vol. 65, no. 1, pp. 427–434, Jan. 2016.
ber of terms involving the calculation of the infinite summation [10] N. Varshney, A. K. Jagannatham, and P. K. Varshney, “Cognitive
MIMO-RF/FSO cooperative relay communication with mobile nodes
in (13) are labeled on the figure, which indicate a comfortable and imperfect channel state information,” IEEE Trans. Cogn. Commun.
convergence rate, especially, for the low modulation orders. Netw., vol. 4, no. 3, pp. 544–555, Sep. 2018.
Further, the new approximate results obtained using (28) show [11] X. Wu, H. Claussen, M. Di Renzo, and H. Haas, “Channel estima-
satisfying accuracy level and outperform (11) for the low mod- tion for spatial modulation,” IEEE Trans. Commun., vol. 62, no. 12,
ulation orders. We can also observe from Fig. 1 that under the pp. 4362–4372, Dec. 2014.
effects of nodes mobility and imperfect CSI estimation, the [12] M. K. Simon and M.-S. Alouini, Digital Communication Over Fading
Channels, vol. 95. Hoboken, NJ, USA: Wiley, 2005.
overall system error performance is degraded and experiences [13] I. Gradshteyn and I. Ryzhik, Table of Integrals, Series and Products,
irreducible floors (plotted using (12)). Fig. 2, shows SER plots 5th ed., A. Jeffrey and D. Zwillinger, Eds. New York, NY, USA:
for the same system considering M-QAM modulation where it Academic, 2007.
608 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Device-free localization (DFL) plays an increasingly theory [3] into DFL, greatly reducing the required measure-
dominant role in some security and military applications, aim- ment used to promise a reliable localization accuracy. In this,
ing at locating targets without carrying any extra devices. Most there have been many types of researches concentrated in CS-
research utilize received signal strength, causing poor accuracy
due to insufficient information. In this letter, we propose DF-
based DFL, coming down to dictionary refinement [5], model
CSPG, a new DFL algorithm using compressive sensing (CS) designing [6], human effort saving [7] and so on. Most of
theory and potential game (PG) approach to exploit channel sta- the existings are based on Received Signal Strength (RSS)
tion information (CSI). It is the first time that the PG approach is for it’s easy to get. But, RSS is always too noisy, which
applied in DFL. At first, the CS-based localization model is built may cause inaccuracy in localization. As an alternative, the
by exploiting CSI and then investigated from a game-theoretic Channel Station Information (CSI) consists of amplitude and
perspective in which the subcarriers are assumed as players.
Then, we prove that the proposed DFL game is an exact PG. phase information from multiple subcarriers. And it is con-
Based on these, the best response theory is utilized to obtain the venient to be measured as long as the system is based on
equilibriums. Finally, simulation results show that DF-CSPG can Orthogonal Frequency Division Multiplexing (OFDM).
achieve more accurate and robust performance. Above all, CS can be applied to reduce the number of
Index Terms—Device-free localization, compressive sensing, wireless devices required to promise the localization accu-
potential game, nash equilibrium, best response. racy while CSI can provide more localization information.
So, the localization algorithm DF-CSPG, based on CS theory
and Potential Game (PG) approach, is proposed. In fact, the
PG approach has been used in some occasions for its advan-
I. I NTRODUCTION tages [8]. However, this is the first time that PG approach is
S AN emerging technique, Device-Free Localization applied in DFL. Firstly, we build the CS-based localization
A (DFL) could utilize the everywhere spread wireless sig-
nal to estimate the locations without requiring any devices to
model using CSI and design it into a nash game. Secondly,
the DFL game is proved to be an Exact Potential Game
be attached to targets [1]. In addition, it is the fundamental of (EPG) with its potential function is the global optimization.
device-free wireless sensing technique, making a great effect According to the definition in [9], we can optimize the utility
on estimating of the motion, activity, and gesture of person. function of DFL game instead of global optimization which is
Nowadays, DFL has drawn considerable attention and played complex. Thirdly, as one kind of the decision rules to achieve
a significant role in security monitoring and emergency rescue. pure Nash Equilibrium (NE) points, Best Response (BR) the-
In DFL, a person within the deployment area influences ory is introduced. Fourthly, CS theory is utilized to optimize
the wireless signal in a predictable way, which renders it fea- the problem obtained from the BR dynamics. Finally, sev-
sible to sense the target location by analysing the wireless eral simulations have been formulated to verify our proposed
patterns and characteristics [2]. Besides, the position accuracy algorithm.
is greatly infected by the quantity of wireless links distributed
in the monitoring area. More wireless links are deployed, II. M ODEL D ESIGNING
more accurate localization will be obtained. However, most
A. CS-Based Localization Model Using CSI
of the wireless devices are energy-constrained, heavily limit-
ing the scale and lifetime of DFL system. Considering the This letter focus on the CS-based multi-target localization
issues, Wang et al. [4] applied Compressive Sensing (CS) using CSI. It can serve both amplitude and phase information
from multiple subcarriers, in which the amplitude information
Manuscript received October 7, 2018; revised November 4, 2018; accepted could be convert into RSS through the research in [10]. As is
November 29, 2018. Date of publication December 5, 2018; date of current
version April 9, 2019. This work was supported in part by the National Natural
illustrated in Fig. 1, the l × l two dimensional area is divided
Science Foundation of China under Grant 61871400 and Grant 61571463, and into N grids and all K targets are located at the grid points.
in part by the Natural Science Foundation of Jiangsu Province under Grant By exploiting the inherent sparsity of localization issue, the
BK20171401. The associate editor coordinating the review of this paper and target location can be represented as a K sparse vector:
approving it for publication was Y. Shen. (Corresponding author: Yan Guo.)
S. Yang, Y. Guo, and N. Li are with the College of Communications
Engineering, Army Engineering University of PLA, Nanjing 210000, China w = [0, 1, 0, . . . , 0, 1]T , (1)
(e-mail: guoyan_1029@sina.com).
D. Fang is with the School of Electronic Engineering and Optoelectronic in which the nonzero coordinates of location vector corre-
Technology, Nanjing University of Science and Technology, Nanjing 210094,
China. sponding to the target locations. Additionally, M wireless links
Digital Object Identifier 10.1109/LWC.2018.2885052 are constructed by 2M uniformly employed wireless nodes,
2162-2345 c 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
YANG et al.: DF-CSPG: PG APPROACH FOR DFL EXPLOITING JOINT SPARSITY 609
So, we design the potential function F using the global Algorithm 1 Algorithm of DF-CSPG
utility function U0 which referencing to the global optimal Require: yi , Ai , σ, ξ, n = 0, τ = 0, wi (0) = 0
objective: 1: Wireless links send the measured CSI to data fusion center.
L 2: while τ < 10 do
F (w1 , w2 , . . . , wL ) = Ui . (12) 3: Update the iteration number n = n + 1.
i=1 4: for i = 1 to L do
L
1−ξ
Then, we certify that the DFL game Γ is an EPG whose 5: Update bi (n) = L−1 Al T Al w
i (n − 1) +
potential function is F. The details are as follows. When player l=1,l=i
k transfers its strategy from wk to wk∗ , the change of potential ξAi T yi .
L
1−ξ
function F is calculated as: 6: Update Di (n) = L−1 Al T Al + ξAi T Ai .
F (wk∗ , w−k ) − F (wk , w−k ) = Uk (wk∗ , w−k ) l=1,l=i
⎧ ⎫ 7: i (n) by CS.
Estimate the location vector w
⎨ L ⎬
− Uk (wk , w−k ) + ∗
[Ui (wi , w−i ) − Ui (wi , w−i )] , (13) 8: end for
⎩ ⎭ L
i=1,i=k 9: if L1 i (n − 1)2 < σ then
wi (n) − w
where the change of potential function includes two parts: one i=1
10: τ =τ +1
is the change of utility function of player k, and the other is the 11: else
sum of the changes of other players’ utility function. In DFL 12: τ =0
games, it is considered that when player k deviate its strategy, 13: end if
player i (i = k) remains in its primary station, making its 14: end while
utility function Ui unchanged. As a result, the element in {} L
of (13) is zero. According to definition 2, the DFL game Γ 15: = L1
Compute the final location vector w i (n)
w
i=1
is proved to be an EPG. So, NESet(Υ) = NESet(Γ), and
we minimize Ui to obtain NE points instead of optimizing F
1−ξ L
which is complex according to Theorem 1.
Let bi = L−1 T T
In game theory, there are several kinds of decision rules to l=1,l=i Al Al wl + ξAi yi and Di =
1−ξ L T T
achieve pure NE points, such as Fictitious Play (FP), Spatial L−1 l=1,l=i Al Al + ξAi Ai , the (14) can be con-
Adaptive Play (SAP), Best Response (BR) and so on. We use verted as:
BR dynamics here.
Lemma 1: For game Γ, the BR dynamics can guarantee the bi = Di wi , (17)
end solution closer to an actual NE in a finite number of steps.
where there might be countless solution of wi . CS reconstruc-
In BR concept, players update their strategy files according
tion algorithm is used here to exploit the sparse solution.
to the best utility in a round-robin. And each time, only one
Based on the above analysing, the new algorithm DF-CSPG
player is chosen to update its strategy while all other players
is summarized as Algorithm 1.
remain unchanged until the criterion is achieved. Based on BR
According to the proposed algorithm, σ = 0.1L is used to
theory, the location vector wi is optimized as follow: control the speed of the convergence and τ = 10 is empirically
wi = arg max Ui (wi , w−i ) set to promise the algorithm has reached the convergence not
wi
⎛ ⎞ on accident.
L
−2(1 − ξ)
= arg min⎝ wl T Al T Al − 2ξyi T Ai ⎠wi
wi L − 1 l=1,l=i IV. S IMULATION R ESULTS
⎛ ⎞ To evaluate the new proposed algorithm, we conduct the
L
1 − ξ
+ wi T ⎝ Al T Al + ξAi T Ai ⎠wi , (14) following simulations. The monitoring scenario is divided into
L − 1 l=1,l=i N = 400 grids and K = 4 targets are randomly located at the
grid points. Besides, M = 40 wireless links are uniformly
where the first-order and second-order derivative of utility
deployed in the monitoring area. To approximate the noise,
function Ui are calculated as follows: the Gaussian White Noise is added, which is represented by
L
∂Ui (wi , w−i ) −2(1 − ξ) Signal-to-Noise Ratio (SNR).
= Al T Al wl
∂wi L−1 We first test the convergence of the new proposed algorithm
l=1,l=i
⎛ ⎞ DF-CSPG. Fig. 2 presents the evolution of the Euclidian dis-
L
T 1 − ξ T T tance of the recovered location vector. As is shown in the
− 2ξAi yi + 2⎝ Al Al + ξAi Ai ⎠wi , (15)
L−1 picture, most of the proposed algorithm DF-CSPG can reach
l=1,l=i
⎛ ⎞ the convergence except the case in which ξ = 0.01. The main
L
∂ 2 Ui (wi , w−i ) 1 − ξ T T reason is that the DF-CSPG with ξ = 0.01 rarely takes the
= 2⎝ Al Al + ξAi Ai ⎠, (16)
∂wi wi T L−1 deviation of personal estimation into consideration. For more
l=1,l=i
details: firstly, the convergence speed is so quick that all the
in which its second-order derivative is easily proved positive algorithms can reach the convergence even when the iteration
semidefinite, meaning that there are pure NE points which times reach 15. Secondly, most algorithms convergence to the
∂Ui (wi ,w−i )
satisfy ∂wi = 0. similar level of Euclidian distance except ξ = 0.01, which
YANG et al.: DF-CSPG: PG APPROACH FOR DFL EXPLOITING JOINT SPARSITY 611
V. C ONCLUSION
In this letter, we investigate the CS-based DFL problem
using CSI. Different from existing algorithms, we solve the
DFL problem from a game-theoretic perspective which states
the DFL as a potential game and each CSI subcarrier as a
player. Accordingly, we prove the DFL game is an EPG.
Following this idea, we develop a new algorithm DF-CSPG,
using CS and BR theory, to find the NE points. At last, sim-
ulations have shown that our methods own better localization
accuracy and robustness.
TABLE I
BASIC I NFORMATION OF 8 S ELECTED C ITIES
TABLE II
P OSITIONS OF THE R IPPLES AND P EAKS IN THE B ETTI C URVES
TABLE III
C ANDIDATE D ISTRIBUTIONS AND THEIR PDF E XPRESSIONS
Fig. 4. Comparisons between the practical PDF and the fitted ones for the
Euler characteristics of BS locations in European cities.
R EFERENCES
Fig. 5. Comparisons between the practical PDF and the fitted ones for the [1] J. Kibiłda, B. Galkin, and L. A. DaSilva, “Modelling multi-operator base
Euler characteristics of BS locations in Asian cities. station deployment patterns in cellular networks,” IEEE Trans. Mobile
Comput., vol. 15, no. 12, pp. 3087–3099, Dec. 2016.
TABLE IV [2] Y. Zhou et al., “Large-scale spatial distribution identification of base
RMSE B ETWEEN E ACH C ANDIDATE D ISTRIBUTION stations in cellular networks,” IEEE Access, vol. 3, pp. 2987–2999, 2016.
AND THE P RACTICAL O NE [3] Z. Zhao, M. Li, R. Li, and Y. Zhou, “Temporal-spatial distribution nature
of traffic and base stations in cellular networks,” IET Commun., vol. 11,
no. 16, pp. 2410–2416, 2017.
[4] Y. Chen, R. Li, Z. Zhao, and H. Zhang, “Study on base station topology
in cellular networks: Take advantage of alpha shapes, Betti numbers,
and Euler characteristics,” arXiv:1808.07356v1, Aug. 2018.
