Research Collection

Doctoral Thesis

Active antenna radio frontends for multiple antenna communication systems

Author(s): Brauner, Thomas

Publication Date: 2004

Permanent Link: https://doi.org/10.3929/ethz-a-004904237

Rights / License: In Copyright - Non-Commercial Use Permitted

This page was generated automatically upon download from the ETH Zurich Research Collection. For more information please consult the Terms of use.

ETH Library ACTIVE ANTENNA RADIO FRONTENDS FOR MULTIPLE ANTENNA COMMUNICATION SYSTEMS

calibration CAL1 line CAL2

attenuator receiver receiver receiver receiver

LO

IF1 IF2 IF3 IF4

Thomas Brauner

DISS. ETH No. 15642

DISS. ETH No. 15642

ACTIVE ANTENNA RADIO FRONTENDS FOR MULTIPLE ANTENNA COMMUNICATION SYSTEMS

A dissertation submitted to the SWISS FEDERAL INSTITUTE OF TECHNOLOGY ZURICH

for the degree of Doctor of Technical Sciences

presented by THOMAS BRAUNER Dipl. Ing., RWTH Aachen Born February 8, 1973 in K¨oln (Germany)

accepted on the recommendation of Prof. Dr. W. B¨achtold, examiner Prof. H. B¨olcskei, Prof. R. K¨ung, coexaminers

2004

Twenty years from now you will be more dis- appointed by the things that you didn’t do than by the ones you did do. So throw off the bow- lines. Sail away from the safe harbor. Catch the trade winds in your sails. Explore. Dream. Discover. — Marc Twain

Contents

Table of contents v Abstract ix Zusammenfassung xi 1 Introduction 1 1.1 Motivation 1 1.2 Organization of this work 3 2 System design 5 2.1 Receiver design 5 2.1.1 Dynamic range 5 2.1.2 Receiver architecture 8 2.2 Multiple antenna system 10 2.2.1 Noise and linearity 10 2.2.2 Antenna combining methods 11 2.2.3 Local oscillator distribution 12 2.2.4 Antenna placement 13 2.3 Noise in multiple antenna systems 16 2.3.1 Signal and noise model 16 2.3.2 Noise correlation 17 2.3.3 Phase noise 18 2.3.4 Correlationofphasenoise 19 2.3.5 System noise model 21 2.4 Testbed architecture 21 3 Integrated circuit design 23 3.1 Process technology 23 3.1.1 Choice of technology 23 3.1.2 TriQuint TQTRx process 25 3.2 Low-noise amplifier 26 3.2.1 Input matching 26 3.2.2 Device scaling 28 3.2.3 Three-stage amplifier 29 vi Contents

3.2.4 Measurement results 30 3.3 Downconverter 32 3.3.1 Resistive mixer design 32 3.3.2 Mixer scaling 36 3.3.3 Integrated downconverter 37 3.3.4 Measurement results 37 3.4 Integrated front-end 42 3.4.1 Architecture 42 3.4.2 Switchable LNA 42 3.4.3 Image filter 44 3.4.4 Layout 45 3.4.5 Experimental results 47 3.5 Power amplifier 54 3.5.1 Design 54 3.5.2 Experimental results 57 3.5.3 Pulsed operation 58 3.6 10.7–11.7 GHz SiGe downconverter 60 3.6.1 IBM6HPSiGeBiCMOSprocess 61 3.6.2 Low-noise amplifier 61 3.6.3 Integrated downconverter 65 3.7 Conclusions 69 4 Passive arrays 71 4.1 Antenna design 71 4.1.1 Choice of antenna structure 71 4.1.2 Aperture-coupled patch antenna 72 4.1.3 Modelling 73 4.1.4 Design method 74 4.1.5 Results 76 4.2 Differential antenna 78 4.2.1 Differential MMIC interface 78 4.2.2 Design 78 4.2.3 Measurement results 79 4.3 Antenna arrays and mutual coupling 82 4.3.1 Non-ideal arrays 82 4.3.2 Classification 83 4.3.3 Coupling compensation 85 4.3.4 Array of aperture-coupled patch antennas 86 4.3.5 Array coupling model 89 4.4 Reduction of coupling in active arrays 90 4.4.1 Interface optimization 90 Contents vii

4.4.2 Experimental verification 92 4.5 Conclusions 95 5 Calibration 97 5.1 Problem formulation 97 5.1.1 Active circuit variations 98 5.1.2 Calibration network precision 100 5.1.3 Calibration requirements 101 5.1.4 Pattern error 103 5.1.5 Statisticalarrayerror 103 5.2 Existing calibration methods 104 5.2.1 Passive array calibration 104 5.2.2 Coupling estimation from far-field 105 5.2.3 Test-tone calibration 108 5.2.4 Hybrid methods 110 5.2.5 Improved Hybrid Calibration 111 5.3 Transmission-line calibration method 112 5.3.1 Description of method 112 5.3.2 Estimationofsystematicerror 114 5.3.3 GaAs Tx/Rx-switch with calibration ability 118 5.3.4 Experimental Results 119 5.4 Dynamic transmitter calibration 122 5.4.1 Instantaneouserror 122 5.4.2 Power amplifier 123 5.4.3 Array calibration 123 5.5 Conclusions 127 6 Active antenna arrays 129 6.1 Linear array 129 6.1.1 Design 129 6.1.2 Experimental results 132 6.1.3 Calibration 136 6.2 Gain and phase stability 139 6.3 Noise correlation 143 6.3.1 Amplitude noise correlation 143 6.3.2 Phasenoisecorrelation 144 6.4 Conformal array 146 6.4.1 Motivation 146 6.4.2 Design 147 6.4.3 Experimental results 151 6.5 Conclusions 157 viii Contents

7 Summary, conclusions and outlook 159 7.1 System design 159 7.2 Integratedcircuitdesign 159 7.3 Passive arrays 160 7.4 Calibration 161 7.5 Active antenna arrays 161 7.6 Conclusions and future work 162 Bibliography 163 Curriculum vitae 171 List of publications 173 Acknowledgments 175 Abstract

Goal of the work presented in this dissertation is to implement and characterize a 5 GHz active antenna array. Planar aperture-coupled patch antennas and monolithically integrated active circuits are combined to yield a compact, robust and easy-to-manufacture multiple antenna fron- tend. The realized hardware is intended for the use in both, in multiple antenna wireless LAN systems and for a multi-dimensional channel sound- ing equipment. To enable the application in the measurement system, the frontend is optimized for low-noise and high linearity. A classical superheterodyne architecture is chosen to maintain flexibility when adopting the system to different environments. An internal calibration signal is provided at all receiver inputs to determine and compensate all variations of the active hardware. A commercially available 0.6 µm GaAs MESFET process is used to in- tegrate the complete RF-frontend, including low-noise amplifiers, a lumped- element image filter and downconverter, onto a single chip of 3.2 mm2. Thereby, a state-of-the-art single side-band noise figure of less than 4 dB and an image rejection of more than 35 dB are reached. An active switch- ing concept is proposed to select between receiving and calibration mode without degrading the noise figure. To evaluate the ability of modern silicon-based technologies, an 11 GHz receiver frontend is demonstrated on a low-cost 47 GHz SiGe process. Using the MOSFET device to form a resistive mixer, a single-sideband noise figure of 7 dB and a input com- pression point of −14 dB can be realized at the same time. Through the development of an equivalent circuit model, an efficient design of the passive antenna structure for given specifications is facili- tated. The model is heuristically extended to include the mutual coupling between adjacent elements. The co-design of the receiver and the antenna structure allows to optimize the common interface. A differential antenna interface and the reduction of mutual coupling by controlling the antenna termination impedance are both experimentally verified. The realized four element active integrated antenna array reaches an excellent long-term stability of transmission gain and phase. For a com- x Abstract plete receiver system, including conversion to the digital baseband, gain variations of less than ±0.1 dB are measured over three days in an office environment. These fluctuations are mainly due to changes of the ambi- ent temperature and are similar for all channels. The resulting distortions of the array pattern, therefore, stay even lower. The dynamic behavior of the transfer functions is studied for the case of jointly switched power amplifiers, which experience strong thermal changes due to self heating. It is found, that fast changes are correlated well for the employed mono- lithically integrated circuits. For typical burst lengths up to several mil- liseconds it is demonstrated that no significant pattern error occurs. On a system level, the noise correlation between the individual chan- nels is studied. Some correlation of the receiver noise was noticed due to correlated spurious occurring in the digital receiver part. A small phase decorrelation with a 1/f-characteristic is found, which is almost negligible for most practical applications. Furthermore, available calibration methods are studied and applied to the active array. The effect of mutual coupling is removed using an inverted coupling matrix. Over a the range of the element beam-width the behavior of the calibrated array can be approximated by the simple geometrical ray-model. A novel transmission line method is proposed, which allows to calibrate the variations of the active circuits without the need for a precise divider network. With the help of the new calibration method and the highly integrated frontend a novel type of conformal active array is demonstrated. The circuitry is first assembled in standard planar technology and then bent to the final shape, which enables a low cost production. Further advantages of this array are a reduced mutual coupling and a wider angular range of operation. Zusammenfassung

Das Ziel der in dieser Dissertation vorgestellten Arbeit ist die Imple- mentierung und Charakterisierung eines 5 GHz aktiven Antennenarrays. Planare aperturgekoppelte Patchantennen werden mit monolitisch inte- grierten aktiven Schaltungen kombiniert um ein kompaktes sowie robu- stes Antennen-Frontend zu erhalten. Das realisierte Ger¨at kann sowohl in drahtlosen Datennetzwerken, als auch in einem Messger¨at zur mehr- dimensionalen Funkkanal-Charakterisierung (channel-sounder) eingesetzt werden. Um eine Anwendung im Messsystem zu erm¨oglichen, ist das Front- end auf niedriges Rauschen sowie eine hohe Linearit¨at optimiert. Eine klassische Uberlagerungsempf¨ ¨anger-Architektur wurde ausgew¨ahlt um die n¨otige Flexibilit¨at sicherzustellen, wenn das System an verschiedene Um- gebungen angepasst werden soll. Die Variationen der aktiven Schaltungen k¨onnen mit Hilfe eines Kalibrationssignals bestimmt und kompensiert wer- den. Mit Hilfe eines kommerziell erh¨altlichen 0.6 µm GaAs MESFET Pro- zesses kann das komplette HF-Frontend, bestehend aus rauscharmen Ver- st¨arkern, einem Filter zur Spiegelfrequenzunterdruckung¨ aus konzentrier- ten Elementen und einem Frequenzumsetzer, auf einen einzigen Chip von 3.3 mm2 Gr¨osse integriert werden. Dabei wird eine state-of-the-art Ein- seitenband-Rauschzahl von weniger als 4 dB, sowie eine Spiegelfrequenz- unterdruckung¨ von mehr als 35 dB erzielt. Ein aktiver Schalter wird vor- geschlagen, um zwischen dem Empfangs- und dem Kalibrationssignal zu w¨ahlen, ohne die Rauschzahl zu verschlechtern. Die M¨oglichkeiten moder- ner siliziumbasierter Technologien werden mit einem 11 GHz Empf¨anger auf einem 47 GHz SiGe Prozess demonstriert. Mit der Verwendung des MOSFETs als resistiven Mischer k¨onnen gleichzeitig eine niedrige Ein- seitenband-Rauschzahl von 7 dB, sowie ein hoher Eingangskompressions- punkt von −14 dB erreicht werden. Durch die Entwicklung einer Ersatzschaltung kann der effiziente Ent- wurf der passiven Antennenstruktur nach gegebenen Spezifikationen ver- einfacht werden. Das Modell wird heuristisch zur Erfassung der gegen- seitigen Antennenkopplung zwischen benachbarten Elementen erweitert. xii Zusammenfassung

Der gleichzeitige Entwurf von Empf¨anger und Antennenstruktur erlaubt die gemeinsame Schnitstelle zu optimieren. Sowohl eine differentielle An- tennenschnittstelle, als auch die Reduzierung der gegenseitigen Antennen- kopplung durch die optimale Wahl der Antennenfusspunktimpedanz wer- den beide experimentell best¨atigt. Die Ubertragungsfunktionen¨ des entwickelten vierelementigen aktiven integrierten Antennenarrays zeigen eine exzellente Langzeitstabilit¨at von Phase und Amplitude. An einem kompletten Empfangssystem, welches auch die Konvertierung ins digitale Basisband beinhaltet, wird eine Ver- st¨arkungs¨anderung von weniger als ±0.1 dB uber¨ drei Tage in einer Buro-¨ umgebung festgestellt. Die Anderungen¨ sind haupts¨achlich von der Um- gebungstemperatur abh¨angig und sehr ¨ahnlich fur¨ alle Kan¨ale. Dadurch bleibt der Einfluss auf das Richtdiagramm gering. Das dynamische Ver- halten der Ubertragungsfunktion¨ von gemeinsam geschalteten Leistungs- verst¨arkern wird untersucht, welche starken Temperaturschwankungen durch Selbsterhitzung unterworfen sind. Es wird festgestellt, dass schnelle An-¨ derungen bei den verwendeten monolithisch integrierten Schaltungen gut korreliert sind. Es kann gezeigt werden, dass w¨ahrend einer typischen Ubertragungsdauer¨ von wenigen Millisekunden keine nennenswerten Feh- ler im Richtdiagramm auftreten. Auf Systemebene wird die Korrelation des Rauschens auf den ein- zelnen K¨analen untersucht. Ein korrelierter Rauschanteil erscheint, der auf korrelierte unerwunschte¨ Nebenprodukte der Signalprozessierung im digitalen Empf¨anger zuruckzuf¨ uhren¨ ist. Eine geringe Dekorrelation der Ubertragungsphasen¨ mit 1/f-Charakteristik kann festgestellt werden, die jedoch fur¨ die meisten praktischen Anwendungen ohne Bedeutung ist. Bekannte Kalibrationsmethoden werden untersucht und auf das aktive Array angewendet. Die Auswirungen der gegenseitigen Antennenkopplung k¨onnen mit Hilfe einer invertierten Koppelmatrix kompensiert werden. Das Verhalten des kalibrierten Arrays kann innerhalb der Keulenbreite der Einzelstrahler durch das einfache strahlenoptische Modell angen¨ahert werden. Es wird eine neuartige Kalibrationsleitungs-Methode vorgeschla- gen, mit welcher die Variationen der aktiven Schaltungen kalibriert werden k¨onnen ohne ein pr¨azises Verteilernetzwerk zu ben¨otigen. Mit Hilfe des neuen Kalibrationsverfahrens und den hochintegrierten HF-Frontend kann eine neue Art von konformen aktiven Arrays demon- striert werden. Die Schaltung wird zu¨achst in gew¨ohnlicher planarer Tech- nologie hergestellt, um dann in die endgultige¨ Form gebogen zu werden, was eine kostengunstige¨ Produktion erm¨oglicht. Weitere Vorteile dieses Antennenarrays sind eine geringere Antennenkopplung und ein gr¨osserer abgedeckter Winkelbereich. 1 Introduction

1.1 Motivation The idea of grouping antennas into an array is almost as old as the history of radio wave transmission itself. The first antenna array reported dates back to 1901. It was designed by Guglielmo Marconi, who intended to use it for his transatlantic wireless communication experiment [1]. Unfortu- nately the two erected arrays at both coasts of the Atlantic were destroyed by strong storms before they could ever be used. Since the 1950s, antenna arrays are attractive for radar systems [2, 3], where the phased array principle allows to point the antenna beam to different positions by changing electrical parameters rather then mechan- ically turning the structure. Due to the high associated costs the use of these arrays was mainly limited to military systems and a multitude of theoretical and experimental works were conducted with this background. With rapidly growing possibilities of digital signal processing, antenna arrays became attractive for wireless communication systems. Replac- ing the simple non-directive antenna by an electronically controlled beam principally offers several advantages: • For the same link distance, less transmit power is needed. Battery power and expensive power amplifiers can be saved. • The higher directivity mitigates multi-path propagation effects. This changes the fading properties and reduces the intersymbol interfer- ence. • Interfering sources can be suppressed by placing pattern zeros into their directions. • It becomes possible to separate signals arriving at different angles. This allows to have several individual communication links simulta- neously and at the identical frequency, as illustrated in Fig. 1.1. • Additional antennas add diversity and multiplexing gain. 2 Introduction

user 1 inl2signal 1 signal 2 spatial filter

user 2

Fig. 1.1 Smart antenna acting as a spatial filter, separating the signals from mobile users 1 and 2 at the base station.

Especially for the ability of sharing the radio resources the now- called smart antenna was proposed to solve the frequency congestion, which fol- lowed the enormous worldwide success of the digital mobile communica- tion standards like GSM. Motivated by these expectations, several exper- imental systems have been built for the DECT [4, 5], the GSM/DCS1800 [6, 7] and other standards. Also first commercial products are available. It is concluded that the system capacity can at least be doubled with existing techniques [8]. The performance is generally limited by the fi- nite suppression of unwanted signals, which has two reasons: hardware imperfections do not allow a perfect signal cancellation and multi-path propagation causes each signal to arrive from several angles, which also prevents the complete separation of signals. In the late 1990s an entirely different approach was proposed: from the standpoint of information theory, a system with multiple transmit and receive antennas provides a higher theoretical capacity to exchange data, regardless of the geometrical arrangement. This capacity can be exploited by applying suitable space-time coding schemes [9,10]. A sys- tem block diagram is shown in Fig. 1.2. The wave propagations are now seen as a multi-dimensional vector channel. It was found that multi-path propagation improves the statistical properties of this channel and, there- fore, turns into an advantage. Channel measurements in realistic scenar- ios [11,12] confirm the potential of these so-called multiple-input multiple- output (MIMO) systems. This new field has triggered a lot of research interest all over the world. It can be anticipated that the commercial application of this new technique will first enhance wireless LANs, where higher data rates are already requested nowadays. To find the optimum system design it is mandatory, among other is- sues, to understand the characteristics of multi-dimensional transmission channels. These can be studied using a suitable channel sounder [13]. Organization of this work 3

a,b a,b space−time space−time decoder a coder H a

b a,b b

S H−1, S−1 vector channel

Fig. 1.2 Multiple-in multiple-out (MIMO) system. The signals are separated by appropriate coding techniques.

For all applications, smart antennas, MIMO systems and channel sound- ing, compact, robust and low-cost antenna arrangements are favorable. At higher frequencies long RF-signal interconnect lines become critical due to losses and reflections and should be avoided. These demands are fulfilled by active integrated antenna arrays, which embed the RF-frontends into the antenna structure [14]. In addition to the avoidance of critical high frequency signal transmission lines, this allows an arbitrary placement of the antenna elements as well as a remote mounting of the whole array. Goal of this work described here was to design and characterize such an active antenna array, using RF-frontends which are monolithically in- tegrated on commercially available semiconductor process. The frequency range of 5.15–5.875 GHz is selected to include both, the 5.8 GHz ISM (industrial, scientific, medical) unlicensed band and the 5.2 GHz band re- served for wireless LAN applications in Europe. In the first place the antenna frontend was projected to meet the requirements of a multi- dimensional channel-sounder, but with a generic design it can be applied to smart antenna or MIMO type communication systems in the same way.

1.2 Organization of this work This work starts with the review of the system specifications and the derivation of a suitable architecture. The developed hardware – based on this architecture – is described in chapter 3 and 4, separated into active and passive components, respectively. The design and measurements are presented. Chapter 5 then focuses the attention on the calibration of antenna arrays, before two complete antenna frontends are reported in chapter 6. In addition to the design, results of the investigated noise correlation and array calibration are given.

2 System design

This chapter describes the design considerations which lead to the final system architecture and to the specifications of the individual building blocks. The standard single channel case is discussed before it is focused on the aspects associated with a multiple antenna system. These include the antenna placement as well as the distribution of the local oscillator signal. The latter leads to a discussion of the noise effects in multiple an- tenna systems, where in particular the correlation between the individual channels plays an important role. Last, a smart antenna test-bed is de- scribed, which utilizes the realized antenna frontend and which is helpful to characterize the developed hardware on a system level, as it will be seen later. 2.1 Receiver design 2.1.1 Dynamic range Fig. 2.1 shows a possible power density spectrum as it could be observed at the input of a wireless mobile communication system. It illustrates one of the basic challenges in radio receiver design: the power of the information carrying signal varies over a broad range. Dependent on the distance between transmitter and receiver it can be close to the noise floor, or high enough to saturate the sensitive input stages. Furthermore, other signals from other communication systems, radars, microwave ovens or unwanted radiation from electrical devices might be present. Often the power of these interfering signals surmount the power of the weak signal-of-interest by orders of magnitude. It is therefore mandatory, that a wireless receiver exhibits a high dynamic range to avoid nonlinear effects caused by strong signals, while keeping it sensitive on the other hand. Generally, nonlinearities lead to two unwanted effects: first, a strong signal exhibiting a certain level drives the circuit into compression. This leads to a decrease of gain, not only for this strong signal, but also for other signals received. This phenomenon is known as blocking. Second, even 6 System design

20

0

−20 signal −40 of interest −60

−80

−100 power [dBm] −120

−140

−160

−180 0 5 10 15 freq [GHz]

Fig. 2.1 Example of a received power density spectrum (solid) and blocking mask of HIPERLAN/2 [15] (dashed line). small nonlinearities give rise to intermodulation products. These products appear at all sum- and difference- frequencies fIM = ±nf1 ± mf2 of two signals at the frequencies f1, f2 and their m-th and n-th harmonics. The most critical are the 3rd order products that appear at the frequencies 2f1 − f2 and 2f2 − f1. If f1 and f2 are closely spaced, these products are also located very close and cannot be simply filtered away. For any practical system, reasonable assumptions have to be made concerning the occurring power levels and interferer. Being able to handle these specified cases, the system should master most of all relevant prac- tical situations. Therefore, wireless standards typically define a blocking mask. To be compliant with the standard the receiver needs to tolerate any signal within this mask and still receive the signal-of-interest. For the present system, the blocking mask defined by the European HIPER- LAN/2 wireless LAN standard [15] is used as a reference. This mask is also shown in Fig. 2.1. For the circuit design it is helpful to use some easy-to-determine uni- versal figures, which describe noise and linearity behavior of the functional blocks. Fig. 2.2 illustrates the common conventions. At the output a certain noise floor is present, which consists of the am- plified thermal noise (Pnoise = kT0∆f) and a contribution of the receiver circuit, described by the noise figure (NF) Receiver design 7

Fig. 2.2 Common conventions to define noise and linearity in an analog system.

P NF = noise,out , (2.1) G · kT0∆f where G is the system gain. The noise figure characterizes the signal- to-noise ratio and, therefore, the signal quality. The required number depends on the used modulation scheme as well as on the tolerated bit- error-rate. The 5 GHz wireless LAN standards HIPERLAN/2 and IEEE 802.11a both allow a noise figure up to 12 dB. To allow the application of this antenna frontend in a channel sounder, it is demanded here, that the noise figure stays as low as possible. This is achieved by a low-noise preamplifier and an appropriate gain planning, which will be described below in section 2.1.2. To specify the linearity, two figures are commonly used: the condition where the output power stays 1 dB below the power expected in linear operation is referred to as the 1 dB-compression point (P1dB). This com- pression point can be directly obtained using a large-signal measurement. Second, the intermodulation behavior is described by the 3rd-order inter- cept point (IP3). This measure is obtained from a measurement at small power levels and is defined as the extrapolated intersection of the cubically growing 3rd-order intermodulation product with the linearly raising out- put power. If the system transfer function is assumed to be a memoryless third-order power series, it can be shown that the IP3 is 9.6 dB higher than the 1 dB-compression point. This is not generally valid, but it holds for most practical microwave systems. Therefore, it is usually sufficient to specify only one of the two values. For this system, a 1 dB input compression point higher than −20 dBm 8 System design Table 2.1 Initial specifications of the individual receivers. Specification Value RFfrequency 5.125–5.875GHz RF dynamic range -75 – -35 dBm IF bandwidth 250 MHz IF output power -40 – 0 dBm converter noise figure < 4dB input 1dB-compression point > -20 dBm linearity/ IP3 > -10 dBm power consumption not critical is demanded. This fulfills the specifications of both wireless LAN stan- dards, HIPERLAN/2 [15] and IEEE 802.11a. In addition to the require- ments of these standards, a high bandwidth at the intermediate frequency (IF) of 250 MHz is needed for the channel sounding application. To give an overview, the RF specifications are summarized in Tab. 2.1. 2.1.2 Receiver architecture The classical superheterodyne architecture, on the one hand, is a very flexible concept, which can be adapted to virtually any wireless receiver. On the other hand, it always requires appropriate filters to suppress the occurring image bands. Sharp transitions between pass- and stop-band might be required as well as a high stop-band attenuation. Both might be difficult to realize with monolithically integrated circuits due to in- ternal coupling and high component variations. For integrated receivers, therefore, different architectures are preferred: The direct conversion receiver avoids the image-frequency problem by converting to an intermediate frequency of zero. As a drawback, it is sensitive to self-mixing effects, which appear as a DC component as well. Compared to a conversion to a higher IF, the direct conversion receiver suffers from a stronger 1/f-noise. These problems are avoided by the low-IF or image reject receiver [16]. The received signal is converted down using two orthogonal LO signals of 0◦ and 90◦ phase difference. This yields a quadrature and in-phase component. This complex signal contains the image and wanted band as negative and positive frequencies, respectively. The complex filtering can be carried out in the analog or the digital domain, alternatively. The image-reject architecture is sensitive to amplitude and phase mismatches in the quadrature signals. This becomes critical if a wide bandwidth is needed. Receiver design 9

RF frontend

LO2

LO1

Fig. 2.3 Simplified scheme of the super-heterodyne receiver concept. The RF front-end is the main focus of this work.

For the current design, a high flexibility and a modular system is wanted. Therefore, the classical superheterodyne architecture is favored. To relax the specifications on the necessary filters, a sufficiently high first IF of 1.45 GHz is chosen. A LO frequency above the signal frequency (fLO = fRF + fIF ) is selected for two reasons: the local oscillators need less relative tuning range for the same RF-bandwidth and the image band moves to a “quiet” band, reducing the attenuation the image filter needs to provide. The first conversion step is followed by a second downconversion to a 150 MHz. Fig. 2.3 shows the simplified block diagram of the com- plete analog processing chain. This work only focusses on the indicated RF-frontend. The frontend itself starts with a coarse frequency pre-selection to re- duce the blocking capability by out-of-band signals. Here, this task is inherently performed by the frequency-selective behavior of the antenna and the subsequent amplifier. This first amplifier is optimized for low noise, as it determines the main contribution to the system noise-figure. The bottleneck which limits the dynamic range is typically represented by the mixer. A resistive mixer is selected here, as passive mixer show a significantly higher linearity. Another advantage for this particular system can be explained looking at Fig. 2.4, where the typical conversion char- acteristics of both types are shown: the active mixer typically achieves a conversion gain while requiring less LO power. At rising LO power the gain drops quickly, resulting in a small range of operation. The resistive mixer shows a constant performance over a broader interval of LO pow- ers. This is advantageous in a multiple channel system, where the LO needs to be distributed and no identical powers can be guaranteed. The missing conversion gain and higher LO power requirements can easily be compensated by additional amplifiers. Starting from a given input compression point P1dB ≈ +5 . . .+10dBm and conversion loss g ≈−8 . . . − 6dB of the mixer, a gain of 20 to 25 dB is needed from the low-noise amplifier and filter section. This avoids a 10 System design

active Gilbert mixer

LO power conversion gain [dB] passive FET mixer

Fig. 2.4 Typical conversion characteristics of passive FET mixer (solid line) and active Gilbert mixer (dashed line). The active mixer provides conversion gain at lower LO power, but the range of operation is limited. compression of the mixer, while the signal level at the output of the mixer remains sufficiently high above the noise floor to keep its influence on the system noise figure negligible. Approximately 20 dB of gain need to be provided by the IF amplifier to reach the required output power-level. RF and IF amplifiers both have to show high linearity, as the received signals generally can show a non-constant envelope modulation. If power consumption is critical, the LO amplifier can be a nonlinear amplifier to reach a higher efficiency. It has to be considered that a nonlinear amplifier could upconvert 1/f-noise, leading to additional uncorrelated phase noise in a multiple antenna system (see also section 2.3). 2.2 Multiple antenna system 2.2.1 Noise and linearity Moving from the single-channel receiver to a multi-channel system, there are two general differences: the signal power is shared by all elements and the increased effective antenna area provides additional antenna gain. This has different implications for the transmit and the receive case. In an N-element transmit array only the n-th fraction of the overall output power needs to be delivered by a single element. Smaller power amplifiers can be chosen, which are easier to fabricate, cheaper and have less power dissipation. If the signal-to-noise ratio is kept constant, or the effective isotropic radiated power (EIRP) is limited by legal regulations, the individual output power even has to be reduced by 1/N 2 to compensate for the array gain. In the receive case, the distribution of the signal energy cannot di- rectly be turned into an advantage: unless the transmit power is lowered, Multiple antenna system 11

A D processor

A space−time D

Fig. 2.5 Digital antenna combining allows space-time processing any receiver has to handle the same power as in a single antenna sys- tem. Furthermore, a high compression point and good linearity should be maintained to reject out-of-band interferers. More options are given by the higher signal-to-noise ratio, which increases with a factor of N: this allows either to tolerate a higher receiver noise figure or to decrease the transmitted power at the other side of the link. The latter is very attractive to enhance the battery lifetime in mobile applications. Here, the noise-figure specifications are left unchanged to ensure the lowest possible noise figure for the channel-sounding application.

2.2.2 Antenna combining methods Beamforming requires to control the amplitude and phase of the individ- ual signals and to add them. This is principally possible at any stage of the processing chain. For a long time it was the only economically pos- sible solution for large array antennas to perform the combining at the RF. This eliminates the need for several parallel downconverter branches and high performance digital signal processing. Although the preference has gradually shifted towards digital combining, some approaches have been presented lately which use analog combining to reduce the power consumption and the complexity of the system [17, 18]. Analog combining does not allow a multi-dimensional channel-sounding or the realization of a space-time coded system. Therefore, in the present system, the signals are entirely converted to the digital baseband as illus- trated in Fig. 2.5. At the expense of an increased hardware effort, this approach shows several significant advantages: the weights can be set with the numerical precision of the signal processor. The data of each channels is available to adapt appropriate weights. Furthermore, it is easily possi- ble to store a number K of samples and compute the received signal r as a weighted sum over different channels sn and time delays kT : 12 System design

N K r(t + K · T )= wn,k · sn(t − kT ), (2.2) n=1 X kX=1 with T being the time between to samples and wn,k the complex weight for channel n and delay kT . This corresponds to a varying directional pattern for each time delay. In a typical mobile wireless scenario, different paths of propagation are observed; each with a certain angle of incident and delay. Using space-time processing, different signals can be separated more precisely than with a static directional pattern achieved with analog antenna combining. Several attempts have been exercised to reduce the hardware necessary for array processing. They include time-domain [19] or frequency-domain [20] multiplexing prior to converting the signal to baseband. However, it is found that the lower component count is paid for by higher bandwidth requirements or a large number of frequency synthesizers, respectively. 2.2.3 Local oscillator distribution For all architectural choices except RF combining, a local oscillator signal has to be provided at all receiver branches. As shown in Fig. 2.6, this can be done by either using a splitting network, or by generating the signal locally at each antenna element. In the latter case it is mandatory to establish a fixed phase relationship between all LO signals. This can be done by deriving the local signal from a reference, either by using frequency multiplication, injection locking or a phase-locked loop. The constant phase relationship has to be maintained for every needed LO signal, including the clock frequency of the analog-digital converter. In this work, the distribution of the high frequency LO signal was selected to guarantee the best LO phase correlation possible. The effect of non- constant phases is discussed in section 2.3.4.

phase reference

common LO individual LO

Fig. 2.6 LO distribution: central or local LO Multiple antenna system 13

2.2.4 Antenna placement Smart antennas Most of the known smart antenna techniques base on Huygens principle of elementary sources and quasi-optical wave propagation. This implies that, if a coherent signal is emitted from several positions ~xn, this leads to constructive and destructive interference patterns. 2π With the propagation constant k = λ the phase shift at the location ~y is given by ∆φ = −k| ~xn − ~y|. (2.3) If each source has its own amplitude and phase, expressed in the complex “weights” wn and gn (φ, θ) is the individual element pattern, the superim- posed signal at point d~ can be calculated as

N g(~y)= wn · gn (φ, θ) · exp(−ik| ~xn − ~y|), (2.4) n=0 X whereby the different path losses are neglected. The most common type of array is the uniform linear array (ULA), because it offers a Fourier relationship between antenna weights and array pattern. It consists of N transmitters, placed in a row with a regular spacing of d between two elements, as seen in Fig. 2.7. In the far-field, the phase shift for the element at position n referred to element at n =0 is ∆φ = −knd · sin(φ). (2.5) The superimposed signal seen from angle φ then is

N−1 g(φ)= wn · gi (φ) · exp(−knd · sin(φ)), (2.6) n=0 X where the directional pattern gi (φ) now is similar for all elements. With the substitution sin(φ) = u this leads to an expression similar to the discrete Fourier transformation between the u- and n-domain:

wave front

φ l ∆

d

Fig. 2.7 Uniform linear array 14 System design

N−1 g(u)= gi (u) · w(n) · exp(−kd · n · u) (2.7) n=0 X As a consequence of this, all relationships known from signal processing in the time and frequency domain can be applied on this transformation between spatial and angular domain. Similar to the Nyquist rate, an antenna spacing of d ≤ λ0/2 is needed to represent the entire angular range −1 ≤ u ≤ 1. Otherwise replica of the array pattern appear, which can not be controlled independently. For this particular antenna array it has to be considered that it does not have a fixed frequency, but a possible range of 5.15 to 5.875 GHz. Strictly it would be reasonable to choose an antenna spacing of λ0/2 ac- cording to the highest occurring frequency. Here, the element spacing is approximately selected for the center of the band at 5.5 GHz, resulting in a center-to-center distance of 27.3 mm. For higher frequencies this causes replica to appear. These replica appear outside of the beam width of the antenna elements and can be tolerated. So far it was assumed that all antennas operate independently from each other. In the realistic case of mutually coupled antennas, this slightly increased distance helps to lower the antenna coupling. The antenna cou- pling will be discussed more detailed later in chapter 4.3.

