WIRELESSPERSONAL COMMUNICATIONS

BLUETOOTH ANDOTHER TECHNOLOGIES

THE KLUWER INTERNATIONAL SERIESIN ENGINEERING AND COMPUTER SCIENCE

WIRELESSPERSONAL COMMUNICATIONS

BLUETOOTH ANDOTHER TECHNOLOGIES

edited by

William H. TranterBrian D. WoernerJeffrey H. Reed

Theodore S. RappaportMax Robert

Virginia Polytechnic Institute & State University

KLUWER ACADEMIC PUBLISHERS NEW YORK, BOSTON, DORDRECHT, LONDON, MOSCOW

0-306-46986-30-792-37214-X

TABLE OF CONTENTS

PREFACE ix

I. FRONTIERS IN PROPAGATION

1.

2.

3.

4.

5.

6.

Statistics of the Temporal Variations in the Wireless TransmissionChannel in Indoor EnvironmentMarin Stoytchev and Hugo Safar

UHF-Radio Propagation Predictor for Temporal Variations InPopulated Indoor EnvironmentsF. Villanese, W. G. Scanlon, and N. E. Evans

An Improved Approach for Performance Evaluation of the Downlink ofDS-CDMA PCS Indoor Systems with Distributed AntennasM.R. Hueda, C. E. Rodriguet, and C. A. Marques

Fast and Enhanced Ray Optical Propagation Modeling for RadioNetwork Planning in Urban and Indoor ScenariosR. Hoppe, P. Wertz, G. Wölfle, and F. M. Landstorfer

Indoor Propagation Analysis Techniques for Characterisation ofUltra-Wideband RF EnvironmentsDavid J. Hall

Propagation Signatures to Characterize Wideband EnvironmentsGregory Martin

1

11

23

35

45

57

67

77

89

II. SPATIAL PROCESSING

7.

8.

9.

Smart Antennas for CDMA Cellular and PCS NetworksScot Gordon, Marty Feuerstein, Donn Harvey, and Michael Zhao

Key Techniques Realizing Smart Antenna Hardware for MicrocellCommunication SystemsKeizo Cho, Kentaro Nishimori, Yasushi Takatori, and Toshikazu Hori

Downlink Capacity Enhancement in GSM System Using MultipleBeam Smart Antenna and SWR ImplementationWei Wang, Mohamed Ahmed, Samy Mahmoud, and Roshdy H. M. Hafez

vi

10.

11.

12.

13.

14.

Generalized Equations for Spatial Correlation for Low to ModerateAngle SpreadR. Michael Buehrer

Exploitation of Internode MlMO Channel Diversity inSpatially-Distributed Multipoint Communication NetworksBrian G. Agee

Design of 16-QAM Space-Time Codes for Rapid Rayleigh Fading ChannelsSalam A. Zummo and Saud A. Al-Semari

Transmit Diversity With More Than Two AntennasR. Michael Buehrer, Robert A. Soni, and Quinn Li

Reduced Complexity Space-Time Optimum ProcessingJens Jelitto, Marcus Bronzel, and Gerhard Fettweis

101

109

121

131

145

III. NETWORK SYSTEM DESIGN

15.

16.

17.

18.

19.

Wireless Personal Communications System Planning UsingCombinatorial OptimisationJoseph K. L. Wong, Michael J. Neve, and Kevin W. Sowerby

Frequency Planning and Adjacent Channel Interference in aDSSS Wireless Local Area Network (WLAN)D. Leskaroski and W. B. Mikhael

Modeling and Simulation of Wireless Packet Erasure ChannelsGünther Liebl, Thomas Stockhammer, and Frank Burkert

Reducing Handover Probability Through Mobile PositioningStamatis Kourtis and Rahim Tafazolli

Multi-user Detection Using the Iteration Algorithm inFast-Fading ChannelsSun-Jin Yeom and Yong-Wan Park

157

169

181

193

205

vii

IV. NEXT GENERATION AND BEYOND

20.

21.

22.

FPGA DSP for Wireless CommunicationChris Dick and fred harris

Signal Processing Requirements of the TDD TerminalStamatis Kourtis, Patrick McAndrew, and Phil Tottle

Frame Quality-Based Versus Forward Power ControlMethods for the cdma2000 Third Generation StandardSteven P. Nicoloso, Mike Mettke, and R. Michael Buehrer

217

229

239

V. BLUETOOTH: A SHORT TUTORIALMax Robert

Introduction

Bluetooth Overview

Technical Overview

Critical Perspective

Conclusion

AppendicesA – Packet Format

B – Packet Description

INDEX

249

249

250

264

265

266

267

271

PREFACE

The papers presented in this book were originally presented at the 10th Virginia Tech/MPRGSymposium on Wireless Personal Communications, which was held on the Virginia Tech campusJune 14-16, 2000. This year’s Symposium was sponsored by Virginia Tech’s Mobile and PortableRadio Research Group (MPRG), the Virginia Tech Division of Continuing Education, and theMPRG Industrial Affiliates Program. The IEEE Virginia Mountain Section and the VirginiaTech Joint Student Chapter of the IEEE Communications and Vehicular Technology Societiesprovided technical co-sponsorship.

Much of the success of our annual symposium, as well as the success of MPRG’s research andeducation program, are directly due to the support of our industrial affiliates. The support that isprovided by the industrial affiliates program allows MPRG to serve the wireless communitythrough research, education, and outreach activities. MPRG’s industrial affiliates include thefollowing organizations: Analog Devices, Inc., Anaren Microwave, Inc., the Army ResearchOffice, AT&T Corporation, BAE Systems, BellSouth Cellular Corporation, Comcast CellularCommunications, Inc., Datum, Inc., Ericsson, Inc., Grayson Wireless, Hughes ElectronicsCorporation, ITT Industries, LGIC, Inc., Lucent Technologies, Inc., Motorola, Inc., Nokia, Inc.,Nortel Networks, Qualcomm, Inc., Raytheon Systems Company, Samsung Advanced Institute ofTechnology, Southwestern Bell, Tantivy Communications, Inc., Tektronix, Inc., TelcordiaTechnologies, Texas Instruments, and Wavtrace, Inc.

In 1999, the Wireless Symposium was expanded to include a tutorial course. This activity provedpopular and was continued in 2000. This year’s tutorial was targeted on the emerging technologyknown as Bluetooth. A summary of the tutorial notes follow and is included as the last section ofthis book.

Twenty-two papers were presented at the 2000 symposium and are divided into five groups. Thefirst group of six papers deals with propagation, both indoor and outdoor. The first paper,Statistics of the Temporal Variations in the Wireless Transmission Channel in IndoorEnvironments by Martin Stoytchev and Hugo Safar, examines the temporal behavior of an indoorchannel at 2.8 GHz. They find that the received power is well defined by a Ricean model andthat an unexpected increase in the received SNR occurs when people are present in the building.The second paper, UHF-Radio Propagation Predictor for Temporal Variations in PopulatedIndoor Environments by F. Villanese, W. G. Scanlon, and N. E. Evans, implements animprovement in site-specific ray-tracing models by including the presence of moving humanbodies in the model. They found large temporal variations due to human body movement. Ifantennas are close to the body, such as one finds in Bluetooth applications, a reduction inreceived power occurs due to antenna-body interaction. The third paper, An Improved Approachfor Performance Evaluation of the Downlink of DS-CDMA PCS Indoor Systems with DistributedAntennas by M. R. Hueda, C. E. Rodriguez, and C. A. Marques, examines the performance of theforward link of a DS-CDMA system having distributed antennas in an indoor channelenvironment. They show the inadequacies of the Gaussian approximation for interference andpropose a more accurate model to describe the effect of IPI (interpath interference) on the biterror rate and the frame error rate. The fourth paper in the propagation group, Fast and EnhancedRay Optical Propagation Modeling for Radio Network Planning in Urban and Indoor Scenariosby R. Hoppe, P. Wertz, G. Wolfle, and F. M. Landstorfer, describe a computationally efficient rayoptical model. The model is implemented for both indoor and urban environments, and themodel performance is compared against measurements. The next contribution, IndoorPropagation Analysis Techniques for Characterization of Ultra-Wideband RF Environments byDavid J. Hall, presents a number of statistics for ultra-wideband RF environments. The path loss

and the delay spread are characterized, and the performance gain as a function of the number ofcorrelators, is discussed. The last paper, Propagation Signatures to Characterize WidebandEnvironments by Gregory Martin, introduces the concept of the propagation signature forsummarizing the large quantity of data collected in a wideband propagation measurementprogram.

The second set of papers deals with the techniques and applications of spatial processing. ScotGordon, Many Feuerstein, Donn Harvey, and Michael Zhao contribute the first of these papers,Smart Antennas for CDMA Cellular and PCS Networks. The authors point out how a smartantenna for CDMA equalizes loading across sectors, which improves sectorization efficiency bybalancing traffic load and decreasing handoff overhead. The following paper, Key TechniquesRealizing Smart Antenna Hardware for Microcell Communication Systems by Keizo Cho,Kentaro Nishimori, Yasuhi Takatori, and Toshikazu Hori, summarizes techniques for selectingthe number of antenna elements, simplifying hardware, reducing cost, and establishing systemcalibration. The next paper, Downlink Capacity Enhancement in GSM System Using MultipleBeam Smart Antenna and SWR Implementation by Wei Wang, Mohamed Ahmed, SamyMahmoud, and Roshdy H. M. Hafez, demonstrates that smart antennas can significantly increasethe network capacity of a CDMA cellular system. This paper is followed by a paper contributedby R. Michael Buehrer entitled Generalized Equations for Spatial Correlation for Low toModerate Angle Spread. The focus of this paper is the development of generalized expressionsapproximating the spatial correlation for three angular wavefront distributions. The next paper inthis section, Exploitation of Internode MIMO Channel Diversity in Spatially-DistributedMultipoint Communication Networks by Brian G. Agee, demonstrates that multiple-input,multiple-output (MIMO) networks can provide significant improvements in capacity over point-to-point links. This paper is followed by Design of 16-QAM Space-Time Codes in Rapid RayleighFading Channels by Salam A. Zummo and Saud A. Al-Semari, which explores the use of space-time codes, based on a 16-QAM signal set, for performance enhancement in a fast fadingenvironment. Two different code design techniques are provided, trellis encoding and I-Qencoding. Coding gains of approximately 3dB are achieved. A new diversity technique, referredto as space-time spreading, is investigated in the paper Transmit Diversity With More Than TwoAntennas by R. Michael Buehrer, Robert A. Soni, and Quinn Li. Their results show that potentialbenefits are achieved by increasing the number of antennas at the mobile station. The last paperin this group, Reduced Complexity Space-Time Optimum Processing by Jens Jelitto, MarcusBronzel, and Gerhard Fettweis, analyzes spatial correlation properties and investigates the signalprocessing requirements for designing space-time optimum receivers.

The next set of contributions, consisting of five papers, treats system design and networkingissues. The first of these papers, Wireless Personal Communications System Planning UsingCombinatorial Optimization, by Joseph K. L. Wong, Michael J. Neve, and Kevin W. Sowerby,considers combinatorial optimization techniques for system design. A simple test problem,consisting of system deployment on a single floor of a building, is presented. The suggestedoptimization technique produces useful results. The next paper, Frequency Planning andAdjacent Channel Interference in a DSSS Wireless Local Area Network (LAN) by D. Leskaroskiand W. B. Mikhael, illustrates a theoretical model of 802.11 DS channelization given thatchannels are arranged in a staggered and overlapped fashion. Emphasis is on the 2.4 GHz ISMfrequency band. The proposed technique results in improved capacity while maintaining minimalinterference. The third paper in this group, Modeling and Simulation of Wireless Packet ErasureChannels by Gunther Liebl, Thomas Stockhammer, and Frank Burkert, presents a model forwireless packet erasure channels. The model can be used for both theoretical analysis and forreal-time simulation of network protocol performance. The fourth paper, Reducing HandoverProbability Through Mobile Positioning by Stamatis Kourtis and Rahim Tafazolli, examines the

x

xi

mechanism by which unnecessary handovers occur and proposes a technique for combatingunnecessary handovers. The final paper in this group is by Sun-Jin Yeom and Yong-Wan Parkand is entitled Multi-user Detection Using the Iteration Algorithm in Fast-Fading Channels. Thiscontribution introduces a new interference cancellation scheme for multi-user DS/CDMA. Theyshow improved performance with reduced complexity.

The last group of papers consists of three contributions. These treat a variety of implementation,signal processing, and third-generation issues. The fust of these papers, FPGA DSP for WirelessCommunication by Chris Dick and fred hams, focuses on reconfigurable software defined radiosdeveloped using FPGA architectures. Particular attention is paid to the implementation of digitalfilters, including multi-rate filters and carrier recovery loops. The second paper in this group,contributed by Stamatis Kourtis, Patrick McAndrew, and Phil Tottle, is titled Signal ProcessingRequirements for the TDD Terminal. The authors of this paper consider technology requirementsfor the 30 PP-TDD terminal. The architecture is described and complexity issues are discussed.The last paper in this group, Frame Quality-Based Versus Eb/No-Based Forward Power ControlMethods for the cdma2000 Third Generation Standard by Steven P. Nicoloso, Mike Mettke, andR. Michael Buehrer, shows that forward loop power control provides substantial capacity gains atlow speeds where forward error control provides only moderate gains. They also consider thedifficult problem of time correlated shadowing.

The final section of this volume consists of a brief tutorial on Bluetooth, a standard defining aninfrastructure for wireless devices with a very short operating range. The tutorial discusses theapplication potential of Bluetooth and the general environment in which Bluetooth is designed tooperate. The protocol stack, the software interface, the RF specification, and Bluetooth’sinteroperability with other communication standards is also discussed.

A successful symposium, and consequently the papers contained herein, result from thesignificant efforts of a dedicated team of people. Our first thanks go to those who submittedpapers and those who attended the symposium. Without a strong technical program, thesymposium could not continue to prosper. The MPRG staff and graduate students also deservespecial thanks. The efforts of Jenny Frank, who year-after-year takes the lead for organizing thesymposium, keeping the faculty on schedule, and tending to the vast quantity of details associatedwith the symposium, are gratefully appreciated.

The editors also wish to thank Jennifer Evans of Kluwer Academic Publishers for her support inbringing this book to press.

Blacksburg, VA William H. TranterBrian D. WoernerJeffrey H. ReedTheodore S. RappaportMax Robert

Statistics of the Temporal Variations in the Wireless TransmissionChannel in Indoor Environments

Marin Stoytchev and Hugo SafarBell Labs, Lucent Technologies, 700 Mountain Avenue, Murray Hill, NJ 07974

AbstractThe temporal behavior of a single transmission channel inside buildings is

studied in a series of CW experiments at a frequency of 2.8 GHz. Measurements of thereceived power show that its fluctuations are well described by a Ricean distribution.Measurements of the electric field, however, show a more complex picture that does notlie within the limits of the Recean model. The joint probability distribution of the real andimaginary parts of the transmission coefficient shows numerous well-distinguished peaksassociated with quasi-static transmission channels that dominate the signal over differenttime intervals. Nevertheless, we find that the phasor associated with the transmitted fieldremains confined in a particular region of the complex plane preserving the character ofthe speckle pattern.

I. Introduction

The temporal behavior of the wireless transmission has long been a subject ofintensive studies. The statistics of the received power in a dynamically changingenvironment in many cases has been found to be well described by the Ricean model. Tothe best of our knowledge, however, these studies have been mostly focused oninvestigating the power associated with the transmitted field and not the field itself. Inefforts to increase the information capacity using multiple-antenna arrays have beensuggested to exploit the spatial diversity of the richly scattering environment. Recenttheoretical studies suggest a linear increase in capacity with the number of transmit andreceive antenna elements in the case of Rayleigh scattering environment [1,2], whichimposes upper bounds on the transmission bit rate per unit bandwidth. Subsequently, thepossibility of increased information capacity was demonstrated experimentally in an

2

indoor propagation environment by employing layered space-time architecture, known asBLAST (Bell Laboratories Layered Space-Time) [3,4].

These studies have brought the focus on the complex electric field as thefundamental quantity of interest. In the case of n transmit and m receive antennas thetransmission is characterized by the complex transmission matrix H, which is of order n xm and relates the m-component receive vector to the n-componenttransmit vector The i-th component of the received field vector is givenby where denote the field radiated from the j-th transmit antenna and isthe complex transmission coefficient between the two antennas. In this case theinformation capacity is given by where is a unitary

matrix of rank m and is the average signal-to-noise-ratio (SNR) for an individualchannel. Clearly, in order to estimate the capacity, knowledge of the individual matrixelements (magnitude and phase of the transmitted electric field between each pair oftransmit and receive antennas) is required.

In this paper we present experimental studies of the temporal behavior of a singletransmit wireless channel inside buildings. Measurements of the received power and ofthe received electric field were performed in metal-wall and in brick-wall environments.The experiments were carried out during different time of the day, when no people werepresent in the building and with people present. Using these measurements we study thestatistics of both the transmitted power and the electric field detected. We find that thestatistics of long-term variations of the received power is well described by a Riceanmodel. Measurements of the field, however, reveal a more complex behavior that cannotbe described by such a model.

II. Experiment

We have made CW measurements of the microwave field and its intensity/powertransmitted through a single channel in buildings with metal and with brick walls. Thefrequency of radiation is 2.8 GHz. In these measurements we use identical ground-planequarter-wavelength monopoles as transmit and receive antennas (TA and RA,

3

respectively). The antennas were placed so as to have parallel polarization. We firstperformed power measurements using a HP 83732B signal synthesizer as a source ofradiation and a HP E418 power meter with HP 8481A/D power sensors for detection.The values of the signal power observed in these measurements are well within thedynamic range, 10 pW - 300 mW, provided by the equipment. For some of the locations,independent measurements of the transmitted field have been made using a HP8722Dvector network analyzer. The experimental setup of the measurements performed withthe vector network analyzer are presented in Fig. 1.

In order to insure an adequate dynamic range of the measurements, we use a lownoise RF amplifier at the transmit end, and, in cases of separation between TA and RAgreater than 10 m, we use an identical amplifier at the receive end as well. Low-losscables with attenuation of 0.07 dB/ft are used to connect the antennas to the signalgenerating and measuring equipment. Regular measurements of the received signal withthe transmit end terminated, which gives the noise level in the local environment, wereperformed to give an average SNR of 30 dB in the power measurements and ofapproximately 60 dB in the measurements of the electric field. In the powermeasurements, the output power has been monitored by recording a reference signal

4

using a directional coupler connected to the output of the amplifier on the transmit end.No fluctuations greater than 0.1 % of the output power were observed. All measurementshave been carried out in the main building of Lucent Technologies, Bell Labs at MurrayHill, NJ. To characterize the propagation channel in different environments, we haveperformed measurements at two different locations of the building, one of which hasentirely metal walls (location A) and the other has brick walls (location B). Schematicrepresentation of location A is given in Fig. 1. The positions of the TA and RA aredenoted by full circles. All measurements are made with constantly open doors of therooms, in which the antennas were located. Both, the transmit and the receive antennasare placed at approximately 150 cm above the ground and in all cases no direct line ofsight exists. To insure that the measurements provide an adequate picture of the changesin the transmission channel we use two identical receive antennas, which are separated bya distance of 20 cm. Their positions are chosen so that they appear, correspondingly, in alocal maximum (“bright” spot) and a local minimum (“dark” spot) of the speckle patterndetermined by the geometry of the building in the absence of people. For simplicity, inthe text that follows we will use the notations RA1 and RA2 to denote the antenna locatedat a spatial maximum and at a minimum of the electric field, respectively. The two RAare connected to a two-input switch that allows independent sampling of the antennas.

III. Results and discussions

III. 1. Power measurements

First, we concentrate on the temporal variations of the transmitted power. Thepower of the received signal is monitored in 24-hour periods. Data is taken every twohours for duration of eight minutes for each antenna individually. We measure the signalusing RA1 first and then RA2 for another eight-minute interval. Typical time traces ofthe received power during day and night are shown in Fig. 2. In the absence of people thechannel is completely stationary within the limits of instrumental response. Daytimemeasurements reveal considerably large rapid temporal fluctuations, which are associated

RA2 5

with the activities of people present in the building. The characteristic time, in which thechannel changes appreciably, is estimated to be approximately 0.1 ms.

We would like to note also the long-term variation of the average received power.

Surprisingly, we find that in both, metal- and brick-wall, environments the power levelduring daytime is on average greater than the power level at night. In some instances,this increase was found to be as large as 10 dB. We emphasize that such an increase inthe average power is observed for both RA1 and RA2 simultaneously. This observationexcludes a simple explanation considering shadowing effects due to a change in the localconfiguration of scatterers that could affect the signal received at a particular point inspace. We associate the increase of the average power level with the presence of peoplein the building, who play the role of scattering objects. In the scattering environment ofthe building with no people present, only certain number of paths that are determined bythe geometry of the local environment reach the point of detection. The presence ofscatterers increases the probability of having more partial waves reaching the point ofdetection, including waves that originally would have escaped without contributing to the

6

received signal. Taking into account that there is no direct line of sight in thesemeasurements, it is then natural to expect that increasing the number of scatterers wouldlead to higher energy levels detected. Studies at wavelengths comparable to thewavelength of radiation used in our measurements show that, in general, the human bodyhas a reflection coefficient of more than 0.5 [5], which supports the plausibility of thescenario suggested above.

The power measurements performed allow us to study not only the behavior ofthe average power, but to obtain also the full distribution functions of the instantaneousreceived power. In Fig. 3 we present the probability density and the cumulativedistribution function of the relative power P/<P> with respect to the ensemble averagevalue <P> obtained from measurements carried at location A. The results show that, inthe absence of people, the probability density is delta-like, as expected for a stationarytransmission channel. When people are present in the building, we find that the powerstatistics is well described by a Ricean model with a K-factor of approximately 10. Thissuggests that the presence of people give relatively small random contributions to thepredominant static phasor determined by the geometry of the local environment. We notethat the statistics of relative power obtained from measurements at all locations is foundto be consistent with the results presented in Fig. 3.

7

III.2. Field measurements

The power measurements presented in the previous section provide a simple wayof studying the characteristics of the transmission channel in a multiply scatteringenvironment. However, such measurements do not provide complete information aboutthe transmitted field. Since the information capacity is determined by the productinvolving products of the complex transmission coefficients, it is necessary to measureboth the magnitude and the phase of the transmitted field. For an adequatecharacterization of the transmission channel, we have performed independent fieldmeasurements. In these measurements, we measure the received field alternating RA1and RA2, as data is taken with one antenna for a time interval of 8 s with a time delaybetween points of 10 ms and the same sequence is repeated with the second antenna. Thepositions of the two receive antennas are again chosen so that one of them appears in abright spot of the speckle pattern and the other one in a dark spot, respectively.

From the measured field, we calculate the joint probability density functionof the real and imaginary parts of the complex transmission coefficient,

and where is the complex field measured at the receiveantenna and is a correction factor, which is obtained by means of a through calibrationand accounts for the instrumental response of the system. As expected, the jointprobability density in the absence of people is a sharp delta-like function in the two-dimensional complex plane. However, when people are present in the building thebehavior of the detected field appears more complex than what is assumed in a Riceanmodel. In Fig. 4, we present the results obtained from measurements over a three-hourperiod during daytime. The distributions appear to be comprised of various number ofwell defined peaks. Such picture is not consistent with the Ricean statistical model, whichassumes that the channel has a single static component and a random component due tomoving objects. The results of the measurements presented here could rather beassociated with a time-varying modification of the Ricean model. This is related to thepattern, in which the local environment changes. To simplify the physical picture, we canassume that the dominant static channel is determined by the position and orientation ofall large objects in the vicinity of the transmit and the receive antennas. It is clear that, for

8

the most part, these objects do not change their position in a completely random patternin space and time. As a simple example, one can consider a person working in thevicinity of either of the antennas for certain duration of time and then changing his/herlocation/position (moving to another part of the room, or leaving it). Thus, for theseparticular time spans one would observe different static phasors associated with wavestraversing different paths that reach the point of detection due to the differentsurrounding. This would give a picture consistent with the one observed here.

A more illustrative picture of the evolution of the complex phasor associated withthe transmitted field is given by the contour maps of the distributions as shown in Fig. 5.There are several distinct features that we would like to draw attention to. First, there aredistinct “focal” points that indicate the position of the peaks in the distribution function,which could be associated with changes in the static channel in support with thesuggestion made above. It is worth noting that the focal points are visibly shifted awayfrom the origin, which indicates well-defined non-zero static component even for the“dark” spot. Second, and more importantly in our opinion, the complex field at both localmaxima and local minima is confined within finite regions of the complex plane, which

9

do not overlap. This indicates that, although the presence of people in the building affectsthe spatial field distribution, the speckle pattern is not changed in a fundamental way, i. e.the “bright” and “dark” spots remain “bright” and “dark”, respectively. In addition to this,we find that the average magnitude of the field observed in both “bright” and “dark”spots, when people are present, is greater than that in the absence of people (datasurrounded by small rectangles in Fig. 4), which is consistent with the results of ourpower measurements. Close examination of all data acquired shows that the increase ofaverage power level during daytime is higher in the dark spots compared to the increasein bright spots. This suggests that scattering of people in the building provides acontribution to the field energy that is comparable and even greater to the average energydetected in local minima, leading to a substantial increase of energy in these spots. Yet,this contribution is small compared to the field at local maxima, and, as a result, thebright spots are less affected.

IV. Summary

We have made independent measurements of the field and of the powertransmitted through a single channel inside buildings with metal and brick walls. Wefind that, while the received power is described by a Ricean distribution, the distributionof the complex transmission coefficients shows a rather complex pattern determined by

10

the static geometry of the local environment. We also find an unexpected increase of theaverage SNR, which is observed when people are present in the building. In addition, thepresence of people in the building brings short-term fluctuations in the transmissionchannels with a characteristic time of approximately 0.1 s. Despite these changes,however, it appears that the nature of the speckle pattern of microwaves propagating inindoor environments is not fundamentally affected by human activities.

We thank H. U. Baranger, S. H. Simon, A. L. Moustakas, P. W. Wolniansky, A.M. Sengupta, R. A. Valenzuela, M. R. Andrews, and P. L. Gammel for valuable andstimulating discussions.

References:

[1] G. J. Foschini and M. Gans, “On Limits of Wireless Communications in a FadingEnvironment when Using Multiple Antennas,” Wireless Personal Communications 6, 311(1998).[2] A. L. Moustakas, H. U. Baranger, L. Balents, A. M. Sengupta, and S. H. Simon,“Communication through a Diffusive Medium: Coherence and Capacity,” Science 287,287 (2000).[3] P. W. Wolniansky, G. J. Foschini, G. D. Golden, R. A. Valenzuela, “V-BLAST: AnArchitecture for Realizing Very High Data Rates Over the Rich-Scattering WirelessChannel,” Proc. ISSSE, Pisa, Italy (1998); G. D. Golden, G. J. Foschini, R. A.Valenzuela, P. W. Wolniansky, “Detection Algorithm and Initial Laboratory Resultsusing the V-BLAST Space-Time Communication Architecture,” Electronics Letters 35,14 (1999).[4] M. Stoytchev and H. Safar, “Indoor Measurements of the Information Capacity ofMultiple-Antenna Arrays,” unpublished work, Bell Labs internal technical memorandum,Nov. 1999.[5] J. C. Lin, “Microwave Noninvasive Sensing of Physiological Signatures,” inElectromagnetic Interaction with Biological Systems, ed. J. C. Lin, Plenum Press, NewYork (1989), pp. 3-26.

UHF-Radio Propagation Predictor For Temporal VariationsIn Populated Indoor Environments

F. Villanese, W. G. Scanlon & N. E. EvansCentre for Communications Engineering, School of Electrical & Mechanical Engineering,

University of Ulster, Shore Road, Newtownabbey, Co. Antrim, N. Ireland, UK, BT37 0QB.Contact: Dr. W. G. Scanlon. w.scanlon@ulst.ac.uk. Tel: +44 28 90368054, Fax: +44 28 90366863.

1. AbstractRay-tracing is often used for site-specific propagation predictions, but current methods generally fail toaccommodate shadowing and reflections caused by the movement of people: these represent significanteffects in indoor radio environments. The method presented here improves site-specific indoorpropagation predictions by including multiple, moving human bodies. The model, based on a hybridimage and ray-shooting approach, alto takes into account electromagnetic antenna-tissue interaction forbody-worn terminals found in personal communications applications. The signal fading caused bypedestrians is demonstrated through simulations of a small-area (150 m2) open-plan office environmentat 2.45 GHz. For a fixed point-to-point link, temporal variations in the received signal were observedover a wide range of ~38dB and were caused solely by the movement of pedestrians. The situationbecomes worse when a mobile terminal is considered: a ~58dB variation in received signal power wasobserved. This variation was caused by combination of three distinct effects: spatial fading, temporalvariations and antenna-body interaction. It was noted that, for antennas in close proximity to the user’sbody (as may be found in personal applications using technology such as Bluetooth), a reduction inaverage received power of 15 dB occurred, directly due to antenna-body interaction.

2. Introduction

Interest in wireless communications has grown extremely rapidly in recent years. The advantages ofwireless systems over cabled networks, such as mobility of users and infrastructural flexibility haveensured that wireless remains at the forefront of communications technology. The requirement for highercapacity and the limited spectrum available has focused attention on the development of systems withsmaller cells (Picocells) and operation at higher frequencies. Within the indoor environment, both currentand emerging Picocell systems are intended for very high traffic density or high bandwidth applications.Practical systems, which often combine voice and data transmission, include Bluetooth, PCS, W-LAN(Wireless Local-Area Networks), W-PBX (Wireless Private-Branch Exchange), and parasitic-cellular.

The indoor environment has unique radio propagation properties that are not encountered with moreconventional, outdoor mobile or fixed-access scenarios. Multipath propagation is a key characteristic ofthe indoor channel and either statistical or deterministic models can be used to describe the variation in

12

received signal strength caused by fading. The existence of multiple propagation paths, each with adistinct time delay, attenuation and phase, gives rise to a highly complex transmission channel. Thecomplex baseband channel impulse response can be represented as:

where, k is the path index, is the path gain, is the phase shift and is the time delay of the path.The transmitted impulse is described by a Dirac function and the received signal h(t) is formed from thevector addition of an infinite number of time delayed rays, each represented by an attenuated and phase-shifted version of the original Dirac waveform. The channel response can also be represented in thefrequency domain by applying a Fourier transform.

Signal fading caused by multipath propagation can be either spatial or temporal. Spatial fading (orfast fading) depends on both the relative locations of transmit and receive antennas within the

environment and the nature of the environment itself (resulting in wave reflection and diffraction). Inspatial fading, significant variations occur as the receiver moves over distances that are more than a half-wavelength. Temporal fading (or slow fading), however, results from changes in the local environmentitself; this effect therefore also applies to stationary terminals. In populated indoor environments such asoffices and retail premises, the movement of people is the main cause of temporal fading. In both classesof fading the parameters and of (1) are functions of time. However, for a determinate observationtime, the channel can be assumed stationary [1]. With spatial fading, the direct correlation withwavelength means that wideband transmissions will suffer from frequency-selective fading and dispersionof the information signal. However, a changing local environment often leads to strong variations in line-of-sight (LOS) wave propagation due to geometrical obstruction. This results in temporal variations thatare independent of frequency, i.e., flat fading, even when considering wideband signals.

At UHF and above, the movement of people within the indoor environment leads to strong variationsin the propagation channel: critical fading may occur in otherwise adequate systems as contributory ray-paths are created or blocked [2]. The influence of the human body itself may also become a limitingfactor in the performance of a radio system when the terminal antennas are placed in close proximity to it,for example in the case of a personal digital assistant (PDA) [3].

The site-specific propagation predictor described below is based on a hybrid ray-image / ray-shootingalgorithm that incorporates temporal fading due to pedestrian movement. Section 3 describes a simplifiedhuman body model that accounts for ray transmission, diffraction and reflection. Where appropriate,antenna-tissue interaction effects can be incorporated using a full three-dimensional radiation pattern: anexample for the 2.45 GHz ISM-band is described in Section 4. The hybrid method itself is outlined in

13

Section 4 and in Section 5 results are presented for representative fixed-link and moving-terminalscenarios.

3. Human Body Model

In electromagnetic scattering, the human body is often represented as a simplified geometric form withhomogeneous dielectric parameters. This approach reduces the complexity of the overall problem andfacilitates rapid simulation of environments with multiple moving bodies. In the simulations presentedbelow, tissue dielectric properties are presumed to tend towards values representing human muscle.However, as frequency increases, wave penetration into biological bodies decreases and tissue parametersshould be adjusted to reflect this. Each human body, whether moving or stationary, is incorporated withinthe ray-tracing model as part of the site-specific topology. Represented as finite length, homogeneous,lossy dielectric cylinders, the bodies are located within the indoor environment at specific co-ordinatesand have an associated (individual) speed and direction of movement.

Figure 1 illustrates the three wave phenomena occurring when there is an interaction between adielectric cylinder and a high frequency electromagnetic wave. The high frequency ray approximationmodel must take account of reflections, diffraction around the body, and ray-transmission.

For example, at 2.45 GHz a body can be approximated as a lossy dielectric cylinder 1.8 m tall, 0.3 min diameter with the electromagnetic characteristics of muscle The attenuationconstant, for an electromagnetic wave traveling in lossy media, and the skin depth, for the bodymodel can be obtained by using the above values and equation (2).

.

14

At 2.45 GHz the theoretical wave penetration depth in muscle is extremely low (22.3 mm), resultingin an attenuation of more than 39 dB after 10 cm of travel. The role of in (2) ensures that greaterattenuation will occur as frequency increases. The scale of direct-ray absorption means that raystransmitted through the body may be neglected without incurring significant errors; such rays areconsidered ‘blocked’ by the body.

Incident waves will also be diffracted by the cylinder model; these ‘creeping’ waves are modeled byimplementing a uniform GTD solution [4, 5]. Investigation of this phenomenon suggested that thediffracted waves will suffer significant levels of attenuation, which will increase with frequency(specifically: 12.3 dB at 2.45 GHz, 16.0 dB at 5.7 GHz and 26.5 dB at 62 GHz). They are therefore onlycalculated when a body is placed in the direct path between the transmitter and receiver. In all other cases,the amplitudes of diffracted rays are negligible when compared to either direct or reflected rays.

Another important propagation mechanism in indoor environments is ray-reflection from the movingbodies within the environment. The reflection coefficient will depend on polarization, frequency andtissue-parameters. Considering the 2.45 GHz cylinder model described earlier, Figure 2 shows thereflection coefficient as a function of the incidence angle for both perpendicular and horizontalpolarization. The values relate to specular reflection only; they were calculated by considering a plane-wave incident on a reflecting surface tangential to the incident point. With coefficients ranging from 0.77to 0.99 for the TE case (vertical polarization), reflection from moving bodies is an important mechanismto consider when modeling indoor UHF-radio propagation.

15

4. Antenna-Body Interaction

Personal communications devices are designed to operate in close proximity to the human body. This canresult in a strong interaction between the terminal antenna and the body itself leading to reducedefficiency and severe radiation pattern fragmentation. These effects depend on parameters such as therelative position between devices and the body, the operating frequency and the subject’s dimensions. Inmany indoor scenarios, pattern fading can lead to degraded link quality, particularly when the direct-ray isblocked by the body itself (i.e., where multipath reflections become dominant). Numerical techniquessuch as FDTD can be used at lower frequencies (below 10 GHz) to generate full three-dimensionalradiation patterns for the coupled antenna-body system. At higher frequencies, the model size becomesprohibitively large; here partial body models or GTD solutions must be used.

Consider a 2.45 GHz radio transceiver proximate to the user’s chest and equipped with amonopole antenna, the RF unit acting as a counterpoise (Figure 3). A whole-body adult-male tissue modelwas constructed, incorporating the transceiver and a full three-dimensional radiation pattern was obtainedusing FDTD simulation. Using an outdoor, elevated range, comparison measurements were made for theazimuthal plane. The radiation efficiency was 49.0% for the FDTD case, and 51.3% for the measurements(calculated using PAG – pattern averaged gain). Within the propagation model, the effects of antenna-body interaction were taken into consideration by substituting the original pattern with the simulatedradiation pattern of the entire antenna-body system. Each calculated ray associated with transmitting orreceiving devices placed in close proximity to the body is then weighed by the coefficients of themodified three-dimensional radiation pattern.

16

5. Hybrid method

A ray-tracing algorithm was developed that was capable of supporting single three-dimensional roomscenarios with both transmitting and receiving located inside. An unlimited number of pedestrians (i.e.,cylinders representing human bodies) can be simulated within the room. Each pedestrian can be moved inall parts of the room in order to simulate temporal variations in received power. For maximum flexibility,the information about the position and the movement of each body present is supplied to the program viaa text file. Any areas with different electromagnetic properties, such as windows and doors, can beincluded in the model.

The algorithm initially determines all possible propagation paths between transmitter and receiverusing the image technique. Rays that intersect obstructing bodies are deleted according to theapproximations outlined in Section 3. When the direct ray is blocked (NLOS case) the program evaluatesthe contribution from waves diffracted around the body. Additional paths between the transmitter andreceiver may be created due to body reflection; these are calculated using an efficient ray-shootingalgorithm (see Figure 4). Ray shooting is inherently computationally expensive, particularly forcomplicated indoor environments. In the hybrid model developed, however, rays are only shot towardsthe part of the cylinder surface visible from the transmitter and then only the rays subsequently arriving atreceiver are considered. The method is capable of modeling bodies with variable position co-ordinates(moving cylinders), by recalculating using ray shooting and updating the list of blocked rays for each timestep; note that the image calculations are not repeated. The algorithm also considers the full radiationpattern of both antennas. To incorporate antenna-body interaction effects, a suitably modified radiationpattern can be passed to the program.

17

6. Representative Simulations

The algorithm was applied to a typical 150m2 open-plan office environment. In the first scenario, shownin Figure 5, a 2.45 GHz link connects a stationary terminal (STA) to an access point (AP) in the presenceof moving pedestrians (shown as A, B, C and D). Each pedestrian is represented as a dielectric cylinder ofdiameter 0.3 m and height 1. 8 m. The area was modeled as a regular cuboid with windows and doorsincluded but all other items of furniture were neglected. Pedestrians B and C are both moving at 0. 8 ms-1,A and D have speeds of 1 ms-1 and 1.1 ms-1 respectively. These conditions are intended to represent aworst-case scenario within an open-plan office, with four people moving at the same time. The AP wasfixed close to the wall at 2.5 m above the floor, and simulated as an ideal vertical dipole antenna(+2.1 dBi) fed from a matched 0 dBm source. The STA, equipped with a +0 dBi vertical monopole, wasplaced in a desk-top position as shown in Figure 5.

The instantaneous received power at the stationary terminal was calculated at 0.01 s intervals as thepedestrians move as indicated in Figure 5. The resulting temporal variation in received power is shown inFigure 6 and, in spite of the fixed position of the terminal, the movement of people causes significantfading. The reduction of the mean power observable between 5 and 7 s (points (i) and (ii)) is due to theblocking of the direct ray by the pedestrian bodies when they pass close to the STA.

18

The received signal fluctuates over a range of ~38 dB, with average power of –50.8 dBm. Statisticalanalysis of the events shown in Figure 6 revealed that the temporal fading was Rayleigh in nature. Thereceived signal varies within –31.8 dB and +10 dB of the median with a practical range of ~25 dB for acumulative probability lying between 1% and 99%. In addition, a FFT (Fast Fourier Transform) wasapplied to the 13 s time series of power values shown in Figure 6 in order to analyze the frequencyspectrum of the fading. The bandwidth of the fading variation was found less than 2 Hz; hence, fadinginduced by moving people about can be considered extremely slow.

The link quality within the indoor environment can be improved using ceiling-mounted access pointinstead of wall-mounted antennas located at center of the room. This was investigated by repeating thesimulation of Figure 5 with a typical ceiling-mounted indoor antenna with +8 dBi gain. As expected, dueto the change in physical layout, the effect of the moving personnel was reduced in this situation. Fadesnow occur in bursts of 0.5–2 s duration and with ~27 dB maximum range, separated by periods duringwhich the received signal remained almost constant.

The goal of many new indoor radio systems is to guarantee continuous wireless connectivityregardless of the specific operating conditions to support mobile applications such as body-worncomputing and PDA networks. However, when one terminal is moving the indoor radio channel suffersfrom more than one source of channel variation. The algorithm was therefore applied to the case in whichthe STA is in motion. A body-worn terminal, such as PDA in a pocket moving at a speed of 1 ms-1 alongthe path shown in Figure 7 was simulated. The PDA was considered as being worn close to the body atchest level and facing the direction of travel, representing the worse-case conditions for a personalcommunication device. The three-dimensional radiation pattern of the receiving antenna was modified

19

according to the position of the PDA, considering both the antenna-body interaction described in Section4 and the orientation of the body. In the posterior body direction, the modified pattern presents antennagains of between –40 dBi and –50 dBi, despite having an overall antenna-body system efficiency close to50%.

The instantaneous received power at the mobile terminal is shown in Figure 8. In this simulation thepedestrians (A, B & C) were excluded in order to investigate the variation of the channel related only withthe movement of the receiver. A LOS situation exists until 6 s since the body, and therefore the mobileterminal, is facing the transmitter. The power fluctuates over a range of ~16 dB and the channel statisticswere observed to be Rician with a Rice factor of –13 dB. Due to the change of the orientation of the bodycarrying the terminal after 6 s, when the host changes direction, an OBS situation is created. Here thebody itself represents an obstruction for the direct ray. Spatial fades are observed with a dynamic range of~38 dB and the channel statistics become Rayleigh in nature. The average value of the received powerdrops distinctly from –56.9 dBm (LOS case) to –71.7 dBm (OBS case).

20

Additional pedestrians (moving as shown in Figure 7) were added in the simulation to determine thetemporal variation of the channel when both receiver position and the environment are changing. Theeffect of the moving pedestrians is shown in Figure 9; for the LOS case the signal now has a dynamicrange of ~25 dB. Despite the existence of a direct ray in the LOS situation, it was observed that thechannel now follows a Rayleigh distribution for the entire (LOS and OBS) observation time. This is dueto the additional reflections and obstruction caused by the presence of the pedestrians. In this case, theaverage power drops from -54.5 dBm (LOS case) to –68 dBm (OBS case). The total dynamic range isabout 58 dB.

7. Discussion and Conclusions

In the fixed-link, temporal variations in the received signal envelope, caused by the movement ofpedestrians, were observed. The signal fluctuates over a wide range of ~38 dB. In [1] Bultitude reported atypical dynamic range of 30 dB for 900 MHz indoor fixed-link measurements. However, the channelcharacteristics are strongly dependent on the topology of the environment and the density of pedestriantraffic. Further work in this project will aim to explore channel variations associated with a range ofdifferent environments and traffic conditions.

The observed fading is extremely slow with a bandwidth of ~2 Hz. The effects of such slow fadingupon a digital link are related with the transmission rate and the burst duration. The burst duration has tobe short enough to consider the channel constant over the burst. Further work is required to understandthe effect of this slow fading upon a system operating with specific transmission characteristics. Despitethe existence of a direct ray between an access point and a terminal (except for instances where bodyshadowing occurs), the observed cumulative distribution of the channel was Rayleigh. If we assume thatthe process is stationary (in the widest sense) and ergodic, it is possible to use the values on abscissa as

21

additional attenuation that have to be considered in a radio link budget design with a desired quality oftransmission. For a Rayleigh channel the additional attenuation for a reliability of 95% is 12 dB. It ispossible to counteract these effects; for example, ceiling-mounted distributed antennas can be used inorder to increase both range and reliability [6]. Distributed antennas split the transmission power amongseveral elements separated in space, in this way less power is wasted in shadow losses and the line-of-sight channel will be present more frequently.

Three main effects must be considered for mobile, body-worn terminals: spatial fading, channelvariations due to the movements of others and antenna-body interaction. As shown in Figure 8,obstruction by the user’s body itself results, for the scenario described, in a reduction of average receivedpower by ~15 dB. This is an additional loss that must be taken into account when determining coveragefor personal communications devices. In the OBS situation, the spatial fading distribution is emphasizeddue of the lack of a strong direct-ray: fades of more than 35 dB were observed. The addition of pedestriantraffic into the simulation (Figure 9) results in a combination of two stochastic processes: the PDA movesthrough the spatial fading determined by the environment, but at the same time the propagationenvironment continually changes because of the pedestrian motion. However, analysis of the resultsindicated that the fading was still Rayleigh in nature.

In conclusion, the combined effects of pedestrian movement, a moving receiver and antenna-bodyinteraction can strongly impair the quality of wireless communication system. The model developedoffers a reliable solution to determine the impact of moving pedestrians for a site-specific indoor radiochannel.

7. References

[1] R. J. C. Bultitude, “Measurement, characterization and modeling of indoor 800/900 MHz radio channels fordigital communications,” IEEE Communication Magazine, vol. 25, 6, June 1987, pp. 5–12.

[2] F. Villanese, W. G . Scanlon, N. E. Evans & E. Gambi, “A hybrid Image/Ray-Shooting UHF radio propagationpredictor for populated environment,” Electronics Litters, vol. 35, 21, 1999, pp. 1804–1805.

[3] W. G. Scanlon, “Body-coupled antennas for medical device communications,” URSI National Radio ScienceMeeting, York, UK, p. 18, 1999.

[4] P. E. Hussar, “A uniform GTD treatment of surface diffraction by impedance and coated cylinders,” IEEE onAntennas and Propagation, Vol. 46, N. 7, July 1998, pp. 998–1008.

[5] D. A. McNamara, C. W. I. Pistorius & J. A. G. Malherbe, Introduction to the uniform theory of diffraction,Norwood, Massachusetts, Artech House, 1990.

[6] S. R. Saunders, Antennas and propagation for wireless communication systems, Guildford, UK, John Wiley &Sons, 1999.

An Improved Approach for Performance Evaluation of the Downlink

of DS-CDMA PCS Indoor Systems with Distributed Antennas

MARIO R. HUEDA, CARMEN RODRIGUEZ and CARLOS MARQUESCentro de Comunicaciones Digitales - F.C.E.F. y Naturales - Universidad Nacional de Córdoba. Casilla de Correo 755 –Correo Central - Córdoba (5000) - Argentina Tel./Fax : 54-351-4334147 (ext. 15) - Email :mhueda@com.uncor.edu

Abstract. This paper studies the performance of the forward link of a DS-CDMA system with distributed antennasover indoor radio channels. The interference component caused by the same cell-site users, or interpathinterference (IPI), is a dominant factor in the performance of these systems. In their analysis it is typicallyassumed that the probability density function (pdf) of the IPI can be approximated by the Gaussian distribution.This is called the Gaussian approximation (GA). We show that the GA for the IPI does not provide an accuratecharacterization of the performance of the downlink of DS-CDMA PCS indoor systems with distributed antennas.The bit error rate (BER) at the output of a practical RAKE receiver is computed using an accurate description forthe pdf of the IPI. We show that when the IPI density is thus modeled, the signal to noise ratio (SNR) at the outputof the combiner is well approximated by a random variable with a Weibull distribution. We also propose ananalytical expression for the frame error rate (FER) of convolutional codes (CC) over slowly fading channels.Results based both on the analytical model and computer simulations for cases involving different numbers ofantenna elements and different levels of interference from other cells are presented. The analytical model is shownto be highly efficient in terms of computational cost while preserving excellent accuracy. Comparison of the FERobtained by the new approach for the IPI with those obtained by the GA show that the latter method incursconsiderable error when used to evaluate performance of DS-CDMA PCS indoor systems with distributedantennas.

Keywords: CDMA, PCS, convolutional codes, diversity combining.

1. Introduction

RAKE receivers are used to optimally combine multipath components in direct sequence code

division multiple access (DS-CDMA) systems [1]. These receivers can resolve individual paths with

temporal resolution (with W equal to the bandwidth of the spread spectrum signal), thus

providing path diversity. For example, the DS-CDMA IS-95 standard with 1.23MHz bandwidth [2] was

originally designed for an outdoor cellular system where the delay spread is usually in the range of

However, the mean delay spread for indoor environments is typically around 100ns. Therefore diversity

24

gain through multipath fading cannot be achieved by a CDMA receiver with an 813ns chip interval. In

these situations a distributed antenna system improves performance by providing multipath diversity [3].A large body of literature exists on the analysis of CDMA systems operating in a fading multipath

environment [4][5]. Particularly, the performance of the forward link of DS-CDMA systems has beenanalyzed in several works [6]-[9]. E. Sousa and V. Da Silva [6] studied the effect of multipath propagation

on the performance of the forward link of a CDMA cellular system. Yang [7] analyzed the performance ofthe forward traffic channel of the IS-95 standard using a modified maximal ratio diversity receiver with asearch engine. Yun et. al. [8] derived the bit error rate in a synchronous (downlink) indoor DS-CDMA

system over a Rician multipath channel, while Weerackody [9] examined the effects of the channel fade

rate on the performance of the downlink of the IS-95 system.In most of the previous works, the interference component from the same cell-site users (interpath

interference, or IPI) has been assumed to be describable by a Gaussian probability density function (pdf).This assumption is justified by the central limit theorem when the number of multipath components is large[9]. The resulting approximation is called "the Gaussian approximation" (GA). Although the GA isadequate for the different situations analyzed in those works, it is not appropriate for the case of the

forward link of a DS-CDMA PCS indoor system with moderately large bandwidth (e.g., 1.23MHz) and adistributed antenna system. In this situation, the number of multipath components resolved by the receiveris limited to the number of antenna elements, which is usually small, therefore the central limit theorem isnot applicable.

Furthermore, low-rate convolutional coding with interleaving is typically the error correction

method used in transmission over fading channels (e.g., IS-95 uses a rate ½, constraint length 9,convolutional code (CC) with interleaving in its forward link). In this case it is difficult to obtain goodestimates of the bit and frame error rate (BER and FER) for CC owing to temporal variations in the signalstrength and the presence of interleaving. For this reason, in most previous studies performance evaluationhas been done via computer simulations. Recent work has proposed techniques to estimate FER for CCover fading channels with interleaving [10][11]. Tralli [10] proposed semi-analytic methods which havebeen found much more efficient than simulation for the accurate evaluation of small error probabilities. Theefficiency improves as fading samples become less correlated, therefore this technique is attractive intransmission over fast fading channels. S. Nanda and K. Rege [11] described an analytic technique thatprovides good estimates of FER for CC over fading channels with interleaving. Although this methodseems promising, its application in transmissions over slowly fading channels employing diversitytechniques has not been reported so far.