[5] R. V. D. Weygaert et al., Alpha, Betti and the Megaparsec Universe: On
the Topology of the Cosmic Web. Heidelberg, Germany: Springer, 2011.
[6] J. Zhu, S. Rosset, H. Zou, and T. Hastie, “ Multi-class AdaBoost,” Stat.
Interface, vol. 2, no. 3, pp. 349–360, 2006.
[7] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
MIMO for next generation wireless systems,” IEEE Commun. Mag.,
vol. 52, no. 2, pp. 186–195, Feb. 2014.
[8] P. Pranav et al., “The topology of the cosmic Web in terms of persistent
Betti numbers,” Monthly Notices Roy. Astronomical Soc., vol. 4, no. 465,
pp. 4281–4310, Mar. 2017.
[9] V. D. Silva and R. Ghrist, “Coverage in sensor networks via persis-
is evidently affirmed because all the Hurst exponents turn out tent homology,” Algebraic Geometric Topol., vol. 7, no. 1, pp. 339–358,
to be very close to 1. Apr. 2007.
[10] P. Masulli and A. E. P. Villa, “The topology of the directed clique com-
plex as a network invariant,” Springerplus, vol. 5, no. 388, pp. 1–12,
IV. L OG -N ORMAL D ISTRIBUTION OF THE 2016.
E ULER C HARACTERISTICS [11] R. Li, Z. Zhao, Y. Zhong, C. Qi, and H. Zhang, “The stochastic geometry
analyses of cellular networks with α-stable self-similarity,” IEEE Trans.
Euler characteristics can be calculated from Betti numbers Commun., to be published, doi: 10.1109/TCOMM.2018.2883099.
according to the Euler-Poincare Formula [5]. Since a clear [12] X. Ge, X. Tian, Y. Qiu, G. Mao, and T. Han, “Small-cell networks with
fractal coverage characteristics,” IEEE Trans. Commun., vol. 66, no. 11,
heavy-tailed property is demonstrated in the PDFs of the Euler pp. 5457–5469, Nov. 2018.
Characteristics, three classical heavy-tailed statistical distribu- [13] M. Ulm, P. Widhalm, and N. Brändle, “Characterization of mobile
tions and widely-used Poisson distribution are selected as the phone localization errors with OpenCellID data,” in Proc. Int. Conf.
candidates to match the PDFs. The candidates and their PDF Adv. Logistics Transp., Valenciennes, France, May 2015, pp. 100–104.
formulas are given in Table III. [14] C. Yuan, Z. Zhao, R. Li, M. Li, and H. Zhang, “The emergence of
scaling law, fractal patterns and small-world in wireless networks,” IEEE
The real PDF and the fitted ones are presented in Fig. 4 Access, vol. 5, pp. 3121–3130, 2017.
for European cities and in Fig. 5 for Asian ones, respectively. [15] X. Ge et al., “Wireless fractal cellular networks,” IEEE Wireless
Moreover, the root mean square error (RMSE) between the real Commun., vol. 23, no. 5, pp. 110–119, Oct. 2016.
PDF and every candidate is listed in Table IV. It is obvious [16] Y. Hao et al., “Wireless fractal ultra-dense cellular networks,” Sensors,
vol. 17, no. 4, p. E841, 2017.
that the RMSE between log-normal distribution and the real [17] S. H. Strogatz, “Complex systems: Romanesque networks,” Nature,
PDF is almost one-order of magnitude smaller than the others vol. 433, no. 7024, pp. 365–366, 2005.
in the same row. As a result, an astounding but well-grounded [18] C. Song, S. Havlin, and H. A. Makse, “Self-similarity of complex
conclusion can be drawn as follows: regardless of geographical networks,” Nature, vol. 433, no. 7024, pp. 392–395, 2005.
[19] D. Benhaiem, M. Joyce, and B. Marcos, “Self-similarity and stable
boundaries, culture differences, and historical limitations, the clustering in a family of scale-free cosmologies,” Monthly Notices Roy.
Euler characteristics of both Asian and European cities entirely Astronomical Soc., vol. 443, no. 3, pp. 2126–2153, 2014.
conform to log-normal distribution. [20] T. Gneiting and M. Schlather, “Stochastic models which separate fractal
dimension and Hurst effect,” Soc. Ind. Appl. Math. Rev., vol. 46, no. 2,
pp. 269–282, 2001.
V. C ONCLUSION AND F UTURE W ORKS [21] M. Fernández-Martínez, M. A. Sánchez-Granero, J. E. T. Segovia, and
I. M. Román-Sánchez, “An accurate algorithm to calculate the Hurst
In this letter, several algebraic geometric tools, namely, exponent of self-similar processes,” Phys. Lett. A, vol. 378, nos. 32–33,
α-Shapes, Betti numbers, and Euler characteristics, have been pp. 2355–2362, 2014.
616 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
⎛ ⎞
N M SNRrn,k N
1 + SNRrn,k
M SNRrn,k
yk = Gn,k pn,k P sk + ⎝ h ∗ ĥ − ⎠ Gn,k pn,k P sk
1 + SNRrn,k M SNRrn,k n,k n,k 1 + SNRrn,k
n=1 n=1
Multiuser Interference: Ik
N → ∞, K
vectors have IID entries and, for K k =1 pn,k ≤ 1. Within the class of large-scale-based power
1 ∗ a.s. ∗ allocations, the most natural are:
1 k =k • Maximal-ratio, pn,k = K
Gn,k
.
h h → E hn,k hn,k = (3)
N k k 0 k = k. k=1 Gn,k
• Max-min (a quasi-convex optimization solved iteratively),
Thus, for N K, a precoder f k ∝ h k leads to (i) a hardened which equalizes the SINRs and maximizes fairness [1].
(over the fading) precoded channel at user k, whereby forward User k observes
pilots are not needed and the decoding can rely on large-scale N
quantities, and (ii) minimal interference onto users k = k. yk = Gn,k h ∗n,k x n + vk (9)
In cell-free networks, the channel gains connecting the N
n=1
APs with a user are no longer IID and hence (i) is not upheld.
where vk ∼ NC (0, σ 2 ).
With hardening-based reception (no
B. Reverse-Link Channel Estimation forward pilots), user k recovers the projection of yk onto [1]
Let Pk be the set of users (including user k) that share the M SNRrn,k
pilot of user k. The simultaneous transmission from the users Gn,k E h ∗n,k f cb = Gn,k pn,k P (10)
n,k
in this set of a pilot of power P r is observed at the nth AP as 1 + SNRrn,k
√
yn = Gn,k h n,k P r + v n (4) where we have invoked
(6). In ∗turn, cbthe projection of yk on
k∈Pk Gn,k h ∗n,k f cb
n,k − Gn,k E[h n,k f n,k ] is self-interference.
Combining (6)–(10), the observation at user k can be written
where v n ∼ NC (0, σ 2 I M ). From y n , the nth AP produces the as (2) at top of this page, from which the SINR emerges as
LMMSE channel estimate ĥ n,k satisfying h n,k = ĥ n,k +h̃ n,k
where h̃ n,k ∼ NC (0, MMSEn,k I ) is uncorrelated error and E |Sk |2
SINRk = 2
cb
(11)
σ + E[|Ek |2 ] + E[|Ik |2 ]
1 + k∈Pk ,k=k SNRrn,k 2
MMSEn,k = . (5) N SNR2n,k pn,k
1 + k∈Pk SNRrn,k
n=1 ρ+SNRn,k
Since the problem we address is not directly related to pilot =M N K , (12)
contamination, and there are ways of keeping such contami- 1+ n=1 SNRn,k k=1 pn,k
nation at bay [1], [11], we disregard it to avoid distractions expressed as function of only the forward-link SNRs via (1).
and the need to posit specific pilot assignments. This amounts In interference-limited conditions, the above reduces to
to Pk containing only user k, from which 2
N
M SNRrn,k n=1 G p
n,k n,k
E ĥ n,k 2 = , SIRcb
k = M N K . (13)
1 + SNRrn,k n=1 Gn,k k=1 pn,k
M
E h̃ n,k 2 = . (6)
1 + SNRrn,k IV. H ARDENING S HORTFALL IN C ELL -F REE N ETWORKS
The fluctuations of the precoded channels over their
C. Forward-Link Data Transmission expected values have two detrimental effects, opposite sides
The precoder applied by the nth AP to beamform to user k of the same coin, on hardening-based receivers: they subtract
n,k ∝ ĥ n,k . Altogether, the nth AP generates the signal
is f cb signal power, turning it into self-interference. These effects are
gauged in Fig. 1, which depicts, as a function of M,
K
E |Ek |2 E |Ek |2
xn = f cb
n,k sk (7) and (14)
k =1 E[|Sk |2 ] E[|Ek |2 ] + E[|Ik |2 ]
with sk the unit-variance symbol intended for user k and further averaged over the user locations, for maximal-ratio
power allocation, η = 4, and N/K = 4 and 10. These ratios
p P quantify the self-interference, respectively as a share of the
n,k ĥ n,k
f cb
n,k = (8) desired signal and of the total interference. For M = 1, self-
E ĥ n,k 2 interference steals about a third of the desired signal and it
represents about two thirds of the interference, and only for
where, by virtue of the normalization by E[ ĥ n,k 2 ], the share substantial M is this largely corrected. For cell-free networks
of power that the nth AP devotes to user k is pn,k with with small M, therefore, self-interference is a major issue.
618 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Fig. 2. CDF of SIR (averaged over the fading) for maximal-ratio power
allocation, η = 4, N /K = 10 and M = 1.
Desired Signal: Sk then we declare an upward fluctuation and all the precoders
intended for user k are scaled down to
N
1 + SNRrn,k √
+ Gn,k P h ∗n,k ĥ n,k pn,k sk +vk ∗ f cb
M SNRrn,k r E n G n,k ĥ n,k n,k
n=1 k=k
f mod
n,k = f cb
n,k ∀n (19)
G ∗ f
n,k ĥ n,k n,k
cb
Multiuser Interference: Ik n
(15) such that the overall gain is pushed back to the target.
If (18) is reversed, the fluctuation is downwards. To correct
free of self-interference. From (15), as a function of the
it with the minimal amount of interference to other users, we
realization of h ∗n,k ĥ n,k ,
identify as nmax the AP having the strongest large-scale gain
to user k and adjust upwards only f nmax ,k , setting it to
E |Sk |2 |h ∗n,k ĥ n,k
genie
SINRk = f mod
nmax ,k
σ 2 + E |Ik |2 |h ∗n,k ĥ n,k
2 rE Gn,k ĥ ∗n,k f cb
n,k − Gn,k ĥ ∗n,k f cb
1 N
√ =
n
n=nmax n,k
ĥ nmax ,k
M n=1 ρ + SNRn,k h ∗n,k ĥ n,k pn,k Gnmax ,k ĥ nmax ,k 2
= .
1 N 2 ∗ (20)
1+ M n=1 SNRn,k E h n,k |h n,k f n,k k=k pn,k
(16) where by
In interference-limited conditions, the above reduces to Gnmax ,k ĥ ∗nmax ,k f mod
nmax ,k
2
N 2 √p
G h ∗
genie n=1 n,k n,k n,k =r E Gn,k ĥ n,k f n,k −
cb
Gn,k ĥ ∗n,k f cb
n,k
SIRk = N . (17)
2 n
n=1 Gn,k h n,k k=k pn,k
n=nmax
(21)
Figs. 2–4 present SIR distributions over all locations for
M = 1, confirming the deficiency of (13) relative to (17) when and the target is met again. Altogether, every user experiences
M is small. a stable overall gain, disturbed only by channel estimation
ATTARIFAR et al.: MODIFIED CONJUGATE BEAMFORMING FOR CELL-FREE MASSIVE MIMO 619
Fig. 3. CDF of SIR (averaged over the fading) for max-min power allocation, Fig. 4. CDF of SIR (averaged over the fading) for max-min power allocation,
η = 4, N/K = 10 and M = 1. η = 4, N/K = 4 and M = 1.
errors, and the multiuser interference is curbed. (Since r<1, VIII. D ISCUSSION
the per-AP transmit powers are lowered and a rescaling of all The proposed modifications preserve the utter simplicity
precoders is advisable if noise is significant; in interference- of conjugate beamforming, free of matrix inversions, at the
limited conditions, this is immaterial.) User k observes expense of fading-rate coordination—needed anyway to com-
bine and decode the reverse-link transmissions—among the
yk = r E Gn,k ĥ ∗n,k f cb
n,k sk
APs. By translating these modifications from f n,k to pn,k ,
n they can be construed as a fading-based power allocation,
Desired Signal: Sk
which can be overlaid onto any existing large-scale-based
allocation.
+ Gn,k h ∗n,k f mod
n,k − r E Gn,k ĥ ∗n,k f cb
n,k sk
n n R EFERENCES
Self−interference: Ek
[1] H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta,
+ Gn,k h ∗n,k f mod
n,k sk +vk (22) “Cell-free massive MIMO versus small cells,” IEEE Trans. Wireless
Commun., vol. 16, no. 3, pp. 1834–1850, Mar. 2017.
n k=k
[2] E. Nayebi, A. Ashikhmin, T. L. Marzetta, and B. D. Rao, “Performance
Mutual Interference: Ik of cell-free massive MIMO systems with MMSE and LSFD receivers,”
in Proc. Asilomar Conf. Signals Syst. Comput., Pacific Grove, CA, USA,
Nov. 2016, pp. 203–207.
where, if ĥ n,k = h n,k (interference-limited conditions with- [3] E. Nayebi, A. Ashikhmin, T. L. Marzetta, H. Yang, and B. D. Rao,
out pilot contamination), Ek indeed vanishes giving “Precoding and power optimization in cell-free massive MIMO
2
systems,” IEEE Trans. Wireless Commun., vol. 16, no. 7, pp. 4445–4459,
r 2M n Gn,k pn,k Jul. 2017.