MIMO systems If the frontend is intended to serve as a part of a MIMO-system, the angle of view has to be changed entirely. It is one reason for the high attrac- tiveness of MIMO-systems that the geometrical antenna arrangement is simply included into the unknown channel response. On the one hand this allows almost all arbitrary antenna placements, as long as a data trans- mission is possible between transmitter and receiver. On the other hand, this makes it more difficult to determine the best configuration. A given antenna setup in its environment can be evaluated as follows: the system of M transmit and N receive antennas can be seen as a vector channel HM×N . Without considering the concrete coding techniques, an upper limit for the instantaneous channel capacity of a stochastic MIMO channel can be given from theoretical considerations [9]:

ρ C = log det I + HHT , (2.8) 2 N  t  where ρ is the average SNR at each receiver, Nt the number of trans- mit antennas and I is the identity matrix. The objectives of antenna Multiple antenna system 15

1

0.5 x) ∆ R(

0

−0.5 0 0.5 1 1.5 2 2.5 3 ∆x/λ

Fig. 2.8 Correlation coefficient of two ideal antennas spaced by ∆x. The re- ceiving signals arrive from angles uniformly distributed between 0 and 2π in the horizontal plane. A first correlation minimum occurs at ∆x = 0.4λ. placement are to achieve a high average capacity and a low outage prob- ability Pr (C ≤ Cmin). The difficulty is, that the capacity resulting from Eqn. 2.8 depends on the actual propagation condition. To optimize the antenna arrangement, the capacity needs to be determined on a statistical base, which is beyond the scope of this work. For simplicity it can be assumed that the fading behavior of the M × N single channels is not correlated. The necessary coherence distance can be estimated, if it is assumed that all waves arrive with a uniformly distributed angle from 0 to 2π in the horizontal plane. Then the the spatial correlation is given by [21]:

2π|∆x| R (∆x)= J , (2.9) x 0 λ   where ∆x is the separation distance between two antennas, λ the wave- length and J0 the first kind Bessel function of 0th order. Fig. 2.8 depicts the resulting graph: the correlation coefficient first drops quickly with increasing antenna spacing and becomes zero at ∆x ≈ 0.4λ. At higher distances the correlation function shows an oscillating behavior with a slowly decreasing envelope. 16 System design

It can be concluded that a linear array with a spacing of λ0/2 at 5.5 GHz does not represent the best solution, but to maintain the com- patibility with beamforming applications it is a good compromise. It has to be remarked that this simplified approach assumes an en- vironment with rich scattering. For a set of antennas mounted with no local scatterers near, like e.g. a base station on a mast top, the coherence distance can be several tens of wavelengths. Also it is interesting to investigate different antenna orientations. It was demonstrated that a compact antenna arrangement can be found which makes use of polarization and pattern diversity, improving the per- formance at the same time [22]. 2.3 Noise in multiple antenna systems 2.3.1 Signal and noise model One of the main advantages of multiple antenna systems is their ability to increase the signal-to-noise ratio in a given situation. Any beamforming operation bases on the addition of the N received signals sn to

N r(t)= sn(t)+ nn(t) , (2.10) n=1 X   where nn describes the additional noise at each receiver. Assuming the signals sn originate from the same source, correlated and perfectly in phase (by adequate pre-processing), the signal amplitudes add1

N 2 2 |rsignal(t)| = sn(t) , (2.11)

n=1 X while the noise contributions are uncorrelated (E {ni · nj } =0 and E{. . .} is denoting the expectation value) sum up in terms of power

N 2 2 |rnoise(t)| = |nn(t)| . (2.12) n=1 X 2 2 If noise and signal powers are equal for all elements E{|nn| } = n0 and 2 2 E{|sn| } = s0 respectively, the signal-to-noise ratio is proportional to the number of elements

1This expression violates the law of conservation of energy, indicating that Eqn. 2.10 cannot be realized using passive components. This affects the calculation of the overall gain. However, the calculation of the signal-to-noise ratio is not affected. Noise in multiple antenna systems 17

mutual coupling c1 G c2 noise waves

Fig. 2.9 Generation of correlated noise by mutual coupling s SNR = N · 0 . (2.13) n0 This demonstrates that the basic property that makes the difference be- tween signal and noise is the correlation between the channels. Therefore it is important to identify and control all effects that lead to correlated noise and decorrelation of signals, respectively.

2.3.2 Noise correlation Several mechanisms are known, that lead to correlated noise: • If noise is received, which is radiated from a single source, e.g. a body with a significantly higher temperature than the receiver, or another device. This kind of correlated noise principally can not be distinguished from a signal and has to be treated as interferer. • The pre-amplifiers at the input produce noise. The main portion of this noise is found at the output, where it is amplified together with the signal. A fraction of the amplifiers noise is also found at the input. Through mutual antenna coupling it reaches the neighbor element where it gets amplified causing correlated noise [23], as sketched in Fig. 2.9. The noise coupling ratio is defined as Gc1Γi , where G is c2 the amplifier gain, c1 and c2 the forward and backward travelling noise waves and Γi the active reflection coefficient of the array. For applications in communication systems one aims to realize low mutual antenna coupling. If it is assumed that the output noise c2 is exceeds the noise wave at the input by the amplifier gain Gc1 ≈ c2, the coupling coefficient is in the order of the s-parameter S21 measured between to antennas in the passive array. Typically this is −15 dB or lower. • The system itself radiates signals or interaction occurs over the com- mon power supply. Notably the digital part with its fast transients and strong clock signal is known to emit unwanted energy. This power couples to the analog part and appears as spurious responses after analog/ digital conversion. Theoretically it is possible to completely suppress it, but in practical applications a certain amount of coupling 18 System design

j · a · φ ⇐⇒ a power density φ f0 frequency

Fig. 2.10 Phase noise equivalence in frequency and time domain

has to be tolerated. Although the resulting spurious responses only account for a small fraction of the overall noise power, it was observed that they can be highly correlated [24]. The influence of this effect is difficult to predict, as it is based on parasitic effects. 2.3.3 Phase noise Phase noise is a phenomenon associated with frequency sources. Noise sources affect the frequency generation process which leads to small de- viations from the ideal sinusoidal waveform. The output of a frequency source showing phase noise can be written as

sLO(t)= s0 · cos 2πf0t + φ(t) , (2.14) where φ(t) is the temporal phase deviation from the ideal signal. In the frequency domain this leads to a broadening of what should be a single line at the frequency of oscillation. A common method of characterizing the phase noise of frequency sources is to give the spectral density (usually referred to 1 Hz bandwidth) at a certain spacing ∆f away from the carrier. Its power density is expressed relatively to the carrier using the logarithmic measure dBc/Hz. For small phase deviations a simple relationship between frequency domain characterization and time domain phase jitter can be derived as π illustrated in Fig. 2.10. Under the assumption that |φ|  2 , the com- plex bandpass representation (omitting exp(j2πf0t)) can be linearized as follows: a · exp(φ(t)) ≈ a + j · a · φ(t) (2.15) The power a2 corresponds to the carrier power. Therefore, the power contained in the noise sidebands must be the orthogonal perturbation. For a given bandwidth B the mean phase deviation |φ|2 is [25] B P (f) |φ|2 ≈ noise df. (2.16) P Z0 carrier Noise in multiple antenna systems 19

I (a + jb) · exp(j2πf t) 0 (a + jb) · exp(−φ(t)) Q 0° 90°

exp(−j(2πf0t + φ(t))

Fig. 2.11 Effect of phase noise on signal down-conversion

This expression is only valid if the spectral density Pnoise is clearly result- ing from phase noise. The presence of LO phase noise in communication systems has two undesired effects: • Closely spaced channels cannot be separated • Phase noise modulates the data signal Fig. 2.11 shows how the data signal is affected: for radio transmission the information is modulated as a complex envelope (a + jb), also referred to as the baseband signal, on a carrier with a frequency f0.

sRF (t) = [a(t)+ jb(t)] · exp(j2πf0t) (2.17) From a system point of view the down-conversion equals a complex mul- tiplication with exp(−j2πf0t). If this LO signal is considered not to be spectrally pure, the phase deviation of the sinusoidal signal translates into the same phase deviation of the complex baseband signal:

sBB(t)= sRF (t) · exp(−j(2πf0t + φ(t))) (2.18) = [a(t)+ jb(t)] · exp(−φ(t)) (2.19) LO phase noise is not the only source which leads to this effect: oscillator drift or a moving transmitter or receiver also cause phase shifts of the received signal. To allow the correct detection of the transmitted symbol, this offset has to be estimated and compensated by an appropriate clock recovery scheme. This typically requires, that the change of the phase difference is significantly slower than the phase changes of the data signal. The absolute boundary on the tolerable phase noise depends strongly on the data rate, the modulation scheme and the tolerated bit error rate and, therefore, cannot be given here. The effect of phase noise on multiple antenna systems is discussed below. 2.3.4 Correlation of phase noise As mentioned in section 2.2.3 in a multiple antenna system it is important to maintain a constant phase relationship between all LO signals. There- fore, it is required to derive all signals from a reference source. A real 20 System design

LO signal with phase noise has a small bandwidth and is coherent over a certain time. The differences of line lengths which occur in a typical LO distribution network in an array system are in the range of wavelengths. This is not sufficient to cause any decorrelated LO phases at the individ- ual receivers. Nevertheless, there are possible sources of additional phase noise: • A phase-locked loop (PLL) is an attractive method to lock a local high-frequency oscillator on a reference signal. Within the loop band- width the oscillator phase noise is suppressed, its phase follows the reference. The phase noise outside the loop bandwidth adds as un- correlated fraction to the reference phase noise. If the phase detector and loop filter are not noiseless, their noise also contributes to the uncorrelated phase noise. • All non-linear circuits like frequency doubler, tripler or amplifiers with high efficiency principally show mixing effects. This leads to an up-conversion of 1/f-noise, usually present in semiconductor devices. It appears as additional phase noise. If the LO signal at the nth branch contains partly correlated phase noise, the phase deviation φn(t) can be split into a fully correlated mean phase deviation φ(t) common for all channels and an independent relative 0 angle φn(t): 0 φn(t)= φ(t)+ φn(t) (2.20) Omitting the amplitude noise, Eqn. 2.10 becomes

N 0 r(t)= sn(t) · exp −j(φ(t)+ φn(t)) , (2.21) n=1 X
where the correlated phase noise can be moved out of the sum:

N 0 r(t) = exp −jφ(t) sn(t) · exp(−jφn(t)) (2.22) n=1
X This demonstrates that the correlated portion of phase noise φ(t) affects the output signal, where it has to be estimated and compensated together with possible frequency offsets and Doppler-shifts. The uncorrelated frac- 0 tion φn(t) disturbs the signal combining process itself, leading to a degra- dation in the array interference cancellation ability [25]. The effect of additional phase noise generated by LO-amplifiers or mix- ers generally is difficult to measure. The phase noise of low-noise signal generators, needed for the measurement, typically exceeds this contribu- tion and covers the effect. Testbed architecture 21

At the time of the design, no publications of experimental results re- garding the phase noise correlation were available. It could be demon- strated that a slight phase decorrelation occurs, however at a level low enough for most practical applications. For details see chapter 6.3. In multi-antenna channel-sounding systems the individual paths are often measured sequentially by switching between the antenna elements. In this case the absolute phase-noise value has to be low to ensure the right estimate of the channel capacity [26].

2.3.5 System noise model In a complex system like in Fig. 2.13 several sources of both, phase noise and amplitude noise, are present. Phase noise is introduced at every down- conversion step, including the analog to digital conversion. The additive noise added by each stage is down-converted together with the signal, thus affected by phase noise as well. This additive noise can be assumed not to be correlated with the phase noise. Therefore, the complex noise-vector n = nI + j · nQ becomes

0 n = (nI + j · nQ) exp(−jφ) (2.23) without changing its statistical properties. For the practical case that phase noise contributions are sufficiently correlated, all correlated additive noise sources can be combined in a noise vector ncorr which gives a different noise behavior at the output, depen- dent on the actual beam-forming. The power of all uncorrelated noise sources nuncorr adds to the system output, independent of the beam- forming operation. The uncorrelated and correlated phase noise powers sum up to a single phase shift φn for each branch and one common phase shift φ at the output, respectively. This leads to the system noise model shown in Fig. 2.12. 2.4 Testbed architecture Fig. 2.13 shows the architecture of the entire “SANTRES” testbed. The classical super-heterodyne architecture with two analog conversion steps to 1.45 GHz and 150 MHz is used, as described above. The signal is band- pass filtered and sub-sampled at a rate of 52 MHz. The digital signal then is further processed and decimated by a programmable factor [27]. Subsequently, the signal is either processed or stored. The whole receiver chain from RF to baseband is carried out four times in parallel. The local oscillator signals are generated using programmable synthesizers. To obtain the best possible phase correlation, these signals are divided and equally distributed to all branches. 22 System design

The first down-converter stages, which are the most critical and sensi- tive parts of the system, are directly integrated together with the antenna array. This reduces the influence of phase instabilities of cables and con- nectors, allowing also remote mounting of the active array. To calibrate the phase and amplitude inequalities of the receiver branches it is possible to apply a calibration signal (not shown in the diagram). The calibration aspects are discussed in chapter 5.

s s s s 1 2 3 4 φ φ φ φ 1 2 3 4

ncorr

spatial processing

φ

nuncorr r

Fig. 2.12 System noise model, reduced to a minimum of correlated and uncor- related additive and phase-noise sources.

I,Q processing and storage

LO 1 and decimation active antenna array digital downconverter

A D

LO 2 sample clock

Fig. 2.13 Simplified block diagram of “SANTRES” system, generation and distribution of the calibration signal is not shown here. 3 Integrated circuit design

In this chapter the design and characterization of monolithically integrated circuits is described, which base on the system level considerations in the previous chapter. The available technologies are reviewed and the choice is motivated. Starting from the concepts of the different key components, the whole RF frontend can be integrated on a single chip. In particular this includes also the image filter, which is realized as a lumped-element passive on-chip filter. In addition to these circuits required for the proposed frontend, fur- ther circuits are reported: a power amplifier is described which is used to investigate the calibration of transmit arrays (see chapter 5.4). To ex- ploit the abilities for future high-frequency low-cost applications, a 11 GHz downconverter is demonstrated on a 47 GHz-ft SiGe process. 3.1 Process technology 3.1.1 Choice of technology Today, a large number of various technologies are available, each with its own advantages and applications. High-frequency analog applications tra- ditionally employ III/V-technologies based on GaAs or InP, which benefit from a high electron mobility and offer transit frequencies beyond 500 GHz (InP PHEMT [28]). The enormous success of personal computers, on the other hand, lead to a rapid development of silicon-based technologies, CMOS in particu- lar. The available complementary device allows simple logical gates and therefore a very high integration of digital logic. Silicon shows excellent mechanical and thermal properties at low material costs. This combina- tion of very high integration with easy handling of large wafers results in low costs for high-volume production, and explains why strong efforts were made to improve this technology. The trend of steadily increasing processor speeds caused an aggres- sive down-scaling of the transistor devices below 100 nm gate-length. The 24 Integrated circuit design

Metal2-4umMetal2-4um Metal2

Dielectric Dielectric

Metal1Metal1 Metal1-2um

Dielectric MIMMetalMIMMetal Metal0 NiCr N+N+N+N+ IsolationImplant N-/P-Channel E,D,GMESFET MIMCapacitor NiCrRNiCrResistor

Semi-InsulatingGaAsSubstrate

Fig. 3.1 TQTRx process cross section (from [29]) shrinking of the devices also leads to high transit frequencies of up to 243 GHz (CMOS SOI [30]), which exceeds the performance of many tran- sistors based on more expensive III/V-materials. This qualifies silicon technologies for the use in high-frequency applications with two constraints: the pre-doped conductive substrate leads to losses, which makes the inte- gration of passive components difficult. Second, the device scaling leads to low breakdown voltages. This makes it difficult to handle high power levels. The third important process family can be seen as an extension to CMOS. In SiGe technologies, germanium is added to the base of a sili- con bipolar transistor to enhance its transport properties. This is usually done as an ”add-on”to a standard CMOS process, maintaining all advan- tages for the integration of digital circuits. Additionally, the new device profits from the advanced technology and with accordingly scaled transit frequencies of 375 GHz (SiGe HBT [31]). This enables the fabrication of very fast mixed-signal circuits at the cost of a high number of processing steps and a comparably complex lithography. The advantages and disadvantages of the technologies depend on the actual application and the framework. For a detailed discussion it is re- ferred to [32,33]. For this work a 0.6 µm GaAs technology was chosen for several reasons: for this well-established general-purpose process advanced device models are available. The comparably simple process only requires 16 lithographic steps, reducing the initial costs. This makes this process ideally suited for prototypes and smaller volume productions, where the mask set is a significant part of the overall cost. Process technology 25

Table 3.1 Process key parameters (from [29])

Element Parameter Value enhancement FET threshold voltage +0.15 V (E-FET) transit frequency 18 GHz transconductance 750 mS/mm

depletion FET pinchoff voltage −0.6 V/ −2.2 V (D- /G- FET) transit frequency 20 GHz (D) transconductance 667 mS/mm (D) 567 mS/mm (G)

spiral inductor inductance 0.38 nH. . .6.98 nH quality factor 20 . . .28 (5.2 GHz)

MIM capacitor capacitance 1200 pF/mm2

NiCrresistor sheetresistance 50Ω/2 (σR < 2%)

3.1.2 TriQuint TQTRx process Fig. 3.1 shows a cross-section through the TriQuint TQTRx-process: fab- rication starts from a semi-insulating GaAs substrate, followed by ion- implantation to form the device channels and the device contacts. Differ- ent types of channel implants (E/D/G) are used to yield one enhancement and two depletion transistors with different threshold and pinch-off volt- ages. A Schottky contact with a minimum line width of 0.6 µm forms the gate. The device is typically operated at gate source voltages below the diode barrier voltage, so that only a small leakage current flows into the gate. The gate metal and the first interconnect metal 0 are created in a lift- off process, which forbids to create empty areas encircled by metal planes (”donuts”). This limits the use of this layer to local connections. The global interconnect layers metal 1 and metal 2 are fabricated by an etch- ing process and do not underly these restrictions. They are significantly thicker and, therefore, show less ohmic losses. These upper metal layers are also used to build spiral inductors. Precise resistors are formed using a thin film of NiCr, resulting in a sheet resistance of 50 Ω/2. Capacitors are realized by a thin isolating layer between two metal plates (MIM). An overview of the key parameters of this process is given in Tab. 3.1. 26 Integrated circuit design

(a) (b)(c) (d)

Fig. 3.2 Possible LNA configurations: a) active matching b) inductive current feedback at source (source degeneration) c) RC-feedback d) cascode

3.2 Low-noise amplifier In the previous chapter 2 it was specified, that the low-noise amplifier should have 20 to 25 dB of gain, a bandwidth of 5.15 to 5.875 GHz and a noise figure as low as possible. To minimize the influence of out-of-band interferers, it is advantageous to use a frequency selective amplifier. To meet these requirements, several individually optimized amplifier stages are cascaded. The design starts with the search of a suitable circuit topol- ogy for the single stage.

3.2.1 Input matching In a common-source configuration, the input of any field effect transistor typically shows a capacitive behavior over a broad frequency range, be- fore extrinsic inductances and the gate resistance start to dominate close to the transit frequency. If the gate is reactively matched to a 50Ω in- put port, this leads to very narrow-band network. Additionally to high losses introduced by the low-Q passive components, the matching network becomes very sensitive to process variations, which makes this straight- forward matching inappropriate for monolithic integration. A possible solution would be to use a common gate configuration de- picted in Fig. 3.2a, which can be designed to provide an active input matching. This topology is often found in broadband amplifiers, but due to higher noise and poor linearity it is not the best choice for low-noise amplifiers. As a solution to the reactive input impedance, inductive current feed- back at the source (see Fig. 3.2b) is widely used, because it unifies several advantages: it stabilizes the device, it linearizes the transfer characteris- tic, and it brings together the optimum impedances for power and noise matching. It can be shown [34] that the idealized field effect transistor transforms the source inductance Ls into a resistance seen at the input: Low-noise amplifier 27

1 gmLs Zin = + jωLs + , (3.1) jωCgs Cgs

resistive | {z } where gm is the transconductance, CGS the gate-source capacitance of the device and ω = 2πf is the angular frequency. If the transistor and the feedback inductor are assumed to be ideal components, the real part resulting from this transformation is noiseless. RC-feedback, as seen in Fig. 3.2c, does not require a bulky spiral induc- tor. On one hand, the needed resistor and capacitor require significantly less area and, therefore, make this type of feedback ideal for integration. On the other hand, the resistor represents an additional source of noise, raising the noise figure of the amplifier. For this reason RC-feedback is avoided in the first stages, where the noise contribution is critical. The cascode configuration, shown in Fig. 3.2d, is often found in low- noise amplifiers. Originally proposed to enhance the gain bandwidth prod- uct by minimizing the effect of the Miller-capacitance, it is attractive also for the use in selective amplifiers, because it shows a high gain and mini- mizes the contribution of the consecutive stages to the overall noise figure. As a drawback, the cascode shows a high output impedance. This makes it impossible to realize a broadband power matching to the next stage. Therefore, a cascode topology cannot be used here. Determining the optimum value Ls for the inductive source degener- ation, it has to be considered that a good input power matching does not necessarily yield the best noise-figure. For the practical implementa- tion on a monolithically integrated circuit, two effects play an important role: a strong feedback also decreases the gain of the stage, leading to an increased contribution of the following stage to the overall noise figure. Second, using monolithically integrated inductors with a limited quality factor Q, the inductor becomes a significant source of noise itself and should be kept as small as possible. This leads to a trade-off between amplifier mismatch and noise figure. For the input stage, whose noise contribution is the most critical, a reduced feedback is chosen. The best results are achieved for a real part of around 25 Ω, resulting in an an input return loss between −10 dB and −8 dB. The matching network can be further simplified by scaling the transistor width. For a device of 150 µm gate width, the reactive input matching can be realized with a single inductor. This results in both, a broad matching bandwidth and low losses at the amplifier input and, hence, a reduced noise figure. 28 Integrated circuit design

3.2.2 Device scaling A constraint for device scaling is set by the power handling capability, which determines the compression and intermodulation behavior of the final amplifier. Contrary to power amplifiers, where the maximum power is limited by current and voltage, the output power is limited by the max- imum possible current through the device. This occurs at the maximum gate-source voltage Vd at the onset of gate diode forward conduction. The range of operation is given by Vd and the threshold voltage Vth. The ab- solute power level and the corresponding voltage range are linked by the effective input impedance of the device, which scales with the gate width. It can be roughly estimated that doubling the gate width leads to half of the input impedance, twice the input power at compression and, there- fore, to a compression point increased by 3 dB. The same argumentation applies to the linearity and the 3rd-order intercept point. It is found that a gate width of 600 µm is sufficient to handle the maxi- mum output power of 5 dBm. Therefore, this transistor size is selected for the last stage. The precedent stages obviously need to handle less power and smaller devices can be chosen. To estimate the influence on the system linearity, the total 3rd-order intercept point of two cascaded amplifiers can be given as [34]

1 G −1 P = + 1 , (3.2) IIP 3tot P P  IIP 31 IIP 32 

where PIIP 31 and PIIP 32 are the individual intercept points and G1 is the gain of the first amplifier. If the stages are simply scaled with their gain

PIIP 32 = G1 · PIIP 31 , the intercept point decreases by 3 dB for each stage that is added. The approximate gain of one stage is 8 to 10 dB, thus three stages are required to reach the specified gain of 20 to 25 dB. Simply scaling with the stage gain would result in a gate width of 6 to 17 µm, much less than the width of 150 µm that was found above for the best matching. Furthermore, the large impedance changes from stage to stage limit the bandwidth of the interstage matching networks. At the cost of a higher power consumption, the 150 µm device at the input is maintained and a gate-width ratio of 2 is chosen to obtain a transition to the 600 µm output stage. In addition to the optimal input matching, this practically enhances the bandwidth and lowers the IP3. The sensitivity to blocking by strong out-of-band interferers is significantly reduced, as these mainly saturate the first amplifier stage. Low-noise amplifier 29

Vdd1 VVdd2 dd3

Vg3 Ω 3.7nH 356fF 0.6nH 1.2pF 1k RF

Vg2 5kΩ out Vg1 890fF Ω 2.0nH 2.85nH Ω 5k 1k RF E−FET 150µm in 1.0nH 255fF 356fF 937fF E−FET 1.0nH1.3nH 2.3nH1.3nH 1.0nH 300µm 0.6nH 0.6nH 255fF 1.0nH 2x E−FET 300µm 12.5pF

Fig. 3.3 Schematic of the three stage low-noise amplifier. Bias generation networks are not shown.

3.2.3 Three-stage amplifier Fig. 3.3 shows the schematic of the three-stage cascaded amplifier. As discussed above, the three scaled devices of 150 to 600 µm gate length form the cores of the amplifier stages. The last 600 µm transistor is formed by two 300 µm devices to stay within the valid range of the transistor models. Feedback is used in all stages to obtain input impedances which can be matched over a broad frequency range. The first two stages only use inductive source feedback, while the last stage also employs RC feedback. The input matching consists of a relatively small 1 nH inductor. The drain supply voltages are blocked separately for each stage to avoid unwanted feedback over the bias. To reach the required bandwidth of approximately 13%, the poles of the amplifier transfer function need to be distributed. If amplifiers with identical poles were cascaded, the resulting bandwidth would drop signif- icantly with each additional stage [35]. Here, these poles are practically adjusted by using different center frequencies for the interstage matching networks. For the use in a system, it is advantageous to generate all needed bias voltages locally on the chip to reduce the complexity of external wiring. Another task of this bias network is compensate for process variations. The strongest variation that occurs in the used FET technology is the shift of the threshold voltage Vth. A good immunity against this is achieved, if the gate-source voltage Vgs is adjusted to yield a constant drain current. The applied bias network, seen in Fig. 3.4, uses a current mirror to supply the corresponding Vgs. 30 Integrated circuit design

Vdd D−FET 50 µm

1.2kΩ current source

Ω 5k Vg

E−FET 5pF 15 µm 30kΩ currrent mirror

Fig. 3.4 Bias generation network: a reference current is created by a high-feedback current source and copied by a current mirror. The diodes are needed to keep the mirroring transistor in the saturated region.

The D-FET shows the lowest Vth variations of all available device types and, therefore, it is used to generate the reference current. A strong resistive feedback further reduces the influence of the threshold voltage on the output current. To obtain a current mirror which is independent of supply voltage variations both devices have to work in the saturated region. In this MESFET technology this can not be achieved connecting gate and drain of the reference transistor. Therefore, diodes are used to obtain an increased drain-source voltage. As only a leakage current flows into the gate, a resistor is added to keep the diodes in forward conduction and guarantee a defined voltage difference.

3.2.4 Measurement results The amplifier was fabricated and bonded on a test substrate. Fig. 3.5 shows the measured s-parameters and noise figure. The gain over the 5–6 GHz band is more than 22 dB, input return loss is lower than −8 dB and output return loss lower than −10 dB. The noise figure is 2.4 dB. The low-noise amplifier consumes 27.6 mA from a 3 V supply (83 mW). Fig. 3.6 shows the gain and intermodulation products at 5.5 GHz ver- sus input power. 1 dB gain compression occurs at an input power −16.5 dBm, leaving enough margin to the specified system input compression point of −20 dBm. The third order intermodulation products, measured with two tones spaced 100 kHz, are fairly low: the extrapolated intercept point IIP3 at the input is at −0.8 dB. Fig. 3.7 shows a chip micrograph of the low-noise amplifier. The layout is dominated by spiral inductors for the source feedback and matching networks. The chip size is 1350 µm × 1650 µm. Low-noise amplifier 31

30

S 21 20 5

10 S 4 11

NF 0 3

−10 2 noise figure [dB] gain [dB]/ return loss [dB] S 22 −20 1

−30 0 2 4 6 8 10 frequency [GHz]

Fig. 3.5 Measured LNA S-parameters S11 (solid line with crosses), S22 (solid line with diamonds), S21 (solid line with dots) and noise figure (dashed line).

40

30 IP 3 20 P 10 1dB

0

−10

−20 gain IM output power [dBm] −30 5

−40

−50 IM 3 −60 −40 −30 −20 −10 0 10 input power [dBm]

Fig. 3.6 Measured compression (solid line), 3rd-order (solid line with crosses) and 5th-order (solid line with dots) intermodulation of LNA and extraction of 1 dB-compression and 3rd-order intermodulation point. 32 Integrated circuit design

bias bias bias

E 150

E 300 2x E300

100 um

Fig. 3.7 Micrograph of three stage low-noise amplifier. Chip size is 1350 µm×1650 µm (2.2 mm2).