This paper studies the forward link of a DS-CDMA system with distributed antennas over indoor

radio channels. IPI dominates the performance of these systems, and we show that the GA for the IPI does

25

not provide an accurate characterization of the performance of them. We propose a more accurate model to

describe the effect of IPI on BER and FER, and show that when the IPI density is modeled as proposedhere, the signal to noise ratio (SNR) at the output of the combiner is well approximated by a random

variable with a Weibull distribution. We also propose an analytical expression for the FER of

convolutional codes (CC) over slowly fading channels. FER of the forward traffic channel of an IS-95system are then computed by computer simulation and by exact integration of the analytic expressionsusing the Weibull approximation to the pdf. The latter method is shown to be the most efficient in terms of

computational cost while preserving excellent accuracy. Comparison of the FER estimates obtained by theWeibull approximation with those obtained by complete system simulation reveals that they provide amuch more accurate result than the GA. In Section 2 we derive expressions for the signal statistics at the

output of the RAKE. Section 3 presents the performance evaluation techniques which are the subject of thispaper for the case of uncoded channels, whereas Section 4 extends these techniques to coded channels.

Finally, concluding remarks are given in Section 5.

2. The Received Signal at the MobileLet be the transmitted signal from the base station (baseband signal):

where K is the number of users (k = 0 denotes the pilot signal), and are the voice activity factor and

the signal amplitude, respectively; is the impulse response of the transmitter

(receiver) filter, which corresponds closely to the chip pulse shape used in the IS-95 standard [2]. is

complex chip sequence for the kth user. is the information sequence, where

is the data symbol, N is the processing gain and elsewhere.

The signal received at the mobile station (baseband signal) can be written as

where L is the total number of multipaths, and are the amplitude and the delay of the lth multipath,

respectively. In this work we assume that are independent zero-mean complex Gaussian random

variables. The delays are assumed to be constants for a particular channel model. z(t) represents the

complex noise (thermal noise + interference from other cells), which is modeled as zero- mean AWGN with

(the superscript * indicates complex conjugation and E{ }denotes the expectation

operator).

26

At the receiver, r(t) is filtered and sampled each seconds (i.e.. M = 2, 4, 8). Then, the

samples are processed by synchronization stages (code acquisition, alignment and tracking) and by the

RAKE fingers (despreaders) [12]. After the searching process, the output of the jth finger at mth instant

for the user is (see Fig. 1)

where J is the total number of RAKE fingers is the signal at the receiver filter

output (the operator * denotes convolution), is the chip sequence of the user, is the

sampling instant and with is the sampling phase. Note that

with is the time delay corresponding to the jth finger. The proper values of and

are determined by synchronization stages.

Since the chip interval is 813ns in IS-95 and the mean delay spread for indoor environments ifaround 100ns, only one path is resolved for each antenna element, thus we can assume

Moreover we consider that the average energy received from each element is equal

(i.e., and the chip timing is perfect. Then the signal at the output of the jth finger reduces

to (see [14] for more details):

The first term on the right in (5) is the desired signal component for the user at the mth instant

is the signal energy of the traffic channel); next is the interference due to the AWGN noise

generated from thermal noise and noise from adjacent cell-site base stations. The third term is the

interference from the same cell-site users due to the presence of the multipaths and imperfect chip timing. It

27

is possible to verify from [14] that and can be well modeled as zero mean gaussian processes, with

and for

is the pilot signal amplitude) is the average transmitted power from the base

station.Most previous works have approximated the last term on the right of (5) by a Gaussian process.

This results from assuming large values of L and using the central limit theorem (CLT) [9]. In this case, thefinger output results in

where is zero-mean AWGN with

For the system analyzed in this work, the number of multipath components “seen” by the receiveris practically the same as the number of antenna elements (e.g., L=2,3 or 4), therefore the CLT is notapplicable. For this reason, we will use the last term of (5) without approximations to model theinterference from the same cell-site users; this provides the improved approach for the estimation of the IPI

(IA-IPI). As we will show in next sections, the results obtained from the IA-IPI (5) are significantly betterthan those resulting from the GA-IPI (6).

3. Performance Analysis

In practical systems, the pilot signal is used by the mobile receiver to demodulate the fingersoutputs (pilot-signal aided demodulator). Then, the demodulated signals are added to obtain the decision

variable. In this section we obtain the average bit-error-rate at the output of a RAKE receiver with pilot-signal aided demodulation, considering both approximations (i.e., the IA-IPI and the classical GA-IPI).

3.1 Performance Analysis Using the GA-IPI

The decision variable at the combiner output of a J-branch RAKE receiver using the GA-GPI isgiven by (for simplicity, the time indexes are omitted)

28

where is the finger output given by (6) and is the path gain estimated by means of the pilot signal.

In this paper we assume perfect estimation of the channel parameters, that is, Then it can be

verified that for a given set of fade values, the decision variable (8) is Gaussian with

and

The SNR per bit at the combiner output for the GA-IPI results in

where is the average SNR per branch defined by

In this case, the pdf of is given by [ 1 ]

while the average bit-error-rate results in

where and

3.2 Performance Analysis Using the IA-IPI: the Weibull Aproximation

Similar to the above analysis, it can be verified that the SNR per bit at the combiner output resultsin

Then, the average bit error rate at the output of the RAKE receiver can be obtained from

29

where is the pdf of Since does not have a closed-form, neither does (16).

However we have verified that for several cases of interest, the SNR given in (15) can be accuratelyapproximated by a random variable with Weibull distribution, that is

Fig. 2 compare the cumulative distributions of resulting from the Weibull approximationwith equivalent cumulative distributions obtained from computer simulations. Parameters and arecomputed by fitting (17) to the pdf obtained from simulations [13]. We consider N=64 and several valuesof the relative interference factor f , which is defined as the ratio of the noise power due to other cell-site

base stations to the interference from the users in the same cell-site, that is, where

[9] (thermal noise is neglected in this work). Results for L=2 and 3 with J=L are presented(i.e., number of fingers = number of multipath components). Note that the Weibull approximation showsexcellent accuracy when compared to the simulation results (we have found similar results for othersimulation conditions not presented in this work (e.g., J=3 and L=4,5)).

3.3 Comparison between and

Fig. 3 shows the cumulative distribution function (cdf) for both approximations and several valuesof J, L and f. We can see that the behavior of improves respect to at low values of (this

effect is more important when f decreases). This behavior of is due mainly to the correlation

30

between the signal power at the output of a given finger with the IPI power at the outputs of the other

fingers.

To explain this result, we consider a simple example of a RAKE receiver with two fingers and two

multipaths (J=L=2) with In this case, the finger outputs can be expressed as (see (5)

and (6))

Note that the average power of the IPI (underline terms) is the same for both approximations (i.e.,

Then, when and are in a deep fading we see that:

a) for the GA-IPI, the signal power at and is small thus is seriously

deteriorated;

b) for the IA-IPI, not only the signal power at and is small but also the IPI power. Thus

the overall SNR at the combiner output will be larger than one obtained with the GA-IPI (i.e.,

Then, from a) and b) it is possible to understand the better behavior obtained with the IA-IPI

respect to the GA-IPI at low values of From Fig. 3 note that the improvement obtained with the IA-IPI

31

decreases as grows. This is because when and are large, not only the signal power at

and is large but also the IPI power, which tends to deteriorate (note that variations of or

do not affect the IPI power at and Nevertheless, since the error probability is strongly

influenced by the behavior of the pdf at low values of the overall system performance predicted by the

IA-IPI is significantly better than that derived from the GA-IPI (this will be verified in the next subsection).

3.4 Numerical Results and Discussion

Next we compare the system performance using the IA-IPI and the GA-IPI. The processing gain isN =64. Several values of number of antenna elements, RAKE fingers and relative interference factor f areconsidered.

The theoretical and simulated BER at the combiner output for different signal levels (16) are

shown in Fig. 4. For the GA-IPI, the theoretical results are derived from (14). For the IA-IPI the theoreticalresults are obtained by numerical integration of the (16) using the Weibull approximation (17) to the pdf.In all cases presented in Fig. 4, we can verify that the theoretical and simulated results are very close. It canbe seen that performance predicted by the IA-IPI is much better than that derived from the GA-IPI, asdiscussed in subsection 3.3. Thus, we conclude that the GA-IPI incurs considerable error when used toevaluate performance of DS-CDMA PCS indoor systems with distributed antennas; the inaccuracy getseven larger when the relative interference factor/decreases. Similar results will be obtained in next section,where performance of the error correction code adopted in the forward link of IS-95 is analyzed.

32

4. Performance Evaluation of Convolutional Codes over Slow FadingChannels

Low-rate convolutional coding and interleaving is the error correction method typically used in

transmissions over fading channels. For example, the IS-95 standard employs a rate ½, constraint length 9,

convolutional code (CC) with interleaving in its forward link. In this section we present analyticalexpressions to evaluate the bit and frame error rate for convolutional codes over very slow fading channels.

Let be the frame error probability of a CC over AWGN channels. Since in indoor

environments the channel varies slowly, the SNR can be considered constant over the duration of all errorevents (the effect of interleaving is neglected). In this case, the average frame error rate can be well

approximated by

where is the SNR at the combiner output and is its pdf. For the GA-IPI the pdf is given by

(13) while for the IA-IPI we use the Weibull approximation (17). We approximate the decoder performanceover gaussian channels using direct interpolation from the values derived from simulations.

Then expression (20) is easily evaluated by numerical integration. As we will show in next section, resultsobtained from (20) show good agreement with values derived from system simulations, even when theinterleaver is included (interleaving is not effective when the channel varies slowly).

4.1 Numerical Results

Next we study the performance of the forward link of the IS-95 standard using distributedantennas. The information bits are framed every 20 ms, and then convolutionally coded with rate R=½,constraint length 9. Interleaving as specified in the IS-95 standard and soft-decision decoding are used. Thedata rate is set to full voice rate of 9600 bps and the symbol rate of is 19200 symbols/s.

Carrier frequency is 1800 MHz and Doppler frequency is The processing gain is N=64.

Different number of antenna elements and values of the relative interference factor f are considered.Fig. 5 shows the FER at the output of the Viterbi decoder for different signal levels. Since

and a rate ½ code is used, the SNR per bit is defined by where J is the

number of RAKE fingers and is the average SNR per branch given by (12). We present simulations

results (with and without interleaving) and theoretical results obtained from (20). In this case, thetheoretical results for the GA-IPI and IA-IPI are obtained by numerical integration of (20) using the pdfgiven in (13) and (17), respectively. As in the subsection 3.4, it is possible to verify that the performance

33

obtained from the IA-IPI is much better than that derived from the GA-IPI. Note that in all cases analyzedthe theoretical values show excellent accuracy when compared to the simulation results. Also note thatsince the channel varies very slowly, the effect of interleaving is not very important at practical values of

FER (e.g., Thus we conclude that the analytical expression (20) can be used to obtain

accurate estimates of the performance of CC over slowly fading channels including diversity andinterleaving.

5. Concluding Remarks

This paper has investigated the performance of the forward link of a DS-CDMA system withdistributed antennas in indoor environments. We have shown that the classical Gaussian approximation forthe interference component caused by the same cell-site users incurs considerable error when used toevaluate performance of these systems. We have proposed a more accurate model to describe the effect ofIPI on BER and FER, and showed that when the IPI density is modeled as proposed here, the signal tonoise ratio (SNR) at the output of the combiner is well approximated by a random variable with a Weibulldistribution. We have also proposed an analytical expression for the FER of convolutional codes (CC) overslowly fading channels. FER’s for the CC used in the IS-95 standard were then obtained by computersimulation and by exact integration of the analytic expressions using the Weibull approximation to the pdf.The latter method has shown to be the most efficient in terms of computational cost while preservingexcellent accuracy (we have verified that with the Weibull approximation the computation time is reduced

34

by about 1 ½ orders of magnitude compared to the simulation of the complete system!). Comparison of the

FER estimates obtained by either numerical integration or the Weibull approximation with those obtainedby complete system simulation reveals that they provide a much more accurate result than the GA.

References[1] J. G. Proakis, Digital Communications, McGraw-Hill, Third Edition. 1995.[2] TIA/EIA/IS-95, “Mobile Station-Base Station Compatibility Standard for Dual-Mode Wide-band Spread

Spectrum Cellular Systems,” Telecommunication Industry Association, July 1993.[3] J. Yang, “Analysis and Simulation of a CDMA PCS Indoor System with Distributed Antennae,” Proceedings of

PIMRC’95, pp. 1123-1127, Toronto, Canada.[4] K. S. Gilhousen. I. M. Jacobs, R. Padovani, A. J. Viterbi, L. A. Weaver ans C. E. Wheatly, “On the Capacity of

Cellular CDMA System,” IEEE Trans, on Veh. Technol., Vol. 40, pp. 303-312, May 1991.[5] A. Salsami and K. S. Gilhousen, “On the System Design Aspects of a Code Division Multiple Access (CDMA)

applied to Digital Cellular and Personal Communications Networks,” in Proc. IEEE Conf. Veh. Technol., St.Louis, MO, May 1991.

[6] V.M DaSilva, E.S. Sousa and V. Jovanovic, “Effect of Multipalh Propagation on the Forward Link of a CDMACellular System,” Wireless Personal Communications, Vol. 1, Nol, pp. 33-41, Kluwer Academic Publishers,1994.

[7] J. Yang, “Diversity Receiver Scheme and System Performance Evaluation for a CDMA System,” IEEE Tram.on Commun., Vol. 47, pp. 272-280, Feb. 1999.

[8] L. C. Yun, M. Couture, J. Camagna and J.P Linnartz, “BER for QPSK DS-CDMA Downlink in an IndoorRicean Dispersive Pico-Cellular Channel,” Proc. 27th IEEE Asilomar Conf. On Signals, Systems andComputers, pp 1417-21, Nov. 1993.

[9] V. Weerackody, “Effect of Time Diversity on the Forward Link of the DS-CDMA Cellular System,” WirelessPersonal Communications, Vol. 7, Issue 2/3, pp. 81-109, Kluwer Academic Publishers, August 1998.

[10] V. Tralli, “Efficient Simulation of Frame and Bit Error Rate in Wireless Systems with Convolutional Codesand Correlated Fading Channels,” Proc. IEEE WCNC’99, New Orleans, Sept. 1999.

[11] S. Nanda and K. Rege, “ Frame Error Rates for Convolutional Codes on Fading Channels and the Concept ofEffective Eb/No,” IEEE Trans. on Vehicular Techn., Vol. 47, pp. 1245-1250, Nov. 1998.

[12] A. Viterbi, Principles of Spread Spectrum Communication, Addison -Wesley, 1995.[ 13] K. Pahlavan and A. Levesque, Wireless Information Networks, John Wiley & Sons, 1995.[14] M. Hueda, G.C. Briones, C. Rodriguez and C. Marques, “MMSEC-RAKE Receivers with Resolution

Reduction of the Diversity Branches: Analysis, Simulation and Applications,” Proc. of IEEE Globecom’99, Riode Janeiro, Dec. 1999.

Fast and Enhanced Ray Optical Propagation Modelingfor Radio Network Planning in Urban and Indoor Scenarios

R. Hoppe, P. Wertz, G. Wölfle, and F. M. Landstorfer

Institut für Hochfrequenztechnik, University of Stuttgart,Pfaffenwaldring 47, 70550 Stuttgart, Germany

e-mail: hoppe@ihf.uni-stuttgart.deWWW: http://www.ihf.uni-stuttgart.de

With the increasing number of subscribers in mobile communications there is a growing interestin propagation models for the mobile radio channel in urban scenarios and inside buildings.Because of the increasing transmission rates propagation models should be able to predict thefield strength coverage as well as the wideband properties for these scenarios.Ray optical modeling of wave propagation is often used for the prediction of the field strength(and delay spread) in wireless mobile communication networks. However, the practical use ofthese deterministic models is limited due to their high computational demands. For large areasin urban or indoor scenarios the computation times are in the range of hours which is too longfor the planning of mobile radio networks.A new method for the acceleration of ray optical models is presented in this paper. It is basedon a single intelligent preprocessing of the database in which the mutual visibility relationsbetween the walls and the edges of the buildings are determined. Therefore the computationtime is reduced to a few minutes on standard PCs. The propagation model is implemented forurban and indoor scenarios and comparisons with measurements show the gain in computationefficiency as well as in achieved prediction accuracy [1].

1. Introduction

The increasing number of subscribers in mobile communications forces network operators to utilize economicalfrequency planning. An adequate solution to cope with the growing capacity demands is the reduction of the cell size.Especially in densely built up areas with high traffic rates microcellular or picocellular networks are used very often.

The planning of these networks requires highly sophisticated propagation models. Depending on the parametersof the base station (location, frequency, transmitting power and characteristic of the antenna) the propagation modelsgive information about the quality of service in the area taken into account. The field strength level is still consideredas the most significant parameter for describing the coverage. However, with the increasing number of digital com-munications other parameters that characterize the radio channel, such as delay spread and impulse response, becomemore and more important.

36

The basis for any propagation model is a database which describes the propagation environment. As microcells areplanned to increase the network capacity in cities, it is obvious to use building oriented databases for urban areas. Inorder to get a more accurate description of wave propagation, the building data are stored in a vector format. Everybuilding is modelled as a vertical cylinder with polygonal ground plane and an uniform height above street level. Withthis approach only vertical walls and horizontal flat roofs are considered. Additionally, the material properties of thebuilding surfaces can be taken into account. Figure 1 shows an example of a building database for urban areas.

For indoor sceanrios the building data are stored in a 3D-vector format including all walls, doors, and windows.All elements inside the building are described in terms of plane elements. Every wall is represented by a plane andits extent and location is defined by its corners. As mentioned above, for each element individual material propertiescan be taken into account With respect to an efficient use it is also possible to import dxf-files, a very common dataformat of CAD–software used in architecture.

2. Propagation Models

There are basically two approaches to the prediction of wave propagation in urban and indoor environments whichdiffer in computational effort and accuracy of the prediction [2].

The so called empirical models (e.g. the model according to Walfisch/Ikegami for urban areas or the Multi–Wall–Model for indoor [3]) consider only the propagation in a vertical plane which contains transmitter and receiver (seefig. 2). For the field strength prediction significant parameters have to be extracted from this vertical section (e.g.average building height or number of penetrated walls). Finally equations containing these parameters have to beoptimised and fined to numerous measurements in order to get a prediction model which is applicable in differentpropagation environments. The main advantage of empirical models is their short computation time. However, theirprediction accuracy is limited due to the fact that only a small number of parameters is taken into account and theinfluence of the distance from the transmitter is over-emphasised. Additionally, waveguiding effects in streets or alongfloors cannot be considered with this empirical approach.

Ray optical 3D propagation modeling has become a widely discussed technique for the prediction of the fieldstrength (and also of the delay spread) in indoor and urban scenarios [4]. This kind of wave propagation modeling isvery accurate because it considers waveguiding effects in street canyons (urban) or corridors (indoor) and additionallyincludes diffraction around corners.

37

There are two basic approaches to searching ray optical propagation paths in an arbitrary vector oriented buildingdatabase: ray tracing and ray launching [5]:

Both of them have their individual advantages and disadvantages. Ray tracing computes valid rays for each receiverpoint individually and guarantees the consideration of each wall as well as a constant resolution. This individual com-putation is more time-consuming than the ray launching approach, where the rays are launched from the transmitterinto all relevant directions discretized into small angular increments. There are problems, however, with consideringdiffracted rays. In ray launching an edge could be neglected because it is located in the middle between two rays, ad-ditionally the diffraction multiplicates the number of launched rays. Different approaches to solve the problems withray launching were presented in the last years [6], but the ray launching has still kept the disadvantage of a variableresolution depending on the distance to the transmitter.

Generally the field strength is computed using Fresnel equations for the reflection and transmission and GTD/UTDfor the diffraction [7]. On the other side empirical diffraction models are available, because they can be calibrated withmeasurements [8].

Ray optical models are very time–consuming, because all possible rays must be determined and therefore manyreflections, transmissions and diffractions have to be computed. Especially 3D models generate a large number ofrays, but only few of them deliver an important contribution to the received electromagnetic energy. Therefore imple-mentations for 2D are also available, but they have a limited accuracy and the computation times are still in the rangeof hours on a standard PC. Several approaches to accelerating these models were presented in the last years and lead toacceleration factors up to 10 [8]. But the computation times of these ray optical models (2D and 3D) are currently stillin the range of hours if many prediction points (large prediction areas) and many interactions are taken into account.

38

3. Intelligent Preprocessing of the Building Database

The visibility relations between walls and edges of the buildings stored in the database an independent of theposition of the basestation. Based on this consideration it is possible to accelerate the time consuming process ofpath finding by a single intelligent preprocessing of the database. This preprocessing is the basic idea of our newapproach. For this purpose all walls of the database are subdivided into tiles, all edges of the database are subdividedinto segments and also the prediction area is subdivided into a grid of receiver points as shown in figure 4.

The discretization of the database leads to a reduction of identical operations, because the ray tracing algorithmdetermines nearly the same rays for neighboring prediction points and for all these points the same computations arenecessary (reflection and diffraction points lie on the same walls and edges [5]).

After discretizing the database, the visibility relationships between all tiles, segments and receiver points an deter-mined in the preprocessing. The visibility relations an given by the line of sight criterion between the centers of thetiles (or segments). This leads to a simplification of the path finding problem, i.e. possible interaction points an thecenters of the tiles and segments, only.

If then is e.g. line of sight between a tile and a receiving point (see figure 4), the four connecting straight linesfrom the receiving point to the corners of the tile an considered. By projecting these four lines into two perpendicularplanes, four angles an determined which give an adequate description of the visibility relation. Similar computationsfor the visibility relations between tiles and tiles, tiles and segments, segments and segments and between segmentsand receiving points are performed in the following steps and an also stored in a file.

The projection of the connecting straight lines is very important, because by this operation a range of possiblereflection (or diffraction) angles for the illuminated tile (or segment) is defined. Also the angles continue on theneighboring tile respectively segment, so a very accurate compulation of the rays is possible even if the tiles orsegments are large (up to 5 or 10 meters, depending on the database) [5].

Tables 1 and 2 show the memory requirements and computation times for different urban scenarios and differentsizes of the tiles and segments. The computation times are smaller than the computation time needed for a singleprediction for the same area with the standard ray tracing (see table 3), because each visibility relation is only computed

39

once in the preprocessing while in the prediction the visibility relation might be considered and be computed severaltimes for different prediction points.

4. Prediction after Database Preprocessing

The result of the preprocessing of the building database is a tree structure containing tiles, segments and receivingpoints of the prediction area, as indicated in fig. 5. In this tree every branch symbolizes a visibility relationship betweentwo elements. For the prediction only the tiles, segments and receiving points, which are visible from the base stationhave to be determined. Additionally, the angles of incidence for the visible tiles and segments have to be calculated.Subsequently path finding can be done similar to the Ray Launching algorithm by recursively processing all visibleelements and checking if the specific conditions for reflection or diffraction are fulfilled. The ray search is stopped, ifa receiving point or a given maximum number of interactions is reached. Finally the field strength is summed up at allpotential receiving points. Preprocessing the building database reduces the time consuming path finding to the searchin a tree structure. A comparison between the number of branches in the first layer (determined in the prediction)with the number of branches in the remaining layers (determined in the preprocessing) in the tree structure givenin fig. 5 indicates the relation between the computational effort in the prediction and the computational effort in thepreprocessing.

Figure 6 shows the situation of a single reflected ray between the base station and an arbitrary receiving point.Both the transmitter and the indicated receiving point are visible from the center of the bright tile. Therefore, only theconditions expressed in equations (1) and (2) for the different angles have to be verified (see fig. 6). If they are fulfilled,a single reflected ray between the base station and the receiving point exists and its interaction point is assumed to liein the centre of the tile considered.

The stored visibility relations in the tree structure (all layers except the first layer) are independent of the transmitter

40

location and can be used for all predictions with the same database. Only the relations in the first layer of the treeare depending on the location of the transmitter and must be computed in the prediction process for each transmitterlocation.

The number of interactions influences the computation time because each new interaction corresponds to a fur-ther layer in the visibility tree. Very good results are achieved with a maximum of six interactions (reflections anddiffractions in different combinations with a maximum of two diffractions in each ray).

41

In contrast to the urban propagation model, the indoor model considers also the penetration of walls. For thispurpose the path loss of each tile is stored together with the visibility relation in the preprocessed file, so the algorithmsfor the prediction in the two different scenarios are nearly similar.

The computation times for different urban scenarios in comparison with a conservative 3D model [8] are presentedin table 3. They are gained with a maximum of 4 interactions (all combinations of reflections and diffractions with amaximum of two diffractions). Indoor scenarios are computed with similar acceleration factors.

The new approach combines the accuracy of ray tracing with the idea of ray launching. Like with ray launching, thenew model follows all rays from the transmitter to the receiver points. But in contrast to ray launching, the accuracyand the resolution are very high, because all rays and their points of interaction are determined in the preprocessingsimilar to ray tracing.

5. Transition between urban and indoor models

When different databases are used for the indoor and urban scenarios, a very simple interface can be implementedfor the prediction model, because the tiles of the walls surrounding the building can be used as interface between thetwo databases.

If the transmitter is placed outside the buildings, all rays and their angles of incidence are stored for each tile on thesurrounding walls of the building. The following computation of the indoor propagation is very easy with the indoortool, because it uses the information of the incident rays on the surrounding tiles and follows these rays on their waythrough the indoor visibility tree.

If the transmitter is placed inside the building, all rays reaching the tiles of the surrounding walls are stored andfollowed up later with the urban tool on their way through the urban visibility tree.

6. Comparison to measurements

In order to show the accuracy of the new prediction approach, comparisons with field strength measurements arepresented. As underlying test scenario, part of the city center of Stuttgart is chosen with two transmitting stationsoperating at 900 MHz. The transmitter antennas were mounted at 6 m respectively 5 m height, well below rooftops,which is a typical microcellular environment. Different measurement routes with both LOS and NLOS conditions areconsidered (see fig. 7 and fig. 9).

Figure 7 shows the differences between prediction and measurement for the first transmitter. The distribution ofthe differences shown in fig. 7 is presented in fig. 8. The mean error is 0.3 dB and the standard deviation is 5.8 dB.The prediction considers all classes of ray paths shown in table 4. For the discretization of the building database the

42

parameters shown in table 5 have been selected.In fig. 9 the differences between prediction and measurement for another transmitting station are presented. Figure

10 shows the distribution of these differences. The prediction has been calculated with identical parameters. In this

43

case the mean error is -3.5 dB and the standard deviation is 7.4 dB.These first comparisons between the new prediction approach and field strength measurements indicate that the

same accuracy as with the standard 3-D Ray Tracing can be obtained. With this new approach the delay spread andthe impulse response can also be predicted.

For indoor environments different benchmarks with measurement campaigns were used in different types of build-ings. New office buildings like those of the University of Stuttgart [9], and other kinds of office buildings were usedfor the comparison. The results concerning accuracy and performance are similar to those of the urban scenarios.

44

7. Conclusions

In this paper a fast and efficient ray optical propagation model for the planning of wireless communication networksis presented. The new approach consists of a single preprocessing of the building database. Therefore, the databasemust be discretized first and then the visibilitiy relations between the resulting elements must be computed and storedin a file. This process reduces the computational effort of the deterministic path finding considerably. The comparisonto measurements shows a good agreement of the new model. The new deterministic approach with intelligent prepro-cessing of the database combines the short computation times of empirical models with the accuracy of ray opticalmodelling. Additionally, it is possible to define a very simple interface between indoor and urban propagation modelsby using the visibility information of the preprocessed databases.

References

[1] WINPROP, Software tool (Incl. demo-version) for the Planning of Mobile Communication Networks and for thePrediction of the Field Strength In Urban and Indoor Environments. http://www.winprop.de, Jan. 2000.

[2] E. Damosso, ed., Digital Mobile Radio: COST 231 View on the Evolution towards 3rd Generation Systems.Bruxelles: Final Report of the COST 231 Project, published by the European Comission, 1998.

[3] A. J. Motley and J. M. Keenan, “Radio coverage in buildings,” Belt System Technical Journal (BTSJ), vol. 8, pp. 19– 24, Jan. 1990.

[4] K. Rizk, R. Valenzuela, S. Fortune, D. Chizhik, and F. Gardiol, “Lateral, Full and Vertical Plane Propagation inMicrocells and Small Cells,” in 48th IEEE International Conference on Vehicular Technology (VTC), (Ottawa),pp. 998–1003, May 1998.

[5] R. Hoppe, G. Wölfle, and F. M. Landstorfer, “Fast 3D Ray Tracing for the Planning of Microcells by IntelligentPreprocessing of the Database,” in 3rd European Personal and Mobile Communications Conference (EPMCC),(Paris), Mar. 1999.

[6] G. Durgin, N. Patwari, and T. S. Rappaport, “An Advanced 3D Ray Launching Method for Wireless PropagationPrediction,” in 47th IEEE International Conference on Vehicular Technology (VTC), (Phoenix, AZ), pp. 785 –789, May 1997.

[7] O. Landron, M. J. Feuerstein, and T. S. Rappaport, “A Comparison of Theoretical and Empirical ReflectionCoefficients for Typical Exterior Wall Surfaces in a Mobile Radio Environment,” IEEE Transactions on Antennasand Propagation, vol. 44, pp. 341–351, Mar. 1996.

[8] G. Wölfle, B. E. Gschwendtner, and F. M. Landstorfer, “Intelligent Ray Tracing – A new Approach for theField Strength Prediction in Microcells,” in 47th IEEE International Conference on Vehicular Technology (VTC),(Phoenix, AZ), pp. 790 – 794, May 1997.

[9] G. Wölfle, F. M. Landstorfer, R. Gahleitner, and E. Bonek, “Extensions to the Field Strength Prediction Tech-nique based on Dominant Paths between Transmitter and Receiver in Indoor Wireless Communications,” in 2ndEuropean Personal and Mobile Communications Conference (EPMCC), (Bonn), pp. 29 – 36, Nov. 1997.

Indoor Propagation Analysis Techniques for Characterizationof Ultra-Wideband RF Environments

by: David J. HallTime Domain Corporation

6700 Odyssey Drive Huntsville, AL 35806 USAEmail: david.hall@tdsi.com

AbstractModels of the propagation channel of Time-Modulated Ultra-wideband (TM-UWB) RF

transmissions were developed using a specially built receiver with two independent receivechannels. One channel was used to track the incoming signal, demodulate data, and collectstatistics on a bit error test pattern of length 32768. The second channel synchronized with thefirst channel and scanned through time to obtain a time domain representation of the incomingwaveform. This scanned waveform is representative of the actual distortion of the transmittedGaussian waveform after being filtered by the environment.

Simple statistics such as path loss as a function of distance in several testenvironments are discussed. This paper also characterizes the delay spread and showsperformance gain as a function of the number of correlators in the system. The optimal relativeplacement in time of multiple correlators, and the marginal energy gained by adding morecorrelators are discussed.

1.0 Overview

This document develops models of the propagation channel of Time Modulated Ultra-Wideband(TM-UWB) RF transmissions. A specially built receiver with two independent receive channels was usedto collect data [1]. One channel was used to track the incoming signal, demodulate data, and collectstatistics on a bit error test pattern of length 32768. The second channel synchronized with the firstchannel and scanned through time to obtain a time domain representation of the incoming waveform.This scanned waveform is representative of the actual distortion of the transmitted Gaussian waveformafter being filtered by the environment. By using this technique, precision measurement of the impulseresponse of the channel was limited only by the maximum communication range of the system.

Simple statistics such as path loss as a function of distance in several test environments arediscussed. Examples of this are given for data captured (1) down long hallways, (2) in classrooms andoffices, and (3) through walls of various materials such as drywall. cinderblock, etc. This paper alsocharacterizes the delay spread and shows performance gain as a function of the number of correlators inthe system.

46

2.0 Test Design

2.1 Environment

Data was collected in at least nine distinct environments characteristic of typical conditions foundin representative buildings.

Specific locations were selected with representative characteristics to provide a wide range ofenvironmental conditions.

2.2 Equipment

The test equipment consisted of a transmitter, receiver, two antennas, a laptop computer, andspeakers. Figure 1 shows a block diagram of the test configuration. Figure 2 shows the full equipmentsetup in operation.

47

The transmitter transmitted a known pseudo random bit error test pattern of length 32768. Aspecially built scanning receiver with two independent receive channels was used to collect data. Onechannel was used to track the incoming signal, demodulate data, and collect statistics on a bit error testpattern of length 32768. The second channel synchronized with the first channel and scanned throughtime to obtain a time domain representation of the incoming waveform. This scanned waveform isrepresentative of the actual distortion of the transmitted Gaussian waveform after being filtered by theenvironment and receiver processing. By using this technique, precision measurement of the impulseresponse of the channel was limited only by the maximum communication range of the system.

Figure 3 is a block diagram of the scanning receiver. Figures 4 and 5 show the receiver inoperation in two different environments.

48

2.2.1 CalibrationAn experiment was performed to investigate correlation of the output of the correlator to the

received power of the system to determine if this correlation was linear throughout the entire dynamicrange of the receiver. Also, once the correction factors relating the energy in the scan waveform to the

49

known received power were identified, it could be established for a given uncalibrated received waveformexactly how much received power was represented.

Calibration was based on establishing a link with a scanning receiver/transmitter pair inside ananechoic chamber. The power spectrum of the transmitter was measured using an average power meterand spectrum analyzer in averaging mode. All component insertion and VSWR parameters weremeasured and taken into account during post processing of the data. The received power was determinedby measurement with a preamplifier and spectrum analyzer in averaging mode. The measured receivedpower was compared to the predicted received power using Friis' transmission equation. Correlationbetween theoretical and measured was within an average 2.9 dB. In addition to the externalmeasurements, received time domain waveforms were processed and saved by the scanning receiver andpost processed by the computer program. Once the average power correction factor was entered into theprogram the time domain waveforms were processed and the difference between predicted average powerand measured was plus or minus 0.6 dB. The mean error was -0.0025 dB and the standard deviation was0.48 dB, establishing that the author’s algorithm for calculating received power was accurate throughoutthe dynamic range of the receiver.

3.0 Data Collection

The data collected was a collected in a series of files containing header information and twocolumns of data. The headers contain information such as position, pulses per bit, time window, numberof samples, transmit and receive attenuations, etc.) The first column contains the demodulated datachannel data, and the second column contains the scan waveform data.

One channel was used to track the incoming signal, demodulate data, and collect statistics on a biterror test pattern of length 32768. The second channel synchronized with the first channel and scannedthrough time to obtain a time domain representation of the incoming waveform.

4.0 Test Results

4.1 Data AnalysisEach file of data collected from each location is a time domain fingerprint of the RF environment

at that location. It reveals the different paths taken by the various RF reflections between the transmitterand the receiver. Figure 6 shows typical waveforms in three different environments.

50

51

4.2 Approximate Path Loss Coefficients

The path loss coefficient (denoted by n) is the number describing the decrease in received powerin dB when plotted as a function of the ratio of the distance between the transmitter and receiver, andsome arbitrary reference distance.

PL(d) is the mean path loss in dB at some distance d between the transmitter and the receiver. isthe mean path loss in dB at reference distance and n is the empirical quantity – the “path lossexponent”.[2] In calculating power, the author assumed 40 correlators were available.

Figure 7 plots all data from all locations, showing the received power in dBm as a function of distance inmeters. The heavy line is a best logarithmic fit to the data points. It corresponds to an n of 1.271. Noticethat the goodness of fit is low because the data is all sampled locations, irrespective of attenuating barriersor environment. When all the data is divided into classes based on environment, much better curve fitsresult.

52

4.3 Root Mean Square Delay Spread

Root mean square (RMS) delay is a metric of the “ringing time” of the pulse It is described asfollows:

where

n=1,2 [3], [4]

53

Figure 8 plots all data from all locations. The best linear curve fit (y = -0.0005x + 11.567) wouldtypically increase, but the standard deviation of this fit is so much larger that no definite conclusions canbe drawn for the data taken as a whole. RMS delay spread doesn’t appear to be a function of distance.However, further analysis will be conducted to see if there is a relationship to number of walls or materialproperties.

4.4 Typical Attenuation through Barriers

A major concern is the effect of various barriers on the signal power. Readings were taken inenvironments with a number of different barriers. The author formed rough estimates based onobservations of applicable power vs. distance plots.

54

4.5 Marginal Value of Correlators

Another issue is identifying the number of correlators needed to capture a specific amount ofenergy in the received waveform. Figure 9 shows the amount of energy captured by some number ofcorrelators divided by the amount of energy captured by 40 correlators. The correlators have alreadybeen sorted so that the first correlator captures the most energy, the second correlator captures the secondmost energy, etc. An attempt was made to perform an exponential fit to the following equation:

is a decay constant, approximately equal to the number of correlators required to capture 1-1/e = 63%of the energy available in the waveform. Generally, the fit was quite good. The typical function for theaverage plus one standard deviation of and minus one standard deviation of are plotted below.Figure 9 is a plot of how much available waveform energy is captured as a function of the number ofcorrelators used to capture it. For example, considering the average values, about 6 correlators arerequired to get to within 3 dB of the total available energy (illustrated by the square series in figure 9).However, 68% of the samples fall between plus one sigma and minus one sigma of the values samples .(These would fall between the diamond series and the triangle series in Figure 9). They will capturewithin 1.3 to 4.5 dB of the available energy.

55

Figure 10 shows the discussed in section 4.3 plotted as a function of RMS Delay Spread. Thefit is quite linear, confirming that the more ringing there is in the waveform, the more correlators arerequired to capture a specific fraction of the available energy.

4.6 Issues

to near the ambient noise level. In the future, at least 100 – 150 nS of data should be captured. Also,longer records will allow better comparisons to narrowband propagation models.• Distances didn’t vary greatly. A wider range of classroom data at varying distances andlocations, not just about a circular pattern around the edge of the room, would have allowed better curvefitting.

5.0 ConclusionsThe test results provided a high level description of the overall stochastic properties of an ultra-

wideband environment using such common metrics as RMS Delay Spread and Path Loss Coefficients, aswell as newer metrics such as the marginal correlator coefficient

Test results that were identified by specific environments enabled the construction of a muchmore accurate stochastic model.

• The longest waveform captured was 70 nS. Sometimes, the ringing still had not decreased down

56

6.0 AcknowledgementsMany thanks to James Mann and his numerous assistants for their patient efforts collecting the

data, Bill Beeler for his assistance in the calibration procedures, Paul Withington and Rachel Reinhardtfor their insights into propagation and UWB in general, and Susan Phelan for her assistance in not onlythe technical writing but a prodigious amount of number crunching. Also Bill Beeler for his patience withthe calibration procedure and giving me extra help on the spreadsheet minutiae.

7.0 References[1] Withington, Reinhardt and Stanley, “Preliminary Results of an Ultra-wideband (Impulse) ScanningReceiver”, Paper S38P3, Milcom 1999, Atlantic City, NJ, November, 1999

[2] Greg Durgin, Theodore S. Rappaport, and Hao Xu, “5.85 GHz Radio Path Loss and Penetration LossMeasurements In and Around Homes and Trees,” IEEE Communications Letters, Vol 2. No. 3, March1998, pp 70-72

[3] Theodore S. Rappaport, Keith Blankenship, and Hao Xu, “Propagation and Radio System DesignIssues in Mobile Radio Systems for the GloMo Project” Mobile and Portable Radio Research Group.Bradley Department of Electrical and Computer Engineering. Virginia Polytechnic Institute and StateUniversity. Revised January 31, 1997.

[4] Kaveh Pahlavan & Rajmani Ganesh & Steven J. Howard, “Wideband Frequency and Time DomainModels for the Indoor Radio Channel,” IEEE Global Telecommunications Conference Globecom , 1991,p 1135-1140 Vol.2.

Propagation Signatures to Characterise Wideband Environments

Gregory MartinVictoria University of Technology

Mobile Communications & Signal Processing GroupPO Box 14428MC, Melbourne City, 8001

Victoria, Australiaemail: greg@cabsav.vu.edu.au

Abstract

A propagation signature is introduced to summarise wideband propagation measurements. Thedisplay indicates the probability of rays occurring in a matrix of excess delay and power bins, andusing a contrasting sequence of colours on a two-dimensional plot, provides a vivid visual sum-mary of propagation conditions in a particular environment.

1. Introduction

Presentation of the large amount of data collected during wideband propagation mea-surement sessions poses a dilemma. Each environment area may be characterised by a large num-ber of power delay profiles (PDPs), which show wide individual variations, but overall representthe propagation statistics for the area. While individual PDPs are interesting to examine, thesescarcely convey the overall statistical picture. On the other hand, an average PDP over an arealoses a lot of the available information. Single parameter measures, such as rms delay spread,average delay, delay window and K factor {Ref. 1}, provide useful summaries of signal time dis-persion. Distribution plots of these single parameter measures give a valuable quantitative pictureof the area statistics, and show the spreads, and maximum values, of these measures. But in addi-tion, there remains a need to present a concise picture of multipath activity.

A novel summary of the distribution of multipath rays over power and excess delaybins is proposed, providing a vivid and graphic fingerprint of the multipath propagation excessdelay statistics for a particular environment. (Note: the method relies crucially on the use ofcolour to convey information; when restricted to grey-scale, much of the impact is lost).

2. The Propagation Signature

A novel graphic summary of the distribution of multipath rays over power and excessdelay bins is introduced in this section. This gives a vivid pictorial signature or fingerprint of themultipath propagation excess delay statistics for a particular environment. Akin to a contour map,and plotted as a two-dimensional graph using colour and position to convey propagation parame-ters, an example of the Propagation Signature is shown in Figure 1.

58

Normalised PDPs are analysed to identify individual rays, which are then sorted intoboth power and excess delay bins. The time bins are 50 samples wide, or in outdoor mode,and range from to ranging from to for indoor mode).

Power bins are 2dB wide, and extend from 0dB to -24dB. Thus there are 59 time binsand 12 power bins, giving a matrix of 708 bins. Because the dynamic range of individual PDPsvaries, the power range is chosen as large as possible, consistent with the desire to minimise thenumber of PDPs discarded on the basis of inadequate dynamic range. Profiles with less than 24dBdynamic range are discarded.

The total number of rays in each bin is divided by the total number of PDPs , giving theaverage number of rays per PDP in each bin. Previously each PDP has been normalised to thestrongest ray, which furnishes the origin for power (0dB) and excess delay time

Of course the strongest ray may not be the first to arrive, so the normalisation schemecan result in rays with negative excess delays. Remember that the time scale is relative to thestrongest ray, not the direct path. By definition, the [0 to 0 to -2dB] bin will have an aver-age number of rays/bin of The average number of rays/bin can be interpreted as the probabil-ity of a ray occurring in the bin. If the ‘probability’ exceeds 1, this implies more than 1 ray/bin onaverage. Even though the channel sounder allows up to twelve resolvable rays per time bin,

59

apart from the origin bin, probabilities >1 have seldom been found.A three-dimensional (3D) mesh plot of ray probability (z axis) versus the power, time

bin matrix (xy axes) is generated, with z ranges depicted with a diffused colour (replaced by grey-scale in this printing) band legend. The graph-plotting software performs linear interpolationsbetween bins, and colour diffusion effectively provides further interpolation. Data for power bin[0 to -2dB] is labelled “0dB” etc., with the final [-22 to -24dB] bin labelled “-22dB”.

A conventional oblique 3D view always obscures part of the mesh plot, no matterwhich viewing angles are chosen. However, viewing the plot from vertically above, with perspec-tive turned off, results in the contour map style presentation shown in Figure 1. Colour banding isassigned on a non-linear scale, to show more detail for low ray probabilities. The result is a dis-tinctive, informative and aesthetically pleasing signature of the propagation environment. Someexamples of the propagation signature follow, based on a selection of outdoor propagation mea-surements in the Australian cities of Adelaide, Melbourne and Sydney, and an indoor propagationexample.

A description of the VUT channel sounder is given in Ref.2, and more details of theoutdoor measurements may be found in Ref.3.

3. Adelaide

Adelaide signatures are shown in Figure 2. From the Adelaide Hotel transmitter loca-tions, tall city buildings are from 1.4 km to 3.5 km to to the south. With theantenna on the roof, measurements were made to the north of the hotel, moving away from thecity, and also in the city itself. The strong direct path predominates, with no powerful paths atlarge excess delays. Some activity occurs at -10dB, some at -16 to -22dB, and at-12 to -24dB. All these excess delays may be explained by reflections from tall city buildings.

Locating the transmitter antenna protruding from a 1st floor window on the south (cityfacing) side of the hotel, and with the receiver moving through the Torrens River valley below,and into the city area, shows strong multipath activity at excess delays of and 10 to

The long excess delays range between -2 to -24dB, and could occur in the Torrens valleywith a reflection from the north edge of the city. Some paths at are also present.

The final measurement location, with a low transmitter antenna in Victoria Square inthe centre of the city, is notable for negative excess delay paths, some at relatively high powerlevels, up to 0dB. These range from to and represent the first arriving ray, showingthat the strongest ray is arriving by a more indirect path. There is only slight low power activity atexcess delays greater than

60

61

4. Melbourne

Inner suburban measurements based in Delbridge Street gives a signature with little

dispersion and no large excess delays using the below-rooftop low transmitter antenna (balcony).Raising the transmitter antenna above rooftop level increases time dispersion markedly. Rays arepresent at (-8dB), equivalent to an excess path of 2.9 km, and at and Thestraight line distance from the Delbridge Street transmitter to the comer of Lonsdale and Swan-ston Streets in the city area, is 3.1 km. For a receiver location on Johnston Street, the excess pathto city buildings in Lonsdale Street is 3.3 km, equivalent to Excess paths across the wholecity may produce excess delays over a range of approximately to The city skyline (seeFigure 4) is thus at the correct distance to explain rays at long delays in the Delbridge Street roofsignature.

62

5. Indoor Measurements

Using a channel sounder with 10ns resolution {Ref.l}, measurements were madeinside a large seven storey concrete university building called D block {Ref.4}. The transmitterwas located on the top floor (Level 7), not far from an outside window which gave views of othercampus buildings. One set of measurements was made on the same floor as the transmitter, mov-ing down a long central corridor (about 70 metres in length). The other set was taken two floorslower, on Level 5, while the transmitter remained in the same position on Level 7.

The signature for Level 7 shows activity up to and some activity at excess delaysof about and The longer delays indicate reflections from external structures. Belowon Level 5, a lot of rays are arriving up to before the strongest rays, and many rays arepresent, sometimes at high power levels, up to after the strongest rays, which is aboutafter the first arriving signals. excess path corresponds to 300 metres, or a a reflector 150metres distant, and several other campus buildings lie within this range. On Level 5, the mostdirect paths are suffering floor and wall attenuation, and the strongest signals arrive after reflec-tion from outside buildings.

63

6. Sydney

The McMahons signature comprises measurements clustered around Blues Point on thenorth side of Sydney Harbour, shadowed below cliffs from the elevated 4th floor McMahonsPoint transmitter location. Because of the heavy shadowing, many of the reflected paths are asstrong or stronger than the direct ray. Specular reflectors abound, including the metal SydneyHarbour bridge structure, Luna Park across Lavender Bay, and high-rise city buildings on thesouth shore of the harbour. The signature shows strong rays between 0 and with the longerdelays corresponding to the north end of the bridge, about 700 metres distant. There are other raysat (-1dB to-6dB), and

The excess path corresponding to the northern edge of the high-rise city on the southside of the harbour is approximately The south edge of the city (University of TechnologySydney building, just south of the downtown city area) could give excess paths ranging up toabout so any rays with excess paths in the range to may be caused by high-rise

64

city buildings. This environment is the most dispersive of any measured in Australia, and there isconsiderable evidence that in many cases, the strongest ray is not the first arrival.

7. Results Summary

Results are summarised in Tables 1 and 2, which give rms delay spread, and meandelay values, not exceeded in 25%, 50% and 90% of cases, and also the maximum values of theseparameters observed during the tests. The maximum straight line separation between transmitterand receiver for each transmitter location is listed. Any power delay profiles not exhibiting at least20 dB signal above the noise floor were discarded prior to calculation of the results presented inthe tables.

While the tables provide a quantitative summary of propagation statistics, the propaga-tion signatures allow a rapid visual comparison of the different propagation environments.

65

8. References

[1] Theodore S.Rappaport, “Wireless Communications Principles and Practice”, IEEE Press, Prentice-Hall, New Jersey, 1996.

[2] G.T. Martin and M. Faulkner, “Delay Spread Measurements at 1890 MHz in Pedestrian Areas of theCentral Business District in the City of Melbourne”. Published in the Proceedings of the IEEE 44th.Vehicular Technology Conference (VTC’94), Stockholm, Sweden, June 8-10, 1994. Volume 1, pages145-149.

[3] G.T. Martin and M. Faulkner, “Wide Band PCS Propagation Measurements in Four AustralianCities”. Published in the Proceedings of the 10th.International IEE Conference on Antennas andPropagation (ICAP’97), Edinburgh UK, 14-17 April, 1997, Volume 2, pages 199-203.

[4] G.T. Martin and M. Faulkner, “PCS Ray Characteristics Between Multiple Floors of a ConcreteBuilding”. Published in the Proceedings of the 47th. International IEEE Conference on VehicularTechnology (VTC’97), Phoenix USA, May 1997, Volume 3, pages 1400-1404.

Smart Antennas for CDMA Cellular and PCS Networks

Scot Gordon, Marty Feuerstein, Donn Harvey, Michael ZhaoMetawave Communications

10735 Willows Road, NERedmond, WA 98052

Abstract

Smart antennas have been in existence for some time, but only in the past few years havethey gained acceptance in commercial cellular and PCS networks. For CDMA, thecommercial smart antenna is non-traditional as it synthesizes sectors of varying azimuthsand beamwidths to equalize loading and improve sectorization efficiency. This can beconceived as beamforming over an aggregate of users which is unlike a more traditionalsmart antenna which may beamform or beam-switch on a per user basis. This paperpresents the uses and advances of this form of smart antenna. Specifically, it looks at itsability to support an increasing number of sectors due to the extra degrees of freedom,and large aperture that an antenna array provides. Further, it outlines how the smartantenna supports continuous optimization by dynamically adjusting the sector sizes andorientations based on traffic loading estimates.

1. Introduction

Explosive growth in the cellular and PCS industry has left operators in a challengingposition; keeping network capacity in line with consumer demand. Recently,unprecedented demand for data services, such as internet browsing and email, has led to anew paradigm in cellular and PCS; wireless data services. This shift from voice-centricnetworks to both voice and data accentuates operators’ difficulty in meeting networkcapacity needs. One solution to this dilemma is the smart antenna. Smart antennas havebeen around for many years for military applications but only recently have gainedacceptance for cellular and PCS applications. Traditional cellular and PCS networks arestatic, configured so that the network cannot adjust to the ever changing traffic patternsand interference environment. Smart antennas, in all forms, control interference in some

68

way. This is typically done through the adjustment of radiation patterns. For example,the smart antenna may switch from a set of fixed beams or adaptively form beams. Thebeamforming or beam switching decision has traditionally been done on a per user basismaximizing a specific mobiles carrier to interference ratio. However, recently a newbreed of smart antennas for CDMA networks, referred to as sector synthesis [1], formsbeams that optimally define sectors. This “beamforming” across an aggregate of usersmaximizes the sectorization efficiency by controlling interference, managing handoffoverhead and equally dividing the load across sectors. Further, this beamforming can bedynamic, reconfiguring itself based on the changing traffic conditions. As another benefitthe increased flexibility provided by a smart antenna provides the cell with a smoothtransition to increased sectorization, which, if configured appropriately, primarily throughsoftware changes, can support three, four, five or six sectors. This paper explores thesebenefits by summarizing results of a four, five and six sector deployment and detailingthe method of dynamic sectorization.

