[Journal] IEEE Transactions on Microwave Theory and Techniques. Vol. 64. No 7B

Citation preview

JULY 2016

VOLUME 64

NUMBER 7

IETMAB

(ISSN 0018-9480)

PART II OF TWO PARTS

SPECIAL ISSUE ON 5G WIRELESS COMMUNICATION SYSTEMS AND TECHNOLOGIES Guest Editorial ...................................................................................... D. Choudhury and T. Inoue

2205

SPECIAL ISSUE PAPERS

3-D Millimeter-Wave Statistical Channel Model for 5G Wireless System Design .......................................... ............................................................................................. M. K. Samimi and T. S. Rappaport Hybrid RF and Digital Beamformer for Cellular Networks: Algorithms, Microwave Architectures, and Measurements ....................................................................................... V. Venkateswaran, F. Pivit, and L. Guan 2-D Beam-Steerable Integrated Lens Antenna System for 5G E-Band Access and Backhaul ...... J. Ala-Laurinaho, J. Aurinsalo, A. Karttunen, M. Kaunisto, A. Lamminen, J. Nurmiharju, A. V. R¨ ais¨ anen, J. S¨ aily, and P. Wainio RF Lens-Embedded Massive MIMO Systems: Fabrication Issues and Codebook Design ................................. .............................................................................. T. Kwon, Y.-G. Lim, B.-W. Min, and C.-B. Chae A Multilayer LTCC Solution for Integrating 5G Access Point Antenna Modules ........................................... ............... F. Foglia Manzillo, M. Ettorre, M. S. Lahti, K. T. Kautio, D. Lelaidier, E. Seguenot, and R. Sauleau A 60-GHz Power Amplifier With AM–PM Distortion Cancellation in 40-nm CMOS .... S. Kulkarni and P. Reynaert

2207 2226 2244 2256 2272 2284

SPECIAL ISSUE ON RADIO-FREQUENCY IDENTIFICATION (RFID), SENSING, AND IMAGING Guest Editorial .............................................................................. A. Georgiadis and M. M. Tentzeris

2292

SPECIAL ISSUE PAPERS

High-Data-Rate Single-Chip Battery-Free Active Millimeter-Wave Identification Tag in 65-nm CMOS Technology .... .......................................................................... P. Burasa, T. Djerafi, N. G. Constantin, and K. Wu Wearable Flexible Lightweight Modular RFID Tag With Integrated Energy Harvester .................................... .......................................... S. Lemey, S. Agneessens, P. Van Torre, K. Baes, J. Vanfleteren, and H. Rogier

2294 2304

(Contents Continued on Back Cover)

(Contents Continued from Front Cover) Toward RCS Magnitude Level Coding for Chipless RFID ..................................................................... .................................................................... O. Rance, R. Siragusa, P. Lemaître-Auger, and E. Perret Temporal Separation Detection for Chipless Depolarizing Frequency-Coded RFID ........................................ ..................................................... A. Ramos, E. Perret, O. Rance, S. Tedjini, A. Lázaro, and D. Girbau Soil Moisture Scatter Radio Networking With Low Power ..................................................................... ....................................................... S.-N. Daskalakis, S. D. Assimonis, E. Kampianakis, and A. Bletsas Harmonic Power Harvesting System for Passive RFID Sensor Tags .......................................................... ........................................................... D. Allane, G. Andia Vera, Y. Duroc, R. Touhami, and S. Tedjini Bidirectional Analysis and Design of RFID Using an Additional Resonant Coil to Enhance Read Range ............. ........................................................................ P. Pérez-Nicoli, A. Rodríguez-Esteva, and F. Silveira

2315 2326 2338 2347 2357

IEEE MICROWAVE THEORY AND TECHNIQUES SOCIETY

The Microwave Theory and Techniques Society is an organization, within the framework of the IEEE, of members with principal professional interests in the field of microwave theory and techniques. All members of the IEEE are eligible for membership in the Society upon payment of the annual Society membership fee of $17.00, plus an annual subscription fee of $28.00 per year for electronic media only or $50.00 per year for electronic and print media. For information on joining, write to the IEEE at the address below. Member copies of Transactions/Journals are for personal use only. ADMINISTRATIVE COMMITTEE

K. W U, President A. A BUNJAILEH S. BARBIN

D. W ILLIAMS, President Elect

T. B RAZIL R. G UPTA

R. H ENDERSON W. H ONG

A. JACOB S. KOUL

J. L ASKAR G. LYONS

M. B OZZI, Secretary

M. M ADIHIAN S. PACHECO

Honorary Life Members T. I TOH R. S PARKS

G. P ONCHAK S. R AMAN

A. A BUNJAILEH , Treasurer

J. R AUTIO J. E. R AYAS -S ANCHEZ

S. R EISING M. S ALAZAR -PALMA

A. S ANADA D. S CHREURS

Distinguished Lecturers

P. S TAECKER

C. C AMPBELL R. H. C AVERLY G. C HATTOPADHYAY J.-C. C HIAO

T.-W. H UANG M. JARRAHI J. J. KOMIAK S. KOUL

A. M ORTAZAWI T. NAGATSUMA J. C. P EDRO L. P IERANTONI

M. S TEER

Past Presidents P. ROBLIN D. S CHREURS N. S HINOHARA

A. S TELZER J. W OOD H. Z IRATH

T. L EE (2015) R. W EIGEL (2014) M. G UPTA (2013)

MTT-S Chapter Chairs Albuquerque: E. FARR Argentina: A. M. H ENZE Atlanta: K. NAISHADHAM Austria: A. S PRINGER Baltimore: I. A HMAD Bangalore/India: K. V INOY Beijing: Z. F ENG Belarus: S. M ALYSHEV Benelux: G. VANDENBOSCH Boston: C. G ALBRAITH Bombay/India: M. V. P ITKE Brasilia: J. B EZERRA/ M. V INICIUS A LVES N UNES Buenaventura: C. S EABURY Buffalo: M. R. G ILLETTE Bulgaria: K. A SPARUHOVA Canada, Atlantic: Z. C HEN Cedar Rapids/Central Iowa: C. G. X IE Central & South Italy: L. TARRICONE Central No. Carolina: Z. X IE Central Texas: J. P RUITT Centro-Norte Brasil: M. V. A LVES N UNES Chengdu: Z. N EI Chicago: D. E RRICOLO Cleveland: M. S CARDELLETTI Columbus: A. O’B RIEN Connecticut: C. B LAIR Croatia: D. B ONEFACIC Czech/Slovakia: J. VOVES Dallas: R. S ANTHAKUMAR Dayton: A. T ERZUOLI Delhi/India: A. BASU

Denver: M. JANEZIC Eastern No. Carolina: T. N ICHOLS Egypt: E. H ASHEESH Finland: V. V IIKARI Florida West Coast: J. WANG Foothills: M. C HERUBIN France: D. BAJON Germany: G. B OECK Greece: R. M AKRI Gujarat/India: S. C HAKRABARTY Harbin: Q. W U Hawaii: K. M IYASHIRO Hong Kong: H. W ONG Houston: S. A. L ONG Houston, College Station: G. H. H UFF Hungary: L. NAGY Huntsville: H. S CHANTZ Hyderabad/India: S. R. N OOKALA India: D. B HATNAGER India/Kolkata: S. S ANKARALINGAM Indonesia: E. T. R AHARDJO Israel: S. AUSTER Japan: N. S UEMATSU Kansai: T. I SHIZAKI Kingston: S. P ODILCHAK Kitchener-Waterloo: R. R. M ANSOUR Lebanon: E. NASSAR Lithuania: B. L EVITAS Long Island/New York: S. PADMANABHAN Los Angeles, Coastal: V. R ADISIC Los Angeles, Metro/San Fernando: T. C ISCO

Macau: C. C. P ONG Madras/India: S. S ALIVAHANAN Malaysia: M. K. M. S ALLEH Malaysia, Penang: B. L. L IM Melbourne: R. B OTSFORD Mexican Council: R. M. RODRIGUEZ -DAGNINO Milwaukee: S. G. J OSHI Monterrey/Mexico: R. M. RODRIGUEZ -DAGNINO Morocco: M. E SSAAIDI Montreal: K. W U Morocco: M. E SSAAIDI Nagoya: J. BAE Nanjing: W. H ONG Nanjing, Hangzhou: L. S UN New Hampshire: E. H. S CHENK New Jersey Coast: J. S INSKY New South Wales: Y. R ANGA New Zealand: A. W ILLIAMSON North Italy: G. O LIVERI North Jersey: A. K. P ODDAR Northern Australia: J. M AZIERSKA Northern Canada: M. DANESHMAN Northern Nevada: B. S. R AWAT Norway: M. U BOSTAD Orange County: H. J. DE L OS S ANTOS Oregon: K. M AYS Orlando: K. K ARNATI Ottawa: Q. Z ENG Philadelphia: A. S. DARYOUSH Phoenix: S. ROCKWELL

D OMINIQUE S CHREURS KU Leuven B-3001 Leuven, Belgium

Editorial Assistants

M ARCIA H ENSLEY USA E NAS K ANDIL Belgium

Sweden: A. RYDBERG Switzerland: M. M ATTES Syracuse: D. M C P HERSON Taegu: Y.-H. J EONG Tainan: H.-H. C HEN Taipei: C. M ENG Thailand: C. P HONGCHAROENPANICH Toronto: G. V. E LEFTHERIADES Tucson: H. X IN Tunisia: A. G HARSALLAH Turkey: B. S AKA Twin Cities: C. F ULLER UK/RI: A. R EZAZADEH Ukraine, East: N. K. S AKHNENKO Ukraine, Kiev: Y. P ROKOPENKO Ukraine, Rep. of Georgia: K. TAVZARASHVILI Ukraine, Vinnitsya: V. M. D UBOVOY Ukraine, West: I. I VASENKO United Arab Emirates: N. K. M ALLAT Uttar Pradesh/India: M. J. A KHTAR Vancouver: S. M C C LAIN Venezuela: J. B. P ENA Victoria: K. G HORBANI Virginia Mountain: T. A. W INSLOW Washington DC/Northern Virginia: T. I VANOV Western Saudi Arabia: A. S HAMIM Winnipeg: P. M OJABI Xian: X. S HI

Associate Editors

Editors-In-Chief

J ENSHAN L IN Univ. of Florida Gainesville, FL 32611-6130 USA

Pikes Peak: K. H U Poland: W. J. K RZYSZTOFIK Portugal: J. C ALDINHAS VAZ Princeton/Central Jersey: W. C URTICE Queensland: K. B IALKOWSKI Rio de Janeiro: J. R. B ERGMANN Rochester: M. S IDLEY Romania: T. P ETRESCU Russia, Moscow: V. A. K ALOSHIN Russia, Nizhny-Novgorad: G. L. PAKHOMOV Russia, Novosibirsk: A. YAROSLAVTSEV Russia, Saratov/Penza: M. D. P ROKHOROV Russia, Saint Petersburg: S. P. ZUBKO Russia, Siberia: V. V. S UHOTIN Russia, Tomsk: D. Z YKOV San Diego: J. T WOMEY Santa Clara Valley/San Francisco: N. S HAMS Seattle: S. E BADI Seoul: C. S EO Serbia and Montenegro: B. M ILOVANOVI C´ Shanghai: J. M AO Singapore: Z. YANG South Africa: A. LYSKO South Australia: T. K AUFMANN South Brazil: J. R. B ERGMANN Southeastern Michigan: T. O ZDEMIR Southern Alberta: E. F EAR Spain: J. I. A LONSO Springfield: P. R. S IQUEIRA Sri Lanka: A. U. A. W. G UNAWARDENA St. Louis: D. BARBOUR

N UNO B ORGES C ARVALHO Universidade de Aveiro Aveiro, Portugal

X. C HEN Nat. Univ. Singapore Singapore

K AMRAN G HORBANI RMIT Univ. Melbourne, Vic., Australia

J ON M ARTENS Anritsu Morgan Hill, CA USA

O LGA B ORIC -L UBECKE Univ. of Hawaii at Manoa Manoa, HI USA

J.-C. C HIAO Univ. of Texas at Arlington Arlington, TX USA

ROBERTO G OMEZ -G ARCIA Univ. Alcala Madrid, Spain

F RANCISCO M ESA Universidad de Sevilla Seville, Spain

JAMES F. B UCKWALTER Univ. of California at Santa Barbara Santa Barbara, CA USA

A LESSANDRA C OSTANZO Univ. Bologna Bologna, Italy

J IASHENG H ONG Heriot-Watt Univ. Edinburgh, UK

L UCA P ERREGRINI Univ. of Pavia Pavia, Italy

S HENG -F UH R. C HANG Nat. Chung Cheng Univ. Chiayi County, Taiwan A. R IDDLE, Editor-in-Chief, IEEE Microwave Magazine N. S. BARKER, Editor-in-Chief, IEEE Microwave and Wireless Component Letters BARRY L. S HOOP, President K AREN BARTLESON, President-Elect PARVIZ FAMOURI, Secretary J ERRY L. H UDGINS, Treasurer H OWARD E. M ICHEL, Past President

G ILLES DAMBRINE Univ. of Lille Lille, France

T.-W. H UANG C ARLOS S AAVEDRA Nat. Taiwan Univ. Queen’s Univ. Taipei, Taiwan Kingston, ON, Canada J. S TAKE, Editor-in-Chief, IEEE Trans. Terahertz Science and Technology R. M IYAMOTO, Web Master

IEEE Officers

S. K. R AMESH, Vice President, Educational Activities S HEILA S. H EMAMI, Vice President, Publication Services and Products WAI -C HOONG W ONG, Vice President, Member and Geographic Activities B RUCE P. K RAEMER, President, Standards Association J OSE M. F. M OURA, Vice President, Technical Activities P ETER A LAN E CKSTEIN, President, IEEE-USA W ILLIAM W. M OSES, Director, Division IV—Electromagnetics and Radiation

IEEE Executive Staff D R . E. JAMES P RENDERGAST, T HOMAS S IEGERT, Business Administration J ULIE E VE C OZIN, Corporate Governance D ONNA H OURICAN, Corporate Strategy JAMIE M OESCH, Educational Activities E ILEEN M. L ACH, General Counsel & Chief Compliance Officer S HANNON J OHNSTON, Human Resources C HRIS B RANTLEY, IEEE-USA

Executive Director & Chief Operating Officer C HERIF A MIRAT, Information Technology PATRICK D. M AHONEY, Marketing C ECELIA JANKOWSKI, Member and Geographic Activities M ICHAEL F ORSTER, Publications KONSTANTINOS K ARACHALIOS, Standards Association M ARY WARD -C ALLAN, Technical Activities

IEEE Periodicals Transactions/Journals Department

Senior Director, Publishing Operations: F RAN Z APPULLA Director, Editorial Services: DAWN M ELLEY Director, Production Services: P ETER M. T UOHY Associate Director, Editorial Services: W ILLIAM A. C OLACCHIO Associate Director, Information Conversion and Editorial Support: K EVIN L ISANKIE Managing Editor: M ONA M ITTRA Senior Editor: C HRISTINA M. R EZES IEEE T RANSACTIONS ON M ICROWAVE T HEORY AND T ECHNIQUES (ISSN 0018-9480) is published monthly by the Institute of Electrical 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, Piscataway, NJ 08854-4141. NJ Telephone: +1 732 981 0060. Price/Publication Information: Individual copies: IEEE Members $20.00 (first copy only), nonmember $167.00 per copy. (Note: Postage and handling charge not included.) Member and nonmember subscription prices available upon request. 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 of $31.00 is paid through the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. For all other copying, reprint, or republication permission, c 2016 by The Institute of Electrical and Electronics Engineers, Inc. write to Copyrights and Permissions Department, IEEE Publications Administration, 445 Hoes Lane, Piscataway, NJ 08854-4141. Copyright  All rights reserved. Periodicals Postage Paid at New York, NY and at additional mailing offices. Postmaster: Send address changes to IEEE T RANSACTIONS ON M ICROWAVE T HEORY AND T ECHNIQUES, IEEE, 445 Hoes Lane, Piscataway, NJ 08854-4141. GST Registration No. 125634188. CPC Sales Agreement #40013087. Return undeliverable Canada addresses to: Pitney Bowes IMEX, P.O. Box 4332, Stanton Rd., Toronto, ON M5W 3J4, Canada. IEEE prohibits discrimination, harassment and bullying. For more information visit http://www.ieee.org/nondiscrimination. Printed in U.S.A.

Digital Object Identifier 10.1109/TMTT.2016.2576179

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

2205

Guest Editorial Special Issue on 5G Wireless Communication Systems and Technologies

E

NORMOUS growth of mobile wireless data service demands in combination with ever-increasing user expectations on quality of experience (QoE) is pushing the frontiers of wireless communication technologies, systems, and networks. Emerging wide area wireless services and usage cases are shaping the next-generation 5G wireless vision and driving the 5G technology requirements. Ultrahigh throughput, enhancement in network capacity, ultra-low latency, ubiquitous connectivity, energy efficiency, high reliability, low-cost devices, and quality of service are just some of the requirements that the next-generation 5G wireless network needs to achieve. At the core of such network are the microwave and millimeter wave technologies that the IEEE Microwave Theory and Techniques Society (IEEE MTT-S) contributes to. Although the emerging requirements for 5G are diverse and complex, there is an opportunity for the IEEE MTT-S to contribute to the larger 5G vision with different aspects of IEEE MTT-S relevant microwave and millimeter-wave technologies. In view of such needs, we have put together this special issue on 5G wireless communication systems and technologies with hopes to bring awareness of this exciting opportunity and help make the architectural evolution possible. The papers in

this Special Issue leverage on both old and new technologies, such as lens antenna, multilayer LTCC, and so on, combined with modern system concepts, such as massive MIMO and millimeter wave 3-D channel modeling in urban environment to support 5G wireless key metrics. We would like to extend special thanks to all the reviewers for their valuable time and effort. We also thank the T-MTT Editors-in-Chief, Jenshan Lin and Dominique Schreurs, as well as Associate Editor Olga Boric-Lubecke, for their support of this Special Issue.

D EBABANI C HOUDHURY, Guest Editor Intel Labs Intel Corporation Hillsboro, OR 97124 USA

TAKAO I NOUE, Guest Editor AWR Group National Instruments Austin, TX 78759-3504 USA

Debabani Choudhury (M’91–SM’06–F’11) received the Ph.D. degree in electrical engineering from IIT Bombay, Mumbai, India, in 1991. She was a Senior Research Staff Member with HRL Laboratories and the Millitech Corporation, where she developed various millimeter-wave and terahertz technologies for imaging and other space and defense applications. She was previously with NASA/JPL, where she was involved in THz/submillimeter-wave devices and components for space-based receiver applications. She has been with the Intel Corporation, Hillsboro OR, USA, since 2006. As a Principal Scientist with Intel Labs, she performs research and develops strategic directions on RF and millimeter-wave technologies and architectures for next-generation 5G wireless applications. She has a broad range of expertise in RF, millimeter-wave, and terahertz circuits, system, and technologies. She holds over 35 patents/patent applications and has authored numerous publications. Dr. Choudhury was the recipient of several NASA recognition awards for her work on heterodyne receivers, devices, multipliers, and guiding structures/modules for space and defense applications. She is the Chair of the IEEE MTT-S Technical Committee on Wireless Communications. _____________________ Digital Object Identifier 10.1109/TMTT.2016.2575999 0018-9480 © 2016 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.

2206

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

Takao Inoue (S’95–M’98–SM’14) received the B.S. and M.S. degrees from Oregon State University, Corvallis, OR, USA, in 1996 and 1998, respectively, and the Ph.D. degree from The University of Texas at Austin, Austin, TX, USA, in 2009, all in electrical engineering. He was with Motorola Inc., and also co-founded Fish Technologies Inc., Tokyo, Japan, which specializes in 4G and beyond wireless system prototyping. He is currently a Wireless Solutions Architect with the AWR Group, National Instruments, Austin, TX, USA. His current research interests include signal processing for wireless communications. Dr. Inoue serves on the IEEE Microwave Theory and Techniques Society (IEEE MTT-S) Technical Committee on Wireless Communications and the COMSOC Technical Committee on Signal Processing and Communications Electronics. He has served as a Steering Committee Member of the IEEE RFIC Symposium and the IEEE Radio Wireless Week, where he was the Technical Program Chair of RWW 2012 and the General Chairman in 2014. He is an active Technical Program Member and Transactions Reviewer for conferences and journals in the IEEE MTT-S and COMSOC.

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

2207

3-D Millimeter-Wave Statistical Channel Model for 5G Wireless System Design Mathew K. Samimi, Student Member, IEEE, and Theodore S. Rappaport, Fellow, IEEE

Abstract— This paper presents a 3-D statistical channel impulse response (IR) model for urban line of sight (LOS) and non-LOS channels developed from 28- and 73-GHz ultrawideband propagation measurements in New York City, useful in the design of 5G wireless systems that will operate in both the ultra-high frequency/microwave and millimeter-wave (mmWave) spectrum to increase channel capacities. A 3GPP-like stochastic IR channel model is developed from measured power delay profiles, angle of departure, and angle of arrival power spectra. The extracted statistics are used to implement a channel model and simulator capable of generating 3-D mmWave temporal and spatial channel parameters for arbitrary mmWave carrier frequency, signal bandwidth, and antenna beamwidth. The model presented here faithfully reproduces realistic IRs of measured urban channels, supporting air interface design of mmWave transceivers, filters, and multi-element antenna arrays. Index Terms— Channel model, 5G, impulse response (IR), millimeter-wave (mmWave) propagation, multipath, 73 GHz, spatial channel model (SCM), spatial lobe (SL), statistical simulator, 3-D ray-tracing, time cluster (TC), time cluster spatial lobe (TCSL), 28 GHz.

I. I NTRODUCTION

T

HE next generation of wireless communications will use systems operating from 500 MHz to 100 GHz [1], [2]. Today’s cellular systems use ultrahigh frequency (UHF) and Microwave bands exploiting multi-user multiple-input multiple-output (MU-MIMO) [3], [4], coordinated multipoint systems [5]–[7], heterogeneous networks [3], [7], and carrier aggregation [8]. However, the incredible demand for broadband wireless mobility will be supplied by moving up to the millimeter-wave (mmWave) spectrum, where a massive amount of raw bandwidth exists [9], [10], and therefore, the design of 5G cellular networks requires channel models that characterize the sub-6 GHz and mmWave spectrum to perform multi-band system-level simulations. A number of mmWave bands are currently being considered for global 5G networks. The 28- and 73-GHz frequency bands for outdoor communications are attractive, as the attenuation loss induced from atmospheric absorption

Manuscript received December 29, 2015; revised April 3, 2016; accepted May 14, 2016. Date of publication June 25, 2016; date of current version July 7, 2016. This work was supported by the National Science Foundation under Grant 1302336, Grant 1320472, and Grant 1555332. Portions of this work were published in [16], [35], and [38]. The authors are with NYU WIRELESS, New York University Tandon School of Engineering, Brooklyn, NY 11201 USA (e-mail: [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TMTT.2016.2574851

is minor (much less than 0.1 dB) over a realistic mmWave cell radius of 200 m [11], while it is significantly higher at 60 GHz (∼4 dB/200 m). Also, the Federal Communications Commission has recently issued new rulemakings to bring these bands into service [12]. The 38- and 60-GHz outdoor ultrawideband channels have been extensively studied at the UT Austin campus [13], and more recently at 28 and 73 GHz at New York University using directional antennas for outdoor mobile communications [14]. This work presents a statistical spatial channel model (SSCM) developed using the time cluster–spatial lobe (TCSL) approach, which augments the existing UHF 3GPP model through the additional model parameters of directional RMS lobe angular spreads for spatial lobes (SLs) [15]–[17]. The TCSL modeling framework is shown here to faithfully reproduce the first- and second-order time and angular statistics of measured channels that use directional antennas in mmWave bands [16], and is suitable for 5G system design of filters, electrically-steered antenna arrays, and mmWave transceivers. The SSCM presented here is based on extensive propagation measurements carried out from 2011 through 2015, and generates multipath parameters for omnidirectional and directional channel impulse responses (CIRs) for links between a transmitter (TX) and receiver (RX) as well as over a local area. We demonstrate that the IR model can be generalized to arbitrary environments and antenna beamwidths to create directional channel IRs in other mmWave frequency bands. The IR channel model is first developed for four distinct frequency scenarios: 1) the 28-GHz non-line of sight (NLOS) urban channel; 2) the 73-GHz NLOS urban channel; 3) the combined 28–73-GHz urban line of sight (LOS) channel; and 4) the combined 28–73-GHz urban NLOS channel. The 28and 73-GHz LOS statistics were lumped into one combined data set to extract combined 28- and 73-GHz statistics, motivated by the fact that measurements clearly show that these two frequencies have virtually identical path loss exponents (PLEs) close to free space (n = 2), and have virtually the same values of average number of resolvable multipath components [12]. The similarity in channel characteristics for LOS channels over mmWave frequencies provides motivation to combine the data to provide a greater statistical sample size when deriving the model statistics. The SSCM presented in this paper has been implemented using a MATLAB program and is publicly available as a freely downloadable software package suitable for link and systemlevel simulations [18].

0018-9480 © 2016 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.

2208

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

The contributions in this paper include the following. 1) A 3-D mmWave SSCM for generating CIRs at particular distances and over local areas based on the TCSL modeling framework, which extends the 3GPP model through directional RMS lobe angular spreads while keeping consistent with the 3GPP modeling framework. The SSCM has been validated using published measurements between 28 and 73 GHz and can be used for arbitrary carrier frequency between 6 and 100 GHz, RF signal bandwidths up to 800 MHz, and arbitrary antenna beamwidths greater than 7°. 2) A modeling extension to enable the spatial consistency in multi-user system-level simulations that account for the birth and death of time clusters (TCs) and spatial lobes (SLs) at mmWave frequencies in a track- or gridbased scenario, using the statistical spatial autocorrelation for the number of TCs and SLs. A. Millimeter-Wave Channel Characterization In the 1990s, mmWave channels were studied and modeled especially for LOS communications in urban macrocellular [19]–[21] and short-range indoor environments [22], [23], including important wideband channel models [24]–[26]. In the past few years, the 28-, 38-, and 70–80-GHz frequency bands have become important candidates for mmWave mobile communications [10], [12], [27]. Recently, mmWave propagation measurement campaigns have been conducted in indoor and dense urban outdoor environments [11], [12], [28]–[31]. Previous results yielded directional and omnidirectional path loss models in dense urban LOS and NLOS environments [12], [27], [32], [33], temporal and spatial channel parameters, such as cluster and angular spread statistics, and statistical distributions at 28 and 73 GHz based on measurements and ray-tracing [1], [2], [14], [29], [32], [34]–[37]. Sun et al. [33] demonstrate that urban microcell (UMi) NLOS omnidirectional PLEs range between 2.9 and 3.4, with corresponding shadow factors ranging from 2.9 to 8.6 dB, when using the 1-m close-in (CI) free space reference distance path loss model for carrier frequencies of 2.9, 28, 29, and 73 GHz. The firstand second-order simulated directional RMS delay spreads were shown to agree well with measured directional RMS delay spreads obtained over different antenna half-power beamwidths (HPBWs) and mmWave bands ranging from 28 to 73 GHz, as shown in Fig. 19 and [16]. A 3-D SSCM for LOS and NLOS mobile communications [16], [38] reproduced measurements of wideband power delay profiles (PDPs) and 3-D angle of departure (AOD) and angle of arrival (AOA) power angular spectra for multi-frequency and arbitrary antenna beamwidth [16]. Initial MIMO network simulations were carried out in [39] using a 2-D wideband mmWave statistical simulator developed from 28-GHz wideband propagation measurements [40], and showed orders of magnitude increase in data rates compared with current 3G and 4G LTE mobile systems when using spatial multiplexing and beamforming at the base station (BS) for LOS and NLOS urban environments.

B. Popular Statistical and Analytical Channel Models 1) 3GPP and WINNER II Models: The geometrybased stochastic 3GPP and WINNER II spatial channel models (SCMs) [41], [42] follow a system-level approach, suitable for link-level or system-level simulations to estimate realistic channels between a BS, and one or more user equipments (UEs), that account for empirical correlations between large-scale parameters. The large-scale parameters denote the omnidirectional RMS delay spread (DS), the azimuth spread (AS), the shadow fading (SF), and the Rician K-factor (for LOS channels), and were shown to exhibit significant correlation [43] for a given base-to-mobile link. The spatial crosscorrelation coefficients between two mobile stations (MSs) for the DS and AS, DS and SF, and AS and SF are set to +0.5, −0.6, and −0.6, respectively [41], [44], in the 3GPP models, based on work in [43]. The 3GPP model also specifies a spatial autocorrelation coefficient of +0.5 for the shadow fading experienced by two MSs, but does not specify the distance range over which the shadow fading correlation is applicable [41]. In the WINNER SCM, the spatial cross-correlation coefficients between two MSs separated by a distance dMS are modeled using a decaying exponential function, which is parameterized using the correlation distance parameter. The correlation distance between two largescale parameters is the distance at which the cross-correlation coefficient is equal to 0.37 (1/e), and the values are provided in [42, Table 4-5], with typical correlation distances ranging from 9 to 14 m for the UMi scenario. In both SCMs, the largescale parameters are generated using correlated Gaussian random variables [41] to recreate the measured joint statistics. The models make the simplifying assumption that each multipath component can be represented by a planar wavefront, characterized by small-scale parameters such as path delays, powers, AOAs, and AODs, extracted from measurement-based statistical distributions. Typical IRs in an NLOS UMi environment are modeled using a clustered delay line model using a fixed number of paths, with N = 6 paths [41], or N = 16 paths [42], where each path is further subdivided into M = 20 equal power rays. Note that the 3GPP and WINNER models adopt different terminologies to refer to a group of traveling multipaths. The 3GPP model defines a path as a time-delayed multipath copy of the transmitted signal that is subdivided into 20 rays [41], where all rays have the same path delay but slight AOD and AOA offsets. In contrast, the WINNER II model defines a cluster as a propagation path diffused in space, either or both in delay and angle domains, and a number of rays (typically 20) constitute a cluster, where the two strongest clusters are subdivided into three subclusters with intra-cluster delays of 0, 5, and 10 ns [42]. Current 3GPP and WINNER [41], [42] models make the presumption that clusters are characterized by a joint delay-angle probability density function, such that a group of traveling multipaths must depart and arrive from a unique AOD-AOA angle combination centered around a mean propagation delay [32]. This was not born out by extensive mmWave field measurements [17]. Cluster properties are usually obtained from high-resolution parameter extraction algorithms, like the SAGE [45] or KPowerMeans [46]

SAMIMI AND RAPPAPORT: 3-D mmWAVE STATISTICAL CHANNEL MODEL FOR 5G WIRELESS SYSTEM DESIGN

algorithms, which require measurements acquired with multielement antenna arrays. The small-scale fading statistics of a path (or cluster) are recreated from the superposition of the path rays, by taking the coherent sum of complex ray amplitudes, each subject to Doppler shifts. The path gain amplitudes are assumed to be Rician and Rayleigh [41], [44] in LOS and NLOS UMi environments, respectively, and the phases of each ray are generated using a uniform distribution between 0 and 2π [41]. These widespread SCMs were used to design today’s 3G and 4G systems, based on 1–6 GHz propagation measurements, for RF signal bandwidths spanning 5–100 MHz (20-ns smallest time resolution) [41], [42]. 2) COST 2100 Models: The COST 2100 models follow a cluster-level approach, where clusters (e.g., scattering objects) are dropped in a simulated environment and can interact with one or more mobile terminals using the concept of visibility region. The visibility region as defined in the COST 2100 model [47], [48] is a key concept of geometric and stochastic propagation models, and represents the space- or time-span over which a group of traveling multipath components are present at a generic radio terminal antenna [48]. Spatial consistency is enabled through the use of visibility regions associated with each cluster of multipath components. One or more clusters of rays are assigned to a visibility region, whose size varies as a mobile terminal moves, thereby allowing for spatial consistency in a simulated environment. Spatial consistency refers to smooth channel transitions between closely separated mobile terminals that experience a similar, but slightly different, scattering environment. Ignoring spatial consistency can overestimate the performance of spatial multiantenna techniques [37]. The COST 2100 models assume Rayleigh fading for path gain amplitudes in NLOS environments to recreate the statistics of small-scale fading. In the COST 2100 model, the time delay of a multipath component is the sum of three delays: 1) the BS-to-scatterer delay; 2) the MS-to-scatterer delay; and 3) the cluster-link delay [47]. 3) MiWEBA Models: The mmWave Evolution for Backhaul and Access (MiWEBA) models [14] employ a quasideterministic channel model to characterize the 60-GHz outdoor multipath channels, by considering the superposition of a few strong deterministic paths and several relatively weak random rays. The deterministic rays are modeled using Friis’ free space path loss equation and path-length geometry determined from the TX and RX heights and the T-R separation distance, while the properties of weaker random rays are generated from measurement-based statistical distributions. These models are based on wideband propagation measurements at 60 GHz and directional antennas. The models are useful for system-level simulations and network access capacity analyses. The model assumes that random clusters arrive with a Poisson process, with exponentially distributed inter-arrival times, while random cluster amplitudes are Rayleigh distributed, with phases generated from independent and identically distributed uniform distributions between 0 and 2π [14]. These models also account for blockage by vehicles and humans by specifying the probability of blockage for deterministic rays and random rays.