3.3 Downconverter In chapter 2 it was stated that a passive resistive mixer is needed to achieve a higher dynamic range. It was assumed that this mixer shows a conver- sion loss of 6 to 8 dB, estimated from experience. The LO power is not specified. For the design of the downconverter it is reasonable to demand the best achievable performance, rather then aiming for a certain conver- sion loss and LO power. Instead, it is more important to consider the linearity requirements from the beginning. Here, an input compression point higher than 5 dBm is needed. The unknown loss and LO values sug- gest to design this mixer in a first step and then add LO and IF amplifiers with adapted characteristics. 3.3.1 Resistive mixer design Theoretically, the best signal conversion is achieved if the RF signal is switched or commutated with the LO frequency. Thereby, an ideal switch would give the best element, as it shows a linear time variant transfer characteristic [36]. In [37] it is shown, that a ”cold”FET (Vds ≈ 0) can be used as a mixer with very low intermodulation. For a small drain-source voltage the device behaves like an almost ideal resistor. The channel resistance Downconverter 33

200

150 V =0.5 V 100 g

50

0

−50 drain current [mA] V =−2.2 V −100 g

−150

−200 −0.4 −0.2 0 0.2 0.4 0.6 drain voltage [V]

Fig. 3.8 IV-characteristic of 600 µm depletion FET in the resistive region. The ”cold”FET behaves like a resistor controlled by the gate voltage. is changed by varying the gate-source voltage Vgs. As an example the simulated I/V-characteristic of a cold 600 µm depletion FET is depicted in Fig. 3.8. Fig. 3.9 shows the simulated input reflection coefficient at the drain in a common-source configuration (Vds = 0). The impedance is influenced by the intrinsic device capacitances: the opened transistor (Vgs > −1.4 V) is dominated by the low channel resistance and shows a resistive input impedance, while the closed transistor (Vgs < −2.2 V) appears like a lossy capacitor. This capacitive effect gains influence with raising frequency. It is obvious that at both frequencies, IF (1.45 GHz) and RF (5.5 GHz), the additional capacitance can not be neglected. To find the right matching impedances, the large signal s-parameters

bj Sfj ,fi (PLO)= (3.3) ai PLO

at the drain is calculated using an harmonic balance simulator. ai and bj are the incident wave at frequency fi and the reflected wave at frequency fj , respectively. Thereby it is assumed that only the LO signal is strong enough to cause nonlinear operation of the device, while ai is very small. This leads to s-parameters which are independent of the actual power of ai, but vary with LO power. The large signal s-parameters are shown in Fig. 3.10 for 1.45 GHz and 5.5 GHz and different LO power level. The calculated curves almost follow the small-signal impedances in Fig. 3.9. This can be explained by the fact, that the variations of the nonlinear 34 Integrated circuit design = −1.4 V = −1.8 V = −2.2 V g g g V V V

1.45 GHz

5.5 GHz

Fig. 3.9 Simulated drain small-signal reflection coefficient for different gate bias Vg. One point every 0.1 V at 1.45 GHz (crosses) and 5.5 GHz (diamonds). Frequency sweeps from 0.1 to 10 GHz at gate bias −2.2 V, −1.9 V and −1.4V (solid lines). device capacitances are negligible compared to the variations of the drain resistance. If Vgs is now changed periodically, the large-signal impedance roughly consists of this almost constant capacitances and a ”mean value”of the drain resistance. The complete matrix S describes the power transfer between the differ- ent frequency components ±mfLO ±nfRF , which – in this case – all share the same physical port. If S is reduced to the main frequency components, s s S = fRF ,fRF f(LO−RF ),fRF (3.4) drain s s  fRF ,f(LO−RF ) f(LO−RF ),f(LO−RF )  is obtained. To maximize the mixer conversion gain, the transfer function from fRF to fIF needs to be optimized, which is accomplished by appropri- ate matching networks. As it can be seen in Fig. 3.11, the RF and IF ports are connected to the drain using a highpass and a lowpass network, respectively. Provided that the effect of double frequency conversion is low

sfRF ,f(LO−RF ) · sf(LO−RF ),fRF  1, (3.5)

Downconverter 35 it can be assumed that the mixer is unilateral, which results in a simple power matching. In this case the source and the load at the two frequencies need to present the complex conjugate impedance to the corresponding large-signal impedance to enable maximum power transfer. The filter networks are designed to present these impedances at the drain.

0 dBm

+15 dBm

−10 dBm

Fig. 3.10 Large signal drain reflection coefficient at 1.5 GHz (solid line) and 5.5 GHz (dashed line) for LO power swept from −10 dBm to +15 dBm (one data point every 5 dB).

RF high−pass IF lo−pass RF Sdrain IF

5.7nH 3.7nH 211fF 300fF Ω V 1.75nH 5k 1.2pF dd

LO G−FET G−FET µ 600µm 50 m 355fF 0.6nH 5k Ω 30k Ω 1kΩ 5pF 5pF

bias generation

Fig. 3.11 Schematic of the resistive mixer. High and lowpass filters are used to match RF input and IF output to the large signal impedance at the drain. The transistor bias voltage is generated internally from the 3 V supply. 36 Integrated circuit design

It is known that an appropriate matching of other mixing products and the LO feedthrough can improve the conversion gain an linearity [36]. For this monolithic integration it was found that the potential advantages are compensated by the additional losses of a more complicated matching network. The gate is matched to the LO input using a parallel resonator. An additional 1 kΩ-resistor is used to lower the quality factor and obtain a larger bandwidth.

3.3.2 Mixer scaling The linear range of the a resistive mixer is limited by the voltage swing on the drain. This can be explained by two different effects: for high drain source voltages Vds the device leaves the resistive region and it can no longer assumed to be linear. Second, for the negative half-waves the role of drain and source are interchanged, the transistor is steered by the gate- drain voltage Vgd. As long as Vd = Vds ≈ 0 V stays small, the change can be neglected. If Vd becomes smaller than 0 V, the gate-drain voltage starts to exceed the gate-source voltage Vgd > Vgs. In contrast to the positive half-wave, the signal at the drain now modulates the device channel. This results in an asymmetric operation and, hence, intermodulation. The maximum drain-source voltage for linear operation depends on the transistor characteristics, on the gate-source voltage swing and as well on the termination impedances of all significant mixing products. It is, there- fore, not possible to give a simple rule. But, once the linearity is evaluated for a device, a scaling rule can be derived: to permit a higher input power associated with the same voltage swing, the effective impedance at the drain has to be lowered. This is done by increasing the transistor width. Doubling the transistor width leads to a 3 dB increase of the third order intermodulation point. It has to be noticed that the required LO power changes with the same ratio. It [38] it is found that the deep-depletion FETs give the highest com- pression point and lowest conversion loss. According to harmonic bal- ance simulations, a gate width of 600 µm is sufficient to reach the needed compression point. As seen in Fig. 3.10 this choice leads to large signal impedances which are already close to the 50 Ω-point and, therefore, are easy to match over a large bandwidth. For the complete mixer, a conversion loss of less than 8 dB and a required LO power of 10 dBm are simulated. Downconverter 37

Vdd

Vdd RF 7nH

4.7nH 1nH Ω 1.7pF 890fF 3k IF 5.7nH 200fF 2pF 1.5pF 600Ω Vg 4.5pF 1pF LO 1.3nH D−FET D−FET 7nH 300µm 300µm

730fF 781fF 1kΩ resistive mixer 5kΩ

LO amplifier IF amplifier

Fig. 3.12 Schematic of integrated downconverter: LO and IF amplifiers are added to the resistive mixer in Fig. 3.11. Bias is not shown here.

3.3.3 Integrated downconverter Starting with the resistive mixer as a core, the downconverter is completed by adding an LO and IF amplifier. The complete circuit schematic is depicted in Fig. 3.12. The IF amplifier compensates for the conversion losses of the passive mixer and provides the demanded signal gain of approximately 20dB. A cascode configuration is chosen to obtain high gain from a single stage. This limits the number of large inductors and of the required chip area. RC-feedback is used to stabilize the amplifier. The LO amplifier is designed to deliver the needed 10 dBm LO power in linear operation to avoid upconversion of 1/f-noise. The main task of this amplifier is to provide isolation in the reverse direction. This reduces possible signal-crosstalk over the LO feeding network, if the downconverter is applied in an antenna array. The additional gain is also advantageous, but no specific value is specified. This allows to sacrifice some gain and to use an input matching network which is partly resistive. A good matching over a large bandwidth can be achieved this way. Two integrated circuits were designed and manufactured, the complete downconverter including the mixer and the two amplifiers and the resistive mixer alone.

3.3.4 Measurement results Both circuits were bonded to on test substrate and measured. Fig. 3.13 shows the conversion gain of the resistive mixer vs. the LO power. The measured conversion gain increases with higher LO power, reaching −6.8 dB 38 Integrated circuit design

−5

−6

−7

−8

−9

−10

−11

−12 conversion gain [dB] −13

−14

−15 0 5 10 15 20 LO power [dBm]

Fig. 3.13 Resistive mixer: measured (solid line) and simulated (dashed line) conversion gain vs. LO power. (fRF = 5.5 GHz, fIF = 1.45 GHz) at 15 dBm. At low LO power, the measured conversion loss is roughly 1 dB higher than simulated. For higher LO levels the harmonic-balance sim- ulation results are found to vary strongly with the number of considered harmonics. The depicted curve was simulated using 13 harmonics. At PLO = 10 dBm the conversion loss is less than 8 dB and varies with 0.2 dB/dB of PLO. Although higher LO power levels lead to less conversion loss, this represents a suitable operation point. The following measurements were performed using a power of 10 dBm. Fig. 3.14 shows measured and simulated conversion gain of the mixer for a simultaneous frequency sweep of LO and RF, keeping the interme- diate frequency constant at 1.45 GHz. The measured conversion loss is between 7.6 and 9 dB over the whole band. The measured conversion loss is in good agreement with the simulations. Fig. 3.15 shows the simulated and measured compression behavior of the resistive mixer at 5.5 GHz. The measured input 1 dB-compression point of 4 dBm is lower than simulated. There are two possible reasons for the overestimation of gain and compression point: the LO matching network causes more losses than simulated, the simulated voltage swing at the gate is higher, decreasing the conversion losses and increasing the compression point. Also, the transistor model [39] assumes symmetrical devices, while the actual FETs have an asymmetrical gate recess. Downconverter 39

10

9.5

9

8.5

8

7.5

7

6.5 conversion loss [dB]

6

5.5

5 5000 5200 5400 5600 5800 6000 freq [MHz]

Fig. 3.14 Resistive mixer: measured (solid line) and simulated (dashed line) frequency dependency of conversion loss with swept RF and LO and con- stant IF. LO power is fixed to 10 dBm. RF power is below compression point (PRF = −10 dBm).

−5

−5.5

−6

−6.5

−7

−7.5

−8

−8.5 conversion gain [dB]

−9

−9.5

−10 −25 −20 −15 −10 −5 0 5 10 RF power [dBm]

Fig. 3.15 Resistive mixer: simulated (solid line) and measured (dashed line) compression behavior at fRF = 5.5 GHz. LO power is fixed to 10 dBm. 40 Integrated circuit design

The restrictions of this transistor model leads to uncertainties in the simulated large-signal behavior for this unusual operating conditions which include the inverted operation. The mixer was designed with sufficient margin and to meet the specifications of the system. Fig. 3.16 shows the conversion gain of the integrated downconverter versus the applied LO power. As expected, the two additional amplifiers shift the curve compared to the resistive mixer alone: the LO amplifier reduces the required LO power by about 6 dB, while the IF amplifier adds 15 dB of gain, turning the conversion loss of the passive mixer into conversion gain. Accordingly, a conversion gain of 7 dB is achieved at a reduced LO power level of 5 dBm. To characterize the linearity of the downconverter, 1 dB-compression point and third-order intermodulation intercept point were measured. RF and LO were swept simultaneously, keeping the intermediate frequency constant at 1.45 GHz. The results are depicted in Fig. 3.17. For a LO power of 10 dBm the compression and intercept point stay almost constant over frequency at around 1 dBm and 8 dBm, respectively. If the LO power is reduced to 0 dBm, compression and intercept point at higher frequencies decrease. This indicates that, at these frequencies, the mixer becomes the limiting element that constrains the system linearity, while at 5 GHz the conversion gain is sufficient to saturate the following amplifier.

10

5 LO− amplifier 0 IF− amplifier −5 conversion gain [dB] −10

−15 −10 −5 0 5 10 15 LO power [dBm]

Fig. 3.16 Integrated downconverter with additional LO and IF amplifier (solid line) compared to single passive mixer (dashed line). The IF amplifier provides conversion gain, while the LO amplifier reduces the needed LO power. Downconverter 41

The power consumption is 59 mA from a 3 V supply (177 mW) for the two amplifiers. The passive mixer itself does not require DC power. Fig. 3.18 shows a chip micrograph of the integrated downconverter. Total chip size is 1650 µm×900 µm.

15

10

IP 5 3 input power [dBm] 0

P 1dB

−5 5000 5200 5400 5600 5800 6000 frequency [MHz]

Fig. 3.17 Integrated downconverter: input referred 1 dB-compression point (crosses) IP3 (circles) for LO power of 0 dBm (dashed line) and 10 dBm (solid line).

mixer

LO amplifier IF amplifier

100 um

Fig. 3.18 Chip micrograph of integrated downconverter. Total chip size is 1650 µm×900 µm (1.5 mm2). 42 Integrated circuit design

array element 2nd & 3rd resistive LNA stage mixer

IF integrated switchable image filter LNA stage resonant amplifiers monolithic receiver

divider CALnetworks LO

Fig. 3.19 Chip architecture of monolithically integrated receiver front-end. 3.4 Integrated front-end 3.4.1 Architecture For the application in active antenna arrays, where the available space below each antenna element is limited, further integration of the system is needed. Fig. 3.19 shows the block diagram of the developed single-chip front-end: it bases on the low-noise amplifier and integrated downcon- verter described in sections 3.2 and 3.3. According to the system architecture an additional filter is needed to suppress possible signals that could be received at the image frequency. This image filter is generally difficult to realize in any integrated tech- nology for two reasons: due to the process variations high-Q high-order filters cannot be build and, second, parasitic substrate coupling and on- chip cross-talk limit the achievable stop-band attenuation. In integrated receivers often image-reject architectures are favored [40–43] to circumvent this problem. On the other hand, these architectures are very sensitive to phase and amplitude imbalances of their complex mixer pairs, which degrade the image-rejection capability. Here, the selected intermediate frequency at 1.45 GHz is sufficiently high to allow the use of a low-order filter. Therefore, it is decided to maintain the heterodyne concept avoid- ing the generation of a precise quadrature LO signal. The image filter is integrated in the signal path between LNA and downconverter. 3.4.2 Switchable LNA The active circuit components typically introduce amplitude and phase changes that vary with ambient temperature, biasing and aging. For some antenna array applications, these variations need to be calibrated. This calibration will be discussed later in chapter 5 in detail. This calibration of the downconverter can be performed by switching between a precisely Integrated front-end 43

Vdd Ω Vg1 5k ant 2.3nH E−FET µ 2.3nH 300 m 8pF 0.9pF 2pF

0.5pF 100fF 0.6nH bondwires to 2nd stage

Vg2

E−FET 300µm cal input inductive matching source degeneration

Fig. 3.20 Circuit schematic of switchable LNA. Two identical input stages se- lect between antenna and calibration signal. The drains are matched for the con- dition that one transistor is biased on the gate and the other device is switched off. known reference signal (calibration mode) or the antenna input (receiving mode). Using a passive switch with a typical insertion loss of 1–2 dB [44] would significantly degrade the noise figure. Therefore a switchable low- noise amplifier is proposed. Thereby, it can be turned into an advantage that, on an integrated circuit, closely spaced transistors show almost iden- tical electrical characteristics. The first LNA stage consists of two symmetric 300 µm enhancement FETs connected at the drains as seen in Fig. 3.20. As in the previously described low-noise amplifier, inductive source feedback is used to improve the linearity and increase the real part of the input impedance. The impedance at the connected drains is strongly dependent on the states of the two transistors. The interstage matching to the following two stages assumes that only one of the two devices is active. One transistor is biased for best linearity at a current density of Id/w=40 µA/µm, the second one is switched off by setting the gate bias to Vgs=0 V. Differences in gain or phase of the two parallel input stages lead to calibration errors. Therefore a completely symmetric layout is used. The transistors are located very close to each other and show the same gate orientation to avoid device mismatching. It needs to be taken into account that the receive mode might be used for a long time, while the calibration requires a comparably short inter- val. Thermal effects on the devices need to be considered. To keep them on the same temperature, the transistors are physically connected at the 44 Integrated circuit design drains. As a rough estimate, the following assessment can be made: if both transistors are thermally independent and the paths of heat trans- portation are limited to the way down through the substrate, the working transistor will heat up due to its losses, while the other stays at ambient temperature. With a drain current of 7 mA at 3 V, a wafer thickness of 25 mil (635 µm) and the thermal conductivity 0.5 W/Kcm of GaAs the computed temperature difference 1 3 V · 7 mA ∆T = · (3.6) 0.5 W/Kcm 635 µm is less than 1 K. This indicates that thermal effects can be neglected in this case. A further crucial point is the large-signal behavior of the switch in off- state. As low-power enhancement FETs are used, a large input voltage exceeding the threshold voltage Vth switches the transistor on, thus de- grading the isolation. As the isolation is especially important for signals with high input power levels, it is important to control this effect. To avoid the use of negative control voltages the transistor width was increased to lower the input impedance and therefore also the voltage swing on the gate. This approach leads to a trade-off between maximum tolerable input power and current consumption. It is assumed that a signal power larger than the aimed 1 dB input compression point does not occur in practical operation, as in the system design it was determined to ensure that the receiver always works in linear operation. By choosing a wide transistor for this input stage, the input impedance is lowered and it is ensured that the critical gate voltage is not exceeded below an input power level of −20 dBm. Corresponding simulations were carried out using the TOM3 large-signal model [39]. 3.4.3 Image filter The image filter consists of a third-order bandpass filter, realized with lumped elements according to Fig. 3.21. Since the filter is sensitive to parasitics, its layout was carefully optimized. The filter was designed for a higher bandwidth to allow for process variations. The limitations in achieving high stop-band attenuation are internal coupling and the absence of good grounding. On the semi-insulating GaAs substrate magnetic coupling is the dominant coupling mechanism. To minimize its effect on the stop-band attenuation, the input and output inductors were separated by large distances. The filter ground is not connected to the amplifier ground to minimize any coupling over a common imperfect ground. By providing a number Integrated front-end 45

test pads

RF 5.7nH 5.7nH RF

in 151fF 151fF out

600pH 2.5pF 600pH

image filter L package p parasitic

Fig. 3.21 Schematic of lumped element image filter. T-type bandpass filter with element values suited for monolithic integration. The implementation is very sensitive to layout and package parasitics. of bondwires to ground, the package parasitic Lp was kept sufficiently low not to degrade the stop-band rejection. GSG-pads were added to permit on-wafer measurements of the filter alone. 3.4.4 Layout High frequency circuit design usually employs the concept of wave ports rather than voltages and currents. S-parameters are used to characterize all devices and elements with respect to input and output reference planes. This requires waveguide structures with a well-defined ground reference. On-chip test-structures typically employ a coplanar layout and all models are derived from these measurements. At frequencies in the lower gigahertz range, the coplanar ground struc- ture usually is omitted to simplify the layout of complex circuits. This forces the ground current to take a different way, changing the component behavior. To maintain the validity of the models, this effects needs to be considered. If the geometrical distance of the additional ground path dmax is significantly smaller than the minimum wavelength λmin involved, with 1 dmax < 10 λmin as a rule-of-thumb, the ground path is not needed to be modelled as a distributed structure and can be replaced by appropriate inductances. Here, the chip is bonded on a carrier substrate. Due to the missing ground, all signals now refer to the waveguide ground on the substrate. The main problem then becomes to provide a good on-chip ground by using as many bond-wire connections to the substrate as possible. To obtain guidelines for the layout of a complex chip, the following simplified estimation is done: If a device ”1” is connected to a row of pads bonded to ground, as seen in Fig. 3.22a, the bond-wires can be considered as inductances Lbond 46 Integrated circuit design

LL 1 21,l 1 1,r 2 1 2 L geom LL1 2 bond L common L (a) (b) (c)

Fig. 3.22 Simplified packaging model: a) typical circuit situation with two elements connected to a row of pads bonded to ground b) bond wires and pad distances are modelled as inductances c) the path from both elements two ground contains one common and two individual inductance. to ideal ground, while the distance between to adjacent pads also results in a second inductance Lgeom, which is much smaller. This leads to the equivalent circuit depicted in Fig. 3.22b. If, in each case, Ll,n and Lr,n are the inductances resulting from the structure to the left and to the right of a bondpad n, the total inductance to ground Lin,1 can be calculated recursively: 1 1 1 1 = + + , (3.7) Lin,1 Lbond Ll,n Lr,n 1 Lr,n = Lgeom + (3.8) 1 + 1 Lbond Lr,n+1 and Ll,n accordingly. With the typical values of Lbond = 0.5 nH and Lgeom =0.052 nH this leads to the curves depicted in Fig. 3.23. It is obvious, that with an increasing number of bond pads the min- imum inductance never falls below a certain value of around 80 pH and that devices should not be connected at edge pads if a low inductance to ground is needed. If several connections to ground are needed, any common inductance in the ground path leads to coupling between the circuit elements. This might degrade the system performance or, in the worst case, cause oscil- lations. It is illustrative to also estimate this coupling. Given that a second device ”2” is now connected to bond pad m, the inductor network in Fig. 3.22b can be simplified to the three-inductor net- work in Fig. 3.22c. Lcommon is the common inductance which causes cou- pling, while L1 + Lcommon equals the previously calculated Lin,1. Starting from equations 3.7 and 3.8 Lcommon is calculated. The resulting curves are added to Fig. 3.23. Component ”1” stays fixed at the position of pad number 10, the position of element number ”2” is Integrated front-end 47

150 10 pads total inductance L +L 1 common 100

inductance [pH] 50 20 pads

L common

0 0 5 10 15 20 bondpad number

Fig. 3.23 Inductance to ground (solid lines) and common inductance (dashed lines) vs. bond pad position for rows of 10 to 20 bond pads. For the common in- ductance calculation it is assumed that one element is connected to pad number 10. varied. If both components are connected to the same pad, the obtained values obviously agrees with the previously calculated inductance. The more the two connections are separated, the less the currents have to share the same way, Lcommon decays. A distance of three pads is sufficient to halve the value. Fig. 3.24 shows the chip micrograph of the receiver front-end. The circuit blocks are grouped along the ground pads. The ground of the symmetric input stage and the consecutive stages are separated to avoid feedback which would cause oscillations. The IF amplifier, which is less sensitive to parasitic inductances due to its lower frequency, is located in the center of the chip. Total chip size is 1650 µm×1950 µm. 3.4.5 Experimental results The fabricated chip was bonded on a Duroid 6010 substrate and tested. Fig. 3.25 shows the measured conversion gain as a function of LO power. It saturates at 26.2 dB for LO levels higher than 0 dBm. Fig. 3.26 shows the on-wafer measurement of the integrated image- filter compared to the simulations. The measured bandwidth is smaller, but still covers the signal band. The filter insertion loss is 6 dB and, therefore, higher than simulated. The stop-band attenuation is better than 35 dB. 48 Integrated circuit design

LNA stage 2 LNA stage 3

IF amplifier image filter

LNA stage 1

LO amplifer resistive mixer

100 um

Fig. 3.24 Chip micrograph of integrated receiver front-end. Total chip size is 1650 µm×1950 µm (3.2 mm2).

27

26

25

24

23

conversion gain [dB] 22

21

20 −15 −10 −5 0 5 10 15 Plo [dB]

Fig. 3.25 Total conversion gain of integrated receiver front-end as a function of LO power. (fRF = 5.5 GHz, fIF = 1.45 GHz, PRF = −30 dBm) Integrated front-end 49

0

−5

−10

−15 band

−20

−25 image

S21 [dB] −30

−35

−40

−45

−50 2 3 4 5 6 7 8 9 10 freq [GHz]

Fig. 3.26 Simulated (dashed line) and measured (solid line) frequency response of the integrated image filter, measured on-wafer.

At frequencies higher than 8 GHz the attenuation rises less than pre- dicted by the simulations. This can be explained by parasitic coupling effects. To measure the frequency behavior of the whole chip, RF and LO frequencies were swept, keeping fIF =1.45 GHz constant. The resulting curves are depicted in Fig. 3.27: over the 5.125 GHz to 5.875 GHz range, the gain is above 20 dB. Image rejection is 36.7dB at 5.125 GHz and raises to 50 dB at 5.875 GHz which is a very good result for an integrated receiver. Over the band of interest, the SSB noise figure is below 3.8 dB with a minimum of 3.3dB at 5.6 GHz. The gain difference between the two input channels stays below 0.7 dB. In order to determine the phase difference, an identical switchable LNA is fabricated on another chip. A vector network analyzer is utilized to de- termine the s-parameters in receive and calibration mode. The measured complex gain curves are shown in Fig. 3.28. The maximal amplitude dif- ference is 0.5 dB, the maximal phase difference 4◦. Thereby it has to be noticed that this also includes possible differences resulting from the testing substrate and unequal bondwire lengths. 50 Integrated circuit design

30 7

20 6 image 10 rejection 5

0 4

−10 3

conversion gain [dB] −20

2 SSB noise figure [dB]

−30 1

−40 5 5.2 5.4 5.6 5.8 6 0 freq [GHz]

Fig. 3.27 Frequency response of integrated receiver front-end: conversion gain in receive (solid line) and calibration mode (dashed line), SSB noise figure (solid line with dots) and gain at the corresponding image frequency (solid line with triangles). (PLO = 10 dBm, PRF = −40 dBm).

20

15

10

5 ch1

0 ch2

−5

−10

−15

−20 −20 −10 0 10 20

Fig. 3.28 Linear complex gain of the switchable low-noise amplifier integrated on a separate chip. S21 of channel 1 (solid line with circles) and channel 2 (dashed line with crosses) from 5 to 6 GHz, one point every 50 MHz. Integrated front-end 51

30

25

20

15 1dB compression gain [dB]

10

5

0 −30 −25 −20 −15 −10 −5 0 0 Pin [dBm]

Fig. 3.29 Large signal behavior of integrated receiver front-end: gain of active (solid line) and deactivated (dashed line) channel. The isolation is reduced at input power levels higher than the 1 dB-compression point. (PLO = 10 dBm)

Fig. 3.29 shows the large signal behavior of the receiver frontend with 10 dBm of LO power: 1 dB gain compression occurs at an input power level of −18 dBm, which is higher than the specified −20dBm. For a reduced LO power of 0 dBm an input compression point of −20 dBm is obtained. The switch isolation of the LNA is 17.65 dB at low power levels, de- creasing to 16.55 dB at the compression point. The expected degradation of the switch isolation occurs at input power levels higher than −15 dBm, which is above the compression point. Signals of this power level are not expected to occur and, therefore, the degradation has no practical signif- icance. For an LO power of 5 dBm the compression point and third-order in- termodulation intercept point (∆f=100 kHz) are measured for different frequencies and depicted in Fig. 3.30. At 5.4 GHz, where the highest con- version gain occurs, the compression point drops slightly below −20 dBm for this LO power level. To investigate the immunity against interferers, the blocking behavior of the circuit is measured: a sinusoidal signal at the center of the RF- band (5.5 GHz) and a power clearly below the input compression point (−50 dBm) is applied to the receiver input. It is then converted by the receiver front-end (fLO = 6.95 GHz, PLO = 5 dBm) to the intermediate 52 Integrated circuit design

−5

−10

IP 3 −15

P input power [dBm] 1dB −20

−25 5000 5200 5400 5600 5800 6000 frequency [MHz]

Fig. 3.30 Input referred 1 dB-compression point (solid line) and third order intermodulation intercept point (dashed line) over frequency. (fIF = 1.45 GHz, PLO = 5 dBm) frequency of 1.45 GHz. A second signal with varying frequency is added at the RF input. Its power is raised to the level where the interference reduces the conversion gain of the regular input signal by 1 dB. The resulting blocking curve is depicted in Fig. 3.31. Inside the 5–6 GHz band, an interferer of approximately −20 dBm is sufficient to degrade the receiver performance. This value is equal to the 1 dB-compression point, which also should be expected, as the interferer undergoes the same processing as the regular signal and, therefore, satu- rates the receiver in the same way. For signals outside this frequency band the gain of the selective low-noise amplifier decreases and the signal is fur- ther attenuated by the image-reject bandpass-filter. Blocking only occurs, if the input stages of the LNA are saturated, requiring significantly more power. The reactive input matching of the very first stage forms a lowpass structure (see Fig. 3.20). This explains why at low frequencies the blocking level almost stays constant at −6 dBm, while it increases rapidly towards higher frequencies. Above 9.6 GHz the blocking could not be measured, because the required power level was higher than the power delivered by the signal generator. In the same graph in Fig. 3.31 the blocking level defined by the Hiper- lan/2 standard [15] is depicted. It cannot be compared directly with the measured curve, because it is defined for the whole system including an- Integrated front-end 53

20

15 LO IF 10 f image 5

0 RF −5

−10

input power [dBm] −15

−20

−25

−30 0 2 4 6 8 10 12 freq [GHz]

Fig. 3.31 Blocking experiment: interferer power-level that is needed to decrease the regular signal conversion gain by 1 dB. Blocking-power level (solid line)vs. frequency and Hiperlan/2 blocking mask [15] (dashed line). (PLO = 5 dBm) tennas, modulation and coding: interferer signals falling within this mask are not allowed to cause more than 10% erroneously received ”protocol data units (PDUs)”. However, the blocking measured in this analog fash- ion is a good indication of the expected performance. For frequencies less than 2.5 GHz the measured blocking curve falls below the mask. This is not considered as a problem for the system, as the antenna acts as a bandpass filter with a significant attenuation at low frequencies. In Tab. 3.2 the circuit performance at 5.5 GHz is compared with re- cently published results of other integrated 5–6 GHz receivers. The re- ceiver front-end reaches state-of-the-art performance. The noise figure is very low if it is considered that both receivers that reach similar val- ues [45,46] require external components whose losses were subtracted from the measurements. The circuit draws 55 mA for the signal path (LNA, mixer and IF am- plifier) and 21 mA for the additional LO amplifier, resulting in a total of 76 mA from a 3 V supply (230 mW). Total chip size is 1650 µm×1950 µm (3.2 mm2). 54 Integrated circuit design

Table 3.2 Circuit performance at 5.5 GHz (PLO = 10 dBm) compared to re- cently published integrated 5–6 GHz receivers

frequency gain im. rej. Pin,1dB NF [40] 5.2 GHz 26 dB >50 dB −18dBm 7.2 dB (vgain) [47] 5.3 GHz 14 dB 36 dB −15dBm 6.8 dB (tuned) [41] 5.2 GHz 43 dB 62 dB −26.5dBm 6.4 dB (vgain) [42] 5 GHz 19 dB >40 dB −22dBm 5.9 dB [43] 5.5 GHz 17 dB 80 dB −14dBm 5.1 dB (5.3–5.8) (tuned) [45] 4.7 GHz 24 dB 35 dB −23dBm 3.2 dB [46] 5.25 GHz 8.7 dB n.a. −21 dBm 3 dB(DSB) thiswork 5.5GHz 26dB 48dB -18dBm 3.3dB (5.2-5.9)

3.5 Power amplifier The power amplifier represents a critical part in most systems, because it is one of the building blocks with the highest power consumption. Its mono- lithic integration is often problematic, because the devices must be driven to their electrical and thermal limits. This opens interesting questions concerning the calibration of transmit arrays, especially when the system transmits in bursts and the amplifiers are switched off during times of inactivity. This motivates the design of a power amplifier to address this issue, although it is not directly needed for the receiving antenna. The implications for transmit arrays will be discussed in chapter 5.4. 3.5.1 Design One advantage of multiple antenna systems is that the total transmit power is shared by all elements. In a N element array this reduces the output power of a single amplifier by N if the total power is kept constant, or by N 2 if the additional antenna gain is considered. In a four element array, therefore, the individual power amplifier output power is reduced by 6 dB or 12 dB, respectively. This leads to medium power amplifiers which are easier to integrate. The output power chosen here is 20 dBm. (For comparison: Hiperlan/2 [15] requires a total radiated power higher than −15dBm, but not more than 23dBm at 5.2 GHz and 30 dBm at 5.8 GHz.) Power amplifier 55

300 I dmax 250 load−line

200

150 supply voltage

100 drain current [mA]

50 V dmax 0 0 1 2 3 4 5 6 drain voltage [V]

Fig. 3.32 I/V characteristic of 1200 µm D-FET for different gate source volt- ages Vgs (solid line) and construction of optimal load line (dashed line) for maximal output power.