2. Sector Synthesis

The majority of all cells in a cdmaOne network are sectorized in an effort to increase thecapacity of the cell. In fact, going from an omni cell to a three sector cell may offernearly a three fold capacity improvement. However, rarely is this near three-fold

increase realized because of unequal and ever changing traffic distributions, and largehandoff regions between sectors. Optimal sectorization efficiency requires that trafficload is evenly distributed across sectors and handoff overhead is minimized. In theworst case, the traffic distribution would be such that one sector carries 100% of the loadwhile the other two sectors remain completely idle providing no increase in capacityassociated with sectorizing the cell. Although this scenario is unreaslistic, it is very likelythat a sectorized cell will have large traffic imbalances across sectors. Further,traditional deployments use sector antennas with beamwidths on the order of 80 or 90degrees. These somewhat small aperture antennas have a slow rolloff of the mainlobecausing energy to bleed into adjacent sectors increasing both interference and handoffoverhead.

69

The sector synthesis concept [1] uses three panel antennas each consisting of an antennaarray. The phase and amplitude of the array elements are software controlled allowingthe operator to change the beamwidths and azimuths of each sector. So, unlike somemore traditional smart antennas where beams are formed on a per user basis, the sectorsynthesis approach beamforms across an aggregate of users. For CDMA, the benefit ofthis approach is the beamforming is done at RF, allowing the smart antenna to beimplemented as a non-invasive add on to an existing base station. Other approaches,such as beam switching and adaptive beamforming, require the smart antenna to be moretightly coupled to the base station because most of the smart antenna functions must beimplemented at the baseband level.

The flexibility offered by the sector synthesis enables the site to be configured so thattraffic loads are evenly distributed. Further, the antenna array has a large aperture withrespect to traditional sector antennas improving the rolloff of the mainlobe and hencedecreasing adjacent sector interference and softer handoff activity. The result isimproved sectorization efficiency. Capacity improvements differ depending on thecharacteristics of the cell, however, the average is about a 40% improvement. Table 1lists some examples from different networks. Detailed capacity calculations and resultsare presented in [2].

70

3. Flexible Sectorization

Load balancing is an effective means of improving the capacity of a cell site given thatthe site is imbalanced to begin with. Although this is most often the case, there areinstances where the cell is reasonably well balanced. In such instances increasing thesectorization of the cell is a viable alternative. Historically, moving beyond a threesector cell has been problematic. Pilot pollution is exascerbated and handoff activityincreases reducing the capacity improvement and while increasing the optimizationeffort. Further, installations where analog and CDMA systems share antennas makes re-orienting and installing additional antennas to support the increased sectorizationimpractical.

The smart antenna is a natural enabler of increased sectorization. Because their exists anantenna array with the ability to form narrow beams, implementing four, five or six sectorconfigurations is generally a straightforward change done through software. Further,with the added ability to sculpt the patterns by phasing and weighting the elements in the

antenna, pilot pollution is minimized.

As an example, we look at a recent deployment where a three sector cell was converted toa four, five and six sector cell. The cell of interest is located in a busy network and isconfigured with two CDMA carriers. Users on the cell are assigned the first CDMAcarrier unless the first carrier, F1, is determined to be overloaded, in which case a user isassigned the second carrier, F2. Thus, any call completions that occur on the F2 areblocked by the F1 carrier. This provides a convenient mechanism to measure thecapacity improvements offered by the different sectorization schemes. Defining ascalls completed on F1 and as call completed on F2 the Grade of Service (GOS) is

A low GOS for a given traffic level indicates that most of the traffic is carried by F1 andlittle blocking would exist if a second carrier were not present. Likewise, a high GOS

71

indicates much heavier loading on the second carrier and a large level of blocking wouldexist if the second carrier were not present.

By trending the GOS versus call completions we get a indication of the efficiency of thecell under the baseline three sector scenario and the smart antenna implementation offour, five and six sectors. A trend line is computed by assuming the blocking rates followthe Erlang B model,

where C is the number of trunked channels offered by a trunked radio system and A is thetotal offered traffic in Erlangs. COMA is not a fixed trunk air interface as the number ofavailable traffic channels varies dependent on, among other things, the forward linkpower requirements and handoff state of those mobiles currently served by the cell.However, over time there should be an average number of “available” traffic channels

and hence the Erlang B provides an approximate GOS measure for CDMA. The numberof trunked channels in the Erlang B model, C, is selected in order to minimize thesquared error between the observed GOS values at a given traffic level and the predictedGOS of the Erlang B model at that same traffic level. The offered traffic, A, is obtainedby equating it to the number of call completions in a given hour, and assumingeach call completions lasts 90 seconds, i.e.,

Figure 1 displays both the observed data and the corresponding trends. Using thecomputed trends, the capacity improvement offered by the four, five, and six sectorconfigurations over that of the baseline three sectors configuration is computed anddisplayed in figure 2. Using this figure, we see the four, five, and six sectorconfigurations have improved the capacity handling capabilities of the F1 carrier. Thecapacity improvement of the six sector ranges from nearly 85% at a 0% GOS and 70% ata 5% GOS. At a reasonable GOS of 2% we see a 73.6% increase in capacity. Likewise,at a 2% GOS the the four and five sector configuration demonstrate a 48.7% and 53.7%increase in capacity respectively.

72

73

Another important measure of performance are quality metrics. Poor performance in theform of access failures or lost call rate is an indicator of pilot pollution, which maybecome more problematic as the number of sectors increases. Table 2 displays these twometrics for the various configurations along with the handoff overhead, which is anotherindicator of pilot pollution. Both lost call rates and access failure rates remain consistentwith the baseline levels with only the five sector configuration showing a slight increasein the lost call rate. Further, the handoff overhead increased only a modest amount in thefive and six sector configurations despite the presence of additional handoff boundaries.These factors combined lead to the conclusion that the smart antenna provides enoughconfiguration flexibility and sector-to-sector isolation to keep pilot pollution undercontrol.

3. Continuous Optimization

Optimal CDMA capacity for a given cell requires that the traffic load be evenlydistributed across all available sectors. Doing so equalizes the utilization of each sectorand maximizes the sectorization efficiency. The static approach of synthesizing sectorsizes and orientations is effective especially when optimized for a particular time of day(i.e. busy hour) when traffic loads tend to be more predictable from day to day.However, the dynamic nature of mobile traffic can lead to changing sector loads on bothshort and long term time scales. In such cases a fixed optimization strategy will be sub-optimal. A CDMA smart antenna provides the capability to adjust its sectorization basedon measurements of traffic loading.

74

These measurements of load include uplink received signal strength and downlink pilot tointerference ratio. The uplink received signal measurements are converted to aloading estimate by comparing the levels to the thermal noise floor providing a noise riseestimate from which the percent loading of pole capacity is

where N is the noise rise [3]. Each panel antenna contains a four element array and hencefour nearly disjoint 30 degree beams can be formed on the uplink through a butler matrix.Because the smart antenna consists of three panel antennas, twelve beams are formed thatspan the circumference of the cell on the uplink. This provides twelve noise riseestimates with 30 degrees of resolution. On the downlink the pilot to interference ratio ismeasured on a per sector basis and hence the resolution of each sector is dependent on the

current size of that sector.

Figure 3 displays the noise rise versus beam number. Reverse link load is highest on

beams 2 and 3. If a single sector is currently serving this region a decision is made toreconfigure the site so that a single sector is either serving only beams two and three oralternatively, a sector boundary is placed between beams two and three sharing the loadbetween two sectors. In this way, we continually optimize the site so the load is evenly

distributed.

In almost all cases, CDMA cells are forward link limited and hence, load balancingbased on a reverse link metric may produce sub-optimal results. The alternative is to usea forward link measure. The pilot to interference ratio is the most appropriate candidateas it provides a direct forward link measure of loading on each sector. Further, somebasestation types will block calls on a particular sector when the falls below somethreshold. By configuring the sectors so that the minimum pilot to interference ratios ofthe available sectors is maximized produces an optimal sectorization scheme andminimizes blocking. Figure 4 shows a temporal plot of filtered samples, indicating

time periods of heavy traffic and sector-to-secotr imbalance. In this case, the gammasector exhibits heavy loading and presents an opportunity to offload traffic into adjacent

75

sectors by adjusting its sector size. Doing this improves the capacity of the cell byreducing the air-interface blocking that would be incurred on the gamma sector.

76

Conclusion

The CDMA smart antenna has proven itself capable of improving sectorization efficiencyby balancing the traffic load and decreasing handoff overhead. For cells which alreadyhave good efficiency the smart antenna provides a seamless transition to four, five or sixsectors demonstrating a 74% capacity improvement over that of the existing (baseline)

three sector system. Further, as networks migrate to a more continuous optimization, thesmart antenna enables sector changes on the fly, responding to loading conditions asmeasured either on the reverse link or the forward link.

References

[1] M. J. Feuerstein, J. T. Elson, M. A. Zhao, S. D. Gordon, “CDMA Smart AntennaPerformance”, 1998 Virginia Tech Symposium, June 1998, Blacksburg, VA.

[2] S. D. Gordon, M. J. Feuerstein, M. A. Zhao, “Methods for Measuring and OptimizingCapacity in CDMA Networks Using Smart Antennas”, 9th Virginia Tech WirelessPersonal Communications Symposium, June 1999, Blacksburg, VA.

[3] K. S. Gilhousen, I. M. Jacobs, R. Padovani, L. A. Weaver and C. A. Wheatley, "Onthe capacity of a cellular CDMA system," IEEE Trans. Veh. Tech. VT-40(2) 1991.

[4] M. J. Feuerstein, “Applications of Smart Antennas in Cellular Networks”, IEEE AP-SInternational Symposium, Special Session on Wireless Antenna Systems &Applications, July 1999, Orlando, FL.

Key techniques realizing smart antenna hardwarefor microcell communication systems

Keizo Cho, Kentaro Nishimori, Yasushi Takatori, and Toshikazu Hori

Nippon Telegraph and Telephone Corporation1-1 Hikarino-oka, Yokosuka, 239-0847 Japan

cho@ieee.org

AbstractThis paper summarizes key techniques in actualizing smart antenna hardware and proposes a new

calibration technique as well as antenna and base station configurations. The effectiveness of the proposedtechniques are evaluated based on field tests employing a prototype of the smart antenna, which receives thesignals of a Japanese commercial microcell system (PHS).

1. IntroductionDue to the recent popularity of multimedia and mobile communications, wireless

communication systems are required to push the limits towards higher data rates, greater reliability,and greater channel capacity [1][2]. Smart antennas have shown great potential toward combatingco-channel interference [3] and increasing the channel capacity [4][5][6], as evidenced by the manystudies in this research area. Previous studies on smart antennas have been mainly focused onalgorithms for controlling antenna patterns [3][4][5][6][7]. Recently, some smart antenna testbedshave been developed and measured results have been collected in actual propagation environments[8][9]. However, there still exist problems that must be overcome to actualize the smart antennahardware for commercial wireless communication systems such as efficient calibration, elementarrangement, and effective base station configurations.

This paper first summarizes the design parameters for actualizing smart antenna hardware forwireless communication systems. Then a new calibration technique, antenna arrangement, andbase station configuration are proposed that are suited to a base station adopting the smart antennatechnique for microcell communication systems. The effectiveness of the proposed techniques isevaluated by measurements using a prototype of the smart antenna and computer simulations.

2. Design parameters of smart antenna hardwareFigure 1 shows a typical configuration of the smart antenna hardware. The antenna comprises

a radiation part, transmitter and receiver part, and a digital signal processing part. Beam forming isnow usually carried out by digital signal processing due to the tremendous progress of the digitalsignal processing devices and the flexibility of the beam control. The considered design parametersof the smart antenna hardware are listed below.

78

Radiation partElement patternInter-element spacingElement arrangementNumber of elements

Transmitters and receiver partsCharacteristics of individual transmitter and receiver (Calibration)Number of branches

Digital signal processing partBeamforming algorithmSynchronizationCalculation complexity

The radiation part should be designed by considering the propagation environment. Reducing thenumber of elements or branches contributes to a decrease in the hardware complexity and cost.Since propagation characteristics depend on the operating environment, an appropriate smart antennaconfiguration should be based on the operating environment and systems. The individualcharacteristics of the transmitter and receiver influence the error in tracking and nullifyingperformance of the smart antennas. Moreover, the characteristics of the transmitter in particularchange over time due to temperature variations [10]. Thus, the transmitter and receiver must befrequently calibrated to equalize the characteristics. In the digital signal processing part, since theminimum mean square error (MMSE) algorithm [11] is generally used for beam control, carrierand timing synchronization are required before the beam control is initiated. However, establishingsynchronization is difficult when relatively strong interference is received at the antenna, for which

79

a smart antenna is generally required. To improve me synchronization performance, pre–beamformingis generally adopted to improve the initial input signal-to-interference and noise ratio (SINR) 12].Furthermore, an algorithm with fewer operations reduces the hardware complexity of the digitalsignal processing devices and power consumption.

In the following section, some techniques are proposed for actualizing the smart antennahardware. First, antenna configurations are described that are suitable for base stations placed atlow and high locations on a street microcell. Next, an automatic calibration method using atransmitting signal (ACT) is presented, which enables real-time calibration. Then, a new basestation configuration is described that is suited for elevated base stations in microcell systems andthe field test results of the proposed configuration are presented for an actual microcell environmentFinally, an antenna arrangement is described that is suited to a smart antenna employing space–divuion-multiple-access (SDMA) [13].

3. Techniques for actualizing smart antenna hardware

Commercial systems require a low cost and simple configuration while the performancereducing the multipath waves depends on the number of elements and antenna configuration. Aconfiguration that is appropriate for an operating environment must be used, and hardware complexitymust be reduced. The number of elements can be determined by considering the propagationenvironment.

Figure 2 shows measured angle of arrival (AOA) when a base station antenna was located atlow antenna height in a street microcell. Figures 2 (a) and 2 (b) represent the measurement scenarioand measured AOA, respectively. As can be seen in Fig. 2 (b), the long delayed waves that degradetransmission performance propagate as shown in Fig 2 (a). In the environment, a broadside array isplaced perpendicular to the street, and approximately 2 wavelengths is selected as the elementspacing suitable for the antenna array of the smart antenna. This configuration can discriminate the

3.1 Base station antenna arrangement for a street microcell

80

desired signal from the multipath waves with a small number of elements. Although the wideelement spacing produces a grating lobe in the radiation pattern, no ray with a long delay is receivedby the grating lobe. Thus, the grating lobe does not degrade transmission performance. Figure 3shows the bit error rate (BER) performance when using the smart antenna with a broadside array.The cases in which the element spacing of 0.5 and 2 wavelengths are shown and the number ofelements is 3. The figure confirms improvement in the BER by the wide element spacing. Figure4 shows the relationship between the element spacing and the input SNR improvement obtainingthe BER of The improvement is based on the input SNR of the antenna with a 0.5 wavelengthelement spacing. As shown in Fig. 4, when the element spacing is more than 2 wavelengths, theimprovement is almost constant and about 4 to 5 dB. Since the antenna size becomes large as theelement spacing increases, the appropriate element spacing is about 2 wavelengths.

Figure 5 shows measured AOAs in the vertical plane when a base station was located on therooftop of a building in a street cell environment. Figure 5 (a) represents the measurementenvironment, and Figs. 5 (b) and 5 (c) represent the measured AOAs when a terminal was locatedat 50 m (Location A) and 300 m (Location B) from the base station antenna, respectively. Asshown in these figures, long delayed waves propagate when terminals are located near base stations.The waves are reflected or diffracted by the surrounding buildings and arrive at the base station atvertical angles between 0 (horizontal) to -10 degrees. When the terminals are located far from thebase stations, such long delayed waves does not reach the base stations. Therefore, controlling thebeam in the vertical plane can suppress long delayed waves without reducing the cell size. Figure6 shows the effectiveness of the proposed antenna configuration. Figure 6 (a) represents the delayspread comparison when using an ordinary collinear antenna, a tilt beam antenna, and the proposed

81

smart antenna. Figure 6 (b) represents the relationship between the distance from the base stationand the propagation loss considering the antenna pattern. As Fig. 6(a) shows, the proposed antennaobtains a similar degree of multipath reduction as the tilt beam antenna, while Fig. 6 (b) shows thatthe proposed antenna enlarges the cell size 1.4 times greater than that of the tilt beam antenna.

3.2 Smart antenna calibrationDigital beamforming (DBF) configurations are generally used to form the smart antennas.

Thus, the difference in the characteristics of the antenna, cables, and transceiver at each branchvaries the weights from those at the base band. Therefore, calibration to equalize the transmissioncharacteristics of all branches is required at the smart antenna. Figure 7 shows the proposed

82

configuration of ACT. The calibration method is suitable for a base station used in time-division-

and receiver, respectively. In this configuration, transmission signals are divided and circulated tothe receivers. Transmission characteristics and are obtained by theloops shown in Figs. 7 (a) and 7 (b), respectively. By dividing by

can be obtained, which is a calibration value required for the smart antenna of TDDsystems [14]. Since ACT uses transmission signals as a reference signal for the calibration, noadditional signal generator is needed for the calibration. In addition, since the receivers are idolwhen transmitting signals in TDD systems, the calibration process can be performed withoutinterrupting services. Figure 8 shows the radiation pattern with and without using ACT. Thepattern was taken by using a prototype of a smart antenna in an anechoic chamber. The photographof the prototype is shown in Fig. 9. As Fig. 8 shows, ACT can direct the null toward the interferenceand improve the desired signal to undesired signal ratio (DUR) to 30 dB greater than that without

duplex (TDD) systems. and represent the transmission characteristics of the i-th transmitter

83

calibration.

3.3 New elevated base station configuration for a microcell system using TDDIn microcell systems, base stations are usually placed at a low height. In this case, however,

the cell formed by the base stations is mainly in the line-of-sight region along a street. Thus,numerous base stations must be placed to cover large service areas. An elevated base station canform a fairly large cell because propagation loss decreases when antennas are elevated. However,since the propagation loss of interfering signals also decreases, co-channel interference becomes aproblem. It is well-known that smart antennas can effectively suppress such co-channelinterference[4][5]. On the other hand, base stations usually broadcast signals at a constant intervalto notify each terminal of the current cell, and the signals must be broadcast in a uniform direction.However, smart antennas nullify the directions of the interfering base stations, causing the basestation employing the smart antenna technique to lose its ability to cover the area of the null direction.Furthermore, the null is automatically formed, making it a problem for cell planning. To overcome

84

this problem, we propose a new base station configuration for use in a TDMA system. Figure 10shows the configuration of the proposed base station. This base station comprises two sets oftransceivers, and each works in a different time slot. The pattern of both sets is controlled by acommon pattern control unit. The proposed base station first nullifies co-channel interference byusing a set of transceivers as shown in Fig. 11. Then, the other set of transceivers constrains itsbeam direction to the null direction of the former transceiver and nullifies other co–channelinterference at different time slots from that nullified by the former transceiver. The latter transceiverworks at a different timing from the former one. Therefore, the proposed base station simultaneouslysuppresses the interference and achieves a circular cell as shown in Fig. 11. Figure 12 shows theeffectiveness of increasing the channel capacity by the proposed base station measured in an actualenvironment when applying it to a Japanese commercial microcell system (PHS). The prototypeshown in Fig. 9 was used for the measurement and the data was taken in Tokyo. As the figureshows, the proposed base station obtains about 5 times the available time slots than that of anordinary base station (omnidirectional antenna). Figure 13 shows the measured radiation pattern ofthe proposed base station configuration in the environment. We confirmed that the proposedconfiguration achieves a circular cell while suppressing co-channel interference.

3.4 Antenna arrangement for SDMASDMA is a promising candidate to improve channel capacity in future wireless communication

systems. Since the discrimination performance of the user in the spatial domain depends on thearray arrangement, an appropriate element arrangement for SDMA should be defined. We proposean equation to determine an appropriate element spacing of a uniformly-spaced linear array (ULA)for SDMA. The situation is illustrated in Fig. 14. The equation is based on the spatial coefficientand is expressed below.

This is called ‘average squared spatial coefficent’ (ASSC) hereafter. The following define thesituation that satisfies the above equation.

The number of users is two.The incoming waves are plane waves.Perfect transmitting power control is assumed.(This means the powers of the incoming waves are identical.)The direction of arrival of the users is uniformly distributed.

Equation (1) represents the dependency of an ASSC on the element spacing of the linear array.Figure 15 shows the dependency of the ASSC on the element spacing when the number of elementsis two or four. The dots in Fig. 15 represent the simulation results. The figure shows that thesimulation results of the ASSC agree with those calculated using Eq. (I). The minimum valuesappear when the element spacing is 0.4 wavelengths and at intervals of 0.5 wavelengths. Figure 16

85

shows the relationship between Eq. (1) and the output SINR employing SDMA. The figure showsthat the element spacing giving the minimum ASSC obtains the maximum output SINR. Therefore,we confirm that the element spacing of the linear array suited for a base station adopting SDMA canbe designed as the spacing giving the minimum value of Eq. (1).

4. SummaryThis paper described the design parameters of smart antenna hardware for microcell

communication systems. As techniques for simplifying hardware and reducing cost to introducethe smart antennas to commercial wireless systems, an efficient calibration method, as well asantenna and base station configurations for microcellular communication systems were proposed.The measured performance of the techniques using the smart antenna testbed confirmed the

86

effectiveness of the proposed techniques. Furthermore, an equation determining an antennaarrangement of a uniformly-spaced linear array for a base station adopting SDMA was derived andthe effectiveness was evaluated by computer simulation.

AcknowledgmentThe authors thank Dr. Hideki Mizuno of Nippon Telegraph and Telephone Corporation (NTT)

for his constant encouragement.

References[I] A. Iera, A. Molinaro, S. Marano, “Wireless broadband applications: The teleservice model andadaptive QoS provisioning”, IEEE Commun. Mag., Oct. 1999 pp.71-75.[2] G. V. Tsoulos, M. A. Beach, J. McGeehan, “Wireless personal communications for the 21stcentury: European technological advances in adaptive antennas”, IEEE Commun. Mag., Sep. 1997,pp. 102-109.[3] J. H. Winters, “Optimum combining in digital mobile radio with cochannel interference”, IEEETrans. Veh. Technol., vol. 33, No.3, pp. 144-155..[4] A. F. Naguib, A. Paulraj, T. Kailath, “Capacity improvement with base–station antenna arraysin cellular CDMA”,IEEE Trans. Veh. Technol., vol. 43, No.3, pp. 691-698.[5] T. Ohgane,”Spectral efficiency improvement by base station antenna pattern control for landmobile cellular systems”,IEICE Trans. Commun., Vol. E77-B, No. 5, pp. 598-605.[6] G. V. Tsoulos, M. A. Beach, S. C. Swales, “DS-CDMA capacity enhancement with adaptiveantennas”. Electron. Lett, Vol. 31. No. 16, pp. 1319-1320,1995.[7] Y. Ogawa, Y. Nagashima, K. Itoh, “An adaptive antenna system for high-speed digital mobilecommunications”,IEICE Trans. Commun., Vol. E75-B,No.5, pp.413-421.[8] G. V. Tsoulos, J. McGeehan, M. A. Beach,“Space division multiple access(SDMA) field trials.Part 1 : racking and BER performance”,IEE Proc. Radar, Sonar Navig., Vol. 145, No. 1, pp.73-78.[9] S. Jeng, G. T. Okamoto, G. Xu, H. Lin, W. J. Vogel,“Experimental evaluation of smart antennasystem performance for wireless comunications”,IEEE Trans. Antennas Propagat., vol. AP-46, No.6,pp.749-757.[10] J. Litva, T. Lo, A Digital Beam Forming in Wireless Communication, Aretech House Publishers,1996.[11] S. Haykin, Adaptive Filter Theory Second Edition, Chapter 5, Prentice Hall, 1991.[12] T. Tanaka, R. Miura, I. Chiba, Y. Karasawa,“An ASIC implementation scheme to realize abeam space CMA adaptive array antenna”,IEICE Trans. Commun., Vol. E78-B, No. 11, pp. 1467-1474.[13] R. H. Roy, “Spatial division multiple access technology and its application to wirelesscommunication systems”, IEEE 47th Veh. Technol. Conf., vol. 2, pp. 730-734,1997.[14] K. Nishimori, K. Cho, Y. Takatori, T. Hori, “A new calibration method of adaptive array for

87

TDD systems”, 1999 IEEE AP-S Dig., Orlando, FL, July 1999, pp.1444-1447, 1999.

Downlink Capacity Enhancement in GSM System UsingMultiple Beam Smart Antenna and SWR Implementation

Wei Wang, Mohamed Ahmed, Samy Mahmoud and Roshdy H.M. HafezDept. of Systems and Computer Engineering

Carleton University1125 Colonel By Dr.,

Ottawa, ON K1S 5B6, CanadaPh: (613) 520-5653 Fax: (613) 520-5758

{mhahmed,wwei, mahmoud, hafez@sce.carleton.ca}

Abstract-Third-generation (3G) wireless systems need strategies to further improve performance,increase data rates and at the same time provide flexible and affordable support for multi-services andmulti-standards. Software radio technology is promising to provide the required flexibility in radio fre-quency (RF), intermediate frequency (IF) and baseband signal processing stages. Smart antenna cangreatly improve system performance, enhance system capacity by making use of spatial processing,exploiting the spatial directivity and reducing co-channel interference. This paper addresses the down-link capacity gain of the multiple beam smart antennas in GSM link Frequency Hopping (FH)-TDMAsystem. The system capacity is studied. Analytical results are compared with the sectorization-only appli-cation. Perfect power control and discontinuous transmission (using voice activity) are taken into consid-eration in the analysis. One possible software radio architecture for a base station with smart antenna isproposed. In this architecture, smart antenna algorithms might be dynamically reconfigured according todifferent environment requirements and the baseband processing might also be dynamically reconfiguredaccording to different standard requirements. In this way, the need for flexibility is satisfied.

Index Terms- Software radio (SWR), capacity analysis, frequency Hopping, cellular system, and smart antennas

90

I. IntroductionRecent research interests in the field of personal mobile radio communications have been moving to the next

generation cellular system to meet the following two main growing demands: (i) higher capacity, quality and vari-able transmission speed of multimedia information, (ii) higher flexibility support for multi-services and multi-stan-dards. Smart antenna and software radio technology are the key solutions to satisfy these demands. Smart antennacan greatly enhance system capacity by reducing co-channel interference, exploiting the spatial directivity of theantenna and making use of spatial processing. Software radio may provide the required flexibility in radio fre-quency (RF), intermediate frequency (IF) and baseband signal processing stages of the system.

A great deal of attention has been given to the performance analysis of smart antennas and software radio respec-tively.([3]-[6], [12]-[15]). Few of the previous works discuss about the architecture synergies between the two tech-niques. Moreover, most of the studies on smart antennas reported in the literature is mainly based on computersimulation rather than analytical methods. which may need simplifications and/or approximations. However ana-lytical methods are also important since they need much less computational time and processing loading comparedto the simulation techniques, in addition to their importance for the validation and verification of the simulationmethods. This make it interesting to study the performance of smart antenna in an analytical way and in the sametime study the synergies of the combination of the two technique.

Slow Frequency Hopping (SFH) is widely used in TDMA (especially in GSM) wireless networks because of itsadvantages namely the interference averaging and the frequency diversity. Since SFH-TDMA systems have softcapacity, a potential capacity enhancement can be achieved using an interference reductions techniques such asPower Control (PC), DTX (Discontinuous Transmission), and smart antennas.In this paper we present the down-link capacity enhancement of multiple beam smart antenna in GSM-like SFH-TDMA cellular system.

The paper is organized as follows: Section II presents the stochastic analysis of the estimation of the outage prob-ability and the system capacity based on the system model. section III. In section IV, the analytical results using anumerical example are presented and compared with simulation results in our previous work [l]-[2]. One possiblesoftware radio architecture for the base station with smart antenna is proposed in section IV. Finally the conclu-sions and future work are given in Section V.

II. Multiple Beam Smart Antenna Performance AnalysisIn this section, analysis is based on the Cumulative Density Function (CDF) of the Carrier-to-Interference ratio

(CIR) in the down link at different loading factor. The maximum loading factor that keeps the outage probabilityunder a certain threshold value (e.g. 2%) is determined and used to estimate the maximum system capacity.i) System and Propagation Model:The propagation model used is given in (1):

91

where is the transmitted power, d is the distance between the BS and MS, is a normal random variable corre-

sponding to the shadowing, F is the fading parameter, and is the directive gain of the smart antenna. The effectof the fading is assumed to be neglected due to slow frequency hopping and other techniques such as equalizationand interleaving etc. are employed.

A very tight frequency plan (1/3) is used for FH carriers. This tight frequency plan can be used without degrad-

ing the QoS provided that SFH and partial loading are employed. non-FH (BCCH) carriers are allocated using (4/12). The carrier hopping is done using the GSM frequency hopping algorithm for the non BCCH channels only

Multiple beams smart antennas[9]-[11] with M-beam are used to cover the whole cell. The MS is connected tothe BS through beam which provides it with the best quality in terms of the received power or the Carrier to Inter-ference Ratio (CIR). Switching from one beam to another is needed when the MS crosses the boundary betweentwo beams. Since it does not need neither complex hardware nor sophisticated weight computation algorithms as inother techniques it is considered as the simplest one. We adopt multiple beams smart antennas because of its sim-plicity and effectiveness with frequency hopping techniques.ii) Theoretical Analysis

In order to estimate the system the capacity, the maximum loading (utilization) factor (Erlang/channel) is firstdetermined. Fig. 1 illustrates the downlink interference to a MS in sector 1 in cell 0 from the 11 cochannel interfer-ing cells. The interference from the six first tier BSs (in cells 1-6) and five second tier BSs (in cells (7-11) is takeninto consideration. BSs facing sector 1 (with solid line cell plot) are introducing higher interference than other BSs(with dashed line cell plot) since the former introduce interference from the main lobe while the latter introduceinterference from the side lobe.1) Multiple beams antenna with no power control:

The interference power at the MS in cell 0 from the jth cell is determined from (1) and using Fig. 2 thatdescribes the problem geometry thus is expressed as

wherej is the cochannel interferes index (j=1,2,....

Aj is a Bernoulli random variable representing the activity of the jth interference mobile station. Aj=l withprobability q that is equal to the product of the loading factor (LF) with the discontinuous transmission fac-tor (DTXF); 1j is the distance between the desired MS and the jth interferer BS;

92

is the shadowing parameter of the jth interferer;

is the antenna gain as a function of the interferer angle and beam angle ,defined as

Because of the interference averaging property of the frequency hopping, we are interested in the mean value of

interference power rather than the instantaneous value. The mean value of the interference power from the jth cellis expressed as:

Where is the PMF of the beam angle. It equals to if users are uniformly distributed,

are the possible values of . E[] is the mean value of the random variable. Thus it is straightforward to show that

is equal to where Rewriting

93

As shown in Figure 2, and are functions of and therefore the total interference is:

where is the distance between the MS and the desired BS and is the angle of the MS at the beam with cen-

tral angle The received carrier power at the MS in cell 0 from its BS is:

The outage probability (Pout) is defined as:

where C is the carrier power. By substituting for C from (6) the outage probability is expressed as:

Using the total probability law the outage probability is given by:

where:

Since is normally distributed it is straightforward to show that:

where:

94

Therefore the outage probability is given by:

The outage probability is determined from (10) by solving the double integration numerically at different loadingvalues to find the maximum loading factor and then the system capacity as presented in the next section.

2) Multiple beams antenna with perfect power control

With perfect power control, each mobile in the jth interfering cell receives a constant power (S) from its BS irre-spective of its location. To keep the received power at each MS constant, the transmitted power at BS variesaccording to each MS location within its cell accordingly the transmitted power is function of and (see figure

2) so that it is expressed as:

where is the shadowing factor of the downlink path;

In the similar analysis way, the final outage probability is given by [2]:

By computing the above integrals and summation numerically, we can solve for the maximum loading factor thatsatisfies the outage probability condition. By determining the maximum loading factor (LFmax) the system capac-ity defined as the number of users per unit area can be evaluated by

where is number of channels per cell, is the average traffic in Erlangs per user, and is cell area.

III. Analytical ResultsIn this section a numerical example is used to apply the analysis presented above. Results are compared with the

simulation result of previous work [2]. The following parameters are used in the analysis and simulation: Cell

95

radius (R) = 2 km, Number of sectors per cell (s) = 3, Number of beams per sector Discontinues Trans-

mission Factor (DTXF) = 0.5, Traffic per user in Erlang

Table (1) lists the outage probability for the different cases mentioned above and at loading values 10-60%. It isapparent that the outage probability is highly reduced by using multiple beam antennas especially when it is com-bined with power control, e.g. with MB Pout is reduced from 1.2x10-1 to 6.7x10-2 without PC and to 2.0x10-2with PC. Table (2) lists the maximum loading factor at different maximum outage probability requirements(pmax),e.g. when pmax=2% the maximum loading factor (with no PC) is increased from 7% to 20% (which isalmost three folds) by using MB without PC and by employing PC without MB it can be increased the maximumloading factor to 18% which is more than 2.5 times the original loading factor. By combining both techniques themaximum loading factor is jumped to 60% which is more than 8 folds gain. This high gain is achieved due to themultiplicative effect of both technique MB and PC. However the capacity gain might be less than in reality becauseof the difficulty to have such perfect power control. The gain due to PC is inversely proportional to the loading fac-tor and it is shown that it can degrade the performance especially without MB, e.g. the maximum loading isreduced.from 40% to 26% by using PC since the assumed perfect power control degrades the CIR

at high loading values.

The outage probability is evaluated by simulation and analytically using (11) and (12) at different loading values.The CDF(CIR) with Multiple Beams (MB) or sectorization only (SC) with and without Power Control (PC) at

96

20%, 40% and 60% loading factor are given in [2].

Fig.3 show that significant enhancement in the performance can be achieved by employing multiple beam anten-nas and power control. it can be shown that the outage probability can be reduced by 40-65% without PC while thereduction is more than 95% with PC particularly at low and medium loading factor.

97

VI. Software radio architecture with smart antennasOne base station architecture with smart antenna is proposed here and served as a good example to illustrate the

great flexibility of software radio.

Here, we assume each channel can be shared by K users. K beams are formed, one for each user. Thus, for a sys-tem with L physical channels, KL users can be supported. Figure 4 shows the function block diagram of the soft-ware radio base station with smart antennas. Each antenna element has its own down converter and ADC. K similarblocks are needed for the subsequent beamforming and demodulation, which implemented in software.

The beam formers for each channel is illustrated in Figure 5. The digital downconverter includes the function offrequency translation and filtering.The received IF signals from each antenna element are translated to a complexbaseband signal by the quadrature multiplier. The digitally controlled oscillator (DCO) block generates the quadra-ture signals for the multipliers. Then the complex baseband signal is low-passed filtered. Finally these signals areadjusted by the weights generated by the smart antenna algorithms in beam former, and combined in the “com-biner” block. The output of the demodulator would be the received signal from the desired user.

The smart antenna algorithm chosen to adjust the weights is depended on the environment. For GSM system, inmacrocell with low traffic, more noise than cochannel interference (CCI), high user mobility and low angular

98

spread. The preferred Algorithm, such as Direction of arrival based beam former (DOB), whose performancedepends on the ration of angular spread with the number of antenna elements M. For the microcell with high traffic,

high CCI, low user mobility, high angular spread and low delay spread, time reference based beamforming is suit-

able. The software implementation of the beam former provides the reconfiguration flexibility for the algorithmschosen according to the environment requirement.

The a baseband processing functions are implemented on a common platform using a parameterized implementa-

tion method, which enables the air interface to be changed fast by only exchanging a set of parameters for thereconfiguration, rather than download the whole software of a system. Figure 5 illustrate such a common basebandprocessing platform for GSM and IS-54/136 TDMA systems. In both of the system, lot of functions like channelcoding, modulation and equalization, are used in a similar way and can be approximately implemented using thesame blocks with different parameters. This makes it possible for a parameterized software implementation. Shareone common transceiver architecture.

The proposed architecture provides the flexibility for smart antenna algorithms to be reconfigured according todifferent environment requirements and for baseband processing functions to be reconfigured according to differ-

ent standards requirements.

V. Conclusions

Smart antennas are investigated and shown to be a powerful tool for increasing wireless network capacity of FH-TDMA cellular systems. A multiple beam antenna combined with power control can dramatically enhance the sys-tem capacity particularly at low and medium loading values. This may serve as an upper bound which can beapproached in reality as technology for perfect power control is improving.

One software radio base station architecture with smart antenna is proposed and serves as a good example toillustrate the great flexibility provided by the software radio solution. In this architecture, the smart antenna algo-rithms can be dynamically reconfigured according to different environment requirements and the baseband pro-

cessing can also be dynamically reconfigured according to different standards requirements.As an extension of this work, SDMA where each channel can be reused in every beam will be addressed for FH-

TDMA. In addition, many detailed design issues related to software radio architecture are still open. These issuesinclude: the evolution of software radio architecture, the feasibility and efficiency evaluation of the integrationwith smart antenna algorithms and the method to better evaluate the processing capacity of the whole architecture.

99

Referencesl-Mohamed Ahmed and Samy Mahmoud, "Capacity Analysis of GSM Systems using Frequency Hopping and Smart Anten-

nas," To be Published in IEEE VTC2000, Tokyo,Japan, May2000.2.-Mohamed Ahmed, Wei Wang, and Samy Mahmoud, "Downlink Capacity Enhancement in GSM Systems with Frequency

Hopping and Multiple Beam Smart Antennas," To be Published in ICC2000.3-T. Chebaro, “Capacity evaluation of a frequency hopped TDMA cellular radio system,” Annaels des Telecommunications,

vol. 51,Mar.-Apr. 1996

4-J. Johansen and B. Vejlgaard, “Capacity analysis of a frequency hopping GSM system,” Master thesis, Alborg University1995.5-G Tunnicliffe, A. Sathyendran and A. Murch, “Performance improvement in GSM networks due to slow frequency hop-ping,” IEEE VTC 1997.6-C.-C. Lee and R. Steele, “Signal-to-interference calculations for modern TDMA cellular communication systems,” IEE Pro-ceedings-Communications, Vol. 142, No. 1 , Feb. 1995.7-S. Channakeshu, A. Hassan, J. Anderson and B. Gudmundson. “Capacity analysis of a TDMA-based slow-frequency-hopped cellular system,” IEEE Trans. on Vehicular Technology, Vol. 45, No. 3, Aug. 1996.8-J. Dornstetter and D. Verhulst, “Cellular efficiency with slow frequency hopping: Analysis of the digital SFH900 mobile

system,” IEEE JSAC, Vol. 5, No. 5, June 1987.

9-L. Godara, “Applications of antenna arrays to mobile communications, Part I: Performance improvement, feasibility, andsystem considerations,” Proc. of the IEEE, Vol. 85, No. 7, July 1997.10-J. Winters, “Smart antennas for wireless systems,” IEEE Personal Communications, Feb. 1998.11-M. Wells, “Increasing the capacity of GSM cellular radio using adaptive antennas,” IEEE Proceedings-Communications,Vol. 143, No. 5, Oct. 199612-J. Mitola III, "Software Radio Architecture and Technology," Proc. 1998 Int'l. Symp. Adv. Radio Tech., Boulder, CO, Sept.

1998.

13-F. Riera-Palou, C. Chaikalis, J.M. Noras, "Reconfigurable mobile terminal requirements for third generation applications", UMTS Terminals and Software Radio (Ref. No. 1999/055), IEE Colloquium, pp: 9/1 -9/6, 1999.

14-A.Wiesler, H. Schober, R. Machauer, F. Jondral, "Software radio structure for umts and second generation mobile com-

munication systems ", Vehicular Technology Conference, 1999. VTC 1999 - Fall. IEEE VTS 50th, Vol. 2, pp: 939 -942, 1999.

15-J. Kennedy, M.C. Sullivan, "Direction finding and smart antennas using software radio architectures" IEEE Communica-tions Magazine, Vol. 33 5, pp: 62 -68. May 1995.16-R. Kohno, " Software antenna and its communication theory for mobile radio communications ", Personal Wireless Com-munications, 1997 IEEE International Conference, pp: 227 -233, 1997.17-T. Turletti, D. Tennenhouse, "Complexity of a software GSM base station ", IEEE Communications Magazine, Vol. 37 2,

pp: 113-117, Feb. 1999.

Generalized Equations for Spatial Correlation for Low toModerate Angle Spread

R. Michael BuehrerBell Laboratories - Lucent Technologies

67 Whippany Rd. Room 3A-220Whippany, NJ 07981

mbuehrer@bell-labs.com

Abstract In this paper we derive generalized formulas for three types of angular energy dis-tributions: a Gaussian angle distribution, the angular energy distribution arising from a Gaussianspatial distribution, and a uniform angular distribution. These generalized equations are param-eterized by where d is the distance between antennas and is the standard deviation of theangular energy distribution and approximate the true correlation with the approximation beingvery good for angle spreads below approximately 25°.

1 IntroductionPrevious work on antenna correlation has relied on numerical integration or infinite series to evaluatethe spatial correlation between two points using the angular energy distribution [1, 2]. As a result,separate curves must be generated for each distribution parameter of interest (e.g., each variance ofa Gaussian distribution). However, it was recently suggested by Chizhik and Gans [3] that when theenergy arriving at a linear array has a distribution in angle which is Gaussian, the spatial correlationfunction can be parameterized by where d is the distance, is the carrier wavelength, andis the standard deviation of the angular distribution. In other words, we can create a generalizedspatial correlation curve which would be useful for practical values of In this paper we showthat indeed for low values of we can derive a generalized equation which approximatesthe spatial correlation for any central angle-of-arrival for a Gaussian angular energy distribution,a uniform angular energy distribution, and a Gaussian spatial distribution.

2 Spatial Correlation

Consider a plane wave signal arriving at an array from angle with respect to the normal bisectingtwo points of interest separated by d meters. The signals seen at the two points can be representedas s1(t) = m(t) and If the power of the message signal m(t) is unity,then Thus, if a signal of interest arriving at an array can be describedby the summation of plane waves arriving from angles with distribution then we know thatthe spatial correlation between two points a distance d apart can be determined as [4]

where is defined relative to the normal.First, let us assume a Gaussian distribution for angular energy which is common for spatial

channel modeling. Thus, the angular distribution function can be represented as

102

where is the standard deviation of the distribution in radians assumed to be small enough thatthere is not significant energy beyond and is the central angle of arrival in radians. It isshown in Appendix A that in the case of defined by (2) the correlation can be approximatedby

where we now see that we can parameterize the spatial correlation byIn Figure 1 (a) we plot the spatial correlation versus using the integral in (1) as well (3) for

and several values of We can see that for up to 20°, (3) is a very good fit. Figure 1 (b)presents the same results for As increases the fit is not as good for moderate values ofFigure 5 (a) plots (3) for four values of Using this single plot, we can determine the correlationfor almost any scenario, provided that the angular spread is Gaussian with a standard deviationless than about 20° - 25°.

3 Antenna SeparationUsing the above equation (3), we can easily show that the required distance to ensure a correlation

can be approximated by

The above equation is plotted for for four values of in Figure (2). From this plot we seethat the classic rule-of-thumb of spacing holds for It can be shown that for

where sigma is in degrees provides a good approximation for required antenna separationto achieve correlation values below 0.5.

103

4 Gaussian Spatial DistributionA second model for a spatial channel which is commonly used is a Gaussian spatial distribution.A Gaussian spatial distribution models the scatterers surrounding the mobile using a bi-variateGaussian distribution in space [5]. In other words, the scatterers have position [x, y] with probability

where is the standard deviation in both the x and y directions and is the center ofthe distribution. We wish to find the distribution of the angle-of-arrival, i.e.,the angular energydistribution in order to determine the spatial correlation. To do so we make the substitutions

and into (5). Making this substitution and integratingover r results in

Further if we define and we can show that

Thus for small values of this is nearly identical to a Gaussian distribution and thus we canuse a Gaussian distribution to model the angular energy distribution. Figure 3 (a) plots (6) with

and as well as a Gaussian angular energy distribution with and

104

(0.07 radians). Thus, we can see that we can approximate the angular energy distributionfor a Gaussian spatial distribution with a Gaussian angular energy distribution and thus use (3) toapproximate the correlation. To make the translation we can use the plot in Figure 3 (b) to convertthe spatial parameters to a standard deviation to use in (1). For example, if we assume a a Gaussianspatial distribution of scatterers with and we find thatUsing Figure 3 (b), we find that this corresponds to a standard deviation of approximatelyAlternatively, we can see from (7) that for large and small we can approximate (7)using a Gaussian with This corresponds to the linear region of Figure 3 (b). Thus, we canuse in equation (3) to approximate spatial correlation.

5 Uniform DistributionAnother common assumption for angular energy distribution is a uniform distribution [2]. Auniform distribution of angular energy is defined as

where is the range of angles about a central angle-of-arrival It is shown in Appendix B thatthe spatial correlation in this case can be approximated by

In Figure 4 the spatial correlation is plotted using a numerical integration of (1) along with theapproximation in (9) for and Again, for low to moderate values of we find that(9) is a very good approximation. Again, the approximation is less accurate for larger values of

105

If we substitute (the standard deviation of a uniform distribution) we can compare theGaussian and uniform distributions for several values of as shown in Figure 5 (a). As expected, theGaussian distribution decreases more slowly in the main lobe, but lacks the secondary correlationpeaks. Otherwise the approximate correlation functions are similar. Additionally, it can be shownthat as gets large, the Gaussian function approaches a uniform distribution due to a ambiguityand the correlation function will develop secondary peaks.

As a final note, we compare the results presented here with results given in [6]. In [6] ageneralized correlation function was derived for general angular energy distributions. Specifically,it is shown that

where is a measure of angular spread defined as

and is the nth complex Fourier coefficient

Using these definitions we can show that the correlation function for a Gaussian distribution canbe approximated as

In fact if the approximation then we arrive at

106

which is very nearly our approximation when we realize that Figure 5 plots thisapproximation of the correlation function versus along with (3) and (9). As can be seen,there is very good agreement down to correlation values of 0.5. The approximation of [6] is slightlyoptimistic for a Gaussian distribution and would be more so for values of This makes sensesince is not directly involved in (13).

6 ConclusionsIn this paper we have derived generalized (i.e.,for multiple values of ) correlation functions forthree distributions of angular energy. The generalized equations allow the correlation to be foundfor any practical standard deviation and distance. We have shown that the approximations aregood for standard deviations of about 25° or less.

AcknowledgmentsThe authors would like to thank our colleagues at Lucent - Bell Laboratories including DirckUptegrove, Jay Tsai, Dmitry Chizhik and Mike Gans.

References[1] W.C.Y. Lee. Effects of correlation between two mobile radio base-station antennas. IEEE

Transactions on Communications, COM-21(11):1214–1223, November 1973.

[2] J. Salz and J.H. Winters. Effect of fading correlation on adaptive arrays in digital communi-cations. In Proceedings of the International Conference on Communications, pages 1768–1774,May 1993.

107

[3] D. Chizhik and M.J. Gans. Angular spread, antenna separation and correlation. Technicalreport, Lucent Technologies - Bell Laboratories.

[4] W.C. Jakes. Microwave Mobile Communications. Wiley-Interscience, 1974.

[5] R. Ertel, P. Cardieri, K.W. Sowerby, T.S. Rappaport, and J.H. Reed. Evolution and applica-bility of spatial channel models for wireless communications. IEEE Personal CommunicationsMagazine, 5(1): 10–22, February 1998.

[6] G.D. Durgin and T.S. Rappaport. Effects of multipath angular spread on the spatial cross-correlation of received voltage envelopes. In Proceedings of IEEE Vehicular Technology Confer-ence, pages 996–1000, 1999.

[7] I.S. Gradshteyn and I.M. Ryzhik. Table of Integrals, Series, and Products. Academic Press,second edition, 1980.

Appendix A: Derivation of Generalized Equation for GaussianDistributionFirst, let us assume a Gaussian distribution for angular energy such that the angular distributionfunction can be represented as

where is the standard deviation of the distribution in radians and is the central angle of arrivalin radians.

Then we know that the spatial correlation can be determined as [4]

Now, substituting (2) into (16) and making a change of variables we get

Now, assuming that is small over the range where is significant, we can approximatethe above with

Evaluating the integral [7] then gives

108

Appendix B: Derivation of Generalized Equation for UniformDistribution

In this appendix we wish to show that the spatial correlation function for a uniform angularenergy distribution can be approximated according to (9). First we assume that the angular

energy is distributed according to

Now for small we can approximate and which gives

Evaluating the integral [7] then gives

Q.E.D

Exploitation of Internode MIMO Channel Diversity inSpatially Distributed Multipoint Communication Networks

Brian G. Agee, Ph.D., P.E.1596 Wawona Drive, San Jose, CA 95125

Telephone: 408.269.3218, E-mail: bgagee@pacbell.uet

AbstractThe multiple-input, multiple-output (MIMO) channel exploitation concept, advanced to date for ro-

bust and efficient diversity communication in point-to-point links, it extended here to spatially-distributedmultipoint networks. The approach exploits the inherent MIMO diversity of multipoint networks to allowdata transport at high efficiency and low aggregate network complexity relative to point-to-point links,without the need for opportunistic multipath between nodes in the network. A general mathematicalframework for the networks it presented, and developed in depth for small networks where all nodes arevisible to each other. Network capacity is then analyzed for a ring and star network, and it it shown thatthe approach can provide significant (factors of 3-to-6) improvements in network and node capacity overequal-cost point-to-point links. Results are germane to commercial applications in wireless LANs, picocellnetworks, and wireless Internet appliances, and to military applications in secure internode systems.

1 IntroductionDistributed networks possess compelling attributes for both commercial and military applications. In thecommercial arena, the emergence of the Internet (itself a highly distributed network) as the first newmass medium since television, combined with the ongoing convergence of communication and computerapplications and services, has fueled development of a wide range of products to deliver broadband dataservices over the “last mile” (via wired or wireless media) to businesses and residences. At the sametime, the explosive demand for mobile and portable data services, and the emerging market for wirelessappliances, has fueled the development of low cost (and so far low capacity) wireless devices to bothconnect “conventional” untethered platforms (handsets, laptops, and PDAs) to themselves and the wiredinfrastructure, and to supervise, monitor, and control distributed networks of embedded processors inemerging “smart appliance” products. In all of these applications, distributed networks can provide strongadvantages over conventional systems, by exploiting the inherent advantages of connectionless data service,or by reducing the power required to communicate to laptops and PDAs at data rates competitive withtethered devices.

Distributed networks also provide multiple advantages in military applications, including collection,analysis, and collation/dissemination of reconnaissance data from beyond the front-line of troops (PLOT);intruder detection and location behind the FLOT; and distributed communication of command, control,and voice/video data between the FLOT and rear echelons. By allowing data transfer through nearbynodes and over “flat” network topologies, distributed networks can reduce an adversaries’ ability to identify,target, or even detect high priority nodes in the network, greatly enhancing their security and survivabilityrelative to conventional point-to-multipoint networks.