2209

4) METIS Models: The METIS models employ a combination of map-based and stochastic channel models to estimate path amplitudes, and are suitable for evaluating massive MIMO and advanced beamforming algorithms through Monte-Carlo simulations [37]. The model uses ray-tracing techniques to obtain large-scale fading characteristics for a site-specific environment, and measurement-based results to model the small-scale fading statistics. The small-scale statistics are generated following the 3GPP and WINNER modeling approaches. 5) SIRCIM/SMRCIM Model: A measurement-based statistical indoor radio-channel impulse response (IR) model (SIRCIM) and outdoor mobile simulator (SMRCIM) were successfully implemented from many thousands of collected CIRs in factories at 1.3 GHz [49], [50], and from outdoor cellular channel PDPs [51], [52]. These CIR models were popular with industry in the early years of digital cellular and WiFi [53]. The SIRCIM and SMRCIM models were based on statistical and geometrical models to synthesize the phases and directions of arrival and departure in an IR model [53], [54]. 6) Kronecker Model: The MIMO channel covariance matrix can be decomposed into the Kronecker product of the transmit and receive covariance matrices, and the corresponding CIRs for a MIMO system can be computed using the high-order single-value decomposition (SVD) method [55]. This method was shown to improve accuracy in predicting channel capacities in the UHF and microwave spectrum [56]. II. W IDEBAND P ROPAGATION M EASUREMENTS A. Measurement System Description A 400-megachips-per-second (Mcps) broadband sliding correlator channel sounder was used to measure the 28- and 73-GHz wideband urban channel over a 800-MHz null-to-null RF bandwidth, where the RX locations were placed on New York City streets and inside one building with T-R separation distances ranging from 20 to 425 m. Recorded outages over all measured RX locations are given in [31] and [38], and can be used to determine the conditional probability of a link being made when using a CIR. The transmitter output power was varied between 11 dBm (for close, LOS receiver locations) up to a maximum of 14.6 dBm at 73 GHz and 30 dBm at 28 GHz (for the NLOS locations). A pair of steerable (rotatable) directional horn antennas of 24.5 dBi, with 10.9° and 8.6° HPBWs in the azimuth and elevation planes, respectively (used at 28 GHz) and 27 dBi, with 7° HPBW in both azimuth and elevation, respectively (used at 73 GHz) were used to perform TX and RX azimuth antenna sweeps at various fixed TX/RX elevation pointing angles for each location, with static PDPs measured in azimuth step increments equal to one HPBW (10° or 7° at 28 and 73 GHz, respectively). Additional channel sounder specifications used during the propagation measurements in New York City can be found in Table I and [11], [57], [58]. Fig. 1 shows a photo of the 73-GHz transmitter used during the measurements. Independent frequency sources at the TX and RX were used in the 28- and 73-GHz channel sounder to provide the

2210

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

TABLE I B ROADBAND S LIDING C ORRELATOR C HANNEL S OUNDER S PECIFICATIONS U SED IN THE 28- AND 73-GHz N EW Y ORK C ITY M EASUREMENTS [11], [12], [57]

Observed jitter was removed with additional LabVIEW postprocessing software designed to provide a real-time trigger alignment of the I and Q waveforms, in which 20 successive PDPs at one measured angle were aligned to the time of arrival of the strongest received multipath component before averaging. The PDPs collected at unique pointing angle directions provided excess time delay PDPs, relative to the first arriving multipath component. Complementary 3-D ray-tracing software was developed to recreate the absolute propagation time of multipath arrivals from TX to RX [40]. B. Measurement Procedure Description 12 000 wideband PDPs were recorded at 28 and 73 GHz in New York City using a 400-Mcps broadband sliding correlator channel sounder and directional steerable horn antennas to recover AOD and AOA statistics. The directional steerable horn antennas were exhaustively rotated in the azimuth and elevation planes, and many thousands of PDPs were collected at distinct azimuth and elevation unique pointing angles at 26 RX and 74 RX outdoor locations at 28 and 73 GHz, respectively. For the 28-GHz measurements, the RX antenna was rotated in the azimuth plane in step increments of 10° (one HPBW), at RX elevation planes of +/− 20° and 0° for a fixed TX azimuth/elevation pointing angle to emulate a practical cellular deployment scenario where BS antennas are usually downtilted toward the street to provide maximum covered street areas. This procedure was repeated for three distinct TX azimuth angles, resulting in a total of nine RX azimuth sweeps. One additional TX azimuth sweep was performed at a fixed −10° downtilt, and for a fixed RX azimuth/elevation pointing angle, where the TX antenna was rotated in azimuthal step increments of 10°. A PDP was acquired at each TX-RX unique pointing angle. A similar procedure was performed for the 73-GHz measurements, where the RX elevation planes were adjusted in 7° step increments, in real time during the field measurements, based on the strongest measured elevation plane (elevation plane with strongest received power). Many thousands of unique TX-RX pointing angle PDPs were measured, thereby capturing the majority of significant and most powerful multipath components to extract accurate path loss and multipath properties. Additional details of the measurement campaigns can be obtained in [12]. III. S YNTHESIZING O MNI PDP S F ROM M EASUREMENTS

Fig. 1. Photo of the 73-GHz transmitter used during the New York City measurements.

PN clocks, and the intermediate frequency, and we note that the local oscillator signals on both TX and RX were not phase-synchronized, causing a lack of absolute time stamping.

A 3-D MATLAB-based ray-tracer was developed to recover absolute propagation time delays of the PDP measurements. The downtown Manhattan environment was modeled in Google SketchUp over an 800 × 800 m2 area, which allowed fast and easy 3-D site-specific modeling using simple geometrical shapes such as cubes. The 3-D information was then exported in XML format, and subsequently extracted to numerically reconstruct and discretize the environment in MATLAB [35]. Fig. 2 shows a typical ray-traced map where each ray that leaves the TX and successfully arrives at the RX is depicted for a measured TX-RX location pair. The predicted AOAs are

SAMIMI AND RAPPAPORT: 3-D mmWAVE STATISTICAL CHANNEL MODEL FOR 5G WIRELESS SYSTEM DESIGN

2211

Fig. 2. 3-D view of the downtown Manhattan area. The rays shown in red leave the TX and successfully arrive at the RX, and represent viable multipath signal propagation paths [38].

Fig. 4. PDP measured at an RX azimuth/elevation of 72°/0° and a TX azimuth/elevation of 269°/−10°, and predicted by the 3-D ray-tracer. Azimuth angles are with respect to a True North 0° angle.

Fig. 3. Azimuthal distribution of total received power (dBm units), also referred to as polar plot, showing the predicted 28-GHz AOAs using 3-D ray-tracing at the Manhattan RX location on Wooster Street [11].

shown in Fig. 3, denoted with black arrows on top of the 28-GHz measured power azimuth spectrum for that RX location. The ray-tracer predicted the strongest AODs and AOAs with an accuracy of ±20° (i.e., ±2 antenna beamwidths), which provided sufficient accuracy to pair the strongest PDPs with the estimated absolute propagation delay at the predicted angles. Figs. 4–7 show the 28-GHz excess delay PDPs corresponding to the strongest measured azimuth AOAs at a 0° RX elevation, found by searching ±2 antenna beamwidths about the predicted azimuth AOAs. The absolute propagation times, as computed from the ray-tracing software, are shown at the bottom of the four figures, and were used to properly timeshift each PDP appropriately over an absolute propagation time axis, to synthesize the equivalent omnidirectional PDP as would have been measured with an omnidirectional antenna with comparable gain of the horns. The resulting omnidirectional PDP is shown in Fig. 8, where each excess delay PDP was appropriately time-shifted using the absolute propagation time (obtained by dividing the ray-traced propagation distance by the speed of light) of the first arriving peak at the corresponding RX azimuth angle. In Figs. 4–7, the absolute propagation times of the

Fig. 5. PDP measured at an RX azimuth/elevation of 2°/0° and a TX azimuth/elevation of 269°/−10°, and predicted by the 3-D ray-tracer. Azimuth angles are with respect to a True North 0° angle.

first arriving peak for AOA 1, 2, 3, and 4 are 381, 407, 1433, and 1500 ns, respectively. Since the four angles are orthogonal without beam overlap, the four excess delay PDPs were shifted and summed in mW/ns to synthesize omnidirectional PDPs. This example showed the method for superimposing PDPs over the RX azimuth plane. The PDPs from strongest TX and RX angles were summed over the azimuth and elevation dimensions, yielding 3-D omnidirectional PDPs with absolute timing. Due to database or antenna pointing errors at some locations, we created omnidirectional PDPs for 3 of 6 measured LOS locations and 13 of 20 measured NLOS locations at 28 GHz, and for 5 of 5 measured LOS locations and 19 out of 25 measured NLOS locations at 73 GHz, using the few strongest predicted AODs and AOAs. The synthesized absolute timing omnidirectional PDPs are obtained from the strongest measured TX and RX pointing angles, which capture strong multipath components traveling in the channel,

2212

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

Fig. 6. PDP measured at an RX azimuth/elevation of 232°/0° and a TX azimuth/elevation of 269°/−10°, and predicted by the 3-D ray-tracer. Azimuth angles are with respect to a True North 0° angle.

Fig. 8. Omnidirectional PDP synthesized using the 3-D ray-tracing absolute time of arrivals and the four excess delay PDPs shown in Figs. 4–7, at TX location Coles 1 (COL1).

wavelengths with different propagation paths are identically and independently distributed (i.i.d.) and uniform between 0 and 2π, such that powers may be added [10], [12], [60]  Pr (θT X,i , φT X, j , θ R X,k , φ R X,m )(d) (1) Pr Omni = i, j k,m

P L[dB](d) = PTX + G t + G r − 10 × log10 (PrOmni )   d + χσ P L[dB](d) = P L FS (d0 ) + 10n log10 d0

Fig. 7. PDP measured at an RX azimuth/elevation of 182°/0° and a TX azimuth/elevation of 269°/−10°, and predicted by the 3-D ray-tracer. Azimuth angles are with respect to a True North 0° angle.

in addition to weaker multipaths as a result of significant local scattering in the environment, to provide realistic omnidirectional channel models. This SSCM provides a realistic and comprehensive model that is capable of modeling the statistics of both strong and weak multipath components. A. Omnidirectional LOS and NLOS Path Loss Models Omnidirectional path loss models are necessary to estimate the total received power for arbitrary distance and antenna pattern (see Step 2 in Section IV-A). Omnidirectional path losses were found by summing the received powers measured at each unique non-overlapping azimuth and elevation antenna pointing angle and recovering the path losses after carefully removing double counts (arising from TX and RX sweeps) and antenna gains [59]. This procedure is valid since adjacent angular beamwidths are orthogonal to each other, and the phases of arriving multipath components with such short

(2) (3)

where i, j, k, and m are indices denoting unique pointing directions in azimuth and elevation at the TX and RX, respectively, (θTX , φTX , θRX , φRX ) denote the TX azimuth and elevation angles, and the RX azimuth and elevation angles, respectively, PTX is the transmit power in dBm, G t and G r are the TX and RX horn antenna gains in dBi, respectively, PLFS (d0 ) is the frequency-dependent free space path loss at distance d0 , n is the average PLE over distance, λ is the carrier wavelength, and χσ is a zero-mean lognormal random variable with standard deviation σ modeling large-scale signal fluctuations. The antenna de-embedding was performed by removing the antenna boresight gain from every measured PDP at each and every TX-RX antenna pointing angle. In the measurements, the horn antennas were incremented in steps of one HPBW in the azimuth plane, thereby minimizing overlapping effects of two adjacent radiation patterns. When considering the equivalent omnidirectional antenna pattern arising from the aggregation of adjacent beamwidths, the effective omnidirectional antenna pattern has a relatively constant gain, as shown in [60, Fig. 1]. The antenna cross-polarization ratios (XPRs) were 21 and 25.4 dB for the 28- and 73-GHz outdoor measurements [61], indicating a minor and negligible impact of crosspolar contributions. Note that impact of antenna sidelobes was also minor, with sidelobe levels −20 dB below boresight gain when beyond one HPBW. The d0 = 1 m CI free space reference distance path loss model is a much simpler path loss model than the

SAMIMI AND RAPPAPORT: 3-D mmWAVE STATISTICAL CHANNEL MODEL FOR 5G WIRELESS SYSTEM DESIGN

2213

path loss channel model. Similar results were obtained for the 28-GHz data set [40]. IV. J OINT T EMPORAL -S PATIAL mmWAVE CIR The radio propagation channel is commonly represented by the superposition of many plane waves, and is parameterized with the double-directional CIR [64]. The omnidirectional CIR → − − → h omni (t,  , ) is expressed as [38] Mn N   − − → → h omni (t,  , ) = am,n e j ϕm,n · δ(t − τm,n ) n=1 m=1

Fig. 9. 73 GHz (mobile) omnidirectional path loss models for the LOS and NLOS environments, obtained from the wideband measurements in Manhattan. The omnidirectional power at each RX location was obtained from all PDPs for all RX and TX pointing angles (and removing double-counted angles), and antenna gains were removed from each PDP.

3GPP/WINNER least-squares best fit regression line (named floating-intercept model, or FI for short). The current 3GPP/WINNER path loss models use two parameters to describe path loss over distance (a slope and an intercept), which are highly dependent upon the number of path loss samples measured and corresponding distances. The CI model, however, uses just one parameter, the PLE, making it more stable across various path loss data sets, frequencies, and environments [12], [27], [33], [62], while remaining steeped in physics to a physical free space path loss value at a distance of 1 m, which can always be considered to be within LOS of a BS. Using the CI model, it is therefore easier to compare path loss results from different research groups. Further, the CI model can be used as a global standard for comparing path loss models across different frequencies or scenarios [12], [27], [33]. The CI and FI models have been shown to perform similarly over identical data sets, with standard deviation errors within a fraction of a decibel [12], [27], [63]. Similarly, when considering a probabilistic path loss model used in system-level simulations, which takes into account the probability of LOS, there is virtually no difference between the CI and FI models [35]. Fig. 9 shows the omnidirectional path losses from the 73 GHz LOS and NLOS measurements, computed from (1) and (2), and corresponding path loss parameters using a d0 = 1 m CI free space reference distance in (3). Using (3), the omnidirectional PLE and shadow factor (SF) in LOS were n LOS = 2.0 and σLOS = 5.2 dB, matching free space propagation (n = 2). In NLOS, we obtained n NLOS = 3.3 and σNLOS = 7.6 dB, showing greater attenuation over distance due to obstructions. To verify the method shown in (8), we also computed the path loss parameters from the synthesized absolute timing PDPs, yielding n Syn = 2.1 and σSyn = 5.7 dB, and n Syn = 3.3 and σSyn = 7.8 dB in LOS and NLOS, respectively, in agreement with the measured data. This indicates that only a few of the strongest AODs and AOAs (up to four such angles) are sufficient to recover an accurate

→ − − → → − − → · δ(  −  m,n ) · δ( − m,n ) (4) → − where t denotes absolute propagation time,  = (θ, φ)TX is → − the vector of azimuth and elevation AODs, and = (θ, φ)RX is the vector of azimuth and elevation AOAs; N and Mn denote the number of TCs (also defined in Section IV-A) and the number of cluster subpaths (SPs), respectively; am,n is the magnitude of the mth SP belonging to the nth TC; ϕm,n and τm,n are the phases and propagation time delays, → − → − respectively;  m,n and m,n are the azimuth/elevation AODs and azimuth/elevation AOAs, respectively, of each multipath component. Note that a SP is an individual multipath component contained in either a SL or TC. The omnidirectional CIR can further be partitioned to yield directional PDPs at a desired TX-RX unique antenna pointing angle, and for arbitrary TX and RX antenna patterns [16]  − → − → am,n e j ϕm,n · δ(t − τm,n ) h dir (t, d , d ) = N

Mn

n=1 m=1

− → − → − → − → ·gTX (d −  m,n ) · gRX ( d − m,n ) (5)

→ − → − where (d , d ) are the desired TX-RX antenna pointing → − → − angle, gTX (  ) and gRX ( ) are the arbitrary 3-D (azimuth and elevation) TX and RX complex amplitude antenna patterns of multi-element antenna arrays, respectively. In (5), the TX and RX antenna patterns amplify the power levels of all multipath components lying close to the desired pointing direction, while effectively setting the power levels of multipath components lying far away from the desired pointing direction to 0. The statistical channel model also produces the joint → − − → AOD-AOA power spectra P(  , ) in 3-D obtained by integrating the magnitude squared of (4) over the propagation time dimension,  ∞ − → − → → − − → |h omni (t,  , )|2 dt (6) P(  , ) = 0

→ − − → P(  , ) =

Mn N  

|am,n |2

n=1 m=1

→ − − → → − − → · δ(  −  m,n ) · δ( − m,n ).

(7)

Previous work modeled the CIR as a function of time [65], azimuth AOA [66], [67], and both AOD and AOA azimuth/elevation angles [42]. Note, however, that [42] models AOD and AOA azimuth and elevation information for the

2214

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

in-building, indoor-to-outdoor, and outdoor-to-indoor scenarios, but only azimuth, and not elevation, for the microcellular environment. Here, each multipath component of the IR model is assigned a unique combination of AOD and AOA azimuth and elevation angles based on measured data to simulate realistic outdoor mmWave channels. This modeling approach supports use of directionality at both the BS and mobile handsets [68]. A. Time Clusters and Spatial Lobes Statistics The SSCM given here uses TCs and SLs to model the omnidirectional CIR and corresponding joint AOD-AOA power spectra, which have been used successfully in modeling mmWave channels [38], [40]. TCs are composed of multipath components traveling close in time, and that arrive from potentially different directions in a short propagation time window. SLs represent main directions of arrival (or departure) where energy arrives over several hundred nanoseconds. This SSCM structure is motivated by field measurements [11], [57] which have shown that multiple TCs can arrive at unique pointing angles, detectable due to high gain directional antennas, and this feature has not been modeled in current 3GPP and WINNER models. These definitions decouple the time and space dimensions by extracting temporal and spatial statistics separately. The definition of TC here considers multipath components traveling close in time, but that can arrive from many lobe angular directions, whereas current 3GPP and WINNER models assume that SPs belonging to a cluster travel along the same propagation path, but arrive at the same time delay over a certain AOA angular spread. The TCSL approach implements a physically based clustering scheme (e.g., the use of a fixed inter-cluster void interval representing the minimum propagation time between likely reflection or scattering objects) derived from field observations, and can be used to extract TC and SL statistics for any ray-tracing or measurement data sets. Note the channel models provided in [40] are 2.5-D and are valid for omnidirectional azimuth planes using unique elevation planes or pointing angles, while the models provided in [16] and [38] and in this paper describe true 3-D azimuth and elevation channels where the TCSL framework models the directionality of the channels through separate TCs that have time-delay statistics, and through SLs which represent the strongest directions of multipath arrival and departure [15]–[17], [38], [40], [69]. The approach used here is an extension of early work where intracluster SPs with distinct delays were successfully used in modeling the indoor environment based on wideband channel measurements [65], [66]. MmWave CIRs and power angular spectra can be partitioned conveniently using the definitions of TCs and SLs, respectively. Fig. 10 shows a typical received omnidirectional PDP at 28 GHz. The time-partitioning methodology is shown in Fig. 10 by delineating the beginning and end times of each TC, using a 25-ns minimum inter-cluster void interval. Sequentially arriving multipath components that occur within 25 ns of each other are assumed to belong to one TC. In defining Fig. 10, two TCs are composed of 8 and 6 SP components with random delays, amplitudes, and AOAs. True

Fig. 10. Typical measured omnidirectional PDP, where two TCs are composed of eight and six SP components, assuming a 25-ns minimum inter-cluster void interval. Using a peak detection algorithm, the SP components are found to have randomly varying AOAs, delays, and amplitudes [10], [16], [40].

to real-world measurements, the total power in one TC is also random, as it is composed of the sum of randomly varying SP powers. In addition, the propagation phases of each multipath component can be taken to be i.i.d. uniform between 0 and 2π [65]. By counting the number of TCs and intra-cluster SPs, and extracting TC and SP delays and power levels from all available measured PDPs, measurement-based statistical distributions are obtained and allow reconstruction of time-varying IRs that embody the statistics of the collected data. The key parameters for generating mmWave PDPs are the number of TCs, the number of intra-cluster SPs, the TC and SP delays, and the TC and SP power levels [38], [40]. Since directional transmissions will play a role in mmWave communications, it is equally important to characterize the spatial angular channel at the transmitter and at the receiver. Fig. 11 shows a typical measured power azimuth spectrum obtained with a horn antenna in an NLOS environment, where each dot corresponds to the total received power (area under PDP) measured at azimuth angles in step increments of 10°. Fig. 11 illustrates the concept of SLs, by showing that energy arrives at distinct mean pointing AOAs over a contiguous range of azimuth angles and a −10-dB power threshold with respect to the maximum received angle power. While five mean pointing AOAs can be identified in Fig. 11 at azimuth angles of 65°, 0°, 240°, 180°, and 130° with greatly varying powers, it is sufficient in practice to keep track of the few strongest mean pointing SL AOAs by defining a −10- or −20-dB power threshold with respect to the maximum peak power in the 3-D spectrum in LOS and NLOS, below which all power levels can be disregarded. Note that we had previously used −10- and −20-dB thresholds in LOS and NLOS environments when considering 2-D polar plots [40]. Polar plots can be easily reconstructed having knowledge of the number of SLs, the mean pointing AOAs, the lobe ASs,

SAMIMI AND RAPPAPORT: 3-D mmWAVE STATISTICAL CHANNEL MODEL FOR 5G WIRELESS SYSTEM DESIGN

Fig. 11. Measured power azimuth spectrum, also called polar plot, showing a −10-dB power threshold and two SLs, where the threshold is with respect to the maximum received angle power. An SL has well-defined properties such as its mean pointing angle, its absolute angle spread, and its RMS angle spread [10], [16], [40].

2215

Fig. 12. Cluster powers extracted from 73-GHz NLOS omnidirectional PDPs using a 25-ns minimum intercluster void interval, with decay time constant

= 56.0 ns and y-intercept P 0 = 0.826.

and the RMS lobe ASs. While this example shows a 2-D polar plot, the work presented here uses 3-D power angular spectra that also include the elevation dimension. To benchmark the accuracy of the resulting SSCM model and simulator against real data, we define key parameters, or primary statistics, as measurement-based statistical distributions used to generate mmWave temporal and spatial channel parameters for a time-varying IR. A good test procedure for checking the accuracy of the channel model can be devised by not just comparing first-order statistics, but by also looking at second-order statistics, or secondary statistics, of the generated IRs, such as the RMS DSs and RMS lobe angular spreads. B. Time Cluster Partitioning The TC partitioning scheme heavily affects the outcome of the temporal channel parameter statistics. In this work, a 25-ns minimum inter-cluster void interval was defined to segment omnidirectional PDPs based on time of arrivals, where consecutive multipath components that occur within a time duration less than 25 ns were assumed to belong to one TC. This simple clustering scheme allows us to resolve intricate multipath channel dynamics within the smallest multipath time resolution offered by the 3GPP and WINNER models (20 ns) [41], [42], and is easily adjustable to resolve temporal statistics over arbitrary time resolutions using a different minimum inter-cluster void interval. The value of 25 ns for the minimum inter-cluster void interval was found to match the measured data, and makes sense from a physical standpoint, since multipath components tend to arrive in clusters at different time delays over many angular directions, most likely due to the free space air gaps between reflectors (buildings, lampposts, streets). The narrowest streets have a typical width of 8 m (25 ns in propagation delay) in New York City, thus

Fig. 13. Intra-cluster SP powers extracted from 73-GHz NLOS omnidirectional PDPs using a 25-ns minimum inter-cluster void interval, with decay time constant γ = 15.4 ns and y-intercept P 0 = 0.442.

physically describing the regularly observed minimum void interval for arriving energy. Fig. 12 shows the 73-GHz NLOS cluster power levels, normalized to the total received power, as a function of cluster excess delays, extracted from absolute timing omnidirectional PDPs using a 25-ns minimum inter-cluster void interval. The mean exponential model, found using a least-squares regression, is parameterized using the cluster decay time constant

, defined as the excess time delay when average cluster power falls to 37% (1/e), and the y-intercept at τ = 0 ns, which physically represents the average power P 0 contained in the first arriving TC. Here, we found = 56.0 ns and P 0 = 0.826. It is worth noticing in Fig. 12 that the large cluster power levels for two measurement locations at τ = 156 ns and τ = 177 ns are carrying 98% and 86% of the total received power for those PDPs, respectively, and this situation causes large RMS DSs. Fig. 13 shows the

2216

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

TABLE II K EY TC AND SL M ODEL PARAMETERS N ECESSARY TO G ENERATE mmWave 3-D CIRs [16], [17]. T HE M ODELING S TEP P ROCEDURES AND D ISTRIBUTION M EANS AND S TANDARD D EVIATIONS A RE P ROVIDED IN S ECTION V-A AND TABLE IV

73-GHz NLOS intra-cluster SP power levels (normalized to the TC power), and the corresponding mean exponential curve model. The curve is parameterized with γ = 15.4 ns and P 0 = 0.442. The smaller γ physically denotes that intra-cluster SP components decay much faster than TCs. Note that similar plots were shown in [38, Figs. 3 and 4] for the 28-GHz NLOS case, with measured cluster and SP decay time constants of

= 49.4 ns and γ = 16.9 ns, which are similar to the 73-GHz NLOS time constants, indicating little dependence upon carrier frequency. When combining the 28- and 73-GHz LOS cluster and SP power levels, we found = 25.9 ns and γ = 16.9 ns, showing that cluster powers decay much faster in LOS than in NLOS environments. Table IV summarizes the ’s and γ ’s as a function of environment and carrier frequency. The choice of the minimum inter-cluster void interval can significantly impact the cluster and SP power levels. While a 25-ns minimum inter-cluster void interval worked well for generating PDPs whose RMS DSs match with measured data, we also investigated the temporal statistics under 2.5- and 20-ns minimum inter-cluster void intervals, which yielded = 24.4 ns and = 29.9 ns, respectively, for the 73-GHz NLOS scenario, compared with = 56.0 ns when using a 25-ns inter-cluster void interval. The ’s can vary greatly for different minimum inter-cluster void intervals. As the minimum inter-cluster void interval increases, the number of TCs in a given PDP must decrease, while the number of intra-cluster SPs must increase, and consequently, the cluster powers must increase yielding a larger , and the intracluster SP powers must decrease yielding a smaller γ . For minimum inter-cluster void intervals of 2.5 and 20 ns, there are many more cluster power levels carrying very little power that shift the mean exponential model curve downward, with the resulting omnidirectional RMS DSs of the simulated PDPs much smaller than the measured data. Table II summarizes the TC model parameters. C. 3-D Spatial Lobe Thresholding The 3-D spatial distribution of received power was reconstructed from the 28- and 73-GHz LOS and NLOS directional received powers by linearly interpolating adjacent power level

Fig. 14. Measured number of AOD lobes extracted from 28-GHz power azimuth spectra, using a −10-dB threshold, with a mean and a standard deviation of 1.6 and 0.8, respectively. The simulated distribution is a Poisson distribution with a mean of 1.8 (see Step 3 in Section V-A).

segments in azimuth and elevation with a 1° resolution and extracting 3-D spatial angular statistics. We used a −10-dB threshold below maximum peak power in the 3-D power spectrum in both LOS and NLOS environments, where all power segments below this threshold were disregarded for further processing. Note that work shown in [15] and [40] sometimes used a −20-dB threshold, but here we use a −10-dB lobe threshold to remain consistent. Fig. 14 shows a typical empirical histogram plot of the number of AOD SLs extracted from 28-GHz NLOS power angular spectra, next to the simulated histograms as explained in Step 3 of Section V, yielding good agreement with the measured distributions. The mean numbers of AOD and AOA SLs were both 1.6 for the 28-GHz NLOS measurements, and 1.5 and 2.5, respectively, for the 73-GHz NLOS measurements, showing very little dependence on carrier frequency from the BS perspective. However, there are more AOA SLs found at 73 GHz, most likely arising from more prominent local scattering than at 28 GHz. In the 28–73-GHz combined LOS scenario, the mean number of AOD lobes

SAMIMI AND RAPPAPORT: 3-D mmWAVE STATISTICAL CHANNEL MODEL FOR 5G WIRELESS SYSTEM DESIGN

was 1.9, compared with 1.5 and 1.6 in the NLOS case at 28 and 73 GHz, respectively, showing that multipath signals can reach the RX from more departing directions. The mean number of AOA SLs was 1.8 in the combined 28–73-GHz LOS scenario, which is comparable to the mean number of AOA SLs found in the combined 28–73-GHz NLOS scenario, suggesting that LOS and NLOS spatial environments appear similar to a receiver. SL statistics are summarized in Table IV as a function of environment and frequency. Table II summarizes the SL model parameters. V. G ENERATING T EMPORAL AND S PATIAL C HANNEL PARAMETERS This section outlines the step procedure for generating 3-D mmWave temporal and spatial channel parameters in (4) and (5), e.g., for producing sample functions for an SSCM, in both LOS and NLOS, generalized to arbitrary carrier frequency, signal bandwidth, and antenna beamwidth [16] by closely following the 3GPP model [41]. The SSCM reproduces the omnidirectional CIR between a single TX and single RX antennas, or single input single output (SISO), by separately modeling the temporal and spatial statistics. The omnidirectional CIR for SISO system can readily be extended to a local area using knowledge of small-scale spatial amplitudes and spatial autocorrelations of the multipath amplitudes over tens of wavelengths to emulate a true MIMO communication link [69], [70]. In Step 12, the SP components are randomly assigned to the SL AODs and AOAs in a uniform random fashion, thereby recoupling time and space to produce an accurate joint spatio-temporal SSCM, while remaining faithful to our definitions of TC and SL. The 3-D SSCM is based on measurements in [11] and [57] and is valid for any noise floor greater than −100 dBm (over an 800-MHz RF bandwidth), or any total link dynamic range less than 180 dB, and for RF bandwidths less than 800 MHz and antennas with beamwidths greater than 7°. The model can be used for bit-error simulations as was done in first-generation digital cellular [71]. The 73-GHz SCM in [34] uses the 3GPP delay-angle cluster, where a cluster corresponds to a group of multipath components with similar propagation delays, AODs, and AOAs. The model parameters are obtained in [32] and [34] making a blind assumption that the 3GPP model framework used for sub-6-GHz channel models represents mmWave channels. The step procedures below generalize the 3GPP delay-angle cluster approach by using the concepts of TCSL approach, which considers the time and angle dimensions separately. The TCSL modeling approach was shown to accurately recreate the measured global and directional RMS delay and angular spreads statistics [16]. In the following steps, DU corresponds to the discrete uniform distribution, and the notation [x] denotes the closest integer to x. Steps 11 and 12 apply to both AOD and AOA SLs. A. Step Procedure for Generating Channel Coefficients Step 1: Generate the TX-RX separation distance d (in 3-D) ranging from 30 to 60 m in LOS and from

2217

TABLE III M EASURED PLE S AND S HADOW FACTORS [12], [59] U SED TO G ENERATE THE O MNIDIRECTIONAL R ECEIVED P OWER IN S TEP 2 OF S ECTION V-A

60 to 200 m in NLOS (based on our field measurements, and may be modified): d ∼ U (dmin , dmax ) where

(8)

 dmin = 30 m, dmax = 60 m, LOS dmin = 60 m, dmax = 200 m, NLOS.