Fig. 3.32 shows the I/V-characteristic of a depletion FET with the load line for maximum output power. Idmax is given by the device itself, Vdmax results from the supply voltage Vdd and the knee voltage Vknee. The output power of a given device can be maximized by increasing the supply voltage until Vmax reaches the breakdown voltage. Here, a supply voltage of 3 V is maintained; the device width is changed to adjust the maximal current. The saturation power at the output can be estimated as 1 P = (V − V ) I (3.9) max 8 dd knee dmax

The device is scaled to a width of 1200 µm, leading to a Pmax of roughly 23 dBm, which leaves some margin for implementation losses. The schematic of the last amplifier stage is shown in Fig. 3.33. The output matching network performs a transformation to present the right load impedance to the intrinsic transistor. Thereby, the transistor para- sitics are taken into account. For more details on this technique it is re- ferred to [48]. The input is matched to the source using conjugate complex matching. Additional parallel and series resistors are needed to guarantee unconditional stability. A special network is designed to enable switching of the amplifier. In principle there are two ways to switch a power amplifier: the supply 56 Integrated circuit design

+3V loadline matching input matching & RF in stabilization RF out

Q1

Q2 PA switch temp sensor

Fig. 3.33 Schematic of the power amplifier output stage. The power ampli- fier can be switched off to save power when the system is not transmitting. A Schottky diode is integrated close to the transistor to monitor the thermal behavior. voltage at the drain is removed, or the gate-source voltage Vgs is reduced below the pinch-off voltage. Switching the large currents at the source or drain introduces significant losses, which decreases the efficiency. If the device is switched by the gate voltage, another difficulty arises: in on-state the gate bias voltage should be set with a low impedance at low frequencies. Otherwise self-biasing and memory effects occur [49], which lead to signal distortion in non-constant envelope systems. On the other hand, the switching signal should be well isolated from the RF signal to avoid unwanted cross-talk. The use of large capacitors has to be avoided to keep the resulting time constants and switching times low. In this amplifier the power device Q1 (D-FET) is selected for its con- venient optimum gate bias point of Vgs = 0 V. This allows to use an aux- iliary transistor Q2 to provide a low-impedance short-circuit from gate to ground, if switched on. The remaining on-state resistance is not critical for the circuit performance, because a certain resistive load is needed at the gate to stabilize the device. For negative control voltages, the aux- iliary transistor closes and the gate of the main device is tied down by large-value resistor. This upper limit of the resistance is determined to avoid unintended switch opening like discussed above in 3.4.2. A diode is integrated in close vicinity to the power device as a tem- perature monitor. With raising temperature its diffusion voltages Vd de- creases. If a constant current is impressed, the actual diode temperature T is determined as V T = T 0 , (3.10) 0 V Power amplifier 57

20

15 S 21

10

5

0

−5

−10 gain [dB]/ return loss [dB] S 22 −15 S 11 −20 0 2 4 6 8 10 frequency [GHz]

Fig. 3.34 Measured S-parameters of the power amplifier: return loss S11 (solid line with crosses), S22 (solid line with diamonds) and gain S21 (dashed line with circles). where V is the measured temperature across the diode and V0 is the known voltage at the reference temperature T0. The power amplifier consists of two stages, whereby the driver stage device width is half of the output device width. The compression behavior is mainly determined by the last stage.

3.5.2 Experimental results The power amplifier was bonded on a test substrate and measured. Fig. 3.34 shows the measured S-parameters. The input matching is better than −15 dB over the 5–6 GHz band. The output matching is around −10 dB and does not show a pronounced matching frequency, which is typical for load-line matched amplifiers. This also explains why the gain peak occurs below the design band. The small-signal gain inside the band is more than 10 dB. At small signal operation the amplifier draws 500 mW froma3V supply. Fig. 3.35 shows the compression behavior at 5.2 GHz. At 1 dB gain compression the amplifier delivers 19 dBm. It is remarkable that the rela- tive phase shift at the compression point is below 0.7◦, which demonstrates that the amplifier can be used up to the compression point without intro- ducing significant AM/PM conversion. 58 Integrated circuit design

20

18 1dB 16 compression 14 gain

12

10

8

6 gain [dB], phase [deg] 4

2 phase 0 −5 0 5 10 15 input power [dBm]

Fig. 3.35 Compression behavior of power amplifier: gain (solid line) and phase (dashed line) vs. input power at 5.2 GHz.

3.5.3 Pulsed operation In the literature a typical gain change of −0.015 dB/◦C for GaAs cir- cuits [50] is reported. Therefore, it can be expected, that the amplifier shows a varying gain, when it is used in pulsed operation exhibiting chang- ing temperatures. To quantify this effect the amplifier is switched on and off with a period of 7 × 250 µs=1.75 ms with a varying duty-cycle, corre- sponding to typical IEEE 802.11a or Hiperlan/2 data bursts. The voltage across the temperature sensor diode is monitored. The resulting gain and temperature are depicted in Fig. 3.36. At a duty-cycle of 0/7 the amplifier is at ambient temperature. With in- creasing duty-cycle the temperature rises linearly to 55 ◦C at continu- ous transmission. With a measured power dissipation of 333 mW of the last stage a thermal resistance of 99 K/W results (calculated thermal re- sistance through the 4 mil GaAs substrate is 197 K/W). Changing the duty-cycle from 1/7to 7/7, the gain drops more than 0.5 dB, resulting in a gain change of 0.019 dB/K with temperature. To evaluate the dominating thermal time-constants, the amplifier is switched on for single pulses of different length, assuming that the am- plifier is at ambient temperature at the beginning of the pulse. Fig. 3.37 shows the temperature graph for a 20 ms pulse. The measured curve agrees very well with the simple exponential model Power amplifier 59

0 60 −0.1

−0.2 50 −0.3

−0.4 40

−0.5 C] ° −0.6 30 T [ rel. gain [dB] −0.7

−0.8 20 −0.9

−1 10 0/7 1/7 2/7 3/7 4/7 5/7 6/7 7/7 duty cycle [1]

Fig. 3.36 Gain (solid line) and temperature variations (dashed line) of the switched power amplifier for different duty cycles.

38

36

34

32

30 C] o

T [ 28

26

24

22

20 −10 0 10 20 30 40 time [ms]

Fig. 3.37 Measured diode temperature (solid line) if the amplifier is switched on for a 20 ms pulse and first order exponential model (dashed line) with a time ◦ constant of τth,1 = 1.6 ms at room temperature T0 = 22 C. 60 Integrated circuit design

t T0 + ∆T · 1 − exp( ) 0 ≤ t ≤ 20 ms τth,1 T (t)=    , (3.11) t−20 ms  T0 + ∆T · exp(− ) t> 20 ms  τth,1 assuming a thermal time constant of τth,1 =1.6 ms and a temperature difference of ∆T = 13 K. For shorter pulse lengths no lower time constant is found. Comparing the maximum temperature of 35 ◦C with the 55 ◦C mea- sured at continuous operation, another time-constant must be present. Repeating the above measurement with longer pulse lengths, a second time constant of τth,2 = 2200 ms is found. The two time-constants can be attributed to the small heat capacity of the device and the thermal resistance of the GaAs substrate and die-attach on the one hand, and the mounting patch and the heat transfer to the environment on the other hand. Knowing this thermal behavior, the temperature change at the begin- ning of a pulse is ∂ ∆T ∆T T (t) = 1 + 2 = 8667 K/s+10K/s, (3.12) ∂t τ τ t=0 th,1 th,2

which leads to an approximate gain change rate of ∂ g(t) = −8676 K/s · 0.017 dB/K= −148 dB/s. (3.13) ∂t t=0

This expected change of gain, although it is fairly low considering the typical 250 µs pulse length, opens the question how a possible calibration will be affected. This behavior can be further investigated using system measurements, which will be reported in chapter 5.4. 3.6 10.7–11.7 GHz SiGe downconverter As already mentioned in section 3.1.1, the advance of silicon-based tech- nologies makes them an interesting for RF applications. To achieve a fast analog circuits, the standard CMOS process can be extended by a SiGe heterojunction bipolar transistor (HBT). This allows the complete inte- gration of mixed-signal circuits. By aggressive vertical and lateral scaling, transit frequencies as high as 375 GHz could be reached [31], making this technology competitive with classical RF processes. In the same way as wireless LANs move now from 2.4 GHz to 5 GHz to achieve higher data rates, it can be foreseen that the wireless local loop 10.7–11.7 GHz SiGe downconverter 61

(WLL), which offers fixed wireless links to the public phone network, will move from 3.5 GHz to 11 GHz for the same reason. At this frequency, a low-noise amplifier with a noise figure as low as 2 dB was demonstrated [51], using an advanced SiGe process with fT = 80 GHz. For the envisioned application, it is advantageous to use a low-ft pro- cess, which relaxes the requirements of the lithography and helps to mini- mize the manufacturing costs. On such processes with ft of about 50 GHz, low noise amplifiers with 3.3 dB noise figure [52] and active mixers [53] were demonstrated, up to 17 and 20 GHz [54, 55]. Even a receiver front- end at 24 GHz [56] is reported. All proposed circuits employ Gilbert-type active mixers, which leads to limited dynamic ranges. Input compression points up to −9 dBm with single sideband noise figures at best of 12 dB are reached. To mitigate the noise contribution of the mixer, a pre-amplifier is needed. The com- pression point is lowered further and the frontend becomes more sensible to intermodulation and blocking phenomena. It was already seen above for the 5 GHz receiver, that a resistive mixer can be used to enhance the dynamic range. On a SiGe process, the MOSFET device can be employed to realize this passive FET mixer, as it will be described below. The goal of this study is to develop a downconverter for the 10.7– 11.7 GHz band with specifications similar to those of the presented 5 GHz receiver.

3.6.1 IBM 6HP SiGe BiCMOS process The IBM 6HP process is based on a 0.25 µm standard CMOS process. Processing steps are added to obtain HBTs, which are available as high- speed or high-breakdown devices for power applications. The high-speed device reaches a transit frequency of 47 GHz. To allow low-loss passive devices in spite of the conductive substrate, a thick dielectric and top-level metal are added as last steps [57]. The key parameters of this process are summarized in Tab. 3.3.

3.6.2 Low-noise amplifier The main difference between the LNA design on GaAs as described in 3.2 and the realization on this SiGe process are the higher losses that are associated with the passive components, while the transit frequency of the devices of 47 GHz is more than two times higher. If the same strategy is followed and common-source stages are cascaded using conjugate complex interstage-matching, it is found that a large part of the gain delivered by the transistors is dissipated in the passive matching networks. At a 62 Integrated circuit design

Table 3.3 IBM-6HP process key parameters (from [57])

Element Parameter Value SiGe HBT current amplification β 100 transit frequency ft 47 GHz max. freq. of oscillation fmax 65 GHz spiral inductor inductance 0.28nH..83nH quality factor 19 @ 10GHz MIMcapacitor capacitanceperarea 0.70 fF/µm2

Vcc

1.38nH 1.27nH

120fF

2x 20µm RF out 55fF RF 2x 20µ m 2x 20µm in

155pH

Fig. 3.38 Circuit schematic of two stage 11 GHz low-noise amplifier. The first stage uses inductive source degeneration for noise matching and a cascode topol- ogy for maximal gain. The second stage uses a single transistor to enable output matching. Biasing networks are not shown here. center frequency of 11.2 GHz, the gain of the first stage is critically low, compromising the overall noise figure. To circumvent this problem, a different circuit concept is used, shown in Fig. 3.38: a cascode topology as input stage provides a sufficiently high gain. Again inductive source degeneration is applied to achieve a higher real part of the input impedance. The problem associated with the cascode structure is the high output impedance. A transformation to the 50 Ω port impedance over a certain bandwidth would require a multi-stage network with high associated losses, consuming the gain advantage of the cascode stage. To achieve a simple output matching, a second single-transistor stage is added. At the interface between the two stages no impedance transformation is introduced, only the reactance is compensated to achieve a gain peak in the center of the band. A coplanar-like layout was chosen: all unused space was filled with 10.7–11.7 GHz SiGe downconverter 63

20

15 S21 10

5

0 S22 NF −5

−10

−15 gain/return loss [dB] S11 −20

−25

−30 8 9 10 11 12 13 14 freq [GHz]

Fig. 3.39 Simulated (dashed lines), measured (solid lines) S-parameters and noise figure NF (dots) of 11 GHz LNA patches of metal, which are connected among each other, forming a ground mesh. Fig. 3.39 shows the simulated and measured S-parameters of the manu- factured amplifier: The gain peak of 14.8 dB is less than the simulated 19 dB. Both input and output return loss are better than −10 dB and agree reasonably well with the simulations. The measured noise figure is lower than 4 dB. The measured compression behavior depicted in Fig. 3.40 indicates a 1 dB-compression point of −12.7 dBm. The power consump- tion is 6.4mA from a 3.3 V supply (21 mW). The chip micrograph is shown in Fig. 3.41. The chip size is 1323 µm×923 µm (1.2 mm2). 64 Integrated circuit design

15

14.5

14

13.5

13

12.5

gain [dB] 12

11.5

11

10.5

10 −30 −25 −20 −15 −10 input power [dBm]

Fig. 3.40 Compression behavior of 11 GHz LNA at 11.2 GHz. Gain vs. input power (solid line) and 1 dB-compression point (dashed line).

V cc

RF RF in out

Fig. 3.41 Chip micrograph of 11 GHz low noise amplifier. The total chip size is 1323 µm×923 µm. 10.7–11.7 GHz SiGe downconverter 65

Vcc

RF 600fF 50Ω 5.1nH IF 2.1nH 10pF 744pH 185fF µ 120 m 10pF

2pF 10Ω resistive mixer LO IF amplifier

Fig. 3.42 Circuit schematic of mixer and IF amplifier for the 11 GHz down- converter. Biasing networks are not shown.

3.6.3 Integrated downconverter The LNA is integrated together with a mixer and IF amplifier to form a simple downconverter to an intermediate frequency of fIF =1.45 GHz. Thereby, the MOSFET device, which is offered by the process, is used as a passive resistive mixer with high linearity. The design procedure is basically the same as for the mixer described in section 3.3, but it becomes more important to control the losses of the RF and IF matching networks. Fig. 3.42 shows the circuit schematic of the mixer: a 600 fF capacitor is used for high-pass RF matching. The low-pass IF matching network poses a design difficulty: the required inductance leads to a large spiral inductor with a low self-resonance frequency. It also introduces high par- asitic losses at the RF frequency, when connected to the drain. To solve this problem, the inductance is realized by two smaller series inductors. The self-resonance of the first inductor is lowered to the center of the RF- band by adding a parallel capacitor. This results in an open circuit at RF, which removes the influence of the parasitics of the second inductor on the signal at the drain. For the design of the IF amplifier, it is assumed that the signal level is high enough and its noise contribution is not critical. To avoid the large- value inductors needed for resonant matching, the amplifier is resistively matched, as also shown in Fig. 3.42. Fig. 3.43 shows the measured conversion gain of the downconverter versus LO power level. For an LO power higher than 3 dBm the conversion gain saturates at about 10 dB. The measured gain and noise figure are depicted in Fig. 3.44. The conversion gain is higher than 8 dB over the 10.7 –11.7 GHz range. The single sideband noise figure is 7 dB at 11.2 GHz and raises to more than 8 dB towards the band edges. The measured noise figure is significantly higher than the simulated of 66 Integrated circuit design approximately 3.5 dB. This is attributed to the fact that both, the low- noise amplifier and the resistive mixer, show significantly higher losses than simulated. It can be estimated that the signal gain from the system input to the output of the mixer is less than 5 dB and, therefore, the contribution of the resistively matched IF amplifier to the overall noise figure is significantly higher than intended. Fig. 3.45 shows the measured compression behavior. The 1 dB input compression point of −14 dBm is higher than simulated. This can be explained by the lower gain of the pre-amplifier. It is expected that the downconverter noise figure can be reduced by adding a simple gain stage between LNA and mixer. In Tab. 3.4 the performances in terms of noise figure and linearity are summarized and compared to other published results. Where no explicit values were given, the input compression points as well as the single side- band noise figures were estimated from the given values. These estimated values are stated in brackets. Fig. 3.46 shows a picture of the downconverter. The mixer and ampli- fier were added to the reused layout of the low-noise amplifier in Fig. 3.41. The total chip size is 1323 µm×1323 µm (1.75 mm2). The power con- sumption is 27.6mA from a 3.3 V supply (80 mW).

11

10.5

10

9.5

9

8.5

8

7.5 conversion gain [dB] 7

6.5

6 −10 −5 0 5 10 15 LO power [dBm]

Fig. 3.43 Conversion gain of integrated downconverter at 11.2GHz vs. LO power. (PRF = −30 dBm) 10.7–11.7 GHz SiGe downconverter 67

12

10

8

6

gain/NF [dB] gain 4 SSB−NF

2

0 10.8 11 11.2 11.4 11.6 freq [GHz]

Fig. 3.44 Conversion gain (solid line) and single-sideband noise figure (dashed line with crosses) of the downconverter vs. frequency. LO power is 5 dBm, RF power −20 dBm.

12

11

10

9

8

7

6

5 conversion gain [dB]

4

3

2 −30 −25 −20 −15 −10 −5 RF input power [dBm]

Fig. 3.45 Compression behavior (solid line) of the downconverter at 11.2 GHz and LO power of 5 dBm and 1 dB-compression point (dashed line). 68 Integrated circuit design

Table 3.4 Published downconverters on 50 GHz-ft SiGe processes. Values in brackets were not given explicitly and are estimations.

work function P1dB,input NF (SSB) [53] 11.2 GHz mixer (−15.5dBm) 12.4 dB [54] 17 GHz mixer (−19.6 dBm) (14.5 dB) [55] 20 GHz mixer (−9 dBm) (20 dB) [56] 24 GHz front-end −32.3 dBm (>9dB ) this work 11 GHz front-end -14 dBm 7 dB

Vrf Vif

RF LO in in

IF out

Fig. 3.46 Chip micrograph of downconverter. Mixer and downconverter are added to LNA (compare Fig. 3.41). Total chip size is 1323 µm×1323 µm. Conclusions 69 3.7 Conclusions This chapter describes the design of a 5–6 GHz monolithically integrated downconverter for the use in antenna arrays. It is found that the chosen classical heterodyne architecture in this case is able to compete with other approaches. In particular it was possible to reach an image rejection of 48 dB which is a good value for integrated circuits. This performance is basically limited by parasitic on-chip coupling. Good linearity and a low system noise-figure could be combined by the use of a resistive mixer. To enable a calibration of the downconverter without introducing a switch and the associated losses at the input, a switchable low-noise am- plifier is proposed. A gain and phase difference of less than 0.5dB and 4◦ is demonstrated. A medium output-power power-amplifier is designed and its thermal behavior is studied. It is used later in chapter 5.4 to study the practical implications in multiple antenna systems. An integrated 11 GHz downconverter is demonstrated on a low-ft SiGe process to explore future directions. A MOSFET device is used as a resistive mixer to obtain low-noise and a good linearity. Although the implementation losses are higher than expected, a reasonable performance is reached.

4 Passive array

This chapter focuses on the design of the passive antenna structure. The choice of aperture-coupled patch antennas is motivated first. An equiva- lent circuit model is developed, which helps to efficiently design this type of antenna for given specifications. It is demonstrated that this antenna can be modified to provide a differential interface towards modern integrated circuit technologies. The integration of antennas into an array is discussed, and the effect of mutual coupling is considered. It is shown that this coupling can be com- pensated for the proposed array. The equivalent circuit model is extended from the single antenna to the whole array by including the mutual cou- pling. Last, it is experimentally demonstrated, that the mutual coupling can be reduced, if the antenna-interface in an active array is optimized.

4.1 Antenna design 4.1.1 Choice of antenna structure To select an appropriate antenna structure, it is necessary to define the specific needs for the application in an active array: • The antenna has to operate over the specified bandwidth, here from 5.15 to5.875 GHz. A frequency selective antenna is preferred over a broadband structure, because it pre-filters all out-of-band interferers. • The antenna should be small in geometrical size. An antenna spacing of λ0/2 from center-to-center must be possible. • The antenna should have a low directivity, so that the array pattern is dominated by the steering function and not by the pattern of the individual elements. • For flexibility the antenna should allow both orthogonal polarizations. • The antenna structure should be easy and cheap to fabricate and it has to be compatible with monolithically integrated circuits (MMICs). 72 Passive arrays

antenna patch substrate

ground plane

slot circuit substrate

matching stub

Fig. 4.1 Exploded sketch of aperture-coupled patch antenna; separate sub- strates can be individually selected for circuit and antenna, magnetic coupling is achieved by a slot in the common ground plane.

All these specifications are fulfilled by the aperture-coupled patch an- tenna [58], although the required relative bandwidth of more than 13% is at the upper limit of what can be achieved with this structure [59].

4.1.2 Aperture-coupled patch antenna The demands on microstrip substrates are contradictory for the design of circuits and antennas [59]: to achieve good radiation properties and a large bandwidth a thick substrate with a low dielectric constant r is needed. If passive structures like filter, coupler and transmission lines are implemented, radiation has to be avoided. This requires a thin high r substrate, which confines the electric field to the dielectric material between conductors and ground plane. This conflict can be resolved by using two independent substrates for circuit and antenna as shown in Fig. 4.1. The antenna patch is printed on a thick, low-r substrate, while the circuit resides on a thin, high r substrate. Both are mounted together and share the same ground plane. There are several ways to excite the patch: capacitive, magnetic coupling or probe feeding. Magnetic coupling is proposed, because it can be achieved by etching a slot in the common ground plane and therefore maintaining a simple planar structure [58]. Numerical tools have to be used to simulate and optimize the antenna, which involves a high computational effort. To gain more insight into the influence of the single parameters, and therefore reduce the number of needed simulation runs, a lumped-element circuit model was developed. Using this modelling approach, the design can be speeded up significantly. Antenna design 73

resonator

magnetic coupling

reference plane

Fig. 4.2 Equivalent circuit of aperture-coupled patch antenna: a resonator is magnetically coupled to the transmission line.

Lp Ls

l el1 Rp Rs l el2

Cp Cs a) single resonator model b) double resonator model

Fig. 4.3 Single (a) and double resonator (b) equivalent circuit to model the electrical behavior of the three dimensional antenna structure

4.1.3 Modelling Microstrip structures can accurately be modelled, including losses, disper- sive behavior and end-effects [60]. These models are widely known and available in all modern microwave circuit design tools. Therefore, it is advantageous to separate the design process into the optimization of the patch and slot structure, and the determination of the microstrip feeding and matching-structure, respectively. This reduces the parameter-space for the computational expensive field simulations. Placing a port at both ends of the microstrip-line, the structure can be considered as a resonator coupled to a transmission line as shown in Fig. 4.2. In [61] the equivalent circuit is simplified to a parallel RLC resonator in series with an inductor, as shown in Fig.4.3a. For the ge- ometries and materials used in this work, a two resonator model as shown in Fig.4.3b models the antenna behavior more accurately over a broader frequency range. Fig.4.4 shows the simulated insertion loss of the antenna structure in comparison with both models: it is obvious, that the second resonance at around 7.5 Ghz cannot be described by the single resonator model. Furthermore the two resonator model is more accurate inside the operating bandwidth from 5 to 6 GHz. So far, all model parameters can be determined from magnitude simu- lations (or measurements) and no precise phase information is necessary. 74 Passive arrays

0

−5

−10

−15 insertion loss [dB]

−20

−25 4 4.5 5 5.5 6 6.5 7 7.5 8 freq [GHz]

Fig. 4.4 Transmission coefficient of simulated antenna sructure(solid), one res- onator model (dashed with boxes) and two resonator model (dashed with circles)

If the full wave simulation is deembedded and the reference planes for both ports are shifted to the center of the slot, the transmission function exhibits an additional phase shift of around 20◦ at 5.5 GHz. In the model this is accounted for by adding two ideal transmission lines of correspond- ing lengths lel1 and lel2 at the input and output.

4.1.4 Design method Considering the antenna structure described above and the two resonator model in Fig. 4.3 it is not possible to postulate a general relationship between the geometrical parameters and the values of the equivalent cir- cuit. Nevertheless, it is possible to evaluate the sensitivity of the circuit values to certain geometrical changes [62]. This provides several useful dependencies: • As expected, varying the patch size changes the resonance frequency of the first resonator formed by Rp, Lp and Cp. This determines the operating frequency of the antenna. • The size, geometry and orientation of the slot determines the cou- pling of the antenna resonator to the microstrip line. An increased coupling leads to a higher value Rp, but also to a higher characteristic −1 impedance X0 = jω0Lp = (jω0Cp) . It is found that the quality factor Q stays nearly constant with different sizes and positions. Antenna design 75

• Increasing the slot size not only gives a higher coupling, but also lowers the resonance frequency of the second resonator, formed by Rs, Ls and Cs. This indicates, that the second resonance is related to the slot structure.

• Different shapes of the slot lead to different quality factors Q of the first resonance.

• Change of thickness or dielectric constant r of the patch substrate leads to both, a change of the quality factor and the first resonators resonance frequency.

From the above, one advantage of this modelling approach becomes visible: the quality factor, which is related to the maximum achievable antenna bandwidth, varies only with different slot shapes or antenna substrates. For a given combination it can be evaluated once, using a single field simulation. Therefore, the following design procedure is suggested: first, adequate slot shape and substrate type are selected for the aimed antenna band- width. From several slot types proposed in [58] the “dogbone” slot is found to provide the highest bandwidth. High bandwidth can also be achieved by choosing a thick, low-r substrate Once the combination is fixed, the resonator coupling is set by adjusting the slot. Generally the coupling is higher for a slot centered in the middle of the patch, where the highest current occurs. A higher coupling is also obtained by increasing the slot size. Moreover, it is advantageous to keep the slot resonance frequency as high as possible. If the slot gets into resonance, it radiates to both sides of the ground plane, thus degrading the front-to-back ratio of the antenna. Next, the antenna center frequency is adjusted by varying the patch size. And last, the length of the matching stub is calculated to compensate the inductive part, that remains at the center frequency of the first resonator. Fig. 4.5 shows a typical simulated input matching of the matched an- tenna structure. It demonstrates the typical design trade-off between re- turn loss and matching bandwidth: the trajectory of the reflection coef- ficient in the Smith chart circles around the matching point. This cir- cle reaches its highest impedance at the patch resonance, the impedance equals Rp of the equivalent circuit. With equivalent coupling, an Rp of 50 Ω and therefore a very low return loss at the center frequency could be achieved. If a certain return loss is tolerated, the Rp can be further increased, resulting in a higher bandwidth. A typical requirement for an- tennas is a voltage standing-wave ratio (VSWR) of lower than 2. This corresponds to an input return loss of approximately −10 dB. 76 Passive arrays

5 GHz

6 GHz

Fig. 4.5 Simulated input reflection coefficient S11 of the antenna structure, the second port is terminated with an open-ended stub of appropriate length.

4.1.5 Results The final antenna is designed for an impedance bandwidth (VSWR≤2) of 5.15–5.875GHz. A 635 µm-thick substrate with an r of 10.2 (Duroid 6010) is used as substrate for the circuit side. The patch is printed on Polyguide substrate with a thickness of 3.125 mm and an r of 2.32. This leads to a line width w50 of 600 µm for a 50 Ω-line. The final layout is shown in Fig. 4.6, the corresponding dimensions are given in Tab. 4.1. As discussed above, a dogbone slot is used to achieve the highest possible bandwidth. Fig. 4.7 shows the measured and simulated return loss of the manu- factured antenna. The desired impedance bandwidth is achieved and the measurement agrees well with the simulation. The return loss has a lo- cal maximum in the center of the band. Improving this return loss by reducing the coupling would lead to a reduced overall bandwidth. Antenna design 77

l inset

r patch l l slot w 50

w slot l stub

Fig. 4.6 Layout of broadband patch antenna using a dogbone slot.

Table 4.1 Final geometry of 5.15–5.875 GHz antenna Dimension Value

lslot 7.6 mm wslot 0.9 mm r 0.81 mm linset 1.35 mm lstub 0.62 mm lpatch 14.5 mm

0

−5

−10

−15 return loss [dB]

−20

−25 4 4.5 5 5.5 6 6.5 7 7.5 8 freq [GHz]

Fig. 4.7 Measured (solid line) and simulated (dashed line) input return loss of the broadband aperture-coupled patch antenna. 78 Passive arrays

plane of double symmetry slot

differential feed

Fig. 4.8 Sketch of differential antenna: a double slot couples the symmetric radiator to a symmetric feed, providing a differential port 4.2 Differential antenna 4.2.1 Differential MMIC interface Aperture-coupled patch antennas are well suited for the use in active ar- rays, because their fabrication is simple due to their planar nature. Ad- ditionally they are compatible with microstrip circuitry and monolithic integrated circuits. To achieve low-cost, robust and compact multiple antenna system, it is desirable to integrate each single antenna element with a monolithic RF-frontend. Those frontends are often implemented using differential techniques, especially when image-rejecting architectures are chosen [47]. Furthermore, using balanced circuits, the crosstalk across common bias lines can be reduced significantly. On-chip single-ended-to-differential con- version usually is linked with losses. If losses are introduced before am- plification of the signal, this results in a higher noise figure of the system. Therefore a differential antenna signal leads to a better performance of the system. It was shown [63] that a suitable antenna structure can be used as an almost lossless power divider. 4.2.2 Design The field pattern of a patch antenna is inherently symmetrical. Therefore, there are several ways to arrange one or two slots to provide a differential interface. The layout shown in Fig. 4.8 is optimized to have a simple feeding network and a low cross-polarization: the slots are placed symmetrically Differential antenna 79

coupling patch slots

patch substrate amplifier

power balun supply structure TB02-037

Fig. 4.9 Layout of the active differential antenna at each side of the patch, so that the feed lines meet in the center below the patch. The feed lines are also centered in the perpendicular direction. This way the antenna structure itself maintains two planes of symmetry. Ideally this leads to zero cross-polarization into the main direction of radiation. To optimize the differential antenna in terms of bandwidth and return loss, the same method as described above in section 4.1.4 is applied to the differential impedances.

Low-noise amplifier The employed low-noise amplifier was fabricated using the Triquint TQTRx 0.6 µm GaAs process (see chapter 3). A standard source degenerated stage was used to form a differential couple. To guarantee common-mode sta- bility, the virtual grounds were loaded with resistors, which do not affect the differential noise performance. The amplifier output is also differential. One amplifier from the same wafer was bonded to a test substrate including two microstrip baluns for single-ended to differential conversion. Measured gain is more than 12 dB, with a noise figure between 3 dB and 3.5 dB.

4.2.3 Measurement results Two different antennas were manufactured: A simple passive antenna and one active antenna using the monolithic low-noise amplifier described above. The layout of the active antenna is shown in Fig. 4.8. To allow mea- surements, both circuits use a microstrip balun to obtain a single-ended output. For later applications this balun would be omitted to interface directly with a balanced mixer. Both structures were measured inside an anechoic chamber using a 80 Passive arrays

Narda 642 standard gain horn. The measured gain in Fig. 4.10 shows about 5.5 dBi gain for the passive structure. The active antenna has 14 dBi to 17 dBi over the 5–6 GHz band. Both curves show some ripple over frequency which is due to the presence of the connectors. For comparison with a standard system, where antenna and amplifier are separated, the transducer gain of the embedded amplifier can be cal- culated if the gain of the active antenna and of a passive reference are known [64]. The computation using the averaged gain curves lead to 10 to 12 dB gain over the band. Fig. 4.11 shows the measured and simulated pattern of the active an- tenna at 5.8 GHz. The agreement between measurement and simulation is good regarding the co-polarization. The antenna pattern was found to be similar over the whole band. The sensitive measurement of the cross- polarization is influenced by the near connectors. Therefore, it differs from the simulated pattern, but the cross-polarization rejection of better than 27 dB is very low. The measured front-to-back ratio is higher than 8 dB. Fig. 4.12 shows a photography of the manufactured active antenna. The size of the whole antenna is 50.8 mm × 50.8 mm, the antenna sub- strate is 25.4 mm × 25.4 mm.

20

active 15

10

passive antenna gain[dBi] 5

0 4.5 5 5.5 6 6.5 7 freq/GHz

Fig. 4.10 Measured gain of active differential antenna (solid line) compared to passive reference (dashed line). Differential antenna 81

90 120 60

150 30

[dB]

−20 −10 180 0 0

cross−pol. 210 330

co−polarization 240 300 270

Fig. 4.11 Measured (solid line) co- and cross-polarized pattern of active differ- ential antenna compared to simulations (dashed line).