This paper describes means for further extending the capabilities and applications of distributed net-works, by exploiting the multiple-input, multiple-output (MIMO) channel structure of the network. MIMOchannel exploitation has recently been advanced as a means for enhancing performance of point-to-pointcommunication links, by exploiting opportunistic spatial, spectral, or polarization diversity (significantpropagation paths) present in those links [2, 3]. Advantages of these systems include near-optimal capacity

110

relative to nondiverse links; robust operation in severe multipath environments, and robust, high-capacitybackhaul over non-LOS transmission paths.

However, these approaches have demonstrated limitations in practice, due primarily to the high marginalcost of exploiting lesser modes of the MIMO channel response matrix. In particular, the approaches de-scribed in [2, 3] require an additional antenna at each of the link to exploit each diversity path, implyinga linear growth in cost for each path exploited by the network. However, the capacity added by thatadditional transceiver pair adds “piecewise linearly” as a function of transceiver power, i.e., the diversitypath can only be exploited at power levels high enough to overcome the higher pathless on that path. Inmany practical systems, this pathloss is much higher than the loss on the dominant path(s), e.g., 20-to-30 dB, in fixed outdoor communication systems. Exploitation of this multipath requires both high power(to permit data transport over the weak path) and complex codecs (to permit data transport at highinformation rates on the dominant path). Moreover, the existence and strength of these diversity pathscan vary widely as a function of time and/or node location, greatly complicating the design of networksemploying this technology.

Spatially distributed networks overcome this limitation, by exploiting the inherent diversity betweeninternode channel responses in the network. This diversity exists regardless of multipath present on anyindividual path in the network, i.e., it does not require high levels of opprotunistic multipath to be ex-ploitable-by the system. Moreover, this diversity can be designed into the network by careful choice ofnetwork topologies during the deployment process, in order to provide truly linear growth in capacity astransceivers are added to the network. As a side benefit, the network can also spatially excise transmissionsfrom compromised nodes and external emitters, allowing secure, high quality service in environments withuncontrollable interference, e.g., Part 15 bands.

This paper provides a mathematical framework for exploiting MIMO diversity in distributed networks,and analyses the network and link (node) capacity achievable for a ring and star topology. Section 2reviews capacity attainable in point-to-point MIMO links, and establishes background for the extensionto multinode networks. Section 3 establishes a structurally constrained model for multinode networks,and computes capacity for small networks where all of the nodes are visible to each other. Section 4analyzes capacity of a four-node ring, five-star node, and two-node point-to-point link operating in thesame geographical area. It is shown that the network can double or triple the throughput of the equivalentpoint-to-point link, at equivalent cost to deploy the links.

2 Background: Point to Point MIMO LinksThe MIMO network model is motivated by MIMO model developed for spatially-diverse point-to-pointlinks. An example link is illustrated in Figure 1, showing generation and transmission of spatially-diversesignal from node 1 to node 2 over propagation paths (e.g., a direct path and 2 reflection pathsin Figure 1). Assuming the nodes employ and spatially separated antennas, respectively, and theinverse bandwidth of is much wider than the delay between antennas at each end of the link, theuplink receive signal can be modelled by

where is the interference received at node 2, and are the andtransmit and receive spatial signature vectors for each propagation path, respectively, andand are the pathloss and group delay over each propagation path. Equation (1) is expressedmore compactly (2), by defining group delay and subsuming the remaining channel response into

MIMO channel operator with frequency response

where and are and transmit andreceive signature matrices comprising the spatial signature vectors for each propagation path, respectively,and If is an OFDM waveform with cyclic prefix greater than maxthen the received signal can be expressed as OFDM tones and symbols,

where and are the transmit and receive data symbols, k is the OFDM tone index denotingdata transmitted at frequency and i is the OFDM symbol index denoting data transmitted withintime interval and where and are the OFDM symbol length (FFTduration) and cyclic buffer length, respectively. If the group delays are also small relative to the Inversesignal bandwidth and the interference is complex-Gaussian and temporally white over the signal passband,then where is the center of the signal passband, and can be modelledas an complex random process with covariance and mean This assumption willbe used throughout this paper.

The capacity of the MIMO point-to-point link for-power constrained systems is given by [1, 2, 3]

for this channel, where is the modulation efficiency of the communication link, are the num-ber of information bearing tones in the uplink basspand, and are the nonzero eigenvalues

and is the fraction of power allocated to mode m of and whereand is the average power used at each transceiver. Optimizing

111

112

over feasibility region yields the “waterfilling” solution,

where satisfies In the absence of multipath, reduces toa rank-1 matrix with nonzero mode and reduces to

where (10) results if is spatially white background noise and isotropic elements areused at both ends of the link and

The modulation efficiency accounts for real-world inefficiencies in the airlink modulation format,including filter shape factors, window inefficiencies, data retransmission events, interlink guard-times (inTDD systems), cyclic prefixes, and security measures in military communication systems. A more useful(I.e., achievable) capacity bound can also be developed by deflating by an SNR gap term in (5)to account for inefficiency in the codec. A modulation efficiency of will be assumed here,where is the duty cycle of the communications uplink.

The capacity can be approached in principle using the linear transceiver structure shown in Figure 2,shown here for a duplex communication link. On the transmit path (node 1), the user encodes the transmitdata into a stream of complex symbol vectors and multiplies the symbol vectors by an

diversity transmit gain matrix to form output tone sequenceThis signal is then passed to a vector airlink modulator and transmitted to the other end of the link.On the receive path (node 2), the dimensional receive signal is demodulated and passed through an

diversity receive combiner matrix to form estimated symbol vectorThe symbol vector is then multiplied by a phase ramp to remove tuning error, and passed onto a vectordecoder to recover the transmitted bit stream. The optimum transmit and receive matrices used on theuplink are given by

where is the inverse Cholesky factor of and where and are the right-side andleft-side eigenmodes of normalized MIMO channel transfer matrix

and where The data received from each combiner port is then given by

i.e., the combined processor and channel decomposes into independent channels, each of which trans-ports a symbol stream with signal-to-interference-and-noise ratio (SINR) The encoderdistributes bits to each channel in accordance with SINR achieved on that channel.

113

As shown in [2, 3], the capacity of this link can be dramatically increased as additional transceivers areadded to the system, if the propagation modes exploited by these transceivers are strong relative to thedominant mode. Conversely, this increase is small if the new modes are much weaker than the dominantmode. Moreover, these modes can be unpredictable or time-varying, raising additional planning issues inlarge network deployments. As a consequence, the cost required to achieve this gain (driven largely by thenumber of RF transceiver chains in the network) is only justified where this high SNR can be achieved, e.g.,indoor or dual polarized communication systems. These issues motivate extension of the MIMO conceptto networks of nodes.

3 Extension to Multipoint NetworksThe mathematical structure provided in Section 2 can be extended to distributed networks, wheredownlink nodes are attempting to communicate with uplink nodes. Example networks are shown inFigure 3, depicting a ring network (Figure 3a) where four uplink and downlink nodes are connected ina closed ring topology such that and a star network (Figure 3b) where a central node iscommunicating with four outlying nodes such that In this analysis, it is assumed thateach node possesses a spatially diverse antenna array, allowing it to simultaneously communicate with everynode in its field of view. However, it is also assumed that nodes in each set cannot communicate directlywith themselves, i.e., independent information is transmitted from each node in set 1, and is independentlyprocessed by each node in set 2.

3.1 Environment ModelUsing notation developed in Section 2, node n2 receive signal can be modelled by

where and are the group delays and MIMO transfer functionsbetween transmit and receive nodes and respectively, is the data set trans-

114

mitted from downlink node and is the interference vector received at nodeand where and are the number of transceivers (antenna elements) employed at nodesand respectively. Equation (16) can be expressed in compact matrix notation by “stacking”and into and dimensional vectors andrespectively, yielding network channel model

where is the MIMO transfer function for the uplink network.Equation (16) has the same structure as the point-to-point link model given in (4). However,

has nonunity rank even in the absence of multipath. In this case, reduces to

where is the spatial signature vector from node to node and isthe spatial signature vector from node to node and where is the pathlessbetween nodes and Equation (18) has rank limited by min

The processor structure given in Figure 2 and (11)-(14) cannot be realized at the network level, sincenodes within a set cannot share information with each other. However, (7)-(8) can provide a usefulupper bound on the total capacity achievable by the network. In addition, Figure 2 motivates structurallyconstrained networks that approach this upper bound, by-treating the transmit and receive weights inFigure 2 as part of the channel. This network model is developed below.

3.2 Structurally Constrained Network ModelConsider a network employing the processor structure given in Figure 2 at each node in the network.Assuming that each transmit node is (in general) communicating with (each) receive node using asingle symbol stream the node transmit signals are modelled by

where is the transmit gain matrix used at node and the decoder inputsymbols are modelled by

115

where is the MIMO processor transfer function for the network andare the total number of links in the network. Both and have the same

structural form as the point-to-point MIMO link described in Figure 2. except that and arestructurally constrained to prevent communication between nodes in the same set. If the interference istemporally white over the signal passband, then can be expressed as

where is the inverse Cholesky transform of and is the average powertransmitted from each transceiver, such that

Equations (25)-(27) can be expressed more compactly by defining and in terms of linkindex q using network address table

where is the number of links deployed by the network and and are the transmitand receive nodes used by link q. Using (28)-(31), the symbols recovered on link are

where is the system processing gain from link into linkFurther defining and constraining the signal-to-interference-and-noise

ratio (SINK) on link can be expressed as

where is the cross-link SNR of link signal present after reception on linkSolving (33) for yields where

116

Equation (32) or (33)-(35) can be used to compute network capacity using information-theoretic or adhoc optimization criteria. A particularly useful ad hoc criterion suggested by this model is

This capacity measure provides a useful lower bound on performance of the structurally constrained net-work. More powerful criteria can also be derived from the channel model given in (32).

A detailed analysis of capacity for general networks is beyond the scope of this paper; however, crite-rion (36)-(37) admits a particularly simple form in small networks where each node may possess enoughtransceivers to spatially resolve every node on the other side of the link.

3.3 Null Steering Solution for Small NetworksConsider the multipath-free environment where is given by (18), and assume thatand Then the off-diagonal terms in can be removed by setting andequal to null-steering solutions

where and are diagonal scaling matrices. Substi-tuting (38)-(39) into (34) yields

where is the Euclidean basis vector, and The extension to environmentscontaining substantive multipath is straightforward.

Substituting (40)-(42) into (35) yields and (36)-(37) reduces to

where is the total number of links that can be established by the network under thenull-steering constraint. This criterion is optimized by the waterfilling solution given in (7)-(8),

where is deflated by an SNR gap term to account for codec inefficiency.

117

Note that this solution requires at least transceivers at each node in downlink set 1, and at leasttransceivers at each node in uplink set 2, such that and and transceivers

are to implement the null-steered network. This number can become unacceptable for large networks.However, the total capacity of this network can grow linearly with in many practical networks, i.e.,the cost added by each transceiver can be offset by an equal improvement on capacity. Moreover, thisresult can hold in environments with little or no multipath, as long as the spatial separation between nodesis acceptably large.

The off-diagonal terms in can also be forced to zero if every downlink node can steer independentnulls to uplink nodes in its field of view and vice verse. That is, if andwhere is the number of uplink nodes visible to node and is the number of downlinknodes visible to node then can be made diagonal and can be expressed as (46). This canprovide a simple extension of this result to wide area mesh and ring networks where each node may bevisible to a small number of other nodes.

3.4 Extension to Reciprocal NetworksThis analysis extends easily to reciprocal networks where the uplink and downlink transmission channelspossess reciprocal symmetry, such that This condition can be obtained fora variety of communication scenarios, including time-division duplex (TDD) networks, simplex networks,and random access packet data networks where transmit and receive operations are all carried out on thesame frequency channel.

In particular, if the received interference is spatially white in both link directions, such thatwhere is the background noise at all receivers, and if the link duty cycles and number of

transceivers are equal in both link directions and then the structurally constrainedad hoc network capacity measure defined in (36)-(37) can be made equal in both link directions, by setting

and at both ends of the link, where are the linear transmitand receive weights used in the downlink.

This result is shown by defining structurally constrained downlink capacity

where is the inverse Cholesky product of and where is the averagedownlink power/transceiver and are the number of active tones transmitted over the downlink. If theinterference is spatially white and the link responses are reciprocal, then and

118

and and are related by

with equality holding in (54) if If and thenand Lastly, if and then and i.e., the adhoc uplink and downlink capacity measures should be identical.

If the received interference is not spatially white, then a reciprocal capacity measure can still be obtainedby constraining and to satisfy

where and are the average interference powersover the network. Defining interference-normalized transmit gains

then cross-link SINRs and can be expressed as

in (35) and (49), respectively, where and and where andare the doubly whitened MIMO link responses, given by

If then whenever the channel responses are reciprocal, and thead hoc uplink and downlink capacity measures are equal if and

4 Network Capacity AnalysesThe capacity achievable by a this network is illustrated in Figure 4, for the four-node ring and star networksshown in Figure 3, and for a baseline point-to-point link between the EW nodes in the network. Pathlossbetween the EW nodes is set to 134 dB, corresponding to two-ray propagation between a pair of nodesseparated by 3 miles and operating at a 5.78 GHz carrier frequency (upper U-NII band) and 50 foot heightabove average terrain (HAAT), with additional shadowing and path variation due to ground scatteringand with low (4 dB median) shadowing. Pathloss between the EW and NS nodes in the ring network,and between all four of outlying nodes and the central node in the star network, ranges between 115 dB

119

118 dB, corresponding to propagation over a 1.4-to-2 mile range, within the first Fresnel zone of each nodeat the 60 foot HAAT’s assumed here.

Each node employs a 12-element circular transmit/receive array with a 6 inch diameter (3 inch separa-tion between antennas) and omnidirectional azimuthal response and 1.76 dBi (Hertzian dipole) elevationgain on each antenna element, connected to a Butler matrix that extracts (or excites) the dominant spa-tial modes of the array after the reception (or before the transmission) operations, and a bank of digitaltransceivers that process the dominant Butler mode(s) using the procedure described In Section 2. Thedigital transceiver is assumed to be TDD with a 40% duty cycle in each direction, an active bandwidthof 1.6 MHz, a modulation efficiency of 32% (including the TDD duty cycle), and an SNR coding gap of2.3 dB.

Capacity is calculated as a function of transceivers transmit powers with the added constraintthat each network employ the same number of transceivers, distributed to satisfy Thisprovides a “cost neutral” measure of capacity for each network, under the assumption that deployment costis driven by the number of RF transceivers employed in the network1. In particular, each network employs24 transceivers in Figure 4, distributed equally at all nodes in the point-to-point link (12 transceivers/node)and ring network (6 transceivers/node), and concentrated at the central node in the star network (12transceivers in the central node, 3 transceivers in the outer nodes). Capacity is presented both in aggregateover the entire network, and separately at each outlying node. In the latter case, this is calculated bothfor unloaded networks where the node has access to the entire network, and for fully loaded and balancednetworks where the node must share that bandwidth simultaneously with every other node in the network.Data concentration is ignored, i.e., the node either has free access to the network (unloaded case), ortransmits, receives, and relays data to allow intemode communication at identical rates (loaded case).

As Figure 4 shows, the ring and star networks provide a dramatic increase in capacity over the point-1 This assumption ignores cost of site acquisition, providing a favorable bias towards network with larger numbers of nodes,

e.g., the star network In Figure 4, if these costs are high.

120

to-point link. For example, the ring and star networks possess 45 Mbps and 28 Mbps in aggregate capacityif 30 dBm PA’s are employed in the network, while the point-to-point link only provides 7.8 Mbps ofcapacity under similar conditions. That is, the star and ring networks provide 3-to-6 times the capacityof the point-to-point link, using the same number of transceivers, and providing potential communicationwith an additional 2-to-3 users!

Examination of the individual node capacities reveal further advantages for the unloaded ring network.In the unloaded case, nodes in the star network and point-to-point links can communicate with other nodesat roughly equal information rates. However, nodes in the ring network can communicate with other nodesat almost three times this rate! This is due in part to the ring network’s ability to transport informationin multiple directions, effectively doubling its throughput rate when bandwidth is available and needed.

An additional advantage can be seen in loaded networks. If the capacity of each link is identical,then every node in the four-node ring can simultaneously communicate with its three counterparts at acontinuous data rate of where is the rate achievable on a single link2. At the 30 dBm transceiverpower limit given above, for example, each node in the ring network can jointly communicate at 7.5 Mbpsunder loaded conditions, as opposed to 22.4 Mbps in the unloaded environment. This rate is very close tothe 7.8 Mbps rate achievable by the point-to-point link.

In contrast, the rate achievable by nodes in the star network drops by a factor of three (in general a factorof under loaded conditions, and only then if the base station is treated as a relay node that doesnot contribute information to the network. At a 30 dBm transceiver power limit, for example, the nodes inthe star network can only jointly communicate at 2.4 Mbps. This result has serious implications for certainapplications, e.g., video teleconferencing, interactive games, and network backhaul, where continuous full-rate communication between nodes may be desirable or even necessary to the intended purpose of thenetwork.

5 ConclusionsThe MIMO channel exploitation concept has been extended to spatially-distributed multipoint networks,and illustrated via capacity analysis for a ring and star network. These analyses demonstrate that MIMOnetworks can provide significant (factors of 3-to-6) improvements in network and node capacity over equal-cost point-to-point links, both at the user end (node capacity) and over the entire network. These resultsestablish both a motivation and a framework for exploring applications of MIMO networks over widerapplications. It is expected that these networks will provide strong advantages in commercial applicationssuch as wireless LANs, picocell networks, and wireless Internet appliances where data is naturally sharedamong many users, and in military applications where the ability to operate over flat and flexible networktopologies, without sacrificing network throughput, detectability, or security, or survivability is of highimportance.

References[1] T. Cover, T. Joy, Elements of Information Theory, Wiley: 1991

[2] G. Raleigh, J. Cioffi, “Spatio-Temporal Coding for Wireless Communication,” IEEE Trans. Comm.,March 1998, Vol. 46, No. 3, pp. 357-366

[3] G. Foschini, M. Gans, “On Limits of Wireless Communications in a Fading Environment When UsingMultiple Antennas,” Wireless Personal Comm., March 1998, Vol. 6, No. 3, pp. 311-335

2 This result generalizes to where is the total number of nodes in the network, if and eachlink achieves the game capacity

Design of 16-QAM Space-Time Codes for Rapid Rayleigh Fading ChannelsSalam A. Zummo and Saud A. Al-Semari

Electrical Engineering DepartmentKing Fahd University of Petroleum & Minerals

Dhahran 31261, Saudi ArabiaTel: +966-3-860-3315, Fax: +966-3-860-2215

E-mail: {sazummo, semari}@kfupm.edu.sa

ABSTRACTThis paper proposes Space-Time (ST) codes for rapid fading channels using the 16-QAM signalconstellation. The design of the proposed codes uses two different methods. The first method uses a hightime diversity trellis encoder, and the second uses the I-Q encoding technique. Both encoding methodsare expected to produce ST codes that perform better than the codes presented in the literature. Optimaland suboptimal decoding algorithms are used to decode the I-Q ST codes. The proposed codes weresimulated over independent and correlated Rayleigh fading channels. Coding gains up to 3 dBs havebeen observed.

1. Introduction

Diversity is a popular method to improve the performance and throughput of wireless systems. Transmittime diversity can be achieved by repeating the transmission of each symbol in different time slots [1]. Itcan be viewed as a repetition code and consumes higher bandwidth [2]. Therefore, substantialperformance improvement can be achieved using more sophisticated codes, utilizing both space and time.The concept of ST codes had appeared first in [3] as the delay diversity system, where different symbolsare simultaneously transmitted via different transmit antennas. Later, ST codes were deigned explicitlyfor quasi-static fading channels [4,5]. Moreover, the performance criteria of ST codes were derived forquasi-static and rapid fading channels in [4,5]. In [6-8], the ST concept was applied to enhance the qualityof transmission at the same bit rate of systems using single transmit antenna. So, the same errorprobability can be achieved at a lower signal-to-noise ratio (SNR). ST coded QPSK schemes arepresented in [9,10] for rapid fading channels.In this paper, 16-QAM ST codes are proposed using two encoding schemes. The first scheme usesconventional trellis encoders, where the second one uses the I-Q encoding technique. The general STsystem model is described in the next section. Then, the proposed codes are presented. After that, theoptimal and suboptimal decoding algorithms used to decode the I-Q 16-QAM ST codes are discussed.The performance of the new codes is compared to that of the 16-QAM ST code presented in [4] for thecases of rapid and correlated fading channels. Finally, conclusions are drawn from the obtained results.

122

2. System Model:A typical system that employs ST coding uses N transmit and M receive antennas. In this work, N hasbeen set to 2. The received signal is a noisy superposition of all transmitted symbols over all transmit

antennas. The signal received at the antenna at time t is given by:

Where is an AWGN modeled as independent samples of a zero-mean Gaussian random process with

variance per dimension. The coefficient is the path gain from the transmit antenna to thereceive antenna at time instant t. The is the transmitted symbol from the transmit antenna.The performance of ST codes having N transmit and M receive antennas is derived in [4] for rapid fadingchannels. Consider a codeword that has been transmittedover l time intervals and was erroneously decoded as The conditional probability of deciding infavor of using maximum liklihood decoding is upper bounded as [4]:

where

and is the average signal energy at each transmit antenna. The pairwise error probability is found to be:

The parameter L, which is the length of the shortest error event with L time intervals, can be referred to asthe Space-Time Minimum Time Diversity (ST-MTD) of the code. It can be visualized as the “branch-wise” Hamming distance (HD) or the MTD in conventional trellis codes, by considering the wholecodeword as one symbol. The quantity multiplied by the SNR term can be referred to as the Space-Time Minimum Square Product Distance (ST-MSPD) and defined as:

The ST-MTD and ST-MSPD are referred to in [4] as Distance and Product criteria, respectively. So,maximizing both of them yields good ST codes suitable for rapid fading channels. The proposed ST codesare presented in the following.

3. The Proposed Codes:Different ST codes were designed in [4] for the quasi-static fading channel. The ST coded QAM scheme,referred to QAM1 here, uses a rate-4/8 trellis encoder to encode the incoming 4 bits to 8 output bits. The

123

8 bits at the output of the encoder are mapped onto two 16-QAM signals and transmitted over twoantennas. The ST-MTD of this code is 2 and its ST-MSPD is 0.16.The first proposed scheme, called QAM2, uses a rate-4/8 trellis encoder. However, it is designed so thatboth the ST-MTD and the ST-MSPD of the code are maximized. To be able to do this, the allowed pair of16-QAM signals to appear at branches departing or emerging into the same state should be different inboth symbols. This can be done by applying the permutation method used for the 4-dimensional MPSKsignal space in [11] with slight modifications.At the beginning, all the possible 16-QAM symbols are listed in order, starting by and ending within a vector. Then a vector is formed by listing all the pairs of same first and second symbolsand denoted by The vector, that has the second column of shifted by i rows is denoted by

Similarly, when the vector is shifted by j rows, it is denoted by Figure 1-a shows the vectors

and as examples of the permuted vectors. The labels of branches leaving each state are taken

as the rows of the vectors having the maximum HD from each other. The trellis diagram of the 16-stateQAM2 code is shown in Figure 1-bSince the MTD of a trellis code is inversely proportional to the number of input bits of the encoder, thenusing different encoders in parallel, such as I-Q encoding, can increase the ST-MTD. I-Q trellis codeswith different throughputs were presented in [12]. These codes show significant coding gains overconventional trellis codes having the same complexity. The proposed structure of the encoder/decoderemploying the ST concept is shown in Figure 2. It uses two rate-2/4 encoders, where each one encodestwo bits per signaling interval. Each encoder outputs two 4-AM signals: the first symbols from bothencoders are mapped onto a 16-QAM signal to be transmitted over the first antenna. The second symbolsare mapped onto the second 16-QAM symbol and transmitted over the second antenna.In order to design codes with the highest possible MTD and MSPD, the 2-dimensional 4-AM signal spaceis partitioned. The partitioning process is done so that the HD and the squared product distance (SPD) inthe generated subsets are higher each time the partitioning is performed. The set partitioning of the 2-dimensional 4-AM signal space is shown in Figure 3. The trellis diagrams of the 4 and 32-state codes areshown in Figure 4. It can be observed that the labels of branches departing or emerging at the same statediffer in both symbols. This maximizes the ST-MTD of the 4, and 32-state codes to 2 and 3, respectively.The ST-MSPD is 16.6 for the 4-state code and 106 for the 32-state code. The large difference in both theST-MTD and the ST-MSPD, compared to those of the previous two codes, and resulting from using the I-Q encoding scheme is clearly observed.The complexity of a trellis code is equal to the total number of branches leaving all states divided by theassociated information bits with each transition [12]. The complexity of the 16-state QAM1 and QAM2codes is 64. The I-Q code that has a similar complexity is the 32-state code. So, for fair comparison, the

124

125

126

16-state QAM1 and QAM2 codes are compared to the 32-state I-Q code. The decoding algorithms used todecode the I-Q ST codes are presented in the following.

4. Suboptimal Decoding AlgorithmTwo decoding algorithms are used to decode the I-Q 16-QAM ST codes: the optimal and a sub-optimaldecoding algorithms. The suboptimal decoding algorithm was proposed in [9,13] to decode the I-Q STcoded QPSK systems and is based on estimating the Q/I components in the I/Q decoders.This is performed by partitioning the 2-dimensional signal space available at the output of the STencoder. The signal space to be partitioned is a 4-dimensional 16-QAM space that consists of possiblesignal pairs. The partitioning is performed so that all pairs in one subset have the same in-phasecomponents. In other words, they are caused by the same 4-AM symbol pair at the output of the I -encoder. Hence, for each 4-AM symbol pair, there are 16 possible 16-QAM signal pairs that could betransmitted from both transmit antennas.The notation denotes the possible 16-QAM pairs that can appear at the output of the I-Q ST encodergiven that the 4-AM symbols at the output of the I-encoder are l and k. In addition, the signal is the 16-QAM signal whose label in the constellation is j, and is transmitted over the transmit antenna. The setpartitioning yields 16 subsets and one set is presented for illustration:

Now, the following 16 metrics are computed at the I-decoder before the trellis:

Where xi and yi are the in-phase and quadrature components of the 16-QAM signal In each 16metrics are compared and the minimum is found accordingly, ending with 16 different metrics. Each ofthem is associated with one of the 4-AM signal pairs that may be at the I-encoder’s output. Since eachencoder has two input bits, there are four possible metrics to be compared at each state in the trellis of theI-decoder. The same principle is applied to the Q-decoder case.The results of the I-Q ST codes show that the suboptimal decoding algorithm used does not perform wellbecause it is trying to guess for the best Q/I components from the received signals in the I/Q decoder. Fora decision in the I-decoder, there are 16 different combinations of the Q component to be compared,which is a large number. This algorithm is called suboptimal because the I-decoder uses sequencedecoding for the I components and symbol-by-symbol decoding for the Q components. In order to get themaximum performance of the I-Q ST codes, an optimal but complex decoding algorithm is used anddiscussed in the following section.

127

128

5. Optimal Decoding AlgorithmTo mitigate for suboptimality of the suboptimal algorithm, the use of the super-trellis decoding isproposed to decode the I-Q ST codes. In this algorithm, the I and Q components are decoded using onetrellis whose size is the square of that of the individual I and Q trellises.The super-trellis is used to decode I-Q ST codes optimally by decoding (not guessing) the I and Qcomponents simultaneously using sequence decoding. The optimal decoding algorithm is used to decodethe 4-state 16-QAM I-Q ST code only. The resultant super-trellis decoding complexity of this code is 64,which is the same as the complexities of the QAM1, QAM2 codes.

6. ResultsThe above ST codes are simulated over independent fading channels. Figure 5 shows the performance ofthe 16-state QAM1, QAM2 and I-Q codes for the cases of one and two receive antennas. The 4-state I-Qcode with super-trellis decoding is the best followed by QAM2, QAM1 codes and the 32-state I-Q codewith suboptimal decoding. This is expected for the 4-state I-Q code since the main controlling parameterof the code in rapid fading channels is the ST-MTD of the code that is highest in the I-Q code.It can be seen that the 32-state I-Q code with suboptimal decoding does not perform well in the singlereceive antenna case. Also, degradation decreases as the SNR in increased, because the guessing processis done in a less noisy environment. This degradation becomes less in the case of two receive antennas,where the 32-state I-Q with suboptimal decoding is the best. In the case of one receive antenna, the gainsof the I-Q code, decoded by the super-trellis method, over the QAM1 and QAM2 codes are 2.5 and 0.5dBs, respectively. In the two receive antennas case, the above gains become around 2 and 0.3 dB,respectively. On the other hand, gains of the 32-state I-Q code, using suboptimal decoding algorithm,over QAM1, QAM2 and the 4-state code with super-trellis decoding are 2.5, 1 and 0.5, respectively.The same codes are tested over a correlated fading channel with a fade rate of 0.01, and the result sare shown in Figure 6. A block interleaver is used to break the memory of the channel. The sametrends observed in the rapid fading channel case are observed in this case. The gains of the best code overworse codes are less in the one receive antenna case. In the two receive antennas case, they do not changemuch because the presence of two receive antennas makes it less dependent on the codes’ parameters.Figure 7 shows the performance of the codes for a slower channel, with a fade rate of 0.005. Theabove interleaver is used which is improper for this channel. The performance trends for the codes are thesame as the previous two channels with decreased gains.

7. Conclusions and DiscussionTwo new ST codes based on the 16-QAM signal constellation for rapid fading channels are proposed.The results showed that the new codes are better in terms of the design criteria. Also, they were tested

129

130

over different fading rates and they showed to be robust in such environments. The optimal andsuboptunal decoding algorithms were used to decode I-Q ST codes. Results indicated that the suboptimalalgorithm is less complex than the optimal one at the expense of degradation in the performance,especially for the case of one receive antenna. More gains are expected from the I-Q ST codes if betterand simple decoding algorithms are used.

Acknowledgements:The authors wish to acknowledge the support of King Fahd University of Petroleum and Mineralsprovided to conduct this research.

REFERENCES

[1] J. Poakis, Digital Communications. New York: McGraw-Hill, Inc., 1989.[2] Sklar, “Rayleigh Fading Channels in Mobile Digital Communication Systems Part II: Mitigation,” IEEE Comm.

Magazine, September 1997.[3] A. Wittneben, “A New Bandwidth Efficient Transmit Antenna Modulation Diversity Scheme for Linear Digital

Modulation,” Proc. IEEE ICC’93.[4] V. Tarokh, N. Seshadri and A.R. Calderbank, “Space-Time Codes For High Data Rate Wireless

Communication: Performance Criterion and Code Construction,” IEEE Trans. Info. Theory, March 98.[5] N. Seshadri, V. Tarokh and A.R. Calderbank, “Space-Time Codes For High Data Rate Wireless

Communication: Code Construction,” Proc. IEEE VTC’97.[6] Siavash Alamouti, “A Simple Transmit Diversity Technique for Wireless Communications,” 1EEE JSAC, Oct.

98.[7] V. Tarokh, S. Alamouti and P. Poon, “New Detection Schemes for Transmit Diversity with no Channel

Estimation,” Proc. IEEE Int. Conf. On Universal Personal Comm. ’98.[8] S. Alamouti, V. Tarokh and P. Poon, “Trellis-Coded Modulation and Transmit Diversity: Design Criteria and

Performane Evaluation,” Proc. IEEE Int. Conf. On Universal Personal Comm. ’98.[9] S. Zummo, “Performance and Design of Space-Time Trellis Codes for Wireless Channels.” M.S. Thesis, King

Fahd University of Petroluem & Minerals, Dec. 1999.[10] S. Zummo and S. Al-Semari, “Space-Time Coded QPSK for Rapid Fading Channels,” Submitted to the

PIMRC’2000.[11] D. Divsalar and M. Simon, “The Design of Trellis Coded MPSK for Fading Channels: Set Partitioning for

Optimum Code Design,” IEEE Trans. On Comm., Sep. 1988.[12] S. Al-Semari and T. Fuja, “I-Q TCM Reliable Communication Over Rayleigh Fading Channel Close to the

Cutoff Rate,” IEEE Trans. On Infor Theory, Jan 1997.[13] S. Zummo and S. Al-Semari. “A Decoding Algorithm for I-Q Space-Time Coded Systems in Fading

Environments,” submitted to the IEEE Vehicular Technology Conference VTC’2000.

Transmit Diversity With More Than Two AntennasR. Michael Buehrer, Robert A. Soni and Quinn Li

Bell Laboratories - Lucent Technologies67 Whippany Rd. Room 3A-220

Whippany, NJ 07981mbuehrer@bell-labs.com

Abstract Recently, a new form of transmit diversity has been developed and included for cdma2000 , thethird generation successor for IS-95 code division multiple access (CDMA) systems. This transmit diversityscheme was developed by Bell Laboratories and uses Space-Time coding techniques. This paper presentsperformance results of this new transmit diversity method, termed space-time spreading, and investigates”open-loop” improvements to the scheme which use more than two antennas. Additionally, we investigatethe performance improvements possible with closed loop techniques.

1 IntroductionRecently, Phase II of the cdma2000 standardization process has been completed where a review of forwardlink antenna techniques was completed. A few different schemes were proposed as possible enhancements forthe system. The standard originally supported a method of transmit diversity known as orthogonal transmitdiversity (OTD). This method offered significant performance gains for rate 1/4 convolutional codes at lowspeeds, but did not offer the same types of gains for rate 1/2 codes. The scheme is open-loop, and makes nouse of user specific data such as location or condition of its channel, other than through user independentpower control.

Through the efforts of Bell Laboratories [1, 2, 3], an additional open-loop scheme was developed whichsignificantly improved performance of weaker convolutional codes or codes with higher rates. This schemewhich we will term “Space-Time Spreading” or STS can offer significant performance gains over the existingform of open-loop transmit diversity. This scheme is similar in concept and performance to the schemesproposed for the UMTS system (W-CDMA) [4], and by Tarokh et al. [5] for TDMA applications. This paperdiscusses the performance of this scheme, and shows some of the performance results which were used to winapproval of this scheme in the cdma2000 standardization process. Additionally, we investigate methods ofimproving the performance of Space-Time Spreading through the use of more than two transmit antennaswith and without feedback (i.e., closed loop techniques).

2 An cdma2000 System ModelFor a system with K mobiles receiving signals from a common base station, the transmitted signal on asingle antenna can be modeled as:

where is the power transmitted to the ith mobile, and are the data signal and unique Walshfunction intended for the ith mobile respectively, is the power of the pilot signal which uses Walsh function0, and p(t) is the covering code for the base station of interest. Further, the Walsh functions are orthogonaland repeat every symbol time i.e.

132

where is the complex multiplicative distortion caused by the wireless channel and n(t) is thermal noiseand all other interference. Mobile i correlates the received signal with the ith Walsh function during the kth

symbol interval after uncovering to achieve the decision statistic

where represents the cumulative effect of the channel over the kth symbol interval, and is thekth transmitted symbol for the ith mobile.

The transmitted symbol can be recovered by using an estimate of the channel distortion obtainablefrom the pilot channel, i.e.,

where f(·) is an appropriate decision function. Alternately, in a coded system may be used directlyas a symbol metric. If the channel is a flat, slow Rayleigh faded channel, in the absence of fast, accuratepower control, the resulting performance of the link will be rather poor due to the lack of diversity. Asa result, it is desirable to have a second antenna at the receiver to allow diversity reception, improvingperformance considerably. However, mobile handsets do not easily allow a second antenna to be added.

3 Transmit Diversity MethodsOne method of achieving diversity performance is to transmit the same signals on multiple carriers. However,this is wasteful of the one resource we cannot afford to waste in mobile communications, namely bandwidth.As an alternative, re-transmitting the same waveform with a chip-level delay, also known as delay diversity,can help performance in some instances, but it can also degrade performance in other instances as it increasesthe amount of self-interference which degrades the performance of a typical Rake receiver.

Orthogonal transmit diversity (OTD) which is available in cdma2000 as an option transmits half of thebits via one antenna and half of the bits via a second antenna spaced approximately 10 wavelengths away.The received stream of coded bits will be

Using a Viterbi decoder, the link-level performance of the forward link becomes a function of the qualityof both channels. Transmissions via channels with slow fading conditions benefit greatly from this method.However, this method offers less performance gain as the speed increases1 and the code rate decreases.Specifically, with high rate codes, the performance gains are reduced since OTD relies on the decoder toobtain the diversity. This problem is alleviated by a technique termed Space-Time Spreading or STS.

4 Space-Time SpreadingBased upon space-time block codes, attributed to Alamouti [6], a signal transmission scheme which utilizedthe multiple orthogonal code structure already available in the standard was developed for cdma2000 [2, 3].

1The interleaver helps compensate for loss of bits during bursty errors, because at high speeds, the duration of the errorstend to be shorter.

At the mobile the following signal is received on a single antenna

133

This method also uses two antennas spaced approximately 10 wavelengths apart. On the first antenna wetransmit

and on the second antenna we transmit,

where and represent the even and odd streams of bits, respectively. Note that a separate Walshcode is required for each transmit antenna to support a pilot on each antenna. Since the data rate has beenreduced by a factor of two by this scheme for each stream of bits, it is possible to use double length Walshcodes and not utilize additional bandwidth or Walsh resources. That is we can convert a single Walsh codeinto two double length Walsh codes using the repetition pattern

At the receiver, we again uncover and correlate with the two Walsh codes. At the output of the two Walshcorrelations we obtain (dropping the dependence on symbol interval)

where and are the effects of the complex channel. This obviously introduces interference terms in thedecision statistics. However, if we have estimates of the channel distortions and from pilot signals 1and 2, we can obtain a signal estimate for the even bits by

134

where we’ve assumed that the channel estimation is exact, and Similarly, we can estimatethe data for the odd stream of bits as

It can be easily shown that this is identical to the decision statistic for two-antenna diversity reception(without the 3dB aperture gain) [7].

The performance of STS was simulated for the cdma2000 standard using a one path Rayleigh fadingchannel model for the fundamental channel [8]. The transmit power fractions (i.e., the required fraction ofthe base station power), for full-rate voice using radio configurations RC3 and RC4 [8] were derivedfrom simulation. represents the energy per chip, and represents the total transmit power spectraldensity. The quantity, represents the ratio of the transmit power spectral density to the out of cellinterference plus any additional thermal noise. It is commonly referred to as “mobile geometry” with lowvalues associated with mobile locations near the edge of the cell and high values associated with mobiles closeto the base station. The geometry is directly related to the signal to noise ratio of the decision statistics.The effects of power control, puncturing, and coding using the interleaver specified in the ballot version ofthe proposal [8] were included. The major simulation parameters are summarized in Table 1.

The simulation results for the fundamental channel are summarized in Figure 1. For both RC3 and RC4,it is clear that STS offers a significant performance advantage. Since RC3 uses 1/4 rate convolutional codeswhile RC4 uses 1/2 rate codes, RC3 uses length 64 Walsh codes while RC4 uses length 128 Walsh codes.Thus, RC4 is less likely to experience a capacity limit due to a Walsh code limitation and may be preferablein situations where the number of Walsh codes is a concern.

As shown in Figure 1 for a geometry of STS offers up to 5 dB performance improvement

135

over OTD at low speeds, and a minimum of 1.5 dB improvement at high speeds for RC4. For RC3, the gainsof STS over OTD are smaller due to the stronger convolutional coding, with STS offering a minimum of 0.5dB improvement over OTD and up to 1dB at low speeds. STS also provides significant improvement overno diversity achieving gains of 1-5dB. Most importantly transmit diversity helps where the system needs itmost, at low speeds. This flattens out the performance curve versus speed and increases capacity.

5 Transmit Diversity with Four AntennasWe have seen that adding an additional antenna at the base station to provide transmit diversity is beneficialfor cdma2000 . The next question we must ask is “Can we improve upon this with additional antennas?”In this section, we discuss three options for extending this diversity scheme to four antennas. Extension tothree antennas is also similar with one column of the transmission matrices being ignored. Note that thisextension is meant to increase diversity performance, that is we can achieve higher orders of diversity byusing more transmit antennas.

To allow four transmit antennas, we first extend the Walsh code for a particular user twice to obtain fourWalsh codes with four times the length, where the extension pattern is

This allows for no code sharing and can be compared to the STS case discussed previously.To help describe the method of transmission, we define the concept of a transmission matrix. The

transmission matrix simply describes the way symbols are transmitted. The rows of the matrix determinethe Walsh codes used and the columns determine the antennas on which the symbols are transmitted. Forexample, we can see that from equations (7) and (8) in STS the transmission matrix is

To obtain a transmission matrix for four transmit antennas, we require an orthogonal matrix with fourcolumns and thus at least four rows [9]. While a 4 × 4 orthogonal matrix with four complex variables doesnot exist [5], we find that a 4 × 4 matrix with three complex variables does exist. One such transmissionmatrix is

The received vector of Walsh outputs is then where is the vector of complex channel distortions.This can be rearranged as To remove the self-interference we apply the channel matrix to

That is and Thus, we can achieve four-fold diversity.However, in order to achieve this, we must reduce the data rate to 3/4 the original rate. This can be seenby noticing that while we use four codes (i.e., the rows of T) we only transmit three symbols.

A second option for using four transmit antennas without reducing the data rate, is to use the transmissionmatrix

136

Which guarantees orthogonality, but only achieves two-fold diversity before the decoder. However, if theinterleaving is done correctly, we see that going into the decoder the metrics are:

Thus, while two-fold diversity is achieved prior to decoding, the Viterbi decoder can see up to four-folddiversity in the path metrics. Thus, while we rely on the decoder to achieve the diversity gain from 2 to 4,we do not lose data rate.

The last option is similar to Option 1 and uses the orthogonal design from [5]. The transmission matrixis

This option also achieves four-fold diversity prior to decoding, but also suffers from a 25% loss in data rate.The main difference between this option and Option 1, is that this allows all four codes to be used on allfour antennas.

All three of these options essentially provide four-fold diversity performance although option 2 will suffersome degradation when puncturing is included on higher rate codes just as in OTD. However, we shouldnote two things. First consider Figure 2 (a). This figure plots the theoretical performance of several optionsfor four transmit antennas. Included are the theoretical performance of four branch diversity, two-branchdiversity with two-branch aperture gain, four-branch aperture, and four-branch diversity and aperture gain.The theoretical performance is well known to be

where L is the degree of diversity and is the SNR per branch without any aperture gain,and N is the amount of aperture gain being achieved. Also shown in Figure 2 (a) for comparison purposesare no diversity and two-branch diversity. We can see that going from two-branch to four-branch diversitydoes not provide huge gains particularly at high BER’s. Additionally, we see that two-branch diversity withtwo-branch aperture gain provides better performance in the region of 1% BER than four-branch diversity.As a second note, consider Figure 1. Notice that at high speeds the advantage of diversity is diminishedcompared to the gains at low speeds. This is due to the diversity obtained in the decoder due to fast fadingand interleaving. Additionally, the benefits of additional diversity in frequency selective fading are much lessthan in flat fading. This suggests that increased diversity performance is not necessarily the best option.However, in order to implement the methods shown in Figure 2 (a) which are more than pure diversity,either feedback or uplink-based estimation is needed. We will discuss this next.

6 Methods which Require FeedbackThe preceding discussion of higher order diversity was based on the premise that we wished to improveperformance without requiring mobile feedback or uplink estimation. In this section we discuss the perfor-mance improvements when mobile feedback is allowed. There are essentially three options for performanceimprovement with four antennas when feedback is allowed. They are

• Four element fully adaptive transmit diversity

• Four element steered beam

137

• Combined transmit diversity and steering (Steered STS)

The first option has the largest potential gains and constitutes using a four element array with large elementspacing. Relying on mobile feedback the transmitter would dynamically adjust the phase and gains of eachantenna to ideally coherently combine all four transmit antennas. This would ideally provide both diversityand aperture gain but obviously requires fast accurate mobile feedback. Such feedback would be limited byinherent delays and would need to be added to the current standard. The second option does not necessarilyrequire a standards change but does not achieve any diversity gain and might be susceptible to deep fades.A steered beam solution achieves approximately a 6dB aperture gain, although the gains could be less in arich scattering environment. This would require either mobile feedback (i.e., a standards change) or uplinkdirection-of-arrival (DOA) estimation. The third option also does not necessarily require feedback and thuswould not require a standards change. It provides two branch diversity as well as two times aperture gain.

To provide some feel for the performance of the above options, simulations were run without power controlor coding and a 1.25ms average pilot filter providing a pilot The DOA for the S-STScase was estimated over one frame at the mobile and fed back to the base. Figure 2 (b) plots the resultsof these simulations for no diversity, STS, combined STS and beam-steering (called Steered STS or STS-Feedback), and a fully adaptive array with four diversity branches. The STS and Steered STS results assumetwo independently faded Rayleigh channels, while the fully adaptive array assumes four independently fadedRayleigh channels. Feedback is perfect, i.e., full precision, no delay and no feedback error. The resultsmatch well with theory. STS provides two-branch diversity gain over the no diversity case (4-5dB at 1%BER). Steered STS provides a 3dB gain over STS and the fully adaptive array achieves both a 3dB apertureimprovement and a diversity improvement over Steered STS.

The preceding simulations assumed that feedback was perfect. In Figure 3 (a) we remove this constraintand allow the feedback to be quantized by 4 bits and bit errors to occur in the feedback process. Further weexamine the performance as the feedback error rate increases from zero to 20%. Of course the performanceof no diversity and STS remains unchanged. Surprisingly, the performance of S-STS also remains unchanged

138

over the error rate of interest. However, we see that the performance of the fully adaptive array degradessignificantly as the feedback error increases. In fact the fully adaptive approach loses all of its advantageover S-STS if the feedback error is 10% or higher. Note that in order to keep the delay to a reasonable levelthe feedback will have to be uncoded (i.e., we cannot wait for the Viterbi decoding of the entire frame toobtain feedback bits). Thus, the error rate on the bits could be fairly high.

The reason that S-STS is relatively immune to the feedback error is that the range of phases required tosteer antennas over is fairly small and thus any error perturbs the beam only a small amount. On theother hand with the fully adaptive approach the antennas require a much larger range of phase adjustments.This makes the range of phase error larger as well as the resulting degradation. Note that the fully adaptiveapproach also requires three times the feedback.

A second major degradation in a feedback system is due to delay. In cdma2000 there is a minimum2.5ms delay (equivalent to two power control groups). This will obviously be a problem for the fully adaptiveapproach as the fading rate increases. Figure 3 (b) presents the performance of the schemes with 4% feedbackerror as the Doppler rate increases from 10Hz to 180Hz. We can see that when feedback error and delayare considered, the fully adaptive approach suffers dramatically. Note that the performance of the fullyadaptive approach degrades beyond that of the no diversity approach due to the feedback errors combinedwith improper weighting. The receiver relies on correct knowledge of the feedback bits in order to correctlyphase the pilot. Thus, feedback error can cripple performance. Note that all suffer degradation at highspeeds due to the pilot filter length.

7 Steered Space-Time SpreadingResults from the previous section lead us to consider the use of STS with beam-steering, i.e., combineddiversity and aperture gain. In this section we describe this idea more fully. Consider a linear antennaarrangement as shown in Figure 4. The base station transmits on M antennas divided into two groups. Group

139

A has antennas2 with inter-element spacing of approximately where is the carrier wavelength.3

Group B is separated from Group A by a distance large enough to insure that the two groups experienceuncorrelated fading and contains elements.

On antennas within Group A the transmitted signal (ignoring other users) is

where i represents the antenna index which is for Group A, represents the transmit power forthe data from antenna i, and are the data stream divided into even and odd streams respectively, isthe power in the dedicated pilot on each antenna, is the Walsh code used for the pilot on that antenna4,

is a complex weight to be described later, and are extended Walsh codes as described earlier,and p(t) is a pseudo-random sector-specific covering code.

On antennas within group B the transmitted signal (ignoring other users) is

where Note that for M = 2 we have Space-Time Spreading [3]. At the receiver we havea single signal

where and are the time-varying multiplicative distortion due to Rayleigh fading seen from groupsA and B respectively, is the distance of the ith antenna from an arbitrary reference,is the angle formed between a line drawn from the base to the mobile and the array baseline, and n(t) istemporally and spatially white complex Gaussian noise. By correlating the received signal (after removingthe long code) with and and assuming that the channel is static over the integration period, weobtain the following correlation outputs:

2 [x] is defined as the largest integer less than or equal to x.3 Note that the exact element spacing is not crucial. However, it must be small enough so that the signals transmitted from

all elements experience highly correlated fading. Additionally, a spacing of much greater than will introduce grating lobeswhich is generally undesirable.

4 As we will discuss later, a pilot per antenna is only necessary if mobile feedback is being used to steer the array. If uplinkestimation is used to steer the array, only one pilot per group is necessary.

140

Now, using the pilots we can obtain estimates for and We can then use the estimates to create decisionstatistics for the even and odd streams as

where f{·} is an appropriate decision function. Expanding the first equation and assuming perfect channelknowledge for simplicity’s sake results in

From this equation we can see two things: (1) we wish and (2) we must set the transmitpowers such that For an even number of antennas, the second condition issatisfied by giving all antennas equal power. The first condition however, must be accomplished by either (a)using information from the uplink to estimate or (b) using mobile feedback. We will discuss the optionsfor mobile feedback in a moment.

Assuming that the two above conditions are met, the decision statistic for is

Defining the SNR as and for the moment assuming that M is even (i.e., we can seethat the SNR is a Chi-Square random variable with 4 degrees of freedom (i.e., two-fold diversity) and anexpected value of

where we have assumed that Thus, we have an improvement of in SNR when compared tothe case of standard STS which sees no aperture gain but merely a diversity gain.

141

7.1 Calculating Weights7.1.1 Using an Uplink Array

A key to the scheme is the set of weights To maximize SNR we must set One methodof setting the weights is to attempt to estimate from uplink information. In the presence of this uplinkarray we can estimate by measuring and using the relation After estimating theweights are set to

However, this assumes that the distance between elements is known, the elements are phase matched, andthere is symmetry between the uplink angle-of-arrival and downlink angle-of-arrival. All of these are eitherreasonable to assume or could be obtained through calibration. In this case, if the main pilot is put on thefirst antenna of each group and the other elements are phased with respect to it, the transmit signals pergroup will arrive at the mobile in phase and thus only one pilot per group is necessary. The mobile stationin such a system would not need to know that beam-steering was being used.

7.1.2 Feedback Options

A second means of calculating the set of weights is to rely on mobile feedback. Since the weights dependultimately on the angle-of-arrival, they must only be updated at the rate at which changes which islikely very slow compared to channel fading rate. There are several possible methods of employing feedback.

The most straightforward method of feedback is to transmit a dedicated pilot on each antenna and feedback the phase of the received pilots. One pilot per group could be used as a reference and the phase ofthe other pilots with respect to the reference pilots are then fed back. This requires M – 2 phase values befed back per update. For q bit quantization and F Hz feedback rate, this method requires bpsfeedback. This method makes no assumptions about the array spacing and is thus robust to imperfectknowledge of the inter-element spacing.