To validate our simulation, we used the distance ranges in Step 1, but for standards work other distances are likely to be valid. Users located near BSs (i.e., small TX-RX separation) will be power controlled in the near field [27], [62]. Step 2: Generate the total received omnidirectional power Pr (dBm) at the RX location according to the environment type: (9) Pr (d)[dBm] = Pt [dBm] − P L(d)[dB]   d + χσ (10) P L[d B](d) = P L FS (d0 ) + 10n log10 d0   4πd0 (11) P L FS (d0 ) = 20 × log10 λ where Pt is the transmit power in dBm, d0 = 1 m, λ is the carrier wavelength, n is the PLE for omnidirectional TX and RX antennas, given in Table III for 28 or 73 GHz in both LOS and NLOS environments, and χσ is the lognormal random variable with 0-dB mean and standard deviation σ [12]. The d0 = 1 m CI free space reference path loss model is a simple physically-based one-parameter (PLE) model [12], [33], which is more stable across frequencies and environments than the traditional FI least-squares regression equation line [12], [27], [33], [62]. Further, the CI and FI models perform similarly over identical data sets, with differences in standard deviations that are within a fraction of a dB, yet the CI model is simpler and based on fundamental propagation physics [12], [27], [33]. Also, the CI model allows the pooling of LOS power statistics at multiple mmWave frequencies without any change in model coefficients, since the PLE n is set equal to 2 for LOS propagation (this is not the case for the FI model). Step 3: Generate the number of TCs N and the numbers of AOD and AOA SLs (L AOD , L AOA ) at the RX location: N ∼ DU [1, 6]   L AOD ∼ min L max , max{1, Poisson(μAOD )}   L AOA ∼ min L max , max{1, Poisson(μAOA )}

(12) (13) (14)

2218

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

TABLE IV K EY F REQUENCY-D EPENDENT PARAMETERS T HAT R EPRODUCE THE M EASURED S TATISTICS OF O MNIDIRECTIONAL C HANNELS U SING THE SSCM P RESENTED H EREIN FOR THE F OLLOWING F REQUENCY S CENARIOS : C OMBINED 28–73-GHz LOS, 28-GHz NLOS, 73-GHz NLOS, AND C OMBINED 28–73-GHz NLOS

where L max = 5 is the maximum allowable number of SLs, μAOD and μAOA are the empirical mean number of AOD and AOA SLs, respectively (see Table IV). At 28 GHz in NLOS, the maximum number of TCs observed was 5, while it was 6 at 73 GHz, using a −10-dB threshold based on work in [38]. We therefore choose 6 to simplify the model across frequency bands. Note that in [38], (L AOD , L AOA ) were conditioned upon N, but since SPs from the same TC can arrive and depart from arbitrary directions, the number of SLs is generalized here to be independent of the number of TCs [16]. Step 4: Generate the number of cluster SP Mn in each TC: Mn ∼ DU [1, 30], n = 1, 2, . . . , N.

(15)

At 28 GHz in NLOS, the maximum and second to maximum number of cluster SPs were found to be 53 and 30, respectively, over all locations, while at 73 GHz the maximum was 30 in NLOS; therefore, 30 is chosen as the upper bound of the uniform distribution for all frequencies. SP components were identified using a peak detection algorithm. Step 5: Generate the intracluster SP excess delays ρm,n in units of nanoseconds:

1+X n

1 × (m − 1) ρm,n (Bbb ) = Bbb m = 1, 2, . . . , Mn , n = 1, 2, . . . , N

(16) (17)

where Bbb = 400 MHz is the baseband bandwidth of the transmitted PN sequence (but can be modified for different baseband bandwidths less than 400 MHz), and X n applies to all Mn intra-cluster SPs and is generated for the nth TC, using a uniform random variable between 0 and X max . Note that 1/Bbb must have units of nanoseconds when generating ρm,n . This step ensures a bandwidth-independent channel model, while reflecting observations that intra-cluster SP delay intervals tend to increase with delay (through the random variable X n ). The upper bound X max is easily adjustable to field measurements (see Table IV).

Step 6: Generate the cluster excess delays τn : τn ∼ Exp(μτ )   τn = sort τn − min τn  0, n=1 τn = τn−1 +ρ Mn−1 ,n−1 + τn + 25, n = 2, . . . , N

(18) (19) (20)

where sor t (·) orders the delay elements τn from smallest to largest, and where μτ is given in Table IV. This step assures no temporal cluster overlap with a 25-ns minimum intercluster void interval, which was found to match the measured data, and makes sense from a physical standpoint, since multipath components tend to arrive in clusters at different time delays [65] over many angular directions, most likely due to the free space air gaps between reflectors (buildings, lampposts, streets, etc). The narrowest streets have a typical spatial width of 8 m (25 ns in propagation delay) in New York City, thus physically describing the regularly observed minimum void interval for arriving energy. Step 7: Generate the TC powers Pn (mW): τn

Zn

Pn = P 0 e− 10 10 Pn Pn = k=N × Pr [mW ]  k=1 Pk Z n ∼ N(0, σ Z ), n = 1, 2, . . . , N

(21) (22) (23)

where P 0 is the average power in the first arriving TC,

is the cluster decay time constant, and Z n is a lognormal random variable with 0-dB mean and standard deviation σ Z (see Table IV). Equation (22) ensures that the sum of cluster powers adds up to the total omnidirectional received power Pr . Note that P 0 cancels out in (22) using (21), but can be used as a secondary statistic to validate the channel model [38]. The 3GPP, WINNER, COST, and METIS models also estimate mean cluster powers using an exponential function over time delay, as in (21) [37], [41], [42], [47].

SAMIMI AND RAPPAPORT: 3-D mmWAVE STATISTICAL CHANNEL MODEL FOR 5G WIRELESS SYSTEM DESIGN

Step 8: Generate the cluster SP powers m,n (mW): −

ρm,n

Um,n

m,n = 0 e γ 10 10 m,n × Pn [mW] m,n = k=N  k=1 k,n Um,n ∼ N(0, σU )

2219

using the SL angles found in Step 11: (24) (25) (26)

θm,n,AOD = θi + (θi )m,n,AOD φm,n,AOD = φi + (φi )m,n,AOD

(32) (33)

θm,n,AOA = θ j + (θ j )m,n,AOA φm,n,AOA = φ j + (φ j )m,n,AOA

(34) (35)

where where 0 is the average power in the first received intra-cluster SP, γ is the SP decay time constant, and Um,n is a lognormal random variable with 0-dB mean and standard deviation σU (see Table IV); m = 1, 2, . . . , Mn and n = 1, 2, . . . , N. Equation (25) ensures that the sum of SP powers adds up to the cluster power. For model validation, the SP path losses were thresholded at 180 dB (maximum measurable path loss [12]). Note that the measurements have much greater temporal and spatial resolution than previous models. Intra-cluster power levels were observed to fall off exponentially over intra-cluster time delay (see [38, Fig. 4]). Step 9: Generate the SP phases ϕm,n (rad): ϕm,n ∼ U (0, 2π)

(27)

where m = 1, . . . , Mn and n = 1, 2, . . . , N. Different from [38] where phases are estimated from frequency and delays, here the SP phases are assumed i.i.d., and uniform between 0 and 2π [65] since each SP may experience a different scattering environment, thus arriving at arbitrary AOA SL. Step 10: Recover absolute time delays tm,n of cluster SPs using the TX-RX separation distance d (Step 1): tm,n = t0 + τn + ρm,n , t0 =

d c

(28)

where m = 1, 2, . . . , Mn , n = 1, 2, . . . , N, and c = 3 × 108 m/s is the speed of light in free space. Step 11a: Generate the mean AOA and AOD azimuth angles θi (◦ ) of the 3-D SLs to avoid overlap of lobe angles: θi ∼ U (θmin , θmax ), i = 1, 2, . . . , L 360(i − 1) 360i , θmax = θmin = L L

(29) (30)

Step 11b: Generate the mean AOA and AOD elevation angles φi (◦ ) of the 3-D SLs: φi ∼ N(μ, σ ), i = 1, 2, . . . , L.

(31)

Values of φi are defined with respect to horizon, namely, a positive and a negative value indicate a direction above and below horizon, respectively. While the 28-GHz measurements used a fixed 10° downtilt at the TX, and considered elevation planes of 0°, and ±20° at the RX, mmWave transceivers will most likely beamform in the strongest directions, as emulated in the 73-GHz measurements [27]. Consequently, the provided elevation angle distributions for all frequency scenarios are extracted from the 73-GHz measurements (see Table IV). Step 12: Generate the AOD angles (θm,n,AOD, φm,n,AOD ) and AOA angles (θm,n,AOA , φm,n,AOA ) of each SP component

i ∼ DU[1, L AOD], j ∼ DU[1, L AOA ] (θi )m,n,AOD ∼ N(0, σθ,AOD )

(36) (37)

(φi )m,n,AOD ∼ N(0, σφ,AOD ) (θ j )m,n,AOA ∼ N(0, σθ,AOA )

(38) (39)

(φ j )m,n,AOA ∼ Laplace(σφ,AOA ).

(40)

This step assigns to each multipath component a single spatial AOD and AOA lobe in a uniform random fashion, in addition to a random angular offset within the SL with distributions specified in (37)–(40). Note that the Laplace distribution in (40) provided a better fit to all data across frequencies and environments than a normal distribution. The 3GPP model uses a uniform distribution from −40° to +40° to generate path azimuth AODs, and for path azimuth AOAs uses a zero-mean normal distribution whose variance is a function of path powers for the UMi scenario [41]. The WINNER models use a wrapped Gaussian distribution to generate path AODs and AOAs [42]. B. SSCM Implementation 1) CIRs at a Particular T-R Separation Distance: To facilitate the implementation of this SSCM, Tables III and IV provide the necessary parameters required in Steps 2, 3, 5–8, 11b, and 12 for generating CIRs at a particular T-R separation distance as a function of the frequency-scenarios considered in this work. The SSCM also incorporates polarization effects, through the use of complex-amplitude antenna patterns [see (5)], which specify the amplitude and phase induced by an antenna beam for (θ, φ) angles. Cross-polarization losses can be accounted for in the omnidirectional path loss (see Step 2 in Section V-A) using available mmWave crosspolarization path loss models, as described in [12] and [27] providing cross-polarization discrimination factors for indoor and outdoor mmWave channels. Finally, XPR statistics of individual multipath components have been measured and modeled by a Gaussian lognormal distribution to account for mmWave channel depolarization [17]. 2) Local Area CIR: Small-scale spatial fading of individual 2.5-ns resolvable multipath components were measured and shown to follow a Rician distribution with K -factors in the range 9–15 and 3–7 dB in vertical-to-vertical (V-V) and vertical-to-horizontal (V-H) LOS channels, respectively, and with K -factors in the range 5–8 and 3–7 dB in V-V and V-H NLOS channels, respectively [70]. The autocorrelation of individual multipath amplitudes was accurately modeled with an exponential function over antenna separation distance [70]. Knowledge of mmWave small-scale fading distributions and autocorrelation of individual multipath components

2220

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

has allowed for realistic 5G MIMO system-level simulations as demonstrated in [69]. The MIMO channel matrix Hl for the lth multipath component, between a transmitter with Nt antenna elements and a receiver with Nr antenna elements, is computed as in [69] 1/2

1/2

Hl = Rr Hw Rt

(41)

where Hw is an Nt × Nr matrix with entries generated from an i.i.d. Rician distribution, and Rt and Rr are the transmit and receive spatial correlation matrices for userdefined antenna arrays (e.g., uniform linear array, uniform rectangular array) [69]. The statistical channel models presented in Section V-A have been implemented in a MATLAB-based simulator, which generates CIRs at a particular T-R separation distance and local area omnidirectional and directional CIRs, with complementary MATLAB source files freely downloadable in [18]. 3) Track or Grid CIR: The time-varying channel coefficients along a moving receiver can be obtained by accounting for the Doppler phase shifts of individual multipaths, based on the multipath AOA and receiver velocity vector. The largescale spatially varying birth and death of TCs and SLs are analogous to the 3GPP birth and death processes defined for delay-angular clusters [41], [42], and can be used to enable realistic MU-MIMO simulations over a user-defined trajectory along a track or grid of points. Knowledge of the autocorrelation for the number of TCs and SLs as a function of MS distance for closely spaced location points with known CIRs offers the possibility of implementing statistical CIRs over a track or grid of points. The spatial autocorrelation between two MSs with separation distance dMS for the number of SLs and the number of TCs can be computed as in [69], ρ(dMS ) =

Fig. 15. Example of simulated 28-GHz NLOS PDP obtained from the MATLAB-based statistical simulator.

E[L(r + dMS )L(r )] − E[L(r + dMS )]E[L(r )] σ [L(r + dMS )]σ [L(r )] (42)

where σ [L(r )] =



E[L(r ) − E[L(r )]]2

(43)

where L(r ) is the number of SLs or the number of TCs at T-R separation r , E[·] is the expectation operator over all possible values of r , and σ [L(r )] is the standard deviation of the number of TC or SL at distance r . C. Sample Output Functions Figs. 15 and 16 show example output functions of a 28GHz NLOS omnidirectional PDP, and corresponding AOA 3-D power spectrum, obtained from a MATLAB-based statistical simulator that implemented the channel models given in (8)–(40). The generated PDP in Fig. 15 is composed of four multipath taps, grouped into two TCs with exponentially decaying amplitudes with cluster decay constant = 49.4 ns and intra-cluster SP decay constant γ = 16.9 ns (see Steps 6 and 7 in Section V-A). Here, a transmit power of 30 dBm is used with a noise threshold set to −100 dBm over an 800-MHz RF bandwidth, and the simulated PDP has a total omnidirectional path loss of 120 dB with a TX-RX separation

Fig. 16. Example of simulated 28-GHz NLOS 3-D AOD power spectrum obtained from the MATLAB-based statistical simulator. This spectrum is associated with the PDP shown in Fig. 15.

distance of 112 m, an RMS DS of 50 ns, and 0-dBi TX and RX antenna gains. The AOA spectrum (Fig. 16) shows the four multipath taps grouped into two AOA SLs according to (34), (35), (39), and (40). Note that the multipaths in Fig. 16 may be convolved with an arbitrary antenna pattern to study the effects of antenna beamwidths on the CIR. The 28- and 73-GHz NLOS cluster and SP power parameters and γ shown in Table III (rows 4 and 5) are similar, with = 49.4 ns (28 GHz) and = 56.0 ns (73 GHz) and with γ = 16.9 ns (28 GHz) and γ = 15.3 ns (73 GHz), obtained with a 25-ns minimum inter-cluster void interval. The ( , γ) parameters were extracted for various values of minimum inter-cluster void interval, namely, 15 and 20 ns, and showed similar parameter values across frequency bands indicating similar channel dynamics at 28 and 73 GHz. However, note that the cluster and SP decay constants ( , γ) obtained with a 25-ns minimum intercluster void interval recreated the measured statistics most accurately.

SAMIMI AND RAPPAPORT: 3-D mmWAVE STATISTICAL CHANNEL MODEL FOR 5G WIRELESS SYSTEM DESIGN

The channel models provided herein are valid for a 2.5-ns multipath time resolution (800 MHz) and antenna beamwidths of 7°–10° [16]. However, it is also possible to synthesize omnidirectional and directional CIRs for bandwidths less than 800 MHz and for angular resolutions greater than 7°. In the time domain, it is sufficient to coherently sum (e.g., convolve) the multipath components that fall within a time bin width equal to the inverse of the desired bandwidth. For instance, for a desired RF channel bandwidth of 400 MHz (200-MHz baseband bandwidth), a user may coherently sum the multipath amplitudes in (4) (valid for a 2.5-ns time resolution) that fall within a bin width of 5 ns. The angular resolution of the resulting IR can be adjusted by convolving the SSCM output with a 3-D complex antenna pattern with HPBW greater than 7°, in (5) and (45) [16]. The first- and second-order statistics, as well as the cumulative distribution function (CDF) tail behavior of simulated directional RMS DSs were properly recreated using past published measurements that used a wide range of beamwidths and bandwidths [11], [13], [59], as shown in Fig. 19.

2221

Fig. 17. Example of 28-GHz simulated omnidirectional LOS and NLOS path loss results obtained from 10 000 sample functions using the SSCM described herein. There is excellent agreement between simulated and measured PLEs and shadow factors.

VI. S IMULATION R ESULTS The statistical channel model presented in (8)–(40) was used to implement a MATLAB-based statistical simulator to verify and validate the accuracy of the simulated temporal and spatial statistics when compared against the measured statistics. A large simulation was carried out for each of the four frequency scenarios presented in this work, in which 10 000 omnidirectional PDPs, and 3-D AOD and AOA power spectra were generated as sample functions of (4). We used simple number generators to obtain the number of TCs, the numbers of AOD and AOA SLs, cluster and SP delays, and cluster and SP powers, as described in Section V-A. We set the TX power to 30 dBm and assume a noise floor of −100 dBm over an 800-MHz RF bandwidth, with the TX and RX antenna gains set to 24.5 and 27 dBi to emulate the 28- and 73-GHz measurements [11], [57] when performing directional simulations. A. Simulated Omnidirectional Path Loss Fig. 17 shows the 28-GHz simulated omnidirectional LOS and NLOS path losses obtained using the statistical channel model presented in (8)–(28). The LOS and NLOS conditions used in the SSCM were determined based on a probabilistic function describing the LOS probability PLOS (d) between a transmitter and a receiver in a dense urban microcellular environment, obtained using 3-D ray-tracing techniques as proposed in [35]  2    dBP − αd − dα ,1 1 − e PLOS (d) = min (44) +e d where dBP = 27 m and α = 71 m. Steps 1–10 of Section V-A were implemented to recover LOS and NLOS wideband PDPs, and the corresponding path losses were computed by subtracting the total received power (dBm) from the TX power (dBm) for each of the simulated profiles. The simulated PLE and large-scale shadow factor were 2.0 and 3.6 dB in LOS, and 3.4 and 9.7 dB in NLOS, respectively, which are in agreement

Fig. 18. Combined 28–73-GHz LOS and NLOS omnidirectional RMS DSs synthesized from absolute timing PDPs, superimposed with the simulation results of RMS DSs obtained from 10 000 sample functions using the SSCM described herein [16].

with the measured parameters shown in Table III and in [27]. The simulated statistics match the measured statistics very well, over a large ensemble of generated channels, and can be used to realistically reproduce the measurements in the context of a system-level simulation. B. Simulated RMS Delay Spreads The RMS DS describes channel temporal dispersion, and is a vital statistic that an SSCM should reproduce faithfully. Fig. 18 shows the simulated omnidirectional RMS DSs, compared with the RMS DSs obtained from the absolute timing PDPs, at both 28 and 73 GHz in LOS and NLOS scenarios. We obtained 18 and 16 ns for the empirical and simulated medians, respectively, for the combined 28–73-GHz LOS scenario, and 32 and 35 ns for the empirical and simulated medians, respectively, for the combined 28–73-GHz NLOS scenario. The lack of measured data considerably skewed the empirical distributions, so the median was chosen

2222

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

Fig. 19. Simulated directional RMS DS CDFs (lines) for various frequencies and for antenna beamwidths, obtained from directional PDPs generated using (5). Values reported from the literature are shown as points [11], [13], [59]. C stands for cellular, and P2P stands for peer-to-peer.

to represent the distribution trend. The empirical and simulated medians in NLOS were 31 and 32 ns at 28 GHz (see [38, Fig. 5]), respectively, and 47 and 39 ns at 73 GHz, respectively, showing good agreement with empirical values. The strong agreement between simulated and empirical distributions validates the temporal component of the SSCM. The SSCM model can also recreate directional PDPs at arbitrary TX-RX pointing angle combination for user-defined azimuth and elevation antenna HPBWs. It is possible to reconstruct the temporal and spatial statistics of arbitrary antenna beamwidths by weighting the multipath component power levels with a desired antenna pattern, such that the multipath components closest to a desired direction are amplified, while those farthest away are effectively set to 0, as shown in (5). When simulating directional PDPs, the multipath components were weighted by G TX (θ, φ) and G RX (θ, φ), the TX and RX antenna patterns, respectively, commonly parameterized as follows [72]:   αθ 2 +βφ 2 G 0 (45) G(θ, φ) = max G 0 e , 100 4 ln(2) 4 ln(2) 41253η α = 2 , β = 2 , G0 = (46) θ3dB φ3dB θ3dB φ3dB where (θ, φ) are the azimuth and elevation angle offsets from the boresight direction in degrees, G 0 is the maximum directive gain (boresight gain) in linear units, (θ3dB , φ3dB ) are the azimuth and elevation HPBWs in degrees, α, β are parameters that depend on the HPBW values, and η = 0.7 is a typical average antenna efficiency. Fig. 19 shows simulated directional RMS DSs obtained from the 28- and 73-GHz channel models presented here, in comparison with reported values in the literature for the 10%, 50%, and 90% CDF points of measured directional RMS DSs at 28, 38, 60, and 73 GHz [13], [31], [73], [74], for antenna beamwidths of 7.3°, 10.9°, 28.8°, and 49.4°. To test (5), we generated omnidirectional sample functions using the presented SSCM, convolved (multiplied) the omnidirectional CIRs with the simulated antenna patterns [see (45)],

Fig. 20. 28-GHz NLOS global AS and elevation spread, obtained from the NYC measurements [12] and the 3-D SSCM.

and computed directional RMS DSs from directional PDPs obtained using (5) for 20 T-R separation distances using antenna HPBWs of 10°, 7°, and 30° in azimuth/elevation at 28 and 73 GHz to emulate the NYC measurements. Note the empirical CDFs consist of all available data, while the simulated directional RMS DSs were obtained from channel models extracted exclusively from up to four strongest AOD and AOA PDP data. Fig. 19 shows the simulated and measured RMS DS distributions match relatively well across antenna beamwidths and many mmWave bands. C. Simulated RMS Angular Spreads The omnidirectional azimuth and elevation spreads describe the degree of angular dispersion at a BS or MS over the entire 4π steradian sphere [41], [42], also termed global angular spreads in [17] and [47]. The AOD and AOA global angular spreads were computed from all available 28 and 73 GHz NLOS measured data, using the total (integrated over delay) received power at unique azimuth/elevation pointing angles, but not requiring absolute multipath time delays, and the equations in [41, Annex A]. These were compared with the simulated angular spreads using the 3-D SSCM, where the SSCM was developed from the statistics of up to four strong measured angles. The simulated and measured mean global angular spreads match relatively well at 28 and 73 GHz, as can be seen from Figs. 20 and 21. The slight differences (slight under-estimation or over-estimation of global ASs using Step 12 in Section V-A) may be due to the model focusing only on the multipath components contained in the strongest several SLs measured at every location, which will in actuality be the strongest components in a practical wireless system. The directional AOD and AOA RMS lobe AS and elevation spread were also computed based on a −10 dB lobe threshold [38] from the 28 and 73 GHz data, and compared with simulated values using the 3-D SSCM. Fig. 22 shows typical measured against simulated AOA RMS lobe angular spreads for the 73-GHz NLOS scenario, showing an excellent match over the empirical and simulated means of 4° and 2° in

SAMIMI AND RAPPAPORT: 3-D mmWAVE STATISTICAL CHANNEL MODEL FOR 5G WIRELESS SYSTEM DESIGN

Fig. 21. 73-GHz NLOS global AS and elevation spread, obtained from the NYC measurements [12] and the 3-D SSCM.

Fig. 22. 73-GHz NLOS AOA RMS lobe AS and elevation spread, measured and simulated, showing good agreement [38]. Only non-zero simulated RMS lobe angular spreads are plotted, to provide a fair comparison with the measurements. Approximately 4% of the simulated data yielded zero-valued RMS lobe angular spreads (e.g., only a single angular segment in a lobe).

azimuth and elevation, respectively, over all measured SLs. Similar agreement was found for the 28-GHz NLOS and LOS data sets, across AOD and AOA SLs, indicating that the model accurately recreates the empirical spatial statistics. Fig. 21 shows only non-zero simulated RMS lobe angular spreads, to remain consistent with field observations that showed that all the measured SLs are composed of at least two multipath components, corresponding to non-zero rms lobe angular spreads. However, approximately 4% of all simulated data was composed of zero-valued RMS lobe angular spreads, physically denoting that 4% of all SLs are composed of just one multipath component (e.g., zero angle spread in a lobe). VII. D ISCUSSION Twelve thousand measured PDPs have been used to create a statistical channel model that reproduces realistic wideband

2223

power IRs and joint AOD-AOA power spectra in 3-D based on the 28 and 73 GHz New York City measurements. L AOD, the average number of AOD SLs at the transmitter, is Poisson distributed and frequency independent with a mean of about 2 and 1.5 in LOS and NLOS environments, respectively, indicating that strong multipath signals can reach the receiver from more departing angles in LOS compared to NLOS. However, L AOA , the average number of AOA SLs at the receiver, is frequency dependent, with a mean of 1.6 and 2.5 at 28 and 73 GHz in NLOS, respectively, implying more pronounced diffuse scattering with increasing carrier frequency. SLs conveniently represent the mmWave radio channel because they implicitly account for directionality, a key differentiator of future wireless cellular and mobile systems operating in the mmWave spectrum compared with today’s UHF and microwave systems. TC and SP powers were shown to decay exponentially with little dependence on the carrier frequency, with typical

and γ of 25 and 17 ns in LOS, and 51 and 16 ns, in NLOS, respectively, when combining the 28 and 73 GHz data sets, using a 25-ns minimum intercluster void interval. In some cases, TCs can carry up to 80% of the total received power at excess delays larger than 100 ns, thus inducing large RMS DSs. VIII. C ONCLUSION This paper presented a 3-D statistical channel model for mmWave LOS and NLOS communications for link, local area, and track CIRs, with arbitrary carrier frequency, signal bandwidth, and antenna beamwidth, invaluable in the design of next-generation 5G mmWave cellular networks. The simplicity of the presented statistical IR model stems from the independent modeling of time and space, allowing for simple statistical distributions to promote ease of use in simulated software and hardware implementations, while reproducing physically based CIRs and 3-D power angular spectra based on real-world measurements. Arbitrary antenna beamwidths and frequency are supported, and we showed good agreement between the model and published RMS DSs. The TCSL model parameters, provided in Table II, may be included as simple extensions to existing 3GPP parameters to account for observed mmWave channel directionality, as given in [17]. A MATLAB-based statistical simulator, available in [18], was implemented to generate a large ensemble of PDPs and 3-D power angular spectra, showing good agreement with field measurements, thereby validating the 3-D SSCM for the design of next-generation wireless systems that will make use of sub-6 GHz and mmWave channel models for the design of filters, multielement antenna arrays, and mmWave transceivers. ACKNOWLEDGMENT The authors would like to thank S. Sun, G. R. MacCartney Jr., and the NYU WIRELESS Industrial Affiliates for their support, and Prof. H. Bertoni and Prof. S. Rangan for valuable discussions. R EFERENCES [1] K. Haneda et al., “5G 3GPP-like channel models for outdoor urban microcellular and macrocellular environments,” in Proc. IEEE 83rd Veh. Technol. Conf. (VTC-Spring), May 2016.

2224

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

[2] K. Haneda et al., “Indoor 5G 3GPP-like channel models for office and shopping mall environments,” in Proc. IEEE Int. Conf. Commun. (ICC) Workshops, May 2016. [3] L. Lu, G. Y. Li, A. L. Swindlehurst, A. Ashikhmin, and R. Zhang, “An overview of massive MIMO: Benefits and challenges,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 742–758, Oct. 2014. [4] 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. [5] V. Garcia, Y. Zhou, and J. Shi, “Coordinated multipoint transmission in dense cellular networks with user-centric adaptive clustering,” IEEE Trans. Wireless Commun., vol. 13, no. 8, pp. 4297–4308, Aug. 2014. [6] V. Jungnickel et al., “The role of small cells, coordinated multipoint, and massive MIMO in 5G,” IEEE Commun. Mag., vol. 52, no. 5, pp. 44–51, May 2014. [7] G. Nigam, P. Minero, and M. Haenggi, “Coordinated multipoint joint transmission in heterogeneous networks,” IEEE Trans. Commun., vol. 62, no. 11, pp. 4134–4146, Nov. 2014. [8] Z. Khan, H. Ahmadi, E. Hossain, M. Coupechoux, L. A. Dasilva, and J. J. Lehtomäki, “Carrier aggregation/channel bonding in next generation cellular networks: Methods and challenges,” IEEE Netw., vol. 28, no. 6, pp. 34–40, Nov. 2014. [9] F. Khan and Z. Pi, “mmWave mobile broadband (MMB): Unleashing the 3–300GHz spectrum,” in Proc. 34th IEEE Sarnoff Symp., May 2011, pp. 1–6. [10] T. S. Rappaport, R. W. Heath, Jr., R. C. Daniels, and J. N. Murdock, Millimeter Wave Wireless Communications. Englewood Cliffs, NJ, USA: Prentice-Hall, 2015. [11] T. S. Rappaport et al., “Millimeter wave mobile communications for 5G cellular: It will work!” IEEE Access, vol. 1, pp. 335–349, May 2013. [12] T. S. Rappaport, G. R. Maccartney, Jr., M. K. Samimi, and S. Sun, “Wideband millimeter-wave propagation measurements and channel models for future wireless communication system design,” IEEE Trans. Commun., vol. 63, no. 9, pp. 3029–3056, Sep. 2015. [13] T. S. Rappaport, F. Gutierrez, Jr., E. Ben-Dor, J. N. Murdock, Y. Qiao, and J. I. Tamir, “Broadband millimeter-wave propagation measurements and models using adaptive-beam antennas for outdoor urban cellular communications,” IEEE Trans. Antennas Propag., vol. 61, no. 4, pp. 1850–1859, Apr. 2013. [14] A. Maltsev et al., “WP5: Propagation, antennas and multiantenna techniques—D5.1: Channel modeling and characterization,” Millim.-Wave Evol. Backhaul Access (MiWEBA), Jun. 2014. [15] M. K. Samimi et al., “28 GHz angle of arrival and angle of departure analysis for outdoor cellular communications using steerable beam antennas in New York City,” in Proc. IEEE 77th Veh. Technol. Conf. (VTC Spring), Jun. 2013, pp. 1–6. [16] M. K. Samimi and T. S. Rappaport, “Statistical channel model with multi-frequency and arbitrary antenna beamwidth for millimeterwave outdoor communications,” in Proc. IEEE Globecom Workshops, Dec. 2015, pp. 1–7. [17] M. K. Samimi and T. S. Rappaport, “Local multipath model parameters for generating 5G millimeter-wave 3GPP-like channel impulse response,” in Proc. 10th Eur. Conf. Antennas Propag. (EuCAP), Apr. 2016, pp. 1–5. [18] NYU WIRELESS. Open Source Downloadable 5G Channel Simulator Software, accessed on Apr. 2016. [Online]. Available: http://bit.ly/1WNPpDX [19] P. Soma, Y. W. -M. Chia, and L. C. Ong, “Modeling and analysis of time varying radio propagation channel for LMDS,” in Proc. IEEE Radio Wireless Conf. (RAWCON), Sep. 2000, pp. 115–118. [20] A. F. Elrefaie and M. Shakouri, “Propagation measurements at 28 GHz for coverage evaluation of local multipoint distribution service,” in Proc. Wireless Commun. Conf., Aug. 1997, pp. 12–17. [21] S. Y. Seidel, “Radio propagation and planning at 28 GHz for local multipoint distribution service (LMDS),” in Proc. IEEE Antennas Propag. Soc. Int. Symp., vol. 2. Jun. 1998, pp. 622–625. [22] S. Geng, J. Kivinen, X. Zhao, and P. Vainikainen, “Millimeter-wave propagation channel characterization for short-range wireless communications,” IEEE Trans. Veh. Technol., vol. 58, no. 1, pp. 3–13, Jan. 2009. [23] H. Xu, V. Kukshya, and T. S. Rappaport, “Spatial and temporal characteristics of 60-GHz indoor channels,” IEEE J. Sel. Areas Commun., vol. 20, no. 3, pp. 620–630, Apr. 2002. [24] P. Soma, L. C. Ong, S. Sun, and M. Y. W. Chia, “Propagation measurements and modeling of LMDS radio channel in Singapore,” IEEE Trans. Veh. Technol., vol. 52, no. 3, pp. 595–606, May 2003.