Fig. 4.12 Circuit and radiator-side photography of active differential antenna. The size of the whole antenna is 50.8 mm × 50.8 mm, the antenna substrate is 25.4 mm × 25.4 mm. 82 Passive arrays 4.3 Antenna arrays and mutual coupling 4.3.1 Non-ideal arrays A practical antenna has certain geometrical extensions and therefore dif- fers from the idealized point source. The field radiated by an antenna can be roughly seperated into three regions: • Reactive near field: this region close to the antenna structure is dom- inated by field components that drop off with 1/r3 or higher orders with increasing distance r from the antenna. It contains electromag- netic energy, which is not radiated to the free space. If any object is introduced into this zone, the directional pattern, but also the feed- point impedance, are changed. • Radiating near field: this region, also called Fresnel zone, does not contain significant amounts of reactive energy. It is characterized by field components with radial dependency. Objects introduced here lead to a change of the directional pattern, but the antenna matching is not noticeably changed. • Radiating far field, or Fraunhofer region: the radiated wave in this region equals an ideal spherical wave: the field pattern does not vary with increasing distance, only the field energy decreases with 1/r2. At a certain distance this wave appears as an almost planar wave-front. The terminals in a communication systems typically are sufficiently sepa- rated to assume far-field operation. The dominant spherical wave justifies the use of the quasi-optical model. The main difficulty arises when the antennas are moved together to fulfil the spatial Nyquist condition d ≤ λ/2: the necessary maximum distance lies in the range of the geometrical antenna size, clearly inside the extension of the reactive near field. The resulting mutual coupling is an important effect in antenna arrays, which can not be neglected and will be discussed more detailed below. Furthermore, in a practical system the antennas have to be mounted on some support with finite extensions. Radiated waves reaching these edges are diffracted. Following the geometrical model, these edges act like parasitic sources, which can not be controlled independently, causing a change of the radiation pattern. It is important for a correct operation of any antenna array to main- tain the precise phase and amplitude relationships of all elements. This is especially a problem in active arrays, where components with certain variations are used. Chapter 5 addresses the possible calibration methods to eliminate this effect. Antenna arrays and mutual coupling 83

4.3.2 Classification The mutual coupling in antenna arrays has various effects, spanning from the displacement of diagram zeros over changes of input matching to the occurrence of blind angles. Antenna arrays showing mutual coupling can be classified into some subcategories. If two antennas are separated in a way, that the reactive near-field (see section 4.3.1) is not noticeably changed by the presence of the neighbor element, it can be assumed that the far field pattern also stays unchanged. Under this assumption of well-behaved antennas, the pattern of an element embedded into an array is identical with the pattern of an isolated element gi:

g(φ) ≈ gi(φ). (4.1)

Well-behaved arrays can be obtained using small antenna elements or large spacings. A second important criteria is, whether a unique radiation mode of each element can be associated with the according terminal voltage. If this condition is fulfilled, an array can be modelled with respect to the electrical states at its ports [65]. Without regarding the actual field distributions it can be distinguished between weak coupling and strong coupling, which refers to a measurable effect at the antenna ports. If

z11 ... z1N . . . Z =  . .. .  (4.2) z ... z  N1 NN    is the impedance matrix of the array, the electrical behavior can be de- scribed by the vector equation

V = ZI, (4.3) with v1 i1 . . V =  .  and I =  .  v i  N   N      being the terminal voltage and current vectors. In a weakly coupled array mutual impedances zij (i 6= j) are present, but they do not influence the impedance seen into one of the ports. In a strongly coupled array this 84 Passive arrays Table 4.2 Categories of antenna arrays with mutual coupling Category Effect well behaved embedded pattern does not change single moded a unique mode pattern can be asso- ciated with each port voltage strong/ weak coupling input impedances depend on steer- ing angle blind angles excitation of surface waves, no radi- ation possible for certain angles is not the case; the mutual impedance also changes the port impedance. The so-called active impedance of the m-th element is defined as

v N i Z = m = n · z . (4.4) m i i mn m n=1 m X According to Eqn. 4.4, the active impedance depends on the actual exci- tation of the array, as the ratio in/im changes with moving steering angle. This can lead to a significant reduction of the radiated power. In the worst case, the active input reflection coefficient Γm reaches |Γm| = 1 for certain angles. In this case a surface wave along the antenna elements is excited and no energy can be radiated. This phenomenon is called blind angle and needs to be avoided in practical applications. The different categories of antenna arrays are summarized in Tab. 4.2. Dependent on the type of mutual coupling, the effects on the array oper- ation are also different (a single-mode relationship is assumed): • A well behaved, weakly coupled array shows deformed patterns, if one array element is measured alone. This is due to the parasitic excitation of the neighbor elements. In a synthesized array pattern this results in shifted and degraded diagram zeros and sidelobes. The coupling effect can be compensated (see section 4.3.3) to obtain the simple model of uncoupled antennas (pattern multiplication). • If the array is weakly coupled, but not well behaved, the far-field pattern of each element might be different. Thus, a transformation of the port signals is not sufficient to obtain the simple array description. In large arrays, all inner elements have an identical environment. This leads to similar embedded element patterns gemb(φ). Using these, instead of the isolated element patterns gi(φ), it is again possible to apply the simple pattern multiplication principle [66]. To achieve Antenna arrays and mutual coupling 85

equal pattern for all elements, further “dummy” antennas can be placed around the array. • Strong coupling and blind angles can not be compensated. The array has to be designed to avoid these effects. Under the assumption that each element is linked with its reactive near- field mode pattern, the mutual coupling is identical for the transmit and receive case. This is not generally true. However, for practical arrays with weak coupling the agreement is found to be sufficiently good. 4.3.3 Coupling compensation Under the assumption that each antenna field can be described by a single mode and an associated port voltage, it is possible to compensate the electrical effect of coupling: looking into the feeding ports, the array can be entirely described by its impedance matrix Z. Thereby, the diagonal elements zii account for the radiation resistance of each antenna, while the off-diagonal elements zij represent the mutual coupling. Thus, the 2 radiated power is proportional to the terminal voltages |vi| [67]. Now using the S-parameters representation, a forward-moving wave vector a occurs at the wave ports of the array. These waves are partly reflected and cause a back-travelling wave vector b = Sa, where the S matrix is just a different representation of Z. Then, the terminal voltage vector under the influence of coupling is

V˜ = Z0 (a + b)= Z0 (1 + S) a. (4.5) It is obvious, that with risingp mutual couplingp the transmission coefficients sij become important. Therefore, there is no longer a simple proportional relationship between the terminal voltage vi and the forward wave ai:

vi = f(a1, · · · ,ai, · · · ,aN ) 6= f(ai) (4.6) As the array S-parameters can be obtained from either measurement or simulation, the coupling matrix C = 1 + S (4.7) can be calculated. This allows to determine a so-called ”pre-conditioned” excitation vector ˆa = C−1a, (4.8) which leads to the wanted terminal voltages:

−1 Vˆ = Z0C (C a) = Z0a (4.9) p =ˆa p | {z } 86 Passive arrays

vertically polarized horizontally polarized

Fig. 4.13 Four element aperture-coupled patch antenna array with vertical or horizontal polarization. Arrows show orientation of radiated E-field.

4.3.4 Array of aperture-coupled patch antennas So far, no assumption has been made on the field pattern of the elements. To verify if aperture-coupled patch antennas can be used to form a well- behaved array (see section 4.3.2), a vertical and a horizontal polarized configuration are investigated as shown in Fig. 4.13. The center-to-center antenna spacing in both cases is 27.3 mm, which is approximately λ0/2 at the center frequency of 5.5 GHz. Both arrays are simulated using a commercial field simulator (Agi- lent HFSS v5.6). The structures are first driven with the ideal excitation vectors for each element: 1 0 0 1 a = , a = , · · · 1  0  2  0   0   0          The resulting far-field patterns at 5.5 GHz are depicted in Fig. 4.14 and 4.15 for vertical and horizontal polarization, respectively. In both cases all patterns show some deformations, which is a typical effect of mutual coupling. The aperture size of single element is not sufficient to explain the spatial frequency of the superimposed ripple. This indicates, that further radiation sources must be present. The simulated S-parameter transmission coefficients si(i+1) between two adjacent elements, which are often used to characterize the strength of mutual coupling, are −14 dB for the vertical polarized and −17 dB for the horizontal polarized array. Using the simulated S matrices, the pre-conditioned excitations ˆa1, ˆa2, · · · are calculated and applied to both arrays. Thereby, the de-embedded S0 matrix is referred to a reference plane in the center of the slot, whereas Eqn.4.9 refers to the phase center of the antenna. Empirically it is found that the deembedding distance needs to be increased by 40◦ electrical length at 5.5 GHz. The same value can be validated, if the coupling ma- trix is estimated from a transformation of the far-field (see chapter 5). Antenna arrays and mutual coupling 87

5

uncompensated compensated

0

−5 norm. gain [dB]

−10

−15 −150 −100 −50 angle0 [deg] 50 100 150

Fig. 4.14 Element pattern of four antenna vertical array: embedded elements (solid lines), elements after compensation (solid lines with stars) and isolated element (dashed line). Compensated patterns agree well with isolated element pattern.

5

uncompensated

0

−5 norm. gain [dB]

−10 compensated

−15 −150 −100 −50 0 50 100 150 angle [deg]

Fig. 4.15 Element pattern of four antenna horizontal array: embedded ele- ments (solid lines), elements after compensation (solid lines with stars) and isolated element (dashed line). 88 Passive arrays

15

10 uncompensated 5 compensated

0

−5

−10

−15

norm. gain [deg] −20

−25

−30 ideal −35 −150 −100 −50 0 50 100 150 angle [deg]

Fig. 4.16 Four element vertical polarized array with uniform distribution, ◦ steered to 30 : uncompensated (solid line), compensated (solid line with stars) and ideal (dashed line) pattern.

The resulting embedded element patterns are added to Fig. 4.14 and Fig. 4.15. The agreement of embedded patterns and isolated pattern is good over the relevant range of −90◦ ≤ φ ≤ +90◦. An array with elements mounted 45◦-slanted shows similar results. For this type of arrays, this justifies the assumption of well-behaved antennas. To demonstrate the effect on beamforming, a uniform excitation |ai| =1 is chosen. To remove the possible cancelling of effects due to the symme- try of the structure, the beam is steered to +30◦ by setting the excitation vector to 1 i a = .  −1   −i      In Fig. 4.16 the ideal far-field pattern calculated by pattern multiplica- tion, and the patterns resulting from an excitation with a and the pre- conditioned ˆa = C−1a are compared. It is seen that mutual coupling leads to a shift and degradation of the pattern zeros and a slight change of the side-lobe levels with respect to the ideal pattern. Pre-conditioning of the excitation vector compensates for these changes and restores reasonable Antenna arrays and mutual coupling 89

feed 1 Cp Cs port 1 lel1 Rs l stub L L M, ∆τ p s

feed 2 LpL s

port 2 R l el1 s l stub

Cp Cs

Fig. 4.17 Extension of aperture-coupled patch antenna model: a retarded cou- pling inductance models the mutual coupling. low pattern zeros. Outside the relevant range of −90◦ ≤ φ ≤ +90◦ the pattern is dominated by diffraction effects.

4.3.5 Array coupling model The compensation of mutual coupling described above assumes a matched antenna element showing a purely resistive input impedance. Strictly, this is only valid in the center of the operating bandwidth. To describe the array behavior over a broader bandwidth the aperture- coupled patch-antenna model proposed in section 4.1.4 is extended to mu- tual coupling: given that the resonance frequency of the second resonator (RS ,LS and CS ) is sufficiently high, the radiated power of each antenna element equals the power dissipated in the resistors RP . These, therefore, are replaced by ”radiation”ports with an impedance of Z = RP as shown in Fig.4.17. To model the mutual coupling different forms and combinations of inductive and capacitive coupling were explored. The best agreement was achieved using a retarded coupling inductance

V2 = (jωM · exp(−jω∆τ)) I1 (4.10) between all inductors LP of adjacent elements. Fitting the model to field simulations, a ∆τ of 33 ps is found, which equals a free-space propagation distance of around 10 mm. It is also observed that the model parameter of the single antenna element do not have to be changed significantly. Fig. 4.18 shows the S-parameter transfer function between two adja- cent antennas gained from this model compared to a full field simulation using HFSS. The model approximates this function over a broader range. 90 Passive arrays

0.2

0.15

0.1 5.8 GHz

5.5 GHz 0.05

0 imag

−0.05

−0.1

−0.15 5.2 GHz −0.2 −0.2 −0.1 0 0.1 0.2 real

Fig. 4.18 S-parameter transfer function between two neighbor elements: field simulation (solid) and lumped-element model (dashed). Data ranges from 4.5 to 6.5 GHz. One symbol every 100 MHz from 5 to 6 GHz.

The best agreement is found at the center of the 5–6 GHz band. At the band edges an magnitude error occurs, which cannot be avoided using this simple model. This simple network-model has the advantage that it can directly be incorporated into any network simulator. It allows to simulate influences like amplifier matching and their effects on mutual coupling in a single step and, therefore, facilitates the design and optimization of active antenna arrays.

4.4 Reduction of coupling in active arrays 4.4.1 Interface optimization To study the influence on mutual coupling, different matching impedances are investigated. For simplicity it is assumed, that all ports are terminated with the same impedance. In this case it is possible to evaluate the effect of mismatching by renormalizing the simulated S-matrix: the corresponding Z-matrix −1 Z = Z0 · (1 + S) (1 − S) , (4.11) Reduction of coupling in active arrays 91

5

25 Ohm

0

−5

90 Ohm norm. gain [dB]

−10 isolated

−15 −80 −60 −40 −20 0 20 40 60 80 angle [deg]

Fig. 4.19 Second element in a four element vertical polarized antenna array: elements terminated with 50 Ω (solid line), 25 Ω (crosses), 90 Ω (diamonds) and isolated element (dashed line). A low termination impedance reduces the coupling. leads to the renormalized matrix S0 using the new normalization impedance Znew: 1 −1 1 S0 = Z + 1 Z − 1 . (4.12) Z Z  new   new  Then, the new coupling matrix C0 is given by

Z C0 = 0 (1 + S0) . (4.13) Z r new In a system with a nominal impedance of Z0 a certain mismatch usually can be tolerated, as long as the power transfer is not significantly degraded. This condition can be fulfilled by a range of termination impedances. A return loss of 10 dB, for example, is approximately achieved by a termina- tion impedance of 25 Ohm or 90 Ohm. The difference between these two terminations and 50 Ohm as a reference value is investigated further. The directional patterns that results from the excitation of the second element in a four element vertical polarized array are calculated for all three impedances. They are depicted in Fig. 4.19, where they are also compared to the isolated element pattern. For all three impedances an effect of mutual coupling is visible that leads to a deformation of the 92 Passive arrays ideally isolated element pattern. Furthermore, the optimum impedances for power matching and minimal mutual coupling do not coincide: the pattern deformation resulting from mutual coupling decreases with lower terminating impedances. Therefore, if a low mutual coupling is desired, the antennas should be terminated at the lowest possible impedance that still permits a sufficient power transfer.

4.4.2 Experimental verification To verify if the predicted reduction of mutual coupling is practically ap- plicable, an experimental setup has been built. To form an active receiving antenna array, monolithically integrated low-noise amplifiers are utilized. Fig. 4.20 shows the complex input reflec- tion coefficient which results from a typical design trade-off in low-noise amplifiers based on field-effect transistors: the device used for the first stage typically shows a capacitive behavior. Feedback is needed to obtain a real input impedance (see chapter 3.2). This, on the other hand, reduces the possible gain of this first stage, which is crucial for the noise perfor- mance of the amplifier. Therefore, a weak feedback is chosen, providing a real input impedance just sufficient to obtain the targeted input return

6 GHz

5 GHz

Fig. 4.20 Input impedance of low-noise amplifier used in the three antenna ac- tive array from 0 to 10 GHz. The small real part results from a design trade-off: to increase the real part, gain needs to be sacrificed. Reduction of coupling in active arrays 93

l trafo

low−noise amplifier

Fig. 4.21 Layout of the three element active receiving array. Different lengths ltrafo lead to a high and a low impedance at the antenna phase center. loss. Integrating this amplifier into an array of vertically polarized antennas, as shown in Fig. 4.21, the actual impedance at the antenna feed depends on the distance ltrafo between amplifier and feed. At 5.5 GHz the ampli- fier has a return loss of −8 dB. To present a purely real impedance at the antenna phase center, line lengths ltrafo of 8mm and 13.5mm are required, if the additional electrical distance of 40◦ from the feeding point is taken into account (see section 4.3.3). The resulting impedances are approximately 22 Ω and 116 Ω, respectively. Both structures are simulated using a network simulator. The coupled array model proposed in section 4.3.5 is used and combined with the mea- sured S-parameters of the low-noise amplifier. The coupling matrix C can be evaluated by simply simulating the transfer functions from the inter- nal antenna ports to the three output ports. To obtain results which are independent of the amplifier gain, the results can be divided by this gain. This leads to quantities which are directly comparable to the measured s21 between two elements in a passive array, which is a usual measure for the characterization of mutual coupling. For the center element, the normalized coupling coefficient is

c12 ccoupl = (4.14) c22

The coupling coefficients for both transformation distances ltrafo are shown in Fig. 4.22. For the center frequency of 5.5 GHz, this model predicts a significantly reduced mutual coupling for the arrangement showing a low impedance at the antenna. This reduction is limited to a certain bandwidth, as the line transformation is frequency dependent. Both arrays were manufactured and their far-field radiation patterns were measured in an anechoic chamber. For comparison with the simula- tions, the coupling matrix C is computed from the far-field data using the 94 Passive arrays

−10

−12 8 mm −14

−16

−18

−20

13.5 mm

magnitude of coupling coeff. [dB] −22

−24 4.5 5 5.5 6 6.5 freq [GHz]

Fig. 4.22 Mutual coupling coefficient c12/c22 simulated (dashed line) and ex- tracted from measurements (solid line) for the two amplifier line lengths.

Fourier transformation method, which is described in chapter 5. The re- sulting coupling coefficients are added to Fig. 4.22. The limited frequency range is due to the lack of measurement data towards higher frequencies and the constrained range of the transformation (d ≤ λ/2). The results agree qualitatively with the simulations, but a growing difference appears at the end of the antenna bandwidth. This is explained by the fact that the circuits, in this roll-off region, show the highest sensitivity to possible tolerances. Any deviations from the idealized model, like edge diffraction, reflections, or disturbance from the connectors, lead to an error in the calculation of C. However, both, simulation and far-field measurement, demonstrate that the right choice of the array termination impedance leads to a sig- nificantly reduced mutual coupling (simulated s21 with a termination of 50Ω is −14 dB) at the center frequency. This is also confirmed by the measured active patterns of the center element in both arrays, shown in Fig. 4.23. The array terminated with the lower impedance clearly shows an improved pattern compared to the other array. Conclusions 95

90 30 120 60 25

150 20 30

15

180 0

210 330

240 300 270

Fig. 4.23 Measured directional pattern of the center element of the three an- tenna array at 5.46 GHz, absolute gain (antenna and amplifier) in dBi. LNA is mounted at different distances to present a low (solid line) or a high (dashed line) impedance. For the low impedance the pattern is less degraded by mutual coupling. 4.5 Conclusions In this chapter the design of the passive array was discussed. A lumped- element equivalent circuit model for aperture-coupled patch antennas sim- plifies the design for given specifications. A symmetric antenna structure is proposed to provide a differential electrical interface, which is attractive for the integration of the antenna with integrated circuits. The mutual antenna coupling is studied and it is found that aperture- coupled patch-antenna arrays can be approximated by loosely coupled isolated elements. This well-behaved array condition allows to remove the effect of the coupling. An extension of the proposed equivalent circuit enables the simulation of mutual coupling using a network simulator. By suitable impedance mismatch on antenna arrays, the mutual coupling can be reduced to a certain extend. This is demonstrated by a standard low-noise amplifier 96 Passive arrays mounted with an optimized electrical distance between amplifier and an- tenna. This method, so far, is limited to a certain bandwidth. It is possible to improve its bandwidth by choosing a different line impedance or by a proper amplifier design. 5 Calibration

This chapter discusses the main aspects of antenna array calibration. It starts with a review of the known methods based on far-field sources and test-tones. A novel method is added, which allows to reduce the com- plexity of the calibration network. Last, the dynamic performance is in- vestigated. This becomes important when parts of the active circuits are switched on and off and alter their characteristics, as it occurs in power amplifiers. 5.1 Problem formulation Most array processing techniques assume a set of independent sensors, which measure the phase and amplitude of the electromagnetic field at their locations without influencing it. Practical antenna systems at giga- hertz frequencies show several deviations from this idealized signal model: • To avoid grating lobes, sensor distances half of the signal wavelength, or lower, are required. The low spacing causes mutual coupling be- tween the antennas. • Finite ground planes introduce metal edges, which diffract arriving waves. They act as parasitic elements and alter the array pattern. • The active hardware, including amplifiers and mixers, exhibit gain and phase variations. These depend on fabrication tolerances, biasing, temperature and aging. • Cables and connectors are a further source of phase and amplitude variations. Generally, their sensitivity to mechanical stress rises with the frequency of operation. A calibration is needed to find a mapping between the real antenna array and and the idealized sensor model. Thereby, also the practical feasibility of a calibration method is an important selection criteria. Two general methods can be distinguished: those, that use separate antennas and calibrate via the air interface and those, which use internal calibration signals. 98 Calibration

L Rq C R l resonant matching

Fig. 5.1 Two element resonant power matching for Rl ≥ Rq

5.1.1 Active circuit variations The practical difficulty to obtain identical and time-constant transfer func- tions is illustrated in the following simplified example: Fig. 5.1 shows a two stage resonant power matching between a source impedance Rq and a load Rl. This circuit represents the typical input matching of a transistor amplifier with the capacitance C, or a fraction of it, being the gate-source capacitance of the device. If Rl ≥ Rq are given, the values L and C for optimal power transfer can be calculated:

1 2 L = RqRl − Rq (5.1) ω0 q 1 1 1 C = − 2 , (5.2) ω0 sRqRl Rl where ω0 = 2πf0 is given by the center frequency f0. Strictly, optimum matching is obtained only for this frequency. The practically usable band- width depends on the quality factor Q of the matching network. In this case it is defined as

ωL Rl Q = = ωCRl = − 1 (5.3) Rq sRq and depends on the impedance transformation ratio Rl/Rq. If it is now assumed that C is dominated by the intrinsic capacitances of a transistor, it may strongly vary from the ideal value. Intuitively it is clear that the transfer function from the source to the load resistor is affected stronger, the higher the quality factor is. To give a quantitative example, the am- plitude and phase changes caused by the capacitance variation are given in Fig. 5.2 and Fig. 5.3 for different quality factors. For element variations of ±25% or more, which are not uncommon in monolithic circuit fabrication, significant amplitude and especially phase errors are obtained. Problem formulation 99

0

−0.2 1 −0.4 2 −0.6 3 −0.8

−1

−1.2

−1.4 amplitude variation [dB] −1.6

−1.8

−2 50 75 100 125 150 element variation [%]

Fig. 5.2 Amplitude changes caused by variations of the capacitor C for different quality factors of the matching network.

40

30

20 ] o 10

0

−10 1 phase variation [ −20 2 −30 3 −40 50 75 100 125 150 element variation [%]

Fig. 5.3 Relative phase changes caused by variations of the capacitor C for different quality factors of the matching network. 100 Calibration

These errors could by decreased, if broadband circuits were used. Nev- ertheless, several advantages as higher stage gain, lower power consump- tion and better selectivity and blocking capabilities suggest to use reso- nant circuits. This requires that the emerging errors are determined and compensated.

5.1.2 Calibration network precision As it will be seen later, a local signal can be used to characterize and compensate the variations of the active hardware. This requires a precise distribution network with identical phase and amplitude responses to pro- vide a reliable reference. This concerns the divider network as well as the coupling to the signal paths. The design of the divider network becomes particularly critical when arbitrary or two-dimensional antenna arrangements are chosen. Meander- lines have to be placed to ensure identical electrical distances. In mi- crostrip realizations this involves the precise modelling of any line discon- tinuities and possible coupling effects. To apply the calibration signal to the active circuit, an active switch was proposed in chapter 3.4.2. There, a reproducibility of ±0.5 dB and ±4◦ was found. Also passive components possess a limited precision. E.g., the uncertainties of microstrip realizations are caused by the local varia- tions of the lithographic and etching process. For comparison, the coupling

0.5 0.4 0.3 0.2 0.1 0 −0.1 −0.2

amplitude variation [dB] −0.3 −0.4 −0.5

90 95 100 105 110 line width [%]

Fig. 5.4 Changes of coupling attenuation of a 10 dB directional coupler that result from a non-precise line width. The phase variation stays below 0.4◦. Problem formulation 101 attenuation of a 10 dB directional coupler on a 635 µm substrate with r = 10.2 is simulated for different line widths and depicted in Fig. 5.4. For a typical precision of ±5%, an amplitude variation of ±0.4 dB occurs, while the phase variation stays well below 0.4◦.

5.1.3 Calibration requirements The tolerable difference between the ideal model and the practical imple- mentation depends strongly on the aimed application. Generally it has to be distinguished between point-to-point links and multi-user systems based on spatial filtering. While it is sufficient for point-to-point links to focus the antenna beam into the right direction, multi-user systems require the rejection of interfering signals. This rejection bases on the cancellation of signals and is very sensitive to any signal difference. Fig. 5.5 shows the tolerated amplitude a and phase φ mismatch to reach a certain rejection level gnotch by subtracting two signals according to

jφ gnotch =1 − a · e . (5.4)

To reach a rejection level of −20 dB, referred to the power of a single signal, an amplitude error of less than 0.83dB or a phase error of less than 5.8◦ can be tolerated. If both errors are present simultaneously, even lower variations have to be demanded. Fig. 5.6 shows the error constellations, which lead to a constant gain error of a beam gbeam that is formed by the summation of two signals

jφ gbeam =1+ a · e . (5.5)

A comparably large mismatch error of 1 dB or 40◦ only results in a gain change of around 0.5 dB. Both graphs give an estimation of the needed precision to reach certain system capabilities. Not all multiple antenna systems require explicit calibration. In a re- ceiving array the individual transfer functions can be estimated by sending a training sequence or by using known statistical properties of the signal. This is favorably done jointly with the estimation of the transmission channel. E.g. in space-time coded systems [9] no explicit geometrical knowledge of the antenna arrangement is required. 102 Calibration

1 10 −10 dB ] ° −15 dB

−20 dB

−25 dB phase error [ 0 10 −30 dB

0 0.5 1 1.5 2 2.5 3 amplitude error [dB]

Fig. 5.5 Tolerable phase and amplitude error to reach a certain signal cancel- lation (solid lines) and asymptotes if only phase or amplitude error is present (dashed lines).

90

80

70

] 60 ° 1dB 50 0.5dB 40

phase error [ 30 −0.5dB

20 −1dB

10

0 −4 −2 0 2 4 6 amplitude error [dB]

Fig. 5.6 Amplitude and phase error curves for a constant sum-signal error. Problem formulation 103

For a transmit antenna this is not possible, unless feedback channel is introduced, or the information gained during the receiving mode is reused. The latter requires a reciprocity of the signal paths for receiving and trans- mitting. Due to the unilateral amplification this not the case in a practical system. If the transmitter and receiver are calibrated with respect to a certain phase center, a virtual reciprocity can be restored; the array il- lumination can be made identical for both signal directions. A feedback path, which would decrease the system capacity, is avoided this way. 5.1.4 Pattern error For a given array illumination vector, an error measure can be defined on the base of the array excitation vectors. It is assumed that gi is the ideal vector that yields the intended pattern and ˆg is the real vector on the array. If it is demanded that only the shapes of the two pattern are identical, a complex scaling factor c can be tolerated, which describes a constant gain and phase difference between the two pattern. The pattern error is defined as the error that resides after the pattern are scaled to match: gi − c · ˆg e = min 2 , (5.6) c i w kg k2 w w w where k . . . k2 denotes the quadratic vector norm. The optimal factor copt can be found via the complex derivative:

∂ ∗ gi − c · ˆg gi − c · ˆg =0! , (5.7) ∂c yielding h

i ˆg∗gi copt = . (5.8) kˆgk2 5.1.5 Statistical array error In an array with several antenna elements, each element contributes to a collective error. The resulting pattern variation depends on the relation- ship between these errors as well as on the actual beam-forming operation. To obtain a general benchmark, a statistical description is useful. In an N-element array, the array pattern is given by a sum of complex weights wn = an exp(jφn). The occurring amplitude and phase deviations ∆an and ∆φn can be incorporated into distorted antenna weights

w˜n(t)= wn · [1 + ∆an(t)] · exp(j∆φn(t)) . (5.9) If its phase and amplitude is small, e.g. after a calibration with limited precision, the error of a single element can be expressed as a complex value en: 104 Calibration

20*log10(N)

10*log10(Ne2)

0 gain [dB]

angle

Fig. 5.7 Level definitions. Main-lobe and notch levels refer to the power of a single element.

w˜n(t)= wn + [en,i(t)+ j · en,q(t)] = wn + en(t). (5.10)

For a uniformly excited array with |wn| = 1, it can be assumed that the gain Gbeam of the main beam is unaffected by small errors, while the 2 statistically achievable null depth Gnotch is limited by the error power |e| averaged over all elements:

Gbeam[dB]=20 · log10(N), (5.11)

2 Gnotch[dB]=10 · log10(N · |e| ), (5.12) Both levels refer to the power of a single element (0 dB) as shown in Fig. 5.7. The zero depth relative to the main beam is also referred to as the “residual side-lobe level” σ2 [68], given as: 1 σ2 = |e|2 (5.13) N The residual side-lobes decrease with a larger number of elements. 5.2 Existing calibration methods 5.2.1 Passive array calibration For a rigorous calibration of densely spaced arrays, mutual coupling effects cannot be ignored for most antenna types. Under the assumption of weak coupling (see chapter 4.3) the ideal element excitations wn and the effective electrical valuesw ˜n at the antenna ports can be related by a coupling matrix C: w˜n = Cwn. (5.14) In this case the calibration task is reduced to the determination of this N × N matrix for an N-element array. The typical diagonal structure allows to calculate the inverse C−1 and to enforce the ideal array behavior. Existing calibration methods 105

Two principal ways are possible to determine this coupling matrix [67]: either from the measured scattering parameters S of the passive array C = I + S, (5.15) or by appropriate transformation of the measured far-field element pat- terns. 5.2.2 Coupling estimation from far-field measurements If a single element m of a well-behaved uniform linear array is excited, the resulting pattern gm is given as

N gm(u)= gisol(u) cmn exp(jnkdu) , (5.16) n=1 X where gisol is the isolated element pattern, k the wave-number, d the ele- ment spacing and u = sin φ. cmn are coupling coefficients, which describe the parasitic excitation of neighbor elements. For a matched passive array, these coefficients are identical with the matrix entries of C in Eqn. 5.14. As Eqn. 5.16 represents a Fourier-relationship, these coefficients cmn can be calculated, if all patterns gm(u) are measured [67]:

π 1 kd gm(u) cmn = exp(−jnkdu) du (5.17) 2π − π gisol(u) Z kd

This approach has two restrictions: first, gisol must be free from diagram zeros to prevent the fraction gm(u)/gisol(u) from being undefined. This is typically fulfilled for antenna elements smaller than or equal to λ0/2. Second, the element spacing d must be smaller than λ0/2. Otherwise the integration interval leads to |u| > 1, which is not defined. Contrary to the impedance-based method according to Eqn. 5.15, this method requires a well-behaved array to yield the correct results. If this prerequisite is fulfilled, this method has an important advantage: it can also be applied to active arrays, where the antenna ports are no longer accessible for measurements. In this case phase and amplitude differences introduced by the amplifiers are incorporated into the coupling matrix C. The different transfer functions are represented by the diagonal elements cii. This method does not consider additional scattering or diffraction. To separate the mutual coupling from these finite substrate effects, the beamspace technique is proposed [69]. It applies the reversely used Woodward- Lawson synthesis technique [70], which originally was intended to synthe- size array patterns. The measured pattern of element n is sampled at M 106 Calibration

gain

u

Fig. 5.8 Beam synthesis after Woodward-Lawson [70]. The array pattern (solid line) is sampled at certain locations (circles) and approximated by orthogonal beams (dashed lines). equidistant points um, as demonstrated in Fig. 5.8, and approximated by a weighted sum of M orthogonal beams

M πM sin 2 (u − um) gn(u) ≈ amn , (5.18) πM (u − u ) m=1 2 m
X where amn are the appropriate normalized [72] sampling values

gn(um) amn = . (5.19) gisol(um)

Each beam equals a uniform aperture illumination lm with a linear phase which depends on its pointing direction um: x lm(x) = rect( ) · exp(−jkumx) (5.20) Mλ0 If the superposition of these illuminations is calculated at the antenna positions, the resulting excitation coefficients are

M cnm = amn exp(−jkumx) (5.21) m=1 X Conceptual, each non-ideal element pattern is synthesized by an appro- priate array illumination. If M = N, this method mathematically equals the discrete represen- tation of Fourier integral in Eqn. 5.17. If the synthesized array now is increased by additional elements M >N, more and smaller orthogonal Existing calibration methods 107 beams lead to a better approximation of the measured pattern. The re- sulting virtual array contains elements beyond the physical array, which account for diffraction effects. The undisturbed coupling matrix is gained by taking the subset of antennas which coincide with the real antenna positions. Both methods, the Fourier and beamspace technique, require measure- ments over a predetermined angular range. The data must be available in a quasi-continuous form for the Fourier-transform technique, or sam- pled at equidistant points in the u-space for the use of the beamspace technique. In some situations it can be advantageous to use arbitrary measurement points for the calibration. In complex networks, these can be in-situ calibrations, where the signals of other cooperating base-stations at well known positions are used. Arbitrary sampling points might also be useful in anechoic chamber measurements, particularly if the measurements inside the beam width are more precise than outside. This can be due to a limited dynamic range of the measurement system or due to scattering and diffraction, which have more influence at angles with lower antenna gain. In these cases, confining the measurements used for calibration can help to improve the accuracy within the practically important angular range. Thereby it has to be assured, that the irregular sampling points are sufficient to describe the diagram over the wanted range [71]. In [72] it is proposed to select M ≥ N measured array vectors

g1(um) g2(um) g(um)=  .  , (5.22) .    g (u )   n m    and the associated ideal vectors

exp(−jkumx1) exp(−jkumx2) i g (um)= gisol(um)  .  , (5.23) .    exp(−jku x )   m n    both sampled at the same directions um. The xn are the element positions. If the measured data is only distorted by mutual coupling, Eqn. 5.14 holds i for every pair of g(um) and g (um). This leads to the following matrix equation: 108 Calibration

i i g(u1) . . . g(um) = C · g (u1) . . . g (um) , (5.24) or     G = C · Gi, (5.25) respectively. For M = N the coupling matrix C can be calculated by simple matrix inversion: −1 C = G Gi (5.26) For M >N the same approach leads to
an overdetermined set of equa- tions. In [72], Eqn. 5.25 is rearranged to

T −1 T T iT G C = G ˜c1 ˜c2 . . . ˜cn = G , (5.27) with (. . .)T being the
transposed matrix. This can be interpreted as N least squares problems

T i min G ˜cn − gn 2 n =1, 2,...,N, (5.28) ˜cn which can be solved numerically by QR-decomposition of GT (see [72] for details). The approach of numerical optimization of the coupling matrix can be extended by additional virtual elements that account for possible edge diffraction [73]. As a drawback, the matrix C then becomes rectan- gular and cannot be inverted and direct correction of the received signals becomes impossible. This constrains the range of possible applications of this method. Last, the concept of the active element pattern [65] needs to be men- tioned: as long as there is a single-mode relationship between each elec- trical port state and the corresponding far-field pattern, all embedded element patterns can be stored numerically. This allows for arbitrary pat- terns of all elements, but the array factor model is not valid anymore. Hence, associated algorithms are computationally expensive. This can be partly mitigated by employing this method just for the outer array ele- ments. The inner elements, which exhibit almost identical patterns, are treated with the conventional array factor model [66]. Tab. 5.1 summarizes the advantages and drawbacks of the reviewed calibration methods for the passive array.