SIRk =
mod (23) [4] T. C. Mai, H. Q. Ngo, M. Egan, and T. Q. Duong, “Pilot power control
mod 2
n k=k Gn,k E f n,k for cell-free massive MIMO,” IEEE Trans. Veh. Technol., vol. 67, no. 11,
pp. 11264–11268, Nov. 2018.
[5] J. Zhang, Y. Wei, E. Björnson, Y. Han, and S. Jin, “Performance analysis
with the denominator following from the independence of and power control of cell-free massive MIMO systems with hardware
n,k and h n,k . The portion r, which must be known by the
f mod impairments,” IEEE Access, vol. 6, pp. 55302–55314, 2018.
users, can be optimized over; note that, in (23), r affects the [6] G. Interdonato, H. Q. Ngo, E. G. Larsson, and P. Frenger, “How much do
numerator and, through f mod
n,k , also the denominator.
downlink pilots improve cell-free massive MIMO?” in Proc. IEEE Glob.
Commun. Conf. (GLOBECOM), Washington, DC, USA, 2016, pp. 1–7.
[7] H. Q. Ngo and E. G. Larsson, “No downlink pilots are needed in
VII. E XAMPLES TDD massive MIMO,” IEEE Trans. Wireless Commun., vol. 16, no. 5,
√ pp. 2921–2935, May 2017.
Figs. 2–3 exemplify how, with N/K = 10 and r = 1/ 2 [8] J. Vieira, F. Rusek, and F. Tufvesson, “Reciprocity calibration meth-
for both maximal-ratio and max-min power allocations, the ods for massive MIMO based on antenna coupling,” in Proc. IEEE
modifications push the SIR close to the (unachievable) genie- Glob. Commun. Conf. (GLOBECOM), Austin, TX, USA, 2014,
aided upper bound, erasing most of the deficit of conjugate pp. 3708–3712.
[9] B. Błaszczyszyn, M. K. Karray, and H. P. Keeler, “Wireless networks
beamforming. For shrinking N/K, self-interference is progres- appear Poissonian due to strong shadowing,” IEEE Trans. Wireless
sively overcome by multiuser interference, yet the modified Commun., vol. 14, no. 8, pp. 4379–4390, Aug. 2015.
beamformer continues to perform satisfactorily close to the [10] G. George, R. K. Mungara, A. Lozano, and M. Haenggi, “Ergodic
genie-aided bound. This is illustrated in Fig. 4, which is the spectral efficiency in MIMO cellular networks,” IEEE Trans. Wireless
counterpart to Fig. 3 for N/K = 4. Commun., vol. 16, no. 5, pp. 2835–2849, May 2017.
[11] O. Y. Bursalioglu, C. Wang, H. Papadopoulos, and G. Caire, “RRH
In terms of η, its increase weakens the hardening and based massive MIMO with ‘on the fly’ pilot contamination control,” in
renders the modified beamforming even more effective. Proc. IEEE Int. Conf. Commun. (ICC), Kuala Lumpur, Malaysia, 2016,
Conversely, a decrease in η reduces the advantage. pp. 1–7.
620 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—Distribution matching (DM) transforms indepen- error propagation. Research is therefore now dedicated to
dent and Bernoulli(1/2) distributed bits into a sequence of find improved DM architectures for short blocklengths, e.g.,
output symbols with a desired distribution. A fixed-to-fixed [8] and [9]. Good performance for short blocklengths is also
length, invertible DM architecture based on shell mapping needed to operate several DMs in parallel to further reduce
(SM) is presented. It is shown that SM for DM (SMDM)
is the optimum DM for the informational divergence metric processing latencies.
and that finding energy optimal sequences is a special case In this letter, we introduce shell-mapping distribution match-
of divergence minimization. Additionally, it is shown how to ing (SMDM), a fixed-to-fixed (f2f) length DM architecture for
find the required SM weight function to approximate arbi- short output blocklengths based on shell mapping. Shell map-
trary output distributions. SMDM is combined with probabilis- ping was developed in the early 1990s [3], [10], [11] and was
tic amplitude shaping to operate close to the Shannon limit. used in the V.34 modem standard to realize shaping gains
SMDM exhibits excellent performance for short blocklengths with trellis coded modulation (TCM). We show that SMDM
as required by ultra-reliable low-latency applications. SMDM
outperforms constant composition DM by 0.7 dB when used
minimizes the informational divergence of f2f length DM if
with 64-QAM at a spectral efficiency of 3 bits/channel use and the self-information of the target distribution is used as the
a 5G low-density parity-check code with a short blocklength weight function for the shell mapping algorithm. Further, we
of 192 bits. show that the dyadic and Maxwell-Boltzmann (MB) distribu-
Index Terms—Probabilistic amplitude shaping, distribution
tions [1] lead to integer weight functions, which significantly
matching, shell mapping, coded modulation. simplify the implementation of shell mapping (SM). Finally,
we explain how to integrate SMDM with PAS to operate close
to the Shannon limit at small blocklengths. Numerical simu-
I. I NTRODUCTION lations with 64-QAM and low-density parity-check (LDPC)
codes from the 5G enhanced mobile broadband (eMBB) stan-
IGHER-ORDER modulation is a key enabler for high
H spectral efficiencies and various approaches have been
considered in the past to close the shaping gap of discrete
dard [12] show a gain of 0.6 dB of SMDM over CCDM for
a spectral efficiency (SE) of 3.0 bits/channel use (bpcu).
constellations with uniformly distributed points (e.g., with
quadrature amplitude modulation (QAM)) [1]–[3]. Recently, II. P RELIMINARIES
probabilistic amplitude shaping (PAS) [4] was proposed that A. Notation
is based on a reverse concatenation architecture [5], placing the We denote random variables with uppercase letters, and
shaping operation before the forward error correction (FEC) their realizations with lowercase letters. Let A be a discrete
encoding. Apart from achieving most of the shaping gain, it random variable with probability mass function (pmf) PA
allows flexible rate adaptation with a single constellation and defined on the set A. If an event A = a occurs with positive
FEC code rate. To convert uniformly distributed input bits to probability, then its self-information is
non-uniformly distributed output symbols, PAS requires a dis-
tribution matcher (DM). Schulte and Böcherer [6] introduced ι(PA (a)) = − log2 (PA (a)) bits. (1)
constant composition distribution matching (CCDM) which The entropy of a random variable A is the expectation of the
is asymptotically optimal, in the sense of a vanishing nor- self-information of A, i.e., we have
malized informational divergence [7, p. 7], for long output
blocklengths. H(PA ) = E[ι(PA (A))] = −PA (a) log2 (PA (a)),
For practical communication systems and new requirements a∈supp(PA )
such as ultra reliable low latency communication (URLLC), (2)
shorter output blocklengths in the range of 10 to 500 symbols
are desirable to minimize the processing latency and limit where supp(PA ) ⊆ A is the support of PA , i.e., the subset of
a in A with positive probability. The informational divergence
Manuscript received September 21, 2018; revised December 11, 2018; of two distributions PÃ and PA on A is
accepted December 12, 2018. Date of publication January 1, 2019; date of
current version April 9, 2019. This work was supported by the German Federal PÃ (a)
Ministry of Education and Research in the framework of an Alexander von D PÃ PA = PÃ (a) log2 . (3)
Humboldt Professorship. The associate editor coordinating the review of this PA (a)
paper and approving it for publication was R. C. de Lamare. (Corresponding a∈supp(PÃ )
author: Patrick Schulte.)
The authors are with the Institute for Communications Engineering, The mutual information of two random variables A and B with
Technical University of Munich, 80333 Munich, Germany (e-mail: joint pmf PAB is
patrick.schulte@tum.de; fabian.steiner@tum.de).
Digital Object Identifier 10.1109/LWC.2018.2890595 I(A; B ) = D(PAB PA × PB ), (4)
2162-2345 c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
SCHULTE AND STEINER: DIVERGENCE-OPTIMAL f2f LENGTH DM WITH SM 621
VI. C ONCLUSION
We introduced an informational divergence optimal f2f
length DM approach based on SM, which shows superior
performance compared to state of the art DMs for short
blocklengths. We showed that the self-information of the
target output distribution can be used as the weight func-
tion for the SM algorithm to synthesize arbitrary output
distributions. Furthermore, we showed that energy efficient
signaling is a special case of divergence minimization. We
gave examples for distributions that result in non-negative,
integer valued SM weight functions favorable for practical
implementations.
Fig. 2. Finite length performance for uniform and shaped signaling using
ACKNOWLEDGMENT
CCDM and SMDM. We target an SE of 3 bpcu with 5G LDPC codes of The authors would like to thank Gerhard Kramer and Georg
blocklength 192. Böcherer for fruitful discussions.
Abstract—The signal to clipping noise ratio (SCR)-based squares approximation algorithm (LSA)-TR [9]. However,
tone reservation scheme is a promising candidate for peak-to- these schemes require additional fast Fourier transform (FFT)
average ratio (PAPR) reduction of orthogonal frequency division and inverse fast Fourier transform (IFFT) operations in each
multiplexing signals, but it suffers from low convergence rate
caused by the nonoptimal scalar scaling factor. In this letter, iteration and thereby inevitably result in an increased compu-
we first derive a scaling SCR (S-SCR) scheme, in which the tation complexity. On the other hand, a complexity reduced
scaling factor is an optimized vector with peak regeneration TR scheme using a gradient algorithm was proposed by
constraints. To further improve convergence rate and eliminate Tellado [7]. However, this signal to clipping noise ratio (SCR)
multiple peaks in each iteration, we then propose the multiple scheme suffers from a low convergence speed caused by
scaling SCR (MS-SCR) scheme with an augmented scaling factor
in matrix form. Numerical results demonstrate the superiority of a non-optimal scaling factor. Yu and Jin [10] presented a time-
proposed schemes compared with the conventional ones, includ- domain kernel matrix (TKM) TR scheme with simultaneous
ing better PAPR reduction performance with fewer iterations multi-peak reduction and rapid convergence speed. However,
and comparable bit error ratio performance. Furthermore, the because of the lack of peak regeneration suppression, its
computational complexity reduction ratio of the proposed S-SCR performance deteriorates sharply when applied to low clipping
and MS-SCR schemes are increased by 38.52% and 27.36%,
respectively. thresholds.
In this letter, we firstly propose a novel SCR-based PAPR
Index Terms—OFDM, PAPR reduction, SCR, tone reservation, reduction scheme, called scaling SCR (S-SCR) scheme, in
scaling factor.
which the optimal scaling factor vector is calculated by the
LSA algorithm with peak regeneration constraints. Further,
an improved Multiple Scaling SCR (MS-SCR) method with
I. I NTRODUCTION an augmented scaling factor in matrix form is proposed to
RTHOGONAL frequency division multiplexing suppress peak regeneration and eliminate multiple peaks in
O (OFDM) signals have been extensively adopted in
wireless communication systems. However, OFDM signals
one iteration. Simulated results show that the two schemes
introduce great improvements on computational complexity,
often suffer from high peak-to-average ratio (PAPR) which convergence speed, and peak regeneration suppression with
leads to significant in-band distortion and out-of-band (OOB) comparable communication performance. This letter is orga-
radiation. Consequently, various PAPR reduction techniques nized as follows. Section II briefly introduces the OFDM
have been proposed, such as clipping [1], coding [2], system and conventional SCR scheme. Section III presents
companding [3], selective mapping (SLM) [4], and partial the two proposed schemes. Simulated results are provided in
transmit sequence (PTS) [5]. Section IV and the conclusion is given in Section V.
The tone reservation (TR) schemes [6]–[10] have gained
great attention because of their low complexity and good bit II. P RELIMINARY BACKGROUNDS
error ratio (BER) performance. The TR scheme reserves a sub-
A. OFDM System and PAPR
set of subcarriers as peak reduction tones (PRTs), in which
no information data are loaded and which is used for gen- In an OFDM system, N orthogonal subcarriers are used to
erating the peak canceling signal. To raise the convergence transmit modulated data symbols X = [X0 , X1 , . . . , XN −1 ]T .
speed, improved TR algorithms were subsequently proposed, An OFDM signal sequence in the time domain can be
such as the adaptive-scaling (AS)-TR [8] and the least generated by an IFFT and is written as
N −1
Manuscript received September 30, 2018; revised November 5, 2018; 1 2πnk
accepted December 3, 2018. Date of publication January 1, 2019; date of x (n) = √ Xk e j LN , n = 0, 1, . . . , LN − 1, (1)
current version April 9, 2019. This work was supported in part by the N k =0
National Natural Science Foundation of China under Grant 61301020, in
part by the Natural Science Foundation of Jiangsu Province under Grant where L is the oversampling factor, which usually satisfies
BK20130772, and in part by the Priority Academic Program Development L ≥ 4.
of Jiangsu Higher Education Institutions. The associate editor coordinating
the review of this paper and approving it for publication was J. Mietzner. In the time domain, the discrete signal x(n) is actually
(Corresponding author: Jingqi Wang.) a sum of N orthogonal subcarriers. Therefore, the OFDM sig-
The authors are with the Ministerial Key Laboratory of JGMT, Nanjing nal occasionally exhibits envelope fluctuations which can be
University of Science and Technology, Nanjing 210094, China (e-mail:
wangjingqi@njust.edu.cn). quantified by the PAPR. The PAPR is defined as the ratio of
Digital Object Identifier 10.1109/LWC.2018.2890596 maximum signal power to average signal power.