[25] P. F. M. Smulders and A. G. Wagemans, “Wideband indoor radio propagation measurements at 58 GHz,” Electron. Lett., vol. 28, no. 13, pp. 1270–1272, Jun. 1992. [26] P. F. M. Smulders and A. G. Wagemans, “A statistical model for the MM-wave indoor radio channel,” in Proc. 3rd IEEE Int. Symp. Pers., Indoor Mobile Radio Commun. (PIMRC), Oct. 1992, pp. 303–307. [27] G. R. Maccartney, Jr., T. S. Rappaport, S. Sun, and S. Deng, “Indoor office wideband millimeter-wave propagation measurements and channel models at 28 and 73 GHz for ultra-dense 5G wireless networks,” IEEE Access, vol. 3, pp. 2388–2424, 2015. [28] W. Roh et al., “Millimeter-wave beamforming as an enabling technology for 5G cellular communications: Theoretical feasibility and prototype results,” IEEE Commun. Mag., vol. 52, no. 2, pp. 106–113, Feb. 2014. [29] K. Haneda, J. Järveläinen, A. Karttunen, M. Kyrö, and J. Putkonen, “Indoor short-range radio propagation measurements at 60 and 70 GHz,” in Proc. 8th Eur. Conf. Antennas Propag. (EuCAP), Apr. 2014, pp. 634–638. [30] R. Müller et al., “Ultra-wideband channel sounder for measurements at 70 GHz,” in Proc. IEEE Veh. Technol. Conf. (VTC), May 2014, pp. 1–5. [31] G. R. Maccartney, Jr., M. K. Samimi, and T. S. Rappaport, “Exploiting directionality for millimeter-wave wireless system improvement,” in Proc. IEEE Int. Conf. Commun. (ICC), Jun. 2015, pp. 2416–2422. [32] S. Hur et al., “Proposal on millimeter-wave channel modeling for 5G cellular system,” IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 454–469, Apr. 2016. [33] S. Sun et al., “Investigation of prediction accuracy, sensitivity, and parameter stability of large-scale propagation path loss models for 5G wireless communications,” IEEE Trans. Veh. Technol., vol. 65, no. 5, pp. 2843–2860, May 2016. [34] T. S. Thomas, H. C. Nguyen, G. R. MacCartney, Jr., and T. S. Rappaport, “3D mmWave channel model proposal,” in Proc. IEEE 80th Veh. Technol. Conf. (VTC Fall), Sep. 2014, pp. 1–6. [35] M. K. Samimi, T. S. Rappaport, and G. R. MacCartney, Jr., “Probabilistic omnidirectional path loss models for millimeter-wave outdoor communications,” IEEE Wireless Commun. Lett., vol. 4, no. 4, pp. 357–360, Aug. 2015. [36] H. C. Nguyen, G. R. MacCartney, Jr., T. Thomas, T. S. Rappaport, B. Vejlgaard, and P. Mogensen, “Evaluation of empirical ray-tracing model for an urban outdoor scenario at 73 GHz E-band,” in Proc. IEEE 80th Veh. Technol. Conf. (VTC-Fall), Sep. 2014, pp. 1–6. [37] V. Nurmela et al., “Deliverable D1.4: METIS channel models,” Mobile Wireless Commun. Enablers Twenty-Twenty Inf. Soc. (METIS), Apr. 2014. [38] M. K. Samimi and T. S. Rappaport, “3-D statistical channel model for millimeter-wave outdoor mobile broadband communications,” in Proc. IEEE Int. Conf. Commun. (ICC), Jun. 2015, pp. 2430–2436. [39] S. Sun, T. S. Rappaport, R. W. Heath, Jr., A. Nix, and S. Rangan, “MIMO for millimeter-wave wireless communications: Beamforming, spatial multiplexing, or both?” IEEE Commun. Mag., vol. 52, no. 12, pp. 110–121, Dec. 2014. [40] M. K. Samimi and T. S. Rappaport, “Ultra-wideband statistical channel model for non line of sight millimeter-wave urban channels,” in Proc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2014, pp. 3483–3489. [41] Spatial Channel Model for Multiple Input Multiple Output (MIMO) Simulations, document 3GPP TR 25.996 V12.0.0, Sep. 2014. [42] P. Kyösti et al., “WINNER II channel models,” Eur. Commission, Tech. Rep. D1.1.2 V1.1, Sep. 2007. [Online]. Available: http://projects.celticinitiative.org/winner+/WINNER2-Deliverables/ [43] A. Algans, K. I. Pedersen, and P. E. Mogensen, “Experimental analysis of the joint statistical properties of azimuth spread, delay spread, and shadow fading,” IEEE J. Sel. Areas Commun., vol. 20, no. 3, pp. 523–531, Apr. 2002. [44] G. Calcev et al., “A wideband spatial channel model for system-wide simulations,” IEEE Trans. Veh. Technol., vol. 56, no. 2, pp. 389–403, Mar. 2007. [45] B. H. Fleury, M. Tschudin, R. Heddergott, D. Dahlhaus, and K. I. Pedersen, “Channel parameter estimation in mobile radio environments using the SAGE algorithm,” IEEE J. Sel. Areas Commun., vol. 17, no. 3, pp. 434–450, Mar. 1999. [46] N. Czink, P. Cera, J. Salo, E. Bonek, J.-P. Nuutinen, and J. Ylitalo, “A framework for automatic clustering of parametric MIMO channel data including path powers,” in Proc. IEEE 64th Veh. Technol. Conf. (VTC-Fall), Sep. 2006, pp. 1–5. [47] L. Liu et al., “The COST 2100 MIMO channel model,” IEEE Wireless Commun., vol. 19, no. 6, pp. 92–99, Dec. 2012.

SAMIMI AND RAPPAPORT: 3-D mmWAVE STATISTICAL CHANNEL MODEL FOR 5G WIRELESS SYSTEM DESIGN

[48] R. Verdone and A. Zanella, (Eds.), Pervasive Mobile and Ambient Wireless Communications: COST Action 2100. London, U.K.: SpringerVerlag London Ltd., 2012. [49] T. S. Rappaport, S. Y. Seidel, and K. Takamizawa, “Statistical channel impulse response models for factory and open plan building radio communicate system design,” IEEE Trans. Commun., vol. 39, no. 5, pp. 794–807, May 1991. [50] T. S. Rappaport, W. Huang, and M. J. Feuerstein, “Performance of decision feedback equalizers in simulated urban and indoor radio channels,” IEICE Trans. Commun., vol. E76-B, no. 2, pp. 78–89, Feb. 1993. [51] T. S. Rappaport, S. Y. Seidel, and R. Singh, “900-MHz multipath propagation measurements for US digital cellular radiotelephone,” IEEE Trans. Veh. Technol., vol. 39, no. 2, pp. 132–139, May 1990. [52] S. Y. Seidel, T. S. Rappaport, S. Jain, M. L. Lord, and R. Singh, “Path loss, scattering and multipath delay statistics in four European cities for digital cellular and microcellular radiotelephone,” IEEE Trans. Veh. Technol., vol. 40, no. 4, pp. 721–730, Nov. 1991. [53] T. S. Rappaport, Wireless Communications: Principles and Practice (Prentice-Hall Commun. Eng. Emerg. Technol. Series), 2nd ed. Upper Saddle River, NJ, USA: Prentice-Hall, 2002. [54] J. E. Nuckols, “Implementation of geometrically based single-bounce models for simulation of angle-of-arrival of multipath delay components in the wireless channel simulation tools, SMRCIM and SIRCIM,” M.S. thesis, Dept. Elect. Eng., Virginia Polytech. Inst. and State Univ., Blacksburg, VA, USA, 1999. [55] K. Yu, M. Bengtsson, B. Ottersten, D. McNamara, P. Karlsson, and M. Beach, “A wideband statistical model for NLOS indoor MIMO channels,” in Proc. IEEE 55th Veh. Technol. Conf. (VTC-Spring), vol. 1. May 2002, pp. 370–374. [56] B. Han and Y. R. Zheng, “Higher rank principal Kronecker model for triply selective fading channels with experimental validation,” IEEE Trans. Veh. Technol., vol. 64, no. 5, pp. 1654–1663, May 2015. [57] G. R. MacCartney, Jr., and T. S. Rappaport, “73 GHz millimeter wave propagation measurements for outdoor urban mobile and backhaul communications in New York City,” in Proc. IEEE Int. Conf. Commun. (ICC), Jun. 2014, pp. 4862–4867. [58] S. Deng, C. J. Slezak, G. R. MacCartney, Jr., and T. S. Rappaport, “Small wavelengths—Big potential: Millimeter wave propagation measurements for 5G,” Microw. J., vol. 57, no. 11, pp. 4–12, 2014. [59] G. R. MacCartney, Jr., M. K. Samimi, and T. S. Rappaport, “Omnidirectional path loss models in New York City at 28 GHz and 73 GHz,” in Proc. IEEE 25th Int. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC), Sep. 2014, pp. 227–231. [60] S. Sun, G. R. MacCartney, Jr., M. K. Samimi, and T. S. Rappaport, “Synthesizing omnidirectional antenna patterns, received power and path loss from directional antennas for 5G millimeter-wave communications,” in Proc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2015, pp. 1–7. [61] T. S. Rappaport and S. Deng, “73 GHz wideband millimeter-wave foliage and ground reflection measurements and models,” in Proc. IEEE Int. Conf. Commun. Workshop (ICCW), Jun. 2015, pp. 1238–1243. [62] T. A. Thomas et al., “A prediction study of path loss models from 2–73.5 GHz in an urban-macro environment,” in Proc. IEEE Veh. Technol. Conf. (VTC-Spring), May 2016. [63] S. Sun et al., “Propagation path loss models for 5G urban micro- and macro-cellular scenarios,” in Proc. IEEE Veh. Technol. Conf. (VTC-Spring), May 2016. [64] M. Steinbauer, A. F. Molisch, and E. Bonek, “The double-directional radio channel,” IEEE Antennas Propag. Mag., vol. 43, no. 4, pp. 51–63, Aug. 2001. [65] A. A. M. Saleh and R. A. Valenzuela, “A statistical model for indoor multipath propagation,” IEEE J. Sel. Areas Commun., vol. 5, no. 2, pp. 128–137, Feb. 1987. [66] Q. Spencer, M. Rice, B. Jeffs, and M. Jensen, “A statistical model for angle of arrival in indoor multipath propagation,” in Proc. IEEE 47th Veh. Technol. Conf. (VTC), vol. 3. May 1997, pp. 1415–1419. [67] R. B. Ertel, P. Cardieri, K. W. Sowerby, T. S. Rappaport, and J. H. Reed, “Overview of spatial channel models for antenna array communication systems,” IEEE Pers. Commun., vol. 5, no. 1, pp. 10–22, Feb. 1998.

2225

[68] F. Gutierrez, Jr., S. Agarwal, K. Parrish, and T. S. Rappaport, “On-chip integrated antenna structures in CMOS for 60 GHz WPAN systems,” IEEE J. Solid-State Circuits, vol. 27, no. 8, pp. 1367–1378, Oct. 2009. [69] M. K. Samimi, S. Sun, and T. S. Rappaport, “MIMO channel modeling and capacity analysis for 5G millimeter-wave wireless systems,” in Proc. 10th Eur. Conf. Antennas Propag. (EuCAP), Apr. 2016, pp. 1–5. [70] M. K. Samimi, G. R. MacCartney, Jr., S. Sun, and T. S. Rappaport, “28 GHz millimeter-wave ultrawideband small-scale fading models in wireless channels,” in Proc. IEEE Veh. Technol. Conf. (VTC-Spring), May 2016. [71] B. Thoma, T. S. Rappaport, and M. D. Kietz, “Simulation of bit error performance and outage probability of /4 DQPSK in frequencyselective indoor radio channels using a measurement-based channel model,” in Proc. IEEE Global Telecommun. Conf. (GLOBECOM), vol. 3. Dec. 1992, pp. 1825–1829. [72] C. A. Balanis, Antenna Theory: Analysis and Design. Hoboken, NJ, USA: Wiley, 2005. [73] T. S. Rappaport, Y. Qiao, J. I. Tamir, J. N. Murdock, and E. Ben-Dor, “Cellular broadband millimeter wave propagation and angle of arrival for adaptive beam steering systems,” in Proc. IEEE Radio Wireless Symp. (RWS), Jan. 2012, pp. 151–154. [74] T. S. Rappaport, E. Ben-Dor, J. N. Murdock, and Y. Qiao, “38 GHz and 60 GHz angle-dependent propagation for cellular & peer-to-peer wireless communications,” in Proc. IEEE Int. Conf. Commun. (ICC), Jun. 2012, pp. 4568–4573.

Mathew K. Samimi (S’13) received the B.S. degree in applied physics from the Fu Foundation School of Engineering and Applied Science, Columbia University, New York, NY, USA, in 2012, and the M.S. and Ph.D. degrees in electrical engineering from the Department of Electrical and Computer Engineering, NYU WIRELESS Research Center, New York University Tandon School of Engineering, Brooklyn, NY, USA, in 2014 and 2016, respectively. His research focuses on developing millimeterwave statistical spatial channel models for next generation ultrawideband mobile cellular systems for dense urban environments.

Theodore S. Rappaport (S’83–M’84–SM’91– F’98) is the David Lee/Ernst Weber Professor of Electrical and Computer Engineering with the New York University (NYU) Tandon School of Engineering, Brooklyn, NY, USA, and the Founding Director of the NYU WIRELESS Research Center. He also holds professorship positions with the Courant Institute of Mathematical Sciences and the NYU Langone School of Medicine. He founded major wireless research centers at the Virginia Polytechnic Institute and State University (MPRG), Blacksburg, VA, USA, The University of Texas at Austin (WNCG), Austin, TX, USA, and at NYU (NYU WIRELESS), and founded two wireless technology companies that were sold to publicly traded firms. He is a highly sought-after technical consultant, having testified before the U.S. Congress and having served the ITU. He has advised more than 100 students, has more than 100 patents issued and pending, and has authored or coauthored several books, including Wireless Communications: Principles and Practice—Second Edition (PrenticeHall, 2002). His latest book, Millimeter Wave Wireless Communications (Pearson/Prentice-Hall, 2015), is the first comprehensive text on the subject.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

1

Hybrid RF and Digital Beamformer for Cellular Networks: Algorithms, Microwave Architectures, and Measurements Vijay Venkateswaran, Member, IEEE, Florian Pivit, Member, IEEE, and Lei Guan, Member, IEEE Abstract— Modern wireless communication networks, particularly the upcoming cellular networks, utilize multiple antennas to improve capacity and signal coverage. In these systems, typically an active transceiver is connected to each antenna. However, this one-to-one mapping between transceivers and antennas will dramatically increase the cost and complexity of a large phased antenna array system. In this paper, firstly we propose a digitally steerable beamformer architecture where a reduced number of transceivers with a digital beamformer (DBF) is connected to an increased number of antennas through an RF beamforming network (RFBN). Then, based on the proposed architecture, we present a methodology to derive the minimum number of transceivers that are required for macro-cell and small-cell base stations. Subsequently, in order to achieve optimal beam patterns with given cellular standard requirements and RF operational constraints, we propose efficient algorithms to jointly design DBF and RFBN. Starting from the proposed algorithms, we specify generic microwave RFBNs for optimal macro-cell and small-cell networks. In order to verify the proposed approaches, we compare the performance of RFBN using simulations and anechoic chamber measurements. Experimental measurement results confirm the robustness and performance of the proposed hybrid DBF-RFBN concept ensuring that theoretical multi-antenna capacity and coverage are achieved at a little incremental cost. Index Terms— Active antenna arrays (AAAs), beamforming, Butler matrix, cellular networks, fifth-generation (5G), hybrid RF and digital beamforming.

I. I NTRODUCTION N A typical cellular base-station, a passive antenna array is usually connected to an RF transceiver in the form of so-called remote radio head (RRH), where each transmitted and received signal is shaped by the same antenna beam. Though this passive architecture is quite simple, it has several disadvantages in terms of its applications in fourth-generation (4G) and future/fifth-generation (5G) wireless communications: 1) it does not allow spatial separation of multiple users, which can be seen as a very efficient way to utilize limited frequency spectrum and 2) it does not improve the signal-to-noise ratio (SNR) at the user

I

Manuscript received January 26, 2016; revised April 28, 2016; accepted May 8, 2016. V. Venkateswaran was with Bell Laboratories, Alcatel-Lucent, Dublin D15 Y6NT, Ireland, where this work was carried out. He is now with Huawei Technologies, Kista 164 94, Sweden (e-mail: [email protected]). F. Pivit and L. Guan are now with Wireless Research, Bell Laboratories, Nokia, Dublin D15 Y6NT, Ireland (e-mail: [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TMTT.2016.2569583

Fig. 1. Adaptive antenna array architecture. (a) Full dimension AAA architecture with Ntrx = Nt transceivers (denoted by RF {.}) and DBF u(θd ) for beamtilt range θd ∈ Rθ . (b) Proposed architecture with two-stage beamforming—adaptive ϑ(θd ) connected to Nt × Ntrx RF beamforming network W and, finally, to Nt antennas.

equipment via the use of advanced beamforming technology, which has been largely accepted as a potential key technology for enabling 5G wireless communications. In order to improve spectral efficiency and reduce interference levels, multi-antenna RF transceiver architectures are being proposed for cellular base stations, where each antenna is connected to a dedicated RF chain and a baseband beamformer [1], as shown in Fig. 1(a). Such antenna arrays with active RF components are commonly referred as active antenna arrays (AAAs). This approach allows us to form multiple beams at the same time to/from the same array in order to either serve multiple users simultaneously or to spatially multiplex signals towards a single user. A full-size AAA architecture, i.e., each antenna element connected to a dedicated RF chain, significantly increases the cost, weight, and overall power consumption because of its inherent one-to-one mapping between antennas and RF transceivers. In order to reduce the complexity of a full-size AAA architecture, we propose a two-stage digitally steerable low-cost beamforming network architecture, where a digital beamformer (DBF) with a reduced number of RF transceivers connected to an increased number of antennas through an RF beamforming network (RFBN). This architectural modification imposes a complex set of performance requirements such as spectral mask, insertion loss, sidelobe suppression, effective radiated power, and so on. Thus, a comprehensive view to efficiently design RF communication systems is required. A. Two-Stage Microwave Beamforming We consider a setup where Ntrx independent transmit signals in the digital baseband are converted to RF using a set of Ntrx

0018-9480 © 2016 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.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 2

RF transceivers. These RF signals are subsequently connected to Nt antennas using an Nt × Ntrx RFBN denoted by W, as shown in Fig. 1(b) (where Nt > Ntrx ). The focus of this paper is to design hybrid RF and digital beamforming networks, where the RF part is predesigned and kept fixed for a given scenario and requirement. Subsequently, a fully adaptive digital part is designed whose weights are modified in order to generate different beams. The hybrid of RF and digital beamforming provides a fully steerable beamforming platform. 1) Existing Microwave Beamforming Networks: Tunable RF beamformer architectures with a reduced number of RF chains have been previously proposed for low-power receivers fully implemented in silicon [2]–[4]. Additionally, two-stage adaptive beamforming for radio astronomy receivers have been proposed [5]. However, these architectures cannot be applied for transmit modes since the power levels (as well as microwave loss) of such designs are much lower than the power levels required in cellular base-stations, which can easily exceed 47 dBm. At high RF power, the technology to implement an RF beamformer is limited to components built in printed circuit boards (PCBs) or suspended strip-line technologies. Some examples of standard passive RF beamformers can be found in [6]–[13]. However, these are nonadaptive designs and, in each case, the given transmitted signal is shaped by the same beam. Absence of a dynamically steerable DBF means that the coverage and SNR improvement is limited. Recently, hybrid beamformers have been designed by combining Butler matrices [6] with digital chains for millimeter-wave applications [14]. However, they do not address the RFBN performance requirements and microwave losses; thus optimal RF performance is not guaranteed. As we transition from fully analog/fully digital approaches to hybrid networks, the hybrid RFBN and DBF must be designed to jointly provide optimal beam pattern while minimizing microwave loss and circuit complexity. Note that state-of-the-art networks [5]–[12] are usually synthesized using empirical approaches, hence, optimal performance of such architectures might not be guaranteed. In addition, standard RF networks such as Butler matrices [6] are not customized for base-station antenna array applications. 2) Objective: Our aim in this paper is optimal design of digitally steerable RFBN-DBF architectures for a wide variety of cellular architectures. While doing so, we aim to answer some fundamental theoretical and practical questions such as the following. • How do we split the RF/analog versus digital functionalities in an adaptive antenna array communication systems? • In other words, what are the minimum number of digital transceivers and RF chains required for optimal performance? • How do we achieve optimal main-lobe and sidelobe level (SLL) performance for a given RFBN-DBF setup? • How do we ensure that the RFBN achieves desired performance with respect to microwave loss, beam-pattern coverage, and recovery in case of transceiver failure?

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

B. Signal-Processing Perspective From a signal-processing perspective, our proposed setup with RFBN and DBF can be placed in the already known category referred to as “beamspace processing” [15]. In beamspace processing, signal processing on large arrays (i.e., Nt = Ntrx ) is performed only on a sub-set of digital elements instead of on the overall array in order to reduce the computational effort. Note that in its traditional sense, beamspace processing is entirely done in the digital domain with full AAA architecture with one-to-one mapping. One way to see the RFBN-DBF arrangement is to perform a part of the processing in analog RF and the rest in digital baseband. While beamspaceprocessing techniques do provide a systematic framework for reduced complexity algorithms, they do not explicitly include microwave and cellular network requirements in design specification and will not lead to the optimal RF system. From state-of-the-art passive RF networks, we know how to design a passive phased array system that connects Ntrx = 1 transceivers with Nt antenna elements [6], [7]. In this case, all phase and amplitude weights are generated in the RF. It is also well known how to generate multiple beams with digital beamforming arrays where Ntrx = Nt and all phase and amplitude weights are generated in the digital domain [1], [15]–[17] (or, in some cases, in the analog domain). The design approaches in both of the above cases are straightforward and well documented. However, it is not trivial to design RFBN used in combination with DBF to generate multiple beams with Ntrx < Nt . For example, it is not obvious to ensure optimal beam-pattern performance for the entire tilt range or cell while minimizing the microwave loss. In most cases, an optimal solution cannot be guaranteed. The signal-processing part of this paper aims to provide a set of rules for identification and placement of RFBN components as well as routing of digital and RF signals in order to ensure that optimal performance is achieved with minimal loss at a reasonable cost. C. Contributions and Outline In this paper, we progressively study various aspects of digitally steerable two-stage hybrid DBF-RFBN design. The RFBN is typically designed for a predefined range of beamtilts, and the DBF is digitally varied to ensure that each beamtilt within the predefined range is achieved. The specific contributions of the paper are as follows. • We propose a framework for two-stage hybrid DBF-RFBN. A hybrid design must explicitly consider RF design challenges while ensuring that optimal digital performance is satisfied. We enumerate some of these constraints from both signal processing and microwave design perspectives. Subsequently, we proceed to specify bounds on the number of transceivers and propose algorithms to optimize the RFBN and DBF. More specifically we recast the practical constraints in a hybrid RFBN-DBF setup as a convex optimization problem, and use an interior point algorithm [18] to find the optimal solution.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. VENKATESWARAN et al.: HYBRID RF AND DBF FOR CELLULAR NETWORKS

•

We represent the RFBN using elementary microwave components such as power dividers, phase shifters in the form of delay lines, and directional couplers (DCs), and analyze the performance of the hybrid DBF-RFBN. Varying the digital weights with a fixed RFBN in the proposed setup will lead to mismatch in RF signals, when the signals are combined at the antenna ports. We quantify the limitations and mismatch/loss typically seen in dynamic hybrid networks and propose architectures to minimize loss without compromising on the overall beam-steering capability. • We propose specific methods to factorize macro-cell and small-cell hybrid beamforming networks trading off RF beam-steering capabilities with overall network losses. Note that a macro-cell network would require a narrow beam with high directivity and minimum loss to enhance coverage across cell edge, whereas a small-cell network would require wide beams serving orthogonal directions. The RFBN design must account for these requirements. 1) Organization: In Section II, we specify the antenna array and signal-processing model in order to formulate the design problem. In Section III, we provide theoretical bounds on the minimum number of transceivers required and propose algorithms to design the optimal RFBN and DBF weights, subject to performance and design constraints. In Sections IV and V, we customize RFBN using microwave components for macro-cell and small-cell networks, respectively. In Section VI, we compare the performance of the proposed architectures and algorithms. In Section VII, we instantiate two RFBNs for macro and small cells operating at a frequency of 2.6 GHz. In Section VIII, we test the designed RFBN setup in anechoic chamber, calibrate the RFBN-DBF setup, and perform detailed beam-pattern measurements for a macro-cell RFBN. Notation: Lower and upper case bold letters denote vectors and matrices. An over-tilde (˜.) denotes RF signals, while time indices (.) and [.], respectively, denote analog and digital signals. Superscripts (.)T , (.) H , (.)† , and ., respectively, denote transpose, Hermitian transpose, pseudo-inverse, and Frobenius norm operations. The matrix I K denotes identity,  denotes point-wise multiplication of vectors, while 0l and 1l , respectively, denote l × 1 vectors of zeros and ones. II. S YSTEM M ODEL AND P ROPOSED A RCHITECTURE

3

Consider our proposed setup with Ntrx transceivers connected to Nt radiating elements through a passive RFBN (for example, Nt = 11 and Ntrx = 5). Details of RFBN implementation will be explained in Sections IV–VII. In this case, an Ntrx ×1 DBF vector ϑ(θd ) = [ϑ1 (θd ), . . . , ϑ Ntrx (θd )]T operates on the data stream s[k], followed by Ntrx digital-RF transformation blocks, Nt × Ntrx RFBN matrix W. The signal is eventually radiated as an Nt × 1 vector, where

x˜ (t) = RF{x[k]},

where x[k] = u(θd )s[k].

(1)

Typically, u(θd ) is designed to produce a main lobe centered at θd .

(2)

y˜ (t) = RF{y[k]} and y[k] = ϑ(θd )s[k].

(3)

We refer to the setup as shown in Fig. 1(b) as the hybrid beamforming network since the RFBN is estimated for a specific architecture at the outset and kept fixed. Subsequently, the DBF ϑ(θd ) is adaptively designed to steer the setup for each θd . The radiated signal x˜ (t) is a function of the antenna array response. Assuming the antenna elements are equally spaced, the array response a(θi ) can be modeled as an Nt × 1 vector, which is a function of angle θi , ⎤ ⎡ 1 ⎢ e j 2π/λ δ cos(θi ) ⎥ ⎥ ⎢ a(θi ) = g(θi ) ⎢ (4) ⎥ .. ⎦ ⎣ . e j 2π/λ δ(Nt −1) cos(θi ) where δ is the spacing between adjacent antennas, λ is the wavelength of the carrier frequency of x(t) ˜ in meters, and g(θi ) is the antenna characteristic for cellular standards [19]. B. Full Dimension AAA With Nt Transceivers The full dimension AAA setup having Ntrx = Nt transceivers and ideal microwave components is used as reference to compare the proposed DBF-RFBN setup. The performance of the full-dimension AAA setup depends on the channel capacity, as well as the adaptive sectorization of the beamformer u(θd ). These requirements as well as the operational constraints are represented using spectral mask parameter θd . The constraints that make up the spectral mask θd will be explained in detail in Section II-C (refer to [C1]–[C6]). In short, the spectral mask includes information regarding the gain and directivity along θd as well as the SLLs. In full-dimension AAA architecture, the objective is to design an adaptive beamformer u(θd ) minimizing the overall mean-squared error (MSE),

A. Data Model Consider an Nt × 1 vector x˜ (t) denoting the RF signal radiated from the antenna array at time t. In the full dimension AAA setup with Nt transceivers, x˜ (t) is obtained using an Nt × 1 DBF vector u(θd ) = [u 1 (θd ), . . . , u Nt (θd )]T operating on a data stream s[k] at time t = kT , followed by Nt “digital to RF” transformation blocks denoted as RF [refer to Fig. 1(a)]. Thus,

x˜ r (t) = W˜y(t)

u0 = arg min θd − A(θ )u(θd )2 u(θd )

where

(5)

⎡

⎤ aT (θi = −π) ⎢ ⎥ .. A(θ ) = ⎣ ⎦ . aT (θ

i

(6)

= π)]

is an Nθ × Nt matrix obtained by stacking the array response vectors and θd is an Nθ × 1 vector. The rows of A(θ ) and length of θd correspond to the spatial resolution. One approach to estimate u(θd ) from (5) is u0 ≈ A(θ )† θd

(7)

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 4

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

using the least squares approach [20]. However, such an approach does not take microwave system design constraints into consideration. The least squares solution would have power amplifiers (PAs) operating in the backoff region resulting in reduced efficiency. Thus, the overall design will be suboptimal from an RF systems perspective. We proceed to design the full-dimensional AAA beamformer weights subject to operational constraints (refer to [C1]–[C6] in Section II-C) using iterative convex optimization techniques [18], [16], [17]. The full-dimensional AAA architecture is not among the main contributions of this paper and serves as reference for performance comparisons. C. Reduced Dimension RFBN Architecture: Problem Formulation As mentioned before, our objective is to reduce the number of transceivers and therefore reduce the cost and power consumed by the antenna array. Thus, we jointly design the optimal RFBN matrix W and DBF vector ϑ(θd ) to satisfy the desired set of spectral mask θd corresponding to all beamtilts, i.e., θd , ∀θd ∈ Rθ = {θ1 , . . . , θ Nθ }. In this case, Nθ denotes the number of sectors. Jointly estimating two parameters (such as W and ϑ(θd )) for requirements (such as θd ) can be solved as a weighted least squares problem [20]. One approach to estimate these parameters is through minimizing the overall MSE, {W, ϑ(θd )} = arg min θd − A(θ )Wϑ(θd )2 W,ϑ(θd )

∀θd ∈ {θ1 , . . . , θ Nθ }.

(8)

In order to minimize the overall cost, while taking practical issues into consideration, the above cost function in (8) must include the following constraints. C1 : The number of transceivers Ntrx is restricted to a minimum. C2 : The SLLs are typically required to be at-least 15 dB below the main lobe. This is to ensure that most of the power is directed towards the desired sector, as well as to limit the interference to neighboring cells/sectors. The 3-dB beamwidth is typically constrained to be less than 5° for a macro-cell and less than 15° for a smallcell setup. Note that term “3-dB beamwidth” in antenna design defines the angular width of the main lobe where power is half. C3 : The RFBN design and spectral mask θd must satisfy constraints [C1] and [C2] over the entire beamtilt range θd ∈ Rθ . C4 : The PAs must operate with maximum possible efficiency. This is ensured by limiting the amplitude

variations of DBF (and consequently PAs) to 0 dB ≤ |ϑk (θd )|2 ≤ 1 dB. Note that linearization and predistortion compensation are not the focus of this paper. C5 : In order to minimize insertion loss and complexity in the RFBN, the number of stages inside the RFBN is limited to three. C6 : The incoming signals at the last stage of the RFBN combiner must be matched in amplitude and phase in order to minimize insertion/microwave loss [21]. The objectives are to: 1) design the RFBN and DBF weights satisfying [C1]–[C6] and 2) translate the designed RFBN weights into a microwave network minimizing microwave loss. We proceed with the RFBN design in the following order. P1a : Given a specific number of antenna elements and array configuration, how many transceivers do we really need? P1b : For an arbitrary Ntrx satisfying [C1]–[C4], how do we obtain optimal beam patterns using RFBN and DBF? P2a : How do we represent the RFBN using microwave components such as power dividers and DCs? What are the necessary conditions for optimal RFBN factorizations? P2b : How does the design vary for macro-cellular and small-cell network? The above two problems [P1] and [P2] form the core of this paper and their solutions are covered in the next three sections. Problem [P2] is subdivided, depending on the requirements of the cellular networks, and a detailed synthesis and analysis of each design is provided in Sections IV–VII. III. A LGORITHMS FOR J OINT O PTIMIZATION OF RFBN AND DBF In this section, we consider problems [P1a] and [P1b], and estimate the RFBN and the DBF weights. A. Bounds on the Number of Transceivers Before we proceed to derive the RFBN and DBF weights, it is important to derive theoretical bounds on the minimum number of transceivers Ntrx for a given [C1]–[C3] and Rθ . Our starting point for optimization is the MSE cost function in (8). Let us assume that we have obtained the optimal beamformer weights u(θd ) for the full dimension AAA where Ntrx = Nt . Note that a detailed design procedure for full dimension AAA is also specified in [16] and [17]. For an ideal u(θd ) ignoring [C1] and [C2] and cost function (5), we can approximate the result as θd ≈ A(θ )u(θd ).

Insert the above expression for θd in the MSE cost function (8). The modified optimization problem is written as (10)–(12), shown at the bottom of this page.

{W, ϑ(θd )} ≈ arg min A(θ )u(θd ) − A(θ )Wϑ(θd )2 W,ϑ(θd )

≈ arg ≈ arg

min

A(θ )2 u(θd ) − Wϑ(θd )2

min

u(θd ) − Wϑ(θd )2

{W,ϑ(θd )} {W,ϑ(θd )}

(9)

∀θd ∈ {Rθ } since θd ≈ A(θ )u(θd )

(10)

∀θd ∈ {Rθ }

(11)

∀θd ∈ {Rθ }

(12)

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. VENKATESWARAN et al.: HYBRID RF AND DBF FOR CELLULAR NETWORKS

The following Lemma characterizes the necessary conditions for optimal RFBN weights in the beamtilt range Rθ . Lemma 1: Consider the scenario [P1a]. Assume that the RFBN is made of ideal lossless components and PA efficiency constraint [C4] is ignored ∀ϑ(θd ). Given Rθ = θ1 , . . . , θ Nθ , the optimal weights of the RFBN must lie in the space spanned by the columns of dominant basis vectors of  = [u(θ1 ), . . . , u(θ Nθ )]: W ∈ col span{}. Proof: Note that W has to provide reasonable performance for all values of θd ∈ Rθ . Stacking the cost function in (12) for the entire beamtilt range θd ∈ Rθ ,  {W, ϑ(θd )} = arg min  u(θ1 ), . . . , u(θ Nθ )  −W ϑ(θ1 ), . . . , ϑ(θ Nθ ) 2 ⇔ arg min  − Wϒ2

(13)

where  and ϒ are, respectively, Nt × Nθ and Ntrx × Nθ matrices with ϒ = [ϑ(θ1 ), . . . , ϑ(θ Nθ )]. Let us assume that Nt ≥ Nθ and compute the singular value decomposition (SVD) of ,  = [U][][V] H

⎡

⎢ = [u1 , . . . , u Ntrx , . . .] ⎣

σ1

⎤ σ2

..