5.2.3 Test-tone calibration In most practical applications it is not possible to obtain reliable far-field data. The transfer function variations caused by the entire analog RF Existing calibration methods 109 Table 5.1 Overview over methods for the calibration of passive antenna arrays.

method advantages drawbacks from S-matrix [67] no anechoic chamber only passive array no difraction deembedding difficult Fourier technique active gain considered only well-behaved arrays [67] no diffraction limited to d>λ0/2 beamspace [69] removes diffraction diffraction not compensated fewer samples only equidistant spacing LS fitting [72] arbitrary points no diffraction 2-D arrays extended matrix [73] arbritrary points C not invertible active element [65] individual elements no C no pattern constraints computational expensive hardware, which often exhibit several dB and tenths of degrees of phase shift, prevail the uncertainties resulting from mutual coupling. As a commonly used pragmatic approach, a local calibration signal is employed. This signal is distributed over a symmetrical network to ensure that all fractions maintain identical phases and amplitudes. Directional couplers or switches are applied to inject the calibration tones at the input of the downconverter chain. The resulting response to this set of identical input signals yields the appropriate values for compensation [74]. This straightforward procedure for the receive case poses difficulties in transmitting arrays. To avoid a superposition of all output signals it has to be assured that only one signal reaches the test receiver. This usually is done by activating only one receiver at a time or by switching the calibration network [75]. The basic principles are identical and, therefore, in the following it is not explicitly distinguished between transmitter and receiver calibration. Several methods are published that use the transmitted signals for joint transmitter and receiver calibration to avoid additional hardware. Switching schemes are proposed for time-division duplex systems [76] and enhanced by frequency converters to calibrate frequency-division duplex systems [77, 78]. All these methods require a multitude of switches and interconnecting transmission lines, whose phase and gain equalities are crucial for the obtainable precision. An interesting method proposes to exploit the mutual coupling effect itself for the calibration of both, the mutual coupling and the active hard- ware [79]. As shown in Fig. 5.9, the calibration signal is generated by 110 Calibration

mutual coupling

cal signal element n element n+1

Fig. 5.9 Calibration using the mutual coupling [79]. The reciprocity of the mu- tual coupling is used to determine the transmitter and receiver transfer functions and the mutual coupling itself. transmitter n. It is received by receiver n belonging to the same antenna element. But, also, it is reaches the neighbored receiver n + 1 via mutual coupling. The mutual coupling can be assumed to be reciprocal, giving the same phase shift and attenuation to a signal travelling from element n + 1 to element n. The gained information from these two situations can be used to calculate all correction values, including the mutual coupling factor between these two elements. It has to be considered that the result- ing coupling factor does not necessarily refer to the antenna phase center. Therefore, this method might not lead to the correct calibration matrix for certain types of antennas or higher frequencies. In these cases it might be useful to introduce an appropriate deembedding distance, as needed for the s-parameter method (compare 4.3.3). One severe drawback of this method is the requirement of simultaneous transmit and receive operation. It has to be considered that the power levels at the transmitter and receiver differ significantly. To ensure linear operation of the receiver and, hence, reliable values for calibration, a high isolation must be guaranteed between these components. This might be impossible to fulfill for compact and lightweight arrays. All methods that use any kind of signal circulation of the own trans- mitted signal are referred to as ”self calibration”methods.

5.2.4 Hybrid methods It is a reasonable assumption that, once an array is fabricated, the mu- tual coupling stays nearly constant, while the dominant changes result from the active hardware. This motivates to combine a fixed compensa- tion of mutual coupling with a repeatedly determined correction of the active hardware, which is gained by signal injection during normal oper- ation. The coupling matrix is either determined by field measurements after fabrication [80], or by electromagnetic simulation of the employed array [81]. Existing calibration methods 111

phase center C=1+S CAL CAL R

(a) (b)

Fig. 5.10 Matrix representation of calibration: a) mutual coupling and unequal transfer functions disturb the beam-forming process. b) these can be modelled as a coupling matrix C and a diagonal gain matrix R.

5.2.5 Improved Hybrid Calibration The hybrid methods proposed in literature assume a calibration model as depicted in Fig. 5.10. A coupling matrix C and a gain matrix R associated with the active components are assumed, which are independent of each other. Furthermore, it is demanded that C is determined by far-field mea- surements and R can be precisely gained from the local calibration signal. It is not considered that the reference network has a limited precision and that possible mismatch might occur at the antenna and amplifier inter- faces, which lead to reflections. If a very precise calibration is needed, the hybrid calibration approach has to be extended to account for this problem. All methods which base on the measured far-field automatically include the complete transfer matrix RC, even including mismatch effects. This can be exploited for a more precise calibration: if the disturbed active- gain matrix R˜ is measured simultaneously or in short succession with the pattern measurements, it is possible to estimate the changes of RC. If t is the actual time and tcal is the time of far-field calibration, the diagonal entries of R are r˜m(t) rm(t) ≈ rm(tcal) · , (5.29) r˜m(tcal) RC accordingly is

r˜1(t)/r˜1(tcal) 0 . R(t) · C(t) ≈  ..  · R(tcal) · C(tcal) 0r ˜ (t)/r˜ (t )  n n cal    (5.30) This procedure removes the deficiencies of the test-tone calibration net- work. However, it is still assumed that any mismatch at the antenna interface stays constant. Hence, the remaining uncertainties result from 112 Calibration

all equal al ar c1 ck cN−1

bl br t1r 1 tk rk tk+1 r k+1 tNN r

a1b 1 a k b k a k+1 b k+1 a N b N

Fig. 5.11 Underlying signal model for transmission line calibration. a change of this matching and from the limited precision of the far-field measurements.

5.3 Transmission-line calibration method As discussed above, the test-tone scheme requires a calibration signal that is distributed via a symmetric power splitting network and injected into each receiver path using either switches or directional coupler. The needed splitting network consumes a lot of space and poses a design challenge for two-dimensional, non-equally spaced or non-planar arrays. Self-calibrating methods usually require a complex switching network [76–78] and demand very high internal isolation to permit simultaneous transmit and receive operation. The proposed method uses a single transmission line that links all branches, whereby the estimation of the line segments themselves is part of the calibration. Therefore the hardware effort is noticeably reduced.

5.3.1 Description of method For the derivation of the proposed method the signal model depicted in Fig. 5.11 is assumed: A ”bus-like”transmission line links all N transceiver branches, where the N-1 line segments are represented as complex transmission factors ck. The individual phase and amplitude responses of the transmitters and receivers are modelled by the transmission constants tk and rk, whereby the whole chain with frequency conversion to baseband is included. ak and bk are the complex transmitted and received signals. The transmitted and received signals of each transceiver are assumed to be coupled with the calibration line at a single point. The calibration refers to this point and therefore the electrical lengths from here to the antennas have to be equal. For this derivation it is assumed that • the coupling is weak, no noteworthy signal travels from antenna k to adjacent receivers (e.g. loss >15 dB).

• no reflections of al, ar occur at the coupling points. Transmission-line calibration method 113

ck

a l tk rk tk+1r k+1

akbk a k+1 bk+1

ck

ar tk rk tk+1 rk+1

ak bk ak+1 bk+1

Fig. 5.12 Signal flow for receiver calibration.

The influence of reflections and a possible practical realization will be discussed later in sections 5.3.2 and 5.3.3. Now the calibration signals al and ar, respectively, are successively applied at both ends of the calibration line like shown in Fig. 5.12. The transfer function to each output k is

k−1 b s = k = r c (5.31) kl a k m l m=1 Y and N−1 bk skr = = rk cm. (5.32) ar mY=k The transmission factor of each line segment is reciprocal and can be calculated by combining the equations 5.31 and 5.32 for k and k+1 and solving for ck: s(k+1)l skr ck = ± · . (5.33) s skl s(k+1)r The sign of the above square root cannot be determined from the set of equations, but it can be estimated if the approximate electrical length is known: f c = exp −j2π l (5.34) est c k,est  0  If lest is not known either, it is possible to estimate it once from two closely cπ spaced (∆f < 2l ) frequencies: 114 Calibration c (f) c l = 6 k · 0 (5.35) est c (f + ∆f) 2π∆f  k  Once the values for ck are known, the relationship between two adjacent receivers is given by

rk+1 1 s(k+1)l s(k+1)r = · = ck (5.36) rk ck skl skr and thus all rk can be deduced with respect to one element the calibration refers to. The calculation of tk follows the same scheme, except that transmitters and receivers are exchanged. It has to be noted that no assumptions were made on al and ar, there- fore they can be generated in an arbitrary manner, e.g. by switching a calibration source to both inputs using lines of individual length. 5.3.2 Estimation of systematic error To ascertain the hardware requirements, it is useful to estimate the sys- tematic errors introduced by the proposed calibration method. These errors can be subdivided into two contributions: • Variations in hardware, such as variations of the realized coupling be- tween calibration line and signal or variations of the electrical length between calibration point and antenna lead to errors in each individ- ual branch. These depend strongly on the particular implementation and can be estimated separately for each element. • The proposed calibration method bases on an idealized signal model that is difficult to realize at microwave frequencies. Notably waves are assumed that travel along the calibration line without reflections. Any practical realization will reflect a fraction of the signal and cause a wave travelling backwards, disturbing the calibration values. To gain some insight into the ability of this method the error caused by mismatches on the calibration line can be estimated. Again the receiving case is regarded, while the transmitting case can be treated in the same way and leads to the same results. The basic ”calibration element”(see Fig. 5.13) can be approximated by the following S-parameter matrix

s11 1 a S = 1 s22 a , (5.37)  a a 0  where ’a’ denotes the large coupling attenuation towards the receiver. For a possible practical realization it is again referred to section 5.3.3. To sim- plify the calculation two reflection factors Γkl and Γkr are defined, that Transmission-line calibration method 115

1 2 S 3 Γ Γ k,l k,r rk

Fig. 5.13 Basic calibration element (receive mode) Γ S11 (k+1),r

ck

a l t t r k rk k+1 k+1

akbk a k+1 bk+1

Fig. 5.14 Simplified error model comprehend all superimposed reflection effects looking into the line struc- ture from receiver k to the left or right, respectively (compare Fig. 5.13). Fig. 5.14 shows the used error model. Assuming that multiple reflec- tions between element k and k+1 can be neglected, the disturbed obser- vations s can be obtained for one pair of antennas:

e s(k+1)l = s(k+1),l 1+Γ(k+1)r (5.38)

2 2
skle = sk,l 1+ cks11 + ckΓ(k+1)r (5.39) skr and s(k+1)r are calculated accordingly.
Equations (5.33)e and (5.36) can be combined to obtain the transmis- sione ratioe of two neighbored elements r s s k+1 = ± (k+1)l (k+1)r . (5.40) r s s k r kl kr Inserting the disturbed observations s leads to the disturbed ratio

2 2 rk+1 rk+1 (1+Γ(k+1)r) · (1e + cks22 + ckΓkl) = ± 2 2 . (5.41) rk rk s(1 + cks11 + ckΓ(k+1)r) · (1+Γkl) e If all doublee reflections s11s22 ≈ 0 and Γmsnn ≈ 0 are neglected, this can be simplified to 116 Calibration

2 2 rk+1 rk+1 1+ cks22 + ckΓkl +Γ(k+1)r ≈± 2 2 . (5.42) rk rk s1+ cks11 + ckΓ(k+1)r +Γkr e  e Under the assumption that| all reflections{z are small, Γ } 1, the result- ing error  can be linearized to ∂ ∂  ≈ Γkl ·  +Γ(k+1)r · . (5.43) ∂Γkl ∂Γ(k+1)r

Assuming thate the ”calibration element”is symmetric and s11 = s22 this results in the following simplified expression:

r r 1 k+1 ≈± k+1 1+ c2 − 1 Γ − Γ . (5.44) r r 2 k kl (k+1)r k k   Accordinge to this expression, the systematic
calibration
error depends on the electricale length of the calibration line segments and the difference between the reflection factors Γkl and Γ(k+1)r seen from the considered receiver pair into the left and right transmission line structure. With periodic spacing one has to keep in mind that for certain ck all small reflections might add up to a strong reflection Γ, so that this Bragg condition should be avoided. For illustration an ideal six-element periodic structure with a single element return loss of −20 dB is simulated. The reflection coefficients Γ2l and Γ3r are shown in Fig. 5.15. For 5 GHz, where the segment length lk is λ/2, the return loss of the structure gets worse than −10 dB as an effect of the periodic arrangement. For the same structure the systematic calibration error r3/r2 of this r3/r2 pair of elements is calculated using both, the complete algorithm and the approximated expression in Eqn. (5.44). The results are come epared in Fig. 5.16. Transmission-line calibration method 117

0

−5

−10

−15

return loss [dB] −20

−25

−30 0 2 4 6 8 10 freq [GHz]

Fig. 5.15 Reflection coefficients Γ2l (dashed line) and Γ3r (solid line) of six element periodic structure with s11 = −20dB and lk = λ/2 at 5 GHz

1

0.8 RL=−20dB

0.6

0.4

0.2

0

−0.2

−0.4 calibration error [dB]

−0.6

−0.8 RL=−25dB −1 0 2 4 6 8 10 freq [GHz]

Fig. 5.16 Estimated (solid lines) and simulated (dashed lines) systematic error for return losses −20 dB and −25 dB 118 Calibration

For return losses s11=s22=−20dB and −25 dB the simulated and ap- proximated errors agree well, except a range around 5 GHz where the assumption of a small reflection Γ3r  1 is violated due to the fulfilled Bragg condition. The maximum error stays below 0.9 dB for s11 = −20dB and below 0.5 dB for s11 = −25dB. Return losses of such magnitudes can be achieved with a transmission line structure. It has to be noted that in the analysis above it never was assumed that |ck| = 1 for the transmission line segments. Introducing attenuation here might help to improve the matching and weaken far end reflections. 5.3.3 GaAs transmit/receive-switch with calibration ability The hardware necessary for this calibration method was integrated to- gether with a transmit/receive-switch for an active antenna array system. The simplified schematic is shown in Fig. 5.17: the needed coupling is achieved by a capacitor with C=150 fF. This value was set to obtain a transfer function |sRX,cal| of −15 dB at 6 GHz which leads to an antenna- to-antenna isolation of −30 dB. This capacitor and the bond-pad capacitances form a low-pass trans- mission line structure together with the inductances of the bond-wires and on-chip lines. T-structure FET switches are used to switch between trans- mit and receive mode, so that the open switch presents a high impedance at the combination point. As the calibration is done with respect to this

antenna

parasitic inductance

CAL CAL

¡ ¡

¡ ¡

¢

£ ¢ £ bond coupling wires

capacitor

¨ ¨

©

¨ ¨

©

¤ ¤

¥

¤ ¤

¥

¦ ¦

§

¦ ¦ §

Tx−switch Rx−switch

monolithic Tx Rx circuit

Fig. 5.17 Schematic of Tx/Rx-switch with calibration capability Transmission-line calibration method 119 point the switches are designed according to their individual demand for transmit/ receive operation. The circuit was fabricated using a 0.6 µm GaAs MESFET process.

5.3.4 Experimental Results To verify the proposed concept a test setup was built, employing four of the transmit/receive-switches described above. Fig. 5.18 shows a picture of the circuit. The calibration ports are linked by a single transmission line. The lines connecting the antenna ports are kept at equal lengths, while the lines connecting transmitter and receiver have individual distances. The necessary transfer functions from the left and right end of the cali- bration line to each transmitter/receiver were measured using a network analyzer. The individual transfer functions to the different antennas were also determined for comparison. Fig. 5.19 shows the estimated phase 6 ck of the three calibration line elements over a frequency range up to 10 GHz. The deviation from the ideal progression at around 5 GHz is due to the mounting pad of the chip. Measurements of the empty structure without chip show a parasitic coupling path at this frequency. For operation at this frequency one would need to choose a different layout.

antenna

        

        

        

        

        

        

ports

        

        

                

                

                

             Tx/Rx−

                

           

        ! ! ! !

                

            

                

           

        ! ! ! !

                

            

                

           

        ! ! ! !

switches

                

            

                

           

        ! ! ! !

                

            

                

           

        ! ! ! !

 

 

                

            

                

          

calibration 

        ! ! ! !

 

 

                

            

                

           

        ! ! ! !

   line 

transmit

        

        

                 

                 

           

           

        

        

                 

                 

           

           

        

       

ports

                 

                 

           

           

        

        

                 

                 

           

           

        

        

" " "

# # #

                 

                 

           

           

" " "

# # #

                  $ $ $ $ $ $ $ $

                  % % % % % % % %

           

           

" " "

# # #

                  $ $ $ $ $ $ $ $

                  % % % % % % % %

           

           

" " "

# # #

                  $ $ $ $ $ $ $ $

                  % % % % % % % %

           

           

" " "

# # #

                  $ $ $ $ $ $ $ $

                  % % % % % % % %

           

           

" " "

# # #

                  $ $ $ $ $ $ $ $

                  % % % % % % % %

           

           

" " "

# # #

                  $ $ $ $ $ $ $ $

( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( (                   % % % % % % % %

) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) )

" " "

# # #

$ $ $ $ $ $ $ $

( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( % % % % % % % %

) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) )

" " "

# # #

$ $ $ $ $ $ $ $

( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( % % % % % % % %

) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) )

" " "

# # #

$ $ $ $ $ $ $ $

( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( % % % % % % % %

) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) ) )

& & & & & & & & & & & & &

' ' ' ' ' ' ' ' ' ' ' '

' receive

" " "

# # # ports

Fig. 5.18 Test substrate (size ∼50×50 mm2) 120 Calibration

150

100

50

0

phase [deg] −50

−100

−150

0 2 4 6 8 10 freq [GHz]

Fig. 5.19 Estimated transmission phase of calibration line segments

From the reference measurements (antenna to receive ports) the re- sulting array factor for a unity distribution is calculated and depicted in Fig. 5.20. The main beam, side lobes and zero positions are significantly shifted compared to the theoretical pattern. The calibrated outputs are obtained by simply multiplying the inverse of the estimate:

−1 bk,cal = rk · bk (5.45) The resulting pattern is also depicted in Fig. 5.20. Except slightly changed side-lobes and finite zero depths it resembles the ideal pattern. For com- parison, also the residual side-lobe level is plotted to demonstrate the statistically expected zero depth. To quantitatively characterize the array calibration, the residual side- lobe level and the pattern zeros at 0◦ and 60◦ are calculated over a broad frequency range. Fig. 5.21 shows the achieved results after calibration. Apart from f=5 GHz, where unwanted coupling occurs, the notch stays well below −20 dB up to 7 GHz. This demonstrates that the proposed method can replace the space- consuming symmetrical divider network, usually used for array calibration. The linking transmission line is part of the calibration itself and, there- fore, does not follow any layout constraints. In applications where the obtained precision is sufficient, it can be used to calibrate antenna arrays of arbitrary arrangement. The method is inherently broadband and easily compatible with monolithic front ends to form compact active arrays. Transmission-line calibration method 121

15

10

5

0

−5

−10

−15

array factor [dB] −20

−25

−30

−35 0 45 90 135 180 225 270 315 360 angle [deg.]

Fig. 5.20 Uncalibrated (line with crosses) and calibrated array factor (solid line) (receiving mode) at 4 GHz compared to the ideal value (dashed line). The circles indicate the residual side-lobe level calculated from the measurements.

−10

−15

−20

−25

−30

−35 zero depth [dB]

−40

−45

−50 0 2 4 6 8 10 freq [GHz]

◦ Fig. 5.21 Zero depth after calibration (receiving mode) for the null at 60 (solid ◦ line) and at 0 (dashed line) and residual side-lobe level (circles). 122 Calibration 5.4 Dynamic transmitter calibration So far, it was always assumed that the hardware introduces slowly varying random changes. It is demanded that the individual transfer functions stay constant during a typical data transmission. This neglects switching of the power amplifiers in time-domain access schemes. As already pointed out in chapter 3.5, switching the power transistors leads to strong thermal changes in the semiconductor devices, exceed- ing those caused by fluctuations of the ambient temperature by orders of magnitude. In contrast to environmental influences, these variations are of deterministic nature and can be expected to be partly correlated. One way to go around this question is to bias the power amplifiers con- stantly, but in practice a shutdown during periods of inactivity is very at- tractive to save battery power and fulfill the strict emission limits. There- fore, the correlation of switched power amplifiers with respect to their application in antenna arrays is investigated in the following. This allows to derive guidelines how and when such a system needs to be calibrated.

5.4.1 Instantaneous error

For the time-variant transmitting case, the transmitted vector wT X is

−1 wT X (t)= C(t) · ˆw = C · Cˆ w(t) , (5.46)   where w is the ideal vector and ˆw and Cˆ are the preconditioned vector and the estimated correction matrix, respectively. In the following the antenna uncertainties are excluded and it is assumed, that a coupling matrix estimation

Cˆ = C(t0) (5.47) can be found, which perfectly compensates the coupling at time t0. The instantaneous array error at all other times t 6= t0 then is

e(t) = min c · wT X (t) − w(t) , (5.48) c w w or w w −1 e(t) = min c · C(t)Cˆ (t0) − 1 w(t) . (5.49) c w  w This error measure, previouslyw discussed in section 5.1.4,wis used to include w w only those errors that lead to a deformation of the wanted antenna pattern. In the case of jointly switched power amplifiers it is expected that the gain and phase changes are partly correlated. These common changes are accounted for with the complex scalar factor c. Dynamic transmitter calibration 123

PA attenuator down− 26dB converter

RF and 5.2GHz signal storage multi−channel receiver

PA LO switch 6.65 GHz

Fig. 5.22 Experimental setup to determine gain and phase correlation of si- multaneously switched power amplifiers.

For universality a uniform distribution with |wn| = 1 is chosen and the average error power 1 |e |2 = ee∗ (5.50) norm N referred to a single element is used in the following. 5.4.2 Power amplifier To conduct the transmit calibration experiment, four of the GaAs power amplifiers presented in chapter 3.5 are used. As described there, these amplifiers deliver 19 dBm output power at the 1 dB-compression point. This is a comparably low value and takes advantage of the inherent power splitting and additional gain that occur in array antennas. It is recalled that two thermal time-constants were found, oneof 1.6 ms and a second of 2200 ms. It was observed that the gain decreases with raising temperature, which is a typical behavior. 5.4.3 Array calibration To evaluate the impact of the thermal changes on the system behavior, it is necessary to drive all amplifiers simultaneously and regard the changes and their correlation behavior during a typical burst. For this purpose the setup shown in Fig. 5.22 is used: The four power amplifiers are driven by the same frequency source and switched simultaneously. A smart antenna test-bed, consisting of MMIC downconverters (see section 3.3.3) and a multichannel receiver, is used to record the output. Attenuators reduce the effect of possible mismatch and, at the same time, ensure a linear operation of the receiver. 124 Calibration

reference pulse on T burst 2T burst

7T burst off

time

Fig. 5.23 Used switching scheme: periodic frame structure with variable num- ber of bursts. The calibration is acquired at the first measured frame and kept constant afterwards.

Fig. 5.23 illustrates the used switching scheme: a frame structure with seven time slots is assumed, each time slot 250 µs long. The frame pattern is repeated periodically. According to the HiperLAN standard, the actual transmit pulse can be a multiple of the 250 µs slot length. Here, this is emulated with a varying duty cycle from 1/7 to 7/7, occupying one or several time slots. The experiment starts with a 250 µs pulse. A period of around 120 ms is recorded, covering 71 pulses. The complex calibration values cn are cal- culated from average gain and phase of the first pulse recorded. Fig. 5.24 shows the graph of pulses no. 70 and 71 after calibration. Although the amplifiers have been turned off several times and the isolation during the off-period varies significantly, the calibrated gain agrees very well with the reference level. The observed gain variation between start and end of the pulse is less than 0.2 dB. The initial overshooting is due to the fast switching. In a real system this would cause unwanted spurious emission. A controlled ramp-up would be used to avoid this transient. Maintaining the same calibration values, now the duty cycle again is varied by increasing the pulse length. As shown in Fig. 5.25, a decreasing gain is observed for all channels, but the changes are no longer identical as in the previous case. The different variations result in an increased calibration error. How- ever, the element error power |e|2 averaged over all bursts stays less than −33 dB, which is fairly low. This corresponds to a residual side-lobe level of less than −39 dB in a four element array. Fig. 5.26 shows the progression of the instantaneous error power dur- ing one burst. For both curves, at the initial duty cycle of 1/7 and the Dynamic transmitter calibration 125

5 pulse 70 pulse 71 0

−5 +0.1 −10

−15 −0.1

rel. gain [dB] −20

−25

−30

−35 0 500 1000 1500 2000 2500 3000 3500 time [usec]

Fig. 5.24 Calibrated gain of all four channels, 120 ms after calibration, constant duty cycle 1/7. The inset shows the gain transients during the last pulse. All four channels are almost identical. increased cycle of 6/7, it can be seen that the error power stays almost con- stant for the duration of the whole pulse. This demonstrates that a fixed set of values cn is sufficient for calibration over a typical burst duration. No time-dependent calibration is needed. It can be concluded that all gain and phase changes taking place in the range of the first thermal time-constant of 1.6 ms are well correlated. The changes associated with the second time constant of 2200 ms show the same tendency, but are individually different for each amplifier. 126 Calibration

0.5 −30

0 −35

4 3 rel. gain [dB] −0.5 1 −40 2 error power [dB]

−1 −45 1/7 2/7 3/7 4/7 5/7 6/7 7/7 duty cycle [1]

Fig. 5.25 Gain variation of all four channels and calibration error for different duty cycles, calibration on duty cycle 1/7.

10

0

−10

−20 duty cycle 6/7 −30

−40

−50

error power [dB] −60

−70

−80 duty cycle 1/7

−90 0 500 1000 1500 2000 time [usec]

Fig. 5.26 Instantaneous element error during pulse 70 for duty cycles 1/7 (solid) and 6/7 (dashed), calibrated on duty cycle 1/7. Conclusions 127

on−chip

heat Rth,1 source τ 1 off−chip

Rth,2 τ 2

Fig. 5.27 Thermal model derived from the measured temperature transients. It is assumed that the second time-constant is determined by the off-chip envi- ronment and, therefore, can differ from amplifier to amplifier in an array.

One possible explanation of this behavior can be derived from the ther- mal model depicted in Fig. 5.27. It can be assumed that the small first time-constant is basically defined by the heat capacity of the power device and the thermal substrate resistance. Both can be attributed to physical properties of the chip. As monolithically integrated circuits are used, these values should be virtually identical for all amplifiers. The second time- constant, in contrast, can be expected to be given by the chip mounting, the further substrate around and the final thermal resistance to the en- vironment. All these factors depend on the position of the chip on the testing substrate and the assembling process. This will result in different thermal resistances Rth,2 and, hence, different junction temperatures and gains of the devices.

5.5 Conclusions In this chapter, various methods for array calibration are reviewed. For a complete calibration, far-field sources are necessary to include all mu- tual coupling and diffraction effects. A local test-tone calibration can be used to compensate for the variations of the active circuitry. Only the mutual coupling self-calibrating technique is suitable to determine these variations and the coupling matrix of the array. This method requires simultaneous transmit and receive operation and, therefore, is difficult to apply to compact systems due to the high isolation needed. A new calibration scheme is proposed that employs a local coupling be- tween all branches provided by a weakly coupled transmission line, which links all elements. The determination of the transmission factors of the line segments is part of the calibration process itself. This allows an arbi- 128 Calibration trary arrangement and replaces the previously needed precise and bulky divider network. To derive guidelines for necessary calibration intervals, the dynamic behavior of switched power amplifiers is studied with respect to application in transmit antenna arrays. On monolithically integrated power amplifiers two thermal time-constants are experimentally found. A a typical burst up to a duration of milliseconds can be calibrated using a single set of correction values. If the duty cycle stays constant, the obtained values can be reused for several pulses. For a changed duty cycle the variations are less correlated and a slight pattern variation occurs. 6 Active antenna arrays

In this chapter, the construction of active antenna arrays is discussed. The monolithically integrated frontend presented in chapter 3 is used to yield compact and robust arrays. For a four element linear array, the different calibration methods are evaluated to gain information about the obtainable precision. The antenna front-end is integrated into a complete multiple-antenna test-bed, including conversion to digital baseband and storage of the re- ceived signal. Amplitude and phase stability as well as the noise perfor- mance of this system are studied with respect to a practical application in a communication system. Last, it is demonstrated that the monolithically integrated RF fron- tends together with the transmission line calibration technique enable the simple fabrication of conformal arrays. 6.1 Linear array 6.1.1 Design In chapter 2.4 the architecture of a multiple-antenna test-bed is discussed. The active antenna array described here is conceived as the RF front-end to this system. The function can be explained from the block diagram depicted in Fig. 6.1: the four antenna signals are processed independently by four receiver branches. These consist of monolithically integrated low- noise amplifiers and downconverters, both documented in chapter 3. Im- age rejection is achieved by a coupled-line bandpass filter in microstrip technology. The measured pass-band attenuation of this filter is less than 1 dB, the entire image band is suppressed by more than 50 dB. At the cost of a reduced image rejection, this circuit combination could be replaced by the much smaller integrated receiver (see section 3.4), which was not available at the time of construction of this array. The LO signal is symmetrically distributed to all downconverters by cascaded 2:1 Wilkinson power-dividers. The same signal is also used for 130 Active antenna arrays

patch antennas directional coupler

LNA attenuator coupled line down− converter filter

LO

CAL IF1 IF2 IF3 IF4

Fig. 6.1 System block diagram of the active antenna front-end. Dashed lines indicate monolithically integrated components. the up-conversion of a signal used for calibration. The difficulty in this signal upconversion is the spectrally close spacing of the resulting mixing products and the LO signal. To drive the mixer, a comparably high power of around 10 dBm is needed. Even if a mixer with a good LO to RF isolation of 40 dB or better is used, the leaking signal exhibits a significant power. The proximity of the wanted and unwanted components makes filtering difficult. In this case the possible RF frequencies reach up to 5.875 GHz and the LO range starts at 6.6 GHz, resulting in a sharp transition from pass-band to stop-band. To mitigate this problem, a high power of the calibration signal of up to 0 dBm is chosen, which is the highest possible level for linear operation of the mixer. This guarantees the maximal output power level compared to the LO spurious. Then, an attenuator is used to attenuate both, the signal of interest and the unwanted spurious. The image frequency is removed by six-section coupled-line filter, identical to those used in the receiver branches. After the signal is symmetrically divided, it is coupled to the receiver branches using directional couplers. Considering the conversion loss of 8 dB, the 15 dB attenuation, the signal distribution (≥ 6 dB), and the coupling attenuation of 15 dB, the power of the calibration signal at the input of each branch is significantly below the input compression point. This is important to avoid calibration errors due to gain and phase changes caused by the compression behavior of the receivers. The complete circuit is printed on a Duroid 6010 microstrip substrate Linear array 131 slot LNA down- patch converter DC- supply IF4 network calibration IF3 IF2 IF1 support mechanical LO MIX CAL filter CAL image mixer attenuator

Fig. 6.2 Layout of the four element linear antenna array seen from the circuit side. Substrate size is 150 ×117 mm.