Another method which requires less feedback is to feedback a single value for the entire array. If theinter-element spacing within each group of elements is the same, the elements should differ by a constantphase While the method is simpler and requires less feedback, it is more sensitive tonon-ideal element spacing. The feedback rate would be qFbps. For small array sizes (eg., M=4) this maynot be a significant savings.

7.2 PerformanceTo investigate the detailed performance of Steered STS in cdma2000 , simulations were run using thecdma2000 standard. The simulation assumptions are given in Table 1 and with the exception that S-STS uses two groups of two antennas, /ie M = 4, only RC3 is simulated and Note thatthe geometry assumed is much worse (i.e., lower ) than in Figure 1 thus the higher required powerfraction than the previous case. The array assumed had four antennas (i.e., two pairs).

We can see that Steered-STS provides significant performance improvement over both STS (approxi-mately 3dB) and no diversity. An important point about S-STS is that even at high mobile speeds, thescheme still achieves significant gains over the baseline, which is not true of transmit diversity in general.That is the gains are not speed dependent. However, we see that the performance of S-STS is fairly flat withrespect to speed.

142

8 ConclusionIn this paper we have investigated several possible extensions of transmit diversity to four antennas forthe Third Generation standard cdma2000 . It is clear that there are potential benefits from increasing thenumber of antennas at the mobile station. These benefits to the cdma2000 downlink will be significant asdata becomes more predominant.

AcknowledgmentsThe authors wish to express their appreciation to their colleagues at Lucent-Bell Labs who have contributedto the work in this paper including Roger Benning, Steve Nicoloso, Nallepilli Ramesh, Steve Allpress, Con-stantinos Papadias, Bert Hochwald and Tom Marzetta.

References[1] C. Papadias, B. Hochwald, T. Marzetta, R.M. Buehrer, and R. Soni , “Space-time spreading for CDMA

systems,” Stanford Sixth Workshop on Smart Antennas for Mobile Communications, July 22-23 1999.

[2] Lucent Technologies, “Performance of Space Time Spreading (STS) for IS-2000,” Contribution 3GPP2-C30-19990914-013, September 1999.

[3] R. Soni, R. Buehrer, and J.-A. Tsai, “Open-loop transmit diversity methods in IS-2000 systems,” inProceedings of the Asilomar Conference on Signals, Systems and Computers, October 1999.

[4] Texas Instruments, “Space time block coding improvements,” ETSI - UMTS Contribution, December1998.

143

[5] V. Tarokh, H. Jafarkhani, and A. Calderbank, “Space-time block codes from orthogonal designs,” IEEETransactions on Information Theory, vol. 45, pp. 1456–1467, July 1999.

[6] S. M. Alamouti, “A simple transmitter diversity scheme for wireless communications,” IEEE J. Select.Areas Commun., vol. 16, October 1998.

[7] J. Proakis, Digital Communications. New York, NY: McGraw-Hill, third ed., 1995.

[8] TR45.5, Physical Layer Standard for cdma2000 Spread Spectrum Systems. TIA/EIA/IS-2000.2, 1999.(Ballot Version).

[9] B. Hochwald and T. Marzetta, “Unitary space-time modulation for multiple-antenna communications in.rayleigh flat fading,” IEEE Transactions on Information Theory, vol. 46, pp. 543–564, March 2000.

Reduced Complexity Space-Time Optimum Processing

Jens Jelitto, Marcus Bronzel, Gerhard FettweisDresden University of Technology, Germany

Abstract

New emerging space-time processing technologies promise a significant performance increaseof wireless communication systems. The particular application and scenario strongly influencesthe amount of possible performance and capacity increase if antenna arrays are deployed at thebasestation (BS) and/or at the mobile terminal (MT). The achievable gain is mainly determinedby the spatial correlation properties of the underlying physical transmission channel.

This paper analyzes the spatial correlation properties for various scenarios and investigatesthe procuring requirements for designing apace-time optimum receivers. One aim is to reducethe spatial signal dimension to the information bearing components applying orthogonal trans-formation techniques. It will be shown that even for virtually uncorrelated spatial channels whichare characterized by high delay and angular spread, the spatial dimension can be reduced sig-nificantly. This enables less complex receiver structures and more robust channel estimationtechniques.

1 Introduction

Multiple antenna concepts are commonly regarded as a promising technology to increase the perfor-mance of wireless communication systems. The concept of Space Division Multiple Access (SDMA)enables higher user capacity within a cell if the users can be separated spatially. Furthermore, spa-tial filtering and cancellation of undesired users results in reduced interference. Other multipleantenna concepts include the steadily growing research fields of beamforming, spatial diversitycombining and space-time processing.

This paper investigates the performance of a single link between a MT and a BS and does nottarget on system capacity issues. Depending on the given wireless channel conditions the differentconcepts such as beamforming, space-diversity combining or fully armed space-time processingwill be more or less suitable. The multipath situation has to be considered in order to selectthe appropriate algorithm. If for instance a strong Line-of-Sight (LOS) component is presentwith only weak temporal and spatial spreading, pure beamforming provides an efficient approach.Space diversity combining would not be appropriate in this scenario since the antenna signals areessentially phase shifted copies of each other and therefore highly correlated. In a scenario withconsiderable multipath delay and angular spread space-time processing promises optimum results.The degree of signal correlation between different antenna elements influences the possible gainwhich can be achieved with such a receiver concept. We investigate the spatial correlation properties

at the receiving antennas for various scenarios, which determine the additional information that canbe gained by every additional spatial dimension (or antenna) and can be used as a measure of therequired spatial receiver complexity. We introduce a linear transformation based on singular valuedecomposition (SVD) [1] in order to reduce the potentially correlated M antenna signal streamsto an uncorrelated signal stream of lower order D, which is then fed into a reduced complexityspace-time receiver (Figure 1). An example will show the potential gains of this approach.

2 Signal Model

This section introduces a signal model for a single input multiple output (SIMO) system [2] usingone transmit antenna and M receive antennas including the channel characteristics. In order tosimplify the discussion linear modulation is assumed. The basic system model is shown in Figure 2.A binary data stream is mapped onto complex symbols The continuous-time representationof the symbol stream can be written as

with as symbol duration. This signal is band-limited by a transmitter (Tx) pulse shapingfilter The complex baseband representation of the filtered transmit signal results in

The band-limited signal s(t) is transmitted over a linear time-variant channel. Wireless communi-cation channels are affected by multipath propagation with a possible LOS component and indirectpaths resulting from reflection, scattering and diffraction at several objects in the propagationenvironment. This channel can be modeled by a physical channel impulse response (CIR) [3]

146

147

with P as the number of propagation paths, each of which is characterized by a delay and acomplex weight Both parameters are generally time-variant, but changes in the path delaysare usually slow compared with the symbol duration.

The received signal x(t) can be obtained from the convolution of the transmitted signal (2) withthe physical CIR (3) as

where n(t) is the additional noise term. Interference i(t) from other users as indicated in Figure 2will not be considered in this paper. From (4) an effective CIR can be derived, includingtransmitter filter and physical channel characteristics with fractional delays [3].

Applying multiple antennas with M elements at the receiver effectively introduces M separatechannels. Here, it is assumed that the multipath components at the antenna elements differ onlyin the path length relative to a reference antenna element resulting in additional path delaysfor every receive antenna m. The overall path delay at antenna m is given by

which results in an effective CIR at antenna m

If the rate of change of the received signal envelope is slow compared to the propagation time acrossthe array (narrowband assumption), which applies for most wireless systems, as long as the signalbandwidth is small relative to the carrier frequency f, the additional delays can be regardedas pure phase shifts,

where The effective CIR at antenna m can then be written as

where the path weights of the P elementary rays are assumed to be identical for all antennas.Combining the M impulse responses of the SIMO channel

148

and collecting the phase shifts in an array response vector

leads to the reformulation of the SIMO CIR in matrix notation,

A is the array response matrix of dimension containing the P vectors as columns. Thematrix is a diagonal matrix containing the values of the pulse shaping filter for allpaths as elements, Finally, c(t) contains the P path weights at timet. Using (4) and (11) the resulting M-dimensional received signal vector can now be written as

3 Spatial Correlation of the Received Signal

The spatial correlation properties of the received signal are strongly influenced by the parameters ofthe effective CIR This includes the effects of the pulse shaping filter as well as the multipathcharacteristics of the physical channel. Additionally, the spatial correlation of the received signalat the antenna array is influenced by the temporal correlation properties of the transmitted symbolsequence The spatial correlation matrix of the received signal x(t) can with (12) bedefined as

The transmitted sequence is i.i.d,

and zero-mean. Furthermore, we assume the noise terms at the antenna elements m to bezero mean, uncorrelated with and temporally and spatially white. With these assumptionsthe spatial correlation matrix as defined in (14) can be simplified to

with the noise covariance matrix and a constant scaling factor determined by and themean symbol energy. This factor can be neglected, since it doesn’t influence the matrix properties.With the assumption of constant path delays and AOA’s the array response matrix and pulse

149

shaping matrices are deterministic, which simplifies (16) to

The signal part of the correlation matrix is determined by the matrix form where Prepresents the inter-signal coherence matrix of dimension corresponding to the number ofmultipath components. The properties of the inter-signal coherence matrix are influenced by thecomplex path weights c(t), which are basically determined by the attenuation of the multipathcomponents, and the characteristics of the pulse shaping filter

Following the P-matrix will be analyzed in more detail for different multipath scenarios withrespect to mobility and delay spread to gain some insight into its structure. The reason is that,besides the influence of angular path distribution represented in the array response matrix A, theinter-signal coherence matrix determines the rank of the spatial correlation matrix and thereforeis important for determining the possible dimension reduction. One limiting case of the P-matrixoccurs when the signal copies arriving over P multipaths are coherent which implies identical pathdelays P will in this case be a rank one matrix. The other limiting case occurs with Pmutually uncorrelated signals, where P will have full rank. In realistic scenarios the rank of P mayvary between these values.

Scenario 1: no Mobility, no Delay Spread A static environment with multipath propagationbut negligible delay spread results in space-selective fading. The complex path weights

can be considered as constant due to the stationary environment. For every delaythe entries of for all paths are identical. Therefore, the pulse shaping matrix can bereplaced by a scalar value for every k, which leads to a P-matrix of the form

Clearly, the inter-signal coherence matrix P has rank one independent from the length of the pulseshaping filter, since the outer product has rank one and is a scalar. In this case,the received signal vector x(t) defined in (12) can be written as

where a(t) defines the spatial signature, which is the weighted sum of the array response vectors,

The spatial correlation matrix for space-selective channels is then given as

150

In the noise-free case this matrix will be of rank one according to P, independent of the angularmultipath distribution. This suggests a beamforming or optimum combining approach to weightthe antenna signals according to the spatial signature, which reduces the signal dimension from Mto 1.

Scenario 2: no Mobility, Delay Spread For scenarios with considerable delay spread,the channel is considered to be space-frequency-selective. This is typically the case for excessdelays of The different path delays cause the pulse shaping matrix

to contain independent entries, whichresults in the inter-signal-coherence matrix

The rank of this matrix is now determined by the number of multipath components andtheir respective delays as well as by the length of the pulse shaping filter. Every sub-matrix

in (22) has rank one. However, the summation of k rank one matrices leadsto a matrix with a rank not higher than k. For filters of finite length the rank of P islimited to min(P, L). Figure 3 shows the magnitudes of a typical P-matrix. Here, the path delaysare uniformly distributed in The weights have random phases and decaylinearly with increasing path delay. The pulse shaping filter was modeled as a root-raised cosinefilter with roll-off factor 0.5. The gray levels indicate the amount of inter-signal coherence betweenthe delayed paths, where black indicates total coherence and white no coherence. In most casesit has been observed, that this matrix has full rank or some rank deficiency. The degree of the

151

rank reduction depends on the given multipath situation. This also holds for the received signalcorrelation matrix

where in the noise free case the rank of this matrix is less or equal to min(M, rank(P)). However,since in most cases this information is not sufficient to be applied efficiently as a dimension reductioncriterion, analyzing the eigenvalue properties of the spatial correlation matrix will provide a valuabletool for estimating the effective signal dimension, as will be shown later.

Scenario 3: Mobility, Delay Spread In this most general case the wireless channel is selectivewith respect to space, frequency, and time. The space-time correlation function for two antennaelements and is defined as

The analysis of this correlation function can only be simplified for particular scenarios. One impor-tant special case occurs when the temporal correlation is decoupled from the spatial correlation.Then the matrix of space-time correlation functions can be written as

with a separable matrix of spatial correlation functions1 and a temporal correlation factorThe P- and matrices can be determined from (17).

4 Subspace Methods and Spatial Dimension Reduction

As stated earlier, the rank of the spatial signal correlation matrix is often not a sufficient measureto determine the appropriate spatial dimension of the receiver. Therefore, an eigenanalysis of thespatial correlation matrices is conducted using SVD [1]. The matrix can be decomposed into

with U and V as left- and right-hand side eigenvalue matrices of Since is hermitian,it follows is a diagonal matrix containing eigenvalues of sorted in descendingorder, An important property of the normalized eigenvalues is, thattheir magnitudes represent a measure of signal energy contained in the signal components afteran orthogonal coordinate transformation (OT). If the spatial correlation matrix has full rank, thesignal energy is distributed across all components. However, applying an OT concentrates thesignal energy within the first components.

1The matrix of spatial correlation functions is related to the spatial correlation (covariance) matrix throughdue to the differences in the definitions for complex valued autocorrelation functions and complex valued

correlation matrices for vector processes [4].

152

Usually in array signal processing the performance gain in noise-limited environments can bedetermined by the SNR gain per antenna branch. With the assumptions of identical averageSNR’s at the M antenna elements and spatially uncorrelated noise, the SNR gain is 10 log M.Figure 4 shows the average SNR performance gain depending on the number of signal branchesconsidered. These results are normalized with respect to the average SNR in each branch. Applyingan orthogonal transform to the received signal shows that the first signal components alreadyprovide most of the possible SNR gain, although the corresponding spatial correlation matrix hasfull rank. The maximum possible SNR gain is the same as without transformation, if all signaldimensions are used. Commonly applied criteria for rank estimation in noisy environments such asthe Akaike information-theoretic criterion (AIC) or Minimum description length criterion (MDL)[5] will not provide a sufficient estimate for the required signal dimension.

The simulation of several scenarios with different numbers of multipath components P, delayspread and angular spread for an antenna array with M = 8 elements has shown that usingmore than 3 signal branches does not provide any significant SNR improvement, as indicated inFigure 4. Basic results from these simulations are:

• for scenarios with small angular spread the signal dimension after OT can bereduced to 3 or less with negligible loss of SNR gain independent of the respective delayspread,

• for scenarios with limited excess delay the signal dimension after OT can bereduced to 3 or less with negligible loss of SNR gain independent of the respective angularspread,

153

• for arbitrary scenarios with unrestricted angular spread and large excess delaymore than 82% of the available signal energy after OT is contained in the first 3 componentsresulting in a loss of SNR gain of less than 0.9 dB.

The results from simulations using arbitrary parameter combinations have been verified usingtap delays and average relative path powers from the GSM and ITU channel models [6, 7] asparameters. The delays have been normalized to the GSM symbol period forthe GSM models and to a chip period of for the ITU models. Thechannel parameters delay spread maximum excess delay and their normalized values aresummarized in Table 1. For each tap of the different channel models, a path angle has beenselected which is uniformly distributed within with varying from to180°. The resulting eigenvalues from 100 trials have been averaged for each value of Table 1shows the corresponding loss of SNR gain for the smallest sum of the 3 largest eigenvalues, which isconsidered the worst case (*). The last column of this table shows the required receiver dimension,for a maximum loss of 0.25 dB SNR gain. The results obtained for the GSM HT and TU scenarioswith a restricted angular spread are even more promising as apparent from Table 1.

These results suggest, that it makes sense to remove the partial coherence properties of multi-path channels through spatial dimension reduction to simplify space-time receiver structures andto improve channel estimation techniques.

5 Reduced Dimension Space-Time Receiver

As an example for the potential benefits of the dimension reduction approach we will discuss channelestimation and space-time MLSE using antenna arrays. An equalizer structure using an space-timeViterbi algorithm (VA) combined with reduced rank channel estimation (RRCE) was presented in

154

[8]. Starting from the received signal vector in (12) we can define a discrete time signal vectorsampled at symbol rate with unknown timing error as

assuming that the channel remains constant during one time frame. Using further the assumptions,that is i.i.d (15) and zero mean and that the noise is spatially and temporally white anduncorrelated with and that the CIR is of finite duration we can reformulate (27) with

for notational convenience as

Here, is the channel matrix andis the input symbol sequence which affects x(l). Considering a block of data of N symbol intervalsthe received signal in matrix notation can be written as

with

As discussed in [8], under the assumption of white noise samples the VA can be implemented as adistance measure with the corresponding branch metric being

The performance of the VA depends on the estimate of H. If S defined in (32) contains knowntraining sequence symbols, the least-squares estimate of H is given by

with as the Moore-Penrose pseudo-inverse of S. In [8] it was proposed to replace this estimateby a subspace based approach. Using the decomposition of the spatial correlation matrix (26),the M-dimensional space is divided into a -dimensional signal subspace and a noise subspace bysplitting the eigenvector matrix as The signal

155

subspace dimension is determined using the AIC or MDL criterion. Applying this decompositiona subspace estimate of the channel matrix was proposed as

with

The main advantage of this estimate is the usage of additional knowledge about the signal subspacein which is estimated from the entire frame and does not rely on the training sequence alone.However, analyzing (35) and (36) of the RRCE receiver suggests a reduced dimension (RD) receiverstructure with the same performance at lower receiver complexity. The received signals are firsttransformed to reduced-dimensional data using the signal subspace eigenvector matrix

where the dimension is now truncated to 3 instead of using the AIC or MDL criteria. All theremaining processing can then be performed with reduced complexity. The channel estimationbased on the reduced data set is given by

which provides the same estimates as obtained with (36). The resulting receiver structure is shownin Figure 5(a). In Figure 5(b) the performance of a full space-time receiver (dashed line) and the RDreceiver (solid line) is compared. The usage of signal subspace information results in a performanceadvantage of the RD receiver depending on the training sequence length. The performance is

156

identical to the RRCE receiver in [8], but at significantly lower receiver complexity. If interferenceis considered the RD receiver may even outperform the RRCE receiver since truncation of thenon-signal components will also reduce interference. However, this depends on the interferencecharacteristics and further research is needed.

6 Conclusion

The spatial correlation properties of the received signal at an antenna array using a signal modelwhich includes the transmit filter have been investigated. The rank of the spatial correlation matrixdoesn’t provide a sufficient measure for determining the possible dimension reduction in scenarioswith partially coherent multipath signals. Here, analyzing the eigenvalue strength associated withthe signal energy distribution after orthogonal transformation is more appropriate. The mainadvantages of the dimension reduction are the reduced spatial receiver complexity and the morerobust channel estimation. However, further research needs to be carried out in order to replacethe SVD by numerically less complex algorithms to derive the reduced dimension transformationmatrix.

References

[1] Gene H. Golub and Charles F. van Loan, Matrix Computations, The Johns Hopkins UniversityPress, third edition, 1996.

[2] Arogyaswami Paulraj and Constantinos B. Papadias, “Space-time processing for wireless com-munications,” IEEE Signal Processing Magazine, vol. 14, no. 6, pp. 49–83, Nov. 1997.

[3] Heinrich Meyr, Marc Moeneclaey, and Stefan A. Fechtel, Digital Communication Receivers:Synchronization, Channel Estimation, And Signal Processing, John Wiley & Sons, Inc., 1998.

[4] Steven M. Kay, Fundamentals of Statistical Signal Processing, Volume II: Detection Theory,Prentice Hall, 1993.

[5] Mati Wax and Thomas Kailath, “Detection of signals by information theoretic criteria,” IEEETransactions on Acoustics, Speech and Signal Processing, vol. 33, no. 2, pp. 387–392, Apr. 1985.

[6] ETSI EN 300 910, “Digital cellular telecommunications system (phase 2+); Radio transmissionand reception,” Tech. Rep. GSM 05.05 version 7.1.0 Release 1998, ETSI, 1999.

[7] Radio Communications Study Group, “Guidelines for evaluation of radio transmission tech-nologies for IMT-2000/FPLMTS,” Tech. Rep. 8/29-E, ITU, 1996.

[8] Ayman F. Naguib, Babak Khalaj, Arogyaswami Paulraj, and Thomas Kailath, “Adaptive chan-nel equalization for TDMA digital cellular communications using antenna arrays,” in Proceed-ings of the 1994 IEEE International Conference on Acoustics, Speech, and Signal Processing,1994, vol. IV, pp. 101–104.

Wireless personal communications system planningusing combinatorial optimisation

Joseph K. L. Wong, Michael J. Neve and Kevin W. SowerbyDepartment of Electrical & Electronic Engineering, The University of Auckland,

Private Bag 92019, Auckland, NEW ZEALANDJoseph.wong@auckland.ac.nz

Summary

This paper reports an investigation into the use of combinatorial optimisation techniques in the design ofwireless communication systems. The formulation of an optimisation strategy appropriate to wirelessindoor communication system design is identified. A simple single-floor test problem is proposed andinvestigated to illustrate the feasibility of the approach. Results show that the optimisation strategy iscapable of delivering meaningful solutions with an appropriate balance between system performance andcost. A number of issues (such as the formulation of cost function and optimisation strategy) areidentified which will require additional research in order to realise a planning tool useable by systemplanners.

1. Introduction

The emerging standards in personal wireless communications require new systems to have large capacityand high efficiency [1-3]. Generally, system deployment is specified by numerous design parameters(such as the number of base stations to be used, the base station locations and the transmission powers),each of which can be regarded as a different dimension in a multi-dimensional design problem. Theinterrelationships between these parameters must be considered if a successful deployment is to berealised. As systems become more complex and operational margins become tighter, ‘traditional’ designtechniques that rely mainly on an engineer’s experience are likely to become less effective. This isprimarily due to the sequential nature of the traditional design process, in which design parameters areusually considered in isolation. Consequently, there is a need for alternative planning strategies thatexplicitly consider the interrelationships between the parameters.

This paper proposes an alternative strategy in which the design problem is viewed as a multi-dimensionalcombinatorial optimisation problem. This approach inherently accounts for the interrelationships betweendifferent dimensions and may be better suited to wireless systems design than the ‘traditional’

approaches. The optimisation process includes three primary components: 1) a set of decision variables;2) a cost function; and 3) an optimisation algorithm.

The decision variables are system parameters such as the number of base stations to be used, the basestation locations and the transmission powers. The specific values assigned to these parameters constitutethe ‘output’ of the optimisation process.

The cost function is a function of the decision variables, the output of which is a number representing thecost (or penally) of a particular wireless system configuration.

Finally, the optimisation algorithm is an iterative procedure that searches for a solution by repeatedlyevaluating the cost function for different combinations of decision variable values.

In this paper, the key decision variables and cost functions associated with the deployment of a wirelesscommunication system are investigated. A Guided Simulated Annealing based algorithm [8,9], is used todetermine the optimum deployment of base stations in a simple DS-CDMA system. The results indicatethe feasibility of the proposed approach to wireless system design.

2. Choosing the Decision Variables

In wireless system design, the decision variables are parameters that have an effect on the performance ofa system – both in terms of the quality of communication and the operational cost. Sets of decisionvariables can be used to specify both the hardware configuration and the selection of operationalstrategies (such as handoff and discontinuous transmission). The optimisation of a complete wirelesscommunication system requires a very large number of decision variables. Computational practicalitiesrequire that the problem be resolved into more manageable portions. Accordingly, the optimisationprocess is tackled in stages. Groups of closely interrelated decision variables should be consideredtogether. Largely independent decision variables can be dealt with separately. For example, only thedecision variables influencing the physical configuration of the wireless system hardware need to beconsidered initially. The decision variables relating to operational processes, such as handoff, can be dealtwith subsequently. In this paper, only the base station configuration stage of the complete systemoptimisation process is considered.

158

159

3. Selection a Cost Function

The cost function is an important component of the optimisation strategy as it quantifies systemperformance. In general, the cost function accepts an n-dimensional trial solution (generated by theoptimisation algorithm) and returns a ‘cost’ as an indication of the relative ‘goodness’ of the trial.Although the choice of a cost function is likely to vary from one system type to another (each of whichmay require different performance measures), a generic cost function for optimising the design of awireless communication system will likely contain both technical and commercial components.

3.1 Cost Function - Technical Component

The momentary signal-to-interference ratio (SIR) on the forward link is generally regarded as anindication of the quality of a wireless CDMA system [4, 5]. The outage probability is derived from theSIR and is one of the most important parameters as it is indicative of the actual system capacity [4, 5]. Inaddition to outage probability, the minisum function and the minimax function of the interference-to-signal ratio can also be used as measures of performance since they correspond to the average andminimum value of the SIR of the population, respectively.

Outage probability is defined as the fraction of users in a service region with an SIR below a certainthreshold. As the capacity of a CDMA system is limited by interference, this threshold corresponds to theminimum SIR level that provides acceptable quality for communications. In general, for a population of nusers, the outage probability can be expressed as

The term minisum generally means minimisation of the sum of the cost [6]. Optimisation using theminisum function tends to lower the overall cost of the system. For the current problem, it can beinterpreted as maximising the average SIR for the whole system. However, as the cost is examinedcollectively, a small portion of the population may suffer from extremely low SIR. The minisum functionfor SIR can be formulated as

160

In contrast to minisum, the minimax function optimises the system in a worst case scenario [6]. Iteffectively improves the worst of the population but at the expense of the average service quality. Such astrategy may fail to provide an acceptable SIR level for the entire population even when there aresufficient resources available. The minimax function is given by

During the optimisation process, as the SIR level changes, the amount of improvement achieved isdependent on which of these measures is used to quantify the change. As an illustration. Fig. 1 shows thechanges for sets of five users’ SIR levels. The dotted line in Fig. 1 indicates an outage threshold of 9dB.The improvement achieved in each of the three performance measures is listed in Table I.

161

3.2 Cost Function - Commercial Component

Besides the technical component, there are other costs associated with installation, maintenance andoperation of the physical equipment. In this paper, the fixed installation cost of the base station isassumed to vary according to the number of base stations required, namely

In addition to the fixed cost associated with each base station, there is a marginal operational cost. In thispaper, it is assumed that this operational cost is directly proportional to the transmission power, that is

Consequently, a first-order cost function for optimising the design of a wireless communication systemcan be proposed as

where are used to weight the various cost components.1

1 This ‘first-order’ cost function is in its most primitive form. The identification of additional cost components, theselection of component weights and the formulation of alternative cost functions are areas requiring further research.

162

4. Choosing an Optimisation Algorithm

In general, an optimisation algorithm is an iterative method that searches for the minimum in the costfunction within the solution space [7]. In this paper, the optimisation algorithm chosen consists of twosearch methods, namely, a brute force search and guided simulated annealing [8, 9].

These two search methods are implemented in two stages. The first (outer) stage of the optimisationprocess searches for the optimal number of base stations while the second (inner) stage aims to determinethe optimal combination of this number of base stations, in terms of their locations and transmitter powerlevels. Because the problem of finding the optimal number of base stations is one-dimensional, and withan appropriate selection of limits, the search space is therefore likely to be small, ‘brute force’ searchingis used in preference to a global search technique. ID contrast, the combinatorial optimisation problem ofdetermining base station locations and transmission powers is much more complex. The correspondingsearch space is likely to contain multiple local minima and therefore a global search algorithm (the guidedsimulated annealing algorithm in this case) should be used. The algorithm for the proposed optimisationalgorithm can be expressed in pseudocode as

1. FOR n = 1 t o maximum number of base stations allowed

(a) WHILE Xmin > minimum defined

i. FOR m = 1 to maximum number of trialsX(m) = trial generated by the Guided Simulated Annealing algorithm

for n base stationsFX(m) = evaluate cost function(X(m))

ii. Xmin = min (FX)

5. A Test Problem

A simple test problem has been formulated to examine the feasibility of the proposed algorithm byconsidering the deployment of base stations into a room with 25 regularly spaced users. The targetsystem employs DS-CDMA modulation with a spreading gain of 128. All the antennas are omni-directional and free space radio signal propagation is assumed [10, p. 71]. In this simple problem, theeffect of shadowing and multi-path fading are ignored. A user experiences outage when its SIR (afterapplying the processing gain) falls below 9dB.

The optimisation process searches for the most favourable base station configuration. In this problem, thebase station configuration is quantified by the number of base stations to be used, the transmission power

163

of each base station (relative to the weakest transmitter power) and their respective coordinates on the xy-plane, as listed in Table 2 (a). The upper and lower limits are chosen to constrain the search within the‘likely’ solution space and to avoid unrealistic solutions. The cost function used in this example is givenby equation (7), with weights as listed in Table 2 (b) (Note that for this ‘first-order’ problem, the minisumand minimax components are not considered.)

The weights for Case 1 are chosen so that all cost components of interest have roughly the same influenceon the search. This case acts as a benchmark for comparison with the two other cases considered later inthis paper. The inclusion of Case 2 and Case 3 permits the examination of the effect of varying theweighting of a single cost component (in this case which represents the base station fixed cost).

6. Results and Discussion

6.1 Base station configuration

Because base station transmissions influence the reception in the cells (service areas) of other basestations, the identification of a globally ‘optimal’ configuration is a non-trivial exercise – even for thissimple scenario. Traditionally, wireless system design has focused on achieving adequate coverage.With such an objective, the base stations are likely to be placed in a regular and/or symmetric topology tosimplify the design process. In relation to the current scenario, the periphery of the room would be acommon choice for locating base station(s) [11]. Two possible ‘traditional’ configurations, with two andthree base stations, are shown in Fig. 2 (a) and 2 (b), respectively. The relative performance of theseconfigurations is summarised in Table 4 along with the performance of the configurations generated by

164

the optimisation algorithm. (Note that the optimised transmission power levels are all within 0.8 dB ofeach other. In calculating the cost associated with the transmission power, the lowest base stationtransmission power has been arbitrarily assumed to be 1mW.)

Fig. 3 presents the configurations of two and three base station systems determined using the optimisationalgorithm with the relative transmission power summarised in Table 3. (The dotted lines indicate the

nominal cell boundaries.) The results presented in Table 4 show that the regular and symmetrictopologies in Fig. 2 result in higher outage probabilities than those from optimised base stationdeployments. An optimised deployment leads to better performance because it minimises the areas inwhich users suffer high interference levels from other base stations.

6.2 Defining the number of base stations

The result set marked with ‘+’ in Fig. 4 (a) illustrates the cost performance of the solution generated bythe optimisation algorithm for a fixed base station cost weighting of 0.02. (The other two sets of resultsin Fig. 4 (a) are for fixed base station costs of 0.04 and 0.01. These results will be discussed in section6.3.) The outage probability component of the total cost tends to decrease as more base stations are added(see Fig. 4 (b)), so the outage probability component is at a maximum when there is only one base station.(The cost value of 0.2 corresponds to an outage probability of 100%.) The inclusion of a second basestation decreases the outage probability to 24% of the value for a single base station. Further reduction inoutage probability is achieved by adding more base stations and the outage probability cost componenteventually reaches zero.

In contrast, both the base station fixed cost component and the transmission power cost componentincrease linearly with the number of base stations deployed. As a result, these costs drive up the total costof the system when the number of base stations is increased beyond three. In essence, the improvement inreception quality (decrease in outage probability) is achieved at a financial cost (more base stations and

165

166

higher transmission power). The optimisation algorithm arrives at a good compromise betweenperformance and operational expense by selecting the three base station configuration as the solution. Ofcourse, this selection has been influenced by the pre-determined relative importance of the various costs,as represented by the cost weights used.

6.3 Influence of weight variations

To assess the influence of cost component weights on the final cost, an experiment has been performed inwhich the fixed base station cost weight is varied. The results are shown in Fig. 4 (a), 4 (c) and 4 (d).The data set marked with ‘X’ and ‘O’ in Fig. 4 (a) represent the cases with doubling and halving theweights, respectively.

167

When the weighting for the base station cost is doubled, the ‘optimal’ base station configuration contains

only two base stations, rather than the three base stations for the original scenario (see Fig. 4 (a)). Thedata set shows a much steeper climb in cost as the number of base stations is increased – indicating that itis undesirable to use more base stations. The additional outage probability cost associated with using twobase stations rather than three, is more than offset by the (financial) cost advantage.

In contrast, when the weighting for the base station cost is halved, the ‘optimal’ base station configurationremains at three base stations. Note that the corresponding cost curve in Fig. 4(a) is flatter, indicating thatthere is less emphasis on reducing the number of base stations. In essence, the emphasis has beentransferred to other cost components (such as outage probability and transmission power).

7. Ongoing Development

The optimisation approach presented in this paper appears promising. However, a number of issues mustbe addressed before a practical planning tool can be realised. The formulation of the cost function and theselection of the optimisation algorithm are of primary concern.

The cost function used in this paper is a simple first-order model that was. designed to demonstrate thefeasibility of the optimisation approach. In reality, a practical cost function must contain realistic costcomponents (in terms of performance measures such as customer satisfaction level) and cost weightsappropriate to the planner’s design objectives. Attention is currently being directed at formulating aguideline for the selection and prioritisation of performance measures.

Although the current optimisation algorithm has demonstrated an ability to deliver sensible solutions, itmust be realised that the problem has been significantly simplified. If the optimisation algorithm is to bepractically useful, it must be augmented to be able to deal with considerably more complex scenarios,such as the three-dimensional indoor environment.

8. Conclusions

This paper has addressed the problem of optimising the design of a wireless communication system. Acombinatorial optimisation procedure is used to specify the optimal system configuration such as thenumber of base stations to be used, their locations and their relative transmission power levels. A simple

168

cost function accounts for the outage probability, operational and installation costs associated with anyparticular system configuration. The technique has been demonstrated to be capable of deliveringmeaningful solutions for a simplistic two-dimensional environment.

Attention is now being focused on formulating a guideline for specifying cost functions and selecting anoptimisation algorithm. Of particular interest is the applicability and sensitivity of the optimisationtechnique to a range of complex three-dimensional indoor environments.

References

[1] D. Grillo, S. T. S. Chia, and E. N. Rouelte, “Special Issue on the European Path Towards MobileSystems,” IEEE Personal Comms. Mag., vol. 2, No.l, Feb 1995.

[2] M. H. Callendar, “Special Issue on IMT2000: Standards Efforts of the ITU,” IEEE PersonalComms. Mag., vol. 4, No. 4, Aug 1997.

[3] E. Buracchini, R. D. Gaudenzi, G. Gallinaro, H. H. Lee, and C. G. Kang, “SatelliteUMTS/IMT2000 W-CDMA Air Interfaces,” IEEE Communications Magazine, vol. 39, No. 9,pp. 116-126. Sept 1999.

[4] R. Prasad, CDMA for wireless personal communications, Boston: Artech House, Chapter 5, 1996.[5] S. G. Glisic, Spread spectrum CDMA systems for wireless communications, Boston: Artech

House, Chapter 6, 7, 1997.[6] H. D. Sherali, C. M. Pendyala, and T. S. Rappaport, “Optimal Location of Transmitters for

Micro-Cellular Radio Communication System Design,” IEEE Journal on Selected Areas inCommunications, vol. 14, No. 4, pp. 662-673, 1996.

[7] B. D. Bunday, Basic Optimisation Methods, Edward Arnold, Chapter 1, 1984.[8] P. P. C. Yip and Y. H. Pao, “Guided evolutionary simulated annealing approach to the quadratic

assignment problem,” IEEE Transactions on Systems, Man and Cybernetics, vol. 24, No. 9,pp. 1383-1387, Sept 1994.

[9] P. P. C. Yip and Y. H. Pao, “Combinatorial optimization with use of guided evolutionarysimulated annealing,” IEEE Transactions on Neural Networks, vol. 6, No. 2, pp. 290-295, Mar1995.

[10] T. S. Rappaport, Wireless communications - Principles and Practice, Prentice-Hall Inc, 1996.[11] M. J. Neve and K. W. Sowerby, “Optimising The Performance of Indoor Wireless

Communication Systems,” IEEE Vehicular Technology Conference, vol. 2, pp. 968-972, May1999.

Frequency Planning and Adjacent Channel Interferencein a DSSS Wireless Local Area Network (WLAN)

D. Leskaroski, W. B. MikhaelElectrical and Computer Engineering Department, University of Central Florida, Orlando, FL 32816

dl375l6@pegasus.cc.ucf.eduResearch Sponsored by: Nortel Networks and Harris/Intersil Corporation

ABSTRACT

The new communications standard for wireless local networks is the IEEE 802.11 standard.802.11b specifies that we use the ISM band at 2.4GHz. The ISM band at this frequency is 83MHz wide.For North America, under FCC regulations, the 83MHz of bandwidth is divided into 11 channels. Everychannel has a frequency bandwidth of 22 MHz [1]. In North America, channels 1 through 11 are usedand only Ch1, Ch6, and Ch11 physically do not have any overlapping frequency ranges. Therefore, itfollows that frequency planning has been done using only these three channels. This greatly reduces thecapacity and utilization of a given Wireless LAN.

This paper will deal with and answer the question of how far attenuated the signal has to be sothat it will not interfere with another adjacent or overlapping channel. The goal of the research was togenerate a theoretical model of the 802.11 DS channelization (frequency re-use and interference profile)given that the 11 channels are arranged in a staggered overlapped fashion.

The result of the research shows how a set of overlapping channels (1 through 6) can bepositioned/attenuated in order to increase user capacity yet remain with minimal channel interference. Inaddition, the paper will provide sufficient evidence that will show that frequency planning can be doneusing four channels instead of three channels.

INTRODUCTION

Wireless products provide a mobile solution to various network configurations. A WLAN is anon-premise data communication system that reduces the need for wired connections and makes newapplications possible, thereby adding new flexibility to networking. Mobile WLAN users can accessinformation and network resources as they attend meetings, collaborate with other users, or move to othercampus locations.

With the availability of cost-effective, standards-based products, use of wireless networking israpidly expending from the factory floor, warehouse and retail stores, to hospital wards, universitycampuses and now too corporate offices. The various environments call for different wireless networkingrequirements in terms of coverage area, user density, traffic patterns, and interference.

170

When speaking about interference regarding wireless communications in many instances it isunderstood to be interference such as multipath, reflection, and scattering. This paper on the other handwill deal with the interference that occurs between two adjacent or overlapping signals.

FREQUENCY CHANNEL ASSIGNMENT

The IEEE802 LAN committee has created a wireless data communication standard that allocatesa given frequency range in the 2.4GHz ISM band for different parts of the world. As shown in table Ione can see the different frequency ranges assigned for North America, Europe, and Japan.

The total bandwidth allocated is of significant importance because it defines the number ofchannels that can be utilized at each geographical area. Every region has its specified authority thatmonitors and regulates the use and division of this bandwidth. For the United States, this acting authorityis the Federal Communications Commission (FCC). The FCC has divided the 2.400-2.4835 GHz range,or the 83MHz band, into 11 channels [1]. The table below will show the frequency channels that aredefined for North America, as well as the channel allocations for Europe and Japan.

171

Table 2 above shows only the center frequency of each channel, and as shown there is 5MHzspacing between every center frequency. A power spectral density function of one of these channels(Channel 1) is shown below.

Every channel has bandwidth of 22MHz, so for example channel 1 will spread from 2401MHz to2423Mhz. As imagined there will be an overlap between the channels. In fact, Channel 1 overlaps all theway to channel 5. The figure below shows what is actually happening between the channels.

172

So, as shown Channel 6 is the first channel that Channel 1 does not overlap with. This is wherefrequency planning using non-overlapping channels comes in place. Channels 1, 6, and 11 are used forfrequency planning since that is the only combination of three channels that do not overlap each other.As shown in the figure above there needs to be a 5 channel separation so that the channels will notinterfere with each other.

PROBLEM STATEMENT

Channel 1 will stretch from 2401-2423MHz. This means channel 1 will interfere with all thechannels up to Channel 6. As it follows, Channel 1 will have the biggest interference with Channel 2, notcounting Channel 1 itself where interference will be over the entire 22MHZ of bandwidth. One of theobjectives of this research is to calculate the interference between two adjacent channels given that theyare one on top of the other.

The first step is to find the overlapping area between two interfering channels. For example, wecan calculate a SNR between the curve area of Channel 1 and the overlap area of any other given channel(CH2…CH5). This way we can find the interference of a given channel with respect to Channel 1. Thecalculation will give us a result that looks similar to the following:

Note: These numbers are not actual calculations. They are just provided to better explain theapproach.

None of these overlaps (interference) is acceptable. We will have to set a threshold of overlap(interference) that will be acceptable (for example, 0.05). Now we can "play" with the power spectraldensity function of Channel 1 in order to get below the desired threshold level. The power of Channel 1will be reduced (attenuated) until desired results are achieved.

173

The final result will be a table as follows:

CHl: -xdb

CH2: -ydb

CH3: -zdb

CH4: -adb

CHS: -bdb

CH6: 0db

x, y, z, a, b will be values in db that will show how much given AP needs to be attenuated so it would not

interfere with another AP. (Note: x, y, z, a, b are not yet determined)

There are two different ways to decrease the overlap between the interfering channels.

1. Reduce the output power of the AP itself, or

2. Increase the distance between the two interfering APs.

The above results (x, y, z, a, b) will give you the values of how much attenuation is needed,

however reducing the power of a given AP may not be desired. Another way of reducing the overlap is to

calculate the distance between two adjacent APs so that there is no interference. The equations and

comments given below will help us create a valid theoretical model.

Free Space Path Loss(FSPL) in indoor environment.

FSPL=(47Td/A)2 (Equation 1)

Where d is the distance in meters between the transmitter and receiver, and λ (lambda) is the wavelength

in meters.

X=c/f (Equation 2)

Free Space Loss = lO'log(FSPL) (Equation 3)

Line of Site Path Loss (PL)

PL=FSL,.f + n 1 * 10*log(dB) (Equation 4)

Where FSL f̂ is the free space loss in dB determined in the far field of the antenna. d,, is the

distance between the transmitter and receiver [4]. The symbol n1 is an attenuation factor which depends

on the attenuation of the environment. For line of site application the n1 factor has been determined to be

close to 1.99. [5]

EXPERIMENTAL DATA AND MEASUREMENTS

The following section will deal with and discuss in detail the affects of the interference that

occurs between two overlapping channels. Even more important, this section will show how far

attenuated the signal has to be from a given adjacent channel so that the signals will not interfere. Over a

13 weeks period, I have completed numerous measurements and created a model that will show how far

174

attenuated overlapping channels need to be so they will not present interference to each other. Theequipment used for the experiment is listed below:

4 Lap Top Computers 2 10baseT Network Access Hubs 2 Nortel AirSurfer Pro Wireless Access Points 4 BayStack 660 Wireless PC Cards

2 10baseT PCMCIA Network Card Adapters 2 Power BackUPS 2 Mobile Carts 1 Measuring Tape Harris LAN Eval Software (used to measure the throughput)

The experiment was conducted in an outdoor environment, in a parking lot of the Engineeringbuilding at the University of Central Florida. The measurements were performed between 12:00am to5:00am. This time was chosen in order to reduce the unwanted interference of people, cars, and otherobstructions. The parking lot is approximately 600 feet by 600 feet making it the ideal place to performthis experiment. A picture of the experiment is show in Figure 3.

175

As shown in the picture, there are two different sets of networks. Set A consisting of TX-A(Transmitter), RX-A (Receiver), Hub A, and AP A constantly set on Channel 1. Set B consists of TX-B(Transmitter), RX-B (Receiver), Hub B, and AP B that will be changing Channels from 1 to 6. AP A wasassigned a SSID called UCF1 and AP B was assigned a SSID called UCF2. This way mobile unit B2(TX-B) can only connect to AP B, and mobile unit A2 (TX-A) can only connect to AP A. For thisexperiment the distance between TX-A and RX-A, as well as the distance between TX-B and RX-B, wasset to be 30 feet. The entire set A was at all time stationary, only set B was being moved away from setA. At every point the following things were recorded:

1. Distance between Set A and Set B (ft)2. Throughput of Set A (Kbits/sec)3. Throughput of Set B (Kbits/sec)

Table 3 and Table 4 show the throughput measured. (NT stands for Not Tested)

176

As shown in the tables, the throughput is presented in percentages. Before the measurementsbegan I measured the throughput of set A without any interference (458 Kbits/sec), and the throughput ofset B without any interference (461 Kbits/sec). The percentage shown in the tables shows a normalizedthroughput performance of each system. As I mention earlier, set A was always set to CH1 and set B wasthe one that was being changed to all six channels. Each table shows what happens to both sets when setA is CH1 and set B is CH1, CH2, CH3, CH4, CH5, CH6. It is important to show both performances,because at times the interference affects only one of the systems. When the two systems are close to eachother (they see each other), one of the systems takes over and starts transmitting and the other set iswaiting for a clear channel. The CCA (Clear Channel Assessment) will not let transmission occur when itcannot find a clear channel to transmit.

The most important parts of the throughput tables 3 and 4 is where both systems are performingabove or close to 80% of their "maximum" throughput. I chose 80% as a threshold, but it is up to the userto choose the performance of the two sets. Table 5 below is little more user friendly and it give a so-called overall performance of the entire experiment with respect of distance. The numbers below arecalculated by finding the average of the tow corresponding numbers from Table 3 and Table 4.

177

The second column of the table above represents the attenuation in dB corresponding to thedistance given in column one. The formula given below was used to calculate the attentions:

Where: PL is the path loss (attenuation) in dB. D is the distance between set A and set B in ft, given in column one. (D is divided by

3.3 to be converted to meters) N1 is the attenuation factor. (Since the experiment was done outdoors and line of site,

a value of 1.99 was used for N1.Using all the data shown above we can derive a more general attenuation table for channel 1

through channel 6. Assuming system A set on channel 1 is at location "0", another network (system B)can be positioned with a given channel if attenuated at:

Now by knowing the actual attenuation values in dB we can calculate how far we can positiongiven APs for different values of N1. Using Equation 5, we have:

where the value of is in meters. Using equation 6 and using a value of 2.99 for N1 the valuesof the last column in Table 5 were derived. Using equation 6, the user can enter a value for N1 thatcorresponds to the environment where the APs will be placed. The equation will show how far apart theyneed to be so they will not interfere with each other.

In order to give a better visual representation of how the performance of system A and B behavedwith respect of distance, I will present the performance graphs at the end of this paper. In reality, thecenterpiece of the entire research are the values given in Table 5, but using the graphs below it is easier tosee how the performance varies with respect of the distance between the transmitter and receiver. Thedistance in the graphs is given in feet.

178

179

180

The graphs of Channel 1 vs. Channel 5 and Channel 1 vs. Channel 6 were not given in this paperbecause the performance in this two cases was close to 100 percent throughput the experiment. In otherwords, the graphs resembled a straight horizontal line at 100 percent for any given distance.

CONCLUSION

At the beginning of the paper it was said that only Channel 1, Channel 6, and Channel 11physically did not overlap each other, so these three channels were the preferred when it came tofrequency planning. From observing all the given measurements in this paper one can see that otherchannels such as Channel 5 and even Channel 4 performed with close to 100 percent throughput veryclose to a system that was set on Channel 1. This paper has provided evidence that when it comes tofrequency planning we can use the combination of four (Ch1, Ch4, Ch7, and Ch11) instead of three (Ch1,Ch6, and Ch11) different channels. Even further, this paper has given information on how it would bepossible to use the entire 11 channels arranged in some type of an overlapping fashion.

For this experiment all the testing has been done for channels 1 through 6. Additionalmeasurements can be done for the rest of the channels; however, at this point it is assumed that this modelis valid for any set of six consecutive channels.

LIST OF REFERENCES

1. IEEE 802.11 Standard for Wireless LAN, Wireless LAN Medium Access Control (MAC) andPhysical Layer (PHY) Specifications, Jan. 1997.

2. B. Garon, "PRISM 1KIT-EVAL Wireless LAN Evaluation Kit User’s Guide", Application NoteAN9790.1, Harris Semiconductors, November 1998.

3. "Frequency Hopping and Direct Sequence Spread Spectrum Radio Technologies", Nortel Networks,September 1998.

4. J.C. Stein, "Indoor Radio WLAN Performance -- Part II: Range Performance in a Dense OfficeEnvironment", Harris Semiconductors, February 1998.

5. T. S. Rappaport, “Wireless Communications: Principles and Practice,” IEEE Press, Prentice-Hall.New Jersey, 1996.

Modeling and Simulation of Wireless PacketErasure Channels

Günther Liebl, Thomas Stockhammer1, and ftank Burkert2

Abstract

We will present a new model for wireless packet erasure channels, which can be used for both theoreticalanalysis and real-time simulation of network protocol performance. The correlation between successive packetlosses will be described by a higher-order Markov process, which can be transformed into an equivalent first-order Markov chain, such that all major performance measures can be derived from the model. Since ourapproach belongs to the class of generative discrete channel models, it can also be used as a stochastic sourceto reproduce an infinite sequence of erasure indicator values with desired stochastic properties. Thus, it canbe easily implemented into network simulation environments in terms of a lossy link element, where bothshort-term and long-term fading on the mobile radio link are taken into account separately. Finally, we willshow some modeling results for a GSM GPRS system, which is a packet-oriented transmission service to belaunched in Europe in spring 2000. In addition, we will demonstrate the effects of different packet sizes andchannel coding rates on end-to-end throughput.

I. INTRODUCTION

N future cellular networks, packet-oriented transmission services will play a major role, sincethey provide an efficient means to adapt to the highly varying data rate of different multimedia

applications. One possible scenario is comprised of a totally heterogeneous network environment,where the current Internet serves as the backbone, and the wireless links provide access pointsfor mobile users. However, compared to the benign channel characteristics of wire-line broadbandnetworks, radio links suffer from severe distortions due to fading, noise, and interference. Hence,despite many advances in channel coding and equalization over the last years, the residual bit errorrate is still a crucial factor: if only one single bit is corrupt, the whole packet must be declaredlost. This results in a significantly higher packet erasure rate than on today’s Internet, whereloss is mainly only due to buffer overflow. Since all common network protocols like TCP or UDPhave been optimized with respect to link congestion, they are extremely sensitive to a significantincrease in packet loss. Hence, a stochastic model is needed to analyze the detrimental effectsof wireless links on end-to-end packet transmission in heterogeneous network environments. Inaddition to theoretical results, real-time system behavior has to be examined with respect to aparticular setting of timers and buffer sizes of the various network protocols. In contrast to already

1Institute for Communications Engineering, Munich University of Technology, Arciutr. 21, 80280 Munich, Germany,email:{liebl,tom}@OLNT.EI.TUM.DE

2Siemens AG – Corporate Technology ZT IK 2, 80730 Munich, Germany, email: Frank.Burkert@MchP.Siemens.DE

182

existing approaches, we have developed a new model which can be implemented into a networksimulation environment.We will start with a very general description of packet erasure channels in section II, where wewill also define all the necessary notation. Section III contains the various steps in developingour model using the idea of higher-order Markov processes and their equivalent first-order Markovchain. In section IV, we will explain why our model is well-suited for integration into existingnetwork simulation environments, and show one possible implementation strategy. After a shortintroduction to GSM GPRS (section V) we will present some performance results for a GSM GPRSsystem in section VI to prove the applicability of the proposed model.