⎥ H ⎦V

(14)

.

where U and V contain the left and right singular vectors and  correspond to their singular values, typically arranged in descending order [20]. For Nt ≥ Nθ and a setup with Ntrx transceivers, choosing W = [u1 , . . . , u Ntrx ]

(15)

would provide the best Ntrx -rank representation of  depending on ϑ. A linear combination of Wϑ(θd ) will provide the best approximation of u(θd ). For this reason, the RFBN design focusing on performance close to full dimension AAA is obtained by choosing Ntrx whenever σ Ntrx +1 tends to 0. The following few remarks are in order. • From the array signal-processing perspective, choosing dominant eigenvectors (usually in the digital domain) is referred to as reduced rank approaches [15], [22]. Such techniques are used to reduce digital post-processing complexity. • Lemma 1 assumes that Nt ≥ Nθ . For Nt < Nθ , choosing Nt extreme values in Rθ and proceeding similarly will give an approximate W. • Lemma 1 approach can be seen as a more systematic approach to estimate the weights of Blass matrix, as in [9], [10]. Note that this bound on the optimal W does not consider microwave losses, linear operating range of PAs, and the number of possible interconnects in the overall network. However, it does provide a starting point for modifications in next sections that consider the above practical issues.

5

B. RFBN Optimization Estimating Ntrx for the range of Rθ via Lemma 1 is the first step in the RFBN design. Once we have established minimum Ntrx , the next step is RFBN and DBF weights satisfying [C1]–[C6] and θd . 1) RFBN Optimization With Constraints [C1]–[C3]: One approach to design DBF-RFBN providing a main lobe at a specific direction while minimizing the overall variance along other directions is referred to as Capon approach [20]. Modifying the Capon approach for our DBF-RFBN setup, we can design the weights of W such that Wϑ(θd ) operating on the antenna array A(θ ) provides a main lobe steered towards the desired sector, while minimizing the overall variance towards other sectors. Mathematically, the above two conditions can be combined and written using RA = A H (θ )A(θ ) as {W, ϑ(θd )} =

min ϑ H(θd)W H RA Wϑ(θd )2

(16)

subject to a H (θd )Wϑ(θd )2 = 1.

(17)

{Wϑ(θd )}

Note that although this cost function is different from (8), the solutions typically converge with each other. The main-lobe constraint can be further enhanced by specifying the 3-dB or half-power beamwidth (θ3, dB ) towards the desired sector in (16), i.e., 1 ϑ H (θd )W H a(θd )2 = 1 ϑ H (θd )W H a(θ3, dB )2 = . 2 (18) Typically, θd − θ3, dB ≤ 5° in a macro-cell setup and θd − θ3, dB ≤ 15° in a small-cell setup. In order to suppress signals over unwanted sectors, we include SLL constraint [C2]. To achieve a specific SLL (say, dB = 20 dB below the main lobe) over a range of angles accounting for sidelobes θ SLL , we explicitly introduce the constraint ϑ H (θd )W H A(θ SLL )2 ≤ 

(19)

where A(θ SLL ) denotes the array response for the sidelobes and  = 10(−dB /10), i.e.,  = 0.01 for dB = 20 dB. 2) Interior Point Optimization: Combining all the above constraints (17)–(19), the central optimization problem becomes {W, ϑ(θd )} = min ϑ H (θd )W H RA Wϑ(θd ) subject to ∀θd ∈ Rθ ϑ H (θd )W H a(θd )2 = 1 from (18) 1 ϑ H (θd )W H a(θ3, dB )2 = from (18) 2 ϑ H (θd )W H A(θ SLL )2 ≤ [, . . . , ]. (20) For the optimized Ntrx from Section III-A, the above cost function explicitly includes the constraints [C1]–[C3]. The above cost function can be recast as a convex optimization problem [18] and solved numerically to obtain the optimal solution. More specifically, the solution is obtained using the interior point algorithm [18]; note that similar techniques

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 6

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

to estimate full-dimension AAA weights u(θd ) have been proposed in [16] and [17]. The following few remarks are in order. • A more comprehensive approach to jointly optimize {W, ϑ(θd )} is done by representing ϑ(θd ) as a function of W, as in [23], ϑ(θd ) ≈ θd [A(θd )W] . †

•

(21)

Note that the interior point optimization is not the main contribution of this paper and for a detailed performance analysis of these approaches refer to [18], [16], and [17].

C. DBF Design for Given RFBN Satisfying [C1]–[C4] Once the optimal RFBN W is designed as in Section III-B for ∀θd ∈ Rθ , the adaptive DBF weights are estimated for each beamtilt θd . Note that the DBF ϑ(θd ) is a function of the RFBN W and array response a(θd ). For a given W, designing DBF weights and minimizing the overall cost in (12) is transformed to ϑ 0 (θd ) = arg min θd − H(θ )ϑ(θd )2 , ϑ(θd )

where H(θ ) = A(θ )W.

(22)

1) DBF Design With PA Constraints: The PAs used for cellular networks should operate with maximum efficiency. Thus the gain and amplitude tapering with DBF should be limited to, say, the 0–1-dB range. Restricting the gain avoids PAs operating in the backoff region, eventually maximizing the PA efficiency. The DBF weights should comply with these output levels as specified in [C4]. We explicitly include these constraints on output power from each transceiver as 1 ∀k ∈ {1, . . . , Ntrx }. (23) Ntrx For details on the optimization of DBF refer to [17]. Some comments are in order regarding the DBF design. • The optimization can be expressed including the per PA power constraint as in [17]. Note that for such algorithms to yield optimal solution, we need to explicitly show that the problem is convex. • Please note that per antenna power constraint is not convex (unlike the inequality and linear constraints as proposed in [17]). As a special case, the expression ϑk (θd ) can be represented using magnitude and phase terms, which, in turn, can be represented as a convex problem. For further details on the applicability of such algorithms refer to [18]. 2) Wideband DBF-RFBN Design: The RFBN and DBF have been designed for a specific frequency. This frequency is usually the center frequency of the respective band, and the antenna array spacing depends on the center frequency. Note that in cellular access, it is common to focus design on a specific frequency, e.g., Band 7 ranges from 2500–2570 MHz (Uplink) to 2620–2690 MHz (downlink), resulting in a relative required bandwidth of (190/2595) = 7.3% [19]. As we transition from a fully analog RFBN, we can improve the operational bandwidth to some extent by utilizing a hybrid RFBN and DBF approach. In this case, we can re-optimize |ϑk (θd )|2 ≈

the DBF weights for the fixed RFBN to operate at a different frequency band as long as the overall bandwidth is less than 15%. IV. RFBN I MPLEMENTATION : M ACRO -C ELL N ETWORK Note that Section III provides some important directions on the design of RFBN, however, it does not represent the RFBN in terms of microwave components and account for practical limitations such as interconnect complexity and loss. It is not possible to directly apply the results of Section III in order to design the microwave RFBN. This section proposes design changes for the macro-cell RFBN and factorizes the RFBN using a combination of microwave components. A. Cellular Network Microwave Beamformer Design Challenges The DBF-RFBN arrangement can be seen as two-stage beamforming towards a specific sector. The first stage, i.e., the DBF, is an adaptive transformation for each beamtilt with a straightforward implementation (say, using a fieldprogrammable gate array (FPGA) as in our experimental setup). The second stage, i.e., RFBN, is made up of microwave components, and its implementation is not trivial, especially when the objectives are to minimize the overall loss and provide distinct beam patterns towards different sectors. Consider RFBNs designed using a combination of commonly used microwave elements such as power dividers [Wilkinson dividers (WDs)], phase shifters (microstrip lines), power combiners, and hybrid DCs [21, Ch. 7]. In existing RFBNs using power combiners to combine distinct signals, mismatch between amplitude and phase of incoming signals will lead to insertion and return loss [21, Ch. 2]. A digitally steerable RFBN with varying weights will always have mismatch in RFBN combiners, eventually leading to significant degradation in performance. To some extent, the loss can be isolated and recirculated using hybrid or four-port DCs. However, a systematic approach to design RFBN in order to minimize loss while satisfying [C1]–[C4] is not known. Additionally, the RFBN has to be designed to account for a specific cellular arrangement. For example, in a macro-cell AAA setup, the range of beamtilt between adjacent sectors is small (Rθ < 20°) and the distance between the mobile user and base station is typically large. For such scenarios, the emphasis is to design a macro-RFBN to achieve a narrow beam while minimizing the overall loss. Alternatively, in a small-cell AAA setup, the beamtilt range is large (30°–60°) and the emphasis is on increasing the angular coverage with orthogonal beams. Thus, depending on the cellular network architecture, the joint design problem {W, ϑ(θd )} can be reclassified as follows. D1 : RFBN designed to form a narrow beam and minimize insertion loss over a relatively small beamtilt range—for macro-cell cases. D2 : RFBN designed to form arbitrary orthogonal beams, say, at {+30°, 0°, − 30°}—for small-cell cases. Note [D1] and [D2] would lead to distinct RFBNs since their requirements vary. We focus the rest of this section for

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. VENKATESWARAN et al.: HYBRID RF AND DBF FOR CELLULAR NETWORKS

7

Fig. 2. Factorization of an Nt × Ntrx RFBN into a bank of power dividers, phase shifters, and DCs connected to Nt antennas These stages are subsequently factorized into banks of three-port or four-port microwave networks. Specific decompositions depend on the cellular architecture (such as macro-cell or small cell) and design objectives (such as insertion loss or beam pattern).

the design of macro-cell W. The subsequent redesign of ϑ(θd ) follows Section III. B. Macro-Cell RFBN Synthesis

Fig. 3. Power-divider and phase-shifter banks of macro-cell RF beamformer: decomposition of power-divider bank into equal-split (green) and unequal-split Wilkinson power dividers (red). For a given Nt × Ntrx arrangement, the ratios of power dividers as well as the phase shifts are obtained by the multi-stage RFBN decomposition algorithm.

We decompose the overall RFBN as shown in Fig. 2 where, for simplicity, components with similar functions are combined, respectively, into a filter bank of power dividers Dfb , a bank of phase shifters Pfb , and a filter bank of DCs Rfb , W = Dfb · Pfb · Rfb = (Dw1 × Dw2 × Dw3 ) · Pfb · (Rc1 × Rc2 ).

(24) (25)

In the above expression the subscripts Dw i denotes stage i filter bank of WDs and Rc i denotes stage i filter bank of hybrid couplers or combiners. A generic Nt × Ntrx RFBN matrix can be represented using a bank of three-port networks containing a maximum of the following. • (Nt − 1) power dividers connected to each transceiver, i.e., Ntrx (Nt − 1) dividers in total. • (Ntrx − 1) combiners connected to each antenna or Nt (Ntrx − 1) combiners in total. • Ntrx Nt phase shifts to achieve the desired beam pattern. Such an arrangement would result in a significant increase in RFBN size and interconnect complexity. From a conventional microwave design perspective and S-parameter representation, we can also denote the overall RFBN matrix W using the Ntrx + Nt port network. Following this methodology, Dfb can be represented using S-parameters as the 2Ntrx (Nt − 1) port network and so on. For sake of simplicity and consistency with Sections II and III, we avoid using the S-parameter representation and stick to an Nt × Ntrx matrix W. 1) Microwave RFBN Factorization of D f b and P f b : The flexibility to steer different beams towards distinct sectors with a DBF-RFBN setup depends on the number of phase shifts and combiners inside the RFBN. This determines the cascade ordering of Dfb , Pfb and Rfb . Since Ntrx Nt , the first stage consists of power dividers in order to increase the number of input signals that are eventually phase shifted and combined for beamforming. In order to minimize the implementation complexity, each PA output is successively factorized into multiple stages of three-port WDs Dfb , as shown in Fig. 3.

Fig. 4. DC bank Rfb for macro-cell RF beamformer factorization: For the 11 × 5 instance, each stage is made up of two four-port rat-race couplers (denoted by the 2 × 2 matrix TR ). These couplers are used to combine RF signals from adjacent ports. The combined signals are subsequently recirculated at the next stage to compensate for insertion losses. For simplicity, only the combiners connecting Antennas 5 and 7 for a 11 × 5 setup is detailed.

The design algorithms implement first two stages of Dfb using equal-split WDs followed by unequal-split or asymmetrical WDs in the third stage. The output signals from Dfb undergo a phase shift denoted by diagonal matrix Pfb . The phase shifts are achieved by varying the lengths of microstrip lines. Note that the phase shifts in Pfb (or P1 in Fig. 2) directly correspond to phase of elements in W. The DBF ϑ(θd ) digitally steers the transmit signal s[k] in (2), followed by Dfb and Pfb to achieve the desired beamtilt and pattern. Subsequently, the output of the phase shifter matrix is combined using Rfb and connected to the antenna array. 2) Low-Complexity Low-Loss RFBN Design: Mismatch in the amplitude and phase of incoming signals at each combiner will lead to insertion and return loss. Claim 1: Consider Scenario [P2], where the RFBN has been factorized into a bank of DCs (Rfb ) As shown in Fig. 4(a), each bank is further divided into several stages

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 8

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

of DCs Rci . Irrespective of the RFBN setup and beamtilt range θd , an Ntrx ×1 vector ϑ(θd ) can be designed to minimize microwave and return loss at Ntrx − 1 combiners of Rfb . Proof: From linear estimation theory, we know that the Ntrx × 1 vector ϑ(θd ) has Ntrx degrees of freedom. From Lemma 1 and (13), one of the Ntrx variable in the DBF can be used to achieve a main lobe pointed towards θd . The amplitude and phase of the remaining (Ntrx − 1) elements in the DBF can be used to align the amplitude and phase at (Ntrx − 1) combiners minimizing insertion loss. Although Claim 1 specifies that one of the Ntrx degrees is sufficient, the beam-pattern performance will be poor if we use only one parameter among Ntrx parameters in ϑ(θd ) to optimize for the beam pattern and the rest, i.e., (Ntrx −1) parameters in ϑ(θd ) are used to match signals at each combiner. Claim 1 acts as a starting point and provides a lower bound on the possible number of combiners in order to minimize loss in Ntrx − 1 ports. However, it is beneficial to use all the available Ntrx degrees of freedom available in ϑ(θd ) to optimize for beam pattern. 3) Multi-Stage RFBN Decomposition: The RFBN decomposition satisfying Claim 1 has at most Ntrx − 1 combiners inside Rfb and can be modeled using an Nt × Ntrx RFBN interconnect matrix S. Note that we fix Nt = 11 and Ntrx = 5, in order to satisfy [C1]–[C3]. One such example of the Nt × Ntrx spatial interconnect map S for a 11 × 5 case is ⎤ ⎡ 13 02 04 06 04 (26) S = ⎣ 04 13 13 13 04 ⎦ . 04 06 04 02 13 

11×5 matrix

Our focus is now to redesign the RFBN weights satisfying S as well as Lemma 1. One approach to redesign RFBN is through the use of a modified version of orthogonal matching pursuit (OMP) [23] or through successive orthogonal projection of optimal W [22]. We skip the exact algorithm details and briefly explain applying the algorithm for our RFBN setup as follows. For a given S as well as SVD of  := [U][][V H ], • from Lemma 1: [u1 , . . . u Ntrx , . . .] = Basis{}, • for k ∈ {1, . . . , Ntrx }, 1) from left singular vectors of  : U = [u1 , U N ], 2) extract RFBN weights satisfying the spatial interconnects: wk = sk  u1 , 3) normalize each column of wk , 4) W = [w1 , . . . wk ], 5) compute the orthogonal projection: U = (I − WW H ), 6) update  := U, • end k, • final RFBN: W = [w1 , . . . w Ntrx ]. We perform RFBN decomposition based on multi-stage Wiener decomposition [22] since it provides a systematic low-complexity implementation of optimal W and converges to optimal solution for increasing Ntrx . While the spatial interconnect map varies as we modify the constraints and claims, the methodology is generic. Note that we have skipped detailed

synthesis and performance analysis of such approaches; for details refer to [22] and [23]. C. R f b Design to Minimize Insertion Loss The outputs from Pfb are combined by Rfb and fed to Nt antennas. For efficient and lossless operation of Rfb , it is necessary that the input signal at each power combiner be matched in terms of amplitude and phase and any mismatch will result in insertion loss. Note that the RFBN is fixed, but the phase and amplitude of ϑ(θd ) is varied for each beamtilt. Such a dynamic arrangement when used with the standard three-port DCs will always result in insertion loss. For this reason, we use hybrid elements, such as rat-race couplers or branch hybrids (four-port networks) [21, Ch. 7] instead of a standard three-port power combiner. Typically in a hybrid coupler, ports 2 and 3 are input ports and the inputs are coupled to a standard output (port 1). A fourth port (also referred to as the reflection port or isolation port) extracts the signals that would have otherwise led to loss whenever there is a mismatch between the input signals [21, p. 480]. Thus, at a given stage of Rc i , any phase or amplitude mismatch can be captured at the output of the isolation port of the hybrid coupler. We exploit the following signal-processing properties of such hybrid couplers to reroute the insertion loss seen in stage i to stage i + 1 and minimize the overall loss. 1) Claim 2: Consider scenario [P2], where the RFBN is subdivided into four-port DCs. 1) The elements of the impulse response of Wϑ(θd ) providing the main lobe along θd matches that of full dimension AAA. 2) The insertion loss in each stage of Rc i is minimized if the number of hybrid couplers does not exceed Ntrx − 1. Proof: For simplicity, consider the full dimension AAA setup with Nt transceivers and array response A(θ ) as specified in (5) with Ntrx × 1 DBF u(θd ) = [u 0 (θd ), . . . , u Nt−1 (θd )]T . Note that a(θi ) has linear phase progression since the antenna elements are uniformly spaced. The transfer function of u(θd ) operating on the antenna array can be written as U (θi ) =

N t −1

u k (θd )e− j 2π/λk cos(θi )

∀θi ∈ [−π, π]. (27)

k=0

The full dimension beamformer u(θd ) can be seen as a spatial fir filter operating on the antenna array and (27) can be seen as the spatial equivalent of a filter transfer function or filter spectral response. From filter design theory, we know that matched filters operating on uniformly sampled antennas will have a symmetric magnitude response along the central antenna element Nt /2. They will also have a linear phase progression, i.e., |u k (θd )| = |u Nt −k (θd )| and 

u k (θd ) −  u k−1 (θd ) =  u k+1 (θd ) −  u k (θd ).

(28)

We extend this linear phase and symmetric magnitude argument for RFBN-DBF setup Wϑ(θd ) operating on A(θ ) with

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. VENKATESWARAN et al.: HYBRID RF AND DBF FOR CELLULAR NETWORKS

9

a transfer function similar to U (θi ) in (27). This proves the first part of Claim 2. Exploiting the shift invariance property, the combined DBF-RFBN Wϑ(θd ) operating on antenna array A(θ ) will also have a linear phase progression. Due to linear phase progression, the signal from the isolation port of the rat-race coupler k is in phase with respect to the signal at the output port of the rat-race coupler Ntrx − k∀, k ∈ {1, . . . , Ntrx − 1}. This property means that the incoming signals connecting antennas 5 and 7 in Fig. 4 (denoted by blue and red arrows) will be matched in phase with the corresponding input signals (denoted by black arrows). Thus, signals from stage i combined at the subsequent stage Rc,i+1 will minimize overall loss. This proves the second part of Claim 2. Remark: The combined matched filter Wϑ(θd ) will maximize the SNR for θd , provide optimal beam pattern, and minimize insertion loss. The hybrid elements can either be branch hybrids (commonly used in Butler matrix implementations [6]) or rat-race hybrids. In our setup, we use rat-race hybrid elements since they can also be seen as radix-2 discrete Fourier transform (DFT) implementations. Higher order DFTs can be obtained using multiple stages of such couplers.

a time. Improving from the the Blass matrix design of [9], Djerafi et al. [10] propose an RF beamforming matrix that provides a set of orthogonal beams with multiple PAs operating at the same time. Thus the existing family of RFBNs to generate distinct/orthogonal beams can be classified into the following. • Lossless Butler matrix implementations [6], where there is only one PA operating at a time, sacrificing the overall radiated power. • Lossy Blass matrix [9], [10], [12] with Ntrx transceivers operating simultaneously, sacrificing insertion loss performance for effective radiated power. The fundamental question is whether we can generate orthogonal beams satisfying [C1]–[C6], while keeping insertion loss to acceptable levels.

V. RFBN I MPLEMENTATION : S MALL -C ELL N ETWORK In a small-cell setup, the number of RF chains are limited to 2 or 3 and the number of antennas are limited to 4–6 (typically as a horizontal arrangement). The objective is to provide coverage along Nθ = 3 fixed beams spaced 30° apart from each other (say, θd ∈ {−30°, 0°, +30°}). Each beam has a wider 3-dB beamwidth (nearly 15°) and the focus is more on improving angular coverage (unlike the macro-cell case where the focus is on minimizing loss). The requirements and overall setup make this design fundamentally different from that of Section IV. A. Connections With State of the Art Existing passive beamformers such as [6] provide distinct (and orthogonal) beams while minimizing microwave loss. The setup [6] can be modeled having Nt inputs and Nt outputs, connecting Nt transceivers and Nt antennas, and implemented using either branch or rat-race hybrid couplers. Typical implementations have Nt = 4 antennas spaced half a wavelength (λ/2) apart. Note that in the majority of such lossless implementations only one transceiver is turned on at a given time to generate the desired beam patterns. Such a design sacrifices radiated power/directivity and flexibility DBF for sake of lossless implementation. In order to incorporate the benefits of the DBF, a systematic approach is to design the RFBN for Rθ = {θ1 , θ2 , θ3 } from the basis vectors of  using a QR decomposition [20] as specified by Lemma 1. The QR decomposition approach is similar to the Blass matrix design of [9]. In [9], the authors propose an RF matrix whose columns correspond to orthonormal vectors of . However they do not account for operational constraints [C1]–[C3]. Subsequently, in order to minimize the insertion loss, they transform the Blass matrix into a modified Butler matrix, with only one PA operating at

B. Generalized Butler Matrix If we implement the RFBN based on [9] and [10], Rfb will have more than Ntrx −1 combiners and Claims 1–2 will not be satisfied. Additionally, the DBF-RFBN arrangement will not always have a linear phase progression due to limited degrees of freedom. Thus, [D1] is not applicable. Given an Nt × Ntrx RFBN, the signal at the antenna element i, i ∈ {1, . . . , Nt } can be routed through Ntrx combiners. These correspond to the number of non-zero elements of the row W(i, :) and we denote this quantity by row weight. One approach to reduce insertion loss is to match the combining signals in amplitude and phase. In cases where this matching is not possible, it is preferable reduce row weight without modifying the overall DBF-RFBN response. From the RFBN designed in Section III-B, we search for the unmatched/ out-of-phase combiners. • Selectively remove the connections corresponding to these unmatched combiners. This operation can be done by zeroing specific entries in the RFBN matrix that exceed a specific mismatch threshold. To zero a specific entry in the RFBN matrix, we use the Givens rotation method [20]. • Subsequently, the updated RFBN is used to re-estimate DBF ϑ(θd ) as in Section III-C such that the input signals at the remaining combiners are matched and constraints [C1]–[C6] are satisfied. As an example, we illustrate the Givens rotation based RFBN design for ϑ(θ1 ) = [1, 1, 1]T , say, at beamtilt θ1 and Nt = 4. The combined DBF–RFBN response can be represented using arbitrary complex variables ∗ as ⎡   ⎤ ⎡ ⎤ ∗∗∗ 1 ⎥ ⎢ ∗∗∗ ⎥⎣1⎦. (29) Wϑ(θ1 ) = ⎢ ⎣∗ − − ∗ ⎦ → ∗ ← 1 ∗∗∗

 

ϑ

W

(θd )

Consider the above expression, where w3,2 (element at row 3 and column 2) is out of phase with the two other signals w3,1 and w3,3 . Givens rotation allows us to remove a specific element (in this case, w3,2 ) using the Nt × Nt matrix

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 10

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

for G(i, j, φ),

⎡

1 0 0 ⎢ 0 cos(φ) sin(φ) G(2, 3, φ) = ⎢ ⎣ 0 − sin(φ) − cos(φ) 0 0 0

⎤ 0 0⎥ ⎥ φ ∈ {−π, π}. (30) 0⎦ 1

A generalization of the above expression G(i, j, φ) is recursively applied to zero elements along the i th row and the j th column, whenever the amplitude and phase mismatch exceeds a particular threshold without modifying the overall transfer function considerably. Thus, ⎡ ⎤ ∗ ∗ ∗ ⎢∗ ∗ ∗⎥ ⎥ W :⇒ G(3, 2, φ1 )W = ⎢ (31) ⎣ ∗ 0 ∗ ⎦ :⇒ ∗ ∗ ∗ 

recursion 1 ⎡ ⎤ ∗ ∗ 0 ⎢∗ ∗ ∗⎥ ⎥ (32) :⇒ G(3, 1, φ2 )W = ⎢ ⎣∗ 0 ∗⎦. ∗ ∗ ∗ 

recursion 2

In practice, we perform successive Givens rotation on the combined response Wϒ, where ϒ = [ϑ(θ1 ), ϑ(θ2 ), ϑ(θ3 )]. The following comments are in order. • Note that after every rotation, the DBF ϑ(θd ) as well as the combined response Wϑ(θd ) has to be re-estimated to account for SLL requirements, PA limitations, etc. • The number of Givens rotation iterations is a tradeoff between the quality of beam patterns and microwave loss. • The Bulter matrix can be interpreted as a special case of Givens rotation, where ϒ is an 4 × 4 identity when Nt = Ntrx = 4. VI. S IMULATION R ESULTS To assess the performance of the joint RFBN-DBF architectures, we have applied it to macro and small-cell multiantenna base stations. We present simulation results for the RFBN setup based on multi-stage RFBN decomposition as outlined in Section IV-B. These results include computing the beam patterns for varying number of transceivers, varying beamtilts, and microwave loss performance. These results are complimented by RFBN implementation and measurements are detailed in Sections VII and VIII, respectively. The performance indicators are usually the following: • radiated energy/directivity along desired beamtilt θd and SLLs; • insertion loss and beamtilt range limitations with RFBNs. The hybrid DBF-RFBN arrangement should be able to beamform and transmit the desired signal towards distinct sectors θd ∈ {0°, . . . , 20°}. We consider a setup with Nt = 10 − 12 antennas radiating at 2.6 GHz. Note that Nt ≥ 10 is chosen to get desired array gain of 18 dB while ensuring that SLLs are 18–20 dB below the main lobe, as specified in (20). Antenna elements are uniformly spaced at a distance 0.8λ. Note that Nyquist sampling criterion

Fig. 5. Hybrid RF and DBF performance with Nt = 11 antennas for multiple beamtilts θd . (a) Comparison of RFBN-DBF performance with optimal DBF performance when used with a full dimension AAA setup containing Nt = Ntrx = 11 transceivers and beamtilt θd = {0°, 10°}. (b) Performance of hybrid RFBN-DBF with non-adaptive RF beamformer plus varying DBF for θd = {0°, 10°}.

would typically limit array spacing to 0.5λ at the given frequency of operation, however, cellular antennas are typically designed for wideband operation and the spacing constraint is relaxed. The increased spacing leads to grating lobes for some beamtilts. Lemma 1 says that we need at least Ntrx = 3 transceivers for a macro-cell scenario with beamtilt range to achieve 18–20-dB SLL for the entire beamtilt range θd ∈ {0°, . . . , 20°}. In practice, we would need 4–5 transceivers to minimize insertion loss and to limit the dynamic range of the PAs. A. Vertical Sectorization Performance Fig. 5(a) shows the beam-pattern performance for an Ntrx = 4 transceiver and Nt = 11 antenna setup. Curve 1 is the reference curve corresponding to the optimal beamformer with Ntrx = 11 transceivers with Ntrx = 11 PAs while ignoring PA efficiency. Note that this is impractical and is shown as the theoretical bound for optimal beamformer design. Curve 2 shows the performance with an 11 × 4 RFBN and an 4 × 1 DBF. The partially adaptive setup provides optimal main-lobe performance while satisfying the 3GPP requirements and converges to curve 1. Both approaches provide 24-dB SLL and θ3,dB = 5° for a radiated power of 10 dB along θd = 5°, satisfying [C1]–[C6] as well as θd .

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. VENKATESWARAN et al.: HYBRID RF AND DBF FOR CELLULAR NETWORKS

11

Fig. 6. Optimal hybrid RFBN-DBF performance for different transceiver configurations for fixed beamtilt θd = 12 and Ntrx = 2, 3, 4, and 5. Notice that the side-lobe levels degrade as Ntrx reduces.

Fig. 5(b) shows the simulation performance for two sectors θd = 0° and θd = 10°. These results are obtained using an interior point algorithm optimizing (11) with  18 dB below 1 and shows that the constraints used in optimization problem are coherent with simulations. For clarity, we have shown only two beamtilts, but the setup can account for the entire range θd ∈ {0°, . . . , 20°}. Comparing curves 1 and 2, we observe that the digitally steerable RFBN-DBF achieves 18.5-dB SLL performance. Note that the directivity along θd is still preserved. Fig. 6 shows the beam patterns for varying RFBN arrangements with Ntrx = {2, . . . , 5} and Nt = 11. In this case, Ntrx is fixed in each case and subsequently the optimal RFBN is designed. The RFBN is initially optimized for the sectors Rθ ∈ {0°, . . . , 10°}. Curves 1–4 show a snapshot of beampattern performance, when we are required to provide a main lobe across θd = 12° outside Rθ . Curves 1 (Ntrx = 4) and 2 (Ntrx = 5) show reasonably good performance with 16- and 18-dB SLL, respectively, while achieving θ3,dB = 5°. As expected, the performance significantly degrades for an RFBN arrangement with Ntrx = 2. To account for design flexibility and address different θd , it is necessary to keep Ntrx ≥ 3. VII. RFBN D ESIGN E XAMPLES This section details the microwave design of macro and small-cell RFBN architecture. The DBF implementation is relatively straightforward and explained in Section VIII. We use Keysight’s Advanced Design System (ADS) version 20111 to implement the RFBN. A. Macro-Cell RF Beamformer Design With Ntrx = 5 and Nt = 11 Fig. 7(a) shows an 11 × 5 RFBN for a macro basestation application. The implementation shows Ntrx = 5 voltage sources corresponding to the PA outputs and Nt = 11 S-parameter ports functioning as the input impedance of the antenna elements. This setup allows us to vary the amplitude 1 [Online]. Available: www.keysight.com

Fig. 7. Macro-cell 11 × 5 RF beamformer instantiation. (a) RF beamformer factorized into two stages of WDs, Dw1 and Dw2 , followed by Pfb (P1) and one stage of power combiners Rc1 . The length of the microstrip line connecting input transceiver port 1 with antenna element (AE) 1. (b) Implementation of asymmetric WD—the ratio of impedance Z A and Z B corresponds to the amplitude of corresponding elements, connecting transceiver 1 with antenna ports 1 and 2.

and phase of input signals with DBF, leading to different beamtilts for a given RFBN. We realize the RFBN using microstrip lines on a typical dielectric substrate (dielectric constant = 3.48, loss tangent = 0.004), while considering the isolation resistance loss as well the microstrip loss. Note that in a real implementation, suspended air-striplines would be used to minimize substrate loss. The PCB solution was chosen for manufacturing simplicity. The RFBN implementation, more specifically, the weights and connections between transceiver chains and antenna elements, the power divider ratios, as well as the phase shifts follow the algorithms and architectures in Sections III and IV satisfying [C1]–[C6]. We factorize the RFBN connections based on Claim 1 and Claim 2 into multiple stages of WDs (Dw1 and Dw2 ), phase-shift matrix (Pfb ),

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 12

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

Fig. 8. Macro-cell 11 × 5 RF beamformer performance. (a) Amplitude tapering and output voltage levels at Nt = 11 antenna ports. (b) Phase tapering as we progress from antenna element 1 to antenna element 11. Note that the phase progression is linear to avoid spatial aliasing and near optimal beamforming.

and DCs (Rc1 ). Note that, for sake of simplicity, the current implementation contains only one stage of direction coupler Rc1 made of Wilkinson combiners. The five input signals are split successively into Ns = 15 signal paths (using some asymmetrical power splitters). The ratio of impedance Z A and Z B along the two branches in Fig. 7(b) have a one-to-one relationship with the amplitude of elements in W. Similarly, the length of the microstrip lines dictating the phase shift corresponds to the phase of the elements of W. For example, φ1 connecting antenna 1 with transceiver 1 is given by the phase of w1,1 ∈ W. To prevent microwave at Rfb , we include additional line lengths for the corresponding strip lines. At this point, for sake of simplicity we depart from the design methodology proposed in Section IV and avoid using hybrid couplers. The 15 signals at the output of Pfb are combined and coupled to 11 antennas using a bank of Wilkinson combiners. 1) Beam-Pattern Performance: The RFBN has been initially designed for Rθ ∈ {0°−15°}. As we aim to increase the range of beamtilts θd for a given RFBN setup, we pay with poor beam-pattern performance and increased insertion loss. For a sanity check for the RFBN design, we excite Ntrx = 5 input signals with voltages of the same magnitudes (with varying phase shifts). The beam-pattern quality for steerable beamtilts depends on the phase progression at RFBN output. From Section IV and Claim 2, we know of the advantages in designing a linear phase RFBN. Fig. 8(a) shows the voltage or amplitude levels at each antenna port for θd = 8° and Fig. 8(b) shows the phase progression of signals at each antenna element for θd = 8°. We observe from Fig. 8(b) that linear phase progression is preserved at the RFBN output. Fig. 9(a) shows beam-pattern performance with main lobe at θd = 0° satisfying [C1]–[C6] when no additional phase progression is applied using the DBF. If we apply a phase progression of −50° from the first input port to second input port, we observe a main lobe at θd = −13.80°. Note that the SLLs of the main lobe (marker m2) comply with 3GPP specs. Similarly, a phase progression of +100° results in a main lobe at θd ≈ 0° in Fig. 9(b). In practice, once we include the

Fig. 9. Macro-cell 11 × 5 RF beamformer beam pattern: beam-pattern performance of a Nt = 11 and Ntrx = 5 design provided with the same amplitude, but with varying phase values at each RFBN output to achieve varying beamtilts. (a) Phase progression 0°. (b) Phase progression −50°, +50°, and +100°.