(r = 10.2, h = 635 µm) with the aperture-coupled patch antennas (com- pare chapter 4) mounted on the backside. Fig. 6.2 shows the final layout. It was taken care that all signal lines are well separated in order to avoid unintended cross-talk. The calibration network, whose precision is important, avoids any crossings with other signal lines. The number of different signals, however, requires certain crossings of transmission lines. In these cases the line carrying the signal with the lower frequency is interrupted and bridged with a bondwire. This keeps all transfer functions as close to each other as possible. 132 Active antenna arrays

Fig. 6.3 Picture of the four element linear array. On top the four antenna substrates with the slanted patch antennas are visible.

The antennas are equally spaced with a distance of 27.3 mm from center to center. This corresponds to half of the free-space wavelength ◦ λ0 at the center frequency of 5.5GHz. A 45 -slanted linear polarization is chosen for this prototype. Each antenna employs two coupling slots for both perpendicular polarizations, out of which only one is used. The second antenna port is terminated with a 50 Ω resistor. The large substrate, which measures 150 × 117 mm, is mechanically supported at several points to ensure mechanical stability. As the most critical part, the fixture of the connectors was reinforced. The mechanical robustness is of great importance for the electrical performance; particu- larly it removes an important source for amplitude and phase variations. Fig. 6.3 shows a picture of the manufactured array.

6.1.2 Experimental results After manufacturing, the active antenna array was characterized inside an anechoic chamber. Thereby, it had to be considered that this array includes a frequency conversion process. A reference mixer was used to calibrate the measurement system; the yielded results then were corrected by the previously measured conversion loss of this mixer. Fig. 6.4 shows the total absolute gain of all four measured channels. This includes the conversion gain of the receiver, but also the gain of the passive antenna structure. From the measurement of a single antenna alone, a passive Linear array 133

35

33

31

29 total gain [dBi]

27 ch1 ch2 ch3 ch4 25 5 5.2 5.4 5.6 5.8 6 freq [GHz]

Fig. 6.4 Measured gain of all four individual channels. This total gain includes the passive antenna gain compared to an isotropic radiator and the additional active conversion gain. antenna gain of around 5.5 dB can be assumed. It is found that all four channels show similar gains. The total gain is around 30 dB, corresponding to approximately 25 dB active conversion gain. Subsequently, the array was rotated to measure the directional pat- terns. The definition of angles is illustrated in Fig. 6.5. A change of ele- vation corresponds to the angle θ, the main-beam of the elements points at θ = 90◦. φ describes the azimuthal rotation of the array. This is the plane in which the beam can be steered by changing the element weights. φ = 0 is the main-beam direction of the antenna elements. Fig. 6.6 shows the directional pattern for different elevations 0◦ ≤ θ ≤ 180◦. All element patterns are similar. For comparison, the pattern of a single element is simulated and depicted. At 5.5 GHz an active gain of 27 dB is assumed. For a certain range around the main beam θ = 90◦ a good agreement is found. For radiation almost parallel to the substrate surface the patterns differ. This can be explained by the effect of the finite substrate, which is not included in the simulation. Also cables and connectors are present that affect the pattern measurements. The patterns measured for azimuthal rotation is significantly different: as seen in Fig. 6.7, all patterns exhibit a strong superimposed gain ripple, which can not be found on the simulated curve of the single element. 134 Active antenna arrays

z θ y

ϕ τ x

Fig. 6.5 Definition of axes and angles: the main direction of radiation is along ◦ ◦ the x-axis at (θ = 90 ,φ = 0 ). The angle θ between main direction and the z-axis is also referred to as elevation. φ is the azimuthal rotation around the z-axis. The array can be steered in the φ-plane.

Furthermore, the limited size of the antenna elements does not explain gain changes occurring with this high spatial frequency. Thus, it is suspected that these pattern deformations are a consequence of mutual coupling among the individual antennas. This coupling and the possible calibration are further discussed in the following section 6.1.3. The linearity of the antenna front-end is characterized outside of an anechoic environment. Using a transmit antenna at a very close distance to the array, high input power levels can be easily reached and possible multi-path propagations can be neglected. This allows both, to perform a two-tone test and to drive the receiver into compression. The attenuation of the radio link is not known, therefore, the compression point and IP3 are determined as output referred values. From the approximate conver- sion gain of the active circuit the intercept and compression points at the antenna/ amplifier interface can be estimated. For all channels compres- sion points of approximately −25 dBm and third-order intercept points of −16 dBm are found this way. The noise figures can not be measured directly by ENR-measurements, they have to be calculated from the known gain and the absolute noise power at the output. This method contains several uncertainties, such as the estimated active gain. The yielded results exhibit measurement errors in the same order of magnitude as the noise figures themselves. Here, it is assumed that the noise figures are similar to the values measured on the receivers alone, which are approximately 3.5 dB. The DC power consumption of the complete front-end is 230 mA from a 3 V supply. All figures-of-merit of this linear active array are summarized in Tab. 6.1. Linear array 135

0 35 30 30 25 60 20 15 90

120

150 180

Fig. 6.6 Measured (solid lines) and simulated (dashed line) elevation patterns at 5.5 GHz. Total gain in dBi. The simulated pattern assumes an active gain of 27 dB.

0 35 −30 30 30 25 −60 60 20 15 −90 90

Fig. 6.7 Measured (solid lines) and simulated (dashed line) azimuthal pattern at 5.5 GHz. The simulated pattern again assumes additional 27 dB active gain. 136 Active antenna arrays Table 6.1 Summary of front-end characteristics Parameter Value layout linear 4 × 1 polarization 45◦-slanted passive gain ≈5.5 dBi active gain 24.5 dB P1dB −25.1 dBm IIP3 −16.3 dBm DC-supply 3 V, 230 mA NF ≈3.5 dB

6.1.3 Calibration As seen above, a significant mutual coupling between the antenna elements is present, which leads to a deformation of the azimuthal element patterns. In the previous chapter several methods were outlined, how to determine and compensate this coupling. Thereby, it was assumed that only one radiation mode is linked to each individual antenna element, which can be fully described by a corresponding port voltage. Furthermore, it was supposed, that the isolated element pattern is not changed by bringing the antenna into the vicinity of others (assumption of well-behaved antennas). To verify if the real array fulfils these conditions, the different calibration methods are applied to the measured far-field patterns. Fig. 6.8 shows the ideal array pattern for a uniform distribution. It is gained by superposition of simulated isolated-element patterns. The beam is tilted 30◦ to avoid a symmetrical situation, which could partly cover the effect of mutual coupling. The measured pattern is depicted in the same Fig. 6.8. The side-lobe and notch positions differ from the ideal angles. If the Fourier transformation technique is used to calculate the coupling matrix C and to correct the received signals, a significantly improved pattern is obtained. Within a range of ±60◦ off broadside, the measured pattern agrees well with the ideal one. An effective zero depth of more than 24 dB is obtained. To evaluate if radiation from the substrate edges has some effect on the pattern, the beamspace technique is applied to the same measure- ment data. Fig. 6.9 shows the calculated virtual array illuminations that synthesize the four individual element patterns. As expected, the main contribution of each element is found at its physical location. The re- spective neighbor elements show a parasitic excitation of −17 to −20 dB relative to the particular element. Apart from these values, which roughly Linear array 137

10

5

0

−5

−10

−15

−20 directivity [dB] −25

−30

−35

−40 −80 −60 −40 −20 0 20 40 60 80 phi [deg]

Fig. 6.8 Uncalibrated (solid line with circles), calibrated (solid line with crosses) and ideal (dashed line) array pattern for a uniform distribution tilted ◦ to 30 . This calibration bases on the Fourier-transform technique [67].

10

5 physical array 0

−5

−10

−15

−20 excitation [dB] −25

−30

−35

−40 2 4 6 8 10 12 14 16 18 20 element number

Fig. 6.9 Relative excitation of the 20-element virtual array gained using the beamspace-technique [69]. Elements 9–12 coincide with the physical array. The corresponding illumination values are the estimated entries of the coupling ma- trix C. 138 Active antenna arrays

coincide with the coupling matrix given by the Fourier transformation technique, no other significant contributions beyond the physical array are found. Contrary to the waveguide array in [69], no noticeable radiation from the substrate edges is present in this array. To compare the performance of the different calibration methods dis- cussed in chapter 5.2.1, it is useful to employ a suitable error measure rather than looking at the array diagram for a fixed steering angle. In Fig. 6.10, the residual error according to chapter 5.1.4 is calculated for a uniform distribution steered from −90◦ to +90◦. Within a range of −60◦ to 60◦, all methods show a significantly lower error power compared to the uncalibrated results. For angles outside, the error power remains high; no improvement is obtained by the calibration. Also it is noteworthy, that the Fourier transformation technique and the beamspace method yield similar results. In contrast to the other methods, the least-squares fitting method offers the flexibility to select arbitrary sampling points. It is found that, concen- trating these points to a small angular range, a very low error power can be achieved within this limited region. If the sampling points are spread more widely, the mean error increases. For sampling points uniformly spaced in

0

−5

−10

−15

error power [dB] −20

−25

−30 −90 −60 −30 0 30 60 90 phi [deg]

Fig. 6.10 Calculated residual error after calibration (compare chapter 5.1.4): uncalibrated (dashed line), Fourier transform technique (solid line), beamspace technique (solid line with circles) and least-squares fit (solid line with crosses). For the last two methods, the symbols are placed on the used sampling points. Gain and phase stability 139

0 0 −30 30 −5

−60 −10 60

−15

−90 90

Fig. 6.11 Azimuthal calibrated element pattern (solid lines) compared to the simulated isolated element pattern (dashed line). Least-squares fit equally ◦ ◦ spaced in u from sin(−40 ) to sin(40 ) is used for calibration. u between −40◦ and 40◦, the calibration results are slightly better than given by the other methods. Using the gained calibration matrix, it is possible to correct the single element pattern. The results are shown in Fig. 6.11. For a certain angle around the main beam direction, the corrected pattern agree well with the isolated element pattern. It can be concluded that the behavior of this realized active array can be approximated by the ideal uncoupled sensor model. For the relevant range of the antenna beam width a good agreement can be obtained. Outside of this range the element pattern differ from the ideal pattern and can not be corrected by calibration. Concentrating the calibration on this reasonable range, the residual error within this region can be further minimized.

6.2 Gain and phase stability To guarantee the correct spatial signal processing, the different processing paths need to have stable phase and amplitude responses with respect to each other. This aspect is investigated on the complete smart antenna test-bed described in chapter 2.4. This system allows to simultaneously receive all four channels and to store the results. A microwave signal generator is used to generate a constant sinusoidal signal that is connected 140 Active antenna arrays to the calibration port of the active antenna front-end. The system is left running for three days in a normal office environment. The recorded results are normalized on the first sample. The observed gain and phase changes are depicted in Fig. 6.12 and Fig. 6.13, respec- tively. The phases are referred to the phase of the first channel, because the calibration signal source and the local receiver oscillators are not syn- chronized, resulting in random absolute phases. It is observed, that the gain and phase changes of ±0.1 dB and ±1◦, respectively, are very small. All channels show a similar progression, which further reduces the effect on the spatial processing. To demonstrate this, the calculated zero depth and residual side-lobe level are shown in Fig. 6.14 The maximum values of −38 dB for the residual side-lobe level and 46 dB for the zero depth are very low, if it is considered that during three days no new calibration is performed. From the gain variation in Fig. 6.12 and the zero depth in Fig. 6.14, par- ticularly, a daily cycle can be observed. This regularity can be explained by a temperature dependence of the analog signal processing section. For verification, the active antenna front-end is placed inside a climatic ex- posure cabinet, which generates abrupt temperature changes. Fig. 6.15 shows the gain changes that result from a sudden temperature change from 0 ◦C to 25 ◦C. The gain variations are more than 1 dB and also the differences between the channels become significant. Obviously, the gain and phase changes are mainly caused by temper- ature variations. Gain and phase stability 141

0.2

0.1

0 gain variation [dB] −0.1

−0.2 0 10 20 30 40 50 60 70 time [hours]

Fig. 6.12 Measured gain variation over three days of all four channels. The values include the complete test-bed and were recorded in a normal office envi- ronment.

1

0

−1 phase variation [deg]

−2 0 10 20 30 40 50 60 70 time [hours]

Fig. 6.13 Test-bed phase variations over three days. Three channels referred to the first channel. 142 Active antenna arrays

−30

−40

−50 zero depth [dB]

−60

−70 0 10 20 30 40 50 60 70 time [hours]

Fig. 6.14 Calculated null depth (solid line) and residual side-lobe level (dashed line) over three days.

0.5

0

4

−0.5 3

2 −1 amplitude variation [dB rel.] 1

−1.5 0 10 20 30 40 50 60 time [min]

Fig. 6.15 Gain variation of the four front-end channels for a forced temperature ◦ ◦ step from 0 C to 25 C. Noise correlation 143 6.3 Noise correlation As discussed in chapter 2.3, the spatial processing typically requires that the amplitude noise contributed by the receivers is uncorrelated. The phase noise, in contrast, should be dominated by the phase noise of the LO. The contribution of the individual branches to the phase noise has to be prevented to avoid any zero leakage.

6.3.1 Amplitude noise correlation To characterize the noise added by the system, the receiver is operated in a shielded environment with no signals present. To study the influence of the different system blocks (compare Fig. 2.13), the system without the frontend is examined first. Fig. 6.16 shows a typical output noise-spectrum when the analog fron- tend is removed. The observed spurious level is common in communication systems. It originates from the unwanted injection of the switching and clock noise emitted from the digital circuits into the analog hardware. Usually it can be tolerated, because its contribution to the total noise power is low. To observe the noise correlation, the angular dependent mean noise power

T N 2 1 −j2π n∆ sin θ P¯ (θ)= a˜ (t)e λ (6.1) noise T n t=1 n=1 X  X  is calculated and depicted in Fig. 6.17, where ∆/λ is 0.5, T the number of time samples and θ the angle into which the array is steered. Compared

−80

−90

−100

−110

[dBFS] −120

−130

−140

−150 −100 −50 0 50 100 freq [kHz]

Fig. 6.16 Noise spectrum of digital downconverter (DDC) with spurious, rela- tive to full scale excitation (dBFS). 144 Active antenna arrays

90 30 120 60

20 150 system 30 10

180 0

210 DDC only 330

240 300 270

Fig. 6.17 Directional pattern of the digital part and the complete system (solid lines) compared to the not coherently added noise levels (dashed lines). to the theoretical pattern without correlation

1 T N 2 P¯ = a˜ (t) (6.2) n,uncorr T n t=1 n=1 X X   the noise origins mainly from the array broadside, indicating strong cor- relation. If the whole system including the frontend is considered, this correlated noise of the DDC is covered by the amplified thermal noise from the system input, but a correlated fraction of the noise remains. As a quantitative measure, the normalized cross-correlation matrix R can be calculated: 0.89 0.22 0.08 0.12 ∗ 0.22 1.14 0.20 0.15 a˜i ∗ a˜j R =   , [ri,j ]= (6.3) 0.08 0.20 0.95 0.37 1 N a˜2  0.12 0.15 0.37 1.02  N n=1 n   P This noise correlation has to be taken into account for the development of the array processing algorithms. 6.3.2 Phase noise correlation To evaluate the influence of phase noise, a sinusoidal signal was applied at the antennas with the maximum allowed power level. This ensures, Noise correlation 145

6

4

2

0

phase [deg.] −2

−4

−6 0 100 200 300 400 500 time [us]

Fig. 6.18 Phase variation of four channels due to phase noise for few samples. The changes are almost identical, but a slight offset between the channels is observed. that the phase noise of the resulting output signals dominates the ampli- tude noise. Fig. 6.18 shows the phase jitter of the different channels over time. To further reduce the effect of amplitude noise, only the phase dif- ferences are considered. Therefore, the pattern zero was calculated from the normalized signal values:

N 1 a˜ (t) s (t)= (−1)n n (6.4) 0 N |a˜ (t)| n=1 n X The resulting diagram over time is shown in Fig. 6.19. The measurable null depth is limited by the remaining amplitude noise power. The corre- sponding noise floor was calculated and is depicted for comparison. The graph shows a significant degradation of the formed zero beyond the noise floor. The main changes occur very slow in time, associated with frequen- cies of tenths of hertz. This indicates that the decorrelating effect origins from 1/f-noise contributions to the individual paths. The achievable zero depth is limited to about −45 dB. This is a low value for most practical applications; it has to be considered that a high power level was used and an amplitude normalization for noise reduction had to be used to make a significant zero-leakage visible. 146 Active antenna arrays

−40

−50

−60

−70

zero depth [dB] −80

−90

−100 0 0.2 0.4 0.6 0.8 1 time [s]

Fig. 6.19 ”Zero-leakage”due to uncorrelated phase noise (solid line) and ther- mal noise floor (dashed line).

6.4 Conformal array 6.4.1 Motivation When designing an array of antennas, it has to be considered that the final array pattern is given not only by the changeable element weights, but also by the directivity of the individual elements. The employed patch antennas show a 3 dB beam-width of around 90◦; beyond that range, the gain drops significantly. If a beam is steered into these directions, the achievable maximum gain is accordingly lower. In particular patch antennas can be realized on a curved surface, which allows to point the individual elements into different directions. With the proper orientations, an antenna array can be build, which can be steered over a wider angular range without this decrease in gain. Fig. 6.20 shows the maximum obtainable gain of a four-element array for both, a linear and a curved arrangement. To reach a gain drop of less than 3 dB over a range of 180◦, an angle of as much as 50◦ between two antenna elements is required. In addition to the more uniform steering capability, this curved array offers a further advantage, if used in a communication system: in an en- vironment with rich scattering, like an indoor scenario, the array offers a kind of “path diversity”. Due to the individual orientations, the link from the transmitter to each antenna favors different propagation paths. This Conformal array 147

20

15

10

gain [dBi] 5

0

−5 −135 −90 −45 0 45 90 135 phi [deg]

Fig. 6.20 Individual element pattern (dashed lines) and maximum obtainable directivity (solid lines) for a linear array (lines without symbols) and a circular arrangement (lines with symbols). The arrows indicate the main-beam direc- tions of the elements. makes it less likely that the received signals at several antennas experi- ence fading due to signal cancellation in the same situation. Therefore, the overall link reliability is improved. The type of antenna arrays on a curved surface is known as confor- mal array. The main application is found in radar systems, whereby the antenna array can be integrated into the arbitrarily shaped surface of an aircraft or other vehicle. Generally it is difficult to manufacture active conformal arrays, as the mounting of active components requires a planar structure. One proposed way is to package the front-end into thin and compact modules, which are mounted onto the curved surface [82]. It is problematic as well to realize a precise symmetrical divider network to provide a calibration signal. Typically these arrays have to be calibrated from far-field measurements [83]. In the following a four element conformal antenna array is presented, which bases on the monolithically integrated receiver front-end presented in chapter 3.4 and the calibration scheme proposed in chapter 5.3. 6.4.2 Design The basic idea for this easy-to-manufacture active conformal array is to use a planar substrate. First, all active and discrete components are mounted 148 Active antenna arrays

calibration CAL1 line CAL2

attenuator receiver receiver receiver receiver

LO

IF1 IF2 IF3 IF4

Fig. 6.21 System block diagram of the conformal active array. Completely integrated active front-ends are used to minimize the area consumption. The transmission-line calibration method is used to calibrate the array. A microstrip crossing provides the needed signal coupling. using standard methods, before the array is bent into the wanted shape. Those areas, that remain planar in the final array, can be used for circuitry like standard microstrip substrates. Inside the regions that are bent no components can be mounted, as these would not remain attached; trans- mission lines have to be placed carefully. This significantly reduces the substrate area which is available for circuit placement. To overcome the problem of the limited space, the monolithically inte- grated receivers described in chapter 3.4 are employed for each of the four receiver branches. As it can be seen in the block diagram in Fig. 6.21, the antennas are directly connected to the monolithic receivers. To avoid a symmetric divider network, the calibration inputs of the receivers are not used here. Although the changes are found to be small, the unused inputs are terminated with 50 Ω loads. Instead of this classical calibration signal distribution, the transmission line calibration method proposed in chapter 5.3 is used. The coupling between the calibration line and the receiver branches is achieved by a simple crossing of the transmission lines, where the interrupted calibration line is bridged by a bond wire. Varying the distance between the line ends and the continuous transmission line, the capacitive coupling between both can be adjusted. Here, gaps of 100 µm length are chosen, resulting in a coupling loss of around 25 dB at 5.5 GHz. The two end capacitances together with the bond-wire inductance act as a discrete transmission line equivalent and, therefore, do not introduce any electrical discontinuity into the calibration line structure. Conformal array 149

patch substrate

d=27.3 mm r=5mm

screws circuit side r=29 mm

subtrate mechanical support

Fig. 6.22 Sketch of the mechanical support for the bent substrate. The antenna elements are arranged on a circle. The substrate is fixed at both sides of the bend by screws and the solid antenna substrate. This prevents a bending of the planar parts, where the components are mounted.

As already mentioned in chapter 5.3, the precision of the proposed cali- bration line method is degraded by reflections, which cause back-travelling waves on the transmission line. To mitigate possible reflections at the end of the line, two 15 dB attenuators are added. To demonstrate the feasibility, a large angle of 50◦ between two el- ements is chosen. As demonstrated in Fig. 6.20, this allows a coverage of 180 degree with a gain decrease of less than 3 dB. Experimentally it was found, that a bending radius of 5 mm is possible without mechan- ically damaging the Duroid 6010 substrate (r = 10.2, h = 635 µm) or the 17 µm thin copper film. This leads to the arrangement depicted in Fig. 6.22: the four patches are arranged as tangents on a circle at the corresponding angles. A mechanical support defines the final shape of the array and the 5 mm bending radius. At both sides of the bend, the substrate is attached to this support to avoid a bending of these planar parts. This is important, as a bending of these parts could cause the active circuitry to fail. In the remaining areas, the material of the supporting structure is removed to allow the placement of undisturbed high frequency transmission lines on the substrate. The distance from center to center of two adjacent patches is set to 27.3 mm, which is approximately half of the free space wavelength λ0 at 5.5 GHz. Fig. 6.23 shows the planar circuit layout before the substrate is bent. The four identical receiver branches can easily be identified. Also the 150 Active antenna arrays CAL2 attenuator IF4 bentregions IF3 calibrationline receiver DC patch IF1 IF2 LO slot mechanicalsupports CAL1

Fig. 6.23 Layout of the circuit substrate for the conformal antenna how it is fabricated in standard planar technology. After assembly, the indicated regions of the substrate are bended and mounted on a mechanical support. The other areas remain planar. Substrate size is 135 ×71 mm. calibration line can be seen, terminated at both ends by an attenuator to reduce the influence of external reflections. The dashed lines indicate the bent regions and the areas which are occupied by the mechanical sup- ports. These ”keep-out”areas require about 30% of the total substrate size. Transmission lines can be placed in the bent regions, whereby a certain phase error has to be expected. Following simple geometrical considera- tions, the electrical length of the conductor is shortened by approximately 2◦ in the present case. The figure shows that a non-symmetrical divider network is chosen to distribute the LO signal. This avoids the placement of power splitters on bended regions. Fig. 6.24 illustrates a further advantage of the proposed array struc- ture: from electromagnetic simulations it is found that – for an identical center-to-center spacing and the same polarization – the conformal array shows a significantly reduced mutual coupling. The coupling coefficients Conformal array 151

−15

−20

−25 coupling coefficient [dB]

−30 4.8 5 5.2 5.4 5.6 5.8 6 6.2 frequency [GHz]

Fig. 6.24 Mutual coupling from simulated s-parameters: coupling of the outer elements (solid lines) and between the inner elements (solid line with crosses). For comparison: simulated coupling coefficient of two inner elements in a verti- cally polarized linear array (dashed line). of a similar planar array reach up to −15 dB, whereas the same parameter stays below −22 dB for the bent structure.

6.4.3 Experimental results The manufactured array was characterized inside an anechoic chamber. The measured channel gains versus frequency are depicted in Fig. 6.25. All curves show a similar progression. Fig. 6.26 shows the same curves, but also the measured gain of the frontend at the corresponding image frequencies. Image rejection is better than 40 dB over the whole band. For an estimation of the expected element patterns, the isolated el- ement pattern is simulated. For this purpose, the simplified structure depicted in Fig. 6.27 is used. It includes one antenna patch, the substrate that supports this patch and the ground plane. The circuit substrate is only present below the antenna, where it is needed to model the aperture coupling and feeding network. The bends toward the adjacent elements are included into the model to consider the diffraction effects. The measured element pattern of elements four and three, pointing at −75◦ and −25◦, respectively, are depicted in Fig. 6.28 and Fig. 6.29. The pattern of element one and two are almost symmetric to those shown. 152 Active antenna arrays

35

30

25

gain [dBi] 20

15

10 5 5.2 5.4 5.6 5.8 6 freq [GHz]

Fig. 6.25 Total frontend gain compared to isotropic radiator. Channel 1 (solid line with crosses), channel 2 (with diamonds), channel 3 (with circles) and chan- nel 4 (with squares).

40

30

20

10

0

−10

gain [dBi] −20

−30

−40

−50

−60 5 5.2 5.4 5.6 5.8 6 freq [GHz]

Fig. 6.26 Frontend gain at signal band (solid lines) and image band (dashed lines). Channel 1 (with crosses), channel 2 (with diamonds), channel 3 (with circles) and channel 4 (with squares). Conformal array 153

Fig. 6.27 Simplified model used to simulate the directional pattern of an iso- lated element. The circuit substrate just extends behind the patch antenna. The bends towards the adjacent elements are modelled to include diffraction effects.

Over a broad range the measured element pattern agree well with the simulated isolated element pattern. Differences are found for angles |φ| > 90◦ and angles that differ significantly more than 90◦ from the main beam direction of the particular element. Fig. 6.30 explains the origin of these constraints: angles beyond ±90◦ correspond to the backside of the array. On this side the array is mechanically mounted. Measurements at these angles are, therefore, obstructed by the supporting structure. Angles of more than 90◦ to 140◦ from the main beam direction are obstructed by the adjacent element. It is found that the range of agreement is much larger than it is for the linear array in section 6.1. To investigate the influence of mutual coupling, the coupling matrix C is determined from the far-field measurements. The method suited best for this purpose is the least-squares fitting method [72], which is already discussed in chapter 5.2.1. It assumes that a coupling matrix C exists, which maps the ideal array vectors gi(φ) on the measured ones g(φ):

g(φ)= C · gi(φ). (6.5)

The matrix C then is found numerically by finding the optimum solution over a certain number of samples at the angles φm. In contrast to the linear array case in Eqn. 5.23, the ideal array vectors are now calculated from the isolated element patterns gisol(φ) as 154 Active antenna arrays

10

5

0

−5 gain [dBi] −10

−15

−20 −135 −90 −45 0 45 90 135 angle [degree]

Fig. 6.28 Measured (solid line with crosses) and calibrated (solid line with circles) pattern of outer element compared to the corresponding isolated element pattern (dashed line) at 5.5 GHz. The symbols are placed at the angles used for calibration. The arrow indicates the direction of the main beam.

10

5

0

−5 gain [dBi] −10

−15

−20 −135 −90 −45 0 45 90 135 angle [degree]

Fig. 6.29 Measured (solid line with crosses) and calibrated (solid line with circles) pattern of inner element at 5.5 GHz. Again, the symbols are placed at the angles used for calibration. The corresponding isolated element pattern ◦ (dashed line) is identical to the one in Fig. 6.28, but rotated by 50 . Conformal array 155

o element 0 orientation shadow boundary

backside center of rotation

Fig. 6.30 Valid range of the single element model. On the backside and behind the shadow boundary the pattern of the single element and the array differ.

◦ gisol(φm − 75 ) ◦ i gisol(φm − 25 ) g (φm)=  ◦  . (6.6) gisol(φm + 25 ) ◦  gisol(φm + 75 )    This way, a coupling matrix can be determined without assuming a Fourier relationship, which is no longer valid for a this curved aperture surface. To concentrate the calibration on the relevant regions, the sampling points are constrained to the front side of the array from −90◦ to +90◦. One sampling point every 15◦ is chosen, resulting in a total of 13 different angles. The determined matrix

0.1 −22.8 −19.6 −23.8 −18.4 −0.1 −20.3 −29.6 20 log |C| = , (6.7)  −24.8 −16.6 0.1 −26.0   −23.2 −19.9 −17.2 1.3      indicates low coupling coefficients below −20 dB, which agrees well with the predicted low mutual coupling, previously shown in Fig. 6.24. The corrected element gain pattern are also depicted in Fig. 6.28 and Fig. 6.29. Compared to the linear array in section 6.1.3, only a minor difference is found between the calibrated and uncalibrated curve. It is worth noticing, that the pattern deviation on the backside of the outer elements (see Fig. 6.28 at angles greater than 45◦) can not be compensated by this method. This is another indication that other effects than mutual coupling are responsible for this deviation. In conformal array processing it is not uncommon to work with subsets of antennas to avoid this shadowing problem. In this particular array it would be reasonable to select a subset of three antennas for angles more than 45◦ away from the main axis. This would be possible without 156 Active antenna arrays

15

10

5

0

−5 array gain [dBi] −10

−15

−20 −135 −90 −45 0 45 90 135 phi [degree]

Fig. 6.31 Array gain (including element gain) of the conformal array at 5.2 GHz. The beam is steered to 0◦ with zeros at ±50◦. Simulated (dashed line with crosses) and measured (solid line with stars) diagram and corrected results obtained by the transmission-line calibration method (solid line with circles) and by far-field measurements (solid line with squares). degrading the total gain significantly, as the contribution of the backward facing element is very low. To study the behavior of the whole array, the antenna outputs are com- bined with suitable element weights. The array is steered to 0◦ to involve all elements and two diagram zeros are formed at −50◦ and +50◦. Due to the different element orientations, the element pattern cannot simply be removed from the calculations of the steering vector. Here, the symmetry is used and the remaining two weights are calculated numerically to ob- tain the beam and the zeros. The simulated array diagram is depicted in Fig. 6.31. The measured diagram, also shown there, does neither show a pronounced main beam, nor a notch at −50◦. This can be explained by the fact, that the individual LO feeding branches have different lengths and, therefore, introduce different phase shifts. Using the transmission-line cal- ibration method, these phase shifts can be determined and compensated. The curve in Fig. 6.31 shows that the calibrated diagram resembles the simulated pattern. Although the zero depth of 14 dB at −50◦ is limited, the main beam is formed well. A good zero depth of more than 23 dB is obtained using a far-field calibration. Conclusions 157

Fig. 6.32 Photography of the manufactured active conformal antenna array.