II. GENERAL DESCRIPTION OF A PACKET ERASURE CHANNEL

In the following, our packet erasure channel of interest shall contain all the components ofan arbitrary mobile radio system, i.e. logical channel structure, channel coding and modulationschemes at the transmitter, demodulation and detection methods at the receiver, and the actualcharacteristics of the underlying physical transmission channel, as depicted in Fig. 1. The input tothe channel is assumed to be a sequence of packets of fixed length, depending on the segmentationand reassembly procedure. At the output, a delayed version of the input sequence is received, inwhich all lost packets are replaced by an erasure indicator.

The reception of successive packets can therefore be described by a binary discrete-time stochasticprocess, i.e. a family of binary random variables where is the denumerable set ofintegers, and i denotes the position of a packet in a transmission sequence. Hence, the sequence

represents a particular realization of the erasure process, as shown in Fig. 2.The binary random variable is often called an erasure indicator variable, and takes on a value

183

in {0,1} according to the following rule:

In general, the probability distribution of the random variable depends on an infinite numberof past events, due to the correlated fading process on typical mobile radio channels. However,in most practical cases,

memory length If we furthermore assume our packet erasure process to be both stationaryand ergodic, it can be completely described via its conditional probability mass function (PMF):

In the following, we will show how Eq. 2 can be rewritten in terms of the state transitions of aMarkov chain, thus enabling us to apply all the well-known results in this field to our problem ofmodeling packet erasure channels.

III. PROPOSED MODEL: THE MARKOV ERASURE CHANNEL (MEC)In this section, we will develop a Markov model for our correlated packet erasure process of

memory order which has an extended state space and a corresponding mapping function betweenerasure indicator variables and states. For sake of compactness, only the major ideas of the approachare mentioned. For a more detailed treatment of this topic, including all relevant proofs, the readeris referred to [3].

A. The Higher-order Markov Erasure Process

Let us define a higher-order stationary and ergodic Markov process with binary state spacewhich has a memory of order Thus, the probability of making a transition from

state to state also depends on previous states. There exists now a set ofconditional probabilities for describing such a transition. Due to the assumed stationarity

of our process, the same set holds for all possible sequence indices and is denoted by

only a finite number of previous erasure events, say

have to be taken into account. Then, we usually speak of a channel of

184

withGiven the following (deterministic) mapping between erasure indicator values and states

we can rewrite Eq. 3 in terms of erasure indicator variables as

Based on the above definition of a higher-order Markov erasure process, we will now try to findan equivalent first-order Markov chain, such that the computation of Eq. 2 only involves the twosuccessive states in the sequence.

B. The Equivalent First-order Markov Chain

We will start by defining a first-order stationary and ergodic Markov chain with an extended

state space The probability of making a transition from state to statehere only depends on these two states, and can be written as

If we furthermore apply the following modified (deterministic) mapping between erasure indicatorvalues and states

it can be proved [3], that the above combination results in the same stochastic properties as thepreviously defined higher-order Markov erasure process.Since the inverse relationship to Eq. 7 is given by

the conditional PMF of Eq. 2 can be represented in terms of a state transition probability:

C. Simplest Non-trivial Example: Markov Erasure Channel of Order 2

We will illustrate the idea behind our proposed model by giving a simple example. Let us considera Markov erasure channel with a total of four states. Fig. 3 shows a sample sequence of erasureindicator values and the respective state transition diagram of the equivalent first-order Markovchain. Suppose the first packet has been received correctly, whereas the second has been lost, i.ethe channel is currently in state 1. With probability it will switch into state 2 during the

185

next transmission, which results in the third packet being correctly received. If the channel thenswitches into state 0 with probability the fourth packet will also pass un-erased.Instead of the state transition diagram, we can also use the transition matrix of the underlyingfirst-order Markov chain to describe the stochastic properties of the Markov erasure channel:

As we can easily realize, there are only two entries in each row, corresponding to the two possibletransitions starting in each state. This is due to the binary nature of the erasure process and thedeterministic mapping between states and erasure indicator variables, and therefore holds for anymemory orderFrom the above example, two aspects of our proposed model can already be observed:• Given a sufficiently long sequence of measured erasure indicator values for a specific mobile radiolink, we are able to parameterize the corresponding Markov model of desired memory order, i.e.compute an approximation for the actual transition probabilities by using the inverse relationshipof Eq. 8.• Once we have set up all (possibly) nonzero entries in the transition matrix, we can reproduce aninfinite sequence of erasure indicator values by cycling through the state transition diagram andusing Eq. 7.

186

IV. A REAL-TIME NETWORK SIMULATION ENVIRONMENT FEATURING A LOSSY WIRELESS LINK

When implementing our real-time network simulation environment, we made use of an existingshareware tool developed by UC Berkeley, the so-called NS version 2.1b5 [1]. It is an object-orientedsimulation tool, for which various common protocol elements like agents, queues, or transmission

links already exist in a basic version, such that we do not have to worry about how to pass onpackets between entities, or perform synchronization between transmitter and receiver parts. Onemajor benefit of NS is the full support of almost any possible version of IP-based traffic, e.g.protocols like TCP and UDP.Since spring 1999, there also exists a real-time enhancement called NSE, which provides a simple

means to investigate interactive protocol behavior: Via regular network interface cards, externalIP-traffic from remote hosts can be fed into a virtual simulation scenario, where single packets canbe dropped or delayed, before they are passed back onto the fixed network. Thus, it is possibleto add a virtual wireless link at any point in a heterogeneous network, and analyze end-to-endperformance of common transmission protocols.

A. A new Lossy Link Element Based on the MEC Model

Fig. 4 shows a simple structure for an end-to-end link in the virtual network. The agents at bothsides are usually modified with respect to our desired system of interest, which will be describedlater on. The queue is usually a fixed element of every transmission link in NS, and is responsiblefor correctly spaced packet transmission according to the target bit rate and delay set for thelink. The NS specification already includes a general procedure for integrating lossy link behavior

into virtual networks. According to an underlying stochastic source, a binary random variable isproduced whenever a packet is transmitted over the link. If its value is “1”, the packet is marked aslost, and the receiving agent can exactly tell which packets have to be dropped. But this mechanismexactly corresponds to our previous definition of erasure indicator values in section II.As we have mentioned in section III, our proposed MEC model has the property of being able toreproduce a (possibly) infinite sequence of erasure indicator values, which has approximately thesame stochastic characteristics as the measured sequence on a wireless link. Hence, our approachbelongs to the class of generative channel models [2], based on which it is fairly easy to construct

187

a new stochastic source with desired properties.

B. Modeling of Different Fading Characteristics

An important new aspect of our simulation environment is the clear distinction between differentfading characteristics. We consider an interference-limited system, where the influence of short-termfading is included in the statistical properties of the physical channel. Hence, for each target carrier-to-interference-ratio (C/I), we can set up a corresponding stochastic description of the erasureprocess. To take into account the long-term variations in the received signal, we can dynamicallyswitch our model parameters with respect to measured C/I-profiles during a simulation run. Thus,the movement of a mobile terminal through a natural environment can be analyzed.

V. GSM GENERAL PACKET RADIO SERVICE (GPRS)Against the background of the growth of both Internet and cellular phone users there is an

evident need for an efficient wireless access to packet switched data networks. Current so-called2nd generation mobile communication systems, e.g. GSM, are not able to serve this purpose.They have been designed on base of a circuit switched radio transmission for narrow band speechcommunications. This results in two major drawbacks for data transmission. Firstly, the availablebandwidth per user is much too small to allow higher data rates and, secondly, data traffic in packetswitched networks, e.g. the Internet, is bursty by nature. Conveying bursty traffic over a circuitswitched bearer results in a highly inefficient utilization of the available radio channels and causesunreasonable high costs for the user. Therefore, the Special Mobile Group (SMG) within the ETSIstandardization body launched in GSM phase 2+ the development of an efficient cellular packetdata service. At the end of 1998, ETSI specified GPRS as a new bearer service for GSM networksto improve and simplify wireless access to packet data networks. It is build atop of the regularGSM protocol stack to facilitate a low complex and easy integration into already established GSMsystems.In the following, we will give a brief overview of the extended system architecture of a GSM GPRSsystem, where we will solely focus on the air interface, i.e. we will not consider session setup andtear down or the delivery and routing of packets between mobile stations and external packet datanetworks (e.g the Internet). Our descriptions are mainly based on [4][5].

A. The GPRS Protocol StackThe basic idea behind the development of the GPRS specification was to ensure the concurrent

existence of current voice transfer with future high-rate packet data transfer in common GSMnetworks. Thus, the GPRS protocol can be considered as a set of extensions of the existing GSMprotocol stack. One important extension is related to channel allocation. In GRPS, a mobilestation (MS) may use multiple time slots of the same TDMA frame. The channel allocation is very

188

fast and flexible. The base station subsystem (BSS) assigns the available resources to the mobilestations according to a so-called capacity-on-demand principle, i.e. multiple users in a cell sharea common physical channel. The allocation of the time slots may differ from TDMA frame toTDMA frame. This enables the system to allocate a channel only when either the mobile stationor the BBS needs to send data packets. Moreover, downlink and uplink channels are assignedindependently (asymmetric transmission). Hence, for bursty traffic this results in a very efficientusage of the valuable resource bandwidth.The GPRS protocol stack can be gathered bom Fig. 5. On top of the network layer, any packetdata protocol can be applied. Most usually this will be the IP Protocol.The Subnetwork Dependent Convergence Protocol (SNDCP) adapts the upper layer protocols tothe functionality of the underlying GRPS layers. It performs segmentation and reassembly of longuser data packets and provides means for header compression and data encryption on the mobilelink to ensure privacy of user communication.The data link layer encompasses three sublayers. Logical Link Control (LLC) is used to establisha logical link between MS and BBS and is based on LAPD, which is also part of the common GSMprotocol stack. It supports point-to-point as well as point-to-multipoint connections. Backwarderror protection is provided in form of a Go-back-N retransmission protocol. The Radio LinkControl (RLC) layer performs segmentation of the LLC packet data units (PDUs) into short blocksof fixed length according to one of the channel coding schemes described in subsection V-B. Ablock check sequence is appended (in dependence on the applied coding scheme this is either aFire code or a CRC code), which allows in combination with sequence numbering the detectionof erroneous or lost packets. In addition, RLC provides an optional Automatic Request (ARQ)protocol to achieve a reliable data transfer when needed. The Medium Access Control (MAC)sublayer performs multiplexing of user data and signaling information.

189

The physical layer consists of two sublayers. The Physical Link Layer (PLL) provides means offorward error correction (FEC) which will be described in more detail later. The Radio FrequencyLayer (RFL) equals the one specified for GSM.Fig. 5 also depicts the hierarchical frame structure and the segmentation/reassembly according tothe GRPS protocol stack as described above.

B. Channel Coding Schemes for GSM GPRS

The purpose of channel coding is to protect the transmitted data against errors. For the originalfull-rate GSM traffic channel for speech, the output bits of the speech encoder are expanded to atotal of 456 coded bits by use of a memory 4 rate 1/2 convolutional code. The encoded bits aresubsequently block-diagonally interleaved and mapped onto 8 successive radio bursts.In GPRS, four different coding schemes (CS) have been defined each delivering 456 bits, since thestructure of the underlying interleaver and radio bursts have not been changed compared to GSM.A procedure called link adaptation can be applied to dynamically switch between the CS afterevery RLC block. This allows to adapt the level of error protection to the channel characteristics.Tab. I details the parameters of the four CS, where data rate denotes the rate that is available perGSM time slot .The convolutional code is the same code as used in the original GSM system.That is, if CS-4 is applied and a single user can use all 8 time slots of a TDMA frame, then themaximum data rate is 171.2 kbit/s. It is important to note that the term payload in this contextdoes not denote the amount of user data that can be mapped onto one RLC block, but the sum ofthe RLC header and the RLC data field as shown in Fig. 6. Thus, for throughput calculations, wehave later to take into account the respective overhead.

VI. RESULTS

In order to verify our new modeling concept for packet erasure channels we applied it to theGSM GPRS system. We will consider the packet transmission at the RLC layer. A RLC PDUis declared as lost (erasure), when the block check sequence indicates an residual error in thereceived packet. Since there exits no analytical means of describing the packet erasure process at

190

this level, the parameterization of the MEC depends on empirical data that can be gathered eitherby measurements or simulations of the GPRS system. Thus, the parameterization involves threesteps. Firstly, we generate a sufficiently long sequence of erasure indicator values (such that theassumption of ergodicity is valid) by simulation of the radio link. Then, we choose the memoryorder for the Markov erasure process and set up the state transition diagram. Finally, by analysisof the generated erasure indicator sequence we approximate the transition probabilities of the MECaccording to section III. It is important to note that this approach requires a very long sequenceof erasure indicator values, since it is impossible to determine all transition probabilities for anarbitrary memory order from a finite set of empirical data. Furthermore, for very small transitionprobabilities, the statistical uncertainty is very high due to the few events that can actually befound in the simulated sample sequence. Note that increasing the memory order results only in abetter approximation of the model, if the data set is sufficiently long.For our simulation of the radio link, we have chosen the following scenario. The fading model for thechannel is based on COST TU03, i.e. we consider a typical urban environment with a pedestrian ata speed of 3 km/h. Frequency hopping after each TDMA frame is enabled and assumed to be idealand all eight time slots are occupied by a single GPRS user. Further, we assume an interferencelimited system, where the received carrier signal is only disturbed by one co-channel interferer in adistant cell. The state of the mobile radio channel can therefore be denoted by the actual carrier-to-interference ratio (C/I) at the receiver.We will verify the validity of our model by means of erasure and erasure free run distributions.An erasure run of length is a sequence of successive erasures whichis delimited by both and i.e. the event Our statistic of interest is theso-called erasure run distribution (ERD), which is the probability that an erasure run is longerthan packets. According to the same concept an erasure-free run of length is defined as asequence of successive non-erasures which is delimited by bothand i.e. the event The corresponding statistic of interest is the so-callederasure-free run distribution EFRD. which is the probability that an erasure-free run is longer than

packets. We started with a MEC of memory oder 2, but found out that it is by far not sufficientto approximate the GPRS packet erasure channel. By stepwise increasing the memory order of the

191

MEC we observed that we get a piecewise linear function (on a logarithmic scale) that successivelyallows the approximation of one more point of the empirical run distributions. As Fig. 7 shows, aMEC of order 9 yields a quite good approximation for both EFRD and ERD.Once the MEC model has been parameterized, it can easily be applied to the network simulator forfurther examinations of the link behaviour. In our case, we were mainly interested in optimizingthe system in terms of throughput maximization and in analyzing the end-to-end performance ofTCP/IP connections in a GPRS environment. Especially interesting is the throughput at the LLClevel. In GRPS, a LLC PDU is segmented into RLC PDU of fixed length. Let us assumethe system is operating without the optional link level ARQ. Then, a LLC frame is lost, if at leastone of the n RLC PDUs has been corrupted while being transmitted. Fig. 8 shows the blockwisethroughput at the LLC layer in dependence on n for a moderate channel with a C/I of 10 dB andfor all applicable coding schemes. The blockwise throughput refers to the payload of a LLC PDU,i.e. LLC and RLC protocol overheads are taken into account. Again, it is presumed that a singleuser can use all eight time slots. It can be seen that a proper choice of both n and the codingscheme is required to achieve an optimum throughput. A false strategy may lead to a throughputof less than 10kbit/s compared to the maximum achievable of about 70kbit/s at this operatingpoint. This clearly shows that such studies are crucial for a system optimization. At this point ittherefore cannot be too strongly emphasized that they would be impossible without an accuratelow-complexity abstract model of the wireless link like the proposed MEC model.

192

VII. CONCLUSIONS

We have presented a new probabilistic model based on Markov chains for wireless packet era-sure channels, that approximates the erasure characteristics of the underlying packet transmissionsystem with a high level of accuracy. Besides accuracy, the proposed MEC model limits the com-plexity to a minimum, such that it is easy and efficient to implement and in addition mathematicallytractable. It has been detailed how the MEC model can be implemented in network simulationenvironments. The validity of our approach has been verified by means of the GSM GPRS system.As our simulation results have shown, the proposed MEC model has proved to be an adequatemodel for wireless packet erasure channels according to GPRS. It can surely be expected that theseresults do also hold for other kind of wireless packet erasure channels.

REFERENCES

[1] K. Fall and K. Varadhan, ns Notes and Documentation, The VINT Project, UC Berkeley, July 1999[2] L. Kanal and A. R. K. Sastry, “Models for channels with memory and their applications to error control”, Proc.

IEEE vol. 66, no. 7, pp. 724–744, July 1978[3] G. Liebl, Modeling, Theoretical Analysis, and Coding for Wireless Packet Erasure Channels, Diploma Thesis,

Inst. for Communications Engineering, Munich University of Technology, 1999[4] J. Cai and D.J. Goodman, “General Packet Radio Service in GSM”, IEEE Communications Magazine vol. 35,

no. 10, pp. 122-131, October 1999[5] C. Bettstetter, H.J. Vogel and J. Eberspaecher, “GSM Phase 2+ - General Packet Radio Service GPRS: Archi-

tecture, Protocols, and Air Interface”, IEEE Communications Surveys vol. 2, no. 3, 1999

Reducing handover probability through mobile positioning

Stamatis Kourtis Dr. Rahim Tafazolli

Motorola University of Surrey69 Buckingham street, Aylesbury, CCSR, University of Surrey, Guidford,

HP20 2NF, UK GU2 5XH, UKStamatis.Kourtis@motorola.com R.Tafazzolli@ee.surrey.ac.uk

Abstract—Because of the two-dimensional layout of cellular networks unnecessary handovers (notnecessarily Ping-Pong handovers) could occur. Fact that has not been taken into consideration so far,since handover algorithms found in the literature consider the simplified one-dimensional case of twobase stations. This paper examines in detail how these unnecessary handovers occur and proposes analgorithm in order to combat them, which takes advantage of the future mobile positioning capabilities ofcellular networks. Simulation results are presented depicting the algorithm’s capacity in decreasing thenumber of the performed handovers during an active call. In addition, the algorithm is largely unaffectedby the mobility behavior of the mobile stations and only at high speeds its performance seems tosubstantially deteriorate.

Introduction

Provided that shadow fading was not existent, a mobile station (MS) travelling away from its serving basestation (BS) towards a neighboring BS, would perform only one handover. This handover would beinitiated at the instant when the received signal strength of the neighboring BS would become strongerthan that of the serving BS. However, the reality of the shadow fading affects significantly this veryidealistic situation.The shadow fading causes a long-term variation of the received signal strength, fact that results in theinitiation of several handovers, commonly known as “Ping-Pong” handovers, instead of just one. To

194

combat this phenomenon, numerous handover algorithms have been proposed in the literature [2]-[4]. Allof them consider the simplified case of two BSs and one MS travelling from one to the other, and theyintroduce various techniques like averaging windows and handover margins in order to minimize thenumber of the unnecessary handovers.More or less these proposals succeed in their goal for the case of two BSs. Nevertheless, in real cellularsystems the handover initiation mechanism is dictated by a large number of BSs typically equal to thenumber of BSs included in the neighboring list of the serving BS. Figure 1 illustrates a possible handoverinitiation scenario where an MS located at point A near the cell edge, is heading towards and iscurrently being served by Furthermore, it is assumed chat at point A all the BSs can provide adequatereceived signal strength with the strongest provided by and consequently, a handover will beexecuted to The MS will continue towards and at point B it is probable that the received signalof will be weaker than that of or and a new handover will be initiated, where the MS couldbe served by either or In the case that the MS would be served by then it is possible thatanother handover to will be requested later. In conclusion, the existence of multiple BSs could resultin the execution of unnecessary handovers (not necessarily Ping-Pong handovers).

The rest of the paper is organized as follows. Initially, a detailed handover analysis is given which derivesformulas regarding the probability of the unnecessary handover. Then, these formulas are used to derivethe algorithm, which aims at minimizing the unnecessary handovers. The system model, which will helpto evaluate the performance of the proposed algorithm, is described and simulation results based on thismodel are presented and discussed. Finally, some important conclusions are drawn.

195

Handover analysis

Part of the handover analysis presented here can be found in [6] but it is repeated here for the sake ofcompleteness. The analysis is formulated for the general case of an MS located at the point M travellingwith a velocity amid a number of BSs which are located at points respectively, and its serving BS isat the point The MS measures the signal strength of the serving and N neighboring cells, where N isequal to the size of the neighboring cell list, at constant time intervals (measurement period). Hence, at

the time instant the MS receives the signal levels:

where and are the parameters of the mean signal strength for the link,dist(M(m), is the distance between the points M(m) and Furthermore, is the shadowingfactor, where and are jointly normally distributed, each with zero mean and with

autocorrelation where is the norm of the MS velocity and d is the decorrelation

length. Consequently, is normally distributed with mean and variance where

is the standard deviation of the shadow fading [1].

Assuming that during the time interval the velocity remains constant, then the next position of the

MS is given by:

Let the MS velocity be equal to at the instant Based on a widely used mobility model [1],

given a probability pvelocity for a velocity update, the MS velocity changes to whenever the MS has

managed to travel in the meantime a distance greater than a predetermined update distance d’:

The serving BS gathers K measurement reports and produces an average estimate of the received signalstrength of the various BSs. These estimates are used to calculate the PBGT0,n decision variables:

Assuming that for the mean signal strength related parameters and then Eq. 4 can be writtenas:

196

The sum is normally distributed with mean zero and variance equal to:

A handover is initiated if the relation

where is the handover margin for a handover from the serving cell to the neighbouring cell, issatisfied for one neighboring BS. Therefore, the probability for a handover at the decision instance mKequals:

where and erfc(x) is the complementary error function.

Assuming that at the time instant 0 an MS emerges and starts to be served by (therefore

then the probability that a handover would be performed at the time instant given

that the MS is continuously served by until then, is given by:

Let the cell residence time of a newly emerged MS be (cell residence time of an MS from the momenta call has been started until the first handover according to [5], [6]). Evidently, if the previous

formula corresponds to the probability density function (pdf) (t) of

Typically, the duration of a voice call follows a negative exponential distribution i.e., the probability thata call will end in time duration t is:

It can be easily seen [5] that the probability for a newly emerged MS to perform a handover during thecall is equal to:

197

Additionally, assuming that at the time instant an MS (already in a call process and served by theis handed over to the which is located at the point then Eq. 9 could be rewritten as:

where and (mK) is derived by Eq. 8 for

If the cell residence time of a handed over MS is (cell residence time between subsequent handovers),

then for the previous formula gives the pdf (t) of

Lastly, the probability for a handed over MS to perform an additional handover during the call is equal to:

Evaluation of the unnecessary handover probability

For the calculation of the unnecessary handover probability, we leave the general case and focus on thesimplified network scenario depicted in Figure 1. An MS emerges at the position of (and

consequently served by moves right towards and at the time instant a handover is initiated.

At that point, the MS will be handed over to either or instead of if the received signalstrength of either of these BSs is greater than that of Following the same steps as in the handoveranalysis section, it can be seen that the probability for the MS not to be handed over to is given by:

The overall probability that the MS would be handed over to an incorrect BS or whilst beingactive, is simply:

198

Minimizing the unnecessary handovers

Hereinafter it is assumed that the network has the ability to determine the exact position, speed anddirection of an MS whenever a handover occurs [7]. Continuing the analysis on the simplified network

scenario, it is assumed that actually at the time instant a handover is initiated and therefore, the MSwill handed over to either or Furthermore, it is assumed that all of the potentially servingBSs could provide an adequate signal strength (received signal power greater than the receiversensitivity). At this point, it is interesting to calculate the probability of a subsequent handover given thatthe MS would be handed over to (k = 1, 2, 3).Under the hypothesis that until the end of the call the MS will not change its speed and direction, it is

possible to compute the (mK), where for all the BSs, and consequently from Eq. 12

the (t). Finally, from Eq. 13 the probability of a subsequent handover after the MS has been

handed over to can be calculated. Among the possible BSs, a handover towards the BS whichcorresponds to

ensures that the probability for a future handover is minimum.Interestingly, regarding the time until the end of the call (residual call time) it is worth noting here thatbecause of the memoryless property of the exponential distribution, it is independent of the time elapsedsince the start of the call. As a result, the probability distribution of the residual call time given the timeelapsed since the start of the call is the same as the original call duration

System model (I) – simplified network

For the simplified network of Figure 1 a cell radius of 1.5 km is assumed, resulting in a BSO – BS2 of

The network is deployed in urban or suburban areas which is characterized by macro-cellularpropagation, the path loss model of which is given by [1]:

199

where the distance d is in km. Also, the shadow fading standard deviation is set to 10 dB, thedecorrelation length equals 20 m. and an averaging window of 15 s is used. The handover margin is set to0 dB. Lastly, voice calls have a duration which is exponentially distributed with mean equal to 120 s.

The probability of incorrect handover (MS is not handed over to as given by Eq. 14 against the MSspeed is drawn in Figure 2. Interestingly, for MS speeds from 10 km/h up to 50 km/h there is a probabilityfor an incorrect handover greater than 6%, with a maximum of 10% at a velocity of about 25 km/h.Figure 3 plots the min against all the possible handover decision instances for a MS with avelocity of 50 km/h (apparently As it was expected, by the time it is likely that a handover

will occur (peak of (t)), handover decisions based on Eq. 16 are capable of identifying the BS in

this case), which is more probable to serve the BS until the very end of the ongoing call.

200

System model (II)

In order to test the ability of the algorithm in minimizing the number of performed handovers, two 17x17-cell networks with a cell radius of 1.5 km and 0.5 km respectively are considered, which are deployed inthe same environment with the previously considered 4-cell network.MS are randomly generated and distributed uniformly throughout the network. Again the mean callduration is 120 s. According to the MS velocity two mobility patterns are considered: pedestrian (0 km/hto 5 km/h) and vehicular (10 km/h to 200 km/h). More specifically, a pedestrian-type MS changes rapidly

its direction (±180°) at the position update, whereas a vehicular-type MS changes its direction to a

relative maximum of ±45°. In both cases, an MS has a 20% likelihood to change its direction at everyposition update and a position update point is declared whenever an MS travels a distance equal to thedecorrelation length from the last position update [1]. Also, it is assumed that there is no error in thedetermination of the MS position and mobility characteristics.Lastly, the receiver sensitivity level is set to –100 dBm.

Simulation results

Simulations were run in order to test the relative performance characteristics of two handover algorithms:Strongest received signal (SRS) and minimization of unnecessary handovers (MUH, Eq. 16).Given that the fundamental assumption of the MUH algorithm is of an MS which does not change itsspeed and direction until the end of the call, it is imperative to test the effectiveness of the MUHalgorithm under this assumption and to compare the results against a typically behaved MS. Table 1quotes the expected number of handovers during a voice call, when the MS speed is 50 km/h.Interestingly, the assumption regarding the invariability of the MS direction does not impact at all theMUH algorithm. The location and mobility characteristics of the MS are used indirectly in the MUHalgorithm to calculate received signal strengths, since essentially this algorithm compares expected signalstrengths for a number of future positions of the MS. The application of shadow fading results indecreasing the dependence of the signal strength calculation on the relative position of the MS with

201

respect the adjacent BSs. Consequently, the MUH algorithm tolerates effectively the mobility nature ofthe MSs. As expected, the same is also observed for the SRS algorithm.

Performance curves of the RSR and the MUH algorithms are drawn in Figure 4 for a range of MS speeds.From 0 km/h to 5 km/h an MS exhibits pedestrian mobility pattern behavior, whereas from 10 km/h up to200 km/h it exhibits vehicular mobility pattern behavior. The cell radius is 1.5 km.The results show that for a significant range of MS velocities, the handover algorithms do not actuallydistinguish between the two mobility patterns. In essence, this indicates that all the handover algorithmsare quite tolerable against the mobility behaviors of the various MS.At low pedestrian MS speeds (below 0.5 km/h), the RSR algorithm exhibits a substantial increase in thenumber of performed handovers. This is due to the fact that at these speeds consecutive signal strengthmeasurements are highly correlated, therefore there is an increased likelihood for additional handovers. Incontrast, the MUH algorithm takes into consideration this correlation and manages to reach its bestperformance at these low speeds.

202

At high vehicular MS speeds, all the algorithms perform more handovers. The substantial degradation ofthe MUH performance is because of the fact that for a 200 km/h MS the mean traveled distance is about6.6 km or almost 4.5 times the cell radius. Thus, after a few position updates BSs, which are supposed tobe in the MS’s locality, could be resulted in being far away from the MS and vice versa. Conversely, theRSR algorithm appears to have a more graceful degradation of its performance. Nevertheless, the RSRalgorithm again performs a greater number of handovers per call than the MUH algorithm. Lastly, if alarger cell radius were chosen, then the performance degradation of all the algorithms at the high MSspeeds may appear smaller or even in some cases, it may not appear at all.Essentially, the presented results of the MUH algorithm are optimal, since no errors are considered for thedetermination of the MS position, speed and direction and the impact of such errors on its performanceworth investigation. Nonetheless, because of the depicted results in Figure 4, it is anticipated that thealgorithm should be largely unaffected from errors of the MS speed, since for a significant range ofspeeds the algorithm gives virtually the same results. On the other hand, errors in the determination of theposition and the direction are quite difficult to understand how they degrade the performance of thealgorithm and detailed evaluation is required.

Conclusions

This paper examined the unnecessary handovers that could occur because of the two-dimensional layoutof a cellular network, fact that has not been taken into consideration so far, since handover algorithmsfound in the literature consider the simplified one-dimensional case of two BSs. A detailed analysis of thephenomenon resulted in an algorithm (minimization of unnecessary handovers or MUH), which seems tocombat efficiently the problem of the unnecessary handovers. In short, this algorithm takes advantage ofthe future ability of the cellular networks to know the position of the MSs, and makes the assumption thatfrom the handover decision instant and onwards until the end of the call the MS will not change its speedand direction. Upon a handover request, an MS is handed over to the BS which gives the minimumprobability of a subsequent handover.In comparison to the commonly used received signal strength (RSR) algorithm, the MUH algorithmreduces the number of performed handovers by 20% and 50% when the cell radius is 0.5 km and 1.5 kmrespectively. Moreover, its fundamental assumption does not effect whatsoever its capability in reducingthe handover probability. Lastly, the algorithm is not affected by the particular mobility characteristics of

203

the MS and its performance deteriorates only when the MS speed is quite high (with regards to the cellradius).Unfortunately, such an algorithm is difficult to implement since several parameters have to be estimatedbeforehand including the path loss exponent, the shadow fading standard deviation and the decorrelationlength. As this algorithm would reside on every BS, and generally in real systems the various BSs arelocated in different environments, different implementations (and hence parameter estimations) of thealgorithm have to be done. Consequently, in theory the MUH algorithm may give very good results,however in practice its implementation has important difficulties.Because of these limitations, two other algorithms are presently evaluated which overcome theimplementation difficulties of the MUH algorithm and perform equally well. This work is currently beingdone and it will be presented in a future paper.

References

[1] ETSI TR 101 112, “Universal Mobile Telecommunications System (UMTS); Selection proceduresfor the choice of radio transmission technologies of the UMTS”, V3.2.0, April 1998

[2] Zonoozi M., Dassanayake P., “Handover delay and hysteresis margin in microcells andmacrocells”, IEEE International Symposium on Personal, Indoor and Mobile RadioCommunications, Vol. 2, pp. 396-400, Helsinki, September 1997

[3] Senadji B., Boashash B., “Estimation of the hysteresis value for handover decision algorithmsusing Bayes criterion”, 1997 International Conference on Information, Communication and SignalProcessing, Vol. 3, pp 1771-1775, Singapore, September 1997

[4] Benvenuto N., Santucci F., “A least squares path-loss estimation approach to handover algorithms”,IEEE Transactions on Vehicular Technology, Vol. 48, No. 2, pp. 437-447, March 1999

[5] Zonoozi M., Dassanayake P., “User mobility modeling and characterisation of mobility patterns”,IEEE Journal on Selected Areas in Communications, Vol. 15, No. 7, pp. 1239-1252. September1997

[6] (to be presented) Kourtis S., Tafazolli R., “Evaluation of handover-related statistics and theapplicability of mobility modelling in their prediction”, IEEE International Symposium onPersonal, Indoor and Mobile Radio Communication, London, September 2000

[7] Reed J.H., Rappaport T.S., Woemer B.D., “Position location using wireless communications onhighways of the future”. IEEE Communications Magazine, Vol. 34, No. 10, pp. 33-41, October1996

Multi-user Detection Using the Iteration Algorithm in Fast-Fading Channels

Sun-Jin Yeom, Yong -Wan ParkDepartment of Information & Communication Engineering, Yeungnam University

#214-1, Daedong, Kyung-San City, Kyungpook, Korea. 712-749TEL: +82-53-810-3523, 810-1539 FAX : +82-53-814-5713, 812-8583

E-Mail: ywpark@ynucc.yeungnam.ac.kr

Abstract

In this paper, we introduce a modified interference cancellation scheme for multi-userdetection in DS/CDMA. Among ICs(Interference Cancellers), PIC(Parallel InterferenceCanceller) requires the more stages to have the better BER(Bit Error Rate), and SIC(SuccessiveInterference Canceller) faces the problems of power reordering and large delays. Most of all, theadaptive detector achieves the good performance using the adaptive filter conducted iterationalgorithm. But it requires many iterations for convergence. To resolve those problems, wepropose a new hybrid interference cancellation scheme combining nonlinear detector and theadaptive filter using CMA( Constant Modulus Algorithm). The proposed interference cancellerhas improved the performance through the received signal applied ranking scheme is fed into theadaptive filter. That is, its structure provides the same BER performance even though it iteratesthe smaller than the conventional AD(adaptive detector) because the signal removed MAI is fedinto the adaptive filter. The proposed IC structure does extract the following characteristics. ; (1)

detector having adaptive filter requires less complexity than nonlinear detector.

I. INTRODUCTION

In the third generation system called IMT-2000, direct-sequence code division multipleaccess(DS/CDMA) has been selected as the vital multiple access technique. To access userssimultaneously and independently, it is applied to the signature waveforms having orthogonality

it has the same BER performance only using smaller iterations than the conventional AD, (2) the

206

[1]. If the signature waveforms were perfectly orthogonal, a bank of single-user detectors wouldachieve optimum demodulation and detection. But the wireless communication systems areseverely affected by MAI(Multiple Access Interference) and near-far problem where a high-powered signal creates significant MAI for low-powered signals because in practice the systemdesign limitations is destroyed the orthogonality of assigned signatures; system complexity,bandwidth limitation and others. To eliminate this problems, several techniques suppressedinterferences have been used in communication system, or investigated or proposed. Most of all,the central mechanism for resource allocation and interference management is power control. Itis used basically as a mechanism to keep the received powers of users equal so that the nearbyusers do not dominate over the far away users. But the power control can’t perfectly remove theMAI because of the characteristics of DS/CDMA, and this adds complexity to the system, andinaccuracies in power control have a detrimental impact on performance.

The multiuser detection schemes have the more fundamental potential of significantlyraising system capacity by removed MAI [2]~[4]. Among them, PIC and SIC appear to beattractive for different reasons. PIC estimates and subtracts out all of the MAI for each user inparallel [2]. And to have the better BER(Bit Error Rate) performance, it requires the morestages. In contrast, SIC takes a serial approach to subtracting out the MAI [5]. By thosereasons, PIC requires more hardware, and SIC faces the problems of power reordering and largedelays. The adaptive detector may solve the assumed knowledge and complexity problems ofnonlinear detectors simultaneously [7],[8]. It realizes a filter that both is matched to the desireduser’s channel-affected signal and removes interference from other users. And it achieves thegood performance using the adaptive filter conducted iteration algorithm. But it requires manyiterations for convergence [9],[10].

In this paper, we propose a new hybrid interference cancellation combining nonlineardetector(SIC & PIC) and the adaptive filter using CMA(Constant Modulus Algorithm) [11]. Theproposed interference canceller has improved the performance through the advanced form ofreceived signal applied ranking scheme is fed into the adaptive detector [12]. That is, it providesthe same BER performance even though it iterates the smaller than the conventional AD becausethe signal removed MAI is fed into the adaptive filter. The paper is organized as follows : Insection II, we describe a simple DS/CDMA system. And in section III, we propose the modelingof the modified interference canceller and its characteristics. Section IV compares theperformances between the proposed IC and the others. The conclusion is included in section V.

207

II. SYSTEM MODELING

We considered the reverse link of a asynchronous DS/CDMA system that K active userstransmit the modulated band-pass signals which are spread by the each pseudo noisecodes(PG=31). The modulated signals are transmitted over the channel which experience MAI,Rayleigh fading(Jake model) and Additive White Gaussian Noise(AWGN). The channel model isshown in Fig. 1. If the DS/CDMA system applied to the perfect power control, several effectsaffected system capacity would be significantly reduced. But in recent because power controltechnique isn’t perfect, it is assumed that power control can perfectly compensate aboutvariations of the received signal due to large scale fading, but the received signal is affected bythe effect of the fast Rayleigh fading. The received signal can be described as

where I is the total number of transmitted symbols, K is the total number of active users, L is thetotal number of multipaths, P is transmitted chip power, b is transmitted binary information {+1,-1}, n(t) is assumed AWGN to be zero mean and have variance of And S(t) is as follows

where s(t) is a spreading code of each user, T is the symbol interval, u(t) is a unit square pulse

shape spanning the interval and c is a time delay and channel coefficient due to the

Rayleigh fading.

208

III. THE PROPOSED SYSTEM

In this paper, we propose a modified interference cancellation combining nonlinear detectorand the adaptive filter using CMA(Constant Modulus Algorithm). The proposed interferencecanceller has improved the performance rather than the conventional AD through the receivedsignal applied ranking scheme is fed into the adaptive filter as shown in Fig. 2. First of all, thesystem designer can randomly assign to user numbers composed each group. However, to get thebetter performance, the system may require the assignment that followed the distribution ofsignal strength because the ranking scheme is effectively applied. But in this paper we assumedthat the same user numbers(U = total users[K] / total groups[G]) is assigned in each group to thesimple application [6].

The key of the proposed system is “generation of the advanced input signal” in Fig. 2. Toimprove the received signal, the system can be proceeded with the ranking scheme before theprocess of adaptive filter. That is, each user signal is correlated with the received signal in amatched filter. And total users based upon the strength is composed the each group. Thisprocedure is shown in Fig. 3. The user signals of each group are then mapped by HD(HardDecision) method and regenerated to generate the input signal of the adaptive filter. Each group

is fed into the equation (3) except for 1st group

209

where g is a specific group among total groups(G), U is users composed each group, f is themappedsignal. The correlated signal is given by:

where MAI is interference which is affected by both signals of different users and self-signal ofthe different paths, N(t) is the correlated noise. The mapped signal is as follows

According to the equation (3)~(5), total users are grouped using the ranking scheme. Togenerate the input signal of each group in the adaptive filter bank, the regeneration signal of f isused. That is, the input signal subtracts the regeneration signals of the groups excluded self-group from the total received signal. And it can be described as

210

After the advanced signal generates, it is fed into the each adaptive filter bank.

Because the blind algorithm doesn’t need the use of initial training and need only the signaturewaveform of the desired user is known [10], we selected the blind algorithm in the proposedsystem. The adaptive filter using CMA(Constant Modulus Algorithm) is shown in Fig. 4.

We define the cost function based on CM(Constant Modulus) criterion as

where the index k designates the k-th bit interval and is a positive constant. Invoking a

gradient search algorithm on J CM , we can obtain the weight vector update recursion as follows;

the detected signal of each adaptive filter is as follows;

The characteristics of the proposed IC are as follows; As the adaptive filter bank makes useof the advanced input signal the proposed IC may has much better BER performance than

the conventional AD by fed into r(t). In particular the number of iterations is reduced rather thanthe conventional AD on condition that they require the same BER performance because of theadvanced input signal to reduce MAI. In aspect of complexity, the drawback of nonlinear multi-user detector is much complexity to do the cancellation.

211

But the proposed IC is composed to the simple hardware to implement the adaptive filter.Furthermore, it can operate in a slow power control. That is to say, the fast power control is notmuch needed for the implementation.

III. SIMULATION

In this section, we analyze the performance between the proposed IC and the other ICs. Thesimulation environment is as follows. In the reverse link, the base station suppressesinterferences through the slow power control, which power of the active users is distributed from0dB to 30dB. And we consider the synchronous system which the number of maximum users isK = 28 and the processing gain is 25

5

–1. The channel model is assumed AWGN and theRayleigh fading that the Doppler frequency is 176Hz which corresponds to a mobile speed of100km/h at 1.9GHz. And in the initial stage of the proposed IC, the number of groups is G=2 andG=4 with the same number of users within each group.

First of all, we prove that the detector having the advanced signal has the better BER

performance than the detector having the received signal r(t). In Fig. 5, to confirm theperformance of only advanced signal we described the BER performance versus the number ofactive users and the SNR between the proposed IC and the conventional detector on conditionthat the number of iterations is 0. In (a), if the total active users are 20, the BER of the

conventional AD is 1.726×10-1, that of the proposed IC with G=2 is 1.359×10-1, and with

G=4 is 1.161×10-1 In Fig. 6, we described the BER performance versus the number of activeusers and the SNR between the proposed IC and the other representative ADs, which are theconventional blind detector(blind-LMS, CMA). In (a), if the total active users are 20 and the

number of iterations is 400, the BER of the blind-LMS detector is 1.706×10-1, that of the CMA

detector is 1.560×10-1, that of the proposed IC with G=2 is 1.036×10-1, and with G=4 is

8.74×10-2. The proposed IC has the better BER performance than the other ADs.In Fig 7, MSE convergence profiles are shown for a system with spreading ratio of 31, four

active transmitters of unequal power, and 6dB. Because of the advanced signal removed

the MAI, the proposed IC has the faster convergence than the other ADs as the number of groupsis increased, and .the smaller vibration about iterations.

212

213

As we analyzed the above results, the proposed IC can be the better BER performance thanthe conventional AD because of the application of cancellation scheme. And we can make certainthat the BER can be better as the number of groups is increased because the more interferedsignals can be cancelled as the system is designed for many groups.

214

IV. CONCLUSION

To do the fast convergence of the conventional adaptive detector, we evaluated the CMAdetector fed into the advanced signal in this paper. In the same simulation condition, the

BER performance of ICs was simulated to compare with each other. Through the pre-generatedsignal of the group form, the adaptive detector of the proposed IC is fed into the input signal thatsubtracts the regeneration signals of the groups excluded self-group from the received signal.Consequently this procedure is efficient for the pre-reduction of MAI before the adaptive filter isprocessed. The characteristics of the proposed IC are as follows. Because MAI is reduced beforethe adaptive filter, the proposed IC has the faster convergence than the other ADs. And Thenumber of iterations is reduced rather than the conventional AD on condition that they requirethe same BER performance because of the advanced input signal to reduce MAI. And theproposed IC is composed to the simple hardware to implement the adaptive filter.

REFERENCES

[1] S. G. Glisic, P. A. Leppanen, Code Division Multiple Access Communications, Kluwer Acade-mic Publishers, 1995

[2] S. Moshavi, “Multi-User Detection for DS-CDMA Communications”, IEEE CommunicationMagazine, Oct. 1996, pp. 124-136

[3] R. Lupas, S. Verdu, “Linear Multiuser Detectors for Synchronous Code-Division Multiple-Access Channels”, IEEE Transaction on Information Theory, Vol. 35, No. 1, Jan. 1989,pp.123-136

[4] S. Verdu, Multiuser Detection, CAMBRIDGE University Press, 1998[5] P. Patel, J. Holtzman, “Analysis of a Simple Successive Interference Cancellation Scheme in

a DS/CDMA System”, IEEE Journal on Selected Areas in Communications, Vol. 12, No. 5,June. 1994, pp.796-807

[6] S. Sun, L. K. Rasmussen, H. Sugimoto, T. J. Lim, “A Hybrid Interference Canceller inCDMA”, IEEE fifth International Symposium on Spread Spectrum Technology &Application., Vol. 1, Sep. 1998, pp. 150-154

[7] Simon Haykin, Adaptive Filter Theroy, Prentice Hall, 1996

215

[8] S. Verdu, Adaptive Multiuser Detection, Kluwer Academic Publishers, 1995[9] G. Woodward, B. S. Vucetic, “Adaptive Detection for DS-CDMA”, Proceedings of The IEEE,

Vol. 86, No. 7, July 1998, pp.1413-1434[10] M. Honig, U. Madhow, S. Verdu, “Blind Adaptive Multiuser Detection”, IEEE Transactions

on Information Theory, Vol.41, No.4, July 1995, pp. 994-960[11] W. Lee, B. R. Vojcic, “Constant Modulus Algorithm for Blind Multiuser Detection”, In

Proceeding of the ISSSTA ’96, Germany, 1996, pp. 1262-1266[12] Sunjin Yeom, Panjong Park, Yongwan Park, “Evaluation of Parallel Interference

Cancellation with the Advanced First-Stage in Rayleigh-Fading Channels”, VTC 2000, toappear.

FPGA DSP for Wireless Communication

Chris DickXilinx Inc., 2100 Logic Drive, San Jose, CA 95124, USA

chris.dick@xilinx.com

fred harrisCollege of Engineering, San Diego State University, San Diego

fred.harris@sdsu.edu

1 Introduction

Software defined radios (SDR) are highly configurable hardware platforms that provide the technologyfor realizing the rapidly expanding third (and future) generation digital wireless communicationinfrastructure. Figure 1 is a generic model of the signal processing subsystem in a SDR. As shown in thefigure, many sophisticated signal processing tasks are performed in a SDR, including advancedcompression algorithms, power control, channel estimation, equalization, forward error control (Viterbi,Reed-Solomon and Turbo coding/decoding) and protocol management.

Digital filters are employed in a number of ways in DSP based transmitters and receivers. Polyphaseinterpolators are used in the transmitter for upsampling a baseband signal to the digital IF (intermediatefrequency), to ensure compliance with the appropriate regulatory bodies spectral requirements, and tomatch the signal's bandwidth to that of the channel. In the receiver section of the system multi-stagepolyphase filters are frequently used in a digital down converter (DDC) to perform channelization anddecimation. Complex filters are also required for estimating channel statistics and performing channel

218

equalization to compensate for multipath effects, and to correct for phase and amplitude distortionintroduced during transmission. Re-sampling filters are also an integral component in all-digital symbolsynchronization loops. Finite impulse response (FIR) differentiators are also commonly used during thedemodulation process, for example, in systems that use frequency modulated waveforms for channelaccess.

While there is a plethora of silicon alternatives available for implementing the various functions in aSDR, field programmable gate arrays (FPGAs) are an attractive option for many of these tasks for reasonsof performance, power consumption and configurability.

This paper provides a brief FPGA architecture review and then describes several signal processingfunctions implemented using FPGA technology. Single-rate and multi-rate FPGA filter mechanizationsare presented. A QPSK carrier recovery circuit is described and its FPGA implementation is described.

2 FPGA Architecture

There are a rich range of FPGAs provided by many semiconductor vendors including Xilinx, Altera,Atmel, AT&T and several others. The architectural approaches are as diverse as there are manufacturers,but some generalizations can be made. Most of the devices are basically organized as an array of logicelements and programmable routing resources used to provide the connectivity between the logicelements, FPGA I/O pins and other resources such as on-chip memory. The structure and complexity ofthe logic elements, as well as the organization and functionality supported by the interconnectionhierarchy, distinguish the devices from each other. Other device features such as block memory and delaylocked loop technology are also significant factors that influence the complexity and performance of analgorithm that is implemented using FPGAs.

A logic element usually consists of 1 or more RAM (random access memory) n-input look-up tables,where n is between 3 and 6, and 1 to several flip-flops. There may also be additional hardware support ineach element to enable high-speed arithmetic operations.

As a specific example, consider the Xilinx series of FPGAs [2]. The logic elements, calledslices, essentially consist of two 4-input look-up tables (LUTs), two flip-flops, several multiplexers andsome additional silicon support that allows the efficient implementation of carry-chains for building high-speed adders, subtracters and shift registers. Two slices form a configurable logic block (CLB). The CLBis the basic tile that is used to build the logic matrix. Some FPGAs supply on-chip block RAM. Figure 2shows the CLB matrix that defines a Xilinx Virtex FPGA. Current generation Virtex silicon provides afamily of devices offering 768 to 32,448 logic slices, and from 8 to 280 variable form factor true dual-portblock memories.

Xilinx XC4000 and Virtex [2] devices also allow the designer to use the logic element LUTs as memory -either ROM or RAM. Constructing memory with this distributed memory approach can yield accessbandwidths in the many tens of gigabytes per second range.

Typical clock frequencies for current generation devices are in the multiple tens of mega-Hertz (100 to200+) range.

3 Digital Filters

The FIR filter is one of the basic building blocks common to nearly all digital signal processing systems.In demanding applications that require a large filter order, high sample rate, or combination of both theseparameters, the arithmetic workload required can be quite substantial. For an O(N-l) filter, N

219

multiplications and N-1 additions are required to compute a single output sample. High-performance real-time filter platforms are of great interest to the signal processing and digital communications community.

A common option for implementing real-time filters is a software programmable signal processing chip.A higher performance, but less flexible alternative is an ASIC solution. A more recent design option is toexploit the parallelism that an FPGA-based hardware system can provide.

3.1 Implementing Digital Filters Using FPGAs

There are numerous options for implementing FIR filters in an FPGA. The most obvious approach, and asmay be expected this is not always the most optimal solution, is to model the technique used in an ASICor instruction-set based DSP (ISDSP). This is to employ a scheduled multiply-accumulate (MAC) unit.Since many signal processing engineers are familiar with these semiconductor technologies for realizingfilters, we will use this as the starting-point for examining FPGA realizations of FIR filters.

3.2 The MAC Based Approach

An inner-product computation may be partitioned over 1 or several MAC units. And this is a commonapproach used by current generation signal processors - both ASICs and ISDSPs. This same method canobviously be used in an FPGA implementation. But in the FPGA environment the designer has virtuallycomplete control of the silicon and can decide how much of this resource is allocated to the inner-productengine. As a reference point, a 16-by-16 MAC (32-bit precision result) occupies 174 logic slices, or 5.6%of an FPGA like the XCV300 [2]. This unit will support a clock frequency of 162 MHz in a –6 speedgrade. This is a medium speed grade component in the context of FPGAs that are currently available: -8being the fastest. One such functional unit provides good performance, but of course the flexible andhighly parallel nature of FPGAs allow for the construction of highly concurrent systems, and inner-product engines with many MAC units may be constructed to produce very high-performance systems.

220

Instead of using functional unit concurreny to increase performance, fitter parallelism may be introducedin a different manner by employing and alternative algorithm. Distributed arithmetic [1] provides onesuch opportunity.