DBF algorithms, the overall performance improves, as shown in Section VI. 2) Microwave Loss Performance: Fig. 10(a) shows the simulated phase mismatch of signals combined at the antenna for

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. VENKATESWARAN et al.: HYBRID RF AND DBF FOR CELLULAR NETWORKS

Fig. 10. Insertion-loss performance comparison. (a) Simulation performance of average phase mismatch between input signals at the last stage of Rfb . Note that the overall insertion loss is directly proportional to the average phase mismatch at the last stage of R f b . (b) Overall loss in radiated power due to insertion loss in the RFBN with one stage Rfb .

an 11 × 5 RFBN and varying beamtilts θd ∈ 0°, . . . , 30°. The insertion loss is proportional to the phase mismatch of signals combined at the antennas. Curve 1 shows the microwave loss for the RFBN implemented using a lossy Blass matrix [9]. Curves 2 and 3, respectively, show the performance of the proposed RFBN from Section V. From curve 3, we observe that, as the beamtilt range increases, it becomes important to use hybrid couplers to compensate for the microwave loss. The impact of the measured phase mismatch and microwave loss is shown in Fig. 10(b). This performance was obtained by subtracting the radiated power from combined input power for varying phase progressions at the input ports. The result shows loss in directivity of the antenna array. This becomes more obvious for increasingly negative or positive phase increments from θd = 8°. B. Small-Cell RFBN Design With Ntr x = 3 and Nt = 6 Fig. 11 shows an Ntrx = 3 and Nt = 6 RFBN design for small-cell setup. The RFBN has been implemented using standard microstrip technology and the connections between RF chains and antenna elements follows Section V. In the Dfb and Rfb stages, the ratios of all the employed power combiners and splitters vary from 0 to 12 dB. These components have been implemented either as unbalanced WDs. The phase shifts have been implemented using standard microstrip-based transmission lines, where the length of the line dictates the phase shift.

13

Fig. 11. Small-cell hybrid RFBN-DBF Nt = 6, Ntrx = 3 to provide distinct beams at −30°, 0°, + 30°.

The small-cell RFBN is more complicated than the macroRFBN design due to orthogonal beam-pattern requirements. This example is composed of five discrete stages (two stages of Dfb , two stages of Rfb , and one stage of Pfb ). The RFBN inputs are generated by Ntrx = 3 transceivers x = (x 1 , x 2 , x 3 ). The 1st stage of the RFBN is composed of three three-way Wilkinson power dividers, which split the signal of each transceiver output into three components. The 2nd stage of the RFBN is composed of nine two-way power dividers, leading to 18 ports after stage 2. The 3rd stage of this RFBN Pfb is composed of 18 static phase-shifting elements that match the phase of the 2nd stage of the RFBN. In the 4th stage of the RFBN, six two-way power combiners are used to combine the amplitude and phase-shift signal from the transceivers x 2 and x 3 . The incoming signals at each combiner is phase matched and subsequently combined at the final stage of the combiners. Finally, the last, i.e., 5th stage of this RFBN consists of 2-to-1 power combiners coupling signals to Ntrx = 6 antennas. 1) Beam-Pattern Performance: A small-cell base station with an RFBN antenna array typically beamforms and transmits the desired signal towards specific sector θd ∈ {−30°, . . . , +30°}. In this case, the PAs typically generate 0.5 W of power. The base-station antennas are spaced 0.5λ apart, and each antenna element has a 3-dB beamwidth of 110°. The focus is to provide orthogonal beam patterns, while sacrificing on the microwave loss performance.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 14

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

Fig. 12. Small-cell joint RFBN-DBF to generate orthogonal beam patterns: Nt = 6, Ntrx = 3 to provide distinct main lobes at −30°, 0°, + 30° with 13-dB sidelobe suppression.

Fig. 13. Overall setup with RFBN implementation on Rogers 4350 substrate comprising of WDs, DCs, and microstrip lines for phase shifts connected to RF and digital transceiver boards.

2) Effect of Number of Transceivers Ntrx and Beamtilt Range θd : Fig. 12 shows the beam patterns for Ntrx = 3, Nt = 6, providing three sectors spaced θd ∈ {−30°, 0°, +30°}. Note that the array response a(θd ), θd = {−30°, 0°30°} for each θd is orthogonal to the other. We observe that it is possible to achieve 13-dB SLL suppression where all the PAs operating at a constant power. Curves 1, 3, and 5 in Fig. 12 compare the measured beam pattern of the ADS implementation with the simulations results and we notice that the RFBN arrangement provides with 13-dB SLL. VIII. A NECHOIC C HAMBER RFBN M EASUREMENTS To assess the capabilities of a macro-cell based RFBN, we implement an Ntrx = 5, Nt = 11 RFBN detailed in Section VII for 2.6 GHz and measure its beam-pattern performance in an anechoic chamber.

Fig. 14. Anechoic chamber measurement setup. During the calibration phase, the signals at each RF chain are compared with reference signal (TX 6) for amplitude and phase offsets, as well as phase drifts over a period of time. Subsequently, the far-field setup is used to receive signals from device-undertest and compute beam patterns.

A. RFBN and DBF Implementation The RFBN as shown in Fig. 13 is implemented with a Roger 4350 substrate material. All power dividers, phase shifters, line crossings, and combiners are optimized using using the High Frequency Structure Simulator (HFSS) version 2011 software from Ansoft,2 a commercial electromagnetic mode solver used to design all network components in order to meet the optimal performance of the individual components. In order to maximize the tilt range, a pre-tilt of 8° was implemented, which is the norm in a majority of cellular basestation implementations. In our system, a Xilinx Kintex-7 FPGA was used as the central signal-processing unit, and multiple complex digital phase shifters were implemented in the FPGA for tuning the amplitude and phase of signals at each transceiver independently. The data signals clocked at 122.88 Ms/s are split into Ntrx = 5 streams and each stream is modified in amplitude and phase according to the weights of ϑ(θd ) (as shown in Fig. 15). Note that the DBF weights for a given RFBN and θd satisfying [C1]–[C4] is generated using the interior point 2 [Online]. Available: http://www.ansys.com/Products/

algorithm specified in Section III-C. The required beamtilt values are in turn provided by higher layer algorithms present in the base station or the core network in order to achieve desired data rates towards the user. B. Measurement Setup and Calibration In order to test the performance of the joint DBF-RFBN, we use multiple synchronized RF transmitters to generate beams from the device-under-test. The testing platform includes a hosting PC with a graphic user interface control panel, an RF transceiver board with multiple transmitters, an RFBN, and an antenna array, as shown in Fig. 14. 1) DBF and RF Chain Calibration: Ideally, the transmission path of each RF transmitter chain would be identical, in other words, the phase and amplitude of multiple transmitter outputs would be exactly the same if the source signal is the same. However, it is very difficult to layout and route multiple transmit chains on the PCB with equal length and the same frequency response. It is also very difficult to guarantee that the connection cables are of equal length and phase aligned.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. VENKATESWARAN et al.: HYBRID RF AND DBF FOR CELLULAR NETWORKS

15

Fig. 15. Calibration of AAA system. The signals at each RF chain is compared with reference signal (TRX 6) for amplitude and phase offsets, as well as phase drifts over a period of time. Once the calibration phase is over, the transmitted signals are modified by the complex weights in ϑ(θd ) steered towards desired user.

Fig. 16. Average amplitude and phase error as we progress through the entire beamscanning range of the network analyzer.

Additionally, the digital to RF transformation blocks, denoted as RF{.}, are not identical and the real physical system contains the following imperfections: 1) the phase lags of the signals from digital baseband RFBN and antenna arrays are different and 2) the amplitude of the signals at the inputs of antenna arrays are not identical. In order to form desired beams with the RFBN, phase and amplitude calibration are fundamentally required. A continuous wave (CW) signal is generated in the FPGA and upconverted to 2.6 GHz. We use a spare transmitter (TX 6) as reference and measure the relative phase difference and amplitude difference between the individual active transmitters TX 1 to TX 5 with reference TX 6 and a vector network analyzer (VNA). In our measurements, a common local oscillator (LO) is used to ensure phase lock between phase-locked loops (PLLs) in the RRH and the VNA. Subsequently we add the corresponding phase and amplitude differences in the digital domain as offsets. After calibration, we sweep the phase from −180◦ to 180° and plot the difference between the input and output amplitude and phase values. Fig. 16 plots the average phase and amplitude error as we sweep the DBF for all possible angles (−π, π). The precision of the calibration setup is a function of this phase and amplitude error. We observe, that the average phase error is less than ±1° and average amplitude error is less than 0.03 dB. In other words, active transmitters are aligned in phase and amplitude after calibration. C. RFBN Performance Evaluation With Antenna Array In a typical macro-cell scenario, the base station is required to at least have a gain of 18 dBi along the main lobe, a 3-dB beam width of nearly 6°, and an SLL suppression in the order of at least −17 dB. In order to achieve vertical sectorization, the RFBN must account for the above set of gain and SLL requirements for the entire beamtilt range Rθ ∈ {0°, . . . , 15°}.

Fig. 17.

Device-under-test in anechoic chamber.

The device-under-test, as shown in Fig. 17, is placed in the anechoic chamber and its beam-pattern performance is measured. A common reference signal (with frequency of 15.36 MHz) is used to synchronize the phase of FPGA output signals with that of the signals received at the network analyzer. Additionally, the beamtilts can be increased or decreased by applying the according amplitude and phase weights in the DBF resulting in a total tilt range from 0° to 16°, while meeting the spatial mask and beamforming requirements required in a macro-cell base stations. Note that for each beamtilt, the power levels input to the PAs must be limited to a range of 0–1 dB. Fig. 18(a) shows the measured beam pattern of the RFBN setup. The DBF gain/amplitude values are limited to

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 16

Fig. 18. Macro-cell joint RFBN-DBF beam-pattern performance for Nt = 11 and Ntrx = 5 at 2.6 GHz and 0.8λ spacing. (a) Optimal setup for Ntrx = 11 transceivers and ideal PAs with infinite dynamic range for beamtilt 8° compared with 11 × 5 RFBN-DBF arrangement. (b) RFBN-DBF setup with beam-tilt θd = 1°.

be within a range of 0–1 dB. Curve 1 shows the theoretical performance bounds when Ntrx = 11 transceivers are used without satisfying constraints [C3]–[C6]. Note that this is unrealistic in practical systems. Curve 2 shows anechoic chamber measurements for beamtilt θd = 8° and we observe that the main lobe is perfectly aligned and that the SLLs are around 18 dB below the main lobe. As the beamtilts are varied to an extreme case for a beamtilt of θd = 1° [see Fig. 18(b)], the main-lobe has a gain of 17 dBi, and the SLL is nearly 14 dB below the main lobe. Fig. 19(a) shows the performance for θd = 4° and shows that the SLL is 16 dB below the main lobe. As we push the DBF-RFBN setup to θd = 14°, the performance starts to degrade as shown in Fig. 19(b), with the SLL 10-dB below the main lobe and grating lobes 8 dB below the main lobe. These are fundamental limitations due to spatial aliasing within the RFBN setup. One important property of the joint DBF–RFBN setup is its robustness. For example, if one element fails, the overall array can provide reasonably good performance with only Ntrx = 4 transceivers operating in the degraded mode. Fig. 20 shows the measurement performance when TX 5 has failed, and only Ntrx = 4 are operational and connected to

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

Fig. 19. Macro-cell joint RFBN-DBF beam-pattern performance for Nt = 11 and Ntrx = 5 at 2.6 GHz and 0.8λ spacing. (a) RFBN-DBF performance for beam-tilt θd = 4°. (b) Beam-pattern performance measurements of macrocell RFBN setup with Nt = 11 and Ntrx = 5 at 2.6 GHz, 0.8λ spacing, and beamtilt = 14°.

Fig. 20. RFBN–DBF for beamtilt θd = 8° where the DBF is dynamically optimized if one of the transceiver fails.

Nt = 9 antennas. Note that in a passive RFBN setup, with only one TRX the system would become instantaneously inoperable, leading to a complete loss of service in the sector. However, re-optimizing the DBF for a degraded mode leads to a reasonably good performance while significantly relaxing maintenance response time and cost. From curve 2, we observe that there is slight degradation in main-lobe energy and increase in the grating lobes due to spatial aliasing.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. VENKATESWARAN et al.: HYBRID RF AND DBF FOR CELLULAR NETWORKS

17

IX. C ONCLUSION Existing RFBNs are constructed in a mainly empirical way, which, depending on the experience of the designer and the complexity of the required network often gives reasonable results. However, this process is very time consuming and does not guarantee the optimal solution in terms of beamforming performance, network complexity, and minimize microwave loss. A thorough understanding of the theoretical bounds, as well as the microwave implementation limitations will lead to the optimal solution enhancing the capacity and coverage of the communications system while operating at a reduced cost and improved reliability. For this reason, we have adopted a holistic approach and proposed RFBN designs to reduce the number of transceivers while accounting for the desirable features in next-generation cellular base stations. Effectively, we have showed how to determine the minimum number of transceiver elements in order to achieve a given set of access requirements and presented a unified view on designing a hybrid beamforming network. Note that the two RFBN design requirements and designs differ a lot in terms of their system requirements. • RFBN for a small-cell base station, generating three static beams with a rather broad beam width in the horizontal direction and tilt range of (−30°, 0°, + 30°). In this design, the focus was on achieving a set of orthogonal beams while maintaining a sidelobe suppression of at least 10 dB. • RFBN for a macro-cell base station antenna with a sharp and narrow vertical beam providing a continuous beamtilt range anywhere between θd ∈ {0°, . . . , 15°}, while maintaining a tight set of spectral mask, SLL, and insertion loss requirements. We have verified its performance in the anechoic chamber. Despite these two very different sets of requirements in both applications, the two derived RFBNs prove benefits in designing through a signal theoretic approach. Especially in the case of a macro-cell base-station antenna, we could show that the overall loss is kept to a minimum, which is essential for applications where the amount of radiated power easily reaches 100 W and more. Note that the overall loss in the RFBN leads to reduced radiated power, as well as additional problems in thermal management. Implementation trials for antenna feed networks based the proposed technique are currently being considered for commercial products featuring such advanced hybrid analog and digital beamforming in LTE-Advanced and 5G. Some of the future research directions are as follows. • In small-cell networks, combiner loss is not critical in terms of thermal management when radiated power levels ≤ 5 W. For example, an insertion loss of 1 dB results in approximately 0.6 W through heat dissipation. However, they affect the communication range and receiver sensitivity. For this reason, future small-cell networks must consider insertion-loss minimization in addition to wide beamtilt ranges. • Another aspect of our future work is the expansion to next-generation communication systems such as

•

•

•

millimeter-wave [14] and large-scale antenna arrays [24]. It is reasonably clear that the additional cost incurred due to increased number of transceivers for large-scale arrays will limit their widespread use. Cost-effective construction of large RFBNs and advanced 2-D beamsteering methods will ensure that the theoretical benefits of massive multiple-input multiple-output (MIMO) and millimeter-wave communication systems are realized in practice. Recently, there has been some interest in RFBN algorithms for 5G systems [25]. However, RFBN algorithms must consider realistic RF challenges as seen in Section IV, and failure to do so will lead to a significant gap between theory and reality [26]. The antenna array elements are designed such that the mutual coupling observed between adjacent antenna elements is less than 20 dB. However, as we proceed towards large-scale arrays, we expect that the mutual coupling between elements of a phased array to play a role in RFBN design. This is future work and outside the scope of this paper. Note that in the existing prototype, we calibrate the setup initially with TRX 6, as shown in Fig. 14. One aspect of our future work is designing algorithms to periodically estimate and adaptively compensate for changes due to temperature, environment, etc. ACKNOWLEDGMENT

The authors are very grateful to T. Kokkinos for his help in small-cell RFBN implementation and D. Morgan, as well as the anonymous reviewers for comments that have significantly improved the presentation quality of this paper. R EFERENCES [1] L. Godara, “Applications of antenna arrays to mobile communications— I: Performance improvement, feasibility, and system considerations,” Proc. IEEE, vol. 85, no. 7, pp. 1031–1060, Jul. 1997. [2] A. Hajimiri, H. Hashemi, A. Natarajan, X. Guang, and A. Babakhani, “Integrated phased arrays systems in silicon,” Proc. IEEE, vol. 93, no. 9, pp. 1637–1655, Sep. 2005. [3] G. M. Kautz, “Phase-only shaped beam synthesis via technique of approximated beamaddition,” IEEE Trans. Antennas Propag., vol. 47, no. 5, pp. 887–894, May 1999. [4] H. Zarei, C. Charles, and D. Allstot, “Reflective-type phase shifters for multiple-antenna transceivers,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 54, no. 8, pp. 1647–1656, Aug. 2007. [5] S. Tingay et al., “The Murchison widefield array: The square kilometre array precursor at low radio frequencies,” Pubs. Astronom. Soc. Australia, vol. 30, p. e007, 2013. [6] J. Butler and R. Lowe, “Beam-forming matrix simplifies design of electrically scanned antennas,” Electron Design, vol. 9, pp. 170–173, Apr. 1961. [7] D. Parker and D. Zimmermann, “Phased arrays-part II: Implementations, applications, and future trends,” IEEE Trans. Microw. Theory Techn., vol. 50, no. 3, pp. 688–698, Mar. 2002. [8] P. E. Haskell, “Phased array antenna system with adjustable electrical tilt,” U.S. Patent 7 450 066, Nov. 2008. [9] S. Mosca, F. Bilotti, A. Toscano, and L. Vegni, “A novel design method for Blass matrix beam-forming networks,” IEEE Trans. Antennas Propag., vol. 50, no. 2, pp. 225–232, Feb. 2002. [10] T. Djerafi, N. J. Fonseca, and K. Wu, “Planar-band 4 4 nolen matrix in SIW technology,” IEEE Trans. Microw. Theory Techn., vol. 58, no. 2, pp. 259–266, Feb. 2010.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 18

[11] M. Nedil, T. A. Denidni, and L. Talbi, “Novel Butler matrix using CPW multilayer technology,” IEEE Trans. Microw. Theory Techn., vol. 54, no. 1, pp. 499–507, Jan. 2006. [12] P. Chen, W. Hong, Z. Kuai, and J. Xu, “A double layer substrate integrated waveguide Blass matrix for beamforming applications,” IEEE Microw. Wireless Compon. Lett., vol. 19, no. 6, pp. 374–376, Jun. 2009. [13] C.-C. Chang, R.-H. Lee, and T.-Y. Shih, “Design of a beam switching/ steering Butler matrix for phased array system,” IEEE Trans. Antennas Propag., vol. 58, no. 2, pp. 367–374, Feb. 2010. [14] W. Roh et al., “Millimeter-wave beamforming as an enabling technology for 5G cellular communications: Theoretical feasibility and prototype results,” IEEE Commun. Mag., vol. 52, no. 2, pp. 106–113, Feb. 2014. [15] B. D. van Veen and R. Roberts, “Partially adaptive beamformer design via output power minimization,” IEEE Trans. Signal Process., vol. SP-11, pp. 1524–1532, Nov. 1987. [16] D. Palomar, J. Cioffi, and M. Lagunas, “Joint Tx–Rx beamforming design for multicarrier MIMO channels: A unified framework for convex optimization,” IEEE Trans. Signal Process., vol. 51, no. 9, pp. 2381–2401, Sep. 2003. [17] W. Yu and T. Lan, “Transmitter optimization for the multi-antenna downlink with per-antenna power constraints,” IEEE Trans. Signal Process., vol. 55, no. 6, pp. 2646–2660, Jun. 2007. [18] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.: Cambridge Univ. Press, 2006. [19] LTE Evolved Universal Terrestrial Radio Access (E-UTRA)—Release 10. pub-ETSI-STD, Eur. Telecommun. Standards Inst. TS 23.401 V 10.4.0 Nov. 2010. [20] T. K. Moon and W. Stirling, Mathematical Methods and Algorithms for Signal Processing. Upper Saddle River, NJ, USA: Prentice-Hall, 2000. [21] D. M. Pozar, Microwave engineering, 2nd ed. New York, NY, USA: Wiley, 2004. [22] J. S. Goldstein, I. S. Reed, and L. L. Scharf, “A multistage representation of the Wiener filter based on orthogonal projections,” IEEE Trans. Inf. Theory, vol. 44, no. 7, pp. 2943–2959, Nov. 1998. [23] V. Venkateswaran and A. J. van der Veen, “Analog beamforming in MIMO communications with phase shift networks and online channel estimation,” IEEE Trans. Signal Process., vol. 58, no. 8, pp. 1431–1443, Aug. 2010. [24] F. Rusek et al., “Scaling up MIMO: Opportunities and challenges with very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60, Jan. 2013. [25] O. E. Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014. [26] A. Garcia-Rodriguez, V. Venkateswaran, P. Rulikowski, and C. Masouros, “Hybrid analog-digital precoding revisited under realistic RF modeling,”, 2016, [Online]. Available: http://arxiv.org/abs/ 1604.08123 Vijay Venkateswaran (M’10) received the M.S. degree in electrical engineering from the University of Arizona, Tucson, AZ, USA, in 2003, and the Ph.D. degree from the Delft University of Technology, Delft, The Netherlands, in 2010. He is currently a Principal Researcher with Huawei Technologies, Kista, Sweden. From 2003 to 2005, he was an R&D Engineer with the Sony Corporation, Tokyo, Japan, where he was involved with data storage technologies. From 2010 to December 2015, he was a Member of Technical Staff with Bell Laboratories, Alcatel-Lucent, where he was involved with wireless technologies. He has held visiting/internship positions with Seagate (USA), Institute Eurecom, and Nanyang Technological University. He has authored or coauthored over 25 journal and conference papers. He has filed over 30 patent applications. His research interests are in the general areas of communication theory, RF systems, digital signal processing, and field-programmable gate array (FPGA) design, and more specifically on a hybrid of RF and digital communications.

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

Florian Pivit (M’06) received the Dipl.-Ing. (M.S.E.E.) and Dr.-Ing. (Ph.D. E.E.) degrees from the University of Karlsruhe, Karlsruhe, Germany, in 2000 and 2005, respectively. His doctoral research concerned base-station antenna technology He is currently a Technical Manager with Nokia Bell Laboratories, Dublin, Ireland. From 1999 to 2000, he was with Anaren Microwave Inc., New York, NY, USA, where he was involved with antenna feed systems for satellite and mobile communications. In 2006, he joined Lucent Technologies Bell Laboratories (now Nokia Bell Laboratories), as a Member of Technical Staff. His research interest is in the area of active antenna array technologies, integrated RF transceiver design, analog and digital beamforming techniques, millimeter-wave technologies, and antenna and filter technologies. Since 2013, he has head the RF Research Department, Nokia Bell Laboratories. He has authored or coauthored over 25 papers, journal publications, and book chapters. He has successfully filed over 30 patent applications and currently holds 16 patents in the field of antenna, filter, and transceiver technologies.

Lei Guan (S’09–M’12) received the B.E. and M.E. degrees from the Harbin Institute of Technology, Harbin, China, in 2006 and 2008, respectively, and the Ph.D. degree in electronic engineering from University College Dublin (UCD), Dublin, Ireland, in 2012. He then spent two years involved with industrialrelated research projects with UCD and Trinity College Dublin, respectively. He is currently a Member of Technical Staff with Nokia Bell Laboratories, Dublin, Ireland. He possesses ten years of R&D experience in field-programmable gate arrays (FPGAs) and digital signal processing. His research interests include RF signal conditioning algorithms for digitally assisted high-efficient RF front-ends such as digital predistortion and crest factor reduction, nonlinear system identification, software-defined radio systems, and wideband RF front-ends. He also has strong interests in high-performance computing platforms like FPGAs and all programmable systems-on-chip (SoCs), and their potentials in next-generation wireless systems, hybrid beamforming antenna array systems, and cloud radio access networks.

2244

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

2-D Beam-Steerable Integrated Lens Antenna System for 5G E-Band Access and Backhaul Juha Ala-Laurinaho, Jouko Aurinsalo, Aki Karttunen, Member, IEEE, Mikko Kaunisto, Antti Lamminen, Juha Nurmiharju, Antti V. Räisänen, Fellow, IEEE, Jussi Säily, and Pekka Wainio Abstract— The new services available through smart devices require very high cellular network capacity. The capacity requirement is expected to increase exponentially with the forthcoming 5G networks. The only available spectrum for truly wideband communication (>1 GHz) is at millimeter wavelengths. The high free space loss can be overcome by using the directive and beam-steerable antennas. This paper describes a design and the measurement results for a lens antenna system for E-band having 2-D beam-steering capability. Continuous beam-switching range of about ±4° × ±17° is demonstrated with the lens having the maximum measured directivity of 36.7 dB. Link budget calculation for backhaul application using the presented lens antenna system is presented and compared with the measurement results of the implemented demo system. Index Terms— Beam-steering, E-band, 5G, lens antenna, millimeter wave (mmW) communications.

I. I NTRODUCTION

T

HE smart device boom has created an ever-increasing need for higher cellular network capacity resulting in an exponential rise in backhaul capacity requirement as well. In particular, the next generation 5G mobile communication system is estimated to provide 10 000 fold increase in wireless data traffic by 2030 [1], [2]. Methods to increase data throughput in both the cellular access and the backhaul networks include spectral efficiency improvements by, e.g., using MIMO techniques, spectrum increase, and network densification, i.e., increasing the number of cells. The use of higher frequency bands, such as 28, 38, 60 GHz, or E-band (71–76 and 81–86 GHz), has been proposed,

Manuscript received September 1, 2015; revised April 1, 2016 and May 17, 2016; accepted May 23, 2016. Date of publication June 20, 2016; date of current version July 7, 2016. This work was supported in part by the Finnish Funding Agency for Technology and Innovation Tekes through the BEAMS Project, and in part by the Academy of Finland through the DYNAMITE Project. J. Ala-Laurinaho and A. V. Räisänen are with the Department of Radio Science and Engineering, Aalto University, Aalto FI-00076, Finland (e-mail: [email protected]; [email protected]). J. Aurinsalo, M. Kaunisto, A. Lamminen, and J. Säily are with the VTT Technical Research Centre of Finland, Espoo 02044, Finland (e-mail: [email protected]; [email protected]; [email protected]; [email protected]). A. Karttunen was with the Department of Radio Science and Engineering, Aalto University, Aalto FI-00076, Finland. He is now with the Department of Electrical Engineering, University of Southern California, Los Angeles, CA 90089-2560 USA (e-mail: [email protected]). J. Nurmiharju is with Vesatel Oy, Espoo 02200, Finland (e-mail: [email protected]). P. Wainio is with Nokia Bell Labs, Nokia, Espoo 00045, Finland (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TMTT.2016.2574317

Fig. 1.

Block diagram of the beam-steerable lens antenna system.

both for access and backhaul to tackle the above mentioned capacity growth [2]–[5]. Especially the E-band in this context provides many advantages. A huge spectral resource, an order of magnitude larger than in traditional microwave bands (6–42 GHz), is available and bandwidths comparable to 10–100-Gb/s fiber connections can be provided with low latency. The E-band is already used for fixed cellular backhaul links due to the worldwide allocation of two 5-GHz bands used for frequency-division duplexed (FDD) uplink and downlink thus enabling multigigabit data transmission per channel per link [6]. Besides FDD, it is also feasible to use time-division duplexing (TDD) in E-band systems. The light licensing regime available for E-band in many countries provides operators lower total cost of ownership and lower cost per transmitted bit than the traditional access and the backhaul bands. High frequency reuse is possible due to very narrow radiating beams. On the other hand, E-band propagation causes a few challenges for communication. Due to weak diffraction and penetration of radio waves in E-band, a clear line of sight is required. In addition, adverse weather conditions, such as heavy rain and wind, affect link performance. Electrical beam steering in two dimensions (in azimuth and in elevation) is required to ease the deployment and to cope with the mast sways of lamp posts and other assumed 5G small cell deployment sites during operation. The beam-steerable lens antenna system shown in Fig. 1 and described in this paper was developed for the TDD access and the backhaul systems and has been used for 5G access system demonstrations and field tests published in [7] and [8] and as the underlying transport media for a self-optimizing wireless backhaul concept [9], [10]. The Nokia millimeter wave (mmW) access and the backhaul demonstrators built based on the beamsteering lens antenna system are shown in Figs. 2 and 3, respectively.

0018-9480 © 2016 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.

ALA-LAURINAHO et al.: 2-D BEAM-STEERABLE ILA SYSTEM FOR 5G E -BAND ACCESS AND BACKHAUL

2245

TABLE I S YSTEM R EQUIREMENTS FOR 5G A CCESS AND BACKHAUL A PPLICATIONS

Fig. 2. Nokia E-band beam tracking demonstrator shown at Brooklyn 5G Summit 2014.

element and array design considerations, the waveguide transition design, switch network, and the complete feed network designs. Section VI discusses the antenna and feeder measurements and shows the measurement results for the complete lens antenna system. System level measurements and link budget analysis are given in Section VII. Finally, conclusions are drawn in Section VIII. II. 5G R EQUIREMENTS AND C HALLENGES FOR E -BAND A NTENNA S YSTEM

Fig. 3.

Nokia E-band backhaul node prototype in vibrating test pole.