2162-2345 c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
WANG et al.: SCR-BASED TR SCHEMES WITH FAST CONVERGENCE FOR PAPR REDUCTION IN OFDM SYSTEM 625
TABLE I
OVERALL C OMPUTATIONAL C OMPLEXITY OF
T HREE S CHEMES AND CCRR S
Fig. 2. PAPR reduces along with the increasing iteration number when A. Complexity Analysis
CCDF = 10-3 (top) and the number of peaks above the threshold with To evaluate computational complexity, iteration numbers of
iteration numbers (bottom) for three schemes.
the proposed schemes that achieve the same PAPR reduction
performance need to be discussed. Fig. 2 shows the PAPR
Then the optimization problem can be modified as and the number of peaks above the threshold decreases along
U −1 2 with increasing iteration number for SCR, S-SCR and MS-
SCR schemes when CR = 2dB. As shown in Fig. 2, for
m m
min γu · cu (n) − f (n) . (11)
n∈S the particular OFDM symbol considered, the PAPR reaches
u=0 2 8.8 dB after 18 and 11 iterations in SCR and S-SCR, respec-
Accordingly, by applying the LSA algorithm, the approxi- tively. Meanwhile, by using the MS-SCR scheme, PAPR can
mate optimization value of the scaling factor γu is obtained as be reduced to 5dB in only one iteration.
|cum (n)||f m (n)| To represent the complexity reduction accurately, we use
γu = n∈S 2
. (12) the computational complexity reduction ratio (CCRR).
m
n∈S |cu (n)|
The peak reduction signal in (10) is generated by mixing complexity of proposed method
CCRR = 1 − × 100%.
kernel signals with many non-ignorable values of amplitudes complexity of previous method
near the maximum peak, so the peak regrowth is serious. To (15)
avoid the undesirable peak regeneration, similar to the scal-
ing vector of the S-SCR scheme, the LN-dimensional scaling Table I lists the overall computation complexity in terms of
factor matrix Γu is designed as real multiplications for the three schemes, we can make a brief
summary that the CCRRs of the proposed S-SCR and MS-SCR
γu i = su
Γ u (i ) = (13) schemes are increased by 38.52% and 27.36% compared with
1 i = su .
the original SCR scheme, respectively.
Thus, the iteratively updating peak reduced signal x̃ (n) can
be denoted as
B. PAPR Reduction Performance
U
−1 m
x̃ m+1 (n) = x̃ m (n) − Γu · x̃ m (su ) − Ae j arg{x̃ (su )} It can be observed from the above simulations that the SCR
u=0 and S-SCR converge after 18 iterations, and the MS-SCR with
× p[((n − su ))LN ]. (14) one iteration can offer better PAPR performance. Moreover,
TKM converges after 3 iterations in [10]. Therefore, in the fol-
Fig. 1 depicts the peak reduced signal of the MS-SCR lowing simulations, the number of iterations for SCR, S-SCR,
scheme in one iteration compared with the original OFDM TKM, and MS-SCR is set to be 18, 18, 3 and 1, respectively.
signal when CR is set to be 2dB. The dashed red line is As shown in Fig. 3, for the same CR, the PAPR reduction
the clipping threshold A. It can be observed that the ampli- performance of the proposed S-SCR and MS-SCR scheme
tudes of the clipped signals are almost all below the predefined is obviously better than the traditional SCR scheme. At the
threshold A after only one iteration. Obviously, the proposed same time, compared with TKM after three iterations, MS-
MS-SCR can efficiently eliminate numerous peaks. SCR obtains 2.7dB reduction gain with only one iteration.
The detailed steps of the MS-SCR scheme can be summa- Furthermore, it is evident from this figure that the PAPR of
rized as Algorithm 1. OFDM with MS-SCR technique is reduced from 6.1, 5.1 to
4.6dB for CR = 3, 2, and 1 dB, respectively. It is obvious that
IV. C OMPLEXITY A NALYSIS AND S IMULATION the multi-peak elimination ability of MS-SCR enhances as the
In this section, a 16-QAM modulated OFDM signal with the clipping ratio reduces.
number of subcarriers N = 256 and the number of reserved Fig. 4 further compares the performance of the four schemes
tones T = 32 is applied to evaluate the proposed schemes. with different relative number of reserved tones T/N for a sub-
Moreover, an oversampling rate L = 4 is used and the PAPR carrier number N = 256 and CR = 2 dB. It is shown
performance is represented by the complementary cumulative that when SCR, TKM, and S-SCR all reach the conver-
distribution function (CCDF) of the PAPR. gence state, the S-SCR scheme achieves better PAPR reduction
WANG et al.: SCR-BASED TR SCHEMES WITH FAST CONVERGENCE FOR PAPR REDUCTION IN OFDM SYSTEM 627
Fig. 3. PAPR reduction performance of the SCR, TKM, S-SCR and MS-SCR V. C ONCLUSION
schemes with CR = 1 dB, 2 dB, and 3 dB.
In this letter, we first proposed an improved SCR-based
PAPR reduction scheme, which utilizes a least square algo-
rithm to calculate the optimal scaling factor for clipping
noise. This proposed S-SCR method greatly improves the
convergence rate and computational complexity. Furthermore,
we expanded the scaling factor vector into matrix form and
reshaped it for optimal peak regeneration suppression. The
derived MS-SCR scheme can eliminate multiple peaks in
one iteration. Simulation results show that the proposed MS-
SCR scheme achieves powerful multi-peak elimination with
rapid convergence rate and comparable BER performance.
It is attractive for practical implementations with low PAPR
requirement. Moreover, the S-SCR scheme shows low com-
putational complexity and great PAPR reduction performance,
Fig. 4. PAPR reduction performance of the SCR, S-SCR, TMK and MS-SCR given a relatively high threshold, and therefore is a suitable
schemes with relative number of reserved tones T/N = 1/16, 1/8, and 1/4.
candidate for highly cost-efficient applications.
R EFERENCES
[1] H. Ochiai and H. Imai, “Performance analysis of deliberately clipped
OFDM signals,” IEEE Trans. Commun., vol. 50, no. 1, pp. 89–101,
Jan. 2002.
[2] S. H. Han and J. H. Lee, “An overview of peak-to-average power
ratio reduction techniques for multicarrier transmission,” IEEE Wireless
Commun., vol. 12, no. 2, pp. 56–65, Apr. 2005.
[3] X. Huang, J. H. Lu, J. L. Zheng, K. B. Letaief, and J. Gu, “Companding
transform for reduction in peak-to-average power ratio of OFDM sig-
nals,” IEEE Trans. Wireless Commun., vol. 3, no. 6, pp. 2030–2039,
Nov. 2004.
[4] S.-J. Heo, H.-S. Noh, J.-S. No, and D.-J. Shin, “A modified SLM scheme
with low complexity for PAPR reduction of OFDM systems,” IEEE
Trans. Broadcast., vol. 53, no. 4, pp. 804–808, Dec. 2007.
[5] R. J. Baxley and G. T. Zhou, “Comparing selected mapping and partial
Fig. 5. BER performance comparison of SCR, S-SCR, TKM and MS-SCR transmit sequence for PAR reduction,” IEEE Trans. Broadcast., vol. 53,
with 16-QAM mapping over AWGN channel through an SSPA model, with no. 4, pp. 797–803, Dec. 2007.
parameter p = 2 and the input back-off (IBO) set to be 5 dB. [6] C. Ni, Y. Ma, and T. Jiang, “A novel adaptive tone reservation scheme
for PAPR reduction in large-scale multi-user MIMO-OFDM systems,”
IEEE Wireless Commun. Lett., vol. 5, no. 5, pp. 480–483, Oct. 2016.
[7] J. Tellado, “Peak to average power reduction for multicarrier modula-
performance. Moreover, MS-SCR is obviously superior to tion,” Ph.D. dissertation, Dept. Elect. Eng., Stanford Univ., Stanford,
other three schemes with only one iteration. CA, USA, 2000.
[8] L. Wang and C. Tellambura, “Analysis of clipping noise and tone-
reservation algorithms for peak reduction in OFDM systems,”
C. BER Performance IEEE Trans. Veh. Technol., vol. 57, no. 3, pp. 1675–1694,
In order to evaluate the overall BER performance for SCR, May 2008.
[9] H. Li, T. Jiang, and Y. Zhou, “An improved tone reservation scheme
TKM, S-SCR and MS-SCR schemes, we consider passing with fast convergence for PAPR reduction in OFDM systems,” IEEE
the clipped OFDM signals through a solid-state power ampli- Trans. Broadcast., vol. 57, no. 4, pp. 902–906, Dec. 2011.
fier (SSPA) model. Fig. 5 compares the BER performance of [10] P. Yu and S. Jin, “A low complexity tone reservation scheme
based on time-domain kernel matrix for PAPR reduction in OFDM
the four schemes with 16-QAM mapping when the parame- systems,” IEEE Trans. Broadcast., vol. 61, no. 4, pp. 710–716,
ters of OFDM are N = 256 and T = 32. To evaluate the BER Dec. 2015.
628 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—In this letter, we deal with the age of information queueing-theoretic standpoint under various service policies,
(AoI) for a sensor network with wireless power transfer (WPT) e.g., [5] and [6]. Recent works employ the notion of AoI
capabilities. Specifically, we study a simple network topology, in energy harvesting communication systems (from natural
where a sensor node harvests energy from radio frequency signals
renewable sources), and investigate transmission policies that
(transmitted by a dedicated energy source) to transmit real-time
status updates. The sensor node generates an update when its minimize AoI-based performance metrics [7]–[10]. On the
capacitor/battery becomes fully charged and transmits by using other hand, the design of WPT-based communication systems
all the available energy without further energy management. with objective to optimize AoI, is a new research area with
The average AoI performance of the considered greedy policy is potential applications. Dong et al. [11] propose a two-way data
derived in closed form and is a function of the capacitor’s size. exchanging system, where a master node transfers energy and
The optimal value of the capacitor that maximizes the freshness information to a slave node, while the slave node uses the
of the information, corresponds to a simple optimization problem
requiring a 1-D search. The derived theoretical results provide energy harvested to power the uplink channel; the average
useful performance bounds for practical WPT networks. uplink AoI is derived in closed form. Although AoI seems to
be a natural design metric for WPT networks, other relevant
Index Terms—Age of information, wireless power transfer,
energy harvesting, sensor networks.
works cannot be found in the literature.
In this letter, we study a basic communication link where a
sensor node with WPT capabilities communicates with a sin-
gle destination. Specifically, the sensor node is equipped with
I. I NTRODUCTION
a capacitor, which is charged via RF radiation by a dedicated
IRELESS power transfer (WPT) via dedicated radio-
W frequency (RF) radiation is a promising technology
for wireless communication systems, which are characterized
energy source. Once the capacitor is charged, the sensor node
transmits status updates containing the most recent information
about parameters of interest by using all the energy stored.
by a massive number of low-power devices such as in the This online transmission policy does not require compli-
Internet-of-Things systems. It can support mobility, energy cated energy management decisions (e.g., energy-depended
multicasting, non-line-of-sight propagation environments and thresholds) and is appropriate for WPT low complexity/low
contributes in the development of smaller, lighter and more power devices. We investigate the freshness of the received
compact devices. From the pioneering work of Varshney [1], information and we provide simple closed form expressions
who has introduced this concept, WPT has been extensively for the average AoI, which depend on the size of the capaci-
studied in the literature for different network architectures, tor. The design of the system introduces an interesting tradeoff:
e.g., [2] and [3]. However, most of the current works focus a small capacitor is charged quickly and thus new updates are
on complex network structures with limited practical interest sent more frequently to minimize the AoI; on the other hand,
and/or use conventional performance metrics, e.g., throughput, a larger capacitor increases the transmit power and boosts
coverage probability, diversity gain, information-energy capac- the successful decoding. The optimal value of the capaci-
ity etc, which do not capture timeliness requirements that arise tor is computed by solving a one-dimensional optimization
from sensing and actuation applications within machine-type problem. It is worth noting that the network topology and
communications. the transmission policy considered are inspired by commer-
A performance metric that captures the freshness of the cial battery-free WPT products, e.g., Powercast [12]; these
received information and is appropriate for applications requir- devices are equipped with supercapacitors that deliver high
ing timely information to accomplish specific tasks (e.g., power bursts when charged. Although our analysis refers
sensor networks, cyberphysical systems, etc), was proposed to a simplistic system model, the derived theoretical results
in [4], i.e., age of information (AoI). It is defined as the can serve as guidelines (performance bounds) for practical
time elapsed since the generation of the freshest status update implementations.
that has reached the destination. Initial works on AoI take
into account traffic burstiness and minimize the AoI from a
II. S YSTEM M ODEL
Manuscript received November 8, 2018; revised December 14, 2018;
accepted December 15, 2018. Date of publication January 9, 2019; date We assume a simple WPT sensor network consisting of
of current version April 9, 2019. This work was supported in part by one energy transmitter (ET), one sensor node, S, and one
the European Regional Development Fund and in part by the Republic
of Cyprus through the Research Promotion Foundation under Project
information receiver (IR); all the nodes are equipped with
INFRASTRUCTURES/1216/0017. The associate editor coordinating the single antennas. The ET is connected to the power grid and
review of this paper and approving it for publication was K. W. Choi. continuously broadcasts an energy signal with power P. The
The author is with the Department of Electrical and Computer Engineering, sensor node has WPT capabilities and harvests energy from
Faculty of Engineering, University of Cyprus, 1678 Nicosia, Cyprus
(e-mail: krikidis@ucy.ac.cy). the received RF signal; the harvested energy is stored in a
Digital Object Identifier 10.1109/LWC.2018.2890605 capacitor of finite-size B. When the capacitor becomes fully
2162-2345 c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
KRIKIDIS: AVERAGE AoI IN WIRELESS POWERED SENSOR NETWORKS 629
Abstract—In this letter, we introduce new user-cooperation of the limitations of BackCom is that if the incident signal is
schemes for wireless devices in a wireless-powered backscatter unavailable, information backscattering (IB) is impossible. To
communication network with the aim to improve communica- fully exploit the advantages of both the HTT and BackCom,
tion and energy efficiency for the whole network. In particular,
BackCom has been introduced in WPCNs [6], [7], where each
we consider two types of wireless devices which can support
different communication modes, i.e., backscatter and harvest- device can choose to operate in either the HTT or BackCom
then-transmit, and they can cooperate to deliver the information mode. However, IoT devices are hardware-constrained devices,
to the access point. To improve energy transmission efficiency for the backscatter and energy harvesting circuits together with an
the devices, energy beamforming is deployed at the power beacon. adaptive switch required to support the HTT and BackCom
We then formulate the weighted sum-rate maximization problem modes may not be available in practice. Hence, the assump-
by jointly optimizing time schedule, power allocation, and energy tion that the devices can support both the two modes may not
beamforming. Due to the non-convex issue of the optimization
problem, we employ the variable substitutions and semidefinite be practical. Furthermore, both user cooperation and energy
relaxation techniques to obtain the optimal solution. Simulation beamforming are not considered in these works, hence the
results show that the proposed cooperation framework can sig- communication and energy efficiency can not be maximized.
nificantly improve the communication efficiency compared with In this letter, we introduce two user cooperation schemes for
non-cooperation approach. the WPCN with BackCom with the aim to optimize commu-
Index Terms—Energy harvesting, backscatter communication, nication and energy efficiency for the network. In particular,
user cooperation, energy beamforming. we consider two wireless devices, denoted by HD and BD,
supported to operate in two different modes, i.e., HTT and
BackCom, respectively. We then consider two important sce-
I. I NTRODUCTION narios, i.e., the HD (BD) is located nearer the AP and can
IRELESS power transfer (WPT) has been considered be served as a relay node to assist the BD (HD) to transmit
W to be a promising way to supply wireless devices
with sustainable energy. In a wireless-powered communication
information due to the low channel quality of the BD (HD).