Fig. 6.32 shows a photograph of the fabricated active conformal an- tenna array. One can clearly see the patch antennas pointing at individual angles. The metal strips in the lower part are used to reinforce the sub- strate and give a good hold to the connectors. The circles in the center belong to the ground vias connected to the receiver chips, which reside on the opposite, non-visible side. 6.5 Conclusions In this chapter, active antenna arrays are presented, which employ the monolithically integrated front-ends that were presented in chapter 3. As intended, an excellent long-term stability of gain and phase transfer func- tions is achieved with the integration of the critical high frequency process- ing into the antenna array. The remaining gain and phase variations are mainly due to changes of the ambient temperature and are well correlated. It is found, that the phase variations caused by phase noise also are well correlated. The ”zero-leakage”resulting from uncorrelated phase noise stays fairly low and does not affect the practical application. This opens 158 Active antenna arrays the way for a more aggressive design. If the specifications are relaxed and a certain amount of uncorrelated phase noise is intentionally tolerated, the system could benefit in other areas. E.g., it is possible to replace the linear LO amplifiers by highly efficient nonlinear class-C amplifiers to reduce the power consumption. It could also be attractive to use phase-locked-loops to generate the LO signals locally for each branch from a common low- frequency reference. This would replace the last remaining high-frequency interconnections. On a four element aperture-coupled patch antenna array with 45◦- slanted polarization it is demonstrated, that the mutual coupling can be described and compensated by a suitable coupling matrix. The behavior of this calibrated array can be described by the simple geometrical array model with a reasonable precision. For this particular array, the agreement is limited to an angular range of less than ±60◦ from the array broadside direction. Beyond that range the individual element pattern differ from the pattern of an isolated element in a way, which can not be explained by mutual coupling. The high integration of the RF hardware allows to construct a new type of active conformal arrays. The individual orientation of the array elements allows to distribute the maximum obtainable beamforming gain more uniformly over a larger angular range. A conformal array is demon- strated, which covers a range of 180 degree with less than 3 dB decrease of gain. This array inherently exhibits a significantly lower mutual coupling. Furthermore, the geometrical array model is valid for a larger range. If subsets of three antennas are chosen and the element facing to the oppo- site side is not considered, the geometrical model can be used over the complete 180◦ range. 7 Summary, conclusions and outlook

The following sections list the main results achieved in this dissertation and indicate some directions for future work on this subject. 7.1 System design Chapter 2 presents the system-level design of a multiple-channel receiver for 5.15 –5.875 GHz. To obtain a modular and flexible system, a super- heterodyne architecture is chosen. To solve the problem of image filtering without the need for external components, a high intermediate frequency is selected and the mirror band is placed to a quiet region. The imple- mentation aspects of a multi-channel system are discussed with a focus on the noise correlation in such a system. For the application in a multi- dimensional channel-sounder, minimum noise contribution of the frontend is required. It is found that the required mixer is the key component to reach a good linearity and a low noise-figure simultaneously. The use of a passive resistive mixer is proposed to reach a high dynamic range. All these considerations lead to the specifications of the single circuit blocks. 7.2 Integrated circuit design Chapter 3 describes the implementation of the active downconverter. The different building blocks are integrated on a commercially available 0.6 µm GaAs MESFET process. The key components are a three-stage low-noise amplifier with 22 dB gain and 2.4 dB noise figure and in input compression point of −16.5 dBm, a resistive mixer with 6.8 dB conversion loss and an input compression point of 4 dBm at a required LO power of 10 dBm and a downconverter with 7 dB conversion gain at a required LO power level of 5 dBm and an input compression point of 1 dBm. The complete frontend is integrated, including a lumped element image filter. For this complete receiver, the following key parameters are measured: at 5.5 GHz the con- version gain is 26 dB for an LO power higher than 0 dBm. The input 160 Summary, conclusions and outlook

1 dB-compression point is higher than −18 dB. Over the band of interest, the single-sideband noise-figure is below 3.8 dB and the image rejection is between 35 dB and 50 dB. Both are remarkable results for integrated receivers. To avoid a passive switch and the associated higher noise figure for system calibration, an active switching concept is proposed. A channel mismatch of less than 0.7 dB over the whole band can be achieved using a symmetrical layout. The whole system draws 76 mA from a 3 V supply and requires 3.2 mm2 of chip area. On the same process, a power amplifier with an integrated temperature monitoring diode is designed to study the dynamic behavior of the calibration in transmit arrays. To evaluate the capability of silicon-based technologies, a 10.7 –11.7 GHz downconverter is designed on a low-ft SiGe process. A gain of 15 dB, a noise figure of less than 4 dB and an input compression point of −12.7 dBm are found for the low-noise amplifier alone. Utilizing the MOSFET as a resistive mixer, a complete receiver is integrated and a conversion gain of 10 dB, a single-sideband noise-figure of 7 dB and an input 1 dB-compression point of −14 dBm are measured. The power consumption is 27.6mA from a 3.3 V supply and the circuit requires an area of 1.75 mm2. The dynamic range of this receiver is significantly larger than other published receiver on comparable technologies, which all base on active mixers.

7.3 Passive arrays Chapter 4 focuses on the passive antenna structure. The choice of aperture- coupled patch antennas is motivated, for which an equivalent circuit model can be found that supports the design process. With the help of this model, an antenna layout is determined, which achieves the required band- width. For compatibility with differential integrated circuits, which are ad- vantageous for mixed-signal applications and on technologies with strong substrate cross-talk, a differential antenna interface is proposed. The mu- tual coupling is discussed and classified. For linear arrays of aperture- coupled patch antennas it is verified by simulations that assumption of a well-behaved array can be made. This allows to describe the far-field pattern of the array by the superposition of identical element patterns, which equal the pattern of a single isolated element. The coupling can be described by a coupling matrix C, which also allows to compensate this effect. To simulate the consequence of a mismatch at the antenna feed- ing points, the equivalent circuit model for the antenna is extended by mutual coupling. It is experimentally demonstrated that the interface be- tween antenna and amplifier in an active array can be optimized to lower the mutual coupling. Calibration 161 7.4 Calibration In chapter 5 various practical aspects of the calibration of antenna arrays are discussed. To assess the required precision, the typical element varia- tions and error effects are reviewed. An overview is given over the available calibration schemes, which can be subdivided into methods which use far- field sources and those which use a local reference signal. The far-field methods include the effect of mutual coupling, but an anechoic environ- ment is needed. Considering the fact that the main changes of the transfer functions are caused by the active hardware, it is attractive to perform a two-step calibration: the passive array is characterized once from measure- ments or simulations and the local calibration signal is used to track the receiver changes. To simplify the calibration network for two-dimensional or arbitrary array arrangements, a new calibration method is proposed that employs a local coupling between all branches provided by a weakly coupled transmission line, which links all elements. The determination of the transmission factors of the line segments is part of the calibration process itself. The dynamic effect of switched power amplifiers on the cali- bration of transmit arrays is studied. For monolithically integrated power amplifiers two thermal time-constants are experimentally found. A typical burst up to a duration of milliseconds can be calibrated using a single set of correction values. If the duty cycle stays constant, the obtained values can be reused for several pulses. For a changed duty cycle the variations are less correlated and a slight pattern variation occurs. 7.5 Active antenna arrays Chapter 6 deals with the final integration of the monolithic receivers pre- sented in chapter 3 and the passive array discussed in chapter 4 and system-level measurements are performed on these active frontends. An excellent long-term stability of gain and phase transfer functions is found. The gain and phase variations are mainly due to changes of the ambi- ent temperature and are well correlated. The phase variations caused by phase noise also are well correlated. The zero-leakage resulting from un- correlated phase noise stays fairly low and does not affect the practical application. For a four-element array with 45◦-slanted polarization it is demonstrated, that a good calibration can be achieved over a range of ±60◦ from the array broadside direction. This coincides approximately with the beam-width of the antenna elements. An active conformal array is designed, which covers a range of 180 degree with less than 3 dB de- crease of gain. This array inherently exhibits a significantly lower mutual coupling. Due to the curved surface, the individual element pattern agree with the isolated element pattern over a larger angular range. 162 Summary, conclusions and outlook 7.6 Conclusions and future work The developed active antenna arrays benefit from the monolithic inte- gration of the critical RF frontends. A compact and robust solution is obtained, which allows to construct uniform linear arrays as well as more freely arranged arrays. The novel transmission line calibration method helps to simplify the design of these arrays. The elimination of high-frequency transmission lines and connectors leads to a very good long-term stability of the individual transfer functions. Thanks to the good reproducibility of monolithically integrated circuits not only similar path responses are achieved, but also all changes are well correlated. This ensures that the negative effect on beamforming is very low and the interferer rejection capability is not significantly degraded. Long time intervals between the calibrations can be tolerated. This is of special interest in transmit arrays: rapid temperature changes in switched power amplifiers could cause a pattern degradation, which can be avoided using identical monolithic amplifiers. The presented system is consequently designed for a low noise-figure and well synchronized phases of the transfer functions. The system-level analysis of the multi-channel phase noise indicates that a very good corre- lation of the phases is achieved, the phase errors of the individual branches are not significant for most applications. For a future system it is advis- able to specify a tolerated phase deviation due to individual phase noise contributions. Adjusting the system architecture, this parameter can be traded off against other advantages: nonlinear LO amplifiers could reduce the power consumption, or an integrated phase-locked loop could be ap- plied to generate the LO signal locally. Both options increase the amount of uncorrelated phase noise. The integration of a phase-locked loop would further reduce the number of off-chip high-frequency connections. This be- comes increasingly attractive at higher operating frequencies, which will be required to address the growing bandwidth demands in the future. The SiGe technology appears suitable to combine the analog frontend and the required digital phase-detector and frequency divider. A differential cir- cuit design, which minimizes the noise cross-talk, could benefit from the proposed differential antenna interface. In the presented system the antenna arrangement was selected for the requirements of classical beamforming, being a suboptimal choice for a MIMO system. With the example of a conformal array it has been demon- strated that arrangements different from uniform linear arrays can be re- alized, which are advantageous in certain scenarios. Bibliography

[1] P.K. Bondyopadhyay, ”The First Application of Array Antenna”, IEEE Int. Conf. on Phased Array Systems and Technology, p. 29, may 2000 [2] D. Parker, D. C. Zimmermann, “Phased-Arrays–Part1: Theory and Architectures”, IEEE T. on Microwave Theory and Techniques, vol. 50, no. 3, p. 678, mar. 2002 [3] D. Parker, D. C. Zimmermann, “Phased-Arrays–Part2: Implemen- tations, Applictions, and Future Trends”, IEEE T. on Microwave Theory and Techniques, vol. 50, no. 3, p. 688, mar. 2002 [4] G. Tsoulos, J. McGeehan, M. Beach, “Space division multiple access (SDMA) field trials. I. Tracking and BER performance”, IEE Proc. Radar, Sonar and Navigation, vol. 145, iss. 1, p. 73, feb. 1998 [5] G. Tsoulos, J. McGeehan, M. Beach, “Space division multiple access (SDMA) field trials. II. Calibration and linearity issues”, IEE Proc. Radar, Sonar and Navigation, vol. 145, iss. 1, p. 79, feb. 1998 [6] S. Anderson, U. Forssen, J. Karlsson, T. Witzschel, P. Fischer, A. Krug, “Ericsson/Mannesmann GSM field-trials with adaptive antennas”, IEEE Vehicular Technology Conference, vol. 3, p. 1587, may 1997 [7] J.J. Monot, J. Thibault, P. Chevalier, F. Pipon, S. Mayrargue, A. Levy, “A fully programmable prototype for the experimentation of the SDMA concept and use of smart antennas for UMTS and GSM/DCS1800 networks”, IEEE Int. Symp. Personal, Indoor and Mobile Radio Communications, vol. 2, p. 534, sept. 1997 [8] A. O. Boukalov, S.-G. H¨aggman, “System Aspects of Smart-Antenna Technology in Cellular Wireless Communications–An Overview”, IEEE T. on Microwave Theory and Techniques, vol. 48, no. 6, p. 919, june. 2000 [9] G. J. Foschini, “Layered space-time architecture for wireless com- munication in a fading envionment when using multi-element an- tennas”, Bell Labs Tech. J., vol. 1, no 2, p. 41, autumn 1996 [10] S. M. Alamouti, “A Simple Transmit Diversity Technique for Wire- 164 BIBLIOGRAPHY

less Communications”, IEEE J. on Selected Areas in Communica- tions, vol. 16, no. 8, oct. 1998 [11] G. J. Foschini, D. Chizhik, M. J. Gans, C. Papadias, R. A. Valen- zuela, “Analysis and Performance of Some Basic Space-Time Archi- tectures”, J. on Selected Areas in Communications, vol. 21, no. 3, april 2003 [12] D. Chizhik, J. Ling, P.W. Wolniansky, R.A. Valenzuela, N. Costa, K. Huber, “Multiple-Input–Multiple-Output Measurements and Modeling in Manhattan”, J. on Selected Areas in Communications, vol. 21, no. 3, p. 321, april 2003 [13] A. Stucki et. al., ”Propsound System Specifications Document: Con- cept and Specifications”, 2001, Elektrobit AG, www.propsound.com [14] J. Lin, T. Itoh, “Active Integrated Antennas”, IEEE T. on Mi- crowave Theory and Techniques, vol. 42, no. 12, p. 2186, dec. 1994 [15] “Broadband Radio Access Networks (BRAN); HIPERLAN Type 2; Physical (PHY) layer”, ETSI TS 101 475, technical specification, april 2000 [16] J. Crols, M.S.J. Steyaert, ”Low-IF Topologies for High-Performance Analog Front Ends of Fully Integrated Receivers”, IEEE Trans. on Circuits and Systems – II: Analog and Digital Signal Processing, vol. 45, no. 3, march 1998 [17] F. Ellinger, W. B¨achtold, “Adaptive Antenna Receiver Module for WLAN at C-Band With Low-Power Consumption”, IEEE Mi- crowave and Wireless Components Letters, vol. 12, no. 9, sept. 2002 [18] S.S. Jeon, Y. Wang, Y. Qian, T. Itoh, “A novel smart antenna system implementation for broad-band wireless communications”, IEEE Trans. on Antennas and Propagation, vol. 50, iss. 5, p. 600, may 2002 [19] J.D. Fredrick, Y. Wang, T. Itoh, “A smart antenna receiver array using a single RF channel and digital beamforming”, IEEE Trans. on Microwave Theory and Techniques, vol. 50, iss. 12, p. 3052, dec. 2002 [20] T. Nishio, Y. Wang, T. Itoh, “Multiple-beam adaptive array archi- tecture using channel-level FDMA concept”, Antennas and Propa- gation Int. Symp., vol. 3, p. 908, june 2003 [21] R. Nabar, ”MIMO wireless communications”, lecture ETH doctoral school, summer semster 2003 [22] C. Waldschmidt, C. Kuhnert, S. Schulteis, W. Wiesbeck,”Compact MIMO-arrays based on polarisation-diversity”, IEEE Antennas and Propagation Int. Symp., vol. 2, p. 499, june 2003 BIBLIOGRAPHY 165

[23] J. P. Weem, Z. Popovi´c, ”A method for determining noise coupling in a phased array antenna”, IEEE MTT-S Int. Microwave Symp. 2001, vol. 1, p. 271 [24] L. C. Howard, D. J. Rabideau, ”Correlation of Nonlinear Distor- tion in Digital Phased Arrays: Measurement and Mitigation”, IEEE MTT-S Int. Microwave Symp. 2002, vol. 1, p. 49 [25] G. W. Slade, ”Effects of local oscillator phase noise on interference rejection capability of CDMA receivers using adaptive antenna ar- rays”, IEEE MTT-S Int. Microwave Symp. 2002, vol. 2, p. 1353 [26] D.S. Baum, H. B¨olcskei, ”Impact of Phase Noise on MIMO Channel Measurement Accuracy”, VTC Fall 2004 [27] J. H. Reed, ”Software Radio”, Prentice Hall, chapter 4, ISBN 0-13- 081158-0, 2002 [28] Y. Yamashita, A. Endoh, K. Shinohara, K. Hikosaka, T. Matsui, S. Hiyamizu, T. Mimura, ”Pseudomorphic In0.52Al0.48As/In0.7Ga0.3As HEMTs With an Ultrahigh fT of 562 GHz”, IEEE Electron Device Letters, vol. 23, no. 10, oct. 2002 [29] TriQuint Semiconductor, ”TQTRx process information”, http://www.tqs.com [30] N. Zamdmer, J. Kim, R. Trzcinski, J.-O. Plochart, S. Narasimha, M. Khare, L. Wagner, S. Chaloux, ”A 243-GHz Ft and 208-GHz Fmax, 90-nm SOI CMOS SoC technology with low-power millimeter- wave digital and RF circuit capability”, IEEE Symposium on VLSI Technology 2004, accepted for publication [31] J.-S. Rieh, B. Jagannath, H. Chen, K. Schonenberg, S.-J. Jeng, M. Khater, D. Ahlgren, G. Freeman, S. Subbanna, ”Perfor- mance and Design Considerations for High Speed SiGe HBTs of fT /fmax=375GHz/210GHz”, Intern. Conf. on Indium Phosphide and Related Materials, p. 374, 2003 [32] J. j. Kucera, ”Highly Integrated RF transceivers”, Diss. ETH Nr. 13361 [33] L.E. Larson, ”Integrated Circuit Technology Options for RFIC’s– Present Status and Future Directions”, IEEE J. of Solid-State Cir- cuits, v. 33, no. 3, p. 387, march 1998 [34] W. B¨achtold, ”Mikrowellenelektronik”, Vieweg Verlag, ISBN 3-528- 03937-X [35] H. J¨ackel, ”Schnelle analoge und digitale Kommunikations- Elektronik I”, lecture script, ETH Zurich [36] S. A. Maas, ”Microwave Mixers”, Norwood, MA, 1993 166 BIBLIOGRAPHY

[37] S. A. Maas, “A GaAs MESFET Mixer with Very Low Intermodula- tion”, IEEE Trans. on Microwave Theory and Techniques, vol. 35, no. 4, april 1987 [38] F. Ellinger, R. Vogt, W. B¨achtold, ”Compact Monolithic Integrated Resistive Mixers With Low Distortion for HIPERLAN, IEEE T. on Microwave Theory and Techniques, vol. 50, no. 1, p. 187, jan. 2002 [39] R. Hallgren, P. Litzenberg, ”TOM3 Capacitance Model: Linking Large- and Small-Signal MESFET Models in SPICE”, IEEE T. on Microwave Theory and Techniques, vol. 47, no. 5, p. 556, may. 1999 [40] T. Lee, H. Samavati, H. Rategh, ”5-GHz CMOS Wireless LANs”, IEEE T. on Microwave Theory and Techniques, vol. 50, no. 1, p. 286, jan. 2002 [41] B. Razavi, ”A 5.2-GHz CMOS Receiver with 62-dB Image Rejec- tion”, IEEE J. of Solid-State Circuits, vol. 36, no. 5, p. 810, may 2001 [42] M. Copeland, S. Voinigescu, D. Marchesan, P. Popescu, M. Maliepaard, ”5-GHz SiGe HBT Monolithic Radio Transceiver with Tunable Filtering”, IEEE T. on Microwave Theory and Tech- niques, vol. 48, no. 2, , p.170, feb. 2000 [43] J.P. Maligeorgos, J.R. Long, ”A low-voltage 5.1-5.8-GHz Image- Reject Receiver with Wide Dynamic Range”, IEEE J. Solid-State Circuits, vol. 35, no. 12, dec. 2000 [44] J. Alves Torres, J. Costa Freire, ”Monolithic Transitors SPST Switch for L-Band”, IEEE T. on Microwave Theory and Techniques, vol. 50, no. 1, p. 51, jan. 2002 [45] A. Italia, E. Ragonese, G. Girlando, G. Palmisano, ”A 5–GHz Mono- lithic Silicon Bipolar Down-converter with a 3.2-dB Noise Figure”, IEEE Radio Frequency Integrated Circuits Symposium 2003, p. 453, june 2003 [46] T.-P. Liu, E. Westerwick, ”5-GHz CMOS Radio Transceiver Front- End Chipset”, J. of Solid-State Circuits, vol. 35, no. 12, dec. 2000 [47] J.R. Long, ”A low-voltage 5.1-5.8-GHz Image-Reject Downconverter RF IC”, IEEE J. Solid-State Circuits, vol. 35, no. 9, Sept. 2000 [48] S.C. Cripps, “RF Power Amplifiers for Wireless Communications”, Artech House, Norwood, ISBN. 0-89006-989-1, 1999 [49] H. Ku, M.D. McKinley, J.S. Kenney, “Quantifying memory effects in RF power amplifiers”, IEEE Trans. on Microwave Theory and Techniques, vol. 50, no. 12, dec. 2002 [50] S.Y. Liao, “Microwave Circuit Analysis and Amplifier Design”, Prentice Hall, 1984 BIBLIOGRAPHY 167

[51] D. Z¨oschg, W. Wilhelm, T.F. Meister, J. Knapp, H.-D. Wohlmuth, K. Aufinger, M. Wurzer, J. B¨ock, H. Sch¨afer, A. Scholz, ”2dB noise figure, 10.5GHz LNA using SiGe bipolar technology”, Electronics Letters, vol. 35, no. 25, dec. 1999 [52] U. Erben, H. Schumacher, A. Sch¨uppen, H. Dietrich, J. Arndt, ”Ap- plication of a Production Ready SiGe HBT Process to 1.9, 5.7 and 10 GHz Low Noise MMIC’s”, Topical Meeting on Silicon Monolithic Integrated Circuits in RF Systems, p. 100, sept. 1998 [53] W. D¨urr, U. Erben, A. Sch¨uppen, H. Dietrich, H. Schumacher, ”Low-power low-noise active mixers for 5.7 and 11.2 GHz using commercially available SiGe HBT MMIX technology”, Electronics Letters, vol. 34, no. 21, oct. 1998 [54] M.W. Lynch, C.D. Holdenried, J.W. Haslett, ”A 17-GHz Direct Down-Conversion Mixer in a 47-GHz SiGe Process”, IEEE Radio Frequency Integrated Circuits Symposium, p. 461, june 2003 [55] K.B. Schad, H. Schumacher, A. Sch¨uppen, ”Low-power active mixer for Ku-Band application using SiGe HBT MMIC technology”, IEEE Radio Frequency Integrated Circuits Symposium, p. 263, june 2000 [56] E. S¨onmez, A. Trasser, K.B. Schad, P. Abele, H. Schumacher, ”A single-chip 24 GHz receiver front-end using a commercially avail- able SiGe HBT foundry process”, IEEE Radio Frequency Integrated Circuits Symposium, p. 159, june 2002 [57] IBM, ”Blue Logic BiCMOS 6HP Technology”, available online: www.chips.ibm.com [58] J.-F. Z¨urcher, ”The SSFIP, a Global Concept for High Performance Broadband Planar Antennas”, Electronics Letters, vol. 24, 1988 [59] J.-F. Z¨urcher, F. Gardiol, ”Broadband patch antennas”, Arctech House, 1995 [60] W.J. Getsinger, ”Measurement and Modeling of the Apparent Char- acteristic Impedance of Microstrip”, IEEE Trans. on Microwave Theory and Techniques, vol. 83, no. 8, aug. 1983 [61] M. Kossel, R. K¨ung, J. Hansen, W. B¨achtold, ”Circular Polarized Aperture-Coupled Patch Antennas for an RFID-System in the 2.4- GHz ISM Band”, Microwave Journal, vol. 42, no. 11, p. 20, 1999 [62] L. Steinemann, “Aperturgekoppelte Patchantenne”, Semesterarbeit IFH, SoSe 2001 [63] W.R. Deal, V. Radisic, Y. Qian, T. Itoh, ”Integrated-Antenna Push- Pull Power Amplifiers”, IEEE T. on Microwave Theory and Tech- niques, vol. 47, no. 8, Aug. 1999 168 BIBLIOGRAPHY

[64] H. An, B.K.J.C. Nauwelaers, A.R. Van De Capelle, R. G. Bosisio, ”A novel measurement technique for amplifier-type active antennas”, 1994 IEEE MTT-S Int. Microwave Symp. Dig. , vol. 3, p. 1473, May 1994 [65] D.M. Pozar, “The Active Element Pattern”, IEEE Trans. on An- tennas and Propagation, vol. 42, no. 8, aug. 1994 [66] D.F. Kelley, W.L. Stutzman, “Array Antenna Pattern Modeling Methods That Include Mutual Coupling Efffects”, IEEE Trans. on Antennas and Propagation, vol. 41, no. 12, dec. 1993 [67] H. Steyskal, J.S. Herd, “Mutual Coupling Compensation in Small Array Antennas”, IEEE Trans. on Antennas and Propagation, vol. 38, no. 12, dec. 1990 [68] D. Parker, D. C. Zimmermann, “Phased Arrays–Part 1: Theory and Architectures”, IEEE Trans. on Microwave Theory and Techniques, vol. 50, no. 3, march 2002 [69] H. Aumann, F.G. Willwerth, ”Phased Array Calibration Using Mea- sured Element Patterns”, IEEE Int. Symp. on Antennas and Prop- agation, vol. 2, p. 918, june 1995 [70] W.L. Stutzman, G.A. Thiele, Antenna Theory and Design, John Wiley & Sons, New York, 1981 [71] R.J. Marks II, ”Advanced Topics in Shannon Sampling and Inter- polation Theory”, Springer-Verlag New York, ISBN 0-387-97906-9 [72] L. Kuehnke, ”Phased Array Calibration Procedures Based on Mea- sured Element Patterns”, IEE Int. Conf. on Antennas and Propa- gation, p. 660, april 2001 [73] A. Kortke, ”A New Calibration Algorithm for Smart Antenna Ar- rays”, IEEE Vehicular Technology Conference, vol. 2, p. 1030, april 2003 [74] G.V. Tsoulos, M.A. Beach, ”Calibration and Linearity issues for an Adaptive Antenna System”, IEEE Vehicular Technology Confer- ence, vol. 1, p. 45, 1995 [75] C.M. Simmonds, M.A. Beach, ”Downlink Calibration Requirements for the TSUNAMI (II) Adaptive Antenna Testbed”, IEEE Int. Conf. on Personal, Indoor and Mobile Radio Communications, vol. 3, p. 1260, 1998 [76] K. Nishimori, K. Cho, Y. Takatori, T. Hori, ”Automatic Calibration Method Using Transmitting Signals of an Adaptive Array for TDD Systems”, IEEE Trans. on Vehicular Technology, vol. 50, no. 6, nov. 2001 BIBLIOGRAPHY 169

[77] Y. Takeushi, H. Hirayama, K. Fukino, T. Murayama, Y. Notsu, A. Hayashi, ”Auto Calibrated and Distributed Local Loop Config- uration of Array Antenna for CDMA Cellular Base Station”, IEE Int. Symp. on Spread-Spectrum Tech. & Appl., p. 666, sept. 2000 [78] K. Nishimori, K. Cho, Y. Takatori, T. Hori, ”Automatic Calibration Method of Adaptive Array Utilizing signal Circulation for Linear Array”, IEEE Vehicular Technology Conference 2001, vol. 3, 2001 [79] H.M. Aumann, A.J. Fenn, F.G. Willwerth, “Phased Array Antenna Calibration and Pattern Prediction Using Mutual Coupling Mea- surements”, IEEE Trans. on Antennas and Propagation, vol. 37, no. 7, july 1989 [80] C. Passmann, F. Hickel, T. Wixforth, ”Investigation of a calibration concept for optimum performance of adaptive antenna systems”, IEEE Vehicular Technology Conference, vol. 1, p. 577, may 1998 [81] K.R. Dandekar, H. Ling, G. Xu, ”Smart antenna array calibration procedure including amplitude and phase mismatch and mutual cou- pling effects”, IEEE Int. Conf. on Personal, Indoor and Wireless Communications, p. 293, dec. 2000 [82] M. Schreiner, H. Leier, W. Menzel, H.-P. Feldle, ”Architecture and interconnect technologies for a novel conformal active phased array radar module”, IEEE Int. Microwave Symposium, vol. 1, p. 567, june 2003 [83] J.D. Hanfling, O.J. Bedigian, ”Calibration and test of an active con- formal phased array”, Antennas and Propagation Society Int. Symp., vol. 3, p. 1142, may 1990

Curriculum vitae

Thomas M. Brauner born February 8th, 1973 in K¨oln, Germany

School education:

1983-1992 Gymnasium Herkenrath, Bergisch-Gladbach, Germany

Diploma in electrical engineering:

1993-1999 Student of electrical engineering at Aachen university of technology (RWTH Aachen), Germany 1996-1997 Student assistant at the Laboratory for High Frequency Technology, RWTH Aachen 1996 Internship at Deutsche Telekom AG, Germany 1997 Internship at ASCOM Systec AG, Switzerland 1999 Diploma thesis on the verification of a deterministic propagation model using channel sounding

Doctor of technical sciences:

1999-2004 Research assistant at the Microwave Electronics Laboratory (IFH) of the Swiss Federal Institute of Technology (ETH), Z¨urich, Switzerland 1999-2002 Teaching assistant at ETH 2004 Dissertation on the implementation of active antenna frontends for multiple antenna systems

List of publications

T.Brauner, R.Vogt, W.B¨achtold, ”A Differential Active Patch Antenna Element for Array Applications”, IEEE Microwave and Wireless Compo- nent Letters, vol. 13, no. 4, p. 161, April 2003

T.Brauner, R.Vogt, W.B¨achtold, ”5-6 GHz Monolithically Integrated Cal- ibratable Low-Noise Downconverter for Smart Antenna Arrays”, 2003 IEEE Radio Frequency Integrated Circuits Symposium (RFIC), p. 435, June 8-10, 2003, Philadelphia, PA, USA

T.Brauner, R.K¨ung, R.Vogt, W.B¨achtold, ”Noise in Smart Antennas for Mobile Communications”, 2003 IEEE Antennas and Propagation Sympo- sium (APS/URSI), vol. 4, p. 184, June 22-27, 2003, Columbus, OH, USA

T.Brauner, R.K¨ung, R.Vogt, W.B¨achtold, ”5-6 GHz Low-Noise Active Antenna Array for Multi-Dimensional Channel-Sounding”, 2003 Interna- tional Microwave and Optoelectronics Conference (IMOC 2003), vol. 1, p. 297, September 2003, Foz do Iguazu, Brazil

T.Brauner, R.Vogt, W.B¨achtold, ”A Versatile Calibration Method for Small Active Antenna Arrays”, European Microwave Conference 2003 (EuMC 2003), p. 797, October 2003, Munich, Germany

T.Brauner, R.Negra, R.Vogt, W.B¨achtold, ”PA Calibration in TDMA An- tenna Arrays”, IEEE Microwave and Wireless Component Letters, ac- cepted for publication

Acknowledgments

I would like to acknowledge the people who supported me during this work. First, I would like to thank my ”Doktorvater” Prof. Werner B¨achtold for giving me the opportunity to work in his research group and profit from a wealth of knowledge and experience. I am also grateful for the valuable inputs from the co-examiners Prof. Helmut B¨olcskei and Prof. Roland K¨ung. A special thank goes to all those people from the IFH, whose technical and organizational support was indispensable for the successful outcome of this work; Martin Lanz and Hansruedi Benedickter for bonding and measurements, Claudio Maccio and Stephen Wheeler for the mechanical constructions and Ray Ballisti, Federico Bonzanigo and Aldo Rossi for the computer support. I would like to thank Urs Lott for sharing his know- how and teaching me the first steps in integrated circuit design and Rolf Vogt, who saved me from an unlucky project. Furthermore, I am grateful to all people involved into the ”SANTRES” project for the successful cooperation, especially Marcel Wattinger from Elektrobit, Peter N¨uchter from Huber & Suhner and Luca Pergola from the field theory group. I especially appreciate all my actual and former colleagues who made this institute an unforgettable place to stay. Especially I want to mention those, who became close friends during this period: Franck, who made at least my rope reach the summit of Dom, Andrea who revealed for me the uncompromising philosophy of italian cooking and the secret ingredients and Esteban for his unconventional views on ordinary things and maybe the best climbing days. But also all the others, which are not listed here, should know that I truly enjoyed the open and inspirative atmosphere they all contributed to. A big thank goes to my parents and my family for their unconditional support. And last I want to thank Laura for always being close, even when there was some euclidian distance between us.