The architecture of the integrated lens antenna (ILA) presented in this paper is based on the earlier designs in [11]. However, the access and backhaul system implementation with an operating feed array and the demonstration of the 2-D beam steering with 64 beams is the original work. In addition, a detailed performance analysis of the designed antenna system is presented based on the simulations and the system level measurements. This paper is organized as follows. Section II describes the main requirements set by the future 5G networks for the E-band antenna system. The E-band antenna solution alternatives are considered in Section III. In Section IV, the design of ILA design is presented. Section V describes the feed antenna

The system requirements for E-band antenna system for 5G access and backhaul applications are summarized in Table I. Then, some of the design challenges and consequences related to these requirements are highlighted and compared with the more conventional antenna systems for the access and backhaul use cases. The free space loss increases 6 dB when frequency doubles, which means that in E-band, path loss is more than 30 dB higher than in traditional mobile bands in 1–3-GHz region. Also higher bandwidth means that less power/Hz is transmitted and this must be compensated to sustain the signal to noise ratio (SNR) needed for receiver. For example, in 1 GHz bandwidth, 10 dB more received power is required than in 100 MHz bandwidth for the same SNR. To compensate for higher path loss either higher transmit power or higher gain antennas can be used. However, increasing transmit power is limited by regulation and technical aspects, such as dc power consumption and cooling requirements. The only way to have reasonable hop lengths in E-band is to use high gain directional antennas in both access and backhaul. In current 3G and 4G access systems, wide beamwidth sector antennas are used. When a user is moving to other

2246

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 64, NO. 7, JULY 2016

sector, the mobile network will do a handover procedure. These handovers do not happen at very fast rate per terminal. In E-band, and in mmW frequencies in general, this is different. If the beamwidth is only a few degrees, the user movement from beam to beam should happen in milliseconds. It is also probable that user movement will block RF signal completely, since mmW signals do not pass the human body or other obstacles. In that case, user traffic must be rerouted to another base station also within a very short timeframe. This unpredictable user movement from beam to beam and to another base station is a major challenge in the medium access control layer design for 5G access. Thus, beam-switching time and beam-steering decision speed are critical design parameters for the 5G access antenna system. High-gain antennas (ETSI 38 dBi and FCC 43 dBi) are used in traditional point-to-point E-band backhaul links. The beamwidth of such high-gain antennas is 1°–2° and in order to have a reliable link performance, backhaul radios have to be installed into specially designed rigid masts. Future small cell access and backhaul radios will be installed to lampposts and other utility poles, which will twist and sway in the wind much more than traditional telecom masts [12]. In order to compensate for the mast sway, the beam must be dynamically steered to mitigate the effect of mast movements. Electrical beam steering is preferred over mechanical beam-steering solutions due to faster reaction speed, lower cost, and higher reliability. Fully electrical beam steering has also other benefits. In the deployment phase of access or backhaul radio units, there is no need for time-consuming precision alignment of the antenna, and only rough manual aligning is needed. Thus, installation cost is reduced. The third benefit is the possibility of pointto-multipoint communication. The backhaul network topology can be designed to include redundant paths, which can be taken into use dynamically using electrical beam steering in the case of link failures or link degradations. In addition, the use of electrically beam-steerable antenna systems enables autonomous backhaul network build up as is depicted in [10]. III. A NTENNA S YSTEM C ONSIDERATIONS There are two conventional alternatives for implementing the electrical beam-steering antennas, namely, switched-beam antenna arrays and phased array antennas. Switched-beam antenna arrays have several available fixed beam patterns. An example of a switched-beam antenna is the ILA where the feed antenna array is attached to the back surface of the lens collimator. Only one of the antenna elements is active at any given time. The beam steering is achieved by switching between the feed antenna elements. The beam direction depends on the location of the active antenna element. In the development of switched-beam lens antennas, the switch network losses are the major obstacle in reaching the specified hop lengths. There is a tradeoff in the design; with a larger switch network, the number of beams is higher and scan range is larger, but losses in the network are higher. A phased array antenna [13] is one alternative for a beamswitching lens antenna. Several interesting scalable phased array structures and designs for mmW frequencies have been

recently proposed [14], [15]. Phased array antenna is an array of antennas, each having adjustable phase and possibly also gain. By shifting the phase of each antenna element to a predefined value, the signals can be added constructively in the desired direction and destructively in other direction. Thereby, antenna gain and direction of the transmission can be adjusted. RF phase shifting is probably the most feasible for mmW applications due to the lowest power consumption and the smallest chip area [16]. However, a phase shifter is needed for each antenna element, which increases the complexity, losses, and cost of a phased array, especially for large antenna arrays. With the beam-switching lens antenna, the overall complexity can be reduced by using switches having multiple outputs. Wafer-scale antenna array integration may become feasible at higher frequency applications; let us say above 100 GHz. At such high frequencies, the antenna element spacing is small enough, for example, 0.5λ = 1.25 mm at 120 GHz, and comparable to a chip size so that there is no space wasted on silicon wafer, which is very important from the cost perspective. On the other hand, chip assembly at such high frequencies is very difficult and parasitic impedances due to flip-chip or wire bonds can be avoided using waferscale integration. 16-element and 256-element phased arrays suitable for wafer-scale integration have been demonstrated for W -band and 60 GHz frequencies in [17] and [18]. Rotman lens [19], [20] and Butler matrix [21], [22] antennas have characteristics of both the switched-beam and phasedarray antennas. They have a number of input ports that are used to generate the desired amplitudes and phases for an array of radiating antennas. The complexity of both Butler matrices and Rotman lenses increase rapidly with increasing input port number. Electrical beam switching is possible by providing a switch network to the input ports. A complete 60-GHz beam-switching antenna based on a Rotman lens is presented in [20]. Butler matrices have also been realized at mmWs [22]. In addition to conventional beam-switched and phased array antennas there are recent developments using metamaterials. Most metamaterials are based on resonant structures and may be narrowband. For example, metamaterial gradient index lenses have been proposed to reduce the inherent scan loss and limited the beam-steering range of dielectric ILAs [23], [24]. In addition, Kymeta Corporation has developed reconfigurable metamaterials surface antenna technology (MSA-T) for Ka band (26.5–40 GHz) applications. Their antenna consists of a standard PCB-like circuit board composed of several thousand subwavelength resonators that can be individually tuned. The PCB-like board is attached to a conventional feed structure [25]. The MSA-T could be scalable also for the E-band frequencies, however, controlling of individual tuning elements could be challenging. IV. I NTEGRATED L ENS A NTENNA D ESIGN The designed antenna presented in this paper is an ILA with a 64-element feed antenna array. The beam selection is controlled by switching between the feed elements, as in [26]–[29]. The feed elements and the beams are aligned

ALA-LAURINAHO et al.: 2-D BEAM-STEERABLE ILA SYSTEM FOR 5G E -BAND ACCESS AND BACKHAUL

2247

TABLE II S IMULATION R ESULTS OF THE 95-mm D IAMETER R EXOLITE L ENS ; R ESULTS FOR L OWER , M IDDLE , AND U PPER F REQUENCIES W ITHIN THE L OWER E -BAND W ITH M INIMUM F EED O FFSET /M AXIMUM O FFSET

Fig. 4. Top: design of the lens antenna with a shaped extension and absorbers around the extension in order to reduce the internal reflections and radiation from the extension. Bottom: schematic of the element placement of the 64-element feed array.

in four rows of 16 producing continuous beam-steering range of about ±4° × ±17°. The antenna is designed to have high gain (G > 28 dBi), with low sidelobes, for the whole lower E-band from 71 to 76 GHz. The diameter of the lens is 95 mm and the material is Rexolite (cross-linked polystyrene), which has a relative permittivity εr of 2.53, and a loss tangent of 1.3 × 10−3 [30]. The lens is designed with an in-house developed ray-tracing simulation program [26], [27], [31]–[33]. The ray-tracing simulation is 3-D and takes accurately into account the 3-D shape of the lens and the feed offsets. The lens design is based on minimizing the reflections from the lens-air interface, and therefore, a very good simulation accuracy is achieved even without ray tracing the reflected rays [31]–[33]. The simulated radiation pattern of the designed aperturecoupled microstrip-line fed patch antenna (ACMPA) is utilized in the simulations conducted during the lens design. The patch antenna element design is presented in Section V-A. The rotationally symmetric lens and the 64-element feed array are shown in Fig. 4. The design of the lens shape and the arrangement of the feeds are presented in Sections IV-A and IV-B, and the simulation results are summarized in Section IV-C. A. Lens Shape The lens shape is designed with the design rules presented in [31]. The design starts with the typical rotationally symmetric elliptical ILA with eccentricity of e = (εr )−1/2 and cylindrical extension [34], [35]. The semiminor axis of the elliptical part is chosen 10% larger than the final lens radius. The design is based on retaining the good focusing properties with small feed offsets of the elliptical lens and designing the rest of the lens shape to avoid total reflections also with feed offsets. The lens shape, shown in Fig. 4, is designed to achieve low reflection loss (25.2%). The module generates a fixed broadside beam, but multibeam operation in H-plane can be easily achieved. In the 50–66-GHz band, the peak gain is 14.25 dBi and the average first side-lobe level in H-plane is −20.6 dB. The proposed technology and the design concept are suited for highly integrated millimeterwave systems, such as access points in the future V -band high data-rate wireless networks. Index Terms— Antenna-in-package (AiP), corporate feed networks (CFNs), fifth-generation (5G), low temperature cofired ceramic (LTCC), millimeter-wave (mm-wave) antennas, mm-wave technologies, transverse electromagnetic (TEM) waveguides.

I. I NTRODUCTION

T

HE exponential growth of mobile data traffic demands for novel technical solutions, providing multigigabit-per-second connections. The large unlicensed spectrum worldwide available in V -band, e.g., the 57–66-GHz band in Europe, could fit the requirements of short-range systems for high capacity millimeter-wave

Manuscript received September 29, 2015; revised April 7, 2016; accepted May 14, 2016. This work was supported by the European Union Seventh Framework Programme (FP7/2007-2013) through the MiWaveS Project under Grant 619563. F. Foglia Manzillo, M. Ettorre, and R. Sauleau are with the Institute of Electronics and Telecommunications of Rennes, University of Rennes 1, Rennes 35042, France (e-mail: [email protected]; [email protected]; [email protected]). M. S. Lahti and K. T. Kautio are with the VTT Technical Research Centre of Finland Ltd., Oulu 90571, Finland (e-mail: [email protected]; [email protected]). D. Lelaidier and E. Seguenot are with Orange Labs Networks, La Turbie 06320, France (e-mail: [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TMTT.2016.2574313

(mm-wave) cells. The latter will coexist in the future fifth-generation (5G) architectures with standard microwave technologies providing basic coverage to mobile users [1]–[3]. The mm-wave technologies and the antennas are key enablers for the development of broadband, high-speed radio access, and backhauling in 5G heterogeneous networks. The design of 60-GHz radios and related technologies have been hot topics in last two decades. However, there are still design and technological issues to achieve antenna requirements for mm-wave mobile networks, such as fractional bandwidth larger than 15%, high efficiency, a gain of ∼15 and 30 dBi for access points and backhauls, respectively, and beam-switching capabilities. Antenna stack-ups suitable for in-package integration with transceiver are demanded. System and antenna-in-package (AiP) solutions [4]–[6] are particularly promising to provide high performances while highly integrated. The expected gain and the scanning capabilities for 5G networks require large antenna apertures. Even though planar printed arrays are very attractive for cost and packaging reasons, their long feed networks lead to a very low efficiency. Slotted waveguide and open cavity arrays [7]–[9] in substrate-integrated waveguide (SIW) achieve gain values suitable for backhauling. However, they usually require large full-corporate feed networks (CFNs), implemented on stacked layers and interconnected by coupling elements, which cause a narrowband behavior and add losses to the overall system. Moreover, accurate fabrication processes are needed to embed CFNs in multilayer substrates. Several technological platforms have been proposed for AiPs, e.g., Teflon [10], fused silica [11], liquid crystal polymer [12], and LTCC [13]–[17]. LTCC process provides an unparalleled design flexibility for multilayer configurations. It allows the designer to stack a large number of dielectric and metal layers, and to distribute stacked and staggered vias. These features are very appealing for the design of low-loss networks in vertical configuration. On the other hand, the high dielectric constant of LTCC materials penalizes the antenna performance [19]. In this paper, a novel substrate-embedded full-CFN in vertical configuration and an LTCC integration platform for advanced AiPs are presented. In the state-of-the-art SIW feed networks, the signal is guided on dispersive waveguide modes, e.g., quasi-TE10 , and coupled to the radiating section through apertures. By contrast, we introduce a nondispersive true-time delay network based on vertical parallel plate waveguide (PPW) lines, realized with stacked via-rows embedded in the substrate. The designed CFN supports a

0018-9480 © 2016 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.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 2

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

Fig. 2. Simulation setup to evaluate the leakage loss of an LTCC-embedded PPW line realized with via-rows. Perfect electric conductors and a loss-less substrate (r = 5.74 and tan δ = 0) are considered.

Fig. 1. (a) 3-D view and (b) cross section in the xz plane of the designed antenna. The basic components of the module are indicated. Dimensions are reported in millimeters.

quasi-transverse electromagnetic (TEM) mode and enables the signal propagation along the vertical direction. It feeds an array of four long slots. The main objective of this paper is the demonstration of the broadband capabilities of the proposed CFN design and LTCC technology for integrated 5G antennas in mm-wave access point links. This goal is accomplished through the design and experimental characterization of the antenna architecture shown in Fig. 1. It exploits the wideband properties of parallel-fed arrays of tightly coupled long slots [20], [21]. A pillbox-based quasi-optical beamformer [22], [23] is used to create a line source that excites the CFN. It allows for a multibeam operation in one plane [20], even if the fabricated module generates only one fixed beam. The LTCC process is based on the Ferro A6-M tape system (r = 5.74 and tan δ = 2.3 × 10−3 ). Despite the stack-up complexity, the process employed ensures high accuracy. This paper is organized as follows. Section II introduces the key concept of vertical substrate-integrated PPW (SI-PPW) and its benefits for the design of nondispersive CFNs. In Section III, the design of each subsystem of the module is detailed. In Section IV, the LTCC process and the design rules are discussed. Section V reports guided and radiated measurements and a comparison with the state of the art. The conclusions are drawn in Section VI. II. V ERTICAL S UBSTRATE -I NTEGRATED PPW S The substrate-integrated PPW CFN is shown in Fig. 1. More details about its design are given in Section III-D and in Fig. 5(a). The wideband features of the proposed embedded CFN rely on the proper operation of vertical substrateintegrated PPW lines. Standard components and CFNs in the hollow PPW technology can be manufactured by means of

computer numerical control milling. However, their implementation in dielectric-embedded structures with planar processes is very challenging. Via-rows are used here to resemble the continuous metallic plates of ideal PPWs. The basic via-made PPW line is analyzed in more detail in Fig. 2. Two via-rows of length ll placed at a distance h l are considered. They are embedded in a substrate having the same relative permittivity of Ferro A6-M material (r = 5.74). The diameter of the vias (0.15 mm) and the pitch (0.30 mm) corresponds to the values of the actual design of the CFN. The leakage losses of the SI-PPW due to the fields propagating outside the rows of vias are discussed in the following. The numerical evaluations are based on full-wave simulations with Ansys HFSS 15. The via-made PPW section is connected to two ideal PPW lines, to be properly excited by waveports (ports 1 and 2), as shown in Fig. 2. In the final system, the values of wl  h l , ll have been selected for the PPW lines (see Section III-D) to approximate the behavior of ideal PPW lines that are infinite along the y-direction. In the structure under analysis, several finite values of wl have been considered and perfect magnetic conductor (PMC) boundaries have been imposed at the edges. In order to estimate the leakage loss, substrate losses are neglected (tan δ = 0), and vias and metals are assumed perfect conductors. The simulated attenuation at 60 GHz for h l = 0.48 mm is very low, about 3 × 10−6 Np/mm2 , i.e., per unit of length of propagation and width of the via-made structure. For example, if wl = 30 mm, as in the actual CFN, the attenuation along the z-direction, due to the leakage, is ∼0.08 dB/m. This value is comparable to typical leakage losses in low-loss SIWs [24]. The overall losses of an SI-PPW are evaluated with reference to a specific geometry employed in the designed CFN. The center-to-center distance of the two parallel rows of vias is h l = 0.48 mm. A finite structure is considered, with a width wl = 30 mm. The diameter and the pitch of the vias are the same as in Fig. 2. The technological parameters of the Ferro A6-M tapes (r = 5.74 and tan δ = 0.0023) and the conductivity σ = 7 × 106 of the golden paste used for metals are assumed in simulations. The SI-PPW is excited by two waveports, with a simulation setup similar to the one shown in Fig. 2, where the PMC boundaries are replaced by radiation boundaries, given the finiteness of the structure along the y-direction. Several lines of different lengths ll are

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. FOGLIA MANZILLO et al.: MULTILAYER LTCC SOLUTION FOR INTEGRATING 5G ACCESS POINT ANTENNA MODULES

TABLE I S IMULATED L OSSES AT 60 GHz OF A S TANDARD PPW LINE AND OF AN SI-PPW L INE IN F ERRO A6-M M ATERIAL

3

realizes a good field confinement by employing three parallel via-fences for each side of the waveguide. The solution of multiple via-fences may be too bulky for the design of substrate-integrated CFN at mm-waves, given the small array spacings. By contrast, the proposed concept of vertical via-made SI-PPW allows the design of CFNs working on a quasi-TEM mode, fully embedded in an LTCC substrate, by using a single via-row for each side of the PPW. III. D ESIGN OF THE A NTENNA M ODULE

considered. The loss per unit length at 60 GHz is then derived. The computed values for each loss mechanism are reported in Table I. They are compared with the attenuation values obtained for a standard PPW line, with continuous metal plates of conductivity σ , filled with Ferro A6-M material. The same geometrical parameters (h l , wl , and ll ) are considered. The total attenuation of the SI-PPW is 0.062 dB/mm. The ohmic dissipation represents the largest contribution to losses (0.039 dB/mm). It is observed that the via-made structure introduces additional ohmic losses, as compared with the standard PPW line (0.021 dB/mm). Given the small vias pitch, the leakage loss of the SI-PPW is almost negligible. The same dielectric losses (0.023 dB/mm) are estimated for the two lines. In conclusion, the increase of loss due to the vias is limited and the total attenuation of an SI-PPW, with tightly spaced vias, is similar to the values measured on other mmwave laminates. As a reference, the attenuation at 60 GHz of an SIW line in a low-loss polytetrafluoroethylene substrate (r = 2.16, tan δ = 7 × 10−4 , and σ = 1.85 × 107 ) is 0.013 dB/mm [25]. Larger losses per unit length are observed for LTCC components. In [18], an attenuation of 0.189 dB/mm is estimated for an SIW line in a lossy LTCC substrate (r = 6.6, tan δ = 0.013, and σ = 3.3 × 107 ). In order to realize the proposed concept of embedded full-CFN in vertical configuration, the via-rows that define the SI-PPWs are distributed on different stacked tapes. The electric field is guided by the nondispersive PPWs through the vertical axis of the module. It is polarized perpendicularly to the viarows delimiting the vertical PPWs, i.e., along the x-direction. Several multilayer SIW power dividers and feeding systems have been presented in the literature [9], [13]. However, they are all based on laminated waveguides that support dispersive waveguide modes, propagating in a direction orthogonal to the vertical axis of the substrate and with the E-field parallel to the via-rows. The vertical coupling is enabled by apertures on the metallic plates of the SIWs and is often effective only in a narrow band. High fabrication accuracy for the conductor patterns may be required. To the best of our knowledge, no full-CFNs in SIW achieving a guided propagation along the vertical axis of a multilayer substrate have been ever demonstrated. This is due to the challenges rectangular SIWs encounter in providing a good propagation along the vertical axis. Indeed, the surface current distribution of a quasiTE10 mode on the broad edge of a rectangular waveguide cannot be effectively supported by via-rows, considering the present technological constraints on the vias-pitch. The vertical SIW structure presented in [15] for a corrugated horn

The antenna system, shown in Fig. 1, is fed on the bottom face by an end-launch connector. A grounded coplanar waveguide (GCPW) to SIW transition transfers the signal to an SIW sectoral horn, which is the input of the quasioptical beamformer. Note that the proposed LTCC platform enables the cointegration of ICs connected to the input feeds of the beamformer. The horn illuminates a parabolic via-made reflector, and a multislot pillbox transition guides the signal toward the input of the PPW CFN. The latter feeds in parallel four long slots, loaded by a matching cover. The final size is 32.5×34×3.4 mm3. The array occupies only 8.4×34 mm2 of the total area, since it comprises only four slots, while the focal length of the parabola is much larger. The present design aims at the validation of the LTCC integration process. It has not been optimized to reduce the number of layers (18 tapes have been used). The module can be redesigned in order to allocate a larger array, e.g., eight or more slots, by using the same number of tapes. Each subsystem is described in the following. Full-wave simulations are performed with Ansys HFSS 15, assuming for the substrate r = 5.74 and tan δ = 0.0023, and a conductivity σ = 7 × 106 for the metals. A. Input Section A coaxial end-launch connector (Southwest Microwave, model 1892-03-A-6) feeds a 50- GCPW line printed in the bottom metalization of the module. A GCPW-to-transition guides the signal from the input line to the horn feeding the beamformer. The details of the transition are shown in Fig. 3(a). The width of the strip and gaps of the GCPW are 0.40 and 0.20 mm, respectively. The GCPW is gradually matched to an SIW line of width 1.70 mm (center-to-center distance of the via-rows), which ensures a single-mode propagation in the 37–75-GHz band. A thickness of 0.40 mm ≈ 0.2λd , where λd is the wavelength in the dielectric at 60 GHz, is chosen for the SIW line in order to reduce its losses. The GCPW line has the same height, i.e., the distance between signal plane and reference ground, so that the input section occupies two LTCC tapes of thickness 0.2 mm. The design is based on a tapered section of length 2 mm ≈ λd . An impedance step in the GCPW line at the interface with the SIW and two λd /4-long coupling stubs ensure both impedance and field matching. The width of the coupling slots is tuned to maximize the field transfer to the SIW line. The via-fences surrounding the GCPW are designed to suppress surface waves and PPW modes. The minimum diameter (0.15 mm) and the pitch (0.30 mm) allowed by the process are chosen to minimize leakage in the SIW line [24]. The simulated

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 4

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

Fig. 5. Radiating section and PPW CFN. (a) Cross section and (b) top view of the array. Dimensions are in millimeters.

Fig. 3. Input coaxial-to-GCPW-to-SIW transition. (a) Geometry and stack-up (inset on the top), dimensions are in millimeters. (b) Simulated results when fed by the V -band end-launch connector.

Fig. 4. Top view of the quasi-optical system. Dimensions are in millimeters.

performances of the transitions are shown in Fig. 3(b). The model of the end-launch connector has been included in the simulations and port 1 refers to the input of the coaxial connector. The reflection coefficient is lower than −10 dB in the band 50–70 GHz, and lower than −18 dB in the design band (57–66 GHz). Within this range, the insertion loss is ∼0.65 dB.

first one, as shown in Fig. 1(b). The common ground shields input and radiating sections. A pillbox transition [22], [23] connects the two PPW regions. It consists of a 180° PPW bend with a parabolic profile and of a series of thin slots contouring the reflector, in the common ground of the PPWs. The slots are placed at a distance of about λd /2 (0.90 mm) from the reflector and have a width of λd /10 = 0.20 mm. Their lengths are designed so that they subtend the same angle (≈2.5°), as seen from the focus of the parabola. The cylindrical wave launched by the horn is transformed into a quasi-plane wave in the upper PPW by the parabola. The design of the reflector determines the radiation characteristics of the array in H -plane (yz plane in Fig. 4). The diameter D of the reflector, equal to the length of the radiating slots, is set to 30 mm. The focal-to-length ratio of the reflector is F/D = 0.67, so that the system has an edge tapering of about −16.5 dB. This value achieves a good tradeoff among illumination, spill-over efficiency of the parabola, and sidelobe levels (SLLs) of the radiated patterns in H -plane. The vias of diameter of 0.20 mm, larger than minimum allowed by the process, are used to implement the reflector, in order to limit their number. The via-to-via distance (0.40 mm) ensures negligible leakage. The overall quasi-optical system has a thickness of 0.80 mm, i.e., four 0.20-mm-thick LTCC tapes.

B. Quasi-Optical Beamformer The quasi-optical system is shown in Fig. 4. The phase center of the H -plane sectoral horn is placed in the focus of a 2-D via-made parabolic reflector. The horn aperture is w = 4.13 mm ≈ 2λd , and the flaring length is lt = 2λd . These parameters minimize the phase aberrations at the aperture plane in the operating band [27]. The input section is embedded in a PPW structure of height 0.40 mm. A second PPW, of the same height, is stacked on the

C. Long Slot Array An accurate design of the array is essential to achieve broadband operation [21], given the high permittivity of the dielectric. The array comprises four slots in a metal plane, as shown in Fig. 5. Each slot is fed along the z-axis by a PPW line, as wide as the slot. Each element can be thus seen as an open-ended waveguide integrated in LTCC and radiating in the upper half-space. If this structure radiated in air, it would

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. FOGLIA MANZILLO et al.: MULTILAYER LTCC SOLUTION FOR INTEGRATING 5G ACCESS POINT ANTENNA MODULES

Fig. 6. Active impedance Z act seen at (a) edge elements and (b) central elements of the four-element array, for different thicknesses t of the cover. The array spacing is d = 1.20 mm and the slot width is a = 0.83 mm. Values are normalized to the characteristic impedance Z 0 of the PPW feedline, having width a along the x-direction and length 30 mm along the y-axis.

suffer from strong reflections at the air–dielectric interface. The real part of the active impedances can be several times higher than the impedance of the PPW feedlines. To overcome this issue, the array can be loaded by a matching dielectric layer made of several LTCC tapes, i.e., the same dielectric used for the module, without any metallic pattern. It can be placed right above the radiating slots and integrated in the same process. Its thickness can be tuned to enhance bandwidth. It also acts as a protective cover. The slot width a, the array period d, and the thickness t of the cover have been jointly optimized following the design guidelines in [21]. Values of a ≤ 0.9 mm have been considered, so that only the fundamental mode propagates in the PPW feedlines of the slots. The optimization has been constrained to the values of d ≤ λ0 /2 to avoid grating lobes and scan blindness phenomena. The final values are a = 0.83 mm, d = 1.20 mm ≈ 0.6λd , and t = 0.40 mm. The length of the slots is equal to the diameter of the parabola D = 30 mm ≈ 15λg . Fig. 6(a) shows that without the matching layer (t = 0), the real part of the active impedance Z act seen at the edge elements would be more than three times higher than the impedance of the PPW feedlines of height a, i.e., Z 0 = ηd a /D, where ηd is the wave impedance in the dielectric. For such an impedance ratio, a good matching cannot be practically achieved without using overmoded PPWs in the CFN. By contrast, for t = 0.40 mm, corresponding to two LTCC tapes, {Z act } < Z 0 both for edge and central elements of the array [see Fig. 6(b)]. Moreover, {Z act } is very low and constant against frequency and can be compensated by the CFN. In particular, the capacitive component of Z act can be canceled by the series feeding line [of width 0.83 mm in Fig. 5(a)], and the shunt capacitor realized by the E-plane step discontinuity [26], between the feeding line and the SI-PPW line of width 0.48 mm [see Fig. 5(a)]. The E-plane 90° bend can also be designed to provide an equivalent reactance [26] and compensate a residual imaginary part of Z act . D. SI-PPW CFN The CFN is designed to match over a wideband the active impedances of the array, seen at the radiating slots, to the

5

Fig. 7. Simulated reflection coefficient and losses of the CFN, considering ideal PPWs (dashed lines) and the actual via-made structure (solid lines). The input port is placed at section A A in Fig. 5(a).

impedance of the input PPW line. The CFN has been initially designed, considering ideal PPWs, with the aid of the equivalent circuits [26] of its basic components, such as E-plane T-junctions, E-plane steps, and bends. Only single-mode PPW lines, i.e., smaller than 0.9 mm in the x-direction, have been considered to achieve a theoretically nondispersive network. The first-pass design has been then optimized by full-wave simulations, considering the actual SI-PPW structures and vias of diameter 0.15 mm. Multistage transformers have been avoided to reduce the number of LTCC layers. The reactive parts of the active impedances of the array are almost canceled by using an E-plane step in series with the feedlines of the slots. A basic T-junction [see T-junction 1 in Fig. 5(a)] and a quarter-wavelength transformer that occupies three layers are used to achieve the impedance level of a reference PPW line of height 0.48 mm. The T-junction 2 [see Fig. 5(a)] comprises a power divider, quarter-wavelength transformers, and 90° bends, and is designed to achieve a broadband matching [28] at the input PPW line of the CFN. Note that to implement the quarter-wavelength transformer in the T-junction, two 0.1-mm-thick layers have been used [see Fig. 5(a)]. These small features ensure the enhancement of the matching bandwidth. The impact of the via-rows on the performances of the CFN has been investigated in simulations. Two simulation setups have been considered. In the first one, the CFN is realized by means of PPWs with continuous metal plates. For vertical PPW lines, i.e., extending along the z-axis, the distance of the parallel plates is equal to the center-to-center distance of the via-rows employed in the final design. All PPW lines have a finite length D = 30 mm along the y-direction. In the second setup, the actual via-made structure, with finite extent along the y-axis, has been analyzed. Dielectric and conductive losses have been included in both setups. Some results are shown in Fig. 7. The reflection coefficient at the input PPW line of the CFN [section A A in Fig. 5(a)] is not significantly affected by the vias. The −10-dB impedance matching bandwidth is larger than the 50–66-GHz band. The simulated losses of the CFN from the input port to the plane of the radiating slots are reported as well. The difference between the values obtained from the two simulation setups are due to additional

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 6

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

TABLE II S OME D ESIGN RULES OF THE LTCC P ROCESS

losses introduced by the via-made PPW, i.e., energy leakage through the via-rows and conductive losses in the vias. A small increase of the loss of ∼0.05 dB is observed at 60 GHz and similar values in the entire band. These results confirm the broadband, low-loss behavior of the via-made CFN. IV. LTCC P ROCESS Among the materials commercially available, Ferro A6-M tape system [29] has been selected, because its dielectric loss and permittivity are the most suitable to achieve good antenna performance. The nominal values of relative permittivity r and loss tangent tan δ are usually given at low frequencies, typically around 2.5 GHz. At mm-waves, these values are different. Measurements on test structures on Ferro A6-M has provided the values of r = 5.74 and tan δ = 2.3 × 10−3 for the loss tangent. The system utilizes screen-printed gold (Au) conductors, with conductivity σ = 7×106 S/m. The prototypes have been manufactured by VTT. Some of the processing parameters are shown in Table II. The overall substrate consists of 18 tapes. The two uppermost ones are blank layers, used for the matching cover. Most tapes have the thickness of 10 mil (≈0.20 mm after cofiring). Two layers have the thickness of 5 mil (≈0.10 mm after cofiring) and have been employed to optimize the design of T-junction 2 in Fig. 5(a). It is important to ensure that these two tape thicknesses have a close shrinkage match to avoid any delamination or warpage problems. The fabrication of SI-PPW full-CFNs in vertical configuration does not demand narrow and very accurate conductor lines, which are frequently requested at mm-waves. For instance, standard feeding systems in SIW require accurate conductor-to-conductor line distances and fine features for the coupling slots, since these parameters significantly affect the performances. Being based on via-rows, the proposed SI-PPW network sidesteps these needs, since simple metal patterns are employed. The focus of the process is in the alignment of stacked vias on different layers. Four antennas have been placed on a 6.5-in square LTCC tape sheet, one in each tape quarter. Via holes were punched and filled with Au paste using stencils. Conductor patterns were carefully aligned to the vias by proper scaling the screen pattern and by printer adjustments. Cavity holes were punched to the tape layers to realize through holes in the substrate for mounting the connector with screws. Tape layers were stacked in a mechanical fixture, and the stack was laminated isostatically. The laminate was green cut into 70 mm × 70 mm size to be cofired at 850 °C. The fired substrates were then diced to the final antenna dimensions.

Fig. 8. Photographs of one of the fabricated antennas, with small differences from the antenna described in this paper. (a) Connectorized antenna module. (b) Zoomed-in cross-sectional view. (c) Cross-sectional view of the CFN, showing the alignment of stacked via-rows.

Fig. 9. Measurement setup for gain patterns measurements. Inset: AUT and holder.

Fig. 10. Measured and simulated reflection coefficients at the input connector of the antenna module.

Four antenna designs, with slightly different geometrical features, have been fabricated. Fig. 8 shows some photos of the prototypes. It should be noticed that the proposed

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. FOGLIA MANZILLO et al.: MULTILAYER LTCC SOLUTION FOR INTEGRATING 5G ACCESS POINT ANTENNA MODULES

7

Fig. 11. Normalized radiation patterns in H -plane at (a) 52, (b) 54, (c) 58, (d) 60, (e) 63, and (f) 66 GHz. Measured co-polarized (Co-pol) and cross-polarized (X-pol) components are shown.

concept requires rather large amount of tightly spaced vias. The alignment of vias ensured by the process is good, as shown in Fig. 8(b) and in the cross-sectional view [Fig. 8(a)] for one of the fabricated antennas. Note that Fig. 8(a) does not refer to the design described through all this paper, for which no pictures of the cross section are available. The sample in Fig. 8(a) features, for matching purposes, a row of vias close to the input of the CFN that has not been employed in the antenna discussed in this paper [Fig. 5(a)]. Metallic catch pads, shown in Fig. 8(b), have been used to improve the electrical contact of via-rows stacked on different layers. They consist of the round pads, one for each via, of diameter 125 μm. The region of the antenna occupied by the CFN is 50 μm thicker than the rest of the module. This is due to the high density of vias, filled with a gold paste, that do not shrink as much as the surrounding ceramic. Moreover, the metal layers, each of typical thickness in the range 4–7 μm, are mainly concentrated in this area. The measured thickness of the substrate varies in the range 3.135–3.185 mm from edge to the center. V. E XPERIMENTAL R ESULTS AND D ISCUSSION The S-parameters, radiation patterns, and gain of the antenna design described in this paper have been measured in the frequency range of 50–66 GHz at Orange Labs La Turbie. A mm-wave configuration (50–75 GHz) of Orange Labs anechoic chamber has been used to perform radiation patterns and gain measurements in far field. The setup is shown in Fig 9. A custom holder for the antennas under test (AUTs) has been fabricated in a foam material (r ≈ 1), in order to precisely control the position of the antenna and its alignment

with the standard gain horn. This setup provides accurate measurements when the distance between the AUT and the 2 /λ is reference horn is larger than 2R f , where R f = 2 Dmax 0 the Fraunhofer distance, Dmax is the largest dimension of the AUT, and λ0 is the free-space wavelength. The larger diagonal of the top surface of the AUT is assumed as maximum dimension (Dmax ). A distance of 2 m, i.e., about 2.05R f at 66 GHz, was used for the measurements shown in the following. A. Measurements Measured and simulated reflection coefficients at the input V -band connector are shown in Fig. 10. They are in very good agreement, thus confirming the reliability of the proposed design and fabrication process. A small frequency shift between numerical data and measurements is observed above 62 GHz. It might be attributed to the deviations of the dielectric constant and fabrication tolerances. Except for a peak of −9.2 dB at 51.2 GHz, the reflection coefficient is lower than −10 dB over the entire frequency range, i.e., the relative impedance bandwidth exceeds 25%. The H -plane cuts (yz plane in Fig. 1) of the normalized patterns are shown in Fig. 11 at several frequencies. Well-shaped patterns with low SLLs are observed over the entire measurement band. Good agreement with simulations is observed, in particular for the main lobe, despite the complexity of the fabrication process. The differences in the positions of the first nulls at some frequencies may be due to fabrication errors and to the drifts of the permittivity of the tapes from the nominal value. The measured normalized patterns in H -plane as a function of

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 8

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

Fig. 12. Radiation patterns measured in H -plane as a function of frequency and elevation angle θ . (a) Co-polarized component. (b) Cross-polarized component. Fig. 15. Antenna realized gain, directivity, and estimated radiation efficiency against frequency. TABLE III E STIMATED L OSSES AT 60 GHz

Fig. 13. First SLLs and HPBW of the radiation patterns in H -plane, across the 50–66-GHz band.