For each scenario, we formulate the weighted sum-rate (WSR)
network (WPCN), wireless devices can first harvest energy optimization problem by jointly optimizing the time schedule,
from a power beacon (PB), and then transmit their information power allocation, and energy beamforming. To deal with the
to the dedicated access point (AP) following the harvest-then- non-convex issue of the optimization problem, we first employ
transmit (HTT) protocol [1]. In [2] and [3], user cooperation the variable substitutions and design the optimal energy beam-
was applied in WPCNs to enhance system performance by forming vector only for IB or information forwarding (IF).
exploiting cooperative diversity. However, since devices in After that, the energy beamforming matrix is derived based
[2] and [3] are the HTT devices, the dedicated energy harvest- on semidefinite relaxation (SDR) [11] for the joint IB and EH
ing (EH) phase is required, which may reduce the duration of which satisfies the rank-one constraint. Simulation results then
the information transmission (IT) phase. show that our proposed cooperation framework can achieve
Recently, backscatter communication (BackCom) has been larger communication efficiency than that of non-cooperation
introduced as a novel communication method for IoT approach.
networks [4]. The BackCom device transmits information to
II. S YSTEM M ODEL AND N OTATIONS
the AP by modulating and reflecting the incident signals,
which requires less circuit power consumption and makes its As illustrated in Fig. 1, we consider the WPCN with
instantaneous harvested energy be sufficient to power its circuit BackCom, including a PB, an AP, and two devices, denoted by
operation [5]. Hence, the dedicated EH phase is not necessary, HD and BD, supported to operate in two different modes, i.e.,
which avoids the limitation of the HTT protocol. However, one HTT and BackCom, respectively. The PB with stable power
supply has N antennas, and the two devices are with single
Manuscript received November 19, 2018; accepted December 26, 2018. antenna. We consider two cases: (i) the HD is located nearer
Date of publication January 1, 2019; date of current version April 9, 2019. the AP than the BD, and it can operate as a relay node, and
This work was supported by the National Natural Science Foundation of
China under Grant 61671252 and Grant 61772287. The associate editor (ii) the BD is located nearer the AP than the HD, and it can
coordinating the review of this paper and approving it for publication was work as a relay node. Note that the relay node also needs to
P. D. Diamantoulakis. (Corresponding author: Bin Lyu.) deliver its own information to the AP. Moreover, we assume
B. Lyu and Z. Yang are with the National Engineering Research Center
of Communication and Network Technology, Nanjing University of Posts
the relay node decodes the information transmitted by the other
and Telecommunications, Nanjing 210003, China (e-mail: blyu@njupt.edu.cn; device more easily than the AP, which is useful for cooperative
yangz@njupt.edu.cn). communication [2]. The channel vectors between the PB and
D. T. Hoang is with the Faculty of Engineering and Information Technology, the BD/HD/AP are denoted as h 0,1 , h 0,2 , and h 0,3 , respec-
University of Technology Sydney, Sydney, NSW 2007, Australia (e-mail:
hoang.dinh@uts.edu.au). tively. The channel variables between the BD-HD, BD-AP,
Digital Object Identifier 10.1109/LWC.2018.2890642 HD-AP, and HD-BD links are denoted as h1,2 , h1,3 , h2,3 , and
2162-2345 c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
LYU et al.: USER COOPERATION IN WIRELESS-POWERED BACKSCATTER COMMUNICATION NETWORKS 633
III. W EIGHTED S UM -R ATE M AXIMIZATION this letter, we use CVX tools [13] to derive the optimal
A. Case i solution. The optimal power allocations are further given by
∗ = e ∗ /τ ∗ and P ∗ = e ∗ /τ ∗ . Then, we compute the
P1,i
We first set
the time and energy constraints for network as 1,i 2 2,i 2,i 3
3 optimal solution ŵ ∗0,i from W ∗i . Note that if W ∗i satis-
follows: C1: i=0 τi ≤ 1, C2: τi ≥ 0, ∀i , C3: P1,i τ2 + fies the rank-one constraint, ŵ ∗0 computed from W ∗i /τ0∗ by
P2,i τ3 ≤ Eh,i , C4: ||ŵ 0,i ||2 ≤ 1, and C5: ||ŵ 1,i ||2 ≤ 1.
eigen-decomposition is the optimal energy beamforming vec-
Then, the optimization problem can be formulated by
tor during τ0 . Hence, we proceed to show that W ∗i always
max ω1 Rbd,i + ω2 Rhd,i has the rank-one property in the following proposition.
ŵ 0,i ,ŵ 1,i ,P i ,τ Proposition 3: The optimal solution W ∗i derived from P3
s.t. C1, C2, C3, C4, C5, (P1) is a rank-one matrix.
Proof: To show W ∗i is a rank-one matrix, we first give the
where τ = [τ0 , τ1 , τ2 , τ3 ], P i = [P1,i , P2,i ], ω1 and ω2
following optimization problem.
denote the nonnegative rate weights for the BD and the
HD, respectively. Problem P1 is not a convex optimization min Tr(W i )
Wi
problem because ŵ 0,i , ŵ 1,i , τ , and P i are coupled with ∗
s.t. e0,i ≤ Tr(h 0,1 h H
0,1 W i ), W i 0,
each other. To solve P1, we introduce some new variables
∗ ∗
and apply the SDR technique [11]. First, we introduce R̄bd,i , e1,i + e2,i ≤ ηP Tr(h 0,2 h H
0,2 W i ). (P4)
R̂bd,i , e0,i , and let e1,i = P1,i τ2 , e2,i = P2,i τ3 , W i =
Denote the optimal solution for Problem P4 as W †i , which
τ0 ŵ 0,i ŵ H
0,i . Hence, we have the following new constraints: is also a feasible solution for P3. The reason is that there are
C6: R̄bd,i ≤ τ1 log2 (1 + ξP |α0,i |2 |h1,3 |2 |h H 2 2
0,1 ŵ 1,i | /σap ) + more constraints in P3 than in P4, which guarantees that a
2 2
τ2 log2 (1 + ξe1,i |h2,3 | /(σap τ2 )), C7: R̄bd,i ≤ τ1 log2 (1 + feasible solution for P3 is also feasible for P4. Hence, we can
ξP |α0,i |2 |h1,2 |2 |h H 2 2
0,1 ŵ 1,i | /σhd ), C8: R̂bd,i = τ0 log2 (1 + derive that Tr(W †i ) ≤ Tr(W ∗i ) ≤ τ0∗ , which shows that W †i
ξP |α0,i |2 |h1,3 |2 e0,i /(σap2 τ )), C9: R
0 hd,i = τ3 log2 (1 +
is a feasible solution for P3. Furthermore, since the objective
ξe2,i |h2,3 | /(σap τ3 )), C10: e0,i ≤ Tr(h 0,1 h H
2 2 function of P3 is a function of e i , τ , R̂bd,i , R̄bd,i and Rhd,i ,
0,1 W i ),
C11: e1,i + e2,i ≤ ηP Tr(h 0,2 h H W ), C12: Tr(W we can derive that {W †i , e ∗i , τ ∗ , R̂bd,i
∗ , R̄ ∗ , R ∗ } is also
0,2 i i ) ≤ τ0 , bd,i hd,i
C13: W i 0, and C14: rank (W i ) = 1. Then, P1 is the optimal solution for P3, i.e., W †i = W ∗i . According to
recast as the theorem given in [14, Th. 3.2], we then show that W †i
is a rank-one matrix. Since there exists an optimal solution
max ω1 (R̂bd,i + R̄bd,i ) + ω2 Rhd,i W †i satisfying (rank (W †i ))2 ≤ 2, we derive that W †i = 0
W i ,ŵ 1,i ,e i ,τ ,R̂bd,i ,R̄bd,i ,Rhd,i
is rank-one. Hence, rank(W ∗i ) = 1.
s.t. C1, C2, C5, C6, C7, C8, C9, C10, C11, C12, C13, C14,
(P2) B. Case ii
where e i = [e0,i , e1,i , e2,i ]. However, Problem P2 is still Similar
2 to the first case, we add the following constraints:
not a convex optimization problem due to the rank-one con- C17: i=0 ti ≤ 1, C18: ti ≥ 0, ∀i , C19: P1,ii t1 ≤ Eh,ii ,
straint and the couple of τ1 and ŵ 1,i . To handle this, we C20: |ŵ 0,ii |2 ≤ 1, and C21: |ŵ 2,ii |2 ≤ 1. Then, the
first give the following proposition. Denote the optimal solu- optimization problem for Case ii is formulated by:
tion for P2 as {W ∗i , ŵ ∗1,i , e ∗i , τ ∗ , R̂bd,i
∗ , R̄ ∗ , R ∗ }, where
bd,i hd,i max ω1 Rbd,ii + ω2 Rhd,ii
e i = [e0,i , e1,i , e2,i ] and τ = [τ0 , τ1∗ , τ2∗ , τ3∗ ].
∗ ∗ ∗ ∗ ∗ ∗ ŵ ii ,P1,ii ,t
Proposition 1: The optimal energy beamforming vector s.t. C17, C18, C19, C20, C21, (P5)
during τ1 is given by ŵ ∗1,i = h 0,1 /||h 0,1 ||.
The proof of Proposition 1 can be done by contra- where ŵ ii = [ŵ 0,ii , ŵ 2,ii ] and t ∗ = [t0∗ , t1∗ , t2∗ ].
Following the similar approach for Case i , Problem
diction theory and is omitted due to the limited space.
P5 can be solved as follows. We introduce auxiliary
Then, based on Proposition 1, C6 and C7 are recast as
variables R̄hd,ii , e0,ii , e1,ii , and let e2,ii = P1,ii t1 ,
C15: R̄bd,i ≤ τ1 log2 (1 + ξP |α0,i |2 |h1,3 |2 ||h 0,1 ||2 /σap 2 )+
2 2 W ii = t0 ŵ 0,ii ŵ H0,ii . Then, we introduce the following
τ2 log2 (1 + ξe1,i |h2,3 | /(σap τ2 )) and C16: R̄bd,i ≤
new constraints C22: t1 log2 (1 + ξe2,ii |h2,3 |2 /(σap 2 t )) +
τ1 log2 (1 + ξP |α0,i |2 |h1,2 |2 ||h 0,1 ||2 /σhd 2 ), respectively. With
2 2 H 2 2
1
Proposition 1, P2 is still non-convex due to the rank-one t2 log2 (1 + ξP |α2,ii | |h1,3 | |h 0,1 ŵ 2,ii | /σap ) ≥ R̄hd,ii ,
constraint. The SDR technique is an efficient approximation C23: t1 log2 (1+ξe2,ii |g1,2 |2 /(σbd
2 t )) ≥ R̄
1 hd,ii , C24: e0,ii ≤
H
Tr(h 0,2 h 0,2 W ii ), C25: e1,ii ≤ Tr(h 0,1 h H
technique to convert the non-convex problem to a convex 0,1 W ii ),
problem [11]. By relaxing C14 following SDR, P2 is recast C26: Rbd,ii = t0 log2 (1 + ξP |α0,ii |2 |h1,3 |2 e1,ii /(σap 2 t )),
0
as follows: C27: e2,ii ≤ ηPe0,ii , C28: Tr(W ii ) ≤ τ0 , C29: W ii 0,
and C30: rank (W ii ) = 1. Then, P5 is recast as:
max ω1 (R̂bd,i + R̄bd,i ) + ω2 Rhd,i
W i ,e i ,τ ,R̂bd,i ,R̄bd,i ,Rhd,i
max ω1 Rbd,ii + ω2 R̄hd,ii
s.t. C1, C2, C8, C9, C10, C11, C12, C13, C15, C16. (P3) W ii ,ŵ 2,ii ,e ii ,t,Rbd,ii ,R̄hd,ii
s.t. C17, C18, C21, C22, C23, C24,
Proposition 2: Problem P3 is a convex problem [12].