3.3 FIR Filters: An Alternative Approach Using Distributed Arithmetic

FPGAs can bring high-performance, power and bandwidth efficiencies, flexibility through re-configuration, in addition to economic benefits to a design. One additional, and very excitingcharacteristic that FPGAs provide, is access to the wide range of creative solutions to DSP tasks that havebeen reported in the open literature over the last few decades. Because FPGAs are like miniature siliconfoundries with extremely short tum-around times, the system architect is free to explore a wide range ofpotential solutions to a problem. This option is often not available using ISDSP. The ISDSP chip designermust define a data-path that solves a large range of problems and provide an adequate level ofperformance. Many novel signal processing algorithms just do not map well onto the pre-defined data-path of a software programmable signal processor.

There are in fact many ways to compute an inner-product. One approach first published in the openliterature by Peled and Liu [1] is called distributed arithmetic (DA).

A generic model of a DA filter is shown in Figure 3. This technique has several characteristics that makeit well suited to implementation in distributed memory based FPGAs. For an excellent tutorialpresentation on distributed arithmetic based DSP for performing FIR, IIR (infinite impulse filter) andFFTs (fast Fourier transforms), the reader is referred to the article by White [3].

Distributed arithmetic based calculations require a series of table look-up operations, additions andsubtractions. All of these functions are highly suitable to FPGA implementation. One interesting propertyof DA filters is that the filter throughput is no longer coupled to the filter length, but instead has adependency on the input sample precision. This is indicated by the linear plots in Figure 4. The figuresays that in a DA FIR filter mechanization, for a given input sample precision B, the sample rate remainsconstant independent of the number of filter taps. For example, for B= 12 and a 100 MHz system clock,the filter sample rate is 8.333 MHz for filter lengths of 10, 20 100, 200,... and so on. For N=200, this is aneffective computation rate of one MAC every 0.6 microseconds, or 1.7 Giga-MACs per second.

For 24-bit input samples, the sample rate is 4.1666 MHz. For many applications the filter coefficient set issymmetrical. Symmetry can be exploited to minimize the logic requirements of the filter implementation.The filter rate is reduced slightly when this is done. For 24-bit input samples, and a 100 MHz clock, thesample rate will be 4 MHz. For a 200-tap filter this still results in an impressive 800 Mega-MAC persecond computation rate.

221

3.4 Increasing the Speed of Multiplication - Parallel Distributed Arithmetic

In its most obvious and direct form, DA based computations are bit-serial in nature. Extensions to thebasic algorithm remove this potential throughput limitation. Processing the data serially, one-bit-at-a-time, can result in modest computation rates. When the input variables are B bits in length, B clock cyclesare required to complete an inner-product calculation. Additional speed may be obtained in several ways.One approach is to partition the input words into L subwords and process these subwords in parallel. Thismethod requires L-times as many memory look-up tables and so comes at a cost of a linear increase instorage requirements. Maximum speed is achieved by factoring the input variables into single bitsubwords. The resulting structure is a fully parallel DA (PDA) FIR filter. With this factoring a new outputsample is computed on each clock cycle. PDA FIR filters provide exceptionally high-performance. Forexample, consider an 80-tap filter using 12-bit precision for both the coefficients and input samples.Using the Xilinx Core Generator [8] to produce the FPGA realization, the filter occupies 2864 logic slicesand comfortably supports a 150 MHz sample rate. This is equivalent to 12 billion multiply-accumulatesper second. The design floorplan is shown in Figure 5.

222

3.5 Why DA for FPGAs?

Several architectural features make DA filters well suited for implementing using Xilinx Virtex FPGAtechnology. The function generators (FGs) in these FPGAs can be efficiently used as shift registers(SRL16 [2]). This functionality is required to implement a bit-serial delay line, or time-skew buffer, thatkeeps a history of the filter input samples. The time-skew buffer tap points are used as address inputs tothe DA look-up table (LUT). Function generators can also be configured as RAM and ROM. Thiscapability may be used to efficiently build very high-speed, and highly parallel (if required), DA look-uptables.

An alternative to implementing the DA LUT in distributed memory is to employ the block memorypresent in recent generation FPGAs. Consider building a 70-tap filter linear phase FIR filter. There areonly 35 unique filter coefficients. A simple single LUT approach requires storage for a prohibitively largenumber of partial product terms. This becomes much more manageable if the 35 address lines arepartitioned in to 4 groups of 8 and a group of 3. Now only four 256-entry and one 8-entry LUTs arerequired. These 5 tables can be stored in on-chip block memory and a simple adder tree employed tocombine their outputs.

3.6 FPGA Multi-rate Filters

Having access to efficient single-rate FPGA FIR filters enables the construction of a wealth of multiratefilters. Polyphase interpolators and declinators like that shown in Figure 6 and Figure 7 respectively, canbe implemented using either a time shared MAC approach or distributed arithmetic.

223

3.6.1 Polyphase Decimators

One method suitable for implementing a polyphase decimator is to employ a distributed arithmeticapproach. Each polyphase segment could be implemented as a separate DA process. All of the segmentsoperate concurrently to produce a down-sampling filter that can support a very high input sample rate.For example, consider an 8-to-l complex (real input samples, complex coefficients) decimating filter withan input sample rate of 80 MHz and an output rate of 10MHz. Each of the filter segments is to have 32-taps and support 10-bit input samples and 12-bit coefficients. Each polyphase sub-filter has a computationrate of 320 MMACs (for a 100 MHz clock). The net performance for the complex filter is 5.12 Giga-MACs. In addition to the raw performance provided by this structure, one other important dimension ofthe problem to consider is the logic requirements. One of the 32-tap filters consumes 126 Virtex logicslices. The complete filter occupies 2,200 Virtex logic slices. An XCV400 FPGA provides 4,800 slices,and can accommodate two of these decimating structures and provide a performance in excess of 10billion MACs per second. The system clock rate assumed in this example is 100 MHz. Current generationFPGA technology can easily support this rate, and can in fact support 200 MHz clock frequencies forcertain arithmetic functions. One final point to note is that filter coefficient symmetry has not beenexploited. Doing so, if possible, would yield more compact realizations.

3.6.2 Polyphase Interpolators

A polyphase interpolator could be implemented in an FPGA using distributed arithmetic or a time-sharedMAC approach. The different mechanizations are appropriate for different problems. For example, in adigital receiver many control loops must operate for carrier and symbol synchronization, as well as forautomatic gain control. In the timing recovery loop, an interpolator is required to adjust the signal samplephase, and for driving the tracking loop itself. While the interpolator may need to support many phases, atany one time only a small number of segments will typically be operational. In his case a conventionalscheduled MAC approach is appropriate. The full set of coefficients for all the filters would be stored inblock or distributed memory, and the required set of coefficients (corresponding to a polyphase arm)would be directed to one or several MACs to form the inner-product calculation.

3.6.3 CIC Filters

There are many alternatives for realizing multirate filters in addition to the structures described above. Forexample, the cascaded-integrator-comb approach first published by Hogenauer [4] is highly suitable forFPGAs because its reliance on adders, registers and subtracters for performing the arithmetic: FPGAs areextremely efficient for realizing these functions.

4 Carrier Recovery Using A QPSK Costas Loop

There are many options for implementing carrier phase and frequency synchronization in a digitalcommunication system. At the heart of all synchronizers is the phase-locked loop (PLL).

4.1 Phase Locked loops

The generic PLL is shown in Figure 8.

224

PLLs are servo control mechanisms whose controlled parameter is the phase of a locally generated replicaof the incoming carrier wave. Phase locked loops have three basic components: a phase detector, voltagecontrolled oscillator (VCO) and a loop filter. The phase detector measures the difference between phaseof the local oscillator and the input carrier. This signal is fed to a loop filter that governs the response ofthe PLL to variations in the error signal. The Loop filter is designed to track changes in the error signal,but not be overly responsive to receiver noise. The loop filter determines the type of disturbances the PLLcan track, for example, a phase or frequency step. A detailed description of PLL operation can be found in

In an all-digital receiver a digital phase-locked loop (DPLL) like that shown in Figure 9 is required. ThisDPLL employs a second order infinite impulse response (IIR) loop filter. The two filter coefficients ki andkp control the filter comer frequency and damping ratio. In the digital implementation, the VCO in Figure8 is replaced with a direct digital synthesizer (DDS). The phase detector is implemented using the arc-tanfunctional unit in the figure.

4.2 QPSK Costas Loop

Communication systems employing QPSK modulation are very common. The basic Costas loop [5] canbe enhanced to perform carrier recovery and symbol detection for a QPSK modulation scheme as shownin Figure 10.

[5].

225

To understand the operation of this loop, consider the scenario when the loop is reasonably near lock. Thesignal on the I processing arm (or rail) after the LPF is close to the data symbol value and the signalon the quadrature arm is close to The slicer enforces this by ignoring small perturbations in thesignal, which could be due to the opposite-rail symbol if the loop is not locked, or the shaping of thepulse, or simply channel noise. The ±1 symbol decisions feed a network that produces from the receivedbaseband signals a phase difference signal. This signal, working with the loop filter and the VCO, operatelike the basic PLL shown in Figure 8.

4.3 QPSK Costas Loop Implementation

To produce a fixed-point arithmetic realization of the QPSK Costas loop in Figure 10 a combination ofMatlab [6] and Simulink [7] where employed. After the quantized model was verified in the Simulinkdomain, a conventional FPGA implementation flow using VHDL andintellectual property cores (like optimized multipliers and gain blocks) is used to produce the final design.

To verify the operation of the loop a system level design was developed that modeled a simple transmitterand channel that simulated a Doppler shift of the transmitter carrier wave. In practice, the Doppler shift isassociated with movement between the transmit and receive platforms, as might be the case with acellular handset user traveling in a car.

The transmitter generated a pseudo random complex sequence that was shaped by a multirate transmitfilter with an excess bandwidth and an interpolation factor of l-to-8. The channel modelintroduced a small frequency translation of the carrier. Therefore, the receiver was presented with a signalthat had a frequency and phase offset compared to the nominal local oscillator. The purpose of the Costasloop is to track the frequency and phase offset to allow coherent demodulation of the transmittedwaveform. The sequence of plots in Figure 11 provide some insight to the operation of the carrierrecovery loop. Figure 11(a) shows the QPSK constellation diagram after the matched filter. Figure 11(b)is the corresponding eye diagram. The eye is clearly open and, in the absence of any channel impairments,the receiver can easily make correct symbol decisions using this waveform. The frequency translationapplied to the transmitted signal causes the constellation to rotate as shown in Figure 11(c). The receivereye diagram shown in Figure 11(d) clearly shows the eye is closed, indicating that valid symbolsdecisions cannot be made. The Doppler shift modeled in this experiment causes a frequency translation ofthe carrier. The corresponding phase slope is linear, with a gradient that corresponds to the magnitude ofthe frequency offset. One way to observe and quantify the performance of the carrier tracking loop is to

226

monitor the phase function of the interfering signal and that of the oscillator in the Costas loop. This isshown in Figure 11(e) and Figure 11(f). We observe that the loop attains lock after a few hundredsamples. The difference between the two phase functions, or phase error, is presented in Figure 11(g).Finally the de-rotated constellation is shown in Figure 11(h).

The quantized model was developed using the Simulink fixed-point blockset. This approach allowed ahigh degree of design compression. After the Simulink floating-point model was completed and verified,approximately 30 minutes was required to generate the quantized solution.

4.4 FPGA Implementation

Several functional units are required to implement the carrier recovery loop. A complex heterodyne isemployed to down-convert the input signal. This is of course recognized as a complex multiplier. Thereare two matched filters, one for each of the I and Q arms. The phase detector is straightforward,consisting of two 1-bit slicers (sign detector), two 2's complementers and a subtracter. The second-orderloop filter is realized using two multipliers, an integrator and an adder. The local replica of the carrierwave is generated by a DDS.

The complex multiplier is implemented using 4 multiplications and two additions. The ADC samples arerepresented using 8-bits, while the heterodyning signal employs 12-bit samples. Each 8x12 multiplieroccupies approximately 81 logic slices. The complete multiplier occupies 344 slices. the recursive natureof the Costas loop demands the use of purely combinatorial multipliers and adders.

Two matched filters are required. One for each of the I and Q processing arms. These filters are 97-tapsymmetrical FIR filters with 12-bit coefficients and support 9-bit precision input samples. The filters weregenerated using the Xilinx Core Generator [8] filter compiler. The implementation employs serialdistributed arithmetic. Taking advantage of the symmetrical coefficient data, each filter occupies 248Virtex FPGA logic slices. Using 9-bit precision input samples, the filter requires 10 clock cycles tocompute a new output. The bit-clock for the filter is a function of the FPGA speed grade. Typical valuesare between 100 and 150 MHz. This translates to a sample throughput of 10 to 15 MSamples/sec.

The multipliers in the loop filter occupy most of the logic resources for this sub-system. The coefficientparameters are represented using 16-bits while the input samples are carried with 8-bit precision. Thecomplete loop filter occupies 80 slices.

The DDS was implemented using a simple phase truncation architecture [9]. Using the quantizedSimulink model, a 1024-point sin/cos look-up table (LUT) with 12-bit precision samples was found to beadequate for the application. Using quarter wave symmetry [9], the sin/cos LUT requires only a singleVirtex block RAM (BRAM). The dual-port nature of the BRAM permits both the I and Q samples to becomputed simultaneously. The DDS phase accumulator consists of a 28-bit adder and register. Thesecomponents occupy a modest 14 logic slices.

227

The complete QPSK Costas loop occupies approximately 1000 logic slices.

5 Conclusion

FPGA based signal processors are being employed in a diverse range of signal processing applications forreasons of performance, economics, flexibility and power consumption.The telecommunication industryhas been quick to embrace FPGA technology. Nearly 50% of all FPGA production finds its way intotelecommunications and network equipment of one sort or another - wireless base stations, switches,

228

provided by FPGAs also allows designers to easily track evolving standards like MPEG, and provide amethodology for dealing with fluid standards such as ADSL.

Even though FPGA DSP systems represent a significant faction of the signal processing arena, we arewitnessing an exponential growth in the insertion of FPGAs in DSP hardware. This explosive growth isenhanced by access to FPGA intellectual property (IP) cores from all the major FPGA suppliers as well as3rd-party IP designers. With these resources, the system implementor is able to focus on the design ratherthan the details of lower-level modules like filters and transforms.

The continuing evolution of communication standards and competitive pressure in the market placedictate that communication system architects must start the engineering design and development cyclewhile standards are still in a fluid state. Third and future generation communication infrastructure mustsupport multiple modulation formats and air interface standards. FPGAs provide the flexibility to achievethis goal, while simultaneously providing high levels of performance. The SDR implementation oftraditionally analog and digital hardware functions opens-up new levels of service quality, channel accessflexibility and cost efficiency.

The software in a SDR defines the system personality, but currently, the implementation is often a mix ofanalog hardware, ASICs, FPGAs and DSP software. The rapid uptake of state-of-the-art semi-conductorprocess technology by FPGA manufacturers is opening-up new opportunities for the effective insertion ofFPGAs in the SDR signal conditioning chain. Functions frequently performed by ASICs and DSPprocessors can now be done by configurable logic. This paper has provided an overview of how severalsignal processing functions can be implemented in an FPGA. The DA implementation of very high-performance single-rate and multi-rate filters was described in addition to the FPGA implementation of aQPSK Costas loop for carrier recovery.

References

[1] Peled and B. Liu, “A New Hardware Realization of Digital Filters”, IEEE Trans. on Acoust., Speech,Signal Processing, vol. 22, pp. 456-462, Dec. 1974.

[2] Xilinx Inc., The Programmable Logic Data Book, 1999.

[3] S. A. White, “Applications of Distributed Arithmetic to Digital Signal Processing ”, IEEE ASSPMagazine, Vol. 6(3), pp. 4-19, July 1989.

[4] E. B. Hogenauer, “An Economical Class of Digital Filters for Decimation and Interpolation”, IEEE.Trans. Acoust., Speech Signal Processing, Vol. 29, No. 2, pp. 155-162, April 1981.

[5] B. Sklar, Digital Communications Fundamentals and Applications, Prentice Hall, Englewood Cliffs,New Jersey, 1988.

[6] The Mathworks Inc, Matlab. Getting Started with Matlab, Natick, Massachusetts, U.S.A, 1999.

[7] The Mathworks Inc, Simulink, Dynamic System Simulation for Matlab, Using Simulink, Natick,Massachusetts, U.S.A, 1999.

[8] Xilinx Core Generator System, http://www.xilinx.com/products/logicore/coregen/index.htm

[9] C. H. Dick and f. j. harris, “Direct Digital Synthesis - Some Options for FPGA Implementation”,SPIE International Symposium On Voice Video and Data Communication: ReconfigurableTechnology: FPGAs for Computing and Applications Stream, Boston, MA, USA, pp. 2-10,September 20-21 1999.

Signal Processing Requirements of the TDD Terminal

Stamatis Kourtis, Patrick McAndrew, Phil Tottle

Motorola Semiconductors69 Buckingham street, Aylesbury, HP20 2NJ, UK

{Stamatis.Kourtis, P.Mcandrew, Phil.Tottle}@motorola.com

Abstract: This paper considers the signal processing requirements of 3GPP-TDD terminal and examinesthe technology requirements for the implementation of the 3GPP-TDD mode, this is followed by adiscussion of how the terminal complexity of the TDD mode compares to that of the FDD mode.The first part of the paper begins with a short explanation of how the 3GPP TDD mode fits into the ITUfamily of standards, the specification of the TDD mode covering the frequency bands, the synchronousnature, and the physical channel structure. The differences in the inner receiver implementation between

the TDD mode and the FDD mode are highlighted.The second part of the paper identifies the key signal processing complexity issues that determine theTDD terminal architecture. This section explains the requirement for terminal synchronisation to thenetwork in the synchronous TDD mode and asynchronous FDD mode. The baseband implementationsection briefly covers the inner receiver functions of the TDD mode: channel estimation, active codedetection and multi-user detection. The latter is basically examined from the joint detection point of view;nevertheless other techniques are presented with particular interest in the single detection algorithms.This leads to a brief comparison of the terminal architectures for the FDD and TDD modes, in particularthe suitability of the air interface to flexible implementation on a DSP.

I INTRODUCTION

The imminent arrival of the third generation cellular has resulted in the creation of many different usagescenarios from video conferencing to Internet access in addition to conventional voice traffic. The TDDmode of 3G potentially provides the flexibility and adaptability to support the different requirements oflatency, asymmetric and variable rate traffic in a single terminal architecture and hence offers manyadvantages to service providers, operators, manufacturers and users.

230

II THE ITU FAMILLY

The ITU has identified 5 standards that are part of the IMT2000 family, the five are:IMT-DS :- W-CDMA (UMTS & Japan CDMA, 3GPP-FDD)IMT-MC:- CDMA2000 (North America CDMA, 3GPP2)IMT-TC:- TD-CDMA (TDD version of UMTS, 3GPP-TDD)IMT-SC:- EDGE (EDGE ETSI & UWC-136)IMT-FT:- DECT

IMT-TC or TD-CDMA, the TDD mode of 3GPP, can efficiently support the asymmetric services, e-commerce, web browsing, and email that are expected to drive revenue generation in 3G networks. TD-CDMA has allocated spectrum in Europe and is expected to get more spectrum in the 2005 timeframe,TD-CDMA has low regulatory and organisational barriers to obtaining future spectrum (it does notrequire inflexible paired allocations). TD-CDMA is supported by China, is designed to inter-workseamlessly with W-CDMA and is expected to fit on the Node-B and user equipment (UE) platforms withlittle cost differential. TD-CDMA is still in development, Release 99 of the standard is available now andthere will be a new release at the end of 2000.

III 3GPP-TDD TERMINAL

Although much commonality now exists between the TDD and FDD modes since the alignment within3GPP, several important key differences remain. Since the outer transceiver functions (source andchannel coding-decoding) are identical between the TDD and FDD modes, the following discussionfocuses on the inner transceiver implementation (fading channel mitigation, modem, carriersynchronisation) which employs disparate algorithms for the two different modes of the 3GPP standard.This is addressed within the context of the frame structure, whilst the issues of complexity for thephysical layer algorithms is discussed in a later section.

231

The TD-CDMA mode fundamentally operates in a synchronous manner. This necessitates the use of bothBS-to-network synchronisation and time-slot synchronisation. Since the frequency allocation is within theunpaired spectrum the uplink and downlink are multiplexed in a time division manner. Thus, users aremultiplexed in a time-slot manner similar to that of GSM. However, like W-CDMA, the spectrum ofeach TD-CDMA user is also spread through the use of orthogonal codes. This allows multiple users to beadditionally grouped within a time-slot; each separated in the code domain.

Figure 2 shows the fundamental frame structure used. Each 10 ms frame consists of 15 time-slots, eachspread over the 5 MHz channel bandwidth by a variable spreading factor (SF=1..16). Each time-slotcarries both the data bits and a midamble sequence. The use of a 3.84 Mc/s chip rate allows the target of2 Mb/s data services. Flexibility for support of asymmetric services is provided within a frame level bythe flexible allocation of uplink and downlink time-slots. Either use of multiple spreading factors and/ormulti-code operation or the use of multi-slot operation may accommodate higher data rate users.More specifically, in the downlink where a fixed scrambling code of 16 is used (unless the total slotcapacity is assigned to one user and thus no spreading is used, SF=1), higher data rates are accommodatedonly by multi-code or multi-slot transmission. The latter, however, should be avoided because of thesubstantial increase in the required complexity and the reduction of the available time for measurements.The utilisation of fixed spreading in the downlink was considered as a good solution, which could enablethe use of inexpensive detection schemes at the UE. In this way, only the codes associated with thisspreading factor are assumed as active at the UE receiver, otherwise all the possible codes from theorthogonal variable spreading factor (OVSF) code tree from SF=2..16 would be taken into consideration.Furthermore, in the uplink, higher data rates should be accommodated primarily by reduced spreadingfactors or otherwise by multi-slot operation. Multi-code operation is highly discouraged due to the

232

resultant high peak-to-average-ratio transmitted signal, which causes the inefficient operation of the UEpower amplifier.Since in general the spreading factors used are lower than that of W-CDMA, recovery of a user throughdespreading (correlation of the received sequence against the known user’s code: the basic principal ofspread spectrum CDMA [2]) may limit system capacity due to the higher interference levels. (Actually, itis worth noting here that the scrambling codes of W-CDMA have been especially designed having inmind a Rake structure at the receiver and the respective scrambling codes of TD-CDMA have beenoptimised for a joint detection structure at the receiver). To mitigate these effects, joint detectiontechniques are utilised for multi-user interference cancellation. These techniques require the knowledgeof both other users codes and their corresponding channel impulse responses. In the case of the downlink,since all transmissions are from one base-station, this may simplify to only one channel (unless transmitdiversity is employed). The implementation complexity of joint detection may be high, due to theinversion of large matrices. In order to simplify the receiver design, an alternative technique has beenproposed for the case of stationary or pedestrian speed mobiles. This is termed joint predistortion andbasically works on the assumption that at low speeds channel reciprocity exists for adjacent time-slots.This allows the base station to pre-equalise the downlink channel based upon measurements of the uplink.Although, this increases the complexity at the base station, the mobile receiver may be implemented byuse of a single finger rake structure.

IV POWER CONTROL

The TDD mode of 3GPP differs from the FDD mode in the complexity of the power control that is usedto improve capacity. Power control is used to reduce the interference level due to other users transmittingon the same frequency, in the FDD fast power control is used with measurements made on one time slotsignalled to the base station on the next transmit timeslot as shown in Figure 3.

233

3GPP-TDD is less sensitive to power control than the FDD mode due to the use of joint detection at theNode-B and the terminal. The received level at the base station is maintained at a reference level throughopen loop power control and no power control commands are transmitted by the base station. This ispossible because the mobile station is capable of performing accurate measurements of the path-loss sincethe transmitted power of the beacon channel (channel used by the terminal for measurement purposes) issignalled through the broadcast channel (BCH). In contrast, such measurements can not be performed bythe Node-B as the terminal does not have a relevant beacon function. Consequently, closed loop powercontrol is used in the downlink whereby a terminal makes signal-to-interference-ratio measurements ofthe received dedicated channel, compares this ration to a predetermined threshold and accordingly itsends the power control commands to the Node-B once per frame.

V INNER RECEIVER FUNCTIONALITY

The inner receiver of the 3GPP-TDD terminal basically performs two operations: channel estimation andjoint detection. Figure 4 depicts the flow chart of the inner receiver under two operation modes: normaland transmit diversity (Tx diversity).

234

VI CHANNEL ESTIMATION

Under normal operation, all the codes transmitted from the BS experience the same radio channel andconsequently only single channel estimation is required at the UE, which can be easily performed bymatched filtering. However, if Tx diversity is employed at the BS, the various codes (corresponding todifferent users) experience different radio conditions and as a result a multiple channel estimation has tobe performed, which can be very efficiently undertaken by operating in the frequency domain (using FFTand IFFT).

VII JOINT DETECTION

The most power processing intensive function of the TDD terminal is the joint detection, which largely ischaracterised by the inversion of the matrix:

where (N is the number of symbols within the data block either 61 or 69 (or twice thesenumbers if both the data pans are considered), Q is the spreading factor which equals 16 for the downlink(the case for Q=l is not examined as this does require only channel equalisation), W is the channel lengthin chips and K the number of the active codes) is the channel convolution matrix, and is the conjugatetranspose of A. As a result, which implies that the complexity of this stage depends on Kwhich ranges from 1 to 16, although it is anticipated that no more than 8 to 10 codes would beconcurrently active at any instance. Lastly, it can be seen [3] that the total complexity of the jointdetection generally varies as a function of

235

In contrast to the uplink case, in the downlink the UE does not have any knowledge of the currently activecodes in the received slot, as this information does not be signalled by the network. This would result inan always worst case joint detection processing of K equal to 16, if some signal pre-processing would nottake place, the objective of which is the determination of the active codes.Because of the two different operation modes of the receiver, this pre-processing can be realised in twototally different ways. Whilst in normal operation (only one midamble is used in the downlink), thedetection of the active codes can be performed from the output of the matched filter operation data(r),

as shown in Figure 4, where A’ is the full version of the channel convolution matrix (K=16) and data(r) is

the data pan of the received burst. This information enables the reduction of matrix A’ to the finally usedmatrix A. On the other hand, given the utilisation of multiple midambles for the case of Tx diversity, theoutput of the multiple channel estimation can clearly identify which codes are active and accordingly thematrix A is created.The comparison of these two approaches readily reveals the superiority of the latter both in terms ofcomplexity and performance, and as such an extension of the utilisation of multiple midambles even forthe case of non-Tx diversity is not precluded. At this point it must be mentioned that due to the non-idealautocorrelation properties of the midamble, some degradation in the channel estimation is expected.Several algorithms can be used for the joint detection including the zero forcing block linear equalisation(ZF-BLE), the minimum mean square error block linear equalisation (MMSE-BLE) and theircorresponding decision feedback realisations [4]. Comparatively, MMSE-BLE is reported to have slightlybetter performance up to 0.5 dB than ZF-BLE, requiring, however, an estimate of the noise power(indicated as other inputs in Figure 4). In addition, a significant performance improvement can beachieved by the use of decision feedback equalisation schemes, although in bad channel conditions theunavoidable error presence causes extensive error propagation, which results in rather poor performance.Alternatively, other non-joint-detection techniques include interference cancellation and channelequalisation algorithms as described in [5]. These techniques generally are lesser computationaldemanding, albeit at a substantially worse error performance.

VIII SINGLE DETECTION ALGORITHMS

The desire to eliminate the pre-processing stage for the determination of the active codes has led to theinvestigation of sub-optimum single detection schemes [6],[7]. In general, the proposed algorithmsperform worse than their joint detection (optimum) counterparts, although the performance differencesare reported to be up to 1 dB. Interestingly, the detection algorithm proposed in [7] introduces the concept

236

of “spectral matching” and all the needed signal processing for the detection is performed in thefrequency domain (FFT and IFFT operations) without requiring any matrix inversion. In addition, thisalgorithm apart from not requiring the knowledge of the other active codes, it does not even need achannel estimation, thereby reducing substantially the required complexity at the UE receiver.

IX RADIO ARCHITECTURE

The radio implementation requires the development of integrated RF transceiver circuits which meet thelinearity and bandwidth needs of the different air interfaces or the re-use of existing components in a“velcro” phone architecture. Current RF transceiver architectures typically comprise several devices ofdifferent technologies chosen to optimally implement the power amplifier (PA), the low noise amplifier(LNA), synthesiser and filtering operations. Package parasitics and thermal dissipation also constrain theintegration options.

237

Multiband radio implementations demand higher PA linearity over a wider bandwidth, higher resolutionwider-band synthesisers, and wide bandwidth LNA, each requiring novel implementation techniques,demanding higher semiconductor performance and further constraining integration prospects [1].The Figure 5 depicts a radio architecture for a TDD terminal that operates at low data rates (144 kb/s),this terminal type could be used in unlicensed spectrum as a medium data rate terminal that is capable ofalso operating in a GSM mode. The shaded blocks represent the RF & IF blocks that need to be added toa GSM terminal to permit it to operate in TDD mode. The two receiver chains of unshaded blocks arerequired for the two GSM frequency bands, the third receiver stage of band pass filter, followed by a lownoise amplifier and mixer is needed for the unlicensed frequency band. The data converter bandwidthsmust be increased to cope with the 5 MHz bandwidth of TD-CDMA. The significant advantage of thismedium data rate terminal is that it may operate in unpaired spectrum and requires a lower power budgetthan a W-CDMA teminal since an antenna switch may be used, as there is no simultaneous transmissionand reception.A W-CDMA-TD-CDMA radio architecture that also supports GSM is shown in Figure 6. The shadedblocks represent the RF and IF blocks that need to be added to a GSM terminal to permit it to operate inW-CDMA mode. A duplexor is required where there is simultaneous reception and transmission from thesame antenna, the duplexor has losses (3dB receive, l.5dB transmit) that are large compared to a switchthat can be used in used in GSM. If slotted mode cannot be used for handover to GSM frequencies anadditional IF receiver chain and A-D are needed plus an additional synthesiser. One additional transmitsynthesiser may be needed because of the variable duplex distance in the FDD frequency bands. Also a

238

circulator may be required in the transmit chain depending on the implementation to allow small powercontrol steps needed in the UMTS FDD standard. In the baseband, the processor must operate on W-CDMA and GSM data resulting in a peak load on the baseband processor that is higher than that of aGSM terminal alone.

X CONCLUSIONS

This paper presented the TDD mode of the 3GPP, and in particular it showed the key importantdifferences with the FDD mode. From the terminal point of view, these differences are primarily focusedon the inner receiver of the terminal where functions such as channel estimation and multi-user detectionsignificantly impact its total complexity. Lastly, radio terminal architectures have been presented whichwill allow the production of low cost TDD terminals.

REFERENCES

[ 1 ] Sheng S. and Brodersen R., 1998, “Low-Power CMOS wireless communications. A widebandCDMA system design”, Kluwer Academic Publishers[2] Peterson R. L, Ziemer R. E. and Borth D. E., 1995, “Introduction to Spread SpectrumCommunications”, Prentice-Hall[3] Kourtis S., McAndrew P. and Tottle P., 1999, “Baseband Complexity of Software Defined Radiofor 2nd & 3rd Generation Air Interfaces”, 4th ACTS Mobile Communications Summit. Vol. 2, pp. 727-732[4] Klein A., Kaleh G. K. and Baier P. W., 1996, “Zero forcing and minimum mean-square-errorequalization for multiuser detection in code-division multiple-access channels”, IEEE Transactions onVehicular Technology, Vol. 45, No. 2, pp. 276-287[5] Steiner B., 1997, “Interference cancellation vs. channel equalization and joint detection for thedownlink of C/TDMA mobile radio concepts”. Proceedings of EPMCC Conference together with 3. ITG-Fachberichte, No. 145, pp. 253-260[6] Klein A., 1997, “Data detection algorithms specially designed for the downlink of CDMA mobileradio systems”, 47th IEEE Vehicular Technology Conference, Vol. 1, pp. 203-207[7] ETSI TDoc SMG2 UMTS L1 362/98, 1998, “Low cost MMSE-BLE-SD algorithm for UTRATDD mode downlink”

Frame Quality-Based Versus Based Forward PowerControl Methods for the cdma2000 Third Generation Standard

Steven P. Nicoloso, Mike Mettke, and R. Michael BuehrerBell Laboratories - Lucent Technologies

67 Whippany Rd. Room 2A-277Whippany, NJ 07981

spn@lucent.com

Abstract Fast forward link power control is one of several improvements offered by thirdgeneration standard cdma2000 over the well-known second generation IS-95. This paper describeseach of the five forward power control methods allowed in cdma2000, wherein the slow frame quality-based algorithms are similar to those of IS-95. We examine the performance of frame quality-basedversus based (fast) power control for a typical voice traffic channel and a high data ratetraffic under various geometries and a wide range of mobile station velocities. We also consider theimpact of Space-Time Spreading (STS) transmit diversity along with power control. Forward linkpower control proves itself to be a worthwhile addition to CDMA, providing substantial gains incapacity where they are needed most, viz., at low speeds where FEC interleaving is weaker. Theeffect of time-correlated shadowing is also considered and experimental results suggest that actuallymay show additional gains hidden by conventional static (Rayleigh fading only) experiments.

1 IntroductionThe third generation (3G) cdma2000 standard [1] provides backwards compatibility with IS-95 [2],but also adds substantial evolutionary technical improvements that will increase system capacityand enhance the reliability of service even during the first (1X) phase of deployment1. Technicalimprovements include turbo codes for high data rate channels, the use of pilot-assisted BPSKmodulation on the reverse link, and, for the forward link, transmit diversity, lower traffic coderates, and the addition of fast (up to 800 bps) power control. A realistic examination of theimprovements provided by fast forward link power control is the primary focus of this work.

We begin in section 2 by describing the various forward power control (FPC) methods providedby cdma2000. Then in section 3, we describe the experimental environment in which the variousFPC methods are studied. Section 4 discusses the trends we see in our simulation study. Finallyin section 5, we summarize our findings. With very few exceptions we find that fast basedpower control provides significant gains for the CDMA forward link and most importantly providescapacity gains where they are needed most.

2 Forward Power Control Methods in cdma2000Forward link power control is accomplished in cdma2000 by puncturing the Reverse Pilot Channelwith the Reverse Power Control Subchannel (RPCSC)2. One quarter (384 of 1536 chips at the IXchip rate of 1.2288 Mcps) of the pilot is punctured with power control information. The standardallows for five different methods of forward link power control, defined by stored mobile stationparameter FPC_MODEs We shall alternatively denote each of these viaFPC Modes 0 through 4. Modes 0 through 2 are based on receiver measurements and aredescribed in section 2.1. Modes 3 and 4 are based upon estimates of frame quality (usually whethera frame error has occurred) and are quite similar to each other. These are outlined in section 2.2.

1The IX portion of cdma2000 uses the same 1.2288 Mcps chip rate as IS-95.2This applies to the 3G proper portions of cdma2000, meaning radio configurations 3 and above. In the legacy

IS-95 portions of cdma2000, FPC remains unchanged.

240

2.1 Description of Based Power ControlSimply viewed, FPC Modes 0 through 2 are variations of based power control methodssimilar to those used on the IS-95 reverse link where during each 1.25 ms power control group(PCG), the ratio is estimated by the receiver. The estimate, is then comparedwith an internally saved threshold setpoint). If is less than the setpoint,the power control bit is set to zero during the next PCG. Otherwise, it is set to one. The basestation interprets a zero as a request for more transmit energy, and a one as a request for less. Thistype of based power control is often called “fast power control” since it operates at a rate ofup to 800 Hz and can in some circumstances actually invert the fading signal envelope seen by thereceiver. What is different between forward power control in cdma2000 and reverse power controlboth in cdma2000 and IS-95 is that for reverse power control methods, the mobile is required tocomply with power control commands received from the base station. In the forward link case,however, there is no such requirement for the base station to obey mobile commands. This is duein part to the relatively tight dynamic range constraints that have to be placed on the portion ofthe total sector transmit power given to forward traffic channels.

In based FPC Mode 0, all power control bits transmitted are based on measurementsof one and only one forward traffic channel, either the Forward Fundamental (FFCH) or ForwardDedicated Control Channel (FDCCH). FFCHs are typically variable rate voice channels with max-imum rates of 9.6 kbps in Radio Configurations 3 and 4 (RC3 and RC4) and 14.4 kbps in RadioConfiguration 5 (RC5). Similarly FDCCHs have rates of 9.6 kbps in RCs 3 through 5 and 14.4kbps in RC5 only. Forward Supplemental channels (FSCHs), having rates as high as 307.2 kbps incdma2000 IX (1036.8 kbps in cdma2000 3X), are provided no power control feedback mechanismin FPC Mode 0. This does not imply that FSCHs are not allowed in FPC Mode 0, but merelysuggests that control of FSCH transmit power is left to base station designers.

Direct feedback power control of (potentially high data rate) FSCH is provided by basedFPC Modes 1 and 2. In each, the 800 Hz reverse power control subchannel is divided up intoprimary and secondary portions. The primary power control subchannel transmits power controlbits based on measurements of either a FFCH or FDCCH. The secondary transmits powercontrol bits based on measurements of a FSCH. In FPC Mode 1, commands are alternated suchthat the effective rate of power control for each of the channels is 400 bps. In FPC Mode 2, thedivision is unequal. During each 20 ms epoch of 16 PCGs numbered 0 through 15, power controlbits sent during PCGs 1, 5, 9, and 13 are based on estimated of the FFCH/FDCCH. Theother 12 are reserved for measurements of the FSCH. This results in an effective 200 bps powercontrol stream for the low data rate channel and 600 bps for the (usually) higher data rate FSCH.

What we have described thus far in this section is only the “inner loop” portion of basedpower control. What happens with the setpoint is referred to as the “outer loop.” Itsbehavior is illustrated in Figure 1. After every frame, the mobile receiver determines whether itmade an error in decoding or not via cyclic redundancy check (CRC) or some other method. If theframe is in error, the mobile may raise its setpoint by some value. If the frame is not in error, themobile may lower its setpoint by some (usually smaller) value. In this manner, a target FER maybe maintained. If, for example, a y% FER is desired, the mobile should raise its setpoint by +AdB when it records an error. The figure illustrates the behavior for a 1% target FER. Down stepa is given by

Over time, this raising and lowering of setpoint is intended to result in the desired target FER. Theonly restriction on the mobile is that it may not raise its setpoint above mobile stored parameter

or reduce it below where chan {FCH, DCCH, SCH} forforward fundamental, dedicated control, and supplemental channels respectively. Maintenance of

241

a desired FER is crucial to the capacity of virtually any CDMA system since one mobile station’ssignal of interest is interference for the rest. Furthermore, power given on the forward link to oneuser’s signal is power that cannot be given to another. Too low an FER is seen as a waste of basestation transmit power and (at least in cdma2000) constitutes a direct drain on capacity.

2.2 Description of Frame Quality-Based Power ControlFPC Modes 3 and 4 are not based on measurements of in the mobile receiver, but ratheron measures of frame quality. Each of 16 power control bits per 20 ms epoch on the Reverse PowerControl Subchannel are all set to the Erasure Indicator Bit (EIB) or Quality Indicator Bit (QIB)for Modes 3 and 4 respectively. Thus frame quality-based power control methods in cdma2000 workat an effective rate of 50 bps. Although this is much slower than the “fast” (800 bps) power controlmethods described in section 2.1, we point out that the reliability of each bit is much higher. TheEIB is used specifically to indicate frame erasures (unrecoverable frame errors) detected for a FFCHor FDCCH. The QIB is a somewhat relaxed version of the EIB, as it defined merely as indicationof detected signal quality on a FFCH or FDCCH. The cdma2000 standard does mandate, however,that the QIB shall be equal to the EIB when a FFCH is present. For the purposes of this work,QIB and EIB-based power control are considered equivalent.

There are no specific requirements in cdma2000 instructing the base station in how to respondto quality indicators transmitted on the reverse power control subchannel. What is envisionedhowever is something very much like the “outer loop” power control described in section 2.1. Inthis case however, instead of an setpoint (or threshold) kept by the receiver, there is a targettransmitted energy per chip kept by the transmitter. is updated either up (A dB)or down (a dB) depending upon the quality metric received during a frame. If

(i.e., it is within the dynamic range allowed for any one signal), then transmitted energyper chip is equal to Otherwise, the minimum or maximum energy is transmitted.

242

Finally, we point out that in a manner similar to that of FPC Mode 0, there is no reportingof frame quality for forward supplemental channels with FPC Modes 3 and 4 as the reportingmechanisms are solely for FFCHs or FDCCHs. It is up therefore up to the base station designers todetermine how quality of service wilt be acheived for FSCHs when using these frame quality-basedforward power control methods. For example, base station designers might determine the ratiopower between a FSCH and FFCH such that the FSCH may be properly power controlled basedsolely on feedback on the FFCH. An idealized version of just this assumption is used below in thedescription of our simulation in section 3.

3 Simulation MethodologyCreation of simulation to accurately estimate forward system capacity required great effort on thepart of far more engineers than are named in this paper. In creating our simulation, we attemptedwherever possible to make no simplifying assumptions. In this section we describe the simulation ofthe transmitted traffic and pilot channels, the mobile receiver, the implementation of the candidateforward power control methods, and finally the simulated channel conditions.

3.1 Base Station Transmitter

In this study of the FPC methods available in cdma2000, we consider the performance of two forwardtraffic channels: the 9.6 kbps FFCH channel and the 153.6 kbps FSCH under Radio Configuration3 (RC3). Therein, rate r = 1/4 forward error correction (FEC) coding is used. The convolutionalencoder used for the FFCH and the FSCH has a constraint length of K = 9. For the FSCH, turbocoding is examined along side convolutional. The turbo encoder consists of two K = 4 constituentencoders, the outputs of which are interleaved and punctured according to the cdma2000 standard.Two out of every 24 interleaved code symbols of the FFCH are pseudo-randomly punctured with theForward Power Control Subchannel (carrying reverse link power control information) as requiredby the standard. No such puncturing is done on the FSCH. Walsh function lengths are 64 and 4respectively for the FFCH and FSCH. We note that the Walsh function length is the spreadingfactor applied after code symbol interleaving. The total spreading gain therefore applied to eachchannel, i.e., the chip rate over the data rate is 256 and 16 respectively.

Both single antenna transmission (no transmit diversity) and Space-Time Spreading (STS) areexamined. STS is an open loop transmit diversity scheme that uses two widely spaced transmitantennas each transmitting a mutually orthogonal version of the same code symbols. Its designand performance are detailed in [3] and [4]. Pilot energies per chip are and

for diversity antennas 0 and 1 respectively. is the maximum sustainableaverage base station transmit power spectral density per sector, i.e., that transmit sector energyrequired for fully loaded system. When simulating non-diversity mode, pilot 1 is not used. In STS,pilot 0 is given a higher energy than pilot 1 due to the expected need of supporting legacy secondgeneration mobiles.

Dynamic range assumptions for transmitted traffic power are that andThis means that the base station may devote up to one-half of its total

power to one traffic channel. It is not clear that in practice base stations would be able to supportsuch a high upper limit. Similarly, it is not clear that in practice such a low lower limitwould be used. We choose a large dynamic range for this study merely to provide as much “y-axis”data as possible. For FSCH simulations, we assume perfect signaling layer and therefore do notsimulate the FFCH or FDCCH along side of it as this requirement has no bearing on the FSCH’sperformance.

243

3.2 Mobile Station ReceiverFor the FFCH (low data rate voice channel) we compare the performance of based FPCMode 0 and QIB-based FPC Mode 4. The target FER in both cases is 5%. For FPC Mode 0, theouter loop step size is 1 dB (updated at 50 Hz) and the inner loop step (operating at 800 Hz) is0.5 dB. Thus when a frame error is detected in the mobile receiver, the setpoint is raised byA = 1 dB. When no frame error is detected, the setpoint is reduced by asgiven by (1). For Mode 4, we allow a wider transmit step size of 3 dB since it is the only controlmechanism. Mobile station estimation for the FFCH is based solely on coherent estimation ofthe bits on the Forward Power Control Subchannel. QIBs in FPC Mode 4 are based on the successor failure of the 12 bit CRC as is the decision to move the setpoint up or down in FPCMode 0. No restrictions are placed upon the value of the mobile setpoint in our simulation.

For the FSCH (high data rate channel) we compare the performance of based FPC Mode2 (600 bps) and QIB-based FPC Mode 4 in a somewhat idealized manner. For Mode 2, due to thedifficulty of estimating non-coherently3 on FSCHs at high speeds, we make the assumptionthat the the mobile receiver has genie-knowledge of its received In a similarly idealizedmanner, we implemented FPC Mode 4 for the FSCH by assuming that the base station transmitteradjusts its power according to the success or failure of decoded FSCH frames. In practice, theseadjusments would be based on the CRC of the FFCH or FDCCH.

All other mobile receiver functionality is simulated in a realizable manner. Receiver channelestimation is performed via moving average filter of the received common pilot channel(s). Channelestimates are then used to create decoder metrics via maximal ratio combining (MRC). For turbodecoding, an additional scale factor of

is required, where is the transmit energy per traffic chip, is the sum of pilot transmit energiesper chip, and is the double-sided spectral density of received noise plus interference. Each of thecomponents for the turbo scaling parameter is estimated by the mobile on a frame-by-frame basis.Finally, for the based FPC methods, we assume a uniform 5% power control bit error rateon the Reverse Power Control Subchannel. Since the frame quality-based methods have a muchhigher reliability, we assume error free feedback of frame erasures.

3.3 Propagation ChannelChannel conditions in our experiments were assumed to be single-path Rayleigh fading. In cases oftransmit diversity, fully independent fading is applied to the transmitted signals from each antenna.Mobile station velocities range from 0.1 km/h to 300 km/h in half-decade steps. Time-correlatedRayleigh fading samples are taken from a standard uniformly spaced Jakes’ oscillator model. Acenter carrier frequency of 1960 MHz is used.

For shadowing experiments, we assume log-normal shadow fading with a standard deviation,The shadow fading is filtered log-domain noise tied to mobile velocity. Log-domain

samples of the shadow fading envelope are given by the recursion found in [5]

where and

3Recall that then is no forward power control puncturing for FSCHs and therefore no possibility of doing coherentestimation. The issue of estimation for high data rate channels in cdma2000 is a rich topic for

investigation in and of itself..

244

where is the sample period (20 ms), is mobile velocity in m/s, and is the correlation distancefor which we assumed 504 m.

Interference at the mobile is controlled by what is known as “geometry”. Geometry is denned bythe ratio of spectral densities of the received sector signal of interest (containing transmitted

traffic energy per chip and other orthogonal signals such as the pilot) and other cell (non-orthogonal) interference plus noise. Non-orthogonal interference is simulated as additive whiteGaussian noise and is assumed to dominate the receiver noise floor (i.e., Geometries of0 and 6 dB were examined for the FFCH. Slightly higher geometries of 3 and 9 dB were used forthe FSCH.

3.4 Figures of MeritForward link CDMA system capacity is most clearly observed by consideration of or theratio of transmit sector energy per chip for one user’s channel to the total signal energy in thatsector. As mentioned in section 3.1, may be seen as the maximum sustainable average outputpower per base station sector. Received (a familiar receiver figure of merit) may thusbe derived from relationship between and However, since forward link capacityis determined by base station transmit power, transmitter figure of merit for a specificgeometry and achieved FER is used primarily in this work.

4 Analysis of ResultsWe begin our analysis by examining the behavior of FPC Modes 0 and 4 for the 9.6 kbps FFCH inFigure 2. Average transmit is plotted versus mobile velocity ranging from 0.1 km/h to 300km/h in half-decade steps. The geometry is with no shadow fading. Results withsingle antenna transmission as well as Space-Time Spreading (STS) are shown. The propagationchannel is single-path Rayleigh fading. For STS, we assume zero correlation between transmitantennas. Target FER for both FPC modes is 5%, and we assume a uniform error rate of 5%individual power control bits for Mode 0 fast power control. We note that the target FER (printedin tiny numbers by each point) was achieved for all cases. The largest gains with fast (Mode 0)power control occur in non-diversity mode at lower speeds. In some cases (e.g., 0.3 and 1.0 km/h)this improvement is around 5 dB. STS alone yields dramatic gains at lower speeds. Thus, theincremental gains seen from fast power control are less dramatic but still substantial. At 1 km/h,for example, FPC Mode 0 power control achieves 3 dB lower We see that in all casesthe quality of the link improves as velocity gets above about 3 km/h which is attributable to theinterleaver’s increased ability to span nulls in the Rayleigh fading envelope. Here the performancedifferences between fast and slow power control begin to disappear.

We see in Figure 2 an actual crossover between 40 and 50 km/h where frame quality-based FPCactually begins to works slightly better. This crossover behavior is quite interesting and meritsdiscussion. As noted above, at high mobile velocities, forward link cdma2000 receiver performancecan improve dramatically depending upon diversity mode. This improvement is due to severalfactors. First and foremost, as speeds get higher the interleaver more perfectly randomizes theobserved fading events from the FEC decoder’s point of view. In addition, channel estimationin fast fading is only slightly degraded since a very strong common pilot channel (typically muchstronger than the low data rate traffic) allows for very short integration periods. For example,Figure 2 shows little or no degradation in performance in moving from 100 km/h to 300 km/hvelocities. In just such circumstances we begin to see the limitations of fast based powercontrol. It simply cannot invert the fast fading envelope and the mechanism that gives it the abilitysave so much transmit power in slow to moderating fading actually begin to cause a slight waste

245

of power.In FPC Modes 0, 1, and 2, there is an irreducible delay between the time the mobile

sends its power control bit and the time that it receives additional power in its signal. In otherwords, by the time the mobile receives the requested change in power, the coherence time of thechannel has potentially passed. This “drawback” is compounded by realities such as imperfectestimation of and relatively high error rates on the feedback. Additionally, since the basestation transmit traffic power is either raised or lowered each PCG and never kept constant (aslong as is between its upper and lower limits), fast FPC induces additional variation andthus added mismatch error in transmit power.

On the other hand, slow frame quality-based power control manages to avoid some of the pitfallsencountered by fast power control at even faster speeds. What is desired ideally in such fast fadingconditions is for the transmitter to make no attempt at inverting the channel. The receiver will doquite well (given a well designed interleaver and reasonably strong pilot) if the transmit power isleft constant. This is very nearly what slow power control provides. Its adjustments are much lessfrequent, and, except in the case of a frame error, much smaller. This is why slow power control isshown to perform slightly better than fast power control at high speeds.

However, we should note that the very features of fast FPC which cause the slight degradationrelative to slow FPC at high speeds save the base station a large amount of power at low speeds.This corresponds to an equally large increase in system capacity exactly where it is most needed.At high speeds, such a small is required anyway that the crossover is practically irrelevantfor most cell sites from a system capacity standpoint. Finally, we point out that when STS isemployed in conjunction with fast power control, the mobile receiver has very low sensitivity tovelocity. In Figure 2 little more than 2 dB separate the very best from very worse performanceover velocities spanning orders of magnitude.