Fig. 14. Measured and simulated radiation patterns in E-plane at (a) 57 and (b) 66 GHz.

the frequency and elevation angle θ highlight the stability of the radiation properties. The co-polarized and cross-polarized components are shown in Fig. 12. In the 50–66-GHz band, the angular tilt of the main beam is lower than 0.6°. The peak cross-polarization level is lower than −23 dB in the range 52–66 GHz. The full half-power beamwidth (HPBW) varies between 9° and 12.6°, while the first SLLs are lower −17.5 dB, with an average value of −20.6 dB, as shown in Fig. 13. The patterns in E-plane (x z plane in Fig. 1) reflect the small size of the array (four elements), much shorter than the length of the overall module along the x-direction. Measured and simulated radiation patterns in E-plane are shown in Fig. 14. The asymmetries are due to the off-center position of the array along the x-axis, while the ripples are mainly due to the limited size of the ground plane of the slots. A radiation box, including the entire system,

has been used in simulations. The patterns presented have been calculated considering a reference system located in the middle of the radiating aperture [see Fig. 5(b)], i.e., offset from the center of the module. The same reference system is assumed as the antenna phase center in the measurements. The discrepancies between numerical and experimental data may be due to possible misalignments between the reference horn and the actual phase center of the module under test, and also to spurious radiation from the input connector. The measured realized gain is compared with simulated values and directivity against frequency in Fig. 15. The maximum measured value is 14.25 dBi at 62.7 GHz. The measured −3 dB gain bandwidth is 54–66 GHz (20%). The average difference between measurements and simulations is 0.68 dB in the 50–66-GHz band. The additional measured loss and some discrepancies at high frequencies are attributed to fabrication errors, such as layer misalignments, shrinkage, and bending, to the assembly of the input connector, to deviation of material losses from nominal values. The radiation efficiency is estimated as the ratio of measured gain and simulated directivity. Its mean value in the measurement band is 46%, and it is never lower than 33%. Despite the large volume of the system, due to very thick stack-up and the pillbox beamformer, these values are in line with the efficiency achieved by more compact state-of-the-art LTCC antennas at mm-waves. B. Discussion and Design Tradeoffs The antenna gain and the radiation efficiency are mainly limited by dielectric and ohmic losses. An estimation of their separate impact at 60 GHz is proposed in Table III. The input

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. FOGLIA MANZILLO et al.: MULTILAYER LTCC SOLUTION FOR INTEGRATING 5G ACCESS POINT ANTENNA MODULES

9

TABLE IV C OMPARISON OF THE S TATE - OF - THE -A RT 60-GHz A NTENNAS IN LTCC

connector is included in all simulation setups considered in the following. First, the overall module is simulated assuming lossy dielectric and perfect conductors. The dielectric losses (0.98 dB) are evaluated as the difference between simulated directivity and gain. The ohmic losses (1.75 dB) are computed by considering the finite conductivity of the metals and lossless dielectric. It is found that the ohmic losses represent the dominant dissipation mechanism. The simulation that includes both dielectric and ohmic losses provides a value almost equal to the sum of the losses previously estimated. The major part of the losses is due to the quasi-optical system (≈1.9 dB at 60 GHz) because of its large size. The insertion loss of the input transition at 60 GHz is ∼0.6 dB. By subtracting these values from the total loss in Table III, a loss of ≈0.2 dB is evaluated for the CFN. It can been observed that only a small amount of the energy is radiated by surface waves. Assuming an input power of 1 W, the power flowing through the four sides of the substrate and through the portion of the top-face not occupied by array is ∼0.03 W at 60 GHz. This further confirms that SI-PPWs made of via-rows (Fig. 2) provide a good field confinement. The previous analysis demonstrates that the limited gain is the main drawback of the proposed antenna solution. The reasons of this limitation are twofold: the losses of the quasioptical system, which are inherently related to the antenna architecture, and the small number of radiating slots due to technological constraints on the number of LTCC tapes. Even though the quasi-optical beamformer is responsible for the largest part of the losses of the entire system, it provides multibeam capabilities and high scanning performances [22], [23]. The losses of the quasi-optical system are proportional to the focal length F and to the diameter D of the reflector. The definition of these quantities is related to the tradeoff among SLLs, beam overlap level in a multibeam system [23], and antenna directivity. A fixed beam design, such as the

one presented in this paper, may benefit from small values of F that lower the losses and ensure low SLLs. However, this choice generally penalizes the scanning performances in multibeam architectures [23]. The implementation of a larger array ensures a higher gain but requires additional LTCC tapes and thicker stack-ups. Furthermore, fabrication costs and risks increase. With reference to the design here discussed, the number of radiating slots can be doubled, for example, by using three more LTCC tapes of thickness 0.2 mm. Indeed, three layers are required to implement an additional stage of power dividers, with geometrical features similar to the T-junction 1 shown in Fig. 5(a). The total thickness of the module would increase to 4 mm, which is the maximum allowed value in the standard LTCC production processes [29]. A thickness of 5 mm may be provided for special applications. It is thus clear the need of saving LTCC tapes in the design of the beamformer and of the CFN. Focusing on the present antenna design, the implementation of the quasi-optical system with only two LTCC tapes, instead of four, would reduce the overall thickness. The additional losses, due to the smaller distance of the metallic parallel plates [30], are estimated to be 1.25 dB at 60 GHz. The realization of an array of eight slots would then lead to a net increase of the gain, considering a rough increase of 3 dB of the directivity. Moreover, the SI-PPW lines between the input T-junction and the power dividers shown T-junction 1 in Fig. 5(a) could be shortened (0.4 mm) at the expenses of a reduced bandwidth. However, by fine tuning the parameters in the input power divider, the antenna performances would not be significantly affected between 57 and 66 GHz. The design of T-junctions in a single layer of thickness 0.2 mm is currently being investigated for the future development of larger arrays fully embedded in LTCC. Finally, it is worth to remind that the antenna architecture and the selected LTCC technology are particularly

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 10

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

suitable for the integration of power amplifiers in a monolithic module. Such active antennas could thus realize high gain values while retaining the advantages of the proposed solution. C. Comparison With the State of the Art A comparison of characteristics and performances between the prototype presented in this paper and the state-of-the-art 60-GHz antennas in LTCC is reported in Table IV. Microstrip arrays [14], [16] are compact and require few LTCC tapes, but they are limited in bandwidth and efficiency. Arrays of wideband elements, such as cavity antennas [9], L-probe fed [17], or metamaterial-enhanced patches [13], can achieve broad impedance and gain bandwidths. They are typically fed by SIW parallel-feed networks that lead to larger sizes and higher losses, thus requiring more radiating elements to achieve a given gain value. In [18], a high-gain (20.4 dBi at 58.7 GHz) postwall parallel plate slot array is presented. The matching and the gain bandwidth are inherently narrow ( Z RV MIN ), the reader resonant frequency is set to 126.7 kHz and no RQR is used (Table II). The maximum read distance using this reader is 17 cm, 1 cm higher than the one presented in Section III. This is due to the fact that the reader current slightly increases because of not using RQR to limit |Z RV |, as was explained in this section. The additional coil resonant frequency is not used to match bit amplitudes, as it will jeopardize the charging phase, which is the limiting one as is going to be shown in Section V-B. B. Additional Coil Design It is not easy to obtain analytic expressions for VC L and V R S as a function of the additional coil radius (r A ). Strictly, its optimum value depends on Q A , the resonance frequencies wres R,A,T , and the distance between coils DRA (reader-additional) and DAT (additional-tag). In addition, the relationship between r A and the coupling factors (kRA and kAT ) should be known. Fig. 5 allows us to find the optimum r A that requires the lowest Q A to fulfill the steady-state amplitude requirements. The first plot of Fig. 5 refers to the charging phase. The minimum Q A that satisfies VC L > 5 V for at least one position of the additional coil is plotted as a function of DRT = DRA + DAT for several r A values. The same was done for the reading phase, but satisfying V R S > 10 mV p . In this section, f res R = 126.7 kHz (as was discussed in Section V-A) and f res A = 134.2 kHz were used.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. PÉREZ-NICOLI et al.: BIDIRECTIONAL ANALYSIS AND DESIGN OF RFID USING AN ADDITIONAL RESONANT COIL

7

Fig. 7. Part of the response signal (measured) received by the reader (11101001). Without RQA , no steady state is achieved and intersymbol interference occurs. Fig. 5. Minimum Q A that fulfills charging phase, low (0)- and high (1)-b steady-state amplitude constraints, for r A values between r A = 5 cm and r A = 25 cm.

Fig. 6. Minimum Q A that fulfills all steady-state amplitude constraints, for r A values between r A = 5 cm and r A = 25 cm.

This analysis determines the optimal values of r A , which require the lowest Q A in order to fulfill the specifications (VC L > 5 V and V R S > 10 mV p ). A higher Q A value may cause intersymbol interference. As can be seen, for the charging phase, slight variations exist for radius between 13 and 25 cm. For the 0 b (134.2 kHz) reading phase, almost no variations exist between the different radius. The 1 b (123.2 kHz) starts presenting differences at long distances (DRT > 45 cm). It is important to highlight that the charging phase is the one that requires higher Q A . It should be mentioned that, strictly, Q [email protected] kHz should fulfill the charging and 0-b constraints and Q [email protected] kHz should fulfill the 1-b constraint. Fig. 5 allows different DRA values for charging and reading. This is useful to see the dependence of each phase with r A , but the design must consider the same additional coil position (DRA ) for any phase. In Fig. 6, the minimum Q A that satisfy all the specification (VC L > 5 V, V R S @123.2 kHz > 10 mV p , and V R S @134.2 kHz > 10 mV p ) for at least one DRA is plotted as a function of DRT . With r A = 13 cm (or lower), for DRT around 45 cm, the limiting phase is the charging one, as the reading phase does not need a high Q A . For higher radius, the 1 b starts limiting at large DRT (around 50 cm, see Figs. 5 and 6). In conclusion, the best performance is obtained with r A = 13 cm. This radius fulfills the steady-state

TABLE II F INAL D ESIGN

amplitude constraints, with the minimum possible Q A , preventing intersymbol interference. Moreover, the best additional coil resonant frequency is fres A = 134.2 kHz. Changing this resonant frequency directly deteriorates the charging phase reducing the read distance, see (8) and (10). If the 1 b had limited the read distance, a different additional coil resonant frequency would have been considered. Achieving high DRT is only possible if Q Amin needed does not neglect the transient response. A three-coil system was built using the typical tag [Fig. 2(c)] and the reader without RQR and f res R = 126.7 kHz. The additional coil has 13-cm radius, 40 turns, 0.9-mm wire section diameter, a quality factor of [email protected]−kHz ([email protected] kHz ), and 134.2-kHz resonant frequency. Initially, RQA is not included. The three-coil system was set with DRA = 40 cm, DAT = 3 cm, and DRT = 43 cm. At such a distance, the tag was correctly charged but its response was not properly read. Fig. 7 shows the signal received at the reader with and without the additional coil. As can be seen, the high Q A affects the transient response. This causes that the steady state is not achieved in each bit time (visible in the dashed line region of the waveform). Intersymbol interference then occurs preventing a successful reading. In order to solve this problem, RQA was introduced. The additional coil quality factor can be reduced but not below the limit shown in Fig. 6. It was found that using RQA = 53 k, which results in Q [email protected] kHz = 32 and Q [email protected] kHz = 29, recovers the signal so that the reader decodes the response. Table II summarizes the design for the additional coil and the modification proposed for the reader. At higher DRT distances (DRT > 43 cm), the minimum Q A needed for successfully charging deteriorates the transient response and it was impossible to read the tag. The system without RQR is capable of energizing the tag up to DRT = 46 cm (DRA = 42 cm and DAT = 4 cm); however the tag response is not correctly decoded.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 8

Fig. 8.

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

Measurement setup.

Fig. 10.

Fig. 9. VC L . (a) Modeled (surface) and measured (dots). (b) Cross sections of (a) for four DRA values.

VI. T HREE -C OIL L INK M EASUREMENTS In this section, the measurements of the three-coil system designed in Section V (Fig. 2, Tables I and II) are presented. The tag circuit shown in Fig. 2(g) was implemented using off-the-shelf discrete elements, which allowed us to measure VC L . The read distance of an actual tag, Fig. 2(c), was measured and compared with the model predictions. Fig. 8 shows the measurement setup. Fig. 9 compares VC L for the model and measurements for different distances between the coils (DRA and DAT ). The measurements were done after 50 ms of the charging phase. As can be seen, the analytical model and measurements are in good agreement, considering the simplifications made. The model does not consider the voltage drop at the rectification

Read limits, modeled and measured with actual tag.

diode and cross coupling between reader and tag kRT . In addition, inaccuracies in the value of the coil resonant frequency, quality, and coupling factors directly alter VC L . Fig. 10 shows the regions where VC L > 5 V (charging limit), and the regions where the 1- and 0-b amplitudes are higher than 10 mV p , for different positions of the coils (DRA and DAT ). All the points (DRA and DAT ) that fulfill the charging limit (VC L > 5 V) also verify the limit for reliable detection of both bits (V R S > 10 mV p ). Thus, the read distance is limited by the charging phase. The dashed lines on Fig. 10 highlight the points with constant DRT = DRA + DAT . Using Fig. 10, it is possible to identify all the different (DRA and DAT ) configurations where the tag is read. The maximum read distance of an actual tag for different distances between reader and additional coil, DRA , is also presented in Fig. 10. The point where the highest DRT distance is achieved is DRA = 40 cm and DAT = 3 cm. In summary, these results show that the system was correctly modeled and verify the correctness of the proposed design approach. The dashed-dotted line in Fig. 10 corresponds to the reading distance DRT = 16 cm of the two-coil system. Any point over this line corresponds to a configuration that has larger read distance than the original two-coil RFID system. Therefore, regarding the cattle identification application, the results show that the distance can be significantly increased both by using the additional coil closer to the reader (like an extension of the reader) or closer to the tag (either a modified tag with an extension or a fixed additional coil close to a place where the animals pass by). VII. C ONCLUSION The enhancement of an HDX, FSK, RFID system through the application of an additional resonant coil was modeled, considering both the phases, charging and reading. It was shown that the high values of the additional coil quality factor Q A , in spite of reducing the system bandwidth, make both bit amplitudes received by the reader to increase. Therefore, we demonstrate that the reduction in bandwidth is not a problem from the point of view of attenuation of

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. PÉREZ-NICOLI et al.: BIDIRECTIONAL ANALYSIS AND DESIGN OF RFID USING AN ADDITIONAL RESONANT COIL

the transmission frequencies in the reading phase. However, the transient response is affected by high additional coil quality factor Q A . The results of this paper demonstrate that a tradeoff exists between achieving high steady-state amplitude and reducing intersymbol interference. The results of the model were applied to the design and test of a three-coil RFID. The test system applied a commercial tag (RI-INL-R9QM widely used in cattle identification), a reader (using a TMS3705), and additional resonant coil (whose radius and quality factor were optimized). The designed system increases almost 2.7× the original RFID read distance (from 16 to 43 cm). The analytic design predictions and measurements are in good agreement, thus validating the presented modeling approach. The design procedure can be applied not only to other RFID system but also to other systems that use the same channel to wirelessly energize and transfer data, such as some implantable medical devices.

Fig. 11. tag.

Additional coil model, including the reflected inductance from the

where (wC T .R L /2)2 = (R L /2/w.L T )2 (w/wresT )4  1 was assumed. Finally, from (22) and (23) ηT(C) = Q L (C) QT

A PPENDIX T HREE -C OIL C HARGING P HASE C ALCULATIONS

QT Q T + Q L (C)

(24)

R L /2 wL T = = Re{Z L } w.L T wL T = . RpT



w wresT

4 (25) (26)

B. Additional Coil Efficiency

A. Tag Efficiency In order to simplify the model, we make two approximations. First, we assume that the diode of the half wave rectifier is ideal and therefore does not dissipate. Hence, VC T p = VC L where VC T p is the peak voltage at C T , and VC L is the dc voltage at C L [see Fig. 2(g)]. Second, we approximate the effect of the diode and its load (C L and R L ) by a resistor Rv L , making the overall model linear and, hence, easily tractable. Rv L is determined, so that the mean power dissipated in Rv L (PRv L ) is equal to the power dissipated in R L (PR L ). This results in Rv L = R L /2 as seen in the following equation: PRv L =

VC2T p 2.Rv L

=

VC2T p RL

=

VC2L RL

= PR L .

(21)

The tag efficiency (22) is calculated as the ratio between the power dissipated in R L over the total power dissipated in R L and RpT [see Fig. 2(g)] ηT(C) =

⇒ Re{Z L } = 

res T

R L /2 2

w.R L /2 2 wres LT

 +1

=

2 L L w2 kAT A T 4  2 w res T T) RpT + (w.L + j wL T + 1/jwCT R L /2 . w

RpT + RpT wL T

+

(w.L T )2 R L /2 .

2 L L w2 kAT A T 4 res T + j wL T (1 − (wresT /w)2 ) w

w

2 wL kAT A 4 resT + j (1 − (wresT /w)2 ) w

w

(w.L T ) R L /2 . 2 wL Q kAT A T

Q L (C)

Q T + Q L (C) + j Q T Q L (C) (1 − (wresT /w)2 )

.

(28)

In addition, RQA can be substituted by an equivalent series resistance, in order to do that Z A (Fig. 11) is calculated in a way similar to the one for (23) 1 Z A = RQA // ⇒ jwC A RQA (w.L A )2  wres A 4 Re{Z A } =  .  2 RQA w w.RQA + 1 2 w L

T

w  (w.L T resT 4 . R L /2 w )2

res A

T

1 −R L /2.jwCT .R L /2  2 (wC T .R L /2) + 1 jwCT

Z TAref =

=

1 R L /2 ⇒ = jwCT jwCT .R L /2 + 1 R L /2  ⇒ Re{Z L } = (wC .R 2 T L /2) +1 ⇒ √ 1 = wresT → C T = w2 1 L C .L T

where MAT is the mutual inductance between the tag and the √ reader, MAT = kAT L A L T . Z L was found in (23), so Z TAref is

(22)

Z L = Rv L //

T

In order to calculate η A(C) , the reflected impedance in the additional coil from the tag Z TAref [Figs. 2(f) and 11] should be computed first VL A = (RpA + j wL A )i A + j wMAT i T  −i A j wMAT ⇒ iT = RpT + j wL T + Z L 2 w2 MAT iA (27) VL A = (RpA + j wL A )i A + RpT + j wL T + Z L

=

IT2 Re{Z L } IT2 RpT + IT2 Re{Z L }

where Re{Z L } is simply calculated as [see Fig. 4(c)]

Im{Z L } =

9

(23)

Im{Z A } =

A

1 −RQA .jwC A .RQA  2 (wC A .RQA ) + 1 jwC A

(29)

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 10

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES

where (wC A .RQA )2 =



  RQA 2 w w.L A wres A

4

 1 was assumed.

Thus, the circuit of Fig. 11 can be simplified as is shown in Fig. 4(b), where RpAeq is defined as (w.L A )2  wres A 4 RpAeq = . + RpA . (30) RQA w The normalized reflected impedance in the additional coil from the tag (Z TAref ) is calculated from (28) Z TAref = Z TAref /RpAeq =

Q T + Q L (C) + j Q T Q L (C) (1 − (wresT /w)2 )

wL A QA = . RpAeq

(31)

Finally, the additional coil efficiency η A(C) is calculated [see Fig. 4(b)] as the ratio between the power dissipated in RpAeq over the total power dissipated in RpAeq and Z TAref 1 + Re{Z TAref }

.

(33)

C. Reader Efficiency Z ARref [Fig. 4(a)] can be obtained in a way similar to the one applied for (27) and (28) 2 L L  w2 kRA R A Z ARref = R +Z 2 pAeq TAref + j wL A (1−(wres A /w) ) ⇒ Z ARref = Z ARref /R p R Z ARref = QR =

2 Q Q kRA A R

1 + Z TAref + j Q A (1 − (wres A /w)2 )

wL R . RpR

(34) (35)

Now, the effect of RQR is analyzed by calculating Z eqR [Fig. 4(a)]. (R p R + Z ARref ) is referred as Z ∗ . The imaginary part of Z ∗ is assumed to be much lower than j wL R . This is usually the case based on operation close to the resonance of the additional coil and low coupling factor kRA , see (34) (Z ∗ + L R j w)RQR Z ∗ + RQR + L R j w (Re{Z ∗ }+ L R j w)RQR  ⇒ Re{Z ∗ }+ RQR + L R j w

Z eqR =

Re{Z eqR } =

RQR [Re{Z ∗ }(Re{Z ∗ } + RQR ) + (L R w)2 ] (Re{Z ∗ } + RQR )2 + (L R w)2

 Re{Z ∗ } + Im{Z eq } =

Re{Z ARref } RpR +

(L R w)2 RQR

+ Rout R + Re{Z ARref }

Re{Z ARref } 1+

RpR Q 2R RQR

+

Rout R RpR

+ Re{Z ARref }

.

(37)

ACKNOWLEDGMENT (32)

Re{Z TAref }

η R(C) = =

2 Q Q Q kAT T A L (C)

η A(C) =

fundamentally altered. As a conclusion, RQR increases the series resistance of L R without altering the imaginary part. Then, the reader efficiency η R(C) is obtained as the power dissipated in Z ARref divided by the total power dissipated in w)2 R p R + (LRRQR , Rout R , and Z ARref

(L R w)2 RQR

RQR L R w(2Re{Z ∗ } + RQR ) j  L R j w. (36) (Re{Z ∗ } + RQR )2 + (L R w)2

It was assumed that RQR  Re{Z ∗ } and (RQR )2  (wL R )2 . These assumptions usually hold for typical designs as the one under study in this paper where high RQR values are used. If this was not the case, the analysis presented is not

The authors would like to thank Prof. J. P. Oliver for valuable discussions and providing radio frequency identification tags, reader, and CST AG for providing the simulation software. R EFERENCES [1] M. Abbak and I. Tekin, “RFID coverage extension using microstrippatch antenna array [wireless corner],” IEEE Antennas Propag. Mag., vol. 51, no. 1, pp. 185–191, Feb. 2009. [2] A. J. S. Boaventura and N. Carvalho, “Extending reading range of commercial RFID readers,” IEEE Trans. Microw. Theory Techn., vol. 61, no. 1, pp. 633–640, Jan. 2013. [3] Radio Frequency Identification of Animals—Technical Concept, ISO/IEC doc. ISO 11785, 1996. [4] A. Kurs, A. Karalis, R. Moffatt, J. D. Joannopoulos, P. Fisher, and M. Soljaˇci´c, “Wireless power transfer via strongly coupled magnetic resonances,” Science, vol. 317, no. 5834, pp. 83–86, Jul. 2007. [5] K. Na, H. Jang, H. Ma, and F. Bien, “Tracking optimal efficiency of magnetic resonance wireless power transfer system for biomedical capsule endoscopy,” IEEE Trans. Microw. Theory Techn., vol. 63, no. 1, pp. 295–304, Jan. 2015. [6] M. Kiani, U.-M. Jow, and M. Ghovanloo, “Design and optimization of a 3-coil inductive link for efficient wireless power transmission,” IEEE Trans. Biomed. Circuits Syst., vol. 5, no. 6, pp. 579–591, Dec. 2011. [7] J. Kim, D. H. Kim, J. Choi, K. H. Kim, and Y. J. Park, “Free-positioning wireless charging system for small electronic devices using a bowlshaped transmitting coil,” IEEE Trans. Microw. Theory Techn., vol. 63, no. 3, pp. 791–800, Mar. 2015. [8] F. Zhang, S. A. Hackworth, W. Fu, C. Li, Z. Mao, and M. Sun, “Relay effect of wireless power transfer using strongly coupled magnetic resonances,” IEEE Trans. Magn., vol. 47, no. 5, pp. 1478–1481, May 2011. [9] S. V. Georgakopoulos and O. Jonah, “Optimized wireless power transfer to RFID sensors via magnetic resonance,” in Proc. IEEE Int. Symp. Antennas Propag., Jul. 2011, pp. 1421–1424. [10] O. Jonah, S. V. Georgakopoulos, and S. Yao, “Strongly coupled resonance magnetic for RFID applications,” in Proc. IEEE Antennas Propag. Soc. Int. Symp., Jul. 2013, pp. 1110–1111. [11] R. Gonalves, N. B. Carvalho, and P. Pinho, “Increasing the RFID readability range using wireless power transmission enhancements,” in Proc. Wireless Power Transf. (WPT), May 2013, pp. 135–138. [12] P. Perez-Nicoli and F. Silveira, “Matching networks for maximum efficiency in two and three coil wireless power transfer systems,” in Proc. IEEE 7th Latin Amer. Symp. Circuits Syst. (LASCAS), Feb./Mar. 2016, pp. 215–218. [13] “International agreement of recording practices,” Int. Committee Animal Rec. (ICAR), Rome, Italy, Tech. Rep., Jun. 2008, pp. 433–434. [14] M. Kiani and M. Ghovanloo, “The circuit theory behind coupledmode magnetic resonance-based wireless power transmission,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 59, no. 9, pp. 2065–2074, Sep. 2012. [15] M.-L. Kung and K.-H. Lin, “Enhanced analysis and design method of dual-band coil module for near-field wireless power transfer systems,” IEEE Trans. Microw. Theory Techn., vol. 63, no. 3, pp. 821–832, Mar. 2015.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. PÉREZ-NICOLI et al.: BIDIRECTIONAL ANALYSIS AND DESIGN OF RFID USING AN ADDITIONAL RESONANT COIL

Pablo Pérez-Nicoli (S’13–M’13) received the Electrical Engineering degree from the Universidad de la República, Montevideo, Uruguay, in 2013, where he is currently pursuing the Ph.D. degree with the Electrical Engineering Department. He joined the Electrical Engineering Department, Universidad de la República, in 2012, where he is currently a Research Assistant. His current research interests include ultralow-power analog integrated circuits design, dc–dc converters, and wireless energy transmission. Agustín Rodríguez-Esteva (GSM’16) received the Electrical Engineering degree from the Universidad de la República, Montevideo, Uruguay, in 2015. He joined the Electrical Engineering Department, Universidad de la República, in 2015, where he is currently a Research Assistant. His current research interests include wireless energy transmission, lowpower embedded systems design, and industrial control applications.

11

Fernando Silveira (S’88–M’90–SM’03) received the Electrical Engineering degree from the Universidad de la República, Montevideo, Uruguay, in 1990, and the M.Sc. and Ph.D. degrees in microelectronics from the Universitè catholique de Louvain, Louvainla-Neuve, Belgium, in 1995 and 2002, respectively. He is currently a Professor with the Electrical Engineering Department, Universidad de la República. He has had multiple industrial activities with CCC Medical Devices and NanoWattICs, including leading the design of an applicationspecified integrated circuit for implantable pacemakers and designing analog circuit modules for implantable devices for various companies worldwide. He has coauthored one book and many technical articles in his research field. His current research interests include design of ultralow-power analog and RF integrated circuits and systems, in particular in biomedical application.

Digital Object Identifier 10.1109/TMTT.2016.2584758

Editors-in-Chief Dominique Schreurs c/o Mrs. Enas Kandil, Editorial Assistant KU Leuven, Div. ESAT-TELEMIC Kasteelpark Arenberg 10 B-3000 Leuven Belgium E-mail: [email protected] or [email protected]

Jenshan Lin c/o Mrs. Marcia Hensley, Editorial Assistant University of Florida 1064 Center Drive, NEB 559 Gainesville, FL 32611 USA E-mail: [email protected] or [email protected]

Information for Authors The IEEE T RANSACTIONS ON M ICROWAVE T HEORY AND T ECHNIQUES focuses on that part of engineering and theory associated with microwave/millimeter-wave components, devices, circuits, and systems involving the generation, modulation, demodulation, control, transmission, and detection of microwave signals. This includes scientific, technical, and industrial, activities. Microwave theory and techniques relates to electromagnetic waves usually in the frequency region between a few MHz and a THz; other spectral regions and wave types are included within the scope of the Society whenever basic microwave theory and techniques can yield useful results. Generally, this occurs in the theory of wave propagation in structures with dimensions comparable to a wavelength, and in the related techniques for analysis and design. I. Paper Submission in Electronic Form Authors need to visit the website http://www.mtt.org/transactions/34-author-information-transactions.html for the author instructions. To reduce time from submission to publication of papers, the editorial office accepts manuscripts only in electronic form as .pdf files and all communications with authors will be via email. The files must not be larger than 1MB and no *.zip files are accepted. Submissions should be submitted through the ScholarOne Manuscripts site at: http://mc.manuscriptcentral.com/tmtt-ieee and use the templates provided under http://www.ieee.org/publications standards/publications/authors/authors journals.html (Template for all Transactions (except IEEE Transactions on Magnetics), two-column template; can also be requested from the editorial office). Figures, graphs and all other necessary information for reviewing the manuscript must be included in this file (as opposed to being attached to it as separate files) and placed at appropriate locations within the text rather than at the end: • The abstract must be self-contained, without abbreviations, footnotes, or references. It should be no more than 250 words. It must be written as one paragraph, and should not contain displayed mathematical equations or tabular material. • IEEE supports the publication of author names in the native language alongside the English versions of the names in the author list of an article. For more information, please visit the IEEE Author Digital Tool Box at: http://www.ieee.org/publications standards/publications/authors/auth names native lang.pdf • Figures should be large enough to be easily readable on a computer screen and on paper when printed out. • A photograph of any component or circuit presented must be included. • If, at the decision of the Editor, the component or circuit can be fabricated, measured characteristics must be included. • All papers with theoretical contributions must have independent verification with measurement-based validation strongly preferred. • Instrument screen captures are not suitable for publication and the data should be replotted. • The print version of the paper will be in black and white, but color figures may be used in the electronic version of the paper. • Axes should be labeled with large lettering. • Whenever possible, theory and corresponding experimental results should be printed on the same graph for easy comparison. • Follow the Guidelines for Author-Supplied Electronic Text and Graphics available for download at the above website. • The minimum paper length is 4 pages, excluding the authors’ photos and biographies. Short papers of three pages or less should be sent to the IEEE M ICROWAVE AND W IRELESS C OMPONENTS L ETTERS. The font size is specified in the templates. TeX and LaTeX users must use scalable fonts rather than bitmapped fonts to allow easy reading of .pdf files on the computer screen. • This publication accepts graphical abstracts, which must be peer reviewed. For more information about graphical abstracts and their specifications, please visit: http://www.ieee.org/publications standards/publications/graphical abstract.pdf • ORCID Required: All IEEE journals require an Open Researcher and Contributor ID (ORCID) for all authors. To create an ORCID, please visit: https://orcid.org/oauth/signin. The author will need a registered ORCID in order to submit a manuscript or review a proof in this journal. Note: Manuscripts that are related to material submitted to or published at conferences are considered only if the content is significantly updated or contains material of substantially complementary nature. Authors must reference all of their previous papers that are similar. Please attach .pdf files of previous papers and clearly state (on a separate page) the difference with respect to the current submission. Failure to disclose prior papers by the authors that are similar will be rejected. II. Final Submission Format After a manuscript has been accepted for publication, the author will be requested to provide an electronic copy of the final version of the manuscript in pdf format; Microsoft Word is the preferred format for this final submission, although TEX and LATEX formats are also acceptable. Note: Although we require a .pdf file of the manuscript for the review process, this format is not acceptable (neither is .ps) for the final submission. Some additional guidelines must, however, be followed for the submission of the final manuscript in electronic form: • Include all macros (/def) that are required to produce your manuscript (TEX and LATEX). • IEEE Transaction/Journal style dictates a 21-pica (3.5 inch) column width. If mathematical expressions are produced with this in mind, they are more aesthetically pleasing in the final version. • Figures and tables must be submitted as separate files in .ps, .eps, .doc or .tiff format III. Open Access This publication is a hybrid journal, allowing either Traditional manuscript submission or Open Access (author-pays OA) manuscrip�submission. Upon submission, if you choose to have your manuscript be an Open Access article, you commit to pay the discounted $1,750 OA fee if your manuscript is accepted for publication in order to enable unrestricted public access. If you would like your manuscript to be a Traditional submission, your article will be available to qualified subscribers and purchasers via IEEE Xplore. No OA payment is required for Traditional submission. IV. Page Charges for Traditional Submissions Papers will be reviewed for their technical merit, and decisions to publish will be made independently of an author’s ability to pay page charges. Page charges of $110 (U.S.) per printed page will be requested on papers of seven printed pages or less. Overlength page charges of $200 per page are mandatory for each page in excess of seven pages. This is effective for any paper published after August 1, 2014 onward. If the author’s organization agrees to honor the total page charge, which includes the page charges on the first seven pages plus the mandatory overlength charge, the author will receive 100 reprints. If the supporting organization honors only the mandatory charge, no free reprints will be sent. Digital Object Identifier 10.1109/TMTT.2016.2576180