According to Proposition 2, the optimal solution for P3 C25, C26, C27, C28, C29, C30,
can be solved by some standard optimization techniques. In (P6)
LYU et al.: USER COOPERATION IN WIRELESS-POWERED BACKSCATTER COMMUNICATION NETWORKS 635
Abstract—Tag cardinality estimation is one of the most crucial effect (CE) [5]. The observed state of the slot, in this case, is
issues in radio frequency identification technology. The issue, assumed to be singleton to differentiate from one-response. In
however, usually faces with challenges in wireless fading environ- addition, a tag might not be detected with a probability in the
ments due to the presence of the so-called capture effect (CE)
and detection error (DE). The aim of this letter is to provide an corresponding one-response slot, which is referred to as the
efficient and accurate estimation method to cope with the CE detection error (DE) [6]. Similarly, the observed state of the
and DE using expectation-maximization algorithm and the stan- slot is called empty, while in other cases, the state is observed
dard Aloha-based protocol. We show that the proposed method as collision. These phenomena has been extensively studied
gives more accurate estimates than a conventional one. Thanks in the literature of RFID both in theoretical and experimental
to this fact, the Aloha frame size used for the tag identification
process can also be optimally selected so that the identification aspects [6]. They are usually hidden from the reader, and there-
efficiency can be improved. Computer simulations are presented fore, affect the estimation accuracy of conventional methods.
to confirm the merit of the proposed method. Several works have been proposed to deal with the cardinal-
Index Terms—RFID, Aloha, capture effect, detection error, EM ity estimation in the presence of the CE [7]–[9]. The method
algorithm, estimation. in [7], i.e., capture-aware backlog estimation (CMEBE), esti-
mates the tag cardinality and the CE probability by minimizing
the norm-2 distance between theoretical and observed vectors
I. I NTRODUCTION of empty, singleton and collision slots. In [8], they are found by
AG CARDINALITY estimation holds a crucial task in using Bayesian approach. Also, in [9], the capture probability
T Radio Frequency Identification (RFID) technology with
many practical applications such as intelligent transporta-
is analyzed in a more accurate approach by considering the
number of contention tags in a time slot and physical layer
tion, indoor stadium, and warehouse systems. The task has parameters. Thanks to the approach, a closed-form solution
been investigated in several previous works [1], [2] with a of the optimal frame length is found, which then improves
frame slotted Aloha (FSA) protocol, which is originally and the identification performance. The key limitation of those
standardly used to detect RFID tags’ Identity (ID) [3]. In works, nevertheless, is that the DE is completely ignored. On
those works, tags randomly transmit their IDs in a frame of the other hand, the cardinality estimation in the presence of
time slots. Then, observations of the number of responses both the CE and DE was recently studied in [10] based on
in each slot, i.e., no response, one response, and multiple the maximum likelihood (ML) approach. In the approach, an
responses [4], can be utilized for the tag cardinality estima- approximation of the likelihood function of the tag cardinal-
tion. The tags’ ID identification process can be significantly ity, CE and DE probabilities, for given observations of slots is
improved with an accurate estimate of the cardinality. determined. Nevertheless, to maximize the likelihood function,
On the other hand, under effects of wireless channel impair- the method adopted an exhaustive search algorithm to check
ments, the observations of time slots may not accurately reflect all possible values of the tag cardinality and the probabili-
the real number of responses. Indeed, due to the channel fading ties. This approach thus resulted in a very high computational
phenomenon, a tag might be detected with a probability in a complexity, and affected the overall performance of the iden-
multiple-response slot, which is well-known as the capture tification process. Although, the complexity could be reduced
with a simple transmission model such as flat Rayleigh fading
Manuscript received October 9, 2018; revised November 23, 2018; accepted where a deterministic relation between CE and DE probabili-
December 7, 2018. Date of publication January 1, 2019; date of current ties could be obtained [10], it would be much more challenging
version April 9, 2019. This work was supported by JSPS Kakenhi under
Project 18K11269. The associate editor coordinating the review of this paper in practical ones.
and approving it for publication was K. Adachi. (Corresponding author: In this letter, we propose a new method employing the
Chuyen T. Nguyen.) Expectation-Maximization (EM) algorithm [11] and FSA pro-
C. T. Nguyen is with the Department of Telecommunication
Systems, School of Electronics and Telecommunications, Hanoi tocol to efficiently and accurately estimate the tag cardinality
University of Science and Technology, Hanoi 100000, Vietnam in presence of both CE and DE. The method includes iterative
(e-mail: chuyen.nguyenthanh@hust.edu.vn). estimation rounds. In each round, the cardinality is first esti-
V.-D. Nguyen is with the Institute of Research and Development, Duy
Tan University, Da Nang 550000, Vietnam, and also with the Department mated by ML approach given expected values of hidden
of ICMC Convergence Technology, Soongsil University, Seoul 06978, South data/observations caused by the CE and DE. The CE and
Korea (e-mail: nguyenvandinh@ssu.ac.kr). DE probabilities can then be found in closed-forms for the
A. T. Pham is with the Department of Computer Engineering, University
of Aizu, Aizuwakamatsu 965-8580, Japan (e-mail: pham@u-aizu.ac.jp). given estimate, which significantly reduces the computational
Digital Object Identifier 10.1109/LWC.2018.2890650 complexity in comparison with the method in [10]. Simulation
2162-2345 c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
NGUYEN et al.: TAG CARDINALITY ESTIMATION USING EXPECTATION-MAXIMIZATION IN ALOHA-BASED RFID SYSTEMS WITH CE AND DE 637
(4)
pE ≈ p0 + βp1 , pS ≈ (1 − β)p1 + αp2 , pC ≈ (1 − α)p2 , (1)
It is noted that, values of n, α and β in (4) are taken from
where p0 , p1 or p2 is a probability that a slot is, respec- the following M-step, while they can be initially set as S + 2C,
tively, no-response, one-response or multiple-response, i.e., 0.5, and 0.5, respectively.
p0 ≈ (1− L1 )n , p1 ≈ nL (1− L1 )n−1 , and p2 ≈ 1−p0 −p1 [10]. 2) M-Step: Given the complete data [E , S , C ; S 1 , C 1 ],
It is noted in (1) that the DE is assumed to not happen the likelihood function of n, α and β denoted by
in multiple-response slots due to signal diversity, which has f (E , S , C , S 1 , C 1 |n, α, β) is written as
also been validated in [10] under the assumption of a sim-
(E + S + C )!
ple Rayleigh fading channel model. More practical models of f E , S , C , S 1 , C 1 |n, α, β =
the DE, CE, and status of each slot should be investigated in (E − S 1 )!S 1 !(S − C 1 )!C 1 !C !
future works. Then, the estimates of n, α, and β denoted by × p0E −S 1 p01
S 1 S −C 1 C 1 C
p10 p12 pC , (5)
638 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
IV. C ONCLUSION
Fig. 4. RMSE of α, for L = 256.
This letter investigated the issue of tag cardinality estimation
with FSA protocol in RFID systems considering impacts of
both CE and DE. The EM approach was utilized to iteratively
estimate the tag cardinality, the CE, and DE probabilities.
Computer simulations confirmed that the proposed method
was guaranteed to converge after only a few iterations and
provided more accurate estimates than that of the conventional
methods. The proposed method was also proven to improve
the efficiency of the identification process.
R EFERENCES
[1] W. Gong, J. Liu, K. Liu, and Y. Liu, “Toward more rigorous and practical
cardinality estimation for large-scale RFID systems,” IEEE/ACM Trans.
Netw., vol. 25, no. 3, pp. 1347–1358, Jun. 2017.
[2] W.-T. Chen, “An accurate tag estimate method for improving the
performance of an RFID anti-collision algorithm based on dynamic
Fig. 5. The total number of slots used to detect n = 400 tags with respect frame length ALOHA,” IEEE Trans. Autom. Sci. Eng., vol. 6, no. 1,
to the DE (dash lines) or CE (solid lines) probability. pp. 9–15, Jan. 2009.
[3] EPC Radio-Frequency Identity Protocols Class-1 Generation-2
UHF RFID Protocol for Communications at 860 Mhz–960 Mhz
an exhaustive search algorithm over all possible values of n, α, Version 1.2.0. Accessed: May 8, 2016. [Online]. Available:
http://www.gs1.org/sites/default/files/docs/epc/uhfc1g2_1_2_0-standard-
and β. Therefore, [10] can be approximately considered as a 20080511.pdf
lower bound of all considered methods. We can see that while [4] C. Qian, H. Ngan, Y. Liu, and L. M. Ni, “Cardinality estimation for
our method can be comparable with CMEBE and Bayesian large-scale RFID systems,” IEEE Trans. Parallel Distrib. Syst., vol. 22,
no. 9, pp. 1441–1454, Sep. 2011.
when β = 0, it significantly outperforms them in term of the [5] H. Salah, H. A. Ahmed, J. Robert, and A. Heuberger, “A time and cap-
estimation accuracy when β > 0. This is because only the CE ture probability aware closed form frame slotted ALOHA frame length
has been taken into account in CMEBE and Bayesian meth- optimization,” IEEE Commun. Lett., vol. 19, no. 11, pp. 2009–2012,
Nov. 2015.
ods. Nevertheless, the performance of EM-based algorithms
[6] C. T. Nguyen, A.-T. H. Bui, V.-D. Nguyen, and A. T. Pham, “Modified
greatly depends on the initial values of estimated parameters. tree-based identification protocols for solving hidden-tag problem in
Indeed, it is seen in Fig. 4 the degraded performance of the RFID systems over fading channels,” IET Commun., vol. 11, no. 7,
proposed method for small values of α (α < 0.15). This fact pp. 1132–1142, May 2017.
[7] B. Li and J. Wang, “Efficient anti-collision algorithm utilizing the cap-
is also observed for α = 0.3, 0.4 where the performance of ture effect for ISO 18000-6C RFID protocol,” IEEE Commun. Lett.,
the proposed method is even better than that of [10]. vol. 15, no. 3, pp. 352–354, Mar. 2011.
We now provide the worst-case per-iteration complexity [8] H. Wu, Y. Wang, and Y. Zeng, “Capture-aware Bayesian RFID tag esti-
mate for large-scale identification,” IEEE/CAA J. Automatica Sinica,
analysis of Algorithm 1 and compare to that of [10]. Recall vol. 5, no. 1, pp. 119–127, Jan. 2018.
that the per-iteration complexity of the method presented [9] H. A. Ahmed, H. Salah, J. Robert, and A. Heuberger, “A closed-
in [10] is O(n 3 ). The complexity of Algorithm 1 is mostly due form solution for ALOHA frame length optimizing multiple collision
recovery coefficients’ reading efficiency,” IEEE Syst. J., vol. 12, no. 1,
to solving (7) (i.e., step 7 of Algorithm 1), which requires the pp. 1047–1050, Mar. 2018.
complexity of O(n). This is to say, given the same convergence [10] C. T. Nguyen, K. Hayashi, M. Kaneko, and H. Sakai, “Maximum
condition as in (3), the proposed algorithm requires signifi- likelihood approach for RFID tag cardinality estimation under capture
effect and detection errors,” IEICE Trans. Commun., vol. E96-B, no. 5,
cantly lower complexity, compared to that of [10], especially pp. 1122–1129, May 2013.
when n is large. [11] G. J. McLachlan and T. Krishnan, The EM Algorithm and Extensions,
Finally, we plot the total number of slots used to detect 2nd ed. Hoboken, NJ, USA: Wiley, 2008.
[12] E. Vahedi, V. W. S. Wong, I. F. Blake, and R. K. Ward, “Probabilistic
n = 400 tags with respect to different values of α (β is set analysis and correction of Chen’s tag estimate method,” IEEE Trans.
by 0.3) or β (α is set by 0.3) in Fig. 5 to see the impact of Autom. Sci. Eng., vol. 8, no. 3, pp. 659–663, Jul. 2011.
640 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
Abstract—This letter proposes a pilot subcarrier allocation estimation and suppression of PN induced CPE and ICI, with
approach for orthogonal frequency division multiplexing systems, low latency and low overhead. The accurate estimation and
which enables construction of symbol-wise phase-noise (PN) esti- suppression of PN enables the use of higher-order MCSs
mates with high efficiency and low overhead in a non-iterative allowing significantly improved spectral efficiency, particularly
manner. The complexity and performance of the overall PN sup- at higher cmW or mmW bands. No prior symbol deci-
pression algorithm together with the proposed pilot allocation sions or information about consecutive OFDM symbols are
approach are evaluated in 5G single-user-multiple-input multiple- needed, resulting in low latency and reduced buffering require-
output (MIMO) and multi-user-MIMO radio links at 28 GHz
ments. The block-wise pilot signal design, with the defined
carrier frequency, showing clear complexity-performance benefits
against a state-of-the-art reference algorithm. estimation and compensation algorithms, can be applied to
single-input single-output (SISO), single-user (SU) multiple-
Index Terms—OFDM, reference signals, MIMO, phase noise, input multiple-output (MIMO), and multi-user (MU)-MIMO
ICI, mitigation, computational complexity, cmW, mmW, 5G. OFDM links.