246

We now turn our attention to the performance of a FSCH operating at 153.6 kbps in a 9 dBgeometry, shown in Figure 3. Simulation conditions are similar to those of Figure 2, except thatthe Reverse Power Control Subchannel provides feedback at 600 bps (FPC Mode 2) and the FECturbo coded. We note first that the gain provided by fast power control is somewhat smaller thanthat seen in for the FFCH, peaking at around 4 dB for mobile speeds of 1 and 3 km/h for bothnon-diversity and STS. As expected, the high data rate FSCH requires a much larger thandid the FFCH. The spreading factor is 16 times smaller. For a non-diversity, FPC Mode 4 system,this high data rate channel requires an of up to -4 dB (around 40% of total base stationsector transmit power) even at this favorable geometry. However, STS applied in conjunction withfast power control alleviates the situation by bringing down to a more desirable -12 dB. Thegain of fast power control versus Mode 4 power control at very slow mobile speeds is somewhatsmaller for the high data rate FSCH than for the FFCH shown in Figure 2 for non diversity mode.At very low speeds, even slow power control is able to track Rayleigh fading. STS gains roughly 4dB over non diversity, especially for mobile speeds below 3 km/h. For very low mobile speeds, wesee a slight increase of with decreasing mobile velocity. The channel estimation averagingperiod in the mobile (we use a moving average filter with a window size of 0.26 ms) is muchshorter than Rayleigh fading period at low speeds. This is a mobile design choice, and behavior inthis region could be improved at the expense of high speed behavior.

Figure 4 shows the results from an experiment with shadow fading. Forward traffic is the 9.6kbps FFCH described above with mobile velocity of 10 km/h with single antenna transmission. Theshadow fading can be seen as inducing changes in geometry during the life of a simulation, a deltagiven by (3) relative to the mean geometry of 6 dB. In this case we ran 20000 FFCH frames of 20ms duration. During each frame the average was measured and plotted as a point relativeto its correspondng effective geometry. To avoid an overburdened figure, only one out of every

247

hundred measurements are plotted. The lines represent a best (log-log) fit to the data. Here we seean consistent 1.5 to 2 dB advantage for fast (FPC Mode 0) power control over slow (FPC Mode4) power control. This is slightly more than the 1 dB advantage observed in Figure 2 at 10 km/h.Although shadowing experiments are much more difficult to conduct via Monte Carlo simulation4,this result is suggestive that fast power control may give even more gains than observed via staticfading (fast fading only) simulation experiments.

5 ConclusionsThe fast forward link power control methods provided by the 3G standard cdma2000 form a veryimportant part of the set of improvements designed to increase CDMA system capacity over IS-95.In this work, we compared the transmit power requirements of the RC3 9.6 kbps FFCH and 153.6kbps FSCH with fast (FPC Mode 0 or 2 respectively) power control and slow (FPC Mode 4) powercontrol. This effectively isolates the differences in power control from other improvements (such aslower rate FEC) provided by cdma2000 relative to IS-95. Therefore, for all results, 3G modes werecompared with other 3G modes. Almost without exception fast based power control providessignificant gains over slow frame quality-based methods. This is especially dramatic at low speedswhere based FPC provides 2 to 5 dB gain. Moreover, it is just at these speeds where theimprovement is needed most. At very high speeds, where link performance is quite good anyway,fast power control is shown to suffer minor degradation relative to slow. Experiments performed

4The forgetting factor suggested by (4) rapidly approaches unity for even moderate speeds. This creates a problemof sampling the random shadow process. Here, all simulation runs were conducted with exactly the same shadowfading samples.

248

with time-correlated shadow fading are suggestive of even greater gains. Quantifying such gainsvia Monte Carlo simulation, however, is a difficult process and is left to future work. Finally, wenote that, combined with transmit diversity (STS), the gains of fast FPC over slow power controlwith no diversity, are even more dramatic, promising equally dramatic capacity improvements forcdma2000 systems.

AcknowledgementsThe authors of this paper wish to acknowledge the support, assistance, guidance, and stimulatingwitty banter provided by our colleagues at Lucent Technologies-Bell Labs. These include RogerBenning, Robert Soni, Dirck Uptegrove, Stephen Allpress, Francis Domininque, and Hongwei Kong.

References[1] TR45.5, Physical Layer Standard for cdma2000 Spread Spectrum Systems. TIA/EIA/IS-2000.2,

1999. (Ballot Version).

[2] EIA/TIA/IS-95A (Electronic Industries Assocation/Telecommunications Industry Associa-tion/Interim Standard 95-A), Mobile Station-Base Station Compatibility Standards for Dual-Mode Wideband Spread Spectrum Cellular System. March, 1995.

[3] R. Soni, R. Buehrer, and J.-A. Tsai, “Open-loop transmit diversity methods in IS-2000 systems,”in Proceedings of the Asilomar Conference on Signals, Systems and Computers, October 1999.

[4] R. M. Buehrer, R. A. Soni, and Q. Li, “Transmit diversity with more than two antennas,” inProceedings of the 10th Annual Virginia Tech Symposium on Wireless Personal Communica-tions, (Blacksburg, VA), June 2000.

[5] M. Gudmundson, “Correlation model for shadow fading in mobile radio systems,” ElectronicsLetters, vol. 27, no. 23, pp. 2145–2146, 1991.

Bluetooth: A Short TutorialMax Robert

Mobile and Portable Radio Research GroupBradley Department of Electrical and Computer Engineering

Virginia Techprobert@vt.edu

IntroductionBluetooth, a standard defining very short-range wireless communications, has been a

topic of considerable interest in the telecommunications industry since its release in the secondquarter of 1999. There has been considerable hype behind the initial release of the standard,accompanied by high expectations for the performance of the devices.

Bluetooth is named after a -century Viking king known for his success in unitingDenmark and Norway during his rule around 960 AD. Just as King Harald Bluetooth is knownfor uniting different people, Bluetooth-enabled devices promise to unite different informationdevices into a single information infrastructure.

Many publications about Bluetooth tout the revolutionary lifestyle Bluetooth will enablewhile leaving technically oriented readers somewhat at a loss as to its true technical capabilities.This short tutorial is intended to provide the reader with a technical overview of all the layersdefined in the Bluetooth standard.

Initially, a general technical description of the Bluetooth standard with details derivedfrom [1] is presented to the reader. The tutorials begins with a description of Bluetooth’sapplication potential and the general environment in which it can operate; this explanation isfollowed by a description of Bluetooth’s role in the protocol stack, its software interface, the RFspecification, the baseband radio specification, the logical link control, the services running overBluetooth, and Bluetooth’s interoperability with other communications standards. Unlessotherwise stated, all technical information presented in this tutorial was collected from [1].

Bluetooth OverviewBluetooth’s origin and short-term market potential must be considered. Bluetooth

emerged through the efforts of a SIG (Special Interest Group) comprised of companies likeEricsson and Motorola. The Bluetooth SIG in the beginning of the year 2000 was comprised ofmore than 1400 members [2]. The expected semiconductor market potential for Bluetooth isexpected to top $3 billion by 2005 [3], while the production of devices is expected to reach 200million by 2003. This market estimate does not take into account revenue from new marketsderived from new Bluetooth-enabled applications.

The Bluetooth specification describes radio devices designed to operate over very shortranges – on the order of 10 meters. The original intent of these links was as a replacement tocables connecting different information devices. The goals of the specification were to describe adevice that is simple and robust, consumes little power, and is very inexpensive to produce, witha target price on the order of five dollars per device.

Bluetooth Capabilities

To cover most home and office short-range applications, a Bluetooth device must supporttwo general types of information: voice and data. The necessary link requirements for voice aredifferent from those for data. While voice can tolerate a certain number of errors in a link, it is

250

highly susceptible to timing errors, including latency and jitter. In order to support voice traffic,Bluetooth needs to support a data stream on the order of tens of kilobits per second without thestrict requirement that the received data be error-free. On the other hand, data transfers, whilegenerally lacking the strict timing requirements posed by voice streams, require information to benearly error-free. Thus, Bluetooth needs the ability to use a strong error-recovery scheme, thoughtiming requirements may be loose.

Given the short ranges involved and the types of links expected, the number of userscapable of interfering with one another is relatively small. While a cellular system may berequired to support thousands of users, a 10-meter network will only support a handful of devices.Bluetooth has been designed with a multiple-access scheme that, while only supporting a limitednumber of devices, greatly simplifies the synchronization task required to support a single cell(known within the Bluetooth specifications as a “piconet”).

Bluetooth Applications

Although the list of applications is limited only by a designer’s imagination, theBluetooth SIG suggests five applications that provide a good illustration of the capabilities of thestandard [4]: a three-in-one phone, an Internet bridge, an interactive conference, a headset, and anautomatic synchronizer.

The three-in-one phone is a phone that can operate over a fixed-line phone line whenwithin range, a mobile phone when outside the home, or as a walkie-talkie with anotherBluetooth-enabled device when within range.

The Internet bridge example allows a mobile computer to interact with another devicewithin Bluetooth range. The device the computer is interacting with has access to the Internet,whether through a fixed line or a mobile phone.

The interactive conference example allows the sharing of documents among severalcomputers during a live conference. Bluetooth-enabled machines can interact with each otherregardless of the available information infrastructure.

A Bluetooth-enabled headset can connect to any Bluetooth-enabled device that requiresvoice input or can provide sound, such as a wireline phone, mobile phone, or a music player.

An automatic synchronizer is an application that allows multiple devices, such as desktopcomputers, laptops, PDAs, and/or mobile phones to synchronize with each other such thatappointments and contact information available in the different devices matches.

Beyond these five, the list of potential applications using this standard is practicallylimitless. Bluetooth has the ability to simplify the wireless connection between two devices to alevel of complexity similar to that required today when connecting two devices using a simplewire.

Technical Overview

Position in the Protocol Stack

The Bluetooth specification covers details of the physical and data layers of thecommunications link. It should be noted that the strict partitioning of the different layers of thetypical protocol stack defined by the OSI model is losing its significance in wirelessimplementations, since it is sometimes desired for the application to the underlying layers inorder to improve performance.

Figure 1 is a diagram of the layers described by the Bluetooth specifications.

251

Software/Interface StackBluetooth devices are operated in a host, such as a PC, and communicate with that host

through a physical bus. Figure 2 described this configuration.

252

The Bluetooth Host is the machine that contains the application, which resides in somesort of computer. The host contains the Host Controller Interface driver, which provides the APIsthat allow the communication between the Bluetooth Host and Bluetooth Device; the host alsocontains the Physical Bus driver. This driver supports the operation of the data connectionbetween the host and the Bluetooth device.

To provide the connection between the host and the Bluetooth device, the Physical Bus isused. The only software that resides in the bus is the firmware necessary to operate the bus.

The Bluetooth Hardware is the device that contains the Bluetooth communicationshardware; its software contains the Host Controller Interface firmware, providing the connectivitybetween the Bluetooth device and the host. In the Bluetooth Hardware also resides the sofwarefor the Baseband controller, allowing the implementation of the baseband radio, which isexplained in the following sections.

Host Controller Interface

The Host Controller Interface (HCI) provides connectivity between the application andthe Bluetooth device; it acts as a uniform interface to access all Bluetooth hardware capabilities.To achieve this goal, the HCI contains a set of commands for the hardware, a handle to possibleevents, and access to errors codes. Included in the HCI are commands such asCreate_Connection and Disconnect, events such as Encryption Change Event and CommandComplete Event, and errors handles such as Hardware Failure and Host Timeout. The HCI is anouter layer around the connection between the device and the host – this layer operates over atransport layer providing the actual connection between the Bluetooth device and the host.

Transport Layer

As mentioned in the previous section, the transport layer is the layer between hostcontroller driver and the host controller; an example of the transport layer is a PC card. The maingoal of this layer is transparency: the host controller driver does not care what the transportmethod is. This transparency allows upgrades to the HCI or host controller without changes tothe transport layer. The transport can be performed over USB, RS232, or UART. USB(Universal Serial Bus) is a serial bus that is typically available in computers such as latptops, theRS232 serial bus is a common bus format that has been standard equipment of computers forseveral years, and UART (Universal Asynchronous Receiver/Transmitter) is a generic serial busdescription that has been available for several years; UART can be set such that it conforms toRS232 specifications.

The previous sections described the connection between a host and the Bluetooth deviceand provided some general information on how commands are passed to the Bluetooth device.The following sections describe the baseband device, its RF specification, and service protocolsthat can be used by the application to easily access and control the Bluetooth device.

Physical layer/RF Specification

Bluetooth devices operate in the ISM (Industrial, Scientific, and Medical) band at 2.4GHz. While the ISM band has been allocated throughout the world, its specific position iscountry-dependent. However, the United States and most of Europe have allocated the spacebetween 2400 and 2483.5 MHz for this band. Bluetooth devices use 79 channels within thisband, each occupying 1 MHz; Japan, Spain, and France have enough bandwidth allocated foronly 23 channels.

253

Bluetooth uses a frequency-hopping, TDD scheme for each channel. The master devicedetermines the frequency-hopping scheme, whose nominal hopping rate is 1600 hops per second,and also sets the piconet clock. In order to guarantee a uniformly-distributed hopping sequence,it is determined by a cyclic code of length while this sequence is repetitive, in a shortwindow of time it looks like a uniform random variable.

Transmissions are performed in slots, with a single packet being transmitted perslot. The frequency hops every time there is a new packet, matching the framing method to thehopping sequence. The multiplexing methodology is tightly controlled by the device, whereTDD is implemented by alternating the master and slave transmission slots, with the mastertransmitting in even-numbered slots, and slave(s) transmitting in odd-numbered slots.

In order to support asymmetric links, devices have the option of transmitting a singlepacket lasting as much as five slots; the center frequency used for each packet does not changeuntil that packet has ended, regardless of the number of slots the packet occupies. This stronglink between the packet structure and the hopping scheme means that the frequency-hopping ratemay drop below 1600 hops per second.

The modulation used is a Gaussian-shaped, binary FSK modulation scheme, with asymbol rate for the channel is 1 Msymbol per second, yielding a maximum raw bit rate of 1Mbps. Since there is considerable overhead in each packet (discussed in later sections), and sincethere is a window of time given to allow the oscillators to settle at the new frequency, the actualdata rate is significantly lower than 1 Mbps. Furthermore, the error-recovery scheme used byBluetooth, also discussed in later sections, performs re-transmissions of selected packets whoseerrors are detected but not corrected, further lowering the data rate.

Baseband SpecificationThe baseband specification provides a description the inner workings of the Bluetooth

radio, including its connection format, error recovery, packet structure, and link parameters.

Connectivity

The basic Bluetooth network is called a piconet, defined as the set of at most seven activedevices operating under the control of a single device. Note that while the limit is seven activedevices, there can be many more devices in other, inactive or passive modes. A set of Bluetoothpiconets is called a scatternet; the piconets in a scatternet do not need to be integrated in any way.

The piconet is centered on a single Bluetooth device called a master, who controls achannel and all the slave devices operating in that channel. To become a master, a devicerequests a connection with another device: if the paged device accepts the link, the calling devicebecomes a master for that link and the responding device becomes a slave.

Every Bluetooth device is exactly the same except for a 48-bit device identifier(BD_ADDR). The application residing above Bluetooth is the part of the system that determineswhich device is master or slave by the simple act of requesting a Bluetooth device to establish alink.

As can be inferred from the simple manner in which a device can become a master or aslave, the network layout for a set of Bluetooth devices is a dynamic environment. This dynamicbehavior is supported by the set of modes in which devices can operate. These modes allow thesharing of devices across multiple piconets in a scatternet. These modes, discussed in a latersection, also give devices the freedom to enter power-saving states.

Figure 3 shows an example of the link style Bluetooth employs: a scatternet consisting oftwo piconets, and a slave device shared between two different piconets in a scatternet.

254

Error Recovery

Bluetooth is intended to operate in very low quality channels, where the raw channelBER is expected to be as high as In order to operate reliably in such a channel, severallayers of error correction have been added to the standard, implemented in three forms of errorrecovery: rate 1/3 FEC (Forward Error Correction), rate 2/3 FEC, and an ARQ scheme.

Rate 1/3 FEC

The rate 1/3 FEC is a 3-bit repetition of every field of the form:This FEC scheme is used on all the packet headers and for the body of the HV1 packet type(packet types are discussed at a later point in the tutorial). The decoding method for this type ofcode is very simple, and can be implemented through a voting method with very little overhead.This simplicity comes at a cost, adding data redundancy that is not bandwidth efficient.

Rate 2/3 FEC

The rate 2/3 FEC is a (15,10) shortened Hamming code with generator polynomialThis format is used on certain types of payloads; a detailed description of the

packets providing this level of protection is given in Appendix B. The payload of these packets isbroken up into 10-bit blocks, and a 5-bit redundancy set is inserted after each corresponding 10-bit block, generating a 15-bit word. The decoding method for this type of code is considerably

255

more complex than the 1/3 rate code. This added complexity provides a very robust error-correction code that is considerably more efficient than the repetition 1/3 rate code.

ARQ Scheme

Bluetooth uses an unnumbered ARQ scheme with an acknowledgement scheme based onsingle ACK or NACK for each received packet. Each packet header contains a single bit, whichis used as an ACK flag.

Given the structured TDD scheme used by Bluetooth, the ARQ scheme can be easilyintegrated; a slave will acknowledge the receipt of the packet from the master in the slotimmediately after the master’s send slot. To have a positive acknowledgement, the minimumrequirement is that HEC in the packet header be correct and the CRC (if present) must check.

Link Types

There are two link types available in Bluetooth, Synchronous Connection-Oriented(SCO) and Asynchronous Connection-Less (ACL) links.

An SCO link is a symmetric, dedicated link between two devices; this dedicated channelis a circuit-switched connection whose target application is voice. The ACL link is anasynchronous link that uses those slots in a piconet that are not dedicated to an SCO link; thetarget application for ACL is data. In ACL links, a slave is limited to transmitting to the masteronly in the slot directly after the slot where the master addressed this particular slave.

Broadcast messages to the whole piconet are possible using ACL by addressing the all-zeros device in the packet header. Another benefit of ACL is that if the master has noinformation to send, and no polling is taking place, then the channel may be idle.

Link Management

The Link Manager Protocol (LMP) performs link management. The LMP performs linkconfiguration, including quality of service support, security, and establishment of the logicalchannels. There are two primary states for Bluetooth supported by the LMP: Standby andConnection. Apart from these states, there are seven states that are interim states designed to addnew slave devices to a piconet. These seven states are Page, Page Scan, Inquiry, Inquiry Scan,Master Response, Slave Response, and Inquiry Response.

Standby State

The standby state is the default power-up state. If the device issues a page or inquirymessage and the message is replied to, then the device enters the Connection state as a master.On the other hand, if the device scans a page or inquiry message and responds to it, then thedevice enters the Connection state as a slave.

Inquiry, Inquiry Scan, and Inquiry Response States

The goal of the inquiry set of states is to allow a device to find out which devices areavailable within transmit/receive range. The Inquiry Scan state is used to scan on a frequencybased on the inquiry access code; the length of time of the scan is a minimum of 18 time slots.

A device that wishes to query a slave scanning the Inquiry state frequencies does so fromwithin the Inquiry state. The frequencies over which the inquiry is broadcast are based on theinquiry access code. The Inquiry state is used only to query the devices in the Inquiry Scan state,

256

so connection state cannot be reached from an inquiry state. The Inquiry Response state is atransition from the Inquiry Scan state, and is only achieved by the slave device responding to therequest from the device in the Inquiry state.

Page Scan State

In this state, the device regularly scans a single hop frequency for at least 11.25ms, or 18TDD slots; this state can be entered from either the Standby or Connection states. To enter thePage Scan state from the Connection state, the device needs to free as much available scan timeas possible, preferably by entering the Hold or Park modes (these modes are discussed at a laterpoint in more detail). The scan frequency is derived from the Bluetooth device address(BD_ADDR). To enter the Slave Response state, a correlation for the device access code isperformed – when the correlation threshold is passed, the device enters the transition in statesoccurs.

Page State

The Page state is entered by a device that establish a connection with a particular slave.Based on the slave’s BR_ADDR, the master determines on which frequency to transmit the page.This state is best explained in the next section, where the paging process is shown.

Slave Response/Master Response States

Figure 4 shows the progression of messages and states while a page request is beingserviced. After receiving a page, the slave device transmits an ID packet (described in AppendixA) 625 µs after the paging ID packet.

257

Upon receiving the ID packet from the slave, the master device enters the MasterResponse state. In this state, it issues an FHS packet (FHS packet types are described inAppendix A) to arrange the piconet hopping sequence for use in the channel and to synchronizethe clocks of the two devices. When the slave device receives the FHS packet, it issues anotherslave ID packet to acknowledge the receipt of the FHS packet.

After issuing the ACK, the slave device switches to the master’s frequency hoppingsequence for the piconet and it enters the Connection state. The master cannot enter theConnection state until the ACK is received from the slave device. Once the ACK has beenreceived, the master enters the Connection state and begins the transmission of data on the masterfrequency-hopping scheme.

Through this process, a connection in a piconet is established between a master and aslave.

Connection State

The Connection state is the state in which data is exchanged between the master and theslave, and the final result of the transaction seen in Figure 4. The first packet sent by the masteris a POLL packet (explained in Appendix B) to make sure that the slave switched to the proper

258

frequency hopping scheme and that the clocks are synchronized. If the master does not receivethe slave’s response or if the slave does not receive the POLL packet from the master, then themaster/slave return to the Page/Page Scan states.

After the POLL packet, the master sends link details to the slave. These details includethe definition of the type of link and the sniff parameters (more details on sniff mode are shown ata later point in the tutorial). After the details have been passed to the slave, the actual connectionbegins with the TDD exchange of packets between the master and the slave.

The Connection state is exited through either a “detach” or “reset” command. The“detach” command terminates the connection but maintains the link parameters, while the resetcommand performs a hard reset of the device, eliminating all the existing configurationinformation.

The Bluetooth device modes, some of which were mentioned above, are shown in thefollowing sections.

Active Mode

In this mode, the slave participates actively on the channel – the master schedules all thetraffic use between the different devices in the piconet. To conserve batteries, slave devices havethe option of sleeping while the master is not addressing them.

Sniff Mode

While in normal ACL operation, the slave needs to listen during every single slot inwhich the master is transmitting; the sniff mode allows the slave to listen to only a specific subsetof the slots in which the master is transmitting. This reduction in duty cycle for the listen modeof the slave can be used to listen to transmissions in another piconet in which the slave device isalso a slave.

Hold Mode

The ACL link to a slave can be placed on hold, allowing the slave to scan, page, inquire,or handle another piconet. During hold mode, it is also possible for the slave device to sleep,conserving power. During hold mode, the slave remains part of the piconet, so it does not lose itsactive member address, a 3-bit address that is used to address each of the active slave members inthe piconet.

Park Mode

In the park mode, the slave remains synchronized to the channel, but it does not activelyparticipate in it. The slave receives two addresses after it has parked. PM_ADDR is an 8-bitparked member address, which can be used by the master to un-park the slave. AR_ADDR is the8-bit access request address, which is used by the slave to perform a slave-initiated un-park. Inthe park mode, it is possible for a slave device to remain a member of a piconet while not takingup an active member address. As mentioned in the previous sub-section, the active memberaddress is a 3-bit address used to identify each of the seven active slave devices in a piconet.Because of this 3-bit limit, it is important to be able to manage the active member list, and allowinactive devices to temporarily cease their active participation in the piconet.

259

Packets

Data sent across a piconet is sent in packet format. While the packets take the generalformat shown in Figure 5, there are several types that are used to support a variety of services.

As shown in figure 5, packets are broken up into three pieces, an Access Code, andHeader, and the packet Payload. Appendix A describes the content of each section of the packet.All packets in Bluetooth Mow this general format.

Packet Types

In order to provide different quality of service (QoS) guarantees to the application,Bluetooth can support a wide variety of packet types. These packet types are broken up into threeclasses: packets supporting the link types SCO, ACL, and both. A description of each availablepacket is presented in Appendix B. In general, the common packets are used to perform controland general maintenance functions. These packets are primarily used for housekeeping.

The SCO and ACL packets are used for data transport, where different packet formats areused to support various service loads as well as multiple levels of QoS. The required support ofasymmetric loads and different levels of error correction/detection in this system means thatmultiple packet formats need to be supported.

Since different types of information can be sent over a Bluetooth link, the format of thepayload needs to change to satisfy each of these payload needs. This format is used to supporteither voice or data.

The voice field is available in only SCO packets. The voice field length 240 bits for theHV packet series and 80 bits for the DV packet. There is no payload header in voice fields.

The data field contains a payload header, payload body, and, with the exception of theAUX1 packet, a CRC code. The payload header is either one or two bytes long, and contains alogical channel indication, flow control on the logical channels, and a payload length indicator.The CRC code is generated by the CRC-CCITT octal polynomial 2 1 0 0 4 1.

Audio Support

Bluetooth uses 64 kbps Continuous Variable Slope Delta Modulation (CVSD) or 64 kbpsPCM using either A-law or law – The PCM coding method follows the ITU-Trecommendations G.711.

The CVSD codec uses syllabic companding to reduce the slope overload effect. In otherwords, the step size used in the delta modulation is adapted as a function of the average signalslope. Using negative steps represented as “1” and positive steps represented as “0”, the inputsignal is encoded.

The minimum and maximum step sizes are bound between 10 and 1280, and theaccumulator maximum and minimum are set to and respectively.

260

The CVSD audio quality requirements are placed on the transmitter, where the 64 kbpslinear PCM input signal needs to have a power spectral density above 4 kHz set 20 dB below thepower spectral density below 4 kHz.

Security Specifications

The Bluetooth standard contains a detailed description of the security process necessaryto perform key management and encryption of data. While this material is important in thedeployment of the devices, it is beyond the scope of this tutorial. However, a brief description ofthe elements making up the encryption process is provided in the following paragraphs.

In order to maintain a level of security in the transactions between devices, there are fourbasic elements: public address unique to each user (BD_ADDR, 48 bits), private user key(authentication, 128 bits), private user key (encryption, 8-128 bits), and a random number (128bits). BD_ADDR is the unique Bluetooth unit 48-bit IEEE address derived from the UUID(Universal Unique Identifier), which is publicly known and can be obtained through deviceinteraction. The secret keys are derived during initialization and are not disclosed. Because ofencryption export restrictions, the encryption private user key is variable-length ranging between8 and 128 bits; the encryption key is derived from the authentication key. A random numbergenerator is used to generate the authentication and encryption keys.

Logical Link ControlThe Logical Link Control and Adaptation Layer Protocol (L2CAP) is layered over the

baseband protocol, which was described in the above sections. The goal of L2CAP is to supportservices, including protocol multiplexing, segmentation/reassembly of packets, quality-of-service(QoS), and group abstraction.

Protocol Multiplexing

Protocol multiplexing is the ability to mix multiple services on top of the Bluetoothdevice. Since the baseband protocol treats all data packets equally, L2CAP is necessary todistinguish the different services running on top of the device so that the needs of each service aremet. The three protocols described in the standard that can be multiplexed by L2CAP are theService Discovery Protocol (SDP), RFCOMM (emulation of a serial link), and TelephonyControl.

Segmentation/Reassembly of Packets

One of the basic limitations of the Bluetooth baseband protocol is that the packets thatmake up its transport service are size-limited. In order to accommodate large packets, it isnecessary to breakup and reassemble those packets as they enter the transmitter and exit thereceiver; L2CAP performs this segmentation/reassembly process.

L2CAP performs an integrity check on the data by leveraging the 16-bit CRC that isalready available in the baseband protocol. Furthermore, the ARQ mechanism used by thebaseband protocol is also used to guarantee data integrity at the receiver.

Quality-of-Servlce (QoS)

L2CAP supports QoS messages between Bluetooth devices. However, the only servicethat is required implementation in each device is the “Best Effort” service; in other words, the

261

only QoS guarantee that is required in the standard is a service with no guarantees. Other QoSservices are optional. These services include: token rate, token bucket size, peak bandwidth,latency, and delay variation. Each of these services are commonly used in network managementto create QoS guarantees.

Group Abstraction

The Bluetooth baseband protocol is based on the concept of a piconet. If an applicationis to optimize its operation such that the piconet structure matches the operation of theapplication, it is necessary for that application to have direct access to the baseband protocol orlink manager. However, this can lead to a very complicated structure. In order to avoid thisproblem, L2CAP provides a Group Abstraction mechanism, such that specific services andapplications provided within a group of devices can be mapped to the piconet architecture. Thismapping process allows the application to take full advantage of the standard’s piconetarchitecture without the application needing to have direct access to the lower layers of theprotocol stack.

Services

Bluetooth is capable of simultaneously supporting multiple services, allowing the full useof the dynamic nature of the baseband system. The following sections describe these services.

Service Discovery Protocol

From a data standpoint, the Service Discovery Protocol (SDP) is one of the moreinteresting services. SDP allows a Bluetooth device to query nearby devices and find out whichservices are available for usage.

The client application interfaces with the SDP client, requesting either a search ofservices or a connection using one specific service. The SDP client interfaces with the SDPserver in another device, which can service the request forwarded by the SDP client. The SDPserver is aware of the services running on the server device through its interaction with the localapplications.

Basic Functionality

SDP functionality is fairly simple – an SDP client requests information from a nearbySDP server. The SDP server returns a Service Record, which contains a list of the serviceattributes. If the client decides that the returned service description meets its current needs, and ifthe authentication process succeeds, then a connection is established.

A separate connection is required to initiate the service. The connection used to performthe SDP process is limited to SDP functionality only, so a new link needs to be established todeliver a particular service.

SDP Description

A Bluetooth device contains at most a single SDP server. This server manages the SDPprotocol for every server application running in that one Bluetooth device. Since Bluetooth isdesigned to be able to support multiple applications, such that a device can be a master, a slave,or both, a single Bluetooth device can contain an SDP client and server at the same time.

262

The Bluetooth SIG set several goals for the SDP, leading to multiple servicerequirements.

The ease with which devices can enter and exit an area means that SDP needs to operatein a dynamic environment. In order to support this dynamic operation, SDP is required to allowdynamic service discovery, where a device is permitted to enter or exit a coverage area at willwith no adverse effects on the other SDP clients/servers.

Since Bluetooth devices are intended to operate over a wide variety of devices with littleor no supervision by a regulatory institution, SDP is required to support the creation of newservices without their registration with a central authority. Furthermore, in order to be able toidentify the services in a device with no conflicts, the services are required to have uniquely-identified services/service classes. These two requirements may seem contradictory; since thereis no regulatory body overseeing the creation of services, how can services be uniquelyidentified? Bluetooth is capable of creating unique identifiers by associating services with aUUID (Universally Unique Identifier), a 128-bit identifier that guarantees that the identifier isunique. Through the use of the UUID, it is possible to guarantee uniqueness while at the sametime avoiding a central authority.

Services operating over SDP are classified as services and service classes. A service is aspecific instance of a service class. The service class defines all the attributes that are possible toinstances of that class, and the service is a specific collection of desired attributes. A serviceclass is usually a subclass of another service class, where a subclass inherits all the attributes ofthe superclass. An example of a subclass structure is seen in Table 1; in this example, aBluetooth device is used to replace the cable connecting a low-bandwidth data source to anadaptive array.

• AdaptiveAntennaArrayClassID• AntennaArrayClassID• AntennaClassID

Table 1 - Example of adaptive antenna ServiceClassIDList

In the example shown above, several classes of services are shown, whereAdaptiveAntennaArrayClassID is a subclass of AntennaArrayClassID, which in turn is a subclassof AntennaClassID. Through such a hierarchy, a related set of services can be established,simplifying the discovery process. This service class hierarchy can lead to a clean organization inthe service structure.

There are several service attributes that can be associated with a service class. Theseattributes include ServiceName (a human-readable text name), ServiceID (derived from theUUID), ServiceClassIDList (list of classes in which the service is an instance). Table 1 is anexample of an adaptive antenna ServiceClassIDList. This set of attributes defines the class.

Since the service class ID may not be sufficient to determine its applicability to thecurrent application needs, Bluetooth requires SDP to allow the search of services based on theservice attribute. Furthermore, since services may be created without central control, a devicemay be unaware that another device within communication range operates a service of interest.To overcome this problem, Bluetooth requires SDP to allow services to be browsed withoutapriori knowledge of service characteristics. Furthermore, a device may be limited in its abilityto handle the full functionality of a particular service, so Bluetooth requires SDP to allow gradualservice discovery, where the service capabilities made available to a client are limited by thatclient’s ability to handle those service attributes. Service caching is also required, since it can beused to significantly reduce traffic in a system.

Since SDP is required to run over a device that is intended to be very inexpensive (andvery simple), SDP is required to be very simple.

263

Finally, the Bluetooth SIG recognized that SDP may not be sufficient to cover all theneeds of the applications running on a Bluetooth device, so it required SDP to allow the usage ofother service discovery protocols. In order to support applications with large packets, multipleprotocols or QoS requirements, SDP is required to support L2CAP functions.

Based on these requirements, SDP has defined a hierarchy of service classes and serviceattributes such that a powerful methodology for supporting a wide variety of services is available.

Emulation/Telephony Protocols

Two emulation/telephony protocols are described in the Bluetooth standard to supportservices: RFCOMM and Telephony Control Protocol

RFCOMM is the emulation of a serial port over the L2CAP protocol. Using RFCOMM,it is possible to support up to 60 simultaneous connections between two Bluetooth devices – theactual maximum of supported connections may actually be less than 60, and is implementation-specific. Using RFCOMM, Bluetooth can act as a replacement for the serial cable.

The telephony control protocol allows the establishment of telephony functionality overBluetooth devices. The protocol allows call control, where speech or data calls betweenBluetooth devices can be established or terminated. Group management allows the easy handlingof groups of Bluetooth devices. Finally, the protocol can support connectionlesscommunications, allowing the exchange of signaling information that is not related to the on-going call.

Integration with Other Wireless Services

The Bluetooth standard allows the interoperability between Bluetooth and other wirelesscommunications protocols; the standard lists three specific protocols: IrDA and WAP.

IrDA

The concept behind IrDA interoperability is to support the development of applicationsthat are compatible with both short-range RF and IR links. To achieve this goal, a technologyoverlap with IrOBEX (Infrared Object Exchange Protocol) was pursued. IrOBEX was defined byIrDA (Infrared Data Association) – in Bluetooth, IrOBEX is referred to as OBEX. Thetechnology overlap is achieved by mapping OBEX over RFCOMM and TCP/IP; the mappingover TCP/IP is optional.

The connection-oriented version of OBEX is mapped over the connection-orientedBluetooth architecture. Using OBEX, it is possible to exchange data objects. This functionalityallows the implementation of simple commands such as Connect, Disconnect, Put, Get, SetPath,and Abort.

WAP

WAP (Wireless Application Protocol) is a standard that was developed to allow a mobilephone to access the Internet. In this interoperability case, Bluetooth is used as the physical layerand link control to connect two WAP-enabled devices; in other words, the physical layer such asGPRS or GSM is replaced with Bluetooth. Using these interoperability guidelines, its is possibleto access the Bluetooth device through a WAP application, providing applications that canoperate over a wide-area network over a PCS infrastructure as well as over a Bluetooth LAN.

264

The WAP interoperability guidelines are supported through a combination of PPP (point-to-point protocol) and RFCOMM. SDP is also supported, allowing WAP to access the dynamicservice discovery environment provided by Bluetooth.

In order for the WAP interoperability to be functional, a WDP (Wireless DatagramProtocol) management entity is needed – WDP is the WAP layer that provides transportcomparable to UDP (socket-based, unreliable delivery). The management entity provides an out-of-band mechanism for controlling the protocol stack, providing support for the detection ofnodes and other events.

Critical PerspectiveInitial work with Bluetooth has pointed to a very robust and flexible standard that should

be able to support a wide variety of services. While the outlook for the performance of thestandard is good, it leaves several questions open as to its ability to co-exist with other devices inthe same spectrum.

Along with Bluetooth, there are several systems intended to operate over the ISM band,including some versions of the IEEE 802.11 wireless networking standard. While Bluetooth isfairly robust, its effect on other devices has the potential for causing a significant disturbance.Unfortunately, predictions done before deployment are based on broad deployment assumptions,and the gravity of this problem cannot be determined until these systems see wide-scaledeployment in a real-world environment and until the applications that have broad appeal aredeveloped and deployed. Since the physical layer of these systems is limited in its ability tochange to account for changes in the RF environment, one of the key issues governingperformance is the application. The supported application determines the traffic that thesedevices will need to support, and hence the amount of energy that these devices are expected tobroadcast.

Recently, there has been a flood of possible applications growing from both theestablished corporate and the entrepreneurial communities; the potential in Bluetooth has beenidentified by many people and has led to no shortage on ideas for applications. Thus, from amarket standpoint, application concepts do not seem to be a barrier to Bluetooth’s success.However, two primary challenges have been identified that will play an important role indetermining the success of Bluetooth as a way of connecting people: application interoperabilityand price.

The concept of application interoperability is the ability for multiple applications to sharea single physical interface without causing confusion within the device and with other Bluetoothdevices. Bluetooth’s Service Discovery Protocol is a mechanism that can help avoid thisproblem. However, the ability of the application developers to properly use the device/protocolwill determine the degree to which this problem is avoided.

The price of the Bluetooth device is a central concern to developers; Bluetooth devicesneed to hit a price point that is low enough to allow these devices to turn into commodities thatcan be added to practically any appliance/tool. The key to the low price point for these devices isthe single-chip solution, since the bulk of the cost would be limited to the initial design ratherthan the fabrication process. Manufacturers such as Ericsson have announced full solutions [5],with a single-chip solution predicted for the near future. Only time will tell if conditions allowthe market to push the cost of these devices to the anticipated price point.

In the year 2000 or 2001, the specifications for Bluetooth 2.0 will be released [1]. Whilethere has been significant speculation towards the capabilities of this new version of the standard,no information has been released to the public concerning the problems it is intended to addressand its expected capabilities.

265

ConclusionBluetooth is a standard that is designed to enable inexpensive, robust, and secure

communications over short distances. This tutorial provided a general description of theBluetooth standard, providing the reader with an overview of the potential as well as thelimitations of this standard. The flexible architecture of Bluetooth opens the door to the designand inexpensive implementation of several applications that add a whole new dimension to short-range communications. The future holds great hope for short-range data communications, andBluetooth goes a long way in setting the path to ubiquitous, integrated data services.

266

Appendix A – Packet FormatThis appendix describes the content of the fields in the general packet format, a picture of

which is seen in Figure 5.

Access Code

The access code is a 72-bit block that is used in all packets except the FHS packet, whichis discussed in Appendix B. The access code is used to perform packet synchronization, DCoffset compensation, and identification.

The Bluetooth receiver performs a sliding correlation and triggers the receiver when thecorrelation exceeds a threshold.

The access code can be one of three different types: Channel Access Code (CAC),Device Access Code (DAC), and Inquiry Access Code (IAC). The access code is broken up intoa 4-bit preamble, a 64-bit sync word, and (occasionally) a 4-bit trailer.

Header

The 1/3 rate error recovery is used in the header of all packets. The packet header is usedfor link control. The six fields making up the header are: AM_ADDR, TYPE, FLOW, ARQN,SEQN, and HEC.

• AM_ADDR: a three-bit field that contains the piconet member address. These threebits are used to address the seven slave devices making up a piconet – the all-zerosaddress is used as a piconet broadcast. The only exception to the all-zeros broadcastis when an FHS packet is sent (FHS packets are discussed in Appendix B).

• TYPE: a four-bit field that contains describes the packet type. The interpretation ofthis field depends on the link type (i.e.: SCO link type).

• FLOW: a single bit that is used by ACL links to perform flow control. This fieldtakes on the value zero when the receiver buffer is full.

• ARQN: a one-bit acknowledgment used in the AQN mechanism.• SEQN: a bit that is inverted every time that a new packet with data and a CRC is

transmitted.• HEC: the header-error-check. This is an 8-bit word that is generated with octal

polynomial 6 4 7. The HEC is calculated for the 10 header bits.

Payload

The payload is a variable-length field that contains the information that is sent in thepacket. The level of error correction and the existence of this field is a function of the packet typethat is transmitted.

267

Appendix B – Packet DescriptionThis appendix has a description of the different types of packets that are available in only

SCO or ACL links, and as well as those that are available in both types of links.

Common Packets

The common packets are those packets that can be used in either SCO or ACL link types.The five packet types are: ID, NULL, POLL, FHS, and DM1.

ID Packet

The ID packet is the device access code (DAC) or inquiry access code (IAC) with noHeader and no Payload. The total length of this packet is 68 bits. This packet is generally usedas a response to paging or inquiry requests.

NULL Packet

The NULL packet consists of the channel access code and the packet header, but nopayload. The total length of this packet is 126 bits. While this packet does not need to beacknowledged, it is used to return the values of the ARQN and FLOW fields.

POLL Packet

The POLL packet is similar to the NULL packet. The primary difference is that thispacket does not affect the ARQN and SEQN fields – this packet would be used by a master topoll a slave device, which would need to respond with an acknowledgment, regardless of whetheror not it has data to transmit.

FHS Packet

The FHS packet is used for control, and it used to reveal piconet information to themember devices. Examples of the revealed information include the Bluetooth device address andthe system clock. The 2/3 rate code is used to protect the payload, which contains 144 bits plus a16-bit CRC code. The packet contains 240 bits and covers a single slot.

DM1 Packet

The DM1 packet can carry regular data, but is part of the common type of packet groupbecause it can be used to support control messages. While this packet is commonly used in ACLlinks, it can be used in SCO links, where it is capable of interrupting a synchronous link in orderto send control information.

SCO Packets

The SCO link is designed to transmit voice, so these packets contain no CRC, and thesepackets are also not retransmitted. These packets are typically used to deliver 64 kbps speechtransmissions.

268

HV1 Packet

HV stands for High-quality Voice, describing the intended primary service intended forthis type of packet. The HV1 packet transports 10 bytes protected by the 1/3 rate error-correctioncode. No payload header is present, and the total payload length is limited to 240 bits. Thispacket can carry 1.25 ms of speech encoded at 64 kbps. To maintain an SCO link using HV1packets, an HV1 packet needs to be sent every two time slots.

HV2 Packet

The HV2 packet is similar to the HV1 packet. The primary difference is that it uses the2/3 rate error-correction code on its payload, allowing it to carry 20 information bytes. Thisreduced redundancy allows each packet to carry up to 2.5 ms of speech; to maintain an SCO linkusing HV2 packets, an HV2 packet needs to be sent every four time slots.

HV3 Packet

The HV3 packet is similar to the HV1 or HV2 packets, with the primary that it carries noerror-correction code, allowing the code to carry 30 information bytes. The unprotected packetcan carry 3.75 ms of speech; to maintain an SCO link using HV3 packets, an HV3 packet needsto be sent every six time slots.

DV Packet

The DV packet contains a mix of voice and data: 80 voice bits and up to 150 data bits.While the voice field carries no error-correction codes, the data is encoded with a 2/3 rate error-correction code. The voice and data content of each packet is treated differently; the voice field isnever retransmitted, while the data field is retransmitted until a transmission with no detectederrors is received.

ACL Packets

ACL packets are used in asynchronous links, and are designed to carry data. While theDM1 packet is designated as a common packet, it is considered to be an ACL packet – DM1 isused in SCO links to carry control information.

DM1 Packet

DM stands for Data-Medium rate. The payload contains 18 information bytes, one ofwhich is a payload header and an added 16-bit CRC code. The payload in this packet type isencoded with the 2/3 rate error-correction code. DM1 packets occupy only one slot.

DM3 Packet

DM3 packets are very similar to DM1 packets. The only difference is that, unlike theDM1 packet, the DM3 packet occupies three slots. The extra length allows a DM3 packet tocarry 123 data bytes, including a 2-byte header, with an added 16-bit CRC code.

269

DM5 Packet

The DMS packet is, like DM3 packets, a variation of the DM1 packet. The onlydifference between DM5 and DM1 is that DM5 packets occupy five slots, allowing it to carry upto 226 information bytes, including a 2-byte payload header and an added 16-bit CRC code.

DH1 Packet

DH stands for Data-High rate. Unlike the DM series of packets, DH packets carry noerror-correction codes. The only error recovery used by DH packets is error detection through a16-bit CRC combined with the ARQ scheme. DH1 packets can carry up to 28 bytes ofinformation.

DH3 Packet

DH3 packets are very similar to the DH1 packets. The only difference is that DH3packets occupy three time slots. DH3 packets carry up to 185 information bytes including a two-byte payload header as well as a 16-bit CRC code.

DH5 Packet

DH5 packets are also a variation of DH1 packets. The only difference between DH1 andDH5 packets is that DH5 packets occupy five time slots. The extra transmission time per DH5packet allows DH5 packet to transport up to 341 information bytes including a two-byte payloadheader, with an added 16-bit CRC code.

AUX1 Packet

The AUX1 packet is a variation of the DH1 packet. AUX1 packets occupy a single timeslot and use no payload CRC code. The lack of error-detection capability allows the AUX1packet to transport 30 information bytes, including a 1-byte payload header. Since no CRC codeis used on the payload, no retransmission is possible if there are errors in the packet payload.

270

Reference:1 – http://www.bluetooth.com/developer/specification/specification.asp2 – http://www.bluetooth.com/3 – http://www.pcsdata.com/CahnersBluetooth.htm4 – http://www.bluetooth.com/bluetoothguide/models/two in one.asp#top5 – http://bluetooth.ericsson.se/ebc/solulion.asp

About the author

Max Robert is a Ph.D. student at Virginia Tech’s Mobile and Portable Radio ResearchGroup and is currently working with Dr. Jeffrey H. Reed. Mr. Robert attended Case WesternReserve University for his undergraduate work, where he received several honors andscholarships, including the Albert W. Smith scholarship. He graduated magoa cum laude with aBachelor of Science degree in Electrical Engineering and Applied Physics in 1996. Mr. Robertreceived his Master of Science degree in Electrical Engineering from Virginia Tech in 1998. Hisresearch focused on joint channel-video coding for MPEG-2 transmissions over high-bit-error-rate channels, for which he received the Paul E. Torgersen Graduate Student Research ExcellenceAward in 1999. Mr. Robert’s Ph.D. research focuses on data network performance andinterference issues. Mr. Robert is a Bradley Fellow of the Bradley Department of Electrical andComputer Engineering at Virginia Tech. Mr. Robert is also Vice-President and co-founder ofDotMobile, Inc., a company that focuses on wireless information technology.

INDEX

Aaccess code 265ACK 255active mode 256adaptive 205-215adaptive antennas 145Adelaide 59Adelaide propagation signatures 60AM_ADDR 265antenna-body interaction 15ARQ 254ARQN 265asynchronous connection-less 255attenuation 53, 173, 177attenuation factors 53, 173, 177audio 259Australian cities, 59AUXI Packet 268average number of rays 58

Bbandwidth 169-172barriers 53base station deployment 162, 163BD_ADDR 253Bluetooth 2.0 264body-worn terminal 18building database 36

Ccalibration 48, 77-87canceller 205-215capacity analysis 89-99CDMA 23-34CDMA planning 162, 163cellular system 89-99center frequency 171, 172channel capacity 77-87channel estimation 153, 234channel interference 169channel model 181-192channel modeling 45class 261clear channel assessment (CCA), 176co-channel interference reduction 77-87commercial cost 160, 161continuous variable slope delta modulation 259convolutional codes 23-34correlator 45, 48, 54cost function 159-161cost weights 161

Ddata 259data collection 49decision variables 158delay spread 43, 45, 52, 53, 55delay time 53diversity 121-130DHI Packet 268diversity combining 23-34dynamic range 58

Eelement arrangement 77-87encryption 259empirical models 37excess delay 57excess delay bins 58

FFDD 232FEC 254Federal Communication Commission (FCC) 169, 170FHS Packet 266field strength prediction 39fixed radio link 17floor and wall attenuation 62FLOW 265free space path loss 173frequency channels 170frequency hopping 89-99frequency planning 169, 172, 180

GGPRS 181-192guided simulated annealing 162group abstraction 260GSM 181-192

Hhandover 193hardware 77-87header 265HEC 265hold mode 256hopping sequence 252host 252host controller interface 252human body modeling 13

272

IIdOBEX 263ID Packet 266IEEE 80.211 169, 170, 263impulse response 43indoor measurements 62indoor propagation 162, 163inner receiver 231, 233inquiry 255interference 169, 170, 172-174, 176, 177inter-signal coherence 149I-Q TCM 121-130IrDA 263ISM 169, 170, 252

Jjoint detection 234

Lline of sight path loss 173link manager protocol 255logical link control and adaptation layer

protocol 259low transmitter antenna 59L2CAP 259

mai 205-215marginal value of correlators 54Markov chain 181-192Master 253maximum straight line separation 64mean delay 64measurements 41Melbourne 59, 61mesh plot of ray probability 59microcell 35mini-max function 159, 160mini-sum function 159, 160modulation 252mobile positioning 198mobile radio channel 43multipath propagation 12, 57multipath rays 57multiple transmit antennas 121-130multiuser detection 205-215

Nnegative excess delays 58network simulation 181-192noise floor 64NULL packet 266

OOBEX 263optimisation, combinatorial 157-168orthogonal coordinate transformation 151outage probability 159, 160outdoor propagation measurements 59overlapping channels 169, 171-174

Ppacket 258packet erasure channel 181 -192page 256park mode 257path loss 45, 51, 55payload 258PCM coding 259PCS 11, 23-34penetration depth 13pedestrian effects 17picocell 35piconet 253ping-pong handover 193POLL Packetpower control 232power delay profiles 57preprocessing 38price 264probability of a ray 58propagation 45, 55, 56propagation environments 64propagation measurements 57propagation modeling 36propagation signature 57propagation statistics 64protocol stack 250

Qquality of service 260

M

273

Rradio architecture 236radio network planning 35rank reduction 151ray launching 37ray-tracing 16, 37Rayleigh fading 18, 121-130reassembly 260receiver 173, 175, 177reduced dimension space-time receiver 153reflection coefficient 14reflections from external structures 62RFCOMM 262RF environment 49Rician fading 19RMS delay 45, 52, 53, 55, 64RS232 serial bus 252

Sscan 255scanning receiver 47-49scatternet 253SDP 260SDP client 261SDP Server 261security 259segmentation 260SEQN 265service discovery protocol 260single detection 235Slave 253slot 231smart antenna 77-87, 89-99sniff mode 256software radio (SWR) 89-99space time codes 121-130space time processing 145spatial correlation 148, 151spatial dimension reduction 151, 153spatial diversity combining 145spatial signature 149spectrum 230standby 255subspace methods 151superclass 262Sydney 59Sydney Harbour 63synchronous connection-oriented 255system planning 162, 163

TTCP/IP 263TD-CDMA 230TDD 230temporal variations 183GPP230throughput 174-176, 178-180time modulated 45transmitter 173, 175, 177transport layer 252TYPE 265

UUART 252UDP 263ultra-wideband 45University of Technology Sydney 63unnecessary handover 194, 197USB 252UUID261

Vvoice 259

WWDP 263wideband propagation 57wireless application protocol 263wireless LAN 11, 169, 180