IV. P HASE N OISE M ITIGATION S TRATEGY IN and NLOS evaluation cases, respectively [9]. Per polarization,
MIMO-OFDM 8 × 8 antenna array is assumed at the BS and 2 × 2 array
The proposed algorithm can be used also in MIMO links. at the UE [9]. Different polarizations are used for different
It is reasonable to assume that in small mobile transceivers spatial MIMO streams. A 5G NR like subframe structure con-
a common oscillator is shared between the antennas, so the sisting of 14 OFDM symbols is assumed, where the first two
PN realization is always the same for each antenna port. In symbols are reserved for control. From the remaining 12 sym-
SU-MIMO, when the same pilot subcarrier block for PN esti- bols, three equally spaced symbols dedicated for demodulation
mation is sent from all the antenna ports of the transmitter, reference signals are used for channel estimation and the rest
the received pilot blocks can be averaged across the received carry user data and PN estimation pilots. In each OFDM sym-
streams after the channel equalization, and the PN estimation bol carrying user data, a contiguous block of b pilot subcarriers
can then be carried out. This is beneficial, because of relatively (proposed method) or 24 scattered, equidistant pilot subcarri-
low complexity compared to stream-by-stream estimation of ers (reference methods) are allocated for PN estimation. The
the PN and reduced noise variance. It is also possible to trans- actual channel estimation is based on the well-known MMSE
mit the PN estimation pilots only from a single antenna port solution. The channel estimates are interpolated in time and
to reduce the pilot overhead. In this case, the averaging gain frequency by using a Wiener interpolator after which a subcar-
is lost and the PN estimation accuracy is defined by the rier wise MMSE equalizer is applied on the data subcarriers,
inter-stream-interference level in high SNR regime. prior to PN mitigation. The radio link performance is evaluated
In multi-user uplink (UL), each user equipment (UE) by block-error ratio (BLER) and throughput. BLER target of
requires their own PN pilot block, because each UE has an 10% is assumed, which is common in mobile radio networks
independent PN realization. Averaging the PN pilots over all using hybrid automatic repeat request (HARQ) on top of the
streams from one UE is possible, assuming that one oscillator channel coding.
is shared by all the transmit antenna ports in the UE, and that The studied cases are: 1) ‘No PN’ case, where no PN
each UE transmits the PN pilot block in all streams. Sharing nor pilots are added, 2) ‘CPE m.’ case, where only CPE is
the same time-frequency resources by each UE in MU-MIMO mitigated [2], 3) ‘Pet’ case where, CPE and ICI are mitigated
is beneficial, because the interfering signal from each UE is in a way proposed in [2] and 4) ‘u;b’ cases where the proposed
known. If different resources are used, the PN pilot block of algorithm is used with parameterization u and b. The ‘Pet’
the desired UE is interfered by a random data by other UEs, algorithm estimates 3 frequency components of the PN from
unless the corresponding subcarriers are muted by other UEs. the both sides of the direct current (DC) bin, with 3 iterations.
112 subcarrier symbol decisions are used per iteration to con-
struct the PN estimate. These decisions are made so that the
V. S IMULATION R ESULTS AND A NALYSIS most reliable subcarriers are used based on the estimated chan-
A. Simulation Scenario and Simulator Description nel amplitude response. These parameters enable reasonable
performance, but the complexity is still significantly higher
The studied scenario follows closely the 3GPP 5G New
than in the proposed algorithm.
Radio (NR) physical layer standardization in [9] and [10], and
For both SU-MIMO and MU-MIMO cases, simulations
thus represents a very timely example. In the simulator, an
are carried out with MATLAB based 5G NR specifications
OFDM signal is generated with fast Fourier transform (FFT)
compliant tool. We consider 2x2 SU-MIMO DL with spatial
length of 2048 and with 1284 active subcarriers. For the active
multiplexing of 2 streams. The SU-MIMO results are evalu-
subcarriers, 64QAM is used as a representative example of
ated with code rates of 3/4 and 4/5 for 64QAM, and turbo
fairly high modulation order. The center frequency is 28 GHz,
codec implementation following [14]. In the decoder, 8 max-
and the subcarrier spacing is 60 kHz, while the total carrier
log-MAP decoding iterations are used. In the DL SU-MIMO
bandwidth is 80 MHz. We use cyclic prefix of 144 samples (at
scenario, the transmitted pilot blocks are averaged across the
baseband sampling rate of 122.88 MHz). We incorporate both
spatial streams. In MU-MIMO, we consider UL case with sim-
the transmitter and receiver PNs that are generated by a charge-
ilar parameters as in the SU-MIMO case, but now there are
pump PLL oscillator [11] tuned to give similar power spectral
2 UEs each utilizing individual spatial stream, so we cannot do
density as used in 5G standardization evaluations [12], with
any averaging as described in Section IV. We assume 3 km/h
PSD shown in Fig. 2. The radio channel modeling is based
UE mobility for all the cases.
on the clustered delay line (CDL) C and D multipath chan-
nel models [13]. The non-line-of-sight (NLOS) CDL-C and
line-of-sight (LOS) CDL-D channels are with the root-mean- B. Results and Analysis
squared (RMS) delay spreads of 300 ns and 50 ns, respectively. In Fig. 3, 2 × 2 SU-MIMO DL case is considered for
In the CDL-D channel, the K-factor is 9 dB. These are selected 64QAM with code rate 3/4 in CDL-C (right set of curves)
since they correspond to evaluation assumptions for cmW and CDL-D (left set of curves) channels. With the proposed
radio link performance in 5G standardization [9]. CDL-D algorithm, only the DC bin and either one (u = 1) or two
50 ns and CDL-C 300 ns offer typical small cell LOS (u = 2) frequency components from the both sides of the
SYRJÄLÄ et al.: PILOT ALLOCATION AND COMPUTATIONALLY EFFICIENT NON-ITERATIVE ESTIMATION OF PN IN OFDM 643
VI. C ONCLUSION
When moving towards cmW and mmW frequencies in the
emerging OFDM-based 5G and beyond systems, both the CPE
and ICI induced by the oscillator PN needs to be considered
Fig. 3. Block-Error Rate as a function of signal-to-noise ratio in 2 × 2 SU- and potentially suppressed through digital signal processing.
MIMO DL case for 64QAM (3/4) in CDL-C and CDL-D channels with 3 km/h
mobility. In the legend entries for the proposed algorithm, u;b defines the
In this letter, to facilitate symbol-by-symbol PN estimation
number of estimated ICI components per side and the pilot block length. and CPE+ICI compensation, a block type pilot allocation
was proposed. Based on the proposed pilot allocation, a low-
TABLE I latency and computationally very efficient, yet highly accurate
T HROUGHPUTS [M BPS ] IN CDL-D C HANNEL W ITH 3 KM / H M OBILITY. PN mitigation algorithm was formulated. It was shown through
L EFT-H AND S IDE C OLUMNS AT 25.5- D B AND 64QAM (3/4) C OMPARE radio link simulations at 28 GHz carrier frequency that despite
SU-MIMO AND MU-MIMO C ASES , AND R IGHT-H AND S IDE C OLUMNS its simplicity, the proposed algorithm outperforms the existing
AT 26.5- D B C OMPARE D IFFERENT C ODE R ATES IN SU-MIMO
state-of-the-art and can suppress the PN effects near to the PN
free reference case in base station and mobile transceivers. The
proposed scheme enables the use of higher MCSs, and was
shown to be able to significantly improve the overall radio
link throughput despite the involved pilot overhead.
R EFERENCES
[1] T. C. W. Schenk, R. W. Van Der Hofstad, E. R. Fledderus, and
P. F. M. Smulders, “Distribution of the ICI term in phase noise
DC bin are estimated, and a single pilot block with vary- impaired OFDM systems,” IEEE Trans. Wireless Commun., vol. 6, no. 4,
ing amounts of subcarriers is used. In CDL-D case, some pp. 1488–1500, Apr. 2007.
[2] D. Petrovic, W. Rave, and G. Fettweis, “Effects of phase noise on OFDM
improvement over the CPE-only mitigation (‘CPE m.’) can be systems with and without PLL: Characterization and compensation,”
achieved with the ‘Pet’ reference algorithm of [2] at higher IEEE Trans. Commun., vol. 55, no. 8, pp. 1607–1616, Aug. 2007.
[3] V. Syrjälä and M. Valkama, “Analysis and mitigation of phase noise
SNR regime. With the proposed algorithm in 2 × 2 SU- and sampling jitter in OFDM radio receivers,” Int. J. Microw. Wireless
MIMO case with 1;12, 1;24, and 2;24 configurations, the Technol., vol. 2, no. 2, pp. 193–202, Apr. 2010.
difference to the ‘no PN’ case is only around 1.7 dB, 1.2 dB [4] N. N. Tchamov, et al., “Enhanced algorithm for digital mitigation of
ICI due to phase noise in OFDM receivers,” IEEE Wireless Commun.
and 0.7 dB, respectively, for the BLER target of 10%. Clearly Lett., vol. 2, no. 1, pp. 6–9, Feb. 2013.
in 2;12 case, the used 12 subcarrier block is too small for [5] V. Syrjälä and M. Valkama, “Iterative receiver signal process-
the u = 2 case, as it gives only performance similar to the ing for joint mitigation of transmitter and receiver phase noise
in OFDM-based cognitive radio link,” in Proc. Int. Conf. Cogn.
reference algorithm. The CDL-C case is clearly more demand- Radio Orient. Wireless Netw., Stockholm, Sweden, Jun. 2012,
ing for all the algorithms due to more frequency selective doi: 10.4108/icst.crowncom.2012.248513.
[6] P. Mathecken, T. Riihonen, S. Werner, and R. Wichman, “Phase noise
channel. Compared to the ‘CPE m.’ case, the considered estimation in OFDM: Utilizing its associated spectral geometry,” IEEE
1;12 configuration offers performance improvement at the Trans. Signal Process., vol. 64, no. 8, pp. 1999–2012, Apr. 2016.
higher SNR range. Furthermore, the 1;24 configuration pro- [7] P. Rabiei, W. Namgoong, and N. Al-Dhahir, “A non-iterative technique
for phase noise ICI mitigation in packet-based OFDM systems,” IEEE
vides performance already within 1 dB of the ideal ‘no PN’ Trans. Signal Process., vol. 58, no. 11, pp. 5945–5950, Nov. 2010.
case at the 10% BLER target. With u = 2, the performance [8] J. Gozalvez, “5G worldwide developments,” IEEE Veh. Technol. Mag.,
is slightly degraded and a larger pilot block size should be vol. 12, no. 1, pp. 4–11, Mar. 2017.
[9] “Study on new radio access technology; physical layer aspects, v14.2.0,”
applied to benefit from larger u. 3GPP, Sophia Antipolis, France, Rep. TR 38.802, Sep. 2017.
In the left two columns of Table I, throughput results are [10] NR; Physical Layer Procedures for Data, V15.1.0, 3GPP Standard TS
given for 2 × 2 SU-MIMO and 2 × 2 MU-MIMO cases 38.214, Apr. 2018.
[11] N. N. Tchamov, Circuit- and System-Level Design of OFDM Receivers
at 25.5-dB SNR in CDL-D channel. Notice that MU-MIMO in the Presence of Phase Noise, D.Sc. dissertation, Dept. Electron.
throughput is evaluated as aggregated throughput of the Commun. Eng., Tampere University of Technology, Tampere, Finland,
2013.
two MU-MIMO users. The performance of the proposed [12] “Study on new radio access technology; RF and co-existence aspects,
algorithm is very good, in general, also in MU-MIMO, while v14.2.0,” 3GPP, Sophia Antipolis, France, Rep. TR 38.803, Sep. 2017.
the computational complexity stays very low. Furthermore, [13] “Study on channel model for frequencies from 0.5 to 100 GHz, v15.0.0,”
3GPP, Sophia Antipolis, France, Rep. TR 38.901, Jun. 2018.
the overhead is very small, e.g., with a block of 12 pilots [14] E-UTRA Multiplexing and Channel Coding, V14.5.1, 3GPP Standard
it is only 0.93%, and even with larger block size of 24, TS 36.212, Jan. 2018.
644 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
where α is the path loss exponent and γc is RF signal to direct In the steady state in which all stations successfully transmitted
current (DC) energy conversion rate. at least once, each station can transmit data with the backoff
If stations are uniformly deployed within an area with the values without collision. Even if there is no central coordi-
radius of Rmax around the PB (i.e., the Probability Density nator, it is possible for a station to transmit data by its own
Function (PDF) fR (r ) = R2r 2 , 0 ≤ r ≤ Rmax ), one can transmission schedule. If the number of stations accessing the
max
derive the PDF of Ehar when α = 2 and γc = 1. network exceeds ωd , collisions may occur in CSMA/ECA. To
P te 1 PTX te solve this problem, research has been conducted to allow the
2 , ≤ ehar < PTX te
TX
2
Rmax ehar 2
Rmax
fEhar (ehar ) = (2) stations to change the contention window size by themselves
1 , ehar = PTX te .
R2 with the assistance of the AP, i.e., CSMA/ECA adjust-CW. In
max
the CSMA/ECA adjust-CW, the AP determines the contention
Since we consider a discrete random variable for Ehar , it can
window size ωmin , based on the number of connected stations
be expressed as a discrete value by quantizing Ehar .
n, and informs it by beacon frames.
d
Ehar = i , if iecon ≤ Ehar ≤ (i + 1)econ , 1 ≤ i ≤ kmax , (3)
qn, qn mod 2 = 1,
ωmin = (7)
where kmax is the maximum number of quantization levels. qn + 1, otherwise,
Then, we can derive the Probability Mass Function (PMF) of
where q is a control parameter (q ≥ 2).
the harvested energy as follows.
(i+1)econ
d C. Proposed MAC Protocol
P[Ehar = i] = fEhar (ehar )dehar , 1 ≤ i ≤ kmax . (4)
iecon
In the proposed protocol, a PB senses the channel status and
The amount of stored energy in a station can be modeled by participates in the backoff competition as the other stations
a random variable Esto . A discrete random variable for Esto do. After winning the competition, the PB can transmit the
can be derived as energy signal to the stations without collisions or interference.
d The average energy signal transmission opportunity of the PB
Esto = j , if (j − 1)econ ≤ Esto ≤ jecon , 1 ≤ j ≤ M . (5)
is adjusted according to the energy states of the stations by
We assume that M > kmax + 1. multiplying the original ωd by an integer factor r. Then, while
A massive number of devices are expected to be deployed the PB can transmit the RF energy signal once and the other
in a wide area. And, it may not be viable to cover such a data stations can transmit data signals r times.
wide area with a single PB since the amount of power trans-
fer decreases due to increased distance. Thus, PBs can be ωd,PB = r (ωmin + 1)/2. (8)
deployed to supply energy to the nodes in sub-areas. And, each
The parameter r is determined according to the energy level
PB is responsible for a small sub-area. If PTX and Rmax of
of the stations. If the energy level of the stations is higher
PBs are the same and sub-areas have the same distribution of
than a high threshold ζhigh , r = rhigh , i.e., ωd,PB =
Ehar , fEhar (ehar ), it is equivalent to a network with a single
rhigh (ωmin + 1)/2. If the energy level is lower than a low
PB in a sub-area.
threshold ζlow , r = rlow , i.e., ωd,PB = rlow (ωmin + 1)/2.
Otherwise r = rmid , i.e., ωd,PB = rmid (ωmin + 1)/2.
B. Carrier Sense Multiple Access With Enhanced Collision
The parameters rlow , rmid , rhigh , ζhigh , and ζlow are
Avoidance (CSMA/ECA)
determined based on the amount of energy consumption
CSMA/CA protocol is a distributed MAC protocol for econ and the average amount of harvested energy Ehar d . The
WLAN. Before transmitting data, a station senses the channel. CSMA/ECA-based stations can transmit data signals r times
If the channel is idle during the DCF Interframe Space (DIFS), more than the PB does and consume r units energy. The
it randomly selects a backoff counter value in [0, 2k (ωmin + stations harvest Ehard energy from the PB. When the sta-
1) − 1], where ωmin is the minimum contention window and tion’s energy level is ζ, the station’s energy level becomes
k ∈ [0, m] is the backoff stage. The backoff counter decreases ζ = ζ − r + Ehard . If a station is in the low energy state, it
by one if the channel is idle during a slot time. When the back- requires quick energy harvesting. Then the average transmis-
off counter value reaches zero, the station transmits data. For sion opportunity of the PB is set to the same as that of the
each unsuccessful transmission by collision, a station increases other stations, i.e., rlow = 1. Thus, ζ will be at least equal
the backoff stage by one up to the maximum backoff stage m to or greater than ζ. In the medium energy state, in order
and repeats the backoff procedure. This algorithm is known to balance the amount of the harvested energy Ehar d and the
as a Binary Exponential Backoff (BEB) mechanism. consumed energy rmid ,
The difference between CSMA/ECA and CSMA/CA is that
d
a station which successfully transmitted data selects a deter- rmid = E har (9)
ministic backoff value and keeps the backoff value afterwards
d
in CSMA/ECA. When a station succeeds in data transmission where · is a rounding function and E har is the average of
and receives an Acknowledgment (ACK) frame, it sets the d .
a random variable Ehar
backoff stage to 0 and selects a deterministic backoff value
k
ωd instead of a random backoff value as in CSMA/CA, d d
max
d
E har = E[Ehar ]= i P[Ehar = i ]. (10)
ωd = (ωmin + 1)/2. (6) i=1
646 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
TABLE I
S YSTEM PARAMETERS waste of energy even for a small number of stations. Since
W2 P-MAC transmits energy signal without considering the
energy states of stations, energy may be wasted and frequent
collisions may occur. Thus unused amount of energy in
W2 P-MAC is larger than that in the proposed mechanism.
The energy efficiency of the stations is shown in Fig. 3 (c).
With the introduction of the PB, WIT and the energy efficiency
may be reduced. Thus, WPT and WIT have a trade-off. In the
proposed protocol, to balance the WPT and WIT, the PB can
control the WPT opportunity according to the energy levels of
the stations. The proposed protocol has almost twice as high
energy efficiency as those of the others due to the control of
the PB’s WPT opportunities.
IV. C ONCLUSION
In this letter, we have proposed a distributed MAC protocol
for both WPT and WIT based on CSMA/ECA in wire-
less powered communication networks. The proposed protocol
attempts to coordinate data and energy transmissions. It not
only reduces the amount of wasted energy but also improves
throughput and energy efficiency by appropriately controlling
the energy transmission opportunity of PB according to the
energy level of stations. The proposed protocol is shown to
improve the power availability of devices and allows efficient
data transfer and power transmission in a wireless network.
As future research, WPT with beamforming could be con-
Fig. 3. Performances of the proposed MAC: (a) Throughput for varying sidered. If the WPT with beamforming technique is applied,
numbers of stations, (b) Wasted energy due to energy transfer by the PB, and performances is expected to be further improved.
(c) Energy efficiency of the stations due to energy harvesting.
R EFERENCES
[1] W. Ejaz, M. Naeem, A. Shahid, A. Anpalagan, and M. Jo, “Efficient
B. Simulations energy management for the Internet of Things in smart cities,” IEEE
Commun. Mag., vol. 55, no. 1, pp. 84–91, Jan. 2017.
In order to evaluate the performance of the proposed pro- [2] P. Kamalinejad et al., “Wireless energy harvesting for the Internet of
tocol, we compare the throughput, wasted energy, and energy Things,” IEEE Commun. Mag., vol. 53, no. 6, pp. 102–108, Jun. 2015.
efficiency of the proposed protocol with those of W2 P-MAC [3] T. X. Doan, T. M. Hoang, T. Q. Duong, and H. Q. Ngo, “Energy
harvesting-based D2D communications in the presence of interference
and the proposed protocol without adjusting PB’s transmission and ambient RF sources,” IEEE Access, vol. 5, pp. 5224–5234, 2017.
opportunity. The wasted energy is defined as the ratio between [4] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, “Wireless networks
total amount of wasted energy and total time, and the energy with RF energy harvesting: A contemporary survey,” IEEE Commun.
Surveys Tuts., vol. 17, no. 2, pp. 757–789, 2nd Quart., 2015.
efficiency is defined as the total number of data bits transmit- [5] J. Ren et al., “RF energy harvesting and transfer in cognitive radio sensor
ted divided by the total energy transmitted by the PB. The networks: Opportunities and challenges,” IEEE Commun. Mag., vol. 56,
parameters used in the simulation and the analysis are shown no. 1, pp. 104–110, Jan. 2018.
[6] S. Lee, L. Liu, and R. Zhang, “Collaborative wireless energy and infor-
in Table I [11]. mation transfer in interference channel,” IEEE Trans. Wireless Commun.,
Throughput performance in Fig. 3(a) increases as the num- vol. 14, no. 1, pp. 545–557, Jan. 2015.
ber of stations increases, which stems from the fact that the [7] R. Gupta, A. K. Chaturvedi, and R. Budhiraja, “Improved rate-energy
tradeoff for energy harvesting interference alignment networks,” IEEE
time interval for the PB decreases as the number of stations Wireless Commun. Lett., vol. 6, no. 3, pp. 410–413, Jun. 2017.
increases. Since the stations retransmit data due to collisions [8] Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
in W2 P-MAC, throughput is the lowest. Proposed schemes Specifications, IEEE Standard 802.11, Dec. 2016.
[9] L. Sanabria-Russo, J. Barcelo, B. Bellalta, and F. Gringoli, “A high
show higher performance than W2 P-MAC because there are efficiency MAC protocol for WLANs: Providing fairness in dense
no retransmissions due to collisions. When the PB adjusts scenarios,” IEEE/ACM Trans. Netw., vol. 25, no. 1, pp. 492–505,
transmission opportunity, it can prevent unnecessary energy Feb. 2017.
[10] H. Lee, Y. Kim, J. H. Ahn, M. Y. Chung, and T.-J. Lee, “Wi-Fi and
transmission so that the channel can be used about 7% more wireless power transfer live together,” IEEE Commun. Lett., vol. 22,
for data transmission. no. 3, pp. 518–521, Mar. 2018.
Fig. 3(b) shows the wasted energy rate. As the number [11] W. Cheng, X. Zhang, and H. Zhang, “Full-duplex spectrum-sensing
and MAC-protocol for multichannel nontime-slotted cognitive radio net-
of stations increases, the wasted energy tends to decrease. works,” IEEE J. Sel. Areas Commun., vol. 33, no. 5, pp. 820–831,
However, the proposed protocol can significantly reduce the May 2015.
648 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 8, NO. 2, APRIL 2019
RN fall outside the admissible rate region RDF-IE . Thus, probability is an upper bound on the outage probability of
using [1, eq. (4)], we have DF-IE. This bound cannot be solved in closed form, requiring
1 1 numerical evaluation.
out
PDF -IE,N = Pr φ(ΓSD ), φ(ΓFL D ) We verified for the numerical examples presented in
Rc Rc
[1, Fig. 2a] that the upper bound of DF-IE in (3) proves
∈/ RDF-IE ΓSFL (2) marginally different from the (incorrect!) outage probability
1 given in [1, eq. (9)]. The curves barely change. In particular,
< Pr ψ(ΓSD ) ≤ h pi (ΓSFi ) ∗ κl the conclusion drawn in [1] still holds true: DF-IE outperforms
Rc
i∈L conventional DF in terms of outage probability, becoming
− η(pL (ΓSFL ), κL ) ∪ more advantageous as more relays are used.
1
ψ(ΓFi D ) ≤ η(pL (ΓSFL ), κL ) R EFERENCES
Rc
∀S⊂L i∈S [1] A. Wolf, D. C. González, M. Dörpinghaus, L. L. Mendes,
J. C. S. Santos Filho, and G. Fettweis, “Outage analysis for decode-and-
− η(pS c (ΓSFS c ), κS c ) − h(κi ) ∪ forward multirelay systems allowing intra-link errors,” IEEE Wireless
i∈S Commun. Lett., vol. 6, no. 6, pp. 758–761, Dec. 2017.
1 [2] A. Wolf, D. C. González, M. Dörpinghaus, J. C. S. Santos Filho,
ψ(ΓFi D ) ≤ 1 + η(pL (ΓSFL ), κL ) − h(κi ) , and G. Fettweis, “On the many-help-one problem with
Rc independently degraded helpers,” 2017. [Online]. Available:
i∈L i∈L
http://arxiv.org/abs/1701.06416v1
S c = L\S, κL ∈ [0, 0.5](N −1) . (3) [3] A. Wolf, D. C. González, M. Dörpinghaus, J. C. S. Santos Filho,
and G. Fettweis, “On the binary lossless many-help-one problem with
independently degraded helpers,” in Proc. IEEE 56th Annu. Allerton
In (3), we substituted the rate constraints from (1). Because (1) Conf. Commun. Control Comput. (Allerton), Monticello, IL, USA,
represents a subset of the admissible rate region, the resulting Oct. 2018, pp. 1–5.
IEEE COMMUNICATIONS SOCIETY
2019 Board of Governors
O FFICERS
President President-Elect VP-Member & Global Activities VP-Publications
K. L ETAIEF V. C HAN N. K ATO X. S HEN
Hong Kong Univ. Science & Tech. MIT Graduate School of Info. Sci. Univ. of Waterloo
Executive Director/Secretary
S.M. B ROOKS, IEEE ComSoc
A PPOINTED O FFICERS
Chief Information Officer Treasurer Director—Educational Services Director—Magazines Director—Member Services Director—Sister & Related Societies
Z. D ING F. TAKAWIRA F. G RANELLI M. FANG S. G UO O. D OBRE
Univ. of California Univ. of Witwatersrand Univ of Trento Memorial Univ.
Director—Tech. Services Director—NA Region
Chief Marketing Officer Director—AP Region Director—EMEA Region T. TALEB W. A LMUHTADI Director—Standards Dev.
R. F ISH S. BAHK A. K SENTINI T. E L -BAWAB
NETovations, LLC Director—Industry Communities Director—On-Line Content Jackson State Univ.
Director—Conf. Dev. Director—Journals I. W ONG Z. N IU
Parliamentarian J. RODRIGUES R. S CHOBER Tsinghua Univ. Director—Standardization
R. D E M ARCA Univ. of Beira Interior Friedrich-Alexander Univ. Director—Industry Outreach Programs Development
Center for Telecommunication A. D UTTA, AT&T N. G OLMIE
Studies Director—Conf. Operations Director—LA Region National Inst. of Standards & Tech.
H. S ARI L. Z AMBENEDETTI
Nanjing Univ. of Posts and
Telecommunications
C OMMITTEE C HAIRS
Awards Dist. Lecturers Selection Finance GITC Strategic Planning
C. X IAO S. M AO F. TAKAWIRA M. VALENTI V. C HAN
Lehigh Univ. Univ. of Witwatersrand West Virginia Univ. MIT
Emerging Technologies
Communications History J. A NDREWS Governance Add Marketing Technical Committees Recertification
D. M ICHELSON Univ. of Texas at Austin M. H ARTMANN R. F ISH - NETovations, LLC N. F ONSECA
Univ. of British Columbia State Univ. Campinas
Fellow Evaluation GIMS Nominations & Elections
ComSoc Young Professional S. B ENEDETTO
N. R AMIREZ K. A SATANI TBD WICE
Politecnico de Torino Kogakuin Univ. A. G ARCIA A RMADA
Operations & Facilities
K. L ETAIEF
Hong Kong Univ. Science & Tech.
Transactions on Commun. Global Commun. Newsletter Communications Systems Network Operations & Mgmt. Wireless Communications
N. A L -D HAHIR, EIC S. B REGNI Integration & Modeling L. G RANVILLE W. Z HANG
Univ. of Texas at Dallas Politecnico di Milano C. V ERIKOUKIS Fed. Univ. Rio Grande do Sul Univ. of South